[2026-10-18 13:23:11,747] train_config.py:341 INFO Saved resolved config to /data/output/ckpts/config.yaml [2026-10-18 13:23:11,752] train_entry.py:129 INFO Resolved config: dataset: chat_template: null defer_tokenization: false eval_data_path: null eval_interval: 500 eval_micro_batch_size: null eval_prompt_key: conversations last_turn_loss_only: true length_group_size: 1 min_loss_tokens: 16 prompt_key: conversations renderer: kimi-k3 shuffle_dataset: true train_data_path: /data/corpus/v7_text_train.jsonl debug: enable_perf_metrics: true max_dump_steps: 5 memory_recorder: torch memory_snapshot_dir: . memory_snapshot_num_steps: null memory_snapshot_path: '' profile_dir_name: /tmp/torchspec_profiles profile_step_end: 0 profile_step_start: 0 profile_target: - train_overall record_memory_history: false save_debug_train_data: null use_pytorch_profiler: false decode: cuda_graph_max_bs: null max_new_tokens: 512 min_new_tokens: 2 stop_token_ids: null max_running_requests: null speculative_algorithm: null speculative_draft_model_path: null speculative_eagle_topk: null speculative_num_draft_tokens: null speculative_num_steps: null temperature: 1.0 top_k: -1 top_p: 1.0 weight_sync_enabled: false weight_sync_interval: 500 inference: aux_hidden_states_layers: - 1 - 29 - 57 inference_batch_size: 4 inference_buffer_threshold: 8 inference_engine_type: vllm inference_fetch_batch: 8 inference_num_gpus: 16 inference_num_gpus_per_engine: 4 inference_num_gpus_per_node: 4 inference_node_ips: - 10.0.3.58 - 10.0.3.59 - 10.0.3.60 - 10.0.3.63 inference_node_selectors: null last_hidden_states_prenorm: true max_sample_pool_size: 64 store_last_hidden_states: true offline: data_path: null num_engines: 1 sglang: tp_size: 8 pp_size: 1 nnodes: 1 mem_fraction_static: 0.8 enable_metrics: false enable_multimodal: false dist_init_addr: null dist_timeout: 60 init_timeout: 300 extra_args: {} vllm: tp_size: 4 pp_size: 1 nnodes: 1 mem_fraction_static: 0.85 enable_metrics: false enable_multimodal: false dist_init_addr: null dist_timeout: 1800 init_timeout: 3600 extra_args: max_model_len: 65536 max_num_batched_tokens: 32768 enable_chunked_prefill: true enforce_eager: true block_size: 64 kv_cache_dtype: fp8_ds_mla tokenizer_mode: kimi_k3 kv_cache_memory_bytes: 17179869184 async_scheduling: true load_format: fastsafetensors trtllm: tp_size: 8 pp_size: 1 nnodes: 1 mem_fraction_static: 0.8 init_timeout: 600 extra_args: {} tokenspeed: tp_size: 1 nnodes: 1 mem_fraction_static: 0.8 init_timeout: 600 extra_args: {} logging: report_to: none use_tensorboard: false use_wandb: false wandb_dir: null wandb_group: null wandb_host: null wandb_key: null wandb_mode: online wandb_project: train_kimi_k3_flash_dspark wandb_random_suffix: true wandb_run_id: null wandb_team: null model: draft_model_config: configs/draft_models/kimi_k3_flash_dspark_mla_v3h6.json initial_draft_model_path: null keep_initial_vocab_mapping: false embedding_key: language_model.model.embed_tokens.weight freeze_lm_head: false lm_head_key: language_model.lm_head.weight norm_key: language_model.model.norm.weight target_model_backend: vllm target_model_path: /data/models/moonshotai/Kimi-K3-Flash trust_remote_code: true mooncake: master_server_address: 10.0.3.61:51661 metadata_server: null metadata_port: 8653 protocol: rdma device_name: mlx5_bond_0 enable_gpu_direct: false hidden_dim: 7168 enable_hard_pin: true kv_lease_ttl_s: 180.0 get_retry_max_wait_seconds: 180.0 global_segment_size: 128GB local_buffer_size: 16GB max_seq_len: 65536 training: attention_backend: flex_attention colocate: false continual_training: false distributed_backend: nccl distributed_timeout_minutes: 60 draft_accumulation_steps: 16 fsdp_reduce_dtype: float32 fsdp_strategy: FULL_SHARD placement_strategy: custom training_node_ips: - 10.0.3.61 - 10.0.3.62 training_node_selectors: null compile_model: false sp_ring_size: 1 sp_ulysses_size: 1 gradient_checkpointing: true learning_rate: 5.0e-05 lk_eta: 3.0 load_path: null loss_type: forward_kl lr_decay_style: cosine lr_wsd_decay_ratio: 0.2 lr_wsd_decay_style: cosine lr_total_steps: null max_concurrent_batches: 1 max_grad_norm: 1.0 max_seq_length: 65536 min_lr: 0.0 weight_decay: 0.0 num_epochs: 2 num_train_steps: null micro_batch_size: 1 prefetch_depth: 2 save_interval: 250 save_per_epoch: true max_checkpoints: 3 seed: 42 train_backend: fsdp train_env_vars: '{}' train_with_decode: false training_num_gpus_per_node: 4 training_num_nodes: 2 ttt_length: 7 ploss_weights: null warmup_ratio: 0.04 wsd_decay_ratio: 0.2 wsd_decay_style: null dflash_block_size: 3 dflash_dpace_alpha: 0.5 dflash_loss_decay_gamma: 7.0 dflash_loss_objective: decay dflash_ce_loss_alpha: 1.0 dflash_l1_loss_alpha: 0.0 dflash_num_anchors: 512 dflash_num_target_layers: 5 dflash2_selector_loss_alpha: 1.0 dspark_num_anchors: 512 dspark_num_target_layers: 3 dspark_loss_decay_gamma: 4.0 dspark_ce_loss_alpha: 0.1 dspark_l1_loss_alpha: 0.9 dspark_confidence_head_alpha: 1.0 cache_dir: /data/output/cache cache_key: null model_download_dir: null output_dir: /data/output/ckpts 2026-10-18 13:23:11,754 INFO worker.py:1683 -- Using address 10.0.3.61:6390 set in the environment variable RAY_ADDRESS 2026-10-18 13:23:11,758 INFO worker.py:1833 -- Connecting to existing Ray cluster at address: 10.0.3.61:6390... 2026-10-18 13:23:11,778 INFO worker.py:2024 -- Connected to Ray cluster. [2026-10-18 13:23:11,914] auto.py:101 INFO Set draft model tie_word_embeddings to False [2026-10-18 13:23:11,919] placement_group.py:334 INFO Ray cluster has 24 GPUs (need 24) [2026-10-18 13:23:11,919] placement_group.py:382 INFO Creating custom unified placement group with 24 GPUs (8 training + 16 inference); training nodes=['10.0.3.61', '10.0.3.62'], inference nodes=['10.0.3.58', '10.0.3.59', '10.0.3.60', '10.0.3.63'] [2026-10-18 13:23:12,520] placement_group.py:294 INFO bundle 0, actual_bundle_index: 0, node: 10.0.3.61, gpu: 0 [2026-10-18 13:23:12,520] placement_group.py:294 INFO bundle 1, actual_bundle_index: 1, node: 10.0.3.61, gpu: 1 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 2, actual_bundle_index: 2, node: 10.0.3.61, gpu: 2 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 3, actual_bundle_index: 3, node: 10.0.3.61, gpu: 3 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 4, actual_bundle_index: 4, node: 10.0.3.62, gpu: 0 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 5, actual_bundle_index: 5, node: 10.0.3.62, gpu: 1 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 6, actual_bundle_index: 6, node: 10.0.3.62, gpu: 2 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 7, actual_bundle_index: 7, node: 10.0.3.62, gpu: 3 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 8, actual_bundle_index: 8, node: 10.0.3.58, gpu: 0 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 9, actual_bundle_index: 9, node: 10.0.3.58, gpu: 1 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 10, actual_bundle_index: 10, node: 10.0.3.58, gpu: 2 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 11, actual_bundle_index: 11, node: 10.0.3.58, gpu: 3 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 12, actual_bundle_index: 12, node: 10.0.3.59, gpu: 0 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 13, actual_bundle_index: 13, node: 10.0.3.59, gpu: 1 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 14, actual_bundle_index: 14, node: 10.0.3.59, gpu: 2 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 15, actual_bundle_index: 15, node: 10.0.3.59, gpu: 3 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 16, actual_bundle_index: 16, node: 10.0.3.60, gpu: 0 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 17, actual_bundle_index: 17, node: 10.0.3.60, gpu: 1 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 18, actual_bundle_index: 18, node: 10.0.3.60, gpu: 2 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 19, actual_bundle_index: 19, node: 10.0.3.60, gpu: 3 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 20, actual_bundle_index: 20, node: 10.0.3.63, gpu: 0 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 21, actual_bundle_index: 21, node: 10.0.3.63, gpu: 1 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 22, actual_bundle_index: 22, node: 10.0.3.63, gpu: 2 [2026-10-18 13:23:12,521] placement_group.py:294 INFO bundle 23, actual_bundle_index: 23, node: 10.0.3.63, gpu: 3 [2026-10-18 13:23:12,523] placement_group.py:403 INFO Placement (strategy=custom): training bundles=[0, 1, 2, 3, 4, 5, 6, 7], inference bundles=[8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23] [2026-10-18 13:23:12,524] ray_actor.py:49 INFO Pinning mooncake_master actor to node 10.0.3.61 (id=1eea48f9...) (AsyncTrainingController pid=8785) [2026-10-18 13:23:19,904] dataset.py:361 INFO Max sequence length allowed for training: 65536 (AsyncTrainingController pid=8785) [2026-10-18 13:23:19,920] dataset.py:426 INFO Tokenizing dataset (cache will be saved to /data/output/cache/tokenized_dataset/dc80602f18a10cd4329cb693c826db50.pt) (MooncakeMaster pid=8851) [2026-10-18 13:23:19,856] utils.py:116 INFO Starting mooncake master on port 51661 [2026-10-18 13:23:21,868] utils.py:316 INFO mooncake master actor started: {'master_addr': '10.0.3.61:51661', 'metadata_port': 8653} (MooncakeMaster pid=8851) [2026-10-18 13:23:21,866] utils.py:142 INFO mooncake master started (PID: 8915) (AsyncTrainingController pid=8785) [2026-10-18 13:23:28,932] dataset.py:439 INFO last_turn_loss_only=True: loss mask will only cover the last assistant turn (AsyncTrainingController pid=8785) Loading samples: 0%| | 0/1153209 [00:00 for this kind of AutoModel: AutoModelForCausalLM. (TrainerActor pid=1753, ip=10.0.3.62) Model type should be one of GPT2Config, AfmoeConfig, ApertusConfig, ArceeConfig, AriaTextConfig, AXK1Config, AXK2Config, BambaConfig, BartConfig, BertConfig, BertGenerationConfig, BigBirdConfig, BigBirdPegasusConfig, BioGptConfig, BitNetConfig, BlenderbotConfig, BlenderbotSmallConfig, BloomConfig, BltConfig, CamembertConfig, CodeGenConfig, CohereConfig, Cohere2Config, Cohere2MoeConfig, CpmAntConfig, CTRLConfig, CwmConfig, Data2VecTextConfig, DbrxConfig, DeepseekV2Config, DeepseekV3Config, DeepseekV32Config, DeepseekV4Config, DiffLlamaConfig, DogeConfig, Dots1Config, ElectraConfig, Emu3Config, ErnieConfig, Ernie4_5Config, Ernie4_5_MoeConfig, Exaone4Config, ExaoneMoeConfig, FalconConfig, FalconH1Config, FalconMambaConfig, FlexOlmoConfig, FuyuConfig, GemmaConfig, Gemma2Config, Gemma3Config, Gemma3TextConfig, Gemma3nConfig, Gemma3nTextConfig, Gemma4Config, Gemma4AssistantConfig, Gemma4TextConfig, Gemma4UnifiedConfig, Gemma4UnifiedAssistantConfig, Gemma4UnifiedTextConfig, GitConfig, GlmConfig, Glm4Config, Glm4MoeConfig, Glm4MoeLiteConfig, GlmMoeDsaConfig, GotOcr2Config, GPT2Config, GPTBigCodeConfig, GPTNeoConfig, GPTNeoXConfig, GPTNeoXJapaneseConfig, GptOssConfig, GPTJConfig, GraniteConfig, GraniteSWAConfig, GraniteMoeConfig, GraniteMoeSWAConfig, GraniteMoeHybridConfig, GraniteMoeSharedConfig, HeliumConfig, HrmTextConfig, HunYuanDenseV1Config, HunYuanMoEV1Config, HYV3Config, HyperCLOVAXConfig, InklingTextConfig, Jais2Config, JambaConfig, JetMoeConfig, LagunaConfig, Lfm2Config, Lfm2MoeConfig, LlamaConfig, Llama4Config, Llama4TextConfig, LongcatFlashConfig, MambaConfig, Mamba2Config, MarianConfig, MBartConfig, MegatronBertConfig, MellumConfig, MiMoV2FlashConfig, MiniCPM3Config, MiniMaxConfig, MiniMaxM2Config, MiniMaxM3VLTextConfig, MinistralConfig, Ministral3Config, MistralConfig, MixtralConfig, MllamaConfig, ModernBertDecoderConfig, MoshiConfig, MptConfig, MusicgenConfig, MusicgenMelodyConfig, MvpConfig, NanoChatConfig, NemotronConfig, NemotronHConfig, OlmoConfig, Olmo2Config, Olmo3Config, OlmoHybridConfig, OlmoeConfig, OpenAIGPTConfig, OPTConfig, PegasusConfig, PersimmonConfig, PhiConfig, Phi3Config, Phi4MultimodalConfig, PhimoeConfig, PLBartConfig, ProphetNetConfig, Qwen2Config, Qwen2MoeConfig, Qwen3Config, Qwen3_5Config, Qwen3_5MoeConfig, Qwen3_5MoeTextConfig, Qwen3_5TextConfig, Qwen3MoeConfig, Qwen3NextConfig, RecurrentGemmaConfig, ReformerConfig, RemBertConfig, RobertaConfig, RobertaPreLayerNormConfig, RoCBertConfig, RoFormerConfig, RwkvConfig, SeedOssConfig, SmolLM3Config, SolarOpenConfig, StableLmConfig, Starcoder2Config, TrOCRConfig, VaultGemmaConfig, WhisperConfig, XGLMConfig, XLMConfig, XLMRobertaConfig, XLMRobertaXLConfig, XLNetConfig, xLSTMConfig, XmodConfig, YoutuConfig, ZambaConfig, Zamba2Config, ZayaConfig. (TrainerActor pid=9310) [2026-10-18 13:27:32,948] dflash_trainer.py:228 INFO [Rank 0] DFlash model initialized with FSDP2 (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:26:53 [api_utils.py:286] non-default args: {'tokenizer_mode': 'kimi_k3', 'trust_remote_code': True, 'load_format': 'fastsafetensors', 'kv_cache_dtype': 'fp8_ds_mla', 'max_model_len': 65536, 'distributed_executor_backend': 'mp', 'master_port': 10004, 'distributed_timeout_seconds': 3600, 'tensor_parallel_size': 4, 'block_size': 64, 'enable_prefix_caching': False, 'gpu_memory_utilization': 0.85, 'kv_cache_memory_bytes': 17179869184, 'max_num_batched_tokens': 32768, 'max_num_seqs': 4, 'disable_log_stats': True, 'enforce_eager': True, 'disable_custom_all_reduce': True, 'enable_chunked_prefill': True, 'speculative_config': {'method': 'extract_hidden_states', 'num_speculative_tokens': 1, 'draft_model_config': {'hf_config': {'eagle_aux_hidden_state_layer_ids': [2, 30, 58, 61]}}}, 'compilation_config': {'mode': None, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': [], 'ir_enable_torch_wrap': None, 'splitting_ops': None, 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': None, 'compile_ranges_endpoints': None, 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': None, 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': None, 'pass_config': {}, 'max_cudagraph_capture_size': None, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': None, 'static_all_moe_layers': []}, 'kv_transfer_config': KVTransferConfig(kv_connector='MooncakeHiddenStatesConnector', engine_id='e27b7385-b32e-456a-9455-0e04e3f9773a', kv_buffer_device='cuda', kv_buffer_size=1000000000.0, kv_role='kv_producer', kv_rank=None, kv_parallel_size=1, kv_ip='127.0.0.1', kv_port=14579, kv_connector_extra_config={}, kv_connector_module_path='torchspec.inference.engine.mooncake_hidden_states_connector', enable_permute_local_kv=False, kv_load_failure_policy='fail'), 'async_scheduling': True, 'model': '/data/models/moonshotai/Kimi-K3-Flash'} (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:26:53 [envs.py:2249] Unknown vLLM environment variable detected: VLLM_USE_FLASHINFER_MOE_FP8 (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:05 [model.py:684] Resolved architecture: KimiK3FlashForConditionalGeneration (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:05 [model.py:2017] Using max model len 65536 (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:27:05 [model.py:981] Model does not support mm_device_do_normalize, forcing mm_device_do_normalize = False. (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:26:54 [api_utils.py:286] non-default args: {'tokenizer_mode': 'kimi_k3', 'trust_remote_code': True, 'load_format': 'fastsafetensors', 'kv_cache_dtype': 'fp8_ds_mla', 'max_model_len': 65536, 'distributed_executor_backend': 'mp', 'master_port': 10009, 'distributed_timeout_seconds': 3600, 'tensor_parallel_size': 4, 'block_size': 64, 'enable_prefix_caching': False, 'gpu_memory_utilization': 0.85, 'kv_cache_memory_bytes': 17179869184, 'max_num_batched_tokens': 32768, 'max_num_seqs': 4, 'disable_log_stats': True, 'enforce_eager': True, 'disable_custom_all_reduce': True, 'enable_chunked_prefill': True, 'speculative_config': {'method': 'extract_hidden_states', 'num_speculative_tokens': 1, 'draft_model_config': {'hf_config': {'eagle_aux_hidden_state_layer_ids': [2, 30, 58, 61]}}}, 'compilation_config': {'mode': None, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': [], 'ir_enable_torch_wrap': None, 'splitting_ops': None, 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': None, 'compile_ranges_endpoints': None, 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': None, 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': None, 'pass_config': {}, 'max_cudagraph_capture_size': None, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': None, 'static_all_moe_layers': []}, 'kv_transfer_config': KVTransferConfig(kv_connector='MooncakeHiddenStatesConnector', engine_id='becd3e77-e6f1-49cc-bd53-d11f79c734e8', kv_buffer_device='cuda', kv_buffer_size=1000000000.0, kv_role='kv_producer', kv_rank=None, kv_parallel_size=1, kv_ip='127.0.0.1', kv_port=14579, kv_connector_extra_config={}, kv_connector_module_path='torchspec.inference.engine.mooncake_hidden_states_connector', enable_permute_local_kv=False, kv_load_failure_policy='fail'), 'async_scheduling': True, 'model': '/data/models/moonshotai/Kimi-K3-Flash'} [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-18 13:26:54 [envs.py:2249] Unknown vLLM environment variable detected: VLLM_USE_FLASHINFER_MOE_FP8 [repeated 3x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:06 [import_utils.py:478] DeepEP v2 requires NCCL >= 2.30.4 but found 2.29.7. deepep_v2 backend will not be available. (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:27:07 [registry.py:259] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead. (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:08 [cache.py:335] Using fp8_ds_mla data type to store kv cache. It reduces the GPU memory footprint and boosts the performance. Meanwhile, it may cause accuracy drop without a proper scaling factor (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:08 [arg_utils.py:2068] Using ray runtime env (env vars redacted): {'env_vars': {'GLOO_SOCKET_IFNAME': '***', 'NCCL_DEBUG': '***', 'RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES': '***', 'RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES': '***', 'RAY_EXPERIMENTAL_NOSET_HABANA_VISIBLE_MODULES': '***', 'RAY_EXPERIMENTAL_NOSET_NEURON_RT_VISIBLE_CORES': '***', 'RAY_EXPERIMENTAL_NOSET_ONEAPI_DEVICE_SELECTOR': '***', 'RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES': '***', 'RAY_EXPERIMENTAL_NOSET_TPU_VISIBLE_CHIPS': '***', 'TORCHSPEC_LOG_LEVEL': '***', 'TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC': '***', 'TP_SOCKET_IFNAME': '***', 'VLLM_USE_V2_MODEL_RUNNER': '***'}} (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:08 [model.py:684] Resolved architecture: KimiK3FlashForConditionalGeneration [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:08 [model.py:2017] Using max model len 65536 [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-18 13:27:08 [model.py:981] Model does not support mm_device_do_normalize, forcing mm_device_do_normalize = False. [repeated 3x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:15 [scheduler.py:242] Chunked prefill is enabled with max_num_batched_tokens=32768. (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:27:15 [vllm.py:1388] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:27:15 [vllm.py:1423] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:15 [kernel.py:310] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:15 [vllm.py:1602] Cudagraph is disabled under eager mode (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:08 [import_utils.py:478] DeepEP v2 requires NCCL >= 2.30.4 but found 2.29.7. deepep_v2 backend will not be available. [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-18 13:27:10 [registry.py:259] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead. [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:11 [cache.py:335] Using fp8_ds_mla data type to store kv cache. It reduces the GPU memory footprint and boosts the performance. Meanwhile, it may cause accuracy drop without a proper scaling factor [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:11 [arg_utils.py:2068] Using ray runtime env (env vars redacted): {'env_vars': {'GLOO_SOCKET_IFNAME': '***', 'NCCL_DEBUG': '***', 'RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES': '***', 'RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES': '***', 'RAY_EXPERIMENTAL_NOSET_HABANA_VISIBLE_MODULES': '***', 'RAY_EXPERIMENTAL_NOSET_NEURON_RT_VISIBLE_CORES': '***', 'RAY_EXPERIMENTAL_NOSET_ONEAPI_DEVICE_SELECTOR': '***', 'RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES': '***', 'RAY_EXPERIMENTAL_NOSET_TPU_VISIBLE_CHIPS': '***', 'TORCHSPEC_LOG_LEVEL': '***', 'TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC': '***', 'TP_SOCKET_IFNAME': '***', 'VLLM_USE_V2_MODEL_RUNNER': '***'}} [repeated 3x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:15 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant, allreduce_rms (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:27:16 [system_utils.py:157] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned; CUDA is initialized (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:20 [scheduler.py:242] Chunked prefill is enabled with max_num_batched_tokens=32768. [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-18 13:27:20 [vllm.py:1388] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-18 13:27:20 [vllm.py:1423] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:20 [kernel.py:310] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:20 [vllm.py:1602] Cudagraph is disabled under eager mode [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:20 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant, allreduce_rms [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-18 13:27:21 [system_utils.py:157] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned; CUDA is initialized [repeated 3x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) INFO 10-18 13:27:23 [import_utils.py:478] DeepEP v2 requires NCCL >= 2.30.4 but found 2.29.7. deepep_v2 backend will not be available. (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Bootstrap: Using eth0:10.0.3.61<0> (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO cudaDriverVersion 13000 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO NCCL version 2.29.7+cuda13.2 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO NCCL git version stable b81d6a5a3 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Comm config Blocking set to 1 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO NET/Plugin: Could not find: libnccl-net.so (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO NCCL_IB_HCA set to mlx5_bond_0 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO NET/IB : Using [0]mlx5_bond_0:1/RoCE [RO]; OOB eth0:10.0.3.61<0> (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Initialized NET plugin IB (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Assigned NET plugin IB to comm (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO GIN/Plugin: Could not find: libnccl-gin.so (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Using network IB (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO DMA-BUF is available on GPU device 0 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO [Rank 0] ncclCommInitRankConfig comm 0x4af1baa0 rank 0 nranks 8 cudaDev 0 nvmlDev 0 busId 806000 commId 0x5b058ae8201a6de4 - Init START (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO RAS client listening socket at 127.0.0.1<28028> (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Bootstrap timings total 0.091208 (create 0.000064, send 0.000239, recv 0.088039, ring 0.000559, delay 0.000002) (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0. (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO MNNVL busId 0x806000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 0 is 0-71. (GPU affinity = 0-71 ; CPU affinity = 0-143). (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO NVLS multicast support is available on dev 0 (NVLS_NCHANNELS 32) (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO comm 0x4af1baa0 rank 0 nRanks 8 nNodes 2 localRanks 4 localRank 0 MNNVL 0 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO NVLS Head 0: 0 4 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 00/04 : 0 3 2 1 4 7 6 5 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 01/04 : 0 3 2 1 4 7 6 5 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 02/04 : 0 3 2 1 4 7 6 5 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 03/04 : 0 3 2 1 4 7 6 5 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Trees [0] 1/4/-1->0->-1 [1] 1/4/-1->0->-1 [2] 1/-1/-1->0->4 [3] 1/-1/-1->0->4 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO P2P Chunksize set to 262144 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 (TrainerActor pid=9310) host-10-0-3-61:9310:9579 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 40 (TrainerActor pid=9310) host-10-0-3-61:9310:9577 [0] NCCL INFO [Proxy Service] Device 0 CPU core 62 (TrainerActor pid=9310) host-10-0-3-61:9310:9581 [0] NCCL INFO [Proxy Progress] Device 0 CPU core 71 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 00/0 : 5[1] -> 0[0] [receive] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 01/0 : 5[1] -> 0[0] [receive] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 02/0 : 5[1] -> 0[0] [receive] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 03/0 : 5[1] -> 0[0] [receive] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 00/0 : 0[0] -> 3[3] via P2P/IPC (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 01/0 : 0[0] -> 3[3] via P2P/IPC (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 02/0 : 0[0] -> 3[3] via P2P/IPC (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 03/0 : 0[0] -> 3[3] via P2P/IPC (TrainerActor pid=9310) host-10-0-3-61:9310:9577 [0] NCCL INFO NCCL_IB_GID_INDEX set by environment to 3. (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 00/0 : 0[0] -> 1[1] via P2P/IPC (TrainerActor pid=9310) host-10-0- (TrainerActor pid=1751, ip=10.0.3.62) host-10-0-3-62:1751 (TrainerActor pid=1752, ip=10.0.3.62) host-10-0-3-62:1752:1752 [1] NCCL INFO Channel 00/0 : 5[1] -> 0[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=1752, ip=10.0.3.62) host-10-0-3-62:1752:1752 [1] NCCL INFO Channel 01/0 : 5[1] -> 0[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=1752, ip=10.0.3.62) host-10-0-3-62:1752:1752 [1] NCCL INFO Channel 02/0 : 5[1] -> 0[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=1752, ip=10.0.3.62) host-10-0-3-62:1752:1752 [1] NCCL INFO Channel 03/0 : 5[1] -> 0[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=1752, ip=10.0.3.62) host-10-0-3-62:1752:1752 [1] NCCL INFO Connected all trees (TrainerActor pid=1752, ip=10.0.3.62) host-10-0-3-62:1752:1752 [1] NCCL INFO NVLS comm 0x1c934b80 headRank (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO NVLS comm 0x4af1baa0 headRank 0 nHeads 1 nvlsRanks 4 buffSize 1048576 nvlsPerRankSize 67108864 nvlsTotalSize 67108864 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 05/0 : 0[0] -> 4[0] [send] via N (TrainerActor pid=9371) host-10-0-3-61:9371:9371 [2] NCCL INFO Connected NVLS tree (TrainerActor pid=9371) host-10-0-3-61:9371:9371 [2] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so (TrainerActor pid=9371) host-10-0-3-61:9371:9371 [2] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512 (TrainerActor pid=9371) host-10-0-3-61:9371:9371 [2] NCCL INFO 4 coll channels, 4 collnet chann (TrainerActor pid=9368) host-10-0- (TrainerActor pid=1751, ip=10.0.3.62) host-10-0-3-62:1751:1751 [0] NCCL INFO Channel (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:1754 [2] NCCL INFO 4 coll channels, 4 collnet channels, 32 nvls channels, 4 p2p channels, 4 p2p channels per pe (VllmEngine pid=1867, ip=10.0.3.60) (EngineCore pid=2062) INFO 10-18 13:27:24 [core.py:123] Initializing a V1 LLM engine (v0.26.1rc1.dev1102+ge9d1398d9) with config: model='/data/models/moonshotai/Kimi-K3-Flash', speculative_config=SpeculativeConfig(method='extract_hidden_states', model=None, num_spec_tokens=1), tokenizer='/data/models/moonshotai/Kimi-K3-Flash', skip_tokenizer_init=False, tokenizer_mode=kimi_k3, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=65536, download_dir=None, load_format=fastsafetensors, tensor_parallel_size=4, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=True, quantization=compressed-tensors, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=fp8_ds_mla, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, per_request_spec_decode_metrics='none', kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=/data/models/moonshotai/Kimi-K3-Flash, enable_prefix_caching=False, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [4681, 32768], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': True, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') (VllmEngine pid=1867, ip=10.0.3.60) (EngineCore pid=2062) INFO 10-18 13:27:24 [multiproc_executor.py:153] DP group leader: node_rank=0, node_rank_within_dp=0, master_addr=127.0.0.1, mq_connect_ip=10.0.3.60 (local), world_size=4, local_world_size=4 (TrainerActor pid=9370) host-10-0-3-61:9370:9370 [3] NCCL INFO 4 coll channels, 4 collnet channels, 32 nvls channels, 4 p2p channels, 4 p2p channels per peer (TrainerActor pid=9370) host-10-0-3-61:9370:9370 [3] NCCL INFO Symmetric memory is not supported. cuMemEnable 0, globalGinSupport 1, globalNicFused 0 cuMemGdrSupport 1 (TrainerActor pid=9370) host-10-0-3-61:9370:9370 [3] NCCL INFO ncclCommInitRankConfig comm 0x24029bd0 rank 3 nranks 8 cudaDev 3 nv (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO ncclCommInitRankConfig comm 0x1fe19a20 rank 7 nranks 8 cudaDev 3 nvmlDev 3 busId 1906000 commId 0x5b058ae8201a6de4 - Init COMPL (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:29 [import_utils.py:478] DeepEP v2 requires NCCL >= 2.30.4 but found 2.29.7. deepep_v2 backend will not be available. [repeated 3x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Bootstrap: Using eth0:10.0.3.62<0> [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO cudaDriverVersion 13000 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO NCCL version 2.29.7+cuda13.2 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO NCCL git version stable b81d6a5a3 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Comm config Blocking set to 1 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO NET/Plugin: Could not find: libnccl-net.so [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO NCCL_IB_HCA set to mlx5_bond_0 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO NET/IB : Using [0]mlx5_bond_0:1/RoCE [RO]; OOB eth0:10.0.3.62<0> [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Initialized NET plugin IB [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Assigned NET plugin IB to comm [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO GIN/Plugin: Could not find: libnccl-gin.so [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Using network IB [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO DMA-BUF is available on GPU device 3 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO [Rank 7] ncclCommInitRankConfig comm 0x1fe19a20 rank 7 nranks 8 cudaDev 3 nvmlDev 3 busId 1906000 commId 0x5b058ae8201a6de4 - Init START [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO RAS client listening socket at 127.0.0.1<28028> [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Bootstrap timings total 0.783088 (create 0.000038, send 0.000413, recv 0.692341, ring 0.088747, delay 0.000000) [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0. [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO MNNVL busId 0x1906000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 3 is 72-143. (GPU affinity = 72-143 ; CPU affinity = 0-143). [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO NVLS multicast support is available on dev 3 (NVLS_NCHANNELS 32) [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO comm 0x1fe19a20 rank 7 nRanks 8 nNodes 2 localRanks 4 localRank 3 MNNVL 0 [repeated 7x across cluster] (TrainerActor pid=9370) host-10-0-3-61:9370:9370 [3] NCCL INFO NVLS Head 0: 0 4 [repeated 3x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Trees [0] -1/-1/-1->7->6 [1] -1/-1/-1->7->6 [2] -1/-1/-1->7->6 [3] -1/-1/-1->7->6 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO P2P Chunksize set to 262144 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:2101 [0] NCCL INFO [Proxy Service UDS] Device 3 CPU core 102 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:2099 [0] NCCL INFO [Proxy Service] Device 3 CPU core 94 [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:2108 [0] NCCL INFO [Proxy Progress] Device 3 CPU core 96 [repeated 7x across cluster] (TrainerActor pid=1751, ip=10.0.3.62) host-10-0-3-62:1751:1751 [0] NCCL INFO Channel 31/0 : 0[0] -> 4[0] [receive] via NET/IB/0/GDRDMA [repeated 68x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Channel 03/0 : 7[3] -> 6[2] via P2P/IPC [repeated 51x across cluster] (TrainerActor pid=9368) host-10-0-3-61:9368:9576 [1] NCCL INFO NCCL_IB_GID_INDEX set by environment to 3. [repeated 3x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 [repeated 7x across cluster] (TrainerActor pid=1751, ip=10.0.3.62) host-10-0-3-62:1751:1751 [0] NCCL INFO Channel 08/0 : 4[0] -> 0[0] [send] via NET/IB/0/GDRDMA [repeated 18x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Connected all trees [repeated 7x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO NVLS comm 0x1fe19a20 headRank -1 nHeads 1 nvlsRanks 4 buffSize 1048576 nvlsPerRankSize 67108864 nvlsTotalSize 67108864 [repeated 5x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Connected NVLS tree [repeated 3x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so [repeated 3x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512 [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (EngineCore pid=1946) INFO 10-18 13:27:32 [core.py:123] Initializing a V1 LLM engine (v0.26.1rc1.dev1102+ge9d1398d9) with config: model='/data/models/moonshotai/Kimi-K3-Flash', speculative_config=SpeculativeConfig(method='extract_hidden_states', model=None, num_spec_tokens=1), tokenizer='/data/models/moonshotai/Kimi-K3-Flash', skip_tokenizer_init=False, tokenizer_mode=kimi_k3, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=65536, download_dir=None, load_format=fastsafetensors, tensor_parallel_size=4, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=True, quantization=compressed-tensors, quantization_config=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=fp8_ds_mla, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, per_request_spec_decode_metrics='none', kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False, jit_monitor_mode='warn', jit_monitor_verbose=False), seed=0, served_model_name=/data/models/moonshotai/Kimi-K3-Flash, enable_prefix_caching=False, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'ir_enable_torch_wrap': False, 'splitting_ops': [], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [4681, 32768], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': True, 'fuse_act_quant': True, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': True, 'enable_qk_norm_rope_fusion': False, 'fuse_rope_kvcache_cat_mla': False, 'fuse_act_padding': False, 'fuse_qk_norm_rope_kvcache': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']), enable_flashinfer_autotune=True, enable_cutedsl_warmup=True, enable_jit_warmup=True, enable_bf16x3_router_gemm=False, moe_backend='auto', linear_backend='auto') [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (EngineCore pid=1946) INFO 10-18 13:27:32 [multiproc_executor.py:153] DP group leader: node_rank=0, node_rank_within_dp=0, master_addr=127.0.0.1, mq_connect_ip=10.0.3.63 (local), world_size=4, local_world_size=4 [repeated 3x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO 4 coll channels, 4 collnet channels, 32 nvls channels, 4 p2p channels, 4 p2p channels per peer (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:1753 [3] NCCL INFO Symmetric memory is not supported. cuMemEnable 0, globalGinSupport 1, globalNicFused 0 cuMemGdrSupport 1 (VllmEngine pid=1809, ip=10.0.3.59) INFO 10-18 13:27:35 [import_utils.py:478] DeepEP v2 requires NCCL >= 2.30.4 but found 2.29.7. deepep_v2 backend will not be available. [repeated 12x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:27:36 [registry.py:259] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead. (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:27:36 [registry.py:259] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead. (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:27:36 [registry.py:259] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead. (VllmEngine pid=1867, ip=10.0.3.60) WARNING 10-18 13:27:36 [registry.py:259] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead. (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) INFO 10-18 13:27:36 [parallel_state.py:1638] world_size=4 rank=0 local_rank=0 distributed_init_method=file:///tmp/vllm_dist_7a8074f046cd4280b55e0068d077bb35 backend=nccl (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2080) INFO 10-18 13:27:36 [parallel_state.py:1638] world_size=4 rank=3 local_rank=3 distributed_init_method=file:///tmp/vllm_dist_7a8074f046cd4280b55e0068d077bb35 backend=nccl (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2078) INFO 10-18 13:27:36 [parallel_state.py:1638] world_size=4 rank=1 local_rank=1 distributed_init_method=file:///tmp/vllm_dist_7a8074f046cd4280b55e0068d077bb35 backend=nccl (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2079) INFO 10-18 13:27:36 [parallel_state.py:1638] world_size=4 rank=2 local_rank=2 distributed_init_method=file:///tmp/vllm_dist_7a8074f046cd4280b55e0068d077bb35 backend=nccl (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) INFO 10-18 13:27:36 [pynccl.py:113] vLLM is using nccl==2.29.7 (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:27:39 [import_utils.py:478] DeepEP v2 requires NCCL >= 2.30.4 but found 2.29.7. deepep_v2 backend will not be available. [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-18 13:27:43 [registry.py:259] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead. [repeated 11x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker pid=2020) INFO 10-18 13:27:39 [parallel_state.py:1638] world_size=4 rank=1 local_rank=1 distributed_init_method=file:///tmp/vllm_dist_6af072270b0147fe80aa8e4f0667f24d backend=nccl [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker pid=2019) INFO 10-18 13:27:39 [pynccl.py:113] vLLM is using nccl==2.29.7 [repeated 2x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Bootstrap: Using eth0:10.0.3.60<0> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO cudaDriverVersion 13000 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO NCCL version 2.29.7+cuda13.2 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO NCCL git version stable b81d6a5a3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO NET/Plugin: Could not find: libnccl-net.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO NCCL_IB_HCA set to mlx5_bond_0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO NET/IB : Using [0]mlx5_bond_0:1/RoCE [RO]; OOB eth0:10.0.3.60<0> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Initialized NET plugin IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Assigned NET plugin IB to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO GIN/Plugin: Could not find: libnccl-gin.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Using network IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO DMA-BUF is available on GPU device 0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO [Rank 0] ncclCommInitRank comm 0x4f938130 rank 0 nranks 4 cudaDev 0 nvmlDev 0 busId 806000 commId 0x5db36afb32ef643b - Init START (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO RAS client listening socket at 127.0.0.1<28028> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Bootstrap timings total 0.016923 (create 0.000047, send 0.000242, recv 0.011568, ring 0.000099, delay 0.000001) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO MNNVL busId 0x806000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 0 is 0-71. (GPU affinity = 0-71 ; CPU affinity = 0-143). (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO NVLS multicast support is available on dev 0 (NVLS_NCHANNELS 24) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO comm 0x4f938130 rank 0 nRanks 4 nNodes 1 localRanks 4 localRank 0 MNNVL 0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 00/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 01/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 02/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 03/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 04/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 05/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 06/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 07/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 08/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 09/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 10/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 11/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 12/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 13/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 14/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 15/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 16/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 17/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 18/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 19/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 20/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 21/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 22/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 23/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 24/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 25/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 26/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 27/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 28/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 29/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 30/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 31/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Trees [0] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-18 13:27:43 [registry.py:259] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead. (VllmEngine pid=1751, ip=10.0.3.63) (Worker pid=1962) INFO 10-18 13:27:45 [parallel_state.py:1638] world_size=4 rank=1 local_rank=1 distributed_init_method=file:///tmp/vllm_dist_0fcd78459745454e9e8798148c136a4b backend=nccl [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker pid=1961) INFO 10-18 13:27:45 [pynccl.py:113] vLLM is using nccl==2.29.7 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO cudaDriverVersion 13000 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Bootstrap: Using eth0:10.0.3.60<0> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO NCCL version 2.29.7+cuda13.2 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO NCCL git version stable b81d6a5a3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO NET/Plugin: Could not find: libnccl-net.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO NCCL_IB_HCA set to mlx5_bond_0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO NET/IB : Using [0]mlx5_bond_0:1/RoCE [RO]; OOB eth0:10.0.3.60<0> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Initialized NET plugin IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Assigned NET plugin IB to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO GIN/Plugin: Could not find: libnccl-gin.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Using network IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO DMA-BUF is available on GPU device 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO [Rank 3] ncclCommInitRank comm 0x25bd0800 rank 3 nranks 4 cudaDev 3 nvmlDev 3 busId 1906000 commId 0x5db36afb32ef643b - Init START (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO RAS client listening socket at 127.0.0.1<28028> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Bootstrap timings total 0.016484 (create 0.000033, send 0.000210, recv 0.000204, ring 0.000409, delay 0.000000) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO MNNVL busId 0x1906000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 3 is 72-143. (GPU affinity = 72-143 ; CPU affinity = 0-143). (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO NVLS multicast support is available on dev 3 (NVLS_NCHANNELS 24) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO comm 0x25bd0800 rank 3 nRanks 4 nNodes 1 localRanks 4 localRank 3 MNNVL 0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Trees [0] -1/-1/-1->3->2 [1] -1/-1/-1->3->2 [2] -1/-1/-1->3->2 [3] -1/-1/-1->3->2 [4] -1/-1/-1->3->2 [5] -1/-1/-1->3->2 [6] -1/-1/-1->3->2 [7] -1/-1/-1->3->2 [8] -1/-1/-1->3->2 [9] -1/-1/-1->3->2 [10] 0/-1/-1->3->2 [11] 0/-1/-1->3->2 [12] 0/-1/-1->3->2 [13] 0/-1/-1->3->2 [14] 0/-1/-1->3->-1 [15] 0/-1/-1->3->-1 [16] -1/-1/-1->3->2 [17] -1/-1/-1->3->2 [18] -1/-1/-1->3->2 [19] -1/-1/-1->3->2 [20] -1/-1/-1->3->2 [21] -1/-1/-1->3->2 [22] -1/-1/-1->3->2 [23] -1/-1/-1->3->2 [24] -1/-1/-1->3->2 [25] -1/-1/-1->3->2 [26] 0/-1/-1->3->2 [27] 0/-1/-1->3->2 [28] 0/-1/-1->3->2 [29] 0/-1/-1->3->2 [30] 0/-1/-1->3->-1 [31] 0/-1/-1->3->-1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO P2P Chunksize set to 524288 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2303 [0] NCCL INFO [Proxy Service] Device 3 CPU core 110 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2305 [0] NCCL INFO [Proxy Service UDS] Device 3 CPU core 143 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO ncclCommInitRank comm 0x25bd0800 rank 3 nranks 4 cudaDev 3 nvmlDev 3 busId 1906000 commId 0x5db36afb32ef643b - Init COMPLETE (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Init timings - ncclCommInitRank: rank 3 nranks 4 total 19.47 (kernels 18.23, alloc 0.09, bootstrap 0.02, allgathers 0.00, topo 0.37, graphs 0.01, connections 0.74, rest 0.01) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 00/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 01/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO cudaDriverVersion 13000 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Bootstrap: Using eth0:10.0.3.60<0> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO NCCL version 2.29.7+cuda13.2 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO NCCL git version stable b81d6a5a3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO NET/Plugin: Could not find: libnccl-net.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO NCCL_IB_HCA set to mlx5_bond_0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO NET/IB : Using [0]mlx5_bond_0:1/RoCE [RO]; OOB eth0:10.0.3.60<0> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Initialized NET plugin IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Assigned NET plugin IB to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO GIN/Plugin: Could not find: libnccl-gin.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Using network IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO DMA-BUF is available on GPU device 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO [Rank 1] ncclCommInitRank comm 0x39140c60 rank 1 nranks 4 cudaDev 1 nvmlDev 1 busId 906000 commId 0x5db36afb32ef643b - Init START (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO RAS client listening socket at 127.0.0.1<28028> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Bootstrap timings total 0.005431 (create 0.000032, send 0.000173, recv 0.000225, ring 0.000462, delay 0.000001) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO MNNVL busId 0x906000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 1 is 0-71. (GPU affinity = 0-71 ; CPU affinity = 0-143). (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO NVLS multicast support is available on dev 1 (NVLS_NCHANNELS 24) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO comm 0x39140c60 rank 1 nRanks 4 nNodes 1 localRanks 4 localRank 1 MNNVL 0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Trees [0] 2/-1/-1->1->0 [1] 2/-1/-1->1->0 [2] 2/-1/-1->1->0 [3] 2/-1/-1->1->0 [4] 2/-1/-1->1->0 [5] 2/-1/-1->1->0 [6] 2/-1/-1->1->0 [7] 2/-1/-1->1->0 [8] 2/-1/-1->1->0 [9] 2/-1/-1->1->0 [10] 2/-1/-1->1->-1 [11] 2/-1/-1->1->-1 [12] -1/-1/-1->1->0 [13] -1/-1/-1->1->0 [14] 2/-1/-1->1->0 [15] 2/-1/-1->1->0 [16] 2/-1/-1->1->0 [17] 2/-1/-1->1->0 [18] 2/-1/-1->1->0 [19] 2/-1/-1->1->0 [20] 2/-1/-1->1->0 [21] 2/-1/-1->1->0 [22] 2/-1/-1->1->0 [23] 2/-1/-1->1->0 [24] 2/-1/-1->1->0 [25] 2/-1/-1->1->0 [26] 2/-1/-1->1->-1 [27] 2/-1/-1->1->-1 [28] -1/-1/-1->1->0 [29] -1/-1/-1->1->0 [30] 2/-1/-1->1->0 [31] 2/-1/-1->1->0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO P2P Chunksize set to 524288 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2307 [0] NCCL INFO [Proxy Service] Device 1 CPU core 34 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2309 [0] NCCL INFO [Proxy Service UDS] Device 1 CPU core 35 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO ncclCommInitRank comm 0x39140c60 rank 1 nranks 4 cudaDev 1 nvmlDev 1 busId 906000 commId 0x5db36afb32ef643b - Init COMPLETE (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Init timings - ncclCommInitRank: rank 1 nranks 4 total 19.47 (kernels 18.25, alloc 0.08, bootstrap 0.01, allgathers 0.00, topo 0.37, graphs 0.01, connections 0.74, rest 0.01) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 00/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 01/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 host-10-0-3-60:2079:2079 [2] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO cudaDriverVersion 13000 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Bootstrap: Using eth0:10.0.3.60<0> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO NCCL version 2.29.7+cuda13.2 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO NCCL git version stable b81d6a5a3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO NET/Plugin: Could not find: libnccl-net.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO NCCL_IB_HCA set to mlx5_bond_0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO NET/IB : Using [0]mlx5_bond_0:1/RoCE [RO]; OOB eth0:10.0.3.60<0> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Initialized NET plugin IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Assigned NET plugin IB to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO GIN/Plugin: Could not find: libnccl-gin.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Using network IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO DMA-BUF is available on GPU device 2 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO [Rank 2] ncclCommInitRank comm 0x420297f0 rank 2 nranks 4 cudaDev 2 nvmlDev 2 busId 1806000 commId 0x5db36afb32ef643b - Init START (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO RAS client listening socket at 127.0.0.1<28028> (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Bootstrap timings total 0.006361 (create 0.000041, send 0.000181, recv 0.000185, ring 0.000092, delay 0.000000) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO MNNVL busId 0x1806000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa(VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2080) Calling super().encode with {'truncation': True, 'max_length': 65537, 'add_special_tokens': True} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2079) Calling super().encode with {'truncation': True, 'max_length': 65537, 'add_special_tokens': True} (TrainerActor pid=9371) Final-norm architecture extraction failed; using a config-driven RMSNorm fallback (hidden_size=7168, eps=1e-05): Unrecognized configuration class for this kind of AutoModel: AutoModelForCausalLM. [repeated 7x across cluster] (TrainerActor pid=9371) Model type should be one of GPT2Config, AfmoeConfig, ApertusConfig, ArceeConfig, AriaTextConfig, AXK1Config, AXK2Config, BambaConfig, BartConfig, BertConfig, BertGenerationConfig, BigBirdConfig, BigBirdPegasusConfig, BioGptConfig, BitNetConfig, BlenderbotConfig, BlenderbotSmallConfig, BloomConfig, BltConfig, CamembertConfig, CodeGenConfig, CohereConfig, Cohere2Config, Cohere2MoeConfig, CpmAntConfig, CTRLConfig, CwmConfig, Data2VecTextConfig, DbrxConfig, DeepseekV2Config, DeepseekV3Config, DeepseekV32Config, DeepseekV4Config, DiffLlamaConfig, DogeConfig, Dots1Config, ElectraConfig, Emu3Config, ErnieConfig, Ernie4_5Config, Ernie4_5_MoeConfig, Exaone4Config, ExaoneMoeConfig, FalconConfig, FalconH1Config, FalconMambaConfig, FlexOlmoConfig, FuyuConfig, GemmaConfig, Gemma2Config, Gemma3Config, Gemma3TextConfig, Gemma3nConfig, Gemma3nTextConfig, Gemma4Config, Gemma4AssistantConfig, Gemma4TextConfig, Gemma4UnifiedConfig, Gemma4UnifiedAssistantConfig, Gemma4UnifiedTextConfig, GitConfig, GlmConfig, Glm4Config, Glm4MoeConfig, Glm4MoeLiteConfig, GlmMoeDsaConfig, GotOcr2Config, GPT2Config, GPTBigCodeConfig, GPTNeoConfig, GPTNeoXConfig, GPTNeoXJapaneseConfig, GptOssConfig, GPTJConfig, GraniteConfig, GraniteSWAConfig, GraniteMoeConfig, GraniteMoeSWAConfig, GraniteMoeHybridConfig, GraniteMoeSharedConfig, HeliumConfig, HrmTextConfig, HunYuanDenseV1Config, HunYuanMoEV1Config, HYV3Config, HyperCLOVAXConfig, InklingTextConfig, Jais2Config, JambaConfig, JetMoeConfig, LagunaConfig, Lfm2Config, Lfm2MoeConfig, LlamaConfig, Llama4Config, Llama4TextConfig, LongcatFlashConfig, MambaConfig, Mamba2Config, MarianConfig, MBartConfig, MegatronBertConfig, MellumConfig, MiMoV2FlashConfig, MiniCPM3Config, MiniMaxConfig, MiniMaxM2Config, MiniMaxM3VLTextConfig, MinistralConfig, Ministral3Config, MistralConfig, MixtralConfig, MllamaConfig, ModernBertDecoderConfig, MoshiConfig, MptConfig, MusicgenConfig, MusicgenMelodyConfig, MvpConfig, NanoChatConfig, NemotronConfig, NemotronHConfig, OlmoConfig, Olmo2Config, Olmo3Config, OlmoHybridConfig, OlmoeConfig, OpenAIGPTConfig, OPTConfig, PegasusConfig, PersimmonConfig, PhiConfig, Phi3Config, Phi4MultimodalConfig, PhimoeConfig, PLBartConfig, ProphetNetConfig, Qwen2Config, Qwen2MoeConfig, Qwen3Config, Qwen3_5Config, Qwen3_5MoeConfig, Qwen3_5MoeTextConfig, Qwen3_5TextConfig, Qwen3MoeConfig, Qwen3NextConfig, RecurrentGemmaConfig, ReformerConfig, RemBertConfig, RobertaConfig, RobertaPreLayerNormConfig, RoCBertConfig, RoFormerConfig, RwkvConfig, SeedOssConfig, SmolLM3Config, SolarOpenConfig, StableLmConfig, Starcoder2Config, TrOCRConfig, VaultGemmaConfig, WhisperConfig, XGLMConfig, XLMConfig, XLMRobertaConfig, XLMRobertaXLConfig, XLNetConfig, xLSTMConfig, XmodConfig, YoutuConfig, ZambaConfig, Zamba2Config, ZayaConfig. [repeated 7x across cluster] (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 13:27:32,948] dflash_trainer.py:228 INFO [Rank 6] DFlash model initialized with FSDP2 [repeated 7x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) Calling super().encode with {'truncation': True, 'max_length': 65537, 'add_special_tokens': True} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2078) Calling super().encode with {'truncation': True, 'max_length': 65537, 'add_special_tokens': True} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2080) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2080) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2080) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2079) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2079) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2079) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2078) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2078) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2078) Calling super().encode with {'add_special_tokens': False} (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) Loading fastsafetensors checkpoint shards: 0% Completed | 0/16 [00:002->1 [1] 3/-1/-1->2->1 [2] 3/-1/-1->2->1 [3] 3/-1/-1->2->1 [4] 3/-1/-1->2->1 [5] 3/-1/-1->2->1 [6] 3/-1/-1->2->1 [7] 3/-1/-1->2->1 [8] 3/-1/-1->2->1 [9] 3/-1/-1->2->1 [10] 3/-1/-1->2->1 [11] 3/-1/-1->2->1 [12] 3/-1/-1->2->-1 [13] 3/-1/-1->2->-1 [14] -1/-1/-1->2->1 [15] -1/-1/-1->2->1 [16] 3/-1/-1->2->1 [17] 3/-1/-1->2->1 [18] 3/-1/-1->2->1 [19] 3/-1/-1->2->1 [20] 3/-1/-1->2->1 [21] 3/-1/-1->2->1 [22] 3/-1/-1->2->1 [23] 3/-1/-1->2->1 [24] 3/-1/-1->2->1 [25] 3/-1/-1->2->1 [26] 3/-1/-1->2->1 [27] 3/-1/-1->2->1 [28] 3/-1/-1->2->-1 [29] 3/-1/-1->2->-1 [30] -1/-1/-1->2->1 [31] -1/-1/-1->2->1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO P2P Chunksize set to 524288 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2302 [0] NCCL INFO [Proxy Service] Device 2 CPU core 103 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2304 [0] NCCL INFO [Proxy Service UDS] Device 2 CPU core 112 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO ncclCommInitRank comm 0x420297f0 rank 2 nranks 4 cudaDev 2 nvmlDev 2 busId 1806000 commId 0x5db36afb32ef643b - Init COMPLETE (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Init timings - ncclCommInitRank: rank 2 nranks 4 total 19.47 (kernels 18.24, alloc 0.09, bootstrap 0.01, allgathers 0.00, topo 0.37, graphs 0.01, connections 0.75, rest 0.00) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 00/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 01/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2 1/-1/-1->0->-1 [1] 1/-1/-1->0->-1 [2] 1/-1/-1->0->-1 [3] 1/-1/-1->0->-1 [4] 1/-1/-1->0->-1 [5] 1/-1/-1->0->-1 [6] 1/-1/-1->0->-1 [7] 1/-1/-1->0->-1 [8] 1/-1/-1->0->-1 [9] 1/-1/-1->0->-1 [10] -1/-1/-1->0->3 [11] -1/-1/-1->0->3 [12] 1/-1/-1->0->3 [13] 1/-1/-1->0->3 [14] 1/-1/-1->0->3 [15] 1/-1/-1->0->3 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1 [24] 1/-1/-1->0->-1 [25] 1/-1/-1->0->-1 [26] -1/-1/-1->0->3 [27] -1/-1/-1->0->3 [28] 1/-1/-1->0->3 [29] 1/-1/-1->0->3 [30] 1/-1/-1->0->3 [31] 1/-1/-1->0->3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO P2P Chunksize set to 524288 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2306 [0] NCCL INFO [Proxy Service] Device 0 CPU core 2 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2308 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO CC Off, workFifoBytes 1048576 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO ncclCommInitRank comm 0x4f938130 rank 0 nranks 4 cudaDev 0 nvmlDev 0 busId 806000 commId 0x5db36afb32ef643b - Init COMPLETE (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Init timings - ncclCommInitRank: rank 0 nranks 4 total 19.28 (kernels 18.05, alloc 0.08, bootstrap 0.02, allgathers 0.00, topo 0.37, graphs 0.01, connections 0.74, rest 0.01) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 00/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 01/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 02/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 03/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 04/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 05/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 06/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 07/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 08/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 09/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 10/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 11/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 12/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 13/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 14/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 15/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 16/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 17/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 18/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 19/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 20/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 21/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 22/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 23/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 24/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 25/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 26/0 : 0[0] -> 1[1] via P2P/CUM (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) INFO 10-18 13:27:56 [cuda_communicator.py:269] Using ['FLASHINFER', 'SYMM_MEM', 'PYNCCL'] all-reduce backends (in dispatch order) for group 'tp:0' out of potential backends: ['NCCL_SYMM_MEM', 'QUICK_REDUCE', 'FLASHINFER', 'AITER_CUSTOM', 'CUSTOM', 'SYMM_MEM', 'PYNCCL']. (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 02/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 03/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 04/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 05/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 06/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 07/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 08/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 09/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 10/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 11/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 12/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 13/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 14/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 15/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 16/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 17/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 18/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 19/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 20/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 21/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 22/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 23/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 24/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 25/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 26/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 27/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 28/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 29/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 30/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 31/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Trees [0] -1/-1/-1->3->2 [1] -1/-1/-1->3->2 [2] -1/-1/-1->3->2 [3] -1/-1/-1->3->2 [4] -1/-1/-1->3->2 [5] -1/-1/-1->3->2 [6] -1/-1/-1->3->2 [7] -1/-1/-1->3->2 [8] -1/-1/-1->3->2 [9] -1/-1/-1->3->2 [10] 0/-1/-1->3->2 [11] 0/-1/-1->3->2 [12] 0/-1/-1->3->2 [13] 0/-1/-1->3->2 [14] 0/-1/-1->3->-1 079:2079 [2] NCCL INFO Channel 02/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 03/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 04/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 05/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 06/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 07/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 08/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 09/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 10/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 11/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 12/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 13/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 14/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 15/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 16/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 17/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 18/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 19/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 20/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 21/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 22/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 23/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 24/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 25/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 26/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 27/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 28/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 29/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 30/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 31/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Trees [0] 3/-1/-1->2->1 [1] 3/-1/-1->2->1 [2] 3/-1/-1->2->1 [3] 3/-1/-1->2->1 [4] 3/-1/-1->2->1 [5] 3/-1/-1->2->1 [6] 3/-1/-1->2->1 [7] 3/-1/-1->2->1 [8] 3/-1/-1->2->1 [9] 3/-1/-1->2->1 [10] 3/-1/-1->2->1 [11] 3/-1/-1->2->1 [12] 3/-1/-1->2->-1 [13] 3/-1/-1->2->-1 [14] -1/-1/-1->2->1 [15] -1/-1/-1->2->1 [16][1] NCCL INFO Channel 02/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 03/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 04/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 05/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 06/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 07/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 08/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 09/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 10/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 11/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 12/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 13/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 14/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 15/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 16/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 17/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 18/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 19/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 20/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 21/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 22/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 23/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 24/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 25/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 26/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 27/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 28/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 29/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 30/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Channel 31/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 27/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 28/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 29/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 30/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 31/0 : 0[0] -> 1[1] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1846:1846 [1 1/-1/-1->0->-1 [1] 1/-1/-1->0->-1 [2] 1/-1/-1->0->-1 [3] 1/-1/-1->0->-1 [4] 1/-1/-1->0->-1 [5] 1/-1/-1->0->-1 [6] 1/-1/-1->0->-1 [7] 1/-1/-1->0->-1 [8] 1/-1/-1->0->-1 [9] 1/-1/-1->0->-1 [10] -1/-1/-1->0->3 [11] -1/-1/-1->0->3 [12] 1/-1/-1->0->3 [13] 1/-1/-1->0->3 [14] 1/-1/-1->0->3 [15] 1/-1/-1->0->3 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1 [24] 1/-1/-1->0->-1 [25] 1/-1/-1->0->-1 [26] -1/-1/-1->0->3 [27] -1/-1/-1->0->3 [28] 1/-1/-1->0->3 [29] 1/-1/-1->0->3 [30] 1/-1/-1->0->3 [31] 1/-1/-1->0->3 (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Channel 26/0 : 0[0] -> 1[1] via P2P/CUMhost-10-0-3-58:1848:1848 [3] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so (VllmEngine pid=1867, ip=10.0.3.60) >0->3 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1 [24] 1/-1/-1->0->-1 [25] 1/-1/-1->0->-1 [26] -1/-1/-1->0->3 [27] -1/-1/-1->0->3 [28] 1/-1/-1->0->3 [29] 1/-1/-1->0->3 [30] 1/-1/-1->0->3 [31] 1/-1/-1->0->3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Connected all rings, use ring (Worker pid=2077) INFO 10-18 13:27:58 [cuda_communicator.py:269] Using ['PYNCCL'] all-reduce backends (in dispatch order) for group 'ep:0' out of potential backends: ['NCCL_SYMM_MEM', 'QUICK_REDUCE', 'FLASHINFER', 'AITER_CUSTOM', 'CUSTOM', 'SYMM_MEM', 'PYNCCL']. (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) INFO 10-18 13:27:58 [parallel_state.py:1982] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2079) WARNING 10-18 13:27:58 [__init__.py:205] min_p and logit_bias parameters won't work with speculative decoding. (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2080) WARNING 10-18 13:27:58 [__init__.py:205] min_p and logit_bias parameters won't work with speculative decoding. (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1 1/-1/-1->0->-1 [1] 1/-1/-1->0->-1 [2] 1/-1/-1->0->-1 [3] 1/-1/-1->0->-1 [4] 1/-1/-1->0->-1 [5] 1/-1/-1->0->-1 [6] 1/-1/-1->0->-1 [7] 1/-1/-1->0->-1 [8] 1/-1/-1->0->-1 [9] 1/-1/-1->0->-1 [10] -1/-1/-1->0->3 [11] -1/-1/-1->0->3 [12] 1/-1/-1->0->3 [13] 1/-1/-1->0->3 [14] 1/-1/-1->0->3 [15] 1/-1/-1->0->3 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1 [24] 1/-1/-1->0->-1 [25] 1/-1/-1->0->-1 [26] -1/-1/-1->0->3 [27] -1/-1/-1->0->3 [28] 1/-1/-1->0->3 [29] 1/-1/-1->0->3 [30] 1/-1/-1->0->3 [31] 1/-1/-1->0->3 (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) INFO 10-18 13:27:59 [topk_topp_sampler.py:62] Using FlashInfer for top-p & top-k sampling. (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2078) WARNING 10-18 13:27:59 [__init__.py:205] min_p and logit_bias parameters won't work with speculative decoding. (VllmEngine pid=1867, ip=10.0.3.60) (Worker pid=2077) WARNING 10-18 13:27:59 [__init__.py:205] min_p and logit_bias parameters won't work with speculative decoding. (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1848:1848 [3] NCCL INFO Trees [0] -1/-1/-1->3->2 [1] -1/-1/-1->3->2 [2] -1/-1/-1->3->2 [3] -1/-1/-1->3->2 [4] -1/-1/-1->3->2 [5] -1/-1/-1->3->2 [6] -1/-1/-1->3->2 [7] -1/-1/-1->3->2 [8] -1/-1/-1->3->2 [9] -1/-1/-1->3->2 [10] 0/-1/-1->3->2 [11] 0/-1/-1->3->2 [12] 0/-1/-1->3->2 [13] 0/-1/-1->3->2 [14] 0/-1/-1->3->-1 ] NCCL INFO Channel 02/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP3 pid=2080) INFO 10-18 13:28:00 [kernel.py:310] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP2 pid=2079) INFO 10-18 13:28:00 [kernel.py:310] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:00 [gpu_model_runner.py:5491] Starting to load model /data/models/moonshotai/Kimi-K3-Flash... (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:00 [cuda.py:551] Using backend AttentionBackendEnum.FLASH_ATTN for vit attention (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:00 [mm_encoder_attention.py:375] Using AttentionBackendEnum.FLASH_ATTN for MMEncoderAttention. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:00 [vllm.py:1388] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:00 [vllm.py:1423] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:00 [kernel.py:310] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:00 [vllm.py:1602] Cudagraph is disabled under eager mode (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:00 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant, allreduce_rms (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2host-10-0-3-59:2020:2020 [1] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so [repeated 7x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Bootstrap: Using eth0:10.0.3.59<0> [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO cudaDriverVersion 13000 [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO NCCL version 2.29.7+cuda13.2 [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO NCCL git version stable b81d6a5a3 [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO NET/Plugin: Could not find: libnccl-net.so [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO NCCL_IB_HCA set to mlx5_bond_0 [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO NET/IB : Using [0]mlx5_bond_0:1/RoCE [RO]; OOB eth0:10.0.3.59<0> [repeated 8x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Initialized NET plugin IB [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Assigned NET plugin IB to comm [repeated 16x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO GIN/Plugin: Could not find: libnccl-gin.so [repeated 8x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Using network IB [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO DMA-BUF is available on GPU device 0 [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO [Rank 0] ncclCommInitRank comm 0x37609ca0 rank 0 nranks 4 cudaDev 0 nvmlDev 0 busId 806000 commId 0xeb33607ed5bc29f7 - Init START [repeated 16x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO RAS client listening socket at 127.0.0.1<28028> [repeated 8x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Bootstrap timings total 0.001082 (create 0.000033, send 0.000204, recv 0.000531, ring 0.000130, delay 0.000000) [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO MNNVL busId 0x806000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 0 is 0-71. (GPU affinity = 0-71 ; CPU affinity = 0-71). [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO NVLS multicast support is available on dev 0 (NVLS_NCHANNELS 24) [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO comm 0x37609ca0 rank 0 nRanks 4 nNodes 1 localRanks 4 localRank 0 MNNVL 0 [repeated 16x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Channel 31/32 : 0 1 2 3 [repeated 128x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Trees [0] 1/-1/-1->0->-1 [1] 1/-1/-1->0->-1 [2] 1/-1/-1->0->-1 [3] 1/-1/-1->0->-1 [4] 1/-1/-1->0->-1 [5] 1/-1/-1->0->-1 [6] 1/-1/-1->0->-1 [7] 1/-1/-1->0->-1 [8] 1/-1/-1->0->-1 [9] 1/-1/-1->0->-1 [10] -1/-1/-1->0->3 [11] -1/-1/-1->0->3 [12] 1/-1/-1->0->3 [13] 1/-1/-1->0->3 [14] 1/-1/-1->0->3 [15] 1/-1/-1- [repeated 12x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Trees [0] 2/-1/-1->1->0 [1] 2/-1/-1->1->0 [2] 2/-1/-1->1->0 [3] 2/-1/-1->1->0 [4] 2/-1/-1->1->0 [5] 2/-1/-1->1->0 [6] 2/-1/-1->1->0 [7] 2/-1/-1->1->0 [8] 2/-1/-1->1->0 [9] 2/-1/-1->1->0 [10] 2/-1/-1->1->-1 [11] 2/-1/-1->1->-1 [12] -1/-1/-1->1->0 [13] -1/-1/-1->1->0 [14] 2/-1/-1->1->0 [15] 2/-1/-1->1->0 [16] 2/-1/-1->1->0 [17] MEM (VllmEngine pid=1635, ip=10.0.3.58) >0->3 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1 [24] 1/-1/-1->0->-1 [25] 1/-1/-1->0->-1 [26] -1/-1/-1->0->3 [27] -1/-1/-1->0->3 [28] 1/-1/-1->0->3 [29] 1/-1/-1->0->3 [30] 1/-1/-1->0->3 [31] 1/-1/-1->0->3 (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Connected all rings, use ring(Worker pid=1845) INFO 10-18 13:28:00 [cuda_communicator.py:269] Using ['PYNCCL'] all-reduce backends (in dispatch order) for group 'ep:0' out of potential backends: ['NCCL_SYMM_MEM', 'QUICK_REDUCE', 'FLASHINFER', 'AITER_CUSTOM', 'CUSTOM', 'SYMM_MEM', 'PYNCCL']. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP1 pid=2078) INFO 10-18 13:28:00 [kernel.py:310] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO P2P Chunksize set to 524288 [repeated 10x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so [repeated 8x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 [repeated 10x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:2189 [0] NCCL INFO [Proxy Service] Device 0 CPU core 71 [repeated 10x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:2194 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 1 [repeated 10x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so [repeated 8x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 [repeated 10x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer [repeated 10x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Symmetric VA size=280GB [repeated 20x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO ncclCommInitRank comm 0x37609ca0 rank 0 nranks 4 cudaDev 0 nvmlDev 0 busId 806000 commId 0xeb33607ed5bc29f7 - Init COMPLETE [repeated 10x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Init timings - ncclCommInitRank: rank 0 nranks 4 total 0.79 (kernels 0.00, alloc 0.00, bootstrap 0.00, allgathers 0.00, topo 0.67, graphs 0.01, connections 0.11, rest 0.00) [repeated 10x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO Channel 31/0 : 0[0] -> 1[1] via P2P/CUMEM [repeated 315x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1845:1845 [0] NCCL INFO CC Off, workFifoBytes 1048576 [repeated 4x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) >0->3 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1 [24] 1/-1/-1->0->-1 [25] 1/-1/-1->0->-1 [26] -1/-1/-1->0->3 [27] -1/-1/-1->0->3 [28] 1/-1/-1->0->3 [29] 1/-1/-1->0->3 [30] 1/-1/-1->0->3 [31] 1/-1/-1->0->3 (VllmEngine pid=1809, ip=10.0.3.59) (Worker pid=2019) INFO 10-18 13:28:00 [cuda_communicator.py:269] Using ['FLASHINFER', 'SYMM_MEM', 'PYNCCL'] all-reduce backends (in dispatch order) for group 'tp:0' out of potential backends: ['NCCL_SYMM_MEM', 'QUICK_REDUCE', 'FLASHINFER', 'AITER_CUSTOM', 'CUSTOM', 'SYMM_MEM', 'PYNCCL']. [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2022:2022 [3] NCCL INFO Trees [0] -1/-1/-1->3->2 [1] -1/-1/-1->3->2 [2] -1/-1/-1->3->2 [3] -1/-1/-1->3->2 [4] -1/-1/-1->3->2 [5] -1/-1/-1->3->2 [6] -1/-1/-1->3->2 [7] -1/-1/-1->3->2 [8] -1/-1/-1->3->2 [9] -1/-1/-1->3->2 [10] 0/-1/-1->3->2 [11] 0/-1/-1->3->2 [12] 0/-1/-1->3->2 [13] 0/-1/-1->3->2 [14] 0/-1/-1->3->-1 [15021:2021 [2] NCCL INFO Channel 02/0 : 2[2] -> 3[3] via P2P/CUMEM [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker pid=2019) INFO 10-18 13:28:01 [parallel_state.py:1982] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker pid=2020) WARNING 10-18 13:28:02 [__init__.py:205] min_p and logit_bias parameters won't work with speculative decoding. [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker pid=2019) INFO 10-18 13:28:02 [topk_topp_sampler.py:62] Using FlashInfer for top-p & top-k sampling. [repeated 2x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1962:1962 [1 1/-1/-1->0->-1 [1] 1/-1/-1->0->-1 [2] 1/-1/-1->0->-1 [3] 1/-1/-1->0->-1 [4] 1/-1/-1->0->-1 [5] 1/-1/-1->0->-1 [6] 1/-1/-1->0->-1 [7] 1/-1/-1->0->-1 [8] 1/-1/-1->0->-1 [9] 1/-1/-1->0->-1 [10] -1/-1/-1->0->3 [11] -1/-1/-1->0->3 [12] 1/-1/-1->0->3 [13] 1/-1/-1->0->3 [14] 1/-1/-1->0->3 [15] 1/-1/-1->0->3 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1 [24] 1/-1/-1->0->-1 [25] 1/-1/-1->0->-1 [26] -1/-1/-1->0->3 [27] -1/-1/-1->0->3 [28] 1/-1/-1->0->3 [29] 1/-1/-1->0->3 [30] 1/-1/-1->0->3 [31] 1/-1/-1->0->3 (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Channel 26/0 : 0[0] -> 1[1] via P2P/Chost-10-0-3-63:1964:1964 [3] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so (VllmEngine pid=1751, ip=10.0.3.63) ho (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Trees [0] 3/-1/-1->2->1 [1] 3/-1/-1->2->1 [2] 3/-1/-1->2->1 [3] 3/-1/-1->2->1 [4] 3/-1/-1->2->1 [5] 3/-1/-1->2->1 [6] 3/-1/-1->2->1 [7] 3/-1/-1->2->1 [8] 3/-1/-1->2->1 [9] 3/-1/-1->2->1 [10] 3/-1/-1->2->1 [11] 3/-1/-1->2->1 [12] 3/-1/-1->2->-1 [13] 3/-1/-1->2->-1 [14] -1/-1/-1->2->1 [15] -1/-1/-1->2->1 [16] ] NCCL INFO Channel 02/0 : 1[1] -> 2[2] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:04 [kernel.py:310] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:03 [gpu_model_runner.py:5491] Starting to load model /data/models/moonshotai/Kimi-K3-Flash... [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:04 [cuda.py:551] Using backend AttentionBackendEnum.FLASH_ATTN for vit attention [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:04 [mm_encoder_attention.py:375] Using AttentionBackendEnum.FLASH_ATTN for MMEncoderAttention. [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:28:04 [vllm.py:1388] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:28:04 [vllm.py:1423] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:04 [vllm.py:1602] Cudagraph is disabled under eager mode [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:04 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant, allreduce_rms [repeated 2x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:196host-10-0-3-63:1962:1962 [1] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Bootstrap: Using eth0:10.0.3.63<0> [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO cudaDriverVersion 13000 [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO NCCL version 2.29.7+cuda13.2 [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO NCCL git version stable b81d6a5a3 [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO NET/Plugin: Could not find: libnccl-net.so [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO NCCL_IB_HCA set to mlx5_bond_0 [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO NET/IB : Using [0]mlx5_bond_0:1/RoCE [RO]; OOB eth0:10.0.3.63<0> [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Initialized NET plugin IB [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Assigned NET plugin IB to comm [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO GIN/Plugin: Could not find: libnccl-gin.so [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Using network IB [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO DMA-BUF is available on GPU device 0 [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO [Rank 0] ncclCommInitRank comm 0x4116b210 rank 0 nranks 4 cudaDev 0 nvmlDev 0 busId 806000 commId 0x9999f29976d49fd0 - Init START [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO RAS client listening socket at 127.0.0.1<28028> [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Bootstrap timings total 0.001484 (create 0.000028, send 0.000182, recv 0.000303, ring 0.000174, delay 0.000000) [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO MNNVL busId 0x806000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 0 is 0-71. (GPU affinity = 0-71 ; CPU affinity = 0-71). [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO NVLS multicast support is available on dev 0 (NVLS_NCHANNELS 24) [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO comm 0x4116b210 rank 0 nRanks 4 nNodes 1 localRanks 4 localRank 0 MNNVL 0 [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Channel 31/32 : 0 1 2 3 [repeated 96x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Trees [0] 1/-1/-1->0->-1 [1] 1/-1/-1->0->-1 [2] 1/-1/-1->0->-1 [3] 1/-1/-1->0->-1 [4] 1/-1/-1->0->-1 [5] 1/-1/-1->0->-1 [6] 1/-1/-1->0->-1 [7] 1/-1/-1->0->-1 [8] 1/-1/-1->0->-1 [9] 1/-1/-1->0->-1 [10] -1/-1/-1->0->3 [11] -1/-1/-1->0->3 [12] 1/-1/-1->0->3 [13] 1/-1/-1->0->3 [14] 1/-1/-1->0->3 [15] 1/-1/- [repeated 7x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO Connected all rings, use ring(Worker pid=2019) INFO 10-18 13:28:01 [cuda_communicator.py:269] Using ['PYNCCL'] all-reduce backends (in dispatch order) for group 'ep:0' out of potential backends: ['NCCL_SYMM_MEM', 'QUICK_REDUCE', 'FLASHINFER', 'AITER_CUSTOM', 'CUSTOM', 'SYMM_MEM', 'PYNCCL']. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=0) role=F(with indexer) (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO P2P Chunksize set to 524288 [repeated 5x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 [repeated 5x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:2191 [0] NCCL INFO [Proxy Service] Device 3 CPU core 83 [repeated 5x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:2193 [0] NCCL INFO [Proxy Service UDS] Device 3 CPU core 84 [repeated 5x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 [repeated 5x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer [repeated 5x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Symmetric VA size=280GB [repeated 10x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO ncclCommInitRank comm 0x13ec2680 rank 3 nranks 4 cudaDev 3 nvmlDev 3 busId 1906000 commId 0xe5cf17c7a7e8afd8 - Init COMPLETE [repeated 5x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Init timings - ncclCommInitRank: rank 3 nranks 4 total 19.77 (kernels 18.31, alloc 0.09, bootstrap 0.02, allgathers 0.00, topo 0.49, graphs 0.01, connections 0.84, rest 0.00) [repeated 5x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Channel 31/0 : 0[0] -> 1[1] via P2P/CUMEM [repeated 157x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO CC Off, workFifoBytes 1048576 [repeated 2x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker pid=1961) INFO 10-18 13:28:05 [cuda_communicator.py:269] Using ['FLASHINFER', 'SYMM_MEM', 'PYNCCL'] all-reduce backends (in dispatch order) for group 'tp:0' out of potential backends: ['NCCL_SYMM_MEM', 'QUICK_REDUCE', 'FLASHINFER', 'AITER_CUSTOM', 'CUSTOM', 'SYMM_MEM', 'PYNCCL']. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [cuda.py:492] Using FLASHMLA_SPARSE attention backend out of potential backends: ['FLASHMLA_SPARSE']. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [selector.py:192] Using FLASH_ATTN MLA prefill backend. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [deep_gemm.py:186] deep_gemm not found in site-packages, trying vendored vllm.third_party.deep_gemm (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [deep_gemm.py:213] DeepGEMM PDL enabled on vllm.third_party.deep_gemm. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [deep_gemm.py:131] DeepGEMM E8M0 enabled on current platform. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=1) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:07 [model.py:271] MoE renormalize is True. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [compressed_tensors_moe_w4a4_mxfp4.py:62] Using CutlassExpertsMxfp4 for MXFP4 MoE (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=2) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=3) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=4) role=F(with indexer) (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1963:1963 [2] NCCL INFO Trees [0] 3/-1/-1->2->1 [1] 3/-1/-1->2->1 [2] 3/-1/-1->2->1 [3] 3/-1/-1->2->1 [4] 3/-1/-1->2->1 [5] 3/-1/-1->2->1 [6] 3/-1/-1->2->1 [7] 3/-1/-1->2->1 [8] 3/-1/-1->2->1 [9] 3/-1/-1->2->1 [10] 3/-1/-1->2->1 [11] 3/-1/-1->2->1 [12] 3/-1/-1->2->-1 [13] 3/-1/-1->2->-1 [14] -1/-1/-1->2->1 [15] -1/-1/-1->2->1 [16] 3] NCCL INFO Channel 02/0 : 1[1] -> 2[2] via P2P/CUMEM [repeated 2x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) 1->0->3 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1 [24] 1/-1/-1->0->-1 [25] 1/-1/-1->0->-1 [26] -1/-1/-1->0->3 [27] -1/-1/-1->0->3 [28] 1/-1/-1->0->3 [29] 1/-1/-1->0->3 [30] 1/-1/-1->0->3 [31] 1/-1/-1->0->3 (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Connected all rings, use rin(Worker pid=1961) INFO 10-18 13:28:07 [cuda_communicator.py:269] Using ['PYNCCL'] all-reduce backends (in dispatch order) for group 'ep:0' out of potential backends: ['NCCL_SYMM_MEM', 'QUICK_REDUCE', 'FLASHINFER', 'AITER_CUSTOM', 'CUSTOM', 'SYMM_MEM', 'PYNCCL']. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=5) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=6) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=7) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=8) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=9) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=10) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=11) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=12) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=13) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=14) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=15) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=16) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=17) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=18) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=19) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=20) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=21) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=22) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=23) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=24) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=25) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=26) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=27) role=S(skip_topk)(VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) Loading fastsafetensors checkpoint shards: 6% Completed | 1/16 [00:07<01:59, 7.97s/it] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) Loading fastsafetensors checkpoint shards: 0% Completed | 0/16 [00:00` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP3 pid=2080) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_011785'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP2 pid=2079) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP3 pid=2080) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP1 pid=2078) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_011785'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP1 pid=2078) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_011785'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP3 pid=1964) Calling super().encode with {'truncation': True, 'max_length': 65537, 'add_special_tokens': True} [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP3 pid=1964) Calling super().encode with {'add_special_tokens': False} [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP1 pid=2020) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_011785'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP1 pid=2020) exe_args, func_types, adapted_args = self.generate_mlir_function_types( [repeated 8x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=28) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=29) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=30) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=31) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=32) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=33) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=34) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=35) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=36) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=37) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=38) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=39) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=40) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=41) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=42) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=43) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=44) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=45) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=46) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=47) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=48) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:07 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=49) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=50) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=51) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=52) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=53) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=54) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=55) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=56) role=F(with indexer) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=57) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=58) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=59) role=S(skip_topk) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:28:08 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=60) role=F(with indexer) (VllmEngine pid=1751, ip=10.0.3.63) (Worker pid=1961) INFO 10-18 13:28:07 [parallel_state.py:1982] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A (VllmEngine pid=1751, ip=10.0.3.63) (Worker pid=1963) WARNING 10-18 13:28:08 [__init__.py:205] min_p and logit_bias parameters won't work with speculative decoding. [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker pid=1961) INFO 10-18 13:28:08 [topk_topp_sampler.py:62] Using FlashInfer for top-p & top-k sampling. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP3 pid=1964) INFO 10-18 13:28:10 [kernel.py:310] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:08 [gpu_model_runner.py:5491] Starting to load model /data/models/moonshotai/Kimi-K3-Flash... (VllmEngine pid=1867, ip=10.0.3.60) 3/-1/-1->2->1 [17] 3/-1/-1->2->1 [18] 3/-1/-1->2->1 [19] 3/-1/-1->2->1 [20] 3/-1/-1->2->1 [21] 3/-1/-1->2->1 [22] 3/-1/-1->2->1 [23] 3/-1/-1->2->1 [24] 3/-1/-1->2->1 [25] 3/-1/-1->2->1 [26] 3/-1/-1->2->1 [27] 3/-1/-1->2->1 [28] 3/-1/-1->2->-1 [29] 3/-1/-1->2->-1 [30] -1/-1/-1->2->1 [31] -1/-1/-1->2->1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Comm config Blocking set to 1 (VllmEngine pid=1867, ip=10.0.3.60) h (VllmEngine pid=1867, ip=10.0.3.60) 7] 2/-1/-1->1->0 [18] 2/-1/-1->1->0 [19] 2/-1/-1->1->0 [20] 2/-1/-1->1->0 [21] 2/-1/-1->1->0 [22] 2/-1/-1->1->0 [23] 2/-1/-1->1->0 [24] 2/-1/-1->1->0 [25] 2/-1/-1->1->0 [26] 2/-1/-1->1->-1 [27] 2/-1/-1->1->-1 [28] -1/-1/-1->1->0 [29] -1/-1/-1->1->0 [30] 2/-1/-1->1->0 [31] 2/-1/-1->1->0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Comm config Blocking set to 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2 (VllmEngine pid=1867, ip=10.0.3.60) [15] 0/-1/-1->3->-1 [16] -1/-1/-1->3->2 [17] -1/-1/-1->3->2 [18] -1/-1/-1->3->2 [19] -1/-1/-1->3->2 [20] -1/-1/-1->3->2 [21] -1/-1/-1->3->2 [22] -1/-1/-1->3->2 [23] -1/-1/-1->3->2 [24] -1/-1/-1->3->2 [25] -1/-1/-1->3->2 [26] 0/-1/-1->3->2 [27] 0/-1/-1->3->2 [28] 0/-1/-1->3->2 [29] 0/-1/-1->3->2 [30] 0/-1/-1->3->-1 [31] 0/-1/-1->3->-1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:09 [cuda.py:551] Using backend AttentionBackendEnum.FLASH_ATTN for vit attention (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:09 [mm_encoder_attention.py:375] Using AttentionBackendEnum.FLASH_ATTN for MMEncoderAttention. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) WARNING 10-18 13:28:09 [vllm.py:1388] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) WARNING 10-18 13:28:09 [vllm.py:1423] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:09 [vllm.py:1602] Cudagraph is disabled under eager mode (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:09 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant, allreduce_rms (VllmEngine pid=1867, ip=10.0.3.60) PXN 0 GDR 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Comm config Blocking set to 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Initialized NET plugin IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Assigned NET plugin IB to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Using network IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO DMA-BUF is available on GPU device 0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO [Rank 0] ncclCommInitRankConfig comm 0x653c9b60 rank 0 nranks 4 cudaDev 0 nvmlDev 0 busId 806000 commId 0x33f6e55a90025d95 - Init START (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Bootstrap timings total 2.719782 (create 0.000037, send 0.000213, recv 2.665014, ring 0.000137, delay 0.000001) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO MNNVL busId 0x806000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 0 is 0-71. (GPU affinity = 0-71 ; CPU affinity = 0-71). (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO NVLS multicast support is available on dev 0 (NVLS_NCHANNELS 24) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO comm 0x653c9b60 rank 0 nRanks 4 nNodes 1 localRanks 4 localRank 0 MNNVL 0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 00/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 01/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 02/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 03/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 04/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 05/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 06/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 07/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 08/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 09/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 10/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 11/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 12/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 13/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 14/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 15/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 16/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 17/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 18/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 19/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 20/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 21/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 22/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 23/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 24/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 25/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 26/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 27/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 28/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 29/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 30/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Channel 31/32 : 0 1 2 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Trees [0] 1/-1/-1->0->-1 [1] 1/-1/-1->0->-1 [2] 1/-1/-1->0->-1 [3] 1/-1/-1->0->-1 [4] 1/-1/-1->0->-1 [5] 1/-1/-1->0->-1 [6] 1/-1/-1->0->-1 [7] 1/-1/-1->0->-1 [8] 1/-1/-1->0->-1 [9] 1/-1/-1->0->-1 [10] -1/-1/-1->0->3 [11] -1/-1/-1->0->3 [12] 1/-1/-1->0->3 [13] 1/-1/-1->0->3 [14] 1/-1/-1->0->3 [15] 1/-1/-1->0->3 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1 [24] 1/-1/-1->0->-1 [25] 1/-1/-1->0->-1 [26] -1/-1/-1->0->3 [27] -1/-1/-1->0->3 [28] 1/-1/-1->0->3 [29] 1/-1/-1->0->3 [30] 1/-1/-1->0->3 [31] 1/-1/-1->0->3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-6 (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:11 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=60) role=F(with indexer) [repeated 38x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO P2P Chunksize set to 524288 [repeated 5x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 [repeated 4x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2420 [0] NCCL INFO [Proxy Service] Device 3 CPU core 124 [repeated 4x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2424 [0] NCCL INFO [Proxy Service UDS] Device 3 CPU core 126 [repeated 4x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 [repeated 4x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer [repeated 4x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Symmetric VA size=280GB [repeated 8x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO ncclCommInitRank comm 0x2ed61eb0 rank 3 nranks 4 cudaDev 3 nvmlDev 3 busId 1906000 commId 0x7d16e9ee8f1c60dd - Init COMPLETE [repeated 4x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Init timings - ncclCommInitRank: rank 3 nranks 4 total 1.01 (kernels 0.00, alloc 0.00, bootstrap 0.00, allgathers 0.00, topo 0.44, graphs 0.01, connections 0.55, rest 0.00) [repeated 4x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Channel 31/0 : 3[3] -> 0[0] via P2P/CUMEM [repeated 128x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO CC Off, workFifoBytes 1048576 (VllmEngine pid=1867, ip=10.0.3.60) ost-10-0-3-60:2079:2079 [2] NCCL INFO Initialized NET plugin IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Assigned NET plugin IB to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Using network IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO DMA-BUF is available on GPU device 2 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO [Rank 2] ncclCommInitRankConfig comm 0x5828f940 rank 2 nranks 4 cudaDev 2 nvmlDev 2 busId 1806000 commId 0x33f6e55a90025d95 - Init START (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Bootstrap timings total 0.236630 (create 0.000057, send 0.000276, recv 0.235840, ring 0.000166, delay 0.000000) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO MNNVL busId 0x1806000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 2 is 72-143. (GPU affinity = 72-143 ; CPU affinity = 72-143). (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO NVLS multicast support is available on dev 2 (NVLS_NCHANNELS 24) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO comm 0x5828f940 rank 2 nRanks 4 nNodes 1 localRanks 4 localRank 2 MNNVL 0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Trees [0] 3/-1/-1->2->1 [1] 3/-1/-1->2->1 [2] 3/-1/-1->2->1 [3] 3/-1/-1->2->1 [4] 3/-1/-1->2->1 [5] 3/-1/-1->2->1 [6] 3/-1/-1->2->1 [7] 3/-1/-1->2->1 [8] 3/-1/-1->2->1 [9] 3/-1/-1->2->1 [10] 3/-1/-1->2->1 [11] 3/-1/-1->2->1 [12] 3/-1/-1->2->-1 [13] 3/-1/-1->2->-1 [14] -1/-1/-1->2->1 [15] -1/-1/-1->2->1 [16] 3/-1/-1->2->1 [17] 3/-1/-1->2->1 [18] 3/-1/-1->2->1 [19] 3/-1/-1->2->1 [20] 3/-1/-1->2->1 [21] 3/-1/-1->2->1 [22] 3/-1/-1->2->1 [23] 3/-1/-1->2->1 [24] 3/-1/-1->2->1 [25] 3/-1/-1->2->1 [26] 3/-1/-1->2->1 [27] 3/-1/-1->2->1 [28] 3/-1/-1->2->-1 [29] 3/-1/-1->2->-1 [30] -1/-1/-1->2->1 [31] -1/-1/-1->2->1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO ncclCommInitRankConfig comm 0x5828f940 rank 2 nranks 4 cudaDev 2 nvmlDev 2 busId 1806000 commId 0x33f6e55a90025d95 - Init COMPLETE (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Init timings - ncclCommInitRankConfig: rank 2 nranks 4 total 1.18 (kernels 0.00, alloc 0.05, bootstrap 0.24, allgathers 0.02, topo 0.69, graphs 0.01, connections 0.18, rest 0.00) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO Channel 13/0 : 2[2] -> 3[3] via P Comm config Blocking set to 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Initialized NET plugin IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Assigned NET plugin IB to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Using network IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO DMA-BUF is available on GPU device 3 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO [Rank 3] ncclCommInitRankConfig comm 0x3bd3c190 rank 3 nranks 4 cudaDev 3 nvmlDev 3 busId 1906000 commId 0x33f6e55a90025d95 - Init START (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Bootstrap timings total 0.001026 (create 0.000047, send 0.000233, recv 0.000453, ring 0.000130, delay 0.000000) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO MNNVL busId 0x1906000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 3 is 72-143. (GPU affinity = 72-143 ; CPU affinity = 72-143). (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO NVLS multicast support is available on dev 3 (NVLS_NCHANNELS 24) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO comm 0x3bd3c190 rank 3 nRanks 4 nNodes 1 localRanks 4 localRank 3 MNNVL 0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Trees [0] -1/-1/-1->3->2 [1] -1/-1/-1->3->2 [2] -1/-1/-1->3->2 [3] -1/-1/-1->3->2 [4] -1/-1/-1->3->2 [5] -1/-1/-1->3->2 [6] -1/-1/-1->3->2 [7] -1/-1/-1->3->2 [8] -1/-1/-1->3->2 [9] -1/-1/-1->3->2 [10] 0/-1/-1->3->2 [11] 0/-1/-1->3->2 [12] 0/-1/-1->3->2 [13] 0/-1/-1->3->2 [14] 0/-1/-1->3->-1 [15] 0/-1/-1->3->-1 [16] -1/-1/-1->3->2 [17] -1/-1/-1->3->2 [18] -1/-1/-1->3->2 [19] -1/-1/-1->3->2 [20] -1/-1/-1->3->2 [21] -1/-1/-1->3->2 [22] -1/-1/-1->3->2 [23] -1/-1/-1->3->2 [24] -1/-1/-1->3->2 [25] -1/-1/-1->3->2 [26] 0/-1/-1->3->2 [27] 0/-1/-1->3->2 [28] 0/-1/-1->3->2 [29] 0/-1/-1->3->2 [30] 0/-1/-1->3->-1 [31] 0/-1/-1->3->-1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO ncclCommInitRankConfig comm 0x3bd3c190 rank 3 nranks 4 cudaDev 3 nvmlDev 3 busId 1906000 commId 0x33f6e55a90025d95 - Init COMPLETE (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 [3] NCCL INFO Init timings - ncclCommInitRankConfig: rank 3 nranks 4 total 0.94 (kernels 0.00, alloc 0.05, bootstrap 0.00, allgathers 0.02, topo 0.68, graphs 0.01, connections 0.18, rest 0.00) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2080:2080 078 [1] NCCL INFO Initialized NET plugin IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Assigned NET plugin IB to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Using network IB (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO DMA-BUF is available on GPU device 1 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO [Rank 1] ncclCommInitRankConfig comm 0x4ea53580 rank 1 nranks 4 cudaDev 1 nvmlDev 1 busId 906000 commId 0x33f6e55a90025d95 - Init START (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Bootstrap timings total 0.055001 (create 0.000050, send 0.000195, recv 0.000306, ring 0.054274, delay 0.000001) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO MNNVL busId 0x906000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 1 is 0-71. (GPU affinity = 0-71 ; CPU affinity = 0-71). (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO NVLS multicast support is available on dev 1 (NVLS_NCHANNELS 24) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO comm 0x4ea53580 rank 1 nRanks 4 nNodes 1 localRanks 4 localRank 1 MNNVL 0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Trees [0] 2/-1/-1->1->0 [1] 2/-1/-1->1->0 [2] 2/-1/-1->1->0 [3] 2/-1/-1->1->0 [4] 2/-1/-1->1->0 [5] 2/-1/-1->1->0 [6] 2/-1/-1->1->0 [7] 2/-1/-1->1->0 [8] 2/-1/-1->1->0 [9] 2/-1/-1->1->0 [10] 2/-1/-1->1->-1 [11] 2/-1/-1->1->-1 [12] -1/-1/-1->1->0 [13] -1/-1/-1->1->0 [14] 2/-1/-1->1->0 [15] 2/-1/-1->1->0 [16] 2/-1/-1->1->0 [17] 2/-1/-1->1->0 [18] 2/-1/-1->1->0 [19] 2/-1/-1->1->0 [20] 2/-1/-1->1->0 [21] 2/-1/-1->1->0 [22] 2/-1/-1->1->0 [23] 2/-1/-1->1->0 [24] 2/-1/-1->1->0 [25] 2/-1/-1->1->0 [26] 2/-1/-1->1->-1 [27] 2/-1/-1->1->-1 [28] -1/-1/-1->1->0 [29] -1/-1/-1->1->0 [30] 2/-1/-1->1->0 [31] 2/-1/-1->1->0 (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO ncclCommInitRankConfig comm 0x4ea53580 rank 1 nranks 4 cudaDev 1 nvmlDev 1 busId 906000 commId 0x33f6e55a90025d95 - Init COMPLETE (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:2078 [1] NCCL INFO Init timings - ncclCommInitRankConfig: rank 1 nranks 4 total 0.96 (kernels 0.00, alloc 0.01, bootstrap 0.06, allgathers 0.02, topo 0.68, graphs 0.01, connections 0.18, rest 0.00) (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2078:20 (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:10 [cuda.py:492] Using FLASHMLA_SPARSE attention backend out of potential backends: ['FLASHMLA_SPARSE']. [repeated 2x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) [15] 0/-1/-1->3->-1 [16] -1/-1/-1->3->2 [17] -1/-1/-1->3->2 [18] -1/-1/-1->3->2 [19] -1/-1/-1->3->2 [20] -1/-1/-1->3->2 [21] -1/-1/-1->3->2 [22] -1/-1/-1->3->2 [23] -1/-1/-1->3->2 [24] -1/-1/-1->3->2 [25] -1/-1/-1->3->2 [26] 0/-1/-1->3->2 [27] 0/-1/-1->3->2 [28] 0/-1/-1->3->2 [29] 0/-1/-1->3->2 [30] 0/-1/-1->3->-1 [31] 0/-1/-1->3->-1 (VllmEngine pid=1635, ip=10.0.3.58) host (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:10 [selector.py:192] Using FLASH_ATTN MLA prefill backend. [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:10 [deep_gemm.py:186] deep_gemm not found in site-packages, trying vendored vllm.third_party.deep_gemm [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:10 [deep_gemm.py:213] DeepGEMM PDL enabled on vllm.third_party.deep_gemm. [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:10 [deep_gemm.py:131] DeepGEMM E8M0 enabled on current platform. [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:28:10 [model.py:271] MoE renormalize is True. [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:10 [compressed_tensors_moe_w4a4_mxfp4.py:62] Using CutlassExpertsMxfp4 for MXFP4 MoE [repeated 2x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1847:1847 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 [repeated 5x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:28:11 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=59) role=S(skip_topk) [repeated 84x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) 2/-1/-1->1->0 [18] 2/-1/-1->1->0 [19] 2/-1/-1->1->0 [20] 2/-1/-1->1->0 [21] 2/-1/-1->1->0 [22] 2/-1/-1->1->0 [23] 2/-1/-1->1->0 [24] 2/-1/-1->1->0 [25] 2/-1/-1->1->0 [26] 2/-1/-1->1->-1 [27] 2/-1/-1->1->-1 [28] -1/-1/-1->1->0 [29] -1/-1/-1->1->0 [30] 2/-1/-1->1->0 [31] 2/-1/-1->1->0 (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3-58:1846:184 (VllmEngine pid=1635, ip=10.0.3.58) host-10-0-3- (VllmEngine pid=1635, ip=10.0.3.58) omm config Blocking set to 1 (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.0.self_attn.indexer.k_norm.bias to param model.layers.0.self_attn.dsa_indexer.k_index_norm.bias (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.0.self_attn.indexer.k_norm.weight to param model.layers.0.self_attn.dsa_indexer.k_index_norm.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.0.self_attn.indexer.weights_proj.weight to param model.layers.0.self_attn.dsa_indexer.wv_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.0.self_attn.indexer.wk.weight to param model.layers.0.self_attn.dsa_indexer.wk_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.0.self_attn.indexer.wq_b.weight to param model.layers.0.self_attn.dsa_indexer.wq_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.1.self_attn.indexer.k_norm.bias to param model.layers.1.self_attn.dsa_indexer.k_index_norm.bias (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.1.self_attn.indexer.k_norm.weight to param model.layers.1.self_attn.dsa_indexer.k_index_norm.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.1.self_attn.indexer.weights_proj.weight to param model.layers.1.self_attn.dsa_indexer.wv_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.1.self_attn.indexer.wk.weight to param model.layers.1.self_attn.dsa_indexer.wk_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.1.self_attn.indexer.wq_b.weight to param model.layers.1.self_attn.dsa_indexer.wq_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.2.self_attn.indexer.k_norm.bias to param model.layers.2.self_attn.dsa_indexer.k_index_norm.bias (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.2.self_attn.indexer.k_norm.weight to param model.layers.2.self_attn.dsa_indexer.k_index_norm.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.2.self_attn.indexer.weights_proj.weight to param model.layers.2.self_attn.dsa_indexer.wv_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.2.self_attn.indexer.wk.weight to param model.layers.2.self_attn.dsa_indexer.wk_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:14 [model.py:1764] rename weight model.layers.2.self_attn.indexer.wq_b.weight to param model.layers.2.self_attn.dsa_indexer.wq_index.weight (VllmEngine pid=1809, ip=10.0.3.59) 3/-1/-1->2->1 [17] 3/-1/-1->2->1 [18] 3/-1/-1->2->1 [19] 3/-1/-1->2->1 [20] 3/-1/-1->2->1 [21] 3/-1/-1->2->1 [22] 3/-1/-1->2->1 [23] 3/-1/-1->2->1 [24] 3/-1/-1->2->1 [25] 3/-1/-1->2->1 [26] 3/-1/-1->2->1 [27] 3/-1/-1->2->1 [28] 3/-1/-1->2->-1 [29] 3/-1/-1->2->-1 [30] -1/-1/-1->2->1 [31] -1/-1/-1->2->1 (VllmEngine pid=1809, ip=10.0.3.59) hos (VllmEngine pid=1809, ip=10.0.3.59) ] 0/-1/-1->3->-1 [16] -1/-1/-1->3->2 [17] -1/-1/-1->3->2 [18] -1/-1/-1->3->2 [19] -1/-1/-1->3->2 [20] -1/-1/-1->3->2 [21] -1/-1/-1->3->2 [22] -1/-1/-1->3->2 [23] -1/-1/-1->3->2 [24] -1/-1/-1->3->2 [25] -1/-1/-1->3->2 [26] 0/-1/-1->3->2 [27] 0/-1/-1->3->2 [28] 0/-1/-1->3->2 [29] 0/-1/-1->3->2 [30] 0/-1/-1->3->-1 [31] 0/-1/-1->3->-1 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:15 [model.py:1764] rename weight model.layers.3.self_attn.indexer.k_norm.bias to param model.layers.3.self_attn.dsa_indexer.k_index_norm.bias (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:15 [model.py:1764] rename weight model.layers.3.self_attn.indexer.k_norm.weight to param model.layers.3.self_attn.dsa_indexer.k_index_norm.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:15 [model.py:1764] rename weight model.layers.3.self_attn.indexer.weights_proj.weight to param model.layers.3.self_attn.dsa_indexer.wv_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:15 [model.py:1764] rename weight model.layers.3.self_attn.indexer.wk.weight to param model.layers.3.self_attn.dsa_indexer.wk_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:15 [model.py:1764] rename weight model.layers.3.self_attn.indexer.wq_b.weight to param model.layers.3.self_attn.dsa_indexer.wq_index.weight (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3- (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO Comm config Blocking set to 1 [repeated 6x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2022:2022 [3] NCCL INFO Com2/-1/-1->1->0 [18] 2/-1/-1->1->0 [19] 2/-1/-1->1->0 [20] 2/-1/-1->1->0 [21] 2/-1/-1->1->0 [22] 2/-1/-1->1->0 [23] 2/-1/-1->1->0 [24] 2/-1/-1->1->0 [25] 2/-1/-1->1->0 [26] 2/-1/-1->1->-1 [27] 2/-1/-1->1->-1 [28] -1/-1/-1->1->0 [29] -1/-1/-1->1->0 [30] 2/-1/-1->1->0 [31] 2/-1/-1->1->0 [repeated 2x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:16 [model.py:1764] rename weight model.layers.4.self_attn.indexer.k_norm.bias to param model.layers.4.self_attn.dsa_indexer.k_index_norm.bias (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:16 [model.py:1764] rename weight model.layers.4.self_attn.indexer.k_norm.weight to param model.layers.4.self_attn.dsa_indexer.k_index_norm.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:16 [model.py:1764] rename weight model.layers.4.self_attn.indexer.weights_proj.weight to param model.layers.4.self_attn.dsa_indexer.wv_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:16 [model.py:1764] rename weight model.layers.4.self_attn.indexer.wk.weight to param model.layers.4.self_attn.dsa_indexer.wk_index.weight (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:16 [model.py:1764] rename weight model.layers.4.self_attn.indexer.wq_b.weight to param model.layers.4.self_attn.dsa_indexer.wq_index.weight (VllmEngine pid=1809, ip=10.0.3.59) m config Blocking set to 1 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 13/0 : 2[2] -> 3[3] via P2P/[1] NCCL INFO Initialized NET plugin IB (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [ (VllmEngine pid=1809, ip=10.0.3.59) PXN 0 GDR 1 [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2022:2022 [3]t-10-0-3-59:2021:2021 [2] NCCL INFO Initialized NET plugin IB [repeated 7x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Assigned NET plugin IB to comm [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Using network IB [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO DMA-BUF is available on GPU device 1 [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO [Rank 1] ncclCommInitRankConfig comm 0x30a5ccd0 rank 1 nranks 4 cudaDev 1 nvmlDev 1 busId 906000 commId 0xd159f757db5ec71a - Init START [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Bootstrap timings total 0.000872 (create 0.000045, send 0.000200, recv 0.000307, ring 0.000160, delay 0.000000) [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO MNNVL busId 0x906000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 1 is 0-71. (GPU affinity = 0-71 ; CPU affinity = 0-71). [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO NVLS multicast support is available on dev 1 (NVLS_NCHANNELS 24) [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO comm 0x30a5ccd0 rank 1 nRanks 4 nNodes 1 localRanks 4 localRank 1 MNNVL 0 [repeated 8x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO Channel 31/32 : 0 1 2 3 [repeated 64x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Trees [0] 2/-1/-1->1->0 [1] 2/-1/-1->1->0 [2] 2/-1/-1->1->0 [3] 2/-1/-1->1->0 [4] 2/-1/-1->1->0 [5] 2/-1/-1->1->0 [6] 2/-1/-1->1->0 [7] 2/-1/-1->1->0 [8] 2/-1/-1->1->0 [9] 2/-1/-1->1->0 [10] 2/-1/-1->1->-1 [11] 2/-1/-1->1->-1 [12] -1/-1/-1->1->0 [13] -1/-1/-1->1->0 [14] 2/-1/-1->1->0 [15] 2/-1/-1->1->0 [16] 2/-1/-1->1->0 [17] 2/-1/-1->1->0 [18] 2/-1/-1->1->0 [19] 2/-1/-1->1->0 [20] 2/-1/-1->1->0 [21] 2/-1/-1->1->0 [22] 2/-1/-1->1->0 [23] 2/-1/-1->1->0 [24] 2/-1/-1->1->0 [25] 2/-1/-1->1->0 [26] 2/-1/-1->1->-1 [27] 2/-1/-1->1->-1 [28] -1/-1/-1->1->0 [29] -1/-1/-1->1->0 [30] 2/-1/-1->1->0 [31] 2/-1/-1->1->0 [repeated 8x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:17 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=60) role=F(with indexer) [repeated 19x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO P2P Chunksize set to 524288 [repeated 17x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 [repeated 15x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2487 [0] NCCL INFO [Proxy Service] Device 1 CPU core 28 [repeated 15x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2491 [0] NCCL INFO [Proxy Service UDS] Device 1 CPU core 30 [repeated 15x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 [repeated 15x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer [repeated 15x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Symmetric VA size=280GB [repeated 30x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO ncclCommInitRank comm 0x23f6b330 rank 1 nranks 4 cudaDev 1 nvmlDev 1 busId 906000 commId 0x33f0bccc04af17b - Init COMPLETE [repeated 6x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Init timings - ncclCommInitRank: rank 1 nranks 4 total 0.71 (kernels 0.00, alloc 0.00, bootstrap 0.00, allgathers 0.00, topo 0.58, graphs 0.01, connections 0.11, rest 0.00) [repeated 6x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Channel 13/0 : 1[1] -> 2[2] via P2P/CUMEM [repeated 313x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO ncclCommInitRankConfig comm 0x30a5ccd0 rank 1 nranks 4 cudaDev 1 nvmlDev 1 busId 906000 commId 0xd159f757db5ec71a - Init COMPLETE [repeated 6x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Init timings - ncclCommInitRankConfig: rank 1 nranks 4 total 1.03 (kernels 0.00, alloc 0.00, bootstrap 0.00, allgathers 0.00, topo 0.49, graphs 0.01, connections 0.52, rest 0.00) [repeated 6x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:16 [cuda.py:492] Using FLASHMLA_SPARSE attention backend out of potential backends: ['FLASHMLA_SPARSE']. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:16 [selector.py:192] Using FLASH_ATTN MLA prefill backend. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:16 [deep_gemm.py:186] deep_gemm not found in site-packages, trying vendored vllm.third_party.deep_gemm (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:16 [deep_gemm.py:213] DeepGEMM PDL enabled on vllm.third_party.deep_gemm. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:16 [deep_gemm.py:131] DeepGEMM E8M0 enabled on current platform. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) WARNING 10-18 13:28:16 [model.py:271] MoE renormalize is True. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:16 [compressed_tensors_moe_w4a4_mxfp4.py:62] Using CutlassExpertsMxfp4 for MXFP4 MoE (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:28:17 [model.py:1120] Kimi K3 Flash DSA index share: index_topk_freq=4, index_skip_topk_offset=5; this layer (layer_id=59) role=S(skip_topk) [repeated 42x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:19 [model.py:1764] rename weight model.layers.8.self_attn.indexer.wq_b.weight to param model.layers.8.self_attn.dsa_indexer.wq_index.weight [repeated 45x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) 2/-1/-1->1->0 [18] 2/-1/-1->1->0 [19] 2/-1/-1->1->0 [20] 2/-1/-1->1->0 [21] 2/-1/-1->1->0 [22] 2/-1/-1->1->0 [23] 2/-1/-1->1->0 [24] 2/-1/-1->1->0 [25] 2/-1/-1->1->0 [26] 2/-1/-1->1->-1 [27] 2/-1/-1->1->-1 [28] -1/-1/-1->1->0 [29] -1/-1/-1->1->0 [30] 2/-1/-1->1->0 [31] 2/-1/-1->1->0 (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1962:1965] 0/-1/-1->3->-1 [16] -1/-1/-1->3->2 [17] -1/-1/-1->3->2 [18] -1/-1/-1->3->2 [19] -1/-1/-1->3->2 [20] -1/-1/-1->3->2 [21] -1/-1/-1->3->2 [22] -1/-1/-1->3->2 [23] -1/-1/-1->3->2 [24] -1/-1/-1->3->2 [25] -1/-1/-1->3->2 [26] 0/-1/-1->3->2 [27] 0/-1/-1->3->2 [28] 0/-1/-1->3->2 [29] 0/-1/-1->3->2 [30] 0/-1/-1->3->-1 [31] 0/-1/-1->3->-1 (VllmEngine pid=1751, ip=10.0.3.63) host- (VllmEngine pid=1751, ip=10.0.3.63) g PXN 0 GDR 1 (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3 (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1963:1963 [2] NCCL INFO Channel 13/0 : 2[2] -> 3[3] via P2P/CUMm config Blocking set to 1 (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] N (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Comm config Blocking set to 1 [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Com/-1/-1->2->1 [17] 3/-1/-1->2->1 [18] 3/-1/-1->2->1 [19] 3/-1/-1->2->1 [20] 3/-1/-1->2->1 [21] 3/-1/-1->2->1 [22] 3/-1/-1->2->1 [23] 3/-1/-1->2->1 [24] 3/-1/-1->2->1 [25] 3/-1/-1->2->1 [26] 3/-1/-1->2->1 [27] 3/-1/-1->2->1 [28] 3/-1/-1->2->-1 [29] 3/-1/-1->2->-1 [30] -1/-1/-1->2->1 [31] -1/-1/-1->2->1 (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Initialized NET plugin IB [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Assigned NET plugin IB to comm [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Assigned RMA plugin GIN_IB_PROXY to comm [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Using network IB [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO DMA-BUF is available on GPU device 3 [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO [Rank 3] ncclCommInitRankConfig comm 0x29951f30 rank 3 nranks 4 cudaDev 3 nvmlDev 3 busId 1906000 commId 0x813d4438a370ffe5 - Init START [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Bootstrap timings total 0.004286 (create 0.000063, send 0.000333, recv 0.000298, ring 0.000177, delay 0.000001) [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO MNNVL busId 0x1906000 fabric UUID 74a91164b7137e3.264d6aedfa0492a2 cliqueId 0x7ffe state 3 healthMask 0x1aa [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 3 is 72-143. (GPU affinity = 72-143 ; CPU affinity = 72-143). [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO NVLS multicast support is available on dev 3 (NVLS_NCHANNELS 24) [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO comm 0x29951f30 rank 3 nRanks 4 nNodes 1 localRanks 4 localRank 3 MNNVL 0 [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1961:1961 [0] NCCL INFO Channel 31/32 : 0 1 2 3 [repeated 32x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Trees [0] -1/-1/-1->3->2 [1] -1/-1/-1->3->2 [2] -1/-1/-1->3->2 [3] -1/-1/-1->3->2 [4] -1/-1/-1->3->2 [5] -1/-1/-1->3->2 [6] -1/-1/-1->3->2 [7] -1/-1/-1->3->2 [8] -1/-1/-1->3->2 [9] -1/-1/-1->3->2 [10] 0/-1/-1->3->2 [11] 0/-1/-1->3->2 [12] 0/-1/-1->3->2 [13] 0/-1/-1->3->2 [14] 0/-1/-1->3->-1 [15] 0/-1/-1->3->-1 [16] -1/-1/-1->3->2 [17] -1/-1/-1->3->2 [18] -1/-1/-1->3->2 [19] -1/-1/-1->3->2 [20] -1/-1/-1->3->2 [21] -1/-1/-1->3->2 [22] -1/-1/-1->3->2 [23] -1/-1/-1->3->2 [24] -1/-1/-1->3->2 [25] -1/-1/-1->3->2 [26] 0/-1/-1->3->2 [27] 0/-1/-1->3->2 [28] 0/-1/-1->3->2 [29] 0/-1/-1->3->2 [30] 0/-1/-1->3->-1 [31] 0/-1/-1->3->-1 [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO P2P Chunksize set to 524288 [repeated 7x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 [repeated 6x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:2427 [0] NCCL INFO [Proxy Service] Device 3 CPU core 86 [repeated 6x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:2431 [0] NCCL INFO [Proxy Service UDS] Device 3 CPU core 87 [repeated 6x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 [repeated 6x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer [repeated 6x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Symmetric VA size=280GB [repeated 12x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1963:1963 [2] NCCL INFO ncclCommInitRank comm 0x2ccd6eb0 rank 2 nranks 4 cudaDev 2 nvmlDev 2 busId 1806000 commId 0x9999f29976d49fd0 - Init COMPLETE [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1963:1963 [2] NCCL INFO Init timings - ncclCommInitRank: rank 2 nranks 4 total 1.36 (kernels 0.00, alloc 0.00, bootstrap 0.00, allgathers 0.00, topo 0.81, graphs 0.01, connections 0.54, rest 0.00) [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Channel 12/0 : 3[3] -> 0[0] via P2P/CUMEM [repeated 136x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO ncclCommInitRankConfig comm 0x29951f30 rank 3 nranks 4 cudaDev 3 nvmlDev 3 busId 1906000 commId 0x813d4438a370ffe5 - Init COMPLETE [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1964:1964 [3] NCCL INFO Init timings - ncclCommInitRankConfig: rank 3 nranks 4 total 0.87 (kernels 0.00, alloc 0.01, bootstrap 0.00, allgathers 0.00, topo 0.52, graphs 0.01, connections 0.32, rest 0.00) [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) host-10-0-3-63:1963:1963 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 [repeated 3x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) WARNING 10-18 13:28:24 [model.py:1764] rename weight model.layers.12.self_attn.indexer.wq_b.weight to param model.layers.12.self_attn.dsa_indexer.wq_index.weight [repeated 45x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:28:30 [model.py:1764] rename weight model.layers.16.self_attn.indexer.wq_b.weight to param model.layers.16.self_attn.dsa_indexer.wq_index.weight [repeated 40x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:28:35 [model.py:1764] rename weight model.layers.28.self_attn.indexer.wq_b.weight to param model.layers.28.self_attn.dsa_indexer.wq_index.weight [repeated 35x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) WARNING 10-18 13:28:41 [model.py:1764] rename weight model.layers.24.self_attn.indexer.wq_b.weight to param model.layers.24.self_attn.dsa_indexer.wq_index.weight [repeated 30x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) WARNING 10-18 13:28:46 [model.py:1764] rename weight model.layers.40.self_attn.indexer.wq_b.weight to param model.layers.40.self_attn.dsa_indexer.wq_index.weight [repeated 40x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:28:52 [model.py:1764] rename weight model.layers.44.self_attn.indexer.wq_b.weight to param model.layers.44.self_attn.dsa_indexer.wq_index.weight [repeated 35x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:28:58 [model.py:1764] rename weight model.layers.52.self_attn.indexer.wq_b.weight to param model.layers.52.self_attn.dsa_indexer.wq_index.weight [repeated 40x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:29:00 [default_loader.py:430] Loading weights took 52.10 seconds (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:29:00 [mxfp4.py:1901] Using MoEPrepareAndFinalizeNoDPEPModular (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:29:00 [mxfp4.py:1902] Using CutlassExpertsMxfp4 (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP2 pid=2079) INFO 10-18 13:29:03 [gpu_model_runner.py:5668] Using auxiliary layers from speculative config: (2, 30, 58, 61) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP3 pid=2080) INFO 10-18 13:29:03 [gpu_model_runner.py:5668] Using auxiliary layers from speculative config: (2, 30, 58, 61) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:29:03 [gpu_model_runner.py:5515] Loading drafter model... (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:29:03 [gpu_model_runner.py:5668] Using auxiliary layers from speculative config: (2, 30, 58, 61) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP1 pid=2078) INFO 10-18 13:29:03 [gpu_model_runner.py:5668] Using auxiliary layers from speculative config: (2, 30, 58, 61) (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:29:04 [model.py:1764] rename weight model.layers.60.self_attn.indexer.wq_b.weight to param model.layers.60.self_attn.dsa_indexer.wq_index.weight [repeated 35x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:29:04 [gpu_model_runner.py:5583] Model loading took 134.73 GiB memory and 62.300860 seconds (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:29:04 [default_loader.py:430] Loading weights took 52.61 seconds [repeated 3x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:29:04 [mxfp4.py:1901] Using MoEPrepareAndFinalizeNoDPEPModular [repeated 2x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:29:04 [mxfp4.py:1902] Using CutlassExpertsMxfp4 [repeated 2x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (EngineCore pid=2062) INFO 10-18 13:29:06 [utils.py:304] Using LBNHC KV cache layout. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:29:06 [gpu_model_runner.py:6658] Encoder cache will be initialized with a budget of 32768 tokens, and profiled with 1 image items of the maximum feature size.(VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) JIT kernel warmup (1 compile keys): 0%| | 0/1 [00:00` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP3 pid=1964) exe_args, func_types, adapted_args = self.generate_mlir_function_types( [repeated 4x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) JIT kernel warmup (1 compile keys): 100%|██████████| 1/1 [00:00<00:00, 10.86kernel/s, ComputeSlotMappingKernel (1 keys)] (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP2 pid=2079) 2026-10-18 13:29:48,624 - INFO - autotuner.py:829 - flashinfer.jit: [Autotuner]: Autotuning process starts ... (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) 2026-10-18 13:29:48,629 - INFO - autotuner.py:829 - flashinfer.jit: [Autotuner]: Autotuning process starts ... (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP3 pid=2080) 2026-10-18 13:29:48,629 - INFO - autotuner.py:829 - flashinfer.jit: [Autotuner]: Autotuning process starts ... (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP1 pid=2078) 2026-10-18 13:29:48,629 - INFO - autotuner.py:829 - flashinfer.jit: [Autotuner]: Autotuning process starts ... (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) JIT kernel warmup (1 compile keys): 0%| | 0/1 [00:00` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP2 pid=2021) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP3 pid=2022) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_017439'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP3 pid=2022) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP1 pid=2020) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_017439'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP1 pid=2020) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_017439'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.377518 2019 client_service.cpp:423] client_id=16809850858888053513-14626931832072339627 (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.377537 2019 client_service.cpp:425] client_id=16809850858888053513-14626931832072339627, host_id=10.0.3.59 (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) [2026-10-18 13:30:19,367] store.py:78 INFO Connecting to Mooncake master at 10.0.3.61:51661, metadata server at http://10.0.3.61:8653/metadata (VllmEngine pid=1809, ip=10.0.3.59) WARNING: Logging before InitGoogleLogging() is written to STDERR (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.375849 2019 client_metric.cpp:102] Client metrics enabled (default enabled) (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.375877 2019 client_metric.cpp:103] Client bandwidth summary enabled via MC_STORE_CLIENT_METRIC_BANDWIDTH (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.377540 2019 client_service.cpp:437] Client metrics enabled but reporting disabled (interval=0) (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.379256 2019 client_service.cpp:1077] Storage root directory is not set. persisting data is disabled. (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.379310 2019 transfer_engine_impl.cpp:1046] Metrics reporting is disabled (set MC_TE_METRIC=1 to enable) (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.379331 2019 transfer_engine_impl.cpp:116] Transfer Engine parseHostNameWithPort. server_name: 10.0.3.59 port: 12708 (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.380533 2019 transfer_metadata_plugin.cpp:1298] Found active interface eth0 with IP 10.0.3.59 (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.380549 2019 transfer_metadata_plugin.cpp:1287] Skipping interface nodelocaldns (not UP or not RUNNING) (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.380636 2019 transfer_metadata_plugin.cpp:1287] Skipping interface kube-ipvs0 (not UP or not RUNNING) [repeated 44x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.380638 2019 transfer_metadata_plugin.cpp:1298] Found active interface vxlan.calico with IP 10.233.103.128 (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.380661 2019 transfer_engine_impl.cpp:183] Transfer Engine RPC using new RPC mapping, listening on 10.0.3.59:15088 (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.381587 2019 client_service.cpp:872] Transfer engine auto discovery is disabled for protocol: rdma (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.381599 2019 client_service.cpp:884] Using specified RDMA devices: mlx5_bond_0 (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.384914 2019 topology.cpp:127] Device mlx5_bond_0 port 1 is available (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.390157 2019 client_service.cpp:894] Topology discovery complete with specified devices. Found 1 HCAs (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.390196 2019 rdma_transport.cpp:159] [RDMA] Relaxed ordering is supported on this host; IBV_ACCESS_RELAXED_ORDERING will be requested for registered memory regions. (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.390210 2019 rdma_context.cpp:207] Using SIEVE endpoint store (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.392702 2019 rdma_context.cpp:1383] Using user-specified GID index: 3 on mlx5_bond_0/1 (with network device) (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.393532 2019 rdma_context.cpp:290] RDMA device: mlx5_bond_0, LID: 0, GID: (GID_Index 3) fd:03:45:c3:00:10:40:e6:10:00:e0:00:00:00:00:01 (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.394656 2019 transfer_task.cpp:965] MC_STORE_MEMCPY not set, auto-detected: non-TCP transport available, memcpy disabled (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.394677 2019 real_client.cpp:851] Successfully created client on port 12708 after 1 attempt(s) (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.394961 2019 real_client.cpp:884] Registering local memory: 17179869184 bytes (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) [2026-10-18 13:30:19,712] store.py:78 INFO Connecting to Mooncake master at 10.0.3.61:51661, metadata server at http://10.0.3.61:8653/metadata (VllmEngine pid=1635, ip=10.0.3.58) WARNING: Logging before InitGoogleLogging() is written to STDERR (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.720280 1845 client_metric.cpp:102] Client metrics enabled (default enabled) (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.720315 1845 client_metric.cpp:103] Client bandwidth summary enabled via MC_STORE_CLIENT_METRIC_BANDWIDTH (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.722033 1845 client_service.cpp:423] client_id=15658426857344834720-11516887181607724222 (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.722051 1845 client_service.cpp:425] client_id=15658426857344834720-11516887181607724222, host_id=10.0.3.58 (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.722056 1845 client_service.cpp:437] Client metrics enabled but reporting disabled (interval=0) (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.723616 1845 client_service.cpp:1077] Storage root directory is not set. persisting data is disabled. (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.723675 1845 transfer_engine_impl.cpp:1046] Metrics reporting is disabled (set MC_TE_METRIC=1 to enable) (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.723695 1845 transfer_engine_impl.cpp:116] Transfer Engine parseHostNameWithPort. server_name: 10.0.3.58 port: 12498 (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.724805 1845 transfer_metadata_plugin.cpp:1298] Found active interface eth0 with IP 10.0.3.58 (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.724822 1845 transfer_metadata_plugin.cpp:1287] Skipping interface nodelocaldns (not UP or not RUNNING) (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.724920 1845 transfer_metadata_plugin.cpp:1298] Found active interface vxlan.calico with IP 10.233.66.192 (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.724939 1845 transfer_engine_impl.cpp:183] Transfer Engine RPC using new RPC mapping, listening on 10.0.3.58:16656 (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.725798 1845 client_service.cpp:872] Transfer engine auto discovery is disabled for protocol: rdma (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.725822 1845 client_service.cpp:884] Using specified RDMA devices: mlx5_bond_0 (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.729101 1845 topology.cpp:127] Device mlx5_bond_0 port 1 is available (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.738656 1845 client_service.cpp:894] Topology discovery complete with specified devices. Found 1 HCAs (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.738704 1845 rdma_transport.cpp:159] [RDMA] Relaxed ordering is supported on this host; IBV_ACCESS_RELAXED_ORDERING will be requested for registered memory regions. (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.738718 1845 rdma_context.cpp:207] Using SIEVE endpoint store (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.741053 1845 rdma_context.cpp:1383] Using user-specified GID index: 3 on mlx5_bond_0/1 (with network device) (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.741811 1845 rdma_context.cpp:290] RDMA device: mlx5_bond_0, LID: 0, GID: (GID_Index 3) fd:03:45:c3:00:10:40:e8:10:00:e0:00:00:00:00:01 (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.742806 1845 transfer_task.cpp:965] MC_STORE_MEMCPY not set, auto-detected: non-TCP transport available, memcpy disabled (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.742828 1845 real_client.cpp:851] Successfully created client on port 12498 after 1 attempt(s) (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.743095 1845 real_client.cpp:884] Registering local memory: 17179869184 bytes (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.880681 2019 real_client.cpp:1026] Mounting segment: 137438953472 bytes, 137438953472 of 137438953472 (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:19.880744 2019 registered_pinned_memory.cpp:98] Store segment pinned memory is disabled, max_bytes=0 (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) [2026-10-18 13:30:22,942] store.py:287 INFO Hard pin enabled for batch_put_from (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP3 pid=1848) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_017439'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. [repeated 4x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP3 pid=1848) exe_args, func_types, adapted_args = self.generate_mlir_function_types( [repeated 4x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) [2026-10-18 13:30:24,705] buffers.py:93 INFO Initialized host buffer pool: 4 x 7.0GB (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:19.724918 1845 transfer_metadata_plugin.cpp:1287] Skipping interface kube-ipvs0 (not UP or not RUNNING) [repeated 44x across cluster] (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) [2026-10-18 13:30:24,762] store.py:116 INFO Async put manager created (pool_size=4) (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) [2026-10-18 13:30:24,763] store.py:125 INFO DtoH copy stream created on cuda:0 (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) [2026-10-18 13:30:24,764] store.py:129 INFO Mooncake Store client initialized (protocol=rdma, device=mlx5_bond_0, gpu_direct=False, pool_size=4) (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:24.945694 3053 rdma_endpoint.cpp:671] Received RDMA ready ACK. (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:20.006181 1845 real_client.cpp:1026] Mounting segment: 137438953472 bytes, 137438953472 of 137438953472 (VllmEngine pid=1635, ip=10.0.3.58) I1018 13:30:20.006244 1845 registered_pinned_memory.cpp:98] Store segment pinned memory is disabled, max_bytes=0 (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) [2026-10-18 13:30:28,163] store.py:78 INFO Connecting to Mooncake master at 10.0.3.61:51661, metadata server at http://10.0.3.61:8653/metadata (VllmEngine pid=1751, ip=10.0.3.63) WARNING: Logging before InitGoogleLogging() is written to STDERR (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.169168 1961 client_metric.cpp:102] Client metrics enabled (default enabled) (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.169196 1961 client_metric.cpp:103] Client bandwidth summary enabled via MC_STORE_CLIENT_METRIC_BANDWIDTH (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.170588 1961 client_service.cpp:423] client_id=13204777775380948812-1541376526236631953 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.170608 1961 client_service.cpp:425] client_id=13204777775380948812-1541376526236631953, host_id=10.0.3.63 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.170611 1961 client_service.cpp:437] Client metrics enabled but reporting disabled (interval=0) (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.172016 1961 client_service.cpp:1077] Storage root directory is not set. persisting data is disabled. (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.172075 1961 transfer_engine_impl.cpp:1046] Metrics reporting is disabled (set MC_TE_METRIC=1 to enable) (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.172091 1961 transfer_engine_impl.cpp:116] Transfer Engine parseHostNameWithPort. server_name: 10.0.3.63 port: 12472 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.173341 1961 transfer_metadata_plugin.cpp:1298] Found active interface eth0 with IP 10.0.3.63 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.173358 1961 transfer_metadata_plugin.cpp:1287] Skipping interface nodelocaldns (not UP or not RUNNING) (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.173439 1961 transfer_metadata_plugin.cpp:1298] Found active interface vxlan.calico with IP 10.233.72.128 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.173460 1961 transfer_engine_impl.cpp:183] Transfer Engine RPC using new RPC mapping, listening on 10.0.3.63:15747 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.174185 1961 client_service.cpp:872] Transfer engine auto discovery is disabled for protocol: rdma (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.174196 1961 client_service.cpp:884] Using specified RDMA devices: mlx5_bond_0 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.177361 1961 topology.cpp:127] Device mlx5_bond_0 port 1 is available (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) [2026-10-18 13:30:23,129] store.py:287 INFO Hard pin enabled for batch_put_from (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_017439'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. [repeated 4x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) exe_args, func_types, adapted_args = self.generate_mlir_function_types( [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.200071 1961 client_service.cpp:894] Topology discovery complete with specified devices. Found 1 HCAs (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.200112 1961 rdma_transport.cpp:159] [RDMA] Relaxed ordering is supported on this host; IBV_ACCESS_RELAXED_ORDERING will be requested for registered memory regions. (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.200125 1961 rdma_context.cpp:207] Using SIEVE endpoint store (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.202562 1961 rdma_context.cpp:1383] Using user-specified GID index: 3 on mlx5_bond_0/1 (with network device) (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.203402 1961 rdma_context.cpp:290] RDMA device: mlx5_bond_0, LID: 0, GID: (GID_Index 3) fd:03:45:c3:00:10:40:e7:10:00:e0:00:00:00:00:01 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.204425 1961 transfer_task.cpp:965] MC_STORE_MEMCPY not set, auto-detected: non-TCP transport available, memcpy disabled (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.204447 1961 real_client.cpp:851] Successfully created client on port 12472 after 1 attempt(s) (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.204772 1961 real_client.cpp:884] Registering local memory: 17179869184 bytes (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.651924 1961 real_client.cpp:1026] Mounting segment: 137438953472 bytes, 137438953472 of 137438953472 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:30:28.651990 1961 registered_pinned_memory.cpp:98] Store segment pinned memory is disabled, max_bytes=0 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.621575 2077 client_service.cpp:423] client_id=7011761144001313754-7609974218382941339 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.621594 2077 client_service.cpp:425] client_id=7011761144001313754-7609974218382941339, host_id=10.0.3.60 (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) [2026-10-18 13:30:24,936] buffers.py:93 INFO Initialized host buffer pool: 4 x 7.0GB (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.624437 2077 transfer_metadata_plugin.cpp:1287] Skipping interface nerdctl0 (not UP or not RUNNING) [repeated 89x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) [2026-10-18 13:30:24,999] store.py:116 INFO Async put manager created (pool_size=4) (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) [2026-10-18 13:30:25,999] store.py:125 INFO DtoH copy stream created on cuda:0 (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) [2026-10-18 13:30:25,000] store.py:129 INFO Mooncake Store client initialized (protocol=rdma, device=mlx5_bond_0, gpu_direct=False, pool_size=4) (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) [2026-10-18 13:30:31,735] store.py:287 INFO Hard pin enabled for batch_put_from (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) [2026-10-18 13:30:35,312] buffers.py:93 INFO Initialized host buffer pool: 4 x 7.0GB (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) [2026-10-18 13:30:30,612] store.py:78 INFO Connecting to Mooncake master at 10.0.3.61:51661, metadata server at http://10.0.3.61:8653/metadata (VllmEngine pid=1867, ip=10.0.3.60) WARNING: Logging before InitGoogleLogging() is written to STDERR (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.619742 2077 client_metric.cpp:102] Client metrics enabled (default enabled) (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.619779 2077 client_metric.cpp:103] Client bandwidth summary enabled via MC_STORE_CLIENT_METRIC_BANDWIDTH (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.621599 2077 client_service.cpp:437] Client metrics enabled but reporting disabled (interval=0) (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.623003 2077 client_service.cpp:1077] Storage root directory is not set. persisting data is disabled. (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.623060 2077 transfer_engine_impl.cpp:1046] Metrics reporting is disabled (set MC_TE_METRIC=1 to enable) (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.623082 2077 transfer_engine_impl.cpp:116] Transfer Engine parseHostNameWithPort. server_name: 10.0.3.60 port: 13601 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.624315 2077 transfer_metadata_plugin.cpp:1298] Found active interface eth0 with IP 10.0.3.60 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.624330 2077 transfer_metadata_plugin.cpp:1287] Skipping interface nodelocaldns (not UP or not RUNNING) (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.624435 2077 transfer_metadata_plugin.cpp:1298] Found active interface vxlan.calico with IP 10.233.122.192 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.624459 2077 transfer_engine_impl.cpp:183] Transfer Engine RPC using new RPC mapping, listening on 10.0.3.60:16822 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.625355 2077 client_service.cpp:872] Transfer engine auto discovery is disabled for protocol: rdma (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.625367 2077 client_service.cpp:884] Using specified RDMA devices: mlx5_bond_0 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.628368 2077 topology.cpp:127] Device mlx5_bond_0 port 1 is available (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.637722 2077 client_service.cpp:894] Topology discovery complete with specified devices. Found 1 HCAs (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.637764 2077 rdma_transport.cpp:159] [RDMA] Relaxed ordering is supported on this host; IBV_ACCESS_RELAXED_ORDERING will be requested for registered memory regions. (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.637781 2077 rdma_context.cpp:207] Using SIEVE endpoint store (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.640064 2077 rdma_context.cpp:1383] Using user-specified GID index: 3 on mlx5_bond_0/1 (with network device) (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.640885 2077 rdma_context.cpp:290] RDMA device: mlx5_bond_0, LID: 0, GID: (GID_Index 3) fd:03:45:c3:00:10:40:ea:10:00:e0:00:00:00:00:01 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.641885 2077 transfer_task.cpp:965] MC_STORE_MEMCPY not set, auto-detected: non-TCP transport available, memcpy disabled (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.641903 2077 real_client.cpp:851] Successfully created client on port 13601 after 1 attempt(s) (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:30.642171 2077 real_client.cpp:884] Registering local memory: 17179869184 bytes (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:31.078325 2077 real_client.cpp:1026] Mounting segment: 137438953472 bytes, 137438953472 of 137438953472 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:30:31.078406 2077 registered_pinned_memory.cpp:98] Store segment pinned memory is disabled, max_bytes=0 (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) [2026-10-18 13:30:35,378] store.py:116 INFO Async put manager created (pool_size=4) (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) [2026-10-18 13:30:35,378] store.py:125 INFO DtoH copy stream created on cuda:0 (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) [2026-10-18 13:30:35,379] store.py:129 INFO Mooncake Store client initialized (protocol=rdma, device=mlx5_bond_0, gpu_direct=False, pool_size=4) (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:35.396201 3053 rdma_endpoint.cpp:671] Received RDMA ready ACK. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) [2026-10-18 13:30:35,338] store.py:287 INFO Hard pin enabled for batch_put_from (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP1 pid=1962) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_017439'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP1 pid=1962) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_017439'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP3 pid=1964) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_017439'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP2 pid=1963) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_None_017439'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP3 pid=1964) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP2 pid=1963) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) [2026-10-18 13:30:37,014] buffers.py:93 INFO Initialized host buffer pool: 4 x 7.0GB (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) [2026-10-18 13:30:37,077] store.py:116 INFO Async put manager created (pool_size=4) (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) [2026-10-18 13:30:37,078] store.py:125 INFO DtoH copy stream created on cuda:0 (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) [2026-10-18 13:30:37,079] store.py:129 INFO Mooncake Store client initialized (protocol=rdma, device=mlx5_bond_0, gpu_direct=False, pool_size=4) (VllmEngine pid=1809, ip=10.0.3.59) I1018 13:30:37.189070 3053 rdma_endpoint.cpp:671] Received RDMA ready ACK. [repeated 4x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:31:24.382970 3085 rdma_endpoint.cpp:671] Received RDMA ready ACK. [repeated 29x across cluster] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:31:28,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:31:33,959] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:31:24.391717 3265 rdma_endpoint.cpp:671] Received RDMA ready ACK. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:31:38,992] inference_manager.py:398 INFO Still waiting for pool capacity (waited 10.1s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:31:44,031] inference_manager.py:398 INFO Still waiting for pool capacity (waited 15.1s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:31:49,069] inference_manager.py:398 INFO Still waiting for pool capacity (waited 20.1s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:31:54,107] inference_manager.py:398 INFO Still waiting for pool capacity (waited 25.2s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:31:59,140] inference_manager.py:398 INFO Still waiting for pool capacity (waited 30.2s): pool_size=64/64 (VllmEngine pid=1751, ip=10.0.3.63) I1018 13:32:02.486405 3085 rdma_endpoint.cpp:671] Received RDMA ready ACK. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:04,176] inference_manager.py:398 INFO Still waiting for pool capacity (waited 35.3s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:09,214] inference_manager.py:398 INFO Still waiting for pool capacity (waited 40.3s): pool_size=64/64 (VllmEngine pid=1867, ip=10.0.3.60) I1018 13:32:07.359006 3265 rdma_endpoint.cpp:671] Received RDMA ready ACK. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:14,249] inference_manager.py:398 INFO Still waiting for pool capacity (waited 45.3s): pool_size=64/64 [2026-10-18 13:32:17,712] loop.py:380 INFO TIMING step=1: step=53.738s data=1.289s compute=53.574s [fwd=44.507s bwd=4.608s opt=4.450s] dispatch=70.051s [2026-10-18 13:32:17,712] loop.py:45 INFO TRAIN_STEP step=1 loss=13.360889 Training: 0%| | 1/18018 [02:03<619:33:41, 123.80s/step, loss=13.361, acc=0.000, acc_len=0.00, thru=0.0, I=0.0, T=2.4, wait=70.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:32:17,707] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=0: forward=44507.1ms backward=4608.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:17,772] inference_manager.py:408 INFO Pool capacity available after 48.9s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:26,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:32:28,918] loop.py:380 INFO TIMING step=2: step=6.491s data=3.300s compute=6.435s [fwd=4.515s bwd=1.908s opt=0.003s] dispatch=4.708s [2026-10-18 13:32:28,918] loop.py:45 INFO TRAIN_STEP step=2 loss=13.334360 Training: 0%| | 2/18018 [02:15<288:05:25, 57.57s/step, loss=13.334, acc=0.000, acc_len=0.00, thru=10.5, I=13.1, T=19.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:28,947] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 13:32:28,914] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1: forward=4515.4ms backward=1908.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:34,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:32:37,145] loop.py:380 INFO TIMING step=3: step=5.803s data=3.458s compute=5.767s [fwd=4.556s bwd=1.201s opt=0.003s] dispatch=2.416s [2026-10-18 13:32:37,146] loop.py:45 INFO TRAIN_STEP step=3 loss=13.324416 Training: 0%| | 3/18018 [02:23<175:20:01, 35.04s/step, loss=13.324, acc=0.000, acc_len=0.00, thru=15.7, I=24.5, T=22.1, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:32:37,141] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2: forward=4555.9ms backward=1200.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:37,592] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:43,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:32:45,562] loop.py:380 INFO TIMING step=4: step=4.770s data=1.471s compute=4.735s [fwd=3.575s bwd=1.150s opt=0.003s] dispatch=3.639s [2026-10-18 13:32:45,562] loop.py:45 INFO TRAIN_STEP step=4 loss=13.341072 (TrainerActor pid=9310) [2026-10-18 13:32:45,558] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3: forward=3575.0ms backward=1149.9ms Training: 0%| | 4/18018 [02:31<122:44:06, 24.53s/step, loss=13.341, acc=0.000, acc_len=0.00, thru=16.1, I=19.1, T=26.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:45,961] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:55,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:32:57,023] loop.py:380 INFO TIMING step=5: step=5.510s data=2.154s compute=5.479s [fwd=4.118s bwd=1.351s opt=0.003s] dispatch=5.942s [2026-10-18 13:32:57,023] loop.py:45 INFO TRAIN_STEP step=5 loss=13.326801 Training: 0%| | 5/18018 [02:43<99:08:54, 19.82s/step, loss=13.327, acc=0.000, acc_len=0.00, thru=11.4, I=12.2, T=23.2, wait=5.9s, pool=64, epoch=1/2] (TrainerActor pid=9310) [2026-10-18 13:32:57,018] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4: forward=4118.1ms backward=1350.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:32:57,067] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 13:33:04,935] loop.py:45 INFO TRAIN_STEP step=6 loss=13.356831 (TrainerActor pid=9310) [2026-10-18 13:33:04,931] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5: forward=3226.8ms backward=1389.1ms Training: 0%| | 6/18018 [02:51<78:54:53, 15.77s/step, loss=13.357, acc=0.000, acc_len=0.00, thru=13.3, I=15.3, T=27.4, wait=3.2s, pool=52, epoch=1/2](VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP1 pid=1962) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP1 pid=1962) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP2 pid=1963) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP3 pid=1964) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP2 pid=1963) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP3 pid=1964) exe_args, func_types, adapted_args = self.generate_mlir_function_types( [2026-10-18 13:33:14,101] loop.py:45 INFO TRAIN_STEP step=7 loss=13.315792 Training: 0%| | 7/18018 [03:00<68:05:09, 13.61s/step, loss=13.316, acc=0.000, acc_len=0.00, thru=11.7, I=16.0, T=25.1, wait=4.0s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:33:22,860] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:33:23,938] loop.py:45 INFO TRAIN_STEP step=8 loss=13.315594 Training: 0%| | 8/18018 [03:10<62:04:30, 12.41s/step, loss=13.316, acc=0.000, acc_len=0.00, thru=17.0, I=14.1, T=31.8, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:33:24,366] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:33:33,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:33:36,767] loop.py:45 INFO TRAIN_STEP step=9 loss=13.265244 Training: 0%| | 9/18018 [03:22<62:43:41, 12.54s/step, loss=13.265, acc=0.000, acc_len=0.00, thru=12.5, I=12.7, T=19.5, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:33:36,998] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 [2026-10-18 13:33:44,595] loop.py:380 INFO TIMING step=10: step=3.954s data=1.731s compute=3.907s [fwd=2.786s bwd=1.110s opt=0.003s] dispatch=3.867s [2026-10-18 13:33:44,596] loop.py:45 INFO TRAIN_STEP step=10 loss=13.270872 Training: 0%| | 10/18018 [03:30<55:27:00, 11.09s/step, loss=13.271, acc=0.000, acc_len=0.00, thru=14.8, I=16.8, T=32.4, wait=3.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:33:51,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:33:52,690] loop.py:45 INFO TRAIN_STEP step=11 loss=13.219999 Training: 0%| | 11/18018 [03:38<50:52:11, 10.17s/step, loss=13.220, acc=0.000, acc_len=0.00, thru=16.9, I=17.9, T=31.8, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:33:52,686] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10: forward=2879.3ms backward=1107.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:33:52,915] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:33:59,066] loop.py:45 INFO TRAIN_STEP step=12 loss=12.768967 Training: 0%| | 12/18018 [03:45<45:05:39, 9.02s/step, loss=12.769, acc=0.000, acc_len=0.00, thru=18.0, I=22.0, T=34.2, wait=2.6s, pool=56, epoch=1/2][2026-10-18 13:34:05,885] loop.py:45 INFO TRAIN_STEP step=13 loss=12.960320 Training: 0%| | 13/18018 [03:51<41:45:51, 8.35s/step, loss=12.960, acc=0.000, acc_len=0.00, thru=17.6, I=17.8, T=32.2, wait=2.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:34:12,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:34:14,748] loop.py:45 INFO TRAIN_STEP step=14 loss=12.892752 Training: 0%| | 14/18018 [04:00<42:32:04, 8.51s/step, loss=12.893, acc=0.000, acc_len=0.00, thru=16.1, I=19.5, T=23.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:34:15,085] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP3 pid=2080) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP2 pid=2079) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP3 pid=2080) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP2 pid=2079) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP1 pid=2078) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP1 pid=2078) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) exe_args, func_types, adapted_args = self.generate_mlir_function_types( [2026-10-18 13:34:21,849] loop.py:380 INFO TIMING step=15: step=4.279s data=2.108s compute=4.244s [fwd=3.113s bwd=1.117s opt=0.003s] dispatch=2.814s [2026-10-18 13:34:21,849] loop.py:45 INFO TRAIN_STEP step=15 loss=12.733692 Training: 0%| | 15/18018 [04:07<40:24:58, 8.08s/step, loss=12.734, acc=0.000, acc_len=0.00, thru=10.1, I=23.9, T=29.9, wait=2.8s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:34:33,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:34:34,133] loop.py:45 INFO TRAIN_STEP step=16 loss=12.648544 Training: 0%| | 16/18018 [04:20<46:44:19, 9.35s/step, loss=12.649, acc=0.000, acc_len=0.00, thru=15.7, I=14.3, T=34.0, wait=8.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:34:34,129] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15: forward=2626.3ms backward=1096.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:34:34,471] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:34:39,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:34:40,560] loop.py:45 INFO TRAIN_STEP step=17 loss=10.525787 Training: 0%| | 17/18018 [04:26<42:20:44, 8.47s/step, loss=10.526, acc=0.022, acc_len=0.02, thru=19.4, I=20.9, T=35.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:34:40,785] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:34:46,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:34:48,253] loop.py:45 INFO TRAIN_STEP step=18 loss=10.384370 Training: 0%| | 18/18018 [04:34<41:10:40, 8.24s/step, loss=10.384, acc=0.024, acc_len=0.03, thru=17.2, I=22.2, T=28.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:34:48,438] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:34:56,070] loop.py:45 INFO TRAIN_STEP step=19 loss=10.402840 Training: 0%| | 19/18018 [04:42<40:32:47, 8.11s/step, loss=10.403, acc=0.025, acc_len=0.03, thru=14.0, I=16.1, T=34.1, wait=4.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:35:02,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:35:02,957] loop.py:380 INFO TIMING step=20: step=3.783s data=1.738s compute=3.735s [fwd=2.614s bwd=1.107s opt=0.003s] dispatch=3.097s [2026-10-18 13:35:02,957] loop.py:45 INFO TRAIN_STEP step=20 loss=10.233594 Training: 0%| | 20/18018 [04:49<38:42:37, 7.74s/step, loss=10.234, acc=0.027, acc_len=0.03, thru=19.0, I=19.4, T=33.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:35:03,026] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 13:35:10,520] loop.py:45 INFO TRAIN_STEP step=21 loss=10.395747 Training: 0%| | 21/18018 [04:56<38:26:10, 7.69s/step, loss=10.396, acc=0.021, acc_len=0.02, thru=14.7, I=18.0, T=31.2, wait=3.5s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:35:10,515] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=20: forward=2834.8ms backward=1211.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:35:16,972] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:35:17,374] loop.py:45 INFO TRAIN_STEP step=22 loss=9.985695 Training: 0%| | 22/18018 [05:03<37:11:14, 7.44s/step, loss=9.986, acc=0.032, acc_len=0.03, thru=19.4, I=17.9, T=34.2, wait=3.1s, pool=64, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:35:17,482] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-18 13:35:25,894] loop.py:45 INFO TRAIN_STEP step=23 loss=11.466491 Training: 0%| | 23/18018 [05:11<38:48:23, 7.76s/step, loss=11.466, acc=0.024, acc_len=0.02, thru=13.2, I=13.9, T=33.4, wait=4.7s, pool=48, epoch=1/2][2026-10-18 13:35:33,205] loop.py:45 INFO TRAIN_STEP step=24 loss=11.673258 Training: 0%| | 24/18018 [05:19<38:07:20, 7.63s/step, loss=11.673, acc=0.023, acc_len=0.02, thru=18.2, I=16.8, T=32.7, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:35:38,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:35:40,161] loop.py:380 INFO TIMING step=25: step=4.166s data=1.967s compute=4.133s [fwd=2.937s bwd=1.181s opt=0.003s] dispatch=2.782s [2026-10-18 13:35:40,161] loop.py:45 INFO TRAIN_STEP step=25 loss=11.252782 Training: 0%| | 25/18018 [05:26<37:06:49, 7.43s/step, loss=11.253, acc=0.019, acc_len=0.02, thru=18.3, I=21.6, T=30.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:35:40,403] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:35:45,688] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:35:46,862] loop.py:45 INFO TRAIN_STEP step=26 loss=10.746106 Training: 0%| | 26/18018 [05:32<36:01:34, 7.21s/step, loss=10.746, acc=0.012, acc_len=0.01, thru=16.6, I=22.9, T=34.5, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:35:46,858] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=25: forward=2605.2ms backward=1061.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:35:47,198] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:35:54,247] loop.py:45 INFO TRAIN_STEP step=27 loss=10.199952 Training: 0%| | 27/18018 [05:40<36:17:13, 7.26s/step, loss=10.200, acc=0.012, acc_len=0.01, thru=15.7, I=18.1, T=34.7, wait=3.7s, pool=48, epoch=1/2][2026-10-18 13:36:02,954] loop.py:45 INFO TRAIN_STEP step=28 loss=9.792153 Training: 0%| | 28/18018 [05:49<38:27:15, 7.70s/step, loss=9.792, acc=0.013, acc_len=0.01, thru=15.9, I=16.3, T=31.9, wait=4.7s, pool=60, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:10,885] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:36:12,216] loop.py:45 INFO TRAIN_STEP step=29 loss=9.516523 Training: 0%| | 29/18018 [05:58<40:47:59, 8.16s/step, loss=9.517, acc=0.010, acc_len=0.01, thru=13.9, I=15.2, T=29.5, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:12,395] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:17,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:36:18,457] loop.py:380 INFO TIMING step=30: step=3.645s data=1.486s compute=3.616s [fwd=2.511s bwd=1.090s opt=0.003s] dispatch=2.589s [2026-10-18 13:36:18,457] loop.py:45 INFO TRAIN_STEP step=30 loss=9.583597 Training: 0%| | 30/18018 [06:04<37:54:49, 7.59s/step, loss=9.584, acc=0.010, acc_len=0.01, thru=18.7, I=22.8, T=35.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:18,506] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:26,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 13:36:27,023] loop.py:45 INFO TRAIN_STEP step=31 loss=9.423443 Training: 0%| | 31/18018 [06:13<39:22:51, 7.88s/step, loss=9.423, acc=0.009, acc_len=0.01, thru=15.2, I=16.7, T=29.1, wait=4.2s, pool=68, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:36:27,018] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=30: forward=3082.3ms backward=1258.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:27,528] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:35,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:36:36,294] loop.py:45 INFO TRAIN_STEP step=32 loss=9.468643 Training: 0%| | 32/18018 [06:22<41:27:25, 8.30s/step, loss=9.469, acc=0.011, acc_len=0.01, thru=14.2, I=15.9, T=29.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:36,550] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:42,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:36:43,408] loop.py:45 INFO TRAIN_STEP step=33 loss=9.178998 Training: 0%| | 33/18018 [06:29<39:40:51, 7.94s/step, loss=9.179, acc=0.012, acc_len=0.01, thru=14.9, I=18.0, T=29.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:43,669] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:50,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:36:52,222] loop.py:45 INFO TRAIN_STEP step=34 loss=9.168322 Training: 0%| | 34/18018 [06:38<40:59:07, 8.20s/step, loss=9.168, acc=0.015, acc_len=0.02, thru=15.0, I=16.6, T=27.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:36:52,513] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 13:36:59,000] loop.py:380 INFO TIMING step=35: step=3.730s data=1.555s compute=3.692s [fwd=2.594s bwd=1.083s opt=0.003s] dispatch=3.040s [2026-10-18 13:36:59,000] loop.py:45 INFO TRAIN_STEP step=35 loss=9.082414 Training: 0%| | 35/18018 [06:45<38:50:42, 7.78s/step, loss=9.082, acc=0.017, acc_len=0.02, thru=14.9, I=17.5, T=34.3, wait=3.0s, pool=40, epoch=1/2][2026-10-18 13:37:06,698] loop.py:45 INFO TRAIN_STEP step=36 loss=9.175807 Training: 0%| | 36/18018 [06:52<38:43:35, 7.75s/step, loss=9.176, acc=0.026, acc_len=0.03, thru=17.9, I=16.6, T=31.6, wait=3.6s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:37:06,694] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=35: forward=2867.4ms backward=1127.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:13,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:37:14,938] loop.py:45 INFO TRAIN_STEP step=37 loss=9.087548 Training: 0%| | 37/18018 [07:01<39:27:15, 7.90s/step, loss=9.088, acc=0.032, acc_len=0.03, thru=17.2, I=17.3, T=31.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:15,009] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:21,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:37:22,447] loop.py:45 INFO TRAIN_STEP step=38 loss=9.019875 Training: 0%| | 38/18018 [07:08<38:51:57, 7.78s/step, loss=9.020, acc=0.035, acc_len=0.04, thru=16.5, I=20.2, T=35.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:22,740] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:29,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:37:30,688] loop.py:45 INFO TRAIN_STEP step=39 loss=9.069032 Training: 0%| | 39/18018 [07:16<39:33:03, 7.92s/step, loss=9.069, acc=0.029, acc_len=0.03, thru=15.9, I=18.4, T=32.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:30,726] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:34,930] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:37:36,723] loop.py:380 INFO TIMING step=40: step=4.255s data=1.761s compute=4.224s [fwd=2.877s bwd=1.331s opt=0.003s] dispatch=1.772s [2026-10-18 13:37:36,723] loop.py:45 INFO TRAIN_STEP step=40 loss=8.936074 Training: 0%| | 40/18018 [07:22<36:43:48, 7.36s/step, loss=8.936, acc=0.030, acc_len=0.03, thru=19.2, I=25.0, T=30.1, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:36,943] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:42,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:37:43,677] loop.py:45 INFO TRAIN_STEP step=41 loss=8.986003 Training: 0%| | 41/18018 [07:29<36:07:24, 7.23s/step, loss=8.986, acc=0.029, acc_len=0.03, thru=16.9, I=22.8, T=31.6, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:37:43,672] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=40: forward=2820.4ms backward=1177.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:44,091] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:49,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:37:50,852] loop.py:45 INFO TRAIN_STEP step=42 loss=8.910391 Training: 0%| | 42/18018 [07:36<36:02:05, 7.22s/step, loss=8.910, acc=0.030, acc_len=0.03, thru=16.9, I=22.5, T=30.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:51,046] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:56,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:37:57,147] loop.py:45 INFO TRAIN_STEP step=43 loss=8.982863 Training: 0%| | 43/18018 [07:43<34:39:18, 6.94s/step, loss=8.983, acc=0.029, acc_len=0.03, thru=18.5, I=24.5, T=34.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:37:57,360] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:38:04,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:38:04,385] loop.py:45 INFO TRAIN_STEP step=44 loss=8.775221 Training: 0%| | 44/18018 [07:50<35:05:34, 7.03s/step, loss=8.775, acc=0.037, acc_len=0.04, thru=17.2, I=18.5, T=32.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:38:04,703] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:38:11,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:38:11,667] loop.py:380 INFO TIMING step=45: step=3.775s data=1.689s compute=3.745s [fwd=2.657s bwd=1.074s opt=0.003s] dispatch=3.500s [2026-10-18 13:38:11,667] loop.py:45 INFO TRAIN_STEP step=45 loss=8.837664 Training: 0%| | 45/18018 [07:57<35:28:14, 7.10s/step, loss=8.838, acc=0.034, acc_len=0.04, thru=17.6, I=20.5, T=33.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:38:11,874] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 13:38:19,869] loop.py:45 INFO TRAIN_STEP step=46 loss=8.746463 Training: 0%| | 46/18018 [08:05<37:06:44, 7.43s/step, loss=8.746, acc=0.031, acc_len=0.03, thru=14.0, I=14.3, T=31.1, wait=4.1s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:38:19,865] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=45: forward=2924.1ms backward=1141.6ms [2026-10-18 13:38:28,401] loop.py:45 INFO TRAIN_STEP step=47 loss=8.665648 Training: 0%| | 47/18018 [08:14<38:45:23, 7.76s/step, loss=8.666, acc=0.029, acc_len=0.03, thru=15.2, I=14.9, T=29.7, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:38:38,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:38:39,291] loop.py:45 INFO TRAIN_STEP step=48 loss=8.759953 Training: 0%| | 48/18018 [08:25<43:26:02, 8.70s/step, loss=8.760, acc=0.023, acc_len=0.02, thru=12.3, I=12.4, T=25.7, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:38:39,774] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 13:38:48,661] loop.py:45 INFO TRAIN_STEP step=49 loss=8.650294 Training: 0%| | 49/18018 [08:34<44:25:57, 8.90s/step, loss=8.650, acc=0.024, acc_len=0.03, thru=12.1, I=12.3, T=27.7, wait=4.7s, pool=48, epoch=1/2][2026-10-18 13:38:57,131] loop.py:380 INFO TIMING step=50: step=4.327s data=1.937s compute=4.299s [fwd=3.065s bwd=1.220s opt=0.003s] dispatch=4.135s [2026-10-18 13:38:57,131] loop.py:45 INFO TRAIN_STEP step=50 loss=8.538673 Training: 0%| | 50/18018 [08:43<43:47:00, 8.77s/step, loss=8.539, acc=0.026, acc_len=0.03, thru=12.9, I=12.9, T=29.6, wait=4.1s, pool=32, epoch=1/2][2026-10-18 13:39:06,659] loop.py:45 INFO TRAIN_STEP step=51 loss=8.465458 Training: 0%| | 51/18018 [08:52<44:54:44, 9.00s/step, loss=8.465, acc=0.030, acc_len=0.03, thru=15.9, I=13.9, T=27.2, wait=4.8s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:39:06,655] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=50: forward=3342.0ms backward=1326.0ms [2026-10-18 13:39:13,497] loop.py:45 INFO TRAIN_STEP step=52 loss=8.376737 Training: 0%| | 52/18018 [08:59<41:40:37, 8.35s/step, loss=8.377, acc=0.033, acc_len=0.03, thru=17.5, I=16.5, T=32.9, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:39:19,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:39:20,932] loop.py:45 INFO TRAIN_STEP step=53 loss=8.554112 Training: 0%| | 53/18018 [09:07<40:18:07, 8.08s/step, loss=8.554, acc=0.033, acc_len=0.03, thru=17.3, I=19.1, T=28.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:39:21,110] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:39:26,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:39:27,299] loop.py:45 INFO TRAIN_STEP step=54 loss=8.381465 Training: 0%| | 54/18018 [09:13<37:44:41, 7.56s/step, loss=8.381, acc=0.037, acc_len=0.04, thru=17.3, I=20.2, T=33.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:39:27,474] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 13:39:33,619] loop.py:380 INFO TIMING step=55: step=3.801s data=1.712s compute=3.765s [fwd=2.642s bwd=1.109s opt=0.003s] dispatch=2.510s [2026-10-18 13:39:33,619] loop.py:45 INFO TRAIN_STEP step=55 loss=8.494708 Training: 0%| | 55/18018 [09:19<35:52:29, 7.19s/step, loss=8.495, acc=0.033, acc_len=0.03, thru=18.1, I=21.0, T=33.7, wait=2.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:39:40,441] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:39:40,741] loop.py:45 INFO TRAIN_STEP step=56 loss=8.356013 Training: 0%| | 56/18018 [09:26<35:46:30, 7.17s/step, loss=8.356, acc=0.032, acc_len=0.03, thru=19.2, I=18.4, T=34.8, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:39:40,738] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=55: forward=2631.4ms backward=1004.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:39:40,943] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:39:47,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:39:49,279] loop.py:45 INFO TRAIN_STEP step=57 loss=8.316700 Training: 0%| | 57/18018 [09:35<37:49:07, 7.58s/step, loss=8.317, acc=0.034, acc_len=0.04, thru=15.4, I=16.9, T=32.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:39:49,474] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:39:55,144] loop.py:45 INFO TRAIN_STEP step=58 loss=8.256960 Training: 0%| | 58/18018 [09:41<35:15:03, 7.07s/step, loss=8.257, acc=0.037, acc_len=0.04, thru=19.3, I=23.5, T=36.4, wait=2.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:03,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:40:03,914] loop.py:45 INFO TRAIN_STEP step=59 loss=8.212938 Training: 0%| | 59/18018 [09:49<37:47:51, 7.58s/step, loss=8.213, acc=0.035, acc_len=0.04, thru=15.3, I=14.8, T=29.7, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:04,187] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:10,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:40:10,143] loop.py:380 INFO TIMING step=60: step=3.756s data=1.607s compute=3.715s [fwd=2.643s bwd=1.060s opt=0.003s] dispatch=2.465s [2026-10-18 13:40:10,144] loop.py:45 INFO TRAIN_STEP step=60 loss=8.159833 Training: 0%| | 60/18018 [09:56<35:46:40, 7.17s/step, loss=8.160, acc=0.036, acc_len=0.04, thru=18.5, I=20.1, T=34.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:10,575] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:18,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:40:18,368] loop.py:45 INFO TRAIN_STEP step=61 loss=8.178251 Training: 0%| | 61/18018 [10:04<37:21:12, 7.49s/step, loss=8.178, acc=0.032, acc_len=0.03, thru=16.1, I=17.7, T=35.6, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:40:18,364] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=60: forward=2528.9ms backward=1015.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:18,520] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 13:40:25,945] loop.py:45 INFO TRAIN_STEP step=62 loss=8.231567 Training: 0%| | 62/18018 [10:12<37:28:55, 7.51s/step, loss=8.232, acc=0.032, acc_len=0.03, thru=17.7, I=18.2, T=35.5, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:34,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:40:34,893] loop.py:45 INFO TRAIN_STEP step=63 loss=8.237432 Training: 0%| | 63/18018 [10:20<39:37:43, 7.95s/step, loss=8.237, acc=0.035, acc_len=0.04, thru=14.9, I=15.5, T=35.1, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:35,188] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:40,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:40:42,128] loop.py:45 INFO TRAIN_STEP step=64 loss=8.102127 Training: 0%| | 64/18018 [10:28<38:33:32, 7.73s/step, loss=8.102, acc=0.037, acc_len=0.04, thru=18.1, I=20.6, T=34.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:42,234] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:47,408] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:40:48,892] loop.py:380 INFO TIMING step=65: step=4.151s data=1.717s compute=4.114s [fwd=2.731s bwd=1.370s opt=0.003s] dispatch=2.605s [2026-10-18 13:40:48,892] loop.py:45 INFO TRAIN_STEP step=65 loss=8.070495 Training: 0%| | 65/18018 [10:34<37:06:45, 7.44s/step, loss=8.070, acc=0.034, acc_len=0.04, thru=18.9, I=25.0, T=30.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:40:48,929] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 13:40:54,632] loop.py:45 INFO TRAIN_STEP step=66 loss=8.006795 Training: 0%| | 66/18018 [10:40<34:33:38, 6.93s/step, loss=8.007, acc=0.036, acc_len=0.04, thru=17.6, I=23.7, T=35.5, wait=2.1s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:40:54,627] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=65: forward=2538.9ms backward=1017.9ms [2026-10-18 13:41:04,987] loop.py:45 INFO TRAIN_STEP step=67 loss=7.990691 Training: 0%| | 67/18018 [10:51<39:40:48, 7.96s/step, loss=7.991, acc=0.039, acc_len=0.04, thru=12.5, I=11.2, T=33.7, wait=6.5s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:41:12,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:41:13,786] loop.py:45 INFO TRAIN_STEP step=68 loss=8.001709 Training: 0%| | 68/18018 [10:59<40:56:17, 8.21s/step, loss=8.002, acc=0.039, acc_len=0.04, thru=17.2, I=16.2, T=29.0, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:41:14,042] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:41:20,658] loop.py:45 INFO TRAIN_STEP step=69 loss=8.018573 Training: 0%| | 69/18018 [11:06<38:56:24, 7.81s/step, loss=8.019, acc=0.034, acc_len=0.04, thru=15.0, I=19.4, T=34.9, wait=3.2s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:41:27,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:41:29,119] loop.py:380 INFO TIMING step=70: step=3.979s data=1.662s compute=3.933s [fwd=2.764s bwd=1.153s opt=0.003s] dispatch=4.469s [2026-10-18 13:41:29,119] loop.py:45 INFO TRAIN_STEP step=70 loss=8.035909 Training: 0%| | 70/18018 [11:15<39:54:13, 8.00s/step, loss=8.036, acc=0.036, acc_len=0.04, thru=17.4, I=17.7, T=32.2, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:41:29,223] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 13:41:35,520] loop.py:45 INFO TRAIN_STEP step=71 loss=7.852804 Training: 0%| | 71/18018 [11:21<37:30:16, 7.52s/step, loss=7.853, acc=0.039, acc_len=0.04, thru=14.6, I=18.2, T=32.1, wait=2.4s, pool=28, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:41:35,515] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=70: forward=2790.2ms backward=1142.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:41:45,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:41:45,481] loop.py:45 INFO TRAIN_STEP step=72 loss=7.980505 Training: 0%| | 72/18018 [11:31<41:08:59, 8.25s/step, loss=7.981, acc=0.033, acc_len=0.04, thru=16.6, I=16.1, T=28.5, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:41:45,540] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:41:50,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:41:52,780] loop.py:45 INFO TRAIN_STEP step=73 loss=7.793852 Training: 0%| | 73/18018 [11:38<39:43:03, 7.97s/step, loss=7.794, acc=0.039, acc_len=0.04, thru=17.7, I=20.6, T=29.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:41:53,284] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 13:42:00,020] loop.py:45 INFO TRAIN_STEP step=74 loss=7.911824 Training: 0%| | 74/18018 [11:46<38:37:40, 7.75s/step, loss=7.912, acc=0.038, acc_len=0.04, thru=12.3, I=17.4, T=32.5, wait=3.3s, pool=44, epoch=1/2][2026-10-18 13:42:08,599] loop.py:380 INFO TIMING step=75: step=4.055s data=1.730s compute=4.017s [fwd=2.826s bwd=1.177s opt=0.003s] dispatch=4.516s [2026-10-18 13:42:08,599] loop.py:45 INFO TRAIN_STEP step=75 loss=7.811233 Training: 0%| | 75/18018 [11:54<39:51:54, 8.00s/step, loss=7.811, acc=0.035, acc_len=0.04, thru=16.5, I=15.5, T=31.6, wait=4.5s, pool=60, epoch=1/2][2026-10-18 13:42:18,057] loop.py:45 INFO TRAIN_STEP step=76 loss=7.830582 Training: 0%| | 76/18018 [12:04<42:02:53, 8.44s/step, loss=7.831, acc=0.040, acc_len=0.04, thru=14.0, I=13.4, T=30.4, wait=5.2s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:42:18,053] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=75: forward=2968.7ms backward=1193.2ms [2026-10-18 13:42:24,423] loop.py:45 INFO TRAIN_STEP step=77 loss=7.851901 Training: 0%| | 77/18018 [12:10<38:56:55, 7.82s/step, loss=7.852, acc=0.039, acc_len=0.04, thru=16.7, I=16.1, T=33.8, wait=2.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:42:31,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:42:33,534] loop.py:45 INFO TRAIN_STEP step=78 loss=7.786313 Training: 0%| | 78/18018 [12:19<40:52:55, 8.20s/step, loss=7.786, acc=0.038, acc_len=0.04, thru=16.6, I=17.1, T=29.6, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:42:33,575] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:42:40,912] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:42:40,991] loop.py:45 INFO TRAIN_STEP step=79 loss=7.795861 Training: 0%| | 79/18018 [12:27<39:45:45, 7.98s/step, loss=7.796, acc=0.040, acc_len=0.04, thru=13.6, I=17.7, T=30.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:42:41,413] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:42:49,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:42:50,279] loop.py:380 INFO TIMING step=80: step=5.125s data=2.479s compute=5.097s [fwd=3.532s bwd=1.551s opt=0.003s] dispatch=4.155s [2026-10-18 13:42:50,280] loop.py:45 INFO TRAIN_STEP step=80 loss=7.879227 Training: 0%| | 80/18018 [12:36<41:43:14, 8.37s/step, loss=7.879, acc=0.038, acc_len=0.04, thru=14.6, I=16.3, T=25.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:42:50,360] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:42:55,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:42:56,030] loop.py:45 INFO TRAIN_STEP step=81 loss=7.781675 Training: 0%| | 81/18018 [12:42<37:47:46, 7.59s/step, loss=7.782, acc=0.036, acc_len=0.04, thru=18.0, I=20.7, T=32.5, wait=1.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:42:56,025] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=80: forward=2827.4ms backward=1065.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:42:56,427] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 13:43:04,118] loop.py:45 INFO TRAIN_STEP step=82 loss=7.790873 Training: 0%| | 82/18018 [12:50<38:32:36, 7.74s/step, loss=7.791, acc=0.038, acc_len=0.04, thru=14.2, I=15.2, T=30.8, wait=3.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:09,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:43:11,149] loop.py:45 INFO TRAIN_STEP step=83 loss=7.882170 Training: 0%| | 83/18018 [12:57<37:29:09, 7.52s/step, loss=7.882, acc=0.036, acc_len=0.04, thru=18.5, I=18.9, T=31.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:11,324] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:18,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:43:18,634] loop.py:45 INFO TRAIN_STEP step=84 loss=7.806170 Training: 0%| | 84/18018 [13:04<37:25:44, 7.51s/step, loss=7.806, acc=0.041, acc_len=0.05, thru=15.7, I=19.3, T=29.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:19,019] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:27,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:43:27,068] loop.py:380 INFO TIMING step=85: step=4.382s data=1.935s compute=4.350s [fwd=3.156s bwd=1.178s opt=0.003s] dispatch=4.043s [2026-10-18 13:43:27,069] loop.py:45 INFO TRAIN_STEP step=85 loss=7.779509 Training: 0%| | 85/18018 [13:13<38:48:01, 7.79s/step, loss=7.780, acc=0.042, acc_len=0.05, thru=14.0, I=14.4, T=29.2, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:27,537] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:34,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:43:36,331] loop.py:45 INFO TRAIN_STEP step=86 loss=7.757675 Training: 0%| | 86/18018 [13:22<40:59:56, 8.23s/step, loss=7.758, acc=0.035, acc_len=0.04, thru=14.6, I=16.8, T=28.6, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:43:36,326] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=85: forward=3178.8ms backward=1245.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:36,738] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:41,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:43:42,790] loop.py:45 INFO TRAIN_STEP step=87 loss=7.560999 Training: 0%| | 87/18018 [13:28<38:20:58, 7.70s/step, loss=7.561, acc=0.038, acc_len=0.04, thru=18.5, I=26.2, T=34.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:43,290] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:43:49,782] loop.py:45 INFO TRAIN_STEP step=88 loss=7.687859 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:49,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 Training: 0%| | 88/18018 [13:35<37:17:25, 7.49s/step, loss=7.688, acc=0.044, acc_len=0.05, thru=17.7, I=21.4, T=31.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:50,206] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:57,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:43:57,786] loop.py:45 INFO TRAIN_STEP step=89 loss=7.972794 Training: 0%| | 89/18018 [13:43<38:03:42, 7.64s/step, loss=7.973, acc=0.041, acc_len=0.05, thru=15.6, I=16.9, T=31.0, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:43:58,233] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:06,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:44:07,090] loop.py:380 INFO TIMING step=90: step=4.015s data=1.699s compute=3.985s [fwd=2.817s bwd=1.155s opt=0.003s] dispatch=5.281s [2026-10-18 13:44:07,090] loop.py:45 INFO TRAIN_STEP step=90 loss=7.646582 Training: 0%| | 90/18018 [13:53<40:32:39, 8.14s/step, loss=7.647, acc=0.040, acc_len=0.05, thru=14.1, I=17.9, T=31.9, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:07,133] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 13:44:14,107] loop.py:45 INFO TRAIN_STEP step=91 loss=7.639168 Training: 1%| | 91/18018 [14:00<38:51:43, 7.80s/step, loss=7.639, acc=0.033, acc_len=0.04, thru=16.1, I=18.9, T=32.2, wait=3.0s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:44:14,102] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=90: forward=2776.1ms backward=1153.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:20,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:44:22,300] loop.py:45 INFO TRAIN_STEP step=92 loss=7.475389 Training: 1%| | 92/18018 [14:08<39:26:17, 7.92s/step, loss=7.475, acc=0.040, acc_len=0.05, thru=16.4, I=18.6, T=28.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:22,565] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:28,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:44:30,280] loop.py:45 INFO TRAIN_STEP step=93 loss=7.711426 Training: 1%| | 93/18018 [14:16<39:31:28, 7.94s/step, loss=7.711, acc=0.041, acc_len=0.05, thru=14.0, I=20.7, T=30.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:30,652] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:36,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:44:37,609] loop.py:45 INFO TRAIN_STEP step=94 loss=7.795483 Training: 1%| | 94/18018 [14:23<38:36:44, 7.76s/step, loss=7.795, acc=0.041, acc_len=0.05, thru=15.8, I=23.4, T=31.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:37,886] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:43,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:44:45,234] loop.py:380 INFO TIMING step=95: step=4.694s data=2.269s compute=4.666s [fwd=3.258s bwd=1.394s opt=0.003s] dispatch=2.924s [2026-10-18 13:44:45,235] loop.py:45 INFO TRAIN_STEP step=95 loss=7.720085 Training: 1%| | 95/18018 [14:31<38:25:00, 7.72s/step, loss=7.720, acc=0.041, acc_len=0.05, thru=15.0, I=21.1, T=27.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:44:45,658] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 13:44:52,364] loop.py:45 INFO TRAIN_STEP step=96 loss=7.641102 Training: 1%| | 96/18018 [14:38<37:32:28, 7.54s/step, loss=7.641, acc=0.037, acc_len=0.05, thru=13.3, I=20.7, T=30.7, wait=2.9s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:44:52,359] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=95: forward=2949.4ms backward=1175.7ms [2026-10-18 13:45:00,840] loop.py:45 INFO TRAIN_STEP step=97 loss=7.489685 Training: 1%| | 97/18018 [14:46<38:56:08, 7.82s/step, loss=7.490, acc=0.040, acc_len=0.06, thru=14.8, I=13.4, T=32.9, wait=4.6s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:10,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:45:11,319] loop.py:45 INFO TRAIN_STEP step=98 loss=7.708146 Training: 1%| | 98/18018 [14:57<42:53:57, 8.62s/step, loss=7.708, acc=0.041, acc_len=0.05, thru=15.7, I=16.0, T=29.3, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:11,519] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:16,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:45:19,598] loop.py:45 INFO TRAIN_STEP step=99 loss=7.794607 Training: 1%| | 99/18018 [15:05<42:23:27, 8.52s/step, loss=7.795, acc=0.049, acc_len=0.06, thru=15.6, I=22.0, T=27.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:20,030] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:26,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:45:27,212] loop.py:380 INFO TIMING step=100: step=3.911s data=1.761s compute=3.877s [fwd=2.758s bwd=1.105s opt=0.003s] dispatch=3.697s [2026-10-18 13:45:27,212] loop.py:45 INFO TRAIN_STEP step=100 loss=7.621095 Training: 1%| | 100/18018 [15:13<41:02:41, 8.25s/step, loss=7.621, acc=0.044, acc_len=0.05, thru=17.8, I=19.8, T=32.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:27,417] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:32,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:45:35,132] loop.py:45 INFO TRAIN_STEP step=101 loss=7.474805 Training: 1%| | 101/18018 [15:21<40:33:08, 8.15s/step, loss=7.475, acc=0.042, acc_len=0.06, thru=16.6, I=22.8, T=28.5, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:45:35,128] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=100: forward=3140.2ms backward=1281.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:35,613] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 13:45:42,136] loop.py:45 INFO TRAIN_STEP step=102 loss=7.421321 Training: 1%| | 102/18018 [15:28<38:50:40, 7.81s/step, loss=7.421, acc=0.039, acc_len=0.06, thru=14.0, I=20.1, T=33.3, wait=3.1s, pool=60, epoch=1/2][2026-10-18 13:45:48,745] loop.py:45 INFO TRAIN_STEP step=103 loss=7.326738 Training: 1%| | 103/18018 [15:34<37:03:07, 7.45s/step, loss=7.327, acc=0.045, acc_len=0.07, thru=17.0, I=16.7, T=32.5, wait=2.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:56,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:45:58,049] loop.py:45 INFO TRAIN_STEP step=104 loss=7.543283 Training: 1%| | 104/18018 [15:44<39:49:28, 8.00s/step, loss=7.543, acc=0.046, acc_len=0.06, thru=16.0, I=19.8, T=25.9, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:45:58,539] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:05,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:46:05,397] loop.py:380 INFO TIMING step=105: step=4.016s data=1.825s compute=3.977s [fwd=2.808s bwd=1.156s opt=0.003s] dispatch=3.324s [2026-10-18 13:46:05,397] loop.py:45 INFO TRAIN_STEP step=105 loss=7.889488 Training: 1%| | 105/18018 [15:51<38:50:50, 7.81s/step, loss=7.889, acc=0.045, acc_len=0.06, thru=14.9, I=19.7, T=31.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:05,627] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:10,459] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:46:13,677] loop.py:45 INFO TRAIN_STEP step=106 loss=7.546590 Training: 1%| | 106/18018 [15:59<39:32:53, 7.95s/step, loss=7.547, acc=0.052, acc_len=0.08, thru=16.8, I=23.5, T=24.9, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:46:13,672] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=105: forward=3857.8ms backward=1222.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:13,986] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 [2026-10-18 13:46:21,120] loop.py:45 INFO TRAIN_STEP step=107 loss=7.327232 Training: 1%| | 107/18018 [16:07<38:47:44, 7.80s/step, loss=7.327, acc=0.053, acc_len=0.08, thru=14.0, I=18.2, T=29.7, wait=3.1s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:30,706] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:46:32,200] loop.py:45 INFO TRAIN_STEP step=108 loss=7.344130 Training: 1%| | 108/18018 [16:18<43:41:21, 8.78s/step, loss=7.344, acc=0.046, acc_len=0.07, thru=14.6, I=15.9, T=24.9, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:32,230] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:41,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:46:41,688] loop.py:45 INFO TRAIN_STEP step=109 loss=7.383254 Training: 1%| | 109/18018 [16:27<44:44:34, 8.99s/step, loss=7.383, acc=0.048, acc_len=0.07, thru=13.6, I=13.9, T=29.0, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:42,037] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:47,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:46:49,084] loop.py:380 INFO TIMING step=110: step=4.292s data=2.005s compute=4.259s [fwd=3.012s bwd=1.234s opt=0.003s] dispatch=3.094s [2026-10-18 13:46:49,085] loop.py:45 INFO TRAIN_STEP step=110 loss=7.676497 Training: 1%| | 110/18018 [16:35<42:21:11, 8.51s/step, loss=7.676, acc=0.051, acc_len=0.08, thru=16.6, I=19.9, T=29.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:49,195] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:54,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:46:55,238] loop.py:45 INFO TRAIN_STEP step=111 loss=7.517990 Training: 1%| | 111/18018 [16:41<38:49:42, 7.81s/step, loss=7.518, acc=0.054, acc_len=0.08, thru=18.4, I=23.7, T=35.9, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:46:55,233] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=110: forward=2494.3ms backward=1022.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:46:55,426] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:00,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:47:01,641] loop.py:45 INFO TRAIN_STEP step=112 loss=7.289061 Training: 1%| | 112/18018 [16:47<36:43:49, 7.38s/step, loss=7.289, acc=0.054, acc_len=0.08, thru=19.3, I=22.6, T=35.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:01,866] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:06,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:47:07,869] loop.py:45 INFO TRAIN_STEP step=113 loss=7.273764 Training: 1%| | 113/18018 [16:53<35:00:30, 7.04s/step, loss=7.274, acc=0.050, acc_len=0.08, thru=19.8, I=24.8, T=33.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:07,979] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 13:47:14,779] loop.py:45 INFO TRAIN_STEP step=114 loss=7.498441 Training: 1%| | 114/18018 [17:00<34:48:54, 7.00s/step, loss=7.498, acc=0.045, acc_len=0.07, thru=18.0, I=19.5, T=32.3, wait=2.9s, pool=60, epoch=1/2][2026-10-18 13:47:24,558] loop.py:380 INFO TIMING step=115: step=4.405s data=1.932s compute=4.374s [fwd=3.179s bwd=1.181s opt=0.003s] dispatch=5.364s [2026-10-18 13:47:24,558] loop.py:45 INFO TRAIN_STEP step=115 loss=7.473534 Training: 1%| | 115/18018 [17:10<38:57:20, 7.83s/step, loss=7.474, acc=0.055, acc_len=0.08, thru=11.6, I=12.6, T=29.1, wait=5.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:31,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:47:33,794] loop.py:45 INFO TRAIN_STEP step=116 loss=7.714788 Training: 1%| | 116/18018 [17:19<41:02:36, 8.25s/step, loss=7.715, acc=0.055, acc_len=0.08, thru=15.4, I=17.9, T=28.7, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:47:33,789] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=115: forward=3177.9ms backward=1228.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:33,980] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:41,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:47:41,667] loop.py:45 INFO TRAIN_STEP step=117 loss=7.720180 Training: 1%| | 117/18018 [17:27<40:28:45, 8.14s/step, loss=7.720, acc=0.055, acc_len=0.08, thru=16.7, I=17.5, T=31.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:42,034] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:47,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:47:48,816] loop.py:45 INFO TRAIN_STEP step=118 loss=7.263744 Training: 1%| | 118/18018 [17:34<38:59:34, 7.84s/step, loss=7.264, acc=0.062, acc_len=0.10, thru=18.7, I=20.1, T=33.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:48,883] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:54,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:47:55,093] loop.py:45 INFO TRAIN_STEP step=119 loss=7.344844 Training: 1%| | 119/18018 [17:41<36:39:20, 7.37s/step, loss=7.345, acc=0.055, acc_len=0.09, thru=19.3, I=22.8, T=34.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:47:55,367] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:00,306] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:48:01,631] loop.py:380 INFO TIMING step=120: step=3.810s data=1.549s compute=3.777s [fwd=2.631s bwd=1.132s opt=0.003s] dispatch=2.721s [2026-10-18 13:48:01,631] loop.py:45 INFO TRAIN_STEP step=120 loss=7.440237 Training: 1%| | 120/18018 [17:47<35:24:37, 7.12s/step, loss=7.440, acc=0.052, acc_len=0.08, thru=18.8, I=22.0, T=33.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:01,813] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:48:09,036] loop.py:45 INFO TRAIN_STEP step=121 loss=7.351808 Training: 1%| | 121/18018 [17:55<35:49:49, 7.21s/step, loss=7.352, acc=0.058, acc_len=0.09, thru=14.4, I=16.4, T=32.6, wait=3.5s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:48:09,031] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=120: forward=2920.7ms backward=962.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:15,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:48:16,744] loop.py:45 INFO TRAIN_STEP step=122 loss=7.520081 Training: 1%| | 122/18018 [18:02<36:34:23, 7.36s/step, loss=7.520, acc=0.057, acc_len=0.09, thru=18.6, I=19.0, T=33.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:17,018] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:23,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:48:24,188] loop.py:45 INFO TRAIN_STEP step=123 loss=7.598552 Training: 1%| | 123/18018 [18:10<36:42:04, 7.38s/step, loss=7.599, acc=0.056, acc_len=0.09, thru=16.8, I=20.3, T=27.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:24,599] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:29,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:48:30,660] loop.py:45 INFO TRAIN_STEP step=124 loss=7.367232 Training: 1%| | 124/18018 [18:16<35:20:21, 7.11s/step, loss=7.367, acc=0.061, acc_len=0.10, thru=17.8, I=20.9, T=34.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:30,881] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP2 pid=1847) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP3 pid=1848) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP2 pid=1847) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP3 pid=1848) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP1 pid=1846) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP1 pid=1846) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:39,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:48:41,695] loop.py:380 INFO TIMING step=125: step=4.227s data=2.013s compute=4.187s [fwd=2.983s bwd=1.189s opt=0.003s] dispatch=6.799s [2026-10-18 13:48:41,695] loop.py:45 INFO TRAIN_STEP step=125 loss=7.147203 Training: 1%| | 125/18018 [18:27<41:11:30, 8.29s/step, loss=7.147, acc=0.063, acc_len=0.10, thru=11.7, I=14.1, T=30.3, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:41,762] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:46,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:48:48,026] loop.py:45 INFO TRAIN_STEP step=126 loss=7.167740 Training: 1%| | 126/18018 [18:34<38:16:17, 7.70s/step, loss=7.168, acc=0.064, acc_len=0.11, thru=18.5, I=23.5, T=32.1, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:48:48,021] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=125: forward=2829.3ms backward=1123.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:48,312] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:54,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:48:55,887] loop.py:45 INFO TRAIN_STEP step=127 loss=7.321270 Training: 1%| | 127/18018 [18:41<38:30:32, 7.75s/step, loss=7.321, acc=0.064, acc_len=0.11, thru=14.1, I=22.8, T=29.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:48:56,028] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:01,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:49:02,680] loop.py:45 INFO TRAIN_STEP step=128 loss=7.279501 Training: 1%| | 128/18018 [18:48<37:05:11, 7.46s/step, loss=7.280, acc=0.065, acc_len=0.11, thru=16.2, I=22.1, T=34.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:02,818] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 13:49:13,153] loop.py:45 INFO TRAIN_STEP step=129 loss=7.257769 Training: 1%| | 129/18018 [18:59<41:34:26, 8.37s/step, loss=7.258, acc=0.066, acc_len=0.11, thru=11.7, I=13.3, T=22.9, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:13,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:13,662] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:19,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:49:20,588] loop.py:380 INFO TIMING step=130: step=4.304s data=1.887s compute=4.265s [fwd=3.016s bwd=1.235s opt=0.003s] dispatch=3.119s [2026-10-18 13:49:20,588] loop.py:45 INFO TRAIN_STEP step=130 loss=7.071975 Training: 1%| | 130/18018 [19:06<40:10:33, 8.09s/step, loss=7.072, acc=0.068, acc_len=0.11, thru=17.0, I=16.6, T=29.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:21,063] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 13:49:27,863] loop.py:45 INFO TRAIN_STEP step=131 loss=7.189581 Training: 1%| | 131/18018 [19:13<38:57:53, 7.84s/step, loss=7.190, acc=0.060, acc_len=0.10, thru=11.7, I=25.3, T=27.0, wait=2.5s, pool=24, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:49:27,858] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=130: forward=3345.1ms backward=1343.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:38,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:49:38,867] loop.py:45 INFO TRAIN_STEP step=132 loss=7.087379 Training: 1%| | 132/18018 [19:24<43:40:38, 8.79s/step, loss=7.087, acc=0.068, acc_len=0.12, thru=17.6, I=13.3, T=30.3, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:39,295] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:47,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:49:48,663] loop.py:45 INFO TRAIN_STEP step=133 loss=7.405940 Training: 1%| | 133/18018 [19:34<45:10:14, 9.09s/step, loss=7.406, acc=0.067, acc_len=0.11, thru=13.7, I=15.3, T=29.8, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:48,709] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:55,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:49:55,873] loop.py:45 INFO TRAIN_STEP step=134 loss=7.407720 Training: 1%| | 134/18018 [19:41<42:21:50, 8.53s/step, loss=7.408, acc=0.069, acc_len=0.12, thru=16.3, I=21.1, T=34.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:49:55,956] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 [2026-10-18 13:50:03,092] loop.py:380 INFO TIMING step=135: step=4.176s data=1.876s compute=4.143s [fwd=2.865s bwd=1.263s opt=0.003s] dispatch=3.033s [2026-10-18 13:50:03,092] loop.py:45 INFO TRAIN_STEP step=135 loss=7.190456 Training: 1%| | 135/18018 [19:49<40:24:41, 8.14s/step, loss=7.190, acc=0.069, acc_len=0.12, thru=14.9, I=17.5, T=30.6, wait=3.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:50:10,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 13:50:12,461] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=135: forward=3063.3ms backward=1245.0ms [2026-10-18 13:50:12,466] loop.py:45 INFO TRAIN_STEP step=136 loss=7.037370 Training: 1%| | 136/18018 [19:58<42:15:16, 8.51s/step, loss=7.037, acc=0.071, acc_len=0.13, thru=17.3, I=18.2, T=29.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:50:12,663] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:50:17,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:50:18,868] loop.py:45 INFO TRAIN_STEP step=137 loss=6.813484 Training: 1%| | 137/18018 [20:04<39:07:01, 7.88s/step, loss=6.813, acc=0.076, acc_len=0.13, thru=16.5, I=25.9, T=30.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:50:19,104] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:50:26,716] loop.py:45 INFO TRAIN_STEP step=138 loss=7.075679 Training: 1%| | 138/18018 [20:12<39:04:22, 7.87s/step, loss=7.076, acc=0.076, acc_len=0.13, thru=14.4, I=18.6, T=27.9, wait=3.3s, pool=60, epoch=1/2][2026-10-18 13:50:38,304] loop.py:45 INFO TRAIN_STEP step=139 loss=7.242049 Training: 1%| | 139/18018 [20:24<44:36:52, 8.98s/step, loss=7.242, acc=0.074, acc_len=0.12, thru=6.8, I=7.0, T=28.7, wait=7.1s, pool=28, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:50:47,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:50:48,516] loop.py:380 INFO TIMING step=140: step=4.577s data=2.119s compute=4.537s [fwd=3.222s bwd=1.299s opt=0.003s] dispatch=5.627s [2026-10-18 13:50:48,516] loop.py:45 INFO TRAIN_STEP step=140 loss=6.930690 Training: 1%| | 140/18018 [20:34<46:26:40, 9.35s/step, loss=6.931, acc=0.083, acc_len=0.14, thru=16.5, I=16.3, T=28.0, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:50:48,948] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:50:55,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:50:56,306] loop.py:45 INFO TRAIN_STEP step=141 loss=6.867904 Training: 1%| | 141/18018 [20:42<44:06:43, 8.88s/step, loss=6.868, acc=0.078, acc_len=0.14, thru=14.1, I=16.8, T=32.0, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:50:56,301] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=140: forward=2864.0ms backward=1086.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:50:56,771] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:01,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:51:03,016] loop.py:45 INFO TRAIN_STEP step=142 loss=6.857378 Training: 1%| | 142/18018 [20:49<40:52:25, 8.23s/step, loss=6.857, acc=0.079, acc_len=0.14, thru=19.0, I=20.6, T=31.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:03,239] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:08,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:51:10,484] loop.py:45 INFO TRAIN_STEP step=143 loss=6.949113 Training: 1%| | 143/18018 [20:56<39:44:01, 8.00s/step, loss=6.949, acc=0.083, acc_len=0.15, thru=15.6, I=20.7, T=30.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:10,519] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:15,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:51:17,901] loop.py:45 INFO TRAIN_STEP step=144 loss=6.924882 Training: 1%| | 144/18018 [21:03<38:51:44, 7.83s/step, loss=6.925, acc=0.089, acc_len=0.15, thru=14.4, I=21.4, T=28.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:17,967] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 [2026-10-18 13:51:27,293] loop.py:380 INFO TIMING step=145: step=4.281s data=1.939s compute=4.234s [fwd=2.907s bwd=1.312s opt=0.003s] dispatch=5.101s [2026-10-18 13:51:27,293] loop.py:45 INFO TRAIN_STEP step=145 loss=6.750978 Training: 1%| | 145/18018 [21:13<41:11:24, 8.30s/step, loss=6.751, acc=0.090, acc_len=0.16, thru=11.6, I=13.5, T=29.9, wait=5.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:35,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:51:36,809] loop.py:45 INFO TRAIN_STEP step=146 loss=6.833025 Training: 1%| | 146/18018 [21:22<43:00:18, 8.66s/step, loss=6.833, acc=0.088, acc_len=0.15, thru=15.8, I=15.4, T=31.5, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:51:36,805] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=145: forward=2877.4ms backward=1126.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:37,189] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:42,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:51:43,832] loop.py:45 INFO TRAIN_STEP step=147 loss=6.797007 Training: 1%| | 147/18018 [21:29<40:33:22, 8.17s/step, loss=6.797, acc=0.094, acc_len=0.17, thru=17.4, I=21.5, T=29.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:43,936] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:49,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:51:50,651] loop.py:45 INFO TRAIN_STEP step=148 loss=6.882034 Training: 1%| | 148/18018 [21:36<38:32:44, 7.77s/step, loss=6.882, acc=0.091, acc_len=0.16, thru=17.9, I=21.9, T=30.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:51,108] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:57,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:51:58,884] loop.py:45 INFO TRAIN_STEP step=149 loss=6.733195 Training: 1%| | 149/18018 [21:44<39:14:24, 7.91s/step, loss=6.733, acc=0.095, acc_len=0.17, thru=13.6, I=19.8, T=27.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:51:59,299] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:52:07,207] loop.py:380 INFO TIMING step=150: step=4.278s data=1.899s compute=4.251s [fwd=3.057s bwd=1.179s opt=0.003s] dispatch=4.036s [2026-10-18 13:52:07,208] loop.py:45 INFO TRAIN_STEP step=150 loss=6.743061 Training: 1%| | 150/18018 [21:53<39:51:42, 8.03s/step, loss=6.743, acc=0.095, acc_len=0.17, thru=14.7, I=17.7, T=29.9, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:52:13,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:52:15,621] loop.py:45 INFO TRAIN_STEP step=151 loss=6.733097 Training: 1%| | 151/18018 [22:01<40:25:24, 8.14s/step, loss=6.733, acc=0.104, acc_len=0.18, thru=16.5, I=19.0, T=24.7, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:52:15,617] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=150: forward=3536.5ms backward=1582.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:52:15,975] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:52:21,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:52:22,897] loop.py:45 INFO TRAIN_STEP step=152 loss=6.895875 Training: 1%| | 152/18018 [22:08<39:07:44, 7.88s/step, loss=6.896, acc=0.098, acc_len=0.17, thru=13.8, I=18.8, T=31.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:52:23,039] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:52:29,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:52:30,354] loop.py:45 INFO TRAIN_STEP step=153 loss=6.515274 Training: 1%| | 153/18018 [22:16<38:29:26, 7.76s/step, loss=6.515, acc=0.110, acc_len=0.20, thru=16.4, I=19.0, T=26.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:52:30,836] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:52:36,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:52:37,746] loop.py:45 INFO TRAIN_STEP step=154 loss=6.397350 Training: 1%| | 154/18018 [22:23<37:56:42, 7.65s/step, loss=6.397, acc=0.113, acc_len=0.20, thru=16.5, I=19.3, T=32.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:52:37,814] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 13:52:44,929] loop.py:380 INFO TIMING step=155: step=4.025s data=1.863s compute=3.992s [fwd=2.821s bwd=1.155s opt=0.003s] dispatch=3.151s [2026-10-18 13:52:44,930] loop.py:45 INFO TRAIN_STEP step=155 loss=6.519794 Training: 1%| | 155/18018 [22:31<37:15:14, 7.51s/step, loss=6.520, acc=0.108, acc_len=0.19, thru=16.5, I=17.6, T=31.8, wait=3.2s, pool=52, epoch=1/2][2026-10-18 13:52:52,115] loop.py:45 INFO TRAIN_STEP step=156 loss=6.516409 Training: 1%| | 156/18018 [22:38<36:46:19, 7.41s/step, loss=6.516, acc=0.115, acc_len=0.21, thru=18.2, I=16.5, T=32.8, wait=3.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:52:52,110] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=155: forward=2694.7ms backward=1167.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:52:59,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:53:01,074] loop.py:45 INFO TRAIN_STEP step=157 loss=6.559089 Training: 1%| | 157/18018 [22:47<39:04:25, 7.88s/step, loss=6.559, acc=0.111, acc_len=0.20, thru=16.7, I=18.9, T=29.8, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:01,144] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:06,324] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:53:08,710] loop.py:45 INFO TRAIN_STEP step=158 loss=6.515202 Training: 1%| | 158/18018 [22:54<38:42:54, 7.80s/step, loss=6.515, acc=0.118, acc_len=0.22, thru=15.3, I=24.1, T=26.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:08,857] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:14,840] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:53:16,482] loop.py:45 INFO TRAIN_STEP step=159 loss=6.342429 Training: 1%| | 159/18018 [23:02<38:39:53, 7.79s/step, loss=6.342, acc=0.121, acc_len=0.22, thru=16.8, I=20.7, T=27.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:16,858] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:23,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:53:23,494] loop.py:380 INFO TIMING step=160: step=3.722s data=1.624s compute=3.691s [fwd=2.660s bwd=1.017s opt=0.003s] dispatch=3.282s [2026-10-18 13:53:23,494] loop.py:45 INFO TRAIN_STEP step=160 loss=6.516238 Training: 1%| | 160/18018 [23:09<37:29:57, 7.56s/step, loss=6.516, acc=0.117, acc_len=0.21, thru=15.7, I=19.0, T=34.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:23,753] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:29,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:53:29,944] loop.py:45 INFO TRAIN_STEP step=161 loss=6.300606 Training: 1%| | 161/18018 [23:16<35:50:43, 7.23s/step, loss=6.301, acc=0.130, acc_len=0.24, thru=19.5, I=20.7, T=36.5, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:53:29,939] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=160: forward=2503.4ms backward=959.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:29,976] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 13:53:38,668] loop.py:45 INFO TRAIN_STEP step=162 loss=6.483133 Training: 1%| | 162/18018 [23:24<38:04:31, 7.68s/step, loss=6.483, acc=0.125, acc_len=0.23, thru=13.3, I=14.0, T=33.0, wait=4.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:46,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:53:47,723] loop.py:45 INFO TRAIN_STEP step=163 loss=6.609027 Training: 1%| | 163/18018 [23:33<40:07:19, 8.09s/step, loss=6.609, acc=0.119, acc_len=0.22, thru=15.9, I=15.9, T=31.9, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:47,940] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:55,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:53:56,556] loop.py:45 INFO TRAIN_STEP step=164 loss=6.429508 Training: 1%| | 164/18018 [23:42<41:13:31, 8.31s/step, loss=6.430, acc=0.130, acc_len=0.24, thru=14.8, I=17.4, T=28.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:53:57,059] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:54:03,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:54:05,000] loop.py:380 INFO TIMING step=165: step=4.330s data=1.937s compute=4.284s [fwd=2.911s bwd=1.358s opt=0.003s] dispatch=4.105s [2026-10-18 13:54:05,000] loop.py:45 INFO TRAIN_STEP step=165 loss=6.178013 Training: 1%| | 165/18018 [23:51<41:25:04, 8.35s/step, loss=6.178, acc=0.141, acc_len=0.26, thru=14.0, I=19.4, T=29.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:54:05,208] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:54:11,866] loop.py:45 INFO TRAIN_STEP step=166 loss=6.421654 Training: 1%| | 166/18018 [23:57<39:12:18, 7.91s/step, loss=6.422, acc=0.131, acc_len=0.24, thru=14.9, I=17.5, T=33.0, wait=3.0s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:54:11,862] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=165: forward=2708.7ms backward=1128.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:54:19,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:54:22,065] loop.py:45 INFO TRAIN_STEP step=167 loss=6.351928 Training: 1%| | 167/18018 [24:08<42:36:46, 8.59s/step, loss=6.352, acc=0.137, acc_len=0.25, thru=15.3, I=19.4, T=27.0, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:54:22,152] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 [2026-10-18 13:54:29,496] loop.py:45 INFO TRAIN_STEP step=168 loss=6.320413 Training: 1%| | 168/18018 [24:15<40:53:01, 8.25s/step, loss=6.320, acc=0.136, acc_len=0.25, thru=13.4, I=17.9, T=34.7, wait=3.7s, pool=60, epoch=1/2][2026-10-18 13:54:37,488] loop.py:45 INFO TRAIN_STEP step=169 loss=6.254141 Training: 1%| | 169/18018 [24:23<40:30:26, 8.17s/step, loss=6.254, acc=0.138, acc_len=0.25, thru=14.5, I=15.2, T=33.7, wait=4.2s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:54:47,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:54:48,232] loop.py:380 INFO TIMING step=170: step=4.376s data=1.978s compute=4.347s [fwd=3.029s bwd=1.302s opt=0.003s] dispatch=6.357s [2026-10-18 13:54:48,232] loop.py:45 INFO TRAIN_STEP step=170 loss=6.262866 Training: 1%| | 170/18018 [24:34<44:19:59, 8.94s/step, loss=6.263, acc=0.142, acc_len=0.27, thru=15.4, I=14.7, T=29.3, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:54:48,569] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:54:54,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:54:55,649] loop.py:45 INFO TRAIN_STEP step=171 loss=6.266986 Training: 1%| | 171/18018 [24:41<42:03:32, 8.48s/step, loss=6.267, acc=0.145, acc_len=0.27, thru=16.1, I=22.7, T=29.8, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:54:55,645] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=170: forward=2895.8ms backward=1348.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:54:56,071] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:55:01,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:55:02,550] loop.py:45 INFO TRAIN_STEP step=172 loss=6.176278 Training: 1%| | 172/18018 [24:48<39:42:07, 8.01s/step, loss=6.176, acc=0.145, acc_len=0.27, thru=16.4, I=23.6, T=32.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:55:02,980] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:55:10,315] loop.py:45 INFO TRAIN_STEP step=173 loss=6.217975 Training: 1%| | 173/18018 [24:56<39:20:28, 7.94s/step, loss=6.218, acc=0.145, acc_len=0.27, thru=15.6, I=18.0, T=34.7, wait=4.1s, pool=60, epoch=1/2][2026-10-18 13:55:18,678] loop.py:45 INFO TRAIN_STEP step=174 loss=6.030471 Training: 1%| | 174/18018 [25:04<39:58:18, 8.06s/step, loss=6.030, acc=0.149, acc_len=0.28, thru=15.8, I=14.6, T=33.0, wait=4.5s, pool=52, epoch=1/2][2026-10-18 13:55:26,151] loop.py:380 INFO TIMING step=175: step=3.953s data=1.741s compute=3.911s [fwd=2.783s bwd=1.112s opt=0.003s] dispatch=3.511s [2026-10-18 13:55:26,151] loop.py:45 INFO TRAIN_STEP step=175 loss=6.082870 Training: 1%| | 175/18018 [25:12<39:05:14, 7.89s/step, loss=6.083, acc=0.151, acc_len=0.28, thru=15.7, I=14.0, T=32.4, wait=3.5s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:55:34,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:55:36,005] loop.py:45 INFO TRAIN_STEP step=176 loss=6.087761 Training: 1%| | 176/18018 [25:22<42:00:41, 8.48s/step, loss=6.088, acc=0.151, acc_len=0.28, thru=15.7, I=16.3, T=27.6, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:55:36,001] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=175: forward=3213.8ms backward=1365.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:55:36,309] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:55:43,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:55:43,630] loop.py:45 INFO TRAIN_STEP step=177 loss=6.140546 Training: 1%| | 177/18018 [25:29<40:44:32, 8.22s/step, loss=6.141, acc=0.149, acc_len=0.28, thru=16.1, I=20.3, T=30.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:55:44,058] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 13:55:50,279] loop.py:45 INFO TRAIN_STEP step=178 loss=5.971210 Training: 1%| | 178/18018 [25:36<38:24:24, 7.75s/step, loss=5.971, acc=0.159, acc_len=0.30, thru=14.1, I=17.7, T=36.0, wait=3.1s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:55:57,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:55:58,228] loop.py:45 INFO TRAIN_STEP step=179 loss=6.107876 Training: 1%| | 179/18018 [25:44<38:42:03, 7.81s/step, loss=6.108, acc=0.161, acc_len=0.30, thru=18.3, I=18.2, T=32.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:55:58,537] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:03,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:56:05,376] loop.py:380 INFO TIMING step=180: step=4.225s data=2.004s compute=4.189s [fwd=2.947s bwd=1.227s opt=0.003s] dispatch=2.912s [2026-10-18 13:56:05,376] loop.py:45 INFO TRAIN_STEP step=180 loss=6.193346 Training: 1%| | 180/18018 [25:51<37:42:39, 7.61s/step, loss=6.193, acc=0.155, acc_len=0.29, thru=16.2, I=23.6, T=30.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:05,653] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 13:56:12,226] loop.py:45 INFO TRAIN_STEP step=181 loss=6.023564 Training: 1%| | 181/18018 [25:58<36:34:35, 7.38s/step, loss=6.024, acc=0.156, acc_len=0.30, thru=16.2, I=21.2, T=31.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:12,198] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 13:56:12,221] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=180: forward=2758.9ms backward=1213.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:12,701] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 13:56:20,021] loop.py:45 INFO TRAIN_STEP step=182 loss=5.847962 Training: 1%| | 182/18018 [26:06<37:11:21, 7.51s/step, loss=5.848, acc=0.169, acc_len=0.33, thru=14.0, I=15.5, T=30.4, wait=3.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:27,712] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:56:28,859] loop.py:45 INFO TRAIN_STEP step=183 loss=6.237046 Training: 1%| | 183/18018 [26:14<39:10:13, 7.91s/step, loss=6.237, acc=0.153, acc_len=0.29, thru=16.6, I=15.3, T=31.3, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:29,222] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:56:35,274] loop.py:45 INFO TRAIN_STEP step=184 loss=6.188788 Training: 1%| | 184/18018 [26:21<36:56:57, 7.46s/step, loss=6.189, acc=0.154, acc_len=0.29, thru=16.0, I=21.4, T=33.2, wait=2.6s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:43,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:56:43,869] loop.py:380 INFO TIMING step=185: step=4.353s data=2.041s compute=4.321s [fwd=3.006s bwd=1.299s opt=0.003s] dispatch=4.233s [2026-10-18 13:56:43,869] loop.py:45 INFO TRAIN_STEP step=185 loss=5.925847 Training: 1%| | 185/18018 [26:29<38:38:14, 7.80s/step, loss=5.926, acc=0.171, acc_len=0.33, thru=18.1, I=16.0, T=29.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:44,073] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:49,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:56:50,509] loop.py:45 INFO TRAIN_STEP step=186 loss=6.023108 Training: 1%| | 186/18018 [26:36<36:54:40, 7.45s/step, loss=6.023, acc=0.165, acc_len=0.31, thru=17.2, I=20.2, T=32.0, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:56:50,505] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=185: forward=2818.7ms backward=1132.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:50,790] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:55,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:56:57,680] loop.py:45 INFO TRAIN_STEP step=187 loss=6.176584 Training: 1%| | 187/18018 [26:43<36:29:24, 7.37s/step, loss=6.177, acc=0.166, acc_len=0.32, thru=17.1, I=23.3, T=33.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:56:57,713] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:03,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:57:04,060] loop.py:45 INFO TRAIN_STEP step=188 loss=6.023778 Training: 1%| | 188/18018 [26:50<35:01:12, 7.07s/step, loss=6.024, acc=0.164, acc_len=0.31, thru=18.6, I=23.4, T=35.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:04,450] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:10,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:57:10,884] loop.py:45 INFO TRAIN_STEP step=189 loss=5.929179 Training: 1%| | 189/18018 [26:56<34:39:23, 7.00s/step, loss=5.929, acc=0.167, acc_len=0.32, thru=18.7, I=19.6, T=36.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:10,945] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:17,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:57:17,583] loop.py:380 INFO TIMING step=190: step=3.947s data=1.789s compute=3.913s [fwd=2.764s bwd=1.135s opt=0.003s] dispatch=2.741s [2026-10-18 13:57:17,583] loop.py:45 INFO TRAIN_STEP step=190 loss=5.877247 Training: 1%| | 190/18018 [27:03<34:12:32, 6.91s/step, loss=5.877, acc=0.171, acc_len=0.33, thru=18.5, I=20.8, T=32.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:18,023] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 13:57:25,875] loop.py:45 INFO TRAIN_STEP step=191 loss=5.964943 Training: 1%| | 191/18018 [27:11<36:16:13, 7.32s/step, loss=5.965, acc=0.168, acc_len=0.33, thru=14.5, I=15.7, T=32.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:25,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 13:57:25,871] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=190: forward=2778.8ms backward=1120.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:26,385] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:31,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:57:32,954] loop.py:45 INFO TRAIN_STEP step=192 loss=6.111207 Training: 1%| | 192/18018 [27:19<35:53:39, 7.25s/step, loss=6.111, acc=0.166, acc_len=0.31, thru=16.6, I=19.8, T=31.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:33,129] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:38,558] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:57:40,316] loop.py:45 INFO TRAIN_STEP step=193 loss=5.853458 Training: 1%| | 193/18018 [27:26<36:03:33, 7.28s/step, loss=5.853, acc=0.173, acc_len=0.33, thru=16.2, I=20.7, T=31.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:40,572] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:46,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:57:47,779] loop.py:45 INFO TRAIN_STEP step=194 loss=5.874990 Training: 1%| | 194/18018 [27:33<36:19:41, 7.34s/step, loss=5.875, acc=0.172, acc_len=0.33, thru=14.4, I=22.4, T=30.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:47,812] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:54,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:57:55,848] loop.py:380 INFO TIMING step=195: step=3.914s data=1.684s compute=3.889s [fwd=2.696s bwd=1.179s opt=0.003s] dispatch=4.145s [2026-10-18 13:57:55,848] loop.py:45 INFO TRAIN_STEP step=195 loss=6.080526 Training: 1%| | 195/18018 [27:41<37:24:46, 7.56s/step, loss=6.081, acc=0.173, acc_len=0.33, thru=13.8, I=19.3, T=32.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:57:56,318] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 13:58:04,133] loop.py:45 INFO TRAIN_STEP step=196 loss=5.987112 Training: 1%| | 196/18018 [27:50<38:29:22, 7.77s/step, loss=5.987, acc=0.174, acc_len=0.34, thru=11.9, I=15.2, T=31.2, wait=4.2s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:58:04,129] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=195: forward=2803.5ms backward=1247.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:10,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:58:12,610] loop.py:45 INFO TRAIN_STEP step=197 loss=5.773764 Training: 1%| | 197/18018 [27:58<39:31:48, 7.99s/step, loss=5.774, acc=0.182, acc_len=0.35, thru=16.4, I=17.5, T=29.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:12,861] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:18,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:58:19,885] loop.py:45 INFO TRAIN_STEP step=198 loss=5.753603 Training: 1%| | 198/18018 [28:05<38:28:27, 7.77s/step, loss=5.754, acc=0.180, acc_len=0.35, thru=16.0, I=26.0, T=32.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:20,064] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:24,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:58:26,023] loop.py:45 INFO TRAIN_STEP step=199 loss=5.623214 Training: 1%| | 199/18018 [28:12<36:02:39, 7.28s/step, loss=5.623, acc=0.191, acc_len=0.37, thru=19.3, I=26.7, T=31.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:26,376] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:32,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:58:34,015] loop.py:380 INFO TIMING step=200: step=3.989s data=1.777s compute=3.955s [fwd=2.764s bwd=1.174s opt=0.003s] dispatch=3.993s [2026-10-18 13:58:34,015] loop.py:45 INFO TRAIN_STEP step=200 loss=5.890394 Training: 1%| | 200/18018 [28:20<37:05:54, 7.50s/step, loss=5.890, acc=0.177, acc_len=0.34, thru=15.3, I=20.0, T=32.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:34,204] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:41,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:58:42,354] loop.py:45 INFO TRAIN_STEP step=201 loss=5.733185 Training: 1%| | 201/18018 [28:28<38:20:43, 7.75s/step, loss=5.733, acc=0.188, acc_len=0.37, thru=13.8, I=18.2, T=31.0, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:58:42,350] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=200: forward=2835.4ms backward=1221.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:42,565] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:50,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:58:50,854] loop.py:45 INFO TRAIN_STEP step=202 loss=5.836637 Training: 1%| | 202/18018 [28:36<39:27:34, 7.97s/step, loss=5.837, acc=0.181, acc_len=0.35, thru=13.8, I=16.0, T=29.0, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:51,071] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:57,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:58:58,168] loop.py:45 INFO TRAIN_STEP step=203 loss=5.646140 Training: 1%| | 203/18018 [28:44<38:28:48, 7.78s/step, loss=5.646, acc=0.189, acc_len=0.37, thru=16.9, I=17.2, T=30.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:58:58,627] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:06,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:59:07,579] loop.py:45 INFO TRAIN_STEP step=204 loss=5.975858 Training: 1%| | 204/18018 [28:53<40:54:30, 8.27s/step, loss=5.976, acc=0.171, acc_len=0.33, thru=13.2, I=15.8, T=28.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:07,876] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:14,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:59:14,614] loop.py:380 INFO TIMING step=205: step=4.034s data=1.579s compute=4.000s [fwd=2.725s bwd=1.260s opt=0.003s] dispatch=2.990s [2026-10-18 13:59:14,614] loop.py:45 INFO TRAIN_STEP step=205 loss=5.803170 Training: 1%| | 205/18018 [29:00<39:04:32, 7.90s/step, loss=5.803, acc=0.178, acc_len=0.35, thru=18.0, I=19.8, T=31.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:14,766] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:21,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:59:21,809] loop.py:45 INFO TRAIN_STEP step=206 loss=5.734872 Training: 1%| | 206/18018 [29:07<38:01:46, 7.69s/step, loss=5.735, acc=0.185, acc_len=0.36, thru=16.9, I=18.9, T=33.3, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:59:21,804] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=205: forward=2656.9ms backward=1133.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:22,220] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:27,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:59:28,180] loop.py:45 INFO TRAIN_STEP step=207 loss=5.540312 Training: 1%| | 207/18018 [29:14<36:04:29, 7.29s/step, loss=5.540, acc=0.196, acc_len=0.39, thru=19.5, I=22.3, T=33.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:28,462] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:33,272] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:59:34,458] loop.py:45 INFO TRAIN_STEP step=208 loss=5.671391 Training: 1%| | 208/18018 [29:20<34:34:12, 6.99s/step, loss=5.671, acc=0.193, acc_len=0.38, thru=18.8, I=22.5, T=34.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:34,781] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:40,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:59:41,566] loop.py:45 INFO TRAIN_STEP step=209 loss=5.628750 Training: 1%| | 209/18018 [29:27<34:44:51, 7.02s/step, loss=5.629, acc=0.199, acc_len=0.39, thru=17.7, I=22.8, T=33.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:41,851] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:47,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:59:47,909] loop.py:380 INFO TIMING step=210: step=3.655s data=1.587s compute=3.623s [fwd=2.602s bwd=1.007s opt=0.003s] dispatch=2.678s [2026-10-18 13:59:47,909] loop.py:45 INFO TRAIN_STEP step=210 loss=5.458834 Training: 1%| | 210/18018 [29:33<33:43:52, 6.82s/step, loss=5.459, acc=0.200, acc_len=0.39, thru=19.1, I=20.5, T=35.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:48,314] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:54,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 13:59:54,805] loop.py:45 INFO TRAIN_STEP step=211 loss=5.706563 Training: 1%| | 211/18018 [29:40<33:50:37, 6.84s/step, loss=5.707, acc=0.186, acc_len=0.36, thru=19.4, I=19.7, T=32.8, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 13:59:54,801] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=210: forward=2747.6ms backward=1107.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:54,907] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 13:59:59,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:00:01,556] loop.py:45 INFO TRAIN_STEP step=212 loss=5.624259 Training: 1%| | 212/18018 [29:47<33:42:29, 6.82s/step, loss=5.624, acc=0.193, acc_len=0.38, thru=17.9, I=22.2, T=33.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:01,643] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:08,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:00:08,928] loop.py:45 INFO TRAIN_STEP step=213 loss=5.577985 Training: 1%| | 213/18018 [29:55<34:31:51, 6.98s/step, loss=5.578, acc=0.196, acc_len=0.38, thru=15.8, I=22.1, T=31.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:09,251] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:13,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:00:15,281] loop.py:45 INFO TRAIN_STEP step=214 loss=5.767814 Training: 1%| | 214/18018 [30:01<33:36:01, 6.79s/step, loss=5.768, acc=0.187, acc_len=0.36, thru=19.2, I=23.8, T=34.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:15,375] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 14:00:22,934] loop.py:380 INFO TIMING step=215: step=4.066s data=1.841s compute=4.035s [fwd=2.836s bwd=1.183s opt=0.003s] dispatch=3.576s [2026-10-18 14:00:22,934] loop.py:45 INFO TRAIN_STEP step=215 loss=5.643243 Training: 1%| | 215/18018 [30:09<34:52:12, 7.05s/step, loss=5.643, acc=0.199, acc_len=0.39, thru=14.5, I=18.5, T=31.5, wait=3.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:30,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:00:30,602] loop.py:45 INFO TRAIN_STEP step=216 loss=5.800221 Training: 1%| | 216/18018 [30:16<35:47:04, 7.24s/step, loss=5.800, acc=0.180, acc_len=0.35, thru=16.4, I=16.2, T=30.8, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:00:30,594] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=215: forward=2930.4ms backward=1180.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:31,105] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:00:38,797] loop.py:45 INFO TRAIN_STEP step=217 loss=5.487625 Training: 1%| | 217/18018 [30:24<37:12:06, 7.52s/step, loss=5.488, acc=0.203, acc_len=0.40, thru=15.4, I=17.6, T=27.8, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:46,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:00:47,466] loop.py:45 INFO TRAIN_STEP step=218 loss=5.483270 Training: 1%| | 218/18018 [30:33<38:53:57, 7.87s/step, loss=5.483, acc=0.206, acc_len=0.40, thru=14.8, I=15.0, T=30.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:47,908] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:56,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:00:57,967] loop.py:45 INFO TRAIN_STEP step=219 loss=5.705129 Training: 1%| | 219/18018 [30:44<42:48:10, 8.66s/step, loss=5.705, acc=0.192, acc_len=0.38, thru=13.0, I=14.4, T=24.3, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:00:58,285] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:05,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:01:06,798] loop.py:380 INFO TIMING step=220: step=4.661s data=2.228s compute=4.618s [fwd=3.293s bwd=1.311s opt=0.003s] dispatch=4.164s [2026-10-18 14:01:06,798] loop.py:45 INFO TRAIN_STEP step=220 loss=5.587208 Training: 1%| | 220/18018 [30:52<43:03:42, 8.71s/step, loss=5.587, acc=0.203, acc_len=0.40, thru=15.0, I=16.4, T=27.5, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:07,304] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:11,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:01:13,996] loop.py:45 INFO TRAIN_STEP step=221 loss=5.681150 Training: 1%| | 221/18018 [31:00<40:48:43, 8.26s/step, loss=5.681, acc=0.195, acc_len=0.39, thru=16.4, I=23.2, T=31.5, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:01:13,991] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=220: forward=2855.5ms backward=1148.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:14,338] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:19,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:01:21,197] loop.py:45 INFO TRAIN_STEP step=222 loss=5.375729 Training: 1%| | 222/18018 [31:07<39:14:45, 7.94s/step, loss=5.376, acc=0.207, acc_len=0.41, thru=15.4, I=25.9, T=30.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:21,226] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:25,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:01:27,816] loop.py:45 INFO TRAIN_STEP step=223 loss=5.379060 Training: 1%| | 223/18018 [31:13<37:17:13, 7.54s/step, loss=5.379, acc=0.213, acc_len=0.43, thru=17.1, I=27.2, T=34.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:28,006] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:01:34,468] loop.py:45 INFO TRAIN_STEP step=224 loss=5.487389 Training: 1%| | 224/18018 [31:20<35:57:44, 7.28s/step, loss=5.487, acc=0.216, acc_len=0.43, thru=17.6, I=25.3, T=37.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:34,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:34,956] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:01:42,276] loop.py:380 INFO TIMING step=225: step=3.982s data=1.760s compute=3.930s [fwd=2.797s bwd=1.121s opt=0.003s] dispatch=3.819s [2026-10-18 14:01:42,277] loop.py:45 INFO TRAIN_STEP step=225 loss=5.357235 Training: 1%| | 225/18018 [31:28<36:45:05, 7.44s/step, loss=5.357, acc=0.221, acc_len=0.44, thru=15.7, I=16.9, T=32.1, wait=3.8s, pool=48, epoch=1/2][2026-10-18 14:01:49,750] loop.py:45 INFO TRAIN_STEP step=226 loss=5.424269 Training: 1%|▏ | 226/18018 [31:35<36:48:18, 7.45s/step, loss=5.424, acc=0.205, acc_len=0.41, thru=16.5, I=16.9, T=32.1, wait=3.5s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:01:49,746] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=225: forward=2788.5ms backward=1142.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:56,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:01:58,817] loop.py:45 INFO TRAIN_STEP step=227 loss=5.675685 Training: 1%|▏ | 227/18018 [31:44<39:12:20, 7.93s/step, loss=5.676, acc=0.183, acc_len=0.36, thru=14.8, I=17.6, T=27.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:01:58,957] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 14:02:05,751] loop.py:45 INFO TRAIN_STEP step=228 loss=5.583908 Training: 1%|▏ | 228/18018 [31:51<37:43:11, 7.63s/step, loss=5.584, acc=0.202, acc_len=0.40, thru=14.5, I=18.3, T=33.7, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:02:10,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:02:12,465] loop.py:45 INFO TRAIN_STEP step=229 loss=5.540579 Training: 1%|▏ | 229/18018 [31:58<36:21:26, 7.36s/step, loss=5.541, acc=0.209, acc_len=0.41, thru=18.6, I=23.6, T=30.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:02:12,747] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 14:02:18,573] loop.py:380 INFO TIMING step=230: step=3.491s data=1.504s compute=3.461s [fwd=2.489s bwd=0.960s opt=0.003s] dispatch=2.609s [2026-10-18 14:02:18,573] loop.py:45 INFO TRAIN_STEP step=230 loss=5.534569 Training: 1%|▏ | 230/18018 [32:04<34:30:17, 6.98s/step, loss=5.535, acc=0.211, acc_len=0.42, thru=15.8, I=23.7, T=36.7, wait=2.6s, pool=40, epoch=1/2][2026-10-18 14:02:30,721] loop.py:45 INFO TRAIN_STEP step=231 loss=5.240324 Training: 1%|▏ | 231/18018 [32:16<42:09:28, 8.53s/step, loss=5.240, acc=0.222, acc_len=0.45, thru=15.2, I=11.6, T=30.5, wait=7.9s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:02:30,714] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=230: forward=2941.3ms backward=1203.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:02:37,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:02:38,408] loop.py:45 INFO TRAIN_STEP step=232 loss=5.431721 Training: 1%|▏ | 232/18018 [32:24<40:53:57, 8.28s/step, loss=5.432, acc=0.212, acc_len=0.42, thru=16.1, I=16.6, T=29.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:02:38,596] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:02:43,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:02:45,017] loop.py:45 INFO TRAIN_STEP step=233 loss=5.271771 Training: 1%|▏ | 233/18018 [32:31<38:25:41, 7.78s/step, loss=5.272, acc=0.222, acc_len=0.45, thru=18.5, I=22.8, T=29.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:02:45,141] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:02:52,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:02:53,254] loop.py:45 INFO TRAIN_STEP step=234 loss=5.742540 Training: 1%|▏ | 234/18018 [32:39<39:06:05, 7.92s/step, loss=5.743, acc=0.200, acc_len=0.39, thru=15.8, I=18.1, T=31.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:02:53,321] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 14:03:01,240] loop.py:380 INFO TIMING step=235: step=4.871s data=2.635s compute=4.843s [fwd=3.638s bwd=1.192s opt=0.003s] dispatch=3.107s [2026-10-18 14:03:01,240] loop.py:45 INFO TRAIN_STEP step=235 loss=5.472419 Training: 1%|▏ | 235/18018 [32:47<39:12:19, 7.94s/step, loss=5.472, acc=0.218, acc_len=0.43, thru=15.1, I=20.3, T=26.3, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:09,744] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:03:11,694] loop.py:45 INFO TRAIN_STEP step=236 loss=5.529464 Training: 1%|▏ | 236/18018 [32:57<42:55:56, 8.69s/step, loss=5.529, acc=0.206, acc_len=0.41, thru=13.5, I=12.8, T=26.1, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:03:11,691] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=235: forward=3651.2ms backward=1180.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:11,759] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:19,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:03:20,535] loop.py:45 INFO TRAIN_STEP step=237 loss=5.372239 Training: 1%|▏ | 237/18018 [33:06<43:08:58, 8.74s/step, loss=5.372, acc=0.218, acc_len=0.44, thru=14.6, I=14.4, T=31.2, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:20,678] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:25,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:03:27,531] loop.py:45 INFO TRAIN_STEP step=238 loss=5.218555 Training: 1%|▏ | 238/18018 [33:13<40:34:06, 8.21s/step, loss=5.219, acc=0.230, acc_len=0.47, thru=17.6, I=23.3, T=28.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:27,617] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:33,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:03:34,285] loop.py:45 INFO TRAIN_STEP step=239 loss=5.458713 Training: 1%|▏ | 239/18018 [33:20<38:24:25, 7.78s/step, loss=5.459, acc=0.220, acc_len=0.44, thru=16.6, I=22.1, T=34.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:34,602] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:39,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:03:40,758] loop.py:380 INFO TIMING step=240: step=4.009s data=1.735s compute=3.979s [fwd=2.806s bwd=1.158s opt=0.003s] dispatch=2.453s [2026-10-18 14:03:40,758] loop.py:45 INFO TRAIN_STEP step=240 loss=5.244076 Training: 1%|▏ | 240/18018 [33:26<36:28:10, 7.39s/step, loss=5.244, acc=0.231, acc_len=0.47, thru=19.2, I=23.5, T=31.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:41,213] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:45,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:03:46,829] loop.py:45 INFO TRAIN_STEP step=241 loss=5.317469 Training: 1%|▏ | 241/18018 [33:32<34:31:14, 6.99s/step, loss=5.317, acc=0.221, acc_len=0.45, thru=19.8, I=25.1, T=35.1, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:03:46,824] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=240: forward=2565.0ms backward=1037.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:46,903] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:51,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:03:52,869] loop.py:45 INFO TRAIN_STEP step=242 loss=5.505305 Training: 1%|▏ | 242/18018 [33:38<33:06:53, 6.71s/step, loss=5.505, acc=0.212, acc_len=0.42, thru=20.2, I=23.7, T=32.0, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:53,248] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:03:59,587] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:04:00,346] loop.py:45 INFO TRAIN_STEP step=243 loss=5.524166 Training: 1%|▏ | 243/18018 [33:46<34:15:12, 6.94s/step, loss=5.524, acc=0.211, acc_len=0.42, thru=15.8, I=19.8, T=32.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:00,595] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:04:06,874] loop.py:45 INFO TRAIN_STEP step=244 loss=5.437367 Training: 1%|▏ | 244/18018 [33:52<33:38:33, 6.81s/step, loss=5.437, acc=0.215, acc_len=0.42, thru=17.1, I=22.0, T=33.6, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:15,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:04:16,285] loop.py:380 INFO TIMING step=245: step=4.349s data=1.965s compute=4.315s [fwd=3.117s bwd=1.183s opt=0.003s] dispatch=5.055s [2026-10-18 14:04:16,285] loop.py:45 INFO TRAIN_STEP step=245 loss=5.341001 Training: 1%|▏ | 245/18018 [34:02<37:29:23, 7.59s/step, loss=5.341, acc=0.219, acc_len=0.44, thru=14.8, I=14.9, T=29.4, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:16,687] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:22,008] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:04:22,840] loop.py:45 INFO TRAIN_STEP step=246 loss=5.434289 Training: 1%|▏ | 246/18018 [34:08<35:56:48, 7.28s/step, loss=5.434, acc=0.218, acc_len=0.44, thru=18.2, I=21.4, T=31.2, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:04:22,835] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=245: forward=2852.1ms backward=1197.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:23,011] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:27,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:04:28,620] loop.py:45 INFO TRAIN_STEP step=247 loss=5.370369 Training: 1%|▏ | 247/18018 [34:14<33:43:10, 6.83s/step, loss=5.370, acc=0.220, acc_len=0.44, thru=19.8, I=23.3, T=36.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:28,723] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:34,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:04:34,910] loop.py:45 INFO TRAIN_STEP step=248 loss=5.314921 Training: 1%|▏ | 248/18018 [34:20<32:55:17, 6.67s/step, loss=5.315, acc=0.227, acc_len=0.46, thru=19.8, I=23.8, T=34.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:35,394] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:40,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:04:41,630] loop.py:45 INFO TRAIN_STEP step=249 loss=5.327773 Training: 1%|▏ | 249/18018 [34:27<32:59:26, 6.68s/step, loss=5.328, acc=0.219, acc_len=0.44, thru=19.0, I=22.6, T=36.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:41,934] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:47,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:04:48,486] loop.py:380 INFO TIMING step=250: step=4.203s data=1.960s compute=4.171s [fwd=2.928s bwd=1.229s opt=0.003s] dispatch=2.645s [2026-10-18 14:04:48,486] loop.py:45 INFO TRAIN_STEP step=250 loss=5.157862 Training: 1%|▏ | 250/18018 [34:34<33:14:33, 6.74s/step, loss=5.158, acc=0.231, acc_len=0.47, thru=18.3, I=23.6, T=30.5, wait=2.6s, pool=64, epoch=1/2][2026-10-18 14:04:48,488] loop.py:419 INFO Saving checkpoint at step 250... (TrainerActor pid=9310) /usr/local/lib/python3.12/dist-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning. (TrainerActor pid=9310) return func(*args, **kwargs) (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:52,433] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 14:04:52,750] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0000251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:04:52,943] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=0/64 [2026-10-18 14:04:59,280] loop.py:45 INFO TRAIN_STEP step=251 loss=5.588046 Training: 1%|▏ | 251/18018 [34:45<39:15:00, 7.95s/step, loss=5.588, acc=0.207, acc_len=0.41, thru=15.8, I=17.6, T=36.6, wait=3.0s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:04:59,276] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=250: forward=2484.1ms backward=962.0ms (TrainerActor pid=9310) [2026-10-18 14:04:52,748] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0000251 [2026-10-18 14:05:07,582] loop.py:45 INFO TRAIN_STEP step=252 loss=5.430867 Training: 1%|▏ | 252/18018 [34:53<39:45:52, 8.06s/step, loss=5.431, acc=0.220, acc_len=0.44, thru=18.1, I=16.8, T=35.2, wait=4.7s, pool=56, epoch=1/2][2026-10-18 14:05:16,256] loop.py:45 INFO TRAIN_STEP step=253 loss=5.133228 Training: 1%|▏ | 253/18018 [35:02<40:40:27, 8.24s/step, loss=5.133, acc=0.225, acc_len=0.46, thru=14.6, I=14.4, T=32.2, wait=4.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:05:22,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:05:22,716] loop.py:45 INFO TRAIN_STEP step=254 loss=5.521304 Training: 1%|▏ | 254/18018 [35:08<38:01:59, 7.71s/step, loss=5.521, acc=0.215, acc_len=0.43, thru=19.4, I=18.8, T=32.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:05:23,164] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:05:30,602] loop.py:380 INFO TIMING step=255: step=3.723s data=1.544s compute=3.695s [fwd=2.654s bwd=1.027s opt=0.003s] dispatch=4.156s [2026-10-18 14:05:30,602] loop.py:45 INFO TRAIN_STEP step=255 loss=5.430842 Training: 1%|▏ | 255/18018 [35:16<38:17:45, 7.76s/step, loss=5.431, acc=0.227, acc_len=0.46, thru=12.1, I=16.0, T=34.4, wait=4.2s, pool=24, epoch=1/2][2026-10-18 14:05:42,391] loop.py:45 INFO TRAIN_STEP step=256 loss=5.102556 Training: 1%|▏ | 256/18018 [35:28<44:15:18, 8.97s/step, loss=5.103, acc=0.241, acc_len=0.49, thru=13.0, I=9.6, T=33.1, wait=7.9s, pool=32, epoch=1/2] (TrainerActor pid=9310) [2026-10-18 14:05:42,386] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=255: forward=2681.2ms backward=1121.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:05:50,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:05:51,187] loop.py:45 INFO TRAIN_STEP step=257 loss=5.276778 Training: 1%|▏ | 257/18018 [35:37<43:59:57, 8.92s/step, loss=5.277, acc=0.228, acc_len=0.46, thru=18.4, I=17.1, T=33.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:05:51,462] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:05:57,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:05:58,567] loop.py:45 INFO TRAIN_STEP step=258 loss=5.450928 Training: 1%|▏ | 258/18018 [35:44<41:43:11, 8.46s/step, loss=5.451, acc=0.213, acc_len=0.43, thru=17.0, I=19.4, T=34.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:05:58,735] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:06:03,904] loop.py:45 INFO TRAIN_STEP step=259 loss=5.109102 Training: 1%|▏ | 259/18018 [35:49<37:05:49, 7.52s/step, loss=5.109, acc=0.242, acc_len=0.49, thru=21.2, I=24.0, T=37.5, wait=1.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:06:11,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:06:12,866] loop.py:380 INFO TIMING step=260: step=4.244s data=1.902s compute=4.210s [fwd=3.031s bwd=1.165s opt=0.003s] dispatch=4.709s [2026-10-18 14:06:12,866] loop.py:45 INFO TRAIN_STEP step=260 loss=5.117305 Training: 1%|▏ | 260/18018 [35:58<39:13:58, 7.95s/step, loss=5.117, acc=0.234, acc_len=0.48, thru=14.9, I=16.4, T=30.2, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:06:13,052] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:06:19,478] loop.py:45 INFO TRAIN_STEP step=261 loss=5.297481 Training: 1%|▏ | 261/18018 [36:05<37:14:33, 7.55s/step, loss=5.297, acc=0.226, acc_len=0.45, thru=15.5, I=20.3, T=33.9, wait=2.8s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:06:19,474] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=260: forward=2644.1ms backward=1081.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:06:29,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:06:29,572] loop.py:45 INFO TRAIN_STEP step=262 loss=5.144284 Training: 1%|▏ | 262/18018 [36:15<41:00:24, 8.31s/step, loss=5.144, acc=0.242, acc_len=0.49, thru=15.6, I=13.7, T=30.0, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:06:29,789] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:06:38,476] loop.py:45 INFO TRAIN_STEP step=263 loss=5.330201 Training: 1%|▏ | 263/18018 [36:24<41:52:28, 8.49s/step, loss=5.330, acc=0.225, acc_len=0.45, thru=12.6, I=12.1, T=29.9, wait=4.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:06:44,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:06:46,216] loop.py:45 INFO TRAIN_STEP step=264 loss=5.252194 Training: 1%|▏ | 264/18018 [36:32<40:45:54, 8.27s/step, loss=5.252, acc=0.231, acc_len=0.46, thru=18.5, I=17.6, T=31.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:06:46,401] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:06:54,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:06:55,457] loop.py:380 INFO TIMING step=265: step=4.491s data=2.138s compute=4.441s [fwd=3.133s bwd=1.294s opt=0.003s] dispatch=4.740s [2026-10-18 14:06:55,457] loop.py:45 INFO TRAIN_STEP step=265 loss=5.249518 Training: 1%|▏ | 265/18018 [36:41<42:12:05, 8.56s/step, loss=5.250, acc=0.224, acc_len=0.45, thru=14.1, I=14.7, T=28.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:06:55,834] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:07:05,095] loop.py:45 INFO TRAIN_STEP step=266 loss=5.118197 Training: 1%|▏ | 266/18018 [36:51<43:47:47, 8.88s/step, loss=5.118, acc=0.245, acc_len=0.50, thru=9.1, I=8.6, T=30.3, wait=5.4s, pool=20, epoch=1/2] (TrainerActor pid=9310) [2026-10-18 14:07:05,091] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=265: forward=2986.7ms backward=1198.9ms (TrainerActor pid=9370) [2026-10-18 14:07:05,132] eagle_store.py:517 WARNING batch_get_buffer missing keys (k8519-b816450c_hs, k8519-b816450c_ids, k8519-b816450c_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:15,316] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9371) [2026-10-18 14:07:05,142] eagle_store.py:517 WARNING batch_get_buffer missing keys (k8518-9ca990e5_hs, k8518-9ca990e5_ids, k8518-9ca990e5_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-18 14:07:17,595] loop.py:45 INFO TRAIN_STEP step=267 loss=4.827014 Training: 1%|▏ | 267/18018 [37:03<49:09:02, 9.97s/step, loss=4.827, acc=0.256, acc_len=0.53, thru=13.0, I=15.2, T=25.7, wait=7.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:17,829] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:24,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:07:25,083] loop.py:45 INFO TRAIN_STEP step=268 loss=5.106143 Training: 1%|▏ | 268/18018 [37:11<45:28:33, 9.22s/step, loss=5.106, acc=0.242, acc_len=0.49, thru=13.7, I=19.2, T=33.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:25,112] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:32,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:07:33,109] loop.py:45 INFO TRAIN_STEP step=269 loss=5.200009 Training: 1%|▏ | 269/18018 [37:19<43:42:09, 8.86s/step, loss=5.200, acc=0.232, acc_len=0.47, thru=16.8, I=19.6, T=31.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:33,396] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:37,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:07:39,038] loop.py:380 INFO TIMING step=270: step=3.648s data=1.563s compute=3.613s [fwd=2.598s bwd=1.001s opt=0.003s] dispatch=2.272s [2026-10-18 14:07:39,038] loop.py:45 INFO TRAIN_STEP step=270 loss=5.144620 Training: 1%|▏ | 270/18018 [37:25<39:21:35, 7.98s/step, loss=5.145, acc=0.234, acc_len=0.47, thru=19.3, I=25.9, T=35.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:39,265] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:45,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:07:46,094] loop.py:45 INFO TRAIN_STEP step=271 loss=5.155900 Training: 2%|▏ | 271/18018 [37:32<37:59:05, 7.71s/step, loss=5.156, acc=0.239, acc_len=0.48, thru=16.8, I=20.9, T=34.0, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:07:46,089] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=270: forward=2662.9ms backward=1054.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:46,428] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:53,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:07:53,796] loop.py:45 INFO TRAIN_STEP step=272 loss=4.957912 Training: 2%|▏ | 272/18018 [37:39<37:58:49, 7.70s/step, loss=4.958, acc=0.255, acc_len=0.52, thru=16.9, I=18.7, T=33.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:07:54,157] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:00,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:08:00,674] loop.py:45 INFO TRAIN_STEP step=273 loss=4.986465 Training: 2%|▏ | 273/18018 [37:46<36:45:13, 7.46s/step, loss=4.986, acc=0.247, acc_len=0.51, thru=18.1, I=19.9, T=32.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:01,008] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:08:08,338] loop.py:45 INFO TRAIN_STEP step=274 loss=4.991483 Training: 2%|▏ | 274/18018 [37:54<37:03:29, 7.52s/step, loss=4.991, acc=0.246, acc_len=0.50, thru=16.2, I=17.2, T=29.6, wait=3.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:14,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:08:15,775] loop.py:380 INFO TIMING step=275: step=4.049s data=1.813s compute=4.019s [fwd=2.867s bwd=1.137s opt=0.003s] dispatch=3.380s [2026-10-18 14:08:15,775] loop.py:45 INFO TRAIN_STEP step=275 loss=5.240593 Training: 2%|▏ | 275/18018 [38:01<36:56:04, 7.49s/step, loss=5.241, acc=0.238, acc_len=0.48, thru=16.8, I=18.1, T=31.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:16,232] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:22,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:08:22,587] loop.py:45 INFO TRAIN_STEP step=276 loss=4.882572 Training: 2%|▏ | 276/18018 [38:08<35:55:28, 7.29s/step, loss=4.883, acc=0.262, acc_len=0.54, thru=17.7, I=21.2, T=32.6, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:08:22,582] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=275: forward=2721.3ms backward=1142.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:22,946] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:30,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:08:31,227] loop.py:45 INFO TRAIN_STEP step=277 loss=5.048437 Training: 2%|▏ | 277/18018 [38:17<37:55:10, 7.69s/step, loss=5.048, acc=0.245, acc_len=0.50, thru=14.9, I=16.4, T=32.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:31,703] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:36,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:08:37,779] loop.py:45 INFO TRAIN_STEP step=278 loss=5.175601 Training: 2%|▏ | 278/18018 [38:23<36:13:56, 7.35s/step, loss=5.176, acc=0.240, acc_len=0.49, thru=17.9, I=24.0, T=30.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:37,981] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:43,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:08:44,382] loop.py:45 INFO TRAIN_STEP step=279 loss=5.122178 Training: 2%|▏ | 279/18018 [38:30<35:07:18, 7.13s/step, loss=5.122, acc=0.236, acc_len=0.48, thru=17.6, I=23.1, T=31.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:44,793] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:52,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:08:54,156] loop.py:380 INFO TIMING step=280: step=4.160s data=1.658s compute=4.125s [fwd=3.023s bwd=1.088s opt=0.003s] dispatch=5.604s [2026-10-18 14:08:54,157] loop.py:45 INFO TRAIN_STEP step=280 loss=4.792323 Training: 2%|▏ | 280/18018 [38:40<39:01:55, 7.92s/step, loss=4.792, acc=0.268, acc_len=0.55, thru=13.7, I=15.2, T=30.8, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:08:54,318] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:09:00,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:09:02,050] loop.py:45 INFO TRAIN_STEP step=281 loss=5.075518 Training: 2%|▏ | 281/18018 [38:48<38:59:18, 7.91s/step, loss=5.076, acc=0.251, acc_len=0.51, thru=14.6, I=22.0, T=30.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:09:02,086] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 14:09:02,045] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=280: forward=2864.1ms backward=1305.6ms [2026-10-18 14:09:11,064] loop.py:45 INFO TRAIN_STEP step=282 loss=5.093936 Training: 2%|▏ | 282/18018 [38:57<40:36:30, 8.24s/step, loss=5.094, acc=0.239, acc_len=0.48, thru=10.3, I=11.8, T=27.3, wait=4.3s, pool=28, epoch=1/2][2026-10-18 14:09:18,912] loop.py:45 INFO TRAIN_STEP step=283 loss=4.732525 Training: 2%|▏ | 283/18018 [39:04<40:01:27, 8.12s/step, loss=4.733, acc=0.267, acc_len=0.55, thru=18.8, I=15.2, T=32.1, wait=3.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:09:24,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:09:26,251] loop.py:45 INFO TRAIN_STEP step=284 loss=4.910635 Training: 2%|▏ | 284/18018 [39:12<38:51:36, 7.89s/step, loss=4.911, acc=0.257, acc_len=0.53, thru=16.7, I=18.2, T=30.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:09:26,370] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 14:09:33,672] loop.py:380 INFO TIMING step=285: step=4.008s data=1.651s compute=3.969s [fwd=2.771s bwd=1.183s opt=0.003s] dispatch=3.405s [2026-10-18 14:09:33,673] loop.py:45 INFO TRAIN_STEP step=285 loss=5.120905 Training: 2%|▏ | 285/18018 [39:19<38:10:03, 7.75s/step, loss=5.121, acc=0.249, acc_len=0.51, thru=14.4, I=18.1, T=31.9, wait=3.4s, pool=48, epoch=1/2][2026-10-18 14:09:41,866] loop.py:45 INFO TRAIN_STEP step=286 loss=4.984416 Training: 2%|▏ | 286/18018 [39:27<38:49:19, 7.88s/step, loss=4.984, acc=0.252, acc_len=0.52, thru=14.8, I=13.7, T=33.5, wait=4.4s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:09:41,861] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=285: forward=2730.8ms backward=1034.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:09:47,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:09:48,612] loop.py:45 INFO TRAIN_STEP step=287 loss=4.929179 Training: 2%|▏ | 287/18018 [39:34<37:08:49, 7.54s/step, loss=4.929, acc=0.252, acc_len=0.51, thru=18.9, I=20.3, T=31.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:09:48,841] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:09:54,569] loop.py:45 INFO TRAIN_STEP step=288 loss=4.819750 Training: 2%|▏ | 288/18018 [39:40<34:48:02, 7.07s/step, loss=4.820, acc=0.271, acc_len=0.56, thru=17.2, I=19.2, T=37.8, wait=2.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:01,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 14:10:01,708] loop.py:45 INFO TRAIN_STEP step=289 loss=4.881805 Training: 2%|▏ | 289/18018 [39:47<34:54:26, 7.09s/step, loss=4.882, acc=0.269, acc_len=0.55, thru=20.0, I=18.8, T=32.5, wait=3.2s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:01,993] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=4/64 [2026-10-18 14:10:09,510] loop.py:380 INFO TIMING step=290: step=3.766s data=1.633s compute=3.736s [fwd=2.708s bwd=1.014s opt=0.003s] dispatch=4.027s [2026-10-18 14:10:09,510] loop.py:45 INFO TRAIN_STEP step=290 loss=5.037731 Training: 2%|▏ | 290/18018 [39:55<35:57:22, 7.30s/step, loss=5.038, acc=0.254, acc_len=0.52, thru=17.9, I=17.8, T=34.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:09,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:10,009] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:16,181] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:10:16,653] loop.py:45 INFO TRAIN_STEP step=291 loss=4.922555 Training: 2%|▏ | 291/18018 [40:02<35:43:13, 7.25s/step, loss=4.923, acc=0.261, acc_len=0.54, thru=18.6, I=18.6, T=33.5, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:10:16,648] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=290: forward=2697.4ms backward=1080.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:16,700] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:23,408] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:10:23,765] loop.py:45 INFO TRAIN_STEP step=292 loss=4.857473 Training: 2%|▏ | 292/18018 [40:09<35:30:40, 7.21s/step, loss=4.857, acc=0.259, acc_len=0.53, thru=18.4, I=18.6, T=34.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:23,914] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:29,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:10:30,538] loop.py:45 INFO TRAIN_STEP step=293 loss=4.909342 Training: 2%|▏ | 293/18018 [40:16<34:51:30, 7.08s/step, loss=4.909, acc=0.259, acc_len=0.53, thru=18.8, I=20.1, T=28.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:30,659] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:35,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:10:36,302] loop.py:45 INFO TRAIN_STEP step=294 loss=4.872016 Training: 2%|▏ | 294/18018 [40:22<32:54:45, 6.69s/step, loss=4.872, acc=0.262, acc_len=0.54, thru=19.6, I=23.7, T=33.4, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:36,471] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:44,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:10:45,979] loop.py:380 INFO TIMING step=295: step=4.648s data=1.623s compute=4.613s [fwd=3.383s bwd=1.215s opt=0.003s] dispatch=5.021s [2026-10-18 14:10:45,979] loop.py:45 INFO TRAIN_STEP step=295 loss=4.973726 Training: 2%|▏ | 295/18018 [40:32<37:20:00, 7.58s/step, loss=4.974, acc=0.255, acc_len=0.52, thru=13.5, I=15.6, T=27.5, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:46,318] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:10:52,717] loop.py:45 INFO TRAIN_STEP step=296 loss=5.002693 Training: 2%|▏ | 296/18018 [40:38<36:04:48, 7.33s/step, loss=5.003, acc=0.249, acc_len=0.51, thru=16.3, I=20.9, T=32.9, wait=2.8s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:10:52,712] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=295: forward=2695.6ms backward=1146.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:10:59,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:10:59,864] loop.py:45 INFO TRAIN_STEP step=297 loss=4.854530 Training: 2%|▏ | 297/18018 [40:45<35:48:39, 7.27s/step, loss=4.855, acc=0.268, acc_len=0.55, thru=18.2, I=18.7, T=28.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:00,040] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:11:06,514] loop.py:45 INFO TRAIN_STEP step=298 loss=4.818944 Training: 2%|▏ | 298/18018 [40:52<34:52:56, 7.09s/step, loss=4.819, acc=0.266, acc_len=0.54, thru=17.6, I=18.8, T=33.5, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:13,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:11:15,009] loop.py:45 INFO TRAIN_STEP step=299 loss=4.891673 Training: 2%|▏ | 299/18018 [41:01<36:57:50, 7.51s/step, loss=4.892, acc=0.270, acc_len=0.55, thru=16.6, I=18.5, T=30.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:15,182] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:20,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:11:21,947] loop.py:380 INFO TIMING step=300: step=4.238s data=1.976s compute=4.208s [fwd=2.950s bwd=1.243s opt=0.003s] dispatch=2.690s [2026-10-18 14:11:21,948] loop.py:45 INFO TRAIN_STEP step=300 loss=4.803105 Training: 2%|▏ | 300/18018 [41:08<36:06:56, 7.34s/step, loss=4.803, acc=0.271, acc_len=0.56, thru=16.9, I=25.2, T=30.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:21,995] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 14:11:28,112] loop.py:45 INFO TRAIN_STEP step=301 loss=4.865217 Training: 2%|▏ | 301/18018 [41:14<34:22:44, 6.99s/step, loss=4.865, acc=0.255, acc_len=0.52, thru=17.5, I=21.2, T=35.7, wait=2.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:11:28,107] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=300: forward=2559.9ms backward=980.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:35,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:11:37,007] loop.py:45 INFO TRAIN_STEP step=302 loss=4.880511 Training: 2%|▏ | 302/18018 [41:23<37:11:48, 7.56s/step, loss=4.881, acc=0.267, acc_len=0.55, thru=15.5, I=16.5, T=31.5, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:37,287] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:41,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:11:43,514] loop.py:45 INFO TRAIN_STEP step=303 loss=4.735393 Training: 2%|▏ | 303/18018 [41:29<35:38:34, 7.24s/step, loss=4.735, acc=0.277, acc_len=0.57, thru=18.0, I=25.0, T=30.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:43,851] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:49,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:11:50,803] loop.py:45 INFO TRAIN_STEP step=304 loss=4.782106 Training: 2%|▏ | 304/18018 [41:36<35:42:27, 7.26s/step, loss=4.782, acc=0.266, acc_len=0.54, thru=15.3, I=19.5, T=33.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:50,882] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:58,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:11:58,928] loop.py:380 INFO TIMING step=305: step=3.947s data=1.662s compute=3.913s [fwd=2.774s bwd=1.126s opt=0.003s] dispatch=4.170s [2026-10-18 14:11:58,929] loop.py:45 INFO TRAIN_STEP step=305 loss=4.823607 Training: 2%|▏ | 305/18018 [41:45<36:59:30, 7.52s/step, loss=4.824, acc=0.265, acc_len=0.54, thru=16.1, I=19.1, T=32.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:11:59,020] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 14:12:05,055] loop.py:45 INFO TRAIN_STEP step=306 loss=4.657749 Training: 2%|▏ | 306/18018 [41:51<34:55:51, 7.10s/step, loss=4.658, acc=0.277, acc_len=0.57, thru=16.8, I=20.9, T=34.2, wait=2.4s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:12:05,050] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=305: forward=2653.4ms backward=1048.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:12:12,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:12:13,083] loop.py:45 INFO TRAIN_STEP step=307 loss=4.773681 Training: 2%|▏ | 307/18018 [41:59<36:18:12, 7.38s/step, loss=4.774, acc=0.275, acc_len=0.56, thru=18.4, I=18.1, T=28.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:12:13,392] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:12:21,369] loop.py:45 INFO TRAIN_STEP step=308 loss=4.856705 Training: 2%|▏ | 308/18018 [42:07<37:38:12, 7.65s/step, loss=4.857, acc=0.272, acc_len=0.56, thru=12.1, I=12.9, T=29.1, wait=3.9s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:12:28,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:12:30,791] loop.py:45 INFO TRAIN_STEP step=309 loss=4.993814 Training: 2%|▏ | 309/18018 [42:16<40:15:04, 8.18s/step, loss=4.994, acc=0.266, acc_len=0.55, thru=16.9, I=15.8, T=28.4, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:12:30,886] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-18 14:12:37,071] loop.py:380 INFO TIMING step=310: step=3.674s data=1.520s compute=3.647s [fwd=2.577s bwd=1.057s opt=0.003s] dispatch=2.596s [2026-10-18 14:12:37,072] loop.py:45 INFO TRAIN_STEP step=310 loss=4.750031 Training: 2%|▏ | 310/18018 [42:23<37:26:20, 7.61s/step, loss=4.750, acc=0.267, acc_len=0.55, thru=17.1, I=19.5, T=34.8, wait=2.6s, pool=60, epoch=1/2][2026-10-18 14:12:45,244] loop.py:45 INFO TRAIN_STEP step=311 loss=4.875297 Training: 2%|▏ | 311/18018 [42:31<38:15:51, 7.78s/step, loss=4.875, acc=0.268, acc_len=0.55, thru=16.1, I=15.9, T=30.2, wait=3.9s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:12:45,239] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=310: forward=3015.4ms backward=1174.6ms [2026-10-18 14:12:52,335] loop.py:45 INFO TRAIN_STEP step=312 loss=4.627792 Training: 2%|▏ | 312/18018 [42:38<37:15:00, 7.57s/step, loss=4.628, acc=0.289, acc_len=0.60, thru=17.9, I=17.2, T=30.7, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:12:58,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:13:00,785] loop.py:45 INFO TRAIN_STEP step=313 loss=4.484722 Training: 2%|▏ | 313/18018 [42:46<38:32:13, 7.84s/step, loss=4.485, acc=0.289, acc_len=0.60, thru=16.0, I=17.9, T=31.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:13:00,866] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:13:09,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:13:11,229] loop.py:45 INFO TRAIN_STEP step=314 loss=4.662237 Training: 2%|▏ | 314/18018 [42:57<42:23:13, 8.62s/step, loss=4.662, acc=0.278, acc_len=0.58, thru=12.0, I=14.0, T=24.2, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:13:11,573] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:13:20,467] loop.py:380 INFO TIMING step=315: step=4.065s data=1.727s compute=4.012s [fwd=2.769s bwd=1.230s opt=0.003s] dispatch=5.162s [2026-10-18 14:13:20,468] loop.py:45 INFO TRAIN_STEP step=315 loss=4.777241 Training: 2%|▏ | 315/18018 [43:06<43:17:35, 8.80s/step, loss=4.777, acc=0.272, acc_len=0.56, thru=13.9, I=13.4, T=31.5, wait=5.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:13:26,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:13:28,467] loop.py:45 INFO TRAIN_STEP step=316 loss=4.858932 Training: 2%|▏ | 316/18018 [43:14<42:06:31, 8.56s/step, loss=4.859, acc=0.267, acc_len=0.55, thru=16.4, I=19.2, T=32.4, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:13:28,462] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=315: forward=2678.5ms backward=1189.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:13:28,688] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:13:35,734] loop.py:45 INFO TRAIN_STEP step=317 loss=4.663144 Training: 2%|▏ | 317/18018 [43:21<40:11:23, 8.17s/step, loss=4.663, acc=0.273, acc_len=0.56, thru=13.0, I=19.5, T=28.6, wait=2.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:13:42,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:13:44,847] loop.py:45 INFO TRAIN_STEP step=318 loss=4.847272 Training: 2%|▏ | 318/18018 [43:30<41:34:27, 8.46s/step, loss=4.847, acc=0.262, acc_len=0.53, thru=16.9, I=17.4, T=29.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:13:44,973] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:13:51,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:13:52,846] loop.py:45 INFO TRAIN_STEP step=319 loss=4.551886 Training: 2%|▏ | 319/18018 [43:38<40:53:51, 8.32s/step, loss=4.552, acc=0.293, acc_len=0.61, thru=14.0, I=19.5, T=28.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:13:52,951] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:00,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:14:01,029] loop.py:380 INFO TIMING step=320: step=3.699s data=1.561s compute=3.654s [fwd=2.541s bwd=1.099s opt=0.003s] dispatch=4.476s [2026-10-18 14:14:01,030] loop.py:45 INFO TRAIN_STEP step=320 loss=4.741898 Training: 2%|▏ | 320/18018 [43:47<40:41:46, 8.28s/step, loss=4.742, acc=0.274, acc_len=0.57, thru=14.2, I=18.6, T=34.6, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:01,135] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:07,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:14:08,599] loop.py:45 INFO TRAIN_STEP step=321 loss=4.659698 Training: 2%|▏ | 321/18018 [43:54<39:38:52, 8.07s/step, loss=4.660, acc=0.284, acc_len=0.59, thru=17.7, I=22.6, T=27.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:14:08,594] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=320: forward=3205.9ms backward=1483.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:09,075] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:16,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:14:17,544] loop.py:45 INFO TRAIN_STEP step=322 loss=4.531878 Training: 2%|▏ | 322/18018 [44:03<40:56:40, 8.33s/step, loss=4.532, acc=0.297, acc_len=0.62, thru=15.1, I=17.0, T=30.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:17,998] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:23,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:14:24,675] loop.py:45 INFO TRAIN_STEP step=323 loss=4.603929 Training: 2%|▏ | 323/18018 [44:10<39:10:28, 7.97s/step, loss=4.604, acc=0.293, acc_len=0.61, thru=17.3, I=24.2, T=29.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:24,705] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:30,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:14:31,125] loop.py:45 INFO TRAIN_STEP step=324 loss=4.602134 Training: 2%|▏ | 324/18018 [44:17<36:55:58, 7.51s/step, loss=4.602, acc=0.281, acc_len=0.58, thru=18.2, I=21.8, T=35.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:31,377] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:37,690] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:14:38,784] loop.py:380 INFO TIMING step=325: step=3.855s data=1.634s compute=3.824s [fwd=2.767s bwd=1.042s opt=0.003s] dispatch=3.795s [2026-10-18 14:14:38,784] loop.py:45 INFO TRAIN_STEP step=325 loss=4.441700 Training: 2%|▏ | 325/18018 [44:24<37:08:26, 7.56s/step, loss=4.442, acc=0.295, acc_len=0.61, thru=16.5, I=20.7, T=33.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:39,199] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:14:46,347] loop.py:45 INFO TRAIN_STEP step=326 loss=4.620017 Training: 2%|▏ | 326/18018 [44:32<37:09:13, 7.56s/step, loss=4.620, acc=0.289, acc_len=0.60, thru=15.8, I=19.0, T=33.3, wait=3.7s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:14:46,342] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=325: forward=2735.1ms backward=1055.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:55,568] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:14:55,997] loop.py:45 INFO TRAIN_STEP step=327 loss=4.662344 Training: 2%|▏ | 327/18018 [44:42<40:13:36, 8.19s/step, loss=4.662, acc=0.287, acc_len=0.60, thru=13.8, I=14.4, T=26.4, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:14:56,082] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:02,008] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:15:03,373] loop.py:45 INFO TRAIN_STEP step=328 loss=4.411085 Training: 2%|▏ | 328/18018 [44:49<39:02:02, 7.94s/step, loss=4.411, acc=0.301, acc_len=0.63, thru=16.6, I=19.3, T=27.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:03,533] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:09,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:15:11,697] loop.py:45 INFO TRAIN_STEP step=329 loss=4.802452 Training: 2%|▏ | 329/18018 [44:57<39:35:36, 8.06s/step, loss=4.802, acc=0.274, acc_len=0.56, thru=14.1, I=20.4, T=27.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:11,799] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-18 14:15:19,549] loop.py:380 INFO TIMING step=330: step=4.292s data=2.023s compute=4.260s [fwd=3.057s bwd=1.189s opt=0.003s] dispatch=3.549s [2026-10-18 14:15:19,550] loop.py:45 INFO TRAIN_STEP step=330 loss=4.473885 Training: 2%|▏ | 330/18018 [45:05<39:17:03, 8.00s/step, loss=4.474, acc=0.302, acc_len=0.63, thru=10.2, I=15.7, T=29.8, wait=3.5s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:30,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:15:31,924] loop.py:45 INFO TRAIN_STEP step=331 loss=4.765279 Training: 2%|▏ | 331/18018 [45:18<45:44:11, 9.31s/step, loss=4.765, acc=0.277, acc_len=0.57, thru=14.7, I=11.5, T=29.8, wait=8.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:15:31,920] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=330: forward=3007.8ms backward=1239.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:32,375] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:38,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:15:40,166] loop.py:45 INFO TRAIN_STEP step=332 loss=4.496670 Training: 2%|▏ | 332/18018 [45:26<44:09:42, 8.99s/step, loss=4.497, acc=0.290, acc_len=0.61, thru=14.6, I=17.1, T=28.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:40,550] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:46,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:15:47,297] loop.py:45 INFO TRAIN_STEP step=333 loss=4.600761 Training: 2%|▏ | 333/18018 [45:33<41:25:13, 8.43s/step, loss=4.601, acc=0.283, acc_len=0.58, thru=15.8, I=21.6, T=31.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:47,631] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:53,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:15:54,948] loop.py:45 INFO TRAIN_STEP step=334 loss=4.512039 Training: 2%|▏ | 334/18018 [45:41<40:16:03, 8.20s/step, loss=4.512, acc=0.305, acc_len=0.64, thru=15.3, I=20.4, T=34.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:55,313] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:15:59,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:16:01,184] loop.py:380 INFO TIMING step=335: step=3.894s data=1.761s compute=3.866s [fwd=2.742s bwd=1.110s opt=0.003s] dispatch=2.334s [2026-10-18 14:16:01,184] loop.py:45 INFO TRAIN_STEP step=335 loss=4.480268 Training: 2%|▏ | 335/18018 [45:47<37:22:40, 7.61s/step, loss=4.480, acc=0.304, acc_len=0.63, thru=19.7, I=25.7, T=32.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:01,479] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:07,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:16:08,370] loop.py:45 INFO TRAIN_STEP step=336 loss=4.513395 Training: 2%|▏ | 336/18018 [45:54<36:45:05, 7.48s/step, loss=4.513, acc=0.292, acc_len=0.61, thru=17.0, I=21.7, T=30.6, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:16:08,364] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=335: forward=2848.7ms backward=1289.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:08,607] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:15,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:16:17,061] loop.py:45 INFO TRAIN_STEP step=337 loss=4.644718 Training: 2%|▏ | 337/18018 [46:03<38:31:37, 7.84s/step, loss=4.645, acc=0.286, acc_len=0.59, thru=14.6, I=18.2, T=29.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:17,385] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:22,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:16:24,499] loop.py:45 INFO TRAIN_STEP step=338 loss=4.521536 Training: 2%|▏ | 338/18018 [46:10<37:55:30, 7.72s/step, loss=4.522, acc=0.300, acc_len=0.63, thru=14.7, I=23.0, T=27.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:24,603] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:30,021] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:16:31,952] loop.py:45 INFO TRAIN_STEP step=339 loss=4.683450 Training: 2%|▏ | 339/18018 [46:18<37:31:35, 7.64s/step, loss=4.683, acc=0.277, acc_len=0.57, thru=15.0, I=21.2, T=29.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:32,041] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:39,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:16:39,623] loop.py:380 INFO TIMING step=340: step=3.755s data=1.642s compute=3.723s [fwd=2.674s bwd=1.034s opt=0.003s] dispatch=3.910s [2026-10-18 14:16:39,624] loop.py:45 INFO TRAIN_STEP step=340 loss=4.386070 Training: 2%|▏ | 340/18018 [46:25<37:34:07, 7.65s/step, loss=4.386, acc=0.305, acc_len=0.64, thru=14.1, I=16.9, T=34.1, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:39,837] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:45,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:16:46,920] loop.py:45 INFO TRAIN_STEP step=341 loss=4.489395 Training: 2%|▏ | 341/18018 [46:33<37:02:42, 7.54s/step, loss=4.489, acc=0.293, acc_len=0.61, thru=18.5, I=21.2, T=29.7, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:16:46,915] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=340: forward=3167.2ms backward=1079.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:47,276] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:53,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:16:53,970] loop.py:45 INFO TRAIN_STEP step=342 loss=4.732924 Training: 2%|▏ | 342/18018 [46:40<36:19:04, 7.40s/step, loss=4.733, acc=0.287, acc_len=0.59, thru=15.8, I=20.5, T=32.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:54,349] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:16:59,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:17:00,962] loop.py:45 INFO TRAIN_STEP step=343 loss=4.429095 Training: 2%|▏ | 343/18018 [46:47<35:42:58, 7.27s/step, loss=4.429, acc=0.293, acc_len=0.62, thru=18.1, I=23.2, T=28.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:01,352] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:09,276] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:17:09,339] loop.py:45 INFO TRAIN_STEP step=344 loss=4.297686 Training: 2%|▏ | 344/18018 [46:55<37:20:12, 7.61s/step, loss=4.298, acc=0.309, acc_len=0.65, thru=16.0, I=16.9, T=24.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:09,778] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:15,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:17:17,998] loop.py:380 INFO TIMING step=345: step=4.651s data=2.200s compute=4.620s [fwd=3.266s bwd=1.339s opt=0.003s] dispatch=4.001s [2026-10-18 14:17:17,998] loop.py:45 INFO TRAIN_STEP step=345 loss=4.595767 Training: 2%|▏ | 345/18018 [47:04<38:53:32, 7.92s/step, loss=4.596, acc=0.299, acc_len=0.63, thru=15.6, I=18.8, T=27.5, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:18,079] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 [2026-10-18 14:17:25,232] loop.py:45 INFO TRAIN_STEP step=346 loss=4.449450 Training: 2%|▏ | 346/18018 [47:11<37:52:41, 7.72s/step, loss=4.449, acc=0.305, acc_len=0.64, thru=11.3, I=17.1, T=29.1, wait=2.8s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:17:25,227] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=345: forward=3141.8ms backward=1211.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:32,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:17:33,526] loop.py:45 INFO TRAIN_STEP step=347 loss=4.298093 Training: 2%|▏ | 347/18018 [47:19<38:43:13, 7.89s/step, loss=4.298, acc=0.311, acc_len=0.65, thru=17.4, I=17.8, T=29.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:33,639] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:41,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:17:42,310] loop.py:45 INFO TRAIN_STEP step=348 loss=4.405825 Training: 2%|▏ | 348/18018 [47:28<40:02:19, 8.16s/step, loss=4.406, acc=0.313, acc_len=0.66, thru=14.8, I=14.8, T=22.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:42,483] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:49,408] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:17:51,809] loop.py:45 INFO TRAIN_STEP step=349 loss=4.441457 Training: 2%|▏ | 349/18018 [47:37<42:00:53, 8.56s/step, loss=4.441, acc=0.304, acc_len=0.63, thru=13.4, I=18.0, T=26.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:51,944] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:56,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:17:59,179] loop.py:380 INFO TIMING step=350: step=4.949s data=2.518s compute=4.914s [fwd=3.397s bwd=1.503s opt=0.003s] dispatch=2.407s [2026-10-18 14:17:59,179] loop.py:45 INFO TRAIN_STEP step=350 loss=4.370047 Training: 2%|▏ | 350/18018 [47:45<40:15:35, 8.20s/step, loss=4.370, acc=0.312, acc_len=0.66, thru=14.0, I=25.8, T=25.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:17:59,601] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:05,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP3 pid=2022) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP2 pid=2021) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP3 pid=2022) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP2 pid=2021) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP1 pid=2020) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP1 pid=2020) exe_args, func_types, adapted_args = self.generate_mlir_function_types( (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl/dsl_packages/cutlass/base_dsl/dsl.py:2309: UserWarning: Argument aux_data (position 19) cannot be converted to a JitArgument for function 'cutlass___call___vllmvllm_flash_attncuteflash_fwd_sm100FlashAttentionForwardSm100_object_at__Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmemoi64i641_Tensorgmem_'. Its type `` is not supported. If the value is known at compile time, annotate it: `aux_data: Constexpr`. Otherwise, implement the `JitArgument` or `DynamicExpression` protocol for type ``, or register a custom argument adapter. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) exe_args, func_types, adapted_args = self.generate_mlir_function_types( [2026-10-18 14:18:06,836] loop.py:45 INFO TRAIN_STEP step=351 loss=4.506801 Training: 2%|▏ | 351/18018 [47:52<39:27:12, 8.04s/step, loss=4.507, acc=0.307, acc_len=0.64, thru=17.7, I=20.5, T=28.6, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:18:06,831] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=350: forward=3164.7ms backward=1264.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:07,231] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:16,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:18:18,555] loop.py:45 INFO TRAIN_STEP step=352 loss=4.497298 Training: 2%|▏ | 352/18018 [48:04<44:51:51, 9.14s/step, loss=4.497, acc=0.296, acc_len=0.62, thru=11.8, I=12.8, T=27.8, wait=7.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:18,734] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:23,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:18:24,674] loop.py:45 INFO TRAIN_STEP step=353 loss=4.422715 Training: 2%|▏ | 353/18018 [48:10<40:24:47, 8.24s/step, loss=4.423, acc=0.313, acc_len=0.66, thru=17.3, I=27.9, T=32.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:25,155] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:18:31,722] loop.py:45 INFO TRAIN_STEP step=354 loss=4.493266 Training: 2%|▏ | 354/18018 [48:17<38:39:42, 7.88s/step, loss=4.493, acc=0.299, acc_len=0.63, thru=16.4, I=21.5, T=32.9, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:40,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:18:40,981] loop.py:380 INFO TIMING step=355: step=4.519s data=2.208s compute=4.487s [fwd=3.203s bwd=1.270s opt=0.003s] dispatch=4.731s [2026-10-18 14:18:40,981] loop.py:45 INFO TRAIN_STEP step=355 loss=4.404696 Training: 2%|▏ | 355/18018 [48:27<40:41:12, 8.29s/step, loss=4.405, acc=0.301, acc_len=0.63, thru=13.9, I=13.9, T=28.3, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:41,113] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:47,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:18:48,412] loop.py:45 INFO TRAIN_STEP step=356 loss=4.634755 Training: 2%|▏ | 356/18018 [48:34<39:25:04, 8.03s/step, loss=4.635, acc=0.295, acc_len=0.61, thru=16.8, I=18.1, T=33.8, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:18:48,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=355: forward=2658.3ms backward=1079.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:48,867] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:54,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:18:55,043] loop.py:45 INFO TRAIN_STEP step=357 loss=4.502819 Training: 2%|▏ | 357/18018 [48:41<37:21:06, 7.61s/step, loss=4.503, acc=0.304, acc_len=0.63, thru=18.9, I=22.7, T=32.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:18:55,411] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:01,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:19:03,127] loop.py:45 INFO TRAIN_STEP step=358 loss=4.345465 Training: 2%|▏ | 358/18018 [48:49<38:02:21, 7.75s/step, loss=4.345, acc=0.315, acc_len=0.67, thru=15.2, I=18.7, T=34.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:03,611] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:09,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:19:10,986] loop.py:45 INFO TRAIN_STEP step=359 loss=4.309777 Training: 2%|▏ | 359/18018 [48:57<38:11:43, 7.79s/step, loss=4.310, acc=0.314, acc_len=0.66, thru=15.7, I=22.9, T=28.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:11,238] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:18,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:19:19,157] loop.py:380 INFO TIMING step=360: step=4.052s data=1.844s compute=4.010s [fwd=2.851s bwd=1.145s opt=0.003s] dispatch=4.109s [2026-10-18 14:19:19,157] loop.py:45 INFO TRAIN_STEP step=360 loss=4.205384 Training: 2%|▏ | 360/18018 [49:05<38:45:22, 7.90s/step, loss=4.205, acc=0.316, acc_len=0.66, thru=14.6, I=17.6, T=31.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:19,591] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:27,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 14:19:28,875] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=360: forward=3142.7ms backward=1260.5ms [2026-10-18 14:19:28,880] loop.py:45 INFO TRAIN_STEP step=361 loss=4.508022 Training: 2%|▏ | 361/18018 [49:14<41:26:01, 8.45s/step, loss=4.508, acc=0.301, acc_len=0.63, thru=13.8, I=14.5, T=28.7, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:29,163] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:19:35,651] loop.py:45 INFO TRAIN_STEP step=362 loss=4.499229 Training: 2%|▏ | 362/18018 [49:21<38:58:12, 7.95s/step, loss=4.499, acc=0.300, acc_len=0.62, thru=16.2, I=23.2, T=30.8, wait=2.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:42,767] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:19:43,584] loop.py:45 INFO TRAIN_STEP step=363 loss=4.379915 Training: 2%|▏ | 363/18018 [49:29<38:56:39, 7.94s/step, loss=4.380, acc=0.311, acc_len=0.65, thru=15.6, I=16.0, T=31.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:43,772] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:50,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:19:53,427] loop.py:45 INFO TRAIN_STEP step=364 loss=4.410893 Training: 2%|▏ | 364/18018 [49:39<41:44:23, 8.51s/step, loss=4.411, acc=0.311, acc_len=0.65, thru=13.4, I=16.4, T=28.6, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:19:53,473] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 [2026-10-18 14:19:59,331] loop.py:380 INFO TIMING step=365: step=4.003s data=1.917s compute=3.969s [fwd=2.759s bwd=1.195s opt=0.003s] dispatch=1.894s [2026-10-18 14:19:59,331] loop.py:45 INFO TRAIN_STEP step=365 loss=4.128307 Training: 2%|▏ | 365/18018 [49:45<37:54:11, 7.73s/step, loss=4.128, acc=0.333, acc_len=0.71, thru=16.6, I=24.0, T=32.0, wait=1.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:07,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:20:08,156] loop.py:45 INFO TRAIN_STEP step=366 loss=4.146953 Training: 2%|▏ | 366/18018 [49:54<39:30:40, 8.06s/step, loss=4.147, acc=0.338, acc_len=0.72, thru=16.1, I=16.6, T=32.1, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:20:08,151] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=365: forward=2790.6ms backward=1149.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:08,190] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:14,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:20:16,463] loop.py:45 INFO TRAIN_STEP step=367 loss=4.618260 Training: 2%|▏ | 367/18018 [50:02<39:52:22, 8.13s/step, loss=4.618, acc=0.286, acc_len=0.58, thru=16.3, I=19.5, T=30.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:16,747] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:22,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:20:22,873] loop.py:45 INFO TRAIN_STEP step=368 loss=4.419880 Training: 2%|▏ | 368/18018 [50:08<37:20:17, 7.62s/step, loss=4.420, acc=0.310, acc_len=0.65, thru=18.7, I=23.0, T=31.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:23,011] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:27,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:20:29,036] loop.py:45 INFO TRAIN_STEP step=369 loss=4.433242 Training: 2%|▏ | 369/18018 [50:15<35:12:10, 7.18s/step, loss=4.433, acc=0.304, acc_len=0.63, thru=19.2, I=21.9, T=34.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:29,383] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:34,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:20:35,675] loop.py:380 INFO TIMING step=370: step=3.705s data=1.676s compute=3.673s [fwd=2.643s bwd=1.015s opt=0.003s] dispatch=2.924s [2026-10-18 14:20:35,675] loop.py:45 INFO TRAIN_STEP step=370 loss=4.144148 Training: 2%|▏ | 370/18018 [50:21<34:24:01, 7.02s/step, loss=4.144, acc=0.332, acc_len=0.71, thru=18.0, I=23.6, T=34.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:35,964] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:20:43,179] loop.py:45 INFO TRAIN_STEP step=371 loss=4.493430 Training: 2%|▏ | 371/18018 [50:29<35:06:54, 7.16s/step, loss=4.493, acc=0.301, acc_len=0.62, thru=15.1, I=18.6, T=33.4, wait=3.7s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:20:43,174] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=370: forward=2681.1ms backward=1112.1ms [2026-10-18 14:20:49,804] loop.py:45 INFO TRAIN_STEP step=372 loss=4.355415 Training: 2%|▏ | 372/18018 [50:35<34:19:15, 7.00s/step, loss=4.355, acc=0.320, acc_len=0.66, thru=16.3, I=14.9, T=32.2, wait=2.6s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:57,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:20:58,432] loop.py:45 INFO TRAIN_STEP step=373 loss=4.507675 Training: 2%|▏ | 373/18018 [50:44<36:42:39, 7.49s/step, loss=4.508, acc=0.306, acc_len=0.64, thru=18.5, I=16.4, T=28.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:20:58,461] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:03,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:21:06,159] loop.py:45 INFO TRAIN_STEP step=374 loss=4.323714 Training: 2%|▏ | 374/18018 [50:52<37:03:19, 7.56s/step, loss=4.324, acc=0.319, acc_len=0.67, thru=17.2, I=20.2, T=28.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:06,267] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:11,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:21:11,464] loop.py:380 INFO TIMING step=375: step=3.529s data=1.541s compute=3.491s [fwd=2.479s bwd=0.997s opt=0.003s] dispatch=1.768s [2026-10-18 14:21:11,464] loop.py:45 INFO TRAIN_STEP step=375 loss=4.167665 Training: 2%|▏ | 375/18018 [50:57<33:44:27, 6.88s/step, loss=4.168, acc=0.333, acc_len=0.70, thru=19.8, I=25.9, T=36.3, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:11,550] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:17,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:21:18,748] loop.py:45 INFO TRAIN_STEP step=376 loss=4.414034 Training: 2%|▏ | 376/18018 [51:04<34:19:24, 7.00s/step, loss=4.414, acc=0.308, acc_len=0.64, thru=18.3, I=21.7, T=33.0, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:21:18,744] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=375: forward=2751.0ms backward=1085.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:18,932] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:24,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:21:25,236] loop.py:45 INFO TRAIN_STEP step=377 loss=4.037084 Training: 2%|▏ | 377/18018 [51:11<33:33:56, 6.85s/step, loss=4.037, acc=0.341, acc_len=0.73, thru=19.0, I=23.6, T=34.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:25,632] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:33,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:21:35,606] loop.py:45 INFO TRAIN_STEP step=378 loss=4.168038 Training: 2%|▏ | 378/18018 [51:21<38:44:19, 7.91s/step, loss=4.168, acc=0.333, acc_len=0.70, thru=12.8, I=15.0, T=25.9, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:35,807] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:42,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:21:43,058] loop.py:45 INFO TRAIN_STEP step=379 loss=4.504890 Training: 2%|▏ | 379/18018 [51:29<38:03:57, 7.77s/step, loss=4.505, acc=0.308, acc_len=0.64, thru=14.0, I=17.0, T=30.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:43,108] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:49,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:21:51,060] loop.py:380 INFO TIMING step=380: step=4.213s data=1.951s compute=4.183s [fwd=2.982s bwd=1.186s opt=0.003s] dispatch=3.781s [2026-10-18 14:21:51,060] loop.py:45 INFO TRAIN_STEP step=380 loss=4.416435 Training: 2%|▏ | 380/18018 [51:37<38:24:21, 7.84s/step, loss=4.416, acc=0.311, acc_len=0.65, thru=15.4, I=21.6, T=30.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:51,336] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:57,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:21:57,542] loop.py:45 INFO TRAIN_STEP step=381 loss=4.119121 Training: 2%|▏ | 381/18018 [51:43<36:24:37, 7.43s/step, loss=4.119, acc=0.336, acc_len=0.71, thru=17.4, I=23.3, T=32.3, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:21:57,537] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=380: forward=2681.6ms backward=1233.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:21:57,793] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:22:04,724] loop.py:45 INFO TRAIN_STEP step=382 loss=4.331885 Training: 2%|▏ | 382/18018 [51:50<36:02:36, 7.36s/step, loss=4.332, acc=0.319, acc_len=0.67, thru=17.0, I=18.5, T=26.8, wait=2.4s, pool=60, epoch=1/2][2026-10-18 14:22:12,961] loop.py:45 INFO TRAIN_STEP step=383 loss=4.273961 Training: 2%|▏ | 383/18018 [51:59<37:19:53, 7.62s/step, loss=4.274, acc=0.328, acc_len=0.69, thru=11.2, I=13.9, T=27.8, wait=3.6s, pool=28, epoch=1/2](TrainerActor pid=9370) [2026-10-18 14:22:13,026] eagle_store.py:517 WARNING batch_get_buffer missing keys (k12259-9ada0628_hs, k12259-9ada0628_ids, k12259-9ada0628_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:22:23,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9371) [2026-10-18 14:22:13,024] eagle_store.py:517 WARNING batch_get_buffer missing keys (k12258-8b214430_hs, k12258-8b214430_ids, k12258-8b214430_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-18 14:22:25,360] loop.py:45 INFO TRAIN_STEP step=384 loss=4.396835 Training: 2%|▏ | 384/18018 [52:11<44:21:14, 9.05s/step, loss=4.397, acc=0.315, acc_len=0.66, thru=13.5, I=17.1, T=28.7, wait=7.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:22:25,563] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:22:30,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:22:31,200] loop.py:380 INFO TIMING step=385: step=3.878s data=1.685s compute=3.841s [fwd=2.659s bwd=1.167s opt=0.003s] dispatch=1.951s [2026-10-18 14:22:31,200] loop.py:45 INFO TRAIN_STEP step=385 loss=4.297085 Training: 2%|▏ | 385/18018 [52:17<39:37:36, 8.09s/step, loss=4.297, acc=0.328, acc_len=0.69, thru=18.5, I=27.1, T=33.0, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:22:31,630] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:22:37,560] loop.py:45 INFO TRAIN_STEP step=386 loss=3.954967 Training: 2%|▏ | 386/18018 [52:23<37:04:51, 7.57s/step, loss=3.955, acc=0.349, acc_len=0.74, thru=15.6, I=20.9, T=32.8, wait=2.4s, pool=28, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:22:37,556] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=385: forward=2708.9ms backward=1138.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:22:48,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:22:48,779] loop.py:45 INFO TRAIN_STEP step=387 loss=4.355574 Training: 2%|▏ | 387/18018 [52:34<42:26:11, 8.66s/step, loss=4.356, acc=0.324, acc_len=0.68, thru=14.5, I=15.7, T=24.0, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:22:49,073] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:22:55,913] loop.py:45 INFO TRAIN_STEP step=388 loss=4.382347 Training: 2%|▏ | 388/18018 [52:41<40:11:07, 8.21s/step, loss=4.382, acc=0.320, acc_len=0.67, thru=16.7, I=17.3, T=31.4, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:02,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:23:05,090] loop.py:45 INFO TRAIN_STEP step=389 loss=4.391429 Training: 2%|▏ | 389/18018 [52:51<41:36:32, 8.50s/step, loss=4.391, acc=0.303, acc_len=0.62, thru=14.9, I=19.8, T=26.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:05,323] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:09,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:23:11,306] loop.py:380 INFO TIMING step=390: step=3.850s data=1.661s compute=3.821s [fwd=2.673s bwd=1.134s opt=0.003s] dispatch=2.358s [2026-10-18 14:23:11,306] loop.py:45 INFO TRAIN_STEP step=390 loss=4.061675 Training: 2%|▏ | 390/18018 [52:57<38:15:36, 7.81s/step, loss=4.062, acc=0.338, acc_len=0.72, thru=16.7, I=29.1, T=33.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:11,610] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:17,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:23:18,634] loop.py:45 INFO TRAIN_STEP step=391 loss=4.115331 Training: 2%|▏ | 391/18018 [53:04<37:32:40, 7.67s/step, loss=4.115, acc=0.338, acc_len=0.71, thru=16.4, I=24.2, T=28.9, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:23:18,630] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=390: forward=3096.6ms backward=1282.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:18,938] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:24,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:23:26,043] loop.py:45 INFO TRAIN_STEP step=392 loss=4.597173 Training: 2%|▏ | 392/18018 [53:12<37:09:30, 7.59s/step, loss=4.597, acc=0.290, acc_len=0.59, thru=15.9, I=19.2, T=29.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:26,498] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:32,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:23:35,154] loop.py:45 INFO TRAIN_STEP step=393 loss=3.855204 Training: 2%|▏ | 393/18018 [53:21<39:23:29, 8.05s/step, loss=3.855, acc=0.350, acc_len=0.75, thru=14.8, I=18.5, T=29.5, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:35,392] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:41,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:23:42,855] loop.py:45 INFO TRAIN_STEP step=394 loss=4.364533 Training: 2%|▏ | 394/18018 [53:28<38:52:55, 7.94s/step, loss=4.365, acc=0.315, acc_len=0.66, thru=13.6, I=21.1, T=28.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:43,199] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:48,839] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:23:49,644] loop.py:380 INFO TIMING step=395: step=3.750s data=1.614s compute=3.715s [fwd=2.617s bwd=1.084s opt=0.003s] dispatch=3.030s [2026-10-18 14:23:49,644] loop.py:45 INFO TRAIN_STEP step=395 loss=4.208299 Training: 2%|▏ | 395/18018 [53:35<37:11:17, 7.60s/step, loss=4.208, acc=0.331, acc_len=0.70, thru=16.1, I=22.0, T=34.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:23:49,853] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:23:57,488] loop.py:45 INFO TRAIN_STEP step=396 loss=4.138024 Training: 2%|▏ | 396/18018 [53:43<37:33:04, 7.67s/step, loss=4.138, acc=0.335, acc_len=0.70, thru=15.5, I=15.0, T=28.5, wait=3.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:23:57,485] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=395: forward=3159.6ms backward=1292.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:06,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:24:08,947] loop.py:45 INFO TRAIN_STEP step=397 loss=4.256989 Training: 2%|▏ | 397/18018 [53:55<43:06:26, 8.81s/step, loss=4.257, acc=0.325, acc_len=0.68, thru=11.0, I=13.4, T=28.1, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:09,345] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 14:24:15,935] loop.py:45 INFO TRAIN_STEP step=398 loss=4.027150 Training: 2%|▏ | 398/18018 [54:02<40:26:07, 8.26s/step, loss=4.027, acc=0.352, acc_len=0.75, thru=14.9, I=19.8, T=34.0, wait=3.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:24,499] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:24:25,101] loop.py:45 INFO TRAIN_STEP step=399 loss=4.185240 Training: 2%|▏ | 399/18018 [54:11<41:45:36, 8.53s/step, loss=4.185, acc=0.335, acc_len=0.70, thru=14.5, I=14.5, T=28.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:25,503] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:32,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:24:32,393] loop.py:380 INFO TIMING step=400: step=4.300s data=2.018s compute=4.263s [fwd=3.056s bwd=1.193s opt=0.003s] dispatch=2.984s [2026-10-18 14:24:32,393] loop.py:45 INFO TRAIN_STEP step=400 loss=3.959841 Training: 2%|▏ | 400/18018 [54:18<39:56:18, 8.16s/step, loss=3.960, acc=0.354, acc_len=0.76, thru=16.5, I=18.5, T=29.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:32,604] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:39,213] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:24:39,344] loop.py:45 INFO TRAIN_STEP step=401 loss=4.055505 Training: 2%|▏ | 401/18018 [54:25<38:09:27, 7.80s/step, loss=4.056, acc=0.342, acc_len=0.72, thru=16.3, I=16.2, T=29.5, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:24:39,340] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=400: forward=3036.1ms backward=1239.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:39,714] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:44,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:24:46,417] loop.py:45 INFO TRAIN_STEP step=402 loss=4.012151 Training: 2%|▏ | 402/18018 [54:32<37:05:29, 7.58s/step, loss=4.012, acc=0.359, acc_len=0.76, thru=16.9, I=18.4, T=26.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:24:46,824] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:24:53,464] loop.py:45 INFO TRAIN_STEP step=403 loss=4.192682 Training: 2%|▏ | 403/18018 [54:39<36:18:27, 7.42s/step, loss=4.193, acc=0.334, acc_len=0.70, thru=13.7, I=16.3, T=31.6, wait=3.0s, pool=44, epoch=1/2][2026-10-18 14:25:00,483] loop.py:45 INFO TRAIN_STEP step=404 loss=3.852786 Training: 2%|▏ | 404/18018 [54:46<35:43:11, 7.30s/step, loss=3.853, acc=0.367, acc_len=0.79, thru=16.4, I=16.4, T=36.0, wait=3.5s, pool=52, epoch=1/2][2026-10-18 14:25:09,711] loop.py:380 INFO TIMING step=405: step=3.876s data=1.565s compute=3.824s [fwd=2.692s bwd=1.117s opt=0.003s] dispatch=5.342s [2026-10-18 14:25:09,711] loop.py:45 INFO TRAIN_STEP step=405 loss=4.192956 Training: 2%|▏ | 405/18018 [54:55<38:32:51, 7.88s/step, loss=4.193, acc=0.324, acc_len=0.67, thru=14.9, I=14.6, T=33.0, wait=5.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:25:16,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:25:16,979] loop.py:45 INFO TRAIN_STEP step=406 loss=4.233407 Training: 2%|▏ | 406/18018 [55:03<37:38:43, 7.69s/step, loss=4.233, acc=0.332, acc_len=0.69, thru=17.5, I=17.5, T=34.9, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:25:16,973] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=405: forward=2520.2ms backward=1081.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:25:17,138] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:25:23,877] loop.py:45 INFO TRAIN_STEP step=407 loss=3.835453 Training: 2%|▏ | 407/18018 [55:09<36:28:23, 7.46s/step, loss=3.835, acc=0.361, acc_len=0.78, thru=17.9, I=17.8, T=32.7, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:25:33,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:25:34,693] loop.py:45 INFO TRAIN_STEP step=408 loss=4.029493 Training: 2%|▏ | 408/18018 [55:20<41:24:08, 8.46s/step, loss=4.029, acc=0.344, acc_len=0.73, thru=13.4, I=13.9, T=25.9, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:25:35,159] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:25:40,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:25:41,446] loop.py:45 INFO TRAIN_STEP step=409 loss=3.985850 Training: 2%|▏ | 409/18018 [55:27<38:53:18, 7.95s/step, loss=3.986, acc=0.354, acc_len=0.75, thru=16.5, I=23.1, T=34.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:25:41,940] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:25:47,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:25:48,691] loop.py:380 INFO TIMING step=410: step=4.052s data=1.728s compute=4.005s [fwd=2.733s bwd=1.257s opt=0.003s] dispatch=3.186s [2026-10-18 14:25:48,692] loop.py:45 INFO TRAIN_STEP step=410 loss=4.251651 Training: 2%|▏ | 410/18018 [55:34<37:51:07, 7.74s/step, loss=4.252, acc=0.339, acc_len=0.71, thru=16.8, I=20.6, T=31.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:25:48,910] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:25:59,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 14:26:00,059] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=410: forward=3358.3ms backward=1356.1ms [2026-10-18 14:26:00,064] loop.py:45 INFO TRAIN_STEP step=411 loss=4.270576 Training: 2%|▏ | 411/18018 [55:46<43:11:05, 8.83s/step, loss=4.271, acc=0.321, acc_len=0.67, thru=11.9, I=11.5, T=26.7, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:26:00,424] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:26:04,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:26:06,432] loop.py:45 INFO TRAIN_STEP step=412 loss=3.955622 Training: 2%|▏ | 412/18018 [55:52<39:34:03, 8.09s/step, loss=3.956, acc=0.362, acc_len=0.77, thru=18.6, I=21.8, T=32.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:26:06,734] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:26:13,807] loop.py:45 INFO TRAIN_STEP step=413 loss=4.133149 Training: 2%|▏ | 413/18018 [55:59<38:30:58, 7.88s/step, loss=4.133, acc=0.338, acc_len=0.71, thru=11.8, I=17.6, T=34.3, wait=3.6s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:26:21,646] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:26:23,303] loop.py:45 INFO TRAIN_STEP step=414 loss=3.985088 Training: 2%|▏ | 414/18018 [56:09<40:53:23, 8.36s/step, loss=3.985, acc=0.349, acc_len=0.74, thru=17.3, I=17.0, T=32.1, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:26:23,660] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:26:28,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:26:29,564] loop.py:380 INFO TIMING step=415: step=3.949s data=1.527s compute=3.910s [fwd=2.639s bwd=1.256s opt=0.003s] dispatch=2.304s [2026-10-18 14:26:29,564] loop.py:45 INFO TRAIN_STEP step=415 loss=4.117328 Training: 2%|▏ | 415/18018 [56:15<37:48:18, 7.73s/step, loss=4.117, acc=0.332, acc_len=0.70, thru=18.9, I=27.4, T=32.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:26:30,029] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:26:35,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:26:36,709] loop.py:45 INFO TRAIN_STEP step=416 loss=3.856195 Training: 2%|▏ | 416/18018 [56:22<36:56:34, 7.56s/step, loss=3.856, acc=0.355, acc_len=0.76, thru=15.5, I=19.9, T=31.8, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:26:36,704] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=415: forward=2812.1ms backward=1160.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:26:36,974] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:26:45,149] loop.py:45 INFO TRAIN_STEP step=417 loss=4.245017 Training: 2%|▏ | 417/18018 [56:31<38:14:19, 7.82s/step, loss=4.245, acc=0.322, acc_len=0.67, thru=10.9, I=12.7, T=26.6, wait=3.6s, pool=36, epoch=1/2][2026-10-18 14:26:53,584] loop.py:45 INFO TRAIN_STEP step=418 loss=4.128896 Training: 2%|▏ | 418/18018 [56:39<39:08:06, 8.00s/step, loss=4.129, acc=0.336, acc_len=0.71, thru=15.6, I=16.0, T=29.9, wait=4.1s, pool=40, epoch=1/2][2026-10-18 14:27:03,299] loop.py:45 INFO TRAIN_STEP step=419 loss=4.335871 Training: 2%|▏ | 419/18018 [56:49<41:38:37, 8.52s/step, loss=4.336, acc=0.310, acc_len=0.64, thru=15.0, I=12.2, T=31.0, wait=5.6s, pool=56, epoch=1/2][2026-10-18 14:27:10,656] loop.py:380 INFO TIMING step=420: step=3.790s data=1.598s compute=3.761s [fwd=2.670s bwd=1.078s opt=0.003s] dispatch=3.557s [2026-10-18 14:27:10,656] loop.py:45 INFO TRAIN_STEP step=420 loss=4.203585 Training: 2%|▏ | 420/18018 [56:56<39:56:11, 8.17s/step, loss=4.204, acc=0.332, acc_len=0.69, thru=16.0, I=16.7, T=33.8, wait=3.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:27:20,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 14:27:21,862] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=420: forward=3236.3ms backward=1253.6ms [2026-10-18 14:27:21,867] loop.py:45 INFO TRAIN_STEP step=421 loss=4.043877 Training: 2%|▏ | 421/18018 [57:07<44:23:38, 9.08s/step, loss=4.044, acc=0.344, acc_len=0.73, thru=13.4, I=15.2, T=28.2, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:27:22,307] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:27:28,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:27:29,065] loop.py:45 INFO TRAIN_STEP step=422 loss=3.944420 Training: 2%|▏ | 422/18018 [57:15<41:37:42, 8.52s/step, loss=3.944, acc=0.352, acc_len=0.76, thru=15.6, I=20.5, T=33.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:27:29,221] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-18 14:27:38,241] loop.py:45 INFO TRAIN_STEP step=423 loss=4.060469 Training: 2%|▏ | 423/18018 [57:24<42:35:31, 8.71s/step, loss=4.060, acc=0.342, acc_len=0.72, thru=11.6, I=12.9, T=26.7, wait=4.4s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:27:46,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:27:48,585] loop.py:45 INFO TRAIN_STEP step=424 loss=4.107656 Training: 2%|▏ | 424/18018 [57:34<44:58:59, 9.20s/step, loss=4.108, acc=0.339, acc_len=0.71, thru=15.2, I=16.4, T=26.0, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:27:48,823] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:27:56,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:27:57,121] loop.py:380 INFO TIMING step=425: step=4.127s data=1.838s compute=4.096s [fwd=2.899s bwd=1.183s opt=0.003s] dispatch=4.398s [2026-10-18 14:27:57,121] loop.py:45 INFO TRAIN_STEP step=425 loss=4.166822 Training: 2%|▏ | 425/18018 [57:43<43:59:48, 9.00s/step, loss=4.167, acc=0.331, acc_len=0.70, thru=15.4, I=14.8, T=31.0, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:27:57,225] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:05,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:28:06,784] loop.py:45 INFO TRAIN_STEP step=426 loss=4.063900 Training: 2%|▏ | 426/18018 [57:52<44:57:52, 9.20s/step, loss=4.064, acc=0.350, acc_len=0.74, thru=13.4, I=14.8, T=28.9, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:28:06,779] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=425: forward=3143.1ms backward=1238.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:07,178] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:12,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:28:13,414] loop.py:45 INFO TRAIN_STEP step=427 loss=4.010778 Training: 2%|▏ | 427/18018 [57:59<41:11:23, 8.43s/step, loss=4.011, acc=0.352, acc_len=0.75, thru=18.1, I=22.1, T=33.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:13,872] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:28:20,345] loop.py:45 INFO TRAIN_STEP step=428 loss=3.769929 Training: 2%|▏ | 428/18018 [58:06<38:59:25, 7.98s/step, loss=3.770, acc=0.362, acc_len=0.78, thru=17.1, I=18.4, T=31.7, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:29,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:28:30,148] loop.py:45 INFO TRAIN_STEP step=429 loss=3.947366 Training: 2%|▏ | 429/18018 [58:16<41:39:48, 8.53s/step, loss=3.947, acc=0.360, acc_len=0.76, thru=14.2, I=15.2, T=28.0, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:30,546] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:35,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:28:36,639] loop.py:380 INFO TIMING step=430: step=3.531s data=1.425s compute=3.497s [fwd=2.486s bwd=0.999s opt=0.003s] dispatch=2.951s [2026-10-18 14:28:36,640] loop.py:45 INFO TRAIN_STEP step=430 loss=4.028769 Training: 2%|▏ | 430/18018 [58:22<38:40:25, 7.92s/step, loss=4.029, acc=0.346, acc_len=0.72, thru=18.2, I=22.1, T=36.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:37,088] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:28:44,528] loop.py:45 INFO TRAIN_STEP step=431 loss=3.736951 Training: 2%|▏ | 431/18018 [58:30<38:37:56, 7.91s/step, loss=3.737, acc=0.374, acc_len=0.81, thru=13.9, I=18.5, T=31.7, wait=3.8s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:28:44,524] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=430: forward=2826.4ms backward=1169.7ms [2026-10-18 14:28:54,090] loop.py:45 INFO TRAIN_STEP step=432 loss=4.004654 Training: 2%|▏ | 432/18018 [58:40<41:03:10, 8.40s/step, loss=4.005, acc=0.352, acc_len=0.75, thru=16.1, I=14.1, T=31.6, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:54,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:28:54,531] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:00,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:29:01,350] loop.py:45 INFO TRAIN_STEP step=433 loss=3.830202 Training: 2%|▏ | 433/18018 [58:47<39:22:43, 8.06s/step, loss=3.830, acc=0.367, acc_len=0.78, thru=18.6, I=20.0, T=33.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:01,653] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:07,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:29:09,197] loop.py:45 INFO TRAIN_STEP step=434 loss=4.147666 Training: 2%|▏ | 434/18018 [58:55<39:03:31, 8.00s/step, loss=4.148, acc=0.340, acc_len=0.71, thru=15.9, I=18.2, T=29.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:09,489] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:29:16,898] loop.py:380 INFO TIMING step=435: step=3.952s data=1.761s compute=3.899s [fwd=2.754s bwd=1.133s opt=0.003s] dispatch=3.741s [2026-10-18 14:29:16,899] loop.py:45 INFO TRAIN_STEP step=435 loss=3.901872 Training: 2%|▏ | 435/18018 [59:02<38:37:30, 7.91s/step, loss=3.902, acc=0.355, acc_len=0.76, thru=14.1, I=15.6, T=32.4, wait=3.7s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:29:23,503] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=435: forward=2775.2ms backward=1142.1ms [2026-10-18 14:29:23,507] loop.py:45 INFO TRAIN_STEP step=436 loss=3.947780 Training: 2%|▏ | 436/18018 [59:09<36:43:06, 7.52s/step, loss=3.948, acc=0.344, acc_len=0.72, thru=17.3, I=20.0, T=32.3, wait=2.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:33,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:29:34,261] loop.py:45 INFO TRAIN_STEP step=437 loss=3.981785 Training: 2%|▏ | 437/18018 [59:20<41:27:19, 8.49s/step, loss=3.982, acc=0.352, acc_len=0.74, thru=14.3, I=12.6, T=27.0, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:34,295] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:40,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:29:41,182] loop.py:45 INFO TRAIN_STEP step=438 loss=4.092972 Training: 2%|▏ | 438/18018 [59:27<39:09:29, 8.02s/step, loss=4.093, acc=0.345, acc_len=0.72, thru=18.3, I=17.9, T=31.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:41,686] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:49,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:29:50,934] loop.py:45 INFO TRAIN_STEP step=439 loss=3.852530 Training: 2%|▏ | 439/18018 [59:37<41:41:54, 8.54s/step, loss=3.853, acc=0.366, acc_len=0.78, thru=13.8, I=15.4, T=26.5, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:50,989] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:55,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:29:56,984] loop.py:380 INFO TIMING step=440: step=3.653s data=1.580s compute=3.623s [fwd=2.628s bwd=0.980s opt=0.003s] dispatch=2.386s [2026-10-18 14:29:56,984] loop.py:45 INFO TRAIN_STEP step=440 loss=3.854809 Training: 2%|▏ | 440/18018 [59:43<38:02:39, 7.79s/step, loss=3.855, acc=0.365, acc_len=0.78, thru=17.5, I=21.6, T=35.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:29:57,325] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:04,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:30:05,134] loop.py:45 INFO TRAIN_STEP step=441 loss=3.824360 Training: 2%|▏ | 441/18018 [59:51<38:34:15, 7.90s/step, loss=3.824, acc=0.368, acc_len=0.79, thru=14.4, I=16.5, T=31.3, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:30:05,130] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=440: forward=2994.5ms backward=1049.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:05,170] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:11,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:30:12,287] loop.py:45 INFO TRAIN_STEP step=442 loss=4.063685 Training: 2%|▏ | 442/18018 [59:58<37:28:16, 7.68s/step, loss=4.064, acc=0.342, acc_len=0.72, thru=18.2, I=19.9, T=33.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:12,650] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:30:20,346] loop.py:45 INFO TRAIN_STEP step=443 loss=3.845421 Training: 2%|▏ | 443/18018 [1:00:06<38:02:04, 7.79s/step, loss=3.845, acc=0.356, acc_len=0.76, thru=11.7, I=15.5, T=30.1, wait=3.8s, pool=20, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 14:30:20,643] eagle_store.py:517 WARNING batch_get_buffer missing keys (k14187-802e3177_hs, k14187-802e3177_ids, k14187-802e3177_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:32,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 14:30:20,864] eagle_store.py:517 WARNING batch_get_buffer missing keys (k14184-9c4f8528_hs, k14184-9c4f8528_ids, k14184-9c4f8528_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-18 14:30:32,726] loop.py:45 INFO TRAIN_STEP step=444 loss=4.056068 Training: 2%|▏ | 444/18018 [1:00:18<44:45:02, 9.17s/step, loss=4.056, acc=0.352, acc_len=0.74, thru=14.6, I=15.6, T=25.5, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:33,093] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:41,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:30:41,665] loop.py:380 INFO TIMING step=445: step=4.532s data=2.314s compute=4.505s [fwd=3.422s bwd=1.070s opt=0.003s] dispatch=4.398s [2026-10-18 14:30:41,665] loop.py:45 INFO TRAIN_STEP step=445 loss=4.062200 Training: 2%|▏ | 445/18018 [1:00:27<44:24:53, 9.10s/step, loss=4.062, acc=0.341, acc_len=0.71, thru=14.2, I=14.4, T=28.2, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:41,698] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:46,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:30:47,355] loop.py:45 INFO TRAIN_STEP step=446 loss=3.816985 Training: 2%|▏ | 446/18018 [1:00:33<39:25:13, 8.08s/step, loss=3.817, acc=0.377, acc_len=0.81, thru=20.0, I=19.8, T=35.5, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:30:47,350] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=445: forward=2550.3ms backward=1014.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:47,516] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:53,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:30:54,286] loop.py:45 INFO TRAIN_STEP step=447 loss=3.882884 Training: 2%|▏ | 447/18018 [1:00:40<37:44:39, 7.73s/step, loss=3.883, acc=0.364, acc_len=0.77, thru=17.3, I=20.6, T=35.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:30:54,418] inference_manager.py:408 INFO Pool capacity available after 1.1s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:00,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:31:01,195] loop.py:45 INFO TRAIN_STEP step=448 loss=3.972039 Training: 2%|▏ | 448/18018 [1:00:47<36:31:53, 7.49s/step, loss=3.972, acc=0.357, acc_len=0.75, thru=17.7, I=20.8, T=29.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:01,549] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:06,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:31:07,839] loop.py:45 INFO TRAIN_STEP step=449 loss=4.014851 Training: 2%|▏ | 449/18018 [1:00:53<35:18:05, 7.23s/step, loss=4.015, acc=0.347, acc_len=0.73, thru=18.2, I=21.8, T=33.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:08,297] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:31:14,639] loop.py:380 INFO TIMING step=450: step=3.742s data=1.578s compute=3.712s [fwd=2.594s bwd=1.104s opt=0.003s] dispatch=3.049s [2026-10-18 14:31:14,639] loop.py:45 INFO TRAIN_STEP step=450 loss=3.701328 Training: 2%|▏ | 450/18018 [1:01:00<34:39:48, 7.10s/step, loss=3.701, acc=0.379, acc_len=0.81, thru=15.6, I=24.2, T=34.2, wait=3.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:23,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:31:24,672] loop.py:45 INFO TRAIN_STEP step=451 loss=3.853390 Training: 3%|▎ | 451/18018 [1:01:10<38:57:07, 7.98s/step, loss=3.853, acc=0.355, acc_len=0.76, thru=15.1, I=13.3, T=31.9, wait=6.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:31:24,668] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=450: forward=2816.2ms backward=1146.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:24,968] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:29,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:31:30,872] loop.py:45 INFO TRAIN_STEP step=452 loss=3.930214 Training: 3%|▎ | 452/18018 [1:01:16<36:20:26, 7.45s/step, loss=3.930, acc=0.366, acc_len=0.78, thru=19.0, I=22.9, T=30.6, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:31,372] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:37,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:31:37,547] loop.py:45 INFO TRAIN_STEP step=453 loss=3.961024 Training: 3%|▎ | 453/18018 [1:01:23<35:12:17, 7.22s/step, loss=3.961, acc=0.362, acc_len=0.76, thru=17.7, I=22.3, T=31.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:37,762] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:43,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:31:44,488] loop.py:45 INFO TRAIN_STEP step=454 loss=3.782332 Training: 3%|▎ | 454/18018 [1:01:30<34:48:06, 7.13s/step, loss=3.782, acc=0.366, acc_len=0.79, thru=18.0, I=19.1, T=33.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:44,997] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:49,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:31:50,632] loop.py:380 INFO TIMING step=455: step=3.628s data=1.513s compute=3.597s [fwd=2.566s bwd=1.017s opt=0.003s] dispatch=2.507s [2026-10-18 14:31:50,632] loop.py:45 INFO TRAIN_STEP step=455 loss=3.687210 Training: 3%|▎ | 455/18018 [1:01:36<33:21:01, 6.84s/step, loss=3.687, acc=0.381, acc_len=0.82, thru=19.5, I=24.4, T=35.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:50,772] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:55,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:31:57,180] loop.py:45 INFO TRAIN_STEP step=456 loss=3.705014 Training: 3%|▎ | 456/18018 [1:01:43<32:55:45, 6.75s/step, loss=3.705, acc=0.381, acc_len=0.82, thru=19.3, I=26.5, T=34.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:31:57,210] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 14:31:57,175] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=455: forward=2631.1ms backward=1087.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:03,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:32:04,195] loop.py:45 INFO TRAIN_STEP step=457 loss=4.024003 Training: 3%|▎ | 457/18018 [1:01:50<33:18:52, 6.83s/step, loss=4.024, acc=0.346, acc_len=0.73, thru=18.1, I=21.6, T=34.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:04,453] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:12,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:32:13,966] loop.py:45 INFO TRAIN_STEP step=458 loss=3.644215 Training: 3%|▎ | 458/18018 [1:02:00<37:37:02, 7.71s/step, loss=3.644, acc=0.391, acc_len=0.85, thru=13.5, I=15.0, T=26.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:14,191] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:32:22,852] loop.py:45 INFO TRAIN_STEP step=459 loss=3.873238 Training: 3%|▎ | 459/18018 [1:02:08<39:19:52, 8.06s/step, loss=3.873, acc=0.359, acc_len=0.77, thru=12.9, I=12.9, T=28.8, wait=4.4s, pool=48, epoch=1/2][2026-10-18 14:32:31,106] loop.py:380 INFO TIMING step=460: step=3.941s data=1.703s compute=3.896s [fwd=2.693s bwd=1.188s opt=0.003s] dispatch=4.305s [2026-10-18 14:32:31,106] loop.py:45 INFO TRAIN_STEP step=460 loss=3.748538 Training: 3%|▎ | 460/18018 [1:02:17<39:36:23, 8.12s/step, loss=3.749, acc=0.377, acc_len=0.81, thru=16.8, I=15.9, T=32.5, wait=4.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:41,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:32:42,903] loop.py:45 INFO TRAIN_STEP step=461 loss=3.761502 Training: 3%|▎ | 461/18018 [1:02:28<44:59:18, 9.22s/step, loss=3.762, acc=0.377, acc_len=0.81, thru=12.4, I=13.8, T=24.1, wait=6.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:32:42,899] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=460: forward=3827.5ms backward=1417.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:43,122] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:48,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:32:49,986] loop.py:45 INFO TRAIN_STEP step=462 loss=3.891084 Training: 3%|▎ | 462/18018 [1:02:36<41:50:53, 8.58s/step, loss=3.891, acc=0.360, acc_len=0.76, thru=16.3, I=19.4, T=32.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:50,384] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:56,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:32:57,066] loop.py:45 INFO TRAIN_STEP step=463 loss=3.736763 Training: 3%|▎ | 463/18018 [1:02:43<39:39:04, 8.13s/step, loss=3.737, acc=0.374, acc_len=0.80, thru=17.3, I=21.9, T=29.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:32:57,236] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:02,032] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:33:03,344] loop.py:45 INFO TRAIN_STEP step=464 loss=4.084898 Training: 3%|▎ | 464/18018 [1:02:49<36:56:15, 7.58s/step, loss=4.085, acc=0.339, acc_len=0.71, thru=18.4, I=22.7, T=33.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:03,540] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:09,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:33:10,723] loop.py:380 INFO TIMING step=465: step=4.346s data=2.075s compute=4.314s [fwd=2.992s bwd=1.308s opt=0.003s] dispatch=3.023s [2026-10-18 14:33:10,723] loop.py:45 INFO TRAIN_STEP step=465 loss=3.666373 Training: 3%|▎ | 465/18018 [1:02:56<36:38:51, 7.52s/step, loss=3.666, acc=0.381, acc_len=0.81, thru=16.7, I=20.6, T=29.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:11,163] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:33:18,324] loop.py:45 INFO TRAIN_STEP step=466 loss=3.876153 Training: 3%|▎ | 466/18018 [1:03:04<36:46:15, 7.54s/step, loss=3.876, acc=0.366, acc_len=0.78, thru=13.7, I=17.2, T=28.7, wait=3.1s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:33:18,320] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=465: forward=3074.6ms backward=1342.2ms [2026-10-18 14:33:27,538] loop.py:45 INFO TRAIN_STEP step=467 loss=4.090560 Training: 3%|▎ | 467/18018 [1:03:13<39:13:05, 8.04s/step, loss=4.091, acc=0.338, acc_len=0.71, thru=15.5, I=14.9, T=29.0, wait=4.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:37,669] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:33:39,782] loop.py:45 INFO TRAIN_STEP step=468 loss=4.037890 Training: 3%|▎ | 468/18018 [1:03:25<45:21:05, 9.30s/step, loss=4.038, acc=0.352, acc_len=0.74, thru=11.6, I=11.6, T=25.3, wait=7.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:40,183] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:46,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:33:47,124] loop.py:45 INFO TRAIN_STEP step=469 loss=3.691828 Training: 3%|▎ | 469/18018 [1:03:33<42:29:07, 8.72s/step, loss=3.692, acc=0.385, acc_len=0.83, thru=14.5, I=18.3, T=28.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:47,610] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:52,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:33:53,973] loop.py:380 INFO TIMING step=470: step=3.963s data=1.657s compute=3.919s [fwd=2.721s bwd=1.184s opt=0.003s] dispatch=2.876s [2026-10-18 14:33:53,973] loop.py:45 INFO TRAIN_STEP step=470 loss=3.497411 Training: 3%|▎ | 470/18018 [1:03:40<39:45:10, 8.16s/step, loss=3.497, acc=0.402, acc_len=0.88, thru=17.6, I=19.4, T=32.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:33:54,455] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:02,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:34:03,844] loop.py:45 INFO TRAIN_STEP step=471 loss=3.740678 Training: 3%|▎ | 471/18018 [1:03:49<42:15:39, 8.67s/step, loss=3.741, acc=0.377, acc_len=0.81, thru=13.6, I=15.3, T=25.9, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:34:03,839] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=470: forward=3537.2ms backward=1355.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:04,116] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:11,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:34:12,359] loop.py:45 INFO TRAIN_STEP step=472 loss=3.942078 Training: 3%|▎ | 472/18018 [1:03:58<42:01:40, 8.62s/step, loss=3.942, acc=0.358, acc_len=0.76, thru=15.5, I=16.3, T=30.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:12,654] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:18,669] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:34:18,784] loop.py:45 INFO TRAIN_STEP step=473 loss=3.699463 Training: 3%|▎ | 473/18018 [1:04:04<38:48:48, 7.96s/step, loss=3.699, acc=0.379, acc_len=0.82, thru=18.8, I=20.3, T=33.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:19,171] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:24,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:34:24,881] loop.py:45 INFO TRAIN_STEP step=474 loss=3.814408 Training: 3%|▎ | 474/18018 [1:04:10<36:04:46, 7.40s/step, loss=3.814, acc=0.367, acc_len=0.78, thru=19.7, I=22.2, T=35.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:25,053] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:31,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:34:31,259] loop.py:380 INFO TIMING step=475: step=4.152s data=1.584s compute=4.123s [fwd=2.711s bwd=1.397s opt=0.003s] dispatch=2.217s [2026-10-18 14:34:31,259] loop.py:45 INFO TRAIN_STEP step=475 loss=3.812929 Training: 3%|▎ | 475/18018 [1:04:17<34:34:42, 7.10s/step, loss=3.813, acc=0.374, acc_len=0.79, thru=19.2, I=21.1, T=30.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:31,659] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:36,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:34:38,506] loop.py:45 INFO TRAIN_STEP step=476 loss=3.863614 Training: 3%|▎ | 476/18018 [1:04:24<34:47:59, 7.14s/step, loss=3.864, acc=0.362, acc_len=0.77, thru=16.8, I=21.0, T=29.7, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:34:38,501] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=475: forward=3018.4ms backward=1247.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:38,844] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:47,621] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:34:47,668] loop.py:45 INFO TRAIN_STEP step=477 loss=3.518122 Training: 3%|▎ | 477/18018 [1:04:33<37:44:57, 7.75s/step, loss=3.518, acc=0.406, acc_len=0.88, thru=14.5, I=13.4, T=31.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:48,125] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:53,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:34:56,286] loop.py:45 INFO TRAIN_STEP step=478 loss=3.642258 Training: 3%|▎ | 478/18018 [1:04:42<39:01:09, 8.01s/step, loss=3.642, acc=0.385, acc_len=0.83, thru=16.1, I=20.3, T=28.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:34:56,361] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:02,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:35:02,591] loop.py:45 INFO TRAIN_STEP step=479 loss=3.603185 Training: 3%|▎ | 479/18018 [1:04:48<36:31:38, 7.50s/step, loss=3.603, acc=0.385, acc_len=0.83, thru=16.7, I=25.3, T=37.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:03,077] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:08,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:35:10,163] loop.py:380 INFO TIMING step=480: step=4.030s data=1.824s compute=3.995s [fwd=2.837s bwd=1.144s opt=0.003s] dispatch=3.534s [2026-10-18 14:35:10,163] loop.py:45 INFO TRAIN_STEP step=480 loss=3.726728 Training: 3%|▎ | 480/18018 [1:04:56<36:37:59, 7.52s/step, loss=3.727, acc=0.375, acc_len=0.80, thru=16.8, I=19.8, T=31.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:10,348] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:35:17,095] loop.py:45 INFO TRAIN_STEP step=481 loss=3.819301 Training: 3%|▎ | 481/18018 [1:05:03<35:46:21, 7.34s/step, loss=3.819, acc=0.369, acc_len=0.79, thru=16.6, I=22.2, T=30.8, wait=2.8s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:35:17,090] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=480: forward=2872.0ms backward=1242.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:24,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:35:26,044] loop.py:45 INFO TRAIN_STEP step=482 loss=3.831071 Training: 3%|▎ | 482/18018 [1:05:12<38:06:59, 7.83s/step, loss=3.831, acc=0.371, acc_len=0.78, thru=15.6, I=16.0, T=26.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:26,271] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:31,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:35:32,607] loop.py:45 INFO TRAIN_STEP step=483 loss=3.759370 Training: 3%|▎ | 483/18018 [1:05:18<36:16:24, 7.45s/step, loss=3.759, acc=0.375, acc_len=0.80, thru=15.3, I=20.8, T=31.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:32,955] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:37,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:35:39,040] loop.py:45 INFO TRAIN_STEP step=484 loss=3.557034 Training: 3%|▎ | 484/18018 [1:05:25<34:47:09, 7.14s/step, loss=3.557, acc=0.390, acc_len=0.84, thru=18.1, I=25.8, T=34.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:39,373] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:44,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:35:46,272] loop.py:380 INFO TIMING step=485: step=3.711s data=1.667s compute=3.678s [fwd=2.575s bwd=1.088s opt=0.003s] dispatch=3.514s [2026-10-18 14:35:46,272] loop.py:45 INFO TRAIN_STEP step=485 loss=3.818279 Training: 3%|▎ | 485/18018 [1:05:32<34:54:50, 7.17s/step, loss=3.818, acc=0.374, acc_len=0.80, thru=16.6, I=24.1, T=34.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:46,724] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:51,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:35:53,119] loop.py:45 INFO TRAIN_STEP step=486 loss=3.526413 Training: 3%|▎ | 486/18018 [1:05:39<34:26:34, 7.07s/step, loss=3.526, acc=0.397, acc_len=0.86, thru=18.5, I=24.1, T=30.4, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:35:53,113] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=485: forward=2924.6ms backward=1236.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:53,399] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:58,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:35:59,644] loop.py:45 INFO TRAIN_STEP step=487 loss=3.757207 Training: 3%|▎ | 487/18018 [1:05:45<33:38:31, 6.91s/step, loss=3.757, acc=0.371, acc_len=0.79, thru=17.8, I=21.6, T=32.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:35:59,939] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:05,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:36:06,410] loop.py:45 INFO TRAIN_STEP step=488 loss=3.763866 Training: 3%|▎ | 488/18018 [1:05:52<33:26:07, 6.87s/step, loss=3.764, acc=0.377, acc_len=0.80, thru=16.6, I=20.0, T=28.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:06,796] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:36:12,636] loop.py:45 INFO TRAIN_STEP step=489 loss=3.912240 Training: 3%|▎ | 489/18018 [1:05:58<32:29:41, 6.67s/step, loss=3.912, acc=0.356, acc_len=0.76, thru=17.2, I=21.4, T=32.2, wait=2.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:18,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:36:19,669] loop.py:380 INFO TIMING step=490: step=4.001s data=1.786s compute=3.969s [fwd=2.840s bwd=1.115s opt=0.003s] dispatch=3.024s [2026-10-18 14:36:19,670] loop.py:45 INFO TRAIN_STEP step=490 loss=3.698991 Training: 3%|▎ | 490/18018 [1:06:05<33:01:07, 6.78s/step, loss=3.699, acc=0.374, acc_len=0.79, thru=17.2, I=20.0, T=32.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:20,107] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:25,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:36:27,193] loop.py:45 INFO TRAIN_STEP step=491 loss=3.598471 Training: 3%|▎ | 491/18018 [1:06:13<34:06:00, 7.00s/step, loss=3.598, acc=0.392, acc_len=0.84, thru=15.4, I=20.9, T=31.0, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:36:27,189] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=490: forward=2906.0ms backward=1176.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:27,392] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:33,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:36:35,067] loop.py:45 INFO TRAIN_STEP step=492 loss=3.681917 Training: 3%|▎ | 492/18018 [1:06:21<35:22:05, 7.26s/step, loss=3.682, acc=0.389, acc_len=0.83, thru=14.4, I=21.9, T=29.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:35,251] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:41,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:36:43,347] loop.py:45 INFO TRAIN_STEP step=493 loss=3.827301 Training: 3%|▎ | 493/18018 [1:06:29<36:50:57, 7.57s/step, loss=3.827, acc=0.365, acc_len=0.77, thru=15.8, I=19.3, T=30.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:36:43,505] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:36:50,636] loop.py:45 INFO TRAIN_STEP step=494 loss=3.261086 Training: 3%|▎ | 494/18018 [1:06:36<36:26:19, 7.49s/step, loss=3.261, acc=0.425, acc_len=0.94, thru=15.3, I=19.5, T=27.7, wait=2.7s, pool=52, epoch=1/2][2026-10-18 14:36:59,664] loop.py:380 INFO TIMING step=495: step=4.048s data=1.798s compute=3.986s [fwd=2.790s bwd=1.180s opt=0.003s] dispatch=4.971s [2026-10-18 14:36:59,664] loop.py:45 INFO TRAIN_STEP step=495 loss=3.662597 Training: 3%|▎ | 495/18018 [1:06:45<38:41:18, 7.95s/step, loss=3.663, acc=0.386, acc_len=0.83, thru=13.6, I=12.2, T=31.6, wait=5.0s, pool=48, epoch=1/2][2026-10-18 14:37:10,929] loop.py:45 INFO TRAIN_STEP step=496 loss=4.116052 Training: 3%|▎ | 496/18018 [1:06:57<43:31:42, 8.94s/step, loss=4.116, acc=0.347, acc_len=0.73, thru=10.5, I=11.6, T=29.8, wait=7.0s, pool=44, epoch=1/2](TrainerActor pid=1752, ip=10.0.3.62) [2026-10-18 14:37:10,993] eagle_store.py:517 WARNING batch_get_buffer missing keys (k15873-86c7e4de_hs, k15873-86c7e4de_ids, k15873-86c7e4de_lhs); sleeping 0.50s. (TrainerActor pid=9310) [2026-10-18 14:37:10,923] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=495: forward=2946.6ms backward=1272.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:37:18,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 14:37:11,043] eagle_store.py:517 WARNING batch_get_buffer missing keys (k15872-949245dd_hs, k15872-949245dd_ids, k15872-949245dd_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-18 14:37:19,602] loop.py:45 INFO TRAIN_STEP step=497 loss=3.697456 Training: 3%|▎ | 497/18018 [1:07:05<43:07:48, 8.86s/step, loss=3.697, acc=0.381, acc_len=0.81, thru=15.6, I=17.2, T=30.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:37:19,631] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:37:25,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:37:26,085] loop.py:45 INFO TRAIN_STEP step=498 loss=3.695068 Training: 3%|▎ | 498/18018 [1:07:12<39:39:23, 8.15s/step, loss=3.695, acc=0.381, acc_len=0.81, thru=17.6, I=21.6, T=30.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:37:26,401] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:37:35,200] loop.py:45 INFO TRAIN_STEP step=499 loss=3.586052 Training: 3%|▎ | 499/18018 [1:07:21<41:03:50, 8.44s/step, loss=3.586, acc=0.395, acc_len=0.85, thru=12.1, I=12.1, T=29.5, wait=4.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:37:42,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:37:44,158] loop.py:380 INFO TIMING step=500: step=5.312s data=2.317s compute=5.275s [fwd=3.848s bwd=1.412s opt=0.003s] dispatch=3.638s [2026-10-18 14:37:44,159] loop.py:45 INFO TRAIN_STEP step=500 loss=3.471007 Training: 3%|▎ | 500/18018 [1:07:30<41:49:11, 8.59s/step, loss=3.471, acc=0.400, acc_len=0.87, thru=17.0, I=17.3, T=24.1, wait=3.6s, pool=64, epoch=1/2][2026-10-18 14:37:44,160] loop.py:419 INFO Saving checkpoint at step 500... (TrainerActor pid=9310) /usr/local/lib/python3.12/dist-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning. (TrainerActor pid=9310) return func(*args, **kwargs) (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:37:47,069] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 14:37:47,539] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0000501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:37:47,586] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:37:53,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 14:37:47,540] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0000501 [2026-10-18 14:37:54,241] loop.py:45 INFO TRAIN_STEP step=501 loss=3.618361 Training: 3%|▎ | 501/18018 [1:07:40<43:59:31, 9.04s/step, loss=3.618, acc=0.387, acc_len=0.83, thru=19.3, I=20.0, T=27.9, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:37:54,236] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=500: forward=3226.0ms backward=1311.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:37:54,721] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:00,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:38:00,964] loop.py:45 INFO TRAIN_STEP step=502 loss=3.541657 Training: 3%|▎ | 502/18018 [1:07:47<40:36:20, 8.35s/step, loss=3.542, acc=0.393, acc_len=0.84, thru=17.1, I=19.5, T=31.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:01,316] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:06,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:38:07,139] loop.py:45 INFO TRAIN_STEP step=503 loss=3.464539 Training: 3%|▎ | 503/18018 [1:07:53<37:26:08, 7.69s/step, loss=3.465, acc=0.394, acc_len=0.85, thru=18.3, I=22.4, T=32.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:07,576] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:14,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:38:15,589] loop.py:45 INFO TRAIN_STEP step=504 loss=3.412164 Training: 3%|▎ | 504/18018 [1:08:01<38:32:07, 7.92s/step, loss=3.412, acc=0.407, acc_len=0.90, thru=14.2, I=16.7, T=29.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:15,770] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:21,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:38:22,425] loop.py:380 INFO TIMING step=505: step=4.301s data=1.924s compute=4.246s [fwd=2.899s bwd=1.332s opt=0.003s] dispatch=2.527s [2026-10-18 14:38:22,426] loop.py:45 INFO TRAIN_STEP step=505 loss=3.933409 Training: 3%|▎ | 505/18018 [1:08:08<36:57:06, 7.60s/step, loss=3.933, acc=0.358, acc_len=0.75, thru=16.5, I=20.7, T=29.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:22,640] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:29,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:38:31,113] loop.py:45 INFO TRAIN_STEP step=506 loss=3.656327 Training: 3%|▎ | 506/18018 [1:08:17<38:32:44, 7.92s/step, loss=3.656, acc=0.381, acc_len=0.81, thru=13.9, I=17.0, T=27.8, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:38:31,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=505: forward=3229.1ms backward=1336.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:31,620] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:37,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:38:38,512] loop.py:45 INFO TRAIN_STEP step=507 loss=3.604270 Training: 3%|▎ | 507/18018 [1:08:24<37:46:22, 7.77s/step, loss=3.604, acc=0.387, acc_len=0.83, thru=15.0, I=21.7, T=32.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:38,740] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:46,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:38:47,065] loop.py:45 INFO TRAIN_STEP step=508 loss=3.754249 Training: 3%|▎ | 508/18018 [1:08:33<38:55:25, 8.00s/step, loss=3.754, acc=0.376, acc_len=0.81, thru=13.8, I=17.2, T=29.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:47,102] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:53,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:38:54,382] loop.py:45 INFO TRAIN_STEP step=509 loss=3.600270 Training: 3%|▎ | 509/18018 [1:08:40<37:55:17, 7.80s/step, loss=3.600, acc=0.398, acc_len=0.85, thru=16.5, I=19.8, T=31.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:38:54,552] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 14:39:01,204] loop.py:380 INFO TIMING step=510: step=4.139s data=1.688s compute=4.095s [fwd=2.836s bwd=1.244s opt=0.003s] dispatch=2.673s [2026-10-18 14:39:01,205] loop.py:45 INFO TRAIN_STEP step=510 loss=3.661232 Training: 3%|▎ | 510/18018 [1:08:47<36:29:40, 7.50s/step, loss=3.661, acc=0.385, acc_len=0.82, thru=14.2, I=17.5, T=30.9, wait=2.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:10,307] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:39:11,954] loop.py:45 INFO TRAIN_STEP step=511 loss=3.831402 Training: 3%|▎ | 511/18018 [1:08:58<41:13:39, 8.48s/step, loss=3.831, acc=0.364, acc_len=0.76, thru=13.9, I=15.6, T=27.8, wait=6.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:39:11,950] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=510: forward=3348.1ms backward=1208.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:12,322] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:19,387] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:39:20,033] loop.py:45 INFO TRAIN_STEP step=512 loss=3.548111 Training: 3%|▎ | 512/18018 [1:09:06<40:38:47, 8.36s/step, loss=3.548, acc=0.401, acc_len=0.87, thru=14.0, I=17.5, T=28.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:20,391] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:27,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:39:27,681] loop.py:45 INFO TRAIN_STEP step=513 loss=3.579031 Training: 3%|▎ | 513/18018 [1:09:13<39:36:22, 8.15s/step, loss=3.579, acc=0.397, acc_len=0.86, thru=16.6, I=16.6, T=35.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:28,046] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:34,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:39:35,549] loop.py:45 INFO TRAIN_STEP step=514 loss=3.553911 Training: 3%|▎ | 514/18018 [1:09:21<39:11:51, 8.06s/step, loss=3.554, acc=0.390, acc_len=0.84, thru=18.0, I=18.8, T=29.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:35,788] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:43,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:39:44,153] loop.py:380 INFO TIMING step=515: step=4.086s data=1.660s compute=4.055s [fwd=2.796s bwd=1.245s opt=0.003s] dispatch=4.511s [2026-10-18 14:39:44,153] loop.py:45 INFO TRAIN_STEP step=515 loss=3.684180 Training: 3%|▎ | 515/18018 [1:09:30<39:59:18, 8.22s/step, loss=3.684, acc=0.388, acc_len=0.83, thru=14.4, I=16.2, T=31.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:44,599] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:39:51,531] loop.py:45 INFO TRAIN_STEP step=516 loss=3.597372 Training: 3%|▎ | 516/18018 [1:09:37<38:44:56, 7.97s/step, loss=3.597, acc=0.396, acc_len=0.85, thru=17.2, I=18.1, T=31.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:51,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 14:39:51,526] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=515: forward=2863.9ms backward=1172.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:39:52,072] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:00,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:40:02,555] loop.py:45 INFO TRAIN_STEP step=517 loss=3.607964 Training: 3%|▎ | 517/18018 [1:09:48<43:12:02, 8.89s/step, loss=3.608, acc=0.392, acc_len=0.84, thru=11.6, I=14.1, T=29.2, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:03,005] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:40:11,178] loop.py:45 INFO TRAIN_STEP step=518 loss=3.725490 Training: 3%|▎ | 518/18018 [1:09:57<42:48:53, 8.81s/step, loss=3.725, acc=0.376, acc_len=0.80, thru=13.7, I=13.4, T=31.0, wait=4.5s, pool=48, epoch=1/2][2026-10-18 14:40:20,483] loop.py:45 INFO TRAIN_STEP step=519 loss=3.801544 Training: 3%|▎ | 519/18018 [1:10:06<43:32:22, 8.96s/step, loss=3.802, acc=0.371, acc_len=0.79, thru=10.6, I=11.2, T=31.4, wait=5.2s, pool=16, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 14:40:20,608] eagle_store.py:517 WARNING batch_get_buffer missing keys (k16615-b59897a3_hs, k16615-b59897a3_ids, k16615-b59897a3_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:28,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 14:40:20,902] eagle_store.py:517 WARNING batch_get_buffer missing keys (k16608-9386fe70_hs, k16608-9386fe70_ids, k16608-9386fe70_lhs); sleeping 0.50s. [repeated 5x across cluster] [2026-10-18 14:40:29,787] loop.py:380 INFO TIMING step=520: step=3.899s data=1.801s compute=3.850s [fwd=2.733s bwd=1.102s opt=0.003s] dispatch=5.395s [2026-10-18 14:40:29,787] loop.py:45 INFO TRAIN_STEP step=520 loss=3.676970 Training: 3%|▎ | 520/18018 [1:10:15<44:02:22, 9.06s/step, loss=3.677, acc=0.387, acc_len=0.83, thru=19.0, I=18.2, T=32.8, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:30,277] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:40:37,507] loop.py:45 INFO TRAIN_STEP step=521 loss=3.323095 Training: 3%|▎ | 521/18018 [1:10:23<42:05:04, 8.66s/step, loss=3.323, acc=0.415, acc_len=0.91, thru=14.9, I=16.9, T=33.2, wait=3.9s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:40:37,504] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=520: forward=2648.9ms backward=1156.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:44,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:40:44,917] loop.py:45 INFO TRAIN_STEP step=522 loss=3.640667 Training: 3%|▎ | 522/18018 [1:10:31<40:15:35, 8.28s/step, loss=3.641, acc=0.376, acc_len=0.81, thru=18.1, I=17.6, T=31.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:45,096] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:49,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:40:52,026] loop.py:45 INFO TRAIN_STEP step=523 loss=3.749202 Training: 3%|▎ | 523/18018 [1:10:38<38:32:35, 7.93s/step, loss=3.749, acc=0.379, acc_len=0.81, thru=16.8, I=21.1, T=29.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:52,390] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:58,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:40:59,197] loop.py:45 INFO TRAIN_STEP step=524 loss=3.573919 Training: 3%|▎ | 524/18018 [1:10:45<37:25:59, 7.70s/step, loss=3.574, acc=0.392, acc_len=0.84, thru=15.2, I=21.1, T=32.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:40:59,278] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 14:41:05,831] loop.py:380 INFO TIMING step=525: step=3.734s data=1.580s compute=3.704s [fwd=2.577s bwd=1.112s opt=0.003s] dispatch=2.893s [2026-10-18 14:41:05,831] loop.py:45 INFO TRAIN_STEP step=525 loss=3.550710 Training: 3%|▎ | 525/18018 [1:10:51<35:52:22, 7.38s/step, loss=3.551, acc=0.402, acc_len=0.87, thru=16.9, I=19.0, T=34.3, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:12,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:41:13,597] loop.py:45 INFO TRAIN_STEP step=526 loss=3.438935 Training: 3%|▎ | 526/18018 [1:10:59<36:25:57, 7.50s/step, loss=3.439, acc=0.404, acc_len=0.88, thru=17.2, I=18.1, T=26.3, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:41:13,592] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=525: forward=3342.1ms backward=1466.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:13,660] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:20,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:41:20,804] loop.py:45 INFO TRAIN_STEP step=527 loss=3.469792 Training: 3%|▎ | 527/18018 [1:11:06<36:00:21, 7.41s/step, loss=3.470, acc=0.409, acc_len=0.88, thru=16.6, I=17.2, T=32.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:20,947] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:26,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:41:27,491] loop.py:45 INFO TRAIN_STEP step=528 loss=3.358276 Training: 3%|▎ | 528/18018 [1:11:13<34:56:49, 7.19s/step, loss=3.358, acc=0.416, acc_len=0.92, thru=18.9, I=20.5, T=30.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:27,735] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:34,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:41:35,780] loop.py:45 INFO TRAIN_STEP step=529 loss=3.622630 Training: 3%|▎ | 529/18018 [1:11:21<36:32:40, 7.52s/step, loss=3.623, acc=0.392, acc_len=0.84, thru=14.5, I=16.3, T=30.2, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:36,049] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:41:42,916] loop.py:380 INFO TIMING step=530: step=4.053s data=1.885s compute=4.017s [fwd=2.868s bwd=1.134s opt=0.003s] dispatch=3.072s [2026-10-18 14:41:42,916] loop.py:45 INFO TRAIN_STEP step=530 loss=3.208705 Training: 3%|▎ | 530/18018 [1:11:29<35:58:32, 7.41s/step, loss=3.209, acc=0.431, acc_len=0.95, thru=13.8, I=18.2, T=31.6, wait=3.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:50,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:41:52,720] loop.py:45 INFO TRAIN_STEP step=531 loss=3.697183 (TrainerActor pid=9310) [2026-10-18 14:41:52,715] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=530: forward=3238.5ms backward=1264.1ms Training: 3%|▎ | 531/18018 [1:11:38<39:28:16, 8.13s/step, loss=3.697, acc=0.386, acc_len=0.83, thru=15.1, I=16.1, T=28.1, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:52,755] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:58,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:41:59,819] loop.py:45 INFO TRAIN_STEP step=532 loss=3.518087 Training: 3%|▎ | 532/18018 [1:11:45<37:58:11, 7.82s/step, loss=3.518, acc=0.392, acc_len=0.84, thru=15.4, I=22.0, T=31.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:41:59,885] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 14:42:07,719] loop.py:45 INFO TRAIN_STEP step=533 loss=3.675662 Training: 3%|▎ | 533/18018 [1:11:53<38:05:31, 7.84s/step, loss=3.676, acc=0.381, acc_len=0.81, thru=12.4, I=14.2, T=28.8, wait=3.4s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:15,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:42:18,519] loop.py:45 INFO TRAIN_STEP step=534 loss=3.598366 Training: 3%|▎ | 534/18018 [1:12:04<42:23:42, 8.73s/step, loss=3.598, acc=0.402, acc_len=0.86, thru=15.0, I=18.7, T=25.1, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:18,742] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:22,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:42:24,292] loop.py:380 INFO TIMING step=535: step=3.594s data=1.571s compute=3.563s [fwd=2.543s bwd=1.005s opt=0.003s] dispatch=2.170s [2026-10-18 14:42:24,293] loop.py:45 INFO TRAIN_STEP step=535 loss=3.192180 Training: 3%|▎ | 535/18018 [1:12:10<38:05:23, 7.84s/step, loss=3.192, acc=0.433, acc_len=0.96, thru=17.6, I=31.2, T=35.6, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:24,560] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:29,839] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:42:31,158] loop.py:45 INFO TRAIN_STEP step=536 loss=3.195667 Training: 3%|▎ | 536/18018 [1:12:17<36:39:42, 7.55s/step, loss=3.196, acc=0.434, acc_len=0.96, thru=17.8, I=26.5, T=33.7, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:42:31,154] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=535: forward=2630.9ms backward=1129.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:31,351] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:37,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:42:38,969] loop.py:45 INFO TRAIN_STEP step=537 loss=3.486494 Training: 3%|▎ | 537/18018 [1:12:25<37:02:26, 7.63s/step, loss=3.486, acc=0.406, acc_len=0.87, thru=15.7, I=22.9, T=26.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:39,041] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 [2026-10-18 14:42:47,024] loop.py:45 INFO TRAIN_STEP step=538 loss=3.547679 Training: 3%|▎ | 538/18018 [1:12:33<37:40:05, 7.76s/step, loss=3.548, acc=0.395, acc_len=0.85, thru=15.6, I=17.2, T=31.2, wait=3.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:54,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:42:55,904] loop.py:45 INFO TRAIN_STEP step=539 loss=3.642838 Training: 3%|▎ | 539/18018 [1:12:41<39:17:37, 8.09s/step, loss=3.643, acc=0.383, acc_len=0.82, thru=14.8, I=16.3, T=30.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:42:56,089] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:43:05,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:43:06,453] loop.py:380 INFO TIMING step=540: step=4.241s data=1.813s compute=4.214s [fwd=2.931s bwd=1.269s opt=0.003s] dispatch=6.297s [2026-10-18 14:43:06,453] loop.py:45 INFO TRAIN_STEP step=540 loss=3.665300 Training: 3%|▎ | 540/18018 [1:12:52<42:51:56, 8.83s/step, loss=3.665, acc=0.386, acc_len=0.82, thru=12.3, I=13.4, T=30.2, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:43:06,929] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:43:13,225] loop.py:45 INFO TRAIN_STEP step=541 loss=3.595415 Training: 3%|▎ | 541/18018 [1:12:59<39:51:55, 8.21s/step, loss=3.595, acc=0.386, acc_len=0.82, thru=14.6, I=18.6, T=32.9, wait=2.9s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:43:13,220] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=540: forward=2677.7ms backward=1158.5ms [2026-10-18 14:43:21,854] loop.py:45 INFO TRAIN_STEP step=542 loss=3.308943 Training: 3%|▎ | 542/18018 [1:13:07<40:28:31, 8.34s/step, loss=3.309, acc=0.421, acc_len=0.92, thru=17.5, I=14.4, T=31.3, wait=4.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:43:28,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:43:30,243] loop.py:45 INFO TRAIN_STEP step=543 loss=3.553004 Training: 3%|▎ | 543/18018 [1:13:16<40:32:47, 8.35s/step, loss=3.553, acc=0.397, acc_len=0.86, thru=15.7, I=18.4, T=29.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:43:30,436] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:43:36,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:43:37,496] loop.py:45 INFO TRAIN_STEP step=544 loss=3.552808 Training: 3%|▎ | 544/18018 [1:13:23<38:56:33, 8.02s/step, loss=3.553, acc=0.394, acc_len=0.85, thru=16.4, I=21.4, T=26.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:43:37,875] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:43:44,499] loop.py:380 INFO TIMING step=545: step=3.925s data=1.784s compute=3.886s [fwd=2.768s bwd=1.103s opt=0.003s] dispatch=3.068s [2026-10-18 14:43:44,499] loop.py:45 INFO TRAIN_STEP step=545 loss=3.440192 Training: 3%|▎ | 545/18018 [1:13:30<37:27:09, 7.72s/step, loss=3.440, acc=0.408, acc_len=0.88, thru=14.7, I=18.5, T=32.6, wait=3.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:43:55,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:43:56,016] loop.py:45 INFO TRAIN_STEP step=546 loss=3.624937 Training: 3%|▎ | 546/18018 [1:13:42<42:59:06, 8.86s/step, loss=3.625, acc=0.392, acc_len=0.84, thru=12.5, I=12.3, T=25.0, wait=6.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:43:56,012] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=545: forward=3639.4ms backward=1441.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:43:56,456] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:01,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:44:02,389] loop.py:45 INFO TRAIN_STEP step=547 loss=3.650591 Training: 3%|▎ | 547/18018 [1:13:48<39:22:02, 8.11s/step, loss=3.651, acc=0.384, acc_len=0.82, thru=18.7, I=18.6, T=34.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:02,819] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:08,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:44:09,296] loop.py:45 INFO TRAIN_STEP step=548 loss=3.313776 Training: 3%|▎ | 548/18018 [1:13:55<37:36:34, 7.75s/step, loss=3.314, acc=0.419, acc_len=0.92, thru=18.4, I=22.4, T=30.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:09,579] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:17,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:44:17,376] loop.py:45 INFO TRAIN_STEP step=549 loss=3.744273 Training: 3%|▎ | 549/18018 [1:14:03<38:05:22, 7.85s/step, loss=3.744, acc=0.376, acc_len=0.79, thru=14.5, I=16.8, T=28.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:17,607] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:23,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:44:24,098] loop.py:380 INFO TIMING step=550: step=4.110s data=1.751s compute=4.086s [fwd=2.888s bwd=1.184s opt=0.003s] dispatch=2.602s [2026-10-18 14:44:24,098] loop.py:45 INFO TRAIN_STEP step=550 loss=3.372522 Training: 3%|▎ | 550/18018 [1:14:10<36:26:38, 7.51s/step, loss=3.373, acc=0.414, acc_len=0.90, thru=18.1, I=16.3, T=31.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:24,396] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:44:32,496] loop.py:45 INFO TRAIN_STEP step=551 loss=3.572453 Training: 3%|▎ | 551/18018 [1:14:18<37:44:01, 7.78s/step, loss=3.572, acc=0.401, acc_len=0.86, thru=12.6, I=13.2, T=32.0, wait=4.4s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:44:32,491] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=550: forward=2795.1ms backward=1151.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:39,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:44:41,195] loop.py:45 INFO TRAIN_STEP step=552 loss=3.377088 Training: 3%|▎ | 552/18018 [1:14:27<39:04:27, 8.05s/step, loss=3.377, acc=0.410, acc_len=0.89, thru=18.3, I=17.2, T=32.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:41,588] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:47,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:44:49,444] loop.py:45 INFO TRAIN_STEP step=553 loss=3.484467 Training: 3%|▎ | 553/18018 [1:14:35<39:21:29, 8.11s/step, loss=3.484, acc=0.403, acc_len=0.86, thru=14.0, I=21.5, T=27.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:44:49,555] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-18 14:44:55,209] loop.py:45 INFO TRAIN_STEP step=554 loss=3.609160 Training: 3%|▎ | 554/18018 [1:14:41<35:56:03, 7.41s/step, loss=3.609, acc=0.390, acc_len=0.82, thru=16.7, I=20.8, T=37.0, wait=2.3s, pool=44, epoch=1/2][2026-10-18 14:45:06,480] loop.py:380 INFO TIMING step=555: step=4.403s data=1.921s compute=4.354s [fwd=3.017s bwd=1.323s opt=0.003s] dispatch=6.861s [2026-10-18 14:45:06,480] loop.py:45 INFO TRAIN_STEP step=555 loss=3.660194 Training: 3%|▎ | 555/18018 [1:14:52<41:33:28, 8.57s/step, loss=3.660, acc=0.388, acc_len=0.83, thru=14.5, I=12.0, T=29.1, wait=6.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:45:11,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:45:12,937] loop.py:45 INFO TRAIN_STEP step=556 loss=3.339756 Training: 3%|▎ | 556/18018 [1:14:59<38:29:12, 7.93s/step, loss=3.340, acc=0.405, acc_len=0.88, thru=17.6, I=21.3, T=32.0, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:45:12,932] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=555: forward=2692.4ms backward=1242.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:45:13,345] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:45:18,834] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:45:19,957] loop.py:45 INFO TRAIN_STEP step=557 loss=3.333967 Training: 3%|▎ | 557/18018 [1:15:06<37:09:12, 7.66s/step, loss=3.334, acc=0.421, acc_len=0.92, thru=16.5, I=22.8, T=29.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:45:20,343] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:45:24,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:45:26,237] loop.py:45 INFO TRAIN_STEP step=558 loss=3.709263 Training: 3%|▎ | 558/18018 [1:15:12<35:08:29, 7.25s/step, loss=3.709, acc=0.384, acc_len=0.81, thru=18.2, I=24.9, T=33.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:45:26,362] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-18 14:45:32,990] loop.py:45 INFO TRAIN_STEP step=559 loss=3.522384 Training: 3%|▎ | 559/18018 [1:15:19<34:25:27, 7.10s/step, loss=3.522, acc=0.396, acc_len=0.86, thru=15.7, I=22.7, T=34.8, wait=3.1s, pool=52, epoch=1/2][2026-10-18 14:45:41,632] loop.py:380 INFO TIMING step=560: step=4.407s data=2.067s compute=4.369s [fwd=3.118s bwd=1.236s opt=0.003s] dispatch=4.226s [2026-10-18 14:45:41,632] loop.py:45 INFO TRAIN_STEP step=560 loss=3.389161 Training: 3%|▎ | 560/18018 [1:15:27<36:40:09, 7.56s/step, loss=3.389, acc=0.414, acc_len=0.89, thru=14.5, I=15.8, T=29.0, wait=4.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:45:47,307] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:45:49,327] loop.py:45 INFO TRAIN_STEP step=561 loss=3.221316 Training: 3%|▎ | 561/18018 [1:15:35<36:51:26, 7.60s/step, loss=3.221, acc=0.430, acc_len=0.94, thru=17.0, I=20.5, T=24.4, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:45:49,322] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=560: forward=3836.9ms backward=1375.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:45:49,828] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 14:45:55,334] loop.py:45 INFO TRAIN_STEP step=562 loss=3.431077 Training: 3%|▎ | 562/18018 [1:15:41<34:32:13, 7.12s/step, loss=3.431, acc=0.412, acc_len=0.89, thru=14.9, I=21.8, T=36.0, wait=2.4s, pool=56, epoch=1/2][2026-10-18 14:46:04,887] loop.py:45 INFO TRAIN_STEP step=563 loss=3.428793 Training: 3%|▎ | 563/18018 [1:15:50<38:04:13, 7.85s/step, loss=3.429, acc=0.410, acc_len=0.89, thru=11.2, I=12.2, T=33.6, wait=5.7s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:46:12,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:46:14,037] loop.py:45 INFO TRAIN_STEP step=564 loss=3.573791 Training: 3%|▎ | 564/18018 [1:16:00<39:57:33, 8.24s/step, loss=3.574, acc=0.395, acc_len=0.84, thru=16.9, I=16.9, T=26.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:46:14,214] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:46:20,793] loop.py:380 INFO TIMING step=565: step=3.923s data=1.858s compute=3.883s [fwd=2.777s bwd=1.091s opt=0.003s] dispatch=2.823s [2026-10-18 14:46:20,793] loop.py:45 INFO TRAIN_STEP step=565 loss=3.445831 Training: 3%|▎ | 565/18018 [1:16:06<37:47:31, 7.80s/step, loss=3.446, acc=0.412, acc_len=0.89, thru=15.8, I=20.3, T=32.6, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:46:29,450] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:46:30,386] loop.py:45 INFO TRAIN_STEP step=566 loss=3.474649 Training: 3%|▎ | 566/18018 [1:16:16<40:24:29, 8.34s/step, loss=3.475, acc=0.413, acc_len=0.88, thru=13.9, I=14.3, T=30.6, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:46:30,459] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (TrainerActor pid=9310) [2026-10-18 14:46:30,381] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=565: forward=2896.3ms backward=1228.7ms [2026-10-18 14:46:36,176] loop.py:45 INFO TRAIN_STEP step=567 loss=3.196303 Training: 3%|▎ | 567/18018 [1:16:22<36:42:11, 7.57s/step, loss=3.196, acc=0.434, acc_len=0.95, thru=17.5, I=19.1, T=36.8, wait=2.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:46:45,893] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:46:46,628] loop.py:45 INFO TRAIN_STEP step=568 loss=3.468349 Training: 3%|▎ | 568/18018 [1:16:32<40:53:17, 8.44s/step, loss=3.468, acc=0.402, acc_len=0.86, thru=14.5, I=15.5, T=33.6, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:46:46,897] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:46:50,930] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:46:53,020] loop.py:45 INFO TRAIN_STEP step=569 loss=3.200456 Training: 3%|▎ | 569/18018 [1:16:39<37:55:00, 7.82s/step, loss=3.200, acc=0.430, acc_len=0.94, thru=19.0, I=25.1, T=31.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:46:53,444] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:46:59,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:47:00,181] loop.py:380 INFO TIMING step=570: step=4.011s data=1.824s compute=3.973s [fwd=2.827s bwd=1.131s opt=0.003s] dispatch=3.141s [2026-10-18 14:47:00,181] loop.py:45 INFO TRAIN_STEP step=570 loss=3.126398 Training: 3%|▎ | 570/18018 [1:16:46<36:56:59, 7.62s/step, loss=3.126, acc=0.444, acc_len=0.98, thru=15.6, I=23.2, T=31.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:00,531] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:04,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:47:06,468] loop.py:45 INFO TRAIN_STEP step=571 loss=3.311497 Training: 3%|▎ | 571/18018 [1:16:52<35:00:14, 7.22s/step, loss=3.311, acc=0.421, acc_len=0.92, thru=18.7, I=24.5, T=30.9, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:47:06,463] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=570: forward=3109.3ms backward=983.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:06,967] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:13,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:47:14,833] loop.py:45 INFO TRAIN_STEP step=572 loss=3.387308 Training: 3%|▎ | 572/18018 [1:17:00<36:39:53, 7.57s/step, loss=3.387, acc=0.415, acc_len=0.89, thru=13.3, I=19.3, T=28.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:15,138] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:22,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:47:23,794] loop.py:45 INFO TRAIN_STEP step=573 loss=3.232788 Training: 3%|▎ | 573/18018 [1:17:09<38:41:29, 7.98s/step, loss=3.233, acc=0.430, acc_len=0.95, thru=14.8, I=15.7, T=29.4, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:24,108] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:29,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:47:30,478] loop.py:45 INFO TRAIN_STEP step=574 loss=3.365899 Training: 3%|▎ | 574/18018 [1:17:16<36:47:48, 7.59s/step, loss=3.366, acc=0.415, acc_len=0.90, thru=17.3, I=24.0, T=33.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:30,713] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:35,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:47:37,818] loop.py:380 INFO TIMING step=575: step=4.588s data=2.181s compute=4.557s [fwd=3.219s bwd=1.323s opt=0.003s] dispatch=2.743s [2026-10-18 14:47:37,818] loop.py:45 INFO TRAIN_STEP step=575 loss=3.398971 Training: 3%|▎ | 575/18018 [1:17:23<36:25:41, 7.52s/step, loss=3.399, acc=0.411, acc_len=0.89, thru=16.5, I=25.4, T=27.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:38,103] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:43,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:47:44,060] loop.py:45 INFO TRAIN_STEP step=576 loss=3.308516 Training: 3%|▎ | 576/18018 [1:17:30<34:34:15, 7.14s/step, loss=3.309, acc=0.424, acc_len=0.92, thru=17.4, I=24.0, T=34.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:44,095] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 14:47:44,055] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=575: forward=2622.6ms backward=1053.1ms [2026-10-18 14:47:51,204] loop.py:45 INFO TRAIN_STEP step=577 loss=3.366990 Training: 3%|▎ | 577/18018 [1:17:37<34:34:45, 7.14s/step, loss=3.367, acc=0.416, acc_len=0.91, thru=16.0, I=18.1, T=32.9, wait=3.2s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:47:59,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:48:00,064] loop.py:45 INFO TRAIN_STEP step=578 loss=3.112904 Training: 3%|▎ | 578/18018 [1:17:46<37:04:55, 7.65s/step, loss=3.113, acc=0.445, acc_len=0.98, thru=16.7, I=16.8, T=28.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:00,563] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:48:07,195] loop.py:45 INFO TRAIN_STEP step=579 loss=3.362973 Training: 3%|▎ | 579/18018 [1:17:53<36:19:11, 7.50s/step, loss=3.363, acc=0.415, acc_len=0.91, thru=13.6, I=16.1, T=32.4, wait=3.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:14,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:48:15,171] loop.py:380 INFO TIMING step=580: step=4.381s data=1.982s compute=4.344s [fwd=3.106s bwd=1.222s opt=0.003s] dispatch=3.586s [2026-10-18 14:48:15,171] loop.py:45 INFO TRAIN_STEP step=580 loss=3.316482 Training: 3%|▎ | 580/18018 [1:18:01<37:00:37, 7.64s/step, loss=3.316, acc=0.422, acc_len=0.92, thru=16.7, I=17.7, T=29.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:15,634] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:22,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:48:23,475] loop.py:45 INFO TRAIN_STEP step=581 loss=3.354425 Training: 3%|▎ | 581/18018 [1:18:09<37:58:28, 7.84s/step, loss=3.354, acc=0.417, acc_len=0.90, thru=14.1, I=15.7, T=28.4, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:48:23,470] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=580: forward=3253.6ms backward=1204.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:23,823] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:29,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:48:30,404] loop.py:45 INFO TRAIN_STEP step=582 loss=3.471578 Training: 3%|▎ | 582/18018 [1:18:16<36:38:52, 7.57s/step, loss=3.472, acc=0.404, acc_len=0.87, thru=17.9, I=18.6, T=29.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:30,910] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:36,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:48:38,005] loop.py:45 INFO TRAIN_STEP step=583 loss=3.045736 Training: 3%|▎ | 583/18018 [1:18:24<36:41:54, 7.58s/step, loss=3.046, acc=0.447, acc_len=1.00, thru=15.2, I=19.0, T=31.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:38,377] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:45,171] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:48:47,084] loop.py:45 INFO TRAIN_STEP step=584 loss=3.420529 Training: 3%|▎ | 584/18018 [1:18:33<38:52:22, 8.03s/step, loss=3.421, acc=0.408, acc_len=0.87, thru=14.7, I=19.0, T=26.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:47,194] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:54,928] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:48:55,749] loop.py:380 INFO TIMING step=585: step=3.734s data=1.506s compute=3.703s [fwd=2.623s bwd=1.065s opt=0.003s] dispatch=4.922s [2026-10-18 14:48:55,750] loop.py:45 INFO TRAIN_STEP step=585 loss=3.167767 Training: 3%|▎ | 585/18018 [1:18:41<39:48:02, 8.22s/step, loss=3.168, acc=0.435, acc_len=0.95, thru=15.0, I=16.1, T=34.3, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:48:55,937] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:01,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:49:02,183] loop.py:45 INFO TRAIN_STEP step=586 loss=3.201046 Training: 3%|▎ | 586/18018 [1:18:48<37:12:07, 7.68s/step, loss=3.201, acc=0.429, acc_len=0.94, thru=19.5, I=21.4, T=34.9, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:49:02,179] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=585: forward=2589.1ms backward=1035.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:02,258] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:07,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:49:08,516] loop.py:45 INFO TRAIN_STEP step=587 loss=3.450570 Training: 3%|▎ | 587/18018 [1:18:54<35:14:32, 7.28s/step, loss=3.451, acc=0.407, acc_len=0.88, thru=19.9, I=24.1, T=32.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:08,752] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:15,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:49:16,184] loop.py:45 INFO TRAIN_STEP step=588 loss=3.388971 Training: 3%|▎ | 588/18018 [1:19:02<35:48:23, 7.40s/step, loss=3.389, acc=0.410, acc_len=0.88, thru=15.6, I=19.0, T=30.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:16,531] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:21,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:49:22,593] loop.py:45 INFO TRAIN_STEP step=589 loss=3.198487 Training: 3%|▎ | 589/18018 [1:19:08<34:22:25, 7.10s/step, loss=3.198, acc=0.429, acc_len=0.94, thru=18.9, I=23.0, T=32.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:22,928] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:28,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:49:29,693] loop.py:380 INFO TIMING step=590: step=4.268s data=1.890s compute=4.234s [fwd=3.028s bwd=1.190s opt=0.003s] dispatch=2.819s [2026-10-18 14:49:29,693] loop.py:45 INFO TRAIN_STEP step=590 loss=3.302960 Training: 3%|▎ | 590/18018 [1:19:15<34:21:55, 7.10s/step, loss=3.303, acc=0.424, acc_len=0.92, thru=17.0, I=22.2, T=30.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:29,874] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:49:36,507] loop.py:45 INFO TRAIN_STEP step=591 loss=3.353022 Training: 3%|▎ | 591/18018 [1:19:22<33:57:11, 7.01s/step, loss=3.353, acc=0.418, acc_len=0.91, thru=15.5, I=17.9, T=35.3, wait=3.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:49:36,503] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=590: forward=2582.7ms backward=1001.6ms [2026-10-18 14:49:49,174] loop.py:45 INFO TRAIN_STEP step=592 loss=3.206153 Training: 3%|▎ | 592/18018 [1:19:35<42:09:51, 8.71s/step, loss=3.206, acc=0.429, acc_len=0.95, thru=12.0, I=10.3, T=28.8, wait=8.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:56,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:49:56,873] loop.py:45 INFO TRAIN_STEP step=593 loss=3.340928 Training: 3%|▎ | 593/18018 [1:19:42<40:41:01, 8.41s/step, loss=3.341, acc=0.419, acc_len=0.90, thru=17.7, I=18.2, T=30.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:49:57,059] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:02,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:50:03,291] loop.py:45 INFO TRAIN_STEP step=594 loss=3.155171 Training: 3%|▎ | 594/18018 [1:19:49<37:48:00, 7.81s/step, loss=3.155, acc=0.442, acc_len=0.97, thru=18.0, I=22.5, T=35.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:03,674] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:50:11,375] loop.py:380 INFO TIMING step=595: step=4.256s data=1.770s compute=4.216s [fwd=3.202s bwd=0.999s opt=0.003s] dispatch=3.819s [2026-10-18 14:50:11,375] loop.py:45 INFO TRAIN_STEP step=595 loss=3.234687 Training: 3%|▎ | 595/18018 [1:19:57<38:11:48, 7.89s/step, loss=3.235, acc=0.431, acc_len=0.94, thru=14.2, I=16.1, T=30.1, wait=3.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:18,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:50:19,414] loop.py:45 INFO TRAIN_STEP step=596 loss=3.592274 Training: 3%|▎ | 596/18018 [1:20:05<38:24:16, 7.94s/step, loss=3.592, acc=0.385, acc_len=0.82, thru=17.8, I=18.8, T=29.7, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:50:19,409] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=595: forward=2975.0ms backward=1285.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:19,654] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:27,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:50:28,371] loop.py:45 INFO TRAIN_STEP step=597 loss=3.399395 Training: 3%|▎ | 597/18018 [1:20:14<39:53:08, 8.24s/step, loss=3.399, acc=0.420, acc_len=0.91, thru=14.7, I=16.3, T=25.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:28,844] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:35,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:50:36,209] loop.py:45 INFO TRAIN_STEP step=598 loss=3.394308 Training: 3%|▎ | 598/18018 [1:20:22<39:17:39, 8.12s/step, loss=3.394, acc=0.410, acc_len=0.88, thru=15.6, I=16.5, T=28.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:36,688] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 14:50:42,483] loop.py:45 INFO TRAIN_STEP step=599 loss=3.229025 Training: 3%|▎ | 599/18018 [1:20:28<36:36:54, 7.57s/step, loss=3.229, acc=0.427, acc_len=0.93, thru=17.7, I=18.4, T=33.8, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:49,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:50:50,715] loop.py:380 INFO TIMING step=600: step=4.367s data=1.919s compute=4.336s [fwd=3.099s bwd=1.221s opt=0.003s] dispatch=3.854s [2026-10-18 14:50:50,715] loop.py:45 INFO TRAIN_STEP step=600 loss=3.307119 Training: 3%|▎ | 600/18018 [1:20:36<37:34:42, 7.77s/step, loss=3.307, acc=0.431, acc_len=0.94, thru=17.8, I=16.8, T=29.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:51,072] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:58,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:50:58,747] loop.py:45 INFO TRAIN_STEP step=601 loss=3.218920 Training: 3%|▎ | 601/18018 [1:20:44<37:57:34, 7.85s/step, loss=3.219, acc=0.437, acc_len=0.95, thru=15.7, I=19.1, T=30.1, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:50:58,742] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=600: forward=3185.2ms backward=1011.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:50:58,954] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:51:05,488] loop.py:45 INFO TRAIN_STEP step=602 loss=3.409687 Training: 3%|▎ | 602/18018 [1:20:51<36:21:13, 7.51s/step, loss=3.410, acc=0.411, acc_len=0.89, thru=16.4, I=16.3, T=34.5, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:12,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:51:13,188] loop.py:45 INFO TRAIN_STEP step=603 loss=3.684301 Training: 3%|▎ | 603/18018 [1:20:59<36:37:15, 7.57s/step, loss=3.684, acc=0.380, acc_len=0.81, thru=17.7, I=18.0, T=33.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:13,218] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:19,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:51:20,654] loop.py:45 INFO TRAIN_STEP step=604 loss=3.333594 Training: 3%|▎ | 604/18018 [1:21:06<36:27:58, 7.54s/step, loss=3.334, acc=0.425, acc_len=0.92, thru=18.4, I=20.3, T=33.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:20,900] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:26,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:51:26,725] loop.py:380 INFO TIMING step=605: step=3.509s data=1.490s compute=3.472s [fwd=2.484s bwd=0.973s opt=0.003s] dispatch=2.555s [2026-10-18 14:51:26,725] loop.py:45 INFO TRAIN_STEP step=605 loss=3.369363 Training: 3%|▎ | 605/18018 [1:21:12<34:20:16, 7.10s/step, loss=3.369, acc=0.419, acc_len=0.91, thru=19.6, I=22.4, T=36.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:26,808] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:32,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:51:34,325] loop.py:45 INFO TRAIN_STEP step=606 loss=3.423956 Training: 3%|▎ | 606/18018 [1:21:20<35:03:31, 7.25s/step, loss=3.424, acc=0.413, acc_len=0.89, thru=17.4, I=21.1, T=34.3, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:51:34,321] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=605: forward=2625.4ms backward=1061.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:34,820] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:40,237] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:51:41,064] loop.py:45 INFO TRAIN_STEP step=607 loss=3.124262 Training: 3%|▎ | 607/18018 [1:21:27<34:19:19, 7.10s/step, loss=3.124, acc=0.443, acc_len=0.98, thru=19.0, I=25.7, T=34.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:41,239] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:46,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:51:47,954] loop.py:45 INFO TRAIN_STEP step=608 loss=3.153395 Training: 3%|▎ | 608/18018 [1:21:34<34:01:07, 7.03s/step, loss=3.153, acc=0.435, acc_len=0.95, thru=18.5, I=23.5, T=31.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:48,385] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:53,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:51:54,406] loop.py:45 INFO TRAIN_STEP step=609 loss=3.216261 Training: 3%|▎ | 609/18018 [1:21:40<33:10:14, 6.86s/step, loss=3.216, acc=0.443, acc_len=0.97, thru=18.1, I=25.2, T=34.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:51:54,660] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:52:00,599] loop.py:380 INFO TIMING step=610: step=3.790s data=1.657s compute=3.751s [fwd=2.637s bwd=1.100s opt=0.003s] dispatch=2.393s [2026-10-18 14:52:00,599] loop.py:45 INFO TRAIN_STEP step=610 loss=3.098745 Training: 3%|▎ | 610/18018 [1:21:46<32:12:04, 6.66s/step, loss=3.099, acc=0.447, acc_len=0.98, thru=18.9, I=24.9, T=33.8, wait=2.4s, pool=56, epoch=1/2][2026-10-18 14:52:09,286] loop.py:45 INFO TRAIN_STEP step=611 loss=3.290042 Training: 3%|▎ | 611/18018 [1:21:55<35:08:57, 7.27s/step, loss=3.290, acc=0.425, acc_len=0.92, thru=11.8, I=14.8, T=31.0, wait=4.6s, pool=28, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:52:09,282] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=610: forward=2825.7ms backward=1234.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:18,927] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:52:20,135] loop.py:45 INFO TRAIN_STEP step=612 loss=3.324084 Training: 3%|▎ | 612/18018 [1:22:06<40:20:02, 8.34s/step, loss=3.324, acc=0.425, acc_len=0.92, thru=15.9, I=15.6, T=31.5, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:20,440] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:26,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:52:27,724] loop.py:45 INFO TRAIN_STEP step=613 loss=3.338792 Training: 3%|▎ | 613/18018 [1:22:13<39:14:15, 8.12s/step, loss=3.339, acc=0.414, acc_len=0.89, thru=16.1, I=19.8, T=31.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:28,152] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:52:34,750] loop.py:45 INFO TRAIN_STEP step=614 loss=3.138441 Training: 3%|▎ | 614/18018 [1:22:20<37:39:05, 7.79s/step, loss=3.138, acc=0.434, acc_len=0.95, thru=15.5, I=18.3, T=29.8, wait=2.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:41,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:52:43,810] loop.py:380 INFO TIMING step=615: step=5.249s data=2.677s compute=5.213s [fwd=3.677s bwd=1.521s opt=0.003s] dispatch=3.804s [2026-10-18 14:52:43,810] loop.py:45 INFO TRAIN_STEP step=615 loss=3.324149 Training: 3%|▎ | 615/18018 [1:22:29<39:29:46, 8.17s/step, loss=3.324, acc=0.422, acc_len=0.91, thru=16.1, I=18.2, T=24.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:47,839] inference_manager.py:398 INFO Still waiting for pool capacity (waited 6.1s): pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:47,839] inference_manager.py:408 INFO Pool capacity available after 6.1s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:53,639] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:52:54,127] loop.py:45 INFO TRAIN_STEP step=616 loss=3.225119 Training: 3%|▎ | 616/18018 [1:22:40<42:36:33, 8.81s/step, loss=3.225, acc=0.429, acc_len=0.93, thru=20.4, I=24.3, T=33.2, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:54,162] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 14:52:54,123] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=615: forward=2703.8ms backward=1110.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:52:59,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:53:00,961] loop.py:45 INFO TRAIN_STEP step=617 loss=3.111031 Training: 3%|▎ | 617/18018 [1:22:47<39:44:08, 8.22s/step, loss=3.111, acc=0.449, acc_len=0.99, thru=17.7, I=20.8, T=35.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:01,124] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:06,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:53:07,904] loop.py:45 INFO TRAIN_STEP step=618 loss=3.145567 Training: 3%|▎ | 618/18018 [1:22:53<37:52:37, 7.84s/step, loss=3.146, acc=0.437, acc_len=0.95, thru=17.7, I=24.8, T=29.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:08,045] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:14,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:53:15,158] loop.py:45 INFO TRAIN_STEP step=619 loss=3.308290 Training: 3%|▎ | 619/18018 [1:23:01<37:01:51, 7.66s/step, loss=3.308, acc=0.423, acc_len=0.91, thru=16.2, I=20.8, T=29.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:15,590] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:53:21,365] loop.py:380 INFO TIMING step=620: step=3.853s data=1.749s compute=3.814s [fwd=2.708s bwd=1.091s opt=0.003s] dispatch=2.346s [2026-10-18 14:53:21,366] loop.py:45 INFO TRAIN_STEP step=620 loss=3.308871 Training: 3%|▎ | 620/18018 [1:23:07<34:55:23, 7.23s/step, loss=3.309, acc=0.416, acc_len=0.90, thru=17.0, I=21.1, T=33.2, wait=2.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:29,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:53:31,196] loop.py:45 INFO TRAIN_STEP step=621 loss=3.493259 Training: 3%|▎ | 621/18018 [1:23:17<38:41:29, 8.01s/step, loss=3.493, acc=0.404, acc_len=0.86, thru=14.4, I=15.3, T=30.2, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:53:31,192] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=620: forward=2899.9ms backward=1259.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:31,271] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:38,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:53:39,593] loop.py:45 INFO TRAIN_STEP step=622 loss=3.197694 Training: 3%|▎ | 622/18018 [1:23:25<39:15:26, 8.12s/step, loss=3.198, acc=0.441, acc_len=0.96, thru=14.0, I=19.1, T=31.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:39,880] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:45,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:53:46,768] loop.py:45 INFO TRAIN_STEP step=623 loss=3.325795 Training: 3%|▎ | 623/18018 [1:23:32<37:52:37, 7.84s/step, loss=3.326, acc=0.417, acc_len=0.91, thru=16.9, I=22.2, T=30.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:46,977] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:53:54,232] loop.py:45 INFO TRAIN_STEP step=624 loss=3.391571 Training: 3%|▎ | 624/18018 [1:23:40<37:20:03, 7.73s/step, loss=3.392, acc=0.413, acc_len=0.89, thru=15.8, I=18.4, T=31.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:54,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:53:54,677] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:54:02,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:54:02,567] loop.py:380 INFO TIMING step=625: step=3.952s data=1.725s compute=3.919s [fwd=2.803s bwd=1.102s opt=0.003s] dispatch=4.373s [2026-10-18 14:54:02,567] loop.py:45 INFO TRAIN_STEP step=625 loss=3.001756 Training: 3%|▎ | 625/18018 [1:23:48<38:12:46, 7.91s/step, loss=3.002, acc=0.445, acc_len=0.98, thru=15.4, I=15.6, T=32.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:54:02,673] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-18 14:54:11,404] loop.py:45 INFO TRAIN_STEP step=626 loss=3.527432 Training: 3%|▎ | 626/18018 [1:23:57<39:33:45, 8.19s/step, loss=3.527, acc=0.400, acc_len=0.85, thru=13.1, I=14.9, T=32.3, wait=4.9s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:54:11,396] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=625: forward=2838.2ms backward=1074.1ms [2026-10-18 14:54:19,373] loop.py:45 INFO TRAIN_STEP step=627 loss=3.399945 Training: 3%|▎ | 627/18018 [1:24:05<39:13:57, 8.12s/step, loss=3.400, acc=0.410, acc_len=0.88, thru=14.8, I=14.1, T=32.4, wait=4.0s, pool=52, epoch=1/2][2026-10-18 14:54:25,862] loop.py:45 INFO TRAIN_STEP step=628 loss=3.270026 Training: 3%|▎ | 628/18018 [1:24:11<36:52:00, 7.63s/step, loss=3.270, acc=0.429, acc_len=0.92, thru=19.6, I=18.5, T=33.4, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:54:33,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:54:33,567] loop.py:45 INFO TRAIN_STEP step=629 loss=3.160244 Training: 3%|▎ | 629/18018 [1:24:19<36:58:20, 7.65s/step, loss=3.160, acc=0.440, acc_len=0.96, thru=17.1, I=16.7, T=29.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:54:33,756] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:54:40,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:54:41,722] loop.py:380 INFO TIMING step=630: step=4.207s data=1.737s compute=4.168s [fwd=3.105s bwd=1.049s opt=0.003s] dispatch=3.938s [2026-10-18 14:54:41,723] loop.py:45 INFO TRAIN_STEP step=630 loss=3.381737 Training: 3%|▎ | 630/18018 [1:24:27<37:41:45, 7.80s/step, loss=3.382, acc=0.412, acc_len=0.89, thru=15.7, I=17.4, T=30.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:54:41,844] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:54:49,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:54:49,786] loop.py:45 INFO TRAIN_STEP step=631 loss=3.295322 Training: 4%|▎ | 631/18018 [1:24:35<38:03:59, 7.88s/step, loss=3.295, acc=0.428, acc_len=0.93, thru=15.4, I=18.1, T=34.8, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:54:49,781] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=630: forward=2584.0ms backward=1043.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:54:50,050] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:54:57,336] loop.py:45 INFO TRAIN_STEP step=632 loss=3.351451 Training: 4%|▎ | 632/18018 [1:24:43<37:35:04, 7.78s/step, loss=3.351, acc=0.409, acc_len=0.88, thru=16.2, I=16.7, T=35.8, wait=4.0s, pool=48, epoch=1/2][2026-10-18 14:55:05,391] loop.py:45 INFO TRAIN_STEP step=633 loss=3.394033 Training: 4%|▎ | 633/18018 [1:24:51<37:58:41, 7.86s/step, loss=3.394, acc=0.417, acc_len=0.89, thru=15.6, I=18.3, T=34.1, wait=4.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:55:12,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:55:12,768] loop.py:45 INFO TRAIN_STEP step=634 loss=3.164837 Training: 4%|▎ | 634/18018 [1:24:58<37:16:13, 7.72s/step, loss=3.165, acc=0.450, acc_len=0.98, thru=16.4, I=17.2, T=36.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:55:13,056] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:55:19,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:55:19,942] loop.py:380 INFO TIMING step=635: step=3.915s data=1.784s compute=3.883s [fwd=2.803s bwd=1.066s opt=0.003s] dispatch=3.249s [2026-10-18 14:55:19,942] loop.py:45 INFO TRAIN_STEP step=635 loss=3.064853 Training: 4%|▎ | 635/18018 [1:25:06<36:28:48, 7.56s/step, loss=3.065, acc=0.451, acc_len=1.00, thru=18.7, I=20.2, T=32.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:55:20,192] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:55:26,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:55:27,057] loop.py:45 INFO TRAIN_STEP step=636 loss=3.143223 Training: 4%|▎ | 636/18018 [1:25:13<35:50:27, 7.42s/step, loss=3.143, acc=0.447, acc_len=0.98, thru=17.1, I=20.5, T=34.6, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:55:27,051] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=635: forward=2610.1ms backward=1042.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:55:27,241] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:55:34,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:55:35,927] loop.py:45 INFO TRAIN_STEP step=637 loss=3.425545 Training: 4%|▎ | 637/18018 [1:25:22<37:56:01, 7.86s/step, loss=3.426, acc=0.418, acc_len=0.90, thru=13.6, I=15.6, T=30.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:55:36,429] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:55:45,391] loop.py:45 INFO TRAIN_STEP step=638 loss=3.320866 Training: 4%|▎ | 638/18018 [1:25:31<40:15:35, 8.34s/step, loss=3.321, acc=0.419, acc_len=0.91, thru=13.8, I=16.8, T=27.5, wait=4.8s, pool=60, epoch=1/2][2026-10-18 14:55:54,012] loop.py:45 INFO TRAIN_STEP step=639 loss=3.272588 Training: 4%|▎ | 639/18018 [1:25:40<40:40:00, 8.42s/step, loss=3.273, acc=0.424, acc_len=0.92, thru=11.9, I=13.6, T=26.9, wait=3.9s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:05,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:56:05,980] loop.py:380 INFO TIMING step=640: step=4.804s data=2.201s compute=4.768s [fwd=3.347s bwd=1.406s opt=0.003s] dispatch=7.152s [2026-10-18 14:56:05,980] loop.py:45 INFO TRAIN_STEP step=640 loss=3.225703 Training: 4%|▎ | 640/18018 [1:25:52<45:47:35, 9.49s/step, loss=3.226, acc=0.429, acc_len=0.93, thru=14.9, I=14.5, T=26.6, wait=7.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:06,142] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:11,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 14:56:12,925] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=640: forward=2832.9ms backward=1143.5ms [2026-10-18 14:56:12,930] loop.py:45 INFO TRAIN_STEP step=641 loss=3.376877 Training: 4%|▎ | 641/18018 [1:25:59<42:07:13, 8.73s/step, loss=3.377, acc=0.413, acc_len=0.89, thru=18.9, I=21.0, T=31.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:13,254] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 14:56:20,275] loop.py:45 INFO TRAIN_STEP step=642 loss=3.277787 Training: 4%|▎ | 642/18018 [1:26:06<40:07:08, 8.31s/step, loss=3.278, acc=0.424, acc_len=0.92, thru=14.5, I=18.9, T=31.9, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:27,293] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:56:29,493] loop.py:45 INFO TRAIN_STEP step=643 loss=3.301201 Training: 4%|▎ | 643/18018 [1:26:15<41:25:40, 8.58s/step, loss=3.301, acc=0.427, acc_len=0.92, thru=14.8, I=18.2, T=28.9, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:29,809] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:37,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:56:38,503] loop.py:45 INFO TRAIN_STEP step=644 loss=3.249545 Training: 4%|▎ | 644/18018 [1:26:24<42:02:32, 8.71s/step, loss=3.250, acc=0.427, acc_len=0.93, thru=14.7, I=17.0, T=32.0, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:38,908] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 14:56:45,500] loop.py:380 INFO TIMING step=645: step=3.828s data=1.672s compute=3.802s [fwd=2.729s bwd=1.058s opt=0.003s] dispatch=3.159s [2026-10-18 14:56:45,500] loop.py:45 INFO TRAIN_STEP step=645 loss=3.073056 Training: 4%|▎ | 645/18018 [1:26:31<39:33:19, 8.20s/step, loss=3.073, acc=0.440, acc_len=0.97, thru=15.2, I=17.8, T=33.4, wait=3.2s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:57,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:56:58,911] loop.py:45 INFO TRAIN_STEP step=646 loss=3.402493 Training: 4%|▎ | 646/18018 [1:26:44<47:06:15, 9.76s/step, loss=3.402, acc=0.418, acc_len=0.90, thru=13.8, I=13.4, T=26.8, wait=8.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:56:58,906] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=645: forward=3332.9ms backward=1381.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:56:59,007] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 14:57:06,873] loop.py:45 INFO TRAIN_STEP step=647 loss=3.141183 Training: 4%|▎ | 647/18018 [1:26:52<44:29:36, 9.22s/step, loss=3.141, acc=0.443, acc_len=0.97, thru=12.9, I=14.1, T=29.4, wait=3.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:57:14,188] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:57:14,352] loop.py:45 INFO TRAIN_STEP step=648 loss=3.204973 Training: 4%|▎ | 648/18018 [1:27:00<41:58:11, 8.70s/step, loss=3.205, acc=0.433, acc_len=0.95, thru=19.5, I=16.0, T=30.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:57:14,690] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:57:23,800] loop.py:45 INFO TRAIN_STEP step=649 loss=3.049590 Training: 4%|▎ | 649/18018 [1:27:09<43:03:20, 8.92s/step, loss=3.050, acc=0.455, acc_len=1.00, thru=13.2, I=13.8, T=29.1, wait=5.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:57:32,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:57:32,738] loop.py:380 INFO TIMING step=650: step=4.401s data=1.941s compute=4.351s [fwd=2.928s bwd=1.408s opt=0.003s] dispatch=4.526s [2026-10-18 14:57:32,738] loop.py:45 INFO TRAIN_STEP step=650 loss=3.336374 Training: 4%|▎ | 650/18018 [1:27:18<43:04:24, 8.93s/step, loss=3.336, acc=0.427, acc_len=0.93, thru=16.0, I=16.1, T=29.1, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:57:32,867] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-18 14:57:40,761] loop.py:45 INFO TRAIN_STEP step=651 loss=3.270342 Training: 4%|▎ | 651/18018 [1:27:26<41:45:28, 8.66s/step, loss=3.270, acc=0.422, acc_len=0.92, thru=14.6, I=13.7, T=31.7, wait=4.0s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:57:40,757] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=650: forward=2904.5ms backward=1088.9ms [2026-10-18 14:57:47,889] loop.py:45 INFO TRAIN_STEP step=652 loss=3.272855 Training: 4%|▎ | 652/18018 [1:27:33<39:32:35, 8.20s/step, loss=3.273, acc=0.427, acc_len=0.92, thru=16.5, I=16.3, T=32.9, wait=3.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:57:53,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:57:54,790] loop.py:45 INFO TRAIN_STEP step=653 loss=3.128866 Training: 4%|▎ | 653/18018 [1:27:40<37:39:58, 7.81s/step, loss=3.129, acc=0.449, acc_len=0.98, thru=18.1, I=19.4, T=33.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:57:54,835] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:57:59,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:58:01,509] loop.py:45 INFO TRAIN_STEP step=654 loss=3.146221 Training: 4%|▎ | 654/18018 [1:27:47<36:05:13, 7.48s/step, loss=3.146, acc=0.440, acc_len=0.96, thru=19.0, I=23.9, T=31.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:01,615] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:07,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:58:08,514] loop.py:380 INFO TIMING step=655: step=4.159s data=1.896s compute=4.129s [fwd=2.943s bwd=1.172s opt=0.003s] dispatch=2.837s [2026-10-18 14:58:08,514] loop.py:45 INFO TRAIN_STEP step=655 loss=3.282378 Training: 4%|▎ | 655/18018 [1:27:54<35:23:50, 7.34s/step, loss=3.282, acc=0.428, acc_len=0.92, thru=16.4, I=22.9, T=30.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:08,717] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:14,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:58:14,906] loop.py:45 INFO TRAIN_STEP step=656 loss=3.281087 Training: 4%|▎ | 656/18018 [1:28:00<34:01:27, 7.05s/step, loss=3.281, acc=0.430, acc_len=0.93, thru=18.2, I=21.4, T=32.5, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:58:14,902] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=655: forward=2808.2ms backward=1090.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:15,082] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:19,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:58:21,200] loop.py:45 INFO TRAIN_STEP step=657 loss=3.163666 Training: 4%|▎ | 657/18018 [1:28:07<32:55:19, 6.83s/step, loss=3.164, acc=0.435, acc_len=0.96, thru=18.7, I=21.7, T=36.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:21,413] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:26,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:58:27,485] loop.py:45 INFO TRAIN_STEP step=658 loss=2.922229 Training: 4%|▎ | 658/18018 [1:28:13<32:08:10, 6.66s/step, loss=2.922, acc=0.469, acc_len=1.04, thru=20.0, I=25.5, T=33.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:27,891] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:32,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:58:33,221] loop.py:45 INFO TRAIN_STEP step=659 loss=3.107189 Training: 4%|▎ | 659/18018 [1:28:19<30:47:23, 6.39s/step, loss=3.107, acc=0.448, acc_len=0.98, thru=20.5, I=26.4, T=36.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:33,304] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:39,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:58:40,476] loop.py:380 INFO TIMING step=660: step=4.116s data=1.717s compute=4.083s [fwd=2.838s bwd=1.230s opt=0.003s] dispatch=3.130s [2026-10-18 14:58:40,476] loop.py:45 INFO TRAIN_STEP step=660 loss=3.172043 Training: 4%|▎ | 660/18018 [1:28:26<32:02:51, 6.65s/step, loss=3.172, acc=0.431, acc_len=0.94, thru=18.1, I=20.8, T=31.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:40,640] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:45,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:58:46,670] loop.py:45 INFO TRAIN_STEP step=661 loss=3.109897 Training: 4%|▎ | 661/18018 [1:28:32<31:23:16, 6.51s/step, loss=3.110, acc=0.442, acc_len=0.97, thru=18.4, I=23.3, T=33.6, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:58:46,664] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=660: forward=2683.9ms backward=1080.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:46,881] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:53,917] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:58:54,735] loop.py:45 INFO TRAIN_STEP step=662 loss=3.062562 Training: 4%|▎ | 662/18018 [1:28:40<33:38:20, 6.98s/step, loss=3.063, acc=0.446, acc_len=0.97, thru=14.8, I=19.0, T=34.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:58:54,924] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:02,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:59:02,957] loop.py:45 INFO TRAIN_STEP step=663 loss=3.362736 Training: 4%|▎ | 663/18018 [1:28:49<35:26:09, 7.35s/step, loss=3.363, acc=0.413, acc_len=0.89, thru=15.0, I=16.5, T=33.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:03,274] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 14:59:10,958] loop.py:45 INFO TRAIN_STEP step=664 loss=3.056599 Training: 4%|▎ | 664/18018 [1:28:57<36:22:58, 7.55s/step, loss=3.057, acc=0.453, acc_len=1.00, thru=15.3, I=16.0, T=33.6, wait=4.2s, pool=60, epoch=1/2][2026-10-18 14:59:18,585] loop.py:380 INFO TIMING step=665: step=4.109s data=1.716s compute=4.080s [fwd=2.902s bwd=1.164s opt=0.003s] dispatch=3.502s [2026-10-18 14:59:18,586] loop.py:45 INFO TRAIN_STEP step=665 loss=3.250518 Training: 4%|▎ | 665/18018 [1:29:04<36:29:24, 7.57s/step, loss=3.251, acc=0.423, acc_len=0.91, thru=16.9, I=17.8, T=31.2, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:25,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:59:27,669] loop.py:45 INFO TRAIN_STEP step=666 loss=2.933482 Training: 4%|▎ | 666/18018 [1:29:13<38:40:30, 8.02s/step, loss=2.933, acc=0.464, acc_len=1.03, thru=14.2, I=16.9, T=27.8, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 14:59:27,665] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=665: forward=3184.5ms backward=1360.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:27,984] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:33,181] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:59:33,531] loop.py:45 INFO TRAIN_STEP step=667 loss=3.026795 Training: 4%|▎ | 667/18018 [1:29:19<35:32:49, 7.38s/step, loss=3.027, acc=0.454, acc_len=1.00, thru=18.1, I=24.0, T=36.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:33,698] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:40,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:59:40,575] loop.py:45 INFO TRAIN_STEP step=668 loss=2.974977 Training: 4%|▎ | 668/18018 [1:29:26<35:03:57, 7.28s/step, loss=2.975, acc=0.460, acc_len=1.01, thru=18.2, I=19.9, T=31.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:40,694] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:46,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:59:48,855] loop.py:45 INFO TRAIN_STEP step=669 loss=3.273422 Training: 4%|▎ | 669/18018 [1:29:34<36:30:45, 7.58s/step, loss=3.273, acc=0.432, acc_len=0.94, thru=15.6, I=20.4, T=27.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:48,886] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:55,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 14:59:55,744] loop.py:380 INFO TIMING step=670: step=4.076s data=1.814s compute=4.050s [fwd=2.912s bwd=1.124s opt=0.003s] dispatch=2.804s [2026-10-18 14:59:55,744] loop.py:45 INFO TRAIN_STEP step=670 loss=3.519223 Training: 4%|▎ | 670/18018 [1:29:41<35:31:01, 7.37s/step, loss=3.519, acc=0.398, acc_len=0.85, thru=14.4, I=21.2, T=31.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 14:59:56,192] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:02,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:00:02,916] loop.py:45 INFO TRAIN_STEP step=671 loss=3.188959 Training: 4%|▎ | 671/18018 [1:29:49<35:13:50, 7.31s/step, loss=3.189, acc=0.431, acc_len=0.94, thru=18.6, I=21.4, T=31.3, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:00:02,911] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=670: forward=2769.4ms backward=1254.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:03,237] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:11,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:00:11,744] loop.py:45 INFO TRAIN_STEP step=672 loss=2.981413 Training: 4%|▎ | 672/18018 [1:29:57<37:25:05, 7.77s/step, loss=2.981, acc=0.449, acc_len=0.99, thru=13.9, I=15.3, T=32.6, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:12,179] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:19,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:00:19,582] loop.py:45 INFO TRAIN_STEP step=673 loss=3.262173 Training: 4%|▎ | 673/18018 [1:30:05<37:31:20, 7.79s/step, loss=3.262, acc=0.428, acc_len=0.92, thru=18.1, I=18.7, T=27.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:19,960] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:26,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:00:27,746] loop.py:45 INFO TRAIN_STEP step=674 loss=3.233732 Training: 4%|▎ | 674/18018 [1:30:13<38:03:49, 7.90s/step, loss=3.234, acc=0.435, acc_len=0.95, thru=15.3, I=17.4, T=29.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:28,126] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:33,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:00:34,095] loop.py:380 INFO TIMING step=675: step=3.729s data=1.604s compute=3.692s [fwd=2.635s bwd=1.042s opt=0.003s] dispatch=2.611s [2026-10-18 15:00:34,095] loop.py:45 INFO TRAIN_STEP step=675 loss=3.163119 Training: 4%|▎ | 675/18018 [1:30:20<35:49:01, 7.43s/step, loss=3.163, acc=0.435, acc_len=0.95, thru=18.2, I=22.8, T=34.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:34,267] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:42,845] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:00:43,125] loop.py:45 INFO TRAIN_STEP step=676 loss=2.991976 Training: 4%|▍ | 676/18018 [1:30:29<38:07:10, 7.91s/step, loss=2.992, acc=0.458, acc_len=1.01, thru=14.6, I=15.4, T=32.0, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:00:43,121] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=675: forward=2747.5ms backward=1202.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:43,347] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:50,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:00:51,857] loop.py:45 INFO TRAIN_STEP step=677 loss=3.221366 Training: 4%|▍ | 677/18018 [1:30:37<39:18:16, 8.16s/step, loss=3.221, acc=0.438, acc_len=0.95, thru=15.2, I=17.6, T=29.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:52,081] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:00:58,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:01:00,122] loop.py:45 INFO TRAIN_STEP step=678 loss=3.035691 Training: 4%|▍ | 678/18018 [1:30:46<39:27:07, 8.19s/step, loss=3.036, acc=0.457, acc_len=1.01, thru=14.2, I=17.6, T=28.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:01:00,473] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:01:07,357] loop.py:45 INFO TRAIN_STEP step=679 loss=3.164548 Training: 4%|▍ | 679/18018 [1:30:53<38:04:12, 7.90s/step, loss=3.165, acc=0.444, acc_len=0.97, thru=16.1, I=18.3, T=29.3, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:01:15,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:01:17,013] loop.py:380 INFO TIMING step=680: step=4.039s data=1.668s compute=3.993s [fwd=2.835s bwd=1.143s opt=0.003s] dispatch=5.606s [2026-10-18 15:01:17,013] loop.py:45 INFO TRAIN_STEP step=680 loss=3.321114 Training: 4%|▍ | 680/18018 [1:31:03<40:35:56, 8.43s/step, loss=3.321, acc=0.430, acc_len=0.92, thru=14.4, I=17.2, T=31.7, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:01:17,123] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-18 15:01:23,579] loop.py:45 INFO TRAIN_STEP step=681 loss=2.956403 Training: 4%|▍ | 681/18018 [1:31:09<37:54:16, 7.87s/step, loss=2.956, acc=0.468, acc_len=1.04, thru=16.3, I=23.6, T=34.0, wait=2.8s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:01:23,575] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=680: forward=2621.6ms backward=1098.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:01:29,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:01:30,573] loop.py:45 INFO TRAIN_STEP step=682 loss=3.056261 Training: 4%|▍ | 682/18018 [1:31:16<36:38:00, 7.61s/step, loss=3.056, acc=0.456, acc_len=1.00, thru=18.2, I=20.9, T=32.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:01:30,670] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 15:01:36,804] loop.py:45 INFO TRAIN_STEP step=683 loss=2.901488 Training: 4%|▍ | 683/18018 [1:31:22<34:38:47, 7.20s/step, loss=2.901, acc=0.477, acc_len=1.06, thru=18.7, I=22.3, T=35.2, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:01:43,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:01:45,811] loop.py:45 INFO TRAIN_STEP step=684 loss=3.157093 Training: 4%|▍ | 684/18018 [1:31:31<37:15:26, 7.74s/step, loss=3.157, acc=0.440, acc_len=0.96, thru=15.4, I=17.9, T=28.5, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:01:45,858] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:01:51,755] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:01:52,137] loop.py:380 INFO TIMING step=685: step=3.639s data=1.605s compute=3.602s [fwd=2.538s bwd=1.050s opt=0.003s] dispatch=2.679s [2026-10-18 15:01:52,137] loop.py:45 INFO TRAIN_STEP step=685 loss=3.062999 Training: 4%|▍ | 685/18018 [1:31:38<35:13:08, 7.31s/step, loss=3.063, acc=0.448, acc_len=0.99, thru=18.0, I=22.5, T=35.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:01:52,262] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-18 15:01:58,763] loop.py:45 INFO TRAIN_STEP step=686 loss=3.258335 Training: 4%|▍ | 686/18018 [1:31:44<34:13:07, 7.11s/step, loss=3.258, acc=0.435, acc_len=0.94, thru=18.2, I=19.8, T=32.8, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:01:58,758] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=685: forward=2697.7ms backward=1138.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:02:04,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:02:04,851] loop.py:45 INFO TRAIN_STEP step=687 loss=3.077525 Training: 4%|▍ | 687/18018 [1:31:50<32:44:48, 6.80s/step, loss=3.078, acc=0.442, acc_len=0.96, thru=19.1, I=19.7, T=33.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:02:04,974] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 15:02:12,012] loop.py:45 INFO TRAIN_STEP step=688 loss=3.122325 Training: 4%|▍ | 688/18018 [1:31:58<33:15:45, 6.91s/step, loss=3.122, acc=0.442, acc_len=0.96, thru=14.1, I=18.1, T=35.0, wait=3.5s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:02:21,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:02:23,447] loop.py:45 INFO TRAIN_STEP step=689 loss=3.191087 Training: 4%|▍ | 689/18018 [1:32:09<39:47:41, 8.27s/step, loss=3.191, acc=0.442, acc_len=0.95, thru=14.5, I=16.2, T=28.9, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:02:23,638] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:02:30,313] loop.py:380 INFO TIMING step=690: step=3.771s data=1.645s compute=3.740s [fwd=2.654s bwd=1.070s opt=0.003s] dispatch=3.086s [2026-10-18 15:02:30,314] loop.py:45 INFO TRAIN_STEP step=690 loss=3.108010 Training: 4%|▍ | 690/18018 [1:32:16<37:46:23, 7.85s/step, loss=3.108, acc=0.446, acc_len=0.98, thru=14.5, I=23.8, T=33.9, wait=3.1s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:02:39,361] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=690: forward=2763.2ms backward=1072.4ms [2026-10-18 15:02:39,366] loop.py:45 INFO TRAIN_STEP step=691 loss=3.100533 Training: 4%|▍ | 691/18018 [1:32:25<39:30:23, 8.21s/step, loss=3.101, acc=0.446, acc_len=0.97, thru=13.7, I=12.4, T=33.0, wait=5.2s, pool=36, epoch=1/2][2026-10-18 15:02:48,224] loop.py:45 INFO TRAIN_STEP step=692 loss=3.241070 Training: 4%|▍ | 692/18018 [1:32:34<40:26:31, 8.40s/step, loss=3.241, acc=0.434, acc_len=0.94, thru=15.7, I=14.3, T=33.4, wait=5.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:02:55,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:02:57,228] loop.py:45 INFO TRAIN_STEP step=693 loss=3.041411 Training: 4%|▍ | 693/18018 [1:32:43<41:18:30, 8.58s/step, loss=3.041, acc=0.456, acc_len=1.00, thru=15.7, I=18.1, T=29.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:02:57,572] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:02,325] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:03:02,768] loop.py:45 INFO TRAIN_STEP step=694 loss=3.264297 Training: 4%|▍ | 694/18018 [1:32:48<36:54:53, 7.67s/step, loss=3.264, acc=0.430, acc_len=0.93, thru=19.8, I=28.8, T=38.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:02,846] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:09,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:03:09,841] loop.py:380 INFO TIMING step=695: step=3.757s data=1.618s compute=3.710s [fwd=2.648s bwd=1.049s opt=0.003s] dispatch=3.306s [2026-10-18 15:03:09,841] loop.py:45 INFO TRAIN_STEP step=695 loss=3.087590 Training: 4%|▍ | 695/18018 [1:32:55<36:02:51, 7.49s/step, loss=3.088, acc=0.443, acc_len=0.97, thru=18.7, I=19.8, T=34.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:10,227] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:03:17,039] loop.py:45 INFO TRAIN_STEP step=696 loss=2.986290 Training: 4%|▍ | 696/18018 [1:33:03<35:37:28, 7.40s/step, loss=2.986, acc=0.459, acc_len=1.01, thru=16.8, I=20.3, T=31.3, wait=3.1s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:03:17,033] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=695: forward=2880.8ms backward=1165.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:25,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:03:27,081] loop.py:45 INFO TRAIN_STEP step=697 loss=2.996020 Training: 4%|▍ | 697/18018 [1:33:13<39:25:51, 8.20s/step, loss=2.996, acc=0.456, acc_len=1.01, thru=15.3, I=14.1, T=29.6, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:27,155] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:32,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:03:34,164] loop.py:45 INFO TRAIN_STEP step=698 loss=3.125175 Training: 4%|▍ | 698/18018 [1:33:20<37:49:08, 7.86s/step, loss=3.125, acc=0.445, acc_len=0.98, thru=16.8, I=23.1, T=29.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:34,195] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:40,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:03:40,912] loop.py:45 INFO TRAIN_STEP step=699 loss=3.019818 Training: 4%|▍ | 699/18018 [1:33:26<36:12:41, 7.53s/step, loss=3.020, acc=0.457, acc_len=1.00, thru=16.5, I=21.6, T=31.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:41,075] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:49,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:03:50,719] loop.py:380 INFO TIMING step=700: step=4.007s data=1.714s compute=3.963s [fwd=2.906s bwd=1.042s opt=0.003s] dispatch=5.791s [2026-10-18 15:03:50,719] loop.py:45 INFO TRAIN_STEP step=700 loss=3.092359 Training: 4%|▍ | 700/18018 [1:33:36<39:30:06, 8.21s/step, loss=3.092, acc=0.449, acc_len=0.98, thru=13.3, I=14.4, T=31.9, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:51,190] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:57,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:03:58,340] loop.py:45 INFO TRAIN_STEP step=701 loss=3.101292 Training: 4%|▍ | 701/18018 [1:33:44<38:38:52, 8.03s/step, loss=3.101, acc=0.447, acc_len=0.98, thru=16.7, I=20.3, T=29.8, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:03:58,335] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=700: forward=2911.7ms backward=1331.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:03:58,679] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:04:07,816] loop.py:45 INFO TRAIN_STEP step=702 loss=3.148124 Training: 4%|▍ | 702/18018 [1:33:53<40:43:22, 8.47s/step, loss=3.148, acc=0.438, acc_len=0.95, thru=13.1, I=13.1, T=27.6, wait=4.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:15,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:04:16,341] loop.py:45 INFO TRAIN_STEP step=703 loss=3.096034 Training: 4%|▍ | 703/18018 [1:34:02<40:48:17, 8.48s/step, loss=3.096, acc=0.453, acc_len=0.98, thru=17.0, I=17.7, T=29.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:16,761] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:24,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:04:24,349] loop.py:45 INFO TRAIN_STEP step=704 loss=3.009374 Training: 4%|▍ | 704/18018 [1:34:10<40:07:09, 8.34s/step, loss=3.009, acc=0.458, acc_len=1.01, thru=15.5, I=17.4, T=29.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:24,561] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:30,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:04:32,328] loop.py:380 INFO TIMING step=705: step=4.263s data=1.974s compute=4.223s [fwd=2.957s bwd=1.252s opt=0.003s] dispatch=3.705s [2026-10-18 15:04:32,328] loop.py:45 INFO TRAIN_STEP step=705 loss=2.861395 Training: 4%|▍ | 705/18018 [1:34:18<39:35:26, 8.23s/step, loss=2.861, acc=0.482, acc_len=1.07, thru=16.6, I=19.0, T=30.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:32,554] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:37,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:04:39,148] loop.py:45 INFO TRAIN_STEP step=706 loss=3.145129 Training: 4%|▍ | 706/18018 [1:34:25<37:33:04, 7.81s/step, loss=3.145, acc=0.445, acc_len=0.97, thru=17.1, I=24.4, T=32.6, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:04:39,144] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=705: forward=2747.7ms backward=1122.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:39,212] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:45,265] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:04:47,249] loop.py:45 INFO TRAIN_STEP step=707 loss=2.918221 Training: 4%|▍ | 707/18018 [1:34:33<37:58:23, 7.90s/step, loss=2.918, acc=0.465, acc_len=1.02, thru=14.6, I=20.8, T=30.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:47,296] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:51,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:04:53,006] loop.py:45 INFO TRAIN_STEP step=708 loss=2.887084 Training: 4%|▍ | 708/18018 [1:34:39<34:52:53, 7.25s/step, loss=2.887, acc=0.469, acc_len=1.03, thru=18.9, I=26.4, T=35.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:53,466] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:59,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:04:59,584] loop.py:45 INFO TRAIN_STEP step=709 loss=3.053348 Training: 4%|▍ | 709/18018 [1:34:45<33:54:14, 7.05s/step, loss=3.053, acc=0.448, acc_len=0.98, thru=19.3, I=22.9, T=33.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:04:59,768] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:04,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:05:05,062] loop.py:380 INFO TIMING step=710: step=3.312s data=1.342s compute=3.285s [fwd=2.357s bwd=0.916s opt=0.003s] dispatch=2.157s [2026-10-18 15:05:05,062] loop.py:45 INFO TRAIN_STEP step=710 loss=2.930848 Training: 4%|▍ | 710/18018 [1:34:51<31:38:06, 6.58s/step, loss=2.931, acc=0.468, acc_len=1.03, thru=20.8, I=22.1, T=38.6, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:05,397] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:11,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 15:05:12,538] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=710: forward=2786.5ms backward=1072.1ms [2026-10-18 15:05:12,542] loop.py:45 INFO TRAIN_STEP step=711 loss=3.209078 Training: 4%|▍ | 711/18018 [1:34:58<32:55:54, 6.85s/step, loss=3.209, acc=0.431, acc_len=0.93, thru=17.9, I=20.8, T=32.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:12,716] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:05:20,967] loop.py:45 INFO TRAIN_STEP step=712 loss=2.875093 Training: 4%|▍ | 712/18018 [1:35:07<35:11:48, 7.32s/step, loss=2.875, acc=0.474, acc_len=1.05, thru=11.6, I=14.5, T=33.0, wait=4.5s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:27,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:05:29,771] loop.py:45 INFO TRAIN_STEP step=713 loss=3.005322 Training: 4%|▍ | 713/18018 [1:35:15<37:20:14, 7.77s/step, loss=3.005, acc=0.468, acc_len=1.03, thru=17.4, I=17.2, T=27.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:29,907] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:35,767] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:05:37,209] loop.py:45 INFO TRAIN_STEP step=714 loss=2.886003 Training: 4%|▍ | 714/18018 [1:35:23<36:51:33, 7.67s/step, loss=2.886, acc=0.476, acc_len=1.05, thru=13.7, I=22.8, T=31.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:37,280] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:44,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:05:45,390] loop.py:380 INFO TIMING step=715: step=4.376s data=1.935s compute=4.342s [fwd=3.157s bwd=1.173s opt=0.003s] dispatch=3.793s [2026-10-18 15:05:45,390] loop.py:45 INFO TRAIN_STEP step=715 loss=2.922626 Training: 4%|▍ | 715/18018 [1:35:31<37:35:31, 7.82s/step, loss=2.923, acc=0.464, acc_len=1.02, thru=14.7, I=19.1, T=29.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:45,851] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:51,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:05:51,884] loop.py:45 INFO TRAIN_STEP step=716 loss=2.883877 Training: 4%|▍ | 716/18018 [1:35:37<35:40:42, 7.42s/step, loss=2.884, acc=0.479, acc_len=1.06, thru=18.4, I=19.9, T=35.5, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:05:51,879] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=715: forward=2535.1ms backward=1020.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:52,158] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:05:59,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:05:59,816] loop.py:45 INFO TRAIN_STEP step=717 loss=3.204177 Training: 4%|▍ | 717/18018 [1:35:45<36:24:42, 7.58s/step, loss=3.204, acc=0.436, acc_len=0.95, thru=17.1, I=18.5, T=28.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:00,014] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:04,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:06:06,212] loop.py:45 INFO TRAIN_STEP step=718 loss=3.065145 Training: 4%|▍ | 718/18018 [1:35:52<34:42:15, 7.22s/step, loss=3.065, acc=0.455, acc_len=0.99, thru=18.0, I=19.4, T=33.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:06,485] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:13,705] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:06:13,890] loop.py:45 INFO TRAIN_STEP step=719 loss=2.851634 Training: 4%|▍ | 719/18018 [1:35:59<35:21:37, 7.36s/step, loss=2.852, acc=0.475, acc_len=1.05, thru=15.0, I=17.3, T=32.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:14,208] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:19,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:06:21,466] loop.py:380 INFO TIMING step=720: step=4.341s data=2.013s compute=4.312s [fwd=3.013s bwd=1.286s opt=0.003s] dispatch=3.227s [2026-10-18 15:06:21,466] loop.py:45 INFO TRAIN_STEP step=720 loss=2.868550 Training: 4%|▍ | 720/18018 [1:36:07<35:40:25, 7.42s/step, loss=2.869, acc=0.471, acc_len=1.05, thru=16.6, I=19.7, T=29.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:21,822] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:31,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:06:31,195] loop.py:45 INFO TRAIN_STEP step=721 loss=2.765199 Training: 4%|▍ | 721/18018 [1:36:17<38:59:33, 8.12s/step, loss=2.765, acc=0.487, acc_len=1.09, thru=13.7, I=12.4, T=29.8, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:06:31,190] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=720: forward=3046.8ms backward=1208.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:31,566] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:06:38,315] loop.py:45 INFO TRAIN_STEP step=722 loss=3.190232 Training: 4%|▍ | 722/18018 [1:36:24<37:33:12, 7.82s/step, loss=3.190, acc=0.445, acc_len=0.96, thru=17.4, I=15.1, T=31.3, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:46,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:06:46,201] loop.py:45 INFO TRAIN_STEP step=723 loss=3.100654 Training: 4%|▍ | 723/18018 [1:36:32<37:39:10, 7.84s/step, loss=3.101, acc=0.448, acc_len=0.97, thru=18.7, I=16.4, T=27.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:46,568] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:52,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:06:53,549] loop.py:45 INFO TRAIN_STEP step=724 loss=3.004768 Training: 4%|▍ | 724/18018 [1:36:39<36:56:39, 7.69s/step, loss=3.005, acc=0.458, acc_len=1.00, thru=18.6, I=18.7, T=30.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:06:54,047] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:00,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:07:02,102] loop.py:380 INFO TIMING step=725: step=4.534s data=2.205s compute=4.507s [fwd=3.150s bwd=1.345s opt=0.003s] dispatch=4.011s [2026-10-18 15:07:02,102] loop.py:45 INFO TRAIN_STEP step=725 loss=3.016098 Training: 4%|▍ | 725/18018 [1:36:48<38:11:08, 7.95s/step, loss=3.016, acc=0.456, acc_len=1.00, thru=14.1, I=16.6, T=28.2, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:02,212] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:09,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:07:10,817] loop.py:45 INFO TRAIN_STEP step=726 loss=2.898911 Training: 4%|▍ | 726/18018 [1:36:56<39:17:24, 8.18s/step, loss=2.899, acc=0.475, acc_len=1.05, thru=14.9, I=15.0, T=33.2, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:07:10,813] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=725: forward=2768.6ms backward=1040.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:10,881] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:17,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:07:18,468] loop.py:45 INFO TRAIN_STEP step=727 loss=2.997416 Training: 4%|▍ | 727/18018 [1:37:04<38:31:30, 8.02s/step, loss=2.997, acc=0.455, acc_len=0.99, thru=16.5, I=18.5, T=29.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:18,808] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:24,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:07:26,987] loop.py:45 INFO TRAIN_STEP step=728 loss=2.903677 Training: 4%|▍ | 728/18018 [1:37:13<39:14:19, 8.17s/step, loss=2.904, acc=0.468, acc_len=1.03, thru=15.2, I=21.5, T=23.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:27,161] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 15:07:34,620] loop.py:45 INFO TRAIN_STEP step=729 loss=3.036659 Training: 4%|▍ | 729/18018 [1:37:20<38:27:40, 8.01s/step, loss=3.037, acc=0.456, acc_len=0.99, thru=15.0, I=15.9, T=32.6, wait=3.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:43,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:07:45,004] loop.py:380 INFO TIMING step=730: step=4.351s data=1.755s compute=4.302s [fwd=2.671s bwd=1.617s opt=0.003s] dispatch=6.025s [2026-10-18 15:07:45,004] loop.py:45 INFO TRAIN_STEP step=730 loss=3.045001 Training: 4%|▍ | 730/18018 [1:37:31<41:53:02, 8.72s/step, loss=3.045, acc=0.458, acc_len=1.00, thru=14.4, I=15.5, T=29.4, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:07:45,248] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:07:52,069] loop.py:45 INFO TRAIN_STEP step=731 loss=2.922311 Training: 4%|▍ | 731/18018 [1:37:38<39:29:32, 8.22s/step, loss=2.922, acc=0.469, acc_len=1.03, thru=14.1, I=19.6, T=31.0, wait=2.9s, pool=28, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:07:52,065] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=730: forward=2910.9ms backward=1172.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:02,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:08:02,812] loop.py:45 INFO TRAIN_STEP step=732 loss=3.268710 Training: 4%|▍ | 732/18018 [1:37:48<43:07:05, 8.98s/step, loss=3.269, acc=0.427, acc_len=0.91, thru=15.8, I=15.3, T=27.8, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:03,160] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:08,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:08:10,888] loop.py:45 INFO TRAIN_STEP step=733 loss=2.833253 Training: 4%|▍ | 733/18018 [1:37:56<41:49:03, 8.71s/step, loss=2.833, acc=0.479, acc_len=1.06, thru=15.3, I=17.0, T=28.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:10,960] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 [2026-10-18 15:08:16,741] loop.py:45 INFO TRAIN_STEP step=734 loss=2.962854 Training: 4%|▍ | 734/18018 [1:38:02<37:42:07, 7.85s/step, loss=2.963, acc=0.467, acc_len=1.03, thru=14.5, I=23.9, T=34.8, wait=2.2s, pool=20, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:25,745] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:08:26,132] loop.py:380 INFO TIMING step=735: step=3.769s data=1.547s compute=3.736s [fwd=2.672s bwd=1.049s opt=0.003s] dispatch=5.610s [2026-10-18 15:08:26,133] loop.py:45 INFO TRAIN_STEP step=735 loss=3.027498 Training: 4%|▍ | 735/18018 [1:38:12<39:54:51, 8.31s/step, loss=3.027, acc=0.455, acc_len=0.99, thru=19.8, I=16.0, T=34.0, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:26,251] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:32,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:08:33,310] loop.py:45 INFO TRAIN_STEP step=736 loss=2.903221 Training: 4%|▍ | 736/18018 [1:38:19<38:16:26, 7.97s/step, loss=2.903, acc=0.467, acc_len=1.03, thru=16.8, I=18.4, T=30.4, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:08:33,306] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=735: forward=3016.4ms backward=1147.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:33,807] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:39,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:08:39,823] loop.py:45 INFO TRAIN_STEP step=737 loss=2.962892 Training: 4%|▍ | 737/18018 [1:38:25<36:10:06, 7.53s/step, loss=2.963, acc=0.465, acc_len=1.02, thru=17.7, I=20.6, T=35.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:40,042] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:48,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:08:50,480] loop.py:45 INFO TRAIN_STEP step=738 loss=2.841967 Training: 4%|▍ | 738/18018 [1:38:36<40:39:52, 8.47s/step, loss=2.842, acc=0.481, acc_len=1.07, thru=12.5, I=14.4, T=31.0, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:50,645] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:56,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:08:56,967] loop.py:45 INFO TRAIN_STEP step=739 loss=2.906250 Training: 4%|▍ | 739/18018 [1:38:43<37:48:11, 7.88s/step, loss=2.906, acc=0.471, acc_len=1.04, thru=17.8, I=22.7, T=37.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:08:57,093] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-18 15:09:04,642] loop.py:380 INFO TIMING step=740: step=4.106s data=1.719s compute=4.050s [fwd=2.729s bwd=1.306s opt=0.003s] dispatch=3.559s [2026-10-18 15:09:04,642] loop.py:45 INFO TRAIN_STEP step=740 loss=3.183993 Training: 4%|▍ | 740/18018 [1:38:50<37:30:35, 7.82s/step, loss=3.184, acc=0.446, acc_len=0.96, thru=17.2, I=18.6, T=31.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:04,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:05,110] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:09:13,855] loop.py:45 INFO TRAIN_STEP step=741 loss=2.922709 Training: 4%|▍ | 741/18018 [1:38:59<39:31:29, 8.24s/step, loss=2.923, acc=0.469, acc_len=1.03, thru=13.8, I=14.2, T=28.1, wait=4.7s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:09:13,851] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=740: forward=3194.8ms backward=1301.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:19,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:09:20,794] loop.py:45 INFO TRAIN_STEP step=742 loss=2.918141 Training: 4%|▍ | 742/18018 [1:39:06<37:39:13, 7.85s/step, loss=2.918, acc=0.470, acc_len=1.03, thru=17.7, I=18.7, T=32.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:21,113] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:26,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:09:27,471] loop.py:45 INFO TRAIN_STEP step=743 loss=2.718935 Training: 4%|▍ | 743/18018 [1:39:13<35:57:53, 7.49s/step, loss=2.719, acc=0.499, acc_len=1.11, thru=17.2, I=21.8, T=32.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:27,612] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-18 15:09:34,283] loop.py:45 INFO TRAIN_STEP step=744 loss=2.915861 Training: 4%|▍ | 744/18018 [1:39:20<34:59:13, 7.29s/step, loss=2.916, acc=0.473, acc_len=1.04, thru=16.8, I=21.8, T=32.3, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:41,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:09:41,885] loop.py:380 INFO TIMING step=745: step=4.234s data=1.892s compute=4.196s [fwd=2.900s bwd=1.282s opt=0.003s] dispatch=3.355s [2026-10-18 15:09:41,885] loop.py:45 INFO TRAIN_STEP step=745 loss=2.979909 Training: 4%|▍ | 745/18018 [1:39:27<35:25:29, 7.38s/step, loss=2.980, acc=0.459, acc_len=1.01, thru=15.7, I=17.4, T=30.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:42,337] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:47,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:09:48,527] loop.py:45 INFO TRAIN_STEP step=746 loss=2.883421 Training: 4%|▍ | 746/18018 [1:39:34<34:21:29, 7.16s/step, loss=2.883, acc=0.472, acc_len=1.04, thru=17.0, I=20.0, T=31.6, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:09:48,522] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=745: forward=2879.8ms backward=1122.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:48,723] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:56,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:09:57,944] loop.py:45 INFO TRAIN_STEP step=747 loss=3.130221 Training: 4%|▍ | 747/18018 [1:39:44<37:36:03, 7.84s/step, loss=3.130, acc=0.447, acc_len=0.96, thru=13.9, I=15.3, T=29.9, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:09:58,100] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:02,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:10:04,736] loop.py:45 INFO TRAIN_STEP step=748 loss=3.011286 Training: 4%|▍ | 748/18018 [1:39:50<36:05:35, 7.52s/step, loss=3.011, acc=0.462, acc_len=1.01, thru=18.1, I=23.9, T=26.9, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:04,974] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:10,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:10:12,194] loop.py:45 INFO TRAIN_STEP step=749 loss=3.103020 Training: 4%|▍ | 749/18018 [1:39:58<36:00:02, 7.50s/step, loss=3.103, acc=0.449, acc_len=0.98, thru=14.8, I=22.8, T=30.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:12,646] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:10:19,121] loop.py:380 INFO TIMING step=750: step=3.794s data=1.671s compute=3.757s [fwd=2.648s bwd=1.094s opt=0.003s] dispatch=3.122s [2026-10-18 15:10:19,122] loop.py:45 INFO TRAIN_STEP step=750 loss=2.962365 Training: 4%|▍ | 750/18018 [1:40:05<35:09:49, 7.33s/step, loss=2.962, acc=0.467, acc_len=1.03, thru=15.3, I=18.8, T=33.7, wait=3.1s, pool=52, epoch=1/2][2026-10-18 15:10:19,125] loop.py:419 INFO Saving checkpoint at step 750... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:19,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 15:10:22,171] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0000751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:22,489] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:30,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 15:10:22,169] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0000751 [2026-10-18 15:10:31,051] loop.py:45 INFO TRAIN_STEP step=751 loss=2.779434 Training: 4%|▍ | 751/18018 [1:40:17<41:46:55, 8.71s/step, loss=2.779, acc=0.488, acc_len=1.08, thru=14.9, I=16.0, T=29.0, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:10:31,046] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=750: forward=3101.3ms backward=1197.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:31,386] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:39,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:10:40,075] loop.py:45 INFO TRAIN_STEP step=752 loss=3.110204 Training: 4%|▍ | 752/18018 [1:40:26<42:13:35, 8.80s/step, loss=3.110, acc=0.445, acc_len=0.97, thru=13.8, I=15.4, T=31.3, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:40,513] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:45,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:10:46,625] loop.py:45 INFO TRAIN_STEP step=753 loss=3.137665 Training: 4%|▍ | 753/18018 [1:40:32<38:59:01, 8.13s/step, loss=3.138, acc=0.448, acc_len=0.97, thru=18.5, I=21.2, T=36.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:46,807] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:51,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:10:53,118] loop.py:45 INFO TRAIN_STEP step=754 loss=2.996383 Training: 4%|▍ | 754/18018 [1:40:39<36:37:36, 7.64s/step, loss=2.996, acc=0.460, acc_len=1.00, thru=18.4, I=25.4, T=33.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:10:53,501] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:10:59,987] loop.py:380 INFO TIMING step=755: step=3.756s data=1.578s compute=3.730s [fwd=2.579s bwd=1.136s opt=0.003s] dispatch=3.104s [2026-10-18 15:10:59,987] loop.py:45 INFO TRAIN_STEP step=755 loss=2.696223 Training: 4%|▍ | 755/18018 [1:40:46<35:31:04, 7.41s/step, loss=2.696, acc=0.494, acc_len=1.11, thru=14.3, I=19.4, T=34.1, wait=3.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:06,845] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:11:08,031] loop.py:45 INFO TRAIN_STEP step=756 loss=3.064713 Training: 4%|▍ | 756/18018 [1:40:54<36:26:07, 7.60s/step, loss=3.065, acc=0.453, acc_len=0.99, thru=17.3, I=18.7, T=33.4, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:11:08,027] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=755: forward=2695.8ms backward=1082.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:08,357] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:13,461] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:11:15,331] loop.py:45 INFO TRAIN_STEP step=757 loss=3.127634 Training: 4%|▍ | 757/18018 [1:41:01<36:00:04, 7.51s/step, loss=3.128, acc=0.445, acc_len=0.96, thru=16.9, I=24.2, T=29.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:15,484] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:11:21,661] loop.py:45 INFO TRAIN_STEP step=758 loss=2.927148 Training: 4%|▍ | 758/18018 [1:41:07<34:18:19, 7.16s/step, loss=2.927, acc=0.467, acc_len=1.03, thru=16.9, I=21.8, T=34.9, wait=2.7s, pool=56, epoch=1/2][2026-10-18 15:11:29,007] loop.py:45 INFO TRAIN_STEP step=759 loss=2.766377 Training: 4%|▍ | 759/18018 [1:41:15<34:34:31, 7.21s/step, loss=2.766, acc=0.483, acc_len=1.07, thru=16.2, I=17.3, T=31.9, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:36,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:11:36,910] loop.py:380 INFO TIMING step=760: step=3.984s data=1.739s compute=3.948s [fwd=2.786s bwd=1.148s opt=0.003s] dispatch=3.911s [2026-10-18 15:11:36,911] loop.py:45 INFO TRAIN_STEP step=760 loss=2.861416 Training: 4%|▍ | 760/18018 [1:41:22<35:34:10, 7.42s/step, loss=2.861, acc=0.474, acc_len=1.04, thru=16.4, I=16.7, T=32.1, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:37,387] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:42,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:11:42,991] loop.py:45 INFO TRAIN_STEP step=761 loss=2.970181 Training: 4%|▍ | 761/18018 [1:41:29<33:38:22, 7.02s/step, loss=2.970, acc=0.465, acc_len=1.02, thru=19.1, I=21.1, T=35.1, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:11:42,987] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=760: forward=2548.0ms backward=1059.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:43,337] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:49,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:11:49,466] loop.py:45 INFO TRAIN_STEP step=762 loss=2.993695 Training: 4%|▍ | 762/18018 [1:41:35<32:51:35, 6.86s/step, loss=2.994, acc=0.466, acc_len=1.02, thru=18.8, I=20.2, T=33.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:49,602] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:54,567] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:11:55,573] loop.py:45 INFO TRAIN_STEP step=763 loss=2.845878 Training: 4%|▍ | 763/18018 [1:41:41<31:46:49, 6.63s/step, loss=2.846, acc=0.470, acc_len=1.03, thru=19.6, I=22.6, T=33.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:11:56,075] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:00,165] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:12:01,888] loop.py:45 INFO TRAIN_STEP step=764 loss=2.871264 Training: 4%|▍ | 764/18018 [1:41:47<31:19:25, 6.54s/step, loss=2.871, acc=0.472, acc_len=1.04, thru=19.5, I=24.9, T=33.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:02,177] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:07,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:12:08,787] loop.py:380 INFO TIMING step=765: step=4.383s data=2.084s compute=4.355s [fwd=3.085s bwd=1.255s opt=0.003s] dispatch=2.507s [2026-10-18 15:12:08,787] loop.py:45 INFO TRAIN_STEP step=765 loss=2.909243 Training: 4%|▍ | 765/18018 [1:41:54<31:50:52, 6.65s/step, loss=2.909, acc=0.467, acc_len=1.03, thru=17.2, I=24.8, T=29.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:08,872] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 15:12:14,698] loop.py:45 INFO TRAIN_STEP step=766 loss=2.985447 Training: 4%|▍ | 766/18018 [1:42:00<30:47:13, 6.42s/step, loss=2.985, acc=0.464, acc_len=1.01, thru=18.4, I=22.8, T=31.8, wait=1.9s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:12:14,693] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=765: forward=2817.3ms backward=1157.9ms [2026-10-18 15:12:20,669] loop.py:45 INFO TRAIN_STEP step=767 loss=2.881294 Training: 4%|▍ | 767/18018 [1:42:06<30:08:13, 6.29s/step, loss=2.881, acc=0.479, acc_len=1.06, thru=15.6, I=22.1, T=35.7, wait=2.4s, pool=24, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:31,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:12:32,150] loop.py:45 INFO TRAIN_STEP step=768 loss=3.175641 Training: 4%|▍ | 768/18018 [1:42:18<37:35:55, 7.85s/step, loss=3.176, acc=0.440, acc_len=0.95, thru=15.7, I=14.3, T=31.7, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:32,247] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:40,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:12:41,626] loop.py:45 INFO TRAIN_STEP step=769 loss=3.023322 Training: 4%|▍ | 769/18018 [1:42:27<39:56:20, 8.34s/step, loss=3.023, acc=0.464, acc_len=1.01, thru=13.7, I=15.1, T=29.5, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:41,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:49,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:12:51,416] loop.py:380 INFO TIMING step=770: step=5.014s data=2.412s compute=4.978s [fwd=3.536s bwd=1.428s opt=0.003s] dispatch=4.764s [2026-10-18 15:12:51,416] loop.py:45 INFO TRAIN_STEP step=770 loss=2.929360 Training: 4%|▍ | 770/18018 [1:42:37<42:01:21, 8.77s/step, loss=2.929, acc=0.474, acc_len=1.04, thru=13.4, I=15.3, T=25.5, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:51,852] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:58,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:12:59,353] loop.py:45 INFO TRAIN_STEP step=771 loss=2.908137 Training: 4%|▍ | 771/18018 [1:42:45<40:49:32, 8.52s/step, loss=2.908, acc=0.475, acc_len=1.04, thru=17.1, I=20.7, T=34.7, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:12:59,349] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=770: forward=2603.2ms backward=1040.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:12:59,548] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:04,926] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:13:06,203] loop.py:45 INFO TRAIN_STEP step=772 loss=2.790315 Training: 4%|▍ | 772/18018 [1:42:52<38:25:13, 8.02s/step, loss=2.790, acc=0.483, acc_len=1.07, thru=18.4, I=23.9, T=32.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:06,435] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:11,903] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:13:12,357] loop.py:45 INFO TRAIN_STEP step=773 loss=2.640664 Training: 4%|▍ | 773/18018 [1:42:58<35:44:02, 7.46s/step, loss=2.641, acc=0.501, acc_len=1.12, thru=19.7, I=24.4, T=33.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:12,407] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:18,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:13:20,084] loop.py:45 INFO TRAIN_STEP step=774 loss=3.049882 Training: 4%|▍ | 774/18018 [1:43:06<36:06:58, 7.54s/step, loss=3.050, acc=0.459, acc_len=1.00, thru=16.0, I=19.1, T=31.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:20,207] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 15:13:26,543] loop.py:380 INFO TIMING step=775: step=4.010s data=1.784s compute=3.977s [fwd=2.834s bwd=1.128s opt=0.003s] dispatch=2.441s [2026-10-18 15:13:26,543] loop.py:45 INFO TRAIN_STEP step=775 loss=2.808378 Training: 4%|▍ | 775/18018 [1:43:12<34:33:33, 7.22s/step, loss=2.808, acc=0.477, acc_len=1.06, thru=17.8, I=22.8, T=31.9, wait=2.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:33,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:13:34,615] loop.py:45 INFO TRAIN_STEP step=776 loss=2.908622 Training: 4%|▍ | 776/18018 [1:43:20<35:47:16, 7.47s/step, loss=2.909, acc=0.466, acc_len=1.03, thru=16.0, I=17.4, T=32.2, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:13:34,611] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=775: forward=2762.6ms backward=1170.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:35,116] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:13:41,589] loop.py:45 INFO TRAIN_STEP step=777 loss=2.918446 Training: 4%|▍ | 777/18018 [1:43:27<35:04:14, 7.32s/step, loss=2.918, acc=0.472, acc_len=1.03, thru=15.8, I=19.5, T=33.2, wait=3.1s, pool=56, epoch=1/2][2026-10-18 15:13:48,899] loop.py:45 INFO TRAIN_STEP step=778 loss=2.842307 Training: 4%|▍ | 778/18018 [1:43:34<35:03:08, 7.32s/step, loss=2.842, acc=0.480, acc_len=1.07, thru=18.3, I=18.2, T=32.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:48,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:49,319] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:55,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:13:56,888] loop.py:45 INFO TRAIN_STEP step=779 loss=2.872111 Training: 4%|▍ | 779/18018 [1:43:42<36:00:41, 7.52s/step, loss=2.872, acc=0.477, acc_len=1.05, thru=15.6, I=17.4, T=30.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:13:56,969] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:04,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:14:04,693] loop.py:380 INFO TIMING step=780: step=4.231s data=1.964s compute=4.196s [fwd=2.962s bwd=1.219s opt=0.003s] dispatch=3.564s [2026-10-18 15:14:04,693] loop.py:45 INFO TRAIN_STEP step=780 loss=2.929118 Training: 4%|▍ | 780/18018 [1:43:50<36:25:12, 7.61s/step, loss=2.929, acc=0.468, acc_len=1.03, thru=15.8, I=18.2, T=30.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:04,731] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:09,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:14:10,566] loop.py:45 INFO TRAIN_STEP step=781 loss=2.760705 Training: 4%|▍ | 781/18018 [1:43:56<33:55:43, 7.09s/step, loss=2.761, acc=0.491, acc_len=1.09, thru=19.4, I=22.1, T=36.2, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:14:10,562] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=780: forward=2493.7ms backward=1003.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:10,599] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:15,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:14:17,033] loop.py:45 INFO TRAIN_STEP step=782 loss=2.723772 Training: 4%|▍ | 782/18018 [1:44:03<33:02:06, 6.90s/step, loss=2.724, acc=0.499, acc_len=1.12, thru=20.0, I=27.2, T=31.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:17,420] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:14:24,677] loop.py:45 INFO TRAIN_STEP step=783 loss=2.635108 Training: 4%|▍ | 783/18018 [1:44:10<34:06:08, 7.12s/step, loss=2.635, acc=0.507, acc_len=1.14, thru=14.5, I=19.8, T=32.5, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:32,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:14:34,299] loop.py:45 INFO TRAIN_STEP step=784 loss=2.688967 Training: 4%|▍ | 784/18018 [1:44:20<37:41:21, 7.87s/step, loss=2.689, acc=0.489, acc_len=1.09, thru=14.7, I=16.8, T=28.3, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:34,420] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:41,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:14:41,360] loop.py:380 INFO TIMING step=785: step=3.738s data=1.616s compute=3.702s [fwd=2.660s bwd=1.027s opt=0.003s] dispatch=3.313s [2026-10-18 15:14:41,360] loop.py:45 INFO TRAIN_STEP step=785 loss=2.812190 Training: 4%|▍ | 785/18018 [1:44:27<36:31:19, 7.63s/step, loss=2.812, acc=0.487, acc_len=1.07, thru=16.0, I=20.4, T=34.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:41,518] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:47,983] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:14:48,305] loop.py:45 INFO TRAIN_STEP step=786 loss=2.994730 Training: 4%|▍ | 786/18018 [1:44:34<35:32:12, 7.42s/step, loss=2.995, acc=0.462, acc_len=1.01, thru=18.3, I=19.5, T=29.2, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:14:48,299] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=785: forward=3138.4ms backward=1198.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:48,498] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:54,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:14:55,717] loop.py:45 INFO TRAIN_STEP step=787 loss=2.944741 Training: 4%|▍ | 787/18018 [1:44:41<35:30:56, 7.42s/step, loss=2.945, acc=0.466, acc_len=1.02, thru=16.5, I=18.8, T=30.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:14:56,088] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:15:02,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:15:03,956] loop.py:45 INFO TRAIN_STEP step=788 loss=2.991309 Training: 4%|▍ | 788/18018 [1:44:50<36:41:20, 7.67s/step, loss=2.991, acc=0.464, acc_len=1.01, thru=15.0, I=19.9, T=28.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:15:04,024] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 [2026-10-18 15:15:11,305] loop.py:45 INFO TRAIN_STEP step=789 loss=2.822205 Training: 4%|▍ | 789/18018 [1:44:57<36:13:57, 7.57s/step, loss=2.822, acc=0.485, acc_len=1.08, thru=13.9, I=15.4, T=32.3, wait=3.4s, pool=56, epoch=1/2][2026-10-18 15:15:18,259] loop.py:380 INFO TIMING step=790: step=4.050s data=1.775s compute=4.018s [fwd=2.792s bwd=1.212s opt=0.003s] dispatch=2.896s [2026-10-18 15:15:18,259] loop.py:45 INFO TRAIN_STEP step=790 loss=2.709960 Training: 4%|▍ | 790/18018 [1:45:04<35:20:33, 7.39s/step, loss=2.710, acc=0.495, acc_len=1.10, thru=18.8, I=15.4, T=31.6, wait=2.9s, pool=48, epoch=1/2][2026-10-18 15:15:26,226] loop.py:45 INFO TRAIN_STEP step=791 loss=3.086889 Training: 4%|▍ | 791/18018 [1:45:12<36:11:05, 7.56s/step, loss=3.087, acc=0.453, acc_len=0.98, thru=16.4, I=14.8, T=29.9, wait=3.7s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:15:26,222] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=790: forward=3046.0ms backward=1189.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:15:32,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:15:34,212] loop.py:45 INFO TRAIN_STEP step=792 loss=2.895506 Training: 4%|▍ | 792/18018 [1:45:20<36:47:13, 7.69s/step, loss=2.896, acc=0.477, acc_len=1.04, thru=16.8, I=19.5, T=29.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:15:34,316] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:15:40,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:15:41,794] loop.py:45 INFO TRAIN_STEP step=793 loss=2.967165 Training: 4%|▍ | 793/18018 [1:45:27<36:37:56, 7.66s/step, loss=2.967, acc=0.464, acc_len=1.01, thru=14.1, I=19.7, T=27.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:15:42,133] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:15:47,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:15:47,945] loop.py:45 INFO TRAIN_STEP step=794 loss=2.911495 Training: 4%|▍ | 794/18018 [1:45:34<34:28:20, 7.21s/step, loss=2.911, acc=0.468, acc_len=1.02, thru=17.9, I=21.6, T=34.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:15:48,003] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 15:15:54,795] loop.py:380 INFO TIMING step=795: step=3.861s data=1.654s compute=3.828s [fwd=2.700s bwd=1.114s opt=0.003s] dispatch=2.978s [2026-10-18 15:15:54,796] loop.py:45 INFO TRAIN_STEP step=795 loss=2.826298 Training: 4%|▍ | 795/18018 [1:45:40<33:57:41, 7.10s/step, loss=2.826, acc=0.478, acc_len=1.05, thru=18.4, I=19.5, T=33.2, wait=3.0s, pool=60, epoch=1/2][2026-10-18 15:16:04,426] loop.py:45 INFO TRAIN_STEP step=796 loss=2.716139 Training: 4%|▍ | 796/18018 [1:45:50<37:35:18, 7.86s/step, loss=2.716, acc=0.497, acc_len=1.10, thru=13.7, I=12.7, T=29.4, wait=5.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:16:04,423] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=795: forward=3151.9ms backward=1156.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:16:10,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:16:11,027] loop.py:45 INFO TRAIN_STEP step=797 loss=2.913840 Training: 4%|▍ | 797/18018 [1:45:57<35:46:58, 7.48s/step, loss=2.914, acc=0.469, acc_len=1.02, thru=17.5, I=18.6, T=32.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:16:11,456] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:16:17,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:16:18,006] loop.py:45 INFO TRAIN_STEP step=798 loss=3.045294 Training: 4%|▍ | 798/18018 [1:46:04<35:03:44, 7.33s/step, loss=3.045, acc=0.446, acc_len=0.97, thru=17.8, I=20.6, T=31.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:16:18,106] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 15:16:26,865] loop.py:45 INFO TRAIN_STEP step=799 loss=2.726180 Training: 4%|▍ | 799/18018 [1:46:12<37:15:11, 7.79s/step, loss=2.726, acc=0.502, acc_len=1.12, thru=13.1, I=14.3, T=29.9, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:16:33,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:16:34,174] loop.py:380 INFO TIMING step=800: step=3.869s data=1.749s compute=3.835s [fwd=2.739s bwd=1.080s opt=0.003s] dispatch=3.431s [2026-10-18 15:16:34,174] loop.py:45 INFO TRAIN_STEP step=800 loss=2.798201 Training: 4%|▍ | 800/18018 [1:46:20<36:33:48, 7.64s/step, loss=2.798, acc=0.480, acc_len=1.06, thru=17.5, I=16.9, T=33.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:16:34,625] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:16:41,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:16:42,044] loop.py:45 INFO TRAIN_STEP step=801 loss=2.926593 Training: 4%|▍ | 801/18018 [1:46:28<36:53:06, 7.71s/step, loss=2.927, acc=0.469, acc_len=1.03, thru=17.0, I=16.9, T=32.1, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:16:42,040] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=800: forward=2755.1ms backward=1185.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:16:42,374] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:16:49,721] loop.py:45 INFO TRAIN_STEP step=802 loss=3.074896 Training: 4%|▍ | 802/18018 [1:46:35<36:50:09, 7.70s/step, loss=3.075, acc=0.443, acc_len=0.96, thru=16.1, I=15.4, T=30.1, wait=3.4s, pool=60, epoch=1/2][2026-10-18 15:16:57,710] loop.py:45 INFO TRAIN_STEP step=803 loss=2.893054 Training: 4%|▍ | 803/18018 [1:46:43<37:14:26, 7.79s/step, loss=2.893, acc=0.474, acc_len=1.05, thru=15.0, I=14.7, T=30.0, wait=3.7s, pool=48, epoch=1/2][2026-10-18 15:17:05,983] loop.py:45 INFO TRAIN_STEP step=804 loss=2.768508 Training: 4%|▍ | 804/18018 [1:46:52<37:56:32, 7.93s/step, loss=2.769, acc=0.490, acc_len=1.09, thru=14.1, I=15.5, T=28.2, wait=3.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:16,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:17:17,124] loop.py:380 INFO TIMING step=805: step=5.089s data=2.298s compute=5.052s [fwd=3.643s bwd=1.394s opt=0.003s] dispatch=6.038s [2026-10-18 15:17:17,124] loop.py:45 INFO TRAIN_STEP step=805 loss=2.870775 Training: 4%|▍ | 805/18018 [1:47:03<42:31:59, 8.90s/step, loss=2.871, acc=0.480, acc_len=1.05, thru=13.5, I=13.2, T=25.2, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:17,336] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:22,954] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:17:24,561] loop.py:45 INFO TRAIN_STEP step=806 loss=2.942428 Training: 4%|▍ | 806/18018 [1:47:10<40:26:21, 8.46s/step, loss=2.942, acc=0.465, acc_len=1.02, thru=17.0, I=16.9, T=26.6, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:17:24,556] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=805: forward=3126.0ms backward=1605.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:24,968] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:30,834] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:17:31,524] loop.py:45 INFO TRAIN_STEP step=807 loss=2.779890 Training: 4%|▍ | 807/18018 [1:47:17<38:17:33, 8.01s/step, loss=2.780, acc=0.487, acc_len=1.08, thru=16.1, I=21.3, T=32.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:31,838] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:17:38,635] loop.py:45 INFO TRAIN_STEP step=808 loss=2.847863 Training: 4%|▍ | 808/18018 [1:47:24<36:59:58, 7.74s/step, loss=2.848, acc=0.475, acc_len=1.04, thru=16.4, I=18.6, T=31.5, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:44,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:17:46,169] loop.py:45 INFO TRAIN_STEP step=809 loss=2.722367 Training: 4%|▍ | 809/18018 [1:47:32<36:42:12, 7.68s/step, loss=2.722, acc=0.493, acc_len=1.10, thru=17.8, I=17.7, T=30.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:46,205] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:51,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:17:53,268] loop.py:380 INFO TIMING step=810: step=4.185s data=1.898s compute=4.148s [fwd=2.944s bwd=1.188s opt=0.003s] dispatch=2.904s [2026-10-18 15:17:53,268] loop.py:45 INFO TRAIN_STEP step=810 loss=2.584658 Training: 4%|▍ | 810/18018 [1:47:39<35:52:18, 7.50s/step, loss=2.585, acc=0.499, acc_len=1.12, thru=15.9, I=21.4, T=30.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:53,489] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:17:58,705] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 15:18:00,069] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=810: forward=2959.2ms backward=1249.6ms [2026-10-18 15:18:00,073] loop.py:45 INFO TRAIN_STEP step=811 loss=2.919283 Training: 5%|▍ | 811/18018 [1:47:46<34:51:49, 7.29s/step, loss=2.919, acc=0.473, acc_len=1.03, thru=16.8, I=24.7, T=30.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:18:00,230] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:18:07,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:18:07,387] loop.py:45 INFO TRAIN_STEP step=812 loss=2.843852 Training: 5%|▍ | 812/18018 [1:47:53<34:53:30, 7.30s/step, loss=2.844, acc=0.482, acc_len=1.06, thru=15.2, I=17.9, T=30.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:18:07,758] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:18:15,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:18:16,106] loop.py:45 INFO TRAIN_STEP step=813 loss=2.821659 Training: 5%|▍ | 813/18018 [1:48:02<36:55:20, 7.73s/step, loss=2.822, acc=0.480, acc_len=1.06, thru=15.8, I=17.1, T=27.9, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:18:16,211] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:18:20,258] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:18:22,178] loop.py:45 INFO TRAIN_STEP step=814 loss=2.921811 Training: 5%|▍ | 814/18018 [1:48:08<34:33:06, 7.23s/step, loss=2.922, acc=0.464, acc_len=1.01, thru=18.6, I=22.5, T=33.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:18:22,282] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-18 15:18:28,956] loop.py:380 INFO TIMING step=815: step=4.212s data=1.972s compute=4.176s [fwd=2.935s bwd=1.226s opt=0.003s] dispatch=2.555s [2026-10-18 15:18:28,956] loop.py:45 INFO TRAIN_STEP step=815 loss=2.933668 Training: 5%|▍ | 815/18018 [1:48:15<33:54:09, 7.09s/step, loss=2.934, acc=0.462, acc_len=1.01, thru=16.4, I=24.7, T=30.4, wait=2.6s, pool=56, epoch=1/2][2026-10-18 15:18:39,978] loop.py:45 INFO TRAIN_STEP step=816 loss=3.074386 Training: 5%|▍ | 816/18018 [1:48:26<39:31:38, 8.27s/step, loss=3.074, acc=0.456, acc_len=0.98, thru=11.5, I=12.4, T=25.3, wait=5.9s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:18:39,973] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=815: forward=3506.5ms backward=1513.4ms [2026-10-18 15:18:49,068] loop.py:45 INFO TRAIN_STEP step=817 loss=3.091032 Training: 5%|▍ | 817/18018 [1:48:35<40:42:02, 8.52s/step, loss=3.091, acc=0.455, acc_len=0.99, thru=14.2, I=13.9, T=31.7, wait=5.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:18:55,098] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:18:55,630] loop.py:45 INFO TRAIN_STEP step=818 loss=2.663884 Training: 5%|▍ | 818/18018 [1:48:41<37:53:21, 7.93s/step, loss=2.664, acc=0.502, acc_len=1.12, thru=18.9, I=19.0, T=34.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:18:56,105] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:04,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:19:04,422] loop.py:45 INFO TRAIN_STEP step=819 loss=2.925810 Training: 5%|▍ | 819/18018 [1:48:50<39:07:25, 8.19s/step, loss=2.926, acc=0.475, acc_len=1.03, thru=14.8, I=17.5, T=32.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:04,522] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:10,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:19:11,644] loop.py:380 INFO TIMING step=820: step=4.386s data=2.055s compute=4.359s [fwd=3.076s bwd=1.268s opt=0.003s] dispatch=2.827s [2026-10-18 15:19:11,645] loop.py:45 INFO TRAIN_STEP step=820 loss=2.592057 Training: 5%|▍ | 820/18018 [1:48:57<37:44:09, 7.90s/step, loss=2.592, acc=0.510, acc_len=1.14, thru=18.6, I=20.7, T=29.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:12,019] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:19:20,975] loop.py:45 INFO TRAIN_STEP step=821 loss=2.975293 Training: 5%|▍ | 821/18018 [1:49:07<39:47:01, 8.33s/step, loss=2.975, acc=0.466, acc_len=1.02, thru=13.4, I=12.2, T=26.9, wait=4.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:19:20,971] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=820: forward=3279.3ms backward=1413.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:28,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:19:30,491] loop.py:45 INFO TRAIN_STEP step=822 loss=2.950653 Training: 5%|▍ | 822/18018 [1:49:16<41:29:06, 8.68s/step, loss=2.951, acc=0.459, acc_len=1.01, thru=14.8, I=16.6, T=28.8, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:30,659] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:35,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:19:36,668] loop.py:45 INFO TRAIN_STEP step=823 loss=2.687013 Training: 5%|▍ | 823/18018 [1:49:22<37:53:26, 7.93s/step, loss=2.687, acc=0.493, acc_len=1.09, thru=17.7, I=23.8, T=34.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:36,912] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:44,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:19:46,622] loop.py:45 INFO TRAIN_STEP step=824 loss=3.048412 Training: 5%|▍ | 824/18018 [1:49:32<40:46:50, 8.54s/step, loss=3.048, acc=0.457, acc_len=0.99, thru=13.2, I=15.2, T=29.1, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:46,917] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:53,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 15:19:55,512] loop.py:380 INFO TIMING step=825: step=5.179s data=2.643s compute=5.145s [fwd=3.571s bwd=1.558s opt=0.003s] dispatch=3.704s [2026-10-18 15:19:55,512] loop.py:45 INFO TRAIN_STEP step=825 loss=2.885172 Training: 5%|▍ | 825/18018 [1:49:41<41:17:04, 8.64s/step, loss=2.885, acc=0.472, acc_len=1.03, thru=15.4, I=18.1, T=24.7, wait=3.7s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:19:55,839] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:03,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:20:03,691] loop.py:45 INFO TRAIN_STEP step=826 loss=2.885523 Training: 5%|▍ | 826/18018 [1:49:49<40:36:51, 8.50s/step, loss=2.886, acc=0.478, acc_len=1.04, thru=13.4, I=17.1, T=33.2, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:20:03,687] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=825: forward=2668.0ms backward=1124.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:03,738] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 15:20:10,065] loop.py:45 INFO TRAIN_STEP step=827 loss=2.794501 Training: 5%|▍ | 827/18018 [1:49:56<37:33:29, 7.87s/step, loss=2.795, acc=0.483, acc_len=1.07, thru=19.0, I=16.9, T=32.5, wait=2.4s, pool=56, epoch=1/2][2026-10-18 15:20:17,218] loop.py:45 INFO TRAIN_STEP step=828 loss=3.030638 Training: 5%|▍ | 828/18018 [1:50:03<36:32:07, 7.65s/step, loss=3.031, acc=0.454, acc_len=0.99, thru=15.7, I=15.6, T=33.7, wait=3.4s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:24,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:20:26,828] loop.py:45 INFO TRAIN_STEP step=829 loss=2.878570 Training: 5%|▍ | 829/18018 [1:50:12<39:20:23, 8.24s/step, loss=2.879, acc=0.478, acc_len=1.06, thru=16.6, I=17.2, T=29.2, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:27,292] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:32,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:20:33,957] loop.py:380 INFO TIMING step=830: step=4.181s data=1.834s compute=4.147s [fwd=2.848s bwd=1.284s opt=0.003s] dispatch=2.940s [2026-10-18 15:20:33,957] loop.py:45 INFO TRAIN_STEP step=830 loss=2.737071 Training: 5%|▍ | 830/18018 [1:50:20<37:44:52, 7.91s/step, loss=2.737, acc=0.491, acc_len=1.09, thru=15.3, I=25.4, T=30.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:34,262] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:39,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:20:40,543] loop.py:45 INFO TRAIN_STEP step=831 loss=2.581273 Training: 5%|▍ | 831/18018 [1:50:26<35:51:26, 7.51s/step, loss=2.581, acc=0.519, acc_len=1.16, thru=17.2, I=23.3, T=33.7, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:20:40,538] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=830: forward=2708.9ms backward=1041.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:40,694] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:45,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:20:46,707] loop.py:45 INFO TRAIN_STEP step=832 loss=2.683472 Training: 5%|▍ | 832/18018 [1:50:32<33:55:31, 7.11s/step, loss=2.683, acc=0.496, acc_len=1.11, thru=19.6, I=24.0, T=33.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:47,002] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:51,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:20:52,845] loop.py:45 INFO TRAIN_STEP step=833 loss=2.858169 Training: 5%|▍ | 833/18018 [1:50:38<32:32:09, 6.82s/step, loss=2.858, acc=0.473, acc_len=1.04, thru=20.0, I=25.0, T=33.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:53,232] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:59,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:20:59,774] loop.py:45 INFO TRAIN_STEP step=834 loss=2.720161 Training: 5%|▍ | 834/18018 [1:50:45<32:41:40, 6.85s/step, loss=2.720, acc=0.487, acc_len=1.09, thru=17.6, I=22.0, T=31.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:20:59,971] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:04,033] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:21:05,719] loop.py:380 INFO TIMING step=835: step=3.592s data=1.491s compute=3.562s [fwd=2.548s bwd=1.000s opt=0.003s] dispatch=2.344s [2026-10-18 15:21:05,720] loop.py:45 INFO TRAIN_STEP step=835 loss=2.939915 Training: 5%|▍ | 835/18018 [1:50:51<31:24:04, 6.58s/step, loss=2.940, acc=0.466, acc_len=1.01, thru=19.6, I=22.5, T=35.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:06,048] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:11,325] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:21:12,299] loop.py:45 INFO TRAIN_STEP step=836 loss=2.832914 Training: 5%|▍ | 836/18018 [1:50:58<31:24:01, 6.58s/step, loss=2.833, acc=0.474, acc_len=1.04, thru=18.4, I=25.3, T=34.3, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:21:12,293] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=835: forward=2546.7ms backward=1138.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:12,334] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 15:21:20,170] loop.py:45 INFO TRAIN_STEP step=837 loss=3.057282 Training: 5%|▍ | 837/18018 [1:51:06<33:14:43, 6.97s/step, loss=3.057, acc=0.458, acc_len=0.99, thru=14.2, I=15.6, T=34.2, wait=4.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:25,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:21:27,987] loop.py:45 INFO TRAIN_STEP step=838 loss=2.997074 Training: 5%|▍ | 838/18018 [1:51:14<34:27:56, 7.22s/step, loss=2.997, acc=0.466, acc_len=1.02, thru=18.6, I=21.1, T=27.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:28,216] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:33,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:21:33,896] loop.py:45 INFO TRAIN_STEP step=839 loss=2.536444 Training: 5%|▍ | 839/18018 [1:51:19<32:34:44, 6.83s/step, loss=2.536, acc=0.515, acc_len=1.16, thru=18.2, I=24.4, T=34.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:34,280] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:21:41,692] loop.py:380 INFO TIMING step=840: step=3.725s data=1.553s compute=3.692s [fwd=2.654s bwd=1.024s opt=0.003s] dispatch=4.063s [2026-10-18 15:21:41,693] loop.py:45 INFO TRAIN_STEP step=840 loss=2.890091 Training: 5%|▍ | 840/18018 [1:51:27<33:57:52, 7.12s/step, loss=2.890, acc=0.479, acc_len=1.05, thru=14.8, I=15.0, T=34.4, wait=4.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:48,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:21:50,450] loop.py:45 INFO TRAIN_STEP step=841 loss=2.861718 Training: 5%|▍ | 841/18018 [1:51:36<36:18:49, 7.61s/step, loss=2.862, acc=0.479, acc_len=1.05, thru=17.1, I=17.7, T=33.1, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:21:50,445] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=840: forward=2735.4ms backward=1075.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:50,504] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:58,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:21:59,038] loop.py:45 INFO TRAIN_STEP step=842 loss=2.765052 Training: 5%|▍ | 842/18018 [1:51:45<37:42:27, 7.90s/step, loss=2.765, acc=0.489, acc_len=1.08, thru=15.0, I=15.3, T=30.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:21:59,157] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-18 15:22:06,132] loop.py:45 INFO TRAIN_STEP step=843 loss=2.835015 Training: 5%|▍ | 843/18018 [1:51:52<36:32:53, 7.66s/step, loss=2.835, acc=0.480, acc_len=1.05, thru=16.2, I=18.1, T=30.8, wait=2.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:22:14,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:22:15,803] loop.py:45 INFO TRAIN_STEP step=844 loss=2.848877 Training: 5%|▍ | 844/18018 [1:52:01<39:25:25, 8.26s/step, loss=2.849, acc=0.484, acc_len=1.07, thru=14.7, I=14.6, T=29.5, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:22:15,901] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:22:20,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:22:21,767] loop.py:380 INFO TIMING step=845: step=3.671s data=1.576s compute=3.630s [fwd=2.563s bwd=1.052s opt=0.003s] dispatch=2.283s [2026-10-18 15:22:21,767] loop.py:45 INFO TRAIN_STEP step=845 loss=2.768733 Training: 5%|▍ | 845/18018 [1:52:07<36:07:39, 7.57s/step, loss=2.769, acc=0.486, acc_len=1.07, thru=19.4, I=23.0, T=34.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:22:22,004] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:22:28,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:22:30,084] loop.py:45 INFO TRAIN_STEP step=846 loss=2.861914 Training: 5%|▍ | 846/18018 [1:52:16<37:11:23, 7.80s/step, loss=2.862, acc=0.482, acc_len=1.06, thru=14.7, I=17.9, T=32.5, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:22:30,078] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=845: forward=2727.7ms backward=1161.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:22:30,488] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:22:35,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:22:38,036] loop.py:45 INFO TRAIN_STEP step=847 loss=2.758632 Training: 5%|▍ | 847/18018 [1:52:24<37:24:45, 7.84s/step, loss=2.759, acc=0.486, acc_len=1.08, thru=15.6, I=26.3, T=22.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:22:38,101] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=48/64 [2026-10-18 15:22:45,227] loop.py:45 INFO TRAIN_STEP step=848 loss=2.656336 Training: 5%|▍ | 848/18018 [1:52:31<36:28:37, 7.65s/step, loss=2.656, acc=0.498, acc_len=1.11, thru=17.4, I=21.0, T=34.0, wait=3.4s, pool=60, epoch=1/2][2026-10-18 15:22:52,626] loop.py:45 INFO TRAIN_STEP step=849 loss=2.496754 Training: 5%|▍ | 849/18018 [1:52:38<36:06:56, 7.57s/step, loss=2.497, acc=0.520, acc_len=1.18, thru=14.8, I=17.9, T=31.3, wait=3.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:01,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:23:03,073] loop.py:380 INFO TIMING step=850: step=5.364s data=3.034s compute=5.335s [fwd=4.151s bwd=1.169s opt=0.003s] dispatch=5.074s [2026-10-18 15:23:03,073] loop.py:45 INFO TRAIN_STEP step=850 loss=2.854080 Training: 5%|▍ | 850/18018 [1:52:49<40:13:31, 8.43s/step, loss=2.854, acc=0.478, acc_len=1.06, thru=15.2, I=15.1, T=23.9, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:03,156] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:09,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:23:10,754] loop.py:45 INFO TRAIN_STEP step=851 loss=2.713081 Training: 5%|▍ | 851/18018 [1:52:56<39:08:42, 8.21s/step, loss=2.713, acc=0.494, acc_len=1.10, thru=14.3, I=19.8, T=30.5, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:23:10,749] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=850: forward=2885.3ms backward=1249.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:10,851] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:15,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:23:16,710] loop.py:45 INFO TRAIN_STEP step=852 loss=2.747543 Training: 5%|▍ | 852/18018 [1:53:02<35:55:22, 7.53s/step, loss=2.748, acc=0.486, acc_len=1.08, thru=19.1, I=25.0, T=33.6, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:16,834] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:21,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:23:23,039] loop.py:45 INFO TRAIN_STEP step=853 loss=2.924748 Training: 5%|▍ | 853/18018 [1:53:09<34:11:48, 7.17s/step, loss=2.925, acc=0.477, acc_len=1.04, thru=19.5, I=27.0, T=31.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:23,546] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:23:31,196] loop.py:45 INFO TRAIN_STEP step=854 loss=2.689044 Training: 5%|▍ | 854/18018 [1:53:17<35:35:58, 7.47s/step, loss=2.689, acc=0.497, acc_len=1.10, thru=12.8, I=16.2, T=34.1, wait=4.4s, pool=56, epoch=1/2][2026-10-18 15:23:37,981] loop.py:380 INFO TIMING step=855: step=3.843s data=1.663s compute=3.809s [fwd=2.717s bwd=1.078s opt=0.003s] dispatch=2.934s [2026-10-18 15:23:37,981] loop.py:45 INFO TRAIN_STEP step=855 loss=2.884969 Training: 5%|▍ | 855/18018 [1:53:24<34:37:28, 7.26s/step, loss=2.885, acc=0.480, acc_len=1.05, thru=17.6, I=18.1, T=33.3, wait=2.9s, pool=60, epoch=1/2][2026-10-18 15:23:43,914] loop.py:45 INFO TRAIN_STEP step=856 loss=2.631427 Training: 5%|▍ | 856/18018 [1:53:29<32:43:07, 6.86s/step, loss=2.631, acc=0.505, acc_len=1.13, thru=19.9, I=21.9, T=35.4, wait=2.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:23:43,909] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=855: forward=2559.5ms backward=1013.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:53,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:23:53,335] loop.py:45 INFO TRAIN_STEP step=857 loss=2.826386 Training: 5%|▍ | 857/18018 [1:53:39<36:22:34, 7.63s/step, loss=2.826, acc=0.473, acc_len=1.05, thru=14.4, I=13.2, T=32.4, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:53,590] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:23:59,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:24:00,054] loop.py:45 INFO TRAIN_STEP step=858 loss=2.753856 Training: 5%|▍ | 858/18018 [1:53:46<35:04:22, 7.36s/step, loss=2.754, acc=0.497, acc_len=1.10, thru=18.2, I=18.9, T=34.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:00,257] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:24:06,703] loop.py:45 INFO TRAIN_STEP step=859 loss=2.767249 Training: 5%|▍ | 859/18018 [1:53:52<34:03:11, 7.14s/step, loss=2.767, acc=0.491, acc_len=1.08, thru=19.2, I=21.5, T=34.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:06,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:07,179] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:24:15,778] loop.py:380 INFO TIMING step=860: step=4.171s data=1.936s compute=4.127s [fwd=2.948s bwd=1.165s opt=0.003s] dispatch=4.896s [2026-10-18 15:24:15,778] loop.py:45 INFO TRAIN_STEP step=860 loss=2.733100 Training: 5%|▍ | 860/18018 [1:54:01<36:48:46, 7.72s/step, loss=2.733, acc=0.494, acc_len=1.09, thru=14.0, I=14.9, T=30.7, wait=4.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:24,140] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:24:25,464] loop.py:45 INFO TRAIN_STEP step=861 loss=2.895273 (TrainerActor pid=9310) [2026-10-18 15:24:25,459] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=860: forward=2839.1ms backward=1699.4ms Training: 5%|▍ | 861/18018 [1:54:11<39:37:05, 8.31s/step, loss=2.895, acc=0.483, acc_len=1.06, thru=13.9, I=15.0, T=27.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:25,649] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:32,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:24:34,140] loop.py:45 INFO TRAIN_STEP step=862 loss=2.740153 Training: 5%|▍ | 862/18018 [1:54:20<40:07:59, 8.42s/step, loss=2.740, acc=0.503, acc_len=1.11, thru=15.1, I=18.3, T=32.0, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:34,585] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:24:41,041] loop.py:45 INFO TRAIN_STEP step=863 loss=2.816594 Training: 5%|▍ | 863/18018 [1:54:27<37:57:21, 7.97s/step, loss=2.817, acc=0.479, acc_len=1.06, thru=17.1, I=21.4, T=33.5, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:47,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:24:48,552] loop.py:45 INFO TRAIN_STEP step=864 loss=2.786896 Training: 5%|▍ | 864/18018 [1:54:34<37:18:27, 7.83s/step, loss=2.787, acc=0.486, acc_len=1.07, thru=18.2, I=18.0, T=27.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:48,765] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:53,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:24:55,605] loop.py:380 INFO TIMING step=865: step=4.160s data=1.880s compute=4.119s [fwd=2.868s bwd=1.236s opt=0.003s] dispatch=2.882s [2026-10-18 15:24:55,605] loop.py:45 INFO TRAIN_STEP step=865 loss=2.707138 Training: 5%|▍ | 865/18018 [1:54:41<36:11:44, 7.60s/step, loss=2.707, acc=0.500, acc_len=1.10, thru=17.0, I=22.3, T=30.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:24:55,840] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:00,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:25:01,304] loop.py:45 INFO TRAIN_STEP step=866 loss=2.709612 Training: 5%|▍ | 866/18018 [1:54:47<33:28:47, 7.03s/step, loss=2.710, acc=0.494, acc_len=1.10, thru=20.1, I=26.3, T=35.8, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:25:01,299] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=865: forward=2508.0ms backward=1018.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:01,352] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:09,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:25:09,950] loop.py:45 INFO TRAIN_STEP step=867 loss=2.673826 Training: 5%|▍ | 867/18018 [1:54:56<35:47:35, 7.51s/step, loss=2.674, acc=0.497, acc_len=1.11, thru=14.0, I=16.4, T=31.6, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:10,092] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:15,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:25:16,149] loop.py:45 INFO TRAIN_STEP step=868 loss=2.979776 Training: 5%|▍ | 868/18018 [1:55:02<33:54:33, 7.12s/step, loss=2.980, acc=0.461, acc_len=1.00, thru=19.0, I=18.6, T=34.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:16,246] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:23,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:25:24,061] loop.py:45 INFO TRAIN_STEP step=869 loss=2.670588 Training: 5%|▍ | 869/18018 [1:55:10<35:02:45, 7.36s/step, loss=2.671, acc=0.498, acc_len=1.10, thru=17.5, I=18.1, T=31.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:24,286] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:25:32,047] loop.py:380 INFO TIMING step=870: step=4.666s data=2.257s compute=4.628s [fwd=3.211s bwd=1.402s opt=0.003s] dispatch=3.310s [2026-10-18 15:25:32,048] loop.py:45 INFO TRAIN_STEP step=870 loss=2.972793 Training: 5%|▍ | 870/18018 [1:55:18<35:56:24, 7.55s/step, loss=2.973, acc=0.470, acc_len=1.02, thru=14.1, I=16.6, T=27.4, wait=3.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:37,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:25:39,319] loop.py:45 INFO TRAIN_STEP step=871 loss=2.737715 (TrainerActor pid=9310) [2026-10-18 15:25:39,314] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=870: forward=2623.0ms backward=1103.2ms Training: 5%|▍ | 871/18018 [1:55:25<35:32:57, 7.46s/step, loss=2.738, acc=0.497, acc_len=1.11, thru=17.2, I=20.8, T=33.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:39,356] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:44,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:25:45,318] loop.py:45 INFO TRAIN_STEP step=872 loss=2.562525 Training: 5%|▍ | 872/18018 [1:55:31<33:27:09, 7.02s/step, loss=2.563, acc=0.513, acc_len=1.15, thru=20.0, I=27.7, T=35.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:45,349] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:52,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:25:52,852] loop.py:45 INFO TRAIN_STEP step=873 loss=2.722866 Training: 5%|▍ | 873/18018 [1:55:38<34:10:55, 7.18s/step, loss=2.723, acc=0.497, acc_len=1.09, thru=16.2, I=18.3, T=33.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:52,999] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:58,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:25:59,721] loop.py:45 INFO TRAIN_STEP step=874 loss=2.886562 Training: 5%|▍ | 874/18018 [1:55:45<33:44:24, 7.08s/step, loss=2.887, acc=0.481, acc_len=1.04, thru=17.4, I=20.7, T=31.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:25:59,907] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:26:08,425] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:26:08,592] loop.py:380 INFO TIMING step=875: step=4.308s data=1.997s compute=4.265s [fwd=3.019s bwd=1.230s opt=0.003s] dispatch=4.552s [2026-10-18 15:26:08,592] loop.py:45 INFO TRAIN_STEP step=875 loss=2.603282 Training: 5%|▍ | 875/18018 [1:55:54<36:17:08, 7.62s/step, loss=2.603, acc=0.507, acc_len=1.14, thru=14.7, I=13.4, T=29.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:26:08,926] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:26:15,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:26:17,316] loop.py:45 INFO TRAIN_STEP step=876 loss=2.764231 Training: 5%|▍ | 876/18018 [1:56:03<37:51:50, 7.95s/step, loss=2.764, acc=0.489, acc_len=1.08, thru=15.4, I=17.0, T=31.5, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:26:17,310] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=875: forward=2807.4ms backward=1194.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:26:17,806] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:26:22,757] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:26:24,545] loop.py:45 INFO TRAIN_STEP step=877 loss=2.779947 Training: 5%|▍ | 877/18018 [1:56:10<36:49:46, 7.74s/step, loss=2.780, acc=0.487, acc_len=1.07, thru=15.6, I=26.7, T=32.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:26:24,771] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:26:30,914] loop.py:45 INFO TRAIN_STEP step=878 loss=2.650938 Training: 5%|▍ | 878/18018 [1:56:16<34:52:21, 7.32s/step, loss=2.651, acc=0.501, acc_len=1.12, thru=17.3, I=22.1, T=38.2, wait=3.0s, pool=48, epoch=1/2][2026-10-18 15:26:40,882] loop.py:45 INFO TRAIN_STEP step=879 loss=2.798568 Training: 5%|▍ | 879/18018 [1:56:26<38:38:51, 8.12s/step, loss=2.799, acc=0.486, acc_len=1.07, thru=12.2, I=14.5, T=30.6, wait=5.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:26:50,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:26:51,454] loop.py:380 INFO TIMING step=880: step=4.436s data=1.964s compute=4.399s [fwd=3.148s bwd=1.237s opt=0.003s] dispatch=6.127s [2026-10-18 15:26:51,454] loop.py:45 INFO TRAIN_STEP step=880 loss=2.851029 Training: 5%|▍ | 880/18018 [1:56:37<42:08:56, 8.85s/step, loss=2.851, acc=0.483, acc_len=1.06, thru=14.1, I=15.0, T=28.9, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:26:51,899] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:27:01,419] loop.py:45 INFO TRAIN_STEP step=881 loss=2.678690 Training: 5%|▍ | 881/18018 [1:56:47<43:44:14, 9.19s/step, loss=2.679, acc=0.506, acc_len=1.12, thru=10.9, I=12.8, T=28.5, wait=5.5s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:27:01,415] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=880: forward=3192.4ms backward=1220.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:27:08,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:27:10,342] loop.py:45 INFO TRAIN_STEP step=882 loss=2.743630 Training: 5%|▍ | 882/18018 [1:56:56<43:21:14, 9.11s/step, loss=2.744, acc=0.497, acc_len=1.10, thru=16.6, I=15.8, T=30.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:27:10,585] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:27:18,183] loop.py:45 INFO TRAIN_STEP step=883 loss=2.818127 Training: 5%|▍ | 883/18018 [1:57:04<41:32:45, 8.73s/step, loss=2.818, acc=0.481, acc_len=1.06, thru=12.1, I=16.1, T=35.4, wait=4.2s, pool=44, epoch=1/2][2026-10-18 15:27:25,429] loop.py:45 INFO TRAIN_STEP step=884 loss=2.732858 Training: 5%|▍ | 884/18018 [1:57:11<39:25:21, 8.28s/step, loss=2.733, acc=0.493, acc_len=1.09, thru=17.4, I=16.8, T=34.1, wait=3.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:27:32,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:27:34,468] loop.py:380 INFO TIMING step=885: step=4.354s data=1.967s compute=4.315s [fwd=2.987s bwd=1.313s opt=0.003s] dispatch=4.675s [2026-10-18 15:27:34,468] loop.py:45 INFO TRAIN_STEP step=885 loss=2.770628 Training: 5%|▍ | 885/18018 [1:57:20<40:29:56, 8.51s/step, loss=2.771, acc=0.488, acc_len=1.08, thru=15.2, I=18.8, T=29.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:27:34,682] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:27:39,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:27:40,654] loop.py:45 INFO TRAIN_STEP step=886 loss=2.665959 Training: 5%|▍ | 886/18018 [1:57:26<37:10:46, 7.81s/step, loss=2.666, acc=0.495, acc_len=1.10, thru=17.7, I=25.7, T=36.5, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:27:40,649] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=885: forward=2486.1ms backward=976.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:27:40,945] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:27:46,752] loop.py:45 INFO TRAIN_STEP step=887 loss=2.681351 Training: 5%|▍ | 887/18018 [1:57:32<34:43:48, 7.30s/step, loss=2.681, acc=0.498, acc_len=1.10, thru=18.4, I=24.1, T=35.7, wait=2.5s, pool=44, epoch=1/2][2026-10-18 15:27:56,351] loop.py:45 INFO TRAIN_STEP step=888 loss=2.688011 Training: 5%|▍ | 888/18018 [1:57:42<38:00:43, 7.99s/step, loss=2.688, acc=0.500, acc_len=1.11, thru=14.8, I=14.5, T=32.0, wait=5.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:04,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:28:06,535] loop.py:45 INFO TRAIN_STEP step=889 loss=2.829448 Training: 5%|▍ | 889/18018 [1:57:52<41:08:37, 8.65s/step, loss=2.829, acc=0.483, acc_len=1.06, thru=13.4, I=14.7, T=26.3, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:06,673] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 [2026-10-18 15:28:13,186] loop.py:380 INFO TIMING step=890: step=4.068s data=1.869s compute=3.992s [fwd=2.794s bwd=1.183s opt=0.003s] dispatch=2.575s [2026-10-18 15:28:13,187] loop.py:45 INFO TRAIN_STEP step=890 loss=2.762120 Training: 5%|▍ | 890/18018 [1:57:59<38:17:27, 8.05s/step, loss=2.762, acc=0.489, acc_len=1.08, thru=15.4, I=17.6, T=31.5, wait=2.6s, pool=60, epoch=1/2][2026-10-18 15:28:20,287] loop.py:45 INFO TRAIN_STEP step=891 loss=2.677877 Training: 5%|▍ | 891/18018 [1:58:06<36:56:40, 7.77s/step, loss=2.678, acc=0.504, acc_len=1.12, thru=14.1, I=15.4, T=32.6, wait=3.2s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:28:20,281] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=890: forward=2768.9ms backward=1113.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:29,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:28:31,427] loop.py:45 INFO TRAIN_STEP step=892 loss=2.858876 Training: 5%|▍ | 892/18018 [1:58:17<41:44:57, 8.78s/step, loss=2.859, acc=0.477, acc_len=1.04, thru=15.6, I=15.2, T=32.7, wait=7.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:31,456] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:36,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:28:37,792] loop.py:45 INFO TRAIN_STEP step=893 loss=2.812675 Training: 5%|▍ | 893/18018 [1:58:23<38:18:39, 8.05s/step, loss=2.813, acc=0.483, acc_len=1.07, thru=19.0, I=23.0, T=34.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:37,866] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:45,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:28:46,891] loop.py:45 INFO TRAIN_STEP step=894 loss=2.588609 Training: 5%|▍ | 894/18018 [1:58:32<39:47:59, 8.37s/step, loss=2.589, acc=0.516, acc_len=1.16, thru=14.2, I=15.9, T=33.1, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:47,108] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:52,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:28:53,200] loop.py:380 INFO TIMING step=895: step=3.910s data=1.814s compute=3.876s [fwd=2.761s bwd=1.101s opt=0.003s] dispatch=2.390s [2026-10-18 15:28:53,200] loop.py:45 INFO TRAIN_STEP step=895 loss=2.718174 Training: 5%|▍ | 895/18018 [1:58:39<36:51:37, 7.75s/step, loss=2.718, acc=0.494, acc_len=1.10, thru=18.6, I=23.4, T=32.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:53,301] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:58,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:28:59,105] loop.py:45 INFO TRAIN_STEP step=896 loss=2.520573 Training: 5%|▍ | 896/18018 [1:58:45<34:13:25, 7.20s/step, loss=2.521, acc=0.513, acc_len=1.15, thru=19.9, I=22.5, T=33.9, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:28:59,101] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=895: forward=2638.0ms backward=1089.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:28:59,269] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:05,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:29:07,404] loop.py:45 INFO TRAIN_STEP step=897 loss=2.710318 Training: 5%|▍ | 897/18018 [1:58:53<35:47:46, 7.53s/step, loss=2.710, acc=0.500, acc_len=1.10, thru=16.5, I=18.7, T=28.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:07,512] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:14,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:29:14,732] loop.py:45 INFO TRAIN_STEP step=898 loss=2.667152 Training: 5%|▍ | 898/18018 [1:59:00<35:30:42, 7.47s/step, loss=2.667, acc=0.501, acc_len=1.11, thru=14.8, I=16.8, T=33.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:15,232] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:29:22,551] loop.py:45 INFO TRAIN_STEP step=899 loss=2.581491 Training: 5%|▍ | 899/18018 [1:59:08<36:00:41, 7.57s/step, loss=2.581, acc=0.519, acc_len=1.16, thru=14.2, I=14.5, T=33.7, wait=4.0s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:31,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:29:32,444] loop.py:380 INFO TIMING step=900: step=4.415s data=1.988s compute=4.381s [fwd=3.149s bwd=1.217s opt=0.003s] dispatch=5.468s [2026-10-18 15:29:32,444] loop.py:45 INFO TRAIN_STEP step=900 loss=2.674812 Training: 5%|▍ | 900/18018 [1:59:18<39:19:06, 8.27s/step, loss=2.675, acc=0.506, acc_len=1.12, thru=16.3, I=17.6, T=29.0, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:32,821] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:40,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:29:41,447] loop.py:45 INFO TRAIN_STEP step=901 loss=2.820480 Training: 5%|▌ | 901/18018 [1:59:27<40:21:54, 8.49s/step, loss=2.820, acc=0.483, acc_len=1.06, thru=14.8, I=15.4, T=32.9, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:29:41,440] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=900: forward=2722.1ms backward=1124.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:41,661] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:47,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:29:47,964] loop.py:45 INFO TRAIN_STEP step=902 loss=2.728554 Training: 5%|▌ | 902/18018 [1:59:34<37:32:48, 7.90s/step, loss=2.729, acc=0.498, acc_len=1.09, thru=18.6, I=20.4, T=34.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:48,107] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:55,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:29:56,300] loop.py:45 INFO TRAIN_STEP step=903 loss=2.826893 Training: 5%|▌ | 903/18018 [1:59:42<38:10:23, 8.03s/step, loss=2.827, acc=0.486, acc_len=1.07, thru=15.0, I=16.8, T=29.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:29:56,423] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:02,741] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:30:03,513] loop.py:45 INFO TRAIN_STEP step=904 loss=2.680117 Training: 5%|▌ | 904/18018 [1:59:49<37:00:38, 7.79s/step, loss=2.680, acc=0.499, acc_len=1.11, thru=17.0, I=18.4, T=30.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:03,746] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:10,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:30:11,213] loop.py:380 INFO TIMING step=905: step=4.240s data=1.948s compute=4.212s [fwd=2.946s bwd=1.251s opt=0.003s] dispatch=3.447s [2026-10-18 15:30:11,213] loop.py:45 INFO TRAIN_STEP step=905 loss=2.710686 Training: 5%|▌ | 905/18018 [1:59:57<36:52:46, 7.76s/step, loss=2.711, acc=0.497, acc_len=1.09, thru=16.4, I=18.7, T=30.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:11,683] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:30:17,400] loop.py:45 INFO TRAIN_STEP step=906 loss=2.655940 Training: 5%|▌ | 906/18018 [2:00:03<34:38:15, 7.29s/step, loss=2.656, acc=0.501, acc_len=1.11, thru=17.7, I=18.5, T=35.8, wait=2.6s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:30:17,395] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=905: forward=2494.3ms backward=1034.9ms [2026-10-18 15:30:24,759] loop.py:45 INFO TRAIN_STEP step=907 loss=2.630254 Training: 5%|▌ | 907/18018 [2:00:10<34:44:25, 7.31s/step, loss=2.630, acc=0.511, acc_len=1.14, thru=16.9, I=17.7, T=33.4, wait=3.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:31,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:30:31,808] loop.py:45 INFO TRAIN_STEP step=908 loss=2.834071 Training: 5%|▌ | 908/18018 [2:00:17<34:22:02, 7.23s/step, loss=2.834, acc=0.479, acc_len=1.05, thru=17.6, I=18.6, T=33.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:32,181] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:39,047] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:30:39,765] loop.py:45 INFO TRAIN_STEP step=909 loss=2.635430 Training: 5%|▌ | 909/18018 [2:00:25<35:23:57, 7.45s/step, loss=2.635, acc=0.505, acc_len=1.12, thru=16.1, I=18.9, T=33.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:40,054] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:44,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:30:46,390] loop.py:380 INFO TIMING step=910: step=4.082s data=1.927s compute=4.052s [fwd=2.843s bwd=1.196s opt=0.003s] dispatch=2.532s [2026-10-18 15:30:46,390] loop.py:45 INFO TRAIN_STEP step=910 loss=2.981977 Training: 5%|▌ | 910/18018 [2:00:32<34:13:27, 7.20s/step, loss=2.982, acc=0.459, acc_len=1.00, thru=19.2, I=23.8, T=31.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:30:46,599] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:30:52,759] loop.py:45 INFO TRAIN_STEP step=911 loss=2.723191 Training: 5%|▌ | 911/18018 [2:00:38<33:01:56, 6.95s/step, loss=2.723, acc=0.494, acc_len=1.09, thru=16.0, I=22.4, T=34.3, wait=2.6s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:30:52,754] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=910: forward=2612.5ms backward=1079.1ms [2026-10-18 15:31:02,199] loop.py:45 INFO TRAIN_STEP step=912 loss=2.897295 Training: 5%|▌ | 912/18018 [2:00:48<36:34:36, 7.70s/step, loss=2.897, acc=0.473, acc_len=1.03, thru=12.7, I=13.0, T=33.8, wait=5.6s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:10,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:31:11,089] loop.py:45 INFO TRAIN_STEP step=913 loss=2.691050 Training: 5%|▌ | 913/18018 [2:00:57<38:16:40, 8.06s/step, loss=2.691, acc=0.499, acc_len=1.11, thru=18.3, I=15.0, T=35.6, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:11,124] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 15:31:18,574] loop.py:45 INFO TRAIN_STEP step=914 loss=2.772670 Training: 5%|▌ | 914/18018 [2:01:04<37:27:48, 7.89s/step, loss=2.773, acc=0.494, acc_len=1.08, thru=15.3, I=16.4, T=32.0, wait=3.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:27,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:31:28,433] loop.py:380 INFO TIMING step=915: step=4.127s data=1.766s compute=4.096s [fwd=2.921s bwd=1.159s opt=0.003s] dispatch=5.720s [2026-10-18 15:31:28,433] loop.py:45 INFO TRAIN_STEP step=915 loss=2.690548 Training: 5%|▌ | 915/18018 [2:01:14<40:16:28, 8.48s/step, loss=2.691, acc=0.503, acc_len=1.11, thru=15.2, I=16.6, T=31.0, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:28,841] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:33,955] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:31:35,813] loop.py:45 INFO TRAIN_STEP step=916 loss=2.678297 Training: 5%|▌ | 916/18018 [2:01:21<38:42:24, 8.15s/step, loss=2.678, acc=0.502, acc_len=1.11, thru=16.4, I=22.8, T=27.2, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:31:35,809] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=915: forward=3323.3ms backward=1334.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:35,972] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:44,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:31:44,589] loop.py:45 INFO TRAIN_STEP step=917 loss=2.746557 Training: 5%|▌ | 917/18018 [2:01:30<39:35:53, 8.34s/step, loss=2.747, acc=0.490, acc_len=1.07, thru=14.9, I=15.1, T=34.2, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:45,028] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:49,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:31:52,058] loop.py:45 INFO TRAIN_STEP step=918 loss=2.597486 Training: 5%|▌ | 918/18018 [2:01:38<38:21:44, 8.08s/step, loss=2.597, acc=0.517, acc_len=1.15, thru=17.0, I=23.6, T=30.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:52,363] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:58,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:31:59,081] loop.py:45 INFO TRAIN_STEP step=919 loss=2.836165 Training: 5%|▌ | 919/18018 [2:01:45<36:51:21, 7.76s/step, loss=2.836, acc=0.477, acc_len=1.05, thru=15.2, I=21.7, T=30.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:31:59,519] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:05,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:32:05,342] loop.py:380 INFO TIMING step=920: step=3.549s data=1.427s compute=3.506s [fwd=2.484s bwd=1.006s opt=0.003s] dispatch=2.704s [2026-10-18 15:32:05,342] loop.py:45 INFO TRAIN_STEP step=920 loss=2.720929 Training: 5%|▌ | 920/18018 [2:01:51<34:43:01, 7.31s/step, loss=2.721, acc=0.498, acc_len=1.10, thru=19.2, I=19.7, T=36.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:05,639] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:10,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 15:32:12,892] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=920: forward=2984.7ms backward=1346.9ms [2026-10-18 15:32:12,896] loop.py:45 INFO TRAIN_STEP step=921 loss=2.649606 Training: 5%|▌ | 921/18018 [2:01:58<35:04:02, 7.38s/step, loss=2.650, acc=0.500, acc_len=1.11, thru=18.2, I=22.5, T=29.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:13,116] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 15:32:22,818] loop.py:45 INFO TRAIN_STEP step=922 loss=2.724040 Training: 5%|▌ | 922/18018 [2:02:08<38:40:38, 8.14s/step, loss=2.724, acc=0.496, acc_len=1.09, thru=13.2, I=12.9, T=29.7, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:22,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:23,294] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:28,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:32:30,003] loop.py:45 INFO TRAIN_STEP step=923 loss=2.734536 Training: 5%|▌ | 923/18018 [2:02:16<37:18:42, 7.86s/step, loss=2.735, acc=0.492, acc_len=1.09, thru=17.0, I=19.4, T=31.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:30,153] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:35,689] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:32:36,424] loop.py:45 INFO TRAIN_STEP step=924 loss=2.475438 Training: 5%|▌ | 924/18018 [2:02:22<35:15:38, 7.43s/step, loss=2.475, acc=0.525, acc_len=1.18, thru=18.0, I=23.7, T=33.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:36,694] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:32:44,635] loop.py:380 INFO TIMING step=925: step=3.508s data=1.423s compute=3.481s [fwd=2.498s bwd=0.968s opt=0.003s] dispatch=4.694s [2026-10-18 15:32:44,635] loop.py:45 INFO TRAIN_STEP step=925 loss=2.675350 Training: 5%|▌ | 925/18018 [2:02:30<36:22:50, 7.66s/step, loss=2.675, acc=0.501, acc_len=1.11, thru=13.1, I=15.1, T=36.5, wait=4.7s, pool=44, epoch=1/2][2026-10-18 15:32:51,936] loop.py:45 INFO TRAIN_STEP step=926 loss=2.815794 Training: 5%|▌ | 926/18018 [2:02:38<35:51:34, 7.55s/step, loss=2.816, acc=0.485, acc_len=1.07, thru=17.7, I=17.3, T=32.1, wait=3.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:32:51,932] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=925: forward=2772.8ms backward=1176.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:57,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:32:58,628] loop.py:45 INFO TRAIN_STEP step=927 loss=2.697633 Training: 5%|▌ | 927/18018 [2:02:44<34:37:52, 7.29s/step, loss=2.698, acc=0.500, acc_len=1.11, thru=18.0, I=19.5, T=35.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:32:59,105] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:04,746] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:33:06,023] loop.py:45 INFO TRAIN_STEP step=928 loss=2.672847 Training: 5%|▌ | 928/18018 [2:02:52<34:46:33, 7.33s/step, loss=2.673, acc=0.502, acc_len=1.11, thru=16.8, I=22.3, T=34.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:06,256] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:12,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:33:13,604] loop.py:45 INFO TRAIN_STEP step=929 loss=2.600759 Training: 5%|▌ | 929/18018 [2:02:59<35:08:05, 7.40s/step, loss=2.601, acc=0.506, acc_len=1.13, thru=16.1, I=20.7, T=29.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:14,088] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:33:19,294] loop.py:380 INFO TIMING step=930: step=3.447s data=1.437s compute=3.422s [fwd=2.462s bwd=0.949s opt=0.003s] dispatch=2.235s [2026-10-18 15:33:19,294] loop.py:45 INFO TRAIN_STEP step=930 loss=2.705548 Training: 5%|▌ | 930/18018 [2:03:05<32:41:45, 6.89s/step, loss=2.706, acc=0.492, acc_len=1.08, thru=19.0, I=21.2, T=37.1, wait=2.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:27,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:33:28,111] loop.py:45 INFO TRAIN_STEP step=931 loss=2.643776 Training: 5%|▌ | 931/18018 [2:03:14<35:26:24, 7.47s/step, loss=2.644, acc=0.504, acc_len=1.12, thru=15.6, I=16.9, T=34.8, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:33:28,106] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=930: forward=2621.2ms backward=1010.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:28,573] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:33:35,606] loop.py:45 INFO TRAIN_STEP step=932 loss=2.633973 Training: 5%|▌ | 932/18018 [2:03:21<35:28:51, 7.48s/step, loss=2.634, acc=0.506, acc_len=1.13, thru=14.5, I=19.1, T=30.7, wait=3.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:43,631] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:33:45,142] loop.py:45 INFO TRAIN_STEP step=933 loss=2.636669 Training: 5%|▌ | 933/18018 [2:03:31<38:24:29, 8.09s/step, loss=2.637, acc=0.514, acc_len=1.15, thru=15.0, I=16.3, T=30.6, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:45,642] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:52,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:33:53,303] loop.py:45 INFO TRAIN_STEP step=934 loss=2.795220 Training: 5%|▌ | 934/18018 [2:03:39<38:30:24, 8.11s/step, loss=2.795, acc=0.486, acc_len=1.06, thru=14.1, I=17.7, T=30.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:33:53,654] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:34:02,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:34:03,716] loop.py:380 INFO TIMING step=935: step=4.480s data=2.058s compute=4.437s [fwd=3.079s bwd=1.346s opt=0.003s] dispatch=5.923s [2026-10-18 15:34:03,716] loop.py:45 INFO TRAIN_STEP step=935 loss=2.543279 Training: 5%|▌ | 935/18018 [2:03:49<41:46:26, 8.80s/step, loss=2.543, acc=0.521, acc_len=1.17, thru=12.2, I=13.7, T=28.6, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:34:03,941] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:34:11,450] loop.py:45 INFO TRAIN_STEP step=936 loss=2.682939 Training: 5%|▌ | 936/18018 [2:03:57<40:15:00, 8.48s/step, loss=2.683, acc=0.507, acc_len=1.12, thru=14.5, I=14.6, T=32.7, wait=3.8s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:34:11,445] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=935: forward=2732.7ms backward=1132.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:34:17,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:34:18,774] loop.py:45 INFO TRAIN_STEP step=937 loss=2.503836 Training: 5%|▌ | 937/18018 [2:04:04<38:36:04, 8.14s/step, loss=2.504, acc=0.528, acc_len=1.19, thru=17.1, I=20.7, T=30.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:34:19,090] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:34:23,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:34:25,496] loop.py:45 INFO TRAIN_STEP step=938 loss=2.561005 Training: 5%|▌ | 938/18018 [2:04:11<36:35:06, 7.71s/step, loss=2.561, acc=0.521, acc_len=1.17, thru=16.6, I=24.9, T=30.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:34:25,777] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:34:32,564] loop.py:45 INFO TRAIN_STEP step=939 loss=2.627327 Training: 5%|▌ | 939/18018 [2:04:18<35:40:05, 7.52s/step, loss=2.627, acc=0.504, acc_len=1.12, thru=15.2, I=19.1, T=31.9, wait=3.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:34:41,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:34:43,506] loop.py:380 INFO TIMING step=940: step=4.247s data=1.961s compute=4.206s [fwd=2.968s bwd=1.224s opt=0.003s] dispatch=6.686s [2026-10-18 15:34:43,507] loop.py:45 INFO TRAIN_STEP step=940 loss=2.650404 Training: 5%|▌ | 940/18018 [2:04:29<40:32:23, 8.55s/step, loss=2.650, acc=0.509, acc_len=1.13, thru=14.2, I=15.5, T=30.1, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:34:43,625] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-18 15:34:51,222] loop.py:45 INFO TRAIN_STEP step=941 loss=2.540271 Training: 5%|▌ | 941/18018 [2:04:37<39:21:09, 8.30s/step, loss=2.540, acc=0.520, acc_len=1.16, thru=12.1, I=14.5, T=29.1, wait=3.3s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:34:51,218] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=940: forward=3137.3ms backward=1223.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:34:58,750] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:35:00,142] loop.py:45 INFO TRAIN_STEP step=942 loss=2.579019 Training: 5%|▌ | 942/18018 [2:04:46<40:14:22, 8.48s/step, loss=2.579, acc=0.516, acc_len=1.15, thru=16.5, I=16.6, T=30.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:00,261] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 15:35:07,985] loop.py:45 INFO TRAIN_STEP step=943 loss=2.762627 Training: 5%|▌ | 943/18018 [2:04:54<39:19:40, 8.29s/step, loss=2.763, acc=0.492, acc_len=1.08, thru=14.6, I=18.3, T=32.8, wait=3.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:13,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:35:15,287] loop.py:45 INFO TRAIN_STEP step=944 loss=2.499181 Training: 5%|▌ | 944/18018 [2:05:01<37:55:03, 7.99s/step, loss=2.499, acc=0.523, acc_len=1.17, thru=17.9, I=20.1, T=32.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:15,321] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:20,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:35:22,122] loop.py:380 INFO TIMING step=945: step=3.979s data=1.858s compute=3.939s [fwd=2.861s bwd=1.065s opt=0.003s] dispatch=2.846s [2026-10-18 15:35:22,122] loop.py:45 INFO TRAIN_STEP step=945 loss=2.542226 Training: 5%|▌ | 945/18018 [2:05:08<36:15:53, 7.65s/step, loss=2.542, acc=0.519, acc_len=1.16, thru=17.4, I=23.5, T=32.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:22,561] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:28,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:35:30,512] loop.py:45 INFO TRAIN_STEP step=946 loss=2.478739 Training: 5%|▌ | 946/18018 [2:05:16<37:19:16, 7.87s/step, loss=2.479, acc=0.523, acc_len=1.17, thru=13.2, I=22.4, T=25.6, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:35:30,507] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=945: forward=3745.9ms backward=1202.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:30,551] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:36,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:35:38,097] loop.py:45 INFO TRAIN_STEP step=947 loss=2.470742 Training: 5%|▌ | 947/18018 [2:05:24<36:54:48, 7.78s/step, loss=2.471, acc=0.526, acc_len=1.18, thru=13.9, I=21.2, T=30.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:38,288] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:47,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:35:47,295] loop.py:45 INFO TRAIN_STEP step=948 loss=2.541138 Training: 5%|▌ | 948/18018 [2:05:33<38:55:09, 8.21s/step, loss=2.541, acc=0.520, acc_len=1.17, thru=14.2, I=16.1, T=30.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:35:47,564] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:35:56,572] loop.py:45 INFO TRAIN_STEP step=949 loss=2.482962 Training: 5%|▌ | 949/18018 [2:05:42<40:26:11, 8.53s/step, loss=2.483, acc=0.526, acc_len=1.18, thru=13.0, I=12.0, T=28.8, wait=4.8s, pool=56, epoch=1/2][2026-10-18 15:36:03,979] loop.py:380 INFO TIMING step=950: step=4.115s data=2.035s compute=4.085s [fwd=2.994s bwd=1.077s opt=0.003s] dispatch=3.285s [2026-10-18 15:36:03,980] loop.py:45 INFO TRAIN_STEP step=950 loss=2.656322 Training: 5%|▌ | 950/18018 [2:05:50<38:50:21, 8.19s/step, loss=2.656, acc=0.505, acc_len=1.12, thru=16.2, I=17.2, T=31.1, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:10,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:36:12,058] loop.py:45 INFO TRAIN_STEP step=951 loss=2.610047 Training: 5%|▌ | 951/18018 [2:05:58<38:40:49, 8.16s/step, loss=2.610, acc=0.514, acc_len=1.14, thru=15.3, I=17.1, T=33.8, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:36:12,053] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=950: forward=2713.1ms backward=1029.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:12,412] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:17,568] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:36:18,942] loop.py:45 INFO TRAIN_STEP step=952 loss=2.514014 Training: 5%|▌ | 952/18018 [2:06:05<36:51:48, 7.78s/step, loss=2.514, acc=0.525, acc_len=1.17, thru=17.5, I=23.5, T=31.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:19,098] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:24,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:36:25,622] loop.py:45 INFO TRAIN_STEP step=953 loss=2.457032 Training: 5%|▌ | 953/18018 [2:06:11<35:18:02, 7.45s/step, loss=2.457, acc=0.527, acc_len=1.19, thru=17.8, I=23.2, T=28.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:26,048] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:30,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:36:32,364] loop.py:45 INFO TRAIN_STEP step=954 loss=2.786415 Training: 5%|▌ | 954/18018 [2:06:18<34:18:00, 7.24s/step, loss=2.786, acc=0.494, acc_len=1.09, thru=17.3, I=26.0, T=30.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:32,541] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:36:37,917] loop.py:380 INFO TIMING step=955: step=3.443s data=1.427s compute=3.410s [fwd=2.432s bwd=0.964s opt=0.003s] dispatch=2.099s [2026-10-18 15:36:37,917] loop.py:45 INFO TRAIN_STEP step=955 loss=2.473068 Training: 5%|▌ | 955/18018 [2:06:24<31:53:54, 6.73s/step, loss=2.473, acc=0.524, acc_len=1.18, thru=17.1, I=24.7, T=37.2, wait=2.1s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:45,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:36:47,507] loop.py:45 INFO TRAIN_STEP step=956 loss=2.461572 Training: 5%|▌ | 956/18018 [2:06:33<35:57:53, 7.59s/step, loss=2.462, acc=0.530, acc_len=1.19, thru=17.1, I=19.7, T=27.3, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:36:47,502] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=955: forward=3242.9ms backward=1399.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:47,746] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:54,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:36:54,999] loop.py:45 INFO TRAIN_STEP step=957 loss=2.362747 Training: 5%|▌ | 957/18018 [2:06:41<35:49:39, 7.56s/step, loss=2.363, acc=0.544, acc_len=1.24, thru=13.5, I=20.5, T=32.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:36:55,265] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:37:01,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:37:04,021] loop.py:45 INFO TRAIN_STEP step=958 loss=2.615043 Training: 5%|▌ | 958/18018 [2:06:50<37:54:16, 8.00s/step, loss=2.615, acc=0.501, acc_len=1.11, thru=13.5, I=19.1, T=25.8, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:37:04,231] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 15:37:10,812] loop.py:45 INFO TRAIN_STEP step=959 loss=2.621171 Training: 5%|▌ | 959/18018 [2:06:56<36:11:04, 7.64s/step, loss=2.621, acc=0.511, acc_len=1.13, thru=13.3, I=22.0, T=34.6, wait=3.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:37:20,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:37:22,301] loop.py:380 INFO TIMING step=960: step=4.145s data=1.887s compute=4.116s [fwd=2.955s bwd=1.147s opt=0.003s] dispatch=7.334s [2026-10-18 15:37:22,301] loop.py:45 INFO TRAIN_STEP step=960 loss=2.666018 Training: 5%|▌ | 960/18018 [2:07:08<41:39:38, 8.79s/step, loss=2.666, acc=0.502, acc_len=1.11, thru=14.4, I=15.0, T=30.9, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:37:22,670] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 15:37:29,352] loop.py:45 INFO TRAIN_STEP step=961 loss=2.507020 Training: 5%|▌ | 961/18018 [2:07:15<39:10:58, 8.27s/step, loss=2.507, acc=0.523, acc_len=1.18, thru=14.2, I=22.6, T=31.3, wait=2.9s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:37:29,347] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=960: forward=2943.9ms backward=1098.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:37:36,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:37:40,474] loop.py:45 INFO TRAIN_STEP step=962 loss=2.565554 Training: 5%|▌ | 962/18018 [2:07:26<43:14:04, 9.13s/step, loss=2.566, acc=0.520, acc_len=1.16, thru=12.5, I=17.1, T=22.5, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:37:40,801] inference_manager.py:408 INFO Pool capacity available after 4.0s, pool_size=0/64 [2026-10-18 15:37:47,136] loop.py:45 INFO TRAIN_STEP step=963 loss=2.526687 Training: 5%|▌ | 963/18018 [2:07:33<39:43:53, 8.39s/step, loss=2.527, acc=0.516, acc_len=1.15, thru=19.6, I=18.2, T=39.2, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:37:52,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:37:54,248] loop.py:45 INFO TRAIN_STEP step=964 loss=2.750856 Training: 5%|▌ | 964/18018 [2:07:40<37:55:00, 8.00s/step, loss=2.751, acc=0.486, acc_len=1.07, thru=18.5, I=21.0, T=31.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:37:54,305] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:37:59,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:38:00,234] loop.py:380 INFO TIMING step=965: step=3.621s data=1.520s compute=3.592s [fwd=2.556s bwd=1.022s opt=0.003s] dispatch=2.354s [2026-10-18 15:38:00,234] loop.py:45 INFO TRAIN_STEP step=965 loss=2.459222 Training: 5%|▌ | 965/18018 [2:07:46<35:02:47, 7.40s/step, loss=2.459, acc=0.533, acc_len=1.19, thru=18.9, I=24.1, T=35.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:38:00,557] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:38:09,721] loop.py:45 INFO TRAIN_STEP step=966 loss=2.735670 Training: 5%|▌ | 966/18018 [2:07:55<38:00:41, 8.02s/step, loss=2.736, acc=0.493, acc_len=1.08, thru=13.2, I=14.2, T=31.1, wait=5.4s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:38:09,716] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=965: forward=2919.1ms backward=1130.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:38:17,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:38:18,128] loop.py:45 INFO TRAIN_STEP step=967 loss=2.612414 Training: 5%|▌ | 967/18018 [2:08:04<38:33:12, 8.14s/step, loss=2.612, acc=0.510, acc_len=1.13, thru=15.3, I=15.3, T=31.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:38:18,415] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:38:24,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:38:24,680] loop.py:45 INFO TRAIN_STEP step=968 loss=2.579461 Training: 5%|▌ | 968/18018 [2:08:10<36:17:36, 7.66s/step, loss=2.579, acc=0.514, acc_len=1.14, thru=18.0, I=19.6, T=36.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:38:24,978] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:38:29,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:38:31,630] loop.py:45 INFO TRAIN_STEP step=969 loss=2.573927 Training: 5%|▌ | 969/18018 [2:08:17<35:16:43, 7.45s/step, loss=2.574, acc=0.514, acc_len=1.14, thru=18.3, I=23.6, T=28.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:38:32,089] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 15:38:41,101] loop.py:380 INFO TIMING step=970: step=4.157s data=1.919s compute=4.123s [fwd=2.869s bwd=1.240s opt=0.003s] dispatch=5.304s [2026-10-18 15:38:41,101] loop.py:45 INFO TRAIN_STEP step=970 loss=2.557336 Training: 5%|▌ | 970/18018 [2:08:27<38:08:43, 8.06s/step, loss=2.557, acc=0.519, acc_len=1.16, thru=11.6, I=10.8, T=30.8, wait=5.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:38:47,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:38:48,225] loop.py:45 INFO TRAIN_STEP step=971 loss=2.591526 Training: 5%|▌ | 971/18018 [2:08:34<36:49:29, 7.78s/step, loss=2.592, acc=0.511, acc_len=1.13, thru=18.4, I=18.3, T=32.9, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:38:48,220] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=970: forward=2753.3ms backward=1088.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:38:48,420] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:38:55,057] loop.py:45 INFO TRAIN_STEP step=972 loss=2.827923 Training: 5%|▌ | 972/18018 [2:08:41<35:28:36, 7.49s/step, loss=2.828, acc=0.483, acc_len=1.05, thru=15.7, I=19.1, T=35.0, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:02,749] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:39:03,373] loop.py:45 INFO TRAIN_STEP step=973 loss=2.696244 Training: 5%|▌ | 973/18018 [2:08:49<36:38:53, 7.74s/step, loss=2.696, acc=0.499, acc_len=1.09, thru=16.8, I=16.7, T=34.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:03,754] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:09,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:39:11,302] loop.py:45 INFO TRAIN_STEP step=974 loss=2.520630 Training: 5%|▌ | 974/18018 [2:08:57<36:54:52, 7.80s/step, loss=2.521, acc=0.517, acc_len=1.16, thru=16.4, I=19.6, T=29.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:11,688] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:16,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:39:18,049] loop.py:380 INFO TIMING step=975: step=4.147s data=1.708s compute=4.120s [fwd=2.743s bwd=1.362s opt=0.003s] dispatch=2.589s [2026-10-18 15:39:18,049] loop.py:45 INFO TRAIN_STEP step=975 loss=2.502046 Training: 5%|▌ | 975/18018 [2:09:04<35:25:03, 7.48s/step, loss=2.502, acc=0.524, acc_len=1.17, thru=17.0, I=24.3, T=30.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:18,330] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:39:23,981] loop.py:45 INFO TRAIN_STEP step=976 loss=2.436844 Training: 5%|▌ | 976/18018 [2:09:10<33:12:49, 7.02s/step, loss=2.437, acc=0.531, acc_len=1.19, thru=19.1, I=22.1, T=35.1, wait=2.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:39:23,975] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=975: forward=2576.2ms backward=1023.0ms [2026-10-18 15:39:32,635] loop.py:45 INFO TRAIN_STEP step=977 loss=2.524295 Training: 5%|▌ | 977/18018 [2:09:18<35:32:19, 7.51s/step, loss=2.524, acc=0.520, acc_len=1.16, thru=13.6, I=15.8, T=29.5, wait=4.3s, pool=36, epoch=1/2][2026-10-18 15:39:42,183] loop.py:45 INFO TRAIN_STEP step=978 loss=2.647810 Training: 5%|▌ | 978/18018 [2:09:28<38:26:10, 8.12s/step, loss=2.648, acc=0.501, acc_len=1.11, thru=15.3, I=14.5, T=32.0, wait=5.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:50,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:39:51,825] loop.py:45 INFO TRAIN_STEP step=979 loss=2.607666 Training: 5%|▌ | 979/18018 [2:09:37<40:35:40, 8.58s/step, loss=2.608, acc=0.517, acc_len=1.16, thru=14.4, I=14.7, T=31.2, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:52,282] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:58,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:39:58,941] loop.py:380 INFO TIMING step=980: step=3.770s data=1.588s compute=3.737s [fwd=2.629s bwd=1.094s opt=0.003s] dispatch=3.337s [2026-10-18 15:39:58,942] loop.py:45 INFO TRAIN_STEP step=980 loss=2.630253 Training: 5%|▌ | 980/18018 [2:09:45<38:31:00, 8.14s/step, loss=2.630, acc=0.501, acc_len=1.11, thru=15.9, I=18.5, T=34.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:39:59,245] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:40:06,082] loop.py:45 INFO TRAIN_STEP step=981 loss=2.456795 Training: 5%|▌ | 981/18018 [2:09:52<37:05:50, 7.84s/step, loss=2.457, acc=0.528, acc_len=1.19, thru=18.6, I=18.5, T=31.5, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:40:06,077] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=980: forward=2814.6ms backward=1202.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:06,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:06,588] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:14,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:40:15,400] loop.py:45 INFO TRAIN_STEP step=982 loss=2.763543 Training: 5%|▌ | 982/18018 [2:10:01<39:11:45, 8.28s/step, loss=2.764, acc=0.492, acc_len=1.08, thru=14.2, I=14.7, T=34.2, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:15,855] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:20,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:40:22,401] loop.py:45 INFO TRAIN_STEP step=983 loss=2.620411 Training: 5%|▌ | 983/18018 [2:10:08<37:22:37, 7.90s/step, loss=2.620, acc=0.508, acc_len=1.13, thru=18.9, I=22.7, T=32.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:22,465] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 15:40:31,068] loop.py:45 INFO TRAIN_STEP step=984 loss=2.722488 Training: 5%|▌ | 984/18018 [2:10:17<38:27:51, 8.13s/step, loss=2.722, acc=0.496, acc_len=1.09, thru=14.0, I=16.1, T=29.3, wait=4.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:39,753] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:40:41,449] loop.py:380 INFO TIMING step=985: step=4.126s data=1.857s compute=4.101s [fwd=2.961s bwd=1.125s opt=0.003s] dispatch=6.244s [2026-10-18 15:40:41,449] loop.py:45 INFO TRAIN_STEP step=985 loss=2.526011 Training: 5%|▌ | 985/18018 [2:10:27<41:39:25, 8.80s/step, loss=2.526, acc=0.523, acc_len=1.17, thru=13.2, I=14.9, T=31.0, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:41,765] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (TrainerActor pid=9310) [2026-10-18 15:40:48,341] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=985: forward=2798.8ms backward=1214.1ms [2026-10-18 15:40:48,347] loop.py:45 INFO TRAIN_STEP step=986 loss=2.512999 Training: 5%|▌ | 986/18018 [2:10:34<38:56:43, 8.23s/step, loss=2.513, acc=0.526, acc_len=1.18, thru=16.6, I=20.7, T=31.4, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:55,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:40:57,133] loop.py:45 INFO TRAIN_STEP step=987 loss=2.636432 Training: 5%|▌ | 987/18018 [2:10:43<39:44:06, 8.40s/step, loss=2.636, acc=0.508, acc_len=1.13, thru=15.8, I=15.8, T=30.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:40:57,343] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:41:05,119] loop.py:45 INFO TRAIN_STEP step=988 loss=2.429702 Training: 5%|▌ | 988/18018 [2:10:51<39:08:35, 8.27s/step, loss=2.430, acc=0.535, acc_len=1.20, thru=14.8, I=16.9, T=32.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:41:05,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:41:05,599] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:41:11,284] loop.py:45 INFO TRAIN_STEP step=989 loss=2.523287 Training: 5%|▌ | 989/18018 [2:10:57<36:08:56, 7.64s/step, loss=2.523, acc=0.518, acc_len=1.16, thru=19.1, I=19.1, T=32.8, wait=2.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:41:22,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:41:23,412] loop.py:380 INFO TIMING step=990: step=4.295s data=1.829s compute=4.228s [fwd=2.949s bwd=1.265s opt=0.003s] dispatch=7.823s [2026-10-18 15:41:23,412] loop.py:45 INFO TRAIN_STEP step=990 loss=2.729028 Training: 5%|▌ | 990/18018 [2:11:09<42:30:50, 8.99s/step, loss=2.729, acc=0.493, acc_len=1.09, thru=16.2, I=11.6, T=29.8, wait=7.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:41:23,464] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 15:41:31,043] loop.py:45 INFO TRAIN_STEP step=991 loss=2.838486 Training: 6%|▌ | 991/18018 [2:11:17<40:35:02, 8.58s/step, loss=2.838, acc=0.484, acc_len=1.05, thru=14.6, I=19.2, T=33.3, wait=3.8s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:41:31,038] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=990: forward=2699.3ms backward=1090.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:41:38,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:41:40,021] loop.py:45 INFO TRAIN_STEP step=992 loss=2.519357 Training: 6%|▌ | 992/18018 [2:11:26<41:08:39, 8.70s/step, loss=2.519, acc=0.524, acc_len=1.17, thru=17.1, I=16.5, T=33.3, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:41:40,266] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:41:46,547] loop.py:45 INFO TRAIN_STEP step=993 loss=2.456186 Training: 6%|▌ | 993/18018 [2:11:32<38:03:25, 8.05s/step, loss=2.456, acc=0.523, acc_len=1.18, thru=18.0, I=25.6, T=34.1, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:41:54,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:41:55,593] loop.py:45 INFO TRAIN_STEP step=994 loss=2.392296 Training: 6%|▌ | 994/18018 [2:11:41<39:28:21, 8.35s/step, loss=2.392, acc=0.537, acc_len=1.21, thru=14.6, I=15.1, T=32.0, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:41:55,778] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:42:03,190] loop.py:380 INFO TIMING step=995: step=3.523s data=1.479s compute=3.488s [fwd=2.541s bwd=0.933s opt=0.003s] dispatch=4.065s [2026-10-18 15:42:03,190] loop.py:45 INFO TRAIN_STEP step=995 loss=2.790141 Training: 6%|▌ | 995/18018 [2:11:49<38:24:17, 8.12s/step, loss=2.790, acc=0.485, acc_len=1.06, thru=13.6, I=15.6, T=36.3, wait=4.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:42:12,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:42:13,240] loop.py:45 INFO TRAIN_STEP step=996 loss=2.547146 Training: 6%|▌ | 996/18018 [2:11:59<41:08:26, 8.70s/step, loss=2.547, acc=0.522, acc_len=1.16, thru=14.6, I=15.1, T=32.3, wait=6.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:42:13,236] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=995: forward=2747.6ms backward=1172.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:42:13,441] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:42:19,997] loop.py:45 INFO TRAIN_STEP step=997 loss=2.518960 Training: 6%|▌ | 997/18018 [2:12:06<38:22:58, 8.12s/step, loss=2.519, acc=0.515, acc_len=1.15, thru=18.2, I=18.6, T=35.4, wait=3.1s, pool=60, epoch=1/2][2026-10-18 15:42:29,174] loop.py:45 INFO TRAIN_STEP step=998 loss=2.622602 Training: 6%|▌ | 998/18018 [2:12:15<39:52:50, 8.44s/step, loss=2.623, acc=0.510, acc_len=1.13, thru=13.2, I=17.1, T=25.8, wait=4.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:42:35,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:42:37,248] loop.py:45 INFO TRAIN_STEP step=999 loss=2.319487 Training: 6%|▌ | 999/18018 [2:12:23<39:22:03, 8.33s/step, loss=2.319, acc=0.552, acc_len=1.25, thru=15.6, I=15.7, T=30.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:42:37,435] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:42:43,374] loop.py:380 INFO TIMING step=1000: step=3.674s data=1.579s compute=3.640s [fwd=2.572s bwd=1.053s opt=0.003s] dispatch=2.441s [2026-10-18 15:42:43,375] loop.py:45 INFO TRAIN_STEP step=1000 loss=2.424096 Training: 6%|▌ | 1000/18018 [2:12:29<36:14:26, 7.67s/step, loss=2.424, acc=0.532, acc_len=1.20, thru=17.6, I=23.4, T=34.8, wait=2.4s, pool=56, epoch=1/2][2026-10-18 15:42:43,378] loop.py:419 INFO Saving checkpoint at step 1000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:42:43,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 15:42:46,920] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0001001 (TrainerActor pid=9310) [2026-10-18 15:42:47,346] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0000251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:42:48,183] inference_manager.py:408 INFO Pool capacity available after 4.6s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:42:53,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 15:42:46,918] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0001001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 15:42:47,710] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0000251 [2026-10-18 15:42:55,498] loop.py:45 INFO TRAIN_STEP step=1001 loss=2.603327 Training: 6%|▌ | 1001/18018 [2:12:41<42:33:38, 9.00s/step, loss=2.603, acc=0.517, acc_len=1.15, thru=17.5, I=21.2, T=29.2, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:42:55,494] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1000: forward=3048.2ms backward=1263.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:42:55,955] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:02,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:43:03,871] loop.py:45 INFO TRAIN_STEP step=1002 loss=2.534266 Training: 6%|▌ | 1002/18018 [2:12:49<41:39:51, 8.81s/step, loss=2.534, acc=0.517, acc_len=1.16, thru=16.2, I=18.6, T=31.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:04,021] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:09,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:43:10,801] loop.py:45 INFO TRAIN_STEP step=1003 loss=2.548320 Training: 6%|▌ | 1003/18018 [2:12:56<38:59:15, 8.25s/step, loss=2.548, acc=0.518, acc_len=1.15, thru=17.4, I=22.7, T=32.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:11,085] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:16,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:43:17,071] loop.py:45 INFO TRAIN_STEP step=1004 loss=2.488066 Training: 6%|▌ | 1004/18018 [2:13:03<36:10:52, 7.66s/step, loss=2.488, acc=0.527, acc_len=1.18, thru=18.1, I=24.8, T=32.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:17,297] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:23,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:43:24,158] loop.py:380 INFO TIMING step=1005: step=4.448s data=2.004s compute=4.417s [fwd=3.112s bwd=1.290s opt=0.003s] dispatch=2.628s [2026-10-18 15:43:24,158] loop.py:45 INFO TRAIN_STEP step=1005 loss=2.589878 Training: 6%|▌ | 1005/18018 [2:13:10<35:22:15, 7.48s/step, loss=2.590, acc=0.511, acc_len=1.13, thru=17.4, I=19.6, T=28.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:24,190] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:29,308] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:43:30,841] loop.py:45 INFO TRAIN_STEP step=1006 loss=2.522022 Training: 6%|▌ | 1006/18018 [2:13:16<34:13:52, 7.24s/step, loss=2.522, acc=0.519, acc_len=1.16, thru=17.2, I=20.6, T=32.8, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:43:30,837] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1005: forward=2735.3ms backward=1115.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:31,324] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:36,395] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:43:37,710] loop.py:45 INFO TRAIN_STEP step=1007 loss=2.478746 Training: 6%|▌ | 1007/18018 [2:13:23<33:42:06, 7.13s/step, loss=2.479, acc=0.519, acc_len=1.17, thru=17.6, I=24.1, T=32.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:37,907] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:43,182] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:43:44,137] loop.py:45 INFO TRAIN_STEP step=1008 loss=2.506474 Training: 6%|▌ | 1008/18018 [2:13:30<32:41:49, 6.92s/step, loss=2.506, acc=0.523, acc_len=1.17, thru=18.2, I=21.9, T=33.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:44,200] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:50,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:43:50,675] loop.py:45 INFO TRAIN_STEP step=1009 loss=2.386184 Training: 6%|▌ | 1009/18018 [2:13:36<32:09:23, 6.81s/step, loss=2.386, acc=0.542, acc_len=1.23, thru=18.8, I=20.7, T=30.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:50,710] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:56,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:43:56,725] loop.py:380 INFO TIMING step=1010: step=3.582s data=1.495s compute=3.545s [fwd=2.521s bwd=1.010s opt=0.003s] dispatch=2.458s [2026-10-18 15:43:56,725] loop.py:45 INFO TRAIN_STEP step=1010 loss=2.726698 Training: 6%|▌ | 1010/18018 [2:13:42<31:04:53, 6.58s/step, loss=2.727, acc=0.492, acc_len=1.08, thru=19.3, I=21.1, T=35.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:43:56,955] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:04,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:44:04,969] loop.py:45 INFO TRAIN_STEP step=1011 loss=2.478559 Training: 6%|▌ | 1011/18018 [2:13:51<33:26:20, 7.08s/step, loss=2.479, acc=0.535, acc_len=1.20, thru=15.8, I=15.8, T=31.0, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:44:04,965] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1010: forward=2888.6ms backward=1197.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:05,202] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:11,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:44:11,955] loop.py:45 INFO TRAIN_STEP step=1012 loss=2.499388 Training: 6%|▌ | 1012/18018 [2:13:58<33:18:33, 7.05s/step, loss=2.499, acc=0.520, acc_len=1.17, thru=19.1, I=18.6, T=30.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:12,198] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:18,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:44:19,172] loop.py:45 INFO TRAIN_STEP step=1013 loss=2.696882 Training: 6%|▌ | 1013/18018 [2:14:05<33:32:21, 7.10s/step, loss=2.697, acc=0.500, acc_len=1.10, thru=16.5, I=20.8, T=28.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:19,536] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:25,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:44:27,056] loop.py:45 INFO TRAIN_STEP step=1014 loss=2.611597 Training: 6%|▌ | 1014/18018 [2:14:13<34:38:53, 7.34s/step, loss=2.612, acc=0.510, acc_len=1.13, thru=14.1, I=17.7, T=33.9, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:27,363] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:33,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:44:34,921] loop.py:380 INFO TIMING step=1015: step=3.809s data=1.652s compute=3.773s [fwd=2.722s bwd=1.037s opt=0.003s] dispatch=4.048s [2026-10-18 15:44:34,921] loop.py:45 INFO TRAIN_STEP step=1015 loss=2.436408 Training: 6%|▌ | 1015/18018 [2:14:21<35:23:44, 7.49s/step, loss=2.436, acc=0.535, acc_len=1.19, thru=15.2, I=20.1, T=33.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:35,301] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:40,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:44:42,259] loop.py:45 INFO TRAIN_STEP step=1016 loss=2.452271 Training: 6%|▌ | 1016/18018 [2:14:28<35:10:23, 7.45s/step, loss=2.452, acc=0.528, acc_len=1.18, thru=15.7, I=22.3, T=30.4, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:44:42,254] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1015: forward=2969.4ms backward=1183.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:42,574] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:48,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:44:48,914] loop.py:45 INFO TRAIN_STEP step=1017 loss=2.496519 Training: 6%|▌ | 1017/18018 [2:14:34<34:02:52, 7.21s/step, loss=2.497, acc=0.527, acc_len=1.17, thru=16.9, I=23.0, T=33.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:49,341] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:58,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:44:58,351] loop.py:45 INFO TRAIN_STEP step=1018 loss=2.517495 Training: 6%|▌ | 1018/18018 [2:14:44<37:12:04, 7.88s/step, loss=2.517, acc=0.527, acc_len=1.18, thru=14.2, I=13.6, T=31.8, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:44:58,811] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:05,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:45:06,407] loop.py:45 INFO TRAIN_STEP step=1019 loss=2.327205 Training: 6%|▌ | 1019/18018 [2:14:52<37:27:00, 7.93s/step, loss=2.327, acc=0.545, acc_len=1.24, thru=15.5, I=16.7, T=29.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:06,637] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:13,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:45:13,688] loop.py:380 INFO TIMING step=1020: step=4.094s data=1.773s compute=4.062s [fwd=2.840s bwd=1.207s opt=0.003s] dispatch=3.178s [2026-10-18 15:45:13,688] loop.py:45 INFO TRAIN_STEP step=1020 loss=2.584645 Training: 6%|▌ | 1020/18018 [2:14:59<36:31:39, 7.74s/step, loss=2.585, acc=0.511, acc_len=1.13, thru=17.9, I=17.9, T=31.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:13,945] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:45:20,046] loop.py:45 INFO TRAIN_STEP step=1021 loss=2.316074 Training: 6%|▌ | 1021/18018 [2:15:06<34:34:27, 7.32s/step, loss=2.316, acc=0.546, acc_len=1.24, thru=16.6, I=17.9, T=35.3, wait=2.7s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:45:20,040] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1020: forward=2545.4ms backward=1028.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:27,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:45:27,770] loop.py:45 INFO TRAIN_STEP step=1022 loss=2.544123 Training: 6%|▌ | 1022/18018 [2:15:13<35:08:33, 7.44s/step, loss=2.544, acc=0.517, acc_len=1.16, thru=18.0, I=17.4, T=31.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:28,014] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:33,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:45:35,963] loop.py:45 INFO TRAIN_STEP step=1023 loss=2.582709 Training: 6%|▌ | 1023/18018 [2:15:22<36:11:51, 7.67s/step, loss=2.583, acc=0.511, acc_len=1.14, thru=15.7, I=18.1, T=30.1, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:36,463] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:41,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:45:42,745] loop.py:45 INFO TRAIN_STEP step=1024 loss=2.563712 Training: 6%|▌ | 1024/18018 [2:15:28<34:56:39, 7.40s/step, loss=2.564, acc=0.509, acc_len=1.13, thru=16.1, I=25.9, T=32.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:43,233] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:48,888] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:45:49,529] loop.py:380 INFO TIMING step=1025: step=3.971s data=1.723s compute=3.935s [fwd=2.793s bwd=1.126s opt=0.003s] dispatch=2.803s [2026-10-18 15:45:49,529] loop.py:45 INFO TRAIN_STEP step=1025 loss=2.549327 Training: 6%|▌ | 1025/18018 [2:15:35<34:03:51, 7.22s/step, loss=2.549, acc=0.518, acc_len=1.16, thru=16.8, I=22.7, T=32.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:49,894] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:45:58,405] loop.py:45 INFO TRAIN_STEP step=1026 loss=2.618501 Training: 6%|▌ | 1026/18018 [2:15:44<36:24:34, 7.71s/step, loss=2.619, acc=0.512, acc_len=1.13, thru=13.7, I=14.4, T=34.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:58,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 15:45:58,400] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1025: forward=2660.9ms backward=1041.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:45:58,879] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:03,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:46:05,049] loop.py:45 INFO TRAIN_STEP step=1027 loss=2.399016 Training: 6%|▌ | 1027/18018 [2:15:51<34:53:49, 7.39s/step, loss=2.399, acc=0.543, acc_len=1.22, thru=18.2, I=21.9, T=33.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:05,386] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:46:13,625] loop.py:45 INFO TRAIN_STEP step=1028 loss=2.467826 Training: 6%|▌ | 1028/18018 [2:15:59<36:34:08, 7.75s/step, loss=2.468, acc=0.536, acc_len=1.20, thru=12.7, I=16.2, T=29.5, wait=4.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:21,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:46:24,147] loop.py:45 INFO TRAIN_STEP step=1029 loss=2.581432 Training: 6%|▌ | 1029/18018 [2:16:10<40:29:31, 8.58s/step, loss=2.581, acc=0.516, acc_len=1.14, thru=12.8, I=14.0, T=29.0, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:24,510] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 15:46:32,450] loop.py:380 INFO TIMING step=1030: step=3.865s data=1.661s compute=3.837s [fwd=2.717s bwd=1.104s opt=0.003s] dispatch=4.428s [2026-10-18 15:46:32,450] loop.py:45 INFO TRAIN_STEP step=1030 loss=2.539531 Training: 6%|▌ | 1030/18018 [2:16:18<40:05:52, 8.50s/step, loss=2.540, acc=0.524, acc_len=1.17, thru=15.1, I=17.3, T=33.1, wait=4.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:40,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:46:42,727] loop.py:45 INFO TRAIN_STEP step=1031 loss=2.312672 Training: 6%|▌ | 1031/18018 [2:16:28<42:36:53, 9.03s/step, loss=2.313, acc=0.552, acc_len=1.25, thru=13.6, I=14.0, T=25.3, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:46:42,722] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1030: forward=3477.4ms backward=1512.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:42,809] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:48,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:46:48,631] loop.py:45 INFO TRAIN_STEP step=1032 loss=2.559948 Training: 6%|▌ | 1032/18018 [2:16:34<38:10:55, 8.09s/step, loss=2.560, acc=0.518, acc_len=1.15, thru=17.8, I=19.7, T=36.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:49,044] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:54,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:46:56,789] loop.py:45 INFO TRAIN_STEP step=1033 loss=2.444350 Training: 6%|▌ | 1033/18018 [2:16:42<38:16:37, 8.11s/step, loss=2.444, acc=0.523, acc_len=1.17, thru=17.2, I=21.2, T=31.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:46:56,912] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:02,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:47:02,976] loop.py:45 INFO TRAIN_STEP step=1034 loss=2.655648 Training: 6%|▌ | 1034/18018 [2:16:49<35:32:45, 7.53s/step, loss=2.656, acc=0.502, acc_len=1.10, thru=18.5, I=24.8, T=36.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:03,375] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:47:09,758] loop.py:380 INFO TIMING step=1035: step=3.756s data=1.694s compute=3.722s [fwd=2.641s bwd=1.066s opt=0.003s] dispatch=3.017s [2026-10-18 15:47:09,758] loop.py:45 INFO TRAIN_STEP step=1035 loss=2.467187 Training: 6%|▌ | 1035/18018 [2:16:55<34:28:45, 7.31s/step, loss=2.467, acc=0.530, acc_len=1.19, thru=16.0, I=20.9, T=34.1, wait=3.0s, pool=32, epoch=1/2][2026-10-18 15:47:18,642] loop.py:45 INFO TRAIN_STEP step=1036 loss=2.580931 Training: 6%|▌ | 1036/18018 [2:17:04<36:42:27, 7.78s/step, loss=2.581, acc=0.519, acc_len=1.15, thru=14.8, I=15.3, T=33.5, wait=5.1s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:47:18,637] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1035: forward=2654.9ms backward=1100.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:27,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:47:28,623] loop.py:45 INFO TRAIN_STEP step=1037 loss=2.634488 Training: 6%|▌ | 1037/18018 [2:17:14<39:48:57, 8.44s/step, loss=2.634, acc=0.509, acc_len=1.13, thru=15.2, I=15.7, T=30.0, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:29,083] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:34,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:47:35,957] loop.py:45 INFO TRAIN_STEP step=1038 loss=2.518403 Training: 6%|▌ | 1038/18018 [2:17:22<38:15:02, 8.11s/step, loss=2.518, acc=0.521, acc_len=1.16, thru=15.2, I=22.0, T=26.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:36,192] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:42,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:47:44,803] loop.py:45 INFO TRAIN_STEP step=1039 loss=2.705909 Training: 6%|▌ | 1039/18018 [2:17:30<39:17:17, 8.33s/step, loss=2.706, acc=0.497, acc_len=1.09, thru=14.9, I=16.0, T=27.0, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:45,121] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:51,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:47:52,765] loop.py:380 INFO TIMING step=1040: step=4.724s data=2.265s compute=4.686s [fwd=3.277s bwd=1.395s opt=0.003s] dispatch=3.229s [2026-10-18 15:47:52,766] loop.py:45 INFO TRAIN_STEP step=1040 loss=2.544614 Training: 6%|▌ | 1040/18018 [2:17:38<38:45:52, 8.22s/step, loss=2.545, acc=0.519, acc_len=1.15, thru=16.8, I=19.8, T=27.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:53,269] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:47:59,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:48:00,170] loop.py:45 INFO TRAIN_STEP step=1041 loss=2.402017 Training: 6%|▌ | 1041/18018 [2:17:46<37:36:34, 7.98s/step, loss=2.402, acc=0.536, acc_len=1.21, thru=15.4, I=20.4, T=33.1, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:48:00,166] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1040: forward=2694.8ms backward=1122.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:00,272] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:07,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:48:07,248] loop.py:45 INFO TRAIN_STEP step=1042 loss=2.543787 Training: 6%|▌ | 1042/18018 [2:17:53<36:20:17, 7.71s/step, loss=2.544, acc=0.522, acc_len=1.16, thru=16.4, I=19.1, T=29.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:07,589] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:15,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:48:17,585] loop.py:45 INFO TRAIN_STEP step=1043 loss=2.428531 Training: 6%|▌ | 1043/18018 [2:18:03<40:03:23, 8.50s/step, loss=2.429, acc=0.531, acc_len=1.19, thru=12.8, I=15.3, T=29.0, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:17,750] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:23,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:48:23,642] loop.py:45 INFO TRAIN_STEP step=1044 loss=2.670479 Training: 6%|▌ | 1044/18018 [2:18:09<36:36:23, 7.76s/step, loss=2.670, acc=0.501, acc_len=1.10, thru=18.7, I=23.5, T=35.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:23,800] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:48:33,461] loop.py:380 INFO TIMING step=1045: step=4.599s data=1.663s compute=4.488s [fwd=3.326s bwd=1.146s opt=0.003s] dispatch=5.211s [2026-10-18 15:48:33,461] loop.py:45 INFO TRAIN_STEP step=1045 loss=2.599176 Training: 6%|▌ | 1045/18018 [2:18:19<39:30:54, 8.38s/step, loss=2.599, acc=0.514, acc_len=1.14, thru=12.8, I=13.7, T=27.8, wait=5.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:41,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:48:42,756] loop.py:45 INFO TRAIN_STEP step=1046 loss=2.525989 Training: 6%|▌ | 1046/18018 [2:18:28<40:48:03, 8.65s/step, loss=2.526, acc=0.521, acc_len=1.16, thru=14.1, I=14.9, T=29.3, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:48:42,752] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1045: forward=2986.8ms backward=1325.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:42,904] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:47,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:48:48,943] loop.py:45 INFO TRAIN_STEP step=1047 loss=2.404347 Training: 6%|▌ | 1047/18018 [2:18:35<37:18:44, 7.91s/step, loss=2.404, acc=0.537, acc_len=1.22, thru=19.3, I=22.4, T=31.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:48,984] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:54,907] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:48:56,303] loop.py:45 INFO TRAIN_STEP step=1048 loss=2.533078 Training: 6%|▌ | 1048/18018 [2:18:42<36:31:29, 7.75s/step, loss=2.533, acc=0.517, acc_len=1.15, thru=15.2, I=19.4, T=31.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:48:56,431] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:01,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:49:02,210] loop.py:45 INFO TRAIN_STEP step=1049 loss=2.386579 Training: 6%|▌ | 1049/18018 [2:18:48<33:55:07, 7.20s/step, loss=2.387, acc=0.531, acc_len=1.20, thru=19.8, I=23.4, T=34.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:02,331] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:11,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:49:11,602] loop.py:380 INFO TIMING step=1050: step=3.790s data=1.592s compute=3.745s [fwd=2.621s bwd=1.110s opt=0.003s] dispatch=5.592s [2026-10-18 15:49:11,602] loop.py:45 INFO TRAIN_STEP step=1050 loss=2.354059 Training: 6%|▌ | 1050/18018 [2:18:57<37:01:06, 7.85s/step, loss=2.354, acc=0.546, acc_len=1.23, thru=13.8, I=13.1, T=33.8, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:12,053] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:19,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:49:21,418] loop.py:45 INFO TRAIN_STEP step=1051 loss=2.585017 Training: 6%|▌ | 1051/18018 [2:19:07<39:47:24, 8.44s/step, loss=2.585, acc=0.521, acc_len=1.16, thru=13.7, I=17.8, T=32.2, wait=5.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:49:21,413] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1050: forward=2798.7ms backward=1123.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:21,901] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 15:49:28,105] loop.py:45 INFO TRAIN_STEP step=1052 loss=2.389930 Training: 6%|▌ | 1052/18018 [2:19:14<37:18:40, 7.92s/step, loss=2.390, acc=0.530, acc_len=1.19, thru=15.4, I=25.4, T=32.9, wait=2.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:36,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:49:38,627] loop.py:45 INFO TRAIN_STEP step=1053 loss=2.401764 Training: 6%|▌ | 1053/18018 [2:19:24<40:59:21, 8.70s/step, loss=2.402, acc=0.534, acc_len=1.21, thru=15.0, I=15.6, T=27.7, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:39,082] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 15:49:45,028] loop.py:45 INFO TRAIN_STEP step=1054 loss=2.542598 Training: 6%|▌ | 1054/18018 [2:19:31<37:44:10, 8.01s/step, loss=2.543, acc=0.524, acc_len=1.17, thru=17.3, I=24.2, T=35.6, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:51,258] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:49:52,283] loop.py:380 INFO TIMING step=1055: step=4.242s data=1.865s compute=4.209s [fwd=2.970s bwd=1.224s opt=0.003s] dispatch=3.006s [2026-10-18 15:49:52,283] loop.py:45 INFO TRAIN_STEP step=1055 loss=2.442337 Training: 6%|▌ | 1055/18018 [2:19:38<36:40:23, 7.78s/step, loss=2.442, acc=0.533, acc_len=1.20, thru=18.6, I=19.4, T=30.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:52,768] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:49:58,535] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:49:59,909] loop.py:45 INFO TRAIN_STEP step=1056 loss=2.432017 Training: 6%|▌ | 1056/18018 [2:19:45<36:26:54, 7.74s/step, loss=2.432, acc=0.531, acc_len=1.20, thru=15.8, I=20.9, T=32.8, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:49:59,903] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1055: forward=2721.5ms backward=1123.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:00,047] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 15:50:05,833] loop.py:45 INFO TRAIN_STEP step=1057 loss=2.431623 Training: 6%|▌ | 1057/18018 [2:19:51<33:52:57, 7.19s/step, loss=2.432, acc=0.534, acc_len=1.19, thru=19.5, I=20.0, T=36.5, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:14,131] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:50:14,918] loop.py:45 INFO TRAIN_STEP step=1058 loss=2.423558 Training: 6%|▌ | 1058/18018 [2:20:01<36:33:31, 7.76s/step, loss=2.424, acc=0.532, acc_len=1.19, thru=15.8, I=16.4, T=31.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:15,136] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:23,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:50:24,348] loop.py:45 INFO TRAIN_STEP step=1059 loss=2.417286 Training: 6%|▌ | 1059/18018 [2:20:10<38:55:06, 8.26s/step, loss=2.417, acc=0.534, acc_len=1.20, thru=13.9, I=15.7, T=26.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:24,382] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 15:50:31,077] loop.py:380 INFO TIMING step=1060: step=4.361s data=1.918s compute=4.323s [fwd=2.969s bwd=1.340s opt=0.003s] dispatch=2.358s [2026-10-18 15:50:31,077] loop.py:45 INFO TRAIN_STEP step=1060 loss=2.340744 Training: 6%|▌ | 1060/18018 [2:20:17<36:44:53, 7.80s/step, loss=2.341, acc=0.548, acc_len=1.25, thru=13.3, I=14.0, T=29.4, wait=2.4s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:39,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:50:40,688] loop.py:45 INFO TRAIN_STEP step=1061 loss=2.657026 Training: 6%|▌ | 1061/18018 [2:20:26<39:18:16, 8.34s/step, loss=2.657, acc=0.508, acc_len=1.12, thru=16.4, I=18.6, T=30.2, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:50:40,683] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1060: forward=2889.4ms backward=1284.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:41,178] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:47,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:50:49,287] loop.py:45 INFO TRAIN_STEP step=1062 loss=2.651999 Training: 6%|▌ | 1062/18018 [2:20:35<39:39:45, 8.42s/step, loss=2.652, acc=0.510, acc_len=1.12, thru=14.0, I=16.3, T=32.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:49,722] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:55,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:50:56,035] loop.py:45 INFO TRAIN_STEP step=1063 loss=2.483675 Training: 6%|▌ | 1063/18018 [2:20:42<37:17:38, 7.92s/step, loss=2.484, acc=0.525, acc_len=1.17, thru=17.6, I=20.9, T=32.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:50:56,290] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:51:02,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:51:03,843] loop.py:45 INFO TRAIN_STEP step=1064 loss=2.599390 Training: 6%|▌ | 1064/18018 [2:20:49<37:08:17, 7.89s/step, loss=2.599, acc=0.510, acc_len=1.13, thru=17.0, I=18.8, T=32.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:51:04,014] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:51:11,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:51:13,411] loop.py:380 INFO TIMING step=1065: step=5.297s data=2.765s compute=5.269s [fwd=4.069s bwd=1.186s opt=0.003s] dispatch=4.261s [2026-10-18 15:51:13,412] loop.py:45 INFO TRAIN_STEP step=1065 loss=2.315564 Training: 6%|▌ | 1065/18018 [2:20:59<39:30:48, 8.39s/step, loss=2.316, acc=0.554, acc_len=1.26, thru=13.7, I=15.8, T=24.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:51:13,616] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:51:19,886] loop.py:45 INFO TRAIN_STEP step=1066 loss=2.502334 Training: 6%|▌ | 1066/18018 [2:21:05<36:48:04, 7.82s/step, loss=2.502, acc=0.520, acc_len=1.16, thru=15.0, I=17.2, T=32.0, wait=2.5s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:51:19,882] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1065: forward=2804.5ms backward=1152.6ms [2026-10-18 15:51:27,952] loop.py:45 INFO TRAIN_STEP step=1067 loss=2.609594 Training: 6%|▌ | 1067/18018 [2:21:14<37:09:15, 7.89s/step, loss=2.610, acc=0.508, acc_len=1.12, thru=16.8, I=15.1, T=30.2, wait=3.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:51:34,745] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:51:36,008] loop.py:45 INFO TRAIN_STEP step=1068 loss=2.618941 Training: 6%|▌ | 1068/18018 [2:21:22<37:23:05, 7.94s/step, loss=2.619, acc=0.508, acc_len=1.12, thru=17.1, I=18.0, T=34.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:51:36,252] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:51:46,154] loop.py:45 INFO TRAIN_STEP step=1069 loss=2.378600 Training: 6%|▌ | 1069/18018 [2:21:32<40:30:07, 8.60s/step, loss=2.379, acc=0.544, acc_len=1.23, thru=11.7, I=13.3, T=30.9, wait=6.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:51:53,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:51:54,005] loop.py:380 INFO TIMING step=1070: step=3.857s data=1.711s compute=3.829s [fwd=2.723s bwd=1.092s opt=0.003s] dispatch=3.983s [2026-10-18 15:51:54,005] loop.py:45 INFO TRAIN_STEP step=1070 loss=2.562470 Training: 6%|▌ | 1070/18018 [2:21:40<39:26:03, 8.38s/step, loss=2.562, acc=0.520, acc_len=1.15, thru=16.2, I=16.3, T=33.2, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:51:54,035] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:51:58,822] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:51:59,944] loop.py:45 INFO TRAIN_STEP step=1071 loss=2.515395 Training: 6%|▌ | 1071/18018 [2:21:46<35:59:31, 7.65s/step, loss=2.515, acc=0.519, acc_len=1.15, thru=19.9, I=23.5, T=32.9, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:51:59,939] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1070: forward=2699.3ms backward=1142.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:00,332] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:05,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:52:06,775] loop.py:45 INFO TRAIN_STEP step=1072 loss=2.360649 Training: 6%|▌ | 1072/18018 [2:21:52<34:50:11, 7.40s/step, loss=2.361, acc=0.538, acc_len=1.21, thru=19.1, I=26.8, T=31.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:07,126] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:12,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:52:13,432] loop.py:45 INFO TRAIN_STEP step=1073 loss=2.607399 Training: 6%|▌ | 1073/18018 [2:21:59<33:47:04, 7.18s/step, loss=2.607, acc=0.513, acc_len=1.13, thru=17.2, I=24.7, T=31.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:13,905] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:18,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:52:20,818] loop.py:45 INFO TRAIN_STEP step=1074 loss=2.452356 Training: 6%|▌ | 1074/18018 [2:22:06<34:04:46, 7.24s/step, loss=2.452, acc=0.521, acc_len=1.17, thru=17.7, I=26.0, T=29.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:21,158] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 15:52:27,017] loop.py:380 INFO TIMING step=1075: step=3.685s data=1.656s compute=3.639s [fwd=2.603s bwd=1.022s opt=0.003s] dispatch=2.503s [2026-10-18 15:52:27,018] loop.py:45 INFO TRAIN_STEP step=1075 loss=2.433985 Training: 6%|▌ | 1075/18018 [2:22:13<32:36:28, 6.93s/step, loss=2.434, acc=0.534, acc_len=1.20, thru=16.4, I=27.2, T=34.7, wait=2.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:35,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:52:37,160] loop.py:45 INFO TRAIN_STEP step=1076 loss=2.350295 Training: 6%|▌ | 1076/18018 [2:22:23<37:08:28, 7.89s/step, loss=2.350, acc=0.550, acc_len=1.25, thru=14.3, I=14.5, T=25.3, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:52:37,156] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1075: forward=3547.6ms backward=1463.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:37,256] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:42,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:52:44,078] loop.py:45 INFO TRAIN_STEP step=1077 loss=2.478714 Training: 6%|▌ | 1077/18018 [2:22:30<35:45:59, 7.60s/step, loss=2.479, acc=0.522, acc_len=1.17, thru=15.4, I=22.3, T=31.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:44,236] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:53,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:52:53,562] loop.py:45 INFO TRAIN_STEP step=1078 loss=2.387893 Training: 6%|▌ | 1078/18018 [2:22:39<38:25:14, 8.16s/step, loss=2.388, acc=0.539, acc_len=1.21, thru=13.8, I=14.6, T=27.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:52:53,997] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 15:53:01,496] loop.py:45 INFO TRAIN_STEP step=1079 loss=2.460057 Training: 6%|▌ | 1079/18018 [2:22:47<38:05:45, 8.10s/step, loss=2.460, acc=0.531, acc_len=1.19, thru=13.7, I=13.5, T=28.0, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:07,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:53:08,960] loop.py:380 INFO TIMING step=1080: step=4.380s data=2.006s compute=4.350s [fwd=3.116s bwd=1.219s opt=0.003s] dispatch=3.072s [2026-10-18 15:53:08,960] loop.py:45 INFO TRAIN_STEP step=1080 loss=2.346725 Training: 6%|▌ | 1080/18018 [2:22:55<37:11:57, 7.91s/step, loss=2.347, acc=0.551, acc_len=1.24, thru=17.8, I=18.2, T=29.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:08,995] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 15:53:16,031] loop.py:45 INFO TRAIN_STEP step=1081 loss=2.530061 Training: 6%|▌ | 1081/18018 [2:23:02<36:01:09, 7.66s/step, loss=2.530, acc=0.518, acc_len=1.15, thru=15.0, I=18.6, T=32.7, wait=3.1s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:53:16,026] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1080: forward=2740.9ms backward=1115.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:23,646] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:53:26,866] loop.py:45 INFO TRAIN_STEP step=1082 loss=2.363158 Training: 6%|▌ | 1082/18018 [2:23:12<40:30:07, 8.61s/step, loss=2.363, acc=0.544, acc_len=1.23, thru=13.3, I=15.6, T=23.7, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:27,172] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:32,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:53:33,263] loop.py:45 INFO TRAIN_STEP step=1083 loss=2.308008 Training: 6%|▌ | 1083/18018 [2:23:19<37:22:30, 7.95s/step, loss=2.308, acc=0.548, acc_len=1.25, thru=14.2, I=24.9, T=32.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:33,497] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:43,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:53:43,744] loop.py:45 INFO TRAIN_STEP step=1084 loss=2.625635 Training: 6%|▌ | 1084/18018 [2:23:29<40:57:17, 8.71s/step, loss=2.626, acc=0.517, acc_len=1.14, thru=12.6, I=12.3, T=25.5, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:44,090] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 15:53:50,807] loop.py:380 INFO TIMING step=1085: step=4.177s data=1.896s compute=4.143s [fwd=2.939s bwd=1.188s opt=0.003s] dispatch=2.876s [2026-10-18 15:53:50,807] loop.py:45 INFO TRAIN_STEP step=1085 loss=2.379924 Training: 6%|▌ | 1085/18018 [2:23:36<38:37:58, 8.21s/step, loss=2.380, acc=0.535, acc_len=1.21, thru=14.9, I=15.8, T=30.6, wait=2.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:57,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:53:59,016] loop.py:45 INFO TRAIN_STEP step=1086 loss=2.314757 Training: 6%|▌ | 1086/18018 [2:23:45<38:37:27, 8.21s/step, loss=2.315, acc=0.550, acc_len=1.25, thru=17.1, I=17.6, T=32.4, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:53:59,012] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1085: forward=2712.5ms backward=1168.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:53:59,095] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 15:54:05,439] loop.py:45 INFO TRAIN_STEP step=1087 loss=2.459621 Training: 6%|▌ | 1087/18018 [2:23:51<36:05:46, 7.68s/step, loss=2.460, acc=0.530, acc_len=1.19, thru=16.6, I=21.3, T=35.4, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:12,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:54:14,722] loop.py:45 INFO TRAIN_STEP step=1088 loss=2.629379 Training: 6%|▌ | 1088/18018 [2:24:00<38:21:53, 8.16s/step, loss=2.629, acc=0.507, acc_len=1.12, thru=15.4, I=16.5, T=30.1, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:15,085] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:19,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:54:21,420] loop.py:45 INFO TRAIN_STEP step=1089 loss=2.494088 Training: 6%|▌ | 1089/18018 [2:24:07<36:18:09, 7.72s/step, loss=2.494, acc=0.529, acc_len=1.18, thru=15.3, I=25.8, T=32.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:21,505] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:26,558] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:54:27,804] loop.py:380 INFO TIMING step=1090: step=4.012s data=1.755s compute=3.980s [fwd=2.775s bwd=1.191s opt=0.003s] dispatch=2.361s [2026-10-18 15:54:27,804] loop.py:45 INFO TRAIN_STEP step=1090 loss=2.386503 Training: 6%|▌ | 1090/18018 [2:24:13<34:24:47, 7.32s/step, loss=2.387, acc=0.532, acc_len=1.20, thru=18.1, I=27.2, T=31.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:28,065] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:35,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:54:36,919] loop.py:45 INFO TRAIN_STEP step=1091 loss=2.462384 Training: 6%|▌ | 1091/18018 [2:24:23<36:56:43, 7.86s/step, loss=2.462, acc=0.528, acc_len=1.19, thru=14.5, I=15.8, T=31.3, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:54:36,914] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1090: forward=2931.1ms backward=1098.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:37,348] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:42,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:54:44,982] loop.py:45 INFO TRAIN_STEP step=1092 loss=2.384693 Training: 6%|▌ | 1092/18018 [2:24:31<37:14:00, 7.92s/step, loss=2.385, acc=0.538, acc_len=1.21, thru=13.7, I=21.5, T=28.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:45,179] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:50,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:54:51,979] loop.py:45 INFO TRAIN_STEP step=1093 loss=2.505742 Training: 6%|▌ | 1093/18018 [2:24:38<35:55:56, 7.64s/step, loss=2.506, acc=0.521, acc_len=1.16, thru=15.3, I=22.4, T=33.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:54:52,410] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:54:59,598] loop.py:45 INFO TRAIN_STEP step=1094 loss=2.411684 Training: 6%|▌ | 1094/18018 [2:24:45<35:53:44, 7.64s/step, loss=2.412, acc=0.536, acc_len=1.20, thru=15.6, I=17.2, T=34.0, wait=3.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:04,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:55:05,386] loop.py:380 INFO TIMING step=1095: step=3.822s data=1.669s compute=3.789s [fwd=2.683s bwd=1.091s opt=0.003s] dispatch=1.957s [2026-10-18 15:55:05,386] loop.py:45 INFO TRAIN_STEP step=1095 loss=2.379262 Training: 6%|▌ | 1095/18018 [2:24:51<33:17:13, 7.08s/step, loss=2.379, acc=0.536, acc_len=1.21, thru=20.4, I=21.7, T=33.5, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:05,437] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:11,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:55:13,214] loop.py:45 INFO TRAIN_STEP step=1096 loss=2.525689 Training: 6%|▌ | 1096/18018 [2:24:59<34:20:17, 7.31s/step, loss=2.526, acc=0.519, acc_len=1.15, thru=16.5, I=19.7, T=27.9, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:55:13,209] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1095: forward=3205.9ms backward=1333.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:13,272] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:19,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:55:20,367] loop.py:45 INFO TRAIN_STEP step=1097 loss=2.477400 Training: 6%|▌ | 1097/18018 [2:25:06<34:07:26, 7.26s/step, loss=2.477, acc=0.532, acc_len=1.19, thru=16.3, I=20.3, T=30.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:20,528] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:25,903] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:55:27,676] loop.py:45 INFO TRAIN_STEP step=1098 loss=2.368692 Training: 6%|▌ | 1098/18018 [2:25:13<34:11:31, 7.27s/step, loss=2.369, acc=0.545, acc_len=1.23, thru=16.9, I=21.9, T=31.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:27,919] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:55:34,727] loop.py:45 INFO TRAIN_STEP step=1099 loss=2.433038 Training: 6%|▌ | 1099/18018 [2:25:20<33:52:14, 7.21s/step, loss=2.433, acc=0.535, acc_len=1.20, thru=15.3, I=23.4, T=34.5, wait=3.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:41,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:55:42,828] loop.py:380 INFO TIMING step=1100: step=3.915s data=1.655s compute=3.876s [fwd=2.775s bwd=1.086s opt=0.003s] dispatch=4.177s [2026-10-18 15:55:42,828] loop.py:45 INFO TRAIN_STEP step=1100 loss=2.391183 Training: 6%|▌ | 1100/18018 [2:25:28<35:07:45, 7.48s/step, loss=2.391, acc=0.533, acc_len=1.20, thru=16.0, I=18.6, T=32.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:43,085] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:47,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:55:48,844] loop.py:45 INFO TRAIN_STEP step=1101 loss=2.461900 Training: 6%|▌ | 1101/18018 [2:25:34<33:04:11, 7.04s/step, loss=2.462, acc=0.532, acc_len=1.19, thru=19.7, I=25.8, T=35.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:48,900] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (TrainerActor pid=9310) [2026-10-18 15:55:48,839] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1100: forward=2565.7ms backward=1050.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:54,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:55:55,187] loop.py:45 INFO TRAIN_STEP step=1102 loss=2.346081 Training: 6%|▌ | 1102/18018 [2:25:41<32:05:11, 6.83s/step, loss=2.346, acc=0.542, acc_len=1.23, thru=19.3, I=22.4, T=32.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:55:55,688] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:56:01,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:56:02,361] loop.py:45 INFO TRAIN_STEP step=1103 loss=2.174783 Training: 6%|▌ | 1103/18018 [2:25:48<32:34:35, 6.93s/step, loss=2.175, acc=0.575, acc_len=1.31, thru=16.9, I=19.3, T=35.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:56:02,822] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:56:08,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:56:09,592] loop.py:45 INFO TRAIN_STEP step=1104 loss=2.479056 Training: 6%|▌ | 1104/18018 [2:25:55<32:59:41, 7.02s/step, loss=2.479, acc=0.528, acc_len=1.18, thru=18.8, I=22.2, T=34.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:56:09,680] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:56:15,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:56:16,303] loop.py:380 INFO TIMING step=1105: step=4.123s data=1.941s compute=4.093s [fwd=2.899s bwd=1.178s opt=0.003s] dispatch=2.577s [2026-10-18 15:56:16,303] loop.py:45 INFO TRAIN_STEP step=1105 loss=2.367884 Training: 6%|▌ | 1105/18018 [2:26:02<32:33:02, 6.93s/step, loss=2.368, acc=0.541, acc_len=1.22, thru=18.4, I=24.0, T=31.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:56:16,738] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:56:22,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:56:23,907] loop.py:45 INFO TRAIN_STEP step=1106 loss=2.537012 Training: 6%|▌ | 1106/18018 [2:26:09<33:30:03, 7.13s/step, loss=2.537, acc=0.525, acc_len=1.16, thru=16.1, I=22.5, T=29.8, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:56:23,902] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1105: forward=2985.4ms backward=1257.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:56:24,219] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 15:56:31,541] loop.py:45 INFO TRAIN_STEP step=1107 loss=2.540279 Training: 6%|▌ | 1107/18018 [2:26:17<34:12:18, 7.28s/step, loss=2.540, acc=0.515, acc_len=1.14, thru=14.9, I=16.7, T=31.7, wait=3.6s, pool=60, epoch=1/2][2026-10-18 15:56:38,495] loop.py:45 INFO TRAIN_STEP step=1108 loss=2.538327 Training: 6%|▌ | 1108/18018 [2:26:24<33:44:39, 7.18s/step, loss=2.538, acc=0.521, acc_len=1.16, thru=17.2, I=16.9, T=33.3, wait=3.1s, pool=52, epoch=1/2][2026-10-18 15:56:45,460] loop.py:45 INFO TRAIN_STEP step=1109 loss=2.623288 Training: 6%|▌ | 1109/18018 [2:26:31<33:26:00, 7.12s/step, loss=2.623, acc=0.505, acc_len=1.11, thru=17.9, I=17.3, T=32.7, wait=3.0s, pool=56, epoch=1/2][2026-10-18 15:56:53,431] loop.py:380 INFO TIMING step=1110: step=3.738s data=1.595s compute=3.701s [fwd=2.656s bwd=1.030s opt=0.003s] dispatch=4.223s [2026-10-18 15:56:53,431] loop.py:45 INFO TRAIN_STEP step=1110 loss=2.243315 Training: 6%|▌ | 1110/18018 [2:26:39<34:38:08, 7.37s/step, loss=2.243, acc=0.554, acc_len=1.26, thru=14.0, I=13.6, T=34.2, wait=4.2s, pool=40, epoch=1/2][2026-10-18 15:57:03,096] loop.py:45 INFO TRAIN_STEP step=1111 loss=2.547032 Training: 6%|▌ | 1111/18018 [2:26:49<37:51:24, 8.06s/step, loss=2.547, acc=0.518, acc_len=1.15, thru=15.4, I=14.4, T=30.4, wait=5.5s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:57:03,092] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1110: forward=3096.9ms backward=1055.5ms [2026-10-18 15:57:10,432] loop.py:45 INFO TRAIN_STEP step=1112 loss=2.363921 Training: 6%|▌ | 1112/18018 [2:26:56<36:49:56, 7.84s/step, loss=2.364, acc=0.539, acc_len=1.22, thru=15.8, I=15.8, T=29.2, wait=2.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:57:17,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:57:19,546] loop.py:45 INFO TRAIN_STEP step=1113 loss=2.567250 Training: 6%|▌ | 1113/18018 [2:27:05<38:37:33, 8.23s/step, loss=2.567, acc=0.516, acc_len=1.15, thru=15.8, I=17.3, T=29.9, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:57:19,599] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:57:23,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:57:25,392] loop.py:45 INFO TRAIN_STEP step=1114 loss=2.330384 Training: 6%|▌ | 1114/18018 [2:27:11<35:16:16, 7.51s/step, loss=2.330, acc=0.543, acc_len=1.23, thru=19.1, I=26.3, T=34.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:57:25,427] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 15:57:34,629] loop.py:380 INFO TIMING step=1115: step=4.439s data=2.168s compute=4.203s [fwd=2.927s bwd=1.260s opt=0.003s] dispatch=4.787s [2026-10-18 15:57:34,629] loop.py:45 INFO TRAIN_STEP step=1115 loss=2.313717 Training: 6%|▌ | 1115/18018 [2:27:20<37:41:48, 8.03s/step, loss=2.314, acc=0.550, acc_len=1.25, thru=13.5, I=13.3, T=28.8, wait=4.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:57:40,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:57:40,881] loop.py:45 INFO TRAIN_STEP step=1116 loss=2.368305 Training: 6%|▌ | 1116/18018 [2:27:26<35:11:43, 7.50s/step, loss=2.368, acc=0.541, acc_len=1.22, thru=19.4, I=17.9, T=32.3, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:57:40,874] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1115: forward=2728.2ms backward=1188.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:57:41,178] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:57:48,403] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:57:49,575] loop.py:45 INFO TRAIN_STEP step=1117 loss=2.575547 Training: 6%|▌ | 1117/18018 [2:27:35<36:52:42, 7.86s/step, loss=2.576, acc=0.512, acc_len=1.13, thru=13.8, I=16.3, T=31.9, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:57:49,914] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:57:56,705] loop.py:45 INFO TRAIN_STEP step=1118 loss=2.509015 Training: 6%|▌ | 1118/18018 [2:27:42<35:51:14, 7.64s/step, loss=2.509, acc=0.519, acc_len=1.16, thru=14.0, I=16.6, T=32.9, wait=3.2s, pool=32, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 15:57:56,784] eagle_store.py:517 WARNING batch_get_buffer missing keys (k35783-b610fa71_hs, k35783-b610fa71_ids, k35783-b610fa71_lhs); sleeping 0.50s. [2026-10-18 15:58:07,036] loop.py:45 INFO TRAIN_STEP step=1119 loss=2.282118 Training: 6%|▌ | 1119/18018 [2:27:53<39:38:37, 8.45s/step, loss=2.282, acc=0.552, acc_len=1.25, thru=10.4, I=17.2, T=30.8, wait=6.2s, pool=20, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:17,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 15:57:56,782] eagle_store.py:517 WARNING batch_get_buffer missing keys (k35782-8417787b_hs, k35782-8417787b_ids, k35782-8417787b_lhs); sleeping 0.50s. [2026-10-18 15:58:18,661] loop.py:380 INFO TIMING step=1120: step=4.163s data=1.928s compute=4.081s [fwd=2.760s bwd=1.306s opt=0.003s] dispatch=7.454s [2026-10-18 15:58:18,661] loop.py:45 INFO TRAIN_STEP step=1120 loss=2.534401 Training: 6%|▌ | 1120/18018 [2:28:04<44:07:15, 9.40s/step, loss=2.534, acc=0.522, acc_len=1.16, thru=16.7, I=14.0, T=30.8, wait=7.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:19,074] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:27,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:58:27,413] loop.py:45 INFO TRAIN_STEP step=1121 loss=2.499403 Training: 6%|▌ | 1121/18018 [2:28:13<43:12:14, 9.20s/step, loss=2.499, acc=0.520, acc_len=1.16, thru=15.3, I=13.6, T=28.8, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:58:27,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1120: forward=3069.3ms backward=1324.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:27,753] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:32,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:58:34,080] loop.py:45 INFO TRAIN_STEP step=1122 loss=2.424911 Training: 6%|▌ | 1122/18018 [2:28:20<39:37:41, 8.44s/step, loss=2.425, acc=0.529, acc_len=1.19, thru=19.0, I=19.6, T=34.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:34,380] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:58:42,194] loop.py:45 INFO TRAIN_STEP step=1123 loss=2.480075 Training: 6%|▌ | 1123/18018 [2:28:28<39:09:40, 8.34s/step, loss=2.480, acc=0.533, acc_len=1.18, thru=14.0, I=16.4, T=34.4, wait=4.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:49,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:58:49,814] loop.py:45 INFO TRAIN_STEP step=1124 loss=2.323793 Training: 6%|▌ | 1124/18018 [2:28:35<38:08:31, 8.13s/step, loss=2.324, acc=0.547, acc_len=1.24, thru=16.5, I=17.5, T=31.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:50,072] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:56,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:58:57,279] loop.py:380 INFO TIMING step=1125: step=4.166s data=1.804s compute=4.136s [fwd=2.888s bwd=1.234s opt=0.003s] dispatch=3.289s [2026-10-18 15:58:57,280] loop.py:45 INFO TRAIN_STEP step=1125 loss=2.448509 Training: 6%|▌ | 1125/18018 [2:28:43<37:12:26, 7.93s/step, loss=2.449, acc=0.529, acc_len=1.19, thru=16.8, I=19.8, T=30.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:58:57,628] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 15:59:03,870] loop.py:45 INFO TRAIN_STEP step=1126 loss=2.502671 Training: 6%|▌ | 1126/18018 [2:28:49<35:19:19, 7.53s/step, loss=2.503, acc=0.530, acc_len=1.19, thru=16.5, I=22.6, T=32.6, wait=2.7s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 15:59:03,866] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1125: forward=2670.0ms backward=1194.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:59:12,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:59:13,483] loop.py:45 INFO TRAIN_STEP step=1127 loss=2.547072 Training: 6%|▋ | 1127/18018 [2:28:59<38:15:23, 8.15s/step, loss=2.547, acc=0.513, acc_len=1.14, thru=16.1, I=15.2, T=29.0, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:59:13,593] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:59:18,171] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:59:19,841] loop.py:45 INFO TRAIN_STEP step=1128 loss=2.680627 Training: 6%|▋ | 1128/18018 [2:29:05<35:43:29, 7.61s/step, loss=2.681, acc=0.500, acc_len=1.09, thru=17.3, I=23.6, T=32.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:59:20,183] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:59:25,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:59:27,420] loop.py:45 INFO TRAIN_STEP step=1129 loss=2.465142 Training: 6%|▋ | 1129/18018 [2:29:13<35:40:24, 7.60s/step, loss=2.465, acc=0.533, acc_len=1.19, thru=14.9, I=20.2, T=33.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:59:27,715] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:59:32,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 15:59:34,149] loop.py:380 INFO TIMING step=1130: step=3.651s data=1.555s compute=3.611s [fwd=2.518s bwd=1.080s opt=0.003s] dispatch=3.067s [2026-10-18 15:59:34,149] loop.py:45 INFO TRAIN_STEP step=1130 loss=2.329456 Training: 6%|▋ | 1130/18018 [2:29:20<34:26:19, 7.34s/step, loss=2.329, acc=0.544, acc_len=1.23, thru=18.1, I=24.8, T=35.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 15:59:34,246] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 15:59:41,036] loop.py:45 INFO TRAIN_STEP step=1131 loss=2.403839 (TrainerActor pid=9310) [2026-10-18 15:59:41,031] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1130: forward=2597.2ms backward=1077.4ms Training: 6%|▋ | 1131/18018 [2:29:27<33:47:56, 7.21s/step, loss=2.404, acc=0.538, acc_len=1.21, thru=16.6, I=20.8, T=34.4, wait=3.2s, pool=52, epoch=1/2][2026-10-18 15:59:49,065] loop.py:45 INFO TRAIN_STEP step=1132 loss=2.415547 Training: 6%|▋ | 1132/18018 [2:29:35<34:57:17, 7.45s/step, loss=2.416, acc=0.535, acc_len=1.19, thru=16.1, I=16.0, T=30.5, wait=3.8s, pool=60, epoch=1/2][2026-10-18 15:59:57,425] loop.py:45 INFO TRAIN_STEP step=1133 loss=2.501530 Training: 6%|▋ | 1133/18018 [2:29:43<36:13:59, 7.73s/step, loss=2.502, acc=0.527, acc_len=1.18, thru=14.8, I=15.6, T=31.7, wait=4.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:03,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:00:05,720] loop.py:45 INFO TRAIN_STEP step=1134 loss=2.361560 Training: 6%|▋ | 1134/18018 [2:29:51<37:01:42, 7.90s/step, loss=2.362, acc=0.539, acc_len=1.21, thru=17.4, I=19.0, T=33.8, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:05,870] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-18 16:00:12,456] loop.py:380 INFO TIMING step=1135: step=3.930s data=1.771s compute=3.896s [fwd=2.767s bwd=1.117s opt=0.003s] dispatch=2.797s [2026-10-18 16:00:12,457] loop.py:45 INFO TRAIN_STEP step=1135 loss=2.307330 Training: 6%|▋ | 1135/18018 [2:29:58<35:23:51, 7.55s/step, loss=2.307, acc=0.544, acc_len=1.23, thru=16.6, I=20.8, T=32.6, wait=2.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:19,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:00:19,879] loop.py:45 INFO TRAIN_STEP step=1136 loss=2.415381 Training: 6%|▋ | 1136/18018 [2:30:05<35:13:05, 7.51s/step, loss=2.415, acc=0.539, acc_len=1.21, thru=18.7, I=18.0, T=30.1, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:00:19,875] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1135: forward=2978.1ms backward=1235.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:20,386] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:26,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:00:27,850] loop.py:45 INFO TRAIN_STEP step=1137 loss=2.474922 Training: 6%|▋ | 1137/18018 [2:30:13<35:51:54, 7.65s/step, loss=2.475, acc=0.526, acc_len=1.18, thru=15.9, I=18.3, T=30.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:27,999] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:34,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:00:34,437] loop.py:45 INFO TRAIN_STEP step=1138 loss=2.645791 Training: 6%|▋ | 1138/18018 [2:30:20<34:22:00, 7.33s/step, loss=2.646, acc=0.513, acc_len=1.12, thru=18.1, I=20.6, T=30.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:34,551] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:38,955] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:00:41,878] loop.py:45 INFO TRAIN_STEP step=1139 loss=2.324056 Training: 6%|▋ | 1139/18018 [2:30:27<34:31:26, 7.36s/step, loss=2.324, acc=0.548, acc_len=1.24, thru=16.6, I=23.0, T=25.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:41,975] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:48,999] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:00:49,306] loop.py:380 INFO TIMING step=1140: step=4.202s data=1.935s compute=4.160s [fwd=2.951s bwd=1.196s opt=0.003s] dispatch=3.216s [2026-10-18 16:00:49,306] loop.py:45 INFO TRAIN_STEP step=1140 loss=2.405325 Training: 6%|▋ | 1140/18018 [2:30:35<34:36:47, 7.38s/step, loss=2.405, acc=0.539, acc_len=1.21, thru=17.5, I=18.5, T=30.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:00:49,502] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 16:00:56,889] loop.py:45 INFO TRAIN_STEP step=1141 loss=2.480601 Training: 6%|▋ | 1141/18018 [2:30:42<34:53:23, 7.44s/step, loss=2.481, acc=0.524, acc_len=1.17, thru=16.2, I=15.5, T=30.4, wait=3.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:00:56,884] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1140: forward=2961.9ms backward=1199.2ms [2026-10-18 16:01:04,276] loop.py:45 INFO TRAIN_STEP step=1142 loss=2.439536 Training: 6%|▋ | 1142/18018 [2:30:50<34:49:00, 7.43s/step, loss=2.440, acc=0.528, acc_len=1.18, thru=16.1, I=15.7, T=29.2, wait=3.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:09,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:01:11,958] loop.py:45 INFO TRAIN_STEP step=1143 loss=2.268568 Training: 6%|▋ | 1143/18018 [2:30:58<35:10:01, 7.50s/step, loss=2.269, acc=0.559, acc_len=1.27, thru=17.2, I=20.8, T=26.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:12,310] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:17,114] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:01:18,022] loop.py:45 INFO TRAIN_STEP step=1144 loss=2.221563 Training: 6%|▋ | 1144/18018 [2:31:04<33:08:40, 7.07s/step, loss=2.222, acc=0.570, acc_len=1.29, thru=17.8, I=28.5, T=34.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:18,122] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:24,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:01:25,586] loop.py:380 INFO TIMING step=1145: step=3.910s data=1.714s compute=3.873s [fwd=2.768s bwd=1.092s opt=0.003s] dispatch=3.644s [2026-10-18 16:01:25,586] loop.py:45 INFO TRAIN_STEP step=1145 loss=2.576708 Training: 6%|▋ | 1145/18018 [2:31:11<33:49:58, 7.22s/step, loss=2.577, acc=0.511, acc_len=1.13, thru=16.0, I=19.4, T=32.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:25,637] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:34,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:01:34,521] loop.py:45 INFO TRAIN_STEP step=1146 loss=2.434162 Training: 6%|▋ | 1146/18018 [2:31:20<36:14:38, 7.73s/step, loss=2.434, acc=0.536, acc_len=1.19, thru=13.3, I=15.0, T=29.5, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:01:34,517] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1145: forward=2974.6ms backward=1310.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:34,893] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 16:01:41,479] loop.py:45 INFO TRAIN_STEP step=1147 loss=2.347873 Training: 6%|▋ | 1147/18018 [2:31:27<35:09:06, 7.50s/step, loss=2.348, acc=0.544, acc_len=1.23, thru=16.8, I=14.9, T=35.2, wait=3.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:49,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:01:50,917] loop.py:45 INFO TRAIN_STEP step=1148 loss=2.359364 Training: 6%|▋ | 1148/18018 [2:31:37<37:52:36, 8.08s/step, loss=2.359, acc=0.544, acc_len=1.23, thru=15.8, I=17.1, T=30.8, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:50,955] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:56,954] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:01:57,947] loop.py:45 INFO TRAIN_STEP step=1149 loss=2.400057 Training: 6%|▋ | 1149/18018 [2:31:44<36:23:36, 7.77s/step, loss=2.400, acc=0.537, acc_len=1.20, thru=17.3, I=21.1, T=34.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:01:57,980] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 16:02:05,852] loop.py:380 INFO TIMING step=1150: step=4.505s data=2.091s compute=4.469s [fwd=3.129s bwd=1.327s opt=0.003s] dispatch=3.390s [2026-10-18 16:02:05,852] loop.py:45 INFO TRAIN_STEP step=1150 loss=2.416078 Training: 6%|▋ | 1150/18018 [2:31:51<36:34:59, 7.81s/step, loss=2.416, acc=0.533, acc_len=1.19, thru=15.0, I=16.3, T=28.4, wait=3.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:02:12,598] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1150: forward=2631.7ms backward=1191.6ms [2026-10-18 16:02:12,603] loop.py:45 INFO TRAIN_STEP step=1151 loss=2.418756 Training: 6%|▋ | 1151/18018 [2:31:58<35:05:54, 7.49s/step, loss=2.419, acc=0.532, acc_len=1.19, thru=15.7, I=17.3, T=33.0, wait=2.9s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:02:18,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:02:21,227] loop.py:45 INFO TRAIN_STEP step=1152 loss=2.439738 Training: 6%|▋ | 1152/18018 [2:32:07<36:41:17, 7.83s/step, loss=2.440, acc=0.535, acc_len=1.19, thru=17.0, I=19.9, T=29.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:02:21,549] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:02:26,708] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:02:28,205] loop.py:45 INFO TRAIN_STEP step=1153 loss=2.371615 Training: 6%|▋ | 1153/18018 [2:32:14<35:29:15, 7.58s/step, loss=2.372, acc=0.544, acc_len=1.22, thru=14.6, I=26.3, T=32.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:02:28,241] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:02:33,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:02:34,084] loop.py:45 INFO TRAIN_STEP step=1154 loss=2.388415 Training: 6%|▋ | 1154/18018 [2:32:20<33:06:06, 7.07s/step, loss=2.388, acc=0.539, acc_len=1.21, thru=20.6, I=27.5, T=33.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:02:34,575] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:02:41,135] loop.py:380 INFO TIMING step=1155: step=3.843s data=1.626s compute=3.815s [fwd=2.734s bwd=1.068s opt=0.003s] dispatch=3.198s [2026-10-18 16:02:41,135] loop.py:45 INFO TRAIN_STEP step=1155 loss=2.424883 Training: 6%|▋ | 1155/18018 [2:32:27<33:04:37, 7.06s/step, loss=2.425, acc=0.535, acc_len=1.20, thru=15.6, I=18.0, T=33.3, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:02:47,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:02:48,784] loop.py:45 INFO TRAIN_STEP step=1156 loss=2.350498 Training: 6%|▋ | 1156/18018 [2:32:34<33:54:00, 7.24s/step, loss=2.350, acc=0.545, acc_len=1.23, thru=18.2, I=19.1, T=32.7, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:02:48,778] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1155: forward=2726.9ms backward=1134.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:02:49,041] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:02:55,712] loop.py:45 INFO TRAIN_STEP step=1157 loss=2.250113 Training: 6%|▋ | 1157/18018 [2:32:41<33:27:41, 7.14s/step, loss=2.250, acc=0.559, acc_len=1.27, thru=16.6, I=20.1, T=29.8, wait=2.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:02,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:03:03,503] loop.py:45 INFO TRAIN_STEP step=1158 loss=2.176836 Training: 6%|▋ | 1158/18018 [2:32:49<34:22:08, 7.34s/step, loss=2.177, acc=0.573, acc_len=1.31, thru=17.8, I=17.4, T=32.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:03,706] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:10,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:03:10,863] loop.py:45 INFO TRAIN_STEP step=1159 loss=2.501559 Training: 6%|▋ | 1159/18018 [2:32:56<34:23:52, 7.35s/step, loss=2.502, acc=0.528, acc_len=1.17, thru=15.6, I=17.5, T=33.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:11,318] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:16,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:03:17,110] loop.py:380 INFO TIMING step=1160: step=3.860s data=1.754s compute=3.808s [fwd=2.683s bwd=1.111s opt=0.003s] dispatch=2.378s [2026-10-18 16:03:17,110] loop.py:45 INFO TRAIN_STEP step=1160 loss=2.459041 Training: 6%|▋ | 1160/18018 [2:33:03<32:51:08, 7.02s/step, loss=2.459, acc=0.531, acc_len=1.18, thru=19.7, I=22.0, T=33.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:17,174] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:23,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:03:25,225] loop.py:45 INFO TRAIN_STEP step=1161 loss=2.367942 Training: 6%|▋ | 1161/18018 [2:33:11<34:23:32, 7.34s/step, loss=2.368, acc=0.543, acc_len=1.22, thru=14.5, I=17.5, T=34.8, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:03:25,220] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1160: forward=2614.3ms backward=1014.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:25,319] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:31,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:03:32,951] loop.py:45 INFO TRAIN_STEP step=1162 loss=2.320926 Training: 6%|▋ | 1162/18018 [2:33:19<34:55:47, 7.46s/step, loss=2.321, acc=0.549, acc_len=1.24, thru=15.8, I=19.9, T=32.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:33,242] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:40,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:03:41,164] loop.py:45 INFO TRAIN_STEP step=1163 loss=2.350310 Training: 6%|▋ | 1163/18018 [2:33:27<35:59:09, 7.69s/step, loss=2.350, acc=0.546, acc_len=1.23, thru=14.5, I=16.1, T=32.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:41,255] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 [2026-10-18 16:03:48,810] loop.py:45 INFO TRAIN_STEP step=1164 loss=2.258007 Training: 6%|▋ | 1164/18018 [2:33:34<35:55:40, 7.67s/step, loss=2.258, acc=0.556, acc_len=1.26, thru=14.5, I=16.0, T=32.7, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:55,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:03:56,375] loop.py:380 INFO TIMING step=1165: step=4.216s data=1.819s compute=4.178s [fwd=2.907s bwd=1.257s opt=0.003s] dispatch=3.338s [2026-10-18 16:03:56,375] loop.py:45 INFO TRAIN_STEP step=1165 loss=2.488694 Training: 6%|▋ | 1165/18018 [2:33:42<35:46:20, 7.64s/step, loss=2.489, acc=0.519, acc_len=1.16, thru=17.4, I=17.9, T=30.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:03:56,459] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-18 16:04:04,020] loop.py:45 INFO TRAIN_STEP step=1166 loss=2.397217 Training: 6%|▋ | 1166/18018 [2:33:50<35:46:13, 7.64s/step, loss=2.397, acc=0.531, acc_len=1.19, thru=13.2, I=15.0, T=34.4, wait=3.9s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:04:04,015] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1165: forward=2621.4ms backward=1045.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:11,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:04:13,169] loop.py:45 INFO TRAIN_STEP step=1167 loss=2.323111 Training: 6%|▋ | 1167/18018 [2:33:59<37:53:10, 8.09s/step, loss=2.323, acc=0.553, acc_len=1.26, thru=17.2, I=17.1, T=31.4, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:13,512] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:04:18,946] loop.py:45 INFO TRAIN_STEP step=1168 loss=2.394666 Training: 6%|▋ | 1168/18018 [2:34:05<34:37:55, 7.40s/step, loss=2.395, acc=0.547, acc_len=1.23, thru=19.4, I=24.4, T=35.8, wait=2.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:23,929] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:04:25,582] loop.py:45 INFO TRAIN_STEP step=1169 loss=2.149887 Training: 6%|▋ | 1169/18018 [2:34:11<33:33:26, 7.17s/step, loss=2.150, acc=0.570, acc_len=1.30, thru=18.9, I=22.8, T=33.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:25,939] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:04:31,691] loop.py:380 INFO TIMING step=1170: step=3.490s data=1.456s compute=3.453s [fwd=2.452s bwd=0.987s opt=0.003s] dispatch=2.611s [2026-10-18 16:04:31,691] loop.py:45 INFO TRAIN_STEP step=1170 loss=2.544866 Training: 6%|▋ | 1170/18018 [2:34:17<32:04:13, 6.85s/step, loss=2.545, acc=0.518, acc_len=1.15, thru=19.3, I=24.6, T=36.7, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:38,441] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:04:40,563] loop.py:45 INFO TRAIN_STEP step=1171 loss=2.257209 Training: 6%|▋ | 1171/18018 [2:34:26<34:54:02, 7.46s/step, loss=2.257, acc=0.559, acc_len=1.26, thru=14.9, I=17.3, T=30.6, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:04:40,558] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1170: forward=2971.5ms backward=1169.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:40,961] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:46,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:04:47,748] loop.py:45 INFO TRAIN_STEP step=1172 loss=2.338075 Training: 7%|▋ | 1172/18018 [2:34:33<34:30:57, 7.38s/step, loss=2.338, acc=0.543, acc_len=1.23, thru=16.1, I=23.2, T=31.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:47,822] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 16:04:53,752] loop.py:45 INFO TRAIN_STEP step=1173 loss=2.505489 Training: 7%|▋ | 1173/18018 [2:34:39<32:35:04, 6.96s/step, loss=2.505, acc=0.525, acc_len=1.17, thru=19.5, I=22.1, T=31.0, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:53,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:54,225] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:04:59,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:05:00,550] loop.py:45 INFO TRAIN_STEP step=1174 loss=2.383588 Training: 7%|▋ | 1174/18018 [2:34:46<32:21:00, 6.91s/step, loss=2.384, acc=0.533, acc_len=1.19, thru=18.2, I=19.3, T=32.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:05:00,836] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:05:05,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:05:06,692] loop.py:380 INFO TIMING step=1175: step=3.929s data=1.766s compute=3.903s [fwd=2.772s bwd=1.116s opt=0.003s] dispatch=2.206s [2026-10-18 16:05:06,693] loop.py:45 INFO TRAIN_STEP step=1175 loss=2.160296 Training: 7%|▋ | 1175/18018 [2:34:52<31:16:08, 6.68s/step, loss=2.160, acc=0.569, acc_len=1.30, thru=19.3, I=23.0, T=32.6, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:05:06,826] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 16:05:15,381] loop.py:45 INFO TRAIN_STEP step=1176 loss=2.422914 Training: 7%|▋ | 1176/18018 [2:35:01<34:04:46, 7.28s/step, loss=2.423, acc=0.536, acc_len=1.20, thru=12.9, I=15.4, T=33.4, wait=4.8s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:05:15,376] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1175: forward=2573.0ms backward=1122.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:05:23,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:05:24,853] loop.py:45 INFO TRAIN_STEP step=1177 loss=2.442953 Training: 7%|▋ | 1177/18018 [2:35:10<37:09:00, 7.94s/step, loss=2.443, acc=0.522, acc_len=1.17, thru=15.2, I=15.7, T=32.2, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:05:24,948] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 16:05:32,014] loop.py:45 INFO TRAIN_STEP step=1178 loss=2.339204 Training: 7%|▋ | 1178/18018 [2:35:18<36:02:51, 7.71s/step, loss=2.339, acc=0.547, acc_len=1.23, thru=17.5, I=17.7, T=30.4, wait=2.9s, pool=60, epoch=1/2][2026-10-18 16:05:39,449] loop.py:45 INFO TRAIN_STEP step=1179 loss=2.396060 Training: 7%|▋ | 1179/18018 [2:35:25<35:39:53, 7.62s/step, loss=2.396, acc=0.535, acc_len=1.20, thru=16.2, I=14.6, T=30.6, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:05:46,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:05:47,140] loop.py:380 INFO TIMING step=1180: step=3.747s data=1.592s compute=3.707s [fwd=2.654s bwd=1.039s opt=0.003s] dispatch=3.937s [2026-10-18 16:05:47,140] loop.py:45 INFO TRAIN_STEP step=1180 loss=2.495531 Training: 7%|▋ | 1180/18018 [2:35:33<35:45:29, 7.65s/step, loss=2.496, acc=0.522, acc_len=1.16, thru=18.4, I=17.2, T=34.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:05:47,291] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:05:55,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:05:55,654] loop.py:45 INFO TRAIN_STEP step=1181 loss=2.454834 Training: 7%|▋ | 1181/18018 [2:35:41<36:58:37, 7.91s/step, loss=2.455, acc=0.530, acc_len=1.19, thru=14.3, I=15.8, T=34.4, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:05:55,650] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1180: forward=2597.6ms backward=1078.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:05:55,822] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:02,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:06:03,455] loop.py:45 INFO TRAIN_STEP step=1182 loss=2.041920 Training: 7%|▋ | 1182/18018 [2:35:49<36:49:40, 7.87s/step, loss=2.042, acc=0.585, acc_len=1.36, thru=16.7, I=19.4, T=32.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:03,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-18 16:06:09,702] loop.py:45 INFO TRAIN_STEP step=1183 loss=2.234478 Training: 7%|▋ | 1183/18018 [2:35:55<34:32:20, 7.39s/step, loss=2.234, acc=0.559, acc_len=1.27, thru=18.5, I=20.5, T=31.6, wait=2.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:15,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:06:16,188] loop.py:45 INFO TRAIN_STEP step=1184 loss=2.366189 Training: 7%|▋ | 1184/18018 [2:36:02<33:16:37, 7.12s/step, loss=2.366, acc=0.545, acc_len=1.22, thru=19.7, I=20.1, T=33.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:16,265] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:23,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:06:24,987] loop.py:380 INFO TIMING step=1185: step=4.243s data=2.014s compute=4.211s [fwd=2.965s bwd=1.231s opt=0.003s] dispatch=4.546s [2026-10-18 16:06:24,987] loop.py:45 INFO TRAIN_STEP step=1185 loss=2.418720 Training: 7%|▋ | 1185/18018 [2:36:11<35:37:58, 7.62s/step, loss=2.419, acc=0.537, acc_len=1.20, thru=13.8, I=17.4, T=30.2, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:25,282] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:31,258] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:06:32,208] loop.py:45 INFO TRAIN_STEP step=1186 loss=2.333268 Training: 7%|▋ | 1186/18018 [2:36:18<35:04:11, 7.50s/step, loss=2.333, acc=0.542, acc_len=1.22, thru=17.2, I=21.8, T=29.9, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:06:32,203] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1185: forward=3043.5ms backward=1187.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:32,265] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:38,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:06:38,456] loop.py:45 INFO TRAIN_STEP step=1187 loss=2.287226 Training: 7%|▋ | 1187/18018 [2:36:24<33:18:46, 7.13s/step, loss=2.287, acc=0.553, acc_len=1.25, thru=18.1, I=20.3, T=31.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:38,598] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:44,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:06:45,377] loop.py:45 INFO TRAIN_STEP step=1188 loss=2.596331 Training: 7%|▋ | 1188/18018 [2:36:31<33:01:17, 7.06s/step, loss=2.596, acc=0.509, acc_len=1.12, thru=18.5, I=19.2, T=34.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:45,733] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:54,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:06:54,498] loop.py:45 INFO TRAIN_STEP step=1189 loss=2.391055 Training: 7%|▋ | 1189/18018 [2:36:40<35:54:22, 7.68s/step, loss=2.391, acc=0.534, acc_len=1.20, thru=14.2, I=14.1, T=31.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:06:54,844] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:01,401] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:07:02,323] loop.py:380 INFO TIMING step=1190: step=3.991s data=1.781s compute=3.952s [fwd=2.814s bwd=1.122s opt=0.003s] dispatch=3.825s [2026-10-18 16:07:02,323] loop.py:45 INFO TRAIN_STEP step=1190 loss=2.240449 Training: 7%|▋ | 1190/18018 [2:36:48<36:06:30, 7.72s/step, loss=2.240, acc=0.562, acc_len=1.28, thru=16.5, I=18.4, T=32.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:02,411] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:09,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:07:10,171] loop.py:45 INFO TRAIN_STEP step=1191 loss=2.418354 Training: 7%|▋ | 1191/18018 [2:36:56<36:16:45, 7.76s/step, loss=2.418, acc=0.532, acc_len=1.19, thru=16.0, I=18.2, T=30.5, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:07:10,167] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1190: forward=2916.8ms backward=1231.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:10,324] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-18 16:07:17,014] loop.py:45 INFO TRAIN_STEP step=1192 loss=2.423936 Training: 7%|▋ | 1192/18018 [2:37:03<34:59:16, 7.49s/step, loss=2.424, acc=0.533, acc_len=1.19, thru=16.9, I=18.1, T=33.4, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:22,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:07:23,737] loop.py:45 INFO TRAIN_STEP step=1193 loss=2.384062 Training: 7%|▋ | 1193/18018 [2:37:09<33:55:02, 7.26s/step, loss=2.384, acc=0.541, acc_len=1.21, thru=19.4, I=20.3, T=34.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:23,886] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:29,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:07:31,508] loop.py:45 INFO TRAIN_STEP step=1194 loss=2.469826 Training: 7%|▋ | 1194/18018 [2:37:17<34:38:06, 7.41s/step, loss=2.470, acc=0.527, acc_len=1.17, thru=15.4, I=23.0, T=28.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:31,713] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:38,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:07:40,049] loop.py:380 INFO TIMING step=1195: step=4.696s data=2.586s compute=4.662s [fwd=3.632s bwd=1.014s opt=0.003s] dispatch=3.834s [2026-10-18 16:07:40,049] loop.py:45 INFO TRAIN_STEP step=1195 loss=2.488004 Training: 7%|▋ | 1195/18018 [2:37:26<36:13:00, 7.75s/step, loss=2.488, acc=0.525, acc_len=1.17, thru=15.3, I=17.8, T=27.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:40,291] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:07:47,412] loop.py:45 INFO TRAIN_STEP step=1196 loss=2.367081 Training: 7%|▋ | 1196/18018 [2:37:33<35:40:07, 7.63s/step, loss=2.367, acc=0.540, acc_len=1.21, thru=13.7, I=16.6, T=33.6, wait=3.5s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:07:47,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1195: forward=2756.2ms backward=1014.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:55,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:07:55,783] loop.py:45 INFO TRAIN_STEP step=1197 loss=2.574745 Training: 7%|▋ | 1197/18018 [2:37:41<36:42:06, 7.85s/step, loss=2.575, acc=0.516, acc_len=1.14, thru=16.9, I=17.1, T=30.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:07:56,032] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:02,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:08:03,579] loop.py:45 INFO TRAIN_STEP step=1198 loss=2.532616 Training: 7%|▋ | 1198/18018 [2:37:49<36:37:13, 7.84s/step, loss=2.533, acc=0.515, acc_len=1.15, thru=16.5, I=20.4, T=32.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:03,735] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:08,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:08:09,956] loop.py:45 INFO TRAIN_STEP step=1199 loss=2.252360 Training: 7%|▋ | 1199/18018 [2:37:56<34:34:12, 7.40s/step, loss=2.252, acc=0.554, acc_len=1.26, thru=18.9, I=24.9, T=33.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:10,280] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:08:18,321] loop.py:380 INFO TIMING step=1200: step=3.679s data=1.509s compute=3.650s [fwd=2.583s bwd=1.052s opt=0.003s] dispatch=4.675s [2026-10-18 16:08:18,321] loop.py:45 INFO TRAIN_STEP step=1200 loss=2.273844 Training: 7%|▋ | 1200/18018 [2:38:04<35:55:10, 7.69s/step, loss=2.274, acc=0.551, acc_len=1.25, thru=14.4, I=16.7, T=34.8, wait=4.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:24,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:08:25,295] loop.py:45 INFO TRAIN_STEP step=1201 loss=2.374575 Training: 7%|▋ | 1201/18018 [2:38:11<34:54:50, 7.47s/step, loss=2.375, acc=0.541, acc_len=1.20, thru=18.1, I=20.3, T=34.0, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:08:25,291] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1200: forward=2668.7ms backward=1043.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:25,707] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:08:32,336] loop.py:45 INFO TRAIN_STEP step=1202 loss=2.294137 Training: 7%|▋ | 1202/18018 [2:38:18<34:18:25, 7.34s/step, loss=2.294, acc=0.554, acc_len=1.25, thru=15.3, I=18.1, T=32.5, wait=3.1s, pool=40, epoch=1/2][2026-10-18 16:08:41,951] loop.py:45 INFO TRAIN_STEP step=1203 loss=2.417785 Training: 7%|▋ | 1203/18018 [2:38:28<37:29:12, 8.03s/step, loss=2.418, acc=0.540, acc_len=1.21, thru=15.8, I=14.5, T=29.4, wait=5.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:48,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:08:50,140] loop.py:45 INFO TRAIN_STEP step=1204 loss=2.241380 Training: 7%|▋ | 1204/18018 [2:38:36<37:42:51, 8.07s/step, loss=2.241, acc=0.565, acc_len=1.28, thru=15.5, I=17.6, T=28.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:50,487] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:56,750] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:08:58,177] loop.py:380 INFO TIMING step=1205: step=4.068s data=1.820s compute=4.020s [fwd=2.850s bwd=1.154s opt=0.003s] dispatch=3.959s [2026-10-18 16:08:58,177] loop.py:45 INFO TRAIN_STEP step=1205 loss=2.319046 Training: 7%|▋ | 1205/18018 [2:38:44<37:39:23, 8.06s/step, loss=2.319, acc=0.552, acc_len=1.24, thru=14.0, I=20.8, T=31.5, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:08:58,263] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:04,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:09:05,287] loop.py:45 INFO TRAIN_STEP step=1206 loss=2.309876 Training: 7%|▋ | 1206/18018 [2:38:51<36:19:12, 7.78s/step, loss=2.310, acc=0.551, acc_len=1.25, thru=16.4, I=19.2, T=29.5, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:09:05,283] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1205: forward=3065.1ms backward=1232.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:05,414] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-18 16:09:11,834] loop.py:45 INFO TRAIN_STEP step=1207 loss=2.389486 Training: 7%|▋ | 1207/18018 [2:38:57<34:35:42, 7.41s/step, loss=2.389, acc=0.538, acc_len=1.21, thru=18.6, I=21.5, T=31.1, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:20,192] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:09:21,123] loop.py:45 INFO TRAIN_STEP step=1208 loss=2.441948 Training: 7%|▋ | 1208/18018 [2:39:07<37:13:30, 7.97s/step, loss=2.442, acc=0.529, acc_len=1.18, thru=14.4, I=14.0, T=31.5, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:21,200] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:27,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:09:28,492] loop.py:45 INFO TRAIN_STEP step=1209 loss=2.453584 Training: 7%|▋ | 1209/18018 [2:39:14<36:22:52, 7.79s/step, loss=2.454, acc=0.530, acc_len=1.18, thru=16.9, I=20.9, T=30.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:28,865] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:35,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:09:36,785] loop.py:380 INFO TIMING step=1210: step=4.692s data=2.202s compute=4.662s [fwd=3.309s bwd=1.337s opt=0.003s] dispatch=3.589s [2026-10-18 16:09:36,785] loop.py:45 INFO TRAIN_STEP step=1210 loss=2.271176 Training: 7%|▋ | 1210/18018 [2:39:22<37:04:51, 7.94s/step, loss=2.271, acc=0.554, acc_len=1.26, thru=14.1, I=17.1, T=27.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:37,048] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:42,535] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:09:43,883] loop.py:45 INFO TRAIN_STEP step=1211 loss=2.524737 Training: 7%|▋ | 1211/18018 [2:39:29<35:53:46, 7.69s/step, loss=2.525, acc=0.521, acc_len=1.16, thru=16.2, I=20.6, T=29.1, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:09:43,878] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1210: forward=3303.0ms backward=1047.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:44,045] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:49,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:09:50,303] loop.py:45 INFO TRAIN_STEP step=1212 loss=2.395815 Training: 7%|▋ | 1212/18018 [2:39:36<34:06:47, 7.31s/step, loss=2.396, acc=0.538, acc_len=1.21, thru=17.7, I=22.2, T=31.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:50,744] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:57,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:09:58,126] loop.py:45 INFO TRAIN_STEP step=1213 loss=2.370895 Training: 7%|▋ | 1213/18018 [2:39:44<34:50:03, 7.46s/step, loss=2.371, acc=0.542, acc_len=1.21, thru=15.6, I=16.8, T=34.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:09:58,467] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 16:10:04,069] loop.py:45 INFO TRAIN_STEP step=1214 loss=2.620733 Training: 7%|▋ | 1214/18018 [2:39:50<32:42:18, 7.01s/step, loss=2.621, acc=0.509, acc_len=1.12, thru=18.6, I=22.5, T=37.1, wait=2.5s, pool=48, epoch=1/2][2026-10-18 16:10:11,182] loop.py:380 INFO TIMING step=1215: step=3.452s data=1.451s compute=3.421s [fwd=2.424s bwd=0.983s opt=0.003s] dispatch=3.652s [2026-10-18 16:10:11,182] loop.py:45 INFO TRAIN_STEP step=1215 loss=2.404242 Training: 7%|▋ | 1215/18018 [2:39:57<32:51:34, 7.04s/step, loss=2.404, acc=0.534, acc_len=1.20, thru=16.4, I=17.2, T=37.1, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:17,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:10:19,082] loop.py:45 INFO TRAIN_STEP step=1216 loss=2.497753 Training: 7%|▋ | 1216/18018 [2:40:05<34:03:16, 7.30s/step, loss=2.498, acc=0.529, acc_len=1.17, thru=17.6, I=19.8, T=31.6, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:10:19,078] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1215: forward=2826.5ms backward=1172.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:19,285] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:10:25,158] loop.py:45 INFO TRAIN_STEP step=1217 loss=2.257653 Training: 7%|▋ | 1217/18018 [2:40:11<32:20:34, 6.93s/step, loss=2.258, acc=0.554, acc_len=1.25, thru=19.3, I=22.7, T=32.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:25,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:25,607] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:31,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:10:32,428] loop.py:45 INFO TRAIN_STEP step=1218 loss=2.514343 Training: 7%|▋ | 1218/18018 [2:40:18<32:49:13, 7.03s/step, loss=2.514, acc=0.518, acc_len=1.15, thru=18.2, I=19.7, T=31.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:32,687] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:40,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:10:42,022] loop.py:45 INFO TRAIN_STEP step=1219 loss=2.497952 Training: 7%|▋ | 1219/18018 [2:40:28<36:24:00, 7.80s/step, loss=2.498, acc=0.526, acc_len=1.17, thru=13.7, I=15.1, T=29.6, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:42,143] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:48,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:10:49,246] loop.py:380 INFO TIMING step=1220: step=3.942s data=1.816s compute=3.906s [fwd=2.797s bwd=1.093s opt=0.003s] dispatch=3.273s [2026-10-18 16:10:49,246] loop.py:45 INFO TRAIN_STEP step=1220 loss=2.289046 Training: 7%|▋ | 1220/18018 [2:40:35<35:35:34, 7.63s/step, loss=2.289, acc=0.557, acc_len=1.26, thru=14.8, I=19.1, T=32.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:49,306] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:56,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:10:57,120] loop.py:45 INFO TRAIN_STEP step=1221 loss=2.393373 Training: 7%|▋ | 1221/18018 [2:40:43<35:56:10, 7.70s/step, loss=2.393, acc=0.540, acc_len=1.21, thru=17.4, I=19.1, T=31.0, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:10:57,115] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1220: forward=2907.0ms backward=1181.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:10:57,390] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:03,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:11:04,508] loop.py:45 INFO TRAIN_STEP step=1222 loss=2.400457 Training: 7%|▋ | 1222/18018 [2:40:50<35:29:36, 7.61s/step, loss=2.400, acc=0.538, acc_len=1.21, thru=16.1, I=18.8, T=33.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:04,835] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:12,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:11:12,516] loop.py:45 INFO TRAIN_STEP step=1223 loss=2.407364 Training: 7%|▋ | 1223/18018 [2:40:58<36:03:03, 7.73s/step, loss=2.407, acc=0.534, acc_len=1.19, thru=16.0, I=18.0, T=29.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:12,929] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:18,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:11:20,132] loop.py:45 INFO TRAIN_STEP step=1224 loss=2.197345 Training: 7%|▋ | 1224/18018 [2:41:06<35:53:41, 7.69s/step, loss=2.197, acc=0.566, acc_len=1.29, thru=15.6, I=18.4, T=28.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:20,190] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:26,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:11:26,887] loop.py:380 INFO TIMING step=1225: step=4.053s data=1.851s compute=4.022s [fwd=2.859s bwd=1.148s opt=0.003s] dispatch=2.691s [2026-10-18 16:11:26,887] loop.py:45 INFO TRAIN_STEP step=1225 loss=2.452077 Training: 7%|▋ | 1225/18018 [2:41:12<34:34:28, 7.41s/step, loss=2.452, acc=0.525, acc_len=1.17, thru=16.9, I=20.6, T=31.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:27,033] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:33,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:11:35,569] loop.py:45 INFO TRAIN_STEP step=1226 loss=2.305741 Training: 7%|▋ | 1226/18018 [2:41:21<36:21:02, 7.79s/step, loss=2.306, acc=0.546, acc_len=1.23, thru=14.5, I=16.5, T=31.2, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:11:35,565] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1225: forward=2786.0ms backward=1259.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:35,623] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:40,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:11:41,727] loop.py:45 INFO TRAIN_STEP step=1227 loss=2.438915 Training: 7%|▋ | 1227/18018 [2:41:27<34:03:38, 7.30s/step, loss=2.439, acc=0.533, acc_len=1.19, thru=18.6, I=26.0, T=32.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:42,214] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:46,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:11:47,920] loop.py:45 INFO TRAIN_STEP step=1228 loss=2.351347 Training: 7%|▋ | 1228/18018 [2:41:34<32:30:31, 6.97s/step, loss=2.351, acc=0.544, acc_len=1.22, thru=19.3, I=26.2, T=36.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:48,003] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:53,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:11:54,315] loop.py:45 INFO TRAIN_STEP step=1229 loss=2.217544 Training: 7%|▋ | 1229/18018 [2:41:40<31:42:02, 6.80s/step, loss=2.218, acc=0.561, acc_len=1.27, thru=19.8, I=24.6, T=33.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:11:54,713] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:00,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:12:01,076] loop.py:380 INFO TIMING step=1230: step=3.902s data=1.750s compute=3.875s [fwd=2.781s bwd=1.079s opt=0.003s] dispatch=2.851s [2026-10-18 16:12:01,076] loop.py:45 INFO TRAIN_STEP step=1230 loss=2.279869 Training: 7%|▋ | 1230/18018 [2:41:47<31:38:46, 6.79s/step, loss=2.280, acc=0.560, acc_len=1.27, thru=17.2, I=23.1, T=32.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:01,254] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:08,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:12:08,901] loop.py:45 INFO TRAIN_STEP step=1231 loss=2.273603 Training: 7%|▋ | 1231/18018 [2:41:54<33:05:58, 7.10s/step, loss=2.274, acc=0.554, acc_len=1.25, thru=16.6, I=19.0, T=35.0, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:12:08,896] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1230: forward=2632.9ms backward=972.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:09,088] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:14,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:12:14,957] loop.py:45 INFO TRAIN_STEP step=1232 loss=2.290750 Training: 7%|▋ | 1232/18018 [2:42:01<31:38:17, 6.79s/step, loss=2.291, acc=0.549, acc_len=1.24, thru=20.1, I=21.6, T=32.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:15,047] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:22,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:12:22,633] loop.py:45 INFO TRAIN_STEP step=1233 loss=2.485908 Training: 7%|▋ | 1233/18018 [2:42:08<32:53:04, 7.05s/step, loss=2.486, acc=0.519, acc_len=1.16, thru=15.4, I=17.1, T=34.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:22,712] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-18 16:12:29,212] loop.py:45 INFO TRAIN_STEP step=1234 loss=2.171306 Training: 7%|▋ | 1234/18018 [2:42:15<32:12:57, 6.91s/step, loss=2.171, acc=0.566, acc_len=1.29, thru=19.0, I=18.5, T=29.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:29,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:29,707] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:35,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:12:36,397] loop.py:380 INFO TIMING step=1235: step=3.422s data=1.373s compute=3.391s [fwd=2.426s bwd=0.950s opt=0.003s] dispatch=3.754s [2026-10-18 16:12:36,397] loop.py:45 INFO TRAIN_STEP step=1235 loss=2.311677 Training: 7%|▋ | 1235/18018 [2:42:22<32:35:53, 6.99s/step, loss=2.312, acc=0.548, acc_len=1.23, thru=18.7, I=19.1, T=37.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:36,624] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:42,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:12:42,358] loop.py:45 INFO TRAIN_STEP step=1236 loss=2.177246 Training: 7%|▋ | 1236/18018 [2:42:28<31:09:25, 6.68s/step, loss=2.177, acc=0.567, acc_len=1.29, thru=20.7, I=23.7, T=37.5, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:12:42,354] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1235: forward=2412.3ms backward=956.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:42,557] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:48,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:12:49,291] loop.py:45 INFO TRAIN_STEP step=1237 loss=2.409621 Training: 7%|▋ | 1237/18018 [2:42:35<31:30:14, 6.76s/step, loss=2.410, acc=0.540, acc_len=1.22, thru=18.8, I=21.6, T=36.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:49,415] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:56,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:12:57,253] loop.py:45 INFO TRAIN_STEP step=1238 loss=2.396466 Training: 7%|▋ | 1238/18018 [2:42:43<33:10:54, 7.12s/step, loss=2.396, acc=0.537, acc_len=1.20, thru=16.1, I=18.5, T=32.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:12:57,738] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:13:06,064] loop.py:45 INFO TRAIN_STEP step=1239 loss=2.424767 Training: 7%|▋ | 1239/18018 [2:42:52<35:33:04, 7.63s/step, loss=2.425, acc=0.534, acc_len=1.20, thru=12.9, I=14.0, T=32.1, wait=4.8s, pool=52, epoch=1/2][2026-10-18 16:13:13,878] loop.py:380 INFO TIMING step=1240: step=3.659s data=1.492s compute=3.622s [fwd=2.576s bwd=1.031s opt=0.003s] dispatch=4.144s [2026-10-18 16:13:13,878] loop.py:45 INFO TRAIN_STEP step=1240 loss=2.615357 Training: 7%|▋ | 1240/18018 [2:42:59<35:48:31, 7.68s/step, loss=2.615, acc=0.513, acc_len=1.13, thru=17.7, I=16.4, T=35.0, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:13:20,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 16:13:20,875] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1240: forward=2753.4ms backward=1147.5ms [2026-10-18 16:13:20,879] loop.py:45 INFO TRAIN_STEP step=1241 loss=2.015572 Training: 7%|▋ | 1241/18018 [2:43:06<34:51:00, 7.48s/step, loss=2.016, acc=0.595, acc_len=1.37, thru=17.3, I=18.3, T=32.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:13:21,231] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:13:27,600] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:13:28,487] loop.py:45 INFO TRAIN_STEP step=1242 loss=2.235308 Training: 7%|▋ | 1242/18018 [2:43:14<35:01:55, 7.52s/step, loss=2.235, acc=0.560, acc_len=1.27, thru=16.8, I=17.9, T=33.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:13:28,615] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:13:35,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:13:36,062] loop.py:45 INFO TRAIN_STEP step=1243 loss=2.251060 Training: 7%|▋ | 1243/18018 [2:43:22<35:06:32, 7.53s/step, loss=2.251, acc=0.561, acc_len=1.27, thru=15.8, I=18.9, T=34.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:13:36,336] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:13:43,341] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:13:43,865] loop.py:45 INFO TRAIN_STEP step=1244 loss=2.393317 Training: 7%|▋ | 1244/18018 [2:43:29<35:28:52, 7.61s/step, loss=2.393, acc=0.536, acc_len=1.20, thru=15.7, I=18.2, T=32.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:13:44,345] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:13:52,084] loop.py:380 INFO TIMING step=1245: step=3.607s data=1.502s compute=3.577s [fwd=2.607s bwd=0.956s opt=0.003s] dispatch=4.603s [2026-10-18 16:13:52,085] loop.py:45 INFO TRAIN_STEP step=1245 loss=2.319944 Training: 7%|▋ | 1245/18018 [2:43:38<36:19:33, 7.80s/step, loss=2.320, acc=0.550, acc_len=1.25, thru=13.7, I=15.0, T=35.5, wait=4.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:13:59,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:13:59,985] loop.py:45 INFO TRAIN_STEP step=1246 loss=2.358165 Training: 7%|▋ | 1246/18018 [2:43:46<36:27:56, 7.83s/step, loss=2.358, acc=0.544, acc_len=1.22, thru=17.1, I=16.8, T=34.3, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:13:59,981] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1245: forward=2694.7ms backward=984.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:00,262] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:14:09,147] loop.py:45 INFO TRAIN_STEP step=1247 loss=2.312566 Training: 7%|▋ | 1247/18018 [2:43:55<38:19:41, 8.23s/step, loss=2.313, acc=0.552, acc_len=1.24, thru=12.9, I=14.3, T=32.0, wait=5.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:16,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:14:17,694] loop.py:45 INFO TRAIN_STEP step=1248 loss=2.175548 Training: 7%|▋ | 1248/18018 [2:44:03<38:46:21, 8.32s/step, loss=2.176, acc=0.571, acc_len=1.31, thru=15.4, I=17.5, T=27.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:17,761] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:24,955] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:14:26,286] loop.py:45 INFO TRAIN_STEP step=1249 loss=2.409681 Training: 7%|▋ | 1249/18018 [2:44:12<39:08:56, 8.40s/step, loss=2.410, acc=0.542, acc_len=1.21, thru=15.1, I=19.0, T=33.2, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:26,464] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:14:33,848] loop.py:380 INFO TIMING step=1250: step=3.727s data=1.554s compute=3.702s [fwd=2.681s bwd=1.006s opt=0.003s] dispatch=3.824s [2026-10-18 16:14:33,848] loop.py:45 INFO TRAIN_STEP step=1250 loss=2.214898 Training: 7%|▋ | 1250/18018 [2:44:19<37:58:05, 8.15s/step, loss=2.215, acc=0.561, acc_len=1.28, thru=15.0, I=18.0, T=34.3, wait=3.8s, pool=52, epoch=1/2][2026-10-18 16:14:33,852] loop.py:419 INFO Saving checkpoint at step 1250... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:34,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 16:14:38,186] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0001251 (TrainerActor pid=9310) [2026-10-18 16:14:38,205] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0000501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:39,254] inference_manager.py:408 INFO Pool capacity available after 4.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:47,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 16:14:38,187] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0001251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 16:14:38,996] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0000501 [2026-10-18 16:14:47,689] loop.py:45 INFO TRAIN_STEP step=1251 loss=2.556572 Training: 7%|▋ | 1251/18018 [2:44:33<45:54:52, 9.86s/step, loss=2.557, acc=0.517, acc_len=1.14, thru=15.2, I=16.3, T=29.5, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:14:47,685] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1250: forward=3041.3ms backward=1241.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:48,061] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:54,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 16:14:54,838] loop.py:45 INFO TRAIN_STEP step=1252 loss=2.322845 Training: 7%|▋ | 1252/18018 [2:44:40<42:07:36, 9.05s/step, loss=2.323, acc=0.546, acc_len=1.23, thru=18.1, I=18.1, T=35.1, wait=3.5s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:14:55,159] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:00,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:15:02,448] loop.py:45 INFO TRAIN_STEP step=1253 loss=2.314011 Training: 7%|▋ | 1253/18018 [2:44:48<40:06:57, 8.61s/step, loss=2.314, acc=0.539, acc_len=1.21, thru=18.2, I=20.6, T=30.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:02,928] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:09,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:15:09,935] loop.py:45 INFO TRAIN_STEP step=1254 loss=2.186338 Training: 7%|▋ | 1254/18018 [2:44:56<38:32:30, 8.28s/step, loss=2.186, acc=0.570, acc_len=1.30, thru=15.7, I=19.9, T=31.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:10,230] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:17,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:15:18,300] loop.py:380 INFO TIMING step=1255: step=4.175s data=1.975s compute=4.138s [fwd=2.945s bwd=1.179s opt=0.003s] dispatch=4.179s [2026-10-18 16:15:18,301] loop.py:45 INFO TRAIN_STEP step=1255 loss=2.114191 Training: 7%|▋ | 1255/18018 [2:45:04<38:39:54, 8.30s/step, loss=2.114, acc=0.580, acc_len=1.32, thru=17.1, I=17.3, T=30.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:18,543] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:27,459] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:15:28,021] loop.py:45 INFO TRAIN_STEP step=1256 loss=2.241150 Training: 7%|▋ | 1256/18018 [2:45:14<40:38:30, 8.73s/step, loss=2.241, acc=0.561, acc_len=1.27, thru=13.5, I=13.5, T=30.2, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:15:28,016] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1255: forward=2993.1ms backward=1202.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:28,464] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:35,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:15:36,370] loop.py:45 INFO TRAIN_STEP step=1257 loss=2.522121 Training: 7%|▋ | 1257/18018 [2:45:22<40:06:24, 8.61s/step, loss=2.522, acc=0.521, acc_len=1.16, thru=15.5, I=18.6, T=25.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:36,539] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:15:42,495] loop.py:45 INFO TRAIN_STEP step=1258 loss=2.390366 Training: 7%|▋ | 1258/18018 [2:45:28<36:37:37, 7.87s/step, loss=2.390, acc=0.533, acc_len=1.19, thru=16.5, I=19.1, T=36.4, wait=2.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:50,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:15:50,663] loop.py:45 INFO TRAIN_STEP step=1259 loss=2.311523 Training: 7%|▋ | 1259/18018 [2:45:36<37:02:37, 7.96s/step, loss=2.312, acc=0.547, acc_len=1.24, thru=17.9, I=16.8, T=33.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:51,091] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:56,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:15:57,516] loop.py:380 INFO TIMING step=1260: step=3.684s data=1.602s compute=3.656s [fwd=2.599s bwd=1.044s opt=0.003s] dispatch=3.162s [2026-10-18 16:15:57,516] loop.py:45 INFO TRAIN_STEP step=1260 loss=2.329750 Training: 7%|▋ | 1260/18018 [2:45:43<35:29:59, 7.63s/step, loss=2.330, acc=0.549, acc_len=1.24, thru=18.2, I=18.7, T=34.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:15:57,980] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:03,183] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:16:04,413] loop.py:45 INFO TRAIN_STEP step=1261 loss=2.263296 Training: 7%|▋ | 1261/18018 [2:45:50<34:28:51, 7.41s/step, loss=2.263, acc=0.551, acc_len=1.25, thru=19.2, I=23.7, T=34.4, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:16:04,409] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1260: forward=2647.3ms backward=1023.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:04,691] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:09,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:16:10,393] loop.py:45 INFO TRAIN_STEP step=1262 loss=2.262343 Training: 7%|▋ | 1262/18018 [2:45:56<32:29:16, 6.98s/step, loss=2.262, acc=0.551, acc_len=1.24, thru=20.1, I=25.0, T=33.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:10,632] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:16,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:16:16,665] loop.py:45 INFO TRAIN_STEP step=1263 loss=2.217802 Training: 7%|▋ | 1263/18018 [2:46:02<31:29:40, 6.77s/step, loss=2.218, acc=0.563, acc_len=1.27, thru=18.9, I=23.9, T=34.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:16,753] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:21,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:16:23,137] loop.py:45 INFO TRAIN_STEP step=1264 loss=2.358718 Training: 7%|▋ | 1264/18018 [2:46:09<31:04:59, 6.68s/step, loss=2.359, acc=0.539, acc_len=1.21, thru=19.0, I=22.9, T=31.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:23,221] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:28,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:16:30,209] loop.py:380 INFO TIMING step=1265: step=4.185s data=1.939s compute=4.152s [fwd=2.959s bwd=1.179s opt=0.003s] dispatch=2.878s [2026-10-18 16:16:30,209] loop.py:45 INFO TRAIN_STEP step=1265 loss=2.326931 Training: 7%|▋ | 1265/18018 [2:46:16<31:37:38, 6.80s/step, loss=2.327, acc=0.539, acc_len=1.22, thru=16.9, I=23.6, T=30.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:30,515] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:35,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:16:36,315] loop.py:45 INFO TRAIN_STEP step=1266 loss=2.247905 Training: 7%|▋ | 1266/18018 [2:46:22<30:39:53, 6.59s/step, loss=2.248, acc=0.562, acc_len=1.27, thru=18.5, I=24.6, T=33.7, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:16:36,310] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1265: forward=2709.0ms backward=1048.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:36,371] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 16:16:42,627] loop.py:45 INFO TRAIN_STEP step=1267 loss=2.146803 Training: 7%|▋ | 1267/18018 [2:46:28<30:16:26, 6.51s/step, loss=2.147, acc=0.572, acc_len=1.30, thru=16.0, I=19.0, T=35.3, wait=2.7s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:51,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:16:52,818] loop.py:45 INFO TRAIN_STEP step=1268 loss=2.173800 Training: 7%|▋ | 1268/18018 [2:46:38<35:24:46, 7.61s/step, loss=2.174, acc=0.567, acc_len=1.30, thru=16.3, I=16.0, T=27.2, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:52,962] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:16:59,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:17:01,343] loop.py:45 INFO TRAIN_STEP step=1269 loss=2.424448 Training: 7%|▋ | 1269/18018 [2:46:47<36:41:23, 7.89s/step, loss=2.424, acc=0.532, acc_len=1.20, thru=13.3, I=17.4, T=26.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:01,602] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:08,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:17:09,334] loop.py:380 INFO TIMING step=1270: step=4.213s data=1.909s compute=4.185s [fwd=2.963s bwd=1.207s opt=0.003s] dispatch=3.767s [2026-10-18 16:17:09,334] loop.py:45 INFO TRAIN_STEP step=1270 loss=2.124813 Training: 7%|▋ | 1270/18018 [2:46:55<36:49:59, 7.92s/step, loss=2.125, acc=0.577, acc_len=1.32, thru=14.1, I=19.9, T=30.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:09,514] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:13,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:17:14,990] loop.py:45 INFO TRAIN_STEP step=1271 loss=2.061446 Training: 7%|▋ | 1271/18018 [2:47:01<33:40:17, 7.24s/step, loss=2.061, acc=0.584, acc_len=1.34, thru=19.8, I=25.6, T=35.2, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:17:14,986] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1270: forward=2573.9ms backward=1018.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:15,203] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:17:23,287] loop.py:45 INFO TRAIN_STEP step=1272 loss=2.208383 Training: 7%|▋ | 1272/18018 [2:47:09<35:09:03, 7.56s/step, loss=2.208, acc=0.564, acc_len=1.29, thru=12.4, I=13.7, T=33.2, wait=4.4s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:31,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:17:33,139] loop.py:45 INFO TRAIN_STEP step=1273 loss=2.404861 Training: 7%|▋ | 1273/18018 [2:47:19<38:21:09, 8.25s/step, loss=2.405, acc=0.538, acc_len=1.20, thru=16.0, I=15.4, T=32.4, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:33,176] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:39,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:17:40,748] loop.py:45 INFO TRAIN_STEP step=1274 loss=2.335875 Training: 7%|▋ | 1274/18018 [2:47:26<37:27:33, 8.05s/step, loss=2.336, acc=0.541, acc_len=1.22, thru=15.2, I=20.0, T=26.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:41,094] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:46,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:17:47,993] loop.py:380 INFO TIMING step=1275: step=4.180s data=1.822s compute=4.148s [fwd=2.964s bwd=1.170s opt=0.003s] dispatch=3.056s [2026-10-18 16:17:47,993] loop.py:45 INFO TRAIN_STEP step=1275 loss=2.271998 Training: 7%|▋ | 1275/18018 [2:47:34<36:19:51, 7.81s/step, loss=2.272, acc=0.559, acc_len=1.27, thru=15.0, I=19.2, T=30.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:17:48,045] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 [2026-10-18 16:17:54,351] loop.py:45 INFO TRAIN_STEP step=1276 loss=2.256982 Training: 7%|▋ | 1276/18018 [2:47:40<34:18:11, 7.38s/step, loss=2.257, acc=0.558, acc_len=1.26, thru=16.5, I=24.3, T=33.5, wait=2.5s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:17:54,346] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1275: forward=2683.5ms backward=1092.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:01,136] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:18:01,764] loop.py:45 INFO TRAIN_STEP step=1277 loss=2.269406 Training: 7%|▋ | 1277/18018 [2:47:47<34:20:54, 7.39s/step, loss=2.269, acc=0.559, acc_len=1.26, thru=18.2, I=17.8, T=35.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:02,140] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:18:07,933] loop.py:45 INFO TRAIN_STEP step=1278 loss=2.152510 Training: 7%|▋ | 1278/18018 [2:47:54<32:38:50, 7.02s/step, loss=2.153, acc=0.570, acc_len=1.30, thru=19.3, I=22.1, T=35.9, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:13,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:18:15,868] loop.py:45 INFO TRAIN_STEP step=1279 loss=2.397877 Training: 7%|▋ | 1279/18018 [2:48:01<33:55:24, 7.30s/step, loss=2.398, acc=0.535, acc_len=1.20, thru=16.6, I=20.1, T=31.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:16,243] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 16:18:24,221] loop.py:380 INFO TIMING step=1280: step=4.002s data=1.710s compute=3.973s [fwd=2.824s bwd=1.134s opt=0.003s] dispatch=4.340s [2026-10-18 16:18:24,221] loop.py:45 INFO TRAIN_STEP step=1280 loss=2.302230 Training: 7%|▋ | 1280/18018 [2:48:10<35:23:45, 7.61s/step, loss=2.302, acc=0.551, acc_len=1.24, thru=14.5, I=14.7, T=32.0, wait=4.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:30,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:18:31,522] loop.py:45 INFO TRAIN_STEP step=1281 loss=2.138538 Training: 7%|▋ | 1281/18018 [2:48:17<34:57:20, 7.52s/step, loss=2.139, acc=0.577, acc_len=1.32, thru=18.4, I=18.1, T=32.6, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:18:31,516] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1280: forward=2693.8ms backward=1159.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:31,552] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 16:18:38,178] loop.py:45 INFO TRAIN_STEP step=1282 loss=2.428075 Training: 7%|▋ | 1282/18018 [2:48:24<33:45:09, 7.26s/step, loss=2.428, acc=0.531, acc_len=1.19, thru=18.4, I=21.5, T=35.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:38,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:38,669] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 16:18:45,704] loop.py:45 INFO TRAIN_STEP step=1283 loss=2.266481 Training: 7%|▋ | 1283/18018 [2:48:31<34:07:10, 7.34s/step, loss=2.266, acc=0.555, acc_len=1.26, thru=15.9, I=18.8, T=35.3, wait=3.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:52,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:18:52,276] loop.py:45 INFO TRAIN_STEP step=1284 loss=2.245286 Training: 7%|▋ | 1284/18018 [2:48:38<33:02:49, 7.11s/step, loss=2.245, acc=0.556, acc_len=1.26, thru=18.9, I=19.6, T=33.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:18:52,562] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 16:18:59,287] loop.py:380 INFO TIMING step=1285: step=3.805s data=1.614s compute=3.764s [fwd=2.706s bwd=1.044s opt=0.003s] dispatch=3.196s [2026-10-18 16:18:59,287] loop.py:45 INFO TRAIN_STEP step=1285 loss=2.205418 Training: 7%|▋ | 1285/18018 [2:48:45<32:54:21, 7.08s/step, loss=2.205, acc=0.565, acc_len=1.28, thru=17.6, I=18.9, T=33.6, wait=3.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:06,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:19:07,731] loop.py:45 INFO TRAIN_STEP step=1286 loss=2.198252 Training: 7%|▋ | 1286/18018 [2:48:53<34:48:34, 7.49s/step, loss=2.198, acc=0.569, acc_len=1.29, thru=16.2, I=16.7, T=34.7, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:19:07,727] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1285: forward=2579.7ms backward=1061.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:08,150] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:13,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:19:14,315] loop.py:45 INFO TRAIN_STEP step=1287 loss=2.254065 Training: 7%|▋ | 1287/18018 [2:49:00<33:32:42, 7.22s/step, loss=2.254, acc=0.555, acc_len=1.25, thru=18.9, I=24.5, T=33.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:14,476] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:19,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:19:20,810] loop.py:45 INFO TRAIN_STEP step=1288 loss=2.110232 Training: 7%|▋ | 1288/18018 [2:49:06<32:31:52, 7.00s/step, loss=2.110, acc=0.576, acc_len=1.32, thru=18.9, I=22.6, T=34.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:21,097] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:26,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:19:27,845] loop.py:45 INFO TRAIN_STEP step=1289 loss=2.131555 Training: 7%|▋ | 1289/18018 [2:49:13<32:34:54, 7.01s/step, loss=2.132, acc=0.573, acc_len=1.31, thru=17.1, I=22.6, T=31.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:28,279] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:32,441] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:19:34,211] loop.py:380 INFO TIMING step=1290: step=3.720s data=1.609s compute=3.689s [fwd=2.645s bwd=1.030s opt=0.003s] dispatch=2.635s [2026-10-18 16:19:34,211] loop.py:45 INFO TRAIN_STEP step=1290 loss=2.132806 Training: 7%|▋ | 1290/18018 [2:49:20<31:40:49, 6.82s/step, loss=2.133, acc=0.566, acc_len=1.29, thru=18.3, I=26.3, T=34.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:34,456] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:19:41,198] loop.py:45 INFO TRAIN_STEP step=1291 loss=2.231329 Training: 7%|▋ | 1291/18018 [2:49:27<31:54:37, 6.87s/step, loss=2.231, acc=0.562, acc_len=1.27, thru=16.8, I=22.0, T=33.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:41,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 16:19:41,194] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1290: forward=2635.0ms backward=1094.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:41,655] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:47,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 16:19:48,520] loop.py:45 INFO TRAIN_STEP step=1292 loss=2.168316 Training: 7%|▋ | 1292/18018 [2:49:34<32:32:44, 7.00s/step, loss=2.168, acc=0.571, acc_len=1.30, thru=18.1, I=19.1, T=33.5, wait=3.5s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:48,815] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:53,690] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:19:55,379] loop.py:45 INFO TRAIN_STEP step=1293 loss=2.186056 Training: 7%|▋ | 1293/18018 [2:49:41<32:20:24, 6.96s/step, loss=2.186, acc=0.567, acc_len=1.29, thru=18.1, I=24.0, T=30.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:19:55,705] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:00,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:20:01,340] loop.py:45 INFO TRAIN_STEP step=1294 loss=2.224584 Training: 7%|▋ | 1294/18018 [2:49:47<30:56:33, 6.66s/step, loss=2.225, acc=0.554, acc_len=1.26, thru=19.1, I=28.6, T=33.6, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:01,511] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:07,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:20:08,606] loop.py:380 INFO TIMING step=1295: step=4.185s data=1.947s compute=4.156s [fwd=2.911s bwd=1.230s opt=0.003s] dispatch=3.072s [2026-10-18 16:20:08,606] loop.py:45 INFO TRAIN_STEP step=1295 loss=2.149504 Training: 7%|▋ | 1295/18018 [2:49:54<31:47:00, 6.84s/step, loss=2.150, acc=0.575, acc_len=1.31, thru=16.9, I=22.1, T=30.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:08,999] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:20:15,072] loop.py:45 INFO TRAIN_STEP step=1296 loss=2.154403 Training: 7%|▋ | 1296/18018 [2:50:01<31:15:32, 6.73s/step, loss=2.154, acc=0.570, acc_len=1.30, thru=16.7, I=18.1, T=34.3, wait=2.7s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:20:15,068] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1295: forward=2559.1ms backward=1135.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:21,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:20:23,001] loop.py:45 INFO TRAIN_STEP step=1297 loss=2.115335 Training: 7%|▋ | 1297/18018 [2:50:09<32:55:46, 7.09s/step, loss=2.115, acc=0.579, acc_len=1.33, thru=18.2, I=18.8, T=29.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:23,418] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:28,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:20:30,396] loop.py:45 INFO TRAIN_STEP step=1298 loss=2.284033 Training: 7%|▋ | 1298/18018 [2:50:16<33:21:01, 7.18s/step, loss=2.284, acc=0.555, acc_len=1.26, thru=15.3, I=23.3, T=32.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:30,815] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:36,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:20:37,731] loop.py:45 INFO TRAIN_STEP step=1299 loss=2.303868 Training: 7%|▋ | 1299/18018 [2:50:23<33:33:58, 7.23s/step, loss=2.304, acc=0.545, acc_len=1.22, thru=15.7, I=19.7, T=32.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:37,875] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:42,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:20:43,985] loop.py:380 INFO TIMING step=1300: step=3.901s data=1.764s compute=3.873s [fwd=2.715s bwd=1.144s opt=0.003s] dispatch=2.343s [2026-10-18 16:20:43,986] loop.py:45 INFO TRAIN_STEP step=1300 loss=2.180123 Training: 7%|▋ | 1300/18018 [2:50:30<32:12:19, 6.94s/step, loss=2.180, acc=0.564, acc_len=1.28, thru=18.6, I=22.4, T=32.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:44,370] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:48,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:20:49,767] loop.py:45 INFO TRAIN_STEP step=1301 loss=2.362562 Training: 7%|▋ | 1301/18018 [2:50:35<30:35:54, 6.59s/step, loss=2.363, acc=0.538, acc_len=1.21, thru=20.7, I=26.1, T=37.4, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:20:49,763] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1300: forward=2360.6ms backward=1019.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:50,133] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:55,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:20:57,770] loop.py:45 INFO TRAIN_STEP step=1302 loss=2.289851 Training: 7%|▋ | 1302/18018 [2:50:43<32:33:52, 7.01s/step, loss=2.290, acc=0.548, acc_len=1.24, thru=16.0, I=24.1, T=33.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:20:58,079] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 16:21:06,176] loop.py:45 INFO TRAIN_STEP step=1303 loss=2.154728 Training: 7%|▋ | 1303/18018 [2:50:52<34:30:06, 7.43s/step, loss=2.155, acc=0.571, acc_len=1.30, thru=14.3, I=13.1, T=29.2, wait=4.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:14,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:21:14,715] loop.py:45 INFO TRAIN_STEP step=1304 loss=2.321655 Training: 7%|▋ | 1304/18018 [2:51:00<36:02:48, 7.76s/step, loss=2.322, acc=0.545, acc_len=1.22, thru=16.7, I=15.9, T=31.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:14,789] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-18 16:21:20,747] loop.py:380 INFO TIMING step=1305: step=3.971s data=1.755s compute=3.937s [fwd=2.847s bwd=1.076s opt=0.003s] dispatch=2.050s [2026-10-18 16:21:20,747] loop.py:45 INFO TRAIN_STEP step=1305 loss=2.439876 Training: 7%|▋ | 1305/18018 [2:51:06<33:37:47, 7.24s/step, loss=2.440, acc=0.526, acc_len=1.19, thru=18.5, I=18.4, T=32.2, wait=2.0s, pool=60, epoch=1/2][2026-10-18 16:21:28,273] loop.py:45 INFO TRAIN_STEP step=1306 loss=2.305892 Training: 7%|▋ | 1306/18018 [2:51:14<34:01:18, 7.33s/step, loss=2.306, acc=0.553, acc_len=1.24, thru=17.4, I=18.1, T=26.5, wait=2.7s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:21:28,269] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1305: forward=3514.3ms backward=1279.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:34,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:21:37,218] loop.py:45 INFO TRAIN_STEP step=1307 loss=2.160056 Training: 7%|▋ | 1307/18018 [2:51:23<36:16:08, 7.81s/step, loss=2.160, acc=0.574, acc_len=1.30, thru=14.6, I=16.7, T=27.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:37,578] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:42,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:21:44,067] loop.py:45 INFO TRAIN_STEP step=1308 loss=2.307200 Training: 7%|▋ | 1308/18018 [2:51:30<34:55:35, 7.52s/step, loss=2.307, acc=0.545, acc_len=1.22, thru=14.9, I=22.2, T=32.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:44,406] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:49,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:21:49,938] loop.py:45 INFO TRAIN_STEP step=1309 loss=2.138649 Training: 7%|▋ | 1309/18018 [2:51:36<32:37:03, 7.03s/step, loss=2.139, acc=0.573, acc_len=1.31, thru=20.0, I=25.0, T=35.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:50,190] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:54,688] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:21:56,569] loop.py:380 INFO TIMING step=1310: step=3.674s data=1.557s compute=3.640s [fwd=2.608s bwd=1.018s opt=0.003s] dispatch=2.950s [2026-10-18 16:21:56,569] loop.py:45 INFO TRAIN_STEP step=1310 loss=2.103161 Training: 7%|▋ | 1310/18018 [2:51:42<32:04:03, 6.91s/step, loss=2.103, acc=0.581, acc_len=1.34, thru=18.8, I=24.3, T=34.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:21:56,713] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:03,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:22:06,516] loop.py:45 INFO TRAIN_STEP step=1311 loss=2.348067 Training: 7%|▋ | 1311/18018 [2:51:52<36:17:30, 7.82s/step, loss=2.348, acc=0.546, acc_len=1.23, thru=13.1, I=17.1, T=23.5, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:22:06,511] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1310: forward=4167.4ms backward=1199.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:06,950] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 16:22:13,335] loop.py:45 INFO TRAIN_STEP step=1312 loss=2.175577 Training: 7%|▋ | 1312/18018 [2:51:59<34:53:48, 7.52s/step, loss=2.176, acc=0.561, acc_len=1.27, thru=13.6, I=18.8, T=29.0, wait=2.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:20,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:22:21,617] loop.py:45 INFO TRAIN_STEP step=1313 loss=2.377558 Training: 7%|▋ | 1313/18018 [2:52:07<35:57:28, 7.75s/step, loss=2.378, acc=0.541, acc_len=1.22, thru=16.4, I=18.1, T=29.8, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:21,761] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:29,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:22:29,237] loop.py:45 INFO TRAIN_STEP step=1314 loss=2.372658 Training: 7%|▋ | 1314/18018 [2:52:15<35:46:23, 7.71s/step, loss=2.373, acc=0.536, acc_len=1.20, thru=14.9, I=18.5, T=30.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:29,502] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:34,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:22:35,802] loop.py:380 INFO TIMING step=1315: step=3.645s data=1.557s compute=3.613s [fwd=2.582s bwd=1.016s opt=0.003s] dispatch=2.911s [2026-10-18 16:22:35,802] loop.py:45 INFO TRAIN_STEP step=1315 loss=2.138911 Training: 7%|▋ | 1315/18018 [2:52:21<34:10:33, 7.37s/step, loss=2.139, acc=0.572, acc_len=1.30, thru=18.0, I=19.8, T=35.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:36,302] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:22:43,355] loop.py:45 INFO TRAIN_STEP step=1316 loss=2.220971 Training: 7%|▋ | 1316/18018 [2:52:29<34:26:28, 7.42s/step, loss=2.221, acc=0.562, acc_len=1.28, thru=16.7, I=21.9, T=31.2, wait=3.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:22:43,350] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1315: forward=2708.0ms backward=1346.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:52,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:22:53,921] loop.py:45 INFO TRAIN_STEP step=1317 loss=2.062826 Training: 7%|▋ | 1317/18018 [2:52:40<38:48:29, 8.37s/step, loss=2.063, acc=0.585, acc_len=1.35, thru=13.1, I=22.3, T=32.1, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:22:54,272] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:00,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:23:01,223] loop.py:45 INFO TRAIN_STEP step=1318 loss=2.278019 Training: 7%|▋ | 1318/18018 [2:52:47<37:19:40, 8.05s/step, loss=2.278, acc=0.551, acc_len=1.24, thru=16.0, I=21.5, T=35.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:01,402] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:07,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:23:07,669] loop.py:45 INFO TRAIN_STEP step=1319 loss=2.324056 Training: 7%|▋ | 1319/18018 [2:52:53<35:05:43, 7.57s/step, loss=2.324, acc=0.549, acc_len=1.24, thru=19.6, I=22.5, T=31.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:08,103] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:13,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:23:14,497] loop.py:380 INFO TIMING step=1320: step=3.862s data=1.756s compute=3.828s [fwd=2.745s bwd=1.069s opt=0.003s] dispatch=2.957s [2026-10-18 16:23:14,497] loop.py:45 INFO TRAIN_STEP step=1320 loss=2.134739 Training: 7%|▋ | 1320/18018 [2:53:00<34:04:08, 7.35s/step, loss=2.135, acc=0.572, acc_len=1.30, thru=18.3, I=21.1, T=33.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:14,621] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 16:23:22,017] loop.py:45 INFO TRAIN_STEP step=1321 loss=2.357658 Training: 7%|▋ | 1321/18018 [2:53:08<34:18:33, 7.40s/step, loss=2.358, acc=0.539, acc_len=1.21, thru=15.5, I=19.7, T=30.5, wait=3.3s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:23:22,012] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1320: forward=2899.3ms backward=1252.3ms [2026-10-18 16:23:31,332] loop.py:45 INFO TRAIN_STEP step=1322 loss=1.969611 Training: 7%|▋ | 1322/18018 [2:53:17<36:58:30, 7.97s/step, loss=1.970, acc=0.601, acc_len=1.39, thru=14.2, I=13.8, T=33.0, wait=5.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:38,401] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:23:40,363] loop.py:45 INFO TRAIN_STEP step=1323 loss=2.336284 Training: 7%|▋ | 1323/18018 [2:53:26<38:26:47, 8.29s/step, loss=2.336, acc=0.543, acc_len=1.21, thru=15.3, I=17.4, T=33.2, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:40,415] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:46,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:23:47,322] loop.py:45 INFO TRAIN_STEP step=1324 loss=2.160346 Training: 7%|▋ | 1324/18018 [2:53:33<36:35:21, 7.89s/step, loss=2.160, acc=0.566, acc_len=1.29, thru=16.1, I=22.9, T=28.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:47,749] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:53,328] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:23:54,655] loop.py:380 INFO TIMING step=1325: step=3.889s data=1.706s compute=3.845s [fwd=2.705s bwd=1.126s opt=0.003s] dispatch=3.436s [2026-10-18 16:23:54,656] loop.py:45 INFO TRAIN_STEP step=1325 loss=2.416991 Training: 7%|▋ | 1325/18018 [2:53:40<35:48:55, 7.72s/step, loss=2.417, acc=0.537, acc_len=1.20, thru=16.6, I=21.2, T=32.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:23:54,838] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:00,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:24:01,656] loop.py:45 INFO TRAIN_STEP step=1326 loss=2.146258 Training: 7%|▋ | 1326/18018 [2:53:47<34:48:24, 7.51s/step, loss=2.146, acc=0.573, acc_len=1.30, thru=17.7, I=23.9, T=35.4, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:24:01,651] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1325: forward=2527.0ms backward=1036.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:01,918] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:07,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:24:08,885] loop.py:45 INFO TRAIN_STEP step=1327 loss=2.076938 Training: 7%|▋ | 1327/18018 [2:53:54<34:25:05, 7.42s/step, loss=2.077, acc=0.580, acc_len=1.33, thru=16.7, I=22.8, T=29.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:09,056] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:24:15,016] loop.py:45 INFO TRAIN_STEP step=1328 loss=2.321745 Training: 7%|▋ | 1328/18018 [2:54:01<32:36:57, 7.04s/step, loss=2.322, acc=0.543, acc_len=1.22, thru=15.7, I=21.2, T=35.4, wait=2.5s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:21,483] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:24:23,374] loop.py:45 INFO TRAIN_STEP step=1329 loss=2.190689 Training: 7%|▋ | 1329/18018 [2:54:09<34:27:24, 7.43s/step, loss=2.191, acc=0.561, acc_len=1.28, thru=17.2, I=18.3, T=25.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:23,501] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:28,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:24:28,925] loop.py:380 INFO TIMING step=1330: step=3.552s data=1.433s compute=3.524s [fwd=2.459s bwd=1.050s opt=0.003s] dispatch=1.988s [2026-10-18 16:24:28,925] loop.py:45 INFO TRAIN_STEP step=1330 loss=2.141834 Training: 7%|▋ | 1330/18018 [2:54:15<31:50:14, 6.87s/step, loss=2.142, acc=0.568, acc_len=1.30, thru=18.6, I=24.1, T=36.0, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:28,960] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:33,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:24:35,341] loop.py:45 INFO TRAIN_STEP step=1331 loss=2.423131 Training: 7%|▋ | 1331/18018 [2:54:21<31:12:16, 6.73s/step, loss=2.423, acc=0.532, acc_len=1.18, thru=19.6, I=24.7, T=30.4, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:24:35,336] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1330: forward=2865.7ms backward=1294.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:35,511] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:40,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:24:41,434] loop.py:45 INFO TRAIN_STEP step=1332 loss=2.357699 Training: 7%|▋ | 1332/18018 [2:54:27<30:19:00, 6.54s/step, loss=2.358, acc=0.536, acc_len=1.20, thru=18.6, I=24.4, T=34.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:24:41,852] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:24:49,714] loop.py:45 INFO TRAIN_STEP step=1333 loss=2.184434 Training: 7%|▋ | 1333/18018 [2:54:35<32:44:02, 7.06s/step, loss=2.184, acc=0.570, acc_len=1.30, thru=11.3, I=15.7, T=31.5, wait=4.2s, pool=28, epoch=1/2][2026-10-18 16:24:58,772] loop.py:45 INFO TRAIN_STEP step=1334 loss=2.283443 Training: 7%|▋ | 1334/18018 [2:54:44<35:30:07, 7.66s/step, loss=2.283, acc=0.551, acc_len=1.24, thru=16.6, I=15.5, T=29.6, wait=4.7s, pool=56, epoch=1/2][2026-10-18 16:25:08,934] loop.py:380 INFO TIMING step=1335: step=4.042s data=1.773s compute=4.006s [fwd=2.858s bwd=1.133s opt=0.003s] dispatch=6.113s [2026-10-18 16:25:08,935] loop.py:45 INFO TRAIN_STEP step=1335 loss=2.327556 Training: 7%|▋ | 1335/18018 [2:54:55<38:58:38, 8.41s/step, loss=2.328, acc=0.548, acc_len=1.23, thru=13.3, I=11.7, T=31.7, wait=6.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:15,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:25:17,449] loop.py:45 INFO TRAIN_STEP step=1336 loss=2.249885 Training: 7%|▋ | 1336/18018 [2:55:03<39:07:24, 8.44s/step, loss=2.250, acc=0.557, acc_len=1.26, thru=16.8, I=18.8, T=30.1, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:25:17,443] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1335: forward=2911.0ms backward=1299.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:17,615] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:23,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:25:23,424] loop.py:45 INFO TRAIN_STEP step=1337 loss=2.077833 Training: 7%|▋ | 1337/18018 [2:55:09<35:41:13, 7.70s/step, loss=2.078, acc=0.582, acc_len=1.34, thru=18.1, I=23.3, T=36.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:23,669] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:30,647] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:25:30,908] loop.py:45 INFO TRAIN_STEP step=1338 loss=2.225284 Training: 7%|▋ | 1338/18018 [2:55:16<35:22:58, 7.64s/step, loss=2.225, acc=0.557, acc_len=1.27, thru=18.1, I=18.0, T=27.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:31,150] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:35,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:25:37,281] loop.py:45 INFO TRAIN_STEP step=1339 loss=2.262895 Training: 7%|▋ | 1339/18018 [2:55:23<33:37:28, 7.26s/step, loss=2.263, acc=0.555, acc_len=1.25, thru=19.0, I=20.1, T=30.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:37,738] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:25:44,645] loop.py:380 INFO TIMING step=1340: step=4.310s data=1.971s compute=4.281s [fwd=2.967s bwd=1.299s opt=0.003s] dispatch=3.045s [2026-10-18 16:25:44,645] loop.py:45 INFO TRAIN_STEP step=1340 loss=2.380302 Training: 7%|▋ | 1340/18018 [2:55:30<33:46:27, 7.29s/step, loss=2.380, acc=0.540, acc_len=1.21, thru=13.9, I=18.3, T=29.7, wait=3.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:51,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:25:53,190] loop.py:45 INFO TRAIN_STEP step=1341 loss=2.101922 Training: 7%|▋ | 1341/18018 [2:55:39<35:30:48, 7.67s/step, loss=2.102, acc=0.580, acc_len=1.33, thru=17.6, I=16.9, T=29.5, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:25:53,186] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1340: forward=2989.6ms backward=1313.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:53,585] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:25:59,459] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:26:00,614] loop.py:45 INFO TRAIN_STEP step=1342 loss=2.106963 Training: 7%|▋ | 1342/18018 [2:55:46<35:10:22, 7.59s/step, loss=2.107, acc=0.577, acc_len=1.32, thru=15.3, I=20.7, T=30.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:00,969] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:06,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:26:07,667] loop.py:45 INFO TRAIN_STEP step=1343 loss=2.287486 Training: 7%|▋ | 1343/18018 [2:55:53<34:25:26, 7.43s/step, loss=2.287, acc=0.551, acc_len=1.24, thru=16.6, I=22.2, T=30.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:07,932] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:13,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:26:14,669] loop.py:45 INFO TRAIN_STEP step=1344 loss=2.134133 Training: 7%|▋ | 1344/18018 [2:56:00<33:49:17, 7.30s/step, loss=2.134, acc=0.572, acc_len=1.31, thru=17.1, I=21.6, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:14,785] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:20,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:26:21,761] loop.py:380 INFO TIMING step=1345: step=4.440s data=2.027s compute=4.405s [fwd=3.069s bwd=1.321s opt=0.003s] dispatch=2.644s [2026-10-18 16:26:21,761] loop.py:45 INFO TRAIN_STEP step=1345 loss=2.265612 Training: 7%|▋ | 1345/18018 [2:56:07<33:31:47, 7.24s/step, loss=2.266, acc=0.552, acc_len=1.24, thru=17.0, I=21.4, T=28.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:22,126] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:26,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:26:27,589] loop.py:45 INFO TRAIN_STEP step=1346 loss=2.197362 Training: 7%|▋ | 1346/18018 [2:56:13<31:33:51, 6.82s/step, loss=2.197, acc=0.564, acc_len=1.28, thru=18.6, I=25.5, T=35.3, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:26:27,584] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1345: forward=2580.1ms backward=994.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:27,948] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:34,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:26:35,644] loop.py:45 INFO TRAIN_STEP step=1347 loss=2.244661 Training: 7%|▋ | 1347/18018 [2:56:21<33:17:08, 7.19s/step, loss=2.245, acc=0.562, acc_len=1.27, thru=15.6, I=17.6, T=32.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:35,832] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:26:41,978] loop.py:45 INFO TRAIN_STEP step=1348 loss=2.238693 Training: 7%|▋ | 1348/18018 [2:56:28<32:05:38, 6.93s/step, loss=2.239, acc=0.560, acc_len=1.27, thru=18.6, I=20.0, T=33.3, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:46,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:26:48,238] loop.py:45 INFO TRAIN_STEP step=1349 loss=2.342379 Training: 7%|▋ | 1349/18018 [2:56:34<31:09:54, 6.73s/step, loss=2.342, acc=0.547, acc_len=1.23, thru=19.6, I=22.4, T=31.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:48,276] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:53,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:26:55,645] loop.py:380 INFO TIMING step=1350: step=3.922s data=1.724s compute=3.887s [fwd=2.704s bwd=1.168s opt=0.003s] dispatch=3.474s [2026-10-18 16:26:55,645] loop.py:45 INFO TRAIN_STEP step=1350 loss=2.235947 Training: 7%|▋ | 1350/18018 [2:56:41<32:05:56, 6.93s/step, loss=2.236, acc=0.561, acc_len=1.26, thru=16.4, I=20.8, T=32.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:26:55,999] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:27:02,370] loop.py:45 INFO TRAIN_STEP step=1351 loss=2.092200 Training: 7%|▋ | 1351/18018 [2:56:48<31:48:27, 6.87s/step, loss=2.092, acc=0.579, acc_len=1.33, thru=17.6, I=22.5, T=31.1, wait=2.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:27:02,365] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1350: forward=2822.4ms backward=1245.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:08,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:27:10,406] loop.py:45 INFO TRAIN_STEP step=1352 loss=2.078448 Training: 8%|▊ | 1352/18018 [2:56:56<33:25:32, 7.22s/step, loss=2.078, acc=0.579, acc_len=1.33, thru=16.2, I=17.7, T=27.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:10,905] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:18,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:27:19,185] loop.py:45 INFO TRAIN_STEP step=1353 loss=2.139662 Training: 8%|▊ | 1353/18018 [2:57:05<35:35:10, 7.69s/step, loss=2.140, acc=0.578, acc_len=1.32, thru=15.5, I=16.2, T=32.2, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:19,682] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:24,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:27:26,085] loop.py:45 INFO TRAIN_STEP step=1354 loss=2.290644 Training: 8%|▊ | 1354/18018 [2:57:12<34:29:44, 7.45s/step, loss=2.291, acc=0.554, acc_len=1.24, thru=17.6, I=20.5, T=29.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:26,254] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 16:27:34,653] loop.py:380 INFO TIMING step=1355: step=3.748s data=1.603s compute=3.720s [fwd=2.655s bwd=1.050s opt=0.003s] dispatch=4.810s [2026-10-18 16:27:34,653] loop.py:45 INFO TRAIN_STEP step=1355 loss=2.226828 Training: 8%|▊ | 1355/18018 [2:57:20<36:02:22, 7.79s/step, loss=2.227, acc=0.560, acc_len=1.26, thru=12.1, I=16.9, T=34.1, wait=4.8s, pool=52, epoch=1/2][2026-10-18 16:27:42,809] loop.py:45 INFO TRAIN_STEP step=1356 loss=2.299126 Training: 8%|▊ | 1356/18018 [2:57:28<36:32:59, 7.90s/step, loss=2.299, acc=0.549, acc_len=1.23, thru=13.9, I=13.5, T=31.0, wait=4.0s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:27:42,805] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1355: forward=2855.7ms backward=1212.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:49,395] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:27:50,421] loop.py:45 INFO TRAIN_STEP step=1357 loss=2.347353 Training: 8%|▊ | 1357/18018 [2:57:36<36:09:12, 7.81s/step, loss=2.347, acc=0.548, acc_len=1.23, thru=18.2, I=17.7, T=31.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:50,902] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:57,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:27:58,085] loop.py:45 INFO TRAIN_STEP step=1358 loss=2.249121 Training: 8%|▊ | 1358/18018 [2:57:44<35:56:44, 7.77s/step, loss=2.249, acc=0.562, acc_len=1.28, thru=15.2, I=18.1, T=34.1, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:27:58,580] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:03,708] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:28:04,344] loop.py:45 INFO TRAIN_STEP step=1359 loss=2.106030 Training: 8%|▊ | 1359/18018 [2:57:50<33:51:08, 7.32s/step, loss=2.106, acc=0.578, acc_len=1.32, thru=19.2, I=22.1, T=35.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:04,715] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:11,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:28:12,836] loop.py:380 INFO TIMING step=1360: step=4.798s data=2.384s compute=4.756s [fwd=3.343s bwd=1.398s opt=0.003s] dispatch=3.685s [2026-10-18 16:28:12,837] loop.py:45 INFO TRAIN_STEP step=1360 loss=2.206563 Training: 8%|▊ | 1360/18018 [2:57:58<35:28:55, 7.67s/step, loss=2.207, acc=0.560, acc_len=1.28, thru=15.1, I=18.7, T=26.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:13,247] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:19,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:28:21,468] loop.py:45 INFO TRAIN_STEP step=1361 loss=2.381524 Training: 8%|▊ | 1361/18018 [2:58:07<36:49:10, 7.96s/step, loss=2.382, acc=0.539, acc_len=1.21, thru=13.7, I=19.1, T=28.7, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:28:21,462] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1360: forward=3070.4ms backward=1326.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:21,557] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:25,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:28:26,946] loop.py:45 INFO TRAIN_STEP step=1362 loss=2.043104 Training: 8%|▊ | 1362/18018 [2:58:13<33:22:17, 7.21s/step, loss=2.043, acc=0.586, acc_len=1.35, thru=19.6, I=27.1, T=34.9, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:27,448] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:33,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:28:35,303] loop.py:45 INFO TRAIN_STEP step=1363 loss=2.058839 Training: 8%|▊ | 1363/18018 [2:58:21<34:57:45, 7.56s/step, loss=2.059, acc=0.586, acc_len=1.35, thru=14.7, I=20.6, T=30.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:35,480] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:41,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:28:42,384] loop.py:45 INFO TRAIN_STEP step=1364 loss=2.225664 Training: 8%|▊ | 1364/18018 [2:58:28<34:17:46, 7.41s/step, loss=2.226, acc=0.558, acc_len=1.26, thru=16.0, I=22.5, T=28.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:42,487] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:47,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:28:49,849] loop.py:380 INFO TIMING step=1365: step=4.906s data=2.332s compute=4.872s [fwd=3.444s bwd=1.413s opt=0.003s] dispatch=2.550s [2026-10-18 16:28:49,850] loop.py:45 INFO TRAIN_STEP step=1365 loss=2.338088 Training: 8%|▊ | 1365/18018 [2:58:35<34:21:55, 7.43s/step, loss=2.338, acc=0.541, acc_len=1.21, thru=16.0, I=21.5, T=26.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:50,126] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:56,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:28:56,623] loop.py:45 INFO TRAIN_STEP step=1366 loss=2.201564 Training: 8%|▊ | 1366/18018 [2:58:42<33:27:21, 7.23s/step, loss=2.202, acc=0.563, acc_len=1.28, thru=15.2, I=20.3, T=31.1, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:28:56,618] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1365: forward=2861.9ms backward=1207.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:28:57,073] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:03,213] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:29:03,575] loop.py:45 INFO TRAIN_STEP step=1367 loss=2.174202 Training: 8%|▊ | 1367/18018 [2:58:49<33:03:46, 7.15s/step, loss=2.174, acc=0.565, acc_len=1.29, thru=17.2, I=20.0, T=33.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:03,722] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 16:29:11,596] loop.py:45 INFO TRAIN_STEP step=1368 loss=2.202385 Training: 8%|▊ | 1368/18018 [2:58:57<34:16:26, 7.41s/step, loss=2.202, acc=0.564, acc_len=1.28, thru=14.3, I=17.1, T=28.4, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:21,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:29:22,825] loop.py:45 INFO TRAIN_STEP step=1369 loss=2.177793 Training: 8%|▊ | 1369/18018 [2:59:08<39:34:07, 8.56s/step, loss=2.178, acc=0.566, acc_len=1.29, thru=13.8, I=12.2, T=26.5, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:22,863] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:28,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:29:29,010] loop.py:380 INFO TIMING step=1370: step=3.732s data=1.574s compute=3.700s [fwd=2.641s bwd=1.044s opt=0.003s] dispatch=2.444s [2026-10-18 16:29:29,010] loop.py:45 INFO TRAIN_STEP step=1370 loss=2.244105 Training: 8%|▊ | 1370/18018 [2:59:15<36:16:38, 7.84s/step, loss=2.244, acc=0.560, acc_len=1.26, thru=17.1, I=21.5, T=34.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:29,059] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:38,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 16:29:40,141] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1370: forward=2960.3ms backward=1092.8ms [2026-10-18 16:29:40,145] loop.py:45 INFO TRAIN_STEP step=1371 loss=2.275774 Training: 8%|▊ | 1371/18018 [2:59:26<40:50:25, 8.83s/step, loss=2.276, acc=0.552, acc_len=1.25, thru=12.0, I=12.7, T=31.1, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:40,386] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:45,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:29:46,513] loop.py:45 INFO TRAIN_STEP step=1372 loss=2.048626 Training: 8%|▊ | 1372/18018 [2:59:32<37:25:11, 8.09s/step, loss=2.049, acc=0.587, acc_len=1.35, thru=18.0, I=23.9, T=33.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:46,859] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:52,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:29:53,030] loop.py:45 INFO TRAIN_STEP step=1373 loss=2.158969 Training: 8%|▊ | 1373/18018 [2:59:39<35:13:50, 7.62s/step, loss=2.159, acc=0.567, acc_len=1.29, thru=18.8, I=22.3, T=30.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:29:53,361] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:29:59,432] loop.py:45 INFO TRAIN_STEP step=1374 loss=2.319624 Training: 8%|▊ | 1374/18018 [2:59:45<33:32:12, 7.25s/step, loss=2.320, acc=0.546, acc_len=1.23, thru=17.7, I=18.7, T=34.1, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:08,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:30:10,706] loop.py:380 INFO TIMING step=1375: step=4.669s data=2.242s compute=4.609s [fwd=3.139s bwd=1.455s opt=0.003s] dispatch=6.597s [2026-10-18 16:30:10,706] loop.py:45 INFO TRAIN_STEP step=1375 loss=2.177901 Training: 8%|▊ | 1375/18018 [2:59:56<39:06:50, 8.46s/step, loss=2.178, acc=0.573, acc_len=1.31, thru=11.3, I=13.7, T=27.4, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:10,882] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:30:18,268] loop.py:45 INFO TRAIN_STEP step=1376 loss=2.419593 (TrainerActor pid=9310) [2026-10-18 16:30:18,263] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1375: forward=2779.6ms backward=1121.3ms Training: 8%|▊ | 1376/18018 [3:00:04<37:51:42, 8.19s/step, loss=2.420, acc=0.533, acc_len=1.19, thru=14.1, I=17.1, T=32.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:18,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:18,697] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 16:30:26,928] loop.py:45 INFO TRAIN_STEP step=1377 loss=2.302794 Training: 8%|▊ | 1377/18018 [3:00:13<38:30:49, 8.33s/step, loss=2.303, acc=0.546, acc_len=1.23, thru=15.6, I=15.4, T=32.1, wait=4.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:34,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:30:35,950] loop.py:45 INFO TRAIN_STEP step=1378 loss=2.286341 Training: 8%|▊ | 1378/18018 [3:00:22<39:27:59, 8.54s/step, loss=2.286, acc=0.557, acc_len=1.26, thru=15.1, I=16.7, T=29.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:36,417] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:42,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:30:42,371] loop.py:45 INFO TRAIN_STEP step=1379 loss=2.193967 Training: 8%|▊ | 1379/18018 [3:00:28<36:31:42, 7.90s/step, loss=2.194, acc=0.566, acc_len=1.29, thru=18.2, I=22.5, T=35.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:42,731] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:52,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:30:52,890] loop.py:380 INFO TIMING step=1380: step=3.946s data=1.572s compute=3.904s [fwd=2.811s bwd=1.077s opt=0.003s] dispatch=6.564s [2026-10-18 16:30:52,890] loop.py:45 INFO TRAIN_STEP step=1380 loss=2.214979 Training: 8%|▊ | 1380/18018 [3:00:38<40:09:08, 8.69s/step, loss=2.215, acc=0.561, acc_len=1.27, thru=13.1, I=12.8, T=32.4, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:53,284] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:59,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:30:59,617] loop.py:45 INFO TRAIN_STEP step=1381 loss=2.096055 (TrainerActor pid=9310) [2026-10-18 16:30:59,612] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1380: forward=2568.1ms backward=1040.3ms Training: 8%|▊ | 1381/18018 [3:00:45<37:26:04, 8.10s/step, loss=2.096, acc=0.582, acc_len=1.34, thru=17.7, I=19.2, T=35.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:30:59,656] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:05,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:31:05,926] loop.py:45 INFO TRAIN_STEP step=1382 loss=2.030863 Training: 8%|▊ | 1382/18018 [3:00:52<34:56:47, 7.56s/step, loss=2.031, acc=0.586, acc_len=1.35, thru=20.4, I=22.7, T=34.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:06,356] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:11,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:31:13,340] loop.py:45 INFO TRAIN_STEP step=1383 loss=2.231098 Training: 8%|▊ | 1383/18018 [3:00:59<34:44:29, 7.52s/step, loss=2.231, acc=0.557, acc_len=1.27, thru=15.5, I=20.6, T=28.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:13,665] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:19,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:31:21,347] loop.py:45 INFO TRAIN_STEP step=1384 loss=2.106499 Training: 8%|▊ | 1384/18018 [3:01:07<35:25:01, 7.67s/step, loss=2.106, acc=0.582, acc_len=1.34, thru=16.7, I=21.0, T=30.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:21,605] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:27,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:31:28,240] loop.py:380 INFO TIMING step=1385: step=3.810s data=1.651s compute=3.783s [fwd=2.686s bwd=1.083s opt=0.003s] dispatch=3.072s [2026-10-18 16:31:28,241] loop.py:45 INFO TRAIN_STEP step=1385 loss=2.192453 Training: 8%|▊ | 1385/18018 [3:01:14<34:20:33, 7.43s/step, loss=2.192, acc=0.565, acc_len=1.29, thru=16.0, I=22.2, T=33.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:28,359] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:34,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:31:34,211] loop.py:45 INFO TRAIN_STEP step=1386 loss=2.186656 (TrainerActor pid=9310) [2026-10-18 16:31:34,207] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1385: forward=2515.5ms backward=1010.7ms Training: 8%|▊ | 1386/18018 [3:01:20<32:18:53, 6.99s/step, loss=2.187, acc=0.565, acc_len=1.28, thru=19.9, I=20.2, T=35.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:34,587] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 16:31:40,813] loop.py:45 INFO TRAIN_STEP step=1387 loss=2.374692 Training: 8%|▊ | 1387/18018 [3:01:26<31:46:02, 6.88s/step, loss=2.375, acc=0.539, acc_len=1.21, thru=18.0, I=23.0, T=31.6, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:50,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:31:51,640] loop.py:45 INFO TRAIN_STEP step=1388 loss=2.159917 Training: 8%|▊ | 1388/18018 [3:01:37<37:14:33, 8.06s/step, loss=2.160, acc=0.568, acc_len=1.30, thru=13.1, I=13.3, T=29.4, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:51,675] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:58,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:31:59,497] loop.py:45 INFO TRAIN_STEP step=1389 loss=2.086793 Training: 8%|▊ | 1389/18018 [3:01:45<36:57:23, 8.00s/step, loss=2.087, acc=0.577, acc_len=1.32, thru=14.8, I=19.5, T=27.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:31:59,671] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:32:04,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:32:06,395] loop.py:380 INFO TIMING step=1390: step=4.673s data=2.093s compute=4.649s [fwd=3.179s bwd=1.454s opt=0.003s] dispatch=2.214s [2026-10-18 16:32:06,395] loop.py:45 INFO TRAIN_STEP step=1390 loss=2.121685 Training: 8%|▊ | 1390/18018 [3:01:52<35:25:34, 7.67s/step, loss=2.122, acc=0.574, acc_len=1.31, thru=16.1, I=24.0, T=27.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:32:06,639] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:32:13,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:32:14,247] loop.py:45 INFO TRAIN_STEP step=1391 loss=2.400532 Training: 8%|▊ | 1391/18018 [3:02:00<35:40:22, 7.72s/step, loss=2.401, acc=0.534, acc_len=1.19, thru=16.8, I=18.1, T=30.8, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:32:14,243] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1390: forward=2981.8ms backward=1131.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:32:14,702] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:32:19,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:32:21,063] loop.py:45 INFO TRAIN_STEP step=1392 loss=2.097435 Training: 8%|▊ | 1392/18018 [3:02:07<34:24:46, 7.45s/step, loss=2.097, acc=0.577, acc_len=1.32, thru=17.5, I=21.5, T=27.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:32:21,531] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:32:29,226] loop.py:45 INFO TRAIN_STEP step=1393 loss=2.136932 Training: 8%|▊ | 1393/18018 [3:02:15<35:23:59, 7.67s/step, loss=2.137, acc=0.572, acc_len=1.31, thru=12.5, I=15.9, T=31.3, wait=4.1s, pool=52, epoch=1/2][2026-10-18 16:32:37,563] loop.py:45 INFO TRAIN_STEP step=1394 loss=2.369251 Training: 8%|▊ | 1394/18018 [3:02:23<36:19:33, 7.87s/step, loss=2.369, acc=0.545, acc_len=1.22, thru=14.3, I=18.5, T=29.7, wait=4.0s, pool=44, epoch=1/2][2026-10-18 16:32:46,780] loop.py:380 INFO TIMING step=1395: step=4.741s data=2.308s compute=4.711s [fwd=3.311s bwd=1.385s opt=0.003s] dispatch=4.467s [2026-10-18 16:32:46,781] loop.py:45 INFO TRAIN_STEP step=1395 loss=2.335080 Training: 8%|▊ | 1395/18018 [3:02:32<38:11:43, 8.27s/step, loss=2.335, acc=0.548, acc_len=1.23, thru=13.3, I=12.2, T=27.0, wait=4.5s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:32:54,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:32:56,874] loop.py:45 INFO TRAIN_STEP step=1396 loss=2.289552 Training: 8%|▊ | 1396/18018 [3:02:42<40:42:50, 8.82s/step, loss=2.290, acc=0.555, acc_len=1.26, thru=17.4, I=15.7, T=28.2, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:32:56,869] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1395: forward=3156.7ms backward=1313.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:32:57,347] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 16:33:03,624] loop.py:45 INFO TRAIN_STEP step=1397 loss=2.150000 Training: 8%|▊ | 1397/18018 [3:02:49<37:50:54, 8.20s/step, loss=2.150, acc=0.573, acc_len=1.30, thru=16.4, I=21.3, T=34.7, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:08,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:33:10,333] loop.py:45 INFO TRAIN_STEP step=1398 loss=2.039705 Training: 8%|▊ | 1398/18018 [3:02:56<35:47:12, 7.75s/step, loss=2.040, acc=0.587, acc_len=1.35, thru=18.9, I=23.3, T=31.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:10,734] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:16,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:33:17,380] loop.py:45 INFO TRAIN_STEP step=1399 loss=2.119000 Training: 8%|▊ | 1399/18018 [3:03:03<34:48:25, 7.54s/step, loss=2.119, acc=0.575, acc_len=1.32, thru=15.3, I=21.0, T=30.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:17,576] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:21,840] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:33:23,334] loop.py:380 INFO TIMING step=1400: step=3.927s data=1.697s compute=3.894s [fwd=2.697s bwd=1.182s opt=0.003s] dispatch=2.018s [2026-10-18 16:33:23,334] loop.py:45 INFO TRAIN_STEP step=1400 loss=2.291188 Training: 8%|▊ | 1400/18018 [3:03:09<32:36:32, 7.06s/step, loss=2.291, acc=0.557, acc_len=1.26, thru=19.1, I=25.6, T=32.6, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:23,365] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 16:33:30,587] loop.py:45 INFO TRAIN_STEP step=1401 loss=2.098483 Training: 8%|▊ | 1401/18018 [3:03:16<32:52:19, 7.12s/step, loss=2.098, acc=0.579, acc_len=1.32, thru=15.4, I=20.8, T=32.9, wait=3.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:33:30,583] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1400: forward=2717.6ms backward=1116.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:36,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:33:37,903] loop.py:45 INFO TRAIN_STEP step=1402 loss=2.259448 Training: 8%|▊ | 1402/18018 [3:03:23<33:08:17, 7.18s/step, loss=2.259, acc=0.556, acc_len=1.26, thru=17.3, I=19.4, T=28.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:37,976] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:43,637] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:33:44,709] loop.py:45 INFO TRAIN_STEP step=1403 loss=2.145536 Training: 8%|▊ | 1403/18018 [3:03:30<32:37:03, 7.07s/step, loss=2.146, acc=0.572, acc_len=1.30, thru=17.1, I=19.7, T=33.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:45,147] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:49,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:33:52,322] loop.py:45 INFO TRAIN_STEP step=1404 loss=2.242985 Training: 8%|▊ | 1404/18018 [3:03:38<33:22:19, 7.23s/step, loss=2.243, acc=0.559, acc_len=1.27, thru=15.7, I=24.3, T=30.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:33:52,358] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 [2026-10-18 16:33:59,593] loop.py:380 INFO TIMING step=1405: step=4.186s data=1.950s compute=4.152s [fwd=2.920s bwd=1.217s opt=0.003s] dispatch=3.073s [2026-10-18 16:33:59,593] loop.py:45 INFO TRAIN_STEP step=1405 loss=2.187891 Training: 8%|▊ | 1405/18018 [3:03:45<33:25:23, 7.24s/step, loss=2.188, acc=0.561, acc_len=1.28, thru=13.3, I=17.8, T=30.6, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:05,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:34:07,594] loop.py:45 INFO TRAIN_STEP step=1406 loss=2.438828 Training: 8%|▊ | 1406/18018 [3:03:53<34:28:18, 7.47s/step, loss=2.439, acc=0.534, acc_len=1.19, thru=18.2, I=19.4, T=31.4, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:34:07,589] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1405: forward=2887.7ms backward=1147.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:07,949] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:13,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:34:15,498] loop.py:45 INFO TRAIN_STEP step=1407 loss=2.429089 Training: 8%|▊ | 1407/18018 [3:04:01<35:04:12, 7.60s/step, loss=2.429, acc=0.529, acc_len=1.19, thru=13.7, I=21.9, T=28.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:15,996] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:21,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:34:23,464] loop.py:45 INFO TRAIN_STEP step=1408 loss=2.339412 Training: 8%|▊ | 1408/18018 [3:04:09<35:34:19, 7.71s/step, loss=2.339, acc=0.542, acc_len=1.22, thru=13.2, I=19.8, T=30.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:23,512] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:28,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:34:29,689] loop.py:45 INFO TRAIN_STEP step=1409 loss=2.292711 Training: 8%|▊ | 1409/18018 [3:04:15<33:31:03, 7.26s/step, loss=2.293, acc=0.553, acc_len=1.24, thru=18.6, I=22.7, T=31.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:29,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:37,265] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:34:37,954] loop.py:380 INFO TIMING step=1410: step=4.084s data=1.746s compute=4.040s [fwd=2.790s bwd=1.237s opt=0.003s] dispatch=4.170s [2026-10-18 16:34:37,954] loop.py:45 INFO TRAIN_STEP step=1410 loss=2.195000 Training: 8%|▊ | 1410/18018 [3:04:24<34:53:47, 7.56s/step, loss=2.195, acc=0.566, acc_len=1.29, thru=14.6, I=17.2, T=31.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:38,271] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:41,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:34:43,689] loop.py:45 INFO TRAIN_STEP step=1411 loss=2.292244 Training: 8%|▊ | 1411/18018 [3:04:29<32:21:59, 7.02s/step, loss=2.292, acc=0.548, acc_len=1.23, thru=19.9, I=24.3, T=32.1, wait=1.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:34:43,684] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1410: forward=2788.4ms backward=1157.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:44,009] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:50,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:34:50,359] loop.py:45 INFO TRAIN_STEP step=1412 loss=2.273401 Training: 8%|▊ | 1412/18018 [3:04:36<31:52:52, 6.91s/step, loss=2.273, acc=0.560, acc_len=1.26, thru=18.2, I=23.1, T=34.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:50,782] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:57,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:34:58,910] loop.py:45 INFO TRAIN_STEP step=1413 loss=2.301764 Training: 8%|▊ | 1413/18018 [3:04:44<34:09:03, 7.40s/step, loss=2.302, acc=0.552, acc_len=1.24, thru=15.4, I=18.7, T=30.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:34:59,064] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:07,265] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:35:07,822] loop.py:45 INFO TRAIN_STEP step=1414 loss=2.289959 Training: 8%|▊ | 1414/18018 [3:04:53<36:14:01, 7.86s/step, loss=2.290, acc=0.550, acc_len=1.24, thru=14.6, I=15.6, T=29.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:08,269] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:14,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:35:14,689] loop.py:380 INFO TIMING step=1415: step=3.526s data=1.416s compute=3.482s [fwd=2.486s bwd=0.983s opt=0.003s] dispatch=3.333s [2026-10-18 16:35:14,689] loop.py:45 INFO TRAIN_STEP step=1415 loss=2.250029 Training: 8%|▊ | 1415/18018 [3:05:00<34:51:52, 7.56s/step, loss=2.250, acc=0.555, acc_len=1.26, thru=18.4, I=19.6, T=36.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:14,945] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:21,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:35:21,952] loop.py:45 INFO TRAIN_STEP step=1416 loss=2.302450 Training: 8%|▊ | 1416/18018 [3:05:08<34:27:09, 7.47s/step, loss=2.302, acc=0.554, acc_len=1.25, thru=18.1, I=20.0, T=28.8, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:35:21,948] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1415: forward=3121.9ms backward=1279.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:22,374] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:29,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:35:30,562] loop.py:45 INFO TRAIN_STEP step=1417 loss=2.150836 Training: 8%|▊ | 1417/18018 [3:05:16<36:01:33, 7.81s/step, loss=2.151, acc=0.574, acc_len=1.31, thru=14.4, I=16.4, T=29.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:30,728] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:38,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:35:39,358] loop.py:45 INFO TRAIN_STEP step=1418 loss=2.213758 Training: 8%|▊ | 1418/18018 [3:05:25<37:22:57, 8.11s/step, loss=2.214, acc=0.567, acc_len=1.28, thru=14.8, I=17.0, T=26.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:35:39,607] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:35:46,675] loop.py:45 INFO TRAIN_STEP step=1419 loss=2.211344 Training: 8%|▊ | 1419/18018 [3:05:32<36:17:16, 7.87s/step, loss=2.211, acc=0.562, acc_len=1.28, thru=16.5, I=17.2, T=34.2, wait=3.6s, pool=60, epoch=1/2][2026-10-18 16:35:55,630] loop.py:380 INFO TIMING step=1420: step=4.315s data=1.880s compute=4.282s [fwd=2.974s bwd=1.295s opt=0.003s] dispatch=4.631s [2026-10-18 16:35:55,630] loop.py:45 INFO TRAIN_STEP step=1420 loss=2.027705 Training: 8%|▊ | 1420/18018 [3:05:41<37:47:31, 8.20s/step, loss=2.028, acc=0.589, acc_len=1.36, thru=11.7, I=13.5, T=29.7, wait=4.6s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:03,576] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:36:04,671] loop.py:45 INFO TRAIN_STEP step=1421 loss=2.243103 Training: 8%|▊ | 1421/18018 [3:05:50<38:57:06, 8.45s/step, loss=2.243, acc=0.560, acc_len=1.27, thru=16.9, I=15.2, T=33.7, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:36:04,666] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1420: forward=2602.1ms backward=1137.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:05,089] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:09,441] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:36:12,156] loop.py:45 INFO TRAIN_STEP step=1422 loss=2.233788 Training: 8%|▊ | 1422/18018 [3:05:58<37:37:03, 8.16s/step, loss=2.234, acc=0.555, acc_len=1.25, thru=17.1, I=24.1, T=25.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:12,461] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 16:36:20,807] loop.py:45 INFO TRAIN_STEP step=1423 loss=2.258856 Training: 8%|▊ | 1423/18018 [3:06:06<38:17:41, 8.31s/step, loss=2.259, acc=0.554, acc_len=1.26, thru=14.4, I=16.1, T=30.7, wait=4.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:27,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:36:28,392] loop.py:45 INFO TRAIN_STEP step=1424 loss=2.208185 Training: 8%|▊ | 1424/18018 [3:06:14<37:17:32, 8.09s/step, loss=2.208, acc=0.563, acc_len=1.28, thru=17.0, I=18.2, T=32.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:28,775] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:34,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:36:34,772] loop.py:380 INFO TIMING step=1425: step=3.632s data=1.492s compute=3.598s [fwd=2.530s bwd=1.054s opt=0.003s] dispatch=2.740s [2026-10-18 16:36:34,772] loop.py:45 INFO TRAIN_STEP step=1425 loss=2.438177 Training: 8%|▊ | 1425/18018 [3:06:20<34:55:26, 7.58s/step, loss=2.438, acc=0.529, acc_len=1.18, thru=18.8, I=21.2, T=35.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:34,940] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 16:36:40,773] loop.py:45 INFO TRAIN_STEP step=1426 loss=2.056401 Training: 8%|▊ | 1426/18018 [3:06:26<32:44:41, 7.10s/step, loss=2.056, acc=0.582, acc_len=1.33, thru=18.4, I=19.0, T=34.4, wait=2.3s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:36:40,768] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1425: forward=2600.4ms backward=1070.2ms [2026-10-18 16:36:47,735] loop.py:45 INFO TRAIN_STEP step=1427 loss=2.317800 Training: 8%|▊ | 1427/18018 [3:06:33<32:32:36, 7.06s/step, loss=2.318, acc=0.547, acc_len=1.23, thru=18.1, I=18.2, T=33.6, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:55,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:36:56,912] loop.py:45 INFO TRAIN_STEP step=1428 loss=2.113960 Training: 8%|▊ | 1428/18018 [3:06:42<35:28:00, 7.70s/step, loss=2.114, acc=0.573, acc_len=1.31, thru=15.1, I=17.1, T=25.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:36:57,399] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:37:04,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:37:06,223] loop.py:45 INFO TRAIN_STEP step=1429 loss=2.099795 Training: 8%|▊ | 1429/18018 [3:06:52<37:41:44, 8.18s/step, loss=2.100, acc=0.582, acc_len=1.33, thru=14.5, I=15.9, T=25.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:37:06,497] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:37:12,910] loop.py:380 INFO TIMING step=1430: step=4.573s data=2.524s compute=4.532s [fwd=3.491s bwd=1.027s opt=0.003s] dispatch=2.107s [2026-10-18 16:37:12,910] loop.py:45 INFO TRAIN_STEP step=1430 loss=2.171216 Training: 8%|▊ | 1430/18018 [3:06:58<35:37:42, 7.73s/step, loss=2.171, acc=0.564, acc_len=1.29, thru=15.4, I=19.1, T=28.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:37:12,860] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:37:13,363] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:37:18,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:37:20,276] loop.py:45 INFO TRAIN_STEP step=1431 loss=2.228658 Training: 8%|▊ | 1431/18018 [3:07:06<35:07:09, 7.62s/step, loss=2.229, acc=0.562, acc_len=1.27, thru=16.2, I=18.3, T=29.3, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:37:20,272] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1430: forward=3135.4ms backward=1182.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:37:20,714] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:37:27,044] loop.py:45 INFO TRAIN_STEP step=1432 loss=2.011809 Training: 8%|▊ | 1432/18018 [3:07:13<33:56:25, 7.37s/step, loss=2.012, acc=0.594, acc_len=1.36, thru=16.9, I=23.3, T=35.1, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:37:31,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:37:33,456] loop.py:45 INFO TRAIN_STEP step=1433 loss=2.158704 Training: 8%|▊ | 1433/18018 [3:07:19<32:37:09, 7.08s/step, loss=2.159, acc=0.570, acc_len=1.30, thru=19.2, I=21.5, T=34.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:37:33,515] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 16:37:42,870] loop.py:45 INFO TRAIN_STEP step=1434 loss=2.194528 Training: 8%|▊ | 1434/18018 [3:07:28<35:50:37, 7.78s/step, loss=2.195, acc=0.564, acc_len=1.27, thru=12.9, I=13.4, T=30.4, wait=5.2s, pool=56, epoch=1/2][2026-10-18 16:37:50,143] loop.py:380 INFO TIMING step=1435: step=3.837s data=1.685s compute=3.799s [fwd=2.711s bwd=1.074s opt=0.003s] dispatch=3.424s [2026-10-18 16:37:50,143] loop.py:45 INFO TRAIN_STEP step=1435 loss=2.261887 Training: 8%|▊ | 1435/18018 [3:07:36<35:08:12, 7.63s/step, loss=2.262, acc=0.561, acc_len=1.27, thru=15.1, I=16.7, T=33.4, wait=3.4s, pool=44, epoch=1/2][2026-10-18 16:37:58,638] loop.py:45 INFO TRAIN_STEP step=1436 loss=2.254244 Training: 8%|▊ | 1436/18018 [3:07:44<36:19:54, 7.89s/step, loss=2.254, acc=0.555, acc_len=1.25, thru=13.0, I=13.7, T=30.1, wait=4.2s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:37:58,634] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1435: forward=2895.3ms backward=1280.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:06,416] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:38:07,079] loop.py:45 INFO TRAIN_STEP step=1437 loss=2.087568 Training: 8%|▊ | 1437/18018 [3:07:53<37:05:48, 8.05s/step, loss=2.088, acc=0.581, acc_len=1.32, thru=18.2, I=16.3, T=32.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:07,421] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:38:14,739] loop.py:45 INFO TRAIN_STEP step=1438 loss=2.061216 Training: 8%|▊ | 1438/18018 [3:08:00<36:32:57, 7.94s/step, loss=2.061, acc=0.587, acc_len=1.35, thru=15.2, I=20.0, T=32.3, wait=3.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:23,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:38:24,557] loop.py:45 INFO TRAIN_STEP step=1439 loss=2.179084 Training: 8%|▊ | 1439/18018 [3:08:10<39:08:41, 8.50s/step, loss=2.179, acc=0.565, acc_len=1.28, thru=15.4, I=14.4, T=30.7, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:24,587] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:29,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:38:30,419] loop.py:380 INFO TIMING step=1440: step=3.822s data=1.753s compute=3.793s [fwd=2.684s bwd=1.095s opt=0.003s] dispatch=2.031s [2026-10-18 16:38:30,419] loop.py:45 INFO TRAIN_STEP step=1440 loss=2.069358 Training: 8%|▊ | 1440/18018 [3:08:16<35:29:51, 7.71s/step, loss=2.069, acc=0.585, acc_len=1.35, thru=19.4, I=22.8, T=33.5, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:30,890] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:38,450] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:38:39,442] loop.py:45 INFO TRAIN_STEP step=1441 loss=2.198975 Training: 8%|▊ | 1441/18018 [3:08:25<37:18:52, 8.10s/step, loss=2.199, acc=0.571, acc_len=1.30, thru=15.0, I=16.2, T=30.7, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:38:39,438] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1440: forward=2926.5ms backward=1198.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:39,482] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:44,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:38:46,027] loop.py:45 INFO TRAIN_STEP step=1442 loss=2.204791 Training: 8%|▊ | 1442/18018 [3:08:32<35:12:41, 7.65s/step, loss=2.205, acc=0.562, acc_len=1.28, thru=17.6, I=23.8, T=32.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:46,410] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:51,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:38:52,748] loop.py:45 INFO TRAIN_STEP step=1443 loss=2.171062 Training: 8%|▊ | 1443/18018 [3:08:38<33:55:48, 7.37s/step, loss=2.171, acc=0.568, acc_len=1.29, thru=16.6, I=23.4, T=32.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:53,174] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:38:59,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:39:00,493] loop.py:45 INFO TRAIN_STEP step=1444 loss=2.119339 Training: 8%|▊ | 1444/18018 [3:08:46<34:26:50, 7.48s/step, loss=2.119, acc=0.580, acc_len=1.32, thru=14.9, I=18.4, T=35.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:00,542] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:05,701] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:39:07,357] loop.py:380 INFO TIMING step=1445: step=3.871s data=1.696s compute=3.841s [fwd=2.800s bwd=1.027s opt=0.003s] dispatch=2.983s [2026-10-18 16:39:07,357] loop.py:45 INFO TRAIN_STEP step=1445 loss=2.355795 Training: 8%|▊ | 1445/18018 [3:08:53<33:35:27, 7.30s/step, loss=2.356, acc=0.537, acc_len=1.21, thru=18.5, I=24.7, T=33.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:07,711] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:14,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:39:15,804] loop.py:45 INFO TRAIN_STEP step=1446 loss=2.136235 Training: 8%|▊ | 1446/18018 [3:09:01<35:10:39, 7.64s/step, loss=2.136, acc=0.570, acc_len=1.30, thru=15.8, I=20.0, T=26.3, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:39:15,798] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1445: forward=3330.6ms backward=1490.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:16,085] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:20,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:39:23,325] loop.py:45 INFO TRAIN_STEP step=1447 loss=2.172057 Training: 8%|▊ | 1447/18018 [3:09:09<35:00:35, 7.61s/step, loss=2.172, acc=0.570, acc_len=1.29, thru=14.1, I=21.6, T=26.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:23,808] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:29,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:39:31,431] loop.py:45 INFO TRAIN_STEP step=1448 loss=2.213758 Training: 8%|▊ | 1448/18018 [3:09:17<35:41:54, 7.76s/step, loss=2.214, acc=0.566, acc_len=1.29, thru=16.8, I=18.0, T=25.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:31,482] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:38,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:39:39,062] loop.py:45 INFO TRAIN_STEP step=1449 loss=2.329906 Training: 8%|▊ | 1449/18018 [3:09:25<35:31:29, 7.72s/step, loss=2.330, acc=0.547, acc_len=1.23, thru=14.8, I=18.5, T=33.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:39,354] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:44,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:39:46,137] loop.py:380 INFO TIMING step=1450: step=3.655s data=1.536s compute=3.619s [fwd=2.598s bwd=1.006s opt=0.003s] dispatch=3.411s [2026-10-18 16:39:46,137] loop.py:45 INFO TRAIN_STEP step=1450 loss=2.419254 Training: 8%|▊ | 1450/18018 [3:09:32<34:37:58, 7.53s/step, loss=2.419, acc=0.539, acc_len=1.20, thru=17.3, I=19.9, T=35.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:46,307] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:51,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:39:52,280] loop.py:45 INFO TRAIN_STEP step=1451 loss=2.102441 Training: 8%|▊ | 1451/18018 [3:09:38<32:43:19, 7.11s/step, loss=2.102, acc=0.583, acc_len=1.33, thru=20.5, I=24.7, T=32.1, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:39:52,276] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1450: forward=2782.3ms backward=1160.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:52,389] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:39:59,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:40:01,117] loop.py:45 INFO TRAIN_STEP step=1452 loss=2.143057 Training: 8%|▊ | 1452/18018 [3:09:47<35:06:24, 7.63s/step, loss=2.143, acc=0.575, acc_len=1.31, thru=14.6, I=17.0, T=31.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:01,324] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:05,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:40:06,809] loop.py:45 INFO TRAIN_STEP step=1453 loss=2.151581 Training: 8%|▊ | 1453/18018 [3:09:52<32:25:45, 7.05s/step, loss=2.152, acc=0.570, acc_len=1.30, thru=19.9, I=26.6, T=32.9, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:06,884] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 16:40:13,048] loop.py:45 INFO TRAIN_STEP step=1454 loss=2.035071 Training: 8%|▊ | 1454/18018 [3:09:59<31:18:37, 6.80s/step, loss=2.035, acc=0.591, acc_len=1.36, thru=19.0, I=23.0, T=33.6, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:20,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:40:21,415] loop.py:380 INFO TIMING step=1455: step=4.371s data=1.906s compute=4.344s [fwd=3.003s bwd=1.327s opt=0.003s] dispatch=3.986s [2026-10-18 16:40:21,416] loop.py:45 INFO TRAIN_STEP step=1455 loss=2.060618 Training: 8%|▊ | 1455/18018 [3:10:07<33:27:55, 7.27s/step, loss=2.061, acc=0.586, acc_len=1.35, thru=16.3, I=17.8, T=29.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:21,857] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:28,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:40:29,600] loop.py:45 INFO TRAIN_STEP step=1456 loss=2.232828 Training: 8%|▊ | 1456/18018 [3:10:15<34:43:06, 7.55s/step, loss=2.233, acc=0.561, acc_len=1.27, thru=14.7, I=18.3, T=29.8, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:40:29,595] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1455: forward=2958.7ms backward=1258.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:29,841] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:36,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:40:37,298] loop.py:45 INFO TRAIN_STEP step=1457 loss=2.148272 Training: 8%|▊ | 1457/18018 [3:10:23<34:55:33, 7.59s/step, loss=2.148, acc=0.573, acc_len=1.30, thru=16.0, I=18.3, T=30.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:37,519] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:42,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:40:44,512] loop.py:45 INFO TRAIN_STEP step=1458 loss=2.158597 Training: 8%|▊ | 1458/18018 [3:10:30<34:24:15, 7.48s/step, loss=2.159, acc=0.570, acc_len=1.29, thru=15.8, I=22.7, T=29.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:44,658] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:50,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:40:51,315] loop.py:45 INFO TRAIN_STEP step=1459 loss=2.051353 Training: 8%|▊ | 1459/18018 [3:10:37<33:27:59, 7.28s/step, loss=2.051, acc=0.587, acc_len=1.35, thru=16.7, I=22.7, T=29.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:40:51,485] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:40:57,329] loop.py:380 INFO TIMING step=1460: step=3.599s data=1.491s compute=3.565s [fwd=2.539s bwd=1.012s opt=0.003s] dispatch=2.406s [2026-10-18 16:40:57,329] loop.py:45 INFO TRAIN_STEP step=1460 loss=2.277132 Training: 8%|▊ | 1460/18018 [3:10:43<31:43:25, 6.90s/step, loss=2.277, acc=0.547, acc_len=1.23, thru=14.9, I=18.3, T=35.6, wait=2.4s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:07,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:41:08,591] loop.py:45 INFO TRAIN_STEP step=1461 loss=2.273642 Training: 8%|▊ | 1461/18018 [3:10:54<37:44:35, 8.21s/step, loss=2.274, acc=0.560, acc_len=1.27, thru=14.9, I=16.2, T=30.2, wait=7.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:41:08,587] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1460: forward=2964.1ms backward=1184.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:08,708] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:14,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:41:15,081] loop.py:45 INFO TRAIN_STEP step=1462 loss=2.175300 Training: 8%|▊ | 1462/18018 [3:11:01<35:22:21, 7.69s/step, loss=2.175, acc=0.567, acc_len=1.28, thru=17.5, I=21.1, T=30.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:15,348] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:20,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:41:22,133] loop.py:45 INFO TRAIN_STEP step=1463 loss=2.164141 Training: 8%|▊ | 1463/18018 [3:11:08<34:29:14, 7.50s/step, loss=2.164, acc=0.570, acc_len=1.29, thru=16.9, I=20.6, T=30.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:22,598] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:27,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:41:28,596] loop.py:45 INFO TRAIN_STEP step=1464 loss=2.104486 Training: 8%|▊ | 1464/18018 [3:11:14<33:03:18, 7.19s/step, loss=2.104, acc=0.578, acc_len=1.33, thru=17.6, I=26.7, T=32.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:29,068] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:38,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:41:39,160] loop.py:380 INFO TIMING step=1465: step=4.585s data=2.198s compute=4.546s [fwd=3.216s bwd=1.315s opt=0.003s] dispatch=5.971s [2026-10-18 16:41:39,160] loop.py:45 INFO TRAIN_STEP step=1465 loss=2.031864 Training: 8%|▊ | 1465/18018 [3:11:25<37:42:48, 8.20s/step, loss=2.032, acc=0.591, acc_len=1.36, thru=12.3, I=13.2, T=27.9, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:39,484] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:44,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:41:46,377] loop.py:45 INFO TRAIN_STEP step=1466 loss=2.257719 Training: 8%|▊ | 1466/18018 [3:11:32<36:20:59, 7.91s/step, loss=2.258, acc=0.554, acc_len=1.26, thru=17.5, I=23.4, T=29.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:41:46,372] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1465: forward=3102.4ms backward=1273.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:46,818] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:52,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:41:52,664] loop.py:45 INFO TRAIN_STEP step=1467 loss=2.073929 Training: 8%|▊ | 1467/18018 [3:11:38<34:06:49, 7.42s/step, loss=2.074, acc=0.584, acc_len=1.35, thru=16.9, I=26.4, T=33.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:53,034] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:58,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:41:59,457] loop.py:45 INFO TRAIN_STEP step=1468 loss=2.089336 Training: 8%|▊ | 1468/18018 [3:11:45<33:15:00, 7.23s/step, loss=2.089, acc=0.578, acc_len=1.32, thru=18.3, I=21.9, T=34.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:41:59,569] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:05,276] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:42:06,571] loop.py:45 INFO TRAIN_STEP step=1469 loss=2.123490 Training: 8%|▊ | 1469/18018 [3:11:52<33:05:02, 7.20s/step, loss=2.123, acc=0.579, acc_len=1.32, thru=17.0, I=23.0, T=28.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:06,785] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:12,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:42:13,408] loop.py:380 INFO TIMING step=1470: step=3.638s data=1.463s compute=3.608s [fwd=2.569s bwd=1.024s opt=0.003s] dispatch=3.189s [2026-10-18 16:42:13,408] loop.py:45 INFO TRAIN_STEP step=1470 loss=1.930147 Training: 8%|▊ | 1470/18018 [3:11:59<32:35:12, 7.09s/step, loss=1.930, acc=0.599, acc_len=1.40, thru=16.8, I=20.5, T=35.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:13,679] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:22,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:42:22,355] loop.py:45 INFO TRAIN_STEP step=1471 loss=2.003864 Training: 8%|▊ | 1471/18018 [3:12:08<35:08:43, 7.65s/step, loss=2.004, acc=0.591, acc_len=1.37, thru=13.9, I=15.4, T=28.1, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:42:22,351] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1470: forward=3232.3ms backward=1266.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:22,533] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:27,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:42:29,507] loop.py:45 INFO TRAIN_STEP step=1472 loss=2.242363 Training: 8%|▊ | 1472/18018 [3:12:15<34:27:43, 7.50s/step, loss=2.242, acc=0.557, acc_len=1.26, thru=17.9, I=20.8, T=28.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:29,649] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:34,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:42:36,227] loop.py:45 INFO TRAIN_STEP step=1473 loss=2.174488 Training: 8%|▊ | 1473/18018 [3:12:22<33:23:06, 7.26s/step, loss=2.174, acc=0.568, acc_len=1.29, thru=16.4, I=25.3, T=31.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:36,712] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:42,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:42:42,452] loop.py:45 INFO TRAIN_STEP step=1474 loss=2.191583 Training: 8%|▊ | 1474/18018 [3:12:28<31:57:02, 6.95s/step, loss=2.192, acc=0.564, acc_len=1.28, thru=18.6, I=24.8, T=35.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:42,860] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:48,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:42:49,015] loop.py:380 INFO TIMING step=1475: step=3.515s data=1.474s compute=3.471s [fwd=2.462s bwd=0.994s opt=0.003s] dispatch=3.038s [2026-10-18 16:42:49,015] loop.py:45 INFO TRAIN_STEP step=1475 loss=2.205346 Training: 8%|▊ | 1475/18018 [3:12:35<31:24:40, 6.84s/step, loss=2.205, acc=0.563, acc_len=1.28, thru=19.4, I=21.2, T=36.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:49,355] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:55,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:42:56,622] loop.py:45 INFO TRAIN_STEP step=1476 loss=1.980426 Training: 8%|▊ | 1476/18018 [3:12:42<32:28:23, 7.07s/step, loss=1.980, acc=0.598, acc_len=1.38, thru=16.5, I=19.9, T=35.6, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:42:56,618] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1475: forward=2435.4ms backward=1099.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:42:56,730] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:04,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:43:05,646] loop.py:45 INFO TRAIN_STEP step=1477 loss=2.176466 Training: 8%|▊ | 1477/18018 [3:12:51<35:10:12, 7.65s/step, loss=2.176, acc=0.571, acc_len=1.29, thru=13.3, I=16.2, T=33.0, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:06,042] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:13,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:43:13,529] loop.py:45 INFO TRAIN_STEP step=1478 loss=2.070399 Training: 8%|▊ | 1478/18018 [3:12:59<35:28:58, 7.72s/step, loss=2.070, acc=0.585, acc_len=1.35, thru=15.6, I=19.0, T=32.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:13,718] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:21,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:43:22,438] loop.py:45 INFO TRAIN_STEP step=1479 loss=2.120061 Training: 8%|▊ | 1479/18018 [3:13:08<37:06:58, 8.08s/step, loss=2.120, acc=0.574, acc_len=1.31, thru=15.7, I=17.9, T=32.8, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:22,564] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:29,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:43:30,113] loop.py:380 INFO TIMING step=1480: step=4.217s data=1.895s compute=4.187s [fwd=2.964s bwd=1.207s opt=0.003s] dispatch=3.448s [2026-10-18 16:43:30,114] loop.py:45 INFO TRAIN_STEP step=1480 loss=2.203948 Training: 8%|▊ | 1480/18018 [3:13:16<36:33:18, 7.96s/step, loss=2.204, acc=0.567, acc_len=1.29, thru=17.1, I=19.2, T=30.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:30,258] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 16:43:37,635] loop.py:45 INFO TRAIN_STEP step=1481 loss=2.397595 Training: 8%|▊ | 1481/18018 [3:13:23<35:57:19, 7.83s/step, loss=2.398, acc=0.539, acc_len=1.21, thru=16.4, I=17.7, T=29.3, wait=3.1s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:43:37,631] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1480: forward=2996.1ms backward=1308.6ms [2026-10-18 16:43:46,049] loop.py:45 INFO TRAIN_STEP step=1482 loss=2.220230 Training: 8%|▊ | 1482/18018 [3:13:32<36:45:39, 8.00s/step, loss=2.220, acc=0.560, acc_len=1.27, thru=13.4, I=14.5, T=27.4, wait=3.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:55,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:43:56,009] loop.py:45 INFO TRAIN_STEP step=1483 loss=2.118116 Training: 8%|▊ | 1483/18018 [3:13:42<39:27:20, 8.59s/step, loss=2.118, acc=0.576, acc_len=1.31, thru=15.2, I=13.5, T=31.3, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:43:56,275] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:03,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:44:04,202] loop.py:45 INFO TRAIN_STEP step=1484 loss=2.139763 Training: 8%|▊ | 1484/18018 [3:13:50<38:54:19, 8.47s/step, loss=2.140, acc=0.573, acc_len=1.31, thru=15.8, I=17.0, T=34.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:04,426] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:09,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:44:11,860] loop.py:380 INFO TIMING step=1485: step=4.859s data=2.236s compute=4.830s [fwd=3.344s bwd=1.471s opt=0.003s] dispatch=2.788s [2026-10-18 16:44:11,860] loop.py:45 INFO TRAIN_STEP step=1485 loss=2.162459 Training: 8%|▊ | 1485/18018 [3:13:57<37:46:58, 8.23s/step, loss=2.162, acc=0.571, acc_len=1.30, thru=17.2, I=22.3, T=26.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:11,901] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:15,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:44:17,373] loop.py:45 INFO TRAIN_STEP step=1486 loss=2.032099 (TrainerActor pid=9310) [2026-10-18 16:44:17,369] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1485: forward=2572.2ms backward=1156.1ms Training: 8%|▊ | 1486/18018 [3:14:03<34:02:21, 7.41s/step, loss=2.032, acc=0.587, acc_len=1.35, thru=18.8, I=26.8, T=33.9, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:17,494] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 16:44:24,288] loop.py:45 INFO TRAIN_STEP step=1487 loss=1.941929 Training: 8%|▊ | 1487/18018 [3:14:10<33:21:08, 7.26s/step, loss=1.942, acc=0.603, acc_len=1.40, thru=15.3, I=19.4, T=31.6, wait=2.9s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:29,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:44:30,711] loop.py:45 INFO TRAIN_STEP step=1488 loss=1.991209 Training: 8%|▊ | 1488/18018 [3:14:16<32:11:29, 7.01s/step, loss=1.991, acc=0.590, acc_len=1.36, thru=19.2, I=20.7, T=36.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:31,214] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:37,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:44:38,164] loop.py:45 INFO TRAIN_STEP step=1489 loss=2.123739 Training: 8%|▊ | 1489/18018 [3:14:24<32:48:12, 7.14s/step, loss=2.124, acc=0.573, acc_len=1.31, thru=16.9, I=21.9, T=33.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:38,460] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:44:44,260] loop.py:380 INFO TIMING step=1490: step=3.624s data=1.514s compute=3.590s [fwd=2.556s bwd=1.020s opt=0.003s] dispatch=2.461s [2026-10-18 16:44:44,261] loop.py:45 INFO TRAIN_STEP step=1490 loss=1.972869 Training: 8%|▊ | 1490/18018 [3:14:30<31:21:21, 6.83s/step, loss=1.973, acc=0.597, acc_len=1.39, thru=17.7, I=19.6, T=35.3, wait=2.5s, pool=44, epoch=1/2][2026-10-18 16:44:52,223] loop.py:45 INFO TRAIN_STEP step=1491 loss=2.253511 Training: 8%|▊ | 1491/18018 [3:14:38<32:54:39, 7.17s/step, loss=2.254, acc=0.561, acc_len=1.27, thru=17.0, I=16.1, T=32.6, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:44:52,218] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1490: forward=2729.1ms backward=1148.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:52,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:52,690] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:57,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:44:58,648] loop.py:45 INFO TRAIN_STEP step=1492 loss=2.043944 Training: 8%|▊ | 1492/18018 [3:14:44<31:53:12, 6.95s/step, loss=2.044, acc=0.586, acc_len=1.35, thru=18.8, I=20.6, T=34.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:44:58,716] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 16:45:04,635] loop.py:45 INFO TRAIN_STEP step=1493 loss=2.192787 Training: 8%|▊ | 1493/18018 [3:14:50<30:33:56, 6.66s/step, loss=2.193, acc=0.565, acc_len=1.28, thru=19.4, I=23.4, T=34.8, wait=2.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:45:11,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:45:12,541] loop.py:45 INFO TRAIN_STEP step=1494 loss=2.194372 Training: 8%|▊ | 1494/18018 [3:14:58<32:16:58, 7.03s/step, loss=2.194, acc=0.565, acc_len=1.28, thru=16.8, I=18.8, T=27.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:45:12,691] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 16:45:18,953] loop.py:380 INFO TIMING step=1495: step=3.463s data=1.409s compute=3.427s [fwd=2.419s bwd=0.993s opt=0.003s] dispatch=2.938s [2026-10-18 16:45:18,954] loop.py:45 INFO TRAIN_STEP step=1495 loss=2.008649 Training: 8%|▊ | 1495/18018 [3:15:05<31:25:23, 6.85s/step, loss=2.009, acc=0.592, acc_len=1.37, thru=16.8, I=19.0, T=37.0, wait=2.9s, pool=60, epoch=1/2][2026-10-18 16:45:28,395] loop.py:45 INFO TRAIN_STEP step=1496 loss=2.265162 Training: 8%|▊ | 1496/18018 [3:15:14<34:59:30, 7.62s/step, loss=2.265, acc=0.559, acc_len=1.26, thru=12.9, I=12.2, T=31.5, wait=5.4s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:45:28,391] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1495: forward=2919.8ms backward=1096.2ms [2026-10-18 16:45:38,906] loop.py:45 INFO TRAIN_STEP step=1497 loss=2.090800 Training: 8%|▊ | 1497/18018 [3:15:24<38:58:06, 8.49s/step, loss=2.091, acc=0.584, acc_len=1.34, thru=11.5, I=11.7, T=30.2, wait=6.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:45:45,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:45:45,982] loop.py:45 INFO TRAIN_STEP step=1498 loss=1.985987 Training: 8%|▊ | 1498/18018 [3:15:32<37:00:52, 8.07s/step, loss=1.986, acc=0.598, acc_len=1.38, thru=19.7, I=15.1, T=30.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:45:46,442] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:45:53,531] loop.py:45 INFO TRAIN_STEP step=1499 loss=1.977531 Training: 8%|▊ | 1499/18018 [3:15:39<36:18:00, 7.91s/step, loss=1.978, acc=0.594, acc_len=1.37, thru=17.4, I=18.5, T=29.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:45:53,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:45:53,961] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:45:59,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:45:59,794] loop.py:380 INFO TIMING step=1500: step=3.682s data=1.536s compute=3.655s [fwd=2.566s bwd=1.074s opt=0.003s] dispatch=2.572s [2026-10-18 16:45:59,794] loop.py:45 INFO TRAIN_STEP step=1500 loss=2.194100 Training: 8%|▊ | 1500/18018 [3:15:45<34:01:44, 7.42s/step, loss=2.194, acc=0.567, acc_len=1.28, thru=19.5, I=19.0, T=34.8, wait=2.6s, pool=64, epoch=1/2][2026-10-18 16:45:59,797] loop.py:419 INFO Saving checkpoint at step 1500... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 16:46:03,093] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0001501 (TrainerActor pid=9310) [2026-10-18 16:46:03,287] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0000751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:04,192] inference_manager.py:408 INFO Pool capacity available after 4.6s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:09,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 16:46:03,091] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0001501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 16:46:04,147] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0000751 [2026-10-18 16:46:11,483] loop.py:45 INFO TRAIN_STEP step=1501 loss=2.308143 Training: 8%|▊ | 1501/18018 [3:15:57<39:54:30, 8.70s/step, loss=2.308, acc=0.553, acc_len=1.24, thru=17.6, I=22.6, T=27.5, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:46:11,478] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1500: forward=3296.3ms backward=1314.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:11,614] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-18 16:46:17,746] loop.py:45 INFO TRAIN_STEP step=1502 loss=2.030388 Training: 8%|▊ | 1502/18018 [3:16:03<36:33:09, 7.97s/step, loss=2.030, acc=0.591, acc_len=1.36, thru=16.9, I=22.8, T=36.9, wait=2.8s, pool=60, epoch=1/2][2026-10-18 16:46:25,674] loop.py:45 INFO TRAIN_STEP step=1503 loss=2.121800 Training: 8%|▊ | 1503/18018 [3:16:11<36:29:54, 7.96s/step, loss=2.122, acc=0.576, acc_len=1.31, thru=16.1, I=18.8, T=27.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:25,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:26,134] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:36,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:46:37,440] loop.py:45 INFO TRAIN_STEP step=1504 loss=2.111576 Training: 8%|▊ | 1504/18018 [3:16:23<41:44:28, 9.10s/step, loss=2.112, acc=0.577, acc_len=1.31, thru=11.2, I=11.6, T=26.7, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:37,779] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:42,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:46:43,960] loop.py:380 INFO TIMING step=1505: step=4.003s data=1.599s compute=3.974s [fwd=2.767s bwd=1.193s opt=0.003s] dispatch=2.506s [2026-10-18 16:46:43,960] loop.py:45 INFO TRAIN_STEP step=1505 loss=2.250367 Training: 8%|▊ | 1505/18018 [3:16:30<38:11:11, 8.33s/step, loss=2.250, acc=0.554, acc_len=1.26, thru=18.3, I=20.9, T=32.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:44,352] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:51,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:46:52,065] loop.py:45 INFO TRAIN_STEP step=1506 loss=2.037538 Training: 8%|▊ | 1506/18018 [3:16:38<37:52:55, 8.26s/step, loss=2.038, acc=0.585, acc_len=1.35, thru=13.9, I=17.5, T=34.6, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:46:52,061] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1505: forward=2604.9ms backward=1036.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:46:52,538] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:46:58,978] loop.py:45 INFO TRAIN_STEP step=1507 loss=2.135385 Training: 8%|▊ | 1507/18018 [3:16:45<36:01:58, 7.86s/step, loss=2.135, acc=0.577, acc_len=1.31, thru=17.9, I=20.2, T=36.0, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:07,860] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:47:09,223] loop.py:45 INFO TRAIN_STEP step=1508 loss=2.149700 Training: 8%|▊ | 1508/18018 [3:16:55<39:18:49, 8.57s/step, loss=2.150, acc=0.578, acc_len=1.31, thru=13.8, I=14.2, T=29.7, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:09,374] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:14,840] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:47:15,515] loop.py:45 INFO TRAIN_STEP step=1509 loss=1.931914 Training: 8%|▊ | 1509/18018 [3:17:01<36:10:25, 7.89s/step, loss=1.932, acc=0.610, acc_len=1.41, thru=18.1, I=22.1, T=31.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:15,844] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:47:22,624] loop.py:380 INFO TIMING step=1510: step=4.077s data=1.876s compute=4.046s [fwd=2.871s bwd=1.161s opt=0.003s] dispatch=3.023s [2026-10-18 16:47:22,624] loop.py:45 INFO TRAIN_STEP step=1510 loss=2.236730 Training: 8%|▊ | 1510/18018 [3:17:08<35:05:51, 7.65s/step, loss=2.237, acc=0.557, acc_len=1.26, thru=16.4, I=17.7, T=31.4, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:30,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:47:30,722] loop.py:45 INFO TRAIN_STEP step=1511 loss=2.223145 Training: 8%|▊ | 1511/18018 [3:17:16<35:42:20, 7.79s/step, loss=2.223, acc=0.560, acc_len=1.26, thru=17.5, I=16.4, T=35.0, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:47:30,718] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1510: forward=2588.6ms backward=1024.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:31,193] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:36,357] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:47:36,796] loop.py:45 INFO TRAIN_STEP step=1512 loss=1.978912 Training: 8%|▊ | 1512/18018 [3:17:22<33:21:00, 7.27s/step, loss=1.979, acc=0.592, acc_len=1.37, thru=20.1, I=20.0, T=35.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:36,868] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:42,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:47:43,259] loop.py:45 INFO TRAIN_STEP step=1513 loss=2.187158 Training: 8%|▊ | 1513/18018 [3:17:29<32:13:59, 7.03s/step, loss=2.187, acc=0.566, acc_len=1.29, thru=19.5, I=21.5, T=32.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:47:43,478] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 16:47:51,700] loop.py:45 INFO TRAIN_STEP step=1514 loss=2.245805 Training: 8%|▊ | 1514/18018 [3:17:37<34:10:14, 7.45s/step, loss=2.246, acc=0.562, acc_len=1.27, thru=14.4, I=15.4, T=34.0, wait=4.7s, pool=56, epoch=1/2][2026-10-18 16:48:00,582] loop.py:380 INFO TIMING step=1515: step=3.842s data=1.615s compute=3.809s [fwd=2.737s bwd=1.057s opt=0.003s] dispatch=5.031s [2026-10-18 16:48:00,582] loop.py:45 INFO TRAIN_STEP step=1515 loss=2.276263 Training: 8%|▊ | 1515/18018 [3:17:46<36:07:45, 7.88s/step, loss=2.276, acc=0.563, acc_len=1.27, thru=12.8, I=12.7, T=33.3, wait=5.0s, pool=32, epoch=1/2][2026-10-18 16:48:11,380] loop.py:45 INFO TRAIN_STEP step=1516 loss=2.414405 Training: 8%|▊ | 1516/18018 [3:17:57<40:08:24, 8.76s/step, loss=2.414, acc=0.538, acc_len=1.19, thru=14.7, I=14.8, T=28.4, wait=6.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:48:11,375] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1515: forward=3188.3ms backward=1262.8ms [2026-10-18 16:48:19,333] loop.py:45 INFO TRAIN_STEP step=1517 loss=2.265486 Training: 8%|▊ | 1517/18018 [3:18:05<39:02:04, 8.52s/step, loss=2.265, acc=0.559, acc_len=1.26, thru=15.0, I=15.6, T=28.9, wait=3.5s, pool=52, epoch=1/2][2026-10-18 16:48:26,855] loop.py:45 INFO TRAIN_STEP step=1518 loss=2.102166 Training: 8%|▊ | 1518/18018 [3:18:12<37:39:50, 8.22s/step, loss=2.102, acc=0.580, acc_len=1.32, thru=15.0, I=15.4, T=31.3, wait=3.4s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:48:36,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:48:37,089] loop.py:45 INFO TRAIN_STEP step=1519 loss=2.118762 Training: 8%|▊ | 1519/18018 [3:18:23<40:25:51, 8.82s/step, loss=2.119, acc=0.582, acc_len=1.33, thru=15.0, I=14.7, T=30.9, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:48:37,284] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:48:45,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:48:45,970] loop.py:380 INFO TIMING step=1520: step=4.580s data=2.127s compute=4.522s [fwd=3.163s bwd=1.344s opt=0.003s] dispatch=4.293s [2026-10-18 16:48:45,970] loop.py:45 INFO TRAIN_STEP step=1520 loss=2.133610 Training: 8%|▊ | 1520/18018 [3:18:32<40:30:52, 8.84s/step, loss=2.134, acc=0.573, acc_len=1.31, thru=14.0, I=16.0, T=27.9, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:48:46,072] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:48:54,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:48:55,609] loop.py:45 INFO TRAIN_STEP step=1521 loss=2.098807 Training: 8%|▊ | 1521/18018 [3:18:41<41:36:35, 9.08s/step, loss=2.099, acc=0.587, acc_len=1.34, thru=13.4, I=14.3, T=29.4, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:48:55,649] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 16:48:55,605] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1520: forward=3130.5ms backward=1172.8ms [2026-10-18 16:49:02,605] loop.py:45 INFO TRAIN_STEP step=1522 loss=2.180719 Training: 8%|▊ | 1522/18018 [3:18:48<38:44:19, 8.45s/step, loss=2.181, acc=0.570, acc_len=1.29, thru=16.2, I=17.3, T=34.8, wait=3.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:12,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:49:12,535] loop.py:45 INFO TRAIN_STEP step=1523 loss=2.170606 Training: 8%|▊ | 1523/18018 [3:18:58<40:45:54, 8.90s/step, loss=2.171, acc=0.572, acc_len=1.29, thru=14.1, I=15.4, T=25.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:12,882] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:18,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:49:21,248] loop.py:45 INFO TRAIN_STEP step=1524 loss=2.058946 Training: 8%|▊ | 1524/18018 [3:19:07<40:30:44, 8.84s/step, loss=2.059, acc=0.585, acc_len=1.34, thru=14.5, I=18.0, T=22.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:21,409] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:27,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:49:27,852] loop.py:380 INFO TIMING step=1525: step=4.156s data=1.943s compute=4.113s [fwd=2.909s bwd=1.189s opt=0.003s] dispatch=2.437s [2026-10-18 16:49:27,852] loop.py:45 INFO TRAIN_STEP step=1525 loss=2.243367 Training: 8%|▊ | 1525/18018 [3:19:13<37:26:02, 8.17s/step, loss=2.243, acc=0.559, acc_len=1.26, thru=14.9, I=22.1, T=30.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:27,889] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:35,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:49:36,392] loop.py:45 INFO TRAIN_STEP step=1526 loss=2.175602 Training: 8%|▊ | 1526/18018 [3:19:22<37:56:18, 8.28s/step, loss=2.176, acc=0.568, acc_len=1.29, thru=14.9, I=16.4, T=28.1, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:49:36,386] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1525: forward=3162.6ms backward=1330.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:36,429] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:42,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:49:43,434] loop.py:45 INFO TRAIN_STEP step=1527 loss=2.125219 Training: 8%|▊ | 1527/18018 [3:19:29<36:13:55, 7.91s/step, loss=2.125, acc=0.576, acc_len=1.32, thru=16.1, I=20.5, T=29.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:43,646] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:47,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:49:49,858] loop.py:45 INFO TRAIN_STEP step=1528 loss=2.236470 Training: 8%|▊ | 1528/18018 [3:19:35<34:11:14, 7.46s/step, loss=2.236, acc=0.562, acc_len=1.28, thru=17.5, I=27.9, T=31.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:50,055] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:56,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:49:57,482] loop.py:45 INFO TRAIN_STEP step=1529 loss=1.987852 Training: 8%|▊ | 1529/18018 [3:19:43<34:24:20, 7.51s/step, loss=1.988, acc=0.592, acc_len=1.36, thru=13.3, I=19.1, T=37.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:49:57,514] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:50:03,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:50:04,209] loop.py:380 INFO TIMING step=1530: step=4.020s data=1.805s compute=3.989s [fwd=2.810s bwd=1.164s opt=0.003s] dispatch=2.698s [2026-10-18 16:50:04,209] loop.py:45 INFO TRAIN_STEP step=1530 loss=2.074211 Training: 8%|▊ | 1530/18018 [3:19:50<33:19:38, 7.28s/step, loss=2.074, acc=0.586, acc_len=1.34, thru=19.4, I=19.1, T=31.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:50:04,374] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:50:11,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:50:11,785] loop.py:45 INFO TRAIN_STEP step=1531 loss=2.081522 (TrainerActor pid=9310) [2026-10-18 16:50:11,781] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1530: forward=2773.3ms backward=1160.2ms Training: 8%|▊ | 1531/18018 [3:19:57<33:44:05, 7.37s/step, loss=2.082, acc=0.582, acc_len=1.34, thru=17.2, I=17.1, T=32.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:50:12,179] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:50:18,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:50:18,574] loop.py:45 INFO TRAIN_STEP step=1532 loss=2.001965 Training: 9%|▊ | 1532/18018 [3:20:04<32:56:21, 7.19s/step, loss=2.002, acc=0.597, acc_len=1.37, thru=18.1, I=18.4, T=35.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:50:18,720] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 16:50:25,877] loop.py:45 INFO TRAIN_STEP step=1533 loss=2.309027 Training: 9%|▊ | 1533/18018 [3:20:11<33:05:13, 7.23s/step, loss=2.309, acc=0.551, acc_len=1.25, thru=15.8, I=16.2, T=32.4, wait=3.3s, pool=48, epoch=1/2][2026-10-18 16:50:34,572] loop.py:45 INFO TRAIN_STEP step=1534 loss=1.857520 Training: 9%|▊ | 1534/18018 [3:20:20<35:06:27, 7.67s/step, loss=1.858, acc=0.622, acc_len=1.45, thru=15.6, I=14.5, T=33.6, wait=4.9s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:50:45,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:50:47,063] loop.py:380 INFO TIMING step=1535: step=4.668s data=2.282s compute=4.580s [fwd=3.135s bwd=1.430s opt=0.003s] dispatch=7.812s [2026-10-18 16:50:47,063] loop.py:45 INFO TRAIN_STEP step=1535 loss=2.273421 Training: 9%|▊ | 1535/18018 [3:20:33<41:43:50, 9.11s/step, loss=2.273, acc=0.560, acc_len=1.26, thru=13.2, I=12.7, T=27.4, wait=7.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:50:47,170] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 16:50:53,760] loop.py:45 INFO TRAIN_STEP step=1536 loss=2.037503 Training: 9%|▊ | 1536/18018 [3:20:39<38:24:31, 8.39s/step, loss=2.038, acc=0.592, acc_len=1.36, thru=15.7, I=19.3, T=32.1, wait=2.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:50:53,755] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1535: forward=2834.9ms backward=1098.0ms [2026-10-18 16:51:02,155] loop.py:45 INFO TRAIN_STEP step=1537 loss=2.232222 Training: 9%|▊ | 1537/18018 [3:20:48<38:24:49, 8.39s/step, loss=2.232, acc=0.564, acc_len=1.28, thru=15.7, I=15.5, T=30.2, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:06,926] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:51:08,556] loop.py:45 INFO TRAIN_STEP step=1538 loss=2.194620 Training: 9%|▊ | 1538/18018 [3:20:54<35:40:35, 7.79s/step, loss=2.195, acc=0.568, acc_len=1.29, thru=18.1, I=22.0, T=31.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:08,941] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:15,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:51:15,602] loop.py:45 INFO TRAIN_STEP step=1539 loss=2.108706 Training: 9%|▊ | 1539/18018 [3:21:01<34:38:55, 7.57s/step, loss=2.109, acc=0.581, acc_len=1.33, thru=16.4, I=21.8, T=32.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:16,020] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:21,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:51:22,203] loop.py:380 INFO TIMING step=1540: step=4.304s data=1.974s compute=4.270s [fwd=2.958s bwd=1.297s opt=0.003s] dispatch=2.287s [2026-10-18 16:51:22,204] loop.py:45 INFO TRAIN_STEP step=1540 loss=2.094660 Training: 9%|▊ | 1540/18018 [3:21:08<33:19:11, 7.28s/step, loss=2.095, acc=0.575, acc_len=1.32, thru=18.6, I=19.9, T=29.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:22,284] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:31,425] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:51:33,424] loop.py:45 INFO TRAIN_STEP step=1541 loss=2.059358 Training: 9%|▊ | 1541/18018 [3:21:19<38:43:37, 8.46s/step, loss=2.059, acc=0.582, acc_len=1.33, thru=11.3, I=13.3, T=24.3, wait=5.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:51:33,419] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1540: forward=3923.2ms backward=1295.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:33,457] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:38,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:51:40,299] loop.py:45 INFO TRAIN_STEP step=1542 loss=2.175283 Training: 9%|▊ | 1542/18018 [3:21:26<36:32:54, 7.99s/step, loss=2.175, acc=0.567, acc_len=1.29, thru=15.4, I=23.0, T=29.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:40,333] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:48,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:51:49,114] loop.py:45 INFO TRAIN_STEP step=1543 loss=1.995221 Training: 9%|▊ | 1543/18018 [3:21:35<37:41:02, 8.23s/step, loss=1.995, acc=0.594, acc_len=1.36, thru=14.6, I=16.1, T=27.1, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:51:49,514] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:51:56,897] loop.py:45 INFO TRAIN_STEP step=1544 loss=2.095647 Training: 9%|▊ | 1544/18018 [3:21:42<37:03:40, 8.10s/step, loss=2.096, acc=0.577, acc_len=1.32, thru=15.0, I=18.8, T=26.2, wait=2.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:07,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:52:09,104] loop.py:380 INFO TIMING step=1545: step=4.852s data=2.471s compute=4.816s [fwd=3.609s bwd=1.193s opt=0.003s] dispatch=7.347s [2026-10-18 16:52:09,105] loop.py:45 INFO TRAIN_STEP step=1545 loss=2.142889 Training: 9%|▊ | 1545/18018 [3:21:55<42:42:00, 9.33s/step, loss=2.143, acc=0.574, acc_len=1.30, thru=12.3, I=12.8, T=26.4, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:09,327] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:15,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:52:16,670] loop.py:45 INFO TRAIN_STEP step=1546 loss=2.085111 Training: 9%|▊ | 1546/18018 [3:22:02<40:16:24, 8.80s/step, loss=2.085, acc=0.586, acc_len=1.34, thru=14.2, I=18.6, T=29.1, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:52:16,665] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1545: forward=3033.7ms backward=1315.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:16,823] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:24,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:52:25,559] loop.py:45 INFO TRAIN_STEP step=1547 loss=2.157490 Training: 9%|▊ | 1547/18018 [3:22:11<40:23:17, 8.83s/step, loss=2.157, acc=0.574, acc_len=1.30, thru=14.7, I=16.7, T=25.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:25,899] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:32,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:52:32,852] loop.py:45 INFO TRAIN_STEP step=1548 loss=2.150459 Training: 9%|▊ | 1548/18018 [3:22:18<38:16:38, 8.37s/step, loss=2.150, acc=0.572, acc_len=1.31, thru=14.6, I=16.2, T=29.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:33,028] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:52:39,211] loop.py:45 INFO TRAIN_STEP step=1549 loss=2.067961 Training: 9%|▊ | 1549/18018 [3:22:25<35:31:11, 7.76s/step, loss=2.068, acc=0.584, acc_len=1.35, thru=19.1, I=19.3, T=33.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:39,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:39,708] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:45,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:52:46,042] loop.py:380 INFO TIMING step=1550: step=3.696s data=1.558s compute=3.667s [fwd=2.581s bwd=1.071s opt=0.003s] dispatch=3.128s [2026-10-18 16:52:46,043] loop.py:45 INFO TRAIN_STEP step=1550 loss=2.054498 Training: 9%|▊ | 1550/18018 [3:22:32<34:14:22, 7.48s/step, loss=2.054, acc=0.581, acc_len=1.33, thru=17.9, I=21.3, T=34.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:46,138] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:51,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:52:53,325] loop.py:45 INFO TRAIN_STEP step=1551 loss=2.288194 Training: 9%|▊ | 1551/18018 [3:22:39<33:57:36, 7.42s/step, loss=2.288, acc=0.550, acc_len=1.24, thru=16.5, I=21.4, T=31.9, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:52:53,321] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1550: forward=2786.3ms backward=1186.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:53,409] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:52:57,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:53:00,179] loop.py:45 INFO TRAIN_STEP step=1552 loss=2.162890 Training: 9%|▊ | 1552/18018 [3:22:46<33:10:38, 7.25s/step, loss=2.163, acc=0.567, acc_len=1.29, thru=16.8, I=25.9, T=29.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:00,229] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:08,983] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:53:09,193] loop.py:45 INFO TRAIN_STEP step=1553 loss=2.076745 Training: 9%|▊ | 1553/18018 [3:22:55<35:35:21, 7.78s/step, loss=2.077, acc=0.587, acc_len=1.34, thru=14.3, I=14.2, T=34.7, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:09,485] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:14,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:53:16,524] loop.py:45 INFO TRAIN_STEP step=1554 loss=2.034560 Training: 9%|▊ | 1554/18018 [3:23:02<34:57:58, 7.65s/step, loss=2.035, acc=0.591, acc_len=1.36, thru=17.0, I=21.6, T=28.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:16,635] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:22,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:53:23,103] loop.py:380 INFO TIMING step=1555: step=3.825s data=1.643s compute=3.785s [fwd=2.652s bwd=1.118s opt=0.003s] dispatch=2.748s [2026-10-18 16:53:23,103] loop.py:45 INFO TRAIN_STEP step=1555 loss=1.986817 Training: 9%|▊ | 1555/18018 [3:23:09<33:30:16, 7.33s/step, loss=1.987, acc=0.599, acc_len=1.39, thru=17.9, I=22.5, T=33.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:23,516] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:28,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:53:29,389] loop.py:45 INFO TRAIN_STEP step=1556 loss=2.035847 Training: 9%|▊ | 1556/18018 [3:23:15<32:04:32, 7.01s/step, loss=2.036, acc=0.587, acc_len=1.35, thru=19.4, I=24.3, T=34.5, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:53:29,384] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1555: forward=2664.5ms backward=1005.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:29,642] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:35,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:53:36,689] loop.py:45 INFO TRAIN_STEP step=1557 loss=2.166737 Training: 9%|▊ | 1557/18018 [3:23:22<32:27:38, 7.10s/step, loss=2.167, acc=0.573, acc_len=1.30, thru=16.2, I=19.8, T=32.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:37,000] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:53:43,465] loop.py:45 INFO TRAIN_STEP step=1558 loss=2.113816 Training: 9%|▊ | 1558/18018 [3:23:29<32:00:58, 7.00s/step, loss=2.114, acc=0.574, acc_len=1.32, thru=16.6, I=18.1, T=34.0, wait=3.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:50,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:53:51,726] loop.py:45 INFO TRAIN_STEP step=1559 loss=2.086490 Training: 9%|▊ | 1559/18018 [3:23:37<33:44:40, 7.38s/step, loss=2.086, acc=0.584, acc_len=1.33, thru=18.4, I=16.8, T=32.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:51,992] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:57,733] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:53:58,559] loop.py:380 INFO TIMING step=1560: step=4.141s data=1.852s compute=4.113s [fwd=2.970s bwd=1.129s opt=0.003s] dispatch=2.681s [2026-10-18 16:53:58,559] loop.py:45 INFO TRAIN_STEP step=1560 loss=2.201304 Training: 9%|▊ | 1560/18018 [3:23:44<32:59:30, 7.22s/step, loss=2.201, acc=0.565, acc_len=1.29, thru=17.6, I=22.1, T=30.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:53:58,740] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:05,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:54:05,297] loop.py:45 INFO TRAIN_STEP step=1561 loss=1.960575 Training: 9%|▊ | 1561/18018 [3:23:51<32:19:47, 7.07s/step, loss=1.961, acc=0.603, acc_len=1.40, thru=17.3, I=18.3, T=31.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:54:05,293] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1560: forward=2761.1ms backward=1225.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:05,571] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 16:54:12,095] loop.py:45 INFO TRAIN_STEP step=1562 loss=2.168413 Training: 9%|▊ | 1562/18018 [3:23:58<31:57:15, 6.99s/step, loss=2.168, acc=0.572, acc_len=1.30, thru=17.4, I=17.6, T=32.0, wait=2.8s, pool=52, epoch=1/2][2026-10-18 16:54:20,309] loop.py:45 INFO TRAIN_STEP step=1563 loss=2.136934 Training: 9%|▊ | 1563/18018 [3:24:06<33:37:48, 7.36s/step, loss=2.137, acc=0.570, acc_len=1.30, thru=16.0, I=15.4, T=32.2, wait=4.2s, pool=60, epoch=1/2][2026-10-18 16:54:27,373] loop.py:45 INFO TRAIN_STEP step=1564 loss=2.236331 Training: 9%|▊ | 1564/18018 [3:24:13<33:13:34, 7.27s/step, loss=2.236, acc=0.561, acc_len=1.27, thru=14.8, I=17.2, T=30.0, wait=2.8s, pool=32, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 16:54:27,460] eagle_store.py:517 WARNING batch_get_buffer missing keys (k50047-bcd963b2_hs, k50047-bcd963b2_ids, k50047-bcd963b2_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:34,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 16:54:27,459] eagle_store.py:517 WARNING batch_get_buffer missing keys (k50046-b0b18666_hs, k50046-b0b18666_ids, k50046-b0b18666_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-18 16:54:36,552] loop.py:380 INFO TIMING step=1565: step=4.388s data=2.083s compute=4.334s [fwd=3.067s bwd=1.252s opt=0.003s] dispatch=4.780s [2026-10-18 16:54:36,552] loop.py:45 INFO TRAIN_STEP step=1565 loss=2.175820 Training: 9%|▊ | 1565/18018 [3:24:22<35:50:32, 7.84s/step, loss=2.176, acc=0.567, acc_len=1.29, thru=17.1, I=17.9, T=29.2, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:36,977] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:42,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:54:43,250] loop.py:45 INFO TRAIN_STEP step=1566 loss=1.967078 Training: 9%|▊ | 1566/18018 [3:24:29<34:15:56, 7.50s/step, loss=1.967, acc=0.601, acc_len=1.38, thru=16.6, I=23.8, T=30.6, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:54:43,245] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1565: forward=2922.1ms backward=1211.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:43,513] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:49,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:54:50,182] loop.py:45 INFO TRAIN_STEP step=1567 loss=2.100059 Training: 9%|▊ | 1567/18018 [3:24:36<33:29:26, 7.33s/step, loss=2.100, acc=0.578, acc_len=1.33, thru=16.3, I=19.6, T=30.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:50,341] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:58,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:54:59,099] loop.py:45 INFO TRAIN_STEP step=1568 loss=2.128831 Training: 9%|▊ | 1568/18018 [3:24:45<35:39:49, 7.80s/step, loss=2.129, acc=0.575, acc_len=1.31, thru=13.5, I=15.7, T=32.2, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:54:59,508] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:06,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:55:07,886] loop.py:45 INFO TRAIN_STEP step=1569 loss=2.110546 Training: 9%|▊ | 1569/18018 [3:24:53<37:00:32, 8.10s/step, loss=2.111, acc=0.576, acc_len=1.31, thru=14.2, I=17.8, T=28.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:08,093] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:14,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:55:14,816] loop.py:380 INFO TIMING step=1570: step=3.627s data=1.477s compute=3.602s [fwd=2.553s bwd=1.034s opt=0.003s] dispatch=3.295s [2026-10-18 16:55:14,817] loop.py:45 INFO TRAIN_STEP step=1570 loss=2.161182 Training: 9%|▊ | 1570/18018 [3:25:00<35:24:22, 7.75s/step, loss=2.161, acc=0.567, acc_len=1.29, thru=17.2, I=22.8, T=35.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:15,031] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:21,903] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:55:23,183] loop.py:45 INFO TRAIN_STEP step=1571 loss=2.229387 Training: 9%|▊ | 1571/18018 [3:25:09<36:14:59, 7.93s/step, loss=2.229, acc=0.561, acc_len=1.27, thru=15.5, I=19.5, T=30.3, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:55:23,178] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1570: forward=2970.3ms backward=1199.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:23,409] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:27,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:55:29,621] loop.py:45 INFO TRAIN_STEP step=1572 loss=2.137932 Training: 9%|▊ | 1572/18018 [3:25:15<34:11:41, 7.49s/step, loss=2.138, acc=0.573, acc_len=1.31, thru=18.1, I=24.8, T=33.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:29,932] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:34,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:55:36,332] loop.py:45 INFO TRAIN_STEP step=1573 loss=2.056500 Training: 9%|▊ | 1573/18018 [3:25:22<33:08:05, 7.25s/step, loss=2.056, acc=0.585, acc_len=1.34, thru=17.0, I=24.5, T=35.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:36,497] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 16:55:43,377] loop.py:45 INFO TRAIN_STEP step=1574 loss=2.105170 Training: 9%|▊ | 1574/18018 [3:25:29<32:50:38, 7.19s/step, loss=2.105, acc=0.578, acc_len=1.32, thru=17.7, I=19.8, T=35.4, wait=3.4s, pool=60, epoch=1/2][2026-10-18 16:55:49,720] loop.py:380 INFO TIMING step=1575: step=3.700s data=1.521s compute=3.659s [fwd=2.562s bwd=1.081s opt=0.003s] dispatch=2.635s [2026-10-18 16:55:49,720] loop.py:45 INFO TRAIN_STEP step=1575 loss=2.223446 Training: 9%|▊ | 1575/18018 [3:25:35<31:41:02, 6.94s/step, loss=2.223, acc=0.568, acc_len=1.29, thru=18.4, I=19.2, T=34.6, wait=2.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:56,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:55:57,831] loop.py:45 INFO TRAIN_STEP step=1576 loss=2.057728 Training: 9%|▊ | 1576/18018 [3:25:43<33:17:23, 7.29s/step, loss=2.058, acc=0.586, acc_len=1.34, thru=15.6, I=17.2, T=33.1, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:55:57,826] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1575: forward=2668.1ms backward=1135.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:55:58,284] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:04,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:56:05,599] loop.py:45 INFO TRAIN_STEP step=1577 loss=2.148234 Training: 9%|▉ | 1577/18018 [3:25:51<33:56:42, 7.43s/step, loss=2.148, acc=0.570, acc_len=1.30, thru=15.1, I=19.8, T=31.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:05,763] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:10,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:56:11,326] loop.py:45 INFO TRAIN_STEP step=1578 loss=1.992078 Training: 9%|▉ | 1578/18018 [3:25:57<31:36:07, 6.92s/step, loss=1.992, acc=0.601, acc_len=1.39, thru=20.2, I=23.1, T=36.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:11,788] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 16:56:17,674] loop.py:45 INFO TRAIN_STEP step=1579 loss=1.962386 Training: 9%|▉ | 1579/18018 [3:26:03<30:49:14, 6.75s/step, loss=1.962, acc=0.598, acc_len=1.38, thru=18.9, I=23.5, T=30.7, wait=2.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:25,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:56:26,014] loop.py:380 INFO TIMING step=1580: step=3.973s data=1.741s compute=3.943s [fwd=2.784s bwd=1.143s opt=0.003s] dispatch=4.356s [2026-10-18 16:56:26,014] loop.py:45 INFO TRAIN_STEP step=1580 loss=2.036498 Training: 9%|▉ | 1580/18018 [3:26:12<32:59:40, 7.23s/step, loss=2.036, acc=0.587, acc_len=1.35, thru=16.4, I=16.3, T=32.2, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:26,172] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:30,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:56:32,386] loop.py:45 INFO TRAIN_STEP step=1581 loss=2.061671 Training: 9%|▉ | 1581/18018 [3:26:18<31:49:31, 6.97s/step, loss=2.062, acc=0.588, acc_len=1.36, thru=18.9, I=22.0, T=33.9, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:56:32,382] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1580: forward=2591.0ms backward=1125.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:32,448] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 16:56:39,063] loop.py:45 INFO TRAIN_STEP step=1582 loss=2.125064 Training: 9%|▉ | 1582/18018 [3:26:25<31:25:25, 6.88s/step, loss=2.125, acc=0.575, acc_len=1.31, thru=17.7, I=19.9, T=36.1, wait=3.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:45,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:56:47,515] loop.py:45 INFO TRAIN_STEP step=1583 loss=2.170438 Training: 9%|▉ | 1583/18018 [3:26:33<33:33:56, 7.35s/step, loss=2.170, acc=0.570, acc_len=1.30, thru=16.4, I=19.4, T=34.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:56:47,706] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:56:53,503] loop.py:45 INFO TRAIN_STEP step=1584 loss=1.988356 Training: 9%|▉ | 1584/18018 [3:26:39<31:42:02, 6.94s/step, loss=1.988, acc=0.594, acc_len=1.37, thru=19.1, I=26.2, T=36.1, wait=2.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:02,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:57:02,906] loop.py:380 INFO TIMING step=1585: step=3.868s data=1.551s compute=3.835s [fwd=2.734s bwd=1.086s opt=0.003s] dispatch=5.523s [2026-10-18 16:57:02,906] loop.py:45 INFO TRAIN_STEP step=1585 loss=2.279202 Training: 9%|▉ | 1585/18018 [3:26:48<35:03:46, 7.68s/step, loss=2.279, acc=0.555, acc_len=1.25, thru=16.2, I=14.4, T=33.1, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:02,968] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:10,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:57:11,022] loop.py:45 INFO TRAIN_STEP step=1586 loss=1.987213 Training: 9%|▉ | 1586/18018 [3:26:57<35:39:21, 7.81s/step, loss=1.987, acc=0.596, acc_len=1.37, thru=15.4, I=15.5, T=29.0, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:57:11,018] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1585: forward=2952.9ms backward=1385.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:11,447] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:17,483] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:57:17,885] loop.py:45 INFO TRAIN_STEP step=1587 loss=2.118874 Training: 9%|▉ | 1587/18018 [3:27:03<34:21:09, 7.53s/step, loss=2.119, acc=0.572, acc_len=1.30, thru=19.0, I=19.4, T=30.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:17,991] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:22,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:57:24,351] loop.py:45 INFO TRAIN_STEP step=1588 loss=2.023316 Training: 9%|▉ | 1588/18018 [3:27:10<32:54:03, 7.21s/step, loss=2.023, acc=0.595, acc_len=1.37, thru=18.3, I=21.6, T=31.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:24,386] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:30,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:57:32,158] loop.py:45 INFO TRAIN_STEP step=1589 loss=2.062820 Training: 9%|▉ | 1589/18018 [3:27:18<33:43:06, 7.39s/step, loss=2.063, acc=0.586, acc_len=1.35, thru=15.5, I=19.8, T=28.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:57:32,492] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 16:57:39,570] loop.py:380 INFO TIMING step=1590: step=4.059s data=1.885s compute=4.021s [fwd=2.841s bwd=1.165s opt=0.003s] dispatch=3.343s [2026-10-18 16:57:39,570] loop.py:45 INFO TRAIN_STEP step=1590 loss=2.254959 Training: 9%|▉ | 1590/18018 [3:27:25<33:44:43, 7.39s/step, loss=2.255, acc=0.560, acc_len=1.26, thru=13.6, I=16.5, T=31.5, wait=3.3s, pool=44, epoch=1/2][2026-10-18 16:57:49,998] loop.py:45 INFO TRAIN_STEP step=1591 loss=2.058507 Training: 9%|▉ | 1591/18018 [3:27:36<37:53:51, 8.31s/step, loss=2.059, acc=0.587, acc_len=1.35, thru=10.6, I=12.4, T=29.6, wait=6.1s, pool=28, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:57:49,994] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1590: forward=2883.0ms backward=1344.6ms [2026-10-18 16:58:00,459] loop.py:45 INFO TRAIN_STEP step=1592 loss=2.123409 Training: 9%|▉ | 1592/18018 [3:27:46<40:50:37, 8.95s/step, loss=2.123, acc=0.583, acc_len=1.33, thru=14.5, I=14.1, T=30.2, wait=6.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:08,866] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:58:10,994] loop.py:45 INFO TRAIN_STEP step=1593 loss=2.088165 Training: 9%|▉ | 1593/18018 [3:27:57<43:00:39, 9.43s/step, loss=2.088, acc=0.583, acc_len=1.33, thru=13.2, I=17.1, T=27.3, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:11,384] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:17,098] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:58:18,253] loop.py:45 INFO TRAIN_STEP step=1594 loss=1.857488 Training: 9%|▉ | 1594/18018 [3:28:04<40:02:17, 8.78s/step, loss=1.857, acc=0.617, acc_len=1.44, thru=14.7, I=23.5, T=34.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:18,605] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:23,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:58:24,576] loop.py:380 INFO TIMING step=1595: step=3.559s data=1.506s compute=3.533s [fwd=2.504s bwd=1.014s opt=0.003s] dispatch=2.755s [2026-10-18 16:58:24,576] loop.py:45 INFO TRAIN_STEP step=1595 loss=1.959866 Training: 9%|▉ | 1595/18018 [3:28:10<36:40:37, 8.04s/step, loss=1.960, acc=0.599, acc_len=1.38, thru=20.3, I=25.8, T=36.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:25,060] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:31,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:58:32,342] loop.py:45 INFO TRAIN_STEP step=1596 loss=1.942430 Training: 9%|▉ | 1596/18018 [3:28:18<36:18:13, 7.96s/step, loss=1.942, acc=0.604, acc_len=1.40, thru=15.6, I=19.1, T=31.5, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:58:32,337] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1595: forward=2805.0ms backward=1195.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:32,377] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:38,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:58:40,711] loop.py:45 INFO TRAIN_STEP step=1597 loss=2.267401 Training: 9%|▉ | 1597/18018 [3:28:26<36:51:50, 8.08s/step, loss=2.267, acc=0.559, acc_len=1.26, thru=16.0, I=19.7, T=32.0, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:40,839] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-18 16:58:46,837] loop.py:45 INFO TRAIN_STEP step=1598 loss=2.263386 Training: 9%|▉ | 1598/18018 [3:28:32<34:11:11, 7.50s/step, loss=2.263, acc=0.551, acc_len=1.24, thru=15.7, I=24.3, T=33.3, wait=2.3s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:54,107] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:58:55,965] loop.py:45 INFO TRAIN_STEP step=1599 loss=2.114088 Training: 9%|▉ | 1599/18018 [3:28:42<36:25:03, 7.98s/step, loss=2.114, acc=0.573, acc_len=1.31, thru=17.2, I=16.6, T=29.4, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:58:56,137] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:59:01,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:59:03,416] loop.py:380 INFO TIMING step=1600: step=4.413s data=1.928s compute=4.377s [fwd=3.057s bwd=1.304s opt=0.003s] dispatch=3.027s [2026-10-18 16:59:03,416] loop.py:45 INFO TRAIN_STEP step=1600 loss=2.133611 Training: 9%|▉ | 1600/18018 [3:28:49<35:40:54, 7.82s/step, loss=2.134, acc=0.574, acc_len=1.30, thru=14.9, I=22.3, T=29.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:59:03,895] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 16:59:10,408] loop.py:45 INFO TRAIN_STEP step=1601 loss=2.142199 Training: 9%|▉ | 1601/18018 [3:28:56<34:32:42, 7.58s/step, loss=2.142, acc=0.576, acc_len=1.31, thru=14.2, I=18.8, T=33.9, wait=3.2s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:59:10,402] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1600: forward=2597.1ms backward=1131.9ms [2026-10-18 16:59:19,277] loop.py:45 INFO TRAIN_STEP step=1602 loss=2.079340 Training: 9%|▉ | 1602/18018 [3:29:05<36:18:33, 7.96s/step, loss=2.079, acc=0.579, acc_len=1.33, thru=15.3, I=14.4, T=32.7, wait=4.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:59:27,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:59:29,744] loop.py:45 INFO TRAIN_STEP step=1603 loss=2.054737 Training: 9%|▉ | 1603/18018 [3:29:15<39:44:10, 8.71s/step, loss=2.055, acc=0.590, acc_len=1.35, thru=13.2, I=15.5, T=30.3, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:59:29,924] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:59:35,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:59:36,665] loop.py:45 INFO TRAIN_STEP step=1604 loss=2.117248 Training: 9%|▉ | 1604/18018 [3:29:22<37:16:39, 8.18s/step, loss=2.117, acc=0.574, acc_len=1.31, thru=16.0, I=24.9, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:59:37,060] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:59:42,670] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 16:59:43,580] loop.py:380 INFO TIMING step=1605: step=4.022s data=1.874s compute=3.985s [fwd=2.815s bwd=1.154s opt=0.003s] dispatch=2.883s [2026-10-18 16:59:43,580] loop.py:45 INFO TRAIN_STEP step=1605 loss=1.905149 Training: 9%|▉ | 1605/18018 [3:29:29<35:33:12, 7.80s/step, loss=1.905, acc=0.604, acc_len=1.40, thru=18.1, I=25.3, T=31.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 16:59:43,683] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 16:59:54,005] loop.py:45 INFO TRAIN_STEP step=1606 loss=2.181667 Training: 9%|▉ | 1606/18018 [3:29:40<39:08:25, 8.59s/step, loss=2.182, acc=0.572, acc_len=1.30, thru=10.7, I=10.0, T=29.9, wait=6.1s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 16:59:54,001] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1605: forward=3086.7ms backward=1152.5ms [2026-10-18 17:00:02,591] loop.py:45 INFO TRAIN_STEP step=1607 loss=2.133429 Training: 9%|▉ | 1607/18018 [3:29:48<39:08:26, 8.59s/step, loss=2.133, acc=0.580, acc_len=1.32, thru=15.4, I=14.8, T=31.0, wait=4.4s, pool=44, epoch=1/2][2026-10-18 17:00:09,181] loop.py:45 INFO TRAIN_STEP step=1608 loss=1.795448 Training: 9%|▉ | 1608/18018 [3:29:55<36:24:26, 7.99s/step, loss=1.795, acc=0.626, acc_len=1.47, thru=17.4, I=15.8, T=34.7, wait=2.9s, pool=48, epoch=1/2][2026-10-18 17:00:21,158] loop.py:45 INFO TRAIN_STEP step=1609 loss=2.464559 Training: 9%|▉ | 1609/18018 [3:30:07<41:51:58, 9.19s/step, loss=2.465, acc=0.534, acc_len=1.19, thru=12.0, I=12.8, T=28.9, wait=7.5s, pool=52, epoch=1/2][2026-10-18 17:00:30,413] loop.py:380 INFO TIMING step=1610: step=4.091s data=1.787s compute=4.058s [fwd=2.888s bwd=1.155s opt=0.003s] dispatch=5.152s [2026-10-18 17:00:30,413] loop.py:45 INFO TRAIN_STEP step=1610 loss=2.038893 Training: 9%|▉ | 1610/18018 [3:30:16<41:57:15, 9.21s/step, loss=2.039, acc=0.588, acc_len=1.34, thru=12.1, I=10.2, T=31.3, wait=5.2s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:00:38,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:00:40,272] loop.py:45 INFO TRAIN_STEP step=1611 loss=2.125027 Training: 9%|▉ | 1611/18018 [3:30:26<42:50:43, 9.40s/step, loss=2.125, acc=0.578, acc_len=1.32, thru=16.7, I=17.0, T=31.2, wait=5.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:00:40,268] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1610: forward=2848.0ms backward=1207.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:00:40,696] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:00:45,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:00:47,960] loop.py:45 INFO TRAIN_STEP step=1612 loss=1.885210 Training: 9%|▉ | 1612/18018 [3:30:34<40:30:00, 8.89s/step, loss=1.885, acc=0.616, acc_len=1.43, thru=15.4, I=23.3, T=25.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:00:48,360] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:00:52,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:00:54,619] loop.py:45 INFO TRAIN_STEP step=1613 loss=1.962343 Training: 9%|▉ | 1613/18018 [3:30:40<37:27:17, 8.22s/step, loss=1.962, acc=0.597, acc_len=1.38, thru=15.3, I=29.7, T=32.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:00:54,787] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:01:01,736] loop.py:45 INFO TRAIN_STEP step=1614 loss=1.987123 Training: 9%|▉ | 1614/18018 [3:30:47<35:56:38, 7.89s/step, loss=1.987, acc=0.595, acc_len=1.37, thru=13.2, I=16.1, T=33.6, wait=3.3s, pool=44, epoch=1/2][2026-10-18 17:01:12,051] loop.py:380 INFO TIMING step=1615: step=4.588s data=1.988s compute=4.552s [fwd=3.235s bwd=1.302s opt=0.003s] dispatch=5.717s [2026-10-18 17:01:12,051] loop.py:45 INFO TRAIN_STEP step=1615 loss=2.247574 Training: 9%|▉ | 1615/18018 [3:30:58<39:15:42, 8.62s/step, loss=2.248, acc=0.568, acc_len=1.28, thru=15.1, I=15.9, T=27.9, wait=5.7s, pool=60, epoch=1/2][2026-10-18 17:01:19,617] loop.py:45 INFO TRAIN_STEP step=1616 loss=1.976775 Training: 9%|▉ | 1616/18018 [3:31:05<37:49:11, 8.30s/step, loss=1.977, acc=0.600, acc_len=1.39, thru=14.1, I=14.7, T=33.2, wait=3.7s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:01:19,611] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1615: forward=2672.6ms backward=1123.6ms [2026-10-18 17:01:27,476] loop.py:45 INFO TRAIN_STEP step=1617 loss=2.169923 Training: 9%|▉ | 1617/18018 [3:31:13<37:12:53, 8.17s/step, loss=2.170, acc=0.570, acc_len=1.30, thru=16.2, I=15.8, T=31.7, wait=3.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:01:33,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:01:34,757] loop.py:45 INFO TRAIN_STEP step=1618 loss=2.131734 Training: 9%|▉ | 1618/18018 [3:31:20<35:59:56, 7.90s/step, loss=2.132, acc=0.576, acc_len=1.31, thru=18.4, I=18.7, T=33.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:01:34,862] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:01:43,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:01:44,392] loop.py:45 INFO TRAIN_STEP step=1619 loss=2.080003 Training: 9%|▉ | 1619/18018 [3:31:30<38:22:07, 8.42s/step, loss=2.080, acc=0.586, acc_len=1.35, thru=14.1, I=13.6, T=30.8, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:01:44,739] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:01:49,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:01:51,728] loop.py:380 INFO TIMING step=1620: step=4.232s data=1.939s compute=4.184s [fwd=2.920s bwd=1.249s opt=0.003s] dispatch=3.092s [2026-10-18 17:01:51,728] loop.py:45 INFO TRAIN_STEP step=1620 loss=2.035482 Training: 9%|▉ | 1620/18018 [3:31:37<36:52:45, 8.10s/step, loss=2.035, acc=0.590, acc_len=1.35, thru=16.8, I=23.8, T=30.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:01:51,890] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:01:58,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:01:59,062] loop.py:45 INFO TRAIN_STEP step=1621 loss=2.083360 Training: 9%|▉ | 1621/18018 [3:31:45<35:50:00, 7.87s/step, loss=2.083, acc=0.579, acc_len=1.32, thru=14.6, I=16.7, T=30.5, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:01:59,058] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1620: forward=2959.9ms backward=1187.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:01:59,183] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-18 17:02:05,747] loop.py:45 INFO TRAIN_STEP step=1622 loss=1.947042 Training: 9%|▉ | 1622/18018 [3:31:51<34:12:52, 7.51s/step, loss=1.947, acc=0.603, acc_len=1.39, thru=16.8, I=18.1, T=33.0, wait=2.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:16,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:02:17,537] loop.py:45 INFO TRAIN_STEP step=1623 loss=2.047777 Training: 9%|▉ | 1623/18018 [3:32:03<40:03:35, 8.80s/step, loss=2.048, acc=0.589, acc_len=1.35, thru=12.1, I=13.1, T=31.6, wait=7.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:17,629] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:24,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:02:24,660] loop.py:45 INFO TRAIN_STEP step=1624 loss=1.976378 Training: 9%|▉ | 1624/18018 [3:32:10<37:45:59, 8.29s/step, loss=1.976, acc=0.601, acc_len=1.38, thru=16.9, I=20.3, T=33.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:24,779] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:33,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:02:34,118] loop.py:380 INFO TIMING step=1625: step=3.991s data=1.679s compute=3.960s [fwd=2.816s bwd=1.129s opt=0.003s] dispatch=5.461s [2026-10-18 17:02:34,118] loop.py:45 INFO TRAIN_STEP step=1625 loss=1.979717 Training: 9%|▉ | 1625/18018 [3:32:20<39:21:23, 8.64s/step, loss=1.980, acc=0.597, acc_len=1.39, thru=13.8, I=13.9, T=32.1, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:34,447] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:40,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:02:40,810] loop.py:45 INFO TRAIN_STEP step=1626 loss=2.364864 Training: 9%|▉ | 1626/18018 [3:32:26<36:41:34, 8.06s/step, loss=2.365, acc=0.542, acc_len=1.22, thru=18.7, I=18.1, T=32.1, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:02:40,806] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1625: forward=2813.8ms backward=1136.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:40,992] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:47,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:02:48,017] loop.py:45 INFO TRAIN_STEP step=1627 loss=1.910614 Training: 9%|▉ | 1627/18018 [3:32:34<35:31:34, 7.80s/step, loss=1.911, acc=0.606, acc_len=1.41, thru=16.8, I=18.6, T=30.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:48,361] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:54,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:02:56,013] loop.py:45 INFO TRAIN_STEP step=1628 loss=2.204837 Training: 9%|▉ | 1628/18018 [3:32:42<35:47:18, 7.86s/step, loss=2.205, acc=0.562, acc_len=1.27, thru=16.1, I=20.7, T=31.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:02:56,102] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-18 17:03:02,203] loop.py:45 INFO TRAIN_STEP step=1629 loss=2.038281 Training: 9%|▉ | 1629/18018 [3:32:48<33:30:08, 7.36s/step, loss=2.038, acc=0.589, acc_len=1.36, thru=17.8, I=23.5, T=32.7, wait=2.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:03:08,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:03:09,382] loop.py:380 INFO TIMING step=1630: step=4.396s data=2.000s compute=4.365s [fwd=2.996s bwd=1.357s opt=0.003s] dispatch=2.774s [2026-10-18 17:03:09,382] loop.py:45 INFO TRAIN_STEP step=1630 loss=2.130471 Training: 9%|▉ | 1630/18018 [3:32:55<33:15:22, 7.31s/step, loss=2.130, acc=0.573, acc_len=1.30, thru=18.7, I=19.3, T=29.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:03:09,643] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:03:16,314] loop.py:45 INFO TRAIN_STEP step=1631 loss=2.388385 Training: 9%|▉ | 1631/18018 [3:33:02<32:44:28, 7.19s/step, loss=2.388, acc=0.534, acc_len=1.19, thru=15.1, I=18.1, T=31.3, wait=2.8s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:03:16,310] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1630: forward=2883.7ms backward=1157.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:03:23,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:03:25,619] loop.py:45 INFO TRAIN_STEP step=1632 loss=2.098011 Training: 9%|▉ | 1632/18018 [3:33:11<35:37:38, 7.83s/step, loss=2.098, acc=0.576, acc_len=1.32, thru=16.1, I=17.1, T=28.2, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:03:25,861] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:03:32,207] loop.py:45 INFO TRAIN_STEP step=1633 loss=2.076756 Training: 9%|▉ | 1633/18018 [3:33:18<33:55:50, 7.46s/step, loss=2.077, acc=0.581, acc_len=1.33, thru=16.6, I=21.6, T=36.3, wait=3.1s, pool=56, epoch=1/2][2026-10-18 17:03:40,749] loop.py:45 INFO TRAIN_STEP step=1634 loss=2.111668 Training: 9%|▉ | 1634/18018 [3:33:26<35:24:38, 7.78s/step, loss=2.112, acc=0.572, acc_len=1.30, thru=15.4, I=14.3, T=29.4, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:03:47,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:03:47,695] loop.py:380 INFO TIMING step=1635: step=3.935s data=1.729s compute=3.898s [fwd=2.814s bwd=1.070s opt=0.003s] dispatch=3.003s [2026-10-18 17:03:47,695] loop.py:45 INFO TRAIN_STEP step=1635 loss=2.156229 Training: 9%|▉ | 1635/18018 [3:33:33<34:16:23, 7.53s/step, loss=2.156, acc=0.570, acc_len=1.30, thru=18.2, I=18.9, T=32.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:03:48,108] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:03:52,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:03:54,940] loop.py:45 INFO TRAIN_STEP step=1636 loss=2.012075 Training: 9%|▉ | 1636/18018 [3:33:41<33:52:50, 7.45s/step, loss=2.012, acc=0.590, acc_len=1.36, thru=16.7, I=21.8, T=27.1, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:03:54,935] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1635: forward=3451.8ms backward=1211.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:03:55,112] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:01,757] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:04:02,639] loop.py:45 INFO TRAIN_STEP step=1637 loss=2.144845 Training: 9%|▉ | 1637/18018 [3:33:48<34:13:21, 7.52s/step, loss=2.145, acc=0.577, acc_len=1.32, thru=17.0, I=18.9, T=31.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:02,770] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:09,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:04:09,267] loop.py:45 INFO TRAIN_STEP step=1638 loss=2.196676 Training: 9%|▉ | 1638/18018 [3:33:55<33:00:04, 7.25s/step, loss=2.197, acc=0.565, acc_len=1.28, thru=19.1, I=21.5, T=31.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:09,527] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 17:04:16,334] loop.py:45 INFO TRAIN_STEP step=1639 loss=2.030097 Training: 9%|▉ | 1639/18018 [3:34:02<32:44:36, 7.20s/step, loss=2.030, acc=0.592, acc_len=1.36, thru=16.4, I=17.5, T=28.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:16,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:16,807] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:26,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:04:26,306] loop.py:380 INFO TIMING step=1640: step=4.711s data=2.295s compute=4.650s [fwd=3.300s bwd=1.335s opt=0.003s] dispatch=5.254s [2026-10-18 17:04:26,306] loop.py:45 INFO TRAIN_STEP step=1640 loss=2.192835 Training: 9%|▉ | 1640/18018 [3:34:12<36:31:43, 8.03s/step, loss=2.193, acc=0.567, acc_len=1.29, thru=13.6, I=12.5, T=27.2, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:26,736] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:32,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:04:33,070] loop.py:45 INFO TRAIN_STEP step=1641 loss=2.146907 Training: 9%|▉ | 1641/18018 [3:34:19<34:48:03, 7.65s/step, loss=2.147, acc=0.575, acc_len=1.31, thru=18.0, I=18.5, T=31.3, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:04:33,066] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1640: forward=2892.8ms backward=1132.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:33,454] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:40,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:04:40,491] loop.py:45 INFO TRAIN_STEP step=1642 loss=2.104390 Training: 9%|▉ | 1642/18018 [3:34:26<34:29:24, 7.58s/step, loss=2.104, acc=0.577, acc_len=1.32, thru=15.7, I=16.9, T=33.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:40,567] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-18 17:04:47,007] loop.py:45 INFO TRAIN_STEP step=1643 loss=2.058738 Training: 9%|▉ | 1643/18018 [3:34:33<33:01:46, 7.26s/step, loss=2.059, acc=0.586, acc_len=1.34, thru=15.7, I=19.3, T=34.1, wait=2.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:52,822] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:04:53,915] loop.py:45 INFO TRAIN_STEP step=1644 loss=2.038226 Training: 9%|▉ | 1644/18018 [3:34:39<32:32:41, 7.16s/step, loss=2.038, acc=0.582, acc_len=1.33, thru=18.4, I=20.4, T=34.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:04:54,329] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:05:01,664] loop.py:380 INFO TIMING step=1645: step=4.053s data=1.763s compute=4.011s [fwd=2.834s bwd=1.162s opt=0.003s] dispatch=3.688s [2026-10-18 17:05:01,664] loop.py:45 INFO TRAIN_STEP step=1645 loss=1.921715 Training: 9%|▉ | 1645/18018 [3:34:47<33:21:12, 7.33s/step, loss=1.922, acc=0.601, acc_len=1.39, thru=15.6, I=18.2, T=31.6, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:09,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:05:11,413] loop.py:45 INFO TRAIN_STEP step=1646 loss=2.158628 Training: 9%|▉ | 1646/18018 [3:34:57<36:38:59, 8.06s/step, loss=2.159, acc=0.564, acc_len=1.28, thru=14.1, I=16.9, T=28.9, wait=5.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:05:11,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1645: forward=3117.4ms backward=1268.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:11,708] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:05:18,253] loop.py:45 INFO TRAIN_STEP step=1647 loss=2.073038 Training: 9%|▉ | 1647/18018 [3:35:04<34:58:52, 7.69s/step, loss=2.073, acc=0.584, acc_len=1.34, thru=15.7, I=19.2, T=31.9, wait=2.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:25,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:05:27,092] loop.py:45 INFO TRAIN_STEP step=1648 loss=2.046628 Training: 9%|▉ | 1648/18018 [3:35:13<36:32:40, 8.04s/step, loss=2.047, acc=0.590, acc_len=1.36, thru=16.6, I=16.9, T=28.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:27,500] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:05:35,951] loop.py:45 INFO TRAIN_STEP step=1649 loss=2.226408 Training: 9%|▉ | 1649/18018 [3:35:22<37:39:44, 8.28s/step, loss=2.226, acc=0.559, acc_len=1.26, thru=13.2, I=15.5, T=30.0, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:35,904] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:36,405] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:43,181] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:05:44,414] loop.py:380 INFO TIMING step=1650: step=3.958s data=1.666s compute=3.915s [fwd=2.773s bwd=1.127s opt=0.003s] dispatch=4.498s [2026-10-18 17:05:44,414] loop.py:45 INFO TRAIN_STEP step=1650 loss=1.953604 Training: 9%|▉ | 1650/18018 [3:35:30<37:54:25, 8.34s/step, loss=1.954, acc=0.599, acc_len=1.38, thru=16.2, I=16.9, T=32.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:44,687] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:50,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:05:51,949] loop.py:45 INFO TRAIN_STEP step=1651 loss=1.982329 Training: 9%|▉ | 1651/18018 [3:35:38<36:48:43, 8.10s/step, loss=1.982, acc=0.597, acc_len=1.38, thru=15.7, I=20.7, T=27.5, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:05:51,944] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1650: forward=3185.9ms backward=1417.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:05:52,187] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:05:58,702] loop.py:45 INFO TRAIN_STEP step=1652 loss=2.027947 Training: 9%|▉ | 1652/18018 [3:35:44<34:58:36, 7.69s/step, loss=2.028, acc=0.592, acc_len=1.36, thru=15.0, I=18.3, T=32.5, wait=2.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:07,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:06:09,054] loop.py:45 INFO TRAIN_STEP step=1653 loss=2.148309 Training: 9%|▉ | 1653/18018 [3:35:55<38:35:58, 8.49s/step, loss=2.148, acc=0.572, acc_len=1.29, thru=14.8, I=15.0, T=27.5, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:09,487] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:06:16,540] loop.py:45 INFO TRAIN_STEP step=1654 loss=2.211237 Training: 9%|▉ | 1654/18018 [3:36:02<37:13:28, 8.19s/step, loss=2.211, acc=0.560, acc_len=1.26, thru=13.3, I=15.3, T=32.5, wait=3.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:22,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:06:23,400] loop.py:380 INFO TIMING step=1655: step=3.903s data=1.719s compute=3.875s [fwd=2.780s bwd=1.080s opt=0.003s] dispatch=2.949s [2026-10-18 17:06:23,400] loop.py:45 INFO TRAIN_STEP step=1655 loss=1.975039 Training: 9%|▉ | 1655/18018 [3:36:09<35:24:39, 7.79s/step, loss=1.975, acc=0.601, acc_len=1.39, thru=19.0, I=19.9, T=32.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:23,530] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:28,044] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:06:29,957] loop.py:45 INFO TRAIN_STEP step=1656 loss=1.883521 Training: 9%|▉ | 1656/18018 [3:36:16<33:43:45, 7.42s/step, loss=1.884, acc=0.610, acc_len=1.42, thru=18.8, I=22.8, T=30.0, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:06:29,952] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1655: forward=2949.8ms backward=1257.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:30,072] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:36,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:06:37,965] loop.py:45 INFO TRAIN_STEP step=1657 loss=2.165262 Training: 9%|▉ | 1657/18018 [3:36:24<34:31:33, 7.60s/step, loss=2.165, acc=0.572, acc_len=1.30, thru=13.3, I=18.5, T=30.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:38,209] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:44,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:06:46,187] loop.py:45 INFO TRAIN_STEP step=1658 loss=2.034903 Training: 9%|▉ | 1658/18018 [3:36:32<35:22:37, 7.78s/step, loss=2.035, acc=0.593, acc_len=1.36, thru=14.4, I=19.8, T=26.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:46,619] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:51,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:06:52,940] loop.py:45 INFO TRAIN_STEP step=1659 loss=2.083962 Training: 9%|▉ | 1659/18018 [3:36:39<33:58:05, 7.48s/step, loss=2.084, acc=0.581, acc_len=1.32, thru=16.2, I=23.7, T=28.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:53,111] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:06:58,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:06:59,857] loop.py:380 INFO TIMING step=1660: step=4.353s data=2.292s compute=4.316s [fwd=3.256s bwd=1.046s opt=0.003s] dispatch=2.552s [2026-10-18 17:06:59,857] loop.py:45 INFO TRAIN_STEP step=1660 loss=2.099050 Training: 9%|▉ | 1660/18018 [3:36:45<33:12:17, 7.31s/step, loss=2.099, acc=0.582, acc_len=1.33, thru=17.3, I=21.5, T=29.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:00,203] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:07:05,603] loop.py:45 INFO TRAIN_STEP step=1661 loss=1.995519 Training: 9%|▉ | 1661/18018 [3:36:51<31:04:24, 6.84s/step, loss=1.996, acc=0.596, acc_len=1.38, thru=18.4, I=23.7, T=35.5, wait=2.1s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:07:05,599] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1660: forward=2556.1ms backward=1002.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:12,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:07:13,476] loop.py:45 INFO TRAIN_STEP step=1662 loss=2.105418 Training: 9%|▉ | 1662/18018 [3:36:59<32:28:56, 7.15s/step, loss=2.105, acc=0.577, acc_len=1.31, thru=16.7, I=18.4, T=32.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:13,845] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:18,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:07:19,979] loop.py:45 INFO TRAIN_STEP step=1663 loss=2.117248 Training: 9%|▉ | 1663/18018 [3:37:06<31:35:46, 6.95s/step, loss=2.117, acc=0.576, acc_len=1.31, thru=18.8, I=23.0, T=33.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:20,408] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:07:26,438] loop.py:45 INFO TRAIN_STEP step=1664 loss=2.125657 Training: 9%|▉ | 1664/18018 [3:37:12<30:55:16, 6.81s/step, loss=2.126, acc=0.577, acc_len=1.31, thru=18.5, I=21.9, T=34.4, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:32,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:07:33,933] loop.py:380 INFO TIMING step=1665: step=4.508s data=2.069s compute=4.471s [fwd=3.162s bwd=1.293s opt=0.003s] dispatch=2.977s [2026-10-18 17:07:33,933] loop.py:45 INFO TRAIN_STEP step=1665 loss=2.020863 Training: 9%|▉ | 1665/18018 [3:37:20<31:51:28, 7.01s/step, loss=2.021, acc=0.592, acc_len=1.37, thru=17.7, I=21.2, T=28.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:34,081] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:39,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:07:40,586] loop.py:45 INFO TRAIN_STEP step=1666 loss=2.100384 Training: 9%|▉ | 1666/18018 [3:37:26<31:21:49, 6.90s/step, loss=2.100, acc=0.580, acc_len=1.33, thru=16.2, I=22.5, T=33.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:07:40,580] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1665: forward=2702.9ms backward=1130.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:40,916] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:46,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:07:47,980] loop.py:45 INFO TRAIN_STEP step=1667 loss=2.027381 Training: 9%|▉ | 1667/18018 [3:37:34<32:01:41, 7.05s/step, loss=2.027, acc=0.584, acc_len=1.34, thru=17.4, I=21.1, T=34.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:48,345] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:53,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:07:54,655] loop.py:45 INFO TRAIN_STEP step=1668 loss=2.134474 Training: 9%|▉ | 1668/18018 [3:37:40<31:30:50, 6.94s/step, loss=2.134, acc=0.570, acc_len=1.30, thru=18.4, I=26.3, T=36.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:07:55,026] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:08:02,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:08:03,262] loop.py:45 INFO TRAIN_STEP step=1669 loss=2.023716 Training: 9%|▉ | 1669/18018 [3:37:49<33:46:56, 7.44s/step, loss=2.024, acc=0.586, acc_len=1.35, thru=15.6, I=17.6, T=32.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:08:03,661] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:08:10,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:08:11,553] loop.py:380 INFO TIMING step=1670: step=4.407s data=1.955s compute=4.371s [fwd=3.002s bwd=1.353s opt=0.003s] dispatch=3.875s [2026-10-18 17:08:11,554] loop.py:45 INFO TRAIN_STEP step=1670 loss=2.052882 Training: 9%|▉ | 1670/18018 [3:37:57<34:56:40, 7.70s/step, loss=2.053, acc=0.588, acc_len=1.35, thru=14.8, I=19.6, T=29.0, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:08:11,683] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 17:08:20,015] loop.py:45 INFO TRAIN_STEP step=1671 loss=2.169778 Training: 9%|▉ | 1671/18018 [3:38:06<35:59:05, 7.92s/step, loss=2.170, acc=0.566, acc_len=1.29, thru=11.2, I=15.4, T=31.5, wait=4.4s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:08:20,010] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1670: forward=2859.2ms backward=1150.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:08:32,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:08:34,304] loop.py:45 INFO TRAIN_STEP step=1672 loss=2.058517 Training: 9%|▉ | 1672/18018 [3:38:20<44:39:08, 9.83s/step, loss=2.059, acc=0.584, acc_len=1.34, thru=11.5, I=13.4, T=23.0, wait=8.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:08:34,556] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 17:08:43,749] loop.py:45 INFO TRAIN_STEP step=1673 loss=2.078223 Training: 9%|▉ | 1673/18018 [3:38:29<44:06:56, 9.72s/step, loss=2.078, acc=0.587, acc_len=1.33, thru=13.5, I=12.8, T=27.4, wait=4.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:08:52,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:08:54,698] loop.py:45 INFO TRAIN_STEP step=1674 loss=2.123433 Training: 9%|▉ | 1674/18018 [3:38:40<45:47:35, 10.09s/step, loss=2.123, acc=0.574, acc_len=1.31, thru=11.8, I=14.6, T=24.4, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:08:54,791] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:00,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:09:01,480] loop.py:380 INFO TIMING step=1675: step=3.905s data=1.812s compute=3.877s [fwd=2.763s bwd=1.100s opt=0.003s] dispatch=2.868s [2026-10-18 17:09:01,480] loop.py:45 INFO TRAIN_STEP step=1675 loss=2.029676 Training: 9%|▉ | 1675/18018 [3:38:47<41:17:28, 9.10s/step, loss=2.030, acc=0.589, acc_len=1.35, thru=15.2, I=22.9, T=32.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:01,970] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:08,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:09:09,356] loop.py:45 INFO TRAIN_STEP step=1676 loss=2.081146 Training: 9%|▉ | 1676/18018 [3:38:55<39:37:35, 8.73s/step, loss=2.081, acc=0.580, acc_len=1.33, thru=15.0, I=17.6, T=29.2, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:09:09,351] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1675: forward=3230.8ms backward=1086.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:09,779] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:14,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:09:15,755] loop.py:45 INFO TRAIN_STEP step=1677 loss=2.079416 Training: 9%|▉ | 1677/18018 [3:39:01<36:27:04, 8.03s/step, loss=2.079, acc=0.582, acc_len=1.33, thru=17.7, I=22.4, T=30.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:16,226] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:20,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:09:22,412] loop.py:45 INFO TRAIN_STEP step=1678 loss=1.951405 Training: 9%|▉ | 1678/18018 [3:39:08<34:34:47, 7.62s/step, loss=1.951, acc=0.602, acc_len=1.39, thru=15.9, I=22.3, T=34.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:22,809] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:31,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:09:31,439] loop.py:45 INFO TRAIN_STEP step=1679 loss=1.947385 Training: 9%|▉ | 1679/18018 [3:39:17<36:29:37, 8.04s/step, loss=1.947, acc=0.604, acc_len=1.39, thru=14.8, I=14.6, T=31.2, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:31,726] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:37,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:09:39,264] loop.py:380 INFO TIMING step=1680: step=4.616s data=2.016s compute=4.577s [fwd=3.186s bwd=1.375s opt=0.003s] dispatch=3.200s [2026-10-18 17:09:39,264] loop.py:45 INFO TRAIN_STEP step=1680 loss=2.142449 Training: 9%|▉ | 1680/18018 [3:39:25<36:11:58, 7.98s/step, loss=2.142, acc=0.580, acc_len=1.32, thru=17.7, I=19.6, T=27.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:39,487] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:46,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:09:47,067] loop.py:45 INFO TRAIN_STEP step=1681 loss=1.941439 Training: 9%|▉ | 1681/18018 [3:39:33<35:57:45, 7.92s/step, loss=1.941, acc=0.601, acc_len=1.39, thru=15.7, I=20.3, T=32.4, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:09:47,062] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1680: forward=2802.6ms backward=1106.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:47,104] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:51,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:09:52,997] loop.py:45 INFO TRAIN_STEP step=1682 loss=2.074876 Training: 9%|▉ | 1682/18018 [3:39:39<33:14:42, 7.33s/step, loss=2.075, acc=0.582, acc_len=1.33, thru=19.8, I=24.5, T=33.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:53,273] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:58,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:09:59,211] loop.py:45 INFO TRAIN_STEP step=1683 loss=2.021578 Training: 9%|▉ | 1683/18018 [3:39:45<31:43:41, 6.99s/step, loss=2.022, acc=0.591, acc_len=1.37, thru=19.3, I=23.6, T=38.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:09:59,262] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:04,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:10:05,221] loop.py:45 INFO TRAIN_STEP step=1684 loss=2.129182 Training: 9%|▉ | 1684/18018 [3:39:51<30:23:14, 6.70s/step, loss=2.129, acc=0.574, acc_len=1.31, thru=20.8, I=23.5, T=32.6, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:05,248] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:10,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:10:11,008] loop.py:380 INFO TIMING step=1685: step=4.033s data=1.752s compute=3.995s [fwd=2.823s bwd=1.157s opt=0.003s] dispatch=1.744s [2026-10-18 17:10:11,008] loop.py:45 INFO TRAIN_STEP step=1685 loss=2.093617 Training: 9%|▉ | 1685/18018 [3:39:57<29:08:46, 6.42s/step, loss=2.094, acc=0.578, acc_len=1.32, thru=20.6, I=23.5, T=31.7, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:11,365] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:10:18,821] loop.py:45 INFO TRAIN_STEP step=1686 loss=2.066038 Training: 9%|▉ | 1686/18018 [3:40:04<31:02:00, 6.84s/step, loss=2.066, acc=0.590, acc_len=1.35, thru=14.6, I=16.4, T=32.8, wait=3.9s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:10:18,816] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1685: forward=2774.1ms backward=1075.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:31,693] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:10:31,906] loop.py:45 INFO TRAIN_STEP step=1687 loss=2.108898 Training: 9%|▉ | 1687/18018 [3:40:17<39:31:56, 8.71s/step, loss=2.109, acc=0.579, acc_len=1.32, thru=15.7, I=9.2, T=30.4, wait=8.9s, pool=64, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:32,195] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:39,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:10:40,516] loop.py:45 INFO TRAIN_STEP step=1688 loss=2.121900 Training: 9%|▉ | 1688/18018 [3:40:26<39:23:16, 8.68s/step, loss=2.122, acc=0.576, acc_len=1.31, thru=15.6, I=16.1, T=31.5, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:40,816] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:47,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:10:47,773] loop.py:45 INFO TRAIN_STEP step=1689 loss=1.986932 Training: 9%|▉ | 1689/18018 [3:40:33<37:26:37, 8.26s/step, loss=1.987, acc=0.596, acc_len=1.37, thru=17.3, I=20.4, T=28.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:48,104] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:54,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:10:55,350] loop.py:380 INFO TIMING step=1690: step=3.784s data=1.568s compute=3.752s [fwd=2.643s bwd=1.094s opt=0.003s] dispatch=3.784s [2026-10-18 17:10:55,351] loop.py:45 INFO TRAIN_STEP step=1690 loss=2.115828 Training: 9%|▉ | 1690/18018 [3:40:41<36:31:00, 8.05s/step, loss=2.116, acc=0.573, acc_len=1.30, thru=15.2, I=17.1, T=33.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:10:55,447] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 17:11:01,645] loop.py:45 INFO TRAIN_STEP step=1691 loss=1.989859 Training: 9%|▉ | 1691/18018 [3:40:47<34:07:34, 7.52s/step, loss=1.990, acc=0.595, acc_len=1.37, thru=18.4, I=21.3, T=36.4, wait=2.8s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:11:01,640] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1690: forward=2482.4ms backward=988.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:08,919] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:11:10,171] loop.py:45 INFO TRAIN_STEP step=1692 loss=2.103966 Training: 9%|▉ | 1692/18018 [3:40:56<35:29:28, 7.83s/step, loss=2.104, acc=0.582, acc_len=1.33, thru=16.1, I=18.0, T=33.9, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:10,429] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:11:17,745] loop.py:45 INFO TRAIN_STEP step=1693 loss=1.938295 Training: 9%|▉ | 1693/18018 [3:41:03<35:08:28, 7.75s/step, loss=1.938, acc=0.598, acc_len=1.38, thru=13.8, I=17.6, T=30.9, wait=3.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:24,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:11:26,688] loop.py:45 INFO TRAIN_STEP step=1694 loss=2.019081 Training: 9%|▉ | 1694/18018 [3:41:12<36:45:54, 8.11s/step, loss=2.019, acc=0.589, acc_len=1.36, thru=16.3, I=15.9, T=28.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:26,969] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:11:34,115] loop.py:380 INFO TIMING step=1695: step=3.670s data=1.499s compute=3.635s [fwd=2.595s bwd=1.024s opt=0.003s] dispatch=3.747s [2026-10-18 17:11:34,115] loop.py:45 INFO TRAIN_STEP step=1695 loss=2.145659 Training: 9%|▉ | 1695/18018 [3:41:20<35:50:05, 7.90s/step, loss=2.146, acc=0.572, acc_len=1.30, thru=13.5, I=15.0, T=34.9, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:40,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:11:41,191] loop.py:45 INFO TRAIN_STEP step=1696 loss=1.972663 Training: 9%|▉ | 1696/18018 [3:41:27<34:42:34, 7.66s/step, loss=1.973, acc=0.598, acc_len=1.37, thru=19.5, I=19.4, T=30.1, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:11:41,187] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1695: forward=2953.5ms backward=1258.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:41,345] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:48,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:11:49,048] loop.py:45 INFO TRAIN_STEP step=1697 loss=1.920024 Training: 9%|▉ | 1697/18018 [3:41:35<34:58:53, 7.72s/step, loss=1.920, acc=0.602, acc_len=1.39, thru=15.6, I=16.1, T=32.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:49,448] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:57,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:11:57,384] loop.py:45 INFO TRAIN_STEP step=1698 loss=2.046560 Training: 9%|▉ | 1698/18018 [3:41:43<35:49:14, 7.90s/step, loss=2.047, acc=0.584, acc_len=1.35, thru=15.9, I=16.0, T=32.5, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:11:57,757] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:04,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:12:05,997] loop.py:45 INFO TRAIN_STEP step=1699 loss=2.208545 Training: 9%|▉ | 1699/18018 [3:41:52<36:47:04, 8.11s/step, loss=2.209, acc=0.564, acc_len=1.28, thru=16.2, I=19.2, T=26.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:06,198] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:10,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:12:12,503] loop.py:380 INFO TIMING step=1700: step=4.140s data=1.884s compute=4.106s [fwd=2.858s bwd=1.234s opt=0.003s] dispatch=2.358s [2026-10-18 17:12:12,503] loop.py:45 INFO TRAIN_STEP step=1700 loss=1.989874 Training: 9%|▉ | 1700/18018 [3:41:58<34:35:49, 7.63s/step, loss=1.990, acc=0.593, acc_len=1.36, thru=18.3, I=23.9, T=30.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:12,844] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:17,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:12:19,026] loop.py:45 INFO TRAIN_STEP step=1701 loss=1.932534 Training: 9%|▉ | 1701/18018 [3:42:05<33:04:57, 7.30s/step, loss=1.933, acc=0.602, acc_len=1.39, thru=18.8, I=26.2, T=34.7, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:12:19,021] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1700: forward=2596.4ms backward=1050.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:19,526] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:24,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:12:25,449] loop.py:45 INFO TRAIN_STEP step=1702 loss=2.140163 Training: 9%|▉ | 1702/18018 [3:42:11<31:53:35, 7.04s/step, loss=2.140, acc=0.575, acc_len=1.31, thru=19.1, I=22.3, T=36.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:25,919] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:31,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:12:31,939] loop.py:45 INFO TRAIN_STEP step=1703 loss=2.057345 Training: 9%|▉ | 1703/18018 [3:42:18<31:08:53, 6.87s/step, loss=2.057, acc=0.584, acc_len=1.34, thru=19.3, I=22.5, T=30.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:32,199] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:38,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:12:40,239] loop.py:45 INFO TRAIN_STEP step=1704 loss=2.230473 Training: 9%|▉ | 1704/18018 [3:42:26<33:05:08, 7.30s/step, loss=2.230, acc=0.561, acc_len=1.26, thru=15.2, I=19.2, T=34.5, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:12:40,575] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:12:47,146] loop.py:380 INFO TIMING step=1705: step=3.715s data=1.586s compute=3.689s [fwd=2.594s bwd=1.080s opt=0.003s] dispatch=3.183s [2026-10-18 17:12:47,147] loop.py:45 INFO TRAIN_STEP step=1705 loss=1.928206 Training: 9%|▉ | 1705/18018 [3:42:33<32:32:57, 7.18s/step, loss=1.928, acc=0.604, acc_len=1.39, thru=14.6, I=22.4, T=34.5, wait=3.2s, pool=36, epoch=1/2][2026-10-18 17:12:56,540] loop.py:45 INFO TRAIN_STEP step=1706 loss=1.967905 Training: 9%|▉ | 1706/18018 [3:42:42<35:33:04, 7.85s/step, loss=1.968, acc=0.596, acc_len=1.38, thru=15.8, I=15.7, T=30.1, wait=5.1s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:12:56,536] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1705: forward=2961.7ms backward=1241.2ms [2026-10-18 17:13:04,485] loop.py:45 INFO TRAIN_STEP step=1707 loss=2.023086 Training: 9%|▉ | 1707/18018 [3:42:50<35:40:53, 7.88s/step, loss=2.023, acc=0.586, acc_len=1.35, thru=15.4, I=15.5, T=33.3, wait=4.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:13:12,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:13:13,087] loop.py:45 INFO TRAIN_STEP step=1708 loss=1.993357 Training: 9%|▉ | 1708/18018 [3:42:59<36:40:04, 8.09s/step, loss=1.993, acc=0.593, acc_len=1.37, thru=15.7, I=14.7, T=31.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:13:13,458] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 17:13:22,178] loop.py:45 INFO TRAIN_STEP step=1709 loss=2.072969 Training: 9%|▉ | 1709/18018 [3:43:08<38:01:11, 8.39s/step, loss=2.073, acc=0.581, acc_len=1.33, thru=12.4, I=13.3, T=34.4, wait=5.4s, pool=36, epoch=1/2][2026-10-18 17:13:32,279] loop.py:380 INFO TIMING step=1710: step=4.249s data=1.823s compute=4.196s [fwd=2.968s bwd=1.213s opt=0.003s] dispatch=5.843s [2026-10-18 17:13:32,280] loop.py:45 INFO TRAIN_STEP step=1710 loss=2.234025 Training: 9%|▉ | 1710/18018 [3:43:18<40:20:39, 8.91s/step, loss=2.234, acc=0.557, acc_len=1.26, thru=14.6, I=13.9, T=30.1, wait=5.8s, pool=52, epoch=1/2][2026-10-18 17:13:43,400] loop.py:45 INFO TRAIN_STEP step=1711 loss=2.188317 Training: 9%|▉ | 1711/18018 [3:43:29<43:20:45, 9.57s/step, loss=2.188, acc=0.571, acc_len=1.30, thru=13.4, I=11.9, T=32.2, wait=7.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:13:43,395] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1710: forward=2684.8ms backward=1210.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:13:43,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:13:43,893] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 17:13:50,227] loop.py:45 INFO TRAIN_STEP step=1712 loss=1.996303 Training: 10%|▉ | 1712/18018 [3:43:36<39:37:23, 8.75s/step, loss=1.996, acc=0.593, acc_len=1.37, thru=16.7, I=15.2, T=35.2, wait=3.2s, pool=36, epoch=1/2](TrainerActor pid=9370) [2026-10-18 17:13:50,308] eagle_store.py:517 WARNING batch_get_buffer missing keys (k54787-826c02d3_hs, k54787-826c02d3_ids, k54787-826c02d3_lhs); sleeping 0.50s. [2026-10-18 17:14:00,487] loop.py:45 INFO TRAIN_STEP step=1713 loss=2.099208 Training: 10%|▉ | 1713/18018 [3:43:46<41:40:26, 9.20s/step, loss=2.099, acc=0.583, acc_len=1.34, thru=14.5, I=14.6, T=30.4, wait=6.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:14:08,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:14:09,179] loop.py:45 INFO TRAIN_STEP step=1714 loss=1.852822 Training: 10%|▉ | 1714/18018 [3:43:55<40:58:40, 9.05s/step, loss=1.853, acc=0.617, acc_len=1.44, thru=16.2, I=14.7, T=32.0, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:14:09,391] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 17:14:16,495] loop.py:380 INFO TIMING step=1715: step=3.720s data=1.474s compute=3.686s [fwd=2.597s bwd=1.075s opt=0.003s] dispatch=3.586s [2026-10-18 17:14:16,495] loop.py:45 INFO TRAIN_STEP step=1715 loss=2.052824 Training: 10%|▉ | 1715/18018 [3:44:02<38:37:11, 8.53s/step, loss=2.053, acc=0.583, acc_len=1.34, thru=16.5, I=15.8, T=34.4, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:14:23,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:14:24,829] loop.py:45 INFO TRAIN_STEP step=1716 loss=2.034809 Training: 10%|▉ | 1716/18018 [3:44:10<38:21:11, 8.47s/step, loss=2.035, acc=0.595, acc_len=1.37, thru=16.0, I=16.8, T=30.7, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:14:24,824] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1715: forward=2932.6ms backward=1189.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:14:25,098] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:14:30,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:14:33,424] loop.py:45 INFO TRAIN_STEP step=1717 loss=2.183531 Training: 10%|▉ | 1717/18018 [3:44:19<38:31:25, 8.51s/step, loss=2.184, acc=0.564, acc_len=1.28, thru=13.8, I=21.2, T=29.9, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:14:33,866] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 17:14:39,543] loop.py:45 INFO TRAIN_STEP step=1718 loss=1.930276 Training: 10%|▉ | 1718/18018 [3:44:25<35:16:42, 7.79s/step, loss=1.930, acc=0.607, acc_len=1.41, thru=15.7, I=24.1, T=35.4, wait=2.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:14:46,755] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:14:48,227] loop.py:45 INFO TRAIN_STEP step=1719 loss=2.006178 Training: 10%|▉ | 1719/18018 [3:44:34<36:29:20, 8.06s/step, loss=2.006, acc=0.594, acc_len=1.37, thru=17.0, I=18.2, T=24.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:14:48,264] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 17:14:56,229] loop.py:380 INFO TIMING step=1720: step=4.044s data=1.765s compute=3.996s [fwd=2.831s bwd=1.149s opt=0.003s] dispatch=3.947s [2026-10-18 17:14:56,230] loop.py:45 INFO TRAIN_STEP step=1720 loss=1.952852 Training: 10%|▉ | 1720/18018 [3:44:42<36:24:34, 8.04s/step, loss=1.953, acc=0.602, acc_len=1.39, thru=12.0, I=14.0, T=31.7, wait=3.9s, pool=44, epoch=1/2][2026-10-18 17:15:05,442] loop.py:45 INFO TRAIN_STEP step=1721 loss=2.000720 Training: 10%|▉ | 1721/18018 [3:44:51<37:59:31, 8.39s/step, loss=2.001, acc=0.593, acc_len=1.37, thru=15.7, I=14.2, T=29.8, wait=4.9s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:15:05,437] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1720: forward=3019.2ms backward=1233.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:12,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:15:14,083] loop.py:45 INFO TRAIN_STEP step=1722 loss=2.167589 Training: 10%|▉ | 1722/18018 [3:45:00<38:19:45, 8.47s/step, loss=2.168, acc=0.570, acc_len=1.30, thru=15.2, I=17.2, T=26.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:14,431] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:20,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:15:21,283] loop.py:45 INFO TRAIN_STEP step=1723 loss=2.050892 Training: 10%|▉ | 1723/18018 [3:45:07<36:36:15, 8.09s/step, loss=2.051, acc=0.583, acc_len=1.34, thru=16.1, I=20.9, T=31.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:21,704] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:25,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:15:28,560] loop.py:45 INFO TRAIN_STEP step=1724 loss=2.012822 Training: 10%|▉ | 1724/18018 [3:45:14<35:30:17, 7.84s/step, loss=2.013, acc=0.589, acc_len=1.34, thru=16.7, I=27.9, T=26.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:28,861] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:34,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:15:36,299] loop.py:380 INFO TIMING step=1725: step=4.546s data=2.152s compute=4.508s [fwd=3.123s bwd=1.369s opt=0.003s] dispatch=3.182s [2026-10-18 17:15:36,299] loop.py:45 INFO TRAIN_STEP step=1725 loss=1.932429 Training: 10%|▉ | 1725/18018 [3:45:22<35:21:36, 7.81s/step, loss=1.932, acc=0.599, acc_len=1.39, thru=17.2, I=23.0, T=28.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:36,439] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-18 17:15:42,146] loop.py:45 INFO TRAIN_STEP step=1726 loss=2.110320 Training: 10%|▉ | 1726/18018 [3:45:28<32:41:10, 7.22s/step, loss=2.110, acc=0.579, acc_len=1.32, thru=18.3, I=22.3, T=36.1, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:15:42,141] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1725: forward=2509.8ms backward=991.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:42,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:42,587] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:47,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:15:48,341] loop.py:45 INFO TRAIN_STEP step=1727 loss=1.979540 Training: 10%|▉ | 1727/18018 [3:45:34<31:17:28, 6.91s/step, loss=1.980, acc=0.598, acc_len=1.38, thru=20.0, I=22.7, T=33.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:48,452] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:53,840] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:15:54,380] loop.py:45 INFO TRAIN_STEP step=1728 loss=2.133285 Training: 10%|▉ | 1728/18018 [3:45:40<30:05:52, 6.65s/step, loss=2.133, acc=0.572, acc_len=1.30, thru=20.1, I=23.7, T=37.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:15:54,845] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:00,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:16:02,418] loop.py:45 INFO TRAIN_STEP step=1729 loss=2.046613 Training: 10%|▉ | 1729/18018 [3:45:48<31:58:49, 7.07s/step, loss=2.047, acc=0.588, acc_len=1.35, thru=16.6, I=22.6, T=28.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:02,835] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:08,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:16:08,952] loop.py:380 INFO TIMING step=1730: step=4.413s data=2.090s compute=4.382s [fwd=3.144s bwd=1.224s opt=0.003s] dispatch=2.111s [2026-10-18 17:16:08,953] loop.py:45 INFO TRAIN_STEP step=1730 loss=1.915621 Training: 10%|▉ | 1730/18018 [3:45:55<31:15:19, 6.91s/step, loss=1.916, acc=0.612, acc_len=1.42, thru=17.3, I=26.8, T=29.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:09,163] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:14,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:16:16,723] loop.py:45 INFO TRAIN_STEP step=1731 loss=2.111795 Training: 10%|▉ | 1731/18018 [3:46:02<32:25:29, 7.17s/step, loss=2.112, acc=0.575, acc_len=1.31, thru=14.8, I=18.7, T=26.3, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:16:16,718] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1730: forward=3473.6ms backward=1349.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:16,822] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:20,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:16:22,373] loop.py:45 INFO TRAIN_STEP step=1732 loss=2.078722 Training: 10%|▉ | 1732/18018 [3:46:08<30:21:40, 6.71s/step, loss=2.079, acc=0.582, acc_len=1.33, thru=18.6, I=25.6, T=32.5, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:22,454] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 17:16:29,949] loop.py:45 INFO TRAIN_STEP step=1733 loss=1.975467 Training: 10%|▉ | 1733/18018 [3:46:16<31:31:57, 6.97s/step, loss=1.975, acc=0.596, acc_len=1.38, thru=14.2, I=21.2, T=30.4, wait=3.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:38,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:16:38,835] loop.py:45 INFO TRAIN_STEP step=1734 loss=1.976671 Training: 10%|▉ | 1734/18018 [3:46:24<34:07:54, 7.55s/step, loss=1.977, acc=0.599, acc_len=1.38, thru=14.8, I=14.0, T=32.9, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:38,947] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:47,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:16:48,064] loop.py:380 INFO TIMING step=1735: step=4.508s data=2.120s compute=4.464s [fwd=3.198s bwd=1.252s opt=0.003s] dispatch=4.711s [2026-10-18 17:16:48,064] loop.py:45 INFO TRAIN_STEP step=1735 loss=2.056591 Training: 10%|▉ | 1735/18018 [3:46:34<36:24:41, 8.05s/step, loss=2.057, acc=0.592, acc_len=1.36, thru=14.1, I=15.1, T=28.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:16:48,240] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:16:55,256] loop.py:45 INFO TRAIN_STEP step=1736 loss=1.975418 Training: 10%|▉ | 1736/18018 [3:46:41<35:14:38, 7.79s/step, loss=1.975, acc=0.595, acc_len=1.37, thru=14.9, I=17.4, T=29.3, wait=2.8s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:16:55,252] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1735: forward=3016.3ms backward=1294.5ms [2026-10-18 17:17:06,069] loop.py:45 INFO TRAIN_STEP step=1737 loss=1.975023 Training: 10%|▉ | 1737/18018 [3:46:52<39:20:35, 8.70s/step, loss=1.975, acc=0.603, acc_len=1.39, thru=10.1, I=11.4, T=30.4, wait=6.6s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:14,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:17:15,977] loop.py:45 INFO TRAIN_STEP step=1738 loss=2.264152 Training: 10%|▉ | 1738/18018 [3:47:02<40:58:41, 9.06s/step, loss=2.264, acc=0.558, acc_len=1.27, thru=16.3, I=16.5, T=31.1, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:16,049] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:20,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:17:21,199] loop.py:45 INFO TRAIN_STEP step=1739 loss=2.208007 Training: 10%|▉ | 1739/18018 [3:47:07<35:46:00, 7.91s/step, loss=2.208, acc=0.562, acc_len=1.27, thru=20.9, I=29.1, T=36.3, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:21,430] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:17:31,202] loop.py:380 INFO TIMING step=1740: step=4.133s data=1.782s compute=4.094s [fwd=2.820s bwd=1.261s opt=0.003s] dispatch=5.863s [2026-10-18 17:17:31,202] loop.py:45 INFO TRAIN_STEP step=1740 loss=2.127357 Training: 10%|▉ | 1740/18018 [3:47:17<38:36:32, 8.54s/step, loss=2.127, acc=0.573, acc_len=1.30, thru=11.1, I=10.6, T=31.0, wait=5.9s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:38,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:17:40,042] loop.py:45 INFO TRAIN_STEP step=1741 loss=1.910965 Training: 10%|▉ | 1741/18018 [3:47:26<39:00:48, 8.63s/step, loss=1.911, acc=0.603, acc_len=1.40, thru=16.8, I=16.4, T=32.9, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:17:40,038] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1740: forward=2790.0ms backward=1056.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:40,137] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:45,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:17:46,268] loop.py:45 INFO TRAIN_STEP step=1742 loss=2.047870 Training: 10%|▉ | 1742/18018 [3:47:32<35:44:52, 7.91s/step, loss=2.048, acc=0.584, acc_len=1.34, thru=18.5, I=26.1, T=35.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:46,715] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:53,032] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:17:53,478] loop.py:45 INFO TRAIN_STEP step=1743 loss=1.990889 Training: 10%|▉ | 1743/18018 [3:47:39<34:48:13, 7.70s/step, loss=1.991, acc=0.602, acc_len=1.39, thru=17.5, I=21.6, T=33.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:17:53,538] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:00,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:18:02,258] loop.py:45 INFO TRAIN_STEP step=1744 loss=2.005000 Training: 10%|▉ | 1744/18018 [3:47:48<36:16:07, 8.02s/step, loss=2.005, acc=0.589, acc_len=1.36, thru=15.0, I=18.0, T=26.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:02,408] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:07,680] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:18:09,559] loop.py:380 INFO TIMING step=1745: step=4.324s data=1.917s compute=4.296s [fwd=3.179s bwd=1.103s opt=0.003s] dispatch=2.966s [2026-10-18 17:18:09,559] loop.py:45 INFO TRAIN_STEP step=1745 loss=2.140153 Training: 10%|▉ | 1745/18018 [3:47:55<35:17:04, 7.81s/step, loss=2.140, acc=0.574, acc_len=1.31, thru=15.6, I=19.3, T=29.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:09,700] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:15,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:18:16,340] loop.py:45 INFO TRAIN_STEP step=1746 loss=2.105470 Training: 10%|▉ | 1746/18018 [3:48:02<33:53:43, 7.50s/step, loss=2.105, acc=0.583, acc_len=1.33, thru=16.9, I=21.2, T=30.0, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:18:16,335] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1745: forward=2982.8ms backward=1237.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:16,769] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:23,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:18:23,854] loop.py:45 INFO TRAIN_STEP step=1747 loss=2.067110 Training: 10%|▉ | 1747/18018 [3:48:09<33:54:37, 7.50s/step, loss=2.067, acc=0.584, acc_len=1.33, thru=16.0, I=17.8, T=30.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:24,272] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:29,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:18:31,526] loop.py:45 INFO TRAIN_STEP step=1748 loss=1.982138 Training: 10%|▉ | 1748/18018 [3:48:17<34:08:22, 7.55s/step, loss=1.982, acc=0.595, acc_len=1.37, thru=16.9, I=20.7, T=28.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:31,941] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 17:18:38,561] loop.py:45 INFO TRAIN_STEP step=1749 loss=1.794582 Training: 10%|▉ | 1749/18018 [3:48:24<33:25:53, 7.40s/step, loss=1.795, acc=0.622, acc_len=1.46, thru=15.4, I=21.4, T=29.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:38,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:39,013] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:44,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:18:45,489] loop.py:380 INFO TIMING step=1750: step=3.808s data=1.669s compute=3.774s [fwd=2.672s bwd=1.088s opt=0.003s] dispatch=3.113s [2026-10-18 17:18:45,489] loop.py:45 INFO TRAIN_STEP step=1750 loss=2.170995 Training: 10%|▉ | 1750/18018 [3:48:31<32:47:49, 7.26s/step, loss=2.171, acc=0.563, acc_len=1.27, thru=17.0, I=18.1, T=33.6, wait=3.1s, pool=64, epoch=1/2][2026-10-18 17:18:45,494] loop.py:419 INFO Saving checkpoint at step 1750... (TrainerActor pid=9310) [2026-10-18 17:18:48,853] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0001751 (TrainerActor pid=9310) [2026-10-18 17:18:48,863] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0001001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:49,579] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:50,080] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:55,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 17:18:48,855] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0001751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 17:18:49,915] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0001001 [2026-10-18 17:18:57,119] loop.py:45 INFO TRAIN_STEP step=1751 loss=2.100774 Training: 10%|▉ | 1751/18018 [3:48:43<38:43:18, 8.57s/step, loss=2.101, acc=0.584, acc_len=1.34, thru=18.2, I=21.7, T=34.2, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:18:57,114] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1750: forward=2633.9ms backward=1035.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:18:57,397] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:19:02,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:19:03,231] loop.py:45 INFO TRAIN_STEP step=1752 loss=2.019158 Training: 10%|▉ | 1752/18018 [3:48:49<35:23:05, 7.83s/step, loss=2.019, acc=0.588, acc_len=1.34, thru=19.8, I=23.5, T=37.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:19:03,385] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:19:11,145] loop.py:45 INFO TRAIN_STEP step=1753 loss=2.159194 Training: 10%|▉ | 1753/18018 [3:48:57<35:29:55, 7.86s/step, loss=2.159, acc=0.571, acc_len=1.30, thru=15.2, I=17.4, T=30.6, wait=3.7s, pool=52, epoch=1/2][2026-10-18 17:19:18,349] loop.py:45 INFO TRAIN_STEP step=1754 loss=2.112589 Training: 10%|▉ | 1754/18018 [3:49:04<34:36:28, 7.66s/step, loss=2.113, acc=0.576, acc_len=1.31, thru=16.0, I=17.4, T=34.6, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:19:24,032] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:19:25,692] loop.py:380 INFO TIMING step=1755: step=4.278s data=1.849s compute=4.244s [fwd=3.024s bwd=1.205s opt=0.003s] dispatch=3.056s [2026-10-18 17:19:25,693] loop.py:45 INFO TRAIN_STEP step=1755 loss=1.992028 Training: 10%|▉ | 1755/18018 [3:49:11<34:10:53, 7.57s/step, loss=1.992, acc=0.599, acc_len=1.38, thru=18.0, I=19.5, T=29.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:19:26,048] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:19:31,977] loop.py:45 INFO TRAIN_STEP step=1756 loss=1.859274 (TrainerActor pid=9310) [2026-10-18 17:19:31,972] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1755: forward=2683.5ms backward=1044.6ms Training: 10%|▉ | 1756/18018 [3:49:18<32:26:29, 7.18s/step, loss=1.859, acc=0.615, acc_len=1.43, thru=16.3, I=20.4, T=33.8, wait=2.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:19:37,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:19:38,615] loop.py:45 INFO TRAIN_STEP step=1757 loss=2.121176 Training: 10%|▉ | 1757/18018 [3:49:24<31:41:52, 7.02s/step, loss=2.121, acc=0.579, acc_len=1.33, thru=18.8, I=20.9, T=34.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:19:38,658] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-18 17:19:44,957] loop.py:45 INFO TRAIN_STEP step=1758 loss=1.985935 Training: 10%|▉ | 1758/18018 [3:49:31<30:46:55, 6.82s/step, loss=1.986, acc=0.593, acc_len=1.37, thru=17.9, I=19.9, T=34.9, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:19:52,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:19:53,422] loop.py:45 INFO TRAIN_STEP step=1759 loss=2.140880 Training: 10%|▉ | 1759/18018 [3:49:39<33:00:56, 7.31s/step, loss=2.141, acc=0.573, acc_len=1.30, thru=18.1, I=18.2, T=34.3, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:19:53,886] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:00,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:20:01,196] loop.py:380 INFO TIMING step=1760: step=3.893s data=1.685s compute=3.868s [fwd=2.778s bwd=1.076s opt=0.003s] dispatch=3.872s [2026-10-18 17:20:01,196] loop.py:45 INFO TRAIN_STEP step=1760 loss=1.996219 Training: 10%|▉ | 1760/18018 [3:49:47<33:38:21, 7.45s/step, loss=1.996, acc=0.593, acc_len=1.36, thru=15.2, I=17.6, T=32.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:01,364] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:07,274] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:20:08,107] loop.py:45 INFO TRAIN_STEP step=1761 loss=1.837649 Training: 10%|▉ | 1761/18018 [3:49:54<32:54:36, 7.29s/step, loss=1.838, acc=0.616, acc_len=1.43, thru=18.1, I=20.5, T=32.3, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:20:08,102] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1760: forward=2732.4ms backward=1167.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:08,281] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:14,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:20:15,726] loop.py:45 INFO TRAIN_STEP step=1762 loss=2.095961 Training: 10%|▉ | 1762/18018 [3:50:01<33:21:23, 7.39s/step, loss=2.096, acc=0.587, acc_len=1.34, thru=16.8, I=20.2, T=29.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:16,092] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:20:22,272] loop.py:45 INFO TRAIN_STEP step=1763 loss=2.023123 Training: 10%|▉ | 1763/18018 [3:50:08<32:13:08, 7.14s/step, loss=2.023, acc=0.591, acc_len=1.36, thru=14.8, I=27.2, T=31.7, wait=2.5s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:30,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:20:31,128] loop.py:45 INFO TRAIN_STEP step=1764 loss=2.094077 Training: 10%|▉ | 1764/18018 [3:50:17<34:32:47, 7.65s/step, loss=2.094, acc=0.584, acc_len=1.34, thru=17.5, I=14.8, T=31.4, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:31,580] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:37,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:20:38,049] loop.py:380 INFO TIMING step=1765: step=4.052s data=1.822s compute=4.017s [fwd=2.823s bwd=1.180s opt=0.003s] dispatch=2.859s [2026-10-18 17:20:38,049] loop.py:45 INFO TRAIN_STEP step=1765 loss=2.061098 Training: 10%|▉ | 1765/18018 [3:50:24<33:33:07, 7.43s/step, loss=2.061, acc=0.585, acc_len=1.33, thru=17.9, I=20.0, T=31.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:38,453] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:43,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:20:44,533] loop.py:45 INFO TRAIN_STEP step=1766 loss=1.916980 Training: 10%|▉ | 1766/18018 [3:50:30<32:16:03, 7.15s/step, loss=1.917, acc=0.602, acc_len=1.39, thru=19.6, I=21.8, T=33.1, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:20:44,528] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1765: forward=2686.8ms backward=1128.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:44,810] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:20:50,441] loop.py:45 INFO TRAIN_STEP step=1767 loss=2.167759 Training: 10%|▉ | 1767/18018 [3:50:36<30:35:14, 6.78s/step, loss=2.168, acc=0.570, acc_len=1.29, thru=18.4, I=21.9, T=33.8, wait=2.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:56,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:20:59,007] loop.py:45 INFO TRAIN_STEP step=1768 loss=2.295277 Training: 10%|▉ | 1768/18018 [3:50:45<33:00:28, 7.31s/step, loss=2.295, acc=0.552, acc_len=1.24, thru=16.8, I=19.9, T=28.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:20:59,111] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 [2026-10-18 17:21:10,651] loop.py:45 INFO TRAIN_STEP step=1769 loss=2.111277 Training: 10%|▉ | 1769/18018 [3:50:56<38:52:12, 8.61s/step, loss=2.111, acc=0.579, acc_len=1.32, thru=7.3, I=7.8, T=31.5, wait=7.6s, pool=40, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:23,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:21:25,052] loop.py:380 INFO TIMING step=1770: step=5.255s data=2.754s compute=5.202s [fwd=3.902s bwd=1.285s opt=0.003s] dispatch=9.138s [2026-10-18 17:21:25,053] loop.py:45 INFO TRAIN_STEP step=1770 loss=2.236650 Training: 10%|▉ | 1770/18018 [3:51:11<46:42:39, 10.35s/step, loss=2.237, acc=0.559, acc_len=1.26, thru=11.7, I=11.3, T=24.4, wait=9.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:25,278] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:30,209] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:21:31,754] loop.py:45 INFO TRAIN_STEP step=1771 loss=2.144994 Training: 10%|▉ | 1771/18018 [3:51:17<41:45:59, 9.25s/step, loss=2.145, acc=0.577, acc_len=1.31, thru=17.0, I=23.2, T=31.2, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:21:31,749] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1770: forward=2882.1ms backward=1176.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:32,223] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:37,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:21:39,325] loop.py:45 INFO TRAIN_STEP step=1772 loss=1.913126 Training: 10%|▉ | 1772/18018 [3:51:25<39:29:12, 8.75s/step, loss=1.913, acc=0.602, acc_len=1.40, thru=15.7, I=22.3, T=32.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:39,446] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:44,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:21:45,675] loop.py:45 INFO TRAIN_STEP step=1773 loss=2.069779 Training: 10%|▉ | 1773/18018 [3:51:31<36:13:51, 8.03s/step, loss=2.070, acc=0.585, acc_len=1.34, thru=19.0, I=26.3, T=34.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:46,028] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:51,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:21:52,177] loop.py:45 INFO TRAIN_STEP step=1774 loss=2.041545 Training: 10%|▉ | 1774/18018 [3:51:38<34:09:57, 7.57s/step, loss=2.042, acc=0.593, acc_len=1.37, thru=18.6, I=23.7, T=33.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:52,340] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:57,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:21:58,453] loop.py:380 INFO TIMING step=1775: step=3.954s data=1.752s compute=3.925s [fwd=2.793s bwd=1.117s opt=0.003s] dispatch=2.312s [2026-10-18 17:21:58,453] loop.py:45 INFO TRAIN_STEP step=1775 loss=1.920138 Training: 10%|▉ | 1775/18018 [3:51:44<32:24:36, 7.18s/step, loss=1.920, acc=0.601, acc_len=1.39, thru=19.4, I=24.4, T=32.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:21:58,618] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:07,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:22:09,063] loop.py:45 INFO TRAIN_STEP step=1776 loss=2.032988 Training: 10%|▉ | 1776/18018 [3:51:55<37:02:43, 8.21s/step, loss=2.033, acc=0.587, acc_len=1.34, thru=12.1, I=13.5, T=27.1, wait=5.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:22:09,058] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1775: forward=3312.0ms backward=1363.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:09,346] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:14,565] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:22:15,952] loop.py:45 INFO TRAIN_STEP step=1777 loss=2.050390 Training: 10%|▉ | 1777/18018 [3:52:02<35:15:20, 7.81s/step, loss=2.050, acc=0.588, acc_len=1.34, thru=15.2, I=21.4, T=32.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:16,081] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:24,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:22:25,209] loop.py:45 INFO TRAIN_STEP step=1778 loss=1.947101 Training: 10%|▉ | 1778/18018 [3:52:11<37:12:12, 8.25s/step, loss=1.947, acc=0.602, acc_len=1.39, thru=14.1, I=14.4, T=35.6, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:25,329] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-18 17:22:32,692] loop.py:45 INFO TRAIN_STEP step=1779 loss=1.994401 Training: 10%|▉ | 1779/18018 [3:52:18<36:10:03, 8.02s/step, loss=1.994, acc=0.596, acc_len=1.37, thru=17.0, I=17.7, T=29.2, wait=3.1s, pool=52, epoch=1/2][2026-10-18 17:22:40,158] loop.py:380 INFO TIMING step=1780: step=4.141s data=2.088s compute=4.110s [fwd=3.140s bwd=0.956s opt=0.003s] dispatch=3.316s [2026-10-18 17:22:40,158] loop.py:45 INFO TRAIN_STEP step=1780 loss=1.984760 Training: 10%|▉ | 1780/18018 [3:52:26<35:25:02, 7.85s/step, loss=1.985, acc=0.596, acc_len=1.37, thru=15.2, I=15.0, T=30.9, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:46,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:22:47,433] loop.py:45 INFO TRAIN_STEP step=1781 loss=2.167704 (TrainerActor pid=9310) [2026-10-18 17:22:47,429] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1780: forward=2805.2ms backward=1066.5ms Training: 10%|▉ | 1781/18018 [3:52:33<34:38:02, 7.68s/step, loss=2.168, acc=0.570, acc_len=1.29, thru=17.8, I=19.1, T=32.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:47,552] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:52,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:22:54,253] loop.py:45 INFO TRAIN_STEP step=1782 loss=2.156878 Training: 10%|▉ | 1782/18018 [3:52:40<33:28:17, 7.42s/step, loss=2.157, acc=0.572, acc_len=1.30, thru=17.6, I=23.7, T=32.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:22:54,433] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:00,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:23:01,075] loop.py:45 INFO TRAIN_STEP step=1783 loss=1.940743 Training: 10%|▉ | 1783/18018 [3:52:47<32:39:19, 7.24s/step, loss=1.941, acc=0.601, acc_len=1.39, thru=18.0, I=22.1, T=30.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:01,219] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:07,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:23:09,205] loop.py:45 INFO TRAIN_STEP step=1784 loss=1.982076 Training: 10%|▉ | 1784/18018 [3:52:55<33:51:25, 7.51s/step, loss=1.982, acc=0.591, acc_len=1.36, thru=14.7, I=21.1, T=24.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:09,556] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:20,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:23:21,532] loop.py:380 INFO TIMING step=1785: step=5.299s data=2.459s compute=5.265s [fwd=3.806s bwd=1.443s opt=0.003s] dispatch=7.018s [2026-10-18 17:23:21,532] loop.py:45 INFO TRAIN_STEP step=1785 loss=2.131972 Training: 10%|▉ | 1785/18018 [3:53:07<40:22:29, 8.95s/step, loss=2.132, acc=0.571, acc_len=1.30, thru=12.7, I=10.7, T=24.2, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:21,740] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:27,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:23:28,194] loop.py:45 INFO TRAIN_STEP step=1786 loss=2.002272 Training: 10%|▉ | 1786/18018 [3:53:14<37:16:12, 8.27s/step, loss=2.002, acc=0.593, acc_len=1.36, thru=17.0, I=18.6, T=30.4, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:23:28,190] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1785: forward=2837.2ms backward=1324.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:28,284] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 17:23:34,587] loop.py:45 INFO TRAIN_STEP step=1787 loss=2.006488 Training: 10%|▉ | 1787/18018 [3:53:20<34:44:18, 7.70s/step, loss=2.006, acc=0.590, acc_len=1.36, thru=17.3, I=20.1, T=32.3, wait=2.4s, pool=52, epoch=1/2][2026-10-18 17:23:42,735] loop.py:45 INFO TRAIN_STEP step=1788 loss=2.082449 Training: 10%|▉ | 1788/18018 [3:53:28<35:19:53, 7.84s/step, loss=2.082, acc=0.587, acc_len=1.35, thru=14.1, I=13.7, T=27.3, wait=3.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:50,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:23:51,261] loop.py:45 INFO TRAIN_STEP step=1789 loss=2.157857 Training: 10%|▉ | 1789/18018 [3:53:37<36:15:49, 8.04s/step, loss=2.158, acc=0.575, acc_len=1.31, thru=17.2, I=16.0, T=28.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:23:51,296] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 17:23:57,742] loop.py:380 INFO TIMING step=1790: step=3.572s data=1.477s compute=3.543s [fwd=2.533s bwd=0.995s opt=0.003s] dispatch=2.898s [2026-10-18 17:23:57,742] loop.py:45 INFO TRAIN_STEP step=1790 loss=1.954024 Training: 10%|▉ | 1790/18018 [3:53:43<34:08:55, 7.58s/step, loss=1.954, acc=0.601, acc_len=1.38, thru=16.5, I=18.8, T=35.8, wait=2.9s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:24:08,325] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1790: forward=2958.3ms backward=1177.2ms [2026-10-18 17:24:08,330] loop.py:45 INFO TRAIN_STEP step=1791 loss=2.004626 Training: 10%|▉ | 1791/18018 [3:53:54<38:13:06, 8.48s/step, loss=2.005, acc=0.597, acc_len=1.37, thru=9.6, I=10.6, T=30.6, wait=6.4s, pool=16, epoch=1/2] (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 17:24:08,379] eagle_store.py:517 WARNING batch_get_buffer missing keys (k57312-a9231193_hs, k57312-a9231193_ids, k57312-a9231193_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:24:16,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9368) [2026-10-18 17:24:08,832] eagle_store.py:517 WARNING batch_get_buffer missing keys (k57317-a2f433e8_hs, k57317-a2f433e8_ids, k57317-a2f433e8_lhs); sleeping 0.50s. [repeated 5x across cluster] [2026-10-18 17:24:17,593] loop.py:45 INFO TRAIN_STEP step=1792 loss=1.908400 Training: 10%|▉ | 1792/18018 [3:54:03<39:16:38, 8.71s/step, loss=1.908, acc=0.608, acc_len=1.42, thru=18.7, I=14.6, T=32.1, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:24:17,753] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-18 17:24:24,963] loop.py:45 INFO TRAIN_STEP step=1793 loss=1.968598 Training: 10%|▉ | 1793/18018 [3:54:11<37:27:21, 8.31s/step, loss=1.969, acc=0.597, acc_len=1.38, thru=14.6, I=17.4, T=30.5, wait=3.2s, pool=44, epoch=1/2][2026-10-18 17:24:34,057] loop.py:45 INFO TRAIN_STEP step=1794 loss=2.219477 Training: 10%|▉ | 1794/18018 [3:54:20<38:30:41, 8.55s/step, loss=2.219, acc=0.564, acc_len=1.27, thru=14.9, I=13.3, T=30.7, wait=4.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:24:42,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:24:43,687] loop.py:380 INFO TIMING step=1795: step=4.734s data=2.301s compute=4.703s [fwd=3.224s bwd=1.463s opt=0.003s] dispatch=4.887s [2026-10-18 17:24:43,688] loop.py:45 INFO TRAIN_STEP step=1795 loss=2.118247 Training: 10%|▉ | 1795/18018 [3:54:29<39:58:40, 8.87s/step, loss=2.118, acc=0.579, acc_len=1.32, thru=14.3, I=16.5, T=27.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:24:43,936] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:24:49,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:24:50,181] loop.py:45 INFO TRAIN_STEP step=1796 loss=1.876811 Training: 10%|▉ | 1796/18018 [3:54:36<36:45:40, 8.16s/step, loss=1.877, acc=0.612, acc_len=1.42, thru=18.1, I=20.9, T=31.0, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:24:50,176] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1795: forward=2847.5ms backward=1236.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:24:50,493] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 17:24:56,160] loop.py:45 INFO TRAIN_STEP step=1797 loss=2.018269 Training: 10%|▉ | 1797/18018 [3:54:42<33:48:46, 7.50s/step, loss=2.018, acc=0.590, acc_len=1.35, thru=18.2, I=19.4, T=35.6, wait=2.4s, pool=56, epoch=1/2][2026-10-18 17:25:04,117] loop.py:45 INFO TRAIN_STEP step=1798 loss=1.990320 Training: 10%|▉ | 1798/18018 [3:54:50<34:25:16, 7.64s/step, loss=1.990, acc=0.595, acc_len=1.37, thru=15.6, I=15.7, T=31.0, wait=3.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:10,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:25:10,936] loop.py:45 INFO TRAIN_STEP step=1799 loss=1.901556 Training: 10%|▉ | 1799/18018 [3:54:57<33:18:33, 7.39s/step, loss=1.902, acc=0.610, acc_len=1.42, thru=18.2, I=18.8, T=31.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:11,435] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:17,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:25:17,992] loop.py:380 INFO TIMING step=1800: step=3.882s data=1.681s compute=3.841s [fwd=2.677s bwd=1.148s opt=0.003s] dispatch=3.165s [2026-10-18 17:25:17,992] loop.py:45 INFO TRAIN_STEP step=1800 loss=2.039188 Training: 10%|▉ | 1800/18018 [3:55:04<32:50:57, 7.29s/step, loss=2.039, acc=0.583, acc_len=1.34, thru=16.6, I=18.1, T=33.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:18,322] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:24,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:25:26,104] loop.py:45 INFO TRAIN_STEP step=1801 loss=2.095500 (TrainerActor pid=9310) [2026-10-18 17:25:26,099] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1800: forward=2942.1ms backward=1131.3ms Training: 10%|▉ | 1801/18018 [3:55:12<33:57:30, 7.54s/step, loss=2.095, acc=0.581, acc_len=1.32, thru=16.4, I=20.0, T=31.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:26,222] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:31,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:25:34,168] loop.py:45 INFO TRAIN_STEP step=1802 loss=2.085844 Training: 10%|█ | 1802/18018 [3:55:20<34:40:06, 7.70s/step, loss=2.086, acc=0.576, acc_len=1.32, thru=14.0, I=22.4, T=28.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:34,244] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:41,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:25:41,754] loop.py:45 INFO TRAIN_STEP step=1803 loss=2.104427 Training: 10%|█ | 1803/18018 [3:55:27<34:30:54, 7.66s/step, loss=2.104, acc=0.578, acc_len=1.32, thru=17.1, I=19.5, T=30.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:41,881] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:47,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:25:49,327] loop.py:45 INFO TRAIN_STEP step=1804 loss=2.008273 Training: 10%|█ | 1804/18018 [3:55:35<34:23:26, 7.64s/step, loss=2.008, acc=0.593, acc_len=1.37, thru=15.9, I=19.6, T=34.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:25:49,811] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:25:57,559] loop.py:380 INFO TIMING step=1805: step=4.213s data=1.947s compute=4.184s [fwd=2.957s bwd=1.212s opt=0.003s] dispatch=4.011s [2026-10-18 17:25:57,559] loop.py:45 INFO TRAIN_STEP step=1805 loss=2.161560 Training: 10%|█ | 1805/18018 [3:55:43<35:11:40, 7.81s/step, loss=2.162, acc=0.568, acc_len=1.28, thru=12.8, I=15.0, T=30.4, wait=4.0s, pool=56, epoch=1/2][2026-10-18 17:26:07,179] loop.py:45 INFO TRAIN_STEP step=1806 loss=2.072559 Training: 10%|█ | 1806/18018 [3:55:53<37:38:07, 8.36s/step, loss=2.073, acc=0.581, acc_len=1.33, thru=12.5, I=12.6, T=27.7, wait=5.0s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:26:07,175] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1805: forward=3186.2ms backward=1371.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:14,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:26:16,040] loop.py:45 INFO TRAIN_STEP step=1807 loss=2.016758 Training: 10%|█ | 1807/18018 [3:56:02<38:18:33, 8.51s/step, loss=2.017, acc=0.593, acc_len=1.37, thru=16.5, I=18.5, T=27.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:16,212] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:26:22,507] loop.py:45 INFO TRAIN_STEP step=1808 loss=1.945559 Training: 10%|█ | 1808/18018 [3:56:08<35:33:01, 7.90s/step, loss=1.946, acc=0.604, acc_len=1.40, thru=14.7, I=19.0, T=33.7, wait=2.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:28,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:26:30,343] loop.py:45 INFO TRAIN_STEP step=1809 loss=2.043168 Training: 10%|█ | 1809/18018 [3:56:16<35:28:08, 7.88s/step, loss=2.043, acc=0.586, acc_len=1.34, thru=17.4, I=18.5, T=29.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:30,812] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:36,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:26:37,684] loop.py:380 INFO TIMING step=1810: step=4.004s data=1.749s compute=3.972s [fwd=2.827s bwd=1.129s opt=0.003s] dispatch=3.327s [2026-10-18 17:26:37,684] loop.py:45 INFO TRAIN_STEP step=1810 loss=2.056979 Training: 10%|█ | 1810/18018 [3:56:23<34:44:37, 7.72s/step, loss=2.057, acc=0.586, acc_len=1.34, thru=14.9, I=20.3, T=32.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:37,791] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:42,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:26:43,823] loop.py:45 INFO TRAIN_STEP step=1811 loss=1.985688 Training: 10%|█ | 1811/18018 [3:56:29<32:36:40, 7.24s/step, loss=1.986, acc=0.595, acc_len=1.36, thru=18.8, I=24.7, T=35.1, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:26:43,818] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1810: forward=2593.7ms backward=1011.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:43,994] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:49,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:26:50,836] loop.py:45 INFO TRAIN_STEP step=1812 loss=2.065884 Training: 10%|█ | 1812/18018 [3:56:36<32:17:41, 7.17s/step, loss=2.066, acc=0.580, acc_len=1.33, thru=17.6, I=21.4, T=32.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:51,342] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:55,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:26:57,502] loop.py:45 INFO TRAIN_STEP step=1813 loss=1.981025 Training: 10%|█ | 1813/18018 [3:56:43<31:36:36, 7.02s/step, loss=1.981, acc=0.594, acc_len=1.37, thru=17.9, I=26.1, T=29.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:26:57,782] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:03,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:27:04,567] loop.py:45 INFO TRAIN_STEP step=1814 loss=2.095848 Training: 10%|█ | 1814/18018 [3:56:50<31:39:50, 7.03s/step, loss=2.096, acc=0.577, acc_len=1.32, thru=14.7, I=21.6, T=31.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:04,967] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:27:12,089] loop.py:380 INFO TIMING step=1815: step=3.973s data=1.828s compute=3.937s [fwd=2.809s bwd=1.114s opt=0.003s] dispatch=3.537s [2026-10-18 17:27:12,090] loop.py:45 INFO TRAIN_STEP step=1815 loss=2.066872 Training: 10%|█ | 1815/18018 [3:56:58<32:19:08, 7.18s/step, loss=2.067, acc=0.584, acc_len=1.34, thru=14.6, I=17.9, T=32.2, wait=3.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:23,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:27:24,140] loop.py:45 INFO TRAIN_STEP step=1816 loss=2.029459 Training: 10%|█ | 1816/18018 [3:57:10<38:53:41, 8.64s/step, loss=2.029, acc=0.588, acc_len=1.35, thru=12.8, I=11.2, T=29.4, wait=7.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:27:24,135] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1815: forward=3058.0ms backward=1234.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:24,216] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 17:27:31,927] loop.py:45 INFO TRAIN_STEP step=1817 loss=2.021573 Training: 10%|█ | 1817/18018 [3:57:18<37:43:57, 8.38s/step, loss=2.022, acc=0.588, acc_len=1.35, thru=13.3, I=14.5, T=31.7, wait=3.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:39,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:27:40,441] loop.py:45 INFO TRAIN_STEP step=1818 loss=1.993430 Training: 10%|█ | 1818/18018 [3:57:26<37:54:37, 8.42s/step, loss=1.993, acc=0.600, acc_len=1.38, thru=17.6, I=17.1, T=31.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:40,579] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:46,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:27:47,491] loop.py:45 INFO TRAIN_STEP step=1819 loss=2.003390 Training: 10%|█ | 1819/18018 [3:57:33<36:02:59, 8.01s/step, loss=2.003, acc=0.591, acc_len=1.36, thru=18.1, I=22.2, T=31.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:47,675] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:54,183] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:27:54,814] loop.py:380 INFO TIMING step=1820: step=4.204s data=1.921s compute=4.177s [fwd=2.967s bwd=1.195s opt=0.003s] dispatch=3.109s [2026-10-18 17:27:54,814] loop.py:45 INFO TRAIN_STEP step=1820 loss=2.085908 Training: 10%|█ | 1820/18018 [3:57:40<35:07:11, 7.81s/step, loss=2.086, acc=0.583, acc_len=1.33, thru=16.2, I=19.9, T=30.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:27:55,189] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:28:02,778] loop.py:45 INFO TRAIN_STEP step=1821 loss=2.105894 Training: 10%|█ | 1821/18018 [3:57:48<35:19:42, 7.85s/step, loss=2.106, acc=0.578, acc_len=1.32, thru=15.8, I=17.2, T=27.2, wait=3.2s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:28:02,773] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1820: forward=3333.3ms backward=1328.3ms [2026-10-18 17:28:11,595] loop.py:45 INFO TRAIN_STEP step=1822 loss=2.041587 Training: 10%|█ | 1822/18018 [3:57:57<36:37:52, 8.14s/step, loss=2.042, acc=0.589, acc_len=1.34, thru=14.6, I=14.9, T=30.7, wait=4.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:20,168] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:28:20,544] loop.py:45 INFO TRAIN_STEP step=1823 loss=1.883465 Training: 10%|█ | 1823/18018 [3:58:06<37:43:08, 8.38s/step, loss=1.883, acc=0.618, acc_len=1.43, thru=15.1, I=13.5, T=27.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:20,680] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-18 17:28:30,368] loop.py:45 INFO TRAIN_STEP step=1824 loss=2.054704 Training: 10%|█ | 1824/18018 [3:58:16<39:39:24, 8.82s/step, loss=2.055, acc=0.586, acc_len=1.34, thru=13.0, I=13.0, T=28.2, wait=5.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:34,002] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:28:36,334] loop.py:380 INFO TIMING step=1825: step=4.026s data=1.743s compute=3.998s [fwd=2.866s bwd=1.116s opt=0.003s] dispatch=1.931s [2026-10-18 17:28:36,334] loop.py:45 INFO TRAIN_STEP step=1825 loss=1.911853 Training: 10%|█ | 1825/18018 [3:58:22<35:48:36, 7.96s/step, loss=1.912, acc=0.612, acc_len=1.42, thru=19.1, I=21.1, T=31.8, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:36,520] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:44,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:28:44,712] loop.py:45 INFO TRAIN_STEP step=1826 loss=1.980296 Training: 10%|█ | 1826/18018 [3:58:30<36:21:59, 8.09s/step, loss=1.980, acc=0.598, acc_len=1.38, thru=15.6, I=14.9, T=30.9, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:28:44,707] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1825: forward=2868.0ms backward=1223.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:45,187] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:49,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:28:51,556] loop.py:45 INFO TRAIN_STEP step=1827 loss=1.906256 Training: 10%|█ | 1827/18018 [3:58:37<34:41:25, 7.71s/step, loss=1.906, acc=0.613, acc_len=1.42, thru=18.6, I=22.3, T=30.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:51,605] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:56,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:28:57,155] loop.py:45 INFO TRAIN_STEP step=1828 loss=2.038736 Training: 10%|█ | 1828/18018 [3:58:43<31:50:18, 7.08s/step, loss=2.039, acc=0.588, acc_len=1.34, thru=20.7, I=27.7, T=36.1, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:28:57,338] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:03,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:29:04,337] loop.py:45 INFO TRAIN_STEP step=1829 loss=1.840099 Training: 10%|█ | 1829/18018 [3:58:50<31:58:19, 7.11s/step, loss=1.840, acc=0.617, acc_len=1.43, thru=17.3, I=21.5, T=34.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:04,486] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:11,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:29:12,316] loop.py:380 INFO TIMING step=1830: step=3.949s data=1.697s compute=3.920s [fwd=2.822s bwd=1.083s opt=0.003s] dispatch=4.022s [2026-10-18 17:29:12,317] loop.py:45 INFO TRAIN_STEP step=1830 loss=1.900212 Training: 10%|█ | 1830/18018 [3:58:58<33:08:40, 7.37s/step, loss=1.900, acc=0.609, acc_len=1.42, thru=15.4, I=18.7, T=32.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:12,739] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:20,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:29:21,667] loop.py:45 INFO TRAIN_STEP step=1831 loss=1.905634 Training: 10%|█ | 1831/18018 [3:59:07<35:48:51, 7.97s/step, loss=1.906, acc=0.604, acc_len=1.40, thru=15.2, I=14.5, T=32.2, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:29:21,662] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1830: forward=2725.6ms backward=1177.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:21,917] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:28,198] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:29:28,846] loop.py:45 INFO TRAIN_STEP step=1832 loss=2.068915 Training: 10%|█ | 1832/18018 [3:59:14<34:44:57, 7.73s/step, loss=2.069, acc=0.581, acc_len=1.33, thru=16.9, I=22.0, T=27.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:29,202] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:34,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:29:36,565] loop.py:45 INFO TRAIN_STEP step=1833 loss=1.901275 Training: 10%|█ | 1833/18018 [3:59:22<34:44:15, 7.73s/step, loss=1.901, acc=0.608, acc_len=1.41, thru=15.6, I=18.4, T=34.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:36,712] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-18 17:29:44,136] loop.py:45 INFO TRAIN_STEP step=1834 loss=1.897654 Training: 10%|█ | 1834/18018 [3:59:30<34:31:15, 7.68s/step, loss=1.898, acc=0.612, acc_len=1.41, thru=14.9, I=18.1, T=30.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:44,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:44,596] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:51,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:29:51,544] loop.py:380 INFO TIMING step=1835: step=3.928s data=1.681s compute=3.904s [fwd=2.820s bwd=1.069s opt=0.003s] dispatch=3.472s [2026-10-18 17:29:51,544] loop.py:45 INFO TRAIN_STEP step=1835 loss=1.940541 Training: 10%|█ | 1835/18018 [3:59:37<34:09:17, 7.60s/step, loss=1.941, acc=0.605, acc_len=1.41, thru=17.3, I=17.7, T=32.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:51,849] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:57,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:29:58,125] loop.py:45 INFO TRAIN_STEP step=1836 loss=2.055471 Training: 10%|█ | 1836/18018 [3:59:44<32:46:48, 7.29s/step, loss=2.055, acc=0.590, acc_len=1.35, thru=19.0, I=21.8, T=33.2, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:29:58,120] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1835: forward=2623.0ms backward=1182.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:29:58,554] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:30:07,701] loop.py:45 INFO TRAIN_STEP step=1837 loss=2.083323 Training: 10%|█ | 1837/18018 [3:59:53<35:51:23, 7.98s/step, loss=2.083, acc=0.579, acc_len=1.32, thru=14.0, I=13.2, T=27.8, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:07,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:08,137] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:14,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:30:16,010] loop.py:45 INFO TRAIN_STEP step=1838 loss=2.132611 Training: 10%|█ | 1838/18018 [4:00:02<36:18:19, 8.08s/step, loss=2.133, acc=0.575, acc_len=1.30, thru=16.4, I=16.8, T=29.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:16,080] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:22,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:30:23,094] loop.py:45 INFO TRAIN_STEP step=1839 loss=1.957693 Training: 10%|█ | 1839/18018 [4:00:09<34:57:42, 7.78s/step, loss=1.958, acc=0.600, acc_len=1.38, thru=16.2, I=21.0, T=30.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:23,450] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:28,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:30:30,142] loop.py:380 INFO TIMING step=1840: step=3.741s data=1.620s compute=3.715s [fwd=2.694s bwd=1.006s opt=0.003s] dispatch=3.297s [2026-10-18 17:30:30,142] loop.py:45 INFO TRAIN_STEP step=1840 loss=1.965970 Training: 10%|█ | 1840/18018 [4:00:16<33:58:29, 7.56s/step, loss=1.966, acc=0.597, acc_len=1.38, thru=17.0, I=21.0, T=34.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:30,351] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:35,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:30:36,495] loop.py:45 INFO TRAIN_STEP step=1841 loss=1.917275 Training: 10%|█ | 1841/18018 [4:00:22<32:20:30, 7.20s/step, loss=1.917, acc=0.602, acc_len=1.39, thru=19.5, I=23.3, T=33.0, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:30:36,490] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1840: forward=2728.2ms backward=1106.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:36,900] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:43,845] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:30:43,942] loop.py:45 INFO TRAIN_STEP step=1842 loss=2.121244 Training: 10%|█ | 1842/18018 [4:00:30<32:40:41, 7.27s/step, loss=2.121, acc=0.579, acc_len=1.32, thru=17.5, I=18.9, T=33.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:44,348] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:48,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:30:50,068] loop.py:45 INFO TRAIN_STEP step=1843 loss=1.932190 Training: 10%|█ | 1843/18018 [4:00:36<31:07:56, 6.93s/step, loss=1.932, acc=0.603, acc_len=1.39, thru=19.9, I=20.8, T=33.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:50,450] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:56,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:30:57,824] loop.py:45 INFO TRAIN_STEP step=1844 loss=2.054113 Training: 10%|█ | 1844/18018 [4:00:43<32:14:38, 7.18s/step, loss=2.054, acc=0.586, acc_len=1.34, thru=16.1, I=21.7, T=31.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:30:58,190] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:31:04,869] loop.py:380 INFO TIMING step=1845: step=3.766s data=1.688s compute=3.733s [fwd=2.607s bwd=1.111s opt=0.003s] dispatch=3.270s [2026-10-18 17:31:04,870] loop.py:45 INFO TRAIN_STEP step=1845 loss=1.940414 Training: 10%|█ | 1845/18018 [4:00:50<32:04:14, 7.14s/step, loss=1.940, acc=0.597, acc_len=1.38, thru=14.4, I=21.2, T=34.0, wait=3.3s, pool=44, epoch=1/2][2026-10-18 17:31:12,661] loop.py:45 INFO TRAIN_STEP step=1846 loss=2.007863 Training: 10%|█ | 1846/18018 [4:00:58<32:56:36, 7.33s/step, loss=2.008, acc=0.589, acc_len=1.35, thru=16.4, I=17.5, T=31.7, wait=3.7s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:31:12,656] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1845: forward=2829.7ms backward=1152.3ms [2026-10-18 17:31:19,279] loop.py:45 INFO TRAIN_STEP step=1847 loss=1.991852 Training: 10%|█ | 1847/18018 [4:01:05<31:58:40, 7.12s/step, loss=1.992, acc=0.594, acc_len=1.37, thru=17.0, I=21.9, T=32.0, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:31:27,341] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:31:28,266] loop.py:45 INFO TRAIN_STEP step=1848 loss=1.936030 Training: 10%|█ | 1848/18018 [4:01:14<34:29:25, 7.68s/step, loss=1.936, acc=0.607, acc_len=1.40, thru=14.8, I=13.7, T=31.7, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:31:28,367] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:31:34,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:31:35,062] loop.py:45 INFO TRAIN_STEP step=1849 loss=2.043774 Training: 10%|█ | 1849/18018 [4:01:21<33:18:01, 7.41s/step, loss=2.044, acc=0.588, acc_len=1.35, thru=18.7, I=20.1, T=33.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:31:35,524] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:31:40,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:31:42,261] loop.py:380 INFO TIMING step=1850: step=4.130s data=1.907s compute=4.098s [fwd=2.926s bwd=1.157s opt=0.003s] dispatch=3.059s [2026-10-18 17:31:42,261] loop.py:45 INFO TRAIN_STEP step=1850 loss=1.896999 Training: 10%|█ | 1850/18018 [4:01:28<33:00:23, 7.35s/step, loss=1.897, acc=0.606, acc_len=1.41, thru=17.0, I=22.1, T=31.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:31:42,637] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:31:47,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:31:48,706] loop.py:45 INFO TRAIN_STEP step=1851 loss=1.968704 Training: 10%|█ | 1851/18018 [4:01:34<31:47:07, 7.08s/step, loss=1.969, acc=0.593, acc_len=1.37, thru=17.2, I=25.0, T=32.3, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:31:48,701] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1850: forward=2773.1ms backward=1143.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:31:49,114] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:31:55,361] loop.py:45 INFO TRAIN_STEP step=1852 loss=2.054522 Training: 10%|█ | 1852/18018 [4:01:41<31:12:53, 6.95s/step, loss=2.055, acc=0.581, acc_len=1.33, thru=17.8, I=20.1, T=35.6, wait=3.1s, pool=56, epoch=1/2][2026-10-18 17:32:04,184] loop.py:45 INFO TRAIN_STEP step=1853 loss=2.135310 Training: 10%|█ | 1853/18018 [4:01:50<33:44:10, 7.51s/step, loss=2.135, acc=0.571, acc_len=1.30, thru=14.9, I=15.0, T=34.1, wait=5.1s, pool=52, epoch=1/2][2026-10-18 17:32:14,168] loop.py:45 INFO TRAIN_STEP step=1854 loss=2.257260 Training: 10%|█ | 1854/18018 [4:02:00<37:03:34, 8.25s/step, loss=2.257, acc=0.561, acc_len=1.27, thru=12.8, I=13.7, T=30.2, wait=5.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:22,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:32:23,198] loop.py:380 INFO TIMING step=1855: step=4.592s data=2.076s compute=4.557s [fwd=3.139s bwd=1.402s opt=0.003s] dispatch=4.430s [2026-10-18 17:32:23,198] loop.py:45 INFO TRAIN_STEP step=1855 loss=2.031434 Training: 10%|█ | 1855/18018 [4:02:09<38:06:20, 8.49s/step, loss=2.031, acc=0.589, acc_len=1.36, thru=16.3, I=15.8, T=27.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:23,670] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:29,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:32:30,292] loop.py:45 INFO TRAIN_STEP step=1856 loss=1.900413 Training: 10%|█ | 1856/18018 [4:02:16<36:13:38, 8.07s/step, loss=1.900, acc=0.610, acc_len=1.41, thru=17.0, I=20.3, T=32.8, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:32:30,286] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1855: forward=2668.8ms backward=1185.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:30,437] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:35,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:32:38,099] loop.py:45 INFO TRAIN_STEP step=1857 loss=1.861440 Training: 10%|█ | 1857/18018 [4:02:24<35:52:19, 7.99s/step, loss=1.861, acc=0.615, acc_len=1.43, thru=15.5, I=21.9, T=26.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:38,469] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:46,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:32:46,538] loop.py:45 INFO TRAIN_STEP step=1858 loss=2.179693 Training: 10%|█ | 1858/18018 [4:02:32<36:28:21, 8.13s/step, loss=2.180, acc=0.562, acc_len=1.28, thru=15.9, I=16.3, T=30.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:46,831] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:51,136] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:32:52,773] loop.py:45 INFO TRAIN_STEP step=1859 loss=1.913519 Training: 10%|█ | 1859/18018 [4:02:38<33:55:20, 7.56s/step, loss=1.914, acc=0.605, acc_len=1.41, thru=19.5, I=22.3, T=32.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:53,149] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:32:59,979] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:33:00,248] loop.py:380 INFO TIMING step=1860: step=3.918s data=1.811s compute=3.884s [fwd=2.746s bwd=1.124s opt=0.003s] dispatch=3.549s [2026-10-18 17:33:00,248] loop.py:45 INFO TRAIN_STEP step=1860 loss=1.724486 Training: 10%|█ | 1860/18018 [4:02:46<33:48:35, 7.53s/step, loss=1.724, acc=0.634, acc_len=1.49, thru=15.4, I=20.1, T=32.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:00,481] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:07,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:33:10,243] loop.py:45 INFO TRAIN_STEP step=1861 loss=2.058096 Training: 10%|█ | 1861/18018 [4:02:56<37:07:24, 8.27s/step, loss=2.058, acc=0.586, acc_len=1.34, thru=13.1, I=15.8, T=27.1, wait=5.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:33:10,238] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1860: forward=3320.6ms backward=1349.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:10,380] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:17,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:33:18,458] loop.py:45 INFO TRAIN_STEP step=1862 loss=1.839046 Training: 10%|█ | 1862/18018 [4:03:04<37:02:49, 8.26s/step, loss=1.839, acc=0.618, acc_len=1.45, thru=15.9, I=18.1, T=27.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:18,585] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:23,757] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:33:24,323] loop.py:45 INFO TRAIN_STEP step=1863 loss=1.920202 Training: 10%|█ | 1863/18018 [4:03:10<33:49:27, 7.54s/step, loss=1.920, acc=0.604, acc_len=1.40, thru=18.6, I=25.2, T=32.7, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:24,763] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:31,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:33:31,926] loop.py:45 INFO TRAIN_STEP step=1864 loss=1.975684 Training: 10%|█ | 1864/18018 [4:03:18<33:54:45, 7.56s/step, loss=1.976, acc=0.597, acc_len=1.37, thru=16.1, I=18.3, T=30.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:32,241] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:33:40,370] loop.py:380 INFO TIMING step=1865: step=4.399s data=2.044s compute=4.368s [fwd=3.061s bwd=1.292s opt=0.003s] dispatch=4.035s [2026-10-18 17:33:40,370] loop.py:45 INFO TRAIN_STEP step=1865 loss=2.081048 Training: 10%|█ | 1865/18018 [4:03:26<35:06:12, 7.82s/step, loss=2.081, acc=0.582, acc_len=1.33, thru=12.5, I=13.0, T=29.1, wait=4.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:46,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:33:48,227] loop.py:45 INFO TRAIN_STEP step=1866 loss=2.028768 (TrainerActor pid=9310) [2026-10-18 17:33:48,223] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1865: forward=2863.1ms backward=1189.1ms Training: 10%|█ | 1866/18018 [4:03:34<35:08:48, 7.83s/step, loss=2.029, acc=0.590, acc_len=1.35, thru=17.2, I=17.7, T=31.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:33:48,505] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:33:56,458] loop.py:45 INFO TRAIN_STEP step=1867 loss=1.970318 Training: 10%|█ | 1867/18018 [4:03:42<35:40:37, 7.95s/step, loss=1.970, acc=0.597, acc_len=1.38, thru=12.6, I=14.9, T=30.5, wait=4.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:02,098] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:34:04,192] loop.py:45 INFO TRAIN_STEP step=1868 loss=2.014410 Training: 10%|█ | 1868/18018 [4:03:50<35:22:52, 7.89s/step, loss=2.014, acc=0.595, acc_len=1.37, thru=17.0, I=18.8, T=29.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:04,618] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:09,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:34:12,358] loop.py:45 INFO TRAIN_STEP step=1869 loss=1.956301 Training: 10%|█ | 1869/18018 [4:03:58<35:45:10, 7.97s/step, loss=1.956, acc=0.599, acc_len=1.38, thru=16.6, I=20.3, T=28.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:12,499] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:20,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:34:21,908] loop.py:380 INFO TIMING step=1870: step=5.184s data=2.509s compute=5.124s [fwd=3.531s bwd=1.579s opt=0.003s] dispatch=4.359s [2026-10-18 17:34:21,909] loop.py:45 INFO TRAIN_STEP step=1870 loss=1.953611 Training: 10%|█ | 1870/18018 [4:04:07<37:52:44, 8.44s/step, loss=1.954, acc=0.601, acc_len=1.39, thru=13.6, I=16.5, T=24.7, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:22,021] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:26,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:34:28,066] loop.py:45 INFO TRAIN_STEP step=1871 loss=2.002715 Training: 10%|█ | 1871/18018 [4:04:14<34:48:09, 7.76s/step, loss=2.003, acc=0.597, acc_len=1.37, thru=17.4, I=23.8, T=31.7, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:34:28,062] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1870: forward=2822.5ms backward=1171.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:28,220] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:34,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:34:34,655] loop.py:45 INFO TRAIN_STEP step=1872 loss=1.899346 Training: 10%|█ | 1872/18018 [4:04:20<33:13:19, 7.41s/step, loss=1.899, acc=0.609, acc_len=1.41, thru=17.0, I=22.6, T=30.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:35,099] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:40,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:34:41,572] loop.py:45 INFO TRAIN_STEP step=1873 loss=1.896883 Training: 10%|█ | 1873/18018 [4:04:27<32:33:46, 7.26s/step, loss=1.897, acc=0.604, acc_len=1.40, thru=17.3, I=20.1, T=32.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:41,962] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:34:50,233] loop.py:45 INFO TRAIN_STEP step=1874 loss=1.922518 Training: 10%|█ | 1874/18018 [4:04:36<34:26:25, 7.68s/step, loss=1.923, acc=0.605, acc_len=1.40, thru=13.4, I=14.8, T=30.3, wait=4.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:57,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:34:59,038] loop.py:380 INFO TIMING step=1875: step=4.242s data=1.969s compute=4.211s [fwd=3.008s bwd=1.189s opt=0.003s] dispatch=4.555s [2026-10-18 17:34:59,038] loop.py:45 INFO TRAIN_STEP step=1875 loss=1.888960 Training: 10%|█ | 1875/18018 [4:04:45<35:57:11, 8.02s/step, loss=1.889, acc=0.611, acc_len=1.42, thru=16.6, I=16.9, T=30.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:34:59,503] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:35:06,098] loop.py:45 INFO TRAIN_STEP step=1876 loss=2.006267 Training: 10%|█ | 1876/18018 [4:04:52<34:39:50, 7.73s/step, loss=2.006, acc=0.591, acc_len=1.35, thru=14.1, I=24.8, T=29.7, wait=2.7s, pool=28, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:35:06,093] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1875: forward=2882.1ms backward=1363.0ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 17:35:06,148] eagle_store.py:517 WARNING batch_get_buffer missing keys (k60027-a35286fd_hs, k60027-a35286fd_ids, k60027-a35286fd_lhs); sleeping 0.50s. [2026-10-18 17:35:19,147] loop.py:45 INFO TRAIN_STEP step=1877 loss=2.044660 Training: 10%|█ | 1877/18018 [4:05:05<41:49:07, 9.33s/step, loss=2.045, acc=0.597, acc_len=1.36, thru=16.1, I=10.0, T=27.4, wait=8.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:35:27,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 17:35:06,158] eagle_store.py:517 WARNING batch_get_buffer missing keys (k60026-a49e578f_hs, k60026-a49e578f_ids, k60026-a49e578f_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-18 17:35:28,531] loop.py:45 INFO TRAIN_STEP step=1878 loss=2.105963 Training: 10%|█ | 1878/18018 [4:05:14<41:53:16, 9.34s/step, loss=2.106, acc=0.579, acc_len=1.32, thru=15.7, I=15.7, T=30.3, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:35:28,655] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 17:35:35,330] loop.py:45 INFO TRAIN_STEP step=1879 loss=2.012292 Training: 10%|█ | 1879/18018 [4:05:21<38:27:43, 8.58s/step, loss=2.012, acc=0.590, acc_len=1.35, thru=16.4, I=19.5, T=33.4, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:35:40,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:35:42,658] loop.py:380 INFO TIMING step=1880: step=4.172s data=1.903s compute=4.140s [fwd=2.982s bwd=1.143s opt=0.003s] dispatch=3.150s [2026-10-18 17:35:42,658] loop.py:45 INFO TRAIN_STEP step=1880 loss=1.881910 Training: 10%|█ | 1880/18018 [4:05:28<36:46:53, 8.21s/step, loss=1.882, acc=0.607, acc_len=1.40, thru=18.1, I=19.8, T=30.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:35:42,758] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-18 17:35:48,636] loop.py:45 INFO TRAIN_STEP step=1881 loss=1.980483 Training: 10%|█ | 1881/18018 [4:05:34<33:47:01, 7.54s/step, loss=1.980, acc=0.595, acc_len=1.38, thru=19.1, I=24.4, T=34.7, wait=2.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:35:48,632] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1880: forward=2574.6ms backward=1073.1ms [2026-10-18 17:35:57,260] loop.py:45 INFO TRAIN_STEP step=1882 loss=2.049853 Training: 10%|█ | 1882/18018 [4:05:43<35:14:22, 7.86s/step, loss=2.050, acc=0.583, acc_len=1.33, thru=12.8, I=13.6, T=30.4, wait=4.4s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:07,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:36:07,864] loop.py:45 INFO TRAIN_STEP step=1883 loss=2.018745 Training: 10%|█ | 1883/18018 [4:05:53<38:55:50, 8.69s/step, loss=2.019, acc=0.592, acc_len=1.36, thru=15.4, I=14.7, T=25.5, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:07,905] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:14,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:36:15,602] loop.py:45 INFO TRAIN_STEP step=1884 loss=2.031705 Training: 10%|█ | 1884/18018 [4:06:01<37:39:02, 8.40s/step, loss=2.032, acc=0.591, acc_len=1.36, thru=15.3, I=16.2, T=28.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:15,850] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:20,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:36:22,525] loop.py:380 INFO TIMING step=1885: step=4.360s data=2.121s compute=4.319s [fwd=2.999s bwd=1.305s opt=0.003s] dispatch=2.552s [2026-10-18 17:36:22,525] loop.py:45 INFO TRAIN_STEP step=1885 loss=2.036823 Training: 10%|█ | 1885/18018 [4:06:08<35:39:41, 7.96s/step, loss=2.037, acc=0.590, acc_len=1.36, thru=16.2, I=22.8, T=29.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:22,935] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:27,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:36:28,775] loop.py:45 INFO TRAIN_STEP step=1886 loss=2.087346 Training: 10%|█ | 1886/18018 [4:06:14<33:21:50, 7.45s/step, loss=2.087, acc=0.581, acc_len=1.32, thru=18.4, I=25.7, T=34.9, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:36:28,771] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1885: forward=2619.9ms backward=1000.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:28,935] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:34,314] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:36:34,791] loop.py:45 INFO TRAIN_STEP step=1887 loss=1.919170 Training: 10%|█ | 1887/18018 [4:06:20<31:26:19, 7.02s/step, loss=1.919, acc=0.609, acc_len=1.41, thru=20.1, I=23.9, T=32.4, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:34,820] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 17:36:41,552] loop.py:45 INFO TRAIN_STEP step=1888 loss=2.040160 Training: 10%|█ | 1888/18018 [4:06:27<31:05:36, 6.94s/step, loss=2.040, acc=0.587, acc_len=1.35, thru=16.4, I=20.0, T=34.7, wait=3.1s, pool=44, epoch=1/2](TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 17:36:41,614] eagle_store.py:517 WARNING batch_get_buffer missing keys (k60414-b3e3e0c6_hs, k60414-b3e3e0c6_ids, k60414-b3e3e0c6_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:48,529] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-18 17:36:41,637] eagle_store.py:517 WARNING batch_get_buffer missing keys (k60413-b0316120_hs, k60413-b0316120_ids, k60413-b0316120_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-18 17:36:49,585] loop.py:45 INFO TRAIN_STEP step=1889 loss=1.860254 Training: 10%|█ | 1889/18018 [4:06:35<32:33:32, 7.27s/step, loss=1.860, acc=0.613, acc_len=1.42, thru=17.1, I=18.1, T=27.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:50,036] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:55,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:36:57,076] loop.py:380 INFO TIMING step=1890: step=4.691s data=2.272s compute=4.659s [fwd=3.296s bwd=1.348s opt=0.003s] dispatch=2.793s [2026-10-18 17:36:57,077] loop.py:45 INFO TRAIN_STEP step=1890 loss=1.975662 Training: 10%|█ | 1890/18018 [4:06:43<32:51:29, 7.33s/step, loss=1.976, acc=0.594, acc_len=1.37, thru=15.1, I=19.3, T=27.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:36:57,543] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:37:03,496] loop.py:45 INFO TRAIN_STEP step=1891 loss=2.082329 Training: 10%|█ | 1891/18018 [4:06:49<31:37:40, 7.06s/step, loss=2.082, acc=0.579, acc_len=1.32, thru=17.2, I=21.4, T=32.2, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:37:03,492] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1890: forward=2778.2ms backward=1145.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:03,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:03,992] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:10,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:37:12,163] loop.py:45 INFO TRAIN_STEP step=1892 loss=1.966357 Training: 11%|█ | 1892/18018 [4:06:58<33:47:16, 7.54s/step, loss=1.966, acc=0.597, acc_len=1.38, thru=14.6, I=17.1, T=26.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:12,333] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:17,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:37:19,984] loop.py:45 INFO TRAIN_STEP step=1893 loss=1.983527 Training: 11%|█ | 1893/18018 [4:07:06<34:09:25, 7.63s/step, loss=1.984, acc=0.596, acc_len=1.37, thru=13.7, I=21.2, T=27.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:20,152] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:27,027] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:37:27,405] loop.py:45 INFO TRAIN_STEP step=1894 loss=1.997004 Training: 11%|█ | 1894/18018 [4:07:13<33:52:55, 7.56s/step, loss=1.997, acc=0.592, acc_len=1.36, thru=14.2, I=18.6, T=26.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:27,544] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:35,979] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:37:37,104] loop.py:380 INFO TIMING step=1895: step=5.204s data=2.575s compute=5.172s [fwd=3.669s bwd=1.488s opt=0.003s] dispatch=4.485s [2026-10-18 17:37:37,105] loop.py:45 INFO TRAIN_STEP step=1895 loss=2.021080 Training: 11%|█ | 1895/18018 [4:07:23<36:44:52, 8.21s/step, loss=2.021, acc=0.590, acc_len=1.36, thru=13.4, I=15.2, T=24.6, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:37,489] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:43,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:37:45,443] loop.py:45 INFO TRAIN_STEP step=1896 loss=1.775489 Training: 11%|█ | 1896/18018 [4:07:31<36:55:30, 8.25s/step, loss=1.775, acc=0.629, acc_len=1.48, thru=13.7, I=18.5, T=25.3, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:37:45,438] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1895: forward=3775.6ms backward=1243.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:45,495] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 [2026-10-18 17:37:52,070] loop.py:45 INFO TRAIN_STEP step=1897 loss=1.888682 Training: 11%|█ | 1897/18018 [4:07:38<34:44:56, 7.76s/step, loss=1.889, acc=0.610, acc_len=1.42, thru=14.5, I=19.0, T=30.3, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:37:58,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:38:00,316] loop.py:45 INFO TRAIN_STEP step=1898 loss=1.966910 Training: 11%|█ | 1898/18018 [4:07:46<35:23:44, 7.90s/step, loss=1.967, acc=0.598, acc_len=1.37, thru=15.2, I=17.7, T=29.5, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:00,775] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:04,972] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:38:07,525] loop.py:45 INFO TRAIN_STEP step=1899 loss=2.083621 Training: 11%|█ | 1899/18018 [4:07:53<34:27:30, 7.70s/step, loss=2.084, acc=0.579, acc_len=1.32, thru=15.1, I=25.8, T=27.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:07,990] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:13,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:38:15,590] loop.py:380 INFO TIMING step=1900: step=5.104s data=2.715s compute=5.067s [fwd=3.712s bwd=1.340s opt=0.003s] dispatch=2.954s [2026-10-18 17:38:15,590] loop.py:45 INFO TRAIN_STEP step=1900 loss=2.007336 Training: 11%|█ | 1900/18018 [4:08:01<34:57:14, 7.81s/step, loss=2.007, acc=0.597, acc_len=1.38, thru=16.9, I=22.8, T=25.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:16,068] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:21,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:38:23,396] loop.py:45 INFO TRAIN_STEP step=1901 loss=2.108895 Training: 11%|█ | 1901/18018 [4:08:09<34:57:13, 7.81s/step, loss=2.109, acc=0.575, acc_len=1.31, thru=13.7, I=23.7, T=26.8, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:38:23,392] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1900: forward=3448.3ms backward=1268.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:23,467] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:31,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:38:31,704] loop.py:45 INFO TRAIN_STEP step=1902 loss=1.864935 Training: 11%|█ | 1902/18018 [4:08:17<35:37:14, 7.96s/step, loss=1.865, acc=0.608, acc_len=1.42, thru=15.6, I=16.6, T=29.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:31,984] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:37,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:38:39,287] loop.py:45 INFO TRAIN_STEP step=1903 loss=2.026999 Training: 11%|█ | 1903/18018 [4:08:25<35:07:04, 7.85s/step, loss=2.027, acc=0.587, acc_len=1.34, thru=16.3, I=18.1, T=27.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:39,320] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:43,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:38:46,197] loop.py:45 INFO TRAIN_STEP step=1904 loss=2.079659 Training: 11%|█ | 1904/18018 [4:08:32<33:51:38, 7.56s/step, loss=2.080, acc=0.582, acc_len=1.33, thru=15.5, I=21.3, T=30.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:46,459] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:54,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:38:55,364] loop.py:380 INFO TIMING step=1905: step=4.356s data=1.969s compute=4.326s [fwd=2.945s bwd=1.365s opt=0.003s] dispatch=4.801s [2026-10-18 17:38:55,364] loop.py:45 INFO TRAIN_STEP step=1905 loss=2.004640 Training: 11%|█ | 1905/18018 [4:08:41<36:00:31, 8.05s/step, loss=2.005, acc=0.593, acc_len=1.36, thru=14.4, I=15.7, T=29.4, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:38:55,647] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:39:02,412] loop.py:45 INFO TRAIN_STEP step=1906 loss=2.115659 (TrainerActor pid=9310) [2026-10-18 17:39:02,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1905: forward=2888.4ms backward=1359.1ms Training: 11%|█ | 1906/18018 [4:08:48<34:39:59, 7.75s/step, loss=2.116, acc=0.577, acc_len=1.31, thru=15.4, I=17.0, T=29.7, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:39:08,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:39:09,379] loop.py:45 INFO TRAIN_STEP step=1907 loss=1.889488 Training: 11%|█ | 1907/18018 [4:08:55<33:37:16, 7.51s/step, loss=1.889, acc=0.615, acc_len=1.42, thru=19.0, I=19.2, T=32.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:39:09,517] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:39:15,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:39:16,859] loop.py:45 INFO TRAIN_STEP step=1908 loss=1.889966 Training: 11%|█ | 1908/18018 [4:09:02<33:34:20, 7.50s/step, loss=1.890, acc=0.610, acc_len=1.41, thru=15.6, I=19.7, T=32.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:39:16,942] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 17:39:23,034] loop.py:45 INFO TRAIN_STEP step=1909 loss=2.029445 Training: 11%|█ | 1909/18018 [4:09:09<31:47:29, 7.10s/step, loss=2.029, acc=0.588, acc_len=1.35, thru=18.1, I=26.2, T=31.2, wait=2.1s, pool=56, epoch=1/2][2026-10-18 17:39:31,519] loop.py:380 INFO TIMING step=1910: step=4.278s data=1.909s compute=4.250s [fwd=3.027s bwd=1.208s opt=0.003s] dispatch=4.196s [2026-10-18 17:39:31,520] loop.py:45 INFO TRAIN_STEP step=1910 loss=1.963780 Training: 11%|█ | 1910/18018 [4:09:17<33:38:39, 7.52s/step, loss=1.964, acc=0.599, acc_len=1.38, thru=12.1, I=15.4, T=29.9, wait=4.2s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:39:39,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:39:39,331] loop.py:45 INFO TRAIN_STEP step=1911 loss=1.893194 Training: 11%|█ | 1911/18018 [4:09:25<34:01:45, 7.61s/step, loss=1.893, acc=0.610, acc_len=1.42, thru=17.8, I=16.0, T=34.3, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:39:39,327] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1910: forward=2621.3ms backward=1061.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:39:39,780] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 17:39:46,006] loop.py:45 INFO TRAIN_STEP step=1912 loss=2.016920 Training: 11%|█ | 1912/18018 [4:09:32<32:46:53, 7.33s/step, loss=2.017, acc=0.592, acc_len=1.36, thru=16.5, I=17.6, T=31.6, wait=2.6s, pool=48, epoch=1/2][2026-10-18 17:39:53,212] loop.py:45 INFO TRAIN_STEP step=1913 loss=1.923966 Training: 11%|█ | 1913/18018 [4:09:39<32:36:49, 7.29s/step, loss=1.924, acc=0.604, acc_len=1.40, thru=15.7, I=15.4, T=32.7, wait=3.3s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:00,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:40:00,611] loop.py:45 INFO TRAIN_STEP step=1914 loss=1.916897 Training: 11%|█ | 1914/18018 [4:09:46<32:45:31, 7.32s/step, loss=1.917, acc=0.610, acc_len=1.41, thru=21.0, I=17.2, T=36.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:00,718] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:06,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:40:07,470] loop.py:380 INFO TIMING step=1915: step=3.999s data=1.831s compute=3.963s [fwd=2.797s bwd=1.151s opt=0.003s] dispatch=2.852s [2026-10-18 17:40:07,470] loop.py:45 INFO TRAIN_STEP step=1915 loss=1.995997 Training: 11%|█ | 1915/18018 [4:09:53<32:08:08, 7.18s/step, loss=1.996, acc=0.590, acc_len=1.35, thru=18.4, I=21.3, T=32.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:07,779] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:13,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 17:40:14,365] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1915: forward=3095.4ms backward=1110.8ms [2026-10-18 17:40:14,369] loop.py:45 INFO TRAIN_STEP step=1916 loss=1.928768 Training: 11%|█ | 1916/18018 [4:10:00<31:44:54, 7.10s/step, loss=1.929, acc=0.597, acc_len=1.38, thru=17.3, I=23.2, T=30.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:14,562] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:40:21,240] loop.py:45 INFO TRAIN_STEP step=1917 loss=2.166778 Training: 11%|█ | 1917/18018 [4:10:07<31:26:23, 7.03s/step, loss=2.167, acc=0.565, acc_len=1.29, thru=15.5, I=19.0, T=35.2, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:28,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:40:28,657] loop.py:45 INFO TRAIN_STEP step=1918 loss=2.182243 Training: 11%|█ | 1918/18018 [4:10:14<31:57:35, 7.15s/step, loss=2.182, acc=0.568, acc_len=1.29, thru=17.6, I=18.1, T=33.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:28,686] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:35,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:40:36,438] loop.py:45 INFO TRAIN_STEP step=1919 loss=1.831175 Training: 11%|█ | 1919/18018 [4:10:22<32:48:43, 7.34s/step, loss=1.831, acc=0.615, acc_len=1.43, thru=16.1, I=19.1, T=31.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:36,630] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:42,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 17:40:43,531] loop.py:380 INFO TIMING step=1920: step=3.920s data=1.735s compute=3.890s [fwd=2.814s bwd=1.061s opt=0.003s] dispatch=3.162s [2026-10-18 17:40:43,532] loop.py:45 INFO TRAIN_STEP step=1920 loss=1.952108 Training: 11%|█ | 1920/18018 [4:10:29<32:28:47, 7.26s/step, loss=1.952, acc=0.597, acc_len=1.38, thru=18.1, I=22.2, T=32.7, wait=3.2s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:43,729] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:49,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:40:51,333] loop.py:45 INFO TRAIN_STEP step=1921 loss=1.946708 Training: 11%|█ | 1921/18018 [4:10:37<33:12:05, 7.43s/step, loss=1.947, acc=0.603, acc_len=1.39, thru=15.0, I=22.3, T=26.5, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:40:51,329] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1920: forward=3357.6ms backward=1419.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:51,406] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:55,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:40:56,886] loop.py:45 INFO TRAIN_STEP step=1922 loss=1.874829 Training: 11%|█ | 1922/18018 [4:10:42<30:41:10, 6.86s/step, loss=1.875, acc=0.608, acc_len=1.42, thru=18.6, I=26.6, T=33.2, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:40:57,159] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:41:03,706] loop.py:45 INFO TRAIN_STEP step=1923 loss=1.981364 Training: 11%|█ | 1923/18018 [4:10:49<30:37:47, 6.85s/step, loss=1.981, acc=0.588, acc_len=1.35, thru=16.6, I=21.2, T=31.9, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:10,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:41:11,680] loop.py:45 INFO TRAIN_STEP step=1924 loss=2.042769 Training: 11%|█ | 1924/18018 [4:10:57<32:07:48, 7.19s/step, loss=2.043, acc=0.591, acc_len=1.35, thru=16.9, I=19.5, T=27.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:12,181] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:16,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:41:18,379] loop.py:380 INFO TIMING step=1925: step=3.783s data=1.596s compute=3.754s [fwd=2.618s bwd=1.121s opt=0.003s] dispatch=2.907s [2026-10-18 17:41:18,380] loop.py:45 INFO TRAIN_STEP step=1925 loss=1.885692 Training: 11%|█ | 1925/18018 [4:11:04<31:28:29, 7.04s/step, loss=1.886, acc=0.609, acc_len=1.41, thru=15.5, I=21.1, T=33.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:18,813] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:24,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:41:25,788] loop.py:45 INFO TRAIN_STEP step=1926 loss=1.974804 Training: 11%|█ | 1926/18018 [4:11:11<31:58:06, 7.15s/step, loss=1.975, acc=0.598, acc_len=1.37, thru=15.7, I=21.7, T=30.9, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:41:25,783] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1925: forward=2870.9ms backward=1218.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:25,824] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:34,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:41:35,742] loop.py:45 INFO TRAIN_STEP step=1927 loss=2.091312 Training: 11%|█ | 1927/18018 [4:11:21<35:43:24, 7.99s/step, loss=2.091, acc=0.583, acc_len=1.33, thru=12.9, I=14.9, T=29.9, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:36,098] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:42,019] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:41:42,368] loop.py:45 INFO TRAIN_STEP step=1928 loss=1.771115 Training: 11%|█ | 1928/18018 [4:11:28<33:53:12, 7.58s/step, loss=1.771, acc=0.628, acc_len=1.47, thru=17.2, I=22.0, T=33.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:42,524] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:50,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:41:50,742] loop.py:45 INFO TRAIN_STEP step=1929 loss=2.189736 Training: 11%|█ | 1929/18018 [4:11:36<34:56:53, 7.82s/step, loss=2.190, acc=0.566, acc_len=1.28, thru=14.9, I=16.2, T=30.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:41:51,193] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:41:57,005] loop.py:380 INFO TIMING step=1930: step=3.671s data=1.644s compute=3.634s [fwd=2.584s bwd=1.035s opt=0.003s] dispatch=2.583s [2026-10-18 17:41:57,005] loop.py:45 INFO TRAIN_STEP step=1930 loss=2.123857 Training: 11%|█ | 1930/18018 [4:11:43<32:51:21, 7.35s/step, loss=2.124, acc=0.576, acc_len=1.31, thru=17.3, I=23.4, T=34.9, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:03,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:42:05,320] loop.py:45 INFO TRAIN_STEP step=1931 loss=1.884432 Training: 11%|█ | 1931/18018 [4:11:51<34:08:44, 7.64s/step, loss=1.884, acc=0.612, acc_len=1.41, thru=15.6, I=18.7, T=32.7, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:42:05,315] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1930: forward=2725.9ms backward=1121.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:05,348] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:10,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:42:12,331] loop.py:45 INFO TRAIN_STEP step=1932 loss=1.903320 Training: 11%|█ | 1932/18018 [4:11:58<33:18:04, 7.45s/step, loss=1.903, acc=0.604, acc_len=1.40, thru=15.9, I=24.6, T=30.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:12,594] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:18,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:42:18,399] loop.py:45 INFO TRAIN_STEP step=1933 loss=2.094174 Training: 11%|█ | 1933/18018 [4:12:04<31:26:35, 7.04s/step, loss=2.094, acc=0.578, acc_len=1.33, thru=18.5, I=26.0, T=33.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:18,603] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:27,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:42:27,928] loop.py:45 INFO TRAIN_STEP step=1934 loss=1.926146 Training: 11%|█ | 1934/18018 [4:12:14<34:46:50, 7.78s/step, loss=1.926, acc=0.606, acc_len=1.40, thru=13.4, I=14.1, T=33.6, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:28,106] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:32,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:42:33,957] loop.py:380 INFO TIMING step=1935: step=3.847s data=1.617s compute=3.817s [fwd=2.553s bwd=1.248s opt=0.003s] dispatch=2.171s [2026-10-18 17:42:33,957] loop.py:45 INFO TRAIN_STEP step=1935 loss=1.803391 Training: 11%|█ | 1935/18018 [4:12:20<32:25:33, 7.26s/step, loss=1.803, acc=0.623, acc_len=1.46, thru=19.8, I=24.9, T=33.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:34,059] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 17:42:41,193] loop.py:45 INFO TRAIN_STEP step=1936 loss=1.830812 (TrainerActor pid=9310) [2026-10-18 17:42:41,188] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1935: forward=3056.5ms backward=1267.4ms Training: 11%|█ | 1936/18018 [4:12:27<32:23:38, 7.25s/step, loss=1.831, acc=0.619, acc_len=1.44, thru=15.3, I=18.8, T=29.3, wait=2.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:47,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:42:48,296] loop.py:45 INFO TRAIN_STEP step=1937 loss=2.142356 Training: 11%|█ | 1937/18018 [4:12:34<32:11:22, 7.21s/step, loss=2.142, acc=0.570, acc_len=1.29, thru=18.7, I=16.6, T=34.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:42:48,801] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:42:55,235] loop.py:45 INFO TRAIN_STEP step=1938 loss=2.058149 Training: 11%|█ | 1938/18018 [4:12:41<31:49:56, 7.13s/step, loss=2.058, acc=0.587, acc_len=1.34, thru=18.1, I=20.7, T=31.9, wait=2.9s, pool=56, epoch=1/2][2026-10-18 17:43:02,968] loop.py:45 INFO TRAIN_STEP step=1939 loss=2.046638 Training: 11%|█ | 1939/18018 [4:12:49<32:38:35, 7.31s/step, loss=2.047, acc=0.582, acc_len=1.33, thru=16.2, I=16.0, T=31.3, wait=3.6s, pool=56, epoch=1/2][2026-10-18 17:43:11,372] loop.py:380 INFO TIMING step=1940: step=4.351s data=2.032s compute=4.317s [fwd=3.035s bwd=1.267s opt=0.003s] dispatch=4.044s [2026-10-18 17:43:11,372] loop.py:45 INFO TRAIN_STEP step=1940 loss=1.918474 Training: 11%|█ | 1940/18018 [4:12:57<34:06:38, 7.64s/step, loss=1.918, acc=0.608, acc_len=1.41, thru=14.8, I=14.7, T=29.4, wait=4.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:16,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:43:18,783] loop.py:45 INFO TRAIN_STEP step=1941 loss=1.974875 Training: 11%|█ | 1941/18018 [4:13:04<33:48:06, 7.57s/step, loss=1.975, acc=0.596, acc_len=1.37, thru=16.7, I=20.3, T=29.0, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:43:18,778] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1940: forward=3086.6ms backward=1262.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:19,147] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:24,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:43:26,931] loop.py:45 INFO TRAIN_STEP step=1942 loss=1.765581 Training: 11%|█ | 1942/18018 [4:13:13<34:34:36, 7.74s/step, loss=1.766, acc=0.624, acc_len=1.47, thru=16.4, I=22.0, T=28.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:27,306] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:33,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:43:34,111] loop.py:45 INFO TRAIN_STEP step=1943 loss=2.047654 Training: 11%|█ | 1943/18018 [4:13:20<33:49:04, 7.57s/step, loss=2.048, acc=0.587, acc_len=1.34, thru=15.8, I=20.0, T=29.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:34,545] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:41,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:43:42,127] loop.py:45 INFO TRAIN_STEP step=1944 loss=1.931185 Training: 11%|█ | 1944/18018 [4:13:28<34:24:32, 7.71s/step, loss=1.931, acc=0.605, acc_len=1.40, thru=16.8, I=17.7, T=32.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:42,487] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:46,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:43:48,402] loop.py:380 INFO TIMING step=1945: step=3.959s data=1.806s compute=3.920s [fwd=2.805s bwd=1.101s opt=0.003s] dispatch=2.308s [2026-10-18 17:43:48,402] loop.py:45 INFO TRAIN_STEP step=1945 loss=1.856998 Training: 11%|█ | 1945/18018 [4:13:34<32:29:30, 7.28s/step, loss=1.857, acc=0.616, acc_len=1.43, thru=19.5, I=22.8, T=32.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:48,642] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:52,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:43:55,614] loop.py:45 INFO TRAIN_STEP step=1946 loss=1.822959 Training: 11%|█ | 1946/18018 [4:13:41<32:24:05, 7.26s/step, loss=1.823, acc=0.619, acc_len=1.45, thru=16.0, I=27.5, T=28.6, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:43:55,609] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1945: forward=3074.0ms backward=1349.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:43:55,807] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:01,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:44:02,281] loop.py:45 INFO TRAIN_STEP step=1947 loss=2.046756 Training: 11%|█ | 1947/18018 [4:13:48<31:36:23, 7.08s/step, loss=2.047, acc=0.588, acc_len=1.34, thru=14.9, I=23.5, T=35.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:02,659] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:09,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:44:09,672] loop.py:45 INFO TRAIN_STEP step=1948 loss=1.965630 Training: 11%|█ | 1948/18018 [4:13:55<32:01:23, 7.17s/step, loss=1.966, acc=0.604, acc_len=1.40, thru=16.2, I=20.2, T=29.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:10,091] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:44:19,537] loop.py:45 INFO TRAIN_STEP step=1949 loss=1.914597 Training: 11%|█ | 1949/18018 [4:14:05<35:37:16, 7.98s/step, loss=1.915, acc=0.607, acc_len=1.40, thru=13.1, I=12.8, T=27.4, wait=5.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:29,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:44:29,438] loop.py:380 INFO TIMING step=1950: step=4.486s data=2.062s compute=4.440s [fwd=3.118s bwd=1.307s opt=0.003s] dispatch=5.407s [2026-10-18 17:44:29,438] loop.py:45 INFO TRAIN_STEP step=1950 loss=2.060985 Training: 11%|█ | 1950/18018 [4:14:15<38:11:35, 8.56s/step, loss=2.061, acc=0.585, acc_len=1.34, thru=13.7, I=14.4, T=28.5, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:29,699] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:35,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:44:35,853] loop.py:45 INFO TRAIN_STEP step=1951 loss=1.879993 Training: 11%|█ | 1951/18018 [4:14:21<35:19:27, 7.91s/step, loss=1.880, acc=0.613, acc_len=1.42, thru=18.5, I=18.7, T=30.1, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:44:35,848] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1950: forward=2974.2ms backward=1232.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:35,892] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:41,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:44:42,368] loop.py:45 INFO TRAIN_STEP step=1952 loss=2.103676 Training: 11%|█ | 1952/18018 [4:14:28<33:26:42, 7.49s/step, loss=2.104, acc=0.577, acc_len=1.31, thru=18.9, I=21.0, T=32.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:42,813] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:49,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:44:50,300] loop.py:45 INFO TRAIN_STEP step=1953 loss=1.899896 Training: 11%|█ | 1953/18018 [4:14:36<34:01:54, 7.63s/step, loss=1.900, acc=0.609, acc_len=1.42, thru=15.8, I=21.3, T=33.1, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:44:50,687] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:44:57,906] loop.py:45 INFO TRAIN_STEP step=1954 loss=1.924917 Training: 11%|█ | 1954/18018 [4:14:43<33:59:58, 7.62s/step, loss=1.925, acc=0.601, acc_len=1.39, thru=14.5, I=16.9, T=31.1, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:07,027] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:45:07,370] loop.py:380 INFO TIMING step=1955: step=4.622s data=2.213s compute=4.593s [fwd=3.280s bwd=1.298s opt=0.003s] dispatch=4.833s [2026-10-18 17:45:07,370] loop.py:45 INFO TRAIN_STEP step=1955 loss=2.039755 Training: 11%|█ | 1955/18018 [4:14:53<36:28:04, 8.17s/step, loss=2.040, acc=0.589, acc_len=1.35, thru=15.6, I=14.5, T=27.7, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:07,528] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:12,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:45:14,672] loop.py:45 INFO TRAIN_STEP step=1956 loss=2.002627 Training: 11%|█ | 1956/18018 [4:15:00<35:18:01, 7.91s/step, loss=2.003, acc=0.593, acc_len=1.37, thru=17.2, I=18.4, T=30.5, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:45:14,667] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1955: forward=2838.6ms backward=1281.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:14,974] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:45:21,713] loop.py:45 INFO TRAIN_STEP step=1957 loss=1.988455 Training: 11%|█ | 1957/18018 [4:15:07<34:07:50, 7.65s/step, loss=1.988, acc=0.595, acc_len=1.37, thru=16.9, I=19.7, T=30.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:21,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:22,166] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:29,736] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:45:30,412] loop.py:45 INFO TRAIN_STEP step=1958 loss=2.012749 Training: 11%|█ | 1958/18018 [4:15:16<35:31:56, 7.96s/step, loss=2.013, acc=0.595, acc_len=1.37, thru=15.2, I=16.0, T=32.8, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:30,740] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:36,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:45:39,081] loop.py:45 INFO TRAIN_STEP step=1959 loss=1.925856 Training: 11%|█ | 1959/18018 [4:15:25<36:28:14, 8.18s/step, loss=1.926, acc=0.602, acc_len=1.40, thru=14.8, I=20.3, T=26.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:39,440] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:45,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:45:46,295] loop.py:380 INFO TIMING step=1960: step=4.623s data=2.297s compute=4.595s [fwd=3.224s bwd=1.356s opt=0.003s] dispatch=2.584s [2026-10-18 17:45:46,296] loop.py:45 INFO TRAIN_STEP step=1960 loss=1.917766 Training: 11%|█ | 1960/18018 [4:15:32<35:11:02, 7.89s/step, loss=1.918, acc=0.606, acc_len=1.40, thru=14.1, I=25.8, T=27.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:46,645] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:51,601] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 17:45:53,433] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1960: forward=2997.3ms backward=1433.4ms [2026-10-18 17:45:53,439] loop.py:45 INFO TRAIN_STEP step=1961 loss=1.976634 Training: 11%|█ | 1961/18018 [4:15:39<34:11:07, 7.66s/step, loss=1.977, acc=0.592, acc_len=1.37, thru=16.2, I=19.1, T=28.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:53,616] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:45:59,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:46:00,533] loop.py:45 INFO TRAIN_STEP step=1962 loss=2.053200 Training: 11%|█ | 1962/18018 [4:15:46<33:25:20, 7.49s/step, loss=2.053, acc=0.584, acc_len=1.34, thru=15.9, I=22.9, T=31.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:00,725] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:46:07,034] loop.py:45 INFO TRAIN_STEP step=1963 loss=2.021502 Training: 11%|█ | 1963/18018 [4:15:53<32:05:30, 7.20s/step, loss=2.022, acc=0.592, acc_len=1.35, thru=17.8, I=21.5, T=32.6, wait=2.6s, pool=56, epoch=1/2][2026-10-18 17:46:13,683] loop.py:45 INFO TRAIN_STEP step=1964 loss=1.944824 Training: 11%|█ | 1964/18018 [4:15:59<31:21:34, 7.03s/step, loss=1.945, acc=0.599, acc_len=1.38, thru=14.9, I=16.8, T=35.4, wait=3.0s, pool=20, epoch=1/2](TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 17:46:14,219] eagle_store.py:517 WARNING batch_get_buffer missing keys (k62852-8c99cd0d_hs, k62852-8c99cd0d_ids, k62852-8c99cd0d_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:21,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-18 17:46:14,226] eagle_store.py:517 WARNING batch_get_buffer missing keys (k62853-951c89a6_hs, k62853-951c89a6_ids, k62853-951c89a6_lhs); sleeping 0.50s. [2026-10-18 17:46:23,200] loop.py:380 INFO TIMING step=1965: step=4.548s data=2.078s compute=4.515s [fwd=3.263s bwd=1.237s opt=0.003s] dispatch=4.957s [2026-10-18 17:46:23,201] loop.py:45 INFO TRAIN_STEP step=1965 loss=2.173663 Training: 11%|█ | 1965/18018 [4:16:09<34:40:54, 7.78s/step, loss=2.174, acc=0.572, acc_len=1.29, thru=18.1, I=16.8, T=28.1, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:23,395] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:28,794] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:46:30,611] loop.py:45 INFO TRAIN_STEP step=1966 loss=1.907217 Training: 11%|█ | 1966/18018 [4:16:16<34:11:17, 7.67s/step, loss=1.907, acc=0.608, acc_len=1.40, thru=14.8, I=21.8, T=30.2, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:46:30,606] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1965: forward=2988.2ms backward=1193.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:30,807] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:36,314] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:46:37,701] loop.py:45 INFO TRAIN_STEP step=1967 loss=1.772961 Training: 11%|█ | 1967/18018 [4:16:23<33:24:43, 7.49s/step, loss=1.773, acc=0.628, acc_len=1.47, thru=15.8, I=23.4, T=32.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:37,829] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:42,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:46:44,404] loop.py:45 INFO TRAIN_STEP step=1968 loss=1.803573 Training: 11%|█ | 1968/18018 [4:16:30<32:21:17, 7.26s/step, loss=1.804, acc=0.618, acc_len=1.44, thru=18.1, I=25.3, T=28.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:44,576] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:46:50,858] loop.py:45 INFO TRAIN_STEP step=1969 loss=2.047770 Training: 11%|█ | 1969/18018 [4:16:36<31:16:48, 7.02s/step, loss=2.048, acc=0.586, acc_len=1.34, thru=16.5, I=23.5, T=33.0, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:58,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:46:58,625] loop.py:380 INFO TIMING step=1970: step=3.736s data=1.677s compute=3.690s [fwd=2.584s bwd=1.090s opt=0.003s] dispatch=4.019s [2026-10-18 17:46:58,625] loop.py:45 INFO TRAIN_STEP step=1970 loss=1.938035 Training: 11%|█ | 1970/18018 [4:16:44<32:16:38, 7.24s/step, loss=1.938, acc=0.602, acc_len=1.39, thru=16.4, I=16.7, T=34.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:46:58,655] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:47:03,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:47:05,333] loop.py:45 INFO TRAIN_STEP step=1971 loss=1.989461 Training: 11%|█ | 1971/18018 [4:16:51<31:33:50, 7.08s/step, loss=1.989, acc=0.593, acc_len=1.37, thru=17.9, I=23.9, T=28.1, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:47:05,328] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1970: forward=3261.9ms backward=1244.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:47:05,704] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 17:47:13,167] loop.py:45 INFO TRAIN_STEP step=1972 loss=2.115815 Training: 11%|█ | 1972/18018 [4:16:59<32:34:02, 7.31s/step, loss=2.116, acc=0.576, acc_len=1.32, thru=12.4, I=17.9, T=30.2, wait=3.6s, pool=56, epoch=1/2][2026-10-18 17:47:20,020] loop.py:45 INFO TRAIN_STEP step=1973 loss=1.842889 Training: 11%|█ | 1973/18018 [4:17:06<31:57:39, 7.17s/step, loss=1.843, acc=0.616, acc_len=1.43, thru=16.8, I=15.8, T=32.2, wait=2.9s, pool=48, epoch=1/2][2026-10-18 17:47:30,745] loop.py:45 INFO TRAIN_STEP step=1974 loss=1.885011 Training: 11%|█ | 1974/18018 [4:17:16<36:42:24, 8.24s/step, loss=1.885, acc=0.614, acc_len=1.43, thru=12.9, I=13.5, T=32.0, wait=6.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:47:36,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:47:38,367] loop.py:380 INFO TIMING step=1975: step=3.863s data=1.673s compute=3.832s [fwd=2.719s bwd=1.098s opt=0.003s] dispatch=3.752s [2026-10-18 17:47:38,368] loop.py:45 INFO TRAIN_STEP step=1975 loss=2.021601 Training: 11%|█ | 1975/18018 [4:17:24<35:53:18, 8.05s/step, loss=2.022, acc=0.589, acc_len=1.35, thru=17.5, I=18.4, T=33.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:47:38,403] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:47:44,750] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:47:46,196] loop.py:45 INFO TRAIN_STEP step=1976 loss=2.025440 Training: 11%|█ | 1976/18018 [4:17:32<35:35:09, 7.99s/step, loss=2.025, acc=0.589, acc_len=1.35, thru=14.9, I=20.0, T=26.6, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:47:46,191] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1975: forward=3305.7ms backward=1464.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:47:46,263] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:47:52,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:47:53,835] loop.py:45 INFO TRAIN_STEP step=1977 loss=2.196581 Training: 11%|█ | 1977/18018 [4:17:39<35:06:58, 7.88s/step, loss=2.197, acc=0.568, acc_len=1.28, thru=14.9, I=19.2, T=29.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:47:53,882] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-18 17:48:00,341] loop.py:45 INFO TRAIN_STEP step=1978 loss=2.010689 Training: 11%|█ | 1978/18018 [4:17:46<33:16:33, 7.47s/step, loss=2.011, acc=0.592, acc_len=1.36, thru=17.6, I=19.7, T=31.6, wait=2.4s, pool=56, epoch=1/2][2026-10-18 17:48:08,632] loop.py:45 INFO TRAIN_STEP step=1979 loss=2.094369 Training: 11%|█ | 1979/18018 [4:17:54<34:22:51, 7.72s/step, loss=2.094, acc=0.577, acc_len=1.32, thru=16.1, I=15.6, T=31.3, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:16,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:48:18,103] loop.py:380 INFO TIMING step=1980: step=4.690s data=2.201s compute=4.629s [fwd=3.243s bwd=1.370s opt=0.003s] dispatch=4.766s [2026-10-18 17:48:18,103] loop.py:45 INFO TRAIN_STEP step=1980 loss=2.151151 Training: 11%|█ | 1980/18018 [4:18:04<36:42:59, 8.24s/step, loss=2.151, acc=0.574, acc_len=1.30, thru=14.0, I=14.9, T=27.3, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:18,376] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:23,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:48:25,864] loop.py:45 INFO TRAIN_STEP step=1981 loss=1.876388 Training: 11%|█ | 1981/18018 [4:18:11<36:04:19, 8.10s/step, loss=1.876, acc=0.609, acc_len=1.41, thru=14.4, I=20.2, T=26.9, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:48:25,859] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1980: forward=3267.7ms backward=1443.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:26,152] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:32,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:48:33,157] loop.py:45 INFO TRAIN_STEP step=1982 loss=2.011574 Training: 11%|█ | 1982/18018 [4:18:19<34:59:43, 7.86s/step, loss=2.012, acc=0.591, acc_len=1.35, thru=14.1, I=21.4, T=31.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:33,299] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:40,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:48:41,170] loop.py:45 INFO TRAIN_STEP step=1983 loss=1.748733 Training: 11%|█ | 1983/18018 [4:18:27<35:12:07, 7.90s/step, loss=1.749, acc=0.631, acc_len=1.48, thru=16.1, I=18.8, T=31.5, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:41,675] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:48:48,804] loop.py:45 INFO TRAIN_STEP step=1984 loss=1.895067 Training: 11%|█ | 1984/18018 [4:18:34<34:50:27, 7.82s/step, loss=1.895, acc=0.612, acc_len=1.42, thru=16.4, I=17.8, T=31.9, wait=3.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:54,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:48:56,274] loop.py:380 INFO TIMING step=1985: step=3.977s data=1.512s compute=3.944s [fwd=2.667s bwd=1.263s opt=0.003s] dispatch=3.484s [2026-10-18 17:48:56,274] loop.py:45 INFO TRAIN_STEP step=1985 loss=1.929311 Training: 11%|█ | 1985/18018 [4:18:42<34:22:07, 7.72s/step, loss=1.929, acc=0.601, acc_len=1.39, thru=18.5, I=18.7, T=32.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:48:56,754] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:03,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:49:04,555] loop.py:45 INFO TRAIN_STEP step=1986 loss=1.885629 Training: 11%|█ | 1986/18018 [4:18:50<35:07:15, 7.89s/step, loss=1.886, acc=0.609, acc_len=1.41, thru=13.4, I=17.8, T=33.4, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:49:04,550] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1985: forward=2542.3ms backward=1156.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:04,778] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:11,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:49:11,876] loop.py:45 INFO TRAIN_STEP step=1987 loss=1.988553 Training: 11%|█ | 1987/18018 [4:18:57<34:21:40, 7.72s/step, loss=1.989, acc=0.594, acc_len=1.36, thru=16.8, I=20.5, T=30.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:12,091] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:17,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:49:19,279] loop.py:45 INFO TRAIN_STEP step=1988 loss=2.004569 Training: 11%|█ | 1988/18018 [4:19:05<33:56:20, 7.62s/step, loss=2.005, acc=0.597, acc_len=1.37, thru=17.8, I=19.6, T=27.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:19,529] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:24,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:49:26,360] loop.py:45 INFO TRAIN_STEP step=1989 loss=1.964226 Training: 11%|█ | 1989/18018 [4:19:12<33:13:01, 7.46s/step, loss=1.964, acc=0.594, acc_len=1.37, thru=16.1, I=22.1, T=30.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:26,402] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:31,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:49:32,745] loop.py:380 INFO TIMING step=1990: step=3.681s data=1.606s compute=3.643s [fwd=2.635s bwd=0.994s opt=0.003s] dispatch=2.692s [2026-10-18 17:49:32,745] loop.py:45 INFO TRAIN_STEP step=1990 loss=1.952074 Training: 11%|█ | 1990/18018 [4:19:18<31:46:35, 7.14s/step, loss=1.952, acc=0.599, acc_len=1.39, thru=18.3, I=24.9, T=34.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:32,842] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:39,457] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:49:39,970] loop.py:45 INFO TRAIN_STEP step=1991 loss=1.882668 Training: 11%|█ | 1991/18018 [4:19:26<31:53:35, 7.16s/step, loss=1.883, acc=0.609, acc_len=1.41, thru=17.3, I=21.7, T=32.1, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:49:39,965] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1990: forward=2804.8ms backward=1141.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:49:40,462] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:49:47,699] loop.py:45 INFO TRAIN_STEP step=1992 loss=1.896627 Training: 11%|█ | 1992/18018 [4:19:33<32:38:40, 7.33s/step, loss=1.897, acc=0.612, acc_len=1.42, thru=16.6, I=18.4, T=29.4, wait=3.4s, pool=60, epoch=1/2][2026-10-18 17:49:57,031] loop.py:45 INFO TRAIN_STEP step=1993 loss=1.918017 Training: 11%|█ | 1993/18018 [4:19:43<35:18:53, 7.93s/step, loss=1.918, acc=0.609, acc_len=1.41, thru=14.0, I=14.1, T=30.1, wait=5.1s, pool=60, epoch=1/2][2026-10-18 17:50:05,226] loop.py:45 INFO TRAIN_STEP step=1994 loss=2.024544 Training: 11%|█ | 1994/18018 [4:19:51<35:39:28, 8.01s/step, loss=2.025, acc=0.589, acc_len=1.36, thru=15.6, I=14.9, T=33.5, wait=4.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:50:10,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:50:11,862] loop.py:380 INFO TIMING step=1995: step=3.802s data=1.549s compute=3.774s [fwd=2.693s bwd=1.066s opt=0.003s] dispatch=2.826s [2026-10-18 17:50:11,863] loop.py:45 INFO TRAIN_STEP step=1995 loss=1.760170 Training: 11%|█ | 1995/18018 [4:19:57<33:49:24, 7.60s/step, loss=1.760, acc=0.628, acc_len=1.48, thru=19.2, I=21.3, T=33.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:50:12,220] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:50:19,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:50:19,437] loop.py:45 INFO TRAIN_STEP step=1996 loss=1.953329 Training: 11%|█ | 1996/18018 [4:20:05<33:47:13, 7.59s/step, loss=1.953, acc=0.600, acc_len=1.38, thru=15.2, I=19.0, T=28.4, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:50:19,433] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=1995: forward=3086.1ms backward=1383.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:50:19,765] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:50:27,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:50:30,418] loop.py:45 INFO TRAIN_STEP step=1997 loss=1.989998 Training: 11%|█ | 1997/18018 [4:20:16<38:18:25, 8.61s/step, loss=1.990, acc=0.597, acc_len=1.37, thru=11.0, I=15.0, T=24.6, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:50:30,727] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:50:35,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:50:37,141] loop.py:45 INFO TRAIN_STEP step=1998 loss=1.905253 Training: 11%|█ | 1998/18018 [4:20:23<35:47:35, 8.04s/step, loss=1.905, acc=0.608, acc_len=1.40, thru=14.9, I=24.6, T=30.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:50:37,464] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:50:45,520] loop.py:45 INFO TRAIN_STEP step=1999 loss=2.064660 Training: 11%|█ | 1999/18018 [4:20:31<36:14:12, 8.14s/step, loss=2.065, acc=0.582, acc_len=1.33, thru=11.2, I=14.9, T=32.5, wait=4.4s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:50:55,141] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:50:56,550] loop.py:380 INFO TIMING step=2000: step=4.385s data=1.948s compute=4.333s [fwd=3.070s bwd=1.248s opt=0.003s] dispatch=6.636s [2026-10-18 17:50:56,551] loop.py:45 INFO TRAIN_STEP step=2000 loss=1.985184 Training: 11%|█ | 2000/18018 [4:20:42<40:05:12, 9.01s/step, loss=1.985, acc=0.595, acc_len=1.37, thru=14.7, I=15.6, T=29.2, wait=6.6s, pool=64, epoch=1/2][2026-10-18 17:50:56,553] loop.py:419 INFO Saving checkpoint at step 2000... (TrainerActor pid=9310) [2026-10-18 17:50:59,941] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0002001 (TrainerActor pid=9310) [2026-10-18 17:50:59,952] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0001251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:00,179] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:01,187] inference_manager.py:408 INFO Pool capacity available after 6.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:08,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 17:50:59,942] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0002001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 17:51:01,062] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0001251 [2026-10-18 17:51:09,018] loop.py:45 INFO TRAIN_STEP step=2001 loss=1.851875 Training: 11%|█ | 2001/18018 [4:20:55<44:42:03, 10.05s/step, loss=1.852, acc=0.619, acc_len=1.44, thru=16.4, I=18.4, T=29.6, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:51:09,014] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2000: forward=2964.0ms backward=1309.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:09,238] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:15,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:51:16,169] loop.py:45 INFO TRAIN_STEP step=2002 loss=1.979466 Training: 11%|█ | 2002/18018 [4:21:02<40:50:07, 9.18s/step, loss=1.979, acc=0.598, acc_len=1.38, thru=17.4, I=20.6, T=34.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:16,587] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:23,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:51:24,355] loop.py:45 INFO TRAIN_STEP step=2003 loss=1.990102 Training: 11%|█ | 2003/18018 [4:21:10<39:30:21, 8.88s/step, loss=1.990, acc=0.595, acc_len=1.36, thru=14.3, I=17.2, T=33.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:24,758] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:51:32,172] loop.py:45 INFO TRAIN_STEP step=2004 loss=1.994242 Training: 11%|█ | 2004/18018 [4:21:18<38:05:02, 8.56s/step, loss=1.994, acc=0.594, acc_len=1.36, thru=15.1, I=18.3, T=28.1, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:38,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:51:40,062] loop.py:380 INFO TIMING step=2005: step=4.554s data=2.360s compute=4.517s [fwd=3.470s bwd=1.032s opt=0.003s] dispatch=3.326s [2026-10-18 17:51:40,062] loop.py:45 INFO TRAIN_STEP step=2005 loss=2.012562 Training: 11%|█ | 2005/18018 [4:21:26<37:11:15, 8.36s/step, loss=2.013, acc=0.591, acc_len=1.35, thru=16.1, I=18.2, T=28.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:40,378] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:51:47,741] loop.py:45 INFO TRAIN_STEP step=2006 loss=1.934922 Training: 11%|█ | 2006/18018 [4:21:33<36:16:23, 8.16s/step, loss=1.935, acc=0.601, acc_len=1.38, thru=14.2, I=16.5, T=30.3, wait=3.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:51:47,735] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2005: forward=2912.6ms backward=1263.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:54,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:51:55,758] loop.py:45 INFO TRAIN_STEP step=2007 loss=1.994734 Training: 11%|█ | 2007/18018 [4:21:41<36:05:18, 8.11s/step, loss=1.995, acc=0.596, acc_len=1.37, thru=16.9, I=17.1, T=28.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:51:55,817] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:02,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:52:03,631] loop.py:45 INFO TRAIN_STEP step=2008 loss=2.058037 Training: 11%|█ | 2008/18018 [4:21:49<35:45:38, 8.04s/step, loss=2.058, acc=0.588, acc_len=1.34, thru=15.7, I=19.3, T=27.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:04,127] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:09,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:52:11,266] loop.py:45 INFO TRAIN_STEP step=2009 loss=1.996193 Training: 11%|█ | 2009/18018 [4:21:57<35:13:13, 7.92s/step, loss=1.996, acc=0.592, acc_len=1.36, thru=14.6, I=19.4, T=27.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:11,324] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 17:52:17,333] loop.py:380 INFO TIMING step=2010: step=3.585s data=1.548s compute=3.544s [fwd=2.541s bwd=0.988s opt=0.003s] dispatch=2.471s [2026-10-18 17:52:17,333] loop.py:45 INFO TRAIN_STEP step=2010 loss=2.131524 Training: 11%|█ | 2010/18018 [4:22:03<32:44:46, 7.36s/step, loss=2.132, acc=0.574, acc_len=1.30, thru=17.7, I=23.6, T=35.7, wait=2.5s, pool=56, epoch=1/2][2026-10-18 17:52:26,080] loop.py:45 INFO TRAIN_STEP step=2011 loss=1.846022 Training: 11%|█ | 2011/18018 [4:22:12<34:35:20, 7.78s/step, loss=1.846, acc=0.612, acc_len=1.42, thru=13.4, I=13.4, T=29.6, wait=4.4s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:52:26,075] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2010: forward=3124.8ms backward=1148.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:33,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:52:35,274] loop.py:45 INFO TRAIN_STEP step=2012 loss=2.196934 Training: 11%|█ | 2012/18018 [4:22:21<36:28:14, 8.20s/step, loss=2.197, acc=0.566, acc_len=1.28, thru=15.7, I=16.8, T=29.2, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:35,581] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:40,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:52:41,719] loop.py:45 INFO TRAIN_STEP step=2013 loss=1.756362 Training: 11%|█ | 2013/18018 [4:22:27<34:07:39, 7.68s/step, loss=1.756, acc=0.629, acc_len=1.48, thru=17.2, I=28.0, T=33.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:42,029] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:51,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:52:52,759] loop.py:45 INFO TRAIN_STEP step=2014 loss=1.923097 Training: 11%|█ | 2014/18018 [4:22:38<38:36:28, 8.68s/step, loss=1.923, acc=0.604, acc_len=1.39, thru=9.8, I=20.1, T=30.1, wait=6.8s, pool=64, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:53,157] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:52:58,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:53:00,548] loop.py:380 INFO TIMING step=2015: step=4.475s data=2.136s compute=4.428s [fwd=3.069s bwd=1.344s opt=0.003s] dispatch=3.305s [2026-10-18 17:53:00,548] loop.py:45 INFO TRAIN_STEP step=2015 loss=1.945755 Training: 11%|█ | 2015/18018 [4:22:46<37:24:52, 8.42s/step, loss=1.946, acc=0.607, acc_len=1.41, thru=15.5, I=20.9, T=28.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:00,603] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:07,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:53:08,572] loop.py:45 INFO TRAIN_STEP step=2016 loss=1.900842 Training: 11%|█ | 2016/18018 [4:22:54<36:53:17, 8.30s/step, loss=1.901, acc=0.612, acc_len=1.42, thru=13.2, I=17.1, T=29.3, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:53:08,566] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2015: forward=3093.4ms backward=1224.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:08,687] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:13,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:53:14,486] loop.py:45 INFO TRAIN_STEP step=2017 loss=2.032729 Training: 11%|█ | 2017/18018 [4:23:00<33:42:11, 7.58s/step, loss=2.033, acc=0.582, acc_len=1.33, thru=19.5, I=23.7, T=33.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:14,977] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:18,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:53:20,488] loop.py:45 INFO TRAIN_STEP step=2018 loss=1.946002 Training: 11%|█ | 2018/18018 [4:23:06<31:35:48, 7.11s/step, loss=1.946, acc=0.603, acc_len=1.39, thru=20.2, I=26.7, T=35.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:20,760] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:27,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:53:27,730] loop.py:45 INFO TRAIN_STEP step=2019 loss=1.960752 Training: 11%|█ | 2019/18018 [4:23:13<31:46:07, 7.15s/step, loss=1.961, acc=0.599, acc_len=1.39, thru=15.8, I=20.6, T=31.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:27,831] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-18 17:53:36,730] loop.py:380 INFO TIMING step=2020: step=4.288s data=1.855s compute=4.245s [fwd=2.941s bwd=1.289s opt=0.003s] dispatch=4.702s [2026-10-18 17:53:36,730] loop.py:45 INFO TRAIN_STEP step=2020 loss=2.026882 Training: 11%|█ | 2020/18018 [4:23:22<34:14:09, 7.70s/step, loss=2.027, acc=0.583, acc_len=1.34, thru=14.1, I=13.8, T=29.8, wait=4.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:42,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:53:43,925] loop.py:45 INFO TRAIN_STEP step=2021 loss=1.994123 Training: 11%|█ | 2021/18018 [4:23:30<33:33:10, 7.55s/step, loss=1.994, acc=0.589, acc_len=1.35, thru=18.3, I=19.9, T=30.7, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:53:43,920] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2020: forward=2951.9ms backward=1166.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:44,399] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:49,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:53:50,746] loop.py:45 INFO TRAIN_STEP step=2022 loss=1.766147 Training: 11%|█ | 2022/18018 [4:23:36<32:34:44, 7.33s/step, loss=1.766, acc=0.633, acc_len=1.48, thru=17.7, I=26.2, T=32.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:53:51,217] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 17:53:58,445] loop.py:45 INFO TRAIN_STEP step=2023 loss=1.813046 Training: 11%|█ | 2023/18018 [4:23:44<33:04:04, 7.44s/step, loss=1.813, acc=0.620, acc_len=1.45, thru=12.2, I=15.8, T=32.6, wait=3.8s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:07,339] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:54:08,444] loop.py:45 INFO TRAIN_STEP step=2024 loss=2.033751 Training: 11%|█ | 2024/18018 [4:23:54<36:28:27, 8.21s/step, loss=2.034, acc=0.588, acc_len=1.35, thru=15.8, I=16.6, T=27.6, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:08,849] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:15,647] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:54:17,059] loop.py:380 INFO TIMING step=2025: step=3.949s data=1.700s compute=3.920s [fwd=2.784s bwd=1.121s opt=0.003s] dispatch=4.655s [2026-10-18 17:54:17,060] loop.py:45 INFO TRAIN_STEP step=2025 loss=1.837004 Training: 11%|█ | 2025/18018 [4:24:03<37:00:44, 8.33s/step, loss=1.837, acc=0.615, acc_len=1.42, thru=14.0, I=18.0, T=32.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:17,159] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:22,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:54:23,230] loop.py:45 INFO TRAIN_STEP step=2026 loss=1.936937 Training: 11%|█ | 2026/18018 [4:24:09<34:07:51, 7.68s/step, loss=1.937, acc=0.599, acc_len=1.38, thru=19.3, I=23.5, T=33.5, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:54:23,226] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2025: forward=2703.5ms backward=1064.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:23,302] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:30,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:54:32,255] loop.py:45 INFO TRAIN_STEP step=2027 loss=1.882175 Training: 11%|█ | 2027/18018 [4:24:18<35:54:56, 8.09s/step, loss=1.882, acc=0.612, acc_len=1.42, thru=13.2, I=16.4, T=26.0, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:32,738] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:38,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:54:40,329] loop.py:45 INFO TRAIN_STEP step=2028 loss=1.889154 Training: 11%|█▏ | 2028/18018 [4:24:26<35:53:55, 8.08s/step, loss=1.889, acc=0.608, acc_len=1.42, thru=13.7, I=17.1, T=29.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:40,707] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 17:54:46,808] loop.py:45 INFO TRAIN_STEP step=2029 loss=1.890848 Training: 11%|█▏ | 2029/18018 [4:24:32<33:45:27, 7.60s/step, loss=1.891, acc=0.610, acc_len=1.41, thru=16.9, I=23.9, T=33.3, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:54,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:54:55,200] loop.py:380 INFO TIMING step=2030: step=4.659s data=2.190s compute=4.630s [fwd=3.242s bwd=1.373s opt=0.003s] dispatch=3.724s [2026-10-18 17:54:55,200] loop.py:45 INFO TRAIN_STEP step=2030 loss=2.017553 Training: 11%|█▏ | 2030/18018 [4:24:41<34:48:41, 7.84s/step, loss=2.018, acc=0.593, acc_len=1.36, thru=16.3, I=17.1, T=27.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:54:55,512] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:00,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:55:01,590] loop.py:45 INFO TRAIN_STEP step=2031 loss=1.903633 Training: 11%|█▏ | 2031/18018 [4:24:47<32:52:39, 7.40s/step, loss=1.904, acc=0.608, acc_len=1.41, thru=17.3, I=22.4, T=31.1, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:55:01,586] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2030: forward=2856.0ms backward=1212.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:02,010] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:55:10,841] loop.py:45 INFO TRAIN_STEP step=2032 loss=1.925884 Training: 11%|█▏ | 2032/18018 [4:24:56<35:20:23, 7.96s/step, loss=1.926, acc=0.606, acc_len=1.40, thru=14.0, I=14.7, T=31.6, wait=5.2s, pool=60, epoch=1/2][2026-10-18 17:55:18,716] loop.py:45 INFO TRAIN_STEP step=2033 loss=2.042550 Training: 11%|█▏ | 2033/18018 [4:25:04<35:13:27, 7.93s/step, loss=2.043, acc=0.584, acc_len=1.33, thru=16.5, I=18.6, T=28.9, wait=3.4s, pool=60, epoch=1/2][2026-10-18 17:55:25,624] loop.py:45 INFO TRAIN_STEP step=2034 loss=1.950146 Training: 11%|█▏ | 2034/18018 [4:25:11<33:51:29, 7.63s/step, loss=1.950, acc=0.603, acc_len=1.38, thru=16.2, I=16.9, T=31.9, wait=2.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:30,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:55:32,299] loop.py:380 INFO TIMING step=2035: step=4.063s data=1.911s compute=4.023s [fwd=2.868s bwd=1.140s opt=0.003s] dispatch=2.602s [2026-10-18 17:55:32,300] loop.py:45 INFO TRAIN_STEP step=2035 loss=1.839740 Training: 11%|█▏ | 2035/18018 [4:25:18<32:35:25, 7.34s/step, loss=1.840, acc=0.613, acc_len=1.43, thru=18.8, I=21.0, T=31.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:32,593] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:38,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:55:39,393] loop.py:45 INFO TRAIN_STEP step=2036 loss=1.993860 Training: 11%|█▏ | 2036/18018 [4:25:25<32:15:28, 7.27s/step, loss=1.994, acc=0.591, acc_len=1.36, thru=17.3, I=22.1, T=33.9, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:55:39,388] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2035: forward=2655.1ms backward=1071.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:39,622] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:44,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:55:45,862] loop.py:45 INFO TRAIN_STEP step=2037 loss=1.909705 Training: 11%|█▏ | 2037/18018 [4:25:31<31:11:48, 7.03s/step, loss=1.910, acc=0.606, acc_len=1.40, thru=19.7, I=24.9, T=31.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:46,116] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:52,098] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:55:52,575] loop.py:45 INFO TRAIN_STEP step=2038 loss=1.909731 Training: 11%|█▏ | 2038/18018 [4:25:38<30:46:28, 6.93s/step, loss=1.910, acc=0.602, acc_len=1.39, thru=17.3, I=23.3, T=30.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:55:52,608] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 17:55:59,916] loop.py:45 INFO TRAIN_STEP step=2039 loss=2.084380 Training: 11%|█▏ | 2039/18018 [4:25:45<31:18:45, 7.05s/step, loss=2.084, acc=0.585, acc_len=1.33, thru=14.9, I=16.5, T=30.1, wait=3.1s, pool=48, epoch=1/2][2026-10-18 17:56:10,367] loop.py:380 INFO TIMING step=2040: step=4.477s data=2.135s compute=4.438s [fwd=3.103s bwd=1.320s opt=0.003s] dispatch=5.966s [2026-10-18 17:56:10,367] loop.py:45 INFO TRAIN_STEP step=2040 loss=2.050969 Training: 11%|█▏ | 2040/18018 [4:25:56<35:50:11, 8.07s/step, loss=2.051, acc=0.584, acc_len=1.33, thru=12.5, I=14.1, T=28.6, wait=6.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:17,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:56:17,940] loop.py:45 INFO TRAIN_STEP step=2041 loss=1.952528 (TrainerActor pid=9310) [2026-10-18 17:56:17,935] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2040: forward=3021.6ms backward=1356.3ms Training: 11%|█▏ | 2041/18018 [4:26:04<35:10:01, 7.92s/step, loss=1.953, acc=0.597, acc_len=1.37, thru=16.8, I=15.9, T=28.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:18,275] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:22,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:56:24,155] loop.py:45 INFO TRAIN_STEP step=2042 loss=1.981164 Training: 11%|█▏ | 2042/18018 [4:26:10<32:53:11, 7.41s/step, loss=1.981, acc=0.595, acc_len=1.37, thru=18.9, I=23.0, T=34.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:24,504] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:30,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:56:32,138] loop.py:45 INFO TRAIN_STEP step=2043 loss=1.935091 Training: 11%|█▏ | 2043/18018 [4:26:18<33:38:58, 7.58s/step, loss=1.935, acc=0.598, acc_len=1.38, thru=14.8, I=20.1, T=28.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:32,425] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:40,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:56:40,497] loop.py:45 INFO TRAIN_STEP step=2044 loss=1.806441 Training: 11%|█▏ | 2044/18018 [4:26:26<34:40:51, 7.82s/step, loss=1.806, acc=0.622, acc_len=1.46, thru=15.9, I=16.6, T=34.0, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:40,546] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:45,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:56:46,978] loop.py:380 INFO TIMING step=2045: step=3.954s data=1.840s compute=3.926s [fwd=2.827s bwd=1.083s opt=0.003s] dispatch=2.514s [2026-10-18 17:56:46,978] loop.py:45 INFO TRAIN_STEP step=2045 loss=2.010907 Training: 11%|█▏ | 2045/18018 [4:26:33<32:54:15, 7.42s/step, loss=2.011, acc=0.595, acc_len=1.37, thru=18.9, I=22.6, T=32.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:47,095] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:52,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:56:53,573] loop.py:45 INFO TRAIN_STEP step=2046 loss=1.905322 Training: 11%|█▏ | 2046/18018 [4:26:39<31:48:24, 7.17s/step, loss=1.905, acc=0.603, acc_len=1.39, thru=17.9, I=23.3, T=33.9, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:56:53,568] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2045: forward=2699.3ms backward=1029.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:53,958] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:56:59,132] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:57:00,007] loop.py:45 INFO TRAIN_STEP step=2047 loss=1.731402 Training: 11%|█▏ | 2047/18018 [4:26:46<30:49:43, 6.95s/step, loss=1.731, acc=0.636, acc_len=1.50, thru=18.9, I=23.6, T=30.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:00,142] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:05,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:57:06,484] loop.py:45 INFO TRAIN_STEP step=2048 loss=2.046727 Training: 11%|█▏ | 2048/18018 [4:26:52<30:11:46, 6.81s/step, loss=2.047, acc=0.590, acc_len=1.35, thru=18.3, I=22.1, T=34.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:06,576] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:11,669] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:57:14,053] loop.py:45 INFO TRAIN_STEP step=2049 loss=1.924967 Training: 11%|█▏ | 2049/18018 [4:27:00<31:12:23, 7.04s/step, loss=1.925, acc=0.602, acc_len=1.39, thru=15.9, I=22.1, T=26.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:14,190] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:19,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:57:20,268] loop.py:380 INFO TIMING step=2050: step=4.141s data=1.898s compute=4.109s [fwd=2.920s bwd=1.173s opt=0.003s] dispatch=2.065s [2026-10-18 17:57:20,269] loop.py:45 INFO TRAIN_STEP step=2050 loss=1.875278 Training: 11%|█▏ | 2050/18018 [4:27:06<30:06:54, 6.79s/step, loss=1.875, acc=0.614, acc_len=1.42, thru=16.9, I=26.5, T=30.9, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:20,600] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 17:57:28,137] loop.py:45 INFO TRAIN_STEP step=2051 loss=1.935507 Training: 11%|█▏ | 2051/18018 [4:27:14<31:33:03, 7.11s/step, loss=1.936, acc=0.605, acc_len=1.40, thru=14.4, I=16.9, T=30.1, wait=3.6s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:57:28,132] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2050: forward=2910.7ms backward=1286.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:35,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:57:37,874] loop.py:45 INFO TRAIN_STEP step=2052 loss=1.980639 Training: 11%|█▏ | 2052/18018 [4:27:23<35:02:14, 7.90s/step, loss=1.981, acc=0.600, acc_len=1.38, thru=14.4, I=15.4, T=24.8, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:37,940] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:45,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:57:46,757] loop.py:45 INFO TRAIN_STEP step=2053 loss=2.067526 Training: 11%|█▏ | 2053/18018 [4:27:32<36:20:38, 8.20s/step, loss=2.068, acc=0.583, acc_len=1.33, thru=14.5, I=16.5, T=31.9, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:47,106] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:57:54,210] loop.py:45 INFO TRAIN_STEP step=2054 loss=1.935934 Training: 11%|█▏ | 2054/18018 [4:27:40<35:21:07, 7.97s/step, loss=1.936, acc=0.603, acc_len=1.39, thru=15.2, I=19.5, T=31.6, wait=3.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:57:59,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:58:01,441] loop.py:380 INFO TIMING step=2055: step=3.930s data=1.708s compute=3.894s [fwd=2.772s bwd=1.107s opt=0.003s] dispatch=3.291s [2026-10-18 17:58:01,441] loop.py:45 INFO TRAIN_STEP step=2055 loss=1.970979 Training: 11%|█▏ | 2055/18018 [4:27:47<34:21:46, 7.75s/step, loss=1.971, acc=0.598, acc_len=1.37, thru=18.4, I=19.8, T=32.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:01,889] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:08,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:58:10,763] loop.py:45 INFO TRAIN_STEP step=2056 loss=1.858620 Training: 11%|█▏ | 2056/18018 [4:27:56<36:27:06, 8.22s/step, loss=1.859, acc=0.619, acc_len=1.44, thru=14.4, I=17.3, T=29.9, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:58:10,758] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2055: forward=2992.7ms backward=1246.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:10,965] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:15,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:58:16,725] loop.py:45 INFO TRAIN_STEP step=2057 loss=1.749577 Training: 11%|█▏ | 2057/18018 [4:28:02<33:26:42, 7.54s/step, loss=1.750, acc=0.632, acc_len=1.48, thru=18.9, I=23.1, T=33.3, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:16,857] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:23,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:58:23,791] loop.py:45 INFO TRAIN_STEP step=2058 loss=1.950762 Training: 11%|█▏ | 2058/18018 [4:28:09<32:48:33, 7.40s/step, loss=1.951, acc=0.597, acc_len=1.38, thru=16.9, I=21.4, T=30.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:24,072] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:29,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:58:30,478] loop.py:45 INFO TRAIN_STEP step=2059 loss=1.848502 Training: 11%|█▏ | 2059/18018 [4:28:16<31:51:25, 7.19s/step, loss=1.849, acc=0.613, acc_len=1.43, thru=17.7, I=21.8, T=33.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:30,570] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:35,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:58:37,369] loop.py:380 INFO TIMING step=2060: step=4.370s data=1.939s compute=4.326s [fwd=2.995s bwd=1.316s opt=0.003s] dispatch=2.511s [2026-10-18 17:58:37,369] loop.py:45 INFO TRAIN_STEP step=2060 loss=1.817955 Training: 11%|█▏ | 2060/18018 [4:28:23<31:27:40, 7.10s/step, loss=1.818, acc=0.620, acc_len=1.45, thru=17.2, I=25.8, T=29.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:37,797] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:45,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:58:45,436] loop.py:45 INFO TRAIN_STEP step=2061 loss=1.907024 Training: 11%|█▏ | 2061/18018 [4:28:31<32:45:03, 7.39s/step, loss=1.907, acc=0.605, acc_len=1.40, thru=16.8, I=17.5, T=29.9, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:58:45,431] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2060: forward=2943.2ms backward=1294.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:45,552] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:51,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:58:52,281] loop.py:45 INFO TRAIN_STEP step=2062 loss=1.935633 Training: 11%|█▏ | 2062/18018 [4:28:38<32:01:28, 7.23s/step, loss=1.936, acc=0.601, acc_len=1.39, thru=18.0, I=19.6, T=31.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:52,652] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:58:59,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:59:01,698] loop.py:45 INFO TRAIN_STEP step=2063 loss=1.923622 Training: 11%|█▏ | 2063/18018 [4:28:47<34:56:01, 7.88s/step, loss=1.924, acc=0.603, acc_len=1.40, thru=14.2, I=19.0, T=28.2, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:01,726] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 [2026-10-18 17:59:08,053] loop.py:45 INFO TRAIN_STEP step=2064 loss=1.902373 Training: 11%|█▏ | 2064/18018 [4:28:54<32:54:22, 7.43s/step, loss=1.902, acc=0.604, acc_len=1.39, thru=14.9, I=23.6, T=31.8, wait=2.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:17,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:59:18,666] loop.py:380 INFO TIMING step=2065: step=4.516s data=2.172s compute=4.417s [fwd=3.024s bwd=1.378s opt=0.003s] dispatch=6.086s [2026-10-18 17:59:18,666] loop.py:45 INFO TRAIN_STEP step=2065 loss=1.718701 Training: 11%|█▏ | 2065/18018 [4:29:04<37:08:33, 8.38s/step, loss=1.719, acc=0.636, acc_len=1.49, thru=15.5, I=15.1, T=28.3, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:18,849] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:24,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:59:24,591] loop.py:45 INFO TRAIN_STEP step=2066 loss=1.921376 Training: 11%|█▏ | 2066/18018 [4:29:10<33:52:23, 7.64s/step, loss=1.921, acc=0.606, acc_len=1.40, thru=18.9, I=20.7, T=31.8, wait=1.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:59:24,587] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2065: forward=2873.4ms backward=1109.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:24,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:29,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 17:59:31,297] loop.py:45 INFO TRAIN_STEP step=2067 loss=2.093874 Training: 11%|█▏ | 2067/18018 [4:29:17<32:37:13, 7.36s/step, loss=2.094, acc=0.584, acc_len=1.33, thru=19.1, I=24.2, T=28.1, wait=2.1s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:31,562] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:37,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:59:38,091] loop.py:45 INFO TRAIN_STEP step=2068 loss=1.879284 Training: 11%|█▏ | 2068/18018 [4:29:24<31:51:58, 7.19s/step, loss=1.879, acc=0.613, acc_len=1.42, thru=15.9, I=21.5, T=33.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:38,148] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:43,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:59:44,877] loop.py:45 INFO TRAIN_STEP step=2069 loss=1.899788 Training: 11%|█▏ | 2069/18018 [4:29:30<31:19:22, 7.07s/step, loss=1.900, acc=0.613, acc_len=1.43, thru=18.8, I=21.3, T=30.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:45,313] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 17:59:51,819] loop.py:380 INFO TIMING step=2070: step=4.028s data=1.817s compute=3.999s [fwd=2.855s bwd=1.129s opt=0.003s] dispatch=2.905s [2026-10-18 17:59:51,819] loop.py:45 INFO TRAIN_STEP step=2070 loss=1.901672 Training: 11%|█▏ | 2070/18018 [4:29:37<31:08:58, 7.03s/step, loss=1.902, acc=0.608, acc_len=1.40, thru=15.8, I=19.2, T=31.8, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:58,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 17:59:59,271] loop.py:45 INFO TRAIN_STEP step=2071 loss=1.881433 Training: 11%|█▏ | 2071/18018 [4:29:45<31:42:30, 7.16s/step, loss=1.881, acc=0.615, acc_len=1.42, thru=16.5, I=19.0, T=30.1, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 17:59:59,266] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2070: forward=2946.9ms backward=1251.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 17:59:59,553] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:04,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:00:05,736] loop.py:45 INFO TRAIN_STEP step=2072 loss=2.097283 Training: 11%|█▏ | 2072/18018 [4:29:51<30:47:01, 6.95s/step, loss=2.097, acc=0.574, acc_len=1.30, thru=17.6, I=22.8, T=31.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:06,196] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:11,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:00:12,646] loop.py:45 INFO TRAIN_STEP step=2073 loss=1.943712 Training: 12%|█▏ | 2073/18018 [4:29:58<30:43:42, 6.94s/step, loss=1.944, acc=0.603, acc_len=1.39, thru=17.4, I=24.1, T=32.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:13,071] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:00:18,855] loop.py:45 INFO TRAIN_STEP step=2074 loss=1.902948 Training: 12%|█▏ | 2074/18018 [4:30:04<29:45:43, 6.72s/step, loss=1.903, acc=0.610, acc_len=1.42, thru=16.4, I=21.0, T=33.4, wait=2.4s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:27,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:00:28,676] loop.py:380 INFO TIMING step=2075: step=4.159s data=1.852s compute=4.132s [fwd=2.973s bwd=1.144s opt=0.003s] dispatch=5.651s [2026-10-18 18:00:28,676] loop.py:45 INFO TRAIN_STEP step=2075 loss=1.968007 Training: 12%|█▏ | 2075/18018 [4:30:14<33:52:46, 7.65s/step, loss=1.968, acc=0.595, acc_len=1.37, thru=15.9, I=18.5, T=30.8, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:28,711] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:34,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 18:00:36,147] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2075: forward=3170.7ms backward=1291.3ms [2026-10-18 18:00:36,153] loop.py:45 INFO TRAIN_STEP step=2076 loss=2.080926 Training: 12%|█▏ | 2076/18018 [4:30:22<33:38:52, 7.60s/step, loss=2.081, acc=0.578, acc_len=1.31, thru=16.8, I=21.8, T=28.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:36,305] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:41,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:00:43,527] loop.py:45 INFO TRAIN_STEP step=2077 loss=1.880592 Training: 12%|█▏ | 2077/18018 [4:30:29<33:20:42, 7.53s/step, loss=1.881, acc=0.610, acc_len=1.42, thru=14.5, I=20.8, T=29.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:43,639] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:48,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:00:50,865] loop.py:45 INFO TRAIN_STEP step=2078 loss=1.998482 Training: 12%|█▏ | 2078/18018 [4:30:36<33:05:23, 7.47s/step, loss=1.998, acc=0.588, acc_len=1.36, thru=14.9, I=23.1, T=29.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:51,196] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:57,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:00:58,083] loop.py:45 INFO TRAIN_STEP step=2079 loss=1.984447 Training: 12%|█▏ | 2079/18018 [4:30:44<32:44:54, 7.40s/step, loss=1.984, acc=0.596, acc_len=1.37, thru=14.5, I=20.9, T=28.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:00:58,349] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:03,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:01:04,937] loop.py:380 INFO TIMING step=2080: step=4.335s data=2.058s compute=4.291s [fwd=2.991s bwd=1.285s opt=0.003s] dispatch=2.508s [2026-10-18 18:01:04,937] loop.py:45 INFO TRAIN_STEP step=2080 loss=1.815642 Training: 12%|█▏ | 2080/18018 [4:30:51<32:01:32, 7.23s/step, loss=1.816, acc=0.623, acc_len=1.46, thru=17.8, I=22.0, T=29.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:05,075] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-18 18:01:10,522] loop.py:45 INFO TRAIN_STEP step=2081 loss=1.840190 Training: 12%|█▏ | 2081/18018 [4:30:56<29:49:51, 6.74s/step, loss=1.840, acc=0.621, acc_len=1.44, thru=19.7, I=23.4, T=35.5, wait=2.0s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:01:10,517] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2080: forward=2502.9ms backward=1044.6ms [2026-10-18 18:01:18,218] loop.py:45 INFO TRAIN_STEP step=2082 loss=1.896970 Training: 12%|█▏ | 2082/18018 [4:31:04<31:06:04, 7.03s/step, loss=1.897, acc=0.610, acc_len=1.42, thru=16.2, I=16.2, T=32.2, wait=3.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:26,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:01:26,645] loop.py:45 INFO TRAIN_STEP step=2083 loss=1.952436 Training: 12%|█▏ | 2083/18018 [4:31:12<32:57:33, 7.45s/step, loss=1.952, acc=0.599, acc_len=1.38, thru=17.0, I=16.1, T=26.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:27,131] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:34,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:01:36,587] loop.py:45 INFO TRAIN_STEP step=2084 loss=1.931498 Training: 12%|█▏ | 2084/18018 [4:31:22<36:16:30, 8.20s/step, loss=1.931, acc=0.603, acc_len=1.40, thru=13.5, I=16.1, T=26.5, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:36,627] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:43,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:01:44,402] loop.py:380 INFO TIMING step=2085: step=4.074s data=1.899s compute=4.037s [fwd=2.868s bwd=1.154s opt=0.003s] dispatch=3.731s [2026-10-18 18:01:44,403] loop.py:45 INFO TRAIN_STEP step=2085 loss=1.852599 Training: 12%|█▏ | 2085/18018 [4:31:30<35:46:01, 8.08s/step, loss=1.853, acc=0.615, acc_len=1.42, thru=14.0, I=20.5, T=31.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:44,543] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-18 18:01:51,567] loop.py:45 INFO TRAIN_STEP step=2086 loss=2.135991 Training: 12%|█▏ | 2086/18018 [4:31:37<34:33:00, 7.81s/step, loss=2.136, acc=0.573, acc_len=1.30, thru=17.4, I=20.2, T=33.5, wait=3.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:01:51,562] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2085: forward=2672.0ms backward=1110.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:57,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:01:59,078] loop.py:45 INFO TRAIN_STEP step=2087 loss=1.961215 Training: 12%|█▏ | 2087/18018 [4:31:45<34:09:01, 7.72s/step, loss=1.961, acc=0.599, acc_len=1.38, thru=16.8, I=18.8, T=32.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:01:59,347] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:04,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:02:06,023] loop.py:45 INFO TRAIN_STEP step=2088 loss=1.961687 Training: 12%|█▏ | 2088/18018 [4:31:52<33:07:25, 7.49s/step, loss=1.962, acc=0.602, acc_len=1.39, thru=17.4, I=20.6, T=34.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:06,431] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:12,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:02:12,971] loop.py:45 INFO TRAIN_STEP step=2089 loss=1.869920 Training: 12%|█▏ | 2089/18018 [4:31:59<32:24:32, 7.32s/step, loss=1.870, acc=0.614, acc_len=1.43, thru=17.9, I=21.2, T=28.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:13,264] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 18:02:20,276] loop.py:380 INFO TIMING step=2090: step=3.873s data=1.679s compute=3.845s [fwd=2.704s bwd=1.126s opt=0.003s] dispatch=3.423s [2026-10-18 18:02:20,276] loop.py:45 INFO TRAIN_STEP step=2090 loss=1.882380 Training: 12%|█▏ | 2090/18018 [4:32:06<32:22:51, 7.32s/step, loss=1.882, acc=0.609, acc_len=1.42, thru=15.2, I=16.1, T=33.1, wait=3.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:29,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:02:30,299] loop.py:45 INFO TRAIN_STEP step=2091 loss=2.142428 Training: 12%|█▏ | 2091/18018 [4:32:16<35:58:13, 8.13s/step, loss=2.142, acc=0.577, acc_len=1.31, thru=14.4, I=14.6, T=33.4, wait=6.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:02:30,294] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2090: forward=2767.6ms backward=1008.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:30,374] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:37,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:02:37,132] loop.py:45 INFO TRAIN_STEP step=2092 loss=1.903634 Training: 12%|█▏ | 2092/18018 [4:32:23<34:14:29, 7.74s/step, loss=1.904, acc=0.607, acc_len=1.40, thru=17.6, I=18.1, T=33.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:37,608] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:43,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:02:43,890] loop.py:45 INFO TRAIN_STEP step=2093 loss=1.960454 Training: 12%|█▏ | 2093/18018 [4:32:29<32:56:24, 7.45s/step, loss=1.960, acc=0.597, acc_len=1.38, thru=18.5, I=20.2, T=33.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:44,050] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:48,834] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:02:49,660] loop.py:45 INFO TRAIN_STEP step=2094 loss=1.840833 Training: 12%|█▏ | 2094/18018 [4:32:35<30:42:47, 6.94s/step, loss=1.841, acc=0.618, acc_len=1.43, thru=20.3, I=22.8, T=34.0, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:49,841] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:54,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:02:56,237] loop.py:380 INFO TIMING step=2095: step=4.111s data=1.967s compute=4.073s [fwd=2.876s bwd=1.182s opt=0.003s] dispatch=2.456s [2026-10-18 18:02:56,238] loop.py:45 INFO TRAIN_STEP step=2095 loss=1.879103 Training: 12%|█▏ | 2095/18018 [4:32:42<30:13:31, 6.83s/step, loss=1.879, acc=0.612, acc_len=1.42, thru=18.9, I=25.5, T=31.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:02:56,444] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:03:05,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:03:06,209] loop.py:45 INFO TRAIN_STEP step=2096 loss=1.872772 Training: 12%|█▏ | 2096/18018 [4:32:52<34:23:02, 7.77s/step, loss=1.873, acc=0.618, acc_len=1.43, thru=13.1, I=13.5, T=33.0, wait=6.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:03:06,204] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2095: forward=2629.9ms backward=1185.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:03:06,671] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:03:14,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:03:15,461] loop.py:45 INFO TRAIN_STEP step=2097 loss=1.910416 Training: 12%|█▏ | 2097/18018 [4:33:01<36:20:44, 8.22s/step, loss=1.910, acc=0.605, acc_len=1.40, thru=13.6, I=15.0, T=30.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:03:15,721] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:03:24,007] loop.py:45 INFO TRAIN_STEP step=2098 loss=2.025618 Training: 12%|█▏ | 2098/18018 [4:33:10<36:46:27, 8.32s/step, loss=2.026, acc=0.587, acc_len=1.34, thru=14.1, I=14.3, T=29.0, wait=4.1s, pool=56, epoch=1/2][2026-10-18 18:03:31,087] loop.py:45 INFO TRAIN_STEP step=2099 loss=1.898083 Training: 12%|█▏ | 2099/18018 [4:33:17<35:08:02, 7.95s/step, loss=1.898, acc=0.608, acc_len=1.41, thru=16.9, I=15.6, T=35.1, wait=3.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:03:39,168] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:03:39,946] loop.py:380 INFO TIMING step=2100: step=3.794s data=1.678s compute=3.749s [fwd=2.602s bwd=1.132s opt=0.003s] dispatch=5.056s [2026-10-18 18:03:39,946] loop.py:45 INFO TRAIN_STEP step=2100 loss=1.873288 Training: 12%|█▏ | 2100/18018 [4:33:26<36:20:43, 8.22s/step, loss=1.873, acc=0.615, acc_len=1.43, thru=16.7, I=15.7, T=33.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:03:40,172] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:03:48,294] loop.py:45 INFO TRAIN_STEP step=2101 loss=1.904568 Training: 12%|█▏ | 2101/18018 [4:33:34<36:30:33, 8.26s/step, loss=1.905, acc=0.605, acc_len=1.40, thru=13.9, I=14.6, T=31.6, wait=4.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:03:48,289] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2100: forward=2861.4ms backward=1127.3ms [2026-10-18 18:03:55,354] loop.py:45 INFO TRAIN_STEP step=2102 loss=1.783880 Training: 12%|█▏ | 2102/18018 [4:33:41<34:55:19, 7.90s/step, loss=1.784, acc=0.627, acc_len=1.46, thru=15.4, I=15.9, T=32.9, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:02,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:04:04,187] loop.py:45 INFO TRAIN_STEP step=2103 loss=2.011827 Training: 12%|█▏ | 2103/18018 [4:33:50<36:09:23, 8.18s/step, loss=2.012, acc=0.595, acc_len=1.37, thru=16.1, I=17.2, T=25.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:04,403] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:04:10,937] loop.py:45 INFO TRAIN_STEP step=2104 loss=2.006294 Training: 12%|█▏ | 2104/18018 [4:33:57<34:15:31, 7.75s/step, loss=2.006, acc=0.590, acc_len=1.35, thru=14.3, I=19.0, T=31.6, wait=2.7s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:18,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:04:21,340] loop.py:380 INFO TIMING step=2105: step=5.304s data=2.290s compute=5.258s [fwd=3.321s bwd=1.921s opt=0.003s] dispatch=5.091s [2026-10-18 18:04:21,341] loop.py:45 INFO TRAIN_STEP step=2105 loss=1.958468 Training: 12%|█▏ | 2105/18018 [4:34:07<37:46:37, 8.55s/step, loss=1.958, acc=0.596, acc_len=1.37, thru=15.4, I=18.7, T=24.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:21,660] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:26,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:04:28,480] loop.py:45 INFO TRAIN_STEP step=2106 loss=1.780563 Training: 12%|█▏ | 2106/18018 [4:34:14<35:54:42, 8.12s/step, loss=1.781, acc=0.631, acc_len=1.48, thru=13.4, I=25.6, T=30.9, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:04:28,474] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2105: forward=2897.4ms backward=1193.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:28,643] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 18:04:35,690] loop.py:45 INFO TRAIN_STEP step=2107 loss=1.812270 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:35,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 Training: 12%|█▏ | 2107/18018 [4:34:21<34:41:45, 7.85s/step, loss=1.812, acc=0.619, acc_len=1.44, thru=14.4, I=20.0, T=29.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:36,106] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:43,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:04:44,547] loop.py:45 INFO TRAIN_STEP step=2108 loss=2.007879 Training: 12%|█▏ | 2108/18018 [4:34:30<36:01:36, 8.15s/step, loss=2.008, acc=0.590, acc_len=1.35, thru=14.9, I=16.3, T=30.4, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:44,873] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:51,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:04:52,459] loop.py:45 INFO TRAIN_STEP step=2109 loss=1.713162 Training: 12%|█▏ | 2109/18018 [4:34:38<35:42:27, 8.08s/step, loss=1.713, acc=0.640, acc_len=1.51, thru=15.4, I=20.2, T=29.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:52,698] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:04:58,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:04:59,935] loop.py:380 INFO TIMING step=2110: step=4.421s data=2.167s compute=4.366s [fwd=3.017s bwd=1.337s opt=0.003s] dispatch=3.045s [2026-10-18 18:04:59,935] loop.py:45 INFO TRAIN_STEP step=2110 loss=2.033848 Training: 12%|█▏ | 2110/18018 [4:34:46<34:54:17, 7.90s/step, loss=2.034, acc=0.585, acc_len=1.34, thru=15.3, I=19.1, T=29.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:00,241] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:04,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:05:06,139] loop.py:45 INFO TRAIN_STEP step=2111 loss=1.968180 Training: 12%|█▏ | 2111/18018 [4:34:52<32:39:25, 7.39s/step, loss=1.968, acc=0.601, acc_len=1.39, thru=17.6, I=24.6, T=33.2, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:05:06,134] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2110: forward=2624.8ms backward=1171.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:06,459] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:13,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:05:13,782] loop.py:45 INFO TRAIN_STEP step=2112 loss=1.926478 Training: 12%|█▏ | 2112/18018 [4:34:59<32:59:05, 7.47s/step, loss=1.926, acc=0.604, acc_len=1.39, thru=14.9, I=20.2, T=26.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:14,248] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:20,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:05:21,054] loop.py:45 INFO TRAIN_STEP step=2113 loss=1.919521 Training: 12%|█▏ | 2113/18018 [4:35:07<32:43:47, 7.41s/step, loss=1.920, acc=0.606, acc_len=1.39, thru=18.3, I=18.1, T=31.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:21,361] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:26,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:05:28,127] loop.py:45 INFO TRAIN_STEP step=2114 loss=1.774947 Training: 12%|█▏ | 2114/18018 [4:35:14<32:16:46, 7.31s/step, loss=1.775, acc=0.628, acc_len=1.47, thru=16.8, I=23.2, T=29.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:28,615] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:35,027] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:05:36,002] loop.py:380 INFO TIMING step=2115: step=4.554s data=2.207s compute=4.527s [fwd=3.220s bwd=1.292s opt=0.003s] dispatch=3.313s [2026-10-18 18:05:36,003] loop.py:45 INFO TRAIN_STEP step=2115 loss=1.988649 Training: 12%|█▏ | 2115/18018 [4:35:22<33:02:09, 7.48s/step, loss=1.989, acc=0.592, acc_len=1.37, thru=13.3, I=20.4, T=28.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:36,037] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:41,576] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:05:43,627] loop.py:45 INFO TRAIN_STEP step=2116 loss=1.943223 Training: 12%|█▏ | 2116/18018 [4:35:29<33:13:23, 7.52s/step, loss=1.943, acc=0.600, acc_len=1.39, thru=15.8, I=20.1, T=30.6, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:05:43,621] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2115: forward=2921.7ms backward=1214.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:05:44,098] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 18:05:50,241] loop.py:45 INFO TRAIN_STEP step=2117 loss=2.046915 Training: 12%|█▏ | 2117/18018 [4:35:36<32:01:08, 7.25s/step, loss=2.047, acc=0.590, acc_len=1.34, thru=15.0, I=22.5, T=33.9, wait=2.8s, pool=48, epoch=1/2][2026-10-18 18:06:00,937] loop.py:45 INFO TRAIN_STEP step=2118 loss=1.984914 Training: 12%|█▏ | 2118/18018 [4:35:47<36:35:07, 8.28s/step, loss=1.985, acc=0.601, acc_len=1.38, thru=13.6, I=14.0, T=26.9, wait=5.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:07,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:06:09,213] loop.py:45 INFO TRAIN_STEP step=2119 loss=1.896977 Training: 12%|█▏ | 2119/18018 [4:35:55<36:34:21, 8.28s/step, loss=1.897, acc=0.606, acc_len=1.40, thru=14.6, I=15.4, T=29.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:09,408] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:14,814] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:06:15,967] loop.py:380 INFO TIMING step=2120: step=3.619s data=1.554s compute=3.574s [fwd=2.515s bwd=1.046s opt=0.003s] dispatch=3.127s [2026-10-18 18:06:15,967] loop.py:45 INFO TRAIN_STEP step=2120 loss=1.797906 Training: 12%|█▏ | 2120/18018 [4:36:02<34:32:55, 7.82s/step, loss=1.798, acc=0.619, acc_len=1.44, thru=17.2, I=21.1, T=35.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:16,324] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:23,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:06:23,921] loop.py:45 INFO TRAIN_STEP step=2121 loss=2.090203 Training: 12%|█▏ | 2121/18018 [4:36:10<34:43:18, 7.86s/step, loss=2.090, acc=0.581, acc_len=1.32, thru=14.9, I=18.2, T=31.5, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:06:23,916] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2120: forward=2892.9ms backward=1123.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:24,018] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:30,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:06:31,149] loop.py:45 INFO TRAIN_STEP step=2122 loss=1.831605 Training: 12%|█▏ | 2122/18018 [4:36:17<33:52:36, 7.67s/step, loss=1.832, acc=0.618, acc_len=1.44, thru=16.9, I=19.9, T=26.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:31,412] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 18:06:37,319] loop.py:45 INFO TRAIN_STEP step=2123 loss=2.106068 Training: 12%|█▏ | 2123/18018 [4:36:23<31:53:11, 7.22s/step, loss=2.106, acc=0.578, acc_len=1.32, thru=15.6, I=17.1, T=33.6, wait=2.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:45,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:06:45,614] loop.py:45 INFO TRAIN_STEP step=2124 loss=1.852181 Training: 12%|█▏ | 2124/18018 [4:36:31<33:18:18, 7.54s/step, loss=1.852, acc=0.618, acc_len=1.44, thru=17.4, I=17.5, T=30.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:45,749] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:53,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:06:53,706] loop.py:380 INFO TIMING step=2125: step=4.702s data=2.488s compute=4.663s [fwd=3.509s bwd=1.139s opt=0.003s] dispatch=3.379s [2026-10-18 18:06:53,707] loop.py:45 INFO TRAIN_STEP step=2125 loss=2.042483 Training: 12%|█▏ | 2125/18018 [4:36:39<34:01:45, 7.71s/step, loss=2.042, acc=0.586, acc_len=1.34, thru=15.9, I=15.2, T=27.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:06:53,785] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:01,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:07:03,807] loop.py:45 INFO TRAIN_STEP step=2126 loss=1.841408 Training: 12%|█▏ | 2126/18018 [4:36:49<37:11:35, 8.43s/step, loss=1.841, acc=0.618, acc_len=1.44, thru=12.8, I=17.3, T=21.7, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:07:03,802] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2125: forward=4670.1ms backward=1183.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:04,080] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:09,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:07:11,870] loop.py:45 INFO TRAIN_STEP step=2127 loss=1.953740 Training: 12%|█▏ | 2127/18018 [4:36:57<36:42:37, 8.32s/step, loss=1.954, acc=0.598, acc_len=1.38, thru=16.4, I=23.0, T=26.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:12,263] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:18,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:07:19,766] loop.py:45 INFO TRAIN_STEP step=2128 loss=1.882102 Training: 12%|█▏ | 2128/18018 [4:37:05<36:09:09, 8.19s/step, loss=1.882, acc=0.617, acc_len=1.44, thru=17.1, I=21.2, T=31.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:19,870] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:27,128] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:07:28,754] loop.py:45 INFO TRAIN_STEP step=2129 loss=2.060584 Training: 12%|█▏ | 2129/18018 [4:37:14<37:12:27, 8.43s/step, loss=2.061, acc=0.584, acc_len=1.33, thru=14.4, I=17.5, T=30.4, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:29,144] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:35,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:07:35,457] loop.py:380 INFO TIMING step=2130: step=3.876s data=1.735s compute=3.834s [fwd=2.637s bwd=1.183s opt=0.003s] dispatch=2.818s [2026-10-18 18:07:35,458] loop.py:45 INFO TRAIN_STEP step=2130 loss=1.841291 Training: 12%|█▏ | 2130/18018 [4:37:21<34:55:09, 7.91s/step, loss=1.841, acc=0.617, acc_len=1.44, thru=17.5, I=22.0, T=33.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:35,601] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:40,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:07:42,959] loop.py:45 INFO TRAIN_STEP step=2131 loss=1.967981 (TrainerActor pid=9310) [2026-10-18 18:07:42,954] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2130: forward=2850.8ms backward=1165.7ms Training: 12%|█▏ | 2131/18018 [4:37:29<34:22:30, 7.79s/step, loss=1.968, acc=0.600, acc_len=1.38, thru=17.8, I=23.2, T=31.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:43,111] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:48,188] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:07:49,605] loop.py:45 INFO TRAIN_STEP step=2132 loss=1.856079 Training: 12%|█▏ | 2132/18018 [4:37:35<32:51:20, 7.45s/step, loss=1.856, acc=0.612, acc_len=1.42, thru=16.4, I=28.3, T=33.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:49,707] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:57,567] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:07:58,726] loop.py:45 INFO TRAIN_STEP step=2133 loss=1.862769 Training: 12%|█▏ | 2133/18018 [4:37:44<35:04:16, 7.95s/step, loss=1.863, acc=0.614, acc_len=1.43, thru=14.2, I=14.8, T=33.1, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:07:59,077] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:08:05,308] loop.py:45 INFO TRAIN_STEP step=2134 loss=1.856980 Training: 12%|█▏ | 2134/18018 [4:37:51<33:15:42, 7.54s/step, loss=1.857, acc=0.619, acc_len=1.44, thru=17.0, I=17.2, T=33.2, wait=2.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:12,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:08:14,404] loop.py:380 INFO TIMING step=2135: step=4.489s data=2.148s compute=4.452s [fwd=3.154s bwd=1.284s opt=0.003s] dispatch=4.597s [2026-10-18 18:08:14,404] loop.py:45 INFO TRAIN_STEP step=2135 loss=2.240403 Training: 12%|█▏ | 2135/18018 [4:38:00<35:19:15, 8.01s/step, loss=2.240, acc=0.561, acc_len=1.26, thru=16.9, I=16.7, T=28.5, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:14,632] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:19,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:08:21,280] loop.py:45 INFO TRAIN_STEP step=2136 loss=2.017802 Training: 12%|█▏ | 2136/18018 [4:38:07<33:49:23, 7.67s/step, loss=2.018, acc=0.592, acc_len=1.36, thru=16.8, I=25.3, T=28.8, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:08:21,275] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2135: forward=3107.6ms backward=1291.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:21,525] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:08:29,260] loop.py:45 INFO TRAIN_STEP step=2137 loss=1.881073 Training: 12%|█▏ | 2137/18018 [4:38:15<34:14:24, 7.76s/step, loss=1.881, acc=0.613, acc_len=1.42, thru=13.3, I=19.7, T=32.0, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:36,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:08:37,743] loop.py:45 INFO TRAIN_STEP step=2138 loss=1.829886 Training: 12%|█▏ | 2138/18018 [4:38:23<35:11:13, 7.98s/step, loss=1.830, acc=0.618, acc_len=1.44, thru=14.5, I=17.2, T=29.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:38,164] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:44,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:08:45,503] loop.py:45 INFO TRAIN_STEP step=2139 loss=1.959829 Training: 12%|█▏ | 2139/18018 [4:38:31<34:53:50, 7.91s/step, loss=1.960, acc=0.603, acc_len=1.38, thru=14.5, I=20.0, T=29.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:45,576] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:50,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:08:52,834] loop.py:380 INFO TIMING step=2140: step=4.557s data=2.190s compute=4.526s [fwd=3.243s bwd=1.269s opt=0.003s] dispatch=2.765s [2026-10-18 18:08:52,834] loop.py:45 INFO TRAIN_STEP step=2140 loss=1.883619 Training: 12%|█▏ | 2140/18018 [4:38:38<34:07:40, 7.74s/step, loss=1.884, acc=0.602, acc_len=1.39, thru=15.7, I=23.1, T=28.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:08:52,864] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-18 18:08:59,669] loop.py:45 INFO TRAIN_STEP step=2141 loss=1.829518 Training: 12%|█▏ | 2141/18018 [4:38:45<32:55:56, 7.47s/step, loss=1.830, acc=0.620, acc_len=1.45, thru=15.7, I=18.7, T=33.6, wait=3.0s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:08:59,664] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2140: forward=2673.9ms backward=1082.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:07,279] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:09:08,071] loop.py:45 INFO TRAIN_STEP step=2142 loss=1.863940 Training: 12%|█▏ | 2142/18018 [4:38:54<34:10:05, 7.75s/step, loss=1.864, acc=0.611, acc_len=1.41, thru=16.8, I=17.9, T=31.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:08,285] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:09:15,525] loop.py:45 INFO TRAIN_STEP step=2143 loss=1.987396 Training: 12%|█▏ | 2143/18018 [4:39:01<33:46:31, 7.66s/step, loss=1.987, acc=0.592, acc_len=1.36, thru=15.9, I=17.2, T=30.6, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:22,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:09:24,105] loop.py:45 INFO TRAIN_STEP step=2144 loss=2.088917 Training: 12%|█▏ | 2144/18018 [4:39:10<34:59:27, 7.94s/step, loss=2.089, acc=0.578, acc_len=1.31, thru=15.4, I=16.1, T=32.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:24,278] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:09:30,785] loop.py:380 INFO TIMING step=2145: step=3.939s data=1.794s compute=3.904s [fwd=2.776s bwd=1.113s opt=0.003s] dispatch=2.732s [2026-10-18 18:09:30,785] loop.py:45 INFO TRAIN_STEP step=2145 loss=1.967253 Training: 12%|█▏ | 2145/18018 [4:39:16<33:19:45, 7.56s/step, loss=1.967, acc=0.597, acc_len=1.37, thru=17.3, I=21.9, T=32.5, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:37,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:09:40,417] loop.py:45 INFO TRAIN_STEP step=2146 loss=1.918740 Training: 12%|█▏ | 2146/18018 [4:39:26<36:04:09, 8.18s/step, loss=1.919, acc=0.608, acc_len=1.41, thru=14.5, I=18.4, T=27.0, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:09:40,412] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2145: forward=3260.4ms backward=1427.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:40,767] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:45,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:09:46,826] loop.py:45 INFO TRAIN_STEP step=2147 loss=1.759938 Training: 12%|█▏ | 2147/18018 [4:39:32<33:43:17, 7.65s/step, loss=1.760, acc=0.625, acc_len=1.46, thru=17.7, I=29.8, T=31.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:47,247] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:55,173] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:09:55,237] loop.py:45 INFO TRAIN_STEP step=2148 loss=1.905658 Training: 12%|█▏ | 2148/18018 [4:39:41<34:43:32, 7.88s/step, loss=1.906, acc=0.608, acc_len=1.41, thru=13.8, I=13.9, T=33.2, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:09:55,675] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 18:10:02,095] loop.py:45 INFO TRAIN_STEP step=2149 loss=1.938627 Training: 12%|█▏ | 2149/18018 [4:39:48<33:22:30, 7.57s/step, loss=1.939, acc=0.599, acc_len=1.38, thru=18.0, I=18.7, T=32.4, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:10:08,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:10:08,722] loop.py:380 INFO TIMING step=2150: step=3.830s data=1.572s compute=3.804s [fwd=2.644s bwd=1.145s opt=0.003s] dispatch=2.789s [2026-10-18 18:10:08,722] loop.py:45 INFO TRAIN_STEP step=2150 loss=1.875795 Training: 12%|█▏ | 2150/18018 [4:39:54<32:07:39, 7.29s/step, loss=1.876, acc=0.615, acc_len=1.42, thru=19.2, I=19.5, T=33.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:10:08,781] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 18:10:14,495] loop.py:45 INFO TRAIN_STEP step=2151 loss=1.817523 Training: 12%|█▏ | 2151/18018 [4:40:00<30:07:07, 6.83s/step, loss=1.818, acc=0.622, acc_len=1.45, thru=19.9, I=20.6, T=34.4, wait=2.0s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:10:14,490] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2150: forward=2651.0ms backward=1009.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:10:23,555] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:10:23,852] loop.py:45 INFO TRAIN_STEP step=2152 loss=1.965914 Training: 12%|█▏ | 2152/18018 [4:40:09<33:27:12, 7.59s/step, loss=1.966, acc=0.595, acc_len=1.37, thru=14.9, I=15.0, T=32.5, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:10:24,057] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:10:30,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:10:31,192] loop.py:45 INFO TRAIN_STEP step=2153 loss=1.932456 Training: 12%|█▏ | 2153/18018 [4:40:17<33:07:19, 7.52s/step, loss=1.932, acc=0.603, acc_len=1.39, thru=18.3, I=18.2, T=35.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:10:31,628] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:10:37,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:10:38,325] loop.py:45 INFO TRAIN_STEP step=2154 loss=1.908785 Training: 12%|█▏ | 2154/18018 [4:40:24<32:36:41, 7.40s/step, loss=1.909, acc=0.601, acc_len=1.39, thru=18.4, I=23.5, T=35.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:10:38,715] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:10:45,380] loop.py:380 INFO TIMING step=2155: step=3.721s data=1.629s compute=3.679s [fwd=2.618s bwd=1.046s opt=0.003s] dispatch=3.326s [2026-10-18 18:10:45,380] loop.py:45 INFO TRAIN_STEP step=2155 loss=1.911620 Training: 12%|█▏ | 2155/18018 [4:40:31<32:09:10, 7.30s/step, loss=1.912, acc=0.607, acc_len=1.41, thru=16.6, I=18.6, T=34.4, wait=3.3s, pool=60, epoch=1/2][2026-10-18 18:10:54,053] loop.py:45 INFO TRAIN_STEP step=2156 loss=1.898389 Training: 12%|█▏ | 2156/18018 [4:40:40<33:58:09, 7.71s/step, loss=1.898, acc=0.607, acc_len=1.41, thru=13.1, I=14.0, T=30.6, wait=4.5s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:10:54,048] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2155: forward=2915.2ms backward=1204.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:03,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:11:04,484] loop.py:45 INFO TRAIN_STEP step=2157 loss=1.819559 Training: 12%|█▏ | 2157/18018 [4:40:50<37:33:53, 8.53s/step, loss=1.820, acc=0.619, acc_len=1.44, thru=16.1, I=15.0, T=32.1, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:04,643] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:11,457] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:11:11,657] loop.py:45 INFO TRAIN_STEP step=2158 loss=1.870104 Training: 12%|█▏ | 2158/18018 [4:40:57<35:46:36, 8.12s/step, loss=1.870, acc=0.615, acc_len=1.43, thru=18.0, I=19.3, T=32.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:11,959] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 18:11:19,556] loop.py:45 INFO TRAIN_STEP step=2159 loss=2.053064 Training: 12%|█▏ | 2159/18018 [4:41:05<35:28:54, 8.05s/step, loss=2.053, acc=0.582, acc_len=1.32, thru=16.2, I=15.6, T=29.3, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:27,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:11:30,129] loop.py:380 INFO TIMING step=2160: step=4.819s data=2.349s compute=4.791s [fwd=3.358s bwd=1.418s opt=0.003s] dispatch=5.742s [2026-10-18 18:11:30,130] loop.py:45 INFO TRAIN_STEP step=2160 loss=1.851615 Training: 12%|█▏ | 2160/18018 [4:41:16<38:48:30, 8.81s/step, loss=1.852, acc=0.617, acc_len=1.44, thru=12.4, I=15.4, T=26.6, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:30,300] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:37,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 18:11:37,204] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2160: forward=3160.7ms backward=1383.5ms [2026-10-18 18:11:37,209] loop.py:45 INFO TRAIN_STEP step=2161 loss=1.917427 Training: 12%|█▏ | 2161/18018 [4:41:23<36:30:57, 8.29s/step, loss=1.917, acc=0.606, acc_len=1.40, thru=14.9, I=20.6, T=27.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:37,566] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:42,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:11:43,620] loop.py:45 INFO TRAIN_STEP step=2162 loss=1.887215 Training: 12%|█▏ | 2162/18018 [4:41:29<34:01:52, 7.73s/step, loss=1.887, acc=0.606, acc_len=1.41, thru=17.4, I=19.9, T=30.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:43,666] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:50,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:11:51,285] loop.py:45 INFO TRAIN_STEP step=2163 loss=1.850873 Training: 12%|█▏ | 2163/18018 [4:41:37<33:56:59, 7.71s/step, loss=1.851, acc=0.613, acc_len=1.43, thru=15.2, I=18.5, T=32.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:51,480] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:57,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:11:58,805] loop.py:45 INFO TRAIN_STEP step=2164 loss=1.843032 Training: 12%|█▏ | 2164/18018 [4:41:44<33:41:54, 7.65s/step, loss=1.843, acc=0.615, acc_len=1.44, thru=16.3, I=19.5, T=30.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:11:58,836] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 18:12:08,779] loop.py:380 INFO TIMING step=2165: step=4.340s data=1.889s compute=4.303s [fwd=3.053s bwd=1.235s opt=0.003s] dispatch=5.623s [2026-10-18 18:12:08,779] loop.py:45 INFO TRAIN_STEP step=2165 loss=1.879015 Training: 12%|█▏ | 2165/18018 [4:41:54<36:45:37, 8.35s/step, loss=1.879, acc=0.612, acc_len=1.41, thru=12.5, I=12.0, T=29.5, wait=5.6s, pool=60, epoch=1/2][2026-10-18 18:12:17,083] loop.py:45 INFO TRAIN_STEP step=2166 loss=2.032256 Training: 12%|█▏ | 2166/18018 [4:42:03<36:42:17, 8.34s/step, loss=2.032, acc=0.591, acc_len=1.35, thru=16.4, I=15.2, T=30.1, wait=4.0s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:12:17,077] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2165: forward=2952.8ms backward=1249.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:23,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:12:24,848] loop.py:45 INFO TRAIN_STEP step=2167 loss=1.978244 Training: 12%|█▏ | 2167/18018 [4:42:10<35:56:45, 8.16s/step, loss=1.978, acc=0.598, acc_len=1.37, thru=17.8, I=18.5, T=28.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:25,138] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:12:31,125] loop.py:45 INFO TRAIN_STEP step=2168 loss=1.946797 Training: 12%|█▏ | 2168/18018 [4:42:17<33:27:00, 7.60s/step, loss=1.947, acc=0.597, acc_len=1.38, thru=17.2, I=19.0, T=33.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:31,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:31,560] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:36,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:12:38,135] loop.py:45 INFO TRAIN_STEP step=2169 loss=1.992319 Training: 12%|█▏ | 2169/18018 [4:42:24<32:40:25, 7.42s/step, loss=1.992, acc=0.588, acc_len=1.35, thru=18.3, I=20.4, T=29.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:38,478] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:44,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:12:46,130] loop.py:380 INFO TIMING step=2170: step=4.203s data=1.936s compute=4.165s [fwd=2.942s bwd=1.208s opt=0.003s] dispatch=3.782s [2026-10-18 18:12:46,130] loop.py:45 INFO TRAIN_STEP step=2170 loss=2.011011 Training: 12%|█▏ | 2170/18018 [4:42:32<33:25:50, 7.59s/step, loss=2.011, acc=0.593, acc_len=1.35, thru=14.2, I=17.9, T=30.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:46,230] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:51,760] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:12:52,751] loop.py:45 INFO TRAIN_STEP step=2171 loss=1.782723 Training: 12%|█▏ | 2171/18018 [4:42:38<32:08:35, 7.30s/step, loss=1.783, acc=0.626, acc_len=1.47, thru=16.4, I=22.1, T=30.9, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:12:52,746] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2170: forward=2871.0ms backward=1221.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:12:52,792] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:13:00,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:13:02,262] loop.py:45 INFO TRAIN_STEP step=2172 loss=1.871286 Training: 12%|█▏ | 2172/18018 [4:42:48<35:03:17, 7.96s/step, loss=1.871, acc=0.609, acc_len=1.41, thru=13.5, I=16.5, T=27.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:13:02,586] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:13:08,791] loop.py:45 INFO TRAIN_STEP step=2173 loss=1.879519 Training: 12%|█▏ | 2173/18018 [4:42:54<33:09:37, 7.53s/step, loss=1.880, acc=0.608, acc_len=1.41, thru=15.3, I=19.2, T=32.1, wait=2.5s, pool=52, epoch=1/2][2026-10-18 18:13:16,991] loop.py:45 INFO TRAIN_STEP step=2174 loss=1.974382 Training: 12%|█▏ | 2174/18018 [4:43:03<34:02:19, 7.73s/step, loss=1.974, acc=0.595, acc_len=1.37, thru=15.3, I=15.6, T=29.3, wait=3.8s, pool=44, epoch=1/2][2026-10-18 18:13:25,454] loop.py:380 INFO TIMING step=2175: step=4.233s data=1.948s compute=4.199s [fwd=3.023s bwd=1.161s opt=0.003s] dispatch=4.219s [2026-10-18 18:13:25,454] loop.py:45 INFO TRAIN_STEP step=2175 loss=1.974605 Training: 12%|█▏ | 2175/18018 [4:43:11<34:59:57, 7.95s/step, loss=1.975, acc=0.597, acc_len=1.37, thru=14.1, I=13.4, T=30.2, wait=4.2s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:13:35,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:13:35,866] loop.py:45 INFO TRAIN_STEP step=2176 loss=1.938788 Training: 12%|█▏ | 2176/18018 [4:43:21<38:14:35, 8.69s/step, loss=1.939, acc=0.604, acc_len=1.40, thru=15.4, I=14.5, T=31.7, wait=6.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:13:35,861] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2175: forward=2895.5ms backward=1099.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:13:36,032] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:13:40,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:13:42,092] loop.py:45 INFO TRAIN_STEP step=2177 loss=2.045013 Training: 12%|█▏ | 2177/18018 [4:43:28<34:59:04, 7.95s/step, loss=2.045, acc=0.587, acc_len=1.35, thru=18.6, I=23.9, T=31.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:13:42,158] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 18:13:50,649] loop.py:45 INFO TRAIN_STEP step=2178 loss=1.947082 Training: 12%|█▏ | 2178/18018 [4:43:36<35:46:57, 8.13s/step, loss=1.947, acc=0.600, acc_len=1.39, thru=14.6, I=14.8, T=33.4, wait=4.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:13:56,416] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:13:57,406] loop.py:45 INFO TRAIN_STEP step=2179 loss=2.035561 Training: 12%|█▏ | 2179/18018 [4:43:43<33:57:56, 7.72s/step, loss=2.036, acc=0.586, acc_len=1.34, thru=19.7, I=20.3, T=34.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:13:57,436] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:04,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:14:06,211] loop.py:380 INFO TIMING step=2180: step=4.247s data=1.835s compute=4.219s [fwd=2.934s bwd=1.270s opt=0.003s] dispatch=4.548s [2026-10-18 18:14:06,211] loop.py:45 INFO TRAIN_STEP step=2180 loss=1.852214 Training: 12%|█▏ | 2180/18018 [4:43:52<35:23:40, 8.05s/step, loss=1.852, acc=0.613, acc_len=1.43, thru=13.5, I=17.1, T=30.1, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:06,312] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:15,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 18:14:16,525] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2180: forward=3121.6ms backward=1343.5ms [2026-10-18 18:14:16,530] loop.py:45 INFO TRAIN_STEP step=2181 loss=1.973884 Training: 12%|█▏ | 2181/18018 [4:44:02<38:23:31, 8.73s/step, loss=1.974, acc=0.598, acc_len=1.37, thru=12.9, I=13.6, T=28.3, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:16,680] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:23,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:14:24,332] loop.py:45 INFO TRAIN_STEP step=2182 loss=1.749900 Training: 12%|█▏ | 2182/18018 [4:44:10<37:10:20, 8.45s/step, loss=1.750, acc=0.632, acc_len=1.48, thru=14.8, I=16.1, T=30.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:24,499] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:30,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:14:31,030] loop.py:45 INFO TRAIN_STEP step=2183 loss=1.860215 Training: 12%|█▏ | 2183/18018 [4:44:17<34:51:18, 7.92s/step, loss=1.860, acc=0.613, acc_len=1.42, thru=17.7, I=19.5, T=28.6, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:31,180] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 18:14:37,857] loop.py:45 INFO TRAIN_STEP step=2184 loss=1.889549 Training: 12%|█▏ | 2184/18018 [4:44:23<33:24:27, 7.60s/step, loss=1.890, acc=0.613, acc_len=1.42, thru=16.8, I=18.0, T=32.6, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:44,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:14:45,478] loop.py:380 INFO TIMING step=2185: step=4.046s data=1.825s compute=4.007s [fwd=2.786s bwd=1.205s opt=0.003s] dispatch=3.565s [2026-10-18 18:14:45,479] loop.py:45 INFO TRAIN_STEP step=2185 loss=1.787100 Training: 12%|█▏ | 2185/18018 [4:44:31<33:26:15, 7.60s/step, loss=1.787, acc=0.621, acc_len=1.45, thru=16.5, I=18.8, T=31.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:45,601] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:55,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:14:56,543] loop.py:45 INFO TRAIN_STEP step=2186 loss=1.940333 Training: 12%|█▏ | 2186/18018 [4:44:42<38:00:15, 8.64s/step, loss=1.940, acc=0.600, acc_len=1.38, thru=11.2, I=12.0, T=29.4, wait=6.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:14:56,538] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2185: forward=2922.3ms backward=1355.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:14:56,990] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:04,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:15:04,356] loop.py:45 INFO TRAIN_STEP step=2187 loss=1.903487 Training: 12%|█▏ | 2187/18018 [4:44:50<36:54:24, 8.39s/step, loss=1.903, acc=0.606, acc_len=1.39, thru=16.4, I=17.4, T=32.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:04,712] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:11,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:15:11,806] loop.py:45 INFO TRAIN_STEP step=2188 loss=1.878963 Training: 12%|█▏ | 2188/18018 [4:44:57<35:39:47, 8.11s/step, loss=1.879, acc=0.611, acc_len=1.41, thru=17.4, I=17.9, T=32.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:12,108] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:18,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:15:19,698] loop.py:45 INFO TRAIN_STEP step=2189 loss=1.931908 Training: 12%|█▏ | 2189/18018 [4:45:05<35:22:07, 8.04s/step, loss=1.932, acc=0.605, acc_len=1.40, thru=16.7, I=19.7, T=31.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:19,786] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:25,648] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:15:27,061] loop.py:380 INFO TIMING step=2190: step=4.117s data=1.882s compute=4.088s [fwd=2.907s bwd=1.167s opt=0.003s] dispatch=3.239s [2026-10-18 18:15:27,061] loop.py:45 INFO TRAIN_STEP step=2190 loss=1.971163 Training: 12%|█▏ | 2190/18018 [4:45:13<34:28:24, 7.84s/step, loss=1.971, acc=0.599, acc_len=1.38, thru=15.7, I=19.6, T=31.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:27,169] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:31,915] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:15:33,167] loop.py:45 INFO TRAIN_STEP step=2191 loss=1.861990 Training: 12%|█▏ | 2191/18018 [4:45:19<32:10:58, 7.32s/step, loss=1.862, acc=0.613, acc_len=1.43, thru=19.2, I=25.7, T=33.0, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:15:33,162] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2190: forward=2699.6ms backward=1129.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:33,425] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:38,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:15:39,083] loop.py:45 INFO TRAIN_STEP step=2192 loss=1.901819 Training: 12%|█▏ | 2192/18018 [4:45:25<30:19:29, 6.90s/step, loss=1.902, acc=0.608, acc_len=1.41, thru=20.2, I=26.2, T=34.6, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:39,522] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:46,529] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:15:47,212] loop.py:45 INFO TRAIN_STEP step=2193 loss=1.937537 Training: 12%|█▏ | 2193/18018 [4:45:33<31:56:47, 7.27s/step, loss=1.938, acc=0.609, acc_len=1.40, thru=15.6, I=20.4, T=26.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:15:47,533] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:15:54,977] loop.py:45 INFO TRAIN_STEP step=2194 loss=1.938905 Training: 12%|█▏ | 2194/18018 [4:45:41<32:36:14, 7.42s/step, loss=1.939, acc=0.603, acc_len=1.39, thru=14.6, I=17.5, T=28.1, wait=3.2s, pool=60, epoch=1/2][2026-10-18 18:16:03,534] loop.py:380 INFO TIMING step=2195: step=4.518s data=2.082s compute=4.477s [fwd=3.140s bwd=1.322s opt=0.003s] dispatch=4.029s [2026-10-18 18:16:03,535] loop.py:45 INFO TRAIN_STEP step=2195 loss=1.926415 Training: 12%|█▏ | 2195/18018 [4:45:49<34:06:09, 7.76s/step, loss=1.926, acc=0.607, acc_len=1.40, thru=15.2, I=14.7, T=28.3, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:16:13,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:16:16,115] loop.py:45 INFO TRAIN_STEP step=2196 loss=1.984464 Training: 12%|█▏ | 2196/18018 [4:46:02<40:27:28, 9.21s/step, loss=1.984, acc=0.595, acc_len=1.37, thru=10.2, I=12.7, T=24.3, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:16:16,167] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=48/64 (TrainerActor pid=9310) [2026-10-18 18:16:16,109] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2195: forward=3604.9ms backward=1616.9ms [2026-10-18 18:16:21,552] loop.py:45 INFO TRAIN_STEP step=2197 loss=1.954447 Training: 12%|█▏ | 2197/18018 [4:46:07<35:29:26, 8.08s/step, loss=1.954, acc=0.595, acc_len=1.37, thru=18.0, I=22.7, T=35.7, wait=1.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:16:27,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:16:28,861] loop.py:45 INFO TRAIN_STEP step=2198 loss=1.822245 Training: 12%|█▏ | 2198/18018 [4:46:14<34:28:25, 7.84s/step, loss=1.822, acc=0.620, acc_len=1.44, thru=17.2, I=19.8, T=27.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:16:29,307] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:16:35,535] loop.py:45 INFO TRAIN_STEP step=2199 loss=1.703507 Training: 12%|█▏ | 2199/18018 [4:46:21<32:55:50, 7.49s/step, loss=1.704, acc=0.641, acc_len=1.51, thru=16.6, I=20.5, T=28.4, wait=2.2s, pool=60, epoch=1/2][2026-10-18 18:16:43,493] loop.py:380 INFO TIMING step=2200: step=3.875s data=1.797s compute=3.824s [fwd=2.652s bwd=1.157s opt=0.003s] dispatch=4.072s [2026-10-18 18:16:43,493] loop.py:45 INFO TRAIN_STEP step=2200 loss=1.788980 Training: 12%|█▏ | 2200/18018 [4:46:29<33:32:13, 7.63s/step, loss=1.789, acc=0.624, acc_len=1.46, thru=15.6, I=14.0, T=33.0, wait=4.1s, pool=56, epoch=1/2][2026-10-18 18:16:50,794] loop.py:45 INFO TRAIN_STEP step=2201 loss=1.962894 (TrainerActor pid=9310) [2026-10-18 18:16:50,789] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2200: forward=3069.7ms backward=1327.8ms Training: 12%|█▏ | 2201/18018 [4:46:36<33:06:06, 7.53s/step, loss=1.963, acc=0.604, acc_len=1.39, thru=16.9, I=19.2, T=28.7, wait=2.8s, pool=60, epoch=1/2][2026-10-18 18:16:59,002] loop.py:45 INFO TRAIN_STEP step=2202 loss=1.929198 Training: 12%|█▏ | 2202/18018 [4:46:45<33:59:08, 7.74s/step, loss=1.929, acc=0.605, acc_len=1.40, thru=14.1, I=13.5, T=29.7, wait=3.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:07,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:17:09,362] loop.py:45 INFO TRAIN_STEP step=2203 loss=2.180593 Training: 12%|█▏ | 2203/18018 [4:46:55<37:26:32, 8.52s/step, loss=2.181, acc=0.569, acc_len=1.28, thru=13.8, I=15.3, T=28.5, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:09,558] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:17:17,667] loop.py:45 INFO TRAIN_STEP step=2204 loss=1.890348 Training: 12%|█▏ | 2204/18018 [4:47:03<37:08:57, 8.46s/step, loss=1.890, acc=0.608, acc_len=1.41, thru=15.3, I=16.2, T=33.9, wait=4.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:25,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:17:26,378] loop.py:380 INFO TIMING step=2205: step=4.087s data=1.823s compute=4.057s [fwd=2.943s bwd=1.099s opt=0.003s] dispatch=4.615s [2026-10-18 18:17:26,378] loop.py:45 INFO TRAIN_STEP step=2205 loss=1.877805 Training: 12%|█▏ | 2205/18018 [4:47:12<37:29:07, 8.53s/step, loss=1.878, acc=0.608, acc_len=1.40, thru=14.9, I=15.1, T=31.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:26,785] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:17:33,530] loop.py:45 INFO TRAIN_STEP step=2206 loss=1.962226 Training: 12%|█▏ | 2206/18018 [4:47:19<35:39:35, 8.12s/step, loss=1.962, acc=0.592, acc_len=1.37, thru=15.2, I=18.1, T=31.6, wait=3.1s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:17:33,525] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2205: forward=2831.4ms backward=1160.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:42,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:17:43,240] loop.py:45 INFO TRAIN_STEP step=2207 loss=1.896461 Training: 12%|█▏ | 2207/18018 [4:47:29<37:45:14, 8.60s/step, loss=1.896, acc=0.606, acc_len=1.40, thru=15.7, I=14.9, T=31.8, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:43,392] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:48,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:17:49,355] loop.py:45 INFO TRAIN_STEP step=2208 loss=1.859587 Training: 12%|█▏ | 2208/18018 [4:47:35<34:28:51, 7.85s/step, loss=1.860, acc=0.617, acc_len=1.43, thru=19.4, I=22.6, T=31.3, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:49,548] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:55,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:17:56,387] loop.py:45 INFO TRAIN_STEP step=2209 loss=1.968395 Training: 12%|█▏ | 2209/18018 [4:47:42<33:24:12, 7.61s/step, loss=1.968, acc=0.599, acc_len=1.38, thru=17.4, I=21.8, T=29.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:17:56,442] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:02,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:18:03,624] loop.py:380 INFO TIMING step=2210: step=4.361s data=2.053s compute=4.328s [fwd=3.105s bwd=1.208s opt=0.003s] dispatch=2.865s [2026-10-18 18:18:03,624] loop.py:45 INFO TRAIN_STEP step=2210 loss=1.935477 Training: 12%|█▏ | 2210/18018 [4:47:49<32:54:54, 7.50s/step, loss=1.935, acc=0.600, acc_len=1.39, thru=18.6, I=18.4, T=29.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:03,659] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:10,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:18:12,116] loop.py:45 INFO TRAIN_STEP step=2211 loss=2.018858 (TrainerActor pid=9310) [2026-10-18 18:18:12,110] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2210: forward=3190.5ms backward=1442.7ms Training: 12%|█▏ | 2211/18018 [4:47:58<34:13:20, 7.79s/step, loss=2.019, acc=0.588, acc_len=1.35, thru=14.2, I=18.0, T=27.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:12,492] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:18:19,463] loop.py:45 INFO TRAIN_STEP step=2212 loss=1.949385 Training: 12%|█▏ | 2212/18018 [4:48:05<33:38:01, 7.66s/step, loss=1.949, acc=0.603, acc_len=1.39, thru=13.0, I=20.9, T=28.1, wait=2.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:27,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:18:28,122] loop.py:45 INFO TRAIN_STEP step=2213 loss=2.007229 Training: 12%|█▏ | 2213/18018 [4:48:14<34:56:38, 7.96s/step, loss=2.007, acc=0.590, acc_len=1.35, thru=16.3, I=16.6, T=27.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:28,527] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:18:36,186] loop.py:45 INFO TRAIN_STEP step=2214 loss=1.993397 Training: 12%|█▏ | 2214/18018 [4:48:22<35:04:46, 7.99s/step, loss=1.993, acc=0.594, acc_len=1.37, thru=13.2, I=16.0, T=33.5, wait=4.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:43,122] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:18:44,040] loop.py:380 INFO TIMING step=2215: step=4.139s data=1.828s compute=4.118s [fwd=2.879s bwd=1.223s opt=0.003s] dispatch=3.706s [2026-10-18 18:18:44,040] loop.py:45 INFO TRAIN_STEP step=2215 loss=1.842931 Training: 12%|█▏ | 2215/18018 [4:48:30<34:53:48, 7.95s/step, loss=1.843, acc=0.621, acc_len=1.44, thru=16.4, I=17.9, T=30.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:44,139] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:50,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:18:50,860] loop.py:45 INFO TRAIN_STEP step=2216 loss=1.761586 Training: 12%|█▏ | 2216/18018 [4:48:36<33:24:35, 7.61s/step, loss=1.762, acc=0.629, acc_len=1.47, thru=17.0, I=21.0, T=31.2, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:18:50,856] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2215: forward=2883.9ms backward=1180.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:51,274] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:57,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:18:58,502] loop.py:45 INFO TRAIN_STEP step=2217 loss=1.941315 Training: 12%|█▏ | 2217/18018 [4:48:44<33:26:55, 7.62s/step, loss=1.941, acc=0.602, acc_len=1.39, thru=16.0, I=19.1, T=29.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:18:58,747] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:04,122] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:19:05,280] loop.py:45 INFO TRAIN_STEP step=2218 loss=1.852648 Training: 12%|█▏ | 2218/18018 [4:48:51<32:19:55, 7.37s/step, loss=1.853, acc=0.617, acc_len=1.43, thru=16.4, I=21.5, T=34.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:05,626] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:12,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:19:15,098] loop.py:45 INFO TRAIN_STEP step=2219 loss=1.951824 Training: 12%|█▏ | 2219/18018 [4:49:01<35:33:40, 8.10s/step, loss=1.952, acc=0.602, acc_len=1.39, thru=13.5, I=17.4, T=25.4, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:15,394] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 18:19:21,365] loop.py:380 INFO TIMING step=2220: step=3.793s data=1.673s compute=3.768s [fwd=2.700s bwd=1.053s opt=0.003s] dispatch=2.464s [2026-10-18 18:19:21,365] loop.py:45 INFO TRAIN_STEP step=2220 loss=1.802417 Training: 12%|█▏ | 2220/18018 [4:49:07<33:08:32, 7.55s/step, loss=1.802, acc=0.623, acc_len=1.46, thru=14.6, I=21.6, T=33.7, wait=2.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:29,307] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:19:30,477] loop.py:45 INFO TRAIN_STEP step=2221 loss=1.884920 Training: 12%|█▏ | 2221/18018 [4:49:16<35:11:28, 8.02s/step, loss=1.885, acc=0.609, acc_len=1.41, thru=16.1, I=16.3, T=31.7, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:19:30,472] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2220: forward=2876.1ms backward=1113.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:30,818] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:36,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:19:38,298] loop.py:45 INFO TRAIN_STEP step=2222 loss=1.961066 Training: 12%|█▏ | 2222/18018 [4:49:24<34:55:47, 7.96s/step, loss=1.961, acc=0.596, acc_len=1.37, thru=16.0, I=21.6, T=29.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:38,397] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 18:19:46,868] loop.py:45 INFO TRAIN_STEP step=2223 loss=1.872572 Training: 12%|█▏ | 2223/18018 [4:49:32<35:43:41, 8.14s/step, loss=1.873, acc=0.615, acc_len=1.43, thru=12.0, I=14.6, T=26.5, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:53,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:19:55,163] loop.py:45 INFO TRAIN_STEP step=2224 loss=1.812367 Training: 12%|█▏ | 2224/18018 [4:49:41<35:55:27, 8.19s/step, loss=1.812, acc=0.619, acc_len=1.44, thru=15.6, I=15.9, T=32.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:19:55,430] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:20:01,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:20:02,598] loop.py:380 INFO TIMING step=2225: step=4.587s data=2.258s compute=4.549s [fwd=3.102s bwd=1.433s opt=0.003s] dispatch=2.839s [2026-10-18 18:20:02,598] loop.py:45 INFO TRAIN_STEP step=2225 loss=1.942934 Training: 12%|█▏ | 2225/18018 [4:49:48<34:55:47, 7.96s/step, loss=1.943, acc=0.603, acc_len=1.40, thru=16.5, I=20.3, T=27.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:20:02,962] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:20:08,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:20:10,033] loop.py:45 INFO TRAIN_STEP step=2226 loss=1.900543 Training: 12%|█▏ | 2226/18018 [4:49:56<34:14:04, 7.80s/step, loss=1.901, acc=0.613, acc_len=1.41, thru=14.9, I=19.6, T=33.0, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:20:10,029] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2225: forward=2657.6ms backward=1172.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:20:10,330] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:20:16,767] loop.py:45 INFO TRAIN_STEP step=2227 loss=1.779546 Training: 12%|█▏ | 2227/18018 [4:50:02<32:49:19, 7.48s/step, loss=1.780, acc=0.624, acc_len=1.46, thru=17.2, I=19.7, T=30.6, wait=2.5s, pool=56, epoch=1/2][2026-10-18 18:20:24,831] loop.py:45 INFO TRAIN_STEP step=2228 loss=1.846035 Training: 12%|█▏ | 2228/18018 [4:50:10<33:35:09, 7.66s/step, loss=1.846, acc=0.617, acc_len=1.42, thru=13.8, I=14.4, T=31.6, wait=4.0s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:20:33,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:20:34,789] loop.py:45 INFO TRAIN_STEP step=2229 loss=1.763028 Training: 12%|█▏ | 2229/18018 [4:50:20<36:36:39, 8.35s/step, loss=1.763, acc=0.630, acc_len=1.47, thru=16.2, I=15.4, T=26.5, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:20:35,087] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:20:40,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:20:41,804] loop.py:380 INFO TIMING step=2230: step=4.184s data=1.913s compute=4.148s [fwd=2.885s bwd=1.251s opt=0.003s] dispatch=2.822s [2026-10-18 18:20:41,804] loop.py:45 INFO TRAIN_STEP step=2230 loss=1.793522 Training: 12%|█▏ | 2230/18018 [4:50:27<34:51:22, 7.95s/step, loss=1.794, acc=0.623, acc_len=1.45, thru=16.5, I=20.0, T=30.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:20:42,193] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:20:48,792] loop.py:45 INFO TRAIN_STEP step=2231 loss=2.110553 Training: 12%|█▏ | 2231/18018 [4:50:34<33:35:30, 7.66s/step, loss=2.111, acc=0.575, acc_len=1.31, thru=16.0, I=19.8, T=32.4, wait=3.0s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:20:48,788] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2230: forward=2792.6ms backward=1106.5ms [2026-10-18 18:20:57,408] loop.py:45 INFO TRAIN_STEP step=2232 loss=1.909310 Training: 12%|█▏ | 2232/18018 [4:50:43<34:50:54, 7.95s/step, loss=1.909, acc=0.607, acc_len=1.39, thru=13.3, I=13.9, T=34.7, wait=4.9s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:05,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:21:06,966] loop.py:45 INFO TRAIN_STEP step=2233 loss=2.038159 Training: 12%|█▏ | 2233/18018 [4:50:53<36:57:54, 8.43s/step, loss=2.038, acc=0.591, acc_len=1.36, thru=16.1, I=17.0, T=23.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:07,227] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:21:14,078] loop.py:45 INFO TRAIN_STEP step=2234 loss=1.917049 Training: 12%|█▏ | 2234/18018 [4:51:00<35:13:33, 8.03s/step, loss=1.917, acc=0.599, acc_len=1.38, thru=15.2, I=16.7, T=29.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:14,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:14,584] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 18:21:21,080] loop.py:380 INFO TIMING step=2235: step=4.118s data=1.927s compute=4.082s [fwd=2.890s bwd=1.178s opt=0.003s] dispatch=2.875s [2026-10-18 18:21:21,080] loop.py:45 INFO TRAIN_STEP step=2235 loss=1.930958 Training: 12%|█▏ | 2235/18018 [4:51:07<33:52:01, 7.72s/step, loss=1.931, acc=0.605, acc_len=1.39, thru=16.4, I=16.7, T=31.1, wait=2.9s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:29,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:21:30,022] loop.py:45 INFO TRAIN_STEP step=2236 loss=2.101261 Training: 12%|█▏ | 2236/18018 [4:51:16<35:27:56, 8.09s/step, loss=2.101, acc=0.572, acc_len=1.30, thru=16.4, I=14.9, T=31.4, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:21:30,017] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2235: forward=2784.2ms backward=1226.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:30,260] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:37,345] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:21:37,519] loop.py:45 INFO TRAIN_STEP step=2237 loss=2.044714 Training: 12%|█▏ | 2237/18018 [4:51:23<34:40:57, 7.91s/step, loss=2.045, acc=0.584, acc_len=1.34, thru=16.5, I=17.1, T=30.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:37,848] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:44,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:21:47,195] loop.py:45 INFO TRAIN_STEP step=2238 loss=2.030977 Training: 12%|█▏ | 2238/18018 [4:51:33<36:59:55, 8.44s/step, loss=2.031, acc=0.587, acc_len=1.34, thru=13.4, I=16.8, T=21.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:47,651] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:53,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:21:55,272] loop.py:45 INFO TRAIN_STEP step=2239 loss=1.990005 Training: 12%|█▏ | 2239/18018 [4:51:41<36:31:19, 8.33s/step, loss=1.990, acc=0.596, acc_len=1.37, thru=16.8, I=16.7, T=31.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:21:55,458] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:02,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:22:03,805] loop.py:380 INFO TIMING step=2240: step=4.757s data=2.524s compute=4.723s [fwd=3.483s bwd=1.225s opt=0.003s] dispatch=3.766s [2026-10-18 18:22:03,805] loop.py:45 INFO TRAIN_STEP step=2240 loss=1.808298 Training: 12%|█▏ | 2240/18018 [4:51:49<36:46:52, 8.39s/step, loss=1.808, acc=0.623, acc_len=1.46, thru=13.4, I=17.6, T=26.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:03,978] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:09,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:22:10,773] loop.py:45 INFO TRAIN_STEP step=2241 loss=2.069943 Training: 12%|█▏ | 2241/18018 [4:51:56<34:54:23, 7.96s/step, loss=2.070, acc=0.585, acc_len=1.33, thru=16.4, I=21.2, T=26.3, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:22:10,768] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2240: forward=3512.7ms backward=1296.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:10,849] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 18:22:17,820] loop.py:45 INFO TRAIN_STEP step=2242 loss=1.964067 Training: 12%|█▏ | 2242/18018 [4:52:03<33:41:47, 7.69s/step, loss=1.964, acc=0.601, acc_len=1.38, thru=13.7, I=17.0, T=29.4, wait=2.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:26,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:22:26,854] loop.py:45 INFO TRAIN_STEP step=2243 loss=1.957727 Training: 12%|█▏ | 2243/18018 [4:52:12<35:27:56, 8.09s/step, loss=1.958, acc=0.604, acc_len=1.38, thru=16.1, I=15.0, T=31.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:27,088] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:31,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:22:33,276] loop.py:45 INFO TRAIN_STEP step=2244 loss=1.851794 Training: 12%|█▏ | 2244/18018 [4:52:19<33:15:53, 7.59s/step, loss=1.852, acc=0.615, acc_len=1.43, thru=18.3, I=24.0, T=29.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:33,311] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:41,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:22:42,007] loop.py:380 INFO TIMING step=2245: step=4.865s data=2.181s compute=4.829s [fwd=3.475s bwd=1.338s opt=0.003s] dispatch=3.856s [2026-10-18 18:22:42,007] loop.py:45 INFO TRAIN_STEP step=2245 loss=1.781473 Training: 12%|█▏ | 2245/18018 [4:52:28<34:45:40, 7.93s/step, loss=1.781, acc=0.626, acc_len=1.46, thru=14.7, I=15.2, T=26.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:42,042] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:47,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:22:47,948] loop.py:45 INFO TRAIN_STEP step=2246 loss=1.847628 Training: 12%|█▏ | 2246/18018 [4:52:34<32:08:13, 7.34s/step, loss=1.848, acc=0.618, acc_len=1.43, thru=18.2, I=20.6, T=30.6, wait=1.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:22:47,943] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2245: forward=2841.7ms backward=1286.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:48,156] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:56,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:22:56,809] loop.py:45 INFO TRAIN_STEP step=2247 loss=1.745072 Training: 12%|█▏ | 2247/18018 [4:52:42<34:08:17, 7.79s/step, loss=1.745, acc=0.634, acc_len=1.48, thru=13.7, I=15.5, T=30.0, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:22:57,308] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:04,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:23:05,352] loop.py:45 INFO TRAIN_STEP step=2248 loss=1.893653 Training: 12%|█▏ | 2248/18018 [4:52:51<35:07:34, 8.02s/step, loss=1.894, acc=0.613, acc_len=1.42, thru=15.1, I=17.2, T=31.8, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:05,393] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:11,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:23:13,079] loop.py:45 INFO TRAIN_STEP step=2249 loss=1.965314 Training: 12%|█▏ | 2249/18018 [4:52:59<34:44:25, 7.93s/step, loss=1.965, acc=0.603, acc_len=1.38, thru=16.9, I=21.2, T=32.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:13,434] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:20,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:23:20,409] loop.py:380 INFO TIMING step=2250: step=4.733s data=2.264s compute=4.701s [fwd=3.269s bwd=1.417s opt=0.003s] dispatch=2.586s [2026-10-18 18:23:20,409] loop.py:45 INFO TRAIN_STEP step=2250 loss=1.828853 Training: 12%|█▏ | 2250/18018 [4:53:06<33:56:55, 7.75s/step, loss=1.829, acc=0.614, acc_len=1.43, thru=16.8, I=20.0, T=27.0, wait=2.6s, pool=64, epoch=1/2][2026-10-18 18:23:20,414] loop.py:419 INFO Saving checkpoint at step 2250... (TrainerActor pid=9310) [2026-10-18 18:23:24,222] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0002251 (TrainerActor pid=9310) [2026-10-18 18:23:24,231] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0001501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:25,101] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:25,603] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:30,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 18:23:24,223] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0002251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 18:23:25,207] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0001501 [2026-10-18 18:23:32,091] loop.py:45 INFO TRAIN_STEP step=2251 loss=1.811506 Training: 12%|█▏ | 2251/18018 [4:53:18<39:06:43, 8.93s/step, loss=1.812, acc=0.621, acc_len=1.45, thru=19.7, I=27.6, T=27.3, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:23:32,085] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2250: forward=3379.6ms backward=1259.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:32,227] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:38,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:23:38,289] loop.py:45 INFO TRAIN_STEP step=2252 loss=1.845414 Training: 12%|█▏ | 2252/18018 [4:53:24<35:30:58, 8.11s/step, loss=1.845, acc=0.618, acc_len=1.43, thru=17.4, I=21.8, T=36.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:38,650] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:44,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:23:45,541] loop.py:45 INFO TRAIN_STEP step=2253 loss=2.005047 Training: 13%|█▎ | 2253/18018 [4:53:31<34:23:25, 7.85s/step, loss=2.005, acc=0.589, acc_len=1.35, thru=18.1, I=20.8, T=33.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:45,803] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:54,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:23:54,879] loop.py:45 INFO TRAIN_STEP step=2254 loss=1.914816 Training: 13%|█▎ | 2254/18018 [4:53:40<36:20:16, 8.30s/step, loss=1.915, acc=0.606, acc_len=1.40, thru=14.1, I=15.2, T=29.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:54,956] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:23:59,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:24:01,212] loop.py:380 INFO TIMING step=2255: step=3.838s data=1.725s compute=3.804s [fwd=2.753s bwd=1.037s opt=0.003s] dispatch=2.485s [2026-10-18 18:24:01,212] loop.py:45 INFO TRAIN_STEP step=2255 loss=1.975005 Training: 13%|█▎ | 2255/18018 [4:53:47<33:45:09, 7.71s/step, loss=1.975, acc=0.596, acc_len=1.37, thru=18.4, I=21.9, T=33.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:01,683] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:06,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:24:07,544] loop.py:45 INFO TRAIN_STEP step=2256 loss=1.892051 Training: 13%|█▎ | 2256/18018 [4:53:53<31:56:29, 7.30s/step, loss=1.892, acc=0.611, acc_len=1.42, thru=18.8, I=26.1, T=32.7, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:24:07,539] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2255: forward=2791.0ms backward=1076.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:07,740] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:24:13,215] loop.py:45 INFO TRAIN_STEP step=2257 loss=1.685205 Training: 13%|█▎ | 2257/18018 [4:53:59<29:48:25, 6.81s/step, loss=1.685, acc=0.642, acc_len=1.51, thru=20.1, I=23.7, T=36.7, wait=2.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:20,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:24:20,722] loop.py:45 INFO TRAIN_STEP step=2258 loss=1.865903 Training: 13%|█▎ | 2258/18018 [4:54:06<30:43:23, 7.02s/step, loss=1.866, acc=0.617, acc_len=1.43, thru=17.8, I=18.8, T=30.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:21,129] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:25,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:24:27,341] loop.py:45 INFO TRAIN_STEP step=2259 loss=1.825283 Training: 13%|█▎ | 2259/18018 [4:54:13<30:11:56, 6.90s/step, loss=1.825, acc=0.617, acc_len=1.44, thru=18.1, I=21.4, T=30.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:27,485] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:32,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:24:33,178] loop.py:380 INFO TIMING step=2260: step=4.013s data=1.866s compute=3.981s [fwd=2.808s bwd=1.156s opt=0.003s] dispatch=1.814s [2026-10-18 18:24:33,179] loop.py:45 INFO TRAIN_STEP step=2260 loss=1.692576 Training: 13%|█▎ | 2260/18018 [4:54:19<28:48:03, 6.58s/step, loss=1.693, acc=0.634, acc_len=1.49, thru=19.6, I=26.6, T=31.9, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:33,611] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:24:40,947] loop.py:45 INFO TRAIN_STEP step=2261 loss=1.956945 Training: 13%|█▎ | 2261/18018 [4:54:27<30:21:36, 6.94s/step, loss=1.957, acc=0.599, acc_len=1.38, thru=14.6, I=17.8, T=34.1, wait=4.0s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:24:40,943] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2260: forward=2631.9ms backward=1072.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:47,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:24:49,138] loop.py:45 INFO TRAIN_STEP step=2262 loss=1.947673 Training: 13%|█▎ | 2262/18018 [4:54:35<32:00:29, 7.31s/step, loss=1.948, acc=0.600, acc_len=1.39, thru=17.0, I=18.3, T=32.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:49,323] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:55,136] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:24:56,458] loop.py:45 INFO TRAIN_STEP step=2263 loss=1.790106 Training: 13%|█▎ | 2263/18018 [4:54:42<32:00:50, 7.32s/step, loss=1.790, acc=0.621, acc_len=1.45, thru=16.9, I=23.0, T=28.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:24:56,645] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:25:04,544] loop.py:45 INFO TRAIN_STEP step=2264 loss=1.875475 Training: 13%|█▎ | 2264/18018 [4:54:50<33:01:21, 7.55s/step, loss=1.875, acc=0.613, acc_len=1.42, thru=13.2, I=15.6, T=34.0, wait=4.3s, pool=56, epoch=1/2][2026-10-18 18:25:14,250] loop.py:380 INFO TIMING step=2265: step=4.656s data=2.146s compute=4.627s [fwd=3.246s bwd=1.366s opt=0.003s] dispatch=5.041s [2026-10-18 18:25:14,251] loop.py:45 INFO TRAIN_STEP step=2265 loss=1.800571 Training: 13%|█▎ | 2265/18018 [4:55:00<35:51:40, 8.20s/step, loss=1.801, acc=0.621, acc_len=1.45, thru=13.7, I=14.9, T=27.5, wait=5.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:22,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:25:23,401] loop.py:45 INFO TRAIN_STEP step=2266 loss=1.786248 Training: 13%|█▎ | 2266/18018 [4:55:09<37:06:33, 8.48s/step, loss=1.786, acc=0.629, acc_len=1.47, thru=14.1, I=13.8, T=28.1, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:25:23,396] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2265: forward=3062.2ms backward=1410.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:23,608] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:25:31,463] loop.py:45 INFO TRAIN_STEP step=2267 loss=1.844337 Training: 13%|█▎ | 2267/18018 [4:55:17<36:33:20, 8.36s/step, loss=1.844, acc=0.615, acc_len=1.43, thru=13.8, I=15.6, T=31.1, wait=3.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:37,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:25:38,466] loop.py:45 INFO TRAIN_STEP step=2268 loss=1.895683 Training: 13%|█▎ | 2268/18018 [4:55:24<34:46:49, 7.95s/step, loss=1.896, acc=0.606, acc_len=1.40, thru=18.2, I=18.6, T=30.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:38,571] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:43,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:25:44,307] loop.py:45 INFO TRAIN_STEP step=2269 loss=1.947931 Training: 13%|█▎ | 2269/18018 [4:55:30<32:00:30, 7.32s/step, loss=1.948, acc=0.604, acc_len=1.39, thru=19.5, I=24.1, T=34.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:44,334] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:48,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:25:49,946] loop.py:380 INFO TIMING step=2270: step=3.663s data=1.601s compute=3.635s [fwd=2.569s bwd=1.051s opt=0.003s] dispatch=1.967s [2026-10-18 18:25:49,946] loop.py:45 INFO TRAIN_STEP step=2270 loss=1.811414 Training: 13%|█▎ | 2270/18018 [4:55:36<29:48:18, 6.81s/step, loss=1.811, acc=0.619, acc_len=1.44, thru=21.8, I=26.4, T=34.9, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:50,117] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:56,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:25:57,126] loop.py:45 INFO TRAIN_STEP step=2271 loss=1.816470 Training: 13%|█▎ | 2271/18018 [4:55:43<30:17:09, 6.92s/step, loss=1.816, acc=0.621, acc_len=1.45, thru=17.0, I=22.5, T=32.1, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:25:57,121] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2270: forward=2813.0ms backward=1132.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:25:57,422] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:04,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:26:05,184] loop.py:45 INFO TRAIN_STEP step=2272 loss=1.878521 Training: 13%|█▎ | 2272/18018 [4:55:51<31:46:25, 7.26s/step, loss=1.879, acc=0.613, acc_len=1.42, thru=15.3, I=16.7, T=33.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:05,225] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:10,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:26:11,338] loop.py:45 INFO TRAIN_STEP step=2273 loss=1.790413 Training: 13%|█▎ | 2273/18018 [4:55:57<30:18:47, 6.93s/step, loss=1.790, acc=0.620, acc_len=1.45, thru=19.6, I=22.7, T=32.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:11,485] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-18 18:26:18,869] loop.py:45 INFO TRAIN_STEP step=2274 loss=2.006511 Training: 13%|█▎ | 2274/18018 [4:56:04<31:05:47, 7.11s/step, loss=2.007, acc=0.595, acc_len=1.37, thru=15.8, I=18.1, T=32.0, wait=3.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:24,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:26:26,354] loop.py:380 INFO TIMING step=2275: step=4.187s data=1.850s compute=4.157s [fwd=2.987s bwd=1.155s opt=0.003s] dispatch=3.290s [2026-10-18 18:26:26,355] loop.py:45 INFO TRAIN_STEP step=2275 loss=1.911390 Training: 13%|█▎ | 2275/18018 [4:56:12<31:35:21, 7.22s/step, loss=1.911, acc=0.610, acc_len=1.41, thru=16.8, I=18.8, T=30.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:26,446] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:32,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:26:33,009] loop.py:45 INFO TRAIN_STEP step=2276 loss=1.666053 Training: 13%|█▎ | 2276/18018 [4:56:19<30:50:20, 7.05s/step, loss=1.666, acc=0.645, acc_len=1.52, thru=18.0, I=21.6, T=28.7, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:26:33,004] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2275: forward=3070.0ms backward=1353.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:33,429] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:38,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:26:40,222] loop.py:45 INFO TRAIN_STEP step=2277 loss=1.856498 Training: 13%|█▎ | 2277/18018 [4:56:26<31:02:56, 7.10s/step, loss=1.856, acc=0.614, acc_len=1.43, thru=17.1, I=19.9, T=31.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:40,468] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:48,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:26:48,413] loop.py:45 INFO TRAIN_STEP step=2278 loss=2.008539 Training: 13%|█▎ | 2278/18018 [4:56:34<32:28:34, 7.43s/step, loss=2.009, acc=0.592, acc_len=1.36, thru=13.6, I=15.1, T=31.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:48,748] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:53,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:26:55,064] loop.py:45 INFO TRAIN_STEP step=2279 loss=1.869004 Training: 13%|█▎ | 2279/18018 [4:56:41<31:27:22, 7.20s/step, loss=1.869, acc=0.614, acc_len=1.42, thru=18.6, I=20.3, T=32.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:26:55,347] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:27:01,379] loop.py:380 INFO TIMING step=2280: step=3.803s data=1.560s compute=3.773s [fwd=2.637s bwd=1.121s opt=0.003s] dispatch=2.501s [2026-10-18 18:27:01,380] loop.py:45 INFO TRAIN_STEP step=2280 loss=1.943540 Training: 13%|█▎ | 2280/18018 [4:56:47<30:17:51, 6.93s/step, loss=1.944, acc=0.604, acc_len=1.39, thru=18.3, I=21.1, T=33.7, wait=2.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:06,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:27:07,724] loop.py:45 INFO TRAIN_STEP step=2281 loss=1.770982 Training: 13%|█▎ | 2281/18018 [4:56:53<29:31:47, 6.76s/step, loss=1.771, acc=0.628, acc_len=1.47, thru=19.9, I=21.6, T=33.9, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:27:07,719] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2280: forward=2674.0ms backward=1054.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:07,905] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:13,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:27:14,598] loop.py:45 INFO TRAIN_STEP step=2282 loss=1.936852 Training: 13%|█▎ | 2282/18018 [4:57:00<29:40:58, 6.79s/step, loss=1.937, acc=0.603, acc_len=1.39, thru=17.4, I=21.0, T=33.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:14,929] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:27:21,976] loop.py:45 INFO TRAIN_STEP step=2283 loss=1.794889 Training: 13%|█▎ | 2283/18018 [4:57:08<30:27:01, 6.97s/step, loss=1.795, acc=0.625, acc_len=1.46, thru=16.9, I=22.0, T=27.9, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:29,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:27:30,471] loop.py:45 INFO TRAIN_STEP step=2284 loss=1.958272 Training: 13%|█▎ | 2284/18018 [4:57:16<32:27:16, 7.43s/step, loss=1.958, acc=0.598, acc_len=1.37, thru=16.2, I=14.6, T=32.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:30,810] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:35,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:27:36,636] loop.py:380 INFO TIMING step=2285: step=3.613s data=1.539s compute=3.582s [fwd=2.561s bwd=1.005s opt=0.003s] dispatch=2.541s [2026-10-18 18:27:36,636] loop.py:45 INFO TRAIN_STEP step=2285 loss=1.871791 Training: 13%|█▎ | 2285/18018 [4:57:22<30:47:47, 7.05s/step, loss=1.872, acc=0.617, acc_len=1.43, thru=19.9, I=22.6, T=35.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:36,737] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:41,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:27:42,874] loop.py:45 INFO TRAIN_STEP step=2286 loss=2.073481 Training: 13%|█▎ | 2286/18018 [4:57:28<29:44:09, 6.80s/step, loss=2.073, acc=0.580, acc_len=1.31, thru=20.2, I=25.4, T=36.6, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:27:42,869] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2285: forward=2466.0ms backward=986.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:43,250] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:48,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:27:50,899] loop.py:45 INFO TRAIN_STEP step=2287 loss=1.701406 Training: 13%|█▎ | 2287/18018 [4:57:36<31:19:55, 7.17s/step, loss=1.701, acc=0.635, acc_len=1.50, thru=14.4, I=24.2, T=26.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:51,392] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:57,107] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:27:57,710] loop.py:45 INFO TRAIN_STEP step=2288 loss=1.872277 Training: 13%|█▎ | 2288/18018 [4:57:43<30:51:34, 7.06s/step, loss=1.872, acc=0.618, acc_len=1.43, thru=14.8, I=23.2, T=33.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:27:58,112] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:02,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:28:04,658] loop.py:45 INFO TRAIN_STEP step=2289 loss=1.885923 Training: 13%|█▎ | 2289/18018 [4:57:50<30:42:31, 7.03s/step, loss=1.886, acc=0.610, acc_len=1.41, thru=16.4, I=22.8, T=32.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:04,886] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:09,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:28:10,859] loop.py:380 INFO TIMING step=2290: step=4.363s data=2.099s compute=4.336s [fwd=2.960s bwd=1.362s opt=0.003s] dispatch=1.829s [2026-10-18 18:28:10,860] loop.py:45 INFO TRAIN_STEP step=2290 loss=1.839938 Training: 13%|█▎ | 2290/18018 [4:57:56<29:37:17, 6.78s/step, loss=1.840, acc=0.620, acc_len=1.44, thru=19.0, I=31.4, T=29.3, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:11,081] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:17,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:28:18,663] loop.py:45 INFO TRAIN_STEP step=2291 loss=1.931232 Training: 13%|█▎ | 2291/18018 [4:58:04<30:57:40, 7.09s/step, loss=1.931, acc=0.603, acc_len=1.39, thru=14.1, I=19.1, T=30.1, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:28:18,658] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2290: forward=2983.7ms backward=1224.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:18,694] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:23,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:28:25,542] loop.py:45 INFO TRAIN_STEP step=2292 loss=1.861471 Training: 13%|█▎ | 2292/18018 [4:58:11<30:41:22, 7.03s/step, loss=1.861, acc=0.614, acc_len=1.42, thru=18.1, I=23.7, T=27.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:25,715] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:30,974] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:28:32,567] loop.py:45 INFO TRAIN_STEP step=2293 loss=1.958168 Training: 13%|█▎ | 2293/18018 [4:58:18<30:40:56, 7.02s/step, loss=1.958, acc=0.597, acc_len=1.37, thru=15.0, I=22.4, T=29.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:32,991] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:38,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:28:39,539] loop.py:45 INFO TRAIN_STEP step=2294 loss=1.940465 Training: 13%|█▎ | 2294/18018 [4:58:25<30:36:52, 7.01s/step, loss=1.940, acc=0.600, acc_len=1.38, thru=16.6, I=23.4, T=29.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:39,821] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:47,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:28:47,396] loop.py:380 INFO TIMING step=2295: step=3.873s data=1.754s compute=3.847s [fwd=2.783s bwd=1.050s opt=0.003s] dispatch=3.974s [2026-10-18 18:28:47,396] loop.py:45 INFO TRAIN_STEP step=2295 loss=1.895954 Training: 13%|█▎ | 2295/18018 [4:58:33<31:43:23, 7.26s/step, loss=1.896, acc=0.605, acc_len=1.40, thru=15.4, I=17.0, T=33.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:47,659] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:52,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:28:54,874] loop.py:45 INFO TRAIN_STEP step=2296 loss=1.783848 Training: 13%|█▎ | 2296/18018 [4:58:40<32:00:09, 7.33s/step, loss=1.784, acc=0.623, acc_len=1.45, thru=17.1, I=20.2, T=28.2, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:28:54,869] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2295: forward=3192.9ms backward=1310.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:28:54,957] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-18 18:29:02,986] loop.py:45 INFO TRAIN_STEP step=2297 loss=1.781046 Training: 13%|█▎ | 2297/18018 [4:58:49<33:01:34, 7.56s/step, loss=1.781, acc=0.626, acc_len=1.46, thru=15.0, I=17.6, T=33.5, wait=4.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:29:10,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:29:10,775] loop.py:45 INFO TRAIN_STEP step=2298 loss=1.797166 Training: 13%|█▎ | 2298/18018 [4:58:56<33:19:12, 7.63s/step, loss=1.797, acc=0.621, acc_len=1.45, thru=16.7, I=16.2, T=32.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:29:10,873] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:29:16,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:29:16,964] loop.py:45 INFO TRAIN_STEP step=2299 loss=1.866158 Training: 13%|█▎ | 2299/18018 [4:59:03<31:25:48, 7.20s/step, loss=1.866, acc=0.618, acc_len=1.43, thru=19.1, I=20.2, T=32.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:29:17,313] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 18:29:26,095] loop.py:380 INFO TIMING step=2300: step=4.534s data=2.165s compute=4.495s [fwd=3.213s bwd=1.266s opt=0.003s] dispatch=4.589s [2026-10-18 18:29:26,096] loop.py:45 INFO TRAIN_STEP step=2300 loss=1.825930 Training: 13%|█▎ | 2300/18018 [4:59:12<33:57:45, 7.78s/step, loss=1.826, acc=0.624, acc_len=1.46, thru=12.0, I=12.1, T=28.2, wait=4.6s, pool=40, epoch=1/2][2026-10-18 18:29:33,865] loop.py:45 INFO TRAIN_STEP step=2301 loss=1.780901 Training: 13%|█▎ | 2301/18018 [4:59:19<33:56:54, 7.78s/step, loss=1.781, acc=0.624, acc_len=1.46, thru=18.6, I=15.7, T=30.0, wait=3.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:29:33,860] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2300: forward=2981.0ms backward=1243.0ms [2026-10-18 18:29:43,132] loop.py:45 INFO TRAIN_STEP step=2302 loss=1.731886 Training: 13%|█▎ | 2302/18018 [4:59:29<35:53:42, 8.22s/step, loss=1.732, acc=0.637, acc_len=1.50, thru=11.3, I=11.8, T=28.7, wait=4.8s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:29:51,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:29:52,584] loop.py:45 INFO TRAIN_STEP step=2303 loss=1.890666 Training: 13%|█▎ | 2303/18018 [4:59:38<37:30:12, 8.59s/step, loss=1.891, acc=0.610, acc_len=1.41, thru=17.1, I=14.7, T=29.2, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:29:52,879] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:29:58,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:29:59,709] loop.py:45 INFO TRAIN_STEP step=2304 loss=1.835151 Training: 13%|█▎ | 2304/18018 [4:59:45<35:35:04, 8.15s/step, loss=1.835, acc=0.621, acc_len=1.44, thru=17.3, I=21.3, T=31.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:29:59,866] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:30:05,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:30:06,530] loop.py:380 INFO TIMING step=2305: step=4.010s data=1.838s compute=3.983s [fwd=2.845s bwd=1.123s opt=0.003s] dispatch=2.800s [2026-10-18 18:30:06,530] loop.py:45 INFO TRAIN_STEP step=2305 loss=1.772654 Training: 13%|█▎ | 2305/18018 [4:59:52<33:50:10, 7.75s/step, loss=1.773, acc=0.624, acc_len=1.47, thru=17.8, I=20.4, T=31.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:30:06,863] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:30:12,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:30:13,141] loop.py:45 INFO TRAIN_STEP step=2306 loss=2.040479 Training: 13%|█▎ | 2306/18018 [4:59:59<32:20:26, 7.41s/step, loss=2.040, acc=0.582, acc_len=1.33, thru=19.1, I=21.0, T=31.3, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:30:13,136] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2305: forward=2864.2ms backward=1177.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:30:13,169] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 18:30:20,087] loop.py:45 INFO TRAIN_STEP step=2307 loss=1.991633 Training: 13%|█▎ | 2307/18018 [5:00:06<31:43:41, 7.27s/step, loss=1.992, acc=0.596, acc_len=1.37, thru=17.2, I=19.1, T=33.2, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:30:26,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:30:27,727] loop.py:45 INFO TRAIN_STEP step=2308 loss=2.022019 Training: 13%|█▎ | 2308/18018 [5:00:13<32:12:42, 7.38s/step, loss=2.022, acc=0.583, acc_len=1.33, thru=18.1, I=19.1, T=31.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:30:28,195] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:30:33,757] loop.py:45 INFO TRAIN_STEP step=2309 loss=1.683925 Training: 13%|█▎ | 2309/18018 [5:00:19<30:26:37, 6.98s/step, loss=1.684, acc=0.637, acc_len=1.50, thru=19.3, I=22.7, T=32.0, wait=2.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:30:41,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:30:42,721] loop.py:380 INFO TIMING step=2310: step=4.085s data=1.842s compute=4.024s [fwd=2.721s bwd=1.288s opt=0.003s] dispatch=4.869s [2026-10-18 18:30:42,722] loop.py:45 INFO TRAIN_STEP step=2310 loss=1.996812 Training: 13%|█▎ | 2310/18018 [5:00:28<33:02:29, 7.57s/step, loss=1.997, acc=0.592, acc_len=1.36, thru=15.5, I=15.8, T=31.3, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:30:42,852] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-18 18:30:49,954] loop.py:45 INFO TRAIN_STEP step=2311 loss=1.815294 Training: 13%|█▎ | 2311/18018 [5:00:36<32:35:39, 7.47s/step, loss=1.815, acc=0.620, acc_len=1.44, thru=15.2, I=19.3, T=33.2, wait=3.4s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:30:49,949] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2310: forward=2767.5ms backward=1046.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:30:57,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:31:00,029] loop.py:45 INFO TRAIN_STEP step=2312 loss=1.953613 Training: 13%|█▎ | 2312/18018 [5:00:46<36:00:00, 8.25s/step, loss=1.954, acc=0.599, acc_len=1.38, thru=15.7, I=15.1, T=28.5, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:00,458] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:06,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:31:06,466] loop.py:45 INFO TRAIN_STEP step=2313 loss=1.825403 Training: 13%|█▎ | 2313/18018 [5:00:52<33:37:25, 7.71s/step, loss=1.825, acc=0.619, acc_len=1.44, thru=17.0, I=22.0, T=35.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:06,530] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:10,760] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:31:12,457] loop.py:45 INFO TRAIN_STEP step=2314 loss=1.763033 Training: 13%|█▎ | 2314/18018 [5:00:58<31:22:40, 7.19s/step, loss=1.763, acc=0.625, acc_len=1.46, thru=21.2, I=24.7, T=31.9, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:12,773] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:31:20,454] loop.py:380 INFO TIMING step=2315: step=3.916s data=1.747s compute=3.878s [fwd=2.726s bwd=1.136s opt=0.003s] dispatch=4.070s [2026-10-18 18:31:20,454] loop.py:45 INFO TRAIN_STEP step=2315 loss=1.837840 Training: 13%|█▎ | 2315/18018 [5:01:06<32:25:28, 7.43s/step, loss=1.838, acc=0.616, acc_len=1.43, thru=13.5, I=16.1, T=32.7, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:25,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:31:27,328] loop.py:45 INFO TRAIN_STEP step=2316 loss=1.827193 Training: 13%|█▎ | 2316/18018 [5:01:13<31:41:29, 7.27s/step, loss=1.827, acc=0.618, acc_len=1.44, thru=18.1, I=19.8, T=31.4, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:31:27,323] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2315: forward=2862.2ms backward=1157.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:27,653] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:33,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:31:35,325] loop.py:45 INFO TRAIN_STEP step=2317 loss=1.798650 Training: 13%|█▎ | 2317/18018 [5:01:21<32:38:46, 7.49s/step, loss=1.799, acc=0.624, acc_len=1.46, thru=14.1, I=19.2, T=34.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:35,435] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:40,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:31:42,102] loop.py:45 INFO TRAIN_STEP step=2318 loss=1.927614 Training: 13%|█▎ | 2318/18018 [5:01:28<31:42:56, 7.27s/step, loss=1.928, acc=0.604, acc_len=1.40, thru=18.0, I=26.1, T=32.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:42,519] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:48,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:31:49,173] loop.py:45 INFO TRAIN_STEP step=2319 loss=1.904449 Training: 13%|█▎ | 2319/18018 [5:01:35<31:27:12, 7.21s/step, loss=1.904, acc=0.604, acc_len=1.39, thru=15.8, I=23.4, T=33.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:31:49,321] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-18 18:31:57,212] loop.py:380 INFO TIMING step=2320: step=4.353s data=2.078s compute=4.324s [fwd=3.013s bwd=1.295s opt=0.003s] dispatch=3.675s [2026-10-18 18:31:57,213] loop.py:45 INFO TRAIN_STEP step=2320 loss=1.874457 Training: 13%|█▎ | 2320/18018 [5:01:43<32:31:57, 7.46s/step, loss=1.874, acc=0.612, acc_len=1.41, thru=13.7, I=16.1, T=29.4, wait=3.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:04,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:32:06,352] loop.py:45 INFO TRAIN_STEP step=2321 loss=1.927888 Training: 13%|█▎ | 2321/18018 [5:01:52<34:43:30, 7.96s/step, loss=1.928, acc=0.604, acc_len=1.40, thru=16.4, I=16.2, T=28.5, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:32:06,346] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2320: forward=2974.9ms backward=1441.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:06,412] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:12,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:32:12,655] loop.py:45 INFO TRAIN_STEP step=2322 loss=1.953884 Training: 13%|█▎ | 2322/18018 [5:01:58<32:33:10, 7.47s/step, loss=1.954, acc=0.599, acc_len=1.38, thru=18.1, I=20.9, T=32.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:12,735] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:19,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:32:19,927] loop.py:45 INFO TRAIN_STEP step=2323 loss=1.847386 Training: 13%|█▎ | 2323/18018 [5:02:06<32:17:37, 7.41s/step, loss=1.847, acc=0.619, acc_len=1.44, thru=17.4, I=18.6, T=31.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:20,041] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:26,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:32:27,230] loop.py:45 INFO TRAIN_STEP step=2324 loss=1.813345 Training: 13%|█▎ | 2324/18018 [5:02:13<32:09:26, 7.38s/step, loss=1.813, acc=0.620, acc_len=1.44, thru=16.8, I=21.8, T=31.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:27,286] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:35,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:32:36,525] loop.py:380 INFO TIMING step=2325: step=4.192s data=1.944s compute=4.153s [fwd=2.907s bwd=1.231s opt=0.003s] dispatch=5.092s [2026-10-18 18:32:36,525] loop.py:45 INFO TRAIN_STEP step=2325 loss=1.882081 Training: 13%|█▎ | 2325/18018 [5:02:22<34:39:51, 7.95s/step, loss=1.882, acc=0.608, acc_len=1.40, thru=13.9, I=15.2, T=30.5, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:36,679] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:43,357] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:32:44,468] loop.py:45 INFO TRAIN_STEP step=2326 loss=2.046617 Training: 13%|█▎ | 2326/18018 [5:02:30<34:39:03, 7.95s/step, loss=2.047, acc=0.581, acc_len=1.33, thru=14.4, I=19.3, T=30.5, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:32:44,463] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2325: forward=2903.4ms backward=1257.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:44,868] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:49,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:32:51,464] loop.py:45 INFO TRAIN_STEP step=2327 loss=1.856881 Training: 13%|█▎ | 2327/18018 [5:02:37<33:24:06, 7.66s/step, loss=1.857, acc=0.616, acc_len=1.44, thru=18.6, I=23.1, T=29.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:51,833] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:57,141] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:32:58,059] loop.py:45 INFO TRAIN_STEP step=2328 loss=1.843009 Training: 13%|█▎ | 2328/18018 [5:02:44<32:00:11, 7.34s/step, loss=1.843, acc=0.618, acc_len=1.43, thru=16.4, I=22.6, T=30.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:32:58,149] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:06,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:33:07,828] loop.py:45 INFO TRAIN_STEP step=2329 loss=1.886993 Training: 13%|█▎ | 2329/18018 [5:02:53<35:10:14, 8.07s/step, loss=1.887, acc=0.610, acc_len=1.41, thru=13.2, I=14.9, T=29.5, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:08,301] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:13,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:33:14,145] loop.py:380 INFO TIMING step=2330: step=3.703s data=1.684s compute=3.670s [fwd=2.583s bwd=1.071s opt=0.003s] dispatch=2.604s [2026-10-18 18:33:14,145] loop.py:45 INFO TRAIN_STEP step=2330 loss=1.834236 Training: 13%|█▎ | 2330/18018 [5:03:00<32:52:43, 7.54s/step, loss=1.834, acc=0.618, acc_len=1.44, thru=18.6, I=22.6, T=34.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:14,653] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:20,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:33:21,329] loop.py:45 INFO TRAIN_STEP step=2331 loss=1.997361 Training: 13%|█▎ | 2331/18018 [5:03:07<32:24:08, 7.44s/step, loss=1.997, acc=0.592, acc_len=1.36, thru=17.1, I=22.3, T=36.0, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:33:21,324] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2330: forward=2513.2ms backward=997.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:21,807] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:27,883] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:33:28,830] loop.py:45 INFO TRAIN_STEP step=2332 loss=1.860851 Training: 13%|█▎ | 2332/18018 [5:03:14<32:29:05, 7.46s/step, loss=1.861, acc=0.616, acc_len=1.43, thru=16.5, I=20.4, T=33.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:28,894] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:33,328] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:33:34,824] loop.py:45 INFO TRAIN_STEP step=2333 loss=1.903300 Training: 13%|█▎ | 2333/18018 [5:03:20<30:34:29, 7.02s/step, loss=1.903, acc=0.610, acc_len=1.41, thru=20.0, I=24.9, T=33.4, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:34,858] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 18:33:41,316] loop.py:45 INFO TRAIN_STEP step=2334 loss=1.915709 Training: 13%|█▎ | 2334/18018 [5:03:27<29:53:02, 6.86s/step, loss=1.916, acc=0.609, acc_len=1.40, thru=16.2, I=23.7, T=32.9, wait=2.6s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:50,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:33:51,137] loop.py:380 INFO TIMING step=2335: step=4.273s data=1.811s compute=4.241s [fwd=3.029s bwd=1.197s opt=0.003s] dispatch=5.539s [2026-10-18 18:33:51,137] loop.py:45 INFO TRAIN_STEP step=2335 loss=2.041189 Training: 13%|█▎ | 2335/18018 [5:03:37<33:45:03, 7.75s/step, loss=2.041, acc=0.586, acc_len=1.34, thru=16.4, I=14.1, T=30.0, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:51,334] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:55,854] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:33:57,142] loop.py:45 INFO TRAIN_STEP step=2336 loss=1.898950 Training: 13%|█▎ | 2336/18018 [5:03:43<31:28:32, 7.23s/step, loss=1.899, acc=0.606, acc_len=1.40, thru=19.4, I=21.2, T=32.7, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:33:57,137] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2335: forward=2771.6ms backward=1093.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:33:57,360] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:34:04,401] loop.py:45 INFO TRAIN_STEP step=2337 loss=1.782859 Training: 13%|█▎ | 2337/18018 [5:03:50<31:30:58, 7.24s/step, loss=1.783, acc=0.622, acc_len=1.45, thru=16.4, I=21.1, T=30.2, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:11,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:34:12,263] loop.py:45 INFO TRAIN_STEP step=2338 loss=1.870467 Training: 13%|█▎ | 2338/18018 [5:03:58<32:19:53, 7.42s/step, loss=1.870, acc=0.617, acc_len=1.43, thru=15.9, I=15.7, T=32.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:12,737] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:17,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:34:18,747] loop.py:45 INFO TRAIN_STEP step=2339 loss=1.862424 Training: 13%|█▎ | 2339/18018 [5:04:04<31:06:07, 7.14s/step, loss=1.862, acc=0.612, acc_len=1.42, thru=18.6, I=21.8, T=32.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:19,104] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:24,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:34:25,032] loop.py:380 INFO TIMING step=2340: step=3.574s data=1.580s compute=3.545s [fwd=2.545s bwd=0.985s opt=0.003s] dispatch=2.700s [2026-10-18 18:34:25,032] loop.py:45 INFO TRAIN_STEP step=2340 loss=1.820432 Training: 13%|█▎ | 2340/18018 [5:04:11<29:58:59, 6.88s/step, loss=1.820, acc=0.618, acc_len=1.44, thru=19.5, I=24.4, T=35.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:25,172] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:34,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:34:35,786] loop.py:45 INFO TRAIN_STEP step=2341 loss=1.894837 Training: 13%|█▎ | 2341/18018 [5:04:21<35:02:02, 8.05s/step, loss=1.895, acc=0.609, acc_len=1.41, thru=11.5, I=13.8, T=29.5, wait=6.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:34:35,781] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2340: forward=3029.2ms backward=1250.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:36,262] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:34:41,896] loop.py:45 INFO TRAIN_STEP step=2342 loss=1.796472 Training: 13%|█▎ | 2342/18018 [5:04:27<32:30:24, 7.47s/step, loss=1.796, acc=0.621, acc_len=1.45, thru=17.6, I=22.6, T=33.7, wait=2.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:50,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:34:51,857] loop.py:45 INFO TRAIN_STEP step=2343 loss=1.814186 Training: 13%|█▎ | 2343/18018 [5:04:37<35:45:53, 8.21s/step, loss=1.814, acc=0.621, acc_len=1.44, thru=14.3, I=14.5, T=29.4, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:34:51,915] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-18 18:34:58,273] loop.py:45 INFO TRAIN_STEP step=2344 loss=1.837598 Training: 13%|█▎ | 2344/18018 [5:04:44<33:24:59, 7.68s/step, loss=1.838, acc=0.615, acc_len=1.42, thru=16.5, I=21.4, T=28.1, wait=1.9s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:10,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:35:11,061] loop.py:380 INFO TIMING step=2345: step=4.522s data=2.023s compute=4.476s [fwd=3.125s bwd=1.337s opt=0.003s] dispatch=8.253s [2026-10-18 18:35:11,061] loop.py:45 INFO TRAIN_STEP step=2345 loss=1.902182 Training: 13%|█▎ | 2345/18018 [5:04:57<40:05:13, 9.21s/step, loss=1.902, acc=0.607, acc_len=1.40, thru=14.4, I=10.9, T=28.3, wait=8.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:11,305] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:18,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:35:19,624] loop.py:45 INFO TRAIN_STEP step=2346 loss=1.993704 Training: 13%|█▎ | 2346/18018 [5:05:05<39:14:46, 9.02s/step, loss=1.994, acc=0.594, acc_len=1.37, thru=16.1, I=18.7, T=28.4, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:35:19,619] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2345: forward=3055.1ms backward=1409.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:19,690] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:24,306] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:35:25,540] loop.py:45 INFO TRAIN_STEP step=2347 loss=1.719732 Training: 13%|█▎ | 2347/18018 [5:05:11<35:11:39, 8.08s/step, loss=1.720, acc=0.637, acc_len=1.50, thru=18.5, I=23.4, T=32.0, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:25,818] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:32,659] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:35:34,160] loop.py:45 INFO TRAIN_STEP step=2348 loss=1.942824 Training: 13%|█▎ | 2348/18018 [5:05:20<35:53:21, 8.25s/step, loss=1.943, acc=0.604, acc_len=1.39, thru=13.7, I=17.9, T=28.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:34,663] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:40,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:35:41,823] loop.py:45 INFO TRAIN_STEP step=2349 loss=1.894466 Training: 13%|█▎ | 2349/18018 [5:05:27<35:07:46, 8.07s/step, loss=1.894, acc=0.607, acc_len=1.41, thru=16.0, I=22.7, T=31.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:42,016] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:49,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:35:49,216] loop.py:380 INFO TIMING step=2350: step=4.154s data=1.809s compute=4.126s [fwd=2.853s bwd=1.258s opt=0.003s] dispatch=3.228s [2026-10-18 18:35:49,216] loop.py:45 INFO TRAIN_STEP step=2350 loss=1.851694 Training: 13%|█▎ | 2350/18018 [5:05:35<34:14:16, 7.87s/step, loss=1.852, acc=0.612, acc_len=1.42, thru=16.2, I=19.8, T=30.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:49,677] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:55,457] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:35:58,296] loop.py:45 INFO TRAIN_STEP step=2351 loss=1.902507 Training: 13%|█▎ | 2351/18018 [5:05:44<35:49:36, 8.23s/step, loss=1.903, acc=0.604, acc_len=1.38, thru=14.9, I=20.3, T=24.0, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:35:58,290] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2350: forward=3745.0ms backward=1546.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:35:58,482] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:04,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:36:07,320] loop.py:45 INFO TRAIN_STEP step=2352 loss=1.828485 Training: 13%|█▎ | 2352/18018 [5:05:53<36:51:16, 8.47s/step, loss=1.828, acc=0.622, acc_len=1.44, thru=14.5, I=19.6, T=26.1, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:07,352] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:11,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:36:13,789] loop.py:45 INFO TRAIN_STEP step=2353 loss=1.950346 Training: 13%|█▎ | 2353/18018 [5:05:59<34:14:28, 7.87s/step, loss=1.950, acc=0.597, acc_len=1.37, thru=16.1, I=28.5, T=29.3, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:14,041] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:21,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:36:21,631] loop.py:45 INFO TRAIN_STEP step=2354 loss=1.945064 Training: 13%|█▎ | 2354/18018 [5:06:07<34:11:59, 7.86s/step, loss=1.945, acc=0.603, acc_len=1.39, thru=16.9, I=18.0, T=32.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:22,043] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:27,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:36:29,219] loop.py:380 INFO TIMING step=2355: step=4.340s data=2.037s compute=4.305s [fwd=3.057s bwd=1.232s opt=0.003s] dispatch=3.240s [2026-10-18 18:36:29,219] loop.py:45 INFO TRAIN_STEP step=2355 loss=1.929387 Training: 13%|█▎ | 2355/18018 [5:06:15<33:50:41, 7.78s/step, loss=1.929, acc=0.599, acc_len=1.38, thru=16.1, I=20.0, T=29.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:29,469] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:34,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:36:36,087] loop.py:45 INFO TRAIN_STEP step=2356 loss=1.629855 Training: 13%|█▎ | 2356/18018 [5:06:22<32:39:08, 7.51s/step, loss=1.630, acc=0.647, acc_len=1.53, thru=17.3, I=23.7, T=29.4, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:36:36,083] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2355: forward=2986.0ms backward=1326.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:36:36,309] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:36:43,793] loop.py:45 INFO TRAIN_STEP step=2357 loss=1.861088 Training: 13%|█▎ | 2357/18018 [5:06:29<32:54:40, 7.57s/step, loss=1.861, acc=0.616, acc_len=1.43, thru=12.8, I=15.2, T=32.7, wait=3.8s, pool=44, epoch=1/2][2026-10-18 18:36:53,631] loop.py:45 INFO TRAIN_STEP step=2358 loss=1.800377 Training: 13%|█▎ | 2358/18018 [5:06:39<35:53:03, 8.25s/step, loss=1.800, acc=0.624, acc_len=1.46, thru=13.4, I=14.9, T=31.2, wait=5.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:01,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:37:02,041] loop.py:45 INFO TRAIN_STEP step=2359 loss=1.893698 Training: 13%|█▎ | 2359/18018 [5:06:48<36:05:13, 8.30s/step, loss=1.894, acc=0.611, acc_len=1.42, thru=16.0, I=17.4, T=33.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:02,154] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 18:37:09,029] loop.py:380 INFO TIMING step=2360: step=3.994s data=1.876s compute=3.963s [fwd=2.806s bwd=1.143s opt=0.003s] dispatch=2.983s [2026-10-18 18:37:09,030] loop.py:45 INFO TRAIN_STEP step=2360 loss=2.027413 Training: 13%|█▎ | 2360/18018 [5:06:55<34:22:36, 7.90s/step, loss=2.027, acc=0.586, acc_len=1.34, thru=15.3, I=16.9, T=32.0, wait=3.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:17,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:37:18,536] loop.py:45 INFO TRAIN_STEP step=2361 loss=1.804365 Training: 13%|█▎ | 2361/18018 [5:07:04<36:27:53, 8.38s/step, loss=1.804, acc=0.622, acc_len=1.45, thru=16.3, I=17.6, T=30.1, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:37:18,530] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2360: forward=2973.2ms backward=1234.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:18,945] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:24,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:37:26,481] loop.py:45 INFO TRAIN_STEP step=2362 loss=1.946872 Training: 13%|█▎ | 2362/18018 [5:07:12<35:53:30, 8.25s/step, loss=1.947, acc=0.603, acc_len=1.39, thru=15.6, I=22.2, T=25.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:26,644] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:32,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:37:33,105] loop.py:45 INFO TRAIN_STEP step=2363 loss=1.889157 Training: 13%|█▎ | 2363/18018 [5:07:19<33:45:47, 7.76s/step, loss=1.889, acc=0.603, acc_len=1.39, thru=16.0, I=21.4, T=33.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:33,353] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:41,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:37:42,841] loop.py:45 INFO TRAIN_STEP step=2364 loss=1.818795 Training: 13%|█▎ | 2364/18018 [5:07:28<36:20:02, 8.36s/step, loss=1.819, acc=0.621, acc_len=1.44, thru=13.2, I=15.8, T=24.1, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:43,165] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:48,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:37:50,514] loop.py:380 INFO TIMING step=2365: step=5.071s data=2.817s compute=5.023s [fwd=3.704s bwd=1.304s opt=0.003s] dispatch=2.591s [2026-10-18 18:37:50,514] loop.py:45 INFO TRAIN_STEP step=2365 loss=1.816637 Training: 13%|█▎ | 2365/18018 [5:07:36<35:26:30, 8.15s/step, loss=1.817, acc=0.621, acc_len=1.44, thru=14.3, I=22.5, T=25.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:50,788] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:58,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:37:58,924] loop.py:45 INFO TRAIN_STEP step=2366 loss=1.872716 Training: 13%|█▎ | 2366/18018 [5:07:45<35:46:27, 8.23s/step, loss=1.873, acc=0.610, acc_len=1.42, thru=15.7, I=17.9, T=27.7, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:37:58,919] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2365: forward=3426.1ms backward=1150.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:37:59,281] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:04,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:38:06,552] loop.py:45 INFO TRAIN_STEP step=2367 loss=1.767400 Training: 13%|█▎ | 2367/18018 [5:07:52<34:59:16, 8.05s/step, loss=1.767, acc=0.632, acc_len=1.47, thru=16.3, I=21.3, T=29.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:07,054] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:13,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:38:15,004] loop.py:45 INFO TRAIN_STEP step=2368 loss=1.709442 Training: 13%|█▎ | 2368/18018 [5:08:01<35:30:48, 8.17s/step, loss=1.709, acc=0.640, acc_len=1.51, thru=16.1, I=19.5, T=28.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:15,433] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:19,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:38:21,475] loop.py:45 INFO TRAIN_STEP step=2369 loss=1.863888 Training: 13%|█▎ | 2369/18018 [5:08:07<33:17:46, 7.66s/step, loss=1.864, acc=0.609, acc_len=1.41, thru=16.7, I=24.7, T=32.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:21,687] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:38:28,731] loop.py:380 INFO TIMING step=2370: step=4.317s data=1.958s compute=4.284s [fwd=3.020s bwd=1.250s opt=0.003s] dispatch=2.931s [2026-10-18 18:38:28,731] loop.py:45 INFO TRAIN_STEP step=2370 loss=1.984620 Training: 13%|█▎ | 2370/18018 [5:08:14<32:46:17, 7.54s/step, loss=1.985, acc=0.590, acc_len=1.35, thru=15.5, I=22.8, T=29.7, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:36,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:38:37,605] loop.py:45 INFO TRAIN_STEP step=2371 loss=1.867649 Training: 13%|█▎ | 2371/18018 [5:08:23<34:30:25, 7.94s/step, loss=1.868, acc=0.613, acc_len=1.42, thru=14.4, I=15.3, T=31.9, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:38:37,600] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2370: forward=2799.9ms backward=1165.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:37,862] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:44,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:38:44,806] loop.py:45 INFO TRAIN_STEP step=2372 loss=1.746719 Training: 13%|█▎ | 2372/18018 [5:08:30<33:32:22, 7.72s/step, loss=1.747, acc=0.635, acc_len=1.49, thru=16.9, I=18.6, T=31.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:45,197] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:51,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:38:51,972] loop.py:45 INFO TRAIN_STEP step=2373 loss=1.882460 Training: 13%|█▎ | 2373/18018 [5:08:38<32:49:24, 7.55s/step, loss=1.882, acc=0.618, acc_len=1.43, thru=17.8, I=17.9, T=30.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:52,145] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:38:58,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:38:59,977] loop.py:45 INFO TRAIN_STEP step=2374 loss=1.903393 Training: 13%|█▎ | 2374/18018 [5:08:46<33:24:27, 7.69s/step, loss=1.903, acc=0.610, acc_len=1.41, thru=15.2, I=16.7, T=30.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:00,437] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:39:06,835] loop.py:380 INFO TIMING step=2375: step=3.964s data=1.768s compute=3.925s [fwd=2.800s bwd=1.110s opt=0.003s] dispatch=2.885s [2026-10-18 18:39:06,836] loop.py:45 INFO TRAIN_STEP step=2375 loss=1.869648 Training: 13%|█▎ | 2375/18018 [5:08:52<32:19:33, 7.44s/step, loss=1.870, acc=0.611, acc_len=1.41, thru=16.0, I=21.1, T=32.3, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:13,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:39:14,948] loop.py:45 INFO TRAIN_STEP step=2376 loss=1.881172 Training: 13%|█▎ | 2376/18018 [5:09:01<33:12:00, 7.64s/step, loss=1.881, acc=0.612, acc_len=1.41, thru=15.2, I=17.0, T=28.1, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:39:14,943] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2375: forward=3110.3ms backward=1393.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:14,979] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:23,188] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:39:24,310] loop.py:45 INFO TRAIN_STEP step=2377 loss=1.778282 Training: 13%|█▎ | 2377/18018 [5:09:10<35:26:24, 8.16s/step, loss=1.778, acc=0.627, acc_len=1.46, thru=13.8, I=15.2, T=26.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:24,700] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:31,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:39:32,136] loop.py:45 INFO TRAIN_STEP step=2378 loss=1.871196 Training: 13%|█▎ | 2378/18018 [5:09:18<35:00:26, 8.06s/step, loss=1.871, acc=0.616, acc_len=1.44, thru=14.5, I=17.3, T=28.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:32,592] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:37,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:39:38,699] loop.py:45 INFO TRAIN_STEP step=2379 loss=1.849368 Training: 13%|█▎ | 2379/18018 [5:09:24<33:03:36, 7.61s/step, loss=1.849, acc=0.615, acc_len=1.43, thru=18.4, I=22.2, T=29.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:38,759] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:43,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:39:45,493] loop.py:380 INFO TIMING step=2380: step=4.033s data=1.779s compute=3.995s [fwd=2.839s bwd=1.141s opt=0.003s] dispatch=2.751s [2026-10-18 18:39:45,493] loop.py:45 INFO TRAIN_STEP step=2380 loss=1.656472 Training: 13%|█▎ | 2380/18018 [5:09:31<31:59:40, 7.37s/step, loss=1.656, acc=0.642, acc_len=1.51, thru=17.9, I=22.4, T=31.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:39:45,622] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-18 18:39:54,823] loop.py:45 INFO TRAIN_STEP step=2381 loss=1.744312 Training: 13%|█▎ | 2381/18018 [5:09:40<34:32:54, 7.95s/step, loss=1.744, acc=0.634, acc_len=1.48, thru=13.5, I=13.0, T=31.9, wait=5.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:39:54,818] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2380: forward=2842.5ms backward=1127.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:00,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:40:02,839] loop.py:45 INFO TRAIN_STEP step=2382 loss=1.906572 Training: 13%|█▎ | 2382/18018 [5:09:48<34:37:51, 7.97s/step, loss=1.907, acc=0.607, acc_len=1.40, thru=17.2, I=19.1, T=29.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:02,946] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:10,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:40:12,403] loop.py:45 INFO TRAIN_STEP step=2383 loss=1.896760 Training: 13%|█▎ | 2383/18018 [5:09:58<36:42:00, 8.45s/step, loss=1.897, acc=0.606, acc_len=1.40, thru=13.6, I=17.5, T=27.8, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:12,770] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:17,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:40:18,666] loop.py:45 INFO TRAIN_STEP step=2384 loss=1.768846 Training: 13%|█▎ | 2384/18018 [5:10:04<33:50:59, 7.79s/step, loss=1.769, acc=0.628, acc_len=1.47, thru=17.1, I=25.1, T=35.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:18,819] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:26,307] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:40:27,557] loop.py:380 INFO TIMING step=2385: step=4.305s data=2.037s compute=4.258s [fwd=2.964s bwd=1.280s opt=0.003s] dispatch=4.575s [2026-10-18 18:40:27,557] loop.py:45 INFO TRAIN_STEP step=2385 loss=1.767642 Training: 13%|█▎ | 2385/18018 [5:10:13<35:16:22, 8.12s/step, loss=1.768, acc=0.631, acc_len=1.47, thru=13.7, I=18.6, T=29.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:27,816] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:34,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:40:34,961] loop.py:45 INFO TRAIN_STEP step=2386 loss=1.707725 Training: 13%|█▎ | 2386/18018 [5:10:21<34:20:07, 7.91s/step, loss=1.708, acc=0.638, acc_len=1.49, thru=15.6, I=17.8, T=33.2, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:40:34,956] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2385: forward=2708.6ms backward=1089.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:35,351] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:41,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:40:42,143] loop.py:45 INFO TRAIN_STEP step=2387 loss=1.807902 Training: 13%|█▎ | 2387/18018 [5:10:28<33:23:23, 7.69s/step, loss=1.808, acc=0.619, acc_len=1.44, thru=17.4, I=22.5, T=33.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:42,475] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:49,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:40:50,765] loop.py:45 INFO TRAIN_STEP step=2388 loss=1.734879 Training: 13%|█▎ | 2388/18018 [5:10:36<34:36:06, 7.97s/step, loss=1.735, acc=0.636, acc_len=1.50, thru=14.2, I=14.8, T=26.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:50,902] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:58,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:40:58,641] loop.py:45 INFO TRAIN_STEP step=2389 loss=1.836308 Training: 13%|█▎ | 2389/18018 [5:10:44<34:28:37, 7.94s/step, loss=1.836, acc=0.617, acc_len=1.43, thru=15.2, I=18.7, T=31.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:40:58,715] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-18 18:41:05,240] loop.py:380 INFO TIMING step=2390: step=3.669s data=1.596s compute=3.641s [fwd=2.589s bwd=1.037s opt=0.003s] dispatch=2.919s [2026-10-18 18:41:05,240] loop.py:45 INFO TRAIN_STEP step=2390 loss=1.986356 Training: 13%|█▎ | 2390/18018 [5:10:51<32:43:24, 7.54s/step, loss=1.986, acc=0.592, acc_len=1.36, thru=16.8, I=17.8, T=34.9, wait=2.9s, pool=52, epoch=1/2][2026-10-18 18:41:13,437] loop.py:45 INFO TRAIN_STEP step=2391 loss=1.918754 Training: 13%|█▎ | 2391/18018 [5:10:59<33:34:46, 7.74s/step, loss=1.919, acc=0.600, acc_len=1.38, thru=14.4, I=14.1, T=32.4, wait=4.2s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:41:13,433] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2390: forward=2775.1ms backward=1125.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:20,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:41:21,696] loop.py:45 INFO TRAIN_STEP step=2392 loss=2.002410 Training: 13%|█▎ | 2392/18018 [5:11:07<34:15:42, 7.89s/step, loss=2.002, acc=0.594, acc_len=1.36, thru=17.4, I=18.0, T=33.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:21,808] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:27,092] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:41:29,237] loop.py:45 INFO TRAIN_STEP step=2393 loss=1.838897 Training: 13%|█▎ | 2393/18018 [5:11:15<33:48:00, 7.79s/step, loss=1.839, acc=0.623, acc_len=1.45, thru=16.2, I=24.0, T=29.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:29,611] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:34,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:41:36,683] loop.py:45 INFO TRAIN_STEP step=2394 loss=1.911664 Training: 13%|█▎ | 2394/18018 [5:11:22<33:21:01, 7.68s/step, loss=1.912, acc=0.604, acc_len=1.40, thru=14.5, I=24.8, T=28.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:37,151] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 18:41:43,243] loop.py:380 INFO TIMING step=2395: step=4.402s data=2.029s compute=4.372s [fwd=3.018s bwd=1.339s opt=0.003s] dispatch=2.149s [2026-10-18 18:41:43,243] loop.py:45 INFO TRAIN_STEP step=2395 loss=1.770865 Training: 13%|█▎ | 2395/18018 [5:11:29<31:53:05, 7.35s/step, loss=1.771, acc=0.628, acc_len=1.47, thru=15.6, I=21.4, T=29.1, wait=2.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:51,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:41:52,679] loop.py:45 INFO TRAIN_STEP step=2396 loss=1.915512 Training: 13%|█▎ | 2396/18018 [5:11:38<34:36:13, 7.97s/step, loss=1.916, acc=0.606, acc_len=1.40, thru=16.0, I=16.5, T=26.5, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:41:52,674] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2395: forward=3406.6ms backward=1384.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:52,946] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:58,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:41:59,551] loop.py:45 INFO TRAIN_STEP step=2397 loss=1.860238 Training: 13%|█▎ | 2397/18018 [5:11:45<33:10:07, 7.64s/step, loss=1.860, acc=0.612, acc_len=1.42, thru=16.9, I=18.2, T=29.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:41:59,988] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:05,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:42:06,378] loop.py:45 INFO TRAIN_STEP step=2398 loss=1.873595 Training: 13%|█▎ | 2398/18018 [5:11:52<32:06:00, 7.40s/step, loss=1.874, acc=0.614, acc_len=1.42, thru=18.0, I=20.1, T=29.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:06,671] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:11,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:42:13,275] loop.py:45 INFO TRAIN_STEP step=2399 loss=1.882477 Training: 13%|█▎ | 2399/18018 [5:11:59<31:26:45, 7.25s/step, loss=1.882, acc=0.613, acc_len=1.43, thru=17.7, I=23.3, T=30.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:13,517] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:42:19,840] loop.py:380 INFO TIMING step=2400: step=3.664s data=1.646s compute=3.635s [fwd=2.585s bwd=1.036s opt=0.003s] dispatch=2.890s [2026-10-18 18:42:19,840] loop.py:45 INFO TRAIN_STEP step=2400 loss=1.891450 Training: 13%|█▎ | 2400/18018 [5:12:05<30:33:16, 7.04s/step, loss=1.891, acc=0.608, acc_len=1.41, thru=14.5, I=22.3, T=34.9, wait=2.9s, pool=36, epoch=1/2][2026-10-18 18:42:29,013] loop.py:45 INFO TRAIN_STEP step=2401 loss=1.994722 Training: 13%|█▎ | 2401/18018 [5:12:15<33:19:35, 7.68s/step, loss=1.995, acc=0.594, acc_len=1.36, thru=15.6, I=15.0, T=34.4, wait=5.4s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:42:29,007] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2400: forward=2689.1ms backward=994.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:35,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:42:35,821] loop.py:45 INFO TRAIN_STEP step=2402 loss=1.949473 Training: 13%|█▎ | 2402/18018 [5:12:21<32:11:05, 7.42s/step, loss=1.949, acc=0.603, acc_len=1.39, thru=19.0, I=20.5, T=33.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:36,149] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:42,904] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:42:43,732] loop.py:45 INFO TRAIN_STEP step=2403 loss=1.814979 Training: 13%|█▎ | 2403/18018 [5:12:29<32:49:26, 7.57s/step, loss=1.815, acc=0.622, acc_len=1.45, thru=15.7, I=16.9, T=28.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:43,917] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:50,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:42:51,444] loop.py:45 INFO TRAIN_STEP step=2404 loss=1.664296 Training: 13%|█▎ | 2404/18018 [5:12:37<33:00:36, 7.61s/step, loss=1.664, acc=0.643, acc_len=1.52, thru=15.0, I=16.9, T=30.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:51,909] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:42:58,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:42:59,651] loop.py:380 INFO TIMING step=2405: step=3.827s data=1.692s compute=3.794s [fwd=2.775s bwd=1.004s opt=0.003s] dispatch=4.370s [2026-10-18 18:42:59,652] loop.py:45 INFO TRAIN_STEP step=2405 loss=1.712983 Training: 13%|█▎ | 2405/18018 [5:12:45<33:46:53, 7.79s/step, loss=1.713, acc=0.638, acc_len=1.50, thru=15.0, I=19.2, T=33.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:00,121] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:05,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:43:06,581] loop.py:45 INFO TRAIN_STEP step=2406 loss=1.892720 Training: 13%|█▎ | 2406/18018 [5:12:52<32:39:41, 7.53s/step, loss=1.893, acc=0.604, acc_len=1.39, thru=17.1, I=24.7, T=29.4, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:43:06,577] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2405: forward=2974.9ms backward=1320.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:07,075] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:12,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:43:13,370] loop.py:45 INFO TRAIN_STEP step=2407 loss=1.830782 Training: 13%|█▎ | 2407/18018 [5:12:59<31:41:33, 7.31s/step, loss=1.831, acc=0.618, acc_len=1.44, thru=16.6, I=22.6, T=35.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:13,760] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:19,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:43:19,578] loop.py:45 INFO TRAIN_STEP step=2408 loss=1.740643 Training: 13%|█▎ | 2408/18018 [5:13:05<30:15:38, 6.98s/step, loss=1.741, acc=0.632, acc_len=1.48, thru=19.8, I=23.3, T=34.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:19,740] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:25,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:43:27,394] loop.py:45 INFO TRAIN_STEP step=2409 loss=1.905468 Training: 13%|█▎ | 2409/18018 [5:13:13<31:20:55, 7.23s/step, loss=1.905, acc=0.607, acc_len=1.40, thru=16.6, I=20.2, T=27.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:27,790] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:32,441] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:43:33,955] loop.py:380 INFO TIMING step=2410: step=4.213s data=2.009s compute=4.166s [fwd=2.896s bwd=1.255s opt=0.003s] dispatch=2.338s [2026-10-18 18:43:33,955] loop.py:45 INFO TRAIN_STEP step=2410 loss=1.835920 Training: 13%|█▎ | 2410/18018 [5:13:20<30:28:24, 7.03s/step, loss=1.836, acc=0.617, acc_len=1.43, thru=18.6, I=25.5, T=30.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:34,455] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:40,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:43:42,349] loop.py:45 INFO TRAIN_STEP step=2411 loss=1.777207 Training: 13%|█▎ | 2411/18018 [5:13:28<32:15:00, 7.44s/step, loss=1.777, acc=0.627, acc_len=1.46, thru=13.3, I=17.5, T=29.0, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:43:42,344] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2410: forward=3085.0ms backward=1291.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:42,501] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:48,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:43:49,185] loop.py:45 INFO TRAIN_STEP step=2412 loss=1.862174 Training: 13%|█▎ | 2412/18018 [5:13:35<31:27:47, 7.26s/step, loss=1.862, acc=0.613, acc_len=1.42, thru=16.4, I=23.2, T=31.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:43:49,619] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:43:58,594] loop.py:45 INFO TRAIN_STEP step=2413 loss=1.950195 Training: 13%|█▎ | 2413/18018 [5:13:44<34:15:26, 7.90s/step, loss=1.950, acc=0.597, acc_len=1.37, thru=12.0, I=11.7, T=30.0, wait=5.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:05,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:44:06,155] loop.py:45 INFO TRAIN_STEP step=2414 loss=1.901142 Training: 13%|█▎ | 2414/18018 [5:13:52<33:48:41, 7.80s/step, loss=1.901, acc=0.610, acc_len=1.41, thru=17.6, I=16.6, T=33.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:06,500] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:12,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:44:12,998] loop.py:380 INFO TIMING step=2415: step=3.970s data=1.806s compute=3.945s [fwd=2.808s bwd=1.122s opt=0.003s] dispatch=2.863s [2026-10-18 18:44:12,998] loop.py:45 INFO TRAIN_STEP step=2415 loss=1.767177 Training: 13%|█▎ | 2415/18018 [5:13:59<32:33:56, 7.51s/step, loss=1.767, acc=0.627, acc_len=1.47, thru=17.7, I=21.3, T=32.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:13,253] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:19,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:44:20,497] loop.py:45 INFO TRAIN_STEP step=2416 loss=2.094041 Training: 13%|█▎ | 2416/18018 [5:14:06<32:32:40, 7.51s/step, loss=2.094, acc=0.577, acc_len=1.31, thru=16.9, I=18.8, T=31.7, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:44:20,492] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2415: forward=2838.8ms backward=1147.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:20,708] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:26,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:44:26,737] loop.py:45 INFO TRAIN_STEP step=2417 loss=1.806923 Training: 13%|█▎ | 2417/18018 [5:14:12<30:53:20, 7.13s/step, loss=1.807, acc=0.624, acc_len=1.45, thru=19.2, I=22.1, T=34.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:27,058] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 18:44:34,854] loop.py:45 INFO TRAIN_STEP step=2418 loss=1.819666 Training: 13%|█▎ | 2418/18018 [5:14:20<32:10:22, 7.42s/step, loss=1.820, acc=0.620, acc_len=1.45, thru=15.3, I=16.0, T=31.6, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:41,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:44:42,805] loop.py:45 INFO TRAIN_STEP step=2419 loss=1.831589 Training: 13%|█▎ | 2419/18018 [5:14:28<32:51:22, 7.58s/step, loss=1.832, acc=0.619, acc_len=1.43, thru=16.7, I=17.9, T=29.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:42,956] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:44:49,857] loop.py:380 INFO TIMING step=2420: step=4.158s data=1.843s compute=4.126s [fwd=2.930s bwd=1.181s opt=0.003s] dispatch=2.884s [2026-10-18 18:44:49,857] loop.py:45 INFO TRAIN_STEP step=2420 loss=1.791104 Training: 13%|█▎ | 2420/18018 [5:14:35<32:09:52, 7.42s/step, loss=1.791, acc=0.624, acc_len=1.46, thru=16.4, I=18.4, T=30.8, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:57,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:44:58,648] loop.py:45 INFO TRAIN_STEP step=2421 loss=1.892853 Training: 13%|█▎ | 2421/18018 [5:14:44<33:56:30, 7.83s/step, loss=1.893, acc=0.611, acc_len=1.41, thru=16.0, I=16.6, T=27.6, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:44:58,643] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2420: forward=3292.9ms backward=1308.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:44:58,793] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:07,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:45:07,964] loop.py:45 INFO TRAIN_STEP step=2422 loss=1.736953 Training: 13%|█▎ | 2422/18018 [5:14:54<35:51:50, 8.28s/step, loss=1.737, acc=0.632, acc_len=1.48, thru=14.0, I=13.9, T=34.8, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:08,388] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:14,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:45:15,247] loop.py:45 INFO TRAIN_STEP step=2423 loss=1.782489 Training: 13%|█▎ | 2423/18018 [5:15:01<34:34:05, 7.98s/step, loss=1.782, acc=0.624, acc_len=1.45, thru=18.9, I=21.6, T=27.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:15,318] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 18:45:24,714] loop.py:45 INFO TRAIN_STEP step=2424 loss=1.847932 Training: 13%|█▎ | 2424/18018 [5:15:10<36:29:50, 8.43s/step, loss=1.848, acc=0.616, acc_len=1.43, thru=12.0, I=11.7, T=29.4, wait=5.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:35,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:45:36,338] loop.py:380 INFO TIMING step=2425: step=4.687s data=2.246s compute=4.642s [fwd=3.279s bwd=1.347s opt=0.003s] dispatch=6.927s [2026-10-18 18:45:36,338] loop.py:45 INFO TRAIN_STEP step=2425 loss=1.831653 Training: 13%|█▎ | 2425/18018 [5:15:22<40:38:59, 9.38s/step, loss=1.832, acc=0.618, acc_len=1.43, thru=12.9, I=13.3, T=27.3, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:36,747] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:43,009] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:45:44,373] loop.py:45 INFO TRAIN_STEP step=2426 loss=1.840592 Training: 13%|█▎ | 2426/18018 [5:15:30<38:53:47, 8.98s/step, loss=1.841, acc=0.614, acc_len=1.42, thru=16.6, I=18.7, T=30.2, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:45:44,369] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2425: forward=2888.9ms backward=1301.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:44,521] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:49,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:45:50,877] loop.py:45 INFO TRAIN_STEP step=2427 loss=1.855751 Training: 13%|█▎ | 2427/18018 [5:15:36<35:40:35, 8.24s/step, loss=1.856, acc=0.610, acc_len=1.42, thru=18.0, I=22.2, T=35.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:51,149] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:55,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:45:56,538] loop.py:45 INFO TRAIN_STEP step=2428 loss=1.699462 Training: 13%|█▎ | 2428/18018 [5:15:42<32:19:30, 7.46s/step, loss=1.699, acc=0.641, acc_len=1.51, thru=21.3, I=27.2, T=35.2, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:45:56,789] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:05,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:46:06,254] loop.py:45 INFO TRAIN_STEP step=2429 loss=1.965203 Training: 13%|█▎ | 2429/18018 [5:15:52<35:14:53, 8.14s/step, loss=1.965, acc=0.597, acc_len=1.37, thru=13.5, I=13.9, T=30.3, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:06,330] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:11,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:46:13,057] loop.py:380 INFO TIMING step=2430: step=4.155s data=1.917s compute=4.113s [fwd=2.921s bwd=1.176s opt=0.003s] dispatch=2.638s [2026-10-18 18:46:13,057] loop.py:45 INFO TRAIN_STEP step=2430 loss=1.931199 Training: 13%|█▎ | 2430/18018 [5:15:59<33:30:33, 7.74s/step, loss=1.931, acc=0.601, acc_len=1.39, thru=17.8, I=22.9, T=30.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:13,312] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:20,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:46:20,572] loop.py:45 INFO TRAIN_STEP step=2431 loss=1.886059 Training: 13%|█▎ | 2431/18018 [5:16:06<33:12:50, 7.67s/step, loss=1.886, acc=0.608, acc_len=1.40, thru=14.7, I=18.7, T=32.4, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:46:20,567] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2430: forward=2701.2ms backward=1198.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:20,852] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:25,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:46:27,020] loop.py:45 INFO TRAIN_STEP step=2432 loss=1.784659 Training: 13%|█▎ | 2432/18018 [5:16:13<31:37:35, 7.31s/step, loss=1.785, acc=0.627, acc_len=1.46, thru=20.4, I=21.4, T=33.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:27,219] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:46:34,301] loop.py:45 INFO TRAIN_STEP step=2433 loss=1.869383 Training: 14%|█▎ | 2433/18018 [5:16:20<31:35:26, 7.30s/step, loss=1.869, acc=0.614, acc_len=1.42, thru=16.1, I=18.9, T=31.7, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:39,971] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:46:41,321] loop.py:45 INFO TRAIN_STEP step=2434 loss=1.861422 Training: 14%|█▎ | 2434/18018 [5:16:27<31:13:52, 7.21s/step, loss=1.861, acc=0.614, acc_len=1.42, thru=17.6, I=19.9, T=31.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:41,485] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:46,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:46:47,975] loop.py:380 INFO TIMING step=2435: step=4.107s data=1.807s compute=4.078s [fwd=2.824s bwd=1.239s opt=0.003s] dispatch=2.536s [2026-10-18 18:46:47,975] loop.py:45 INFO TRAIN_STEP step=2435 loss=1.863641 Training: 14%|█▎ | 2435/18018 [5:16:34<30:30:01, 7.05s/step, loss=1.864, acc=0.615, acc_len=1.43, thru=18.2, I=26.8, T=31.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:46:48,446] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:46:55,141] loop.py:45 INFO TRAIN_STEP step=2436 loss=1.674683 Training: 14%|█▎ | 2436/18018 [5:16:41<30:39:22, 7.08s/step, loss=1.675, acc=0.645, acc_len=1.52, thru=12.9, I=15.4, T=31.8, wait=3.1s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:46:55,136] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2435: forward=2786.5ms backward=1195.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:05,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:47:06,212] loop.py:45 INFO TRAIN_STEP step=2437 loss=1.930623 Training: 14%|█▎ | 2437/18018 [5:16:52<35:49:52, 8.28s/step, loss=1.931, acc=0.600, acc_len=1.39, thru=14.1, I=14.3, T=29.0, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:06,278] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:12,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:47:13,288] loop.py:45 INFO TRAIN_STEP step=2438 loss=1.940711 Training: 14%|█▎ | 2438/18018 [5:16:59<34:16:02, 7.92s/step, loss=1.941, acc=0.602, acc_len=1.39, thru=16.4, I=18.8, T=31.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:13,324] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:21,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:47:22,492] loop.py:45 INFO TRAIN_STEP step=2439 loss=2.005315 Training: 14%|█▎ | 2439/18018 [5:17:08<35:55:53, 8.30s/step, loss=2.005, acc=0.595, acc_len=1.36, thru=14.0, I=14.7, T=26.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:22,977] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:47:29,608] loop.py:380 INFO TIMING step=2440: step=4.149s data=1.901s compute=4.102s [fwd=2.874s bwd=1.212s opt=0.003s] dispatch=2.958s [2026-10-18 18:47:29,608] loop.py:45 INFO TRAIN_STEP step=2440 loss=1.725340 Training: 14%|█▎ | 2440/18018 [5:17:15<34:23:20, 7.95s/step, loss=1.725, acc=0.636, acc_len=1.49, thru=14.5, I=18.2, T=30.9, wait=3.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:40,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:47:40,289] loop.py:45 INFO TRAIN_STEP step=2441 loss=1.847258 Training: 14%|█▎ | 2441/18018 [5:17:26<37:56:11, 8.77s/step, loss=1.847, acc=0.616, acc_len=1.43, thru=13.3, I=13.5, T=30.8, wait=6.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:47:40,285] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2440: forward=2977.7ms backward=1137.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:40,688] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 18:47:47,118] loop.py:45 INFO TRAIN_STEP step=2442 loss=1.864907 Training: 14%|█▎ | 2442/18018 [5:17:33<35:25:15, 8.19s/step, loss=1.865, acc=0.615, acc_len=1.42, thru=17.3, I=17.8, T=33.4, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:55,323] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:47:56,637] loop.py:45 INFO TRAIN_STEP step=2443 loss=2.005721 Training: 14%|█▎ | 2443/18018 [5:17:42<37:08:41, 8.59s/step, loss=2.006, acc=0.591, acc_len=1.36, thru=14.5, I=16.1, T=31.4, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:47:56,833] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:02,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:48:03,932] loop.py:45 INFO TRAIN_STEP step=2444 loss=1.844114 Training: 14%|█▎ | 2444/18018 [5:17:50<35:28:00, 8.20s/step, loss=1.844, acc=0.615, acc_len=1.42, thru=15.3, I=20.2, T=33.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:04,144] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:09,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:48:10,245] loop.py:380 INFO TIMING step=2445: step=3.699s data=1.546s compute=3.668s [fwd=2.604s bwd=1.048s opt=0.003s] dispatch=2.604s [2026-10-18 18:48:10,245] loop.py:45 INFO TRAIN_STEP step=2445 loss=1.796152 Training: 14%|█▎ | 2445/18018 [5:17:56<33:01:11, 7.63s/step, loss=1.796, acc=0.627, acc_len=1.46, thru=19.4, I=24.1, T=34.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:10,400] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:16,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:48:17,097] loop.py:45 INFO TRAIN_STEP step=2446 loss=1.953444 Training: 14%|█▎ | 2446/18018 [5:18:03<32:00:13, 7.40s/step, loss=1.953, acc=0.603, acc_len=1.38, thru=17.4, I=22.2, T=31.0, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:48:17,093] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2445: forward=2945.0ms backward=1146.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:17,272] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:27,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:48:28,754] loop.py:45 INFO TRAIN_STEP step=2447 loss=1.934230 Training: 14%|█▎ | 2447/18018 [5:18:14<37:31:28, 8.68s/step, loss=1.934, acc=0.606, acc_len=1.39, thru=11.4, I=11.3, T=27.2, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:29,180] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:35,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:48:36,485] loop.py:45 INFO TRAIN_STEP step=2448 loss=1.873707 Training: 14%|█▎ | 2448/18018 [5:18:22<36:17:46, 8.39s/step, loss=1.874, acc=0.614, acc_len=1.42, thru=14.7, I=20.8, T=29.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:36,923] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:42,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:48:42,929] loop.py:45 INFO TRAIN_STEP step=2449 loss=1.854217 Training: 14%|█▎ | 2449/18018 [5:18:29<33:46:10, 7.81s/step, loss=1.854, acc=0.614, acc_len=1.42, thru=18.1, I=22.9, T=32.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:43,223] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:49,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:48:49,953] loop.py:380 INFO TIMING step=2450: step=4.503s data=2.238s compute=4.474s [fwd=3.106s bwd=1.352s opt=0.003s] dispatch=2.510s [2026-10-18 18:48:49,953] loop.py:45 INFO TRAIN_STEP step=2450 loss=1.809991 Training: 14%|█▎ | 2450/18018 [5:18:36<32:44:48, 7.57s/step, loss=1.810, acc=0.622, acc_len=1.45, thru=17.6, I=19.7, T=28.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:50,119] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:55,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:48:56,189] loop.py:45 INFO TRAIN_STEP step=2451 loss=1.844654 Training: 14%|█▎ | 2451/18018 [5:18:42<31:00:51, 7.17s/step, loss=1.845, acc=0.615, acc_len=1.43, thru=18.6, I=19.4, T=33.6, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:48:56,184] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2450: forward=2678.5ms backward=1080.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:48:56,300] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 18:49:03,354] loop.py:45 INFO TRAIN_STEP step=2452 loss=1.811491 Training: 14%|█▎ | 2452/18018 [5:18:49<30:59:57, 7.17s/step, loss=1.811, acc=0.624, acc_len=1.45, thru=17.3, I=19.6, T=31.5, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:10,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:49:12,893] loop.py:45 INFO TRAIN_STEP step=2453 loss=1.817793 Training: 14%|█▎ | 2453/18018 [5:18:58<34:04:20, 7.88s/step, loss=1.818, acc=0.622, acc_len=1.45, thru=14.0, I=17.3, T=29.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:13,277] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:18,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:49:19,825] loop.py:45 INFO TRAIN_STEP step=2454 loss=1.792119 Training: 14%|█▎ | 2454/18018 [5:19:05<32:50:21, 7.60s/step, loss=1.792, acc=0.624, acc_len=1.45, thru=15.5, I=26.0, T=33.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:20,193] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:49:26,112] loop.py:380 INFO TIMING step=2455: step=3.773s data=1.699s compute=3.741s [fwd=2.655s bwd=1.072s opt=0.003s] dispatch=2.504s [2026-10-18 18:49:26,112] loop.py:45 INFO TRAIN_STEP step=2455 loss=1.835029 Training: 14%|█▎ | 2455/18018 [5:19:12<31:08:29, 7.20s/step, loss=1.835, acc=0.615, acc_len=1.43, thru=16.9, I=25.0, T=33.9, wait=2.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:36,274] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:49:37,131] loop.py:45 INFO TRAIN_STEP step=2456 loss=1.930079 Training: 14%|█▎ | 2456/18018 [5:19:23<36:05:05, 8.35s/step, loss=1.930, acc=0.604, acc_len=1.39, thru=13.8, I=12.7, T=30.0, wait=6.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:49:37,126] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2455: forward=3008.7ms backward=1191.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:37,278] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:42,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:49:43,936] loop.py:45 INFO TRAIN_STEP step=2457 loss=1.798531 Training: 14%|█▎ | 2457/18018 [5:19:30<34:04:56, 7.88s/step, loss=1.799, acc=0.623, acc_len=1.44, thru=16.5, I=19.8, T=30.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:43,963] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 18:49:51,064] loop.py:45 INFO TRAIN_STEP step=2458 loss=1.954597 Training: 14%|█▎ | 2458/18018 [5:19:37<33:05:52, 7.66s/step, loss=1.955, acc=0.597, acc_len=1.38, thru=15.7, I=17.2, T=32.8, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:57,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:49:59,182] loop.py:45 INFO TRAIN_STEP step=2459 loss=1.777824 Training: 14%|█▎ | 2459/18018 [5:19:45<33:41:42, 7.80s/step, loss=1.778, acc=0.628, acc_len=1.46, thru=18.3, I=19.4, T=29.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:49:59,215] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:04,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:50:05,803] loop.py:380 INFO TIMING step=2460: step=4.147s data=1.970s compute=4.116s [fwd=2.894s bwd=1.208s opt=0.003s] dispatch=2.464s [2026-10-18 18:50:05,803] loop.py:45 INFO TRAIN_STEP step=2460 loss=1.852047 Training: 14%|█▎ | 2460/18018 [5:19:51<32:10:12, 7.44s/step, loss=1.852, acc=0.617, acc_len=1.44, thru=17.4, I=28.1, T=30.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:06,115] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:11,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:50:12,490] loop.py:45 INFO TRAIN_STEP step=2461 loss=1.681494 Training: 14%|█▎ | 2461/18018 [5:19:58<31:11:15, 7.22s/step, loss=1.681, acc=0.640, acc_len=1.50, thru=17.1, I=23.9, T=32.2, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:50:12,485] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2460: forward=2765.7ms backward=1154.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:12,587] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:18,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:50:19,385] loop.py:45 INFO TRAIN_STEP step=2462 loss=1.742923 Training: 14%|█▎ | 2462/18018 [5:20:05<30:46:05, 7.12s/step, loss=1.743, acc=0.627, acc_len=1.46, thru=16.9, I=19.2, T=30.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:19,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:27,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:50:28,732] loop.py:45 INFO TRAIN_STEP step=2463 loss=1.717597 Training: 14%|█▎ | 2463/18018 [5:20:14<33:38:59, 7.79s/step, loss=1.718, acc=0.636, acc_len=1.50, thru=13.7, I=15.7, T=29.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:28,862] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:33,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:50:34,987] loop.py:45 INFO TRAIN_STEP step=2464 loss=1.874454 Training: 14%|█▎ | 2464/18018 [5:20:21<31:39:46, 7.33s/step, loss=1.874, acc=0.612, acc_len=1.42, thru=18.8, I=21.5, T=32.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:35,185] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 18:50:42,845] loop.py:380 INFO TIMING step=2465: step=4.106s data=1.917s compute=4.075s [fwd=2.879s bwd=1.182s opt=0.003s] dispatch=3.742s [2026-10-18 18:50:42,845] loop.py:45 INFO TRAIN_STEP step=2465 loss=1.917474 Training: 14%|█▎ | 2465/18018 [5:20:28<32:20:48, 7.49s/step, loss=1.917, acc=0.605, acc_len=1.39, thru=13.8, I=16.3, T=31.2, wait=3.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:49,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:50:50,437] loop.py:45 INFO TRAIN_STEP step=2466 loss=1.831435 Training: 14%|█▎ | 2466/18018 [5:20:36<32:28:41, 7.52s/step, loss=1.831, acc=0.619, acc_len=1.44, thru=18.0, I=18.4, T=33.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:50,466] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 18:50:50,432] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2465: forward=2733.5ms backward=1099.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:57,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:50:59,275] loop.py:45 INFO TRAIN_STEP step=2467 loss=1.725812 Training: 14%|█▎ | 2467/18018 [5:20:45<34:11:11, 7.91s/step, loss=1.726, acc=0.634, acc_len=1.48, thru=14.0, I=17.4, T=30.5, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:50:59,321] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:03,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:51:05,773] loop.py:45 INFO TRAIN_STEP step=2468 loss=1.812463 Training: 14%|█▎ | 2468/18018 [5:20:51<32:21:10, 7.49s/step, loss=1.812, acc=0.618, acc_len=1.44, thru=18.7, I=24.6, T=33.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:05,973] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:11,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:51:12,348] loop.py:45 INFO TRAIN_STEP step=2469 loss=1.891071 Training: 14%|█▎ | 2469/18018 [5:20:58<31:09:54, 7.22s/step, loss=1.891, acc=0.608, acc_len=1.41, thru=18.6, I=24.8, T=28.9, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:12,383] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 18:51:18,934] loop.py:380 INFO TIMING step=2470: step=3.911s data=1.747s compute=3.874s [fwd=2.765s bwd=1.095s opt=0.003s] dispatch=2.664s [2026-10-18 18:51:18,934] loop.py:45 INFO TRAIN_STEP step=2470 loss=2.019268 Training: 14%|█▎ | 2470/18018 [5:21:05<30:20:39, 7.03s/step, loss=2.019, acc=0.581, acc_len=1.33, thru=16.9, I=18.7, T=32.7, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:28,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:51:28,877] loop.py:45 INFO TRAIN_STEP step=2471 loss=1.912076 Training: 14%|█▎ | 2471/18018 [5:21:14<34:07:26, 7.90s/step, loss=1.912, acc=0.602, acc_len=1.38, thru=13.3, I=13.9, T=29.7, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:51:28,872] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2470: forward=3095.9ms backward=1170.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:29,005] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:33,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:51:35,571] loop.py:45 INFO TRAIN_STEP step=2472 loss=1.770863 Training: 14%|█▎ | 2472/18018 [5:21:21<32:33:14, 7.54s/step, loss=1.771, acc=0.627, acc_len=1.46, thru=17.7, I=22.0, T=29.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:35,756] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:43,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:51:43,331] loop.py:45 INFO TRAIN_STEP step=2473 loss=1.734802 Training: 14%|█▎ | 2473/18018 [5:21:29<32:50:23, 7.61s/step, loss=1.735, acc=0.634, acc_len=1.48, thru=16.9, I=20.4, T=31.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:43,615] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:48,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:51:51,541] loop.py:45 INFO TRAIN_STEP step=2474 loss=1.737493 Training: 14%|█▎ | 2474/18018 [5:21:37<33:37:24, 7.79s/step, loss=1.737, acc=0.631, acc_len=1.48, thru=14.8, I=21.4, T=23.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:51,594] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:57,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:51:58,398] loop.py:380 INFO TIMING step=2475: step=3.598s data=1.560s compute=3.559s [fwd=2.547s bwd=0.997s opt=0.003s] dispatch=3.247s [2026-10-18 18:51:58,398] loop.py:45 INFO TRAIN_STEP step=2475 loss=1.797672 Training: 14%|█▎ | 2475/18018 [5:21:44<32:24:57, 7.51s/step, loss=1.798, acc=0.620, acc_len=1.45, thru=18.9, I=20.7, T=35.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:51:58,653] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:52:04,504] loop.py:45 INFO TRAIN_STEP step=2476 loss=1.895378 Training: 14%|█▎ | 2476/18018 [5:21:50<30:35:50, 7.09s/step, loss=1.895, acc=0.607, acc_len=1.40, thru=19.6, I=23.3, T=32.6, wait=2.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:52:04,499] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2475: forward=2885.4ms backward=988.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:10,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:52:11,534] loop.py:45 INFO TRAIN_STEP step=2477 loss=1.830200 Training: 14%|█▎ | 2477/18018 [5:21:57<30:31:19, 7.07s/step, loss=1.830, acc=0.615, acc_len=1.43, thru=17.9, I=19.8, T=29.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:11,592] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:16,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:52:19,009] loop.py:45 INFO TRAIN_STEP step=2478 loss=1.877297 Training: 14%|█▍ | 2478/18018 [5:22:05<31:02:37, 7.19s/step, loss=1.877, acc=0.609, acc_len=1.41, thru=15.6, I=24.7, T=25.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:19,465] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:25,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 18:52:28,026] loop.py:45 INFO TRAIN_STEP step=2479 loss=1.746175 Training: 14%|█▍ | 2479/18018 [5:22:14<33:24:21, 7.74s/step, loss=1.746, acc=0.629, acc_len=1.47, thru=15.4, I=18.3, T=24.1, wait=3.7s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:28,378] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:33,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:52:34,374] loop.py:380 INFO TIMING step=2480: step=3.730s data=1.638s compute=3.701s [fwd=2.628s bwd=1.058s opt=0.003s] dispatch=2.607s [2026-10-18 18:52:34,374] loop.py:45 INFO TRAIN_STEP step=2480 loss=1.770673 Training: 14%|█▍ | 2480/18018 [5:22:20<31:36:07, 7.32s/step, loss=1.771, acc=0.621, acc_len=1.44, thru=15.3, I=21.0, T=34.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:34,534] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:40,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:52:42,129] loop.py:45 INFO TRAIN_STEP step=2481 loss=1.980547 (TrainerActor pid=9310) [2026-10-18 18:52:42,123] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2480: forward=3256.5ms backward=1387.0ms Training: 14%|█▍ | 2481/18018 [5:22:28<32:09:29, 7.45s/step, loss=1.981, acc=0.591, acc_len=1.36, thru=16.3, I=21.2, T=27.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:42,226] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 18:52:49,068] loop.py:45 INFO TRAIN_STEP step=2482 loss=1.940478 Training: 14%|█▍ | 2482/18018 [5:22:35<31:29:36, 7.30s/step, loss=1.940, acc=0.602, acc_len=1.39, thru=14.7, I=19.7, T=33.4, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:56,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:52:57,745] loop.py:45 INFO TRAIN_STEP step=2483 loss=1.758645 Training: 14%|█▍ | 2483/18018 [5:22:43<33:16:48, 7.71s/step, loss=1.759, acc=0.629, acc_len=1.47, thru=15.1, I=17.5, T=29.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:52:58,070] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:53:05,901] loop.py:45 INFO TRAIN_STEP step=2484 loss=1.904093 Training: 14%|█▍ | 2484/18018 [5:22:51<33:50:59, 7.84s/step, loss=1.904, acc=0.607, acc_len=1.39, thru=13.0, I=17.1, T=30.3, wait=3.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:14,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:53:16,770] loop.py:380 INFO TIMING step=2485: step=4.628s data=2.001s compute=4.598s [fwd=3.176s bwd=1.408s opt=0.003s] dispatch=6.232s [2026-10-18 18:53:16,771] loop.py:45 INFO TRAIN_STEP step=2485 loss=1.831360 Training: 14%|█▍ | 2485/18018 [5:23:02<37:45:50, 8.75s/step, loss=1.831, acc=0.616, acc_len=1.43, thru=13.0, I=14.1, T=27.7, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:17,201] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:25,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:53:25,693] loop.py:45 INFO TRAIN_STEP step=2486 loss=1.773502 Training: 14%|█▍ | 2486/18018 [5:23:11<37:58:47, 8.80s/step, loss=1.774, acc=0.625, acc_len=1.46, thru=15.1, I=15.0, T=33.4, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:53:25,688] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2485: forward=2700.3ms backward=1087.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:25,980] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:33,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:53:35,410] loop.py:45 INFO TRAIN_STEP step=2487 loss=1.903204 Training: 14%|█▍ | 2487/18018 [5:23:21<39:09:40, 9.08s/step, loss=1.903, acc=0.611, acc_len=1.41, thru=13.3, I=17.0, T=29.5, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:35,670] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 18:53:42,313] loop.py:45 INFO TRAIN_STEP step=2488 loss=1.807278 Training: 14%|█▍ | 2488/18018 [5:23:28<36:20:39, 8.42s/step, loss=1.807, acc=0.624, acc_len=1.45, thru=12.3, I=16.2, T=33.3, wait=3.0s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:48,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:53:50,305] loop.py:45 INFO TRAIN_STEP step=2489 loss=1.770273 Training: 14%|█▍ | 2489/18018 [5:23:36<35:47:01, 8.30s/step, loss=1.770, acc=0.628, acc_len=1.47, thru=18.4, I=20.6, T=33.0, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:50,573] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:55,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:53:57,083] loop.py:380 INFO TIMING step=2490: step=3.934s data=1.755s compute=3.908s [fwd=2.741s bwd=1.153s opt=0.003s] dispatch=2.834s [2026-10-18 18:53:57,084] loop.py:45 INFO TRAIN_STEP step=2490 loss=1.798912 Training: 14%|█▍ | 2490/18018 [5:23:43<33:49:05, 7.84s/step, loss=1.799, acc=0.624, acc_len=1.46, thru=16.9, I=29.6, T=32.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:53:57,493] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:02,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:54:03,568] loop.py:45 INFO TRAIN_STEP step=2491 loss=1.762360 Training: 14%|█▍ | 2491/18018 [5:23:49<32:03:44, 7.43s/step, loss=1.762, acc=0.626, acc_len=1.47, thru=18.9, I=25.3, T=33.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:54:03,563] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2490: forward=2650.9ms backward=1104.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:03,748] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:08,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:54:10,035] loop.py:45 INFO TRAIN_STEP step=2492 loss=1.791345 Training: 14%|█▍ | 2492/18018 [5:23:56<30:48:33, 7.14s/step, loss=1.791, acc=0.628, acc_len=1.47, thru=18.6, I=23.2, T=31.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:10,470] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:18,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:54:19,333] loop.py:45 INFO TRAIN_STEP step=2493 loss=1.894488 Training: 14%|█▍ | 2493/18018 [5:24:05<33:35:30, 7.79s/step, loss=1.894, acc=0.608, acc_len=1.40, thru=14.4, I=16.3, T=26.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:19,708] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:24,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:54:26,286] loop.py:45 INFO TRAIN_STEP step=2494 loss=1.818828 Training: 14%|█▍ | 2494/18018 [5:24:12<32:30:39, 7.54s/step, loss=1.819, acc=0.620, acc_len=1.44, thru=16.2, I=20.6, T=32.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:26,398] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:31,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:54:32,820] loop.py:380 INFO TIMING step=2495: step=3.960s data=1.742s compute=3.910s [fwd=2.715s bwd=1.180s opt=0.003s] dispatch=2.562s [2026-10-18 18:54:32,820] loop.py:45 INFO TRAIN_STEP step=2495 loss=1.820815 Training: 14%|█▍ | 2495/18018 [5:24:18<31:12:23, 7.24s/step, loss=1.821, acc=0.617, acc_len=1.43, thru=17.4, I=23.8, T=32.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:33,098] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:39,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:54:40,720] loop.py:45 INFO TRAIN_STEP step=2496 loss=1.807473 Training: 14%|█▍ | 2496/18018 [5:24:26<32:03:43, 7.44s/step, loss=1.807, acc=0.624, acc_len=1.46, thru=15.2, I=18.8, T=24.2, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:54:40,714] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2495: forward=3471.6ms backward=1774.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:41,083] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:48,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:54:49,264] loop.py:45 INFO TRAIN_STEP step=2497 loss=1.766818 Training: 14%|█▍ | 2497/18018 [5:24:35<33:29:29, 7.77s/step, loss=1.767, acc=0.629, acc_len=1.48, thru=14.0, I=17.3, T=28.2, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:49,403] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:57,665] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:54:58,866] loop.py:45 INFO TRAIN_STEP step=2498 loss=1.981225 Training: 14%|█▍ | 2498/18018 [5:24:44<35:51:34, 8.32s/step, loss=1.981, acc=0.598, acc_len=1.38, thru=13.2, I=13.4, T=28.5, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:54:59,169] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:04,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:55:05,283] loop.py:45 INFO TRAIN_STEP step=2499 loss=1.764055 Training: 14%|█▍ | 2499/18018 [5:24:51<33:24:01, 7.75s/step, loss=1.764, acc=0.629, acc_len=1.47, thru=18.1, I=21.5, T=31.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:05,547] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:11,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:55:12,455] loop.py:380 INFO TIMING step=2500: step=4.090s data=1.923s compute=4.052s [fwd=2.837s bwd=1.200s opt=0.003s] dispatch=3.072s [2026-10-18 18:55:12,455] loop.py:45 INFO TRAIN_STEP step=2500 loss=1.849239 Training: 14%|█▍ | 2500/18018 [5:24:58<32:39:11, 7.58s/step, loss=1.849, acc=0.616, acc_len=1.43, thru=16.5, I=20.8, T=31.3, wait=3.1s, pool=64, epoch=1/2][2026-10-18 18:55:12,458] loop.py:419 INFO Saving checkpoint at step 2500... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 18:55:15,834] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0002501 (TrainerActor pid=9310) [2026-10-18 18:55:15,840] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0001751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:16,301] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:17,309] inference_manager.py:408 INFO Pool capacity available after 6.0s, pool_size=0/64 [2026-10-18 18:55:23,501] loop.py:45 INFO TRAIN_STEP step=2501 loss=1.680586 Training: 14%|█▍ | 2501/18018 [5:25:09<37:08:30, 8.62s/step, loss=1.681, acc=0.646, acc_len=1.52, thru=16.8, I=20.7, T=32.4, wait=2.6s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:55:23,496] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2500: forward=2715.5ms backward=1189.8ms (TrainerActor pid=9310) [2026-10-18 18:55:15,833] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0002501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 18:55:16,898] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0001751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:32,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:55:32,823] loop.py:45 INFO TRAIN_STEP step=2502 loss=1.819599 Training: 14%|█▍ | 2502/18018 [5:25:18<38:03:01, 8.83s/step, loss=1.820, acc=0.624, acc_len=1.46, thru=16.0, I=15.3, T=28.2, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:33,077] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:39,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:55:41,083] loop.py:45 INFO TRAIN_STEP step=2503 loss=1.866716 Training: 14%|█▍ | 2503/18018 [5:25:27<37:18:48, 8.66s/step, loss=1.867, acc=0.613, acc_len=1.42, thru=15.2, I=18.2, T=30.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:41,378] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:55:48,460] loop.py:45 INFO TRAIN_STEP step=2504 loss=1.764917 Training: 14%|█▍ | 2504/18018 [5:25:34<35:39:24, 8.27s/step, loss=1.765, acc=0.631, acc_len=1.47, thru=14.9, I=17.5, T=33.2, wait=3.5s, pool=52, epoch=1/2][2026-10-18 18:55:55,998] loop.py:380 INFO TIMING step=2505: step=3.902s data=1.703s compute=3.871s [fwd=2.776s bwd=1.080s opt=0.003s] dispatch=3.624s [2026-10-18 18:55:55,998] loop.py:45 INFO TRAIN_STEP step=2505 loss=1.852777 Training: 14%|█▍ | 2505/18018 [5:25:42<34:41:51, 8.05s/step, loss=1.853, acc=0.617, acc_len=1.43, thru=17.5, I=17.9, T=32.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:55,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:55:56,434] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:01,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 18:56:03,462] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2505: forward=2766.2ms backward=1049.0ms [2026-10-18 18:56:03,468] loop.py:45 INFO TRAIN_STEP step=2506 loss=1.773246 Training: 14%|█▍ | 2506/18018 [5:25:49<33:56:38, 7.88s/step, loss=1.773, acc=0.633, acc_len=1.48, thru=17.4, I=20.4, T=33.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:03,706] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:08,927] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:56:10,171] loop.py:45 INFO TRAIN_STEP step=2507 loss=1.930996 Training: 14%|█▍ | 2507/18018 [5:25:56<32:25:33, 7.53s/step, loss=1.931, acc=0.599, acc_len=1.38, thru=18.4, I=26.3, T=32.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:10,439] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:56:18,903] loop.py:45 INFO TRAIN_STEP step=2508 loss=1.676673 Training: 14%|█▍ | 2508/18018 [5:26:04<33:58:46, 7.89s/step, loss=1.677, acc=0.642, acc_len=1.52, thru=11.2, I=15.1, T=33.0, wait=4.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:26,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:56:27,652] loop.py:45 INFO TRAIN_STEP step=2509 loss=1.924514 Training: 14%|█▍ | 2509/18018 [5:26:13<35:05:35, 8.15s/step, loss=1.925, acc=0.608, acc_len=1.40, thru=16.6, I=16.6, T=30.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:27,878] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:35,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:56:35,788] loop.py:380 INFO TIMING step=2510: step=4.180s data=1.783s compute=4.149s [fwd=2.887s bwd=1.247s opt=0.003s] dispatch=3.947s [2026-10-18 18:56:35,789] loop.py:45 INFO TRAIN_STEP step=2510 loss=1.786946 Training: 14%|█▍ | 2510/18018 [5:26:21<35:04:48, 8.14s/step, loss=1.787, acc=0.622, acc_len=1.45, thru=15.3, I=18.7, T=30.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:35,850] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:39,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:56:41,341] loop.py:45 INFO TRAIN_STEP step=2511 loss=1.692244 Training: 14%|█▍ | 2511/18018 [5:26:27<31:43:49, 7.37s/step, loss=1.692, acc=0.643, acc_len=1.52, thru=20.1, I=22.8, T=34.2, wait=1.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:56:41,337] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2510: forward=2684.8ms backward=1015.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:56:41,500] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:56:47,809] loop.py:45 INFO TRAIN_STEP step=2512 loss=1.663028 Training: 14%|█▍ | 2512/18018 [5:26:33<30:33:52, 7.10s/step, loss=1.663, acc=0.639, acc_len=1.51, thru=15.7, I=24.2, T=29.4, wait=2.1s, pool=36, epoch=1/2][2026-10-18 18:56:59,918] loop.py:45 INFO TRAIN_STEP step=2513 loss=1.943596 Training: 14%|█▍ | 2513/18018 [5:26:46<37:02:21, 8.60s/step, loss=1.944, acc=0.597, acc_len=1.37, thru=12.9, I=12.5, T=28.6, wait=7.6s, pool=56, epoch=1/2][2026-10-18 18:57:08,142] loop.py:45 INFO TRAIN_STEP step=2514 loss=2.019470 Training: 14%|█▍ | 2514/18018 [5:26:54<36:33:17, 8.49s/step, loss=2.019, acc=0.591, acc_len=1.36, thru=15.0, I=15.4, T=30.4, wait=4.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:57:13,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:57:15,085] loop.py:380 INFO TIMING step=2515: step=4.296s data=2.081s compute=4.268s [fwd=2.973s bwd=1.280s opt=0.003s] dispatch=2.635s [2026-10-18 18:57:15,085] loop.py:45 INFO TRAIN_STEP step=2515 loss=1.785090 Training: 14%|█▍ | 2515/18018 [5:27:01<34:33:11, 8.02s/step, loss=1.785, acc=0.623, acc_len=1.46, thru=17.2, I=19.3, T=29.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:57:15,352] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:57:22,320] loop.py:45 INFO TRAIN_STEP step=2516 loss=1.772721 Training: 14%|█▍ | 2516/18018 [5:27:08<33:31:57, 7.79s/step, loss=1.773, acc=0.630, acc_len=1.48, thru=15.4, I=18.4, T=30.0, wait=3.0s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:57:22,315] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2515: forward=3009.6ms backward=1217.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:57:29,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:57:30,548] loop.py:45 INFO TRAIN_STEP step=2517 loss=1.714972 Training: 14%|█▍ | 2517/18018 [5:27:16<34:05:56, 7.92s/step, loss=1.715, acc=0.636, acc_len=1.50, thru=16.2, I=16.3, T=31.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:57:30,704] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 18:57:37,959] loop.py:45 INFO TRAIN_STEP step=2518 loss=2.038225 Training: 14%|█▍ | 2518/18018 [5:27:24<33:26:25, 7.77s/step, loss=2.038, acc=0.589, acc_len=1.34, thru=14.9, I=18.4, T=33.3, wait=3.6s, pool=48, epoch=1/2][2026-10-18 18:57:46,045] loop.py:45 INFO TRAIN_STEP step=2519 loss=1.772208 Training: 14%|█▍ | 2519/18018 [5:27:32<33:51:05, 7.86s/step, loss=1.772, acc=0.629, acc_len=1.47, thru=13.4, I=17.2, T=33.0, wait=4.2s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:57:56,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:57:58,098] loop.py:380 INFO TIMING step=2520: step=5.112s data=2.471s compute=5.061s [fwd=3.518s bwd=1.527s opt=0.003s] dispatch=6.932s [2026-10-18 18:57:58,099] loop.py:45 INFO TRAIN_STEP step=2520 loss=1.835854 Training: 14%|█▍ | 2520/18018 [5:27:44<39:15:40, 9.12s/step, loss=1.836, acc=0.625, acc_len=1.45, thru=12.4, I=13.2, T=25.0, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:57:58,183] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:03,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:58:06,349] loop.py:45 INFO TRAIN_STEP step=2521 loss=1.788872 Training: 14%|█▍ | 2521/18018 [5:27:52<38:08:07, 8.86s/step, loss=1.789, acc=0.624, acc_len=1.45, thru=13.6, I=20.3, T=26.0, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:58:06,344] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2520: forward=3462.5ms backward=1419.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:06,547] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:13,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:58:13,618] loop.py:45 INFO TRAIN_STEP step=2522 loss=1.771492 Training: 14%|█▍ | 2522/18018 [5:27:59<36:04:53, 8.38s/step, loss=1.771, acc=0.623, acc_len=1.45, thru=18.1, I=17.0, T=27.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:13,992] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:21,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:58:22,076] loop.py:45 INFO TRAIN_STEP step=2523 loss=1.824690 Training: 14%|█▍ | 2523/18018 [5:28:08<36:10:29, 8.40s/step, loss=1.825, acc=0.620, acc_len=1.44, thru=16.0, I=16.0, T=28.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:22,164] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:27,258] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:58:28,632] loop.py:45 INFO TRAIN_STEP step=2524 loss=1.711320 Training: 14%|█▍ | 2524/18018 [5:28:14<33:47:17, 7.85s/step, loss=1.711, acc=0.632, acc_len=1.48, thru=17.7, I=21.5, T=34.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:28,779] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:34,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:58:35,491] loop.py:380 INFO TIMING step=2525: step=4.242s data=1.993s compute=4.199s [fwd=2.950s bwd=1.233s opt=0.003s] dispatch=2.606s [2026-10-18 18:58:35,491] loop.py:45 INFO TRAIN_STEP step=2525 loss=1.921265 Training: 14%|█▍ | 2525/18018 [5:28:21<32:30:20, 7.55s/step, loss=1.921, acc=0.603, acc_len=1.40, thru=17.9, I=22.2, T=30.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:35,754] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 18:58:41,982] loop.py:45 INFO TRAIN_STEP step=2526 loss=1.857490 Training: 14%|█▍ | 2526/18018 [5:28:28<31:07:47, 7.23s/step, loss=1.857, acc=0.620, acc_len=1.44, thru=17.8, I=19.5, T=33.5, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:58:41,977] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2525: forward=2706.8ms backward=1062.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:48,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:58:49,575] loop.py:45 INFO TRAIN_STEP step=2527 loss=1.928240 Training: 14%|█▍ | 2527/18018 [5:28:35<31:35:31, 7.34s/step, loss=1.928, acc=0.606, acc_len=1.40, thru=17.4, I=18.9, T=29.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:49,940] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:56,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:58:57,383] loop.py:45 INFO TRAIN_STEP step=2528 loss=1.781684 Training: 14%|█▍ | 2528/18018 [5:28:43<32:11:36, 7.48s/step, loss=1.782, acc=0.627, acc_len=1.46, thru=14.5, I=16.8, T=32.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:58:57,494] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 18:59:04,410] loop.py:45 INFO TRAIN_STEP step=2529 loss=1.797641 Training: 14%|█▍ | 2529/18018 [5:28:50<31:36:15, 7.35s/step, loss=1.798, acc=0.626, acc_len=1.45, thru=16.9, I=20.2, T=28.6, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:10,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:59:12,546] loop.py:380 INFO TIMING step=2530: step=3.950s data=1.778s compute=3.895s [fwd=2.719s bwd=1.161s opt=0.003s] dispatch=4.174s [2026-10-18 18:59:12,546] loop.py:45 INFO TRAIN_STEP step=2530 loss=1.890556 Training: 14%|█▍ | 2530/18018 [5:28:58<32:37:22, 7.58s/step, loss=1.891, acc=0.605, acc_len=1.40, thru=16.9, I=18.0, T=32.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:12,810] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:19,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:59:19,427] loop.py:45 INFO TRAIN_STEP step=2531 loss=1.797730 Training: 14%|█▍ | 2531/18018 [5:29:05<31:42:52, 7.37s/step, loss=1.798, acc=0.622, acc_len=1.45, thru=17.4, I=21.7, T=33.3, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:59:19,422] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2530: forward=2757.0ms backward=1045.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:19,604] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:26,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:59:26,773] loop.py:45 INFO TRAIN_STEP step=2532 loss=1.784193 Training: 14%|█▍ | 2532/18018 [5:29:12<31:40:31, 7.36s/step, loss=1.784, acc=0.623, acc_len=1.45, thru=18.8, I=18.1, T=32.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:27,228] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:33,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:59:35,707] loop.py:45 INFO TRAIN_STEP step=2533 loss=1.714706 Training: 14%|█▍ | 2533/18018 [5:29:21<33:42:11, 7.84s/step, loss=1.715, acc=0.639, acc_len=1.50, thru=15.4, I=19.1, T=31.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:35,740] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:42,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:59:42,268] loop.py:45 INFO TRAIN_STEP step=2534 loss=1.782300 Training: 14%|█▍ | 2534/18018 [5:29:28<32:03:24, 7.45s/step, loss=1.782, acc=0.622, acc_len=1.45, thru=18.4, I=23.8, T=33.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:42,506] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:46,660] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:59:48,655] loop.py:380 INFO TIMING step=2535: step=4.408s data=2.188s compute=4.363s [fwd=3.084s bwd=1.263s opt=0.003s] dispatch=1.969s [2026-10-18 18:59:48,655] loop.py:45 INFO TRAIN_STEP step=2535 loss=1.885839 Training: 14%|█▍ | 2535/18018 [5:29:34<30:40:46, 7.13s/step, loss=1.886, acc=0.610, acc_len=1.40, thru=18.9, I=24.1, T=29.0, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:48,690] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:53,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 18:59:56,653] loop.py:45 INFO TRAIN_STEP step=2536 loss=1.781024 Training: 14%|█▍ | 2536/18018 [5:29:42<31:47:26, 7.39s/step, loss=1.781, acc=0.629, acc_len=1.47, thru=13.2, I=22.8, T=25.1, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 18:59:56,648] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2535: forward=3491.2ms backward=1571.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 18:59:56,804] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:03,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:00:04,133] loop.py:45 INFO TRAIN_STEP step=2537 loss=1.858703 Training: 14%|█▍ | 2537/18018 [5:29:50<31:54:12, 7.42s/step, loss=1.859, acc=0.616, acc_len=1.42, thru=17.5, I=20.7, T=35.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:04,540] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:00:12,810] loop.py:45 INFO TRAIN_STEP step=2538 loss=1.747273 Training: 14%|█▍ | 2538/18018 [5:29:58<33:31:28, 7.80s/step, loss=1.747, acc=0.629, acc_len=1.47, thru=12.1, I=15.1, T=32.3, wait=4.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:19,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:00:20,490] loop.py:45 INFO TRAIN_STEP step=2539 loss=1.988454 Training: 14%|█▍ | 2539/18018 [5:30:06<33:22:20, 7.76s/step, loss=1.988, acc=0.593, acc_len=1.37, thru=17.6, I=18.1, T=30.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:20,701] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:27,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:00:29,025] loop.py:380 INFO TIMING step=2540: step=4.328s data=2.043s compute=4.291s [fwd=2.993s bwd=1.283s opt=0.003s] dispatch=4.196s [2026-10-18 19:00:29,025] loop.py:45 INFO TRAIN_STEP step=2540 loss=1.820372 Training: 14%|█▍ | 2540/18018 [5:30:15<34:22:03, 7.99s/step, loss=1.820, acc=0.623, acc_len=1.45, thru=15.3, I=19.3, T=29.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:29,344] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:34,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:00:36,412] loop.py:45 INFO TRAIN_STEP step=2541 loss=1.909717 Training: 14%|█▍ | 2541/18018 [5:30:22<33:34:59, 7.81s/step, loss=1.910, acc=0.607, acc_len=1.40, thru=15.3, I=24.7, T=28.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:00:36,407] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2540: forward=3110.1ms backward=1388.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:36,692] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:41,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:00:43,090] loop.py:45 INFO TRAIN_STEP step=2542 loss=1.826052 Training: 14%|█▍ | 2542/18018 [5:30:29<32:07:03, 7.47s/step, loss=1.826, acc=0.618, acc_len=1.44, thru=15.1, I=24.2, T=33.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:43,183] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:47,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:00:49,444] loop.py:45 INFO TRAIN_STEP step=2543 loss=1.783855 Training: 14%|█▍ | 2543/18018 [5:30:35<30:40:30, 7.14s/step, loss=1.784, acc=0.621, acc_len=1.45, thru=19.3, I=25.2, T=30.6, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:00:49,475] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 19:00:55,084] loop.py:45 INFO TRAIN_STEP step=2544 loss=1.728534 Training: 14%|█▍ | 2544/18018 [5:30:41<28:44:34, 6.69s/step, loss=1.729, acc=0.633, acc_len=1.49, thru=19.9, I=27.7, T=31.5, wait=1.6s, pool=60, epoch=1/2][2026-10-18 19:01:03,027] loop.py:380 INFO TIMING step=2545: step=4.101s data=1.883s compute=4.072s [fwd=2.915s bwd=1.143s opt=0.003s] dispatch=3.833s [2026-10-18 19:01:03,027] loop.py:45 INFO TRAIN_STEP step=2545 loss=1.735607 Training: 14%|█▍ | 2545/18018 [5:30:49<30:21:36, 7.06s/step, loss=1.736, acc=0.633, acc_len=1.49, thru=13.7, I=13.0, T=31.2, wait=3.8s, pool=44, epoch=1/2][2026-10-18 19:01:09,910] loop.py:45 INFO TRAIN_STEP step=2546 loss=1.674187 Training: 14%|█▍ | 2546/18018 [5:30:55<30:07:29, 7.01s/step, loss=1.674, acc=0.644, acc_len=1.51, thru=18.1, I=17.5, T=30.2, wait=2.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:01:09,905] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2545: forward=2897.7ms backward=1291.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:01:16,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:01:17,440] loop.py:45 INFO TRAIN_STEP step=2547 loss=1.866539 Training: 14%|█▍ | 2547/18018 [5:31:03<30:47:40, 7.17s/step, loss=1.867, acc=0.616, acc_len=1.43, thru=17.5, I=17.9, T=29.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:01:17,911] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:01:23,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:01:24,470] loop.py:45 INFO TRAIN_STEP step=2548 loss=1.821020 Training: 14%|█▍ | 2548/18018 [5:31:10<30:37:08, 7.13s/step, loss=1.821, acc=0.618, acc_len=1.43, thru=16.9, I=20.4, T=31.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:01:24,925] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:01:31,971] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:01:33,563] loop.py:45 INFO TRAIN_STEP step=2549 loss=1.926734 Training: 14%|█▍ | 2549/18018 [5:31:19<33:09:19, 7.72s/step, loss=1.927, acc=0.600, acc_len=1.38, thru=13.7, I=16.8, T=28.6, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:01:33,987] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:01:40,315] loop.py:380 INFO TIMING step=2550: step=4.482s data=2.160s compute=4.455s [fwd=3.109s bwd=1.330s opt=0.003s] dispatch=2.258s [2026-10-18 19:01:40,315] loop.py:45 INFO TRAIN_STEP step=2550 loss=1.722050 Training: 14%|█▍ | 2550/18018 [5:31:26<31:54:24, 7.43s/step, loss=1.722, acc=0.634, acc_len=1.48, thru=15.7, I=18.6, T=28.6, wait=2.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:01:47,665] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:01:48,232] loop.py:45 INFO TRAIN_STEP step=2551 loss=1.821176 Training: 14%|█▍ | 2551/18018 [5:31:34<32:32:27, 7.57s/step, loss=1.821, acc=0.618, acc_len=1.44, thru=18.0, I=15.7, T=28.8, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:01:48,226] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2550: forward=3125.4ms backward=1269.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:01:48,669] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 19:01:55,735] loop.py:45 INFO TRAIN_STEP step=2552 loss=1.622913 Training: 14%|█▍ | 2552/18018 [5:31:41<32:26:47, 7.55s/step, loss=1.623, acc=0.654, acc_len=1.55, thru=15.7, I=17.4, T=32.2, wait=3.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:04,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:02:05,925] loop.py:45 INFO TRAIN_STEP step=2553 loss=1.883319 Training: 14%|█▍ | 2553/18018 [5:31:52<35:50:43, 8.34s/step, loss=1.883, acc=0.608, acc_len=1.41, thru=13.8, I=14.0, T=31.7, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:06,316] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:12,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:02:14,107] loop.py:45 INFO TRAIN_STEP step=2554 loss=1.807468 Training: 14%|█▍ | 2554/18018 [5:32:00<35:37:57, 8.30s/step, loss=1.807, acc=0.617, acc_len=1.43, thru=14.9, I=20.9, T=32.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:14,144] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:21,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:02:21,887] loop.py:380 INFO TIMING step=2555: step=3.812s data=1.662s compute=3.785s [fwd=2.664s bwd=1.105s opt=0.003s] dispatch=3.958s [2026-10-18 19:02:21,887] loop.py:45 INFO TRAIN_STEP step=2555 loss=1.859987 Training: 14%|█▍ | 2555/18018 [5:32:07<34:57:51, 8.14s/step, loss=1.860, acc=0.612, acc_len=1.42, thru=14.9, I=16.3, T=33.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:22,291] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 19:02:29,074] loop.py:45 INFO TRAIN_STEP step=2556 loss=1.802342 Training: 14%|█▍ | 2556/18018 [5:32:15<33:44:03, 7.85s/step, loss=1.802, acc=0.622, acc_len=1.45, thru=16.0, I=16.2, T=32.4, wait=3.2s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:02:29,069] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2555: forward=2816.6ms backward=1085.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:35,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:02:37,190] loop.py:45 INFO TRAIN_STEP step=2557 loss=1.916569 Training: 14%|█▍ | 2557/18018 [5:32:23<34:04:07, 7.93s/step, loss=1.917, acc=0.605, acc_len=1.40, thru=18.2, I=18.9, T=29.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:37,441] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:42,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:02:43,454] loop.py:45 INFO TRAIN_STEP step=2558 loss=1.647386 Training: 14%|█▍ | 2558/18018 [5:32:29<31:54:59, 7.43s/step, loss=1.647, acc=0.645, acc_len=1.52, thru=17.7, I=22.3, T=32.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:43,514] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:48,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:02:50,356] loop.py:45 INFO TRAIN_STEP step=2559 loss=1.730205 Training: 14%|█▍ | 2559/18018 [5:32:36<31:13:54, 7.27s/step, loss=1.730, acc=0.633, acc_len=1.49, thru=18.4, I=22.7, T=30.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:50,422] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:55,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:02:56,331] loop.py:380 INFO TIMING step=2560: step=3.834s data=1.757s compute=3.807s [fwd=2.669s bwd=1.123s opt=0.003s] dispatch=2.132s [2026-10-18 19:02:56,332] loop.py:45 INFO TRAIN_STEP step=2560 loss=1.708095 Training: 14%|█▍ | 2560/18018 [5:32:42<29:33:39, 6.88s/step, loss=1.708, acc=0.637, acc_len=1.50, thru=19.6, I=26.3, T=33.4, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:02:56,557] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:02,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:03:04,868] loop.py:45 INFO TRAIN_STEP step=2561 loss=1.809527 Training: 14%|█▍ | 2561/18018 [5:32:50<31:41:04, 7.38s/step, loss=1.810, acc=0.619, acc_len=1.43, thru=13.7, I=20.3, T=26.5, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:03:04,863] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2560: forward=3303.7ms backward=1477.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:05,335] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:11,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:03:13,374] loop.py:45 INFO TRAIN_STEP step=2562 loss=1.804899 Training: 14%|█▍ | 2562/18018 [5:32:59<33:08:04, 7.72s/step, loss=1.805, acc=0.619, acc_len=1.45, thru=15.9, I=18.4, T=28.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:13,500] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:17,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:03:18,889] loop.py:45 INFO TRAIN_STEP step=2563 loss=1.833675 Training: 14%|█▍ | 2563/18018 [5:33:04<30:17:40, 7.06s/step, loss=1.834, acc=0.619, acc_len=1.43, thru=19.8, I=24.0, T=34.6, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:18,972] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:27,710] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:03:28,977] loop.py:45 INFO TRAIN_STEP step=2564 loss=1.720963 Training: 14%|█▍ | 2564/18018 [5:33:15<34:11:46, 7.97s/step, loss=1.721, acc=0.634, acc_len=1.49, thru=12.8, I=14.6, T=25.8, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:29,217] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:33,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:03:35,175] loop.py:380 INFO TIMING step=2565: step=4.205s data=1.757s compute=4.170s [fwd=2.793s bwd=1.364s opt=0.003s] dispatch=1.984s [2026-10-18 19:03:35,175] loop.py:45 INFO TRAIN_STEP step=2565 loss=1.788765 Training: 14%|█▍ | 2565/18018 [5:33:21<31:55:00, 7.44s/step, loss=1.789, acc=0.628, acc_len=1.46, thru=17.9, I=20.2, T=30.4, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:35,560] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:03:42,919] loop.py:45 INFO TRAIN_STEP step=2566 loss=1.918744 Training: 14%|█▍ | 2566/18018 [5:33:29<32:18:47, 7.53s/step, loss=1.919, acc=0.605, acc_len=1.40, thru=12.9, I=18.3, T=31.9, wait=3.7s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:03:42,913] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2565: forward=2850.1ms backward=1110.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:51,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:03:52,184] loop.py:45 INFO TRAIN_STEP step=2567 loss=1.769756 Training: 14%|█▍ | 2567/18018 [5:33:38<34:32:47, 8.05s/step, loss=1.770, acc=0.631, acc_len=1.47, thru=15.8, I=15.8, T=28.3, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:52,323] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:59,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:03:59,673] loop.py:45 INFO TRAIN_STEP step=2568 loss=1.880380 Training: 14%|█▍ | 2568/18018 [5:33:45<33:49:28, 7.88s/step, loss=1.880, acc=0.608, acc_len=1.41, thru=16.1, I=16.4, T=29.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:03:59,940] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:06,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:04:08,637] loop.py:45 INFO TRAIN_STEP step=2569 loss=1.832125 Training: 14%|█▍ | 2569/18018 [5:33:54<35:12:52, 8.21s/step, loss=1.832, acc=0.618, acc_len=1.43, thru=14.9, I=19.5, T=24.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:08,709] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:13,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:04:14,627] loop.py:380 INFO TIMING step=2570: step=3.852s data=1.723s compute=3.814s [fwd=2.645s bwd=1.155s opt=0.003s] dispatch=2.130s [2026-10-18 19:04:14,628] loop.py:45 INFO TRAIN_STEP step=2570 loss=1.791511 Training: 14%|█▍ | 2570/18018 [5:34:00<32:21:47, 7.54s/step, loss=1.792, acc=0.624, acc_len=1.45, thru=15.9, I=23.8, T=33.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:14,771] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:20,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:04:21,638] loop.py:45 INFO TRAIN_STEP step=2571 loss=1.715835 Training: 14%|█▍ | 2571/18018 [5:34:07<31:40:31, 7.38s/step, loss=1.716, acc=0.633, acc_len=1.48, thru=17.9, I=20.6, T=29.8, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:04:21,633] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2570: forward=2934.5ms backward=1310.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:22,013] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:27,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:04:28,901] loop.py:45 INFO TRAIN_STEP step=2572 loss=1.918783 Training: 14%|█▍ | 2572/18018 [5:34:14<31:31:08, 7.35s/step, loss=1.919, acc=0.601, acc_len=1.38, thru=15.7, I=20.7, T=31.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:29,083] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:04:34,917] loop.py:45 INFO TRAIN_STEP step=2573 loss=1.644134 Training: 14%|█▍ | 2573/18018 [5:34:21<29:48:13, 6.95s/step, loss=1.644, acc=0.646, acc_len=1.53, thru=18.3, I=21.2, T=35.7, wait=2.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:41,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:04:42,852] loop.py:45 INFO TRAIN_STEP step=2574 loss=1.888554 Training: 14%|█▍ | 2574/18018 [5:34:28<31:04:31, 7.24s/step, loss=1.889, acc=0.612, acc_len=1.42, thru=18.2, I=18.2, T=31.5, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:43,142] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:52,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:04:52,470] loop.py:380 INFO TIMING step=2575: step=4.263s data=2.008s compute=4.217s [fwd=2.972s bwd=1.230s opt=0.003s] dispatch=5.345s [2026-10-18 19:04:52,470] loop.py:45 INFO TRAIN_STEP step=2575 loss=1.836039 Training: 14%|█▍ | 2575/18018 [5:34:38<34:07:45, 7.96s/step, loss=1.836, acc=0.618, acc_len=1.45, thru=13.7, I=13.1, T=30.0, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:52,815] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:04:58,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:05:00,802] loop.py:45 INFO TRAIN_STEP step=2576 loss=1.779898 Training: 14%|█▍ | 2576/18018 [5:34:46<34:36:44, 8.07s/step, loss=1.780, acc=0.626, acc_len=1.46, thru=15.5, I=18.9, T=24.7, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:05:00,798] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2575: forward=3470.3ms backward=1662.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:05:01,074] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:05:06,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:05:07,381] loop.py:45 INFO TRAIN_STEP step=2577 loss=1.769622 Training: 14%|█▍ | 2577/18018 [5:34:53<32:41:33, 7.62s/step, loss=1.770, acc=0.625, acc_len=1.46, thru=14.9, I=20.7, T=34.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:05:07,847] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:05:14,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:05:15,606] loop.py:45 INFO TRAIN_STEP step=2578 loss=1.851439 Training: 14%|█▍ | 2578/18018 [5:35:01<33:28:00, 7.80s/step, loss=1.851, acc=0.614, acc_len=1.42, thru=14.8, I=19.5, T=29.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:05:15,765] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 19:05:22,349] loop.py:45 INFO TRAIN_STEP step=2579 loss=1.821342 Training: 14%|█▍ | 2579/18018 [5:35:08<32:05:54, 7.48s/step, loss=1.821, acc=0.622, acc_len=1.44, thru=15.4, I=21.6, T=29.6, wait=2.4s, pool=48, epoch=1/2][2026-10-18 19:05:31,734] loop.py:380 INFO TIMING step=2580: step=4.035s data=1.767s compute=3.994s [fwd=2.894s bwd=1.085s opt=0.003s] dispatch=5.340s [2026-10-18 19:05:31,734] loop.py:45 INFO TRAIN_STEP step=2580 loss=1.787501 Training: 14%|█▍ | 2580/18018 [5:35:17<34:32:34, 8.06s/step, loss=1.788, acc=0.633, acc_len=1.48, thru=14.4, I=13.7, T=31.7, wait=5.3s, pool=56, epoch=1/2][2026-10-18 19:05:39,173] loop.py:45 INFO TRAIN_STEP step=2581 loss=1.916952 Training: 14%|█▍ | 2581/18018 [5:35:25<33:44:43, 7.87s/step, loss=1.917, acc=0.604, acc_len=1.40, thru=16.1, I=16.7, T=30.3, wait=3.2s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:05:39,168] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2580: forward=2957.1ms backward=1215.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:05:50,637] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:05:51,985] loop.py:45 INFO TRAIN_STEP step=2582 loss=1.799302 Training: 14%|█▍ | 2582/18018 [5:35:38<40:06:05, 9.35s/step, loss=1.799, acc=0.626, acc_len=1.46, thru=12.5, I=12.1, T=28.2, wait=8.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:05:52,146] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:05:57,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:05:59,489] loop.py:45 INFO TRAIN_STEP step=2583 loss=1.875421 Training: 14%|█▍ | 2583/18018 [5:35:45<37:43:20, 8.80s/step, loss=1.875, acc=0.607, acc_len=1.40, thru=16.6, I=22.4, T=24.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:05:59,843] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:06:07,703] loop.py:45 INFO TRAIN_STEP step=2584 loss=1.768207 Training: 14%|█▍ | 2584/18018 [5:35:53<36:58:04, 8.62s/step, loss=1.768, acc=0.623, acc_len=1.45, thru=11.3, I=13.2, T=32.8, wait=4.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:15,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:06:16,459] loop.py:380 INFO TIMING step=2585: step=4.660s data=2.145s compute=4.624s [fwd=3.239s bwd=1.370s opt=0.003s] dispatch=4.086s [2026-10-18 19:06:16,459] loop.py:45 INFO TRAIN_STEP step=2585 loss=1.945133 Training: 14%|█▍ | 2585/18018 [5:36:02<37:08:07, 8.66s/step, loss=1.945, acc=0.596, acc_len=1.37, thru=16.7, I=16.4, T=27.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:16,956] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:22,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:06:23,031] loop.py:45 INFO TRAIN_STEP step=2586 loss=1.699191 Training: 14%|█▍ | 2586/18018 [5:36:09<34:26:35, 8.03s/step, loss=1.699, acc=0.641, acc_len=1.50, thru=17.5, I=18.8, T=36.9, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:06:23,026] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2585: forward=2473.6ms backward=947.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:23,499] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 19:06:31,202] loop.py:45 INFO TRAIN_STEP step=2587 loss=1.764305 Training: 14%|█▍ | 2587/18018 [5:36:17<34:37:00, 8.08s/step, loss=1.764, acc=0.629, acc_len=1.48, thru=16.9, I=17.9, T=30.8, wait=4.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:36,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:06:39,855] loop.py:45 INFO TRAIN_STEP step=2588 loss=1.773052 Training: 14%|█▍ | 2588/18018 [5:36:25<35:21:37, 8.25s/step, loss=1.773, acc=0.623, acc_len=1.46, thru=15.7, I=20.9, T=21.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:40,009] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:45,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:06:46,540] loop.py:45 INFO TRAIN_STEP step=2589 loss=1.788784 Training: 14%|█▍ | 2589/18018 [5:36:32<33:20:36, 7.78s/step, loss=1.789, acc=0.617, acc_len=1.44, thru=19.6, I=21.8, T=32.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:46,694] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:51,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:06:53,432] loop.py:380 INFO TIMING step=2590: step=4.328s data=2.151s compute=4.300s [fwd=3.172s bwd=1.114s opt=0.003s] dispatch=2.555s [2026-10-18 19:06:53,432] loop.py:45 INFO TRAIN_STEP step=2590 loss=1.826315 Training: 14%|█▍ | 2590/18018 [5:36:39<32:12:01, 7.51s/step, loss=1.826, acc=0.614, acc_len=1.42, thru=18.2, I=22.1, T=29.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:06:53,905] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:07:03,261] loop.py:45 INFO TRAIN_STEP step=2591 loss=1.759531 Training: 14%|█▍ | 2591/18018 [5:36:49<35:10:35, 8.21s/step, loss=1.760, acc=0.630, acc_len=1.48, thru=13.3, I=15.8, T=30.1, wait=5.6s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:07:03,255] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2590: forward=2960.3ms backward=1247.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:12,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:07:14,081] loop.py:45 INFO TRAIN_STEP step=2592 loss=1.724483 Training: 14%|█▍ | 2592/18018 [5:37:00<38:31:51, 8.99s/step, loss=1.724, acc=0.636, acc_len=1.49, thru=13.2, I=12.0, T=27.2, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:14,413] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:24,132] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:07:24,534] loop.py:45 INFO TRAIN_STEP step=2593 loss=1.888070 Training: 14%|█▍ | 2593/18018 [5:37:10<40:24:17, 9.43s/step, loss=1.888, acc=0.609, acc_len=1.40, thru=12.7, I=12.0, T=28.4, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:24,641] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-18 19:07:31,851] loop.py:45 INFO TRAIN_STEP step=2594 loss=1.963400 Training: 14%|█▍ | 2594/18018 [5:37:17<37:41:03, 8.80s/step, loss=1.963, acc=0.598, acc_len=1.38, thru=14.6, I=15.1, T=27.6, wait=2.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:39,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:07:40,495] loop.py:380 INFO TIMING step=2595: step=4.113s data=1.820s compute=4.064s [fwd=2.890s bwd=1.159s opt=0.003s] dispatch=4.523s [2026-10-18 19:07:40,495] loop.py:45 INFO TRAIN_STEP step=2595 loss=1.924358 Training: 14%|█▍ | 2595/18018 [5:37:26<37:29:16, 8.75s/step, loss=1.924, acc=0.602, acc_len=1.39, thru=17.0, I=15.0, T=31.1, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:40,787] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:46,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:07:47,399] loop.py:45 INFO TRAIN_STEP step=2596 loss=1.953996 Training: 14%|█▍ | 2596/18018 [5:37:33<35:06:54, 8.20s/step, loss=1.954, acc=0.595, acc_len=1.36, thru=17.2, I=20.6, T=27.6, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:07:47,394] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2595: forward=3205.8ms backward=1375.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:47,743] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:52,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:07:54,151] loop.py:45 INFO TRAIN_STEP step=2597 loss=1.770019 Training: 14%|█▍ | 2597/18018 [5:37:40<33:15:10, 7.76s/step, loss=1.770, acc=0.626, acc_len=1.46, thru=16.0, I=22.7, T=32.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:07:54,451] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:00,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:08:00,794] loop.py:45 INFO TRAIN_STEP step=2598 loss=1.736857 Training: 14%|█▍ | 2598/18018 [5:37:46<31:48:50, 7.43s/step, loss=1.737, acc=0.630, acc_len=1.47, thru=17.7, I=23.7, T=33.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:01,038] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 19:08:08,032] loop.py:45 INFO TRAIN_STEP step=2599 loss=1.796559 Training: 14%|█▍ | 2599/18018 [5:37:54<31:33:54, 7.37s/step, loss=1.797, acc=0.621, acc_len=1.45, thru=18.2, I=19.4, T=33.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:08,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:08,531] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:13,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:08:14,737] loop.py:380 INFO TIMING step=2600: step=3.716s data=1.622s compute=3.683s [fwd=2.656s bwd=1.013s opt=0.003s] dispatch=2.982s [2026-10-18 19:08:14,738] loop.py:45 INFO TRAIN_STEP step=2600 loss=1.808253 Training: 14%|█▍ | 2600/18018 [5:38:00<30:42:50, 7.17s/step, loss=1.808, acc=0.621, acc_len=1.45, thru=19.5, I=18.9, T=34.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:15,134] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:22,888] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:08:23,901] loop.py:45 INFO TRAIN_STEP step=2601 loss=1.714032 Training: 14%|█▍ | 2601/18018 [5:38:09<33:16:03, 7.77s/step, loss=1.714, acc=0.639, acc_len=1.50, thru=14.6, I=16.4, T=32.2, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:08:23,896] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2600: forward=2788.1ms backward=1144.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:24,399] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:29,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:08:30,630] loop.py:45 INFO TRAIN_STEP step=2602 loss=1.676465 Training: 14%|█▍ | 2602/18018 [5:38:16<31:56:03, 7.46s/step, loss=1.676, acc=0.642, acc_len=1.51, thru=18.0, I=21.8, T=30.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:30,997] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 19:08:38,551] loop.py:45 INFO TRAIN_STEP step=2603 loss=1.972466 Training: 14%|█▍ | 2603/18018 [5:38:24<32:31:26, 7.60s/step, loss=1.972, acc=0.598, acc_len=1.37, thru=15.3, I=15.7, T=27.5, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:44,141] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:08:45,636] loop.py:45 INFO TRAIN_STEP step=2604 loss=1.979460 Training: 14%|█▍ | 2604/18018 [5:38:31<31:52:02, 7.44s/step, loss=1.979, acc=0.598, acc_len=1.37, thru=19.2, I=17.4, T=29.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:45,667] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:51,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:08:53,121] loop.py:380 INFO TIMING step=2605: step=3.960s data=1.804s compute=3.904s [fwd=2.724s bwd=1.165s opt=0.003s] dispatch=3.516s [2026-10-18 19:08:53,122] loop.py:45 INFO TRAIN_STEP step=2605 loss=1.754045 Training: 14%|█▍ | 2605/18018 [5:38:39<31:55:11, 7.46s/step, loss=1.754, acc=0.632, acc_len=1.47, thru=15.0, I=18.3, T=32.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:53,395] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:08:59,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:09:00,291] loop.py:45 INFO TRAIN_STEP step=2606 loss=1.915434 Training: 14%|█▍ | 2606/18018 [5:38:46<31:32:59, 7.37s/step, loss=1.915, acc=0.609, acc_len=1.40, thru=16.9, I=21.6, T=34.1, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:09:00,287] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2605: forward=2640.0ms backward=1060.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:00,694] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:05,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:09:06,343] loop.py:45 INFO TRAIN_STEP step=2607 loss=1.810507 Training: 14%|█▍ | 2607/18018 [5:38:52<29:51:29, 6.97s/step, loss=1.811, acc=0.622, acc_len=1.45, thru=20.3, I=26.5, T=34.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:06,741] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:12,071] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:09:12,817] loop.py:45 INFO TRAIN_STEP step=2608 loss=1.791378 Training: 14%|█▍ | 2608/18018 [5:38:58<29:12:48, 6.82s/step, loss=1.791, acc=0.623, acc_len=1.44, thru=19.0, I=23.7, T=31.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:13,076] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:20,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:09:22,304] loop.py:45 INFO TRAIN_STEP step=2609 loss=1.964014 Training: 14%|█▍ | 2609/18018 [5:39:08<32:37:37, 7.62s/step, loss=1.964, acc=0.595, acc_len=1.36, thru=13.9, I=15.6, T=30.1, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:22,471] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:27,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:09:28,168] loop.py:380 INFO TIMING step=2610: step=3.641s data=1.549s compute=3.612s [fwd=2.613s bwd=0.985s opt=0.003s] dispatch=2.215s [2026-10-18 19:09:28,168] loop.py:45 INFO TRAIN_STEP step=2610 loss=1.945678 Training: 14%|█▍ | 2610/18018 [5:39:14<30:22:01, 7.10s/step, loss=1.946, acc=0.600, acc_len=1.38, thru=19.0, I=23.8, T=35.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:28,198] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:34,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:09:36,591] loop.py:45 INFO TRAIN_STEP step=2611 loss=1.818919 Training: 14%|█▍ | 2611/18018 [5:39:22<32:04:20, 7.49s/step, loss=1.819, acc=0.619, acc_len=1.44, thru=15.2, I=20.5, T=27.8, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:09:36,585] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2610: forward=3159.2ms backward=1391.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:36,819] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:41,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:09:42,966] loop.py:45 INFO TRAIN_STEP step=2612 loss=1.985245 Training: 14%|█▍ | 2612/18018 [5:39:29<30:38:02, 7.16s/step, loss=1.985, acc=0.594, acc_len=1.35, thru=18.2, I=23.6, T=32.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:43,204] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:09:51,262] loop.py:45 INFO TRAIN_STEP step=2613 loss=1.915436 Training: 15%|█▍ | 2613/18018 [5:39:37<32:05:21, 7.50s/step, loss=1.915, acc=0.605, acc_len=1.40, thru=13.9, I=16.6, T=33.7, wait=4.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:58,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:09:58,675] loop.py:45 INFO TRAIN_STEP step=2614 loss=1.694072 Training: 15%|█▍ | 2614/18018 [5:39:44<31:58:49, 7.47s/step, loss=1.694, acc=0.644, acc_len=1.52, thru=17.8, I=16.8, T=28.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:09:58,719] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 19:10:04,582] loop.py:380 INFO TIMING step=2615: step=3.857s data=1.689s compute=3.823s [fwd=2.694s bwd=1.114s opt=0.003s] dispatch=2.039s [2026-10-18 19:10:04,583] loop.py:45 INFO TRAIN_STEP step=2615 loss=1.884436 Training: 15%|█▍ | 2615/18018 [5:39:50<29:57:52, 7.00s/step, loss=1.884, acc=0.609, acc_len=1.41, thru=16.2, I=18.3, T=33.2, wait=2.0s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:14,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:10:14,619] loop.py:45 INFO TRAIN_STEP step=2616 loss=1.623825 Training: 15%|█▍ | 2616/18018 [5:40:00<33:51:32, 7.91s/step, loss=1.624, acc=0.650, acc_len=1.53, thru=16.6, I=15.2, T=28.2, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:10:14,614] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2615: forward=3187.3ms backward=1288.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:14,760] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 19:10:21,277] loop.py:45 INFO TRAIN_STEP step=2617 loss=1.949942 Training: 15%|█▍ | 2617/18018 [5:40:07<32:14:30, 7.54s/step, loss=1.950, acc=0.601, acc_len=1.38, thru=16.9, I=16.4, T=31.7, wait=2.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:28,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:10:29,827] loop.py:45 INFO TRAIN_STEP step=2618 loss=1.795205 Training: 15%|█▍ | 2618/18018 [5:40:15<33:32:33, 7.84s/step, loss=1.795, acc=0.626, acc_len=1.46, thru=17.4, I=18.0, T=30.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:30,013] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:36,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:10:36,650] loop.py:45 INFO TRAIN_STEP step=2619 loss=1.653899 Training: 15%|█▍ | 2619/18018 [5:40:22<32:14:04, 7.54s/step, loss=1.654, acc=0.642, acc_len=1.51, thru=18.0, I=20.6, T=31.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:36,778] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-18 19:10:42,495] loop.py:380 INFO TIMING step=2620: step=3.853s data=1.710s compute=3.818s [fwd=2.671s bwd=1.132s opt=0.003s] dispatch=1.981s [2026-10-18 19:10:42,496] loop.py:45 INFO TRAIN_STEP step=2620 loss=1.772930 Training: 15%|█▍ | 2620/18018 [5:40:28<30:04:02, 7.03s/step, loss=1.773, acc=0.628, acc_len=1.47, thru=17.7, I=20.5, T=33.2, wait=2.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:48,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:10:49,617] loop.py:45 INFO TRAIN_STEP step=2621 loss=1.885807 Training: 15%|█▍ | 2621/18018 [5:40:35<30:10:37, 7.06s/step, loss=1.886, acc=0.612, acc_len=1.42, thru=18.1, I=19.3, T=32.9, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:10:49,613] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2620: forward=2715.9ms backward=1129.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:50,065] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:56,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:10:58,179] loop.py:45 INFO TRAIN_STEP step=2622 loss=1.795938 Training: 15%|█▍ | 2622/18018 [5:40:44<32:06:32, 7.51s/step, loss=1.796, acc=0.627, acc_len=1.46, thru=15.8, I=19.7, T=31.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:10:58,352] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:03,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:11:04,360] loop.py:45 INFO TRAIN_STEP step=2623 loss=1.765779 Training: 15%|█▍ | 2623/18018 [5:40:50<30:24:10, 7.11s/step, loss=1.766, acc=0.625, acc_len=1.46, thru=18.6, I=24.0, T=33.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:04,799] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:11,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:11:12,815] loop.py:45 INFO TRAIN_STEP step=2624 loss=1.852752 Training: 15%|█▍ | 2624/18018 [5:40:58<32:07:45, 7.51s/step, loss=1.853, acc=0.612, acc_len=1.41, thru=14.9, I=17.7, T=27.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:13,285] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:18,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:11:20,267] loop.py:380 INFO TIMING step=2625: step=4.221s data=1.865s compute=4.184s [fwd=2.967s bwd=1.202s opt=0.003s] dispatch=3.219s [2026-10-18 19:11:20,267] loop.py:45 INFO TRAIN_STEP step=2625 loss=1.973002 Training: 15%|█▍ | 2625/18018 [5:41:06<32:02:50, 7.50s/step, loss=1.973, acc=0.592, acc_len=1.36, thru=15.8, I=18.8, T=30.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:20,419] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 19:11:27,087] loop.py:45 INFO TRAIN_STEP step=2626 loss=1.909898 Training: 15%|█▍ | 2626/18018 [5:41:13<31:10:31, 7.29s/step, loss=1.910, acc=0.602, acc_len=1.39, thru=16.4, I=17.9, T=31.4, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:11:27,083] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2625: forward=2813.0ms backward=1211.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:32,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:11:34,141] loop.py:45 INFO TRAIN_STEP step=2627 loss=1.818950 Training: 15%|█▍ | 2627/18018 [5:41:20<30:52:14, 7.22s/step, loss=1.819, acc=0.622, acc_len=1.45, thru=17.8, I=20.9, T=30.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:34,544] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 19:11:42,082] loop.py:45 INFO TRAIN_STEP step=2628 loss=1.741778 Training: 15%|█▍ | 2628/18018 [5:41:28<31:47:29, 7.44s/step, loss=1.742, acc=0.632, acc_len=1.48, thru=15.9, I=15.4, T=30.0, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:48,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:11:49,581] loop.py:45 INFO TRAIN_STEP step=2629 loss=1.741833 Training: 15%|█▍ | 2629/18018 [5:41:35<31:52:09, 7.46s/step, loss=1.742, acc=0.633, acc_len=1.48, thru=18.8, I=17.9, T=34.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:49,814] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:55,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:11:56,008] loop.py:380 INFO TIMING step=2630: step=3.926s data=1.792s compute=3.899s [fwd=2.801s bwd=1.083s opt=0.003s] dispatch=2.493s [2026-10-18 19:11:56,008] loop.py:45 INFO TRAIN_STEP step=2630 loss=1.913047 Training: 15%|█▍ | 2630/18018 [5:41:42<30:33:04, 7.15s/step, loss=1.913, acc=0.603, acc_len=1.39, thru=19.3, I=22.6, T=32.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:11:56,493] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (TrainerActor pid=9310) [2026-10-18 19:12:03,057] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2630: forward=2576.0ms backward=1099.9ms [2026-10-18 19:12:03,062] loop.py:45 INFO TRAIN_STEP step=2631 loss=1.839535 Training: 15%|█▍ | 2631/18018 [5:41:49<30:25:40, 7.12s/step, loss=1.840, acc=0.618, acc_len=1.43, thru=16.0, I=17.6, T=34.4, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:13,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:12:14,317] loop.py:45 INFO TRAIN_STEP step=2632 loss=1.938969 Training: 15%|█▍ | 2632/18018 [5:42:00<35:43:52, 8.36s/step, loss=1.939, acc=0.602, acc_len=1.40, thru=13.4, I=12.9, T=27.4, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:14,392] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:18,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:12:20,198] loop.py:45 INFO TRAIN_STEP step=2633 loss=1.798388 Training: 15%|█▍ | 2633/18018 [5:42:06<32:33:00, 7.62s/step, loss=1.798, acc=0.624, acc_len=1.45, thru=19.3, I=20.3, T=32.9, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:20,431] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:27,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:12:28,235] loop.py:45 INFO TRAIN_STEP step=2634 loss=1.679640 Training: 15%|█▍ | 2634/18018 [5:42:14<33:05:04, 7.74s/step, loss=1.680, acc=0.638, acc_len=1.50, thru=15.0, I=18.3, T=30.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:28,379] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 19:12:34,986] loop.py:380 INFO TIMING step=2635: step=3.724s data=1.702s compute=3.666s [fwd=2.631s bwd=1.020s opt=0.003s] dispatch=3.019s [2026-10-18 19:12:34,987] loop.py:45 INFO TRAIN_STEP step=2635 loss=1.818925 Training: 15%|█▍ | 2635/18018 [5:42:21<31:48:49, 7.45s/step, loss=1.819, acc=0.621, acc_len=1.44, thru=17.5, I=18.3, T=34.4, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:41,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:12:43,037] loop.py:45 INFO TRAIN_STEP step=2636 loss=1.785414 (TrainerActor pid=9310) [2026-10-18 19:12:43,033] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2635: forward=2810.4ms backward=1110.8ms Training: 15%|█▍ | 2636/18018 [5:42:29<32:35:19, 7.63s/step, loss=1.785, acc=0.622, acc_len=1.45, thru=16.0, I=18.7, T=32.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:43,300] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:49,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:12:50,698] loop.py:45 INFO TRAIN_STEP step=2637 loss=1.766608 Training: 15%|█▍ | 2637/18018 [5:42:36<32:37:39, 7.64s/step, loss=1.767, acc=0.630, acc_len=1.47, thru=14.9, I=20.5, T=30.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:50,860] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:56,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:12:57,649] loop.py:45 INFO TRAIN_STEP step=2638 loss=1.825135 Training: 15%|█▍ | 2638/18018 [5:42:43<31:44:48, 7.43s/step, loss=1.825, acc=0.619, acc_len=1.44, thru=17.7, I=21.2, T=31.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:12:57,905] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:04,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:13:04,777] loop.py:45 INFO TRAIN_STEP step=2639 loss=1.731960 Training: 15%|█▍ | 2639/18018 [5:42:50<31:21:29, 7.34s/step, loss=1.732, acc=0.633, acc_len=1.48, thru=17.0, I=20.0, T=32.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:04,934] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:10,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:13:12,592] loop.py:380 INFO TIMING step=2640: step=3.966s data=1.775s compute=3.914s [fwd=2.786s bwd=1.112s opt=0.003s] dispatch=3.839s [2026-10-18 19:13:12,593] loop.py:45 INFO TRAIN_STEP step=2640 loss=1.864751 Training: 15%|█▍ | 2640/18018 [5:42:58<31:57:50, 7.48s/step, loss=1.865, acc=0.611, acc_len=1.41, thru=16.2, I=19.3, T=32.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:12,995] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:19,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:13:19,872] loop.py:45 INFO TRAIN_STEP step=2641 loss=1.845716 Training: 15%|█▍ | 2641/18018 [5:43:05<31:42:04, 7.42s/step, loss=1.846, acc=0.616, acc_len=1.42, thru=15.8, I=22.4, T=31.2, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:13:19,868] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2640: forward=2837.5ms backward=1210.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:19,970] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:27,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:13:28,442] loop.py:45 INFO TRAIN_STEP step=2642 loss=1.874694 Training: 15%|█▍ | 2642/18018 [5:43:14<33:10:21, 7.77s/step, loss=1.875, acc=0.613, acc_len=1.41, thru=14.3, I=14.6, T=29.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:28,562] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 19:13:35,210] loop.py:45 INFO TRAIN_STEP step=2643 loss=1.847301 Training: 15%|█▍ | 2643/18018 [5:43:21<31:53:21, 7.47s/step, loss=1.847, acc=0.616, acc_len=1.42, thru=16.6, I=18.0, T=29.9, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:41,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:13:42,165] loop.py:45 INFO TRAIN_STEP step=2644 loss=1.909078 Training: 15%|█▍ | 2644/18018 [5:43:28<31:13:49, 7.31s/step, loss=1.909, acc=0.603, acc_len=1.39, thru=17.6, I=19.2, T=33.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:42,529] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:50,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:13:53,373] loop.py:380 INFO TIMING step=2645: step=4.686s data=2.096s compute=4.655s [fwd=3.231s bwd=1.407s opt=0.003s] dispatch=6.512s [2026-10-18 19:13:53,373] loop.py:45 INFO TRAIN_STEP step=2645 loss=1.737151 Training: 15%|█▍ | 2645/18018 [5:43:39<36:13:15, 8.48s/step, loss=1.737, acc=0.634, acc_len=1.49, thru=11.3, I=14.1, T=27.3, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:53,427] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:13:59,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:14:00,082] loop.py:45 INFO TRAIN_STEP step=2646 loss=1.940309 Training: 15%|█▍ | 2646/18018 [5:43:46<33:56:54, 7.95s/step, loss=1.940, acc=0.598, acc_len=1.38, thru=16.0, I=22.7, T=31.1, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:14:00,074] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2645: forward=2831.5ms backward=1239.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:14:00,353] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:14:06,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:14:08,114] loop.py:45 INFO TRAIN_STEP step=2647 loss=1.524956 Training: 15%|█▍ | 2647/18018 [5:43:54<34:02:48, 7.97s/step, loss=1.525, acc=0.664, acc_len=1.58, thru=14.9, I=17.4, T=28.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:14:08,302] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:14:14,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:14:15,840] loop.py:45 INFO TRAIN_STEP step=2648 loss=1.926328 Training: 15%|█▍ | 2648/18018 [5:44:01<33:43:48, 7.90s/step, loss=1.926, acc=0.602, acc_len=1.39, thru=14.6, I=17.9, T=30.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:14:15,894] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-18 19:14:21,667] loop.py:45 INFO TRAIN_STEP step=2649 loss=1.773307 Training: 15%|█▍ | 2649/18018 [5:44:07<31:04:21, 7.28s/step, loss=1.773, acc=0.627, acc_len=1.46, thru=17.6, I=24.0, T=34.8, wait=2.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:14:29,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:14:31,364] loop.py:380 INFO TIMING step=2650: step=4.609s data=2.157s compute=4.583s [fwd=3.261s bwd=1.307s opt=0.003s] dispatch=5.076s [2026-10-18 19:14:31,364] loop.py:45 INFO TRAIN_STEP step=2650 loss=1.764831 Training: 15%|█▍ | 2650/18018 [5:44:17<34:09:59, 8.00s/step, loss=1.765, acc=0.629, acc_len=1.47, thru=14.5, I=14.8, T=27.8, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:14:31,482] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 19:14:40,066] loop.py:45 INFO TRAIN_STEP step=2651 loss=1.901048 Training: 15%|█▍ | 2651/18018 [5:44:26<35:03:24, 8.21s/step, loss=1.901, acc=0.612, acc_len=1.41, thru=13.5, I=16.7, T=28.3, wait=4.2s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:14:40,062] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2650: forward=3182.4ms backward=1300.4ms [2026-10-18 19:14:49,926] loop.py:45 INFO TRAIN_STEP step=2652 loss=1.753015 Training: 15%|█▍ | 2652/18018 [5:44:36<37:10:00, 8.71s/step, loss=1.753, acc=0.625, acc_len=1.47, thru=11.7, I=12.7, T=28.4, wait=5.3s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:14:57,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:15:00,386] loop.py:45 INFO TRAIN_STEP step=2653 loss=1.948357 Training: 15%|█▍ | 2653/18018 [5:44:46<39:24:25, 9.23s/step, loss=1.948, acc=0.598, acc_len=1.37, thru=15.8, I=15.0, T=26.4, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:00,706] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:05,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:15:06,538] loop.py:45 INFO TRAIN_STEP step=2654 loss=1.639334 Training: 15%|█▍ | 2654/18018 [5:44:52<35:27:37, 8.31s/step, loss=1.639, acc=0.648, acc_len=1.53, thru=17.1, I=25.3, T=34.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:06,722] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:11,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:15:13,201] loop.py:380 INFO TIMING step=2655: step=3.805s data=1.721s compute=3.752s [fwd=2.652s bwd=1.085s opt=0.003s] dispatch=2.846s [2026-10-18 19:15:13,201] loop.py:45 INFO TRAIN_STEP step=2655 loss=1.895661 Training: 15%|█▍ | 2655/18018 [5:44:59<33:21:01, 7.81s/step, loss=1.896, acc=0.611, acc_len=1.41, thru=18.1, I=23.0, T=33.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:13,389] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:19,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:15:20,440] loop.py:45 INFO TRAIN_STEP step=2656 loss=1.732241 Training: 15%|█▍ | 2656/18018 [5:45:06<32:36:42, 7.64s/step, loss=1.732, acc=0.633, acc_len=1.49, thru=17.2, I=22.0, T=29.4, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:15:20,435] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2655: forward=3035.7ms backward=1265.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:20,573] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-18 19:15:27,806] loop.py:45 INFO TRAIN_STEP step=2657 loss=1.788158 Training: 15%|█▍ | 2657/18018 [5:45:13<32:15:20, 7.56s/step, loss=1.788, acc=0.626, acc_len=1.45, thru=16.8, I=17.9, T=30.8, wait=3.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:34,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:15:35,649] loop.py:45 INFO TRAIN_STEP step=2658 loss=1.907822 Training: 15%|█▍ | 2658/18018 [5:45:21<32:36:49, 7.64s/step, loss=1.908, acc=0.607, acc_len=1.40, thru=17.4, I=18.4, T=31.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:35,713] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:41,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:15:42,004] loop.py:45 INFO TRAIN_STEP step=2659 loss=1.781990 Training: 15%|█▍ | 2659/18018 [5:45:28<30:57:48, 7.26s/step, loss=1.782, acc=0.624, acc_len=1.45, thru=18.5, I=21.5, T=34.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:42,146] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:47,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:15:48,875] loop.py:380 INFO TIMING step=2660: step=4.057s data=1.853s compute=4.031s [fwd=2.845s bwd=1.172s opt=0.003s] dispatch=2.804s [2026-10-18 19:15:48,875] loop.py:45 INFO TRAIN_STEP step=2660 loss=1.815037 Training: 15%|█▍ | 2660/18018 [5:45:34<30:28:04, 7.14s/step, loss=1.815, acc=0.622, acc_len=1.45, thru=17.8, I=21.3, T=31.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:48,932] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:57,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:15:58,486] loop.py:45 INFO TRAIN_STEP step=2661 loss=1.624970 Training: 15%|█▍ | 2661/18018 [5:45:44<33:37:26, 7.88s/step, loss=1.625, acc=0.651, acc_len=1.54, thru=13.4, I=13.8, T=28.3, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:15:58,481] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2660: forward=3159.1ms backward=1318.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:15:58,513] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:04,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:16:05,191] loop.py:45 INFO TRAIN_STEP step=2662 loss=1.734425 Training: 15%|█▍ | 2662/18018 [5:45:51<32:06:55, 7.53s/step, loss=1.734, acc=0.636, acc_len=1.49, thru=18.2, I=18.1, T=31.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:05,450] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:11,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:16:13,583] loop.py:45 INFO TRAIN_STEP step=2663 loss=1.702634 Training: 15%|█▍ | 2663/18018 [5:45:59<33:13:10, 7.79s/step, loss=1.703, acc=0.637, acc_len=1.50, thru=15.1, I=18.7, T=30.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:13,909] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:16:22,583] loop.py:45 INFO TRAIN_STEP step=2664 loss=1.628891 Training: 15%|█▍ | 2664/18018 [5:46:08<34:46:03, 8.15s/step, loss=1.629, acc=0.651, acc_len=1.55, thru=13.8, I=14.6, T=30.9, wait=4.9s, pool=56, epoch=1/2][2026-10-18 19:16:29,868] loop.py:380 INFO TIMING step=2665: step=4.180s data=1.950s compute=4.150s [fwd=2.956s bwd=1.179s opt=0.003s] dispatch=3.094s [2026-10-18 19:16:29,868] loop.py:45 INFO TRAIN_STEP step=2665 loss=1.770619 Training: 15%|█▍ | 2665/18018 [5:46:15<33:39:12, 7.89s/step, loss=1.771, acc=0.627, acc_len=1.46, thru=13.9, I=14.6, T=30.6, wait=3.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:39,241] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:16:40,010] loop.py:45 INFO TRAIN_STEP step=2666 loss=1.819314 Training: 15%|█▍ | 2666/18018 [5:46:26<36:32:01, 8.57s/step, loss=1.819, acc=0.616, acc_len=1.43, thru=15.0, I=15.7, T=29.4, wait=5.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:16:40,005] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2665: forward=3076.9ms backward=1217.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:40,245] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:45,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:16:46,094] loop.py:45 INFO TRAIN_STEP step=2667 loss=1.958300 Training: 15%|█▍ | 2667/18018 [5:46:32<33:21:14, 7.82s/step, loss=1.958, acc=0.599, acc_len=1.37, thru=18.5, I=21.6, T=32.4, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:46,285] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:55,797] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:16:56,486] loop.py:45 INFO TRAIN_STEP step=2668 loss=1.850832 Training: 15%|█▍ | 2668/18018 [5:46:42<36:38:33, 8.59s/step, loss=1.851, acc=0.615, acc_len=1.43, thru=12.5, I=13.4, T=26.0, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:16:56,802] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 19:17:04,211] loop.py:45 INFO TRAIN_STEP step=2669 loss=1.848004 Training: 15%|█▍ | 2669/18018 [5:46:50<35:31:27, 8.33s/step, loss=1.848, acc=0.611, acc_len=1.41, thru=16.1, I=18.7, T=30.1, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:10,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:17:11,168] loop.py:380 INFO TIMING step=2670: step=4.059s data=1.847s compute=4.032s [fwd=2.883s bwd=1.134s opt=0.003s] dispatch=2.889s [2026-10-18 19:17:11,168] loop.py:45 INFO TRAIN_STEP step=2670 loss=1.878542 Training: 15%|█▍ | 2670/18018 [5:46:57<33:45:43, 7.92s/step, loss=1.879, acc=0.613, acc_len=1.42, thru=17.7, I=19.5, T=31.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:11,628] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:16,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:17:18,036] loop.py:45 INFO TRAIN_STEP step=2671 loss=1.916280 Training: 15%|█▍ | 2671/18018 [5:47:04<32:24:59, 7.60s/step, loss=1.916, acc=0.605, acc_len=1.40, thru=18.1, I=23.9, T=31.2, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:17:18,031] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2670: forward=2831.2ms backward=1222.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:18,319] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:17:24,140] loop.py:45 INFO TRAIN_STEP step=2672 loss=1.739974 Training: 15%|█▍ | 2672/18018 [5:47:10<30:29:45, 7.15s/step, loss=1.740, acc=0.632, acc_len=1.48, thru=18.6, I=25.6, T=32.6, wait=2.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:31,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:17:32,496] loop.py:45 INFO TRAIN_STEP step=2673 loss=1.882084 Training: 15%|█▍ | 2673/18018 [5:47:18<32:01:52, 7.51s/step, loss=1.882, acc=0.611, acc_len=1.41, thru=14.2, I=16.5, T=31.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:32,914] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:39,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:17:40,294] loop.py:45 INFO TRAIN_STEP step=2674 loss=1.751023 Training: 15%|█▍ | 2674/18018 [5:47:26<32:23:34, 7.60s/step, loss=1.751, acc=0.633, acc_len=1.48, thru=14.8, I=18.3, T=29.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:40,571] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:46,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:17:49,435] loop.py:380 INFO TIMING step=2675: step=5.107s data=2.545s compute=5.062s [fwd=3.493s bwd=1.554s opt=0.003s] dispatch=4.023s [2026-10-18 19:17:49,435] loop.py:45 INFO TRAIN_STEP step=2675 loss=1.922811 Training: 15%|█▍ | 2675/18018 [5:47:35<34:21:39, 8.06s/step, loss=1.923, acc=0.610, acc_len=1.41, thru=14.4, I=20.8, T=25.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:49,634] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:58,293] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:17:58,935] loop.py:45 INFO TRAIN_STEP step=2676 loss=1.882773 Training: 15%|█▍ | 2676/18018 [5:47:45<36:11:43, 8.49s/step, loss=1.883, acc=0.611, acc_len=1.41, thru=13.8, I=14.7, T=26.7, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:17:58,931] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2675: forward=3359.5ms backward=1389.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:17:59,297] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:04,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:18:06,951] loop.py:45 INFO TRAIN_STEP step=2677 loss=1.935964 Training: 15%|█▍ | 2677/18018 [5:47:53<35:35:09, 8.35s/step, loss=1.936, acc=0.606, acc_len=1.39, thru=15.6, I=18.3, T=29.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:07,372] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:14,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:18:15,320] loop.py:45 INFO TRAIN_STEP step=2678 loss=1.802748 Training: 15%|█▍ | 2678/18018 [5:48:01<35:36:13, 8.36s/step, loss=1.803, acc=0.621, acc_len=1.45, thru=16.1, I=16.7, T=29.9, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:15,481] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 19:18:22,558] loop.py:45 INFO TRAIN_STEP step=2679 loss=1.816711 Training: 15%|█▍ | 2679/18018 [5:48:08<34:10:24, 8.02s/step, loss=1.817, acc=0.623, acc_len=1.45, thru=16.1, I=15.3, T=33.0, wait=3.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:30,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:18:31,636] loop.py:380 INFO TIMING step=2680: step=4.219s data=1.845s compute=4.182s [fwd=2.901s bwd=1.266s opt=0.003s] dispatch=4.849s [2026-10-18 19:18:31,636] loop.py:45 INFO TRAIN_STEP step=2680 loss=1.667003 Training: 15%|█▍ | 2680/18018 [5:48:17<35:31:28, 8.34s/step, loss=1.667, acc=0.644, acc_len=1.52, thru=16.4, I=18.2, T=30.3, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:31,684] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:37,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:18:38,050] loop.py:45 INFO TRAIN_STEP step=2681 loss=1.709391 Training: 15%|█▍ | 2681/18018 [5:48:24<33:03:38, 7.76s/step, loss=1.709, acc=0.632, acc_len=1.48, thru=18.4, I=22.7, T=32.6, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:18:38,045] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2680: forward=2729.7ms backward=1150.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:38,168] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-18 19:18:44,728] loop.py:45 INFO TRAIN_STEP step=2682 loss=1.762382 Training: 15%|█▍ | 2682/18018 [5:48:30<31:40:40, 7.44s/step, loss=1.762, acc=0.629, acc_len=1.47, thru=17.0, I=19.9, T=33.0, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:51,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:18:52,209] loop.py:45 INFO TRAIN_STEP step=2683 loss=1.845271 Training: 15%|█▍ | 2683/18018 [5:48:38<31:43:51, 7.45s/step, loss=1.845, acc=0.618, acc_len=1.44, thru=17.0, I=17.7, T=29.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:52,462] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:18:59,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:19:01,080] loop.py:45 INFO TRAIN_STEP step=2684 loss=1.926829 Training: 15%|█▍ | 2684/18018 [5:48:47<33:32:55, 7.88s/step, loss=1.927, acc=0.605, acc_len=1.39, thru=14.1, I=17.9, T=29.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:01,239] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:10,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:19:10,849] loop.py:380 INFO TIMING step=2685: step=4.575s data=2.117s compute=4.545s [fwd=3.190s bwd=1.339s opt=0.003s] dispatch=5.184s [2026-10-18 19:19:10,849] loop.py:45 INFO TRAIN_STEP step=2685 loss=1.705096 Training: 15%|█▍ | 2685/18018 [5:48:56<35:57:46, 8.44s/step, loss=1.705, acc=0.640, acc_len=1.50, thru=13.3, I=13.4, T=28.0, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:11,026] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:17,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:19:18,875] loop.py:45 INFO TRAIN_STEP step=2686 loss=1.788788 Training: 15%|█▍ | 2686/18018 [5:49:04<35:25:41, 8.32s/step, loss=1.789, acc=0.631, acc_len=1.47, thru=15.0, I=16.2, T=30.0, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:19:18,870] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2685: forward=2989.6ms backward=1225.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:19,335] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:26,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:19:26,410] loop.py:45 INFO TRAIN_STEP step=2687 loss=1.769742 Training: 15%|█▍ | 2687/18018 [5:49:12<34:25:29, 8.08s/step, loss=1.770, acc=0.623, acc_len=1.46, thru=15.7, I=18.3, T=32.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:26,741] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:32,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:19:34,238] loop.py:45 INFO TRAIN_STEP step=2688 loss=1.832801 Training: 15%|█▍ | 2688/18018 [5:49:20<34:05:46, 8.01s/step, loss=1.833, acc=0.616, acc_len=1.43, thru=17.4, I=19.3, T=29.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:34,485] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:39,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:19:40,458] loop.py:45 INFO TRAIN_STEP step=2689 loss=1.743633 Training: 15%|█▍ | 2689/18018 [5:49:26<31:48:32, 7.47s/step, loss=1.744, acc=0.630, acc_len=1.47, thru=19.2, I=21.2, T=34.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:40,485] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:47,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:19:48,020] loop.py:380 INFO TIMING step=2690: step=3.984s data=1.781s compute=3.951s [fwd=2.830s bwd=1.107s opt=0.003s] dispatch=3.570s [2026-10-18 19:19:48,020] loop.py:45 INFO TRAIN_STEP step=2690 loss=1.912644 Training: 15%|█▍ | 2690/18018 [5:49:34<31:55:39, 7.50s/step, loss=1.913, acc=0.601, acc_len=1.38, thru=16.1, I=18.3, T=32.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:19:48,152] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-18 19:19:54,415] loop.py:45 INFO TRAIN_STEP step=2691 loss=1.715392 Training: 15%|█▍ | 2691/18018 [5:49:40<30:30:41, 7.17s/step, loss=1.715, acc=0.634, acc_len=1.48, thru=18.9, I=19.8, T=35.3, wait=2.8s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:19:54,410] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2690: forward=2527.6ms backward=1053.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:20:01,108] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:20:02,536] loop.py:45 INFO TRAIN_STEP step=2692 loss=1.928941 Training: 15%|█▍ | 2692/18018 [5:49:48<31:43:55, 7.45s/step, loss=1.929, acc=0.604, acc_len=1.38, thru=16.6, I=18.4, T=25.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:20:02,635] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 19:20:10,728] loop.py:45 INFO TRAIN_STEP step=2693 loss=1.871036 Training: 15%|█▍ | 2693/18018 [5:49:56<32:40:11, 7.67s/step, loss=1.871, acc=0.616, acc_len=1.42, thru=13.3, I=14.5, T=30.3, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:20:19,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:20:21,132] loop.py:45 INFO TRAIN_STEP step=2694 loss=1.971583 Training: 15%|█▍ | 2694/18018 [5:50:07<36:09:24, 8.49s/step, loss=1.972, acc=0.598, acc_len=1.37, thru=13.2, I=15.2, T=26.9, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:20:21,427] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:20:29,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:20:29,186] loop.py:380 INFO TIMING step=2695: step=4.611s data=2.342s compute=4.577s [fwd=3.345s bwd=1.218s opt=0.003s] dispatch=3.432s [2026-10-18 19:20:29,186] loop.py:45 INFO TRAIN_STEP step=2695 loss=1.924579 Training: 15%|█▍ | 2695/18018 [5:50:15<35:35:15, 8.36s/step, loss=1.925, acc=0.602, acc_len=1.38, thru=13.7, I=16.0, T=27.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:20:29,614] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 19:20:35,862] loop.py:45 INFO TRAIN_STEP step=2696 loss=1.837337 Training: 15%|█▍ | 2696/18018 [5:50:21<33:26:07, 7.86s/step, loss=1.837, acc=0.619, acc_len=1.43, thru=16.2, I=15.6, T=33.9, wait=2.9s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:20:35,857] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2695: forward=2664.0ms backward=1066.7ms [2026-10-18 19:20:44,460] loop.py:45 INFO TRAIN_STEP step=2697 loss=1.882164 Training: 15%|█▍ | 2697/18018 [5:50:30<34:22:47, 8.08s/step, loss=1.882, acc=0.612, acc_len=1.42, thru=14.9, I=15.3, T=29.6, wait=4.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:20:53,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:20:54,865] loop.py:45 INFO TRAIN_STEP step=2698 loss=1.764239 Training: 15%|█▍ | 2698/18018 [5:50:40<37:20:57, 8.78s/step, loss=1.764, acc=0.626, acc_len=1.46, thru=16.4, I=14.3, T=27.1, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:20:55,247] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:02,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:21:03,408] loop.py:45 INFO TRAIN_STEP step=2699 loss=1.823069 Training: 15%|█▍ | 2699/18018 [5:50:49<37:02:55, 8.71s/step, loss=1.823, acc=0.622, acc_len=1.44, thru=13.7, I=17.2, T=29.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:03,913] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:08,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:21:09,523] loop.py:380 INFO TIMING step=2700: step=3.652s data=1.564s compute=3.622s [fwd=2.581s bwd=1.027s opt=0.003s] dispatch=2.454s [2026-10-18 19:21:09,524] loop.py:45 INFO TRAIN_STEP step=2700 loss=1.795172 Training: 15%|█▍ | 2700/18018 [5:50:55<33:44:26, 7.93s/step, loss=1.795, acc=0.625, acc_len=1.46, thru=19.2, I=22.2, T=35.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:09,734] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:17,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:21:17,394] loop.py:45 INFO TRAIN_STEP step=2701 loss=1.627797 Training: 15%|█▍ | 2701/18018 [5:51:03<33:39:42, 7.91s/step, loss=1.628, acc=0.645, acc_len=1.53, thru=16.9, I=17.8, T=29.6, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:21:17,389] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2700: forward=2987.3ms backward=1291.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:17,570] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:23,688] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:21:23,867] loop.py:45 INFO TRAIN_STEP step=2702 loss=1.698809 Training: 15%|█▍ | 2702/18018 [5:51:09<31:49:29, 7.48s/step, loss=1.699, acc=0.639, acc_len=1.50, thru=18.7, I=18.1, T=32.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:24,190] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:32,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:21:34,731] loop.py:45 INFO TRAIN_STEP step=2703 loss=1.803896 Training: 15%|█▌ | 2703/18018 [5:51:20<36:08:30, 8.50s/step, loss=1.804, acc=0.621, acc_len=1.45, thru=11.8, I=13.7, T=26.4, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:34,902] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:41,091] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:21:42,028] loop.py:45 INFO TRAIN_STEP step=2704 loss=1.778310 Training: 15%|█▌ | 2704/18018 [5:51:28<34:36:32, 8.14s/step, loss=1.778, acc=0.629, acc_len=1.47, thru=15.1, I=19.5, T=28.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:42,103] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:47,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:21:49,735] loop.py:380 INFO TIMING step=2705: step=5.148s data=2.525s compute=5.119s [fwd=3.519s bwd=1.586s opt=0.003s] dispatch=2.548s [2026-10-18 19:21:49,735] loop.py:45 INFO TRAIN_STEP step=2705 loss=1.776451 Training: 15%|█▌ | 2705/18018 [5:51:35<34:03:33, 8.01s/step, loss=1.776, acc=0.625, acc_len=1.46, thru=16.5, I=20.4, T=24.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:21:50,022] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 19:21:57,489] loop.py:45 INFO TRAIN_STEP step=2706 loss=1.776062 Training: 15%|█▌ | 2706/18018 [5:51:43<33:43:55, 7.93s/step, loss=1.776, acc=0.630, acc_len=1.47, thru=16.1, I=16.0, T=31.9, wait=3.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:21:57,484] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2705: forward=2817.3ms backward=1154.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:05,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:22:07,179] loop.py:45 INFO TRAIN_STEP step=2707 loss=1.944269 Training: 15%|█▌ | 2707/18018 [5:51:53<35:58:37, 8.46s/step, loss=1.944, acc=0.601, acc_len=1.38, thru=14.5, I=17.8, T=28.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:07,620] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:13,098] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:22:14,280] loop.py:45 INFO TRAIN_STEP step=2708 loss=1.838084 Training: 15%|█▌ | 2708/18018 [5:52:00<34:14:31, 8.05s/step, loss=1.838, acc=0.619, acc_len=1.43, thru=14.4, I=23.0, T=34.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:14,608] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:20,459] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:22:21,260] loop.py:45 INFO TRAIN_STEP step=2709 loss=1.678311 Training: 15%|█▌ | 2709/18018 [5:52:07<32:52:15, 7.73s/step, loss=1.678, acc=0.640, acc_len=1.50, thru=17.5, I=23.0, T=30.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:21,469] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:27,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:22:28,061] loop.py:380 INFO TIMING step=2710: step=4.430s data=1.963s compute=4.401s [fwd=3.164s bwd=1.222s opt=0.003s] dispatch=2.361s [2026-10-18 19:22:28,061] loop.py:45 INFO TRAIN_STEP step=2710 loss=1.817979 Training: 15%|█▌ | 2710/18018 [5:52:14<31:41:06, 7.45s/step, loss=1.818, acc=0.620, acc_len=1.44, thru=16.2, I=18.4, T=28.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:28,473] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:35,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 19:22:35,473] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2710: forward=2752.4ms backward=1131.5ms [2026-10-18 19:22:35,479] loop.py:45 INFO TRAIN_STEP step=2711 loss=1.799803 Training: 15%|█▌ | 2711/18018 [5:52:21<31:38:19, 7.44s/step, loss=1.800, acc=0.622, acc_len=1.44, thru=16.3, I=17.7, T=32.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:35,527] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:41,893] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:22:42,789] loop.py:45 INFO TRAIN_STEP step=2712 loss=1.957545 Training: 15%|█▌ | 2712/18018 [5:52:28<31:28:22, 7.40s/step, loss=1.958, acc=0.593, acc_len=1.37, thru=17.8, I=19.9, T=32.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:42,912] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:53,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:22:53,837] loop.py:45 INFO TRAIN_STEP step=2713 loss=1.716385 Training: 15%|█▌ | 2713/18018 [5:52:39<36:07:06, 8.50s/step, loss=1.716, acc=0.635, acc_len=1.48, thru=12.0, I=17.3, T=31.2, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:54,276] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:22:59,192] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:23:00,256] loop.py:45 INFO TRAIN_STEP step=2714 loss=1.808423 Training: 15%|█▌ | 2714/18018 [5:52:46<33:27:56, 7.87s/step, loss=1.808, acc=0.623, acc_len=1.45, thru=18.5, I=23.2, T=30.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:00,701] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:06,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:23:06,853] loop.py:380 INFO TIMING step=2715: step=4.070s data=1.852s compute=4.041s [fwd=2.842s bwd=1.185s opt=0.003s] dispatch=2.518s [2026-10-18 19:23:06,853] loop.py:45 INFO TRAIN_STEP step=2715 loss=1.648805 Training: 15%|█▌ | 2715/18018 [5:52:52<31:50:18, 7.49s/step, loss=1.649, acc=0.648, acc_len=1.53, thru=17.0, I=19.8, T=31.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:07,250] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:12,943] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:23:13,415] loop.py:45 INFO TRAIN_STEP step=2716 loss=1.765455 Training: 15%|█▌ | 2716/18018 [5:52:59<30:39:17, 7.21s/step, loss=1.765, acc=0.632, acc_len=1.48, thru=18.5, I=20.9, T=33.3, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:23:13,410] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2715: forward=2680.4ms backward=1121.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:13,448] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:18,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:23:20,892] loop.py:45 INFO TRAIN_STEP step=2717 loss=1.890150 Training: 15%|█▌ | 2717/18018 [5:53:06<30:59:13, 7.29s/step, loss=1.890, acc=0.607, acc_len=1.40, thru=17.5, I=22.1, T=29.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:21,393] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:28,307] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:23:29,579] loop.py:45 INFO TRAIN_STEP step=2718 loss=1.890723 Training: 15%|█▌ | 2718/18018 [5:53:15<32:46:10, 7.71s/step, loss=1.891, acc=0.608, acc_len=1.40, thru=15.6, I=18.5, T=30.6, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:29,816] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:35,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:23:35,314] loop.py:45 INFO TRAIN_STEP step=2719 loss=1.865318 Training: 15%|█▌ | 2719/18018 [5:53:21<30:14:43, 7.12s/step, loss=1.865, acc=0.610, acc_len=1.41, thru=20.0, I=21.1, T=33.3, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:35,537] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:42,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:23:42,661] loop.py:380 INFO TIMING step=2720: step=4.523s data=2.089s compute=4.491s [fwd=3.157s bwd=1.319s opt=0.003s] dispatch=2.814s [2026-10-18 19:23:42,661] loop.py:45 INFO TRAIN_STEP step=2720 loss=1.734852 Training: 15%|█▌ | 2720/18018 [5:53:28<30:32:11, 7.19s/step, loss=1.735, acc=0.632, acc_len=1.49, thru=17.2, I=17.8, T=28.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:42,820] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:51,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:23:51,903] loop.py:45 INFO TRAIN_STEP step=2721 loss=1.840492 Training: 15%|█▌ | 2721/18018 [5:53:37<33:09:21, 7.80s/step, loss=1.840, acc=0.615, acc_len=1.42, thru=14.0, I=13.9, T=27.9, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:23:51,898] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2720: forward=3212.9ms backward=1311.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:52,251] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:23:59,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:24:00,714] loop.py:45 INFO TRAIN_STEP step=2722 loss=1.849883 Training: 15%|█▌ | 2722/18018 [5:53:46<34:26:27, 8.11s/step, loss=1.850, acc=0.617, acc_len=1.43, thru=14.4, I=17.1, T=25.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:01,055] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:07,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:24:07,699] loop.py:45 INFO TRAIN_STEP step=2723 loss=1.791278 Training: 15%|█▌ | 2723/18018 [5:53:53<33:00:52, 7.77s/step, loss=1.791, acc=0.628, acc_len=1.46, thru=17.1, I=21.1, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:08,033] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 19:24:15,716] loop.py:45 INFO TRAIN_STEP step=2724 loss=1.778386 Training: 15%|█▌ | 2724/18018 [5:54:01<33:19:15, 7.84s/step, loss=1.778, acc=0.629, acc_len=1.46, thru=14.0, I=15.0, T=33.2, wait=4.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:24,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:24:25,749] loop.py:380 INFO TIMING step=2725: step=4.145s data=1.807s compute=4.113s [fwd=2.963s bwd=1.136s opt=0.003s] dispatch=5.878s [2026-10-18 19:24:25,750] loop.py:45 INFO TRAIN_STEP step=2725 loss=1.929409 Training: 15%|█▌ | 2725/18018 [5:54:11<36:06:36, 8.50s/step, loss=1.929, acc=0.604, acc_len=1.39, thru=15.8, I=15.5, T=30.9, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:26,210] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:31,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:24:32,806] loop.py:45 INFO TRAIN_STEP step=2726 loss=1.791656 Training: 15%|█▌ | 2726/18018 [5:54:18<34:16:07, 8.07s/step, loss=1.792, acc=0.627, acc_len=1.46, thru=17.4, I=21.8, T=31.0, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:24:32,801] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2725: forward=2857.1ms backward=1225.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:33,004] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:37,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:24:39,046] loop.py:45 INFO TRAIN_STEP step=2727 loss=1.616455 Training: 15%|█▌ | 2727/18018 [5:54:25<31:56:15, 7.52s/step, loss=1.616, acc=0.647, acc_len=1.53, thru=18.0, I=26.5, T=28.3, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:39,159] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:45,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:24:46,098] loop.py:45 INFO TRAIN_STEP step=2728 loss=1.764338 Training: 15%|█▌ | 2728/18018 [5:54:32<31:20:19, 7.38s/step, loss=1.764, acc=0.628, acc_len=1.47, thru=15.2, I=21.3, T=29.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:46,223] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:51,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:24:52,902] loop.py:45 INFO TRAIN_STEP step=2729 loss=1.767786 Training: 15%|█▌ | 2729/18018 [5:54:38<30:36:24, 7.21s/step, loss=1.768, acc=0.629, acc_len=1.47, thru=18.2, I=23.0, T=33.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:53,108] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:24:57,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:25:00,255] loop.py:380 INFO TIMING step=2730: step=4.982s data=2.243s compute=4.945s [fwd=3.472s bwd=1.458s opt=0.003s] dispatch=2.359s [2026-10-18 19:25:00,255] loop.py:45 INFO TRAIN_STEP step=2730 loss=1.780331 Training: 15%|█▌ | 2730/18018 [5:54:46<30:47:27, 7.25s/step, loss=1.780, acc=0.625, acc_len=1.45, thru=16.1, I=25.1, T=25.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:25:00,294] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 [2026-10-18 19:25:07,806] loop.py:45 INFO TRAIN_STEP step=2731 loss=1.779639 Training: 15%|█▌ | 2731/18018 [5:54:53<31:10:05, 7.34s/step, loss=1.780, acc=0.623, acc_len=1.45, thru=14.9, I=14.3, T=30.6, wait=3.4s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:25:07,801] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2730: forward=2949.0ms backward=1186.0ms [2026-10-18 19:25:15,591] loop.py:45 INFO TRAIN_STEP step=2732 loss=1.803786 Training: 15%|█▌ | 2732/18018 [5:55:01<31:44:02, 7.47s/step, loss=1.804, acc=0.625, acc_len=1.45, thru=15.4, I=15.5, T=31.7, wait=3.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:25:25,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:25:26,127] loop.py:45 INFO TRAIN_STEP step=2733 loss=1.658219 Training: 15%|█▌ | 2733/18018 [5:55:12<35:38:02, 8.39s/step, loss=1.658, acc=0.647, acc_len=1.53, thru=14.5, I=13.3, T=27.6, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:25:26,297] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:25:33,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:25:35,616] loop.py:45 INFO TRAIN_STEP step=2734 loss=1.800862 Training: 15%|█▌ | 2734/18018 [5:55:21<37:01:42, 8.72s/step, loss=1.801, acc=0.622, acc_len=1.44, thru=13.6, I=16.8, T=25.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:25:35,860] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 19:25:41,402] loop.py:380 INFO TIMING step=2735: step=3.858s data=1.715s compute=3.831s [fwd=2.710s bwd=1.106s opt=0.003s] dispatch=1.917s [2026-10-18 19:25:41,402] loop.py:45 INFO TRAIN_STEP step=2735 loss=1.634334 Training: 15%|█▌ | 2735/18018 [5:55:27<33:17:13, 7.84s/step, loss=1.634, acc=0.645, acc_len=1.52, thru=18.2, I=22.6, T=33.2, wait=1.9s, pool=60, epoch=1/2][2026-10-18 19:25:50,761] loop.py:45 INFO TRAIN_STEP step=2736 loss=1.678449 Training: 15%|█▌ | 2736/18018 [5:55:36<35:13:02, 8.30s/step, loss=1.678, acc=0.641, acc_len=1.50, thru=14.1, I=15.3, T=31.4, wait=5.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:25:50,756] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2735: forward=2846.8ms backward=1184.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:25:59,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:25:59,822] loop.py:45 INFO TRAIN_STEP step=2737 loss=1.762712 Training: 15%|█▌ | 2737/18018 [5:55:45<36:11:19, 8.53s/step, loss=1.763, acc=0.625, acc_len=1.46, thru=14.6, I=13.1, T=32.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:25:59,915] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:05,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:26:06,675] loop.py:45 INFO TRAIN_STEP step=2738 loss=1.915987 Training: 15%|█▌ | 2738/18018 [5:55:52<34:03:19, 8.02s/step, loss=1.916, acc=0.602, acc_len=1.38, thru=18.3, I=20.5, T=33.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:07,149] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:12,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:26:13,079] loop.py:45 INFO TRAIN_STEP step=2739 loss=1.730560 Training: 15%|█▌ | 2739/18018 [5:55:59<31:59:36, 7.54s/step, loss=1.731, acc=0.634, acc_len=1.48, thru=19.6, I=23.7, T=34.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:13,323] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:19,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:26:20,387] loop.py:380 INFO TIMING step=2740: step=3.956s data=1.836s compute=3.913s [fwd=2.724s bwd=1.175s opt=0.003s] dispatch=3.341s [2026-10-18 19:26:20,387] loop.py:45 INFO TRAIN_STEP step=2740 loss=1.686285 Training: 15%|█▌ | 2740/18018 [5:56:06<31:41:42, 7.47s/step, loss=1.686, acc=0.643, acc_len=1.51, thru=18.8, I=21.2, T=32.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:20,805] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:26,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:26:28,719] loop.py:45 INFO TRAIN_STEP step=2741 loss=1.651806 Training: 15%|█▌ | 2741/18018 [5:56:14<32:47:29, 7.73s/step, loss=1.652, acc=0.648, acc_len=1.53, thru=14.9, I=21.7, T=28.6, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:26:28,714] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2740: forward=3117.7ms backward=1319.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:29,211] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 19:26:35,996] loop.py:45 INFO TRAIN_STEP step=2742 loss=1.785384 Training: 15%|█▌ | 2742/18018 [5:56:22<32:13:09, 7.59s/step, loss=1.785, acc=0.624, acc_len=1.45, thru=12.9, I=20.3, T=33.7, wait=3.5s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:43,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:26:44,779] loop.py:45 INFO TRAIN_STEP step=2743 loss=1.826501 Training: 15%|█▌ | 2743/18018 [5:56:30<33:43:50, 7.95s/step, loss=1.827, acc=0.622, acc_len=1.45, thru=17.4, I=16.7, T=32.6, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:44,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:51,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:26:52,052] loop.py:45 INFO TRAIN_STEP step=2744 loss=1.777727 Training: 15%|█▌ | 2744/18018 [5:56:38<32:52:04, 7.75s/step, loss=1.778, acc=0.628, acc_len=1.46, thru=17.2, I=21.4, T=32.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:52,216] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:57,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:26:59,370] loop.py:380 INFO TIMING step=2745: step=4.087s data=1.838s compute=4.052s [fwd=2.845s bwd=1.192s opt=0.003s] dispatch=3.221s [2026-10-18 19:26:59,371] loop.py:45 INFO TRAIN_STEP step=2745 loss=1.906837 Training: 15%|█▌ | 2745/18018 [5:56:45<32:19:12, 7.62s/step, loss=1.907, acc=0.606, acc_len=1.39, thru=17.2, I=21.8, T=31.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:26:59,571] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:06,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:27:08,001] loop.py:45 INFO TRAIN_STEP step=2746 loss=1.878416 Training: 15%|█▌ | 2746/18018 [5:56:54<33:36:22, 7.92s/step, loss=1.878, acc=0.610, acc_len=1.40, thru=15.2, I=15.4, T=31.2, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:27:07,997] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2745: forward=2917.6ms backward=1142.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:08,472] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:15,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:27:15,287] loop.py:45 INFO TRAIN_STEP step=2747 loss=1.783636 Training: 15%|█▌ | 2747/18018 [5:57:01<32:47:45, 7.73s/step, loss=1.784, acc=0.622, acc_len=1.45, thru=17.0, I=21.3, T=33.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:15,701] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:22,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:27:24,719] loop.py:45 INFO TRAIN_STEP step=2748 loss=1.879598 Training: 15%|█▌ | 2748/18018 [5:57:10<34:57:19, 8.24s/step, loss=1.880, acc=0.605, acc_len=1.40, thru=13.9, I=15.8, T=29.4, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:25,150] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:30,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:27:33,022] loop.py:45 INFO TRAIN_STEP step=2749 loss=1.842783 Training: 15%|█▌ | 2749/18018 [5:57:19<35:02:08, 8.26s/step, loss=1.843, acc=0.611, acc_len=1.41, thru=16.3, I=22.0, T=26.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:33,132] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:39,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:27:41,806] loop.py:380 INFO TIMING step=2750: step=4.384s data=1.963s compute=4.353s [fwd=3.010s bwd=1.327s opt=0.003s] dispatch=4.388s [2026-10-18 19:27:41,806] loop.py:45 INFO TRAIN_STEP step=2750 loss=1.849974 Training: 15%|█▌ | 2750/18018 [5:57:27<35:41:54, 8.42s/step, loss=1.850, acc=0.615, acc_len=1.42, thru=14.8, I=18.8, T=29.2, wait=4.4s, pool=64, epoch=1/2][2026-10-18 19:27:41,810] loop.py:419 INFO Saving checkpoint at step 2750... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:44,868] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 19:27:45,308] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0002751 (TrainerActor pid=9310) [2026-10-18 19:27:45,316] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0002001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:46,879] inference_manager.py:408 INFO Pool capacity available after 7.1s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:51,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 19:27:45,307] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0002751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 19:27:46,411] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0002001 [2026-10-18 19:27:53,227] loop.py:45 INFO TRAIN_STEP step=2751 loss=1.615948 Training: 15%|█▌ | 2751/18018 [5:57:39<39:31:05, 9.32s/step, loss=1.616, acc=0.651, acc_len=1.54, thru=20.2, I=26.8, T=31.2, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:27:53,221] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2750: forward=2857.0ms backward=1201.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:53,370] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:27:59,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:28:00,499] loop.py:45 INFO TRAIN_STEP step=2752 loss=1.774209 Training: 15%|█▌ | 2752/18018 [5:57:46<36:54:35, 8.70s/step, loss=1.774, acc=0.626, acc_len=1.45, thru=15.6, I=22.3, T=31.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:00,630] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 19:28:06,902] loop.py:45 INFO TRAIN_STEP step=2753 loss=1.772227 Training: 15%|█▌ | 2753/18018 [5:57:52<33:59:02, 8.01s/step, loss=1.772, acc=0.624, acc_len=1.45, thru=16.6, I=22.3, T=32.2, wait=2.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:14,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:28:15,910] loop.py:45 INFO TRAIN_STEP step=2754 loss=1.880436 Training: 15%|█▌ | 2754/18018 [5:58:01<35:14:42, 8.31s/step, loss=1.880, acc=0.610, acc_len=1.41, thru=15.7, I=17.2, T=29.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:16,209] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:28:22,478] loop.py:380 INFO TIMING step=2755: step=4.048s data=1.857s compute=4.005s [fwd=2.826s bwd=1.164s opt=0.003s] dispatch=2.509s [2026-10-18 19:28:22,478] loop.py:45 INFO TRAIN_STEP step=2755 loss=1.865629 Training: 15%|█▌ | 2755/18018 [5:58:08<33:01:26, 7.79s/step, loss=1.866, acc=0.610, acc_len=1.41, thru=15.7, I=21.7, T=31.6, wait=2.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:30,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:28:31,099] loop.py:45 INFO TRAIN_STEP step=2756 loss=1.874284 Training: 15%|█▌ | 2756/18018 [5:58:17<34:04:48, 8.04s/step, loss=1.874, acc=0.611, acc_len=1.41, thru=16.9, I=16.4, T=30.0, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:28:31,093] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2755: forward=2947.8ms backward=1276.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:31,207] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:36,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:28:38,205] loop.py:45 INFO TRAIN_STEP step=2757 loss=1.799265 Training: 15%|█▌ | 2757/18018 [5:58:24<32:53:15, 7.76s/step, loss=1.799, acc=0.623, acc_len=1.45, thru=16.9, I=21.2, T=31.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:38,388] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:44,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:28:45,560] loop.py:45 INFO TRAIN_STEP step=2758 loss=1.767966 Training: 15%|█▌ | 2758/18018 [5:58:31<32:22:22, 7.64s/step, loss=1.768, acc=0.623, acc_len=1.45, thru=14.8, I=21.6, T=31.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:45,988] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:50,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:28:51,649] loop.py:45 INFO TRAIN_STEP step=2759 loss=1.723349 Training: 15%|█▌ | 2759/18018 [5:58:37<30:24:13, 7.17s/step, loss=1.723, acc=0.636, acc_len=1.49, thru=18.4, I=25.1, T=33.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:52,054] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:28:59,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:28:59,964] loop.py:380 INFO TIMING step=2760: step=4.625s data=2.279s compute=4.581s [fwd=3.198s bwd=1.367s opt=0.003s] dispatch=3.680s [2026-10-18 19:28:59,964] loop.py:45 INFO TRAIN_STEP step=2760 loss=1.857448 Training: 15%|█▌ | 2760/18018 [5:58:46<31:51:10, 7.52s/step, loss=1.857, acc=0.614, acc_len=1.42, thru=13.3, I=18.1, T=27.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:00,025] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:05,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:29:07,182] loop.py:45 INFO TRAIN_STEP step=2761 loss=1.734337 Training: 15%|█▌ | 2761/18018 [5:58:53<31:28:19, 7.43s/step, loss=1.734, acc=0.632, acc_len=1.47, thru=16.9, I=20.9, T=29.5, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:29:07,176] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2760: forward=3000.2ms backward=1289.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:07,440] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:11,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:29:13,233] loop.py:45 INFO TRAIN_STEP step=2762 loss=1.634735 Training: 15%|█▌ | 2762/18018 [5:58:59<29:43:20, 7.01s/step, loss=1.635, acc=0.646, acc_len=1.52, thru=18.0, I=28.3, T=34.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:13,572] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:18,494] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:29:20,128] loop.py:45 INFO TRAIN_STEP step=2763 loss=1.956098 Training: 15%|█▌ | 2763/18018 [5:59:06<29:34:07, 6.98s/step, loss=1.956, acc=0.594, acc_len=1.36, thru=16.9, I=25.4, T=34.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:20,505] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:26,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:29:27,253] loop.py:45 INFO TRAIN_STEP step=2764 loss=1.644892 Training: 15%|█▌ | 2764/18018 [5:59:13<29:45:24, 7.02s/step, loss=1.645, acc=0.645, acc_len=1.53, thru=17.3, I=20.7, T=28.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:27,689] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:32,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:29:33,919] loop.py:380 INFO TIMING step=2765: step=3.946s data=1.793s compute=3.894s [fwd=2.728s bwd=1.151s opt=0.003s] dispatch=2.710s [2026-10-18 19:29:33,920] loop.py:45 INFO TRAIN_STEP step=2765 loss=1.877635 Training: 15%|█▌ | 2765/18018 [5:59:20<29:18:08, 6.92s/step, loss=1.878, acc=0.610, acc_len=1.41, thru=16.5, I=20.3, T=32.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:33,957] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:38,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:29:40,436] loop.py:45 INFO TRAIN_STEP step=2766 loss=1.641455 Training: 15%|█▌ | 2766/18018 [5:59:26<28:47:25, 6.80s/step, loss=1.641, acc=0.650, acc_len=1.53, thru=18.7, I=23.4, T=31.5, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:29:40,431] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2765: forward=2829.7ms backward=1193.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:29:40,883] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:29:49,586] loop.py:45 INFO TRAIN_STEP step=2767 loss=1.753645 Training: 15%|█▌ | 2767/18018 [5:59:35<31:46:46, 7.50s/step, loss=1.754, acc=0.631, acc_len=1.47, thru=13.8, I=13.6, T=28.1, wait=4.6s, pool=56, epoch=1/2][2026-10-18 19:29:59,389] loop.py:45 INFO TRAIN_STEP step=2768 loss=1.921340 Training: 15%|█▌ | 2768/18018 [5:59:45<34:42:20, 8.19s/step, loss=1.921, acc=0.608, acc_len=1.39, thru=13.4, I=14.6, T=28.9, wait=5.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:06,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:30:09,186] loop.py:45 INFO TRAIN_STEP step=2769 loss=1.827744 Training: 15%|█▌ | 2769/18018 [5:59:55<36:44:23, 8.67s/step, loss=1.828, acc=0.620, acc_len=1.45, thru=14.4, I=15.4, T=30.4, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:09,502] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 19:30:15,240] loop.py:380 INFO TIMING step=2770: step=3.589s data=1.582s compute=3.562s [fwd=2.569s bwd=0.979s opt=0.003s] dispatch=2.456s [2026-10-18 19:30:15,240] loop.py:45 INFO TRAIN_STEP step=2770 loss=1.782738 Training: 15%|█▌ | 2770/18018 [6:00:01<33:24:24, 7.89s/step, loss=1.783, acc=0.625, acc_len=1.46, thru=17.8, I=23.9, T=35.7, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:22,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:30:22,964] loop.py:45 INFO TRAIN_STEP step=2771 loss=1.782560 Training: 15%|█▌ | 2771/18018 [6:00:09<33:11:57, 7.84s/step, loss=1.783, acc=0.627, acc_len=1.46, thru=16.9, I=18.1, T=29.8, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:30:22,960] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2770: forward=2978.7ms backward=1272.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:23,096] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:28,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:30:29,138] loop.py:45 INFO TRAIN_STEP step=2772 loss=1.922449 Training: 15%|█▌ | 2772/18018 [6:00:15<31:04:55, 7.34s/step, loss=1.922, acc=0.606, acc_len=1.40, thru=18.6, I=21.4, T=33.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:29,304] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:33,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:30:35,183] loop.py:45 INFO TRAIN_STEP step=2773 loss=1.880372 Training: 15%|█▌ | 2773/18018 [6:00:21<29:26:08, 6.95s/step, loss=1.880, acc=0.608, acc_len=1.41, thru=19.7, I=24.4, T=33.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:35,326] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:41,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:30:43,676] loop.py:45 INFO TRAIN_STEP step=2774 loss=1.882192 Training: 15%|█▌ | 2774/18018 [6:00:29<31:23:40, 7.41s/step, loss=1.882, acc=0.607, acc_len=1.40, thru=15.4, I=18.5, T=27.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:43,785] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:49,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:30:51,090] loop.py:380 INFO TIMING step=2775: step=4.909s data=2.510s compute=4.871s [fwd=3.277s bwd=1.580s opt=0.003s] dispatch=2.495s [2026-10-18 19:30:51,091] loop.py:45 INFO TRAIN_STEP step=2775 loss=1.861963 Training: 15%|█▌ | 2775/18018 [6:00:37<31:23:30, 7.41s/step, loss=1.862, acc=0.609, acc_len=1.40, thru=15.0, I=22.7, T=26.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:51,348] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:56,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:30:58,683] loop.py:45 INFO TRAIN_STEP step=2776 loss=1.817837 Training: 15%|█▌ | 2776/18018 [6:00:44<31:36:56, 7.47s/step, loss=1.818, acc=0.621, acc_len=1.44, thru=14.3, I=19.8, T=30.4, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:30:58,678] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2775: forward=2743.9ms backward=1416.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:30:59,020] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:05,773] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:31:06,918] loop.py:45 INFO TRAIN_STEP step=2777 loss=1.619618 Training: 15%|█▌ | 2777/18018 [6:00:53<32:35:26, 7.70s/step, loss=1.620, acc=0.653, acc_len=1.55, thru=16.2, I=18.6, T=31.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:07,283] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:13,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:31:15,331] loop.py:45 INFO TRAIN_STEP step=2778 loss=1.812320 Training: 15%|█▌ | 2778/18018 [6:01:01<33:29:45, 7.91s/step, loss=1.812, acc=0.622, acc_len=1.44, thru=13.9, I=20.7, T=29.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:15,614] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:21,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:31:22,862] loop.py:45 INFO TRAIN_STEP step=2779 loss=1.773996 Training: 15%|█▌ | 2779/18018 [6:01:08<33:00:36, 7.80s/step, loss=1.774, acc=0.629, acc_len=1.47, thru=14.0, I=19.8, T=31.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:23,252] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:30,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:31:31,866] loop.py:380 INFO TIMING step=2780: step=4.613s data=2.194s compute=4.571s [fwd=3.210s bwd=1.346s opt=0.003s] dispatch=4.380s [2026-10-18 19:31:31,866] loop.py:45 INFO TRAIN_STEP step=2780 loss=1.706593 Training: 15%|█▌ | 2780/18018 [6:01:17<34:32:21, 8.16s/step, loss=1.707, acc=0.634, acc_len=1.48, thru=14.9, I=17.9, T=27.7, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:31,924] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:37,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:31:40,094] loop.py:45 INFO TRAIN_STEP step=2781 loss=1.885934 Training: 15%|█▌ | 2781/18018 [6:01:26<34:37:13, 8.18s/step, loss=1.886, acc=0.607, acc_len=1.40, thru=13.6, I=22.7, T=28.9, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:31:40,088] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2780: forward=3062.9ms backward=1302.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:40,123] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:45,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:31:46,620] loop.py:45 INFO TRAIN_STEP step=2782 loss=1.692528 Training: 15%|█▌ | 2782/18018 [6:01:32<32:31:06, 7.68s/step, loss=1.693, acc=0.640, acc_len=1.50, thru=16.3, I=25.4, T=31.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:46,905] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:52,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:31:54,511] loop.py:45 INFO TRAIN_STEP step=2783 loss=1.750483 Training: 15%|█▌ | 2783/18018 [6:01:40<32:46:57, 7.75s/step, loss=1.750, acc=0.627, acc_len=1.47, thru=14.9, I=20.0, T=25.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:31:54,729] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:02,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:32:03,670] loop.py:45 INFO TRAIN_STEP step=2784 loss=1.714313 Training: 15%|█▌ | 2784/18018 [6:01:49<34:34:16, 8.17s/step, loss=1.714, acc=0.637, acc_len=1.49, thru=14.3, I=16.8, T=29.0, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:04,127] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:32:12,276] loop.py:380 INFO TIMING step=2785: step=3.826s data=1.685s compute=3.794s [fwd=2.694s bwd=1.085s opt=0.003s] dispatch=4.772s [2026-10-18 19:32:12,276] loop.py:45 INFO TRAIN_STEP step=2785 loss=1.837857 Training: 15%|█▌ | 2785/18018 [6:01:58<35:07:28, 8.30s/step, loss=1.838, acc=0.622, acc_len=1.45, thru=13.3, I=13.9, T=33.5, wait=4.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:20,483] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:32:21,358] loop.py:45 INFO TRAIN_STEP step=2786 loss=1.918968 Training: 15%|█▌ | 2786/18018 [6:02:07<36:06:52, 8.54s/step, loss=1.919, acc=0.602, acc_len=1.38, thru=16.1, I=16.2, T=32.9, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:32:21,353] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2785: forward=2758.9ms backward=1084.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:21,502] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:26,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:32:27,681] loop.py:45 INFO TRAIN_STEP step=2787 loss=1.845379 Training: 15%|█▌ | 2787/18018 [6:02:13<33:18:07, 7.87s/step, loss=1.845, acc=0.619, acc_len=1.44, thru=18.6, I=22.9, T=33.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:27,783] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:33,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:32:34,443] loop.py:45 INFO TRAIN_STEP step=2788 loss=1.923688 Training: 15%|█▌ | 2788/18018 [6:02:20<31:53:30, 7.54s/step, loss=1.924, acc=0.604, acc_len=1.39, thru=17.4, I=20.4, T=29.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:34,654] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:43,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:32:45,194] loop.py:45 INFO TRAIN_STEP step=2789 loss=1.889638 Training: 15%|█▌ | 2789/18018 [6:02:31<35:57:58, 8.50s/step, loss=1.890, acc=0.609, acc_len=1.41, thru=11.2, I=14.8, T=27.4, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:45,667] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:49,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:32:52,712] loop.py:380 INFO TIMING step=2790: step=4.888s data=2.405s compute=4.858s [fwd=3.445s bwd=1.397s opt=0.003s] dispatch=2.620s [2026-10-18 19:32:52,712] loop.py:45 INFO TRAIN_STEP step=2790 loss=1.611060 Training: 15%|█▌ | 2790/18018 [6:02:38<34:42:54, 8.21s/step, loss=1.611, acc=0.654, acc_len=1.55, thru=14.6, I=27.5, T=26.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:52,877] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:57,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:32:58,319] loop.py:45 INFO TRAIN_STEP step=2791 loss=1.770227 Training: 15%|█▌ | 2791/18018 [6:02:44<31:24:47, 7.43s/step, loss=1.770, acc=0.626, acc_len=1.46, thru=18.1, I=24.7, T=36.7, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:32:58,314] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2790: forward=2447.3ms backward=997.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:32:58,487] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:05,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:33:06,175] loop.py:45 INFO TRAIN_STEP step=2792 loss=1.768959 Training: 15%|█▌ | 2792/18018 [6:02:52<31:57:30, 7.56s/step, loss=1.769, acc=0.631, acc_len=1.47, thru=17.5, I=18.5, T=31.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:06,301] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:11,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:33:13,136] loop.py:45 INFO TRAIN_STEP step=2793 loss=1.798411 Training: 16%|█▌ | 2793/18018 [6:02:59<31:12:03, 7.38s/step, loss=1.798, acc=0.628, acc_len=1.46, thru=17.5, I=19.2, T=32.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:13,197] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:19,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:33:20,947] loop.py:45 INFO TRAIN_STEP step=2794 loss=1.737627 Training: 16%|█▌ | 2794/18018 [6:03:07<31:44:56, 7.51s/step, loss=1.738, acc=0.634, acc_len=1.48, thru=14.9, I=20.1, T=30.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:21,191] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:25,494] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:33:28,121] loop.py:380 INFO TIMING step=2795: step=4.475s data=2.134s compute=4.438s [fwd=2.950s bwd=1.472s opt=0.003s] dispatch=2.688s [2026-10-18 19:33:28,122] loop.py:45 INFO TRAIN_STEP step=2795 loss=1.773451 Training: 16%|█▌ | 2795/18018 [6:03:14<31:19:27, 7.41s/step, loss=1.773, acc=0.623, acc_len=1.45, thru=16.4, I=29.8, T=28.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:28,513] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:35,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:33:35,523] loop.py:45 INFO TRAIN_STEP step=2796 loss=1.718501 Training: 16%|█▌ | 2796/18018 [6:03:21<31:18:44, 7.41s/step, loss=1.719, acc=0.640, acc_len=1.50, thru=18.3, I=20.0, T=30.3, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:33:35,518] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2795: forward=2893.3ms backward=1289.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:35,978] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:42,744] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:33:44,107] loop.py:45 INFO TRAIN_STEP step=2797 loss=1.738482 Training: 16%|█▌ | 2797/18018 [6:03:30<32:48:26, 7.76s/step, loss=1.738, acc=0.634, acc_len=1.49, thru=15.2, I=16.5, T=31.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:44,266] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:50,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:33:51,401] loop.py:45 INFO TRAIN_STEP step=2798 loss=1.896970 Training: 16%|█▌ | 2798/18018 [6:03:37<32:12:54, 7.62s/step, loss=1.897, acc=0.600, acc_len=1.39, thru=17.3, I=20.6, T=28.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:51,439] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:56,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:33:58,088] loop.py:45 INFO TRAIN_STEP step=2799 loss=1.810862 Training: 16%|█▌ | 2799/18018 [6:03:44<31:01:46, 7.34s/step, loss=1.811, acc=0.622, acc_len=1.44, thru=16.9, I=19.5, T=29.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:33:58,509] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:05,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:34:06,219] loop.py:380 INFO TIMING step=2800: step=4.481s data=2.158s compute=4.443s [fwd=3.116s bwd=1.312s opt=0.003s] dispatch=3.639s [2026-10-18 19:34:06,219] loop.py:45 INFO TRAIN_STEP step=2800 loss=1.868244 Training: 16%|█▌ | 2800/18018 [6:03:52<32:01:39, 7.58s/step, loss=1.868, acc=0.614, acc_len=1.42, thru=13.6, I=16.3, T=28.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:06,677] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:12,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:34:14,387] loop.py:45 INFO TRAIN_STEP step=2801 loss=1.757756 Training: 16%|█▌ | 2801/18018 [6:04:00<32:46:43, 7.75s/step, loss=1.758, acc=0.626, acc_len=1.46, thru=16.3, I=18.2, T=30.1, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:34:14,383] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2800: forward=2991.7ms backward=1200.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:14,750] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:20,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:34:21,180] loop.py:45 INFO TRAIN_STEP step=2802 loss=1.810555 Training: 16%|█▌ | 2802/18018 [6:04:07<31:33:18, 7.47s/step, loss=1.811, acc=0.621, acc_len=1.44, thru=15.3, I=21.0, T=30.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:21,449] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:26,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:34:27,635] loop.py:45 INFO TRAIN_STEP step=2803 loss=1.748284 Training: 16%|█▌ | 2803/18018 [6:04:13<30:16:15, 7.16s/step, loss=1.748, acc=0.630, acc_len=1.48, thru=19.0, I=22.0, T=33.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:27,684] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:34,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:34:35,263] loop.py:45 INFO TRAIN_STEP step=2804 loss=1.910155 Training: 16%|█▌ | 2804/18018 [6:04:21<30:51:35, 7.30s/step, loss=1.910, acc=0.602, acc_len=1.39, thru=17.2, I=19.9, T=30.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:35,528] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:40,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:34:42,510] loop.py:380 INFO TIMING step=2805: step=4.514s data=1.711s compute=4.483s [fwd=3.353s bwd=1.116s opt=0.003s] dispatch=2.723s [2026-10-18 19:34:42,511] loop.py:45 INFO TRAIN_STEP step=2805 loss=1.755186 Training: 16%|█▌ | 2805/18018 [6:04:28<30:47:17, 7.29s/step, loss=1.755, acc=0.631, acc_len=1.48, thru=16.9, I=21.8, T=28.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:42,541] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:49,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 19:34:51,702] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2805: forward=2934.5ms backward=1238.0ms [2026-10-18 19:34:51,709] loop.py:45 INFO TRAIN_STEP step=2806 loss=1.763577 Training: 16%|█▌ | 2806/18018 [6:04:37<33:12:39, 7.86s/step, loss=1.764, acc=0.630, acc_len=1.47, thru=14.0, I=16.1, T=30.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:51,781] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:59,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:34:59,449] loop.py:45 INFO TRAIN_STEP step=2807 loss=1.776615 Training: 16%|█▌ | 2807/18018 [6:04:45<33:03:28, 7.82s/step, loss=1.777, acc=0.624, acc_len=1.45, thru=14.0, I=17.4, T=33.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:34:59,793] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:06,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:35:06,348] loop.py:45 INFO TRAIN_STEP step=2808 loss=1.848578 Training: 16%|█▌ | 2808/18018 [6:04:52<31:52:51, 7.55s/step, loss=1.849, acc=0.610, acc_len=1.41, thru=18.8, I=17.6, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:06,811] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 19:35:14,021] loop.py:45 INFO TRAIN_STEP step=2809 loss=1.743229 Training: 16%|█▌ | 2809/18018 [6:05:00<32:02:26, 7.58s/step, loss=1.743, acc=0.631, acc_len=1.48, thru=15.8, I=15.7, T=33.1, wait=3.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:22,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:35:22,847] loop.py:380 INFO TIMING step=2810: step=4.813s data=2.421s compute=4.768s [fwd=3.341s bwd=1.413s opt=0.003s] dispatch=4.005s [2026-10-18 19:35:22,848] loop.py:45 INFO TRAIN_STEP step=2810 loss=1.819572 Training: 16%|█▌ | 2810/18018 [6:05:08<33:36:57, 7.96s/step, loss=1.820, acc=0.623, acc_len=1.45, thru=15.9, I=16.1, T=26.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:23,229] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:28,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:35:30,309] loop.py:45 INFO TRAIN_STEP step=2811 loss=1.740587 Training: 16%|█▌ | 2811/18018 [6:05:16<32:59:03, 7.81s/step, loss=1.741, acc=0.632, acc_len=1.48, thru=16.4, I=19.1, T=29.3, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:35:30,304] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2810: forward=3020.0ms backward=1297.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:30,501] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:37,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:35:38,698] loop.py:45 INFO TRAIN_STEP step=2812 loss=1.739900 Training: 16%|█▌ | 2812/18018 [6:05:24<33:43:02, 7.98s/step, loss=1.740, acc=0.630, acc_len=1.47, thru=15.6, I=18.0, T=29.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:39,124] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:45,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:35:45,829] loop.py:45 INFO TRAIN_STEP step=2813 loss=1.701140 Training: 16%|█▌ | 2813/18018 [6:05:31<32:38:16, 7.73s/step, loss=1.701, acc=0.641, acc_len=1.51, thru=17.3, I=20.4, T=30.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:46,060] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:50,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:35:52,525] loop.py:45 INFO TRAIN_STEP step=2814 loss=1.669839 Training: 16%|█▌ | 2814/18018 [6:05:38<31:19:43, 7.42s/step, loss=1.670, acc=0.643, acc_len=1.51, thru=18.2, I=25.0, T=28.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:52,857] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:35:58,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:35:59,544] loop.py:380 INFO TIMING step=2815: step=4.131s data=1.970s compute=4.099s [fwd=2.902s bwd=1.182s opt=0.003s] dispatch=2.878s [2026-10-18 19:35:59,544] loop.py:45 INFO TRAIN_STEP step=2815 loss=1.668441 Training: 16%|█▌ | 2815/18018 [6:05:45<30:49:10, 7.30s/step, loss=1.668, acc=0.644, acc_len=1.52, thru=14.9, I=25.2, T=31.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:00,024] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:04,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:36:06,639] loop.py:45 INFO TRAIN_STEP step=2816 loss=1.772948 Training: 16%|█▌ | 2816/18018 [6:05:52<30:33:38, 7.24s/step, loss=1.773, acc=0.631, acc_len=1.46, thru=16.0, I=21.8, T=29.4, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:36:06,635] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2815: forward=3032.9ms backward=1272.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:06,992] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:11,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:36:12,643] loop.py:45 INFO TRAIN_STEP step=2817 loss=1.712762 Training: 16%|█▌ | 2817/18018 [6:05:58<28:59:58, 6.87s/step, loss=1.713, acc=0.639, acc_len=1.50, thru=19.2, I=27.5, T=34.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:12,769] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:21,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:36:21,706] loop.py:45 INFO TRAIN_STEP step=2818 loss=1.868169 Training: 16%|█▌ | 2818/18018 [6:06:07<31:46:36, 7.53s/step, loss=1.868, acc=0.615, acc_len=1.42, thru=14.3, I=15.0, T=31.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:21,756] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 19:36:29,348] loop.py:45 INFO TRAIN_STEP step=2819 loss=2.017900 Training: 16%|█▌ | 2819/18018 [6:06:15<31:55:42, 7.56s/step, loss=2.018, acc=0.591, acc_len=1.35, thru=14.5, I=16.7, T=29.9, wait=3.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:34,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:36:36,493] loop.py:380 INFO TIMING step=2820: step=3.766s data=1.684s compute=3.716s [fwd=2.593s bwd=1.108s opt=0.003s] dispatch=3.363s [2026-10-18 19:36:36,493] loop.py:45 INFO TRAIN_STEP step=2820 loss=1.773503 Training: 16%|█▌ | 2820/18018 [6:06:22<31:23:19, 7.44s/step, loss=1.774, acc=0.627, acc_len=1.45, thru=18.1, I=19.2, T=34.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:36,613] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:43,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:36:44,487] loop.py:45 INFO TRAIN_STEP step=2821 loss=1.762295 Training: 16%|█▌ | 2821/18018 [6:06:30<32:05:38, 7.60s/step, loss=1.762, acc=0.629, acc_len=1.47, thru=13.6, I=19.5, T=30.4, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:36:44,481] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2820: forward=2914.4ms backward=1242.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:44,753] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:51,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:36:53,859] loop.py:45 INFO TRAIN_STEP step=2822 loss=1.792068 Training: 16%|█▌ | 2822/18018 [6:06:39<34:19:54, 8.13s/step, loss=1.792, acc=0.624, acc_len=1.45, thru=14.1, I=19.2, T=27.2, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:36:54,280] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:02,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:37:04,855] loop.py:45 INFO TRAIN_STEP step=2823 loss=1.773793 Training: 16%|█▌ | 2823/18018 [6:06:50<37:57:18, 8.99s/step, loss=1.774, acc=0.629, acc_len=1.47, thru=11.2, I=14.5, T=29.7, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:05,089] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:10,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:37:11,109] loop.py:45 INFO TRAIN_STEP step=2824 loss=1.724111 Training: 16%|█▌ | 2824/18018 [6:06:57<34:29:14, 8.17s/step, loss=1.724, acc=0.637, acc_len=1.49, thru=17.6, I=25.2, T=32.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:11,143] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:16,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:37:17,339] loop.py:380 INFO TIMING step=2825: step=4.252s data=2.008s compute=4.223s [fwd=3.000s bwd=1.209s opt=0.003s] dispatch=1.967s [2026-10-18 19:37:17,339] loop.py:45 INFO TRAIN_STEP step=2825 loss=1.718161 Training: 16%|█▌ | 2825/18018 [6:07:03<32:01:36, 7.59s/step, loss=1.718, acc=0.635, acc_len=1.49, thru=19.0, I=21.8, T=30.1, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:17,542] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:25,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:37:25,281] loop.py:45 INFO TRAIN_STEP step=2826 loss=1.693411 Training: 16%|█▌ | 2826/18018 [6:07:11<32:28:15, 7.69s/step, loss=1.693, acc=0.642, acc_len=1.50, thru=15.7, I=18.4, T=28.7, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:37:25,276] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2825: forward=3050.0ms backward=1361.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:25,537] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:31,897] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:37:34,319] loop.py:45 INFO TRAIN_STEP step=2827 loss=1.760546 Training: 16%|█▌ | 2827/18018 [6:07:20<34:10:14, 8.10s/step, loss=1.761, acc=0.626, acc_len=1.46, thru=14.2, I=16.9, T=27.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:34,417] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 [2026-10-18 19:37:40,600] loop.py:45 INFO TRAIN_STEP step=2828 loss=1.623194 Training: 16%|█▌ | 2828/18018 [6:07:26<31:52:04, 7.55s/step, loss=1.623, acc=0.646, acc_len=1.52, thru=14.6, I=28.4, T=29.8, wait=2.0s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:50,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:37:51,540] loop.py:45 INFO TRAIN_STEP step=2829 loss=1.944327 Training: 16%|█▌ | 2829/18018 [6:07:37<36:09:10, 8.57s/step, loss=1.944, acc=0.602, acc_len=1.38, thru=15.1, I=13.9, T=29.7, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:51,854] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:56,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:37:58,356] loop.py:380 INFO TIMING step=2830: step=4.329s data=2.021s compute=4.301s [fwd=3.032s bwd=1.253s opt=0.003s] dispatch=2.477s [2026-10-18 19:37:58,356] loop.py:45 INFO TRAIN_STEP step=2830 loss=1.895242 Training: 16%|█▌ | 2830/18018 [6:07:44<33:55:54, 8.04s/step, loss=1.895, acc=0.605, acc_len=1.39, thru=18.1, I=23.7, T=29.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:37:58,687] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:05,968] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:38:07,392] loop.py:45 INFO TRAIN_STEP step=2831 loss=1.746389 Training: 16%|█▌ | 2831/18018 [6:07:53<35:11:00, 8.34s/step, loss=1.746, acc=0.631, acc_len=1.47, thru=14.7, I=17.3, T=24.6, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:38:07,387] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2830: forward=3780.7ms backward=1381.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:07,476] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 19:38:16,491] loop.py:45 INFO TRAIN_STEP step=2832 loss=1.729087 Training: 16%|█▌ | 2832/18018 [6:08:02<36:08:37, 8.57s/step, loss=1.729, acc=0.632, acc_len=1.48, thru=12.9, I=17.6, T=21.6, wait=3.2s, pool=52, epoch=1/2][2026-10-18 19:38:24,887] loop.py:45 INFO TRAIN_STEP step=2833 loss=1.895914 Training: 16%|█▌ | 2833/18018 [6:08:10<35:55:23, 8.52s/step, loss=1.896, acc=0.615, acc_len=1.42, thru=15.1, I=12.6, T=30.5, wait=4.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:33,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:38:33,935] loop.py:45 INFO TRAIN_STEP step=2834 loss=1.721585 Training: 16%|█▌ | 2834/18018 [6:08:20<36:35:38, 8.68s/step, loss=1.722, acc=0.634, acc_len=1.48, thru=16.1, I=15.4, T=31.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:34,109] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:40,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:38:41,562] loop.py:380 INFO TIMING step=2835: step=4.681s data=2.150s compute=4.639s [fwd=3.180s bwd=1.444s opt=0.003s] dispatch=2.937s [2026-10-18 19:38:41,563] loop.py:45 INFO TRAIN_STEP step=2835 loss=1.878588 Training: 16%|█▌ | 2835/18018 [6:08:27<35:16:01, 8.36s/step, loss=1.879, acc=0.606, acc_len=1.40, thru=17.2, I=19.0, T=27.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:41,844] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:46,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:38:48,889] loop.py:45 INFO TRAIN_STEP step=2836 loss=1.774861 Training: 16%|█▌ | 2836/18018 [6:08:34<33:57:06, 8.05s/step, loss=1.775, acc=0.627, acc_len=1.46, thru=14.5, I=19.6, T=27.2, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:38:48,883] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2835: forward=3240.2ms backward=1416.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:49,098] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:55,416] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:38:55,526] loop.py:45 INFO TRAIN_STEP step=2837 loss=1.752584 Training: 16%|█▌ | 2837/18018 [6:08:41<32:09:49, 7.63s/step, loss=1.753, acc=0.630, acc_len=1.46, thru=15.9, I=19.7, T=33.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:38:55,919] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:04,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:39:04,473] loop.py:45 INFO TRAIN_STEP step=2838 loss=1.960320 Training: 16%|█▌ | 2838/18018 [6:08:50<33:49:52, 8.02s/step, loss=1.960, acc=0.603, acc_len=1.38, thru=14.7, I=14.5, T=28.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:04,746] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:09,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:39:10,872] loop.py:45 INFO TRAIN_STEP step=2839 loss=1.756645 Training: 16%|█▌ | 2839/18018 [6:08:56<31:46:28, 7.54s/step, loss=1.757, acc=0.627, acc_len=1.46, thru=18.6, I=20.8, T=30.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:11,999] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:16,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:39:16,754] loop.py:380 INFO TIMING step=2840: step=4.121s data=1.906s compute=4.088s [fwd=2.889s bwd=1.183s opt=0.003s] dispatch=1.750s [2026-10-18 19:39:16,754] loop.py:45 INFO TRAIN_STEP step=2840 loss=1.664231 Training: 16%|█▌ | 2840/18018 [6:09:02<29:40:47, 7.04s/step, loss=1.664, acc=0.642, acc_len=1.51, thru=19.6, I=23.8, T=31.1, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:17,198] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:22,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:39:24,333] loop.py:45 INFO TRAIN_STEP step=2841 loss=1.833467 Training: 16%|█▌ | 2841/18018 [6:09:10<30:21:41, 7.20s/step, loss=1.833, acc=0.616, acc_len=1.43, thru=16.8, I=21.2, T=29.1, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:39:24,328] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2840: forward=2961.0ms backward=1385.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:24,395] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:32,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:39:32,779] loop.py:45 INFO TRAIN_STEP step=2842 loss=1.785640 Training: 16%|█▌ | 2842/18018 [6:09:18<31:55:53, 7.57s/step, loss=1.786, acc=0.620, acc_len=1.44, thru=15.3, I=15.1, T=33.5, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:33,076] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:39,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:39:40,508] loop.py:45 INFO TRAIN_STEP step=2843 loss=1.814386 Training: 16%|█▌ | 2843/18018 [6:09:26<32:07:30, 7.62s/step, loss=1.814, acc=0.626, acc_len=1.45, thru=18.1, I=19.1, T=31.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:40,544] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:45,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:39:47,352] loop.py:45 INFO TRAIN_STEP step=2844 loss=1.909794 Training: 16%|█▌ | 2844/18018 [6:09:33<31:08:16, 7.39s/step, loss=1.910, acc=0.606, acc_len=1.39, thru=17.8, I=23.7, T=29.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:47,674] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:53,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:39:54,946] loop.py:380 INFO TIMING step=2845: step=4.046s data=1.834s compute=4.020s [fwd=2.785s bwd=1.220s opt=0.003s] dispatch=3.538s [2026-10-18 19:39:54,946] loop.py:45 INFO TRAIN_STEP step=2845 loss=1.699629 Training: 16%|█▌ | 2845/18018 [6:09:41<31:23:57, 7.45s/step, loss=1.700, acc=0.639, acc_len=1.51, thru=15.1, I=21.0, T=31.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:39:55,081] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 19:40:03,108] loop.py:45 INFO TRAIN_STEP step=2846 loss=1.803050 Training: 16%|█▌ | 2846/18018 [6:09:49<32:17:55, 7.66s/step, loss=1.803, acc=0.623, acc_len=1.45, thru=14.2, I=16.7, T=29.4, wait=3.8s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:40:03,104] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2845: forward=3047.3ms backward=1255.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:40:08,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:40:10,490] loop.py:45 INFO TRAIN_STEP step=2847 loss=1.854542 Training: 16%|█▌ | 2847/18018 [6:09:56<31:56:21, 7.58s/step, loss=1.855, acc=0.615, acc_len=1.43, thru=18.2, I=17.9, T=30.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:40:10,727] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:40:16,722] loop.py:45 INFO TRAIN_STEP step=2848 loss=1.595856 Training: 16%|█▌ | 2848/18018 [6:10:02<30:13:57, 7.17s/step, loss=1.596, acc=0.651, acc_len=1.54, thru=16.0, I=26.2, T=35.3, wait=2.6s, pool=48, epoch=1/2][2026-10-18 19:40:27,283] loop.py:45 INFO TRAIN_STEP step=2849 loss=2.001122 Training: 16%|█▌ | 2849/18018 [6:10:13<34:30:47, 8.19s/step, loss=2.001, acc=0.592, acc_len=1.35, thru=13.3, I=11.8, T=29.7, wait=6.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:40:33,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:40:34,945] loop.py:380 INFO TIMING step=2850: step=5.051s data=2.452s compute=5.014s [fwd=3.516s bwd=1.483s opt=0.003s] dispatch=2.600s [2026-10-18 19:40:34,946] loop.py:45 INFO TRAIN_STEP step=2850 loss=1.776440 Training: 16%|█▌ | 2850/18018 [6:10:21<33:50:26, 8.03s/step, loss=1.776, acc=0.623, acc_len=1.45, thru=17.3, I=20.0, T=25.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:40:35,155] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:40:41,316] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:40:42,395] loop.py:45 INFO TRAIN_STEP step=2851 loss=1.657877 Training: 16%|█▌ | 2851/18018 [6:10:28<33:06:18, 7.86s/step, loss=1.658, acc=0.642, acc_len=1.51, thru=15.4, I=20.1, T=31.4, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:40:42,391] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2850: forward=2819.7ms backward=1172.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:40:42,826] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:40:50,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:40:51,995] loop.py:45 INFO TRAIN_STEP step=2852 loss=1.984604 Training: 16%|█▌ | 2852/18018 [6:10:38<35:18:17, 8.38s/step, loss=1.985, acc=0.594, acc_len=1.37, thru=14.0, I=14.8, T=29.2, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:40:52,387] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:40:58,516] loop.py:45 INFO TRAIN_STEP step=2853 loss=1.651542 Training: 16%|█▌ | 2853/18018 [6:10:44<32:57:06, 7.82s/step, loss=1.652, acc=0.648, acc_len=1.53, thru=16.8, I=20.5, T=31.4, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:41:03,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:41:05,959] loop.py:45 INFO TRAIN_STEP step=2854 loss=1.756235 Training: 16%|█▌ | 2854/18018 [6:10:52<32:28:04, 7.71s/step, loss=1.756, acc=0.631, acc_len=1.47, thru=17.8, I=21.5, T=30.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:41:06,146] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 19:41:12,629] loop.py:380 INFO TIMING step=2855: step=3.586s data=1.508s compute=3.557s [fwd=2.530s bwd=1.014s opt=0.003s] dispatch=3.076s [2026-10-18 19:41:12,629] loop.py:45 INFO TRAIN_STEP step=2855 loss=1.677019 Training: 16%|█▌ | 2855/18018 [6:10:58<31:09:16, 7.40s/step, loss=1.677, acc=0.641, acc_len=1.51, thru=14.3, I=21.0, T=35.7, wait=3.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:41:20,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:41:21,565] loop.py:45 INFO TRAIN_STEP step=2856 loss=1.779618 Training: 16%|█▌ | 2856/18018 [6:11:07<33:05:50, 7.86s/step, loss=1.780, acc=0.628, acc_len=1.46, thru=16.1, I=17.1, T=27.9, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:41:21,560] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2855: forward=3148.7ms backward=1388.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:41:21,619] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:41:27,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:41:29,437] loop.py:45 INFO TRAIN_STEP step=2857 loss=1.760160 Training: 16%|█▌ | 2857/18018 [6:11:15<33:06:46, 7.86s/step, loss=1.760, acc=0.626, acc_len=1.46, thru=14.1, I=17.5, T=29.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:41:29,600] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:41:36,938] loop.py:45 INFO TRAIN_STEP step=2858 loss=1.690413 Training: 16%|█▌ | 2858/18018 [6:11:23<32:39:13, 7.75s/step, loss=1.690, acc=0.640, acc_len=1.51, thru=14.9, I=21.3, T=33.2, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:41:44,517] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:41:45,734] loop.py:45 INFO TRAIN_STEP step=2859 loss=1.652805 Training: 16%|█▌ | 2859/18018 [6:11:31<33:58:12, 8.07s/step, loss=1.653, acc=0.651, acc_len=1.53, thru=15.0, I=16.7, T=29.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:41:46,028] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:41:54,725] loop.py:380 INFO TIMING step=2860: step=4.002s data=1.797s compute=3.964s [fwd=2.830s bwd=1.119s opt=0.003s] dispatch=4.978s [2026-10-18 19:41:54,726] loop.py:45 INFO TRAIN_STEP step=2860 loss=1.710780 Training: 16%|█▌ | 2860/18018 [6:11:40<35:08:03, 8.34s/step, loss=1.711, acc=0.637, acc_len=1.49, thru=14.3, I=14.9, T=32.0, wait=5.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:01,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 19:42:02,429] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2860: forward=2778.0ms backward=1092.1ms [2026-10-18 19:42:02,434] loop.py:45 INFO TRAIN_STEP step=2861 loss=1.672095 Training: 16%|█▌ | 2861/18018 [6:11:48<34:19:45, 8.15s/step, loss=1.672, acc=0.643, acc_len=1.51, thru=16.5, I=17.8, T=32.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:02,499] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 19:42:09,036] loop.py:45 INFO TRAIN_STEP step=2862 loss=1.640923 Training: 16%|█▌ | 2862/18018 [6:11:55<32:21:49, 7.69s/step, loss=1.641, acc=0.646, acc_len=1.52, thru=18.6, I=20.8, T=34.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:08,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:09,498] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:15,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:42:16,840] loop.py:45 INFO TRAIN_STEP step=2863 loss=1.924664 Training: 16%|█▌ | 2863/18018 [6:12:02<32:30:32, 7.72s/step, loss=1.925, acc=0.603, acc_len=1.39, thru=16.1, I=19.2, T=28.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:16,923] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:27,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:42:27,960] loop.py:45 INFO TRAIN_STEP step=2864 loss=1.890712 Training: 16%|█▌ | 2864/18018 [6:12:14<36:48:03, 8.74s/step, loss=1.891, acc=0.611, acc_len=1.40, thru=10.9, I=11.6, T=26.8, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:28,303] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:34,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:42:35,183] loop.py:380 INFO TIMING step=2865: step=3.807s data=1.660s compute=3.772s [fwd=2.628s bwd=1.129s opt=0.003s] dispatch=3.405s [2026-10-18 19:42:35,183] loop.py:45 INFO TRAIN_STEP step=2865 loss=1.683424 Training: 16%|█▌ | 2865/18018 [6:12:21<34:52:36, 8.29s/step, loss=1.683, acc=0.640, acc_len=1.50, thru=16.6, I=18.1, T=33.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:35,672] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:42:44,914] loop.py:45 INFO TRAIN_STEP step=2866 loss=1.782506 Training: 16%|█▌ | 2866/18018 [6:12:31<36:42:18, 8.72s/step, loss=1.783, acc=0.626, acc_len=1.45, thru=13.4, I=12.8, T=30.4, wait=5.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:42:44,909] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2865: forward=2940.5ms backward=1217.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:52,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:42:53,966] loop.py:45 INFO TRAIN_STEP step=2867 loss=1.806298 Training: 16%|█▌ | 2867/18018 [6:12:40<37:06:57, 8.82s/step, loss=1.806, acc=0.624, acc_len=1.45, thru=14.6, I=15.5, T=31.3, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:42:54,392] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:00,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:43:02,020] loop.py:45 INFO TRAIN_STEP step=2868 loss=1.754886 Training: 16%|█▌ | 2868/18018 [6:12:48<36:08:49, 8.59s/step, loss=1.755, acc=0.626, acc_len=1.47, thru=15.1, I=21.1, T=32.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:02,079] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:10,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:43:11,235] loop.py:45 INFO TRAIN_STEP step=2869 loss=1.694643 Training: 16%|█▌ | 2869/18018 [6:12:57<36:56:10, 8.78s/step, loss=1.695, acc=0.638, acc_len=1.50, thru=14.0, I=14.2, T=28.9, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:11,321] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:18,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:43:20,244] loop.py:380 INFO TIMING step=2870: step=4.698s data=2.146s compute=4.666s [fwd=3.274s bwd=1.376s opt=0.003s] dispatch=4.300s [2026-10-18 19:43:20,244] loop.py:45 INFO TRAIN_STEP step=2870 loss=1.896357 Training: 16%|█▌ | 2870/18018 [6:13:06<37:13:32, 8.85s/step, loss=1.896, acc=0.610, acc_len=1.41, thru=14.5, I=17.4, T=27.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:20,467] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:27,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:43:27,841] loop.py:45 INFO TRAIN_STEP step=2871 loss=1.686361 Training: 16%|█▌ | 2871/18018 [6:13:13<35:38:45, 8.47s/step, loss=1.686, acc=0.639, acc_len=1.50, thru=14.7, I=18.4, T=33.2, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:43:27,835] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2870: forward=2704.0ms backward=1098.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:28,226] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:34,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:43:35,404] loop.py:45 INFO TRAIN_STEP step=2872 loss=1.718767 Training: 16%|█▌ | 2872/18018 [6:13:21<34:29:49, 8.20s/step, loss=1.719, acc=0.639, acc_len=1.49, thru=18.5, I=18.7, T=32.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:35,666] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:41,323] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:43:42,531] loop.py:45 INFO TRAIN_STEP step=2873 loss=1.813181 Training: 16%|█▌ | 2873/18018 [6:13:28<33:08:27, 7.88s/step, loss=1.813, acc=0.621, acc_len=1.44, thru=16.9, I=21.0, T=34.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:42,831] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:50,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:43:50,375] loop.py:45 INFO TRAIN_STEP step=2874 loss=1.753006 Training: 16%|█▌ | 2874/18018 [6:13:36<33:05:44, 7.87s/step, loss=1.753, acc=0.630, acc_len=1.46, thru=15.3, I=17.5, T=34.0, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:43:50,615] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 19:43:56,924] loop.py:380 INFO TIMING step=2875: step=3.937s data=1.731s compute=3.898s [fwd=2.753s bwd=1.131s opt=0.003s] dispatch=2.602s [2026-10-18 19:43:56,924] loop.py:45 INFO TRAIN_STEP step=2875 loss=1.731016 Training: 16%|█▌ | 2875/18018 [6:13:43<31:25:43, 7.47s/step, loss=1.731, acc=0.637, acc_len=1.49, thru=18.8, I=18.4, T=32.5, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:05,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:44:06,641] loop.py:45 INFO TRAIN_STEP step=2876 loss=1.886917 Training: 16%|█▌ | 2876/18018 [6:13:52<34:15:40, 8.15s/step, loss=1.887, acc=0.610, acc_len=1.41, thru=14.2, I=15.3, T=25.6, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:44:06,636] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2875: forward=3412.6ms backward=1524.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:06,760] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:15,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:44:15,929] loop.py:45 INFO TRAIN_STEP step=2877 loss=1.822736 Training: 16%|█▌ | 2877/18018 [6:14:02<35:41:56, 8.49s/step, loss=1.823, acc=0.615, acc_len=1.42, thru=14.0, I=14.5, T=29.4, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:15,960] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:20,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:44:22,084] loop.py:45 INFO TRAIN_STEP step=2878 loss=1.698187 Training: 16%|█▌ | 2878/18018 [6:14:08<32:45:15, 7.79s/step, loss=1.698, acc=0.638, acc_len=1.49, thru=19.5, I=19.5, T=32.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:22,386] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:27,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:44:29,923] loop.py:45 INFO TRAIN_STEP step=2879 loss=1.887105 Training: 16%|█▌ | 2879/18018 [6:14:16<32:48:59, 7.80s/step, loss=1.887, acc=0.605, acc_len=1.40, thru=15.6, I=21.4, T=28.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:29,984] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:34,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:44:36,930] loop.py:380 INFO TIMING step=2880: step=4.063s data=1.909s compute=4.021s [fwd=2.844s bwd=1.162s opt=0.003s] dispatch=2.933s [2026-10-18 19:44:36,930] loop.py:45 INFO TRAIN_STEP step=2880 loss=1.911560 Training: 16%|█▌ | 2880/18018 [6:14:23<31:48:35, 7.56s/step, loss=1.912, acc=0.604, acc_len=1.39, thru=15.2, I=25.9, T=31.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:37,067] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=16/64 [2026-10-18 19:44:43,348] loop.py:45 INFO TRAIN_STEP step=2881 loss=1.621011 Training: 16%|█▌ | 2881/18018 [6:14:29<30:21:23, 7.22s/step, loss=1.621, acc=0.654, acc_len=1.54, thru=16.0, I=21.4, T=34.2, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:44:43,343] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2880: forward=2648.4ms backward=1053.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:51,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:44:52,992] loop.py:45 INFO TRAIN_STEP step=2882 loss=1.874264 Training: 16%|█▌ | 2882/18018 [6:14:39<33:24:50, 7.95s/step, loss=1.874, acc=0.609, acc_len=1.41, thru=14.5, I=16.7, T=28.6, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:44:53,371] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:45:00,192] loop.py:45 INFO TRAIN_STEP step=2883 loss=1.686767 Training: 16%|█▌ | 2883/18018 [6:14:46<32:28:11, 7.72s/step, loss=1.687, acc=0.635, acc_len=1.49, thru=16.1, I=20.6, T=28.7, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:45:09,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:45:10,662] loop.py:45 INFO TRAIN_STEP step=2884 loss=1.761459 Training: 16%|█▌ | 2884/18018 [6:14:56<35:55:54, 8.55s/step, loss=1.761, acc=0.630, acc_len=1.47, thru=12.6, I=12.8, T=29.6, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:45:10,904] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 19:45:19,481] loop.py:380 INFO TIMING step=2885: step=4.402s data=2.015s compute=4.322s [fwd=3.029s bwd=1.277s opt=0.003s] dispatch=4.408s [2026-10-18 19:45:19,481] loop.py:45 INFO TRAIN_STEP step=2885 loss=1.774942 Training: 16%|█▌ | 2885/18018 [6:15:05<36:16:15, 8.63s/step, loss=1.775, acc=0.624, acc_len=1.45, thru=13.3, I=14.3, T=29.1, wait=4.4s, pool=60, epoch=1/2][2026-10-18 19:45:29,363] loop.py:45 INFO TRAIN_STEP step=2886 loss=1.867046 Training: 16%|█▌ | 2886/18018 [6:15:15<37:51:02, 9.00s/step, loss=1.867, acc=0.613, acc_len=1.42, thru=11.3, I=12.4, T=31.6, wait=5.8s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:45:29,358] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2885: forward=2733.9ms backward=1257.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:45:38,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:45:38,893] loop.py:45 INFO TRAIN_STEP step=2887 loss=1.663920 Training: 16%|█▌ | 2887/18018 [6:15:24<38:30:43, 9.16s/step, loss=1.664, acc=0.643, acc_len=1.51, thru=15.9, I=15.0, T=32.8, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:45:39,021] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:45:45,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:45:46,987] loop.py:45 INFO TRAIN_STEP step=2888 loss=1.808644 Training: 16%|█▌ | 2888/18018 [6:15:33<37:09:42, 8.84s/step, loss=1.809, acc=0.621, acc_len=1.44, thru=14.9, I=20.2, T=30.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:45:47,176] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:45:52,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:45:54,512] loop.py:45 INFO TRAIN_STEP step=2889 loss=1.628680 Training: 16%|█▌ | 2889/18018 [6:15:40<35:29:50, 8.45s/step, loss=1.629, acc=0.647, acc_len=1.53, thru=14.5, I=22.4, T=28.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:45:54,856] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:46:01,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:46:02,355] loop.py:380 INFO TIMING step=2890: step=3.869s data=1.779s compute=3.831s [fwd=2.676s bwd=1.139s opt=0.003s] dispatch=3.963s [2026-10-18 19:46:02,355] loop.py:45 INFO TRAIN_STEP step=2890 loss=1.680840 Training: 16%|█▌ | 2890/18018 [6:15:48<34:43:53, 8.27s/step, loss=1.681, acc=0.637, acc_len=1.49, thru=14.0, I=18.5, T=33.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:46:02,688] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:46:09,534] loop.py:45 INFO TRAIN_STEP step=2891 loss=1.737728 Training: 16%|█▌ | 2891/18018 [6:15:55<33:21:37, 7.94s/step, loss=1.738, acc=0.630, acc_len=1.47, thru=16.8, I=18.9, T=31.4, wait=3.1s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:46:09,529] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2890: forward=2764.8ms backward=1240.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:46:15,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:46:16,912] loop.py:45 INFO TRAIN_STEP step=2892 loss=1.861254 Training: 16%|█▌ | 2892/18018 [6:16:02<32:39:15, 7.77s/step, loss=1.861, acc=0.610, acc_len=1.41, thru=18.2, I=19.5, T=30.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:46:17,106] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:46:23,452] loop.py:45 INFO TRAIN_STEP step=2893 loss=1.659075 Training: 16%|█▌ | 2893/18018 [6:16:09<31:05:51, 7.40s/step, loss=1.659, acc=0.643, acc_len=1.51, thru=15.4, I=23.5, T=33.5, wait=2.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:46:32,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:46:33,542] loop.py:45 INFO TRAIN_STEP step=2894 loss=1.891546 Training: 16%|█▌ | 2894/18018 [6:16:19<34:29:10, 8.21s/step, loss=1.892, acc=0.614, acc_len=1.41, thru=14.8, I=14.9, T=28.4, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:46:33,594] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 19:46:40,466] loop.py:380 INFO TIMING step=2895: step=4.560s data=2.090s compute=4.531s [fwd=3.284s bwd=1.232s opt=0.003s] dispatch=2.353s [2026-10-18 19:46:40,466] loop.py:45 INFO TRAIN_STEP step=2895 loss=1.653945 Training: 16%|█▌ | 2895/18018 [6:16:26<32:51:49, 7.82s/step, loss=1.654, acc=0.647, acc_len=1.52, thru=16.6, I=22.9, T=28.1, wait=2.4s, pool=60, epoch=1/2][2026-10-18 19:46:48,874] loop.py:45 INFO TRAIN_STEP step=2896 loss=1.795978 Training: 16%|█▌ | 2896/18018 [6:16:34<33:35:41, 8.00s/step, loss=1.796, acc=0.620, acc_len=1.44, thru=13.6, I=13.3, T=26.7, wait=3.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:46:48,869] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2895: forward=3372.7ms backward=1376.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:46:57,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:46:57,903] loop.py:45 INFO TRAIN_STEP step=2897 loss=1.810081 Training: 16%|█▌ | 2897/18018 [6:16:43<34:53:44, 8.31s/step, loss=1.810, acc=0.620, acc_len=1.45, thru=14.9, I=13.9, T=28.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:46:58,317] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:02,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:47:04,188] loop.py:45 INFO TRAIN_STEP step=2898 loss=1.690217 Training: 16%|█▌ | 2898/18018 [6:16:50<32:20:34, 7.70s/step, loss=1.690, acc=0.640, acc_len=1.50, thru=18.8, I=23.5, T=32.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:04,531] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:10,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:47:11,777] loop.py:45 INFO TRAIN_STEP step=2899 loss=1.790288 Training: 16%|█▌ | 2899/18018 [6:16:57<32:12:07, 7.67s/step, loss=1.790, acc=0.622, acc_len=1.45, thru=14.7, I=20.4, T=27.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:12,031] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:20,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:47:21,456] loop.py:380 INFO TIMING step=2900: step=4.190s data=1.911s compute=4.145s [fwd=2.883s bwd=1.249s opt=0.003s] dispatch=5.478s [2026-10-18 19:47:21,456] loop.py:45 INFO TRAIN_STEP step=2900 loss=1.717717 Training: 16%|█▌ | 2900/18018 [6:17:07<34:43:56, 8.27s/step, loss=1.718, acc=0.638, acc_len=1.49, thru=13.6, I=15.6, T=30.5, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:21,732] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:47:27,811] loop.py:45 INFO TRAIN_STEP step=2901 loss=1.737618 Training: 16%|█▌ | 2901/18018 [6:17:13<32:18:54, 7.70s/step, loss=1.738, acc=0.632, acc_len=1.48, thru=18.2, I=19.2, T=30.9, wait=2.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:47:27,806] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2900: forward=2893.2ms backward=1189.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:35,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:47:36,357] loop.py:45 INFO TRAIN_STEP step=2902 loss=1.798185 Training: 16%|█▌ | 2902/18018 [6:17:22<33:23:03, 7.95s/step, loss=1.798, acc=0.626, acc_len=1.45, thru=16.8, I=16.1, T=30.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:36,456] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 19:47:42,414] loop.py:45 INFO TRAIN_STEP step=2903 loss=1.809176 Training: 16%|█▌ | 2903/18018 [6:17:28<30:59:44, 7.38s/step, loss=1.809, acc=0.623, acc_len=1.45, thru=18.4, I=20.4, T=35.8, wait=2.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:48,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:47:50,221] loop.py:45 INFO TRAIN_STEP step=2904 loss=1.822203 Training: 16%|█▌ | 2904/18018 [6:17:36<31:31:58, 7.51s/step, loss=1.822, acc=0.615, acc_len=1.42, thru=17.9, I=20.7, T=28.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:50,255] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:55,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:47:56,588] loop.py:380 INFO TIMING step=2905: step=3.959s data=1.810s compute=3.923s [fwd=2.740s bwd=1.168s opt=0.003s] dispatch=2.397s [2026-10-18 19:47:56,588] loop.py:45 INFO TRAIN_STEP step=2905 loss=1.771307 Training: 16%|█▌ | 2905/18018 [6:17:42<30:05:23, 7.17s/step, loss=1.771, acc=0.629, acc_len=1.47, thru=17.2, I=23.4, T=32.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:47:56,625] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 19:48:03,249] loop.py:45 INFO TRAIN_STEP step=2906 loss=1.830813 Training: 16%|█▌ | 2906/18018 [6:17:49<29:26:50, 7.01s/step, loss=1.831, acc=0.616, acc_len=1.43, thru=15.2, I=20.3, T=31.1, wait=2.5s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:48:03,244] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2905: forward=2922.9ms backward=1149.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:09,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:48:12,313] loop.py:45 INFO TRAIN_STEP step=2907 loss=1.771675 Training: 16%|█▌ | 2907/18018 [6:17:58<32:01:32, 7.63s/step, loss=1.772, acc=0.628, acc_len=1.46, thru=17.5, I=19.5, T=26.9, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:12,640] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:17,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:48:19,219] loop.py:45 INFO TRAIN_STEP step=2908 loss=1.720110 Training: 16%|█▌ | 2908/18018 [6:18:05<31:06:47, 7.41s/step, loss=1.720, acc=0.640, acc_len=1.50, thru=14.0, I=25.2, T=33.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:19,421] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:23,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:48:25,114] loop.py:45 INFO TRAIN_STEP step=2909 loss=1.729248 Training: 16%|█▌ | 2909/18018 [6:18:11<29:12:08, 6.96s/step, loss=1.729, acc=0.636, acc_len=1.48, thru=20.5, I=26.3, T=35.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:25,335] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:31,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:48:32,023] loop.py:380 INFO TIMING step=2910: step=3.797s data=1.714s compute=3.762s [fwd=2.704s bwd=1.045s opt=0.003s] dispatch=3.100s [2026-10-18 19:48:32,023] loop.py:45 INFO TRAIN_STEP step=2910 loss=1.856180 Training: 16%|█▌ | 2910/18018 [6:18:18<29:08:16, 6.94s/step, loss=1.856, acc=0.613, acc_len=1.41, thru=17.3, I=21.7, T=33.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:32,484] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:38,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:48:39,151] loop.py:45 INFO TRAIN_STEP step=2911 loss=1.862716 Training: 16%|█▌ | 2911/18018 [6:18:25<29:22:09, 7.00s/step, loss=1.863, acc=0.612, acc_len=1.41, thru=18.1, I=21.3, T=30.1, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:48:39,147] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2910: forward=3009.7ms backward=1190.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:39,519] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:46,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:48:47,076] loop.py:45 INFO TRAIN_STEP step=2912 loss=1.656071 Training: 16%|█▌ | 2912/18018 [6:18:33<30:31:57, 7.28s/step, loss=1.656, acc=0.643, acc_len=1.51, thru=14.9, I=18.7, T=35.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:47,308] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:53,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:48:53,852] loop.py:45 INFO TRAIN_STEP step=2913 loss=1.763763 Training: 16%|█▌ | 2913/18018 [6:18:39<29:53:57, 7.13s/step, loss=1.764, acc=0.628, acc_len=1.47, thru=18.1, I=20.9, T=33.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:54,023] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:48:59,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:49:00,399] loop.py:45 INFO TRAIN_STEP step=2914 loss=1.756048 Training: 16%|█▌ | 2914/18018 [6:18:46<29:10:11, 6.95s/step, loss=1.756, acc=0.629, acc_len=1.47, thru=18.9, I=21.1, T=33.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:00,648] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:06,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:49:07,230] loop.py:380 INFO TIMING step=2915: step=4.295s data=2.049s compute=4.259s [fwd=2.928s bwd=1.318s opt=0.003s] dispatch=2.526s [2026-10-18 19:49:07,231] loop.py:45 INFO TRAIN_STEP step=2915 loss=1.621111 Training: 16%|█▌ | 2915/18018 [6:18:53<29:00:48, 6.92s/step, loss=1.621, acc=0.651, acc_len=1.54, thru=17.6, I=19.8, T=29.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:07,469] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 19:49:15,197] loop.py:45 INFO TRAIN_STEP step=2916 loss=1.826226 Training: 16%|█▌ | 2916/18018 [6:19:01<30:20:01, 7.23s/step, loss=1.826, acc=0.621, acc_len=1.44, thru=16.0, I=16.8, T=26.2, wait=3.1s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:49:15,192] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2915: forward=3281.5ms backward=1536.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:21,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:49:22,221] loop.py:45 INFO TRAIN_STEP step=2917 loss=1.722824 Training: 16%|█▌ | 2917/18018 [6:19:08<30:04:30, 7.17s/step, loss=1.723, acc=0.635, acc_len=1.49, thru=17.8, I=17.7, T=31.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:22,283] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:29,767] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:49:30,746] loop.py:45 INFO TRAIN_STEP step=2918 loss=1.817694 Training: 16%|█▌ | 2918/18018 [6:19:16<31:46:36, 7.58s/step, loss=1.818, acc=0.616, acc_len=1.42, thru=13.6, I=16.9, T=22.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:30,780] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:37,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:49:37,967] loop.py:45 INFO TRAIN_STEP step=2919 loss=1.659728 Training: 16%|█▌ | 2919/18018 [6:19:24<31:19:45, 7.47s/step, loss=1.660, acc=0.644, acc_len=1.51, thru=15.2, I=16.8, T=33.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:38,148] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 19:49:44,957] loop.py:380 INFO TIMING step=2920: step=3.659s data=1.640s compute=3.630s [fwd=2.575s bwd=1.042s opt=0.003s] dispatch=3.321s [2026-10-18 19:49:44,957] loop.py:45 INFO TRAIN_STEP step=2920 loss=1.626100 Training: 16%|█▌ | 2920/18018 [6:19:31<30:43:16, 7.33s/step, loss=1.626, acc=0.648, acc_len=1.52, thru=17.0, I=18.9, T=35.0, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:52,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:49:54,862] loop.py:45 INFO TRAIN_STEP step=2921 loss=1.838213 Training: 16%|█▌ | 2921/18018 [6:19:40<33:57:56, 8.10s/step, loss=1.838, acc=0.617, acc_len=1.43, thru=14.3, I=17.5, T=27.5, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:49:54,857] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2920: forward=3303.3ms backward=1301.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:49:55,092] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 19:50:01,537] loop.py:45 INFO TRAIN_STEP step=2922 loss=1.678378 Training: 16%|█▌ | 2922/18018 [6:19:47<32:10:37, 7.67s/step, loss=1.678, acc=0.643, acc_len=1.51, thru=14.9, I=20.3, T=31.6, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:06,326] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:50:07,994] loop.py:45 INFO TRAIN_STEP step=2923 loss=1.776585 Training: 16%|█▌ | 2923/18018 [6:19:54<30:38:21, 7.31s/step, loss=1.777, acc=0.628, acc_len=1.46, thru=18.7, I=22.0, T=31.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:08,342] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:50:14,419] loop.py:45 INFO TRAIN_STEP step=2924 loss=1.587446 Training: 16%|█▌ | 2924/18018 [6:20:00<29:31:29, 7.04s/step, loss=1.587, acc=0.661, acc_len=1.57, thru=18.2, I=22.5, T=31.4, wait=2.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:23,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:50:23,658] loop.py:380 INFO TIMING step=2925: step=4.184s data=1.916s compute=4.142s [fwd=2.959s bwd=1.169s opt=0.003s] dispatch=5.048s [2026-10-18 19:50:23,658] loop.py:45 INFO TRAIN_STEP step=2925 loss=1.702033 Training: 16%|█▌ | 2925/18018 [6:20:09<32:17:13, 7.70s/step, loss=1.702, acc=0.637, acc_len=1.49, thru=15.1, I=15.0, T=30.6, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:23,845] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:28,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:50:30,250] loop.py:45 INFO TRAIN_STEP step=2926 loss=1.747110 Training: 16%|█▌ | 2926/18018 [6:20:16<30:53:29, 7.37s/step, loss=1.747, acc=0.629, acc_len=1.47, thru=18.1, I=20.1, T=30.1, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:50:30,245] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2925: forward=3019.3ms backward=1171.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:30,670] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:50:38,491] loop.py:45 INFO TRAIN_STEP step=2927 loss=1.685025 Training: 16%|█▌ | 2927/18018 [6:20:24<31:59:09, 7.63s/step, loss=1.685, acc=0.635, acc_len=1.50, thru=11.0, I=14.2, T=32.7, wait=4.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:47,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:50:47,640] loop.py:45 INFO TRAIN_STEP step=2928 loss=1.695257 Training: 16%|█▋ | 2928/18018 [6:20:33<33:53:41, 8.09s/step, loss=1.695, acc=0.638, acc_len=1.50, thru=16.4, I=16.6, T=28.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:47,925] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:55,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:50:58,127] loop.py:45 INFO TRAIN_STEP step=2929 loss=1.772157 Training: 16%|█▋ | 2929/18018 [6:20:44<36:54:29, 8.81s/step, loss=1.772, acc=0.625, acc_len=1.46, thru=13.0, I=15.1, T=27.6, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:50:58,451] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:03,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:51:03,903] loop.py:380 INFO TIMING step=2930: step=3.626s data=1.604s compute=3.600s [fwd=2.584s bwd=1.002s opt=0.003s] dispatch=2.142s [2026-10-18 19:51:03,903] loop.py:45 INFO TRAIN_STEP step=2930 loss=1.572915 Training: 16%|█▋ | 2930/18018 [6:20:49<33:05:55, 7.90s/step, loss=1.573, acc=0.657, acc_len=1.56, thru=18.8, I=25.5, T=35.3, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:04,094] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:12,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:51:12,634] loop.py:45 INFO TRAIN_STEP step=2931 loss=1.608706 Training: 16%|█▋ | 2931/18018 [6:20:58<34:08:35, 8.15s/step, loss=1.609, acc=0.653, acc_len=1.54, thru=14.8, I=15.9, T=31.7, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:51:12,630] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2930: forward=2813.9ms backward=1179.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:13,059] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:19,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:51:20,252] loop.py:45 INFO TRAIN_STEP step=2932 loss=1.834051 Training: 16%|█▋ | 2932/18018 [6:21:06<33:28:34, 7.99s/step, loss=1.834, acc=0.612, acc_len=1.42, thru=17.6, I=20.4, T=31.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:20,282] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 19:51:27,727] loop.py:45 INFO TRAIN_STEP step=2933 loss=1.801727 Training: 16%|█▋ | 2933/18018 [6:21:13<32:49:50, 7.83s/step, loss=1.802, acc=0.626, acc_len=1.45, thru=18.0, I=17.4, T=32.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:27,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:28,234] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:33,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:51:35,767] loop.py:45 INFO TRAIN_STEP step=2934 loss=1.842519 Training: 16%|█▋ | 2934/18018 [6:21:21<33:05:12, 7.90s/step, loss=1.843, acc=0.615, acc_len=1.42, thru=17.7, I=19.8, T=31.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:35,992] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:40,904] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:51:42,629] loop.py:380 INFO TIMING step=2935: step=3.668s data=1.608s compute=3.636s [fwd=2.620s bwd=1.001s opt=0.003s] dispatch=3.183s [2026-10-18 19:51:42,629] loop.py:45 INFO TRAIN_STEP step=2935 loss=1.831301 Training: 16%|█▋ | 2935/18018 [6:21:28<31:46:54, 7.59s/step, loss=1.831, acc=0.619, acc_len=1.43, thru=16.8, I=25.9, T=34.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:51:42,917] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:51:49,033] loop.py:45 INFO TRAIN_STEP step=2936 loss=1.603402 Training: 16%|█▋ | 2936/18018 [6:21:35<30:17:37, 7.23s/step, loss=1.603, acc=0.654, acc_len=1.55, thru=18.3, I=22.3, T=34.4, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:51:49,027] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2935: forward=2612.9ms backward=1058.9ms [2026-10-18 19:51:57,330] loop.py:45 INFO TRAIN_STEP step=2937 loss=1.978181 Training: 16%|█▋ | 2937/18018 [6:21:43<31:37:50, 7.55s/step, loss=1.978, acc=0.591, acc_len=1.35, thru=14.4, I=13.9, T=32.0, wait=4.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:04,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:52:05,250] loop.py:45 INFO TRAIN_STEP step=2938 loss=1.796071 Training: 16%|█▋ | 2938/18018 [6:21:51<32:05:45, 7.66s/step, loss=1.796, acc=0.624, acc_len=1.45, thru=18.1, I=16.5, T=33.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:05,486] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:11,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:52:12,694] loop.py:45 INFO TRAIN_STEP step=2939 loss=1.878892 Training: 16%|█▋ | 2939/18018 [6:21:58<31:49:12, 7.60s/step, loss=1.879, acc=0.609, acc_len=1.40, thru=16.6, I=20.0, T=31.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:13,050] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:19,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:52:19,840] loop.py:380 INFO TIMING step=2940: step=4.679s data=2.244s compute=4.651s [fwd=3.215s bwd=1.421s opt=0.003s] dispatch=2.457s [2026-10-18 19:52:19,841] loop.py:45 INFO TRAIN_STEP step=2940 loss=1.706586 Training: 16%|█▋ | 2940/18018 [6:22:05<31:14:56, 7.46s/step, loss=1.707, acc=0.638, acc_len=1.49, thru=16.5, I=23.4, T=27.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:20,252] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:26,009] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:52:27,482] loop.py:45 INFO TRAIN_STEP step=2941 loss=1.880679 Training: 16%|█▋ | 2941/18018 [6:22:13<31:28:29, 7.52s/step, loss=1.881, acc=0.609, acc_len=1.41, thru=15.7, I=19.2, T=28.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:27,527] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (TrainerActor pid=9310) [2026-10-18 19:52:27,476] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2940: forward=3140.5ms backward=1257.2ms [2026-10-18 19:52:36,719] loop.py:45 INFO TRAIN_STEP step=2942 loss=1.742649 Training: 16%|█▋ | 2942/18018 [6:22:22<33:38:24, 8.03s/step, loss=1.743, acc=0.632, acc_len=1.47, thru=11.8, I=12.6, T=28.5, wait=4.7s, pool=44, epoch=1/2][2026-10-18 19:52:44,615] loop.py:45 INFO TRAIN_STEP step=2943 loss=1.808571 Training: 16%|█▋ | 2943/18018 [6:22:30<33:27:39, 7.99s/step, loss=1.809, acc=0.620, acc_len=1.44, thru=15.5, I=15.4, T=29.3, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:51,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:52:52,733] loop.py:45 INFO TRAIN_STEP step=2944 loss=1.893157 Training: 16%|█▋ | 2944/18018 [6:22:38<33:37:16, 8.03s/step, loss=1.893, acc=0.613, acc_len=1.41, thru=16.5, I=17.3, T=31.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:52,838] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:58,888] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:52:59,807] loop.py:380 INFO TIMING step=2945: step=4.434s data=2.335s compute=4.398s [fwd=3.301s bwd=1.082s opt=0.003s] dispatch=2.629s [2026-10-18 19:52:59,807] loop.py:45 INFO TRAIN_STEP step=2945 loss=1.835989 Training: 16%|█▋ | 2945/18018 [6:22:45<32:25:05, 7.74s/step, loss=1.836, acc=0.623, acc_len=1.45, thru=16.6, I=21.4, T=28.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:52:59,905] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:53:06,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:53:07,988] loop.py:45 INFO TRAIN_STEP step=2946 loss=1.893479 Training: 16%|█▋ | 2946/18018 [6:22:54<32:57:59, 7.87s/step, loss=1.893, acc=0.607, acc_len=1.41, thru=14.9, I=17.8, T=26.2, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:53:07,982] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2945: forward=3683.8ms backward=1142.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:53:08,285] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:53:15,456] loop.py:45 INFO TRAIN_STEP step=2947 loss=1.736814 Training: 16%|█▋ | 2947/18018 [6:23:01<32:27:15, 7.75s/step, loss=1.737, acc=0.629, acc_len=1.47, thru=15.1, I=19.0, T=32.5, wait=3.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:53:23,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:53:25,189] loop.py:45 INFO TRAIN_STEP step=2948 loss=1.809115 Training: 16%|█▋ | 2948/18018 [6:23:11<34:56:24, 8.35s/step, loss=1.809, acc=0.620, acc_len=1.44, thru=14.0, I=14.7, T=27.9, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:53:25,425] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:53:32,549] loop.py:45 INFO TRAIN_STEP step=2949 loss=1.552460 Training: 16%|█▋ | 2949/18018 [6:23:18<33:41:43, 8.05s/step, loss=1.552, acc=0.662, acc_len=1.58, thru=13.9, I=16.5, T=31.5, wait=3.3s, pool=48, epoch=1/2][2026-10-18 19:53:42,819] loop.py:380 INFO TIMING step=2950: step=4.541s data=2.028s compute=4.504s [fwd=3.194s bwd=1.295s opt=0.003s] dispatch=5.720s [2026-10-18 19:53:42,819] loop.py:45 INFO TRAIN_STEP step=2950 loss=1.767923 Training: 16%|█▋ | 2950/18018 [6:23:28<36:29:00, 8.72s/step, loss=1.768, acc=0.626, acc_len=1.46, thru=13.7, I=14.3, T=28.2, wait=5.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:53:50,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:53:52,444] loop.py:45 INFO TRAIN_STEP step=2951 loss=1.806118 Training: 16%|█▋ | 2951/18018 [6:23:38<37:37:11, 8.99s/step, loss=1.806, acc=0.628, acc_len=1.46, thru=14.3, I=15.1, T=29.4, wait=5.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:53:52,439] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2950: forward=3042.4ms backward=1256.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:53:52,708] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:53:59,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:53:59,370] loop.py:45 INFO TRAIN_STEP step=2952 loss=1.787759 Training: 16%|█▋ | 2952/18018 [6:23:45<35:01:44, 8.37s/step, loss=1.788, acc=0.624, acc_len=1.46, thru=16.3, I=21.1, T=30.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:53:59,668] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 19:54:06,843] loop.py:45 INFO TRAIN_STEP step=2953 loss=1.736630 Training: 16%|█▋ | 2953/18018 [6:23:52<33:53:51, 8.10s/step, loss=1.737, acc=0.627, acc_len=1.46, thru=14.8, I=16.0, T=28.5, wait=3.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:12,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:54:14,080] loop.py:45 INFO TRAIN_STEP step=2954 loss=1.750051 Training: 16%|█▋ | 2954/18018 [6:24:00<32:48:57, 7.84s/step, loss=1.750, acc=0.629, acc_len=1.46, thru=18.6, I=19.8, T=27.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:14,532] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 19:54:24,424] loop.py:380 INFO TIMING step=2955: step=5.053s data=2.468s compute=5.008s [fwd=3.462s bwd=1.530s opt=0.003s] dispatch=5.281s [2026-10-18 19:54:24,425] loop.py:45 INFO TRAIN_STEP step=2955 loss=1.869038 Training: 16%|█▋ | 2955/18018 [6:24:10<35:57:01, 8.59s/step, loss=1.869, acc=0.610, acc_len=1.41, thru=11.3, I=12.0, T=25.3, wait=5.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:33,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:54:35,097] loop.py:45 INFO TRAIN_STEP step=2956 loss=1.821195 Training: 16%|█▋ | 2956/18018 [6:24:21<38:33:39, 9.22s/step, loss=1.821, acc=0.623, acc_len=1.45, thru=12.8, I=14.7, T=25.0, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:54:35,093] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2955: forward=3514.8ms backward=1540.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:35,298] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:39,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:54:41,273] loop.py:45 INFO TRAIN_STEP step=2957 loss=1.712601 Training: 16%|█▋ | 2957/18018 [6:24:27<34:44:32, 8.30s/step, loss=1.713, acc=0.636, acc_len=1.48, thru=16.9, I=20.0, T=31.4, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:41,419] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:47,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:54:48,401] loop.py:45 INFO TRAIN_STEP step=2958 loss=1.718477 Training: 16%|█▋ | 2958/18018 [6:24:34<33:15:49, 7.95s/step, loss=1.718, acc=0.634, acc_len=1.49, thru=17.1, I=20.0, T=31.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:48,860] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:53,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:54:55,216] loop.py:45 INFO TRAIN_STEP step=2959 loss=1.708134 Training: 16%|█▋ | 2959/18018 [6:24:41<31:50:14, 7.61s/step, loss=1.708, acc=0.634, acc_len=1.48, thru=17.5, I=21.9, T=30.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:54:55,349] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:01,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:55:02,217] loop.py:380 INFO TIMING step=2960: step=4.246s data=1.849s compute=4.203s [fwd=2.922s bwd=1.265s opt=0.003s] dispatch=2.744s [2026-10-18 19:55:02,217] loop.py:45 INFO TRAIN_STEP step=2960 loss=1.835850 Training: 16%|█▋ | 2960/18018 [6:24:48<31:04:02, 7.43s/step, loss=1.836, acc=0.617, acc_len=1.43, thru=16.4, I=20.2, T=30.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:02,538] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:08,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:55:09,894] loop.py:45 INFO TRAIN_STEP step=2961 loss=1.828485 Training: 16%|█▋ | 2961/18018 [6:24:55<31:22:42, 7.50s/step, loss=1.828, acc=0.619, acc_len=1.44, thru=16.2, I=20.0, T=27.0, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:55:09,888] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2960: forward=3271.7ms backward=1421.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:10,049] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:13,733] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:55:15,613] loop.py:45 INFO TRAIN_STEP step=2962 loss=1.677160 Training: 16%|█▋ | 2962/18018 [6:25:01<29:08:28, 6.97s/step, loss=1.677, acc=0.641, acc_len=1.50, thru=19.2, I=24.9, T=32.5, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:15,757] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:22,188] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:55:23,543] loop.py:45 INFO TRAIN_STEP step=2963 loss=1.641676 Training: 16%|█▋ | 2963/18018 [6:25:09<30:20:41, 7.26s/step, loss=1.642, acc=0.640, acc_len=1.51, thru=14.1, I=20.1, T=26.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:23,694] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:28,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:55:31,122] loop.py:45 INFO TRAIN_STEP step=2964 loss=1.805365 Training: 16%|█▋ | 2964/18018 [6:25:17<30:44:56, 7.35s/step, loss=1.805, acc=0.622, acc_len=1.45, thru=15.0, I=16.9, T=27.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:31,451] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:37,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:55:38,768] loop.py:380 INFO TIMING step=2965: step=4.125s data=1.872s compute=4.093s [fwd=2.885s bwd=1.192s opt=0.003s] dispatch=3.510s [2026-10-18 19:55:38,768] loop.py:45 INFO TRAIN_STEP step=2965 loss=1.776567 Training: 16%|█▋ | 2965/18018 [6:25:24<31:06:40, 7.44s/step, loss=1.777, acc=0.630, acc_len=1.47, thru=13.4, I=18.0, T=31.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:39,256] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 19:55:44,841] loop.py:45 INFO TRAIN_STEP step=2966 loss=1.753136 Training: 16%|█▋ | 2966/18018 [6:25:30<29:23:41, 7.03s/step, loss=1.753, acc=0.628, acc_len=1.46, thru=18.5, I=21.4, T=33.5, wait=2.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:55:44,836] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2965: forward=2646.3ms backward=1127.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:50,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:55:52,690] loop.py:45 INFO TRAIN_STEP step=2967 loss=1.792499 Training: 16%|█▋ | 2967/18018 [6:25:38<30:25:21, 7.28s/step, loss=1.792, acc=0.623, acc_len=1.45, thru=18.1, I=21.2, T=27.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:55:52,778] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 [2026-10-18 19:55:58,638] loop.py:45 INFO TRAIN_STEP step=2968 loss=1.715273 Training: 16%|█▋ | 2968/18018 [6:25:44<28:44:59, 6.88s/step, loss=1.715, acc=0.636, acc_len=1.49, thru=16.4, I=22.6, T=35.2, wait=2.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:08,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:56:10,161] loop.py:45 INFO TRAIN_STEP step=2969 loss=1.810408 Training: 16%|█▋ | 2969/18018 [6:25:56<34:34:43, 8.27s/step, loss=1.810, acc=0.622, acc_len=1.45, thru=13.7, I=15.4, T=27.2, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:10,233] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:15,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:56:17,598] loop.py:380 INFO TIMING step=2970: step=4.271s data=2.030s compute=4.234s [fwd=3.000s bwd=1.220s opt=0.003s] dispatch=3.153s [2026-10-18 19:56:17,599] loop.py:45 INFO TRAIN_STEP step=2970 loss=1.811649 Training: 16%|█▋ | 2970/18018 [6:26:03<33:31:43, 8.02s/step, loss=1.812, acc=0.622, acc_len=1.44, thru=14.9, I=22.7, T=30.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:17,714] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:23,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:56:24,893] loop.py:45 INFO TRAIN_STEP step=2971 loss=1.689056 Training: 16%|█▋ | 2971/18018 [6:26:10<32:36:49, 7.80s/step, loss=1.689, acc=0.636, acc_len=1.50, thru=16.0, I=20.9, T=29.3, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:56:24,888] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2970: forward=3017.8ms backward=1296.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:24,981] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:30,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:56:31,702] loop.py:45 INFO TRAIN_STEP step=2972 loss=1.794199 Training: 16%|█▋ | 2972/18018 [6:26:17<31:22:06, 7.51s/step, loss=1.794, acc=0.623, acc_len=1.45, thru=16.9, I=17.8, T=32.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:31,903] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:37,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:56:37,760] loop.py:45 INFO TRAIN_STEP step=2973 loss=1.683046 Training: 17%|█▋ | 2973/18018 [6:26:23<29:33:04, 7.07s/step, loss=1.683, acc=0.640, acc_len=1.50, thru=19.7, I=22.5, T=31.2, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:38,027] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:44,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:56:45,596] loop.py:45 INFO TRAIN_STEP step=2974 loss=1.994537 Training: 17%|█▋ | 2974/18018 [6:26:31<30:30:29, 7.30s/step, loss=1.995, acc=0.594, acc_len=1.36, thru=14.8, I=18.5, T=32.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:56:45,725] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 19:56:51,171] loop.py:380 INFO TIMING step=2975: step=3.664s data=1.609s compute=3.634s [fwd=2.573s bwd=1.046s opt=0.003s] dispatch=1.900s [2026-10-18 19:56:51,171] loop.py:45 INFO TRAIN_STEP step=2975 loss=1.737479 Training: 17%|█▋ | 2975/18018 [6:26:37<28:20:33, 6.78s/step, loss=1.737, acc=0.633, acc_len=1.48, thru=20.4, I=23.8, T=34.9, wait=1.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:00,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:57:00,505] loop.py:45 INFO TRAIN_STEP step=2976 loss=1.757106 Training: 17%|█▋ | 2976/18018 [6:26:46<31:32:16, 7.55s/step, loss=1.757, acc=0.632, acc_len=1.47, thru=13.9, I=14.2, T=30.7, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:57:00,500] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2975: forward=2956.9ms backward=1170.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:00,615] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:06,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:57:06,933] loop.py:45 INFO TRAIN_STEP step=2977 loss=1.784607 Training: 17%|█▋ | 2977/18018 [6:26:53<30:07:57, 7.21s/step, loss=1.785, acc=0.622, acc_len=1.45, thru=18.5, I=20.5, T=32.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:07,097] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:11,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:57:12,726] loop.py:45 INFO TRAIN_STEP step=2978 loss=1.830342 Training: 17%|█▋ | 2978/18018 [6:26:58<28:20:59, 6.79s/step, loss=1.830, acc=0.614, acc_len=1.42, thru=20.4, I=24.5, T=32.8, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:12,986] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:18,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:57:19,911] loop.py:45 INFO TRAIN_STEP step=2979 loss=1.747669 Training: 17%|█▋ | 2979/18018 [6:27:05<28:50:55, 6.91s/step, loss=1.748, acc=0.630, acc_len=1.47, thru=16.9, I=21.3, T=31.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:20,304] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:26,926] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:57:27,976] loop.py:380 INFO TIMING step=2980: step=4.042s data=1.835s compute=3.983s [fwd=2.785s bwd=1.182s opt=0.003s] dispatch=4.013s [2026-10-18 19:57:27,976] loop.py:45 INFO TRAIN_STEP step=2980 loss=1.881900 Training: 17%|█▋ | 2980/18018 [6:27:14<30:17:51, 7.25s/step, loss=1.882, acc=0.608, acc_len=1.41, thru=14.1, I=19.2, T=31.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:28,433] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:33,181] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:57:34,461] loop.py:45 INFO TRAIN_STEP step=2981 loss=1.769597 Training: 17%|█▋ | 2981/18018 [6:27:20<29:20:11, 7.02s/step, loss=1.770, acc=0.628, acc_len=1.46, thru=19.0, I=25.4, T=31.2, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:57:34,456] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2980: forward=2804.7ms backward=1239.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:34,691] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 19:57:42,266] loop.py:45 INFO TRAIN_STEP step=2982 loss=1.678764 Training: 17%|█▋ | 2982/18018 [6:27:28<30:18:48, 7.26s/step, loss=1.679, acc=0.637, acc_len=1.50, thru=14.1, I=15.0, T=32.3, wait=3.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:47,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:57:49,453] loop.py:45 INFO TRAIN_STEP step=2983 loss=1.763895 Training: 17%|█▋ | 2983/18018 [6:27:35<30:13:25, 7.24s/step, loss=1.764, acc=0.624, acc_len=1.46, thru=18.3, I=20.7, T=29.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:57:49,767] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 19:57:56,878] loop.py:45 INFO TRAIN_STEP step=2984 loss=1.744189 Training: 17%|█▋ | 2984/18018 [6:27:42<30:27:33, 7.29s/step, loss=1.744, acc=0.631, acc_len=1.47, thru=13.8, I=19.8, T=31.2, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:05,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:58:06,273] loop.py:380 INFO TIMING step=2985: step=4.259s data=1.865s compute=4.228s [fwd=3.076s bwd=1.137s opt=0.003s] dispatch=5.124s [2026-10-18 19:58:06,273] loop.py:45 INFO TRAIN_STEP step=2985 loss=1.894883 Training: 17%|█▋ | 2985/18018 [6:27:52<33:05:14, 7.92s/step, loss=1.895, acc=0.614, acc_len=1.42, thru=13.9, I=13.7, T=30.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:06,448] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:11,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:58:14,379] loop.py:45 INFO TRAIN_STEP step=2986 loss=1.774176 Training: 17%|█▋ | 2986/18018 [6:28:00<33:18:46, 7.98s/step, loss=1.774, acc=0.623, acc_len=1.45, thru=14.5, I=20.2, T=25.8, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:58:14,374] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2985: forward=3537.6ms backward=1372.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:14,852] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 19:58:22,857] loop.py:45 INFO TRAIN_STEP step=2987 loss=1.653390 Training: 17%|█▋ | 2987/18018 [6:28:08<33:56:05, 8.13s/step, loss=1.653, acc=0.641, acc_len=1.51, thru=12.0, I=13.5, T=29.8, wait=4.2s, pool=32, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 19:58:22,963] eagle_store.py:517 WARNING batch_get_buffer missing keys (k95591-957a07b1_hs, k95591-957a07b1_ids, k95591-957a07b1_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:31,919] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 19:58:22,958] eagle_store.py:517 WARNING batch_get_buffer missing keys (k95590-a10c100f_hs, k95590-a10c100f_ids, k95590-a10c100f_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-18 19:58:33,244] loop.py:45 INFO TRAIN_STEP step=2988 loss=1.813627 Training: 17%|█▋ | 2988/18018 [6:28:19<36:45:58, 8.81s/step, loss=1.814, acc=0.625, acc_len=1.45, thru=16.2, I=16.2, T=24.8, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:33,435] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:40,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:58:41,266] loop.py:45 INFO TRAIN_STEP step=2989 loss=1.817221 Training: 17%|█▋ | 2989/18018 [6:28:27<35:46:46, 8.57s/step, loss=1.817, acc=0.618, acc_len=1.43, thru=14.1, I=15.4, T=29.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:41,411] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-18 19:58:47,484] loop.py:380 INFO TIMING step=2990: step=3.811s data=1.755s compute=3.765s [fwd=2.697s bwd=1.052s opt=0.003s] dispatch=2.398s [2026-10-18 19:58:47,485] loop.py:45 INFO TRAIN_STEP step=2990 loss=1.609415 Training: 17%|█▋ | 2990/18018 [6:28:33<32:49:55, 7.87s/step, loss=1.609, acc=0.651, acc_len=1.54, thru=17.7, I=18.4, T=33.6, wait=2.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:53,688] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:58:55,813] loop.py:45 INFO TRAIN_STEP step=2991 loss=1.590320 Training: 17%|█▋ | 2991/18018 [6:28:41<33:24:29, 8.00s/step, loss=1.590, acc=0.661, acc_len=1.57, thru=17.8, I=19.5, T=28.3, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:58:55,807] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2990: forward=3136.6ms backward=1337.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:58:56,204] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:59:02,323] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:59:03,213] loop.py:45 INFO TRAIN_STEP step=2992 loss=1.828722 Training: 17%|█▋ | 2992/18018 [6:28:49<32:39:04, 7.82s/step, loss=1.829, acc=0.623, acc_len=1.45, thru=14.8, I=21.1, T=31.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:59:03,328] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:59:09,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:59:09,777] loop.py:45 INFO TRAIN_STEP step=2993 loss=1.771935 Training: 17%|█▋ | 2993/18018 [6:28:55<31:04:18, 7.44s/step, loss=1.772, acc=0.627, acc_len=1.46, thru=18.4, I=20.9, T=32.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:59:10,181] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:59:17,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:59:17,747] loop.py:45 INFO TRAIN_STEP step=2994 loss=1.816342 Training: 17%|█▋ | 2994/18018 [6:29:03<31:43:51, 7.60s/step, loss=1.816, acc=0.623, acc_len=1.44, thru=15.4, I=15.9, T=28.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:59:17,871] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-18 19:59:24,796] loop.py:380 INFO TIMING step=2995: step=4.012s data=1.795s compute=3.976s [fwd=2.773s bwd=1.188s opt=0.003s] dispatch=3.026s [2026-10-18 19:59:24,796] loop.py:45 INFO TRAIN_STEP step=2995 loss=1.551466 Training: 17%|█▋ | 2995/18018 [6:29:10<31:02:00, 7.44s/step, loss=1.551, acc=0.664, acc_len=1.58, thru=16.9, I=16.9, T=31.9, wait=3.0s, pool=56, epoch=1/2][2026-10-18 19:59:33,138] loop.py:45 INFO TRAIN_STEP step=2996 loss=1.841444 Training: 17%|█▋ | 2996/18018 [6:29:19<32:09:56, 7.71s/step, loss=1.841, acc=0.618, acc_len=1.42, thru=16.0, I=16.0, T=31.0, wait=4.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 19:59:33,133] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=2995: forward=2848.7ms backward=1207.2ms [2026-10-18 19:59:39,868] loop.py:45 INFO TRAIN_STEP step=2997 loss=1.719429 Training: 17%|█▋ | 2997/18018 [6:29:25<30:56:28, 7.42s/step, loss=1.719, acc=0.634, acc_len=1.49, thru=16.8, I=18.4, T=31.9, wait=2.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:59:46,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 19:59:48,662] loop.py:45 INFO TRAIN_STEP step=2998 loss=1.915278 Training: 17%|█▋ | 2998/18018 [6:29:34<32:39:50, 7.83s/step, loss=1.915, acc=0.607, acc_len=1.39, thru=15.8, I=17.9, T=26.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 19:59:49,094] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 19:59:55,982] loop.py:45 INFO TRAIN_STEP step=2999 loss=1.618888 Training: 17%|█▋ | 2999/18018 [6:29:42<32:01:22, 7.68s/step, loss=1.619, acc=0.651, acc_len=1.54, thru=12.5, I=20.0, T=31.1, wait=3.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:02,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:00:04,513] loop.py:380 INFO TIMING step=3000: step=4.749s data=2.223s compute=4.721s [fwd=3.216s bwd=1.490s opt=0.003s] dispatch=3.771s [2026-10-18 20:00:04,513] loop.py:45 INFO TRAIN_STEP step=3000 loss=1.641303 Training: 17%|█▋ | 3000/18018 [6:29:50<33:05:24, 7.93s/step, loss=1.641, acc=0.646, acc_len=1.52, thru=16.3, I=19.3, T=27.0, wait=3.8s, pool=64, epoch=1/2][2026-10-18 20:00:04,517] loop.py:419 INFO Saving checkpoint at step 3000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:07,741] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 20:00:07,981] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0003001 (TrainerActor pid=9310) [2026-10-18 20:00:07,990] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0002251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:09,251] inference_manager.py:408 INFO Pool capacity available after 6.6s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:15,352] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 20:00:07,981] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0003001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 20:00:09,012] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0002251 [2026-10-18 20:00:16,443] loop.py:45 INFO TRAIN_STEP step=3001 loss=1.797784 Training: 17%|█▋ | 3001/18018 [6:30:02<38:05:26, 9.13s/step, loss=1.798, acc=0.620, acc_len=1.44, thru=17.8, I=21.3, T=30.1, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:00:16,437] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3000: forward=3009.1ms backward=1196.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:16,864] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:21,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:00:22,526] loop.py:45 INFO TRAIN_STEP step=3002 loss=1.669391 Training: 17%|█▋ | 3002/18018 [6:30:08<34:16:30, 8.22s/step, loss=1.669, acc=0.640, acc_len=1.50, thru=19.4, I=22.9, T=32.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:22,563] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 20:00:28,812] loop.py:45 INFO TRAIN_STEP step=3003 loss=1.646835 Training: 17%|█▋ | 3003/18018 [6:30:14<31:51:24, 7.64s/step, loss=1.647, acc=0.648, acc_len=1.52, thru=17.0, I=21.0, T=31.4, wait=2.2s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:37,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:00:39,168] loop.py:45 INFO TRAIN_STEP step=3004 loss=1.587356 Training: 17%|█▋ | 3004/18018 [6:30:25<35:15:19, 8.45s/step, loss=1.587, acc=0.654, acc_len=1.54, thru=14.7, I=14.8, T=29.0, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:39,229] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:46,276] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:00:47,364] loop.py:380 INFO TIMING step=3005: step=4.354s data=2.018s compute=4.324s [fwd=3.031s bwd=1.278s opt=0.003s] dispatch=3.830s [2026-10-18 20:00:47,364] loop.py:45 INFO TRAIN_STEP step=3005 loss=1.890257 Training: 17%|█▋ | 3005/18018 [6:30:33<34:55:40, 8.38s/step, loss=1.890, acc=0.616, acc_len=1.43, thru=13.7, I=17.0, T=29.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:00:47,786] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:00:54,876] loop.py:45 INFO TRAIN_STEP step=3006 loss=1.854296 Training: 17%|█▋ | 3006/18018 [6:30:40<33:50:50, 8.12s/step, loss=1.854, acc=0.612, acc_len=1.42, thru=14.8, I=21.4, T=29.7, wait=3.2s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:00:54,871] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3005: forward=2900.7ms backward=1351.8ms [2026-10-18 20:01:06,066] loop.py:45 INFO TRAIN_STEP step=3007 loss=1.686574 Training: 17%|█▋ | 3007/18018 [6:30:52<37:41:14, 9.04s/step, loss=1.687, acc=0.642, acc_len=1.50, thru=12.7, I=13.0, T=28.0, wait=6.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:12,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:01:15,658] loop.py:45 INFO TRAIN_STEP step=3008 loss=1.790201 Training: 17%|█▋ | 3008/18018 [6:31:01<38:22:45, 9.20s/step, loss=1.790, acc=0.623, acc_len=1.45, thru=15.3, I=17.5, T=25.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:15,961] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:20,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:01:22,219] loop.py:45 INFO TRAIN_STEP step=3009 loss=1.708454 Training: 17%|█▋ | 3009/18018 [6:31:08<35:04:05, 8.41s/step, loss=1.708, acc=0.634, acc_len=1.49, thru=14.8, I=23.9, T=31.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:22,292] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 20:01:32,156] loop.py:380 INFO TIMING step=3010: step=4.712s data=2.051s compute=4.661s [fwd=3.162s bwd=1.483s opt=0.003s] dispatch=5.216s [2026-10-18 20:01:32,156] loop.py:45 INFO TRAIN_STEP step=3010 loss=1.674464 Training: 17%|█▋ | 3010/18018 [6:31:18<36:58:24, 8.87s/step, loss=1.674, acc=0.642, acc_len=1.51, thru=13.0, I=12.8, T=27.2, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:32,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:32,629] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:01:40,206] loop.py:45 INFO TRAIN_STEP step=3011 loss=1.675151 Training: 17%|█▋ | 3011/18018 [6:31:26<35:56:54, 8.62s/step, loss=1.675, acc=0.646, acc_len=1.51, thru=14.4, I=15.8, T=26.5, wait=3.2s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:01:40,201] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3010: forward=3439.3ms backward=1345.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:46,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:01:48,053] loop.py:45 INFO TRAIN_STEP step=3012 loss=1.616019 Training: 17%|█▋ | 3012/18018 [6:31:34<34:58:33, 8.39s/step, loss=1.616, acc=0.653, acc_len=1.54, thru=15.6, I=16.9, T=25.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:48,447] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:54,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:01:54,642] loop.py:45 INFO TRAIN_STEP step=3013 loss=1.673123 Training: 17%|█▋ | 3013/18018 [6:31:40<32:43:00, 7.85s/step, loss=1.673, acc=0.640, acc_len=1.50, thru=17.3, I=21.3, T=31.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:01:55,051] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:00,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:02:01,088] loop.py:45 INFO TRAIN_STEP step=3014 loss=1.760177 Training: 17%|█▋ | 3014/18018 [6:31:47<30:57:49, 7.43s/step, loss=1.760, acc=0.625, acc_len=1.45, thru=18.7, I=20.1, T=34.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:01,338] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:02:07,726] loop.py:380 INFO TIMING step=3015: step=4.054s data=1.856s compute=4.022s [fwd=2.855s bwd=1.152s opt=0.003s] dispatch=2.574s [2026-10-18 20:02:07,726] loop.py:45 INFO TRAIN_STEP step=3015 loss=1.595942 Training: 17%|█▋ | 3015/18018 [6:31:53<29:58:08, 7.19s/step, loss=1.596, acc=0.651, acc_len=1.54, thru=17.0, I=19.0, T=31.6, wait=2.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:15,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:02:16,949] loop.py:45 INFO TRAIN_STEP step=3016 loss=1.696783 Training: 17%|█▋ | 3016/18018 [6:32:03<32:30:22, 7.80s/step, loss=1.697, acc=0.639, acc_len=1.50, thru=15.7, I=17.2, T=28.0, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:02:16,944] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3015: forward=3317.5ms backward=1213.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:17,434] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:24,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:02:25,504] loop.py:45 INFO TRAIN_STEP step=3017 loss=1.626752 Training: 17%|█▋ | 3017/18018 [6:32:11<33:27:06, 8.03s/step, loss=1.627, acc=0.649, acc_len=1.53, thru=14.1, I=17.7, T=30.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:25,539] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:30,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:02:31,700] loop.py:45 INFO TRAIN_STEP step=3018 loss=1.587895 Training: 17%|█▋ | 3018/18018 [6:32:17<31:09:23, 7.48s/step, loss=1.588, acc=0.656, acc_len=1.55, thru=18.9, I=26.0, T=31.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:31,804] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 20:02:37,426] loop.py:45 INFO TRAIN_STEP step=3019 loss=1.679705 Training: 17%|█▋ | 3019/18018 [6:32:23<28:58:06, 6.95s/step, loss=1.680, acc=0.638, acc_len=1.50, thru=18.4, I=24.5, T=32.1, wait=1.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:46,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:02:46,941] loop.py:380 INFO TIMING step=3020: step=4.255s data=1.938s compute=4.218s [fwd=3.039s bwd=1.164s opt=0.003s] dispatch=5.250s [2026-10-18 20:02:46,941] loop.py:45 INFO TRAIN_STEP step=3020 loss=1.876062 Training: 17%|█▋ | 3020/18018 [6:32:33<32:10:01, 7.72s/step, loss=1.876, acc=0.610, acc_len=1.41, thru=15.1, I=14.5, T=30.1, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:47,418] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:54,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:02:55,556] loop.py:45 INFO TRAIN_STEP step=3021 loss=1.793261 Training: 17%|█▋ | 3021/18018 [6:32:41<33:16:58, 7.99s/step, loss=1.793, acc=0.625, acc_len=1.45, thru=14.7, I=18.4, T=28.1, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:02:55,550] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3020: forward=3144.5ms backward=1342.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:02:55,684] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 20:03:01,617] loop.py:45 INFO TRAIN_STEP step=3022 loss=1.723063 Training: 17%|█▋ | 3022/18018 [6:32:47<30:52:09, 7.41s/step, loss=1.723, acc=0.633, acc_len=1.48, thru=15.2, I=18.9, T=36.4, wait=2.5s, pool=40, epoch=1/2][2026-10-18 20:03:11,988] loop.py:45 INFO TRAIN_STEP step=3023 loss=1.838248 Training: 17%|█▋ | 3023/18018 [6:32:58<34:34:01, 8.30s/step, loss=1.838, acc=0.621, acc_len=1.44, thru=15.2, I=13.2, T=27.0, wait=5.6s, pool=60, epoch=1/2][2026-10-18 20:03:23,066] loop.py:45 INFO TRAIN_STEP step=3024 loss=1.811822 Training: 17%|█▋ | 3024/18018 [6:33:09<38:02:21, 9.13s/step, loss=1.812, acc=0.618, acc_len=1.43, thru=8.4, I=9.7, T=22.7, wait=5.4s, pool=28, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:03:31,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:03:33,679] loop.py:380 INFO TIMING step=3025: step=4.244s data=1.881s compute=4.203s [fwd=2.966s bwd=1.223s opt=0.003s] dispatch=6.357s [2026-10-18 20:03:33,680] loop.py:45 INFO TRAIN_STEP step=3025 loss=1.691860 Training: 17%|█▋ | 3025/18018 [6:33:19<39:53:10, 9.58s/step, loss=1.692, acc=0.642, acc_len=1.50, thru=15.9, I=18.1, T=30.2, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:03:33,734] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:03:39,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:03:40,803] loop.py:45 INFO TRAIN_STEP step=3026 loss=1.729270 Training: 17%|█▋ | 3026/18018 [6:33:26<36:48:53, 8.84s/step, loss=1.729, acc=0.637, acc_len=1.49, thru=15.8, I=25.2, T=31.4, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:03:40,799] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3025: forward=2842.9ms backward=1190.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:03:41,201] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:03:48,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:03:48,731] loop.py:45 INFO TRAIN_STEP step=3027 loss=1.762168 Training: 17%|█▋ | 3027/18018 [6:33:34<35:40:30, 8.57s/step, loss=1.762, acc=0.630, acc_len=1.46, thru=14.5, I=19.5, T=32.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:03:48,788] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 20:03:55,950] loop.py:45 INFO TRAIN_STEP step=3028 loss=1.769168 Training: 17%|█▋ | 3028/18018 [6:33:42<33:59:19, 8.16s/step, loss=1.769, acc=0.621, acc_len=1.44, thru=17.4, I=18.0, T=31.4, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:02,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:04:04,553] loop.py:45 INFO TRAIN_STEP step=3029 loss=1.669910 Training: 17%|█▋ | 3029/18018 [6:33:50<34:32:08, 8.29s/step, loss=1.670, acc=0.641, acc_len=1.51, thru=16.0, I=16.9, T=26.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:04,918] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:11,114] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:04:11,872] loop.py:380 INFO TIMING step=3030: step=4.281s data=1.970s compute=4.253s [fwd=2.995s bwd=1.244s opt=0.003s] dispatch=3.028s [2026-10-18 20:04:11,872] loop.py:45 INFO TRAIN_STEP step=3030 loss=1.716564 Training: 17%|█▋ | 3030/18018 [6:33:57<33:18:54, 8.00s/step, loss=1.717, acc=0.635, acc_len=1.49, thru=15.3, I=20.3, T=29.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:12,122] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:19,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:04:19,801] loop.py:45 INFO TRAIN_STEP step=3031 loss=1.717935 Training: 17%|█▋ | 3031/18018 [6:34:05<33:13:15, 7.98s/step, loss=1.718, acc=0.633, acc_len=1.49, thru=15.4, I=15.8, T=30.4, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:04:19,796] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3030: forward=3007.3ms backward=1164.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:19,948] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:25,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:04:26,633] loop.py:45 INFO TRAIN_STEP step=3032 loss=1.619225 Training: 17%|█▋ | 3032/18018 [6:34:12<31:47:11, 7.64s/step, loss=1.619, acc=0.655, acc_len=1.55, thru=18.4, I=19.3, T=28.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:26,697] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:33,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:04:35,421] loop.py:45 INFO TRAIN_STEP step=3033 loss=1.802690 Training: 17%|█▋ | 3033/18018 [6:34:21<33:13:22, 7.98s/step, loss=1.803, acc=0.620, acc_len=1.44, thru=13.7, I=16.6, T=28.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:35,792] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:40,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:04:41,629] loop.py:45 INFO TRAIN_STEP step=3034 loss=1.810880 Training: 17%|█▋ | 3034/18018 [6:34:27<31:00:20, 7.45s/step, loss=1.811, acc=0.619, acc_len=1.43, thru=18.6, I=24.1, T=35.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:41,992] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:48,403] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:04:49,347] loop.py:380 INFO TIMING step=3035: step=4.283s data=2.066s compute=4.232s [fwd=2.925s bwd=1.292s opt=0.003s] dispatch=3.425s [2026-10-18 20:04:49,347] loop.py:45 INFO TRAIN_STEP step=3035 loss=1.784525 Training: 17%|█▋ | 3035/18018 [6:34:35<31:20:13, 7.53s/step, loss=1.785, acc=0.623, acc_len=1.44, thru=16.1, I=18.1, T=29.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:04:49,416] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 20:04:55,868] loop.py:45 INFO TRAIN_STEP step=3036 loss=1.761821 Training: 17%|█▋ | 3036/18018 [6:34:41<30:04:38, 7.23s/step, loss=1.762, acc=0.626, acc_len=1.46, thru=14.1, I=17.3, T=32.3, wait=2.6s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:04:55,863] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3035: forward=2798.3ms backward=1106.4ms (TrainerActor pid=9310) [2026-10-18 20:04:55,934] eagle_store.py:517 WARNING batch_get_buffer missing keys (k97152-931ad68a_hs, k97152-931ad68a_ids, k97152-931ad68a_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:05:02,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9371) [2026-10-18 20:04:55,937] eagle_store.py:517 WARNING batch_get_buffer missing keys (k97154-bf63153a_hs, k97154-bf63153a_ids, k97154-bf63153a_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-18 20:05:03,929] loop.py:45 INFO TRAIN_STEP step=3037 loss=1.684774 Training: 17%|█▋ | 3037/18018 [6:34:50<31:07:03, 7.48s/step, loss=1.685, acc=0.637, acc_len=1.49, thru=17.6, I=17.5, T=29.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:05:04,314] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:05:11,894] loop.py:45 INFO TRAIN_STEP step=3038 loss=1.759653 Training: 17%|█▋ | 3038/18018 [6:34:57<31:43:19, 7.62s/step, loss=1.760, acc=0.631, acc_len=1.48, thru=14.1, I=15.1, T=30.1, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:05:19,314] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:05:19,893] loop.py:45 INFO TRAIN_STEP step=3039 loss=1.773455 Training: 17%|█▋ | 3039/18018 [6:35:05<32:11:15, 7.74s/step, loss=1.773, acc=0.630, acc_len=1.47, thru=17.4, I=16.2, T=27.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:05:20,317] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:05:25,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:05:27,190] loop.py:380 INFO TIMING step=3040: step=4.759s data=2.426s compute=4.731s [fwd=3.210s bwd=1.506s opt=0.003s] dispatch=2.529s [2026-10-18 20:05:27,190] loop.py:45 INFO TRAIN_STEP step=3040 loss=1.788472 Training: 17%|█▋ | 3040/18018 [6:35:13<31:38:17, 7.60s/step, loss=1.788, acc=0.624, acc_len=1.44, thru=17.5, I=20.2, T=26.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:05:27,348] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-18 20:05:35,046] loop.py:45 INFO TRAIN_STEP step=3041 loss=1.724423 Training: 17%|█▋ | 3041/18018 [6:35:21<31:57:02, 7.68s/step, loss=1.724, acc=0.635, acc_len=1.49, thru=12.6, I=15.9, T=31.6, wait=3.8s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:05:35,041] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3040: forward=2828.5ms backward=1180.9ms [2026-10-18 20:05:44,529] loop.py:45 INFO TRAIN_STEP step=3042 loss=1.811535 Training: 17%|█▋ | 3042/18018 [6:35:30<34:12:00, 8.22s/step, loss=1.812, acc=0.621, acc_len=1.44, thru=14.4, I=14.5, T=31.6, wait=5.4s, pool=56, epoch=1/2][2026-10-18 20:05:52,406] loop.py:45 INFO TRAIN_STEP step=3043 loss=1.617159 Training: 17%|█▋ | 3043/18018 [6:35:38<33:46:24, 8.12s/step, loss=1.617, acc=0.653, acc_len=1.54, thru=16.1, I=16.5, T=33.0, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:00,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:06:01,210] loop.py:45 INFO TRAIN_STEP step=3044 loss=2.011602 Training: 17%|█▋ | 3044/18018 [6:35:47<34:37:05, 8.32s/step, loss=2.012, acc=0.588, acc_len=1.34, thru=14.2, I=15.7, T=31.2, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:01,522] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:06,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:06:07,796] loop.py:380 INFO TIMING step=3045: step=3.974s data=1.783s compute=3.939s [fwd=2.804s bwd=1.120s opt=0.003s] dispatch=2.603s [2026-10-18 20:06:07,796] loop.py:45 INFO TRAIN_STEP step=3045 loss=1.692700 Training: 17%|█▋ | 3045/18018 [6:35:53<32:26:55, 7.80s/step, loss=1.693, acc=0.638, acc_len=1.49, thru=17.4, I=21.6, T=32.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:08,184] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:06:14,869] loop.py:45 INFO TRAIN_STEP step=3046 loss=1.793687 Training: 17%|█▋ | 3046/18018 [6:36:00<31:32:16, 7.58s/step, loss=1.794, acc=0.623, acc_len=1.44, thru=16.5, I=21.7, T=32.5, wait=3.1s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:06:14,864] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3045: forward=2776.6ms backward=1121.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:23,538] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:06:24,294] loop.py:45 INFO TRAIN_STEP step=3047 loss=1.746810 Training: 17%|█▋ | 3047/18018 [6:36:10<33:50:03, 8.14s/step, loss=1.747, acc=0.636, acc_len=1.48, thru=14.0, I=14.0, T=31.1, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:24,546] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:06:32,673] loop.py:45 INFO TRAIN_STEP step=3048 loss=1.747859 Training: 17%|█▋ | 3048/18018 [6:36:18<34:07:58, 8.21s/step, loss=1.748, acc=0.632, acc_len=1.47, thru=14.1, I=14.5, T=30.4, wait=4.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:38,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:06:40,238] loop.py:45 INFO TRAIN_STEP step=3049 loss=1.853283 Training: 17%|█▋ | 3049/18018 [6:36:26<33:19:53, 8.02s/step, loss=1.853, acc=0.612, acc_len=1.42, thru=17.5, I=18.2, T=31.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:40,404] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:46,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:06:47,567] loop.py:380 INFO TIMING step=3050: step=4.085s data=1.820s compute=4.059s [fwd=2.895s bwd=1.150s opt=0.003s] dispatch=3.234s [2026-10-18 20:06:47,567] loop.py:45 INFO TRAIN_STEP step=3050 loss=1.724752 Training: 17%|█▋ | 3050/18018 [6:36:33<32:28:24, 7.81s/step, loss=1.725, acc=0.633, acc_len=1.48, thru=15.3, I=20.8, T=31.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:47,611] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:53,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:06:53,939] loop.py:45 INFO TRAIN_STEP step=3051 loss=1.707523 Training: 17%|█▋ | 3051/18018 [6:36:40<30:40:26, 7.38s/step, loss=1.708, acc=0.635, acc_len=1.49, thru=19.4, I=21.3, T=34.2, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:06:53,934] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3050: forward=2639.2ms backward=1057.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:06:54,354] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:07:03,857] loop.py:45 INFO TRAIN_STEP step=3052 loss=1.801757 Training: 17%|█▋ | 3052/18018 [6:36:49<33:50:21, 8.14s/step, loss=1.802, acc=0.621, acc_len=1.44, thru=13.1, I=12.7, T=28.7, wait=5.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:15,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:07:16,872] loop.py:45 INFO TRAIN_STEP step=3053 loss=1.812238 Training: 17%|█▋ | 3053/18018 [6:37:02<39:55:01, 9.60s/step, loss=1.812, acc=0.620, acc_len=1.44, thru=11.3, I=11.5, T=30.6, wait=8.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:17,235] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:22,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:07:23,567] loop.py:45 INFO TRAIN_STEP step=3054 loss=1.672487 Training: 17%|█▋ | 3054/18018 [6:37:09<36:17:28, 8.73s/step, loss=1.672, acc=0.642, acc_len=1.51, thru=17.6, I=23.3, T=31.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:23,643] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:31,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:07:32,589] loop.py:380 INFO TIMING step=3055: step=4.852s data=2.288s compute=4.814s [fwd=3.283s bwd=1.517s opt=0.003s] dispatch=4.159s [2026-10-18 20:07:32,589] loop.py:45 INFO TRAIN_STEP step=3055 loss=1.789599 Training: 17%|█▋ | 3055/18018 [6:37:18<36:39:04, 8.82s/step, loss=1.790, acc=0.623, acc_len=1.45, thru=13.2, I=17.0, T=26.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:33,076] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:39,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:07:39,765] loop.py:45 INFO TRAIN_STEP step=3056 loss=1.744984 Training: 17%|█▋ | 3056/18018 [6:37:25<34:36:09, 8.33s/step, loss=1.745, acc=0.629, acc_len=1.46, thru=16.4, I=20.9, T=32.8, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:07:39,760] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3055: forward=2765.3ms backward=1098.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:40,131] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:47,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:07:48,156] loop.py:45 INFO TRAIN_STEP step=3057 loss=1.784111 Training: 17%|█▋ | 3057/18018 [6:37:34<34:40:46, 8.34s/step, loss=1.784, acc=0.627, acc_len=1.45, thru=15.6, I=16.7, T=27.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:48,339] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:53,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:07:55,219] loop.py:45 INFO TRAIN_STEP step=3058 loss=1.796055 Training: 17%|█▋ | 3058/18018 [6:37:41<33:04:45, 7.96s/step, loss=1.796, acc=0.627, acc_len=1.46, thru=17.9, I=20.6, T=28.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:07:55,286] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:01,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:08:02,623] loop.py:45 INFO TRAIN_STEP step=3059 loss=1.871101 Training: 17%|█▋ | 3059/18018 [6:37:48<32:23:06, 7.79s/step, loss=1.871, acc=0.609, acc_len=1.40, thru=14.6, I=20.3, T=28.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:02,661] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:06,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:08:10,245] loop.py:380 INFO TIMING step=3060: step=5.789s data=3.313s compute=5.757s [fwd=4.339s bwd=1.402s opt=0.003s] dispatch=1.823s [2026-10-18 20:08:10,246] loop.py:45 INFO TRAIN_STEP step=3060 loss=1.644020 Training: 17%|█▋ | 3060/18018 [6:37:56<32:10:02, 7.74s/step, loss=1.644, acc=0.648, acc_len=1.52, thru=14.5, I=24.7, T=22.1, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:10,620] inference_manager.py:408 INFO Pool capacity available after 4.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:18,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:08:18,705] loop.py:45 INFO TRAIN_STEP step=3061 loss=1.797234 Training: 17%|█▋ | 3061/18018 [6:38:04<33:03:36, 7.96s/step, loss=1.797, acc=0.624, acc_len=1.45, thru=15.8, I=15.4, T=32.1, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:08:18,699] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3060: forward=2758.7ms backward=1169.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:18,896] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:24,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:08:25,295] loop.py:45 INFO TRAIN_STEP step=3062 loss=1.669574 Training: 17%|█▋ | 3062/18018 [6:38:11<31:21:07, 7.55s/step, loss=1.670, acc=0.640, acc_len=1.50, thru=19.3, I=19.3, T=30.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:25,425] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:31,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:08:32,452] loop.py:45 INFO TRAIN_STEP step=3063 loss=1.728321 Training: 17%|█▋ | 3063/18018 [6:38:18<30:52:03, 7.43s/step, loss=1.728, acc=0.634, acc_len=1.49, thru=17.3, I=19.4, T=30.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:32,697] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:39,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:08:40,049] loop.py:45 INFO TRAIN_STEP step=3064 loss=1.739388 Training: 17%|█▋ | 3064/18018 [6:38:26<31:04:25, 7.48s/step, loss=1.739, acc=0.635, acc_len=1.49, thru=14.6, I=17.2, T=25.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:40,185] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-18 20:08:46,599] loop.py:380 INFO TIMING step=3065: step=4.194s data=1.975s compute=4.167s [fwd=2.992s bwd=1.161s opt=0.003s] dispatch=2.346s [2026-10-18 20:08:46,599] loop.py:45 INFO TRAIN_STEP step=3065 loss=1.716016 Training: 17%|█▋ | 3065/18018 [6:38:32<29:54:28, 7.20s/step, loss=1.716, acc=0.633, acc_len=1.48, thru=17.5, I=18.0, T=30.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:46,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:47,098] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:51,834] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:08:53,126] loop.py:45 INFO TRAIN_STEP step=3066 loss=1.733074 Training: 17%|█▋ | 3066/18018 [6:38:39<29:04:12, 7.00s/step, loss=1.733, acc=0.630, acc_len=1.47, thru=18.5, I=20.7, T=28.4, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:08:53,121] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3065: forward=3179.4ms backward=1268.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:08:53,342] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:08:58,910] loop.py:45 INFO TRAIN_STEP step=3067 loss=1.666242 Training: 17%|█▋ | 3067/18018 [6:38:44<27:33:15, 6.63s/step, loss=1.666, acc=0.639, acc_len=1.50, thru=16.9, I=20.1, T=34.8, wait=2.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:07,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:09:07,459] loop.py:45 INFO TRAIN_STEP step=3068 loss=2.004794 Training: 17%|█▋ | 3068/18018 [6:38:53<29:56:14, 7.21s/step, loss=2.005, acc=0.587, acc_len=1.34, thru=16.9, I=16.3, T=31.6, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:07,630] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:12,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:09:14,819] loop.py:45 INFO TRAIN_STEP step=3069 loss=1.866010 Training: 17%|█▋ | 3069/18018 [6:39:00<30:07:24, 7.25s/step, loss=1.866, acc=0.614, acc_len=1.41, thru=18.4, I=21.3, T=32.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:14,980] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:19,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:09:21,409] loop.py:380 INFO TIMING step=3070: step=3.952s data=1.720s compute=3.916s [fwd=2.739s bwd=1.161s opt=0.003s] dispatch=2.627s [2026-10-18 20:09:21,409] loop.py:45 INFO TRAIN_STEP step=3070 loss=1.583072 Training: 17%|█▋ | 3070/18018 [6:39:07<29:17:39, 7.06s/step, loss=1.583, acc=0.657, acc_len=1.56, thru=18.7, I=29.6, T=32.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:21,737] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:26,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:09:28,536] loop.py:45 INFO TRAIN_STEP step=3071 loss=1.613756 Training: 17%|█▋ | 3071/18018 [6:39:14<29:22:47, 7.08s/step, loss=1.614, acc=0.652, acc_len=1.54, thru=14.7, I=25.0, T=30.9, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:09:28,531] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3070: forward=2853.6ms backward=1226.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:28,963] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:36,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:09:38,913] loop.py:45 INFO TRAIN_STEP step=3072 loss=1.696435 Training: 17%|█▋ | 3072/18018 [6:39:24<33:29:22, 8.07s/step, loss=1.696, acc=0.638, acc_len=1.50, thru=12.9, I=17.9, T=26.6, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:39,207] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:45,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:09:46,908] loop.py:45 INFO TRAIN_STEP step=3073 loss=1.598587 Training: 17%|█▋ | 3073/18018 [6:39:32<33:23:51, 8.04s/step, loss=1.599, acc=0.655, acc_len=1.55, thru=16.6, I=18.4, T=28.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:46,954] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:54,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:09:55,895] loop.py:45 INFO TRAIN_STEP step=3074 loss=1.690907 Training: 17%|█▋ | 3074/18018 [6:39:41<34:34:04, 8.33s/step, loss=1.691, acc=0.638, acc_len=1.49, thru=13.3, I=16.4, T=29.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:09:56,306] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:01,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:10:02,397] loop.py:380 INFO TIMING step=3075: step=3.598s data=1.609s compute=3.558s [fwd=2.511s bwd=1.032s opt=0.003s] dispatch=2.895s [2026-10-18 20:10:02,397] loop.py:45 INFO TRAIN_STEP step=3075 loss=1.839861 Training: 17%|█▋ | 3075/18018 [6:39:48<32:17:42, 7.78s/step, loss=1.840, acc=0.620, acc_len=1.43, thru=17.9, I=23.6, T=35.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:02,449] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:09,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:10:10,678] loop.py:45 INFO TRAIN_STEP step=3076 loss=1.922589 Training: 17%|█▋ | 3076/18018 [6:39:56<32:54:50, 7.93s/step, loss=1.923, acc=0.604, acc_len=1.39, thru=15.4, I=18.0, T=31.6, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:10:10,673] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3075: forward=2816.6ms backward=1183.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:11,161] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 20:10:20,337] loop.py:45 INFO TRAIN_STEP step=3077 loss=1.749649 Training: 17%|█▋ | 3077/18018 [6:40:06<35:03:56, 8.45s/step, loss=1.750, acc=0.630, acc_len=1.46, thru=10.5, I=10.2, T=27.6, wait=5.0s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:27,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:10:29,094] loop.py:45 INFO TRAIN_STEP step=3078 loss=1.704405 Training: 17%|█▋ | 3078/18018 [6:40:15<35:26:52, 8.54s/step, loss=1.704, acc=0.634, acc_len=1.49, thru=18.2, I=18.3, T=27.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:29,381] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:35,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:10:36,135] loop.py:45 INFO TRAIN_STEP step=3079 loss=1.623349 Training: 17%|█▋ | 3079/18018 [6:40:22<33:34:36, 8.09s/step, loss=1.623, acc=0.653, acc_len=1.54, thru=16.1, I=22.4, T=28.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:36,254] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:44,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:10:44,844] loop.py:380 INFO TIMING step=3080: step=4.099s data=1.923s compute=4.065s [fwd=2.888s bwd=1.163s opt=0.003s] dispatch=4.599s [2026-10-18 20:10:44,844] loop.py:45 INFO TRAIN_STEP step=3080 loss=1.835794 Training: 17%|█▋ | 3080/18018 [6:40:30<34:20:30, 8.28s/step, loss=1.836, acc=0.616, acc_len=1.43, thru=14.2, I=15.3, T=31.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:45,282] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:49,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:10:50,975] loop.py:45 INFO TRAIN_STEP step=3081 loss=1.743404 Training: 17%|█▋ | 3081/18018 [6:40:37<31:40:19, 7.63s/step, loss=1.743, acc=0.632, acc_len=1.47, thru=19.2, I=21.1, T=30.8, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:10:50,970] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3080: forward=2881.1ms backward=1230.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:51,140] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:57,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:10:58,226] loop.py:45 INFO TRAIN_STEP step=3082 loss=1.878598 Training: 17%|█▋ | 3082/18018 [6:40:44<31:11:38, 7.52s/step, loss=1.879, acc=0.608, acc_len=1.40, thru=15.6, I=20.2, T=31.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:10:58,666] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:05,928] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:11:06,441] loop.py:45 INFO TRAIN_STEP step=3083 loss=1.682876 Training: 17%|█▋ | 3083/18018 [6:40:52<32:03:21, 7.73s/step, loss=1.683, acc=0.642, acc_len=1.51, thru=14.5, I=17.3, T=30.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:06,932] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:14,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:11:15,514] loop.py:45 INFO TRAIN_STEP step=3084 loss=1.870357 Training: 17%|█▋ | 3084/18018 [6:41:01<33:43:52, 8.13s/step, loss=1.870, acc=0.607, acc_len=1.41, thru=14.0, I=18.1, T=30.0, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:15,628] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:21,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:11:21,831] loop.py:380 INFO TIMING step=3085: step=4.186s data=2.015s compute=4.147s [fwd=2.939s bwd=1.194s opt=0.003s] dispatch=2.119s [2026-10-18 20:11:21,831] loop.py:45 INFO TRAIN_STEP step=3085 loss=1.742056 Training: 17%|█▋ | 3085/18018 [6:41:07<31:28:07, 7.59s/step, loss=1.742, acc=0.632, acc_len=1.48, thru=19.0, I=21.7, T=30.6, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:21,997] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:11:29,571] loop.py:45 INFO TRAIN_STEP step=3086 loss=1.887151 Training: 17%|█▋ | 3086/18018 [6:41:15<31:39:36, 7.63s/step, loss=1.887, acc=0.608, acc_len=1.40, thru=13.6, I=14.3, T=29.8, wait=3.4s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:11:29,566] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3085: forward=2986.4ms backward=1239.6ms [2026-10-18 20:11:37,823] loop.py:45 INFO TRAIN_STEP step=3087 loss=1.693098 Training: 17%|█▋ | 3087/18018 [6:41:23<32:25:35, 7.82s/step, loss=1.693, acc=0.635, acc_len=1.48, thru=16.7, I=15.1, T=30.7, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:43,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:11:45,460] loop.py:45 INFO TRAIN_STEP step=3088 loss=1.595241 Training: 17%|█▋ | 3088/18018 [6:41:31<32:12:02, 7.76s/step, loss=1.595, acc=0.655, acc_len=1.55, thru=17.4, I=18.8, T=27.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:45,814] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:50,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:11:52,621] loop.py:45 INFO TRAIN_STEP step=3089 loss=1.623298 Training: 17%|█▋ | 3089/18018 [6:41:38<31:26:50, 7.58s/step, loss=1.623, acc=0.649, acc_len=1.53, thru=15.2, I=23.1, T=27.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:11:53,061] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 20:12:00,414] loop.py:380 INFO TIMING step=3090: step=4.687s data=2.172s compute=4.653s [fwd=3.196s bwd=1.442s opt=0.003s] dispatch=3.096s [2026-10-18 20:12:00,415] loop.py:45 INFO TRAIN_STEP step=3090 loss=1.745136 Training: 17%|█▋ | 3090/18018 [6:41:46<31:42:25, 7.65s/step, loss=1.745, acc=0.631, acc_len=1.47, thru=13.6, I=17.1, T=27.3, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:12:06,188] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:12:07,103] loop.py:45 INFO TRAIN_STEP step=3091 loss=1.656647 Training: 17%|█▋ | 3091/18018 [6:41:53<30:30:47, 7.36s/step, loss=1.657, acc=0.642, acc_len=1.51, thru=17.8, I=18.2, T=31.9, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:12:07,097] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3090: forward=2835.0ms backward=1127.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:12:07,205] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:12:11,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:12:13,415] loop.py:45 INFO TRAIN_STEP step=3092 loss=1.700300 Training: 17%|█▋ | 3092/18018 [6:41:59<29:12:25, 7.04s/step, loss=1.700, acc=0.641, acc_len=1.50, thru=19.1, I=24.1, T=33.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:12:13,871] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:12:19,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:12:20,336] loop.py:45 INFO TRAIN_STEP step=3093 loss=1.769924 Training: 17%|█▋ | 3093/18018 [6:42:06<29:03:07, 7.01s/step, loss=1.770, acc=0.624, acc_len=1.45, thru=18.2, I=22.9, T=32.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:12:20,499] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:12:28,729] loop.py:45 INFO TRAIN_STEP step=3094 loss=1.610573 Training: 17%|█▋ | 3094/18018 [6:42:14<30:46:26, 7.42s/step, loss=1.611, acc=0.650, acc_len=1.53, thru=13.9, I=13.9, T=32.6, wait=4.5s, pool=52, epoch=1/2][2026-10-18 20:12:35,562] loop.py:380 INFO TIMING step=3095: step=3.712s data=1.580s compute=3.681s [fwd=2.615s bwd=1.052s opt=0.003s] dispatch=3.111s [2026-10-18 20:12:35,562] loop.py:45 INFO TRAIN_STEP step=3095 loss=1.685505 Training: 17%|█▋ | 3095/18018 [6:42:21<30:02:18, 7.25s/step, loss=1.686, acc=0.640, acc_len=1.50, thru=18.6, I=18.6, T=34.5, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:12:45,033] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:12:45,535] loop.py:45 INFO TRAIN_STEP step=3096 loss=1.730454 Training: 17%|█▋ | 3096/18018 [6:42:31<33:25:27, 8.06s/step, loss=1.730, acc=0.634, acc_len=1.48, thru=15.1, I=13.4, T=30.5, wait=5.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:12:45,529] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3095: forward=2948.3ms backward=1191.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:12:46,037] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:12:52,241] loop.py:45 INFO TRAIN_STEP step=3097 loss=1.609674 Training: 17%|█▋ | 3097/18018 [6:42:38<31:44:12, 7.66s/step, loss=1.610, acc=0.652, acc_len=1.54, thru=16.4, I=20.0, T=32.8, wait=2.8s, pool=48, epoch=1/2][2026-10-18 20:13:00,761] loop.py:45 INFO TRAIN_STEP step=3098 loss=1.742434 Training: 17%|█▋ | 3098/18018 [6:42:46<32:48:28, 7.92s/step, loss=1.742, acc=0.633, acc_len=1.48, thru=11.6, I=13.8, T=31.9, wait=4.5s, pool=24, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:08,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:13:09,687] loop.py:45 INFO TRAIN_STEP step=3099 loss=1.909847 Training: 17%|█▋ | 3099/18018 [6:42:55<34:03:37, 8.22s/step, loss=1.910, acc=0.605, acc_len=1.39, thru=18.4, I=16.5, T=34.5, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:10,010] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:13:16,077] loop.py:380 INFO TIMING step=3100: step=3.682s data=1.582s compute=3.654s [fwd=2.617s bwd=1.022s opt=0.003s] dispatch=2.697s [2026-10-18 20:13:16,077] loop.py:45 INFO TRAIN_STEP step=3100 loss=1.742611 Training: 17%|█▋ | 3100/18018 [6:43:02<31:47:04, 7.67s/step, loss=1.743, acc=0.631, acc_len=1.47, thru=18.0, I=25.7, T=34.8, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:21,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:13:22,234] loop.py:45 INFO TRAIN_STEP step=3101 loss=1.769022 Training: 17%|█▋ | 3101/18018 [6:43:08<29:53:56, 7.22s/step, loss=1.769, acc=0.628, acc_len=1.46, thru=20.5, I=21.4, T=33.6, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:13:22,230] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3100: forward=2679.3ms backward=1084.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:22,613] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:27,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:13:28,501] loop.py:45 INFO TRAIN_STEP step=3102 loss=1.699488 Training: 17%|█▋ | 3102/18018 [6:43:14<28:43:12, 6.93s/step, loss=1.699, acc=0.637, acc_len=1.49, thru=19.6, I=23.6, T=33.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:28,604] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 20:13:35,317] loop.py:45 INFO TRAIN_STEP step=3103 loss=1.621517 Training: 17%|█▋ | 3103/18018 [6:43:21<28:34:14, 6.90s/step, loss=1.622, acc=0.647, acc_len=1.53, thru=16.5, I=19.5, T=30.4, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:42,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:13:42,562] loop.py:45 INFO TRAIN_STEP step=3104 loss=1.808754 Training: 17%|█▋ | 3104/18018 [6:43:28<29:00:16, 7.00s/step, loss=1.809, acc=0.619, acc_len=1.43, thru=16.9, I=17.1, T=31.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:42,650] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:50,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:13:52,494] loop.py:380 INFO TIMING step=3105: step=4.299s data=1.909s compute=4.258s [fwd=2.995s bwd=1.248s opt=0.003s] dispatch=5.623s [2026-10-18 20:13:52,494] loop.py:45 INFO TRAIN_STEP step=3105 loss=1.904464 Training: 17%|█▋ | 3105/18018 [6:43:38<32:38:34, 7.88s/step, loss=1.904, acc=0.606, acc_len=1.40, thru=13.3, I=15.2, T=29.8, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:13:52,906] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:00,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:14:00,804] loop.py:45 INFO TRAIN_STEP step=3106 loss=1.865641 Training: 17%|█▋ | 3106/18018 [6:43:46<33:10:41, 8.01s/step, loss=1.866, acc=0.611, acc_len=1.40, thru=13.3, I=17.6, T=31.0, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:14:00,798] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3105: forward=2898.6ms backward=1177.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:00,959] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 20:14:08,677] loop.py:45 INFO TRAIN_STEP step=3107 loss=1.748498 Training: 17%|█▋ | 3107/18018 [6:43:54<33:00:17, 7.97s/step, loss=1.748, acc=0.630, acc_len=1.47, thru=12.4, I=13.5, T=31.2, wait=3.8s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:18,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:14:19,149] loop.py:45 INFO TRAIN_STEP step=3108 loss=1.782480 Training: 17%|█▋ | 3108/18018 [6:44:05<36:06:41, 8.72s/step, loss=1.782, acc=0.622, acc_len=1.44, thru=15.8, I=16.0, T=28.8, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:19,585] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:25,324] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:14:27,340] loop.py:45 INFO TRAIN_STEP step=3109 loss=1.663043 Training: 17%|█▋ | 3109/18018 [6:44:13<35:27:15, 8.56s/step, loss=1.663, acc=0.645, acc_len=1.51, thru=15.8, I=19.8, T=28.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:27,843] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:33,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:14:33,885] loop.py:380 INFO TIMING step=3110: step=3.561s data=1.530s compute=3.523s [fwd=2.515s bwd=0.993s opt=0.003s] dispatch=2.976s [2026-10-18 20:14:33,885] loop.py:45 INFO TRAIN_STEP step=3110 loss=1.722950 Training: 17%|█▋ | 3110/18018 [6:44:19<32:56:56, 7.96s/step, loss=1.723, acc=0.634, acc_len=1.48, thru=15.6, I=22.8, T=35.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:34,095] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:42,339] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:14:42,633] loop.py:45 INFO TRAIN_STEP step=3111 loss=1.653541 Training: 17%|█▋ | 3111/18018 [6:44:28<33:55:41, 8.19s/step, loss=1.654, acc=0.645, acc_len=1.52, thru=16.1, I=18.0, T=27.9, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:14:42,628] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3110: forward=3127.2ms backward=1393.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:42,840] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:48,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:14:48,915] loop.py:45 INFO TRAIN_STEP step=3112 loss=1.803451 Training: 17%|█▋ | 3112/18018 [6:44:34<31:33:07, 7.62s/step, loss=1.803, acc=0.622, acc_len=1.44, thru=19.2, I=18.2, T=33.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:49,330] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:53,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:14:55,658] loop.py:45 INFO TRAIN_STEP step=3113 loss=1.934214 Training: 17%|█▋ | 3113/18018 [6:44:41<30:27:31, 7.36s/step, loss=1.934, acc=0.601, acc_len=1.37, thru=18.5, I=23.5, T=33.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:14:55,980] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:15:01,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:15:02,132] loop.py:45 INFO TRAIN_STEP step=3114 loss=1.655718 Training: 17%|█▋ | 3114/18018 [6:44:48<29:21:45, 7.09s/step, loss=1.656, acc=0.642, acc_len=1.51, thru=18.2, I=23.0, T=36.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:15:02,511] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:15:10,864] loop.py:380 INFO TIMING step=3115: step=3.850s data=1.615s compute=3.808s [fwd=2.614s bwd=1.179s opt=0.003s] dispatch=4.872s [2026-10-18 20:15:10,864] loop.py:45 INFO TRAIN_STEP step=3115 loss=1.828778 Training: 17%|█▋ | 3115/18018 [6:44:56<31:23:41, 7.58s/step, loss=1.829, acc=0.619, acc_len=1.43, thru=13.7, I=14.5, T=33.2, wait=4.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:15:16,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:15:17,538] loop.py:45 INFO TRAIN_STEP step=3116 loss=1.683917 Training: 17%|█▋ | 3116/18018 [6:45:03<30:15:51, 7.31s/step, loss=1.684, acc=0.639, acc_len=1.50, thru=18.9, I=19.2, T=32.8, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:15:17,533] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3115: forward=2724.8ms backward=1132.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:15:17,728] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:15:25,496] loop.py:45 INFO TRAIN_STEP step=3117 loss=1.805569 Training: 17%|█▋ | 3117/18018 [6:45:11<31:04:00, 7.51s/step, loss=1.806, acc=0.621, acc_len=1.44, thru=13.8, I=15.9, T=32.8, wait=4.1s, pool=40, epoch=1/2][2026-10-18 20:15:36,369] loop.py:45 INFO TRAIN_STEP step=3118 loss=1.710302 Training: 17%|█▋ | 3118/18018 [6:45:22<35:14:39, 8.52s/step, loss=1.710, acc=0.639, acc_len=1.50, thru=14.2, I=13.2, T=31.5, wait=6.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:15:44,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:15:44,107] loop.py:45 INFO TRAIN_STEP step=3119 loss=1.758960 Training: 17%|█▋ | 3119/18018 [6:45:30<34:16:32, 8.28s/step, loss=1.759, acc=0.632, acc_len=1.47, thru=16.9, I=15.3, T=29.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:15:44,547] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:15:52,723] loop.py:380 INFO TIMING step=3120: step=3.868s data=1.649s compute=3.837s [fwd=2.735s bwd=1.087s opt=0.003s] dispatch=4.740s [2026-10-18 20:15:52,723] loop.py:45 INFO TRAIN_STEP step=3120 loss=1.770848 Training: 17%|█▋ | 3120/18018 [6:45:38<34:41:19, 8.38s/step, loss=1.771, acc=0.621, acc_len=1.44, thru=15.7, I=15.2, T=33.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:15:52,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:15:53,214] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:15:59,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 20:16:01,169] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3120: forward=2922.5ms backward=1231.0ms [2026-10-18 20:16:01,175] loop.py:45 INFO TRAIN_STEP step=3121 loss=1.701234 Training: 17%|█▋ | 3121/18018 [6:45:47<34:46:22, 8.40s/step, loss=1.701, acc=0.640, acc_len=1.50, thru=15.2, I=19.4, T=30.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:01,474] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:07,834] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:16:08,147] loop.py:45 INFO TRAIN_STEP step=3122 loss=1.515216 Training: 17%|█▋ | 3122/18018 [6:45:54<32:59:39, 7.97s/step, loss=1.515, acc=0.664, acc_len=1.58, thru=16.2, I=21.7, T=29.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:08,336] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:16,155] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:16:17,501] loop.py:45 INFO TRAIN_STEP step=3123 loss=1.702057 Training: 17%|█▋ | 3123/18018 [6:46:03<34:42:27, 8.39s/step, loss=1.702, acc=0.638, acc_len=1.49, thru=13.8, I=15.0, T=30.6, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:17,663] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:24,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:16:25,908] loop.py:45 INFO TRAIN_STEP step=3124 loss=1.921922 Training: 17%|█▋ | 3124/18018 [6:46:11<34:43:42, 8.39s/step, loss=1.922, acc=0.599, acc_len=1.38, thru=14.1, I=18.8, T=28.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:26,268] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:33,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:16:34,670] loop.py:380 INFO TIMING step=3125: step=4.221s data=1.916s compute=4.189s [fwd=2.930s bwd=1.243s opt=0.003s] dispatch=4.530s [2026-10-18 20:16:34,670] loop.py:45 INFO TRAIN_STEP step=3125 loss=1.751345 Training: 17%|█▋ | 3125/18018 [6:46:20<35:10:46, 8.50s/step, loss=1.751, acc=0.631, acc_len=1.47, thru=15.2, I=18.4, T=30.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:35,027] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (TrainerActor pid=9310) [2026-10-18 20:16:41,798] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3125: forward=3163.7ms backward=1418.9ms [2026-10-18 20:16:41,802] loop.py:45 INFO TRAIN_STEP step=3126 loss=1.627496 Training: 17%|█▋ | 3126/18018 [6:46:27<33:28:41, 8.09s/step, loss=1.627, acc=0.653, acc_len=1.54, thru=15.4, I=23.0, T=27.7, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:49,171] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:16:49,737] loop.py:45 INFO TRAIN_STEP step=3127 loss=1.744949 Training: 17%|█▋ | 3127/18018 [6:46:35<33:16:34, 8.04s/step, loss=1.745, acc=0.632, acc_len=1.48, thru=16.2, I=15.8, T=30.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:16:50,176] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:16:57,049] loop.py:45 INFO TRAIN_STEP step=3128 loss=1.637874 Training: 17%|█▋ | 3128/18018 [6:46:43<32:22:06, 7.83s/step, loss=1.638, acc=0.645, acc_len=1.52, thru=15.6, I=18.3, T=30.7, wait=3.1s, pool=48, epoch=1/2][2026-10-18 20:17:05,314] loop.py:45 INFO TRAIN_STEP step=3129 loss=1.807418 Training: 17%|█▋ | 3129/18018 [6:46:51<32:54:39, 7.96s/step, loss=1.807, acc=0.623, acc_len=1.45, thru=15.4, I=14.4, T=30.0, wait=4.0s, pool=60, epoch=1/2][2026-10-18 20:17:12,387] loop.py:380 INFO TIMING step=3130: step=4.257s data=1.971s compute=4.215s [fwd=2.891s bwd=1.308s opt=0.003s] dispatch=2.805s [2026-10-18 20:17:12,387] loop.py:45 INFO TRAIN_STEP step=3130 loss=1.750275 Training: 17%|█▋ | 3130/18018 [6:46:58<31:48:32, 7.69s/step, loss=1.750, acc=0.629, acc_len=1.47, thru=17.4, I=17.5, T=30.1, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:17:20,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:17:21,615] loop.py:45 INFO TRAIN_STEP step=3131 loss=1.670855 Training: 17%|█▋ | 3131/18018 [6:47:07<33:42:50, 8.15s/step, loss=1.671, acc=0.642, acc_len=1.50, thru=14.7, I=14.7, T=29.6, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:17:21,610] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3130: forward=3078.4ms backward=1198.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:17:21,941] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:17:28,726] loop.py:45 INFO TRAIN_STEP step=3132 loss=1.806095 Training: 17%|█▋ | 3132/18018 [6:47:14<32:25:02, 7.84s/step, loss=1.806, acc=0.627, acc_len=1.45, thru=16.4, I=22.8, T=29.9, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:17:34,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:17:34,879] loop.py:45 INFO TRAIN_STEP step=3133 loss=1.715965 Training: 17%|█▋ | 3133/18018 [6:47:20<30:19:27, 7.33s/step, loss=1.716, acc=0.636, acc_len=1.49, thru=17.9, I=19.1, T=32.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:17:35,327] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:17:41,756] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:17:42,711] loop.py:45 INFO TRAIN_STEP step=3134 loss=1.676671 Training: 17%|█▋ | 3134/18018 [6:47:28<30:56:20, 7.48s/step, loss=1.677, acc=0.640, acc_len=1.50, thru=16.7, I=19.8, T=30.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:17:42,773] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:17:51,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:17:52,740] loop.py:380 INFO TIMING step=3135: step=5.005s data=2.365s compute=4.965s [fwd=3.558s bwd=1.391s opt=0.003s] dispatch=5.015s [2026-10-18 20:17:52,740] loop.py:45 INFO TRAIN_STEP step=3135 loss=1.762592 Training: 17%|█▋ | 3135/18018 [6:47:38<34:05:47, 8.25s/step, loss=1.763, acc=0.627, acc_len=1.46, thru=12.9, I=13.2, T=25.6, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:17:52,839] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:17:59,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:17:59,698] loop.py:45 INFO TRAIN_STEP step=3136 loss=1.758288 Training: 17%|█▋ | 3136/18018 [6:47:45<32:29:25, 7.86s/step, loss=1.758, acc=0.629, acc_len=1.47, thru=17.6, I=19.4, T=34.4, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:17:59,693] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3135: forward=2635.8ms backward=1042.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:00,023] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:05,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:18:07,060] loop.py:45 INFO TRAIN_STEP step=3137 loss=1.723735 Training: 17%|█▋ | 3137/18018 [6:47:53<31:52:25, 7.71s/step, loss=1.724, acc=0.633, acc_len=1.48, thru=16.5, I=18.7, T=29.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:07,499] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:14,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:18:14,358] loop.py:45 INFO TRAIN_STEP step=3138 loss=1.655705 Training: 17%|█▋ | 3138/18018 [6:48:00<31:21:31, 7.59s/step, loss=1.656, acc=0.647, acc_len=1.52, thru=16.6, I=19.0, T=31.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:14,749] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:19,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:18:21,883] loop.py:45 INFO TRAIN_STEP step=3139 loss=1.701262 Training: 17%|█▋ | 3139/18018 [6:48:07<31:17:00, 7.57s/step, loss=1.701, acc=0.637, acc_len=1.49, thru=17.4, I=22.6, T=27.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:22,122] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:28,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:18:30,034] loop.py:380 INFO TIMING step=3140: step=4.327s data=2.083s compute=4.299s [fwd=3.029s bwd=1.256s opt=0.003s] dispatch=3.812s [2026-10-18 20:18:30,034] loop.py:45 INFO TRAIN_STEP step=3140 loss=1.869263 Training: 17%|█▋ | 3140/18018 [6:48:16<31:59:57, 7.74s/step, loss=1.869, acc=0.606, acc_len=1.39, thru=16.2, I=19.4, T=29.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:30,463] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 20:18:37,217] loop.py:45 INFO TRAIN_STEP step=3141 loss=1.717946 Training: 17%|█▋ | 3141/18018 [6:48:23<31:18:22, 7.58s/step, loss=1.718, acc=0.635, acc_len=1.49, thru=13.1, I=15.9, T=33.3, wait=3.3s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:18:37,212] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3140: forward=2731.4ms backward=1066.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:46,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:18:48,249] loop.py:45 INFO TRAIN_STEP step=3142 loss=1.849892 Training: 17%|█▋ | 3142/18018 [6:48:34<35:35:10, 8.61s/step, loss=1.850, acc=0.618, acc_len=1.43, thru=14.5, I=16.2, T=28.3, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:48,693] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:53,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:18:54,849] loop.py:45 INFO TRAIN_STEP step=3143 loss=1.767526 Training: 17%|█▋ | 3143/18018 [6:48:40<33:05:24, 8.01s/step, loss=1.768, acc=0.627, acc_len=1.46, thru=16.7, I=22.5, T=32.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:18:54,916] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:00,646] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:19:01,764] loop.py:45 INFO TRAIN_STEP step=3144 loss=1.577178 Training: 17%|█▋ | 3144/18018 [6:48:47<31:44:08, 7.68s/step, loss=1.577, acc=0.659, acc_len=1.56, thru=18.1, I=21.0, T=30.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:02,157] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:08,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:19:09,120] loop.py:380 INFO TIMING step=3145: step=4.106s data=1.871s compute=4.058s [fwd=2.771s bwd=1.271s opt=0.003s] dispatch=3.239s [2026-10-18 20:19:09,120] loop.py:45 INFO TRAIN_STEP step=3145 loss=1.864695 Training: 17%|█▋ | 3145/18018 [6:48:55<31:19:45, 7.58s/step, loss=1.865, acc=0.612, acc_len=1.41, thru=16.1, I=19.4, T=31.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:09,502] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:14,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:19:16,428] loop.py:45 INFO TRAIN_STEP step=3146 loss=1.766773 Training: 17%|█▋ | 3146/18018 [6:49:02<30:59:06, 7.50s/step, loss=1.767, acc=0.628, acc_len=1.47, thru=16.9, I=20.7, T=29.6, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:19:16,424] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3145: forward=3053.8ms backward=1221.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:16,522] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:22,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:19:23,854] loop.py:45 INFO TRAIN_STEP step=3147 loss=1.664925 Training: 17%|█▋ | 3147/18018 [6:49:09<30:53:22, 7.48s/step, loss=1.665, acc=0.640, acc_len=1.50, thru=15.3, I=20.7, T=30.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:23,939] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:29,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:19:30,965] loop.py:45 INFO TRAIN_STEP step=3148 loss=1.612305 Training: 17%|█▋ | 3148/18018 [6:49:17<30:25:55, 7.37s/step, loss=1.612, acc=0.652, acc_len=1.55, thru=17.5, I=21.7, T=31.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:31,456] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:38,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:19:39,905] loop.py:45 INFO TRAIN_STEP step=3149 loss=1.806729 Training: 17%|█▋ | 3149/18018 [6:49:25<32:22:50, 7.84s/step, loss=1.807, acc=0.624, acc_len=1.45, thru=13.2, I=16.7, T=28.2, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:40,329] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:46,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:19:47,430] loop.py:380 INFO TIMING step=3150: step=3.677s data=1.643s compute=3.637s [fwd=2.577s bwd=1.045s opt=0.003s] dispatch=3.839s [2026-10-18 20:19:47,430] loop.py:45 INFO TRAIN_STEP step=3150 loss=1.795933 Training: 17%|█▋ | 3150/18018 [6:49:33<31:59:21, 7.75s/step, loss=1.796, acc=0.618, acc_len=1.44, thru=16.3, I=20.2, T=34.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:47,595] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:52,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:19:53,383] loop.py:45 INFO TRAIN_STEP step=3151 loss=1.675094 Training: 17%|█▋ | 3151/18018 [6:49:39<29:45:51, 7.21s/step, loss=1.675, acc=0.642, acc_len=1.50, thru=20.3, I=21.8, T=33.0, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:19:53,377] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3150: forward=2723.9ms backward=1098.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:19:53,499] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-18 20:19:59,941] loop.py:45 INFO TRAIN_STEP step=3152 loss=1.577627 Training: 17%|█▋ | 3152/18018 [6:49:46<28:57:44, 7.01s/step, loss=1.578, acc=0.654, acc_len=1.55, thru=18.1, I=22.0, T=34.1, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:07,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:20:07,722] loop.py:45 INFO TRAIN_STEP step=3153 loss=1.731007 Training: 17%|█▋ | 3153/18018 [6:49:53<29:54:24, 7.24s/step, loss=1.731, acc=0.632, acc_len=1.47, thru=16.4, I=16.8, T=33.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:08,050] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:12,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:20:15,203] loop.py:45 INFO TRAIN_STEP step=3154 loss=1.665979 Training: 18%|█▊ | 3154/18018 [6:50:01<30:11:52, 7.31s/step, loss=1.666, acc=0.643, acc_len=1.51, thru=18.2, I=22.1, T=26.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:15,338] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:21,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:20:22,785] loop.py:380 INFO TIMING step=3155: step=4.172s data=1.923s compute=4.141s [fwd=2.937s bwd=1.189s opt=0.003s] dispatch=3.403s [2026-10-18 20:20:22,785] loop.py:45 INFO TRAIN_STEP step=3155 loss=1.709450 Training: 18%|█▊ | 3155/18018 [6:50:08<30:31:50, 7.39s/step, loss=1.709, acc=0.636, acc_len=1.49, thru=17.2, I=17.9, T=30.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:22,816] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:29,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:20:30,360] loop.py:45 INFO TRAIN_STEP step=3156 loss=1.699254 Training: 18%|█▊ | 3156/18018 [6:50:16<30:45:05, 7.45s/step, loss=1.699, acc=0.636, acc_len=1.49, thru=16.1, I=18.0, T=29.6, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:20:30,355] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3155: forward=3048.1ms backward=1234.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:30,634] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:38,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:20:38,857] loop.py:45 INFO TRAIN_STEP step=3157 loss=1.753581 Training: 18%|█▊ | 3157/18018 [6:50:24<32:02:55, 7.76s/step, loss=1.754, acc=0.626, acc_len=1.46, thru=15.3, I=16.5, T=32.1, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:39,062] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:45,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:20:45,643] loop.py:45 INFO TRAIN_STEP step=3158 loss=1.710228 Training: 18%|█▊ | 3158/18018 [6:50:31<30:50:11, 7.47s/step, loss=1.710, acc=0.633, acc_len=1.48, thru=18.4, I=20.8, T=30.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:46,011] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:51,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:20:51,835] loop.py:45 INFO TRAIN_STEP step=3159 loss=1.627622 Training: 18%|█▊ | 3159/18018 [6:50:37<29:14:55, 7.09s/step, loss=1.628, acc=0.649, acc_len=1.53, thru=18.7, I=20.4, T=32.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:52,319] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:20:59,799] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:20:59,860] loop.py:380 INFO TIMING step=3160: step=4.205s data=2.011s compute=4.161s [fwd=2.908s bwd=1.238s opt=0.003s] dispatch=3.811s [2026-10-18 20:20:59,860] loop.py:45 INFO TRAIN_STEP step=3160 loss=1.821155 Training: 18%|█▊ | 3160/18018 [6:50:45<30:24:27, 7.37s/step, loss=1.821, acc=0.620, acc_len=1.43, thru=16.7, I=17.5, T=30.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:21:00,301] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:21:06,925] loop.py:45 INFO TRAIN_STEP step=3161 loss=1.671336 Training: 18%|█▊ | 3161/18018 [6:50:53<30:02:04, 7.28s/step, loss=1.671, acc=0.643, acc_len=1.51, thru=16.7, I=15.6, T=33.3, wait=3.2s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:21:06,920] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3160: forward=2675.7ms backward=1127.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:21:17,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:21:19,228] loop.py:45 INFO TRAIN_STEP step=3162 loss=1.884989 Training: 18%|█▊ | 3162/18018 [6:51:05<36:15:13, 8.79s/step, loss=1.885, acc=0.612, acc_len=1.42, thru=13.6, I=13.1, T=29.3, wait=7.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:21:19,514] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:21:27,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:21:27,227] loop.py:45 INFO TRAIN_STEP step=3163 loss=1.720508 Training: 18%|█▊ | 3163/18018 [6:51:13<35:16:32, 8.55s/step, loss=1.721, acc=0.636, acc_len=1.49, thru=13.6, I=17.6, T=27.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:21:27,644] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:21:36,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:21:37,116] loop.py:45 INFO TRAIN_STEP step=3164 loss=1.844131 Training: 18%|█▊ | 3164/18018 [6:51:23<36:56:04, 8.95s/step, loss=1.844, acc=0.619, acc_len=1.43, thru=13.2, I=15.1, T=25.3, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:21:37,367] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:21:41,293] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:21:44,556] loop.py:380 INFO TIMING step=3165: step=5.202s data=2.585s compute=5.166s [fwd=3.502s bwd=1.650s opt=0.003s] dispatch=2.228s [2026-10-18 20:21:44,556] loop.py:45 INFO TRAIN_STEP step=3165 loss=1.694925 Training: 18%|█▊ | 3165/18018 [6:51:30<35:03:32, 8.50s/step, loss=1.695, acc=0.638, acc_len=1.49, thru=14.5, I=21.9, T=24.6, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:21:44,823] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 [2026-10-18 20:21:51,870] loop.py:45 INFO TRAIN_STEP step=3166 loss=1.774101 Training: 18%|█▊ | 3166/18018 [6:51:37<33:35:53, 8.14s/step, loss=1.774, acc=0.626, acc_len=1.46, thru=15.3, I=15.9, T=34.3, wait=3.6s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:21:51,865] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3165: forward=2573.4ms backward=1113.2ms [2026-10-18 20:21:59,860] loop.py:45 INFO TRAIN_STEP step=3167 loss=1.677991 Training: 18%|█▊ | 3167/18018 [6:51:45<33:24:12, 8.10s/step, loss=1.678, acc=0.638, acc_len=1.50, thru=17.9, I=17.2, T=30.9, wait=3.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:07,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:22:09,691] loop.py:45 INFO TRAIN_STEP step=3168 loss=1.692334 Training: 18%|█▊ | 3168/18018 [6:51:55<35:32:42, 8.62s/step, loss=1.692, acc=0.639, acc_len=1.50, thru=13.5, I=16.7, T=28.9, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:10,028] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:16,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:22:16,883] loop.py:45 INFO TRAIN_STEP step=3169 loss=1.597472 Training: 18%|█▊ | 3169/18018 [6:52:02<33:46:46, 8.19s/step, loss=1.597, acc=0.657, acc_len=1.55, thru=14.5, I=20.1, T=29.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:17,065] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:22:25,679] loop.py:380 INFO TIMING step=3170: step=4.829s data=2.602s compute=4.796s [fwd=3.572s bwd=1.209s opt=0.003s] dispatch=3.957s [2026-10-18 20:22:25,679] loop.py:45 INFO TRAIN_STEP step=3170 loss=1.712404 Training: 18%|█▊ | 3170/18018 [6:52:11<34:31:36, 8.37s/step, loss=1.712, acc=0.638, acc_len=1.49, thru=14.0, I=14.8, T=26.5, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:33,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:22:35,246] loop.py:45 INFO TRAIN_STEP step=3171 loss=1.846766 Training: 18%|█▊ | 3171/18018 [6:52:21<36:00:05, 8.73s/step, loss=1.847, acc=0.611, acc_len=1.42, thru=14.0, I=15.8, T=28.1, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:22:35,240] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3170: forward=3346.2ms backward=1161.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:35,749] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:40,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:22:42,655] loop.py:45 INFO TRAIN_STEP step=3172 loss=1.726949 Training: 18%|█▊ | 3172/18018 [6:52:28<34:21:58, 8.33s/step, loss=1.727, acc=0.634, acc_len=1.48, thru=14.4, I=22.1, T=27.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:42,803] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:48,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:22:48,789] loop.py:45 INFO TRAIN_STEP step=3173 loss=1.458515 Training: 18%|█▊ | 3173/18018 [6:52:34<31:38:37, 7.67s/step, loss=1.459, acc=0.672, acc_len=1.61, thru=18.1, I=22.5, T=34.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:48,871] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:54,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:22:55,761] loop.py:45 INFO TRAIN_STEP step=3174 loss=1.773027 Training: 18%|█▊ | 3174/18018 [6:52:41<30:46:29, 7.46s/step, loss=1.773, acc=0.625, acc_len=1.45, thru=18.1, I=21.9, T=30.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:22:55,945] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:02,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:23:04,222] loop.py:380 INFO TIMING step=3175: step=4.302s data=1.982s compute=4.257s [fwd=2.947s bwd=1.294s opt=0.003s] dispatch=4.149s [2026-10-18 20:23:04,222] loop.py:45 INFO TRAIN_STEP step=3175 loss=1.770152 Training: 18%|█▊ | 3175/18018 [6:52:50<32:00:23, 7.76s/step, loss=1.770, acc=0.630, acc_len=1.46, thru=14.2, I=16.5, T=29.8, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:04,275] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:08,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:23:11,644] loop.py:45 INFO TRAIN_STEP step=3176 loss=1.849283 (TrainerActor pid=9310) [2026-10-18 20:23:11,639] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3175: forward=3560.7ms backward=1336.3ms Training: 18%|█▊ | 3176/18018 [6:52:57<31:34:53, 7.66s/step, loss=1.849, acc=0.612, acc_len=1.41, thru=14.9, I=24.0, T=25.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:11,786] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:17,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:23:18,987] loop.py:45 INFO TRAIN_STEP step=3177 loss=1.630306 Training: 18%|█▊ | 3177/18018 [6:53:05<31:11:10, 7.56s/step, loss=1.630, acc=0.650, acc_len=1.53, thru=17.8, I=22.5, T=29.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:19,082] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:25,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:23:26,841] loop.py:45 INFO TRAIN_STEP step=3178 loss=1.684186 Training: 18%|█▊ | 3178/18018 [6:53:12<31:32:39, 7.65s/step, loss=1.684, acc=0.639, acc_len=1.50, thru=13.5, I=18.6, T=28.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:26,899] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 20:23:33,619] loop.py:45 INFO TRAIN_STEP step=3179 loss=1.675751 Training: 18%|█▊ | 3179/18018 [6:53:19<30:27:28, 7.39s/step, loss=1.676, acc=0.642, acc_len=1.52, thru=14.2, I=15.6, T=29.6, wait=2.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:42,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:23:43,561] loop.py:380 INFO TIMING step=3180: step=4.450s data=1.867s compute=4.428s [fwd=3.027s bwd=1.387s opt=0.003s] dispatch=5.484s [2026-10-18 20:23:43,562] loop.py:45 INFO TRAIN_STEP step=3180 loss=1.739524 Training: 18%|█▊ | 3180/18018 [6:53:29<33:36:49, 8.16s/step, loss=1.740, acc=0.632, acc_len=1.47, thru=15.4, I=17.1, T=28.8, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:43,843] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:49,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 20:23:51,388] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3180: forward=3199.7ms backward=1315.2ms [2026-10-18 20:23:51,393] loop.py:45 INFO TRAIN_STEP step=3181 loss=1.755849 Training: 18%|█▊ | 3181/18018 [6:53:37<33:12:41, 8.06s/step, loss=1.756, acc=0.627, acc_len=1.46, thru=14.7, I=21.1, T=28.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:51,632] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:58,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:23:58,703] loop.py:45 INFO TRAIN_STEP step=3182 loss=1.653409 Training: 18%|█▊ | 3182/18018 [6:53:44<32:16:56, 7.83s/step, loss=1.653, acc=0.645, acc_len=1.52, thru=14.2, I=19.0, T=31.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:23:59,000] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:06,066] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:24:06,223] loop.py:45 INFO TRAIN_STEP step=3183 loss=1.794143 Training: 18%|█▊ | 3183/18018 [6:53:52<31:53:46, 7.74s/step, loss=1.794, acc=0.624, acc_len=1.45, thru=17.8, I=16.5, T=30.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:06,568] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:24:13,851] loop.py:45 INFO TRAIN_STEP step=3184 loss=1.723720 Training: 18%|█▊ | 3184/18018 [6:53:59<31:45:40, 7.71s/step, loss=1.724, acc=0.631, acc_len=1.47, thru=15.2, I=15.4, T=28.6, wait=3.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:20,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:24:20,797] loop.py:380 INFO TIMING step=3185: step=4.167s data=1.910s compute=4.140s [fwd=2.947s bwd=1.179s opt=0.003s] dispatch=2.764s [2026-10-18 20:24:20,797] loop.py:45 INFO TRAIN_STEP step=3185 loss=1.832772 Training: 18%|█▊ | 3185/18018 [6:54:06<30:48:30, 7.48s/step, loss=1.833, acc=0.616, acc_len=1.42, thru=18.1, I=18.1, T=30.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:21,162] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:27,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:24:28,394] loop.py:45 INFO TRAIN_STEP step=3186 loss=1.668722 Training: 18%|█▊ | 3186/18018 [6:54:14<30:57:14, 7.51s/step, loss=1.669, acc=0.641, acc_len=1.50, thru=16.1, I=19.3, T=28.9, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:24:28,389] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3185: forward=3073.9ms backward=1304.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:28,895] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:35,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:24:36,218] loop.py:45 INFO TRAIN_STEP step=3187 loss=1.797925 Training: 18%|█▊ | 3187/18018 [6:54:22<31:20:17, 7.61s/step, loss=1.798, acc=0.624, acc_len=1.44, thru=15.7, I=19.5, T=31.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:36,575] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:24:43,358] loop.py:45 INFO TRAIN_STEP step=3188 loss=1.648687 Training: 18%|█▊ | 3188/18018 [6:54:29<30:45:27, 7.47s/step, loss=1.649, acc=0.646, acc_len=1.53, thru=16.1, I=17.4, T=30.9, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:51,237] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:24:52,785] loop.py:45 INFO TRAIN_STEP step=3189 loss=1.714512 Training: 18%|█▊ | 3189/18018 [6:54:38<33:10:39, 8.05s/step, loss=1.715, acc=0.637, acc_len=1.49, thru=15.0, I=16.8, T=28.6, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:53,248] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:24:59,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:25:01,062] loop.py:380 INFO TIMING step=3190: step=4.617s data=1.952s compute=4.591s [fwd=3.209s bwd=1.368s opt=0.003s] dispatch=3.653s [2026-10-18 20:25:01,062] loop.py:45 INFO TRAIN_STEP step=3190 loss=1.713059 Training: 18%|█▊ | 3190/18018 [6:54:47<33:27:03, 8.12s/step, loss=1.713, acc=0.636, acc_len=1.48, thru=13.7, I=19.6, T=27.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:01,217] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:09,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:25:09,611] loop.py:45 INFO TRAIN_STEP step=3191 loss=1.674442 Training: 18%|█▊ | 3191/18018 [6:54:55<33:58:37, 8.25s/step, loss=1.674, acc=0.642, acc_len=1.51, thru=15.2, I=15.3, T=25.9, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:25:09,606] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3190: forward=3572.8ms backward=1313.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:09,825] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:14,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:25:16,702] loop.py:45 INFO TRAIN_STEP step=3192 loss=1.762790 Training: 18%|█▊ | 3192/18018 [6:55:02<32:32:36, 7.90s/step, loss=1.763, acc=0.625, acc_len=1.46, thru=16.8, I=19.1, T=27.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:16,839] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:22,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:25:23,945] loop.py:45 INFO TRAIN_STEP step=3193 loss=1.624448 Training: 18%|█▊ | 3193/18018 [6:55:10<31:43:32, 7.70s/step, loss=1.624, acc=0.649, acc_len=1.52, thru=15.2, I=21.6, T=26.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:24,390] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:28,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:25:30,972] loop.py:45 INFO TRAIN_STEP step=3194 loss=1.657832 Training: 18%|█▊ | 3194/18018 [6:55:17<30:53:18, 7.50s/step, loss=1.658, acc=0.643, acc_len=1.52, thru=16.8, I=24.4, T=30.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:31,363] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:36,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:25:37,373] loop.py:380 INFO TIMING step=3195: step=3.706s data=1.672s compute=3.677s [fwd=2.614s bwd=1.049s opt=0.003s] dispatch=2.688s [2026-10-18 20:25:37,374] loop.py:45 INFO TRAIN_STEP step=3195 loss=1.705818 Training: 18%|█▊ | 3195/18018 [6:55:23<29:31:41, 7.17s/step, loss=1.706, acc=0.635, acc_len=1.48, thru=17.4, I=27.7, T=34.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:37,612] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:44,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:25:45,455] loop.py:45 INFO TRAIN_STEP step=3196 loss=1.650389 Training: 18%|█▊ | 3196/18018 [6:55:31<30:38:55, 7.44s/step, loss=1.650, acc=0.646, acc_len=1.52, thru=15.8, I=19.5, T=29.2, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:25:45,450] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3195: forward=3122.6ms backward=1221.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:45,921] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:54,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:25:54,645] loop.py:45 INFO TRAIN_STEP step=3197 loss=1.675287 Training: 18%|█▊ | 3197/18018 [6:55:40<32:48:14, 7.97s/step, loss=1.675, acc=0.641, acc_len=1.50, thru=14.7, I=15.1, T=25.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:25:55,004] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:01,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:26:01,280] loop.py:45 INFO TRAIN_STEP step=3198 loss=1.680515 Training: 18%|█▊ | 3198/18018 [6:55:47<31:09:17, 7.57s/step, loss=1.681, acc=0.641, acc_len=1.50, thru=18.8, I=16.0, T=33.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:01,725] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:26:08,086] loop.py:45 INFO TRAIN_STEP step=3199 loss=1.666930 Training: 18%|█▊ | 3199/18018 [6:55:54<30:12:44, 7.34s/step, loss=1.667, acc=0.643, acc_len=1.51, thru=18.2, I=19.5, T=31.5, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:17,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:26:19,858] loop.py:380 INFO TIMING step=3200: step=4.501s data=2.053s compute=4.465s [fwd=3.128s bwd=1.321s opt=0.003s] dispatch=7.264s [2026-10-18 20:26:19,858] loop.py:45 INFO TRAIN_STEP step=3200 loss=1.701021 Training: 18%|█▊ | 3200/18018 [6:56:05<35:41:11, 8.67s/step, loss=1.701, acc=0.638, acc_len=1.50, thru=12.5, I=13.0, T=28.4, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:19,892] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:24,904] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:26:26,747] loop.py:45 INFO TRAIN_STEP step=3201 loss=1.737335 Training: 18%|█▊ | 3201/18018 [6:56:12<33:29:09, 8.14s/step, loss=1.737, acc=0.634, acc_len=1.48, thru=15.2, I=24.9, T=28.7, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:26:26,742] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3200: forward=3003.6ms backward=1406.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:26,918] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:33,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:26:35,063] loop.py:45 INFO TRAIN_STEP step=3202 loss=1.678710 Training: 18%|█▊ | 3202/18018 [6:56:21<33:42:07, 8.19s/step, loss=1.679, acc=0.646, acc_len=1.52, thru=15.7, I=19.3, T=29.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:35,517] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:42,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:26:43,136] loop.py:45 INFO TRAIN_STEP step=3203 loss=1.760387 Training: 18%|█▊ | 3203/18018 [6:56:29<33:33:22, 8.15s/step, loss=1.760, acc=0.627, acc_len=1.46, thru=14.0, I=18.7, T=32.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:43,612] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:48,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:26:49,685] loop.py:45 INFO TRAIN_STEP step=3204 loss=1.810289 Training: 18%|█▊ | 3204/18018 [6:56:35<31:34:33, 7.67s/step, loss=1.810, acc=0.618, acc_len=1.43, thru=18.7, I=20.2, T=31.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:50,014] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:56,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:26:57,436] loop.py:380 INFO TIMING step=3205: step=4.173s data=1.843s compute=4.140s [fwd=2.905s bwd=1.219s opt=0.003s] dispatch=3.568s [2026-10-18 20:26:57,436] loop.py:45 INFO TRAIN_STEP step=3205 loss=1.682385 Training: 18%|█▊ | 3205/18018 [6:56:43<31:40:04, 7.70s/step, loss=1.682, acc=0.640, acc_len=1.50, thru=15.5, I=19.3, T=30.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:26:57,522] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:27:04,615] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:27:06,533] loop.py:45 INFO TRAIN_STEP step=3206 loss=1.768363 Training: 18%|█▊ | 3206/18018 [6:56:52<33:23:40, 8.12s/step, loss=1.768, acc=0.629, acc_len=1.46, thru=14.2, I=17.7, T=25.5, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:27:06,527] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3205: forward=3495.6ms backward=1443.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:27:06,637] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:27:13,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:27:13,573] loop.py:45 INFO TRAIN_STEP step=3207 loss=1.663289 Training: 18%|█▊ | 3207/18018 [6:56:59<32:03:58, 7.79s/step, loss=1.663, acc=0.645, acc_len=1.52, thru=15.2, I=18.2, T=30.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:27:14,038] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:27:22,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:27:23,018] loop.py:45 INFO TRAIN_STEP step=3208 loss=1.652791 Training: 18%|█▊ | 3208/18018 [6:57:09<34:05:56, 8.29s/step, loss=1.653, acc=0.641, acc_len=1.50, thru=13.2, I=16.0, T=31.9, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:27:23,479] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:27:31,599] loop.py:45 INFO TRAIN_STEP step=3209 loss=1.644872 Training: 18%|█▊ | 3209/18018 [6:57:17<34:27:27, 8.38s/step, loss=1.645, acc=0.644, acc_len=1.51, thru=11.7, I=14.2, T=31.2, wait=4.5s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:27:42,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:27:44,455] loop.py:380 INFO TIMING step=3210: step=4.039s data=1.683s compute=4.013s [fwd=2.861s bwd=1.138s opt=0.003s] dispatch=8.809s [2026-10-18 20:27:44,455] loop.py:45 INFO TRAIN_STEP step=3210 loss=1.743383 Training: 18%|█▊ | 3210/18018 [6:57:30<39:59:04, 9.72s/step, loss=1.743, acc=0.630, acc_len=1.47, thru=16.0, I=13.2, T=31.7, wait=8.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:27:44,580] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (TrainerActor pid=9310) [2026-10-18 20:27:51,119] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3210: forward=2693.8ms backward=1103.6ms [2026-10-18 20:27:51,123] loop.py:45 INFO TRAIN_STEP step=3211 loss=1.625996 Training: 18%|█▊ | 3211/18018 [6:57:37<36:12:49, 8.80s/step, loss=1.626, acc=0.647, acc_len=1.52, thru=17.4, I=22.7, T=33.2, wait=2.8s, pool=60, epoch=1/2][2026-10-18 20:27:57,391] loop.py:45 INFO TRAIN_STEP step=3212 loss=1.856036 Training: 18%|█▊ | 3212/18018 [6:57:43<33:04:52, 8.04s/step, loss=1.856, acc=0.610, acc_len=1.40, thru=19.1, I=19.4, T=33.1, wait=2.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:08,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:28:09,011] loop.py:45 INFO TRAIN_STEP step=3213 loss=1.668924 Training: 18%|█▊ | 3213/18018 [6:57:55<37:29:31, 9.12s/step, loss=1.669, acc=0.644, acc_len=1.51, thru=14.5, I=11.9, T=30.6, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:09,128] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:14,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:28:15,521] loop.py:45 INFO TRAIN_STEP step=3214 loss=1.854450 Training: 18%|█▊ | 3214/18018 [6:58:01<34:16:29, 8.33s/step, loss=1.854, acc=0.615, acc_len=1.42, thru=18.1, I=22.3, T=32.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:15,555] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:19,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:28:21,195] loop.py:380 INFO TIMING step=3215: step=3.872s data=1.765s compute=3.835s [fwd=2.710s bwd=1.110s opt=0.003s] dispatch=1.793s [2026-10-18 20:28:21,195] loop.py:45 INFO TRAIN_STEP step=3215 loss=1.625469 Training: 18%|█▊ | 3215/18018 [6:58:07<30:59:11, 7.54s/step, loss=1.625, acc=0.651, acc_len=1.53, thru=20.4, I=29.1, T=33.1, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:21,694] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:25,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:28:28,560] loop.py:45 INFO TRAIN_STEP step=3216 loss=1.732218 Training: 18%|█▊ | 3216/18018 [6:58:14<30:46:39, 7.49s/step, loss=1.732, acc=0.637, acc_len=1.49, thru=15.2, I=27.3, T=25.9, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:28:28,555] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3215: forward=3371.5ms backward=1525.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:28,983] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:35,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:28:36,246] loop.py:45 INFO TRAIN_STEP step=3217 loss=1.753946 Training: 18%|█▊ | 3217/18018 [6:58:22<31:01:16, 7.55s/step, loss=1.754, acc=0.631, acc_len=1.47, thru=17.6, I=19.1, T=27.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:36,315] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:43,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:28:44,901] loop.py:45 INFO TRAIN_STEP step=3218 loss=1.779672 Training: 18%|█▊ | 3218/18018 [6:58:30<32:23:15, 7.88s/step, loss=1.780, acc=0.628, acc_len=1.47, thru=13.7, I=15.9, T=30.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:45,064] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:50,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:28:51,706] loop.py:45 INFO TRAIN_STEP step=3219 loss=1.738807 Training: 18%|█▊ | 3219/18018 [6:58:37<31:03:50, 7.56s/step, loss=1.739, acc=0.629, acc_len=1.46, thru=16.2, I=21.9, T=31.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:51,982] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:58,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:28:58,322] loop.py:380 INFO TIMING step=3220: step=3.843s data=1.672s compute=3.812s [fwd=2.697s bwd=1.100s opt=0.003s] dispatch=2.762s [2026-10-18 20:28:58,322] loop.py:45 INFO TRAIN_STEP step=3220 loss=1.650150 Training: 18%|█▊ | 3220/18018 [6:58:44<29:53:58, 7.27s/step, loss=1.650, acc=0.647, acc_len=1.52, thru=19.3, I=20.4, T=33.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:28:58,606] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:29:04,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:29:05,588] loop.py:45 INFO TRAIN_STEP step=3221 loss=1.582781 Training: 18%|█▊ | 3221/18018 [6:58:51<29:53:23, 7.27s/step, loss=1.583, acc=0.653, acc_len=1.54, thru=17.0, I=19.7, T=29.2, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:29:05,583] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3220: forward=3009.6ms backward=1326.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:29:05,902] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:29:10,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:29:12,214] loop.py:45 INFO TRAIN_STEP step=3222 loss=1.715601 Training: 18%|█▊ | 3222/18018 [6:58:58<29:05:27, 7.08s/step, loss=1.716, acc=0.638, acc_len=1.49, thru=17.5, I=25.4, T=33.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:29:12,715] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 20:29:20,747] loop.py:45 INFO TRAIN_STEP step=3223 loss=1.796471 Training: 18%|█▊ | 3223/18018 [6:59:06<30:53:04, 7.51s/step, loss=1.796, acc=0.621, acc_len=1.44, thru=13.4, I=16.3, T=31.1, wait=4.4s, pool=44, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 20:29:20,838] eagle_store.py:517 WARNING batch_get_buffer missing keys (k103135-ab5c2cd6_hs, k103135-ab5c2cd6_ids, k103135-ab5c2cd6_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:29:28,141] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 20:29:20,831] eagle_store.py:517 WARNING batch_get_buffer missing keys (k103134-914d7a75_hs, k103134-914d7a75_ids, k103134-914d7a75_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-18 20:29:30,532] loop.py:45 INFO TRAIN_STEP step=3224 loss=1.601350 Training: 18%|█▊ | 3224/18018 [6:59:16<33:40:49, 8.20s/step, loss=1.601, acc=0.654, acc_len=1.54, thru=15.5, I=16.9, T=28.4, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:29:30,663] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=16/64 [2026-10-18 20:29:36,831] loop.py:380 INFO TIMING step=3225: step=3.861s data=1.691s compute=3.821s [fwd=2.699s bwd=1.107s opt=0.003s] dispatch=2.429s [2026-10-18 20:29:36,831] loop.py:45 INFO TRAIN_STEP step=3225 loss=1.812722 Training: 18%|█▊ | 3225/18018 [6:59:22<31:20:21, 7.63s/step, loss=1.813, acc=0.619, acc_len=1.44, thru=14.3, I=19.8, T=33.2, wait=2.4s, pool=36, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 20:29:36,911] eagle_store.py:517 WARNING batch_get_buffer missing keys (k103203-b4e4e400_hs, k103203-b4e4e400_ids, k103203-b4e4e400_lhs); sleeping 0.50s. (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 20:29:36,905] eagle_store.py:517 WARNING batch_get_buffer missing keys (k103200-900bb554_hs, k103200-900bb554_ids, k103200-900bb554_lhs); sleeping 0.50s. [2026-10-18 20:29:46,869] loop.py:45 INFO TRAIN_STEP step=3226 loss=1.700573 Training: 18%|█▊ | 3226/18018 [6:59:32<34:18:32, 8.35s/step, loss=1.701, acc=0.637, acc_len=1.49, thru=13.6, I=15.0, T=31.0, wait=5.9s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:29:46,864] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3225: forward=2854.5ms backward=1200.5ms (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-18 20:29:36,952] eagle_store.py:517 WARNING batch_get_buffer missing keys (k103201-a64480b9_hs, k103201-a64480b9_ids, k103201-a64480b9_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-18 20:29:55,063] loop.py:45 INFO TRAIN_STEP step=3227 loss=1.624432 Training: 18%|█▊ | 3227/18018 [6:59:41<34:06:43, 8.30s/step, loss=1.624, acc=0.650, acc_len=1.53, thru=17.1, I=15.7, T=30.3, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:03,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:30:05,017] loop.py:45 INFO TRAIN_STEP step=3228 loss=1.801891 Training: 18%|█▊ | 3228/18018 [6:59:51<36:08:44, 8.80s/step, loss=1.802, acc=0.620, acc_len=1.43, thru=14.1, I=14.1, T=27.3, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:05,474] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:10,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:30:11,355] loop.py:45 INFO TRAIN_STEP step=3229 loss=1.754269 Training: 18%|█▊ | 3229/18018 [6:59:57<33:06:43, 8.06s/step, loss=1.754, acc=0.632, acc_len=1.48, thru=16.7, I=19.9, T=32.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:11,401] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:18,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:30:20,237] loop.py:380 INFO TIMING step=3230: step=4.327s data=1.975s compute=4.284s [fwd=3.005s bwd=1.264s opt=0.003s] dispatch=4.545s [2026-10-18 20:30:20,237] loop.py:45 INFO TRAIN_STEP step=3230 loss=1.886118 Training: 18%|█▊ | 3230/18018 [7:00:06<34:07:13, 8.31s/step, loss=1.886, acc=0.610, acc_len=1.40, thru=13.9, I=16.9, T=29.6, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:20,739] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:27,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:30:28,618] loop.py:45 INFO TRAIN_STEP step=3231 loss=1.635810 Training: 18%|█▊ | 3231/18018 [7:00:14<34:12:44, 8.33s/step, loss=1.636, acc=0.646, acc_len=1.51, thru=13.3, I=17.7, T=30.8, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:30:28,613] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3230: forward=2911.0ms backward=1191.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:28,648] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 20:30:34,504] loop.py:45 INFO TRAIN_STEP step=3232 loss=1.680471 Training: 18%|█▊ | 3232/18018 [7:00:20<31:11:51, 7.60s/step, loss=1.680, acc=0.643, acc_len=1.51, thru=17.1, I=18.0, T=36.1, wait=2.3s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:44,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:30:46,247] loop.py:45 INFO TRAIN_STEP step=3233 loss=1.722617 Training: 18%|█▊ | 3233/18018 [7:00:32<36:18:29, 8.84s/step, loss=1.723, acc=0.639, acc_len=1.50, thru=12.8, I=14.7, T=27.3, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:46,366] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:52,670] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:30:53,801] loop.py:45 INFO TRAIN_STEP step=3234 loss=1.825462 Training: 18%|█▊ | 3234/18018 [7:00:39<34:43:06, 8.45s/step, loss=1.825, acc=0.623, acc_len=1.44, thru=14.4, I=19.5, T=31.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:30:54,180] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:31:02,019] loop.py:380 INFO TIMING step=3235: step=3.779s data=1.625s compute=3.740s [fwd=2.669s bwd=1.056s opt=0.003s] dispatch=4.430s [2026-10-18 20:31:02,019] loop.py:45 INFO TRAIN_STEP step=3235 loss=1.703434 Training: 18%|█▊ | 3235/18018 [7:00:48<34:25:33, 8.38s/step, loss=1.703, acc=0.639, acc_len=1.49, thru=13.3, I=16.4, T=33.9, wait=4.4s, pool=56, epoch=1/2][2026-10-18 20:31:09,069] loop.py:45 INFO TRAIN_STEP step=3236 loss=1.755084 Training: 18%|█▊ | 3236/18018 [7:00:55<32:47:01, 7.98s/step, loss=1.755, acc=0.632, acc_len=1.47, thru=17.2, I=19.3, T=33.5, wait=3.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:31:09,064] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3235: forward=2665.5ms backward=1101.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:14,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:31:16,875] loop.py:45 INFO TRAIN_STEP step=3237 loss=1.556926 Training: 18%|█▊ | 3237/18018 [7:01:02<32:33:29, 7.93s/step, loss=1.557, acc=0.659, acc_len=1.56, thru=16.1, I=20.5, T=29.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:17,260] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:23,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:31:23,813] loop.py:45 INFO TRAIN_STEP step=3238 loss=1.789691 Training: 18%|█▊ | 3238/18018 [7:01:09<31:20:13, 7.63s/step, loss=1.790, acc=0.622, acc_len=1.44, thru=15.3, I=22.3, T=31.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:24,134] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:31,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:31:32,418] loop.py:45 INFO TRAIN_STEP step=3239 loss=1.894014 Training: 18%|█▊ | 3239/18018 [7:01:18<32:31:53, 7.92s/step, loss=1.894, acc=0.611, acc_len=1.40, thru=14.3, I=15.2, T=29.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:32,491] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:38,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:31:39,594] loop.py:380 INFO TIMING step=3240: step=4.090s data=1.880s compute=4.052s [fwd=2.869s bwd=1.169s opt=0.003s] dispatch=3.076s [2026-10-18 20:31:39,594] loop.py:45 INFO TRAIN_STEP step=3240 loss=1.665670 Training: 18%|█▊ | 3240/18018 [7:01:25<31:36:31, 7.70s/step, loss=1.666, acc=0.647, acc_len=1.52, thru=17.2, I=18.1, T=31.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:39,937] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:31:47,324] loop.py:45 INFO TRAIN_STEP step=3241 loss=1.648586 Training: 18%|█▊ | 3241/18018 [7:01:33<31:38:32, 7.71s/step, loss=1.649, acc=0.647, acc_len=1.53, thru=14.8, I=16.1, T=32.0, wait=3.7s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:31:47,319] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3240: forward=2826.1ms backward=1116.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:53,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:31:55,139] loop.py:45 INFO TRAIN_STEP step=3242 loss=1.809117 Training: 18%|█▊ | 3242/18018 [7:01:41<31:46:18, 7.74s/step, loss=1.809, acc=0.616, acc_len=1.42, thru=17.9, I=18.9, T=30.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:55,291] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:31:59,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:32:01,065] loop.py:45 INFO TRAIN_STEP step=3243 loss=1.691290 Training: 18%|█▊ | 3243/18018 [7:01:47<29:32:04, 7.20s/step, loss=1.691, acc=0.636, acc_len=1.49, thru=19.3, I=24.1, T=34.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:01,449] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:06,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:32:08,082] loop.py:45 INFO TRAIN_STEP step=3244 loss=1.811812 Training: 18%|█▊ | 3244/18018 [7:01:54<29:18:40, 7.14s/step, loss=1.812, acc=0.626, acc_len=1.45, thru=17.3, I=24.1, T=29.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:08,555] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:13,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:32:14,675] loop.py:380 INFO TIMING step=3245: step=3.668s data=1.600s compute=3.643s [fwd=2.588s bwd=1.040s opt=0.003s] dispatch=2.915s [2026-10-18 20:32:14,675] loop.py:45 INFO TRAIN_STEP step=3245 loss=1.675588 Training: 18%|█▊ | 3245/18018 [7:02:00<28:38:01, 6.98s/step, loss=1.676, acc=0.643, acc_len=1.52, thru=18.0, I=23.7, T=34.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:14,950] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:21,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:32:22,241] loop.py:45 INFO TRAIN_STEP step=3246 loss=1.596127 Training: 18%|█▊ | 3246/18018 [7:02:08<29:21:28, 7.15s/step, loss=1.596, acc=0.650, acc_len=1.53, thru=16.1, I=19.9, T=28.0, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:32:22,236] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3245: forward=3205.8ms backward=1329.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:22,434] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:28,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:32:28,776] loop.py:45 INFO TRAIN_STEP step=3247 loss=1.799765 Training: 18%|█▊ | 3247/18018 [7:02:14<28:35:28, 6.97s/step, loss=1.800, acc=0.617, acc_len=1.43, thru=17.0, I=22.0, T=30.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:29,120] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:32:42,099] loop.py:45 INFO TRAIN_STEP step=3248 loss=1.817664 Training: 18%|█▊ | 3248/18018 [7:02:28<36:24:34, 8.87s/step, loss=1.818, acc=0.621, acc_len=1.44, thru=9.9, I=8.6, T=26.9, wait=8.6s, pool=56, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:49,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:32:50,379] loop.py:45 INFO TRAIN_STEP step=3249 loss=1.871621 Training: 18%|█▊ | 3249/18018 [7:02:36<35:40:35, 8.70s/step, loss=1.872, acc=0.613, acc_len=1.42, thru=15.5, I=16.6, T=30.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:50,542] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:54,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:32:56,222] loop.py:380 INFO TIMING step=3250: step=3.869s data=1.802s compute=3.832s [fwd=2.684s bwd=1.134s opt=0.003s] dispatch=1.963s [2026-10-18 20:32:56,222] loop.py:45 INFO TRAIN_STEP step=3250 loss=1.737842 Training: 18%|█▊ | 3250/18018 [7:02:42<32:09:50, 7.84s/step, loss=1.738, acc=0.633, acc_len=1.48, thru=19.2, I=28.0, T=33.1, wait=2.0s, pool=64, epoch=1/2][2026-10-18 20:32:56,227] loop.py:419 INFO Saving checkpoint at step 3250... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:32:59,251] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 20:32:59,644] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0003251 (TrainerActor pid=9310) [2026-10-18 20:32:59,654] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0002501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:00,759] inference_manager.py:408 INFO Pool capacity available after 6.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:06,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 20:32:59,643] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0003251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 20:33:00,516] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0002501 [2026-10-18 20:33:07,177] loop.py:45 INFO TRAIN_STEP step=3251 loss=1.697825 Training: 18%|█▊ | 3251/18018 [7:02:53<35:59:31, 8.77s/step, loss=1.698, acc=0.637, acc_len=1.50, thru=19.9, I=21.0, T=35.0, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:33:07,172] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3250: forward=2601.8ms backward=1004.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:07,518] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:33:14,291] loop.py:45 INFO TRAIN_STEP step=3252 loss=1.819626 Training: 18%|█▊ | 3252/18018 [7:03:00<33:56:44, 8.28s/step, loss=1.820, acc=0.621, acc_len=1.44, thru=17.5, I=18.2, T=33.2, wait=3.3s, pool=60, epoch=1/2][2026-10-18 20:33:21,370] loop.py:45 INFO TRAIN_STEP step=3253 loss=1.600378 Training: 18%|█▊ | 3253/18018 [7:03:07<32:28:16, 7.92s/step, loss=1.600, acc=0.650, acc_len=1.54, thru=18.5, I=17.2, T=32.6, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:27,926] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:33:29,910] loop.py:45 INFO TRAIN_STEP step=3254 loss=1.739014 Training: 18%|█▊ | 3254/18018 [7:03:15<33:14:08, 8.10s/step, loss=1.739, acc=0.626, acc_len=1.46, thru=17.1, I=18.7, T=29.1, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:29,942] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:34,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:33:36,569] loop.py:380 INFO TIMING step=3255: step=4.283s data=2.029s compute=4.251s [fwd=3.052s bwd=1.184s opt=0.003s] dispatch=2.365s [2026-10-18 20:33:36,569] loop.py:45 INFO TRAIN_STEP step=3255 loss=1.672177 Training: 18%|█▊ | 3255/18018 [7:03:22<31:27:26, 7.67s/step, loss=1.672, acc=0.643, acc_len=1.51, thru=16.4, I=24.2, T=29.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:36,668] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:43,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:33:44,051] loop.py:45 INFO TRAIN_STEP step=3256 loss=1.758632 Training: 18%|█▊ | 3256/18018 [7:03:30<31:13:20, 7.61s/step, loss=1.759, acc=0.629, acc_len=1.47, thru=15.2, I=21.2, T=26.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:33:44,045] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3255: forward=3285.1ms backward=1452.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:44,237] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:49,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:33:51,551] loop.py:45 INFO TRAIN_STEP step=3257 loss=1.637332 Training: 18%|█▊ | 3257/18018 [7:03:37<31:04:41, 7.58s/step, loss=1.637, acc=0.648, acc_len=1.52, thru=14.9, I=20.5, T=27.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:51,772] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:57,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:33:59,321] loop.py:45 INFO TRAIN_STEP step=3258 loss=1.771891 Training: 18%|█▊ | 3258/18018 [7:03:45<31:18:40, 7.64s/step, loss=1.772, acc=0.627, acc_len=1.46, thru=17.0, I=20.1, T=32.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:33:59,487] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:07,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:34:08,004] loop.py:45 INFO TRAIN_STEP step=3259 loss=1.712761 Training: 18%|█▊ | 3259/18018 [7:03:54<32:35:44, 7.95s/step, loss=1.713, acc=0.641, acc_len=1.49, thru=15.0, I=15.5, T=30.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:08,250] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:14,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:34:16,495] loop.py:380 INFO TIMING step=3260: step=5.666s data=2.403s compute=5.618s [fwd=4.132s bwd=1.471s opt=0.003s] dispatch=2.815s [2026-10-18 20:34:16,495] loop.py:45 INFO TRAIN_STEP step=3260 loss=1.777554 Training: 18%|█▊ | 3260/18018 [7:04:02<33:15:34, 8.11s/step, loss=1.778, acc=0.626, acc_len=1.45, thru=14.3, I=18.1, T=22.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:16,665] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:23,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:34:23,601] loop.py:45 INFO TRAIN_STEP step=3261 loss=1.778153 Training: 18%|█▊ | 3261/18018 [7:04:09<32:01:06, 7.81s/step, loss=1.778, acc=0.621, acc_len=1.45, thru=16.0, I=19.4, T=30.5, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:34:23,596] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3260: forward=2806.8ms backward=1334.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:23,702] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:29,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:34:30,886] loop.py:45 INFO TRAIN_STEP step=3262 loss=1.732952 Training: 18%|█▊ | 3262/18018 [7:04:16<31:22:01, 7.65s/step, loss=1.733, acc=0.633, acc_len=1.48, thru=17.4, I=20.5, T=30.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:31,203] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:35,926] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:34:37,555] loop.py:45 INFO TRAIN_STEP step=3263 loss=1.739669 Training: 18%|█▊ | 3263/18018 [7:04:23<30:09:27, 7.36s/step, loss=1.740, acc=0.633, acc_len=1.48, thru=17.2, I=24.1, T=33.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:37,939] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:43,114] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:34:45,213] loop.py:45 INFO TRAIN_STEP step=3264 loss=1.659616 Training: 18%|█▊ | 3264/18018 [7:04:31<30:31:20, 7.45s/step, loss=1.660, acc=0.646, acc_len=1.52, thru=14.0, I=22.0, T=28.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:45,625] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 20:34:51,884] loop.py:380 INFO TIMING step=3265: step=3.711s data=1.616s compute=3.686s [fwd=2.590s bwd=1.081s opt=0.003s] dispatch=2.951s [2026-10-18 20:34:51,884] loop.py:45 INFO TRAIN_STEP step=3265 loss=1.739161 Training: 18%|█▊ | 3265/18018 [7:04:37<29:33:59, 7.21s/step, loss=1.739, acc=0.635, acc_len=1.48, thru=14.0, I=19.6, T=34.5, wait=3.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:34:58,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:35:00,113] loop.py:45 INFO TRAIN_STEP step=3266 loss=1.719235 Training: 18%|█▊ | 3266/18018 [7:04:46<30:48:50, 7.52s/step, loss=1.719, acc=0.639, acc_len=1.49, thru=17.2, I=17.3, T=33.5, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:35:00,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3265: forward=2729.1ms backward=1052.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:00,253] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 20:35:10,051] loop.py:45 INFO TRAIN_STEP step=3267 loss=1.620715 Training: 18%|█▊ | 3267/18018 [7:04:56<33:47:03, 8.25s/step, loss=1.621, acc=0.650, acc_len=1.53, thru=11.0, I=11.5, T=28.6, wait=5.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:18,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:35:19,064] loop.py:45 INFO TRAIN_STEP step=3268 loss=1.782614 Training: 18%|█▊ | 3268/18018 [7:05:05<34:43:14, 8.47s/step, loss=1.783, acc=0.627, acc_len=1.45, thru=15.9, I=13.8, T=28.8, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:19,234] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:35:24,859] loop.py:45 INFO TRAIN_STEP step=3269 loss=1.814099 Training: 18%|█▊ | 3269/18018 [7:05:10<31:25:46, 7.67s/step, loss=1.814, acc=0.622, acc_len=1.44, thru=16.1, I=17.8, T=34.6, wait=2.1s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:34,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:35:35,560] loop.py:380 INFO TIMING step=3270: step=4.754s data=2.272s compute=4.709s [fwd=3.242s bwd=1.453s opt=0.003s] dispatch=5.937s [2026-10-18 20:35:35,561] loop.py:45 INFO TRAIN_STEP step=3270 loss=1.710463 Training: 18%|█▊ | 3270/18018 [7:05:21<35:08:58, 8.58s/step, loss=1.710, acc=0.633, acc_len=1.48, thru=15.3, I=16.2, T=26.9, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:35,902] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:42,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:35:42,853] loop.py:45 INFO TRAIN_STEP step=3271 loss=1.722168 Training: 18%|█▊ | 3271/18018 [7:05:28<33:33:53, 8.19s/step, loss=1.722, acc=0.634, acc_len=1.48, thru=16.4, I=17.1, T=29.3, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:35:42,846] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3270: forward=3060.4ms backward=1257.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:43,272] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:49,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:35:49,612] loop.py:45 INFO TRAIN_STEP step=3272 loss=1.615502 Training: 18%|█▊ | 3272/18018 [7:05:35<31:47:55, 7.76s/step, loss=1.616, acc=0.655, acc_len=1.55, thru=18.0, I=20.8, T=28.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:49,935] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:57,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:35:59,378] loop.py:45 INFO TRAIN_STEP step=3273 loss=1.575100 Training: 18%|█▊ | 3273/18018 [7:05:45<34:15:41, 8.36s/step, loss=1.575, acc=0.653, acc_len=1.55, thru=13.3, I=14.6, T=26.2, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:35:59,534] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:05,279] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:36:06,036] loop.py:45 INFO TRAIN_STEP step=3274 loss=1.709867 Training: 18%|█▊ | 3274/18018 [7:05:52<32:09:39, 7.85s/step, loss=1.710, acc=0.639, acc_len=1.49, thru=16.2, I=20.6, T=31.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:06,287] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:10,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:36:12,460] loop.py:380 INFO TIMING step=3275: step=3.836s data=1.709s compute=3.801s [fwd=2.704s bwd=1.082s opt=0.003s] dispatch=2.578s [2026-10-18 20:36:12,460] loop.py:45 INFO TRAIN_STEP step=3275 loss=1.651402 Training: 18%|█▊ | 3275/18018 [7:05:58<30:24:02, 7.42s/step, loss=1.651, acc=0.644, acc_len=1.51, thru=17.2, I=23.9, T=33.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:12,611] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:18,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:36:19,473] loop.py:45 INFO TRAIN_STEP step=3276 loss=1.699909 Training: 18%|█▊ | 3276/18018 [7:06:05<29:53:49, 7.30s/step, loss=1.700, acc=0.640, acc_len=1.50, thru=16.5, I=23.3, T=29.6, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:36:19,468] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3275: forward=3003.7ms backward=1263.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:19,765] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:26,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:36:27,918] loop.py:45 INFO TRAIN_STEP step=3277 loss=1.710272 Training: 18%|█▊ | 3277/18018 [7:06:14<31:17:56, 7.64s/step, loss=1.710, acc=0.638, acc_len=1.49, thru=14.6, I=18.9, T=29.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:28,242] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:34,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:36:35,535] loop.py:45 INFO TRAIN_STEP step=3278 loss=1.775867 Training: 18%|█▊ | 3278/18018 [7:06:21<31:15:56, 7.64s/step, loss=1.776, acc=0.626, acc_len=1.46, thru=15.6, I=19.8, T=34.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:35,839] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:40,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:36:41,660] loop.py:45 INFO TRAIN_STEP step=3279 loss=1.563186 Training: 18%|█▊ | 3279/18018 [7:06:27<29:24:18, 7.18s/step, loss=1.563, acc=0.658, acc_len=1.56, thru=19.9, I=22.7, T=33.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:41,875] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:36:47,480] loop.py:380 INFO TIMING step=3280: step=3.358s data=1.364s compute=3.332s [fwd=2.355s bwd=0.964s opt=0.003s] dispatch=2.453s [2026-10-18 20:36:47,480] loop.py:45 INFO TRAIN_STEP step=3280 loss=1.734244 Training: 18%|█▊ | 3280/18018 [7:06:33<27:44:04, 6.77s/step, loss=1.734, acc=0.634, acc_len=1.47, thru=19.8, I=22.0, T=38.1, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:53,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:36:55,781] loop.py:45 INFO TRAIN_STEP step=3281 loss=1.712897 Training: 18%|█▊ | 3281/18018 [7:06:41<29:36:18, 7.23s/step, loss=1.713, acc=0.634, acc_len=1.48, thru=16.6, I=21.1, T=27.6, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:36:55,777] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3280: forward=3220.3ms backward=1376.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:36:56,023] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:03,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:37:03,572] loop.py:45 INFO TRAIN_STEP step=3282 loss=1.514856 Training: 18%|█▊ | 3282/18018 [7:06:49<30:17:20, 7.40s/step, loss=1.515, acc=0.665, acc_len=1.58, thru=17.0, I=15.6, T=28.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:03,937] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:11,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:37:13,055] loop.py:45 INFO TRAIN_STEP step=3283 loss=1.842433 Training: 18%|█▊ | 3283/18018 [7:06:59<32:50:46, 8.02s/step, loss=1.842, acc=0.619, acc_len=1.43, thru=14.2, I=15.2, T=24.9, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:13,342] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:20,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:37:20,762] loop.py:45 INFO TRAIN_STEP step=3284 loss=1.627395 Training: 18%|█▊ | 3284/18018 [7:07:06<32:27:00, 7.93s/step, loss=1.627, acc=0.653, acc_len=1.54, thru=16.1, I=17.7, T=25.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:20,949] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:37:28,354] loop.py:380 INFO TIMING step=3285: step=4.325s data=2.060s compute=4.274s [fwd=3.003s bwd=1.256s opt=0.003s] dispatch=3.259s [2026-10-18 20:37:28,354] loop.py:45 INFO TRAIN_STEP step=3285 loss=1.780913 Training: 18%|█▊ | 3285/18018 [7:07:14<32:02:01, 7.83s/step, loss=1.781, acc=0.624, acc_len=1.45, thru=15.7, I=15.4, T=29.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:28,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:28,811] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:37:36,116] loop.py:45 INFO TRAIN_STEP step=3286 loss=1.671402 Training: 18%|█▊ | 3286/18018 [7:07:22<31:57:12, 7.81s/step, loss=1.671, acc=0.643, acc_len=1.51, thru=14.5, I=14.2, T=33.9, wait=4.0s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:37:36,111] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3285: forward=2679.7ms backward=1044.5ms [2026-10-18 20:37:46,115] loop.py:45 INFO TRAIN_STEP step=3287 loss=1.700888 Training: 18%|█▊ | 3287/18018 [7:07:32<34:38:26, 8.47s/step, loss=1.701, acc=0.641, acc_len=1.50, thru=14.3, I=14.4, T=31.0, wait=5.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:54,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:37:56,869] loop.py:45 INFO TRAIN_STEP step=3288 loss=1.693957 Training: 18%|█▊ | 3288/18018 [7:07:42<37:26:49, 9.15s/step, loss=1.694, acc=0.641, acc_len=1.50, thru=13.3, I=15.3, T=24.1, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:37:56,918] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:03,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:38:05,918] loop.py:45 INFO TRAIN_STEP step=3289 loss=1.769658 Training: 18%|█▊ | 3289/18018 [7:07:52<37:19:05, 9.12s/step, loss=1.770, acc=0.624, acc_len=1.45, thru=14.2, I=19.0, T=26.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:06,133] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:12,039] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:38:13,366] loop.py:380 INFO TIMING step=3290: step=5.264s data=2.794s compute=5.234s [fwd=3.793s bwd=1.426s opt=0.003s] dispatch=2.175s [2026-10-18 20:38:13,367] loop.py:45 INFO TRAIN_STEP step=3290 loss=1.767653 Training: 18%|█▊ | 3290/18018 [7:07:59<35:15:53, 8.62s/step, loss=1.768, acc=0.627, acc_len=1.46, thru=13.5, I=22.2, T=24.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:13,557] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:19,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:38:20,159] loop.py:45 INFO TRAIN_STEP step=3291 loss=1.775869 Training: 18%|█▊ | 3291/18018 [7:08:06<33:01:06, 8.07s/step, loss=1.776, acc=0.627, acc_len=1.46, thru=15.8, I=18.4, T=31.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:38:20,154] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3290: forward=2843.5ms backward=1157.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:20,295] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:27,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:38:29,675] loop.py:45 INFO TRAIN_STEP step=3292 loss=1.682791 Training: 18%|█▊ | 3292/18018 [7:08:15<34:47:22, 8.50s/step, loss=1.683, acc=0.635, acc_len=1.49, thru=13.7, I=16.0, T=27.8, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:29,990] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:34,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:38:36,672] loop.py:45 INFO TRAIN_STEP step=3293 loss=1.688915 Training: 18%|█▊ | 3293/18018 [7:08:22<32:56:12, 8.05s/step, loss=1.689, acc=0.642, acc_len=1.51, thru=16.2, I=21.8, T=29.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:36,895] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 20:38:44,671] loop.py:45 INFO TRAIN_STEP step=3294 loss=1.634397 Training: 18%|█▊ | 3294/18018 [7:08:30<32:51:58, 8.04s/step, loss=1.634, acc=0.650, acc_len=1.54, thru=13.8, I=17.3, T=28.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:44,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:45,156] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:51,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:38:53,509] loop.py:380 INFO TIMING step=3295: step=4.229s data=1.871s compute=4.189s [fwd=2.963s bwd=1.211s opt=0.003s] dispatch=4.601s [2026-10-18 20:38:53,509] loop.py:45 INFO TRAIN_STEP step=3295 loss=1.681496 Training: 18%|█▊ | 3295/18018 [7:08:39<33:50:58, 8.28s/step, loss=1.681, acc=0.642, acc_len=1.50, thru=14.5, I=18.0, T=30.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:53,803] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:38:59,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:39:00,344] loop.py:45 INFO TRAIN_STEP step=3296 loss=1.736884 Training: 18%|█▊ | 3296/18018 [7:08:46<32:04:45, 7.84s/step, loss=1.737, acc=0.631, acc_len=1.47, thru=16.5, I=23.9, T=32.5, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:39:00,339] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3295: forward=2797.3ms backward=1097.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:00,794] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:07,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:39:07,835] loop.py:45 INFO TRAIN_STEP step=3297 loss=1.752892 Training: 18%|█▊ | 3297/18018 [7:08:53<31:38:34, 7.74s/step, loss=1.753, acc=0.626, acc_len=1.46, thru=15.8, I=18.4, T=30.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:08,164] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:14,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:39:15,308] loop.py:45 INFO TRAIN_STEP step=3298 loss=1.709081 Training: 18%|█▊ | 3298/18018 [7:09:01<31:18:52, 7.66s/step, loss=1.709, acc=0.638, acc_len=1.49, thru=16.6, I=19.5, T=28.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:15,772] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:39:22,137] loop.py:45 INFO TRAIN_STEP step=3299 loss=1.700102 Training: 18%|█▊ | 3299/18018 [7:09:08<30:17:39, 7.41s/step, loss=1.700, acc=0.641, acc_len=1.50, thru=17.2, I=18.2, T=32.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:22,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:22,637] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:28,525] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:39:29,625] loop.py:380 INFO TIMING step=3300: step=4.086s data=1.887s compute=4.056s [fwd=2.905s bwd=1.137s opt=0.003s] dispatch=3.395s [2026-10-18 20:39:29,625] loop.py:45 INFO TRAIN_STEP step=3300 loss=1.617771 Training: 18%|█▊ | 3300/18018 [7:09:15<30:23:23, 7.43s/step, loss=1.618, acc=0.648, acc_len=1.53, thru=17.0, I=18.1, T=31.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:30,034] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:38,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:39:38,836] loop.py:45 INFO TRAIN_STEP step=3301 loss=1.728024 Training: 18%|█▊ | 3301/18018 [7:09:24<32:34:16, 7.97s/step, loss=1.728, acc=0.627, acc_len=1.47, thru=14.6, I=16.8, T=27.4, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:39:38,831] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3300: forward=3193.7ms backward=1431.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:38,872] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:47,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:39:49,003] loop.py:45 INFO TRAIN_STEP step=3302 loss=1.692453 Training: 18%|█▊ | 3302/18018 [7:09:35<35:15:46, 8.63s/step, loss=1.692, acc=0.638, acc_len=1.49, thru=13.3, I=13.6, T=28.0, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:49,068] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:55,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:39:55,675] loop.py:45 INFO TRAIN_STEP step=3303 loss=1.626497 Training: 18%|█▊ | 3303/18018 [7:09:41<32:51:59, 8.04s/step, loss=1.626, acc=0.644, acc_len=1.52, thru=15.8, I=20.4, T=30.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:39:55,907] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:01,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:40:02,579] loop.py:45 INFO TRAIN_STEP step=3304 loss=1.673312 Training: 18%|█▊ | 3304/18018 [7:09:48<31:28:14, 7.70s/step, loss=1.673, acc=0.638, acc_len=1.49, thru=18.2, I=20.6, T=29.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:02,922] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:07,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:40:09,716] loop.py:380 INFO TIMING step=3305: step=4.626s data=2.040s compute=4.596s [fwd=3.160s bwd=1.421s opt=0.003s] dispatch=2.500s [2026-10-18 20:40:09,716] loop.py:45 INFO TRAIN_STEP step=3305 loss=1.653908 Training: 18%|█▊ | 3305/18018 [7:09:55<30:46:33, 7.53s/step, loss=1.654, acc=0.649, acc_len=1.53, thru=16.9, I=22.6, T=27.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:10,184] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:16,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:40:18,023] loop.py:45 INFO TRAIN_STEP step=3306 loss=1.557772 Training: 18%|█▊ | 3306/18018 [7:10:04<31:43:35, 7.76s/step, loss=1.558, acc=0.658, acc_len=1.56, thru=16.3, I=19.6, T=24.6, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:40:18,018] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3305: forward=3554.5ms backward=1606.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:18,298] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:24,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:40:25,906] loop.py:45 INFO TRAIN_STEP step=3307 loss=1.787991 Training: 18%|█▊ | 3307/18018 [7:10:11<31:52:23, 7.80s/step, loss=1.788, acc=0.623, acc_len=1.45, thru=14.5, I=19.5, T=26.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:26,107] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:32,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:40:34,587] loop.py:45 INFO TRAIN_STEP step=3308 loss=1.671818 Training: 18%|█▊ | 3308/18018 [7:10:20<32:57:03, 8.06s/step, loss=1.672, acc=0.644, acc_len=1.51, thru=13.2, I=21.8, T=23.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:34,620] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:42,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:40:43,806] loop.py:45 INFO TRAIN_STEP step=3309 loss=1.830973 Training: 18%|█▊ | 3309/18018 [7:10:29<34:21:45, 8.41s/step, loss=1.831, acc=0.617, acc_len=1.42, thru=14.0, I=15.3, T=31.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:43,976] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:49,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:40:50,136] loop.py:380 INFO TIMING step=3310: step=3.699s data=1.606s compute=3.670s [fwd=2.605s bwd=1.050s opt=0.003s] dispatch=2.621s [2026-10-18 20:40:50,136] loop.py:45 INFO TRAIN_STEP step=3310 loss=1.702277 Training: 18%|█▊ | 3310/18018 [7:10:36<31:48:44, 7.79s/step, loss=1.702, acc=0.639, acc_len=1.50, thru=19.0, I=22.3, T=34.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:50,203] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:55,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:40:56,999] loop.py:45 INFO TRAIN_STEP step=3311 loss=1.527017 Training: 18%|█▊ | 3311/18018 [7:10:43<30:40:39, 7.51s/step, loss=1.527, acc=0.664, acc_len=1.58, thru=18.5, I=23.2, T=30.0, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:40:56,994] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3310: forward=2969.0ms backward=1230.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:40:57,476] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:02,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:41:03,428] loop.py:45 INFO TRAIN_STEP step=3312 loss=1.688526 Training: 18%|█▊ | 3312/18018 [7:10:49<29:21:01, 7.18s/step, loss=1.689, acc=0.641, acc_len=1.50, thru=17.7, I=24.0, T=31.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:03,477] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:11,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:41:12,423] loop.py:45 INFO TRAIN_STEP step=3313 loss=1.747582 Training: 18%|█▊ | 3313/18018 [7:10:58<31:34:03, 7.73s/step, loss=1.748, acc=0.626, acc_len=1.46, thru=13.6, I=16.5, T=29.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:12,551] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:19,866] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:41:21,289] loop.py:45 INFO TRAIN_STEP step=3314 loss=1.719560 Training: 18%|█▊ | 3314/18018 [7:11:07<32:57:26, 8.07s/step, loss=1.720, acc=0.638, acc_len=1.50, thru=14.7, I=15.4, T=27.1, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:21,386] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:28,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:41:29,761] loop.py:380 INFO TIMING step=3315: step=4.159s data=1.844s compute=4.130s [fwd=2.970s bwd=1.145s opt=0.003s] dispatch=4.305s [2026-10-18 20:41:29,761] loop.py:45 INFO TRAIN_STEP step=3315 loss=1.744701 Training: 18%|█▊ | 3315/18018 [7:11:15<33:27:06, 8.19s/step, loss=1.745, acc=0.639, acc_len=1.49, thru=13.3, I=18.7, T=30.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:30,155] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:35,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:41:36,138] loop.py:45 INFO TRAIN_STEP step=3316 loss=1.579543 Training: 18%|█▊ | 3316/18018 [7:11:22<31:13:32, 7.65s/step, loss=1.580, acc=0.655, acc_len=1.55, thru=18.5, I=25.5, T=33.3, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:41:36,133] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3315: forward=2674.0ms backward=1116.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:36,434] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:44,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:41:45,228] loop.py:45 INFO TRAIN_STEP step=3317 loss=1.676639 Training: 18%|█▊ | 3317/18018 [7:11:31<32:59:28, 8.08s/step, loss=1.677, acc=0.642, acc_len=1.51, thru=13.6, I=15.1, T=26.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:41:45,728] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:41:52,099] loop.py:45 INFO TRAIN_STEP step=3318 loss=1.589940 Training: 18%|█▊ | 3318/18018 [7:11:38<31:30:40, 7.72s/step, loss=1.590, acc=0.655, acc_len=1.55, thru=16.4, I=16.8, T=33.2, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:01,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:42:03,513] loop.py:45 INFO TRAIN_STEP step=3319 loss=1.860674 Training: 18%|█▊ | 3319/18018 [7:11:49<36:02:08, 8.83s/step, loss=1.861, acc=0.614, acc_len=1.42, thru=11.2, I=13.0, T=25.8, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:04,011] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:11,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:42:13,146] loop.py:380 INFO TIMING step=3320: step=4.068s data=1.710s compute=3.905s [fwd=2.732s bwd=1.157s opt=0.003s] dispatch=5.557s [2026-10-18 20:42:13,146] loop.py:45 INFO TRAIN_STEP step=3320 loss=1.653782 Training: 18%|█▊ | 3320/18018 [7:11:59<37:01:29, 9.07s/step, loss=1.654, acc=0.644, acc_len=1.51, thru=14.0, I=16.6, T=31.5, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:13,606] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 20:42:20,209] loop.py:45 INFO TRAIN_STEP step=3321 loss=1.783820 Training: 18%|█▊ | 3321/18018 [7:12:06<34:33:58, 8.47s/step, loss=1.784, acc=0.624, acc_len=1.44, thru=15.4, I=20.8, T=32.3, wait=3.1s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:42:20,204] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3320: forward=2773.3ms backward=1133.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:28,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:42:29,144] loop.py:45 INFO TRAIN_STEP step=3322 loss=1.686455 Training: 18%|█▊ | 3322/18018 [7:12:15<35:08:05, 8.61s/step, loss=1.686, acc=0.640, acc_len=1.50, thru=15.9, I=16.3, T=28.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:29,303] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:42:35,871] loop.py:45 INFO TRAIN_STEP step=3323 loss=1.639331 Training: 18%|█▊ | 3323/18018 [7:12:21<32:49:56, 8.04s/step, loss=1.639, acc=0.650, acc_len=1.53, thru=18.1, I=19.3, T=28.1, wait=2.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:46,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:42:47,442] loop.py:45 INFO TRAIN_STEP step=3324 loss=1.913003 Training: 18%|█▊ | 3324/18018 [7:12:33<37:09:00, 9.10s/step, loss=1.913, acc=0.602, acc_len=1.38, thru=12.1, I=12.1, T=24.2, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:47,630] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:51,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:42:53,711] loop.py:380 INFO TIMING step=3325: step=4.214s data=1.907s compute=4.178s [fwd=2.871s bwd=1.293s opt=0.003s] dispatch=2.044s [2026-10-18 20:42:53,712] loop.py:45 INFO TRAIN_STEP step=3325 loss=1.619797 Training: 18%|█▊ | 3325/18018 [7:12:39<33:40:39, 8.25s/step, loss=1.620, acc=0.653, acc_len=1.54, thru=15.9, I=22.6, T=30.4, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:42:54,056] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:43:00,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:43:01,101] loop.py:45 INFO TRAIN_STEP step=3326 loss=1.761383 Training: 18%|█▊ | 3326/18018 [7:12:47<32:37:09, 7.99s/step, loss=1.761, acc=0.630, acc_len=1.46, thru=14.7, I=21.6, T=30.4, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:43:01,096] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3325: forward=2823.0ms backward=1338.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:43:01,271] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:43:08,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:43:10,135] loop.py:45 INFO TRAIN_STEP step=3327 loss=1.819446 Training: 18%|█▊ | 3327/18018 [7:12:56<33:53:34, 8.31s/step, loss=1.819, acc=0.621, acc_len=1.43, thru=13.4, I=17.3, T=27.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:43:10,293] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:43:18,438] loop.py:45 INFO TRAIN_STEP step=3328 loss=1.632970 Training: 18%|█▊ | 3328/18018 [7:13:04<33:53:13, 8.30s/step, loss=1.633, acc=0.647, acc_len=1.51, thru=11.6, I=15.9, T=30.0, wait=4.0s, pool=40, epoch=1/2][2026-10-18 20:43:27,451] loop.py:45 INFO TRAIN_STEP step=3329 loss=1.705601 Training: 18%|█▊ | 3329/18018 [7:13:13<34:45:05, 8.52s/step, loss=1.706, acc=0.638, acc_len=1.49, thru=15.8, I=15.3, T=28.3, wait=4.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:43:35,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:43:36,944] loop.py:380 INFO TIMING step=3330: step=4.125s data=1.877s compute=4.094s [fwd=2.951s bwd=1.128s opt=0.003s] dispatch=5.360s [2026-10-18 20:43:36,944] loop.py:45 INFO TRAIN_STEP step=3330 loss=1.680916 Training: 18%|█▊ | 3330/18018 [7:13:23<35:56:50, 8.81s/step, loss=1.681, acc=0.643, acc_len=1.51, thru=14.1, I=13.2, T=31.0, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:43:36,988] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:43:43,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:43:44,924] loop.py:45 INFO TRAIN_STEP step=3331 loss=1.694675 Training: 18%|█▊ | 3331/18018 [7:13:31<34:55:42, 8.56s/step, loss=1.695, acc=0.644, acc_len=1.50, thru=16.1, I=22.1, T=30.8, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:43:44,919] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3330: forward=2907.4ms backward=1208.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:43:45,265] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 20:43:51,876] loop.py:45 INFO TRAIN_STEP step=3332 loss=1.666680 Training: 18%|█▊ | 3332/18018 [7:13:37<32:57:13, 8.08s/step, loss=1.667, acc=0.644, acc_len=1.51, thru=16.0, I=20.0, T=32.4, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:43:59,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:44:00,694] loop.py:45 INFO TRAIN_STEP step=3333 loss=1.776319 Training: 18%|█▊ | 3333/18018 [7:13:46<33:51:18, 8.30s/step, loss=1.776, acc=0.629, acc_len=1.46, thru=15.7, I=16.4, T=23.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:44:01,197] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:44:06,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:44:08,208] loop.py:45 INFO TRAIN_STEP step=3334 loss=1.648746 Training: 19%|█▊ | 3334/18018 [7:13:54<32:53:34, 8.06s/step, loss=1.649, acc=0.650, acc_len=1.53, thru=14.8, I=21.6, T=27.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:44:08,665] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:44:16,306] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:44:17,740] loop.py:380 INFO TIMING step=3335: step=4.463s data=2.080s compute=4.421s [fwd=3.093s bwd=1.313s opt=0.003s] dispatch=5.060s [2026-10-18 20:44:17,740] loop.py:45 INFO TRAIN_STEP step=3335 loss=1.776772 Training: 19%|█▊ | 3335/18018 [7:14:03<34:41:14, 8.50s/step, loss=1.777, acc=0.624, acc_len=1.45, thru=14.1, I=16.1, T=28.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:44:17,824] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 20:44:24,323] loop.py:45 INFO TRAIN_STEP step=3336 loss=1.718984 Training: 19%|█▊ | 3336/18018 [7:14:10<32:20:09, 7.93s/step, loss=1.719, acc=0.631, acc_len=1.47, thru=17.4, I=21.4, T=31.3, wait=2.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:44:24,319] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3335: forward=2844.2ms backward=1185.0ms [2026-10-18 20:44:31,276] loop.py:45 INFO TRAIN_STEP step=3337 loss=1.669692 Training: 19%|█▊ | 3337/18018 [7:14:17<31:08:13, 7.64s/step, loss=1.670, acc=0.640, acc_len=1.50, thru=17.2, I=17.4, T=31.8, wait=2.9s, pool=56, epoch=1/2][2026-10-18 20:44:40,922] loop.py:45 INFO TRAIN_STEP step=3338 loss=1.780233 Training: 19%|█▊ | 3338/18018 [7:14:27<33:35:39, 8.24s/step, loss=1.780, acc=0.622, acc_len=1.45, thru=12.2, I=11.3, T=27.6, wait=5.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:44:48,974] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:44:50,717] loop.py:45 INFO TRAIN_STEP step=3339 loss=1.771219 Training: 19%|█▊ | 3339/18018 [7:14:36<35:29:56, 8.71s/step, loss=1.771, acc=0.623, acc_len=1.45, thru=15.5, I=18.5, T=27.3, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:44:50,993] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:44:57,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:44:58,026] loop.py:380 INFO TIMING step=3340: step=4.477s data=2.079s compute=4.446s [fwd=3.105s bwd=1.326s opt=0.003s] dispatch=2.821s [2026-10-18 20:44:58,026] loop.py:45 INFO TRAIN_STEP step=3340 loss=1.689951 Training: 19%|█▊ | 3340/18018 [7:14:44<33:47:08, 8.29s/step, loss=1.690, acc=0.637, acc_len=1.48, thru=16.2, I=20.1, T=28.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:44:58,427] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:02,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:45:04,357] loop.py:45 INFO TRAIN_STEP step=3341 loss=1.665164 Training: 19%|█▊ | 3341/18018 [7:14:50<31:23:37, 7.70s/step, loss=1.665, acc=0.647, acc_len=1.51, thru=17.9, I=21.6, T=32.2, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:45:04,352] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3340: forward=2833.8ms backward=1099.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:04,449] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 20:45:13,830] loop.py:45 INFO TRAIN_STEP step=3342 loss=1.701900 Training: 19%|█▊ | 3342/18018 [7:14:59<33:33:36, 8.23s/step, loss=1.702, acc=0.636, acc_len=1.49, thru=11.1, I=10.7, T=33.6, wait=5.7s, pool=40, epoch=1/2][2026-10-18 20:45:21,288] loop.py:45 INFO TRAIN_STEP step=3343 loss=1.733397 Training: 19%|█▊ | 3343/18018 [7:15:07<32:36:36, 8.00s/step, loss=1.733, acc=0.633, acc_len=1.48, thru=17.2, I=13.6, T=31.8, wait=3.4s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:28,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:45:29,439] loop.py:45 INFO TRAIN_STEP step=3344 loss=1.759906 Training: 19%|█▊ | 3344/18018 [7:15:15<32:47:35, 8.05s/step, loss=1.760, acc=0.628, acc_len=1.47, thru=18.6, I=16.3, T=31.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:29,709] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:35,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:45:36,141] loop.py:380 INFO TIMING step=3345: step=3.618s data=1.498s compute=3.591s [fwd=2.536s bwd=1.040s opt=0.003s] dispatch=3.074s [2026-10-18 20:45:36,142] loop.py:45 INFO TRAIN_STEP step=3345 loss=1.673623 Training: 19%|█▊ | 3345/18018 [7:15:22<31:09:00, 7.64s/step, loss=1.674, acc=0.645, acc_len=1.51, thru=18.6, I=20.1, T=35.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:36,389] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:41,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:45:43,333] loop.py:45 INFO TRAIN_STEP step=3346 loss=1.706872 Training: 19%|█▊ | 3346/18018 [7:15:29<30:35:47, 7.51s/step, loss=1.707, acc=0.637, acc_len=1.49, thru=18.3, I=21.8, T=29.5, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:45:43,328] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3345: forward=2954.6ms backward=1342.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:43,429] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:48,706] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:45:50,040] loop.py:45 INFO TRAIN_STEP step=3347 loss=1.763355 Training: 19%|█▊ | 3347/18018 [7:15:36<29:36:54, 7.27s/step, loss=1.763, acc=0.626, acc_len=1.45, thru=16.9, I=21.8, T=29.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:50,216] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:55,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:45:56,687] loop.py:45 INFO TRAIN_STEP step=3348 loss=1.719417 Training: 19%|█▊ | 3348/18018 [7:15:42<28:51:17, 7.08s/step, loss=1.719, acc=0.634, acc_len=1.48, thru=17.7, I=23.2, T=31.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:45:57,033] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:02,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:46:04,045] loop.py:45 INFO TRAIN_STEP step=3349 loss=1.818679 Training: 19%|█▊ | 3349/18018 [7:15:50<29:11:22, 7.16s/step, loss=1.819, acc=0.619, acc_len=1.43, thru=15.6, I=22.9, T=31.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:04,073] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 20:46:10,130] loop.py:380 INFO TIMING step=3350: step=3.754s data=1.719s compute=3.721s [fwd=2.669s bwd=1.039s opt=0.003s] dispatch=2.322s [2026-10-18 20:46:10,131] loop.py:45 INFO TRAIN_STEP step=3350 loss=1.692708 Training: 19%|█▊ | 3350/18018 [7:15:56<27:52:14, 6.84s/step, loss=1.693, acc=0.642, acc_len=1.50, thru=17.7, I=20.8, T=34.1, wait=2.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:17,198] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:46:18,891] loop.py:45 INFO TRAIN_STEP step=3351 loss=1.740072 Training: 19%|█▊ | 3351/18018 [7:16:04<30:12:43, 7.42s/step, loss=1.740, acc=0.635, acc_len=1.48, thru=16.9, I=17.5, T=29.4, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:46:18,886] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3350: forward=3021.0ms backward=1273.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:19,204] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:25,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:46:26,535] loop.py:45 INFO TRAIN_STEP step=3352 loss=1.629331 Training: 19%|█▊ | 3352/18018 [7:16:12<30:29:32, 7.48s/step, loss=1.629, acc=0.653, acc_len=1.54, thru=14.5, I=19.8, T=28.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:26,840] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:46:34,446] loop.py:45 INFO TRAIN_STEP step=3353 loss=1.684669 Training: 19%|█▊ | 3353/18018 [7:16:20<31:00:35, 7.61s/step, loss=1.685, acc=0.641, acc_len=1.50, thru=15.1, I=17.8, T=26.4, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:41,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:46:42,730] loop.py:45 INFO TRAIN_STEP step=3354 loss=1.518231 Training: 19%|█▊ | 3354/18018 [7:16:28<31:49:56, 7.81s/step, loss=1.518, acc=0.666, acc_len=1.59, thru=15.9, I=17.9, T=30.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:42,790] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 20:46:50,638] loop.py:380 INFO TIMING step=3355: step=4.190s data=1.823s compute=4.153s [fwd=2.875s bwd=1.264s opt=0.003s] dispatch=3.707s [2026-10-18 20:46:50,638] loop.py:45 INFO TRAIN_STEP step=3355 loss=1.700328 Training: 19%|█▊ | 3355/18018 [7:16:36<31:56:23, 7.84s/step, loss=1.700, acc=0.641, acc_len=1.50, thru=14.1, I=16.6, T=30.6, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:57,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:46:58,212] loop.py:45 INFO TRAIN_STEP step=3356 loss=1.701338 Training: 19%|█▊ | 3356/18018 [7:16:44<31:36:53, 7.76s/step, loss=1.701, acc=0.637, acc_len=1.50, thru=18.4, I=17.2, T=32.9, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:46:58,207] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3355: forward=2803.7ms backward=1043.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:46:58,285] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:05,680] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:47:06,195] loop.py:45 INFO TRAIN_STEP step=3357 loss=1.674079 Training: 19%|█▊ | 3357/18018 [7:16:52<31:52:42, 7.83s/step, loss=1.674, acc=0.637, acc_len=1.49, thru=16.8, I=17.4, T=30.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:06,685] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:47:13,271] loop.py:45 INFO TRAIN_STEP step=3358 loss=1.641038 Training: 19%|█▊ | 3358/18018 [7:16:59<30:57:29, 7.60s/step, loss=1.641, acc=0.651, acc_len=1.53, thru=15.3, I=17.4, T=33.8, wait=3.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:24,076] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:47:24,602] loop.py:45 INFO TRAIN_STEP step=3359 loss=1.660527 Training: 19%|█▊ | 3359/18018 [7:17:10<35:30:41, 8.72s/step, loss=1.661, acc=0.645, acc_len=1.51, thru=12.9, I=13.6, T=28.8, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:25,080] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:31,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:47:32,091] loop.py:380 INFO TIMING step=3360: step=3.729s data=1.565s compute=3.687s [fwd=2.593s bwd=1.079s opt=0.003s] dispatch=3.751s [2026-10-18 20:47:32,092] loop.py:45 INFO TRAIN_STEP step=3360 loss=1.778148 Training: 19%|█▊ | 3360/18018 [7:17:18<34:00:24, 8.35s/step, loss=1.778, acc=0.629, acc_len=1.46, thru=15.9, I=17.8, T=34.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:32,309] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:38,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:47:39,496] loop.py:45 INFO TRAIN_STEP step=3361 loss=1.790990 Training: 19%|█▊ | 3361/18018 [7:17:25<32:50:42, 8.07s/step, loss=1.791, acc=0.614, acc_len=1.42, thru=16.4, I=19.9, T=29.8, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:47:39,491] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3360: forward=2923.7ms backward=1325.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:39,785] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:47,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:47:49,901] loop.py:45 INFO TRAIN_STEP step=3362 loss=1.707907 Training: 19%|█▊ | 3362/18018 [7:17:35<35:41:52, 8.77s/step, loss=1.708, acc=0.638, acc_len=1.49, thru=12.3, I=15.6, T=24.7, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:47:50,168] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 20:47:58,453] loop.py:45 INFO TRAIN_STEP step=3363 loss=1.755932 Training: 19%|█▊ | 3363/18018 [7:17:44<35:26:04, 8.70s/step, loss=1.756, acc=0.625, acc_len=1.45, thru=15.0, I=13.9, T=30.9, wait=4.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:02,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:48:04,743] loop.py:45 INFO TRAIN_STEP step=3364 loss=1.732293 Training: 19%|█▊ | 3364/18018 [7:17:50<32:28:57, 7.98s/step, loss=1.732, acc=0.629, acc_len=1.47, thru=19.4, I=20.4, T=30.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:04,928] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 20:48:11,290] loop.py:380 INFO TIMING step=3365: step=4.012s data=1.847s compute=3.989s [fwd=2.883s bwd=1.094s opt=0.003s] dispatch=2.524s [2026-10-18 20:48:11,290] loop.py:45 INFO TRAIN_STEP step=3365 loss=1.604998 Training: 19%|█▊ | 3365/18018 [7:17:57<30:43:35, 7.55s/step, loss=1.605, acc=0.654, acc_len=1.55, thru=16.4, I=19.0, T=31.9, wait=2.5s, pool=56, epoch=1/2][2026-10-18 20:48:18,548] loop.py:45 INFO TRAIN_STEP step=3366 loss=1.747144 Training: 19%|█▊ | 3366/18018 [7:18:04<30:22:18, 7.46s/step, loss=1.747, acc=0.630, acc_len=1.47, thru=18.2, I=16.8, T=33.8, wait=3.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:48:18,544] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3365: forward=2694.5ms backward=1049.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:23,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:48:25,841] loop.py:45 INFO TRAIN_STEP step=3367 loss=1.704456 Training: 19%|█▊ | 3367/18018 [7:18:11<30:09:48, 7.41s/step, loss=1.704, acc=0.636, acc_len=1.49, thru=18.2, I=20.7, T=28.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:26,333] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:31,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:48:32,422] loop.py:45 INFO TRAIN_STEP step=3368 loss=1.591773 Training: 19%|█▊ | 3368/18018 [7:18:18<29:08:49, 7.16s/step, loss=1.592, acc=0.653, acc_len=1.54, thru=16.8, I=25.1, T=35.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:32,808] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:48:39,056] loop.py:45 INFO TRAIN_STEP step=3369 loss=1.652224 Training: 19%|█▊ | 3369/18018 [7:18:25<28:29:50, 7.00s/step, loss=1.652, acc=0.647, acc_len=1.52, thru=19.1, I=21.6, T=34.2, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:44,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:48:46,126] loop.py:380 INFO TIMING step=3370: step=4.390s data=2.115s compute=4.352s [fwd=3.024s bwd=1.313s opt=0.003s] dispatch=2.672s [2026-10-18 20:48:46,126] loop.py:45 INFO TRAIN_STEP step=3370 loss=1.469465 Training: 19%|█▊ | 3370/18018 [7:18:32<28:34:50, 7.02s/step, loss=1.469, acc=0.675, acc_len=1.61, thru=18.7, I=21.6, T=29.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:46,262] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:53,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:48:54,175] loop.py:45 INFO TRAIN_STEP step=3371 loss=1.709330 (TrainerActor pid=9310) [2026-10-18 20:48:54,170] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3370: forward=2756.3ms backward=1085.0ms Training: 19%|█▊ | 3371/18018 [7:18:40<29:49:37, 7.33s/step, loss=1.709, acc=0.637, acc_len=1.49, thru=13.3, I=18.4, T=32.8, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:48:54,631] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:00,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:49:01,958] loop.py:45 INFO TRAIN_STEP step=3372 loss=1.674220 Training: 19%|█▊ | 3372/18018 [7:18:48<30:22:43, 7.47s/step, loss=1.674, acc=0.643, acc_len=1.51, thru=16.2, I=21.0, T=27.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:02,254] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:06,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:49:07,956] loop.py:45 INFO TRAIN_STEP step=3373 loss=1.698114 Training: 19%|█▊ | 3373/18018 [7:18:54<28:34:47, 7.03s/step, loss=1.698, acc=0.638, acc_len=1.50, thru=18.8, I=24.3, T=34.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:08,035] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 20:49:15,049] loop.py:45 INFO TRAIN_STEP step=3374 loss=1.702068 Training: 19%|█▊ | 3374/18018 [7:19:01<28:39:58, 7.05s/step, loss=1.702, acc=0.640, acc_len=1.49, thru=15.8, I=17.7, T=28.2, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:19,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 20:49:22,374] loop.py:380 INFO TIMING step=3375: step=4.453s data=2.170s compute=4.408s [fwd=3.030s bwd=1.364s opt=0.003s] dispatch=2.858s [2026-10-18 20:49:22,374] loop.py:45 INFO TRAIN_STEP step=3375 loss=1.679139 Training: 19%|█▊ | 3375/18018 [7:19:08<28:59:59, 7.13s/step, loss=1.679, acc=0.640, acc_len=1.49, thru=18.3, I=20.9, T=28.7, wait=2.9s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:22,766] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:27,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:49:28,830] loop.py:45 INFO TRAIN_STEP step=3376 loss=1.711984 Training: 19%|█▊ | 3376/18018 [7:19:14<28:10:39, 6.93s/step, loss=1.712, acc=0.633, acc_len=1.47, thru=16.9, I=27.2, T=31.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:28,870] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 20:49:28,824] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3375: forward=2831.0ms backward=1238.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:33,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:49:35,271] loop.py:45 INFO TRAIN_STEP step=3377 loss=1.681496 Training: 19%|█▊ | 3377/18018 [7:19:21<27:34:51, 6.78s/step, loss=1.681, acc=0.640, acc_len=1.50, thru=18.6, I=22.9, T=28.5, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:35,599] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:40,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:49:42,137] loop.py:45 INFO TRAIN_STEP step=3378 loss=1.754325 Training: 19%|█▊ | 3378/18018 [7:19:28<27:40:53, 6.81s/step, loss=1.754, acc=0.630, acc_len=1.48, thru=15.7, I=21.9, T=29.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:42,469] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:49,538] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:49:51,170] loop.py:45 INFO TRAIN_STEP step=3379 loss=1.612720 Training: 19%|█▉ | 3379/18018 [7:19:37<30:23:32, 7.47s/step, loss=1.613, acc=0.654, acc_len=1.55, thru=14.7, I=16.1, T=28.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:51,544] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:56,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:49:58,575] loop.py:380 INFO TIMING step=3380: step=4.389s data=2.010s compute=4.357s [fwd=3.118s bwd=1.225s opt=0.003s] dispatch=3.008s [2026-10-18 20:49:58,575] loop.py:45 INFO TRAIN_STEP step=3380 loss=1.665624 Training: 19%|█▉ | 3380/18018 [7:19:44<30:18:22, 7.45s/step, loss=1.666, acc=0.641, acc_len=1.49, thru=16.0, I=22.7, T=29.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:49:59,000] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:03,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:50:05,827] loop.py:45 INFO TRAIN_STEP step=3381 loss=1.803177 Training: 19%|█▉ | 3381/18018 [7:19:51<30:03:33, 7.39s/step, loss=1.803, acc=0.619, acc_len=1.43, thru=15.4, I=27.1, T=29.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:05,856] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 20:50:05,821] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3380: forward=3106.6ms backward=1239.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:13,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:50:14,977] loop.py:45 INFO TRAIN_STEP step=3382 loss=1.776148 Training: 19%|█▉ | 3382/18018 [7:20:01<32:12:09, 7.92s/step, loss=1.776, acc=0.619, acc_len=1.44, thru=14.1, I=17.9, T=28.0, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:15,055] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:20,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:50:22,625] loop.py:45 INFO TRAIN_STEP step=3383 loss=1.799083 Training: 19%|█▉ | 3383/18018 [7:20:08<31:52:03, 7.84s/step, loss=1.799, acc=0.625, acc_len=1.45, thru=14.1, I=22.9, T=30.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:22,761] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:30,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:50:31,494] loop.py:45 INFO TRAIN_STEP step=3384 loss=1.846210 Training: 19%|█▉ | 3384/18018 [7:20:17<33:07:19, 8.15s/step, loss=1.846, acc=0.621, acc_len=1.44, thru=14.7, I=17.4, T=28.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:31,611] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:36,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:50:38,372] loop.py:380 INFO TIMING step=3385: step=4.709s data=2.336s compute=4.663s [fwd=3.304s bwd=1.344s opt=0.003s] dispatch=2.159s [2026-10-18 20:50:38,372] loop.py:45 INFO TRAIN_STEP step=3385 loss=1.706543 Training: 19%|█▉ | 3385/18018 [7:20:24<31:34:10, 7.77s/step, loss=1.707, acc=0.637, acc_len=1.49, thru=16.9, I=24.7, T=27.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:38,728] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (TrainerActor pid=9310) [2026-10-18 20:50:44,474] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3385: forward=2684.7ms backward=1114.7ms [2026-10-18 20:50:44,478] loop.py:45 INFO TRAIN_STEP step=3386 loss=1.645989 Training: 19%|█▉ | 3386/18018 [7:20:30<29:32:19, 7.27s/step, loss=1.646, acc=0.645, acc_len=1.51, thru=15.6, I=19.9, T=33.2, wait=2.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:51,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:50:53,436] loop.py:45 INFO TRAIN_STEP step=3387 loss=1.788554 Training: 19%|█▉ | 3387/18018 [7:20:39<31:35:58, 7.78s/step, loss=1.789, acc=0.628, acc_len=1.45, thru=16.6, I=19.5, T=30.6, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:53,833] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:50:58,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:51:00,329] loop.py:45 INFO TRAIN_STEP step=3388 loss=1.553120 Training: 19%|█▉ | 3388/18018 [7:20:46<30:31:27, 7.51s/step, loss=1.553, acc=0.659, acc_len=1.56, thru=16.8, I=24.5, T=30.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:00,436] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:08,108] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:51:08,325] loop.py:45 INFO TRAIN_STEP step=3389 loss=1.597674 Training: 19%|█▉ | 3389/18018 [7:20:54<31:06:43, 7.66s/step, loss=1.598, acc=0.658, acc_len=1.56, thru=14.1, I=17.9, T=28.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:08,609] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:14,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:51:15,419] loop.py:380 INFO TIMING step=3390: step=3.999s data=1.844s compute=3.968s [fwd=2.841s bwd=1.111s opt=0.003s] dispatch=3.086s [2026-10-18 20:51:15,420] loop.py:45 INFO TRAIN_STEP step=3390 loss=1.691041 Training: 19%|█▉ | 3390/18018 [7:21:01<30:25:32, 7.49s/step, loss=1.691, acc=0.637, acc_len=1.48, thru=16.1, I=17.2, T=32.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:15,668] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:21,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:51:21,683] loop.py:45 INFO TRAIN_STEP step=3391 loss=1.627507 Training: 19%|█▉ | 3391/18018 [7:21:07<28:55:43, 7.12s/step, loss=1.628, acc=0.651, acc_len=1.52, thru=19.4, I=19.7, T=35.5, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:51:21,678] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3390: forward=2585.0ms backward=981.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:22,108] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:29,402] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:51:31,325] loop.py:45 INFO TRAIN_STEP step=3392 loss=1.644576 Training: 19%|█▉ | 3392/18018 [7:21:17<32:00:17, 7.88s/step, loss=1.645, acc=0.650, acc_len=1.53, thru=13.9, I=16.0, T=28.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:31,421] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:36,441] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:51:38,024] loop.py:45 INFO TRAIN_STEP step=3393 loss=1.691721 Training: 19%|█▉ | 3393/18018 [7:21:24<30:33:55, 7.52s/step, loss=1.692, acc=0.640, acc_len=1.50, thru=16.0, I=21.7, T=30.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:38,456] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:43,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:51:45,209] loop.py:45 INFO TRAIN_STEP step=3394 loss=1.693315 Training: 19%|█▉ | 3394/18018 [7:21:31<30:09:00, 7.42s/step, loss=1.693, acc=0.638, acc_len=1.50, thru=16.6, I=20.8, T=29.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:51:45,323] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-18 20:51:51,171] loop.py:380 INFO TIMING step=3395: step=3.710s data=1.677s compute=3.674s [fwd=2.614s bwd=1.046s opt=0.003s] dispatch=2.242s [2026-10-18 20:51:51,172] loop.py:45 INFO TRAIN_STEP step=3395 loss=1.773064 Training: 19%|█▉ | 3395/18018 [7:21:37<28:22:11, 6.98s/step, loss=1.773, acc=0.624, acc_len=1.45, thru=16.9, I=23.5, T=34.5, wait=2.2s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:00,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:52:01,163] loop.py:45 INFO TRAIN_STEP step=3396 loss=1.739941 Training: 19%|█▉ | 3396/18018 [7:21:47<32:01:50, 7.89s/step, loss=1.740, acc=0.631, acc_len=1.47, thru=14.9, I=13.6, T=30.1, wait=5.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:52:01,158] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3395: forward=2990.0ms backward=1209.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:01,370] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:52:08,998] loop.py:45 INFO TRAIN_STEP step=3397 loss=1.846066 Training: 19%|█▉ | 3397/18018 [7:21:55<31:57:53, 7.87s/step, loss=1.846, acc=0.616, acc_len=1.42, thru=14.6, I=15.7, T=34.4, wait=4.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:17,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:52:18,529] loop.py:45 INFO TRAIN_STEP step=3398 loss=1.694142 Training: 19%|█▉ | 3398/18018 [7:22:04<33:59:20, 8.37s/step, loss=1.694, acc=0.639, acc_len=1.49, thru=15.9, I=17.2, T=30.7, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:18,734] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:27,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:52:27,224] loop.py:45 INFO TRAIN_STEP step=3399 loss=1.608417 Training: 19%|█▉ | 3399/18018 [7:22:13<34:23:00, 8.47s/step, loss=1.608, acc=0.657, acc_len=1.55, thru=15.1, I=15.1, T=27.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:27,652] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:33,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:52:34,756] loop.py:380 INFO TIMING step=3400: step=4.126s data=1.782s compute=4.093s [fwd=2.858s bwd=1.220s opt=0.003s] dispatch=3.396s [2026-10-18 20:52:34,756] loop.py:45 INFO TRAIN_STEP step=3400 loss=1.661856 Training: 19%|█▉ | 3400/18018 [7:22:20<33:14:17, 8.19s/step, loss=1.662, acc=0.643, acc_len=1.51, thru=14.8, I=17.6, T=31.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:35,253] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:42,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:52:42,908] loop.py:45 INFO TRAIN_STEP step=3401 loss=1.699276 Training: 19%|█▉ | 3401/18018 [7:22:28<33:11:45, 8.18s/step, loss=1.699, acc=0.641, acc_len=1.50, thru=14.3, I=19.4, T=29.3, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:52:42,903] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3400: forward=3079.2ms backward=1238.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:52:43,381] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:52:53,830] loop.py:45 INFO TRAIN_STEP step=3402 loss=1.936752 Training: 19%|█▉ | 3402/18018 [7:22:39<36:32:30, 9.00s/step, loss=1.937, acc=0.603, acc_len=1.38, thru=9.7, I=17.7, T=31.3, wait=6.8s, pool=44, epoch=1/2] [2026-10-18 20:53:04,474] loop.py:45 INFO TRAIN_STEP step=3403 loss=1.854920 Training: 19%|█▉ | 3403/18018 [7:22:50<38:32:15, 9.49s/step, loss=1.855, acc=0.611, acc_len=1.41, thru=14.4, I=12.3, T=30.9, wait=6.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:09,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:53:11,560] loop.py:45 INFO TRAIN_STEP step=3404 loss=1.705585 Training: 19%|█▉ | 3404/18018 [7:22:57<35:36:09, 8.77s/step, loss=1.706, acc=0.638, acc_len=1.50, thru=17.8, I=19.8, T=31.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:11,936] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:17,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:53:18,036] loop.py:380 INFO TIMING step=3405: step=3.876s data=1.803s compute=3.851s [fwd=2.741s bwd=1.096s opt=0.003s] dispatch=2.593s [2026-10-18 20:53:18,036] loop.py:45 INFO TRAIN_STEP step=3405 loss=1.646144 Training: 19%|█▉ | 3405/18018 [7:23:04<32:48:36, 8.08s/step, loss=1.646, acc=0.643, acc_len=1.52, thru=18.3, I=22.7, T=33.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:18,391] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:25,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:53:27,276] loop.py:45 INFO TRAIN_STEP step=3406 loss=1.562563 Training: 19%|█▉ | 3406/18018 [7:23:13<34:13:00, 8.43s/step, loss=1.563, acc=0.663, acc_len=1.57, thru=13.3, I=16.0, T=27.4, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:53:27,271] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3405: forward=3258.8ms backward=1360.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:27,496] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:33,868] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:53:34,139] loop.py:45 INFO TRAIN_STEP step=3407 loss=1.652244 Training: 19%|█▉ | 3407/18018 [7:23:20<32:18:16, 7.96s/step, loss=1.652, acc=0.647, acc_len=1.52, thru=16.9, I=19.3, T=35.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:34,369] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:39,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:53:40,371] loop.py:45 INFO TRAIN_STEP step=3408 loss=1.659898 Training: 19%|█▉ | 3408/18018 [7:23:26<30:11:54, 7.44s/step, loss=1.660, acc=0.645, acc_len=1.51, thru=19.9, I=21.9, T=35.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:40,620] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:46,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:53:47,913] loop.py:45 INFO TRAIN_STEP step=3409 loss=1.830805 Training: 19%|█▉ | 3409/18018 [7:23:33<30:19:17, 7.47s/step, loss=1.831, acc=0.617, acc_len=1.43, thru=16.2, I=19.0, T=28.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:48,314] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:55,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:53:55,745] loop.py:380 INFO TIMING step=3410: step=5.080s data=2.451s compute=5.051s [fwd=3.632s bwd=1.404s opt=0.003s] dispatch=2.742s [2026-10-18 20:53:55,746] loop.py:45 INFO TRAIN_STEP step=3410 loss=1.674962 Training: 19%|█▉ | 3410/18018 [7:23:41<30:45:29, 7.58s/step, loss=1.675, acc=0.640, acc_len=1.50, thru=15.8, I=19.3, T=25.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:53:56,092] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:02,814] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:54:04,530] loop.py:45 INFO TRAIN_STEP step=3411 loss=1.594453 Training: 19%|█▉ | 3411/18018 [7:23:50<32:13:15, 7.94s/step, loss=1.594, acc=0.652, acc_len=1.53, thru=14.7, I=19.4, T=27.9, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:54:04,525] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3410: forward=3186.7ms backward=1360.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:04,825] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:09,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:54:11,845] loop.py:45 INFO TRAIN_STEP step=3412 loss=1.551778 Training: 19%|█▉ | 3412/18018 [7:23:57<31:27:31, 7.75s/step, loss=1.552, acc=0.663, acc_len=1.57, thru=15.1, I=22.6, T=29.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:12,216] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:18,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:54:19,353] loop.py:45 INFO TRAIN_STEP step=3413 loss=1.629125 Training: 19%|█▉ | 3413/18018 [7:24:05<31:09:17, 7.68s/step, loss=1.629, acc=0.647, acc_len=1.52, thru=14.7, I=22.3, T=33.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:19,832] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:27,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:54:29,544] loop.py:45 INFO TRAIN_STEP step=3414 loss=1.824057 Training: 19%|█▉ | 3414/18018 [7:24:15<34:12:35, 8.43s/step, loss=1.824, acc=0.617, acc_len=1.43, thru=13.2, I=15.7, T=24.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:29,973] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:34,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:54:36,455] loop.py:380 INFO TIMING step=3415: step=4.201s data=1.936s compute=4.149s [fwd=2.848s bwd=1.286s opt=0.003s] dispatch=2.701s [2026-10-18 20:54:36,455] loop.py:45 INFO TRAIN_STEP step=3415 loss=1.617314 Training: 19%|█▉ | 3415/18018 [7:24:22<32:21:37, 7.98s/step, loss=1.617, acc=0.650, acc_len=1.54, thru=15.1, I=21.7, T=30.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:36,532] inference_manager.py:408 INFO Pool capacity available after 1.6s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:41,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 20:54:43,561] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3415: forward=3053.9ms backward=1640.3ms [2026-10-18 20:54:43,565] loop.py:45 INFO TRAIN_STEP step=3416 loss=1.798199 Training: 19%|█▉ | 3416/18018 [7:24:29<31:17:52, 7.72s/step, loss=1.798, acc=0.619, acc_len=1.43, thru=16.0, I=25.8, T=27.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:44,056] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 20:54:51,462] loop.py:45 INFO TRAIN_STEP step=3417 loss=1.634845 Training: 19%|█▉ | 3417/18018 [7:24:37<31:30:51, 7.77s/step, loss=1.635, acc=0.641, acc_len=1.51, thru=17.3, I=17.0, T=31.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:51,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:51,919] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:54:59,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:55:01,015] loop.py:45 INFO TRAIN_STEP step=3418 loss=1.718772 Training: 19%|█▉ | 3418/18018 [7:24:47<33:41:04, 8.31s/step, loss=1.719, acc=0.635, acc_len=1.48, thru=14.1, I=16.8, T=31.4, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:55:01,280] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:55:08,365] loop.py:45 INFO TRAIN_STEP step=3419 loss=1.830173 Training: 19%|█▉ | 3419/18018 [7:24:54<32:31:16, 8.02s/step, loss=1.830, acc=0.618, acc_len=1.42, thru=13.1, I=20.1, T=34.2, wait=3.6s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:55:15,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:55:17,345] loop.py:380 INFO TIMING step=3420: step=4.295s data=1.979s compute=4.264s [fwd=3.022s bwd=1.228s opt=0.003s] dispatch=4.673s [2026-10-18 20:55:17,345] loop.py:45 INFO TRAIN_STEP step=3420 loss=1.583389 Training: 19%|█▉ | 3420/18018 [7:25:03<33:40:58, 8.31s/step, loss=1.583, acc=0.656, acc_len=1.55, thru=17.9, I=16.3, T=29.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:55:17,414] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 20:55:25,933] loop.py:45 INFO TRAIN_STEP step=3421 loss=1.667038 Training: 19%|█▉ | 3421/18018 [7:25:12<34:01:26, 8.39s/step, loss=1.667, acc=0.644, acc_len=1.51, thru=12.7, I=14.3, T=31.3, wait=4.5s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:55:25,928] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3420: forward=2838.8ms backward=1186.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:55:33,032] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:55:35,517] loop.py:45 INFO TRAIN_STEP step=3422 loss=1.576902 Training: 19%|█▉ | 3422/18018 [7:25:21<35:28:17, 8.75s/step, loss=1.577, acc=0.654, acc_len=1.55, thru=15.8, I=18.5, T=24.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:55:35,548] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:55:43,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:55:43,903] loop.py:45 INFO TRAIN_STEP step=3423 loss=1.635309 Training: 19%|█▉ | 3423/18018 [7:25:29<35:01:42, 8.64s/step, loss=1.635, acc=0.644, acc_len=1.51, thru=15.3, I=15.4, T=26.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:55:44,262] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:55:53,729] loop.py:45 INFO TRAIN_STEP step=3424 loss=1.646271 Training: 19%|█▉ | 3424/18018 [7:25:39<36:28:04, 9.00s/step, loss=1.646, acc=0.646, acc_len=1.52, thru=8.4, I=11.7, T=30.6, wait=5.6s, pool=16, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:04,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:56:06,777] loop.py:380 INFO TIMING step=3425: step=5.091s data=2.425s compute=5.050s [fwd=3.553s bwd=1.481s opt=0.003s] dispatch=7.947s [2026-10-18 20:56:06,777] loop.py:45 INFO TRAIN_STEP step=3425 loss=1.770288 Training: 19%|█▉ | 3425/18018 [7:25:52<41:23:41, 10.21s/step, loss=1.770, acc=0.633, acc_len=1.48, thru=15.2, I=18.4, T=25.1, wait=7.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:06,881] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:13,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:56:14,077] loop.py:45 INFO TRAIN_STEP step=3426 loss=1.598552 Training: 19%|█▉ | 3426/18018 [7:26:00<37:51:04, 9.34s/step, loss=1.599, acc=0.650, acc_len=1.53, thru=15.2, I=20.5, T=28.2, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:56:14,072] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3425: forward=3084.2ms backward=1392.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:14,367] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:22,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:56:22,513] loop.py:45 INFO TRAIN_STEP step=3427 loss=1.811069 Training: 19%|█▉ | 3427/18018 [7:26:08<36:45:03, 9.07s/step, loss=1.811, acc=0.624, acc_len=1.45, thru=15.0, I=18.0, T=29.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:22,688] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 20:56:30,297] loop.py:45 INFO TRAIN_STEP step=3428 loss=1.597079 Training: 19%|█▉ | 3428/18018 [7:26:16<35:11:21, 8.68s/step, loss=1.597, acc=0.655, acc_len=1.54, thru=14.8, I=15.1, T=27.5, wait=3.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:38,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:56:39,125] loop.py:45 INFO TRAIN_STEP step=3429 loss=1.725207 Training: 19%|█▉ | 3429/18018 [7:26:25<35:21:44, 8.73s/step, loss=1.725, acc=0.632, acc_len=1.47, thru=15.8, I=15.5, T=29.0, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:39,562] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:44,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:56:46,234] loop.py:380 INFO TIMING step=3430: step=4.373s data=1.997s compute=4.345s [fwd=3.015s bwd=1.314s opt=0.003s] dispatch=2.726s [2026-10-18 20:56:46,234] loop.py:45 INFO TRAIN_STEP step=3430 loss=1.652597 Training: 19%|█▉ | 3430/18018 [7:26:32<33:23:40, 8.24s/step, loss=1.653, acc=0.640, acc_len=1.49, thru=16.8, I=22.8, T=29.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:46,309] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:51,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:56:52,131] loop.py:45 INFO TRAIN_STEP step=3431 loss=1.722761 Training: 19%|█▉ | 3431/18018 [7:26:38<30:32:33, 7.54s/step, loss=1.723, acc=0.639, acc_len=1.50, thru=18.6, I=24.6, T=33.2, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:56:52,125] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3430: forward=2714.3ms backward=1102.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:52,622] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:56:59,906] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:57:01,170] loop.py:45 INFO TRAIN_STEP step=3432 loss=1.886850 Training: 19%|█▉ | 3432/18018 [7:26:47<32:21:48, 7.99s/step, loss=1.887, acc=0.606, acc_len=1.40, thru=13.8, I=17.2, T=28.2, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:57:01,414] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:57:06,494] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:57:08,033] loop.py:45 INFO TRAIN_STEP step=3433 loss=1.714313 Training: 19%|█▉ | 3433/18018 [7:26:54<30:59:41, 7.65s/step, loss=1.714, acc=0.633, acc_len=1.47, thru=17.2, I=23.2, T=27.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:57:08,504] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 20:57:14,542] loop.py:45 INFO TRAIN_STEP step=3434 loss=1.532750 Training: 19%|█▉ | 3434/18018 [7:27:00<29:36:22, 7.31s/step, loss=1.533, acc=0.662, acc_len=1.58, thru=16.9, I=20.7, T=34.2, wait=2.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:57:23,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:57:24,793] loop.py:380 INFO TIMING step=3435: step=4.623s data=2.191s compute=4.592s [fwd=3.291s bwd=1.286s opt=0.003s] dispatch=5.617s [2026-10-18 20:57:24,793] loop.py:45 INFO TRAIN_STEP step=3435 loss=1.643137 Training: 19%|█▉ | 3435/18018 [7:27:10<33:10:50, 8.19s/step, loss=1.643, acc=0.649, acc_len=1.52, thru=14.0, I=12.8, T=27.7, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:57:24,972] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:57:33,654] loop.py:45 INFO TRAIN_STEP step=3436 loss=1.736381 Training: 19%|█▉ | 3436/18018 [7:27:19<33:59:29, 8.39s/step, loss=1.736, acc=0.632, acc_len=1.48, thru=12.5, I=16.0, T=27.3, wait=4.2s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:57:33,650] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3435: forward=3257.0ms backward=1391.0ms [2026-10-18 20:57:42,485] loop.py:45 INFO TRAIN_STEP step=3437 loss=1.673106 Training: 19%|█▉ | 3437/18018 [7:27:28<34:31:24, 8.52s/step, loss=1.673, acc=0.641, acc_len=1.50, thru=15.0, I=14.4, T=29.0, wait=4.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:57:47,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:57:49,460] loop.py:45 INFO TRAIN_STEP step=3438 loss=1.721280 Training: 19%|█▉ | 3438/18018 [7:27:35<32:38:18, 8.06s/step, loss=1.721, acc=0.636, acc_len=1.48, thru=17.5, I=21.1, T=28.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:57:49,736] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:57:55,753] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:57:57,031] loop.py:45 INFO TRAIN_STEP step=3439 loss=1.647596 Training: 19%|█▉ | 3439/18018 [7:27:43<32:02:37, 7.91s/step, loss=1.648, acc=0.644, acc_len=1.51, thru=13.7, I=22.0, T=29.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:57:57,261] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 20:58:03,716] loop.py:380 INFO TIMING step=3440: step=3.966s data=1.852s compute=3.935s [fwd=2.762s bwd=1.160s opt=0.003s] dispatch=2.710s [2026-10-18 20:58:03,717] loop.py:45 INFO TRAIN_STEP step=3440 loss=1.729051 Training: 19%|█▉ | 3440/18018 [7:27:49<30:33:02, 7.54s/step, loss=1.729, acc=0.633, acc_len=1.48, thru=15.7, I=20.4, T=32.3, wait=2.7s, pool=52, epoch=1/2][2026-10-18 20:58:11,876] loop.py:45 INFO TRAIN_STEP step=3441 loss=1.751147 Training: 19%|█▉ | 3441/18018 [7:27:57<31:17:43, 7.73s/step, loss=1.751, acc=0.631, acc_len=1.46, thru=14.6, I=13.8, T=27.4, wait=3.5s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:58:11,871] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3440: forward=3220.1ms backward=1402.4ms [2026-10-18 20:58:19,685] loop.py:45 INFO TRAIN_STEP step=3442 loss=1.779334 Training: 19%|█▉ | 3442/18018 [7:28:05<31:23:25, 7.75s/step, loss=1.779, acc=0.624, acc_len=1.45, thru=17.0, I=17.6, T=28.8, wait=3.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:58:28,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:58:29,293] loop.py:45 INFO TRAIN_STEP step=3443 loss=1.811671 Training: 19%|█▉ | 3443/18018 [7:28:15<33:38:43, 8.31s/step, loss=1.812, acc=0.618, acc_len=1.43, thru=14.6, I=14.0, T=29.8, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:58:29,389] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:58:37,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:58:38,161] loop.py:45 INFO TRAIN_STEP step=3444 loss=1.841056 Training: 19%|█▉ | 3444/18018 [7:28:24<34:19:06, 8.48s/step, loss=1.841, acc=0.618, acc_len=1.43, thru=13.5, I=14.4, T=30.9, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:58:38,571] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:58:44,613] loop.py:380 INFO TIMING step=3445: step=3.698s data=1.556s compute=3.663s [fwd=2.609s bwd=1.039s opt=0.003s] dispatch=2.744s [2026-10-18 20:58:44,613] loop.py:45 INFO TRAIN_STEP step=3445 loss=1.730064 Training: 19%|█▉ | 3445/18018 [7:28:30<31:51:12, 7.87s/step, loss=1.730, acc=0.635, acc_len=1.47, thru=18.6, I=20.6, T=34.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:58:44,614] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:58:45,116] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:58:53,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:58:55,804] loop.py:45 INFO TRAIN_STEP step=3446 loss=1.699741 Training: 19%|█▉ | 3446/18018 [7:28:41<35:53:22, 8.87s/step, loss=1.700, acc=0.639, acc_len=1.49, thru=12.1, I=14.7, T=30.7, wait=7.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:58:55,799] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3445: forward=2779.4ms backward=1315.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:58:55,916] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 [2026-10-18 20:59:02,177] loop.py:45 INFO TRAIN_STEP step=3447 loss=1.602597 Training: 19%|█▉ | 3447/18018 [7:28:48<32:51:21, 8.12s/step, loss=1.603, acc=0.647, acc_len=1.53, thru=17.2, I=22.9, T=33.6, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:08,328] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:59:09,910] loop.py:45 INFO TRAIN_STEP step=3448 loss=1.644669 Training: 19%|█▉ | 3448/18018 [7:28:55<32:23:11, 8.00s/step, loss=1.645, acc=0.644, acc_len=1.51, thru=18.1, I=19.3, T=27.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:10,338] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:15,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:59:16,942] loop.py:45 INFO TRAIN_STEP step=3449 loss=1.620466 Training: 19%|█▉ | 3449/18018 [7:29:03<31:12:27, 7.71s/step, loss=1.620, acc=0.649, acc_len=1.53, thru=15.5, I=21.4, T=31.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:17,446] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:23,039] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:59:23,200] loop.py:380 INFO TIMING step=3450: step=4.044s data=1.900s compute=4.020s [fwd=2.818s bwd=1.186s opt=0.003s] dispatch=2.206s [2026-10-18 20:59:23,201] loop.py:45 INFO TRAIN_STEP step=3450 loss=1.531973 Training: 19%|█▉ | 3450/18018 [7:29:09<29:26:26, 7.28s/step, loss=1.532, acc=0.665, acc_len=1.58, thru=19.3, I=21.0, T=31.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:23,541] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:29,357] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:59:30,007] loop.py:45 INFO TRAIN_STEP step=3451 loss=1.756252 Training: 19%|█▉ | 3451/18018 [7:29:16<28:52:21, 7.14s/step, loss=1.756, acc=0.627, acc_len=1.45, thru=18.1, I=19.7, T=33.9, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 20:59:30,001] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3450: forward=2586.0ms backward=1133.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:30,361] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 20:59:36,726] loop.py:45 INFO TRAIN_STEP step=3452 loss=1.671867 Training: 19%|█▉ | 3452/18018 [7:29:22<28:22:00, 7.01s/step, loss=1.672, acc=0.640, acc_len=1.50, thru=16.1, I=19.0, T=34.7, wait=3.0s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:42,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:59:44,086] loop.py:45 INFO TRAIN_STEP step=3453 loss=1.680974 Training: 19%|█▉ | 3453/18018 [7:29:30<28:47:07, 7.11s/step, loss=1.681, acc=0.635, acc_len=1.49, thru=18.8, I=18.2, T=30.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:44,407] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:50,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:59:51,354] loop.py:45 INFO TRAIN_STEP step=3454 loss=1.650514 Training: 19%|█▉ | 3454/18018 [7:29:37<28:58:07, 7.16s/step, loss=1.651, acc=0.645, acc_len=1.51, thru=14.8, I=20.2, T=33.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:51,739] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:57,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 20:59:58,892] loop.py:380 INFO TIMING step=3455: step=3.925s data=1.747s compute=3.885s [fwd=2.707s bwd=1.163s opt=0.003s] dispatch=3.604s [2026-10-18 20:59:58,893] loop.py:45 INFO TRAIN_STEP step=3455 loss=1.632640 Training: 19%|█▉ | 3455/18018 [7:29:44<29:25:38, 7.27s/step, loss=1.633, acc=0.651, acc_len=1.54, thru=16.5, I=20.8, T=32.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 20:59:59,065] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:04,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:00:05,113] loop.py:45 INFO TRAIN_STEP step=3456 loss=1.617648 Training: 19%|█▉ | 3456/18018 [7:29:51<28:08:40, 6.96s/step, loss=1.618, acc=0.648, acc_len=1.52, thru=19.3, I=23.7, T=30.0, wait=1.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:00:05,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3455: forward=3004.4ms backward=1216.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:05,218] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 21:00:11,125] loop.py:45 INFO TRAIN_STEP step=3457 loss=1.616707 Training: 19%|█▉ | 3457/18018 [7:29:57<26:59:50, 6.67s/step, loss=1.617, acc=0.648, acc_len=1.53, thru=16.9, I=21.5, T=35.2, wait=2.4s, pool=44, epoch=1/2][2026-10-18 21:00:19,933] loop.py:45 INFO TRAIN_STEP step=3458 loss=1.632250 Training: 19%|█▉ | 3458/18018 [7:30:06<29:34:51, 7.31s/step, loss=1.632, acc=0.643, acc_len=1.52, thru=16.6, I=14.1, T=31.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:19,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:20,363] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:25,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:00:27,687] loop.py:45 INFO TRAIN_STEP step=3459 loss=1.653838 Training: 19%|█▉ | 3459/18018 [7:30:13<30:06:47, 7.45s/step, loss=1.654, acc=0.646, acc_len=1.51, thru=17.4, I=20.8, T=23.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:27,947] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:34,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:00:35,138] loop.py:380 INFO TIMING step=3460: step=4.481s data=2.073s compute=4.448s [fwd=3.076s bwd=1.356s opt=0.003s] dispatch=2.961s [2026-10-18 21:00:35,138] loop.py:45 INFO TRAIN_STEP step=3460 loss=1.662975 Training: 19%|█▉ | 3460/18018 [7:30:21<30:07:06, 7.45s/step, loss=1.663, acc=0.645, acc_len=1.50, thru=13.7, I=16.4, T=28.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:35,359] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:42,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:00:43,041] loop.py:45 INFO TRAIN_STEP step=3461 loss=1.617834 Training: 19%|█▉ | 3461/18018 [7:30:29<30:40:10, 7.58s/step, loss=1.618, acc=0.648, acc_len=1.52, thru=16.1, I=16.9, T=33.6, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:00:43,036] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3460: forward=2699.7ms backward=1053.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:43,275] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:48,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:00:50,554] loop.py:45 INFO TRAIN_STEP step=3462 loss=1.645871 Training: 19%|█▉ | 3462/18018 [7:30:36<30:34:48, 7.56s/step, loss=1.646, acc=0.645, acc_len=1.52, thru=17.4, I=22.9, T=25.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:50,974] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:56,978] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:00:58,349] loop.py:45 INFO TRAIN_STEP step=3463 loss=1.738473 Training: 19%|█▉ | 3463/18018 [7:30:44<30:51:32, 7.63s/step, loss=1.738, acc=0.632, acc_len=1.47, thru=13.6, I=19.9, T=30.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:00:58,501] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 21:01:06,526] loop.py:45 INFO TRAIN_STEP step=3464 loss=1.601161 Training: 19%|█▉ | 3464/18018 [7:30:52<31:31:07, 7.80s/step, loss=1.601, acc=0.654, acc_len=1.55, thru=13.6, I=18.1, T=32.8, wait=4.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:01:14,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:01:16,219] loop.py:380 INFO TIMING step=3465: step=4.630s data=2.141s compute=4.601s [fwd=3.338s bwd=1.247s opt=0.003s] dispatch=5.051s [2026-10-18 21:01:16,219] loop.py:45 INFO TRAIN_STEP step=3465 loss=1.735217 Training: 19%|█▉ | 3465/18018 [7:31:02<33:48:57, 8.37s/step, loss=1.735, acc=0.627, acc_len=1.46, thru=14.0, I=14.1, T=27.6, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:01:16,254] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:01:21,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:01:22,823] loop.py:45 INFO TRAIN_STEP step=3466 loss=1.553828 Training: 19%|█▉ | 3466/18018 [7:31:08<31:40:39, 7.84s/step, loss=1.554, acc=0.657, acc_len=1.56, thru=17.8, I=22.7, T=30.7, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:01:22,818] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3465: forward=2926.6ms backward=1195.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:01:23,060] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:01:28,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:01:29,178] loop.py:45 INFO TRAIN_STEP step=3467 loss=1.600373 Training: 19%|█▉ | 3467/18018 [7:31:15<29:52:36, 7.39s/step, loss=1.600, acc=0.653, acc_len=1.54, thru=18.8, I=24.7, T=31.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:01:29,266] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 21:01:36,085] loop.py:45 INFO TRAIN_STEP step=3468 loss=1.694107 Training: 19%|█▉ | 3468/18018 [7:31:22<29:17:17, 7.25s/step, loss=1.694, acc=0.643, acc_len=1.51, thru=14.8, I=16.7, T=31.8, wait=2.9s, pool=44, epoch=1/2][2026-10-18 21:01:44,658] loop.py:45 INFO TRAIN_STEP step=3469 loss=1.748002 Training: 19%|█▉ | 3469/18018 [7:31:30<30:53:34, 7.64s/step, loss=1.748, acc=0.635, acc_len=1.49, thru=14.8, I=13.4, T=29.6, wait=4.2s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:01:51,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:01:52,383] loop.py:380 INFO TIMING step=3470: step=3.916s data=1.779s compute=3.875s [fwd=2.784s bwd=1.076s opt=0.003s] dispatch=3.801s [2026-10-18 21:01:52,384] loop.py:45 INFO TRAIN_STEP step=3470 loss=1.590857 Training: 19%|█▉ | 3470/18018 [7:31:38<30:59:28, 7.67s/step, loss=1.591, acc=0.657, acc_len=1.56, thru=18.6, I=16.3, T=32.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:01:52,700] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:01:58,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:02:01,354] loop.py:45 INFO TRAIN_STEP step=3471 loss=1.678295 Training: 19%|█▉ | 3471/18018 [7:31:47<32:33:59, 8.06s/step, loss=1.678, acc=0.643, acc_len=1.50, thru=13.7, I=20.0, T=27.9, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:02:01,349] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3470: forward=3204.2ms backward=1346.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:01,461] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:07,092] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:02:08,032] loop.py:45 INFO TRAIN_STEP step=3472 loss=1.646865 Training: 19%|█▉ | 3472/18018 [7:31:54<30:53:20, 7.64s/step, loss=1.647, acc=0.643, acc_len=1.51, thru=15.2, I=23.2, T=32.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:08,104] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:13,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:02:14,410] loop.py:45 INFO TRAIN_STEP step=3473 loss=1.911275 Training: 19%|█▉ | 3473/18018 [7:32:00<29:21:06, 7.26s/step, loss=1.911, acc=0.604, acc_len=1.39, thru=18.9, I=22.3, T=34.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:14,787] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:20,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:02:22,121] loop.py:45 INFO TRAIN_STEP step=3474 loss=1.734329 Training: 19%|█▉ | 3474/18018 [7:32:08<29:53:19, 7.40s/step, loss=1.734, acc=0.631, acc_len=1.47, thru=16.1, I=20.9, T=28.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:22,216] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:30,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:02:30,324] loop.py:380 INFO TIMING step=3475: step=4.706s data=2.304s compute=4.668s [fwd=3.243s bwd=1.409s opt=0.003s] dispatch=3.488s [2026-10-18 21:02:30,324] loop.py:45 INFO TRAIN_STEP step=3475 loss=1.702312 Training: 19%|█▉ | 3475/18018 [7:32:16<30:51:45, 7.64s/step, loss=1.702, acc=0.642, acc_len=1.51, thru=13.7, I=15.2, T=27.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:30,598] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:35,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:02:36,898] loop.py:45 INFO TRAIN_STEP step=3476 loss=1.696620 Training: 19%|█▉ | 3476/18018 [7:32:22<29:34:07, 7.32s/step, loss=1.697, acc=0.635, acc_len=1.48, thru=18.1, I=18.7, T=32.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:36,928] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 21:02:36,892] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3475: forward=2642.4ms backward=1222.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:42,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:02:44,333] loop.py:45 INFO TRAIN_STEP step=3477 loss=1.766397 Training: 19%|█▉ | 3477/18018 [7:32:30<29:42:18, 7.35s/step, loss=1.766, acc=0.629, acc_len=1.46, thru=15.7, I=21.5, T=27.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:44,797] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:51,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:02:52,518] loop.py:45 INFO TRAIN_STEP step=3478 loss=1.667194 Training: 19%|█▉ | 3478/18018 [7:32:38<30:42:45, 7.60s/step, loss=1.667, acc=0.641, acc_len=1.50, thru=16.6, I=18.2, T=32.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:52,658] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:56,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:02:58,560] loop.py:45 INFO TRAIN_STEP step=3479 loss=1.604082 Training: 19%|█▉ | 3479/18018 [7:32:44<28:49:01, 7.14s/step, loss=1.604, acc=0.651, acc_len=1.53, thru=19.3, I=26.0, T=30.8, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:02:59,003] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:06,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:03:07,226] loop.py:380 INFO TIMING step=3480: step=4.016s data=1.822s compute=3.971s [fwd=2.741s bwd=1.215s opt=0.003s] dispatch=4.640s [2026-10-18 21:03:07,226] loop.py:45 INFO TRAIN_STEP step=3480 loss=1.832931 Training: 19%|█▉ | 3480/18018 [7:32:53<30:40:05, 7.59s/step, loss=1.833, acc=0.617, acc_len=1.43, thru=15.6, I=16.3, T=31.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:07,731] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:03:14,656] loop.py:45 INFO TRAIN_STEP step=3481 loss=1.644467 Training: 19%|█▉ | 3481/18018 [7:33:00<30:28:01, 7.54s/step, loss=1.644, acc=0.649, acc_len=1.53, thru=15.1, I=16.6, T=33.9, wait=3.6s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:03:14,651] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3480: forward=2699.9ms backward=1023.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:20,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:03:22,253] loop.py:45 INFO TRAIN_STEP step=3482 loss=1.769721 Training: 19%|█▉ | 3482/18018 [7:33:08<30:31:39, 7.56s/step, loss=1.770, acc=0.624, acc_len=1.45, thru=18.5, I=21.2, T=30.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:22,449] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:03:31,025] loop.py:45 INFO TRAIN_STEP step=3483 loss=1.843190 Training: 19%|█▉ | 3483/18018 [7:33:17<31:59:31, 7.92s/step, loss=1.843, acc=0.613, acc_len=1.41, thru=14.5, I=12.5, T=26.8, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:39,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:03:40,220] loop.py:45 INFO TRAIN_STEP step=3484 loss=1.593900 Training: 19%|█▉ | 3484/18018 [7:33:26<33:31:52, 8.31s/step, loss=1.594, acc=0.658, acc_len=1.56, thru=14.8, I=15.4, T=28.0, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:40,327] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:46,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:03:48,162] loop.py:380 INFO TIMING step=3485: step=4.314s data=2.003s compute=4.265s [fwd=2.875s bwd=1.375s opt=0.003s] dispatch=3.618s [2026-10-18 21:03:48,162] loop.py:45 INFO TRAIN_STEP step=3485 loss=1.694413 Training: 19%|█▉ | 3485/18018 [7:33:34<33:05:18, 8.20s/step, loss=1.694, acc=0.638, acc_len=1.49, thru=17.0, I=19.6, T=29.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:48,536] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:55,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:03:55,919] loop.py:45 INFO TRAIN_STEP step=3486 loss=1.725292 Training: 19%|█▉ | 3486/18018 [7:33:42<32:33:14, 8.06s/step, loss=1.725, acc=0.635, acc_len=1.48, thru=13.3, I=17.9, T=32.9, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:03:55,914] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3485: forward=2679.1ms backward=1163.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:03:56,220] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:01,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:04:02,190] loop.py:45 INFO TRAIN_STEP step=3487 loss=1.812684 Training: 19%|█▉ | 3487/18018 [7:33:48<30:22:52, 7.53s/step, loss=1.813, acc=0.617, acc_len=1.43, thru=18.9, I=21.0, T=35.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:02,468] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:08,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:04:09,701] loop.py:45 INFO TRAIN_STEP step=3488 loss=1.642748 Training: 19%|█▉ | 3488/18018 [7:33:55<30:21:37, 7.52s/step, loss=1.643, acc=0.650, acc_len=1.53, thru=17.9, I=20.2, T=31.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:09,809] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:15,352] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:04:17,210] loop.py:45 INFO TRAIN_STEP step=3489 loss=1.618475 Training: 19%|█▉ | 3489/18018 [7:34:03<30:20:28, 7.52s/step, loss=1.618, acc=0.648, acc_len=1.53, thru=16.1, I=22.6, T=28.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:17,369] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:23,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:04:23,545] loop.py:380 INFO TIMING step=3490: step=3.843s data=1.724s compute=3.815s [fwd=2.703s bwd=1.097s opt=0.003s] dispatch=2.483s [2026-10-18 21:04:23,545] loop.py:45 INFO TRAIN_STEP step=3490 loss=1.763848 Training: 19%|█▉ | 3490/18018 [7:34:09<28:54:29, 7.16s/step, loss=1.764, acc=0.627, acc_len=1.45, thru=17.5, I=22.7, T=33.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:23,795] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:28,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:04:30,505] loop.py:45 INFO TRAIN_STEP step=3491 loss=1.790852 Training: 19%|█▉ | 3491/18018 [7:34:16<28:39:29, 7.10s/step, loss=1.791, acc=0.622, acc_len=1.44, thru=19.7, I=21.2, T=29.9, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:04:30,500] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3490: forward=2949.8ms backward=1270.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:30,911] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:37,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:04:38,576] loop.py:45 INFO TRAIN_STEP step=3492 loss=1.612758 Training: 19%|█▉ | 3492/18018 [7:34:24<29:49:51, 7.39s/step, loss=1.613, acc=0.648, acc_len=1.52, thru=14.0, I=18.1, T=32.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:38,870] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:46,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:04:46,450] loop.py:45 INFO TRAIN_STEP step=3493 loss=1.716691 Training: 19%|█▉ | 3493/18018 [7:34:32<30:24:41, 7.54s/step, loss=1.717, acc=0.635, acc_len=1.49, thru=16.5, I=18.3, T=29.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:46,639] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:53,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:04:55,467] loop.py:45 INFO TRAIN_STEP step=3494 loss=1.642882 Training: 19%|█▉ | 3494/18018 [7:34:41<32:11:52, 7.98s/step, loss=1.643, acc=0.650, acc_len=1.53, thru=14.0, I=16.4, T=28.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:04:55,910] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:00,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:05:02,552] loop.py:380 INFO TIMING step=3495: step=4.195s data=1.868s compute=4.161s [fwd=2.878s bwd=1.270s opt=0.003s] dispatch=2.880s [2026-10-18 21:05:02,552] loop.py:45 INFO TRAIN_STEP step=3495 loss=1.652866 Training: 19%|█▉ | 3495/18018 [7:34:48<31:06:38, 7.71s/step, loss=1.653, acc=0.644, acc_len=1.52, thru=15.9, I=23.8, T=30.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:02,993] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:08,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:05:09,025] loop.py:45 INFO TRAIN_STEP step=3496 loss=1.589336 Training: 19%|█▉ | 3496/18018 [7:34:55<29:36:44, 7.34s/step, loss=1.589, acc=0.654, acc_len=1.54, thru=18.3, I=25.8, T=33.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:05:09,019] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3495: forward=2650.7ms backward=1098.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:09,242] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 21:05:15,837] loop.py:45 INFO TRAIN_STEP step=3497 loss=1.571733 Training: 19%|█▉ | 3497/18018 [7:35:01<28:58:19, 7.18s/step, loss=1.572, acc=0.659, acc_len=1.55, thru=17.6, I=21.8, T=25.4, wait=1.8s, pool=52, epoch=1/2][2026-10-18 21:05:24,643] loop.py:45 INFO TRAIN_STEP step=3498 loss=1.721709 Training: 19%|█▉ | 3498/18018 [7:35:10<30:55:52, 7.67s/step, loss=1.722, acc=0.640, acc_len=1.50, thru=14.7, I=12.4, T=27.1, wait=4.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:33,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:05:35,541] loop.py:45 INFO TRAIN_STEP step=3499 loss=1.674363 Training: 19%|█▉ | 3499/18018 [7:35:21<34:50:13, 8.64s/step, loss=1.674, acc=0.643, acc_len=1.51, thru=13.0, I=13.7, T=25.4, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:35,962] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:40,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:05:42,524] loop.py:380 INFO TIMING step=3500: step=4.245s data=1.949s compute=4.190s [fwd=2.885s bwd=1.291s opt=0.003s] dispatch=2.728s [2026-10-18 21:05:42,524] loop.py:45 INFO TRAIN_STEP step=3500 loss=1.643439 Training: 19%|█▉ | 3500/18018 [7:35:28<32:49:58, 8.14s/step, loss=1.643, acc=0.642, acc_len=1.50, thru=16.6, I=22.4, T=30.2, wait=2.7s, pool=64, epoch=1/2][2026-10-18 21:05:42,529] loop.py:419 INFO Saving checkpoint at step 3500... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:45,670] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 21:05:46,018] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0003501 (TrainerActor pid=9310) [2026-10-18 21:05:46,028] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0002751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:47,187] inference_manager.py:408 INFO Pool capacity available after 6.6s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:53,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 21:05:46,016] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0003501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 21:05:47,036] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0002751 [2026-10-18 21:05:54,658] loop.py:45 INFO TRAIN_STEP step=3501 loss=1.710329 Training: 19%|█▉ | 3501/18018 [7:35:40<37:39:36, 9.34s/step, loss=1.710, acc=0.637, acc_len=1.50, thru=17.1, I=20.6, T=32.8, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:05:54,652] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3500: forward=2672.9ms backward=1177.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:05:54,785] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 21:06:01,925] loop.py:45 INFO TRAIN_STEP step=3502 loss=1.587944 Training: 19%|█▉ | 3502/18018 [7:35:48<35:08:53, 8.72s/step, loss=1.588, acc=0.656, acc_len=1.55, thru=17.2, I=18.7, T=30.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:01,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:02,377] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 21:06:11,000] loop.py:45 INFO TRAIN_STEP step=3503 loss=1.792314 Training: 19%|█▉ | 3503/18018 [7:35:57<35:34:46, 8.82s/step, loss=1.792, acc=0.627, acc_len=1.46, thru=13.6, I=13.5, T=31.8, wait=5.0s, pool=44, epoch=1/2][2026-10-18 21:06:21,966] loop.py:45 INFO TRAIN_STEP step=3504 loss=1.607106 Training: 19%|█▉ | 3504/18018 [7:36:08<38:10:05, 9.47s/step, loss=1.607, acc=0.654, acc_len=1.54, thru=12.9, I=12.6, T=31.0, wait=6.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:32,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:06:32,593] loop.py:380 INFO TIMING step=3505: step=4.849s data=2.171s compute=4.819s [fwd=3.391s bwd=1.413s opt=0.003s] dispatch=5.770s [2026-10-18 21:06:32,593] loop.py:45 INFO TRAIN_STEP step=3505 loss=1.707785 Training: 19%|█▉ | 3505/18018 [7:36:18<39:34:03, 9.81s/step, loss=1.708, acc=0.639, acc_len=1.50, thru=13.0, I=13.4, T=26.4, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:32,681] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:38,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:06:41,106] loop.py:45 INFO TRAIN_STEP step=3506 loss=1.747584 Training: 19%|█▉ | 3506/18018 [7:36:27<37:59:34, 9.42s/step, loss=1.748, acc=0.632, acc_len=1.48, thru=14.5, I=20.5, T=29.6, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:06:41,100] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3505: forward=3086.1ms backward=1195.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:41,194] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 [2026-10-18 21:06:48,657] loop.py:45 INFO TRAIN_STEP step=3507 loss=1.557129 Training: 19%|█▉ | 3507/18018 [7:36:34<35:43:13, 8.86s/step, loss=1.557, acc=0.660, acc_len=1.57, thru=13.2, I=18.8, T=32.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:48,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:49,151] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:54,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:06:55,933] loop.py:45 INFO TRAIN_STEP step=3508 loss=1.676959 Training: 19%|█▉ | 3508/18018 [7:36:42<33:48:03, 8.39s/step, loss=1.677, acc=0.639, acc_len=1.51, thru=17.0, I=19.1, T=35.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:06:56,391] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:04,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:07:05,558] loop.py:45 INFO TRAIN_STEP step=3509 loss=1.793310 Training: 19%|█▉ | 3509/18018 [7:36:51<35:18:00, 8.76s/step, loss=1.793, acc=0.619, acc_len=1.43, thru=14.0, I=15.7, T=30.5, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:05,597] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:12,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:07:12,454] loop.py:380 INFO TIMING step=3510: step=4.327s data=2.009s compute=4.292s [fwd=2.975s bwd=1.303s opt=0.003s] dispatch=2.558s [2026-10-18 21:07:12,454] loop.py:45 INFO TRAIN_STEP step=3510 loss=1.746349 Training: 19%|█▉ | 3510/18018 [7:36:58<33:02:32, 8.20s/step, loss=1.746, acc=0.632, acc_len=1.47, thru=18.1, I=19.3, T=29.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:12,669] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:19,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:07:21,931] loop.py:45 INFO TRAIN_STEP step=3511 loss=1.638850 Training: 19%|█▉ | 3511/18018 [7:37:08<34:35:14, 8.58s/step, loss=1.639, acc=0.648, acc_len=1.53, thru=13.8, I=16.7, T=23.6, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:07:21,926] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3510: forward=4006.1ms backward=1358.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:22,144] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:29,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:07:30,487] loop.py:45 INFO TRAIN_STEP step=3512 loss=1.587056 Training: 19%|█▉ | 3512/18018 [7:37:16<34:33:07, 8.57s/step, loss=1.587, acc=0.653, acc_len=1.54, thru=15.3, I=16.7, T=23.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:30,943] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:07:36,406] loop.py:45 INFO TRAIN_STEP step=3513 loss=1.753807 Training: 19%|█▉ | 3513/18018 [7:37:22<31:20:21, 7.78s/step, loss=1.754, acc=0.630, acc_len=1.46, thru=16.6, I=20.3, T=35.1, wait=2.3s, pool=56, epoch=1/2][2026-10-18 21:07:44,771] loop.py:45 INFO TRAIN_STEP step=3514 loss=1.677262 Training: 20%|█▉ | 3514/18018 [7:37:30<32:04:06, 7.96s/step, loss=1.677, acc=0.636, acc_len=1.49, thru=15.4, I=13.9, T=34.2, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:51,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:07:53,592] loop.py:380 INFO TIMING step=3515: step=4.469s data=2.087s compute=4.435s [fwd=3.115s bwd=1.306s opt=0.003s] dispatch=4.324s [2026-10-18 21:07:53,592] loop.py:45 INFO TRAIN_STEP step=3515 loss=1.725380 Training: 20%|█▉ | 3515/18018 [7:37:39<33:04:59, 8.21s/step, loss=1.725, acc=0.630, acc_len=1.46, thru=16.5, I=18.5, T=28.6, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:07:53,803] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:00,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:08:01,502] loop.py:45 INFO TRAIN_STEP step=3516 loss=1.684920 Training: 20%|█▉ | 3516/18018 [7:37:47<32:42:52, 8.12s/step, loss=1.685, acc=0.639, acc_len=1.49, thru=13.7, I=18.6, T=31.4, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:08:01,496] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3515: forward=2889.0ms backward=1130.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:02,003] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 21:08:10,470] loop.py:45 INFO TRAIN_STEP step=3517 loss=1.756847 Training: 20%|█▉ | 3517/18018 [7:37:56<33:44:20, 8.38s/step, loss=1.757, acc=0.631, acc_len=1.47, thru=13.0, I=14.8, T=30.0, wait=4.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:15,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:08:18,266] loop.py:45 INFO TRAIN_STEP step=3518 loss=1.659151 Training: 20%|█▉ | 3518/18018 [7:38:04<33:01:57, 8.20s/step, loss=1.659, acc=0.642, acc_len=1.51, thru=15.5, I=19.5, T=28.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:18,740] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:25,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:08:27,172] loop.py:45 INFO TRAIN_STEP step=3519 loss=1.781647 Training: 20%|█▉ | 3519/18018 [7:38:13<33:53:12, 8.41s/step, loss=1.782, acc=0.628, acc_len=1.46, thru=15.2, I=17.9, T=31.0, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:27,330] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:34,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:08:35,844] loop.py:380 INFO TIMING step=3520: step=4.529s data=2.155s compute=4.493s [fwd=3.121s bwd=1.358s opt=0.003s] dispatch=4.131s [2026-10-18 21:08:35,844] loop.py:45 INFO TRAIN_STEP step=3520 loss=1.724993 Training: 20%|█▉ | 3520/18018 [7:38:21<34:11:41, 8.49s/step, loss=1.725, acc=0.633, acc_len=1.47, thru=15.0, I=18.4, T=28.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:36,286] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:41,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:08:42,872] loop.py:45 INFO TRAIN_STEP step=3521 loss=1.683074 Training: 20%|█▉ | 3521/18018 [7:38:28<32:25:33, 8.05s/step, loss=1.683, acc=0.639, acc_len=1.49, thru=16.6, I=22.3, T=33.0, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:08:42,867] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3520: forward=2728.4ms backward=1102.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:43,206] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:49,927] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:08:50,338] loop.py:45 INFO TRAIN_STEP step=3522 loss=1.662843 Training: 20%|█▉ | 3522/18018 [7:38:36<31:42:47, 7.88s/step, loss=1.663, acc=0.640, acc_len=1.50, thru=15.7, I=18.6, T=31.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:50,429] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:56,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:08:57,096] loop.py:45 INFO TRAIN_STEP step=3523 loss=1.719256 Training: 20%|█▉ | 3523/18018 [7:38:43<30:21:45, 7.54s/step, loss=1.719, acc=0.635, acc_len=1.48, thru=18.4, I=18.3, T=34.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:08:57,149] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:03,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:09:04,843] loop.py:45 INFO TRAIN_STEP step=3524 loss=1.806757 Training: 20%|█▉ | 3524/18018 [7:38:50<30:36:32, 7.60s/step, loss=1.807, acc=0.622, acc_len=1.43, thru=18.0, I=21.4, T=28.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:05,032] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:09,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:09:11,853] loop.py:380 INFO TIMING step=3525: step=4.618s data=2.084s compute=4.582s [fwd=3.278s bwd=1.289s opt=0.003s] dispatch=2.381s [2026-10-18 21:09:11,853] loop.py:45 INFO TRAIN_STEP step=3525 loss=1.658449 Training: 20%|█▉ | 3525/18018 [7:38:57<29:53:31, 7.43s/step, loss=1.658, acc=0.643, acc_len=1.52, thru=16.1, I=25.5, T=27.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:11,970] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:17,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:09:19,127] loop.py:45 INFO TRAIN_STEP step=3526 loss=1.750265 Training: 20%|█▉ | 3526/18018 [7:39:05<29:42:18, 7.38s/step, loss=1.750, acc=0.628, acc_len=1.46, thru=15.2, I=21.6, T=30.3, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:09:19,122] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3525: forward=2950.9ms backward=1224.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:19,244] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:25,674] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:09:26,181] loop.py:45 INFO TRAIN_STEP step=3527 loss=1.640259 Training: 20%|█▉ | 3527/18018 [7:39:12<29:18:34, 7.28s/step, loss=1.640, acc=0.647, acc_len=1.52, thru=17.9, I=21.3, T=29.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:26,680] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:31,860] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:09:33,025] loop.py:45 INFO TRAIN_STEP step=3528 loss=1.587566 Training: 20%|█▉ | 3528/18018 [7:39:19<28:46:58, 7.15s/step, loss=1.588, acc=0.654, acc_len=1.54, thru=18.3, I=20.8, T=32.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:33,370] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:38,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:09:40,029] loop.py:45 INFO TRAIN_STEP step=3529 loss=1.846734 Training: 20%|█▉ | 3529/18018 [7:39:26<28:36:09, 7.11s/step, loss=1.847, acc=0.611, acc_len=1.41, thru=17.5, I=25.0, T=30.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:40,271] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:44,753] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:09:46,324] loop.py:380 INFO TIMING step=3530: step=4.224s data=1.979s compute=4.186s [fwd=2.928s bwd=1.242s opt=0.003s] dispatch=2.062s [2026-10-18 21:09:46,325] loop.py:45 INFO TRAIN_STEP step=3530 loss=1.657485 Training: 20%|█▉ | 3530/18018 [7:39:32<27:37:03, 6.86s/step, loss=1.657, acc=0.645, acc_len=1.51, thru=18.1, I=26.2, T=30.3, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:09:46,758] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:09:56,089] loop.py:45 INFO TRAIN_STEP step=3531 loss=1.748832 Training: 20%|█▉ | 3531/18018 [7:39:42<31:07:32, 7.73s/step, loss=1.749, acc=0.628, acc_len=1.46, thru=11.6, I=11.8, T=29.0, wait=5.3s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:09:56,085] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3530: forward=3150.1ms backward=1206.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:03,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:10:04,012] loop.py:45 INFO TRAIN_STEP step=3532 loss=1.824774 Training: 20%|█▉ | 3532/18018 [7:39:50<31:20:54, 7.79s/step, loss=1.825, acc=0.616, acc_len=1.42, thru=17.2, I=17.5, T=27.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:04,511] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:09,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:10:10,976] loop.py:45 INFO TRAIN_STEP step=3533 loss=1.510197 Training: 20%|█▉ | 3533/18018 [7:39:57<30:20:55, 7.54s/step, loss=1.510, acc=0.669, acc_len=1.59, thru=16.3, I=20.9, T=30.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:11,150] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:17,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:10:18,689] loop.py:45 INFO TRAIN_STEP step=3534 loss=1.685264 Training: 20%|█▉ | 3534/18018 [7:40:04<30:33:09, 7.59s/step, loss=1.685, acc=0.641, acc_len=1.50, thru=15.3, I=19.8, T=32.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:18,760] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 21:10:26,429] loop.py:380 INFO TIMING step=3535: step=3.992s data=1.800s compute=3.960s [fwd=2.812s bwd=1.132s opt=0.003s] dispatch=3.737s [2026-10-18 21:10:26,429] loop.py:45 INFO TRAIN_STEP step=3535 loss=1.582956 Training: 20%|█▉ | 3535/18018 [7:40:12<30:43:23, 7.64s/step, loss=1.583, acc=0.658, acc_len=1.56, thru=15.0, I=15.7, T=32.1, wait=3.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:33,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:10:35,151] loop.py:45 INFO TRAIN_STEP step=3536 loss=1.649881 Training: 20%|█▉ | 3536/18018 [7:40:21<32:02:01, 7.96s/step, loss=1.650, acc=0.649, acc_len=1.53, thru=15.8, I=16.8, T=28.8, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:10:35,147] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3535: forward=3095.0ms backward=1293.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:35,535] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:42,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:10:44,024] loop.py:45 INFO TRAIN_STEP step=3537 loss=1.616921 Training: 20%|█▉ | 3537/18018 [7:40:30<33:07:46, 8.24s/step, loss=1.617, acc=0.650, acc_len=1.53, thru=15.1, I=19.1, T=28.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:44,196] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:48,839] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:10:51,542] loop.py:45 INFO TRAIN_STEP step=3538 loss=1.621205 Training: 20%|█▉ | 3538/18018 [7:40:37<32:15:30, 8.02s/step, loss=1.621, acc=0.651, acc_len=1.53, thru=14.3, I=24.7, T=27.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:51,859] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:57,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:10:58,765] loop.py:45 INFO TRAIN_STEP step=3539 loss=1.842084 Training: 20%|█▉ | 3539/18018 [7:40:44<31:17:51, 7.78s/step, loss=1.842, acc=0.612, acc_len=1.42, thru=14.0, I=22.1, T=31.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:10:58,930] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 21:11:05,988] loop.py:380 INFO TIMING step=3540: step=3.520s data=1.458s compute=3.477s [fwd=2.439s bwd=1.023s opt=0.003s] dispatch=3.693s [2026-10-18 21:11:05,988] loop.py:45 INFO TRAIN_STEP step=3540 loss=1.754488 Training: 20%|█▉ | 3540/18018 [7:40:52<30:37:08, 7.61s/step, loss=1.754, acc=0.626, acc_len=1.45, thru=13.9, I=14.1, T=36.4, wait=3.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:15,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:11:17,344] loop.py:45 INFO TRAIN_STEP step=3541 loss=1.703090 Training: 20%|█▉ | 3541/18018 [7:41:03<35:07:53, 8.74s/step, loss=1.703, acc=0.644, acc_len=1.50, thru=13.9, I=14.3, T=26.1, wait=6.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:11:17,340] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3540: forward=3333.9ms backward=1491.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:17,463] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 21:11:24,508] loop.py:45 INFO TRAIN_STEP step=3542 loss=1.683259 Training: 20%|█▉ | 3542/18018 [7:41:10<33:13:54, 8.26s/step, loss=1.683, acc=0.638, acc_len=1.50, thru=15.2, I=17.3, T=34.0, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:29,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:11:32,031] loop.py:45 INFO TRAIN_STEP step=3543 loss=1.700396 Training: 20%|█▉ | 3543/18018 [7:41:18<32:20:18, 8.04s/step, loss=1.700, acc=0.638, acc_len=1.49, thru=16.9, I=21.0, T=28.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:32,427] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:37,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:11:39,015] loop.py:45 INFO TRAIN_STEP step=3544 loss=1.676731 Training: 20%|█▉ | 3544/18018 [7:41:25<31:03:25, 7.72s/step, loss=1.677, acc=0.639, acc_len=1.49, thru=15.6, I=24.7, T=28.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:39,520] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:45,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:11:46,657] loop.py:380 INFO TIMING step=3545: step=4.024s data=1.833s compute=3.982s [fwd=2.822s bwd=1.145s opt=0.003s] dispatch=3.608s [2026-10-18 21:11:46,657] loop.py:45 INFO TRAIN_STEP step=3545 loss=1.699468 Training: 20%|█▉ | 3545/18018 [7:41:32<30:57:26, 7.70s/step, loss=1.699, acc=0.635, acc_len=1.49, thru=15.1, I=20.8, T=31.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:46,738] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:52,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:11:54,000] loop.py:45 INFO TRAIN_STEP step=3546 loss=1.745497 Training: 20%|█▉ | 3546/18018 [7:41:40<30:31:24, 7.59s/step, loss=1.745, acc=0.630, acc_len=1.47, thru=16.2, I=20.4, T=30.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:11:54,033] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 21:11:53,994] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3545: forward=2986.8ms backward=1156.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:02,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:12:02,479] loop.py:45 INFO TRAIN_STEP step=3547 loss=1.672750 Training: 20%|█▉ | 3547/18018 [7:41:48<31:35:16, 7.86s/step, loss=1.673, acc=0.641, acc_len=1.50, thru=15.2, I=15.3, T=34.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:02,712] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:10,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:12:11,463] loop.py:45 INFO TRAIN_STEP step=3548 loss=1.750339 Training: 20%|█▉ | 3548/18018 [7:41:57<32:56:45, 8.20s/step, loss=1.750, acc=0.633, acc_len=1.47, thru=14.5, I=15.1, T=32.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:11,797] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:17,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:12:19,566] loop.py:45 INFO TRAIN_STEP step=3549 loss=1.783075 Training: 20%|█▉ | 3549/18018 [7:42:05<32:49:34, 8.17s/step, loss=1.783, acc=0.622, acc_len=1.44, thru=14.9, I=18.9, T=27.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:20,028] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:26,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:12:27,067] loop.py:380 INFO TIMING step=3550: step=4.332s data=1.955s compute=4.292s [fwd=3.004s bwd=1.273s opt=0.003s] dispatch=3.163s [2026-10-18 21:12:27,067] loop.py:45 INFO TRAIN_STEP step=3550 loss=1.638048 Training: 20%|█▉ | 3550/18018 [7:42:13<32:01:22, 7.97s/step, loss=1.638, acc=0.646, acc_len=1.51, thru=14.5, I=20.8, T=29.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:27,354] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:33,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:12:33,975] loop.py:45 INFO TRAIN_STEP step=3551 loss=1.709592 Training: 20%|█▉ | 3551/18018 [7:42:20<30:44:38, 7.65s/step, loss=1.710, acc=0.636, acc_len=1.49, thru=18.0, I=19.1, T=32.3, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:12:33,970] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3550: forward=2763.8ms backward=1148.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:34,320] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:38,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:12:40,388] loop.py:45 INFO TRAIN_STEP step=3552 loss=1.525274 Training: 20%|█▉ | 3552/18018 [7:42:26<29:15:04, 7.28s/step, loss=1.525, acc=0.664, acc_len=1.58, thru=18.7, I=23.8, T=31.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:40,554] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:46,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:12:46,989] loop.py:45 INFO TRAIN_STEP step=3553 loss=1.598892 Training: 20%|█▉ | 3553/18018 [7:42:33<28:25:40, 7.08s/step, loss=1.599, acc=0.653, acc_len=1.54, thru=16.2, I=21.5, T=32.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:47,440] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:53,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:12:54,706] loop.py:45 INFO TRAIN_STEP step=3554 loss=1.850928 Training: 20%|█▉ | 3554/18018 [7:42:40<29:12:11, 7.27s/step, loss=1.851, acc=0.617, acc_len=1.42, thru=17.8, I=20.4, T=29.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:12:54,926] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 21:13:02,595] loop.py:380 INFO TIMING step=3555: step=4.331s data=2.137s compute=4.276s [fwd=3.007s bwd=1.255s opt=0.003s] dispatch=3.547s [2026-10-18 21:13:02,595] loop.py:45 INFO TRAIN_STEP step=3555 loss=1.717281 Training: 20%|█▉ | 3555/18018 [7:42:48<29:56:39, 7.45s/step, loss=1.717, acc=0.635, acc_len=1.48, thru=14.7, I=15.2, T=29.6, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:11,279] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:13:12,842] loop.py:45 INFO TRAIN_STEP step=3556 loss=1.844715 Training: 20%|█▉ | 3556/18018 [7:42:58<33:18:36, 8.29s/step, loss=1.845, acc=0.613, acc_len=1.41, thru=14.4, I=16.5, T=27.6, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:13:12,837] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3555: forward=3133.8ms backward=1458.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:13,289] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:17,912] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:13:18,713] loop.py:45 INFO TRAIN_STEP step=3557 loss=1.632951 Training: 20%|█▉ | 3557/18018 [7:43:04<30:23:39, 7.57s/step, loss=1.633, acc=0.648, acc_len=1.52, thru=18.6, I=23.5, T=35.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:18,917] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:23,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:13:25,152] loop.py:45 INFO TRAIN_STEP step=3558 loss=1.626881 Training: 20%|█▉ | 3558/18018 [7:43:11<29:02:00, 7.23s/step, loss=1.627, acc=0.649, acc_len=1.53, thru=19.5, I=24.1, T=32.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:25,452] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:31,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:13:33,606] loop.py:45 INFO TRAIN_STEP step=3559 loss=1.599230 Training: 20%|█▉ | 3559/18018 [7:43:19<30:30:19, 7.60s/step, loss=1.599, acc=0.655, acc_len=1.55, thru=13.7, I=18.0, T=30.5, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:33,977] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:38,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:13:40,357] loop.py:380 INFO TIMING step=3560: step=4.019s data=1.892s compute=3.989s [fwd=2.851s bwd=1.123s opt=0.003s] dispatch=2.723s [2026-10-18 21:13:40,358] loop.py:45 INFO TRAIN_STEP step=3560 loss=1.626864 Training: 20%|█▉ | 3560/18018 [7:43:26<29:29:26, 7.34s/step, loss=1.627, acc=0.649, acc_len=1.53, thru=16.4, I=24.4, T=31.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:40,429] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:46,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:13:46,584] loop.py:45 INFO TRAIN_STEP step=3561 loss=1.709559 Training: 20%|█▉ | 3561/18018 [7:43:32<28:08:28, 7.01s/step, loss=1.710, acc=0.636, acc_len=1.48, thru=19.0, I=26.5, T=33.9, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:13:46,579] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3560: forward=2666.7ms backward=1056.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:46,924] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:51,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:13:53,371] loop.py:45 INFO TRAIN_STEP step=3562 loss=1.593774 Training: 20%|█▉ | 3562/18018 [7:43:39<27:52:23, 6.94s/step, loss=1.594, acc=0.655, acc_len=1.54, thru=18.1, I=21.6, T=31.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:53,825] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:13:59,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:14:01,232] loop.py:45 INFO TRAIN_STEP step=3563 loss=1.737674 Training: 20%|█▉ | 3563/18018 [7:43:47<28:58:39, 7.22s/step, loss=1.738, acc=0.631, acc_len=1.47, thru=14.3, I=21.0, T=27.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:01,428] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:06,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:14:07,110] loop.py:45 INFO TRAIN_STEP step=3564 loss=1.514237 Training: 20%|█▉ | 3564/18018 [7:43:53<27:21:51, 6.82s/step, loss=1.514, acc=0.665, acc_len=1.58, thru=18.8, I=23.8, T=35.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:07,551] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:13,601] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:14:14,218] loop.py:380 INFO TIMING step=3565: step=3.383s data=1.398s compute=3.347s [fwd=2.371s bwd=0.961s opt=0.003s] dispatch=3.715s [2026-10-18 21:14:14,218] loop.py:45 INFO TRAIN_STEP step=3565 loss=1.724903 Training: 20%|█▉ | 3565/18018 [7:44:00<27:42:48, 6.90s/step, loss=1.725, acc=0.633, acc_len=1.47, thru=17.4, I=21.8, T=37.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:14,608] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 21:14:21,281] loop.py:45 INFO TRAIN_STEP step=3566 loss=1.703419 Training: 20%|█▉ | 3566/18018 [7:44:07<27:54:16, 6.95s/step, loss=1.703, acc=0.638, acc_len=1.48, thru=17.8, I=19.4, T=32.5, wait=3.1s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:14:21,276] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3565: forward=2774.1ms backward=1109.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:28,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:14:30,907] loop.py:45 INFO TRAIN_STEP step=3567 loss=1.747400 Training: 20%|█▉ | 3567/18018 [7:44:16<31:07:20, 7.75s/step, loss=1.747, acc=0.626, acc_len=1.46, thru=14.2, I=16.6, T=27.2, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:31,404] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:38,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:14:40,321] loop.py:45 INFO TRAIN_STEP step=3568 loss=1.592150 Training: 20%|█▉ | 3568/18018 [7:44:26<33:07:24, 8.25s/step, loss=1.592, acc=0.657, acc_len=1.55, thru=14.4, I=16.9, T=29.4, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:40,463] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:45,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:14:46,308] loop.py:45 INFO TRAIN_STEP step=3569 loss=1.729035 Training: 20%|█▉ | 3569/18018 [7:44:32<30:23:32, 7.57s/step, loss=1.729, acc=0.626, acc_len=1.45, thru=19.3, I=20.5, T=34.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:46,550] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:53,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:14:54,966] loop.py:380 INFO TIMING step=3570: step=4.050s data=1.900s compute=4.010s [fwd=2.823s bwd=1.172s opt=0.003s] dispatch=4.599s [2026-10-18 21:14:54,966] loop.py:45 INFO TRAIN_STEP step=3570 loss=1.796189 Training: 20%|█▉ | 3570/18018 [7:44:41<31:41:47, 7.90s/step, loss=1.796, acc=0.619, acc_len=1.43, thru=14.1, I=18.8, T=31.6, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:14:55,108] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:01,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:15:01,330] loop.py:45 INFO TRAIN_STEP step=3571 loss=1.640422 Training: 20%|█▉ | 3571/18018 [7:44:47<29:50:45, 7.44s/step, loss=1.640, acc=0.645, acc_len=1.51, thru=18.6, I=21.6, T=34.4, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:15:01,325] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3570: forward=2636.4ms backward=1038.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:01,779] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:08,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:15:08,755] loop.py:45 INFO TRAIN_STEP step=3572 loss=1.704110 Training: 20%|█▉ | 3572/18018 [7:44:54<29:49:44, 7.43s/step, loss=1.704, acc=0.637, acc_len=1.49, thru=18.3, I=17.0, T=29.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:08,932] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:15,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:15:16,039] loop.py:45 INFO TRAIN_STEP step=3573 loss=1.745352 Training: 20%|█▉ | 3573/18018 [7:45:02<29:39:01, 7.39s/step, loss=1.745, acc=0.629, acc_len=1.46, thru=17.7, I=16.6, T=33.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:16,441] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:21,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:15:23,123] loop.py:45 INFO TRAIN_STEP step=3574 loss=1.763435 Training: 20%|█▉ | 3574/18018 [7:45:09<29:16:52, 7.30s/step, loss=1.763, acc=0.627, acc_len=1.46, thru=18.7, I=21.7, T=31.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:23,160] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:28,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:15:30,239] loop.py:380 INFO TIMING step=3575: step=4.619s data=2.159s compute=4.578s [fwd=3.129s bwd=1.434s opt=0.003s] dispatch=2.486s [2026-10-18 21:15:30,239] loop.py:45 INFO TRAIN_STEP step=3575 loss=1.671327 Training: 20%|█▉ | 3575/18018 [7:45:16<29:03:30, 7.24s/step, loss=1.671, acc=0.643, acc_len=1.51, thru=17.4, I=24.0, T=27.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:30,360] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:36,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:15:39,109] loop.py:45 INFO TRAIN_STEP step=3576 loss=1.625013 Training: 20%|█▉ | 3576/18018 [7:45:25<31:00:49, 7.73s/step, loss=1.625, acc=0.651, acc_len=1.53, thru=14.7, I=17.4, T=29.2, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:15:39,104] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3575: forward=3127.3ms backward=1204.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:39,538] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:44,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:15:46,317] loop.py:45 INFO TRAIN_STEP step=3577 loss=1.600856 Training: 20%|█▉ | 3577/18018 [7:45:32<30:22:58, 7.57s/step, loss=1.601, acc=0.656, acc_len=1.55, thru=13.5, I=21.1, T=29.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:46,697] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:51,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:15:54,184] loop.py:45 INFO TRAIN_STEP step=3578 loss=1.701314 Training: 20%|█▉ | 3578/18018 [7:45:40<30:44:06, 7.66s/step, loss=1.701, acc=0.639, acc_len=1.49, thru=14.5, I=22.9, T=26.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:54,317] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:15:59,457] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:16:00,599] loop.py:45 INFO TRAIN_STEP step=3579 loss=1.643109 Training: 20%|█▉ | 3579/18018 [7:45:46<29:13:54, 7.29s/step, loss=1.643, acc=0.643, acc_len=1.51, thru=16.2, I=24.9, T=32.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:00,968] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:07,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:16:08,937] loop.py:380 INFO TIMING step=3580: step=4.764s data=2.322s compute=4.735s [fwd=3.204s bwd=1.516s opt=0.003s] dispatch=3.563s [2026-10-18 21:16:08,937] loop.py:45 INFO TRAIN_STEP step=3580 loss=1.595914 Training: 20%|█▉ | 3580/18018 [7:45:55<30:29:32, 7.60s/step, loss=1.596, acc=0.652, acc_len=1.54, thru=14.2, I=18.6, T=26.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:09,249] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:14,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 21:16:15,049] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3580: forward=2632.0ms backward=1027.9ms [2026-10-18 21:16:15,056] loop.py:45 INFO TRAIN_STEP step=3581 loss=1.603517 Training: 20%|█▉ | 3581/18018 [7:46:01<28:42:26, 7.16s/step, loss=1.604, acc=0.653, acc_len=1.54, thru=19.8, I=22.1, T=34.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:15,370] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:19,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:16:21,820] loop.py:45 INFO TRAIN_STEP step=3582 loss=1.678742 Training: 20%|█▉ | 3582/18018 [7:46:07<28:13:43, 7.04s/step, loss=1.679, acc=0.638, acc_len=1.49, thru=18.9, I=25.9, T=31.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:22,310] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:28,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:16:29,287] loop.py:45 INFO TRAIN_STEP step=3583 loss=1.764078 Training: 20%|█▉ | 3583/18018 [7:46:15<28:44:29, 7.17s/step, loss=1.764, acc=0.624, acc_len=1.45, thru=14.8, I=23.5, T=29.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:29,516] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:35,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:16:37,471] loop.py:45 INFO TRAIN_STEP step=3584 loss=1.564346 Training: 20%|█▉ | 3584/18018 [7:46:23<29:57:30, 7.47s/step, loss=1.564, acc=0.661, acc_len=1.57, thru=14.8, I=21.6, T=29.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:37,974] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:16:46,472] loop.py:380 INFO TIMING step=3585: step=3.561s data=1.500s compute=3.526s [fwd=2.525s bwd=0.986s opt=0.003s] dispatch=5.432s [2026-10-18 21:16:46,472] loop.py:45 INFO TRAIN_STEP step=3585 loss=1.679152 Training: 20%|█▉ | 3585/18018 [7:46:32<31:47:41, 7.93s/step, loss=1.679, acc=0.639, acc_len=1.50, thru=13.2, I=14.2, T=35.9, wait=5.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:56,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:16:56,940] loop.py:45 INFO TRAIN_STEP step=3586 loss=1.738501 Training: 20%|█▉ | 3586/18018 [7:46:43<34:50:46, 8.69s/step, loss=1.739, acc=0.631, acc_len=1.47, thru=14.1, I=14.1, T=32.9, wait=6.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:16:56,936] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3585: forward=2724.7ms backward=1115.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:16:57,198] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:02,689] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:17:04,395] loop.py:45 INFO TRAIN_STEP step=3587 loss=1.690395 Training: 20%|█▉ | 3587/18018 [7:46:50<33:21:25, 8.32s/step, loss=1.690, acc=0.638, acc_len=1.49, thru=16.6, I=20.7, T=32.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:04,705] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:10,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:17:11,421] loop.py:45 INFO TRAIN_STEP step=3588 loss=1.591252 Training: 20%|█▉ | 3588/18018 [7:46:57<31:47:41, 7.93s/step, loss=1.591, acc=0.655, acc_len=1.54, thru=15.9, I=20.9, T=31.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:11,509] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:19,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:17:20,769] loop.py:45 INFO TRAIN_STEP step=3589 loss=1.762277 Training: 20%|█▉ | 3589/18018 [7:47:06<33:29:51, 8.36s/step, loss=1.762, acc=0.626, acc_len=1.45, thru=14.1, I=16.3, T=29.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:20,864] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 21:17:27,636] loop.py:380 INFO TIMING step=3590: step=3.952s data=1.711s compute=3.927s [fwd=2.783s bwd=1.129s opt=0.003s] dispatch=2.904s [2026-10-18 21:17:27,636] loop.py:45 INFO TRAIN_STEP step=3590 loss=1.615360 Training: 20%|█▉ | 3590/18018 [7:47:13<31:42:06, 7.91s/step, loss=1.615, acc=0.650, acc_len=1.54, thru=14.2, I=15.7, T=32.4, wait=2.9s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:36,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:17:38,012] loop.py:45 INFO TRAIN_STEP step=3591 loss=1.690977 Training: 20%|█▉ | 3591/18018 [7:47:24<34:39:48, 8.65s/step, loss=1.691, acc=0.642, acc_len=1.50, thru=14.8, I=17.5, T=29.0, wait=5.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:17:38,007] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3590: forward=3126.2ms backward=1238.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:38,335] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:17:46,839] loop.py:45 INFO TRAIN_STEP step=3592 loss=1.622826 Training: 20%|█▉ | 3592/18018 [7:47:32<34:52:39, 8.70s/step, loss=1.623, acc=0.650, acc_len=1.53, thru=14.6, I=16.4, T=27.0, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:55,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:17:55,633] loop.py:45 INFO TRAIN_STEP step=3593 loss=1.734459 Training: 20%|█▉ | 3593/18018 [7:47:41<34:58:48, 8.73s/step, loss=1.734, acc=0.631, acc_len=1.47, thru=14.5, I=12.8, T=31.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:17:55,696] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:03,483] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:18:04,419] loop.py:45 INFO TRAIN_STEP step=3594 loss=1.806602 Training: 20%|█▉ | 3594/18018 [7:47:50<35:02:42, 8.75s/step, loss=1.807, acc=0.620, acc_len=1.44, thru=15.2, I=17.0, T=36.0, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:04,500] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:10,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:18:11,057] loop.py:380 INFO TIMING step=3595: step=4.195s data=2.058s compute=4.163s [fwd=2.839s bwd=1.309s opt=0.003s] dispatch=2.433s [2026-10-18 21:18:11,057] loop.py:45 INFO TRAIN_STEP step=3595 loss=1.615078 Training: 20%|█▉ | 3595/18018 [7:47:57<32:30:38, 8.11s/step, loss=1.615, acc=0.649, acc_len=1.52, thru=19.5, I=21.8, T=30.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:11,266] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:16,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:18:18,336] loop.py:45 INFO TRAIN_STEP step=3596 loss=1.688215 Training: 20%|█▉ | 3596/18018 [7:48:04<31:30:13, 7.86s/step, loss=1.688, acc=0.641, acc_len=1.50, thru=16.5, I=21.6, T=28.4, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:18:18,331] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3595: forward=3111.4ms backward=1347.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:18,531] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:25,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:18:26,635] loop.py:45 INFO TRAIN_STEP step=3597 loss=1.638669 Training: 20%|█▉ | 3597/18018 [7:48:12<32:01:17, 7.99s/step, loss=1.639, acc=0.648, acc_len=1.52, thru=15.8, I=18.9, T=30.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:27,136] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:32,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:18:34,270] loop.py:45 INFO TRAIN_STEP step=3598 loss=1.685174 Training: 20%|█▉ | 3598/18018 [7:48:20<31:35:29, 7.89s/step, loss=1.685, acc=0.641, acc_len=1.50, thru=15.0, I=19.9, T=32.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:34,404] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 21:18:40,626] loop.py:45 INFO TRAIN_STEP step=3599 loss=1.621488 Training: 20%|█▉ | 3599/18018 [7:48:26<29:44:46, 7.43s/step, loss=1.621, acc=0.648, acc_len=1.52, thru=17.2, I=19.6, T=30.5, wait=2.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:49,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:18:50,090] loop.py:380 INFO TIMING step=3600: step=4.326s data=1.962s compute=4.296s [fwd=3.045s bwd=1.236s opt=0.003s] dispatch=5.130s [2026-10-18 21:18:50,090] loop.py:45 INFO TRAIN_STEP step=3600 loss=1.681771 Training: 20%|█▉ | 3600/18018 [7:48:36<32:11:41, 8.04s/step, loss=1.682, acc=0.639, acc_len=1.50, thru=15.8, I=14.2, T=29.6, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:50,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:56,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:18:57,865] loop.py:45 INFO TRAIN_STEP step=3601 loss=1.750502 Training: 20%|█▉ | 3601/18018 [7:48:43<31:52:35, 7.96s/step, loss=1.751, acc=0.627, acc_len=1.45, thru=16.5, I=19.4, T=28.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:18:57,925] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (TrainerActor pid=9310) [2026-10-18 21:18:57,860] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3600: forward=3081.7ms backward=1319.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:08,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:19:08,842] loop.py:45 INFO TRAIN_STEP step=3602 loss=1.777009 Training: 20%|█▉ | 3602/18018 [7:48:54<35:29:53, 8.86s/step, loss=1.777, acc=0.627, acc_len=1.45, thru=13.4, I=12.3, T=24.8, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:09,012] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:15,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:19:16,041] loop.py:45 INFO TRAIN_STEP step=3603 loss=1.817501 Training: 20%|█▉ | 3603/18018 [7:49:02<33:29:35, 8.36s/step, loss=1.818, acc=0.621, acc_len=1.44, thru=15.7, I=15.9, T=31.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:16,187] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:22,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:19:23,100] loop.py:45 INFO TRAIN_STEP step=3604 loss=1.701294 Training: 20%|██ | 3604/18018 [7:49:09<31:55:28, 7.97s/step, loss=1.701, acc=0.635, acc_len=1.49, thru=18.3, I=19.2, T=28.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:23,402] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:29,940] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:19:31,783] loop.py:380 INFO TIMING step=3605: step=4.174s data=1.729s compute=4.134s [fwd=2.871s bwd=1.247s opt=0.003s] dispatch=4.498s [2026-10-18 21:19:31,783] loop.py:45 INFO TRAIN_STEP step=3605 loss=1.693017 Training: 20%|██ | 3605/18018 [7:49:17<32:46:30, 8.19s/step, loss=1.693, acc=0.637, acc_len=1.49, thru=14.4, I=19.2, T=30.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:31,956] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:37,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:19:39,192] loop.py:45 INFO TRAIN_STEP step=3606 loss=1.805241 Training: 20%|██ | 3606/18018 [7:49:25<31:50:15, 7.95s/step, loss=1.805, acc=0.619, acc_len=1.43, thru=15.2, I=21.9, T=29.5, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:19:39,187] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3605: forward=3027.5ms backward=1267.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:39,516] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:44,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:19:46,875] loop.py:45 INFO TRAIN_STEP step=3607 loss=1.688144 Training: 20%|██ | 3607/18018 [7:49:32<31:30:45, 7.87s/step, loss=1.688, acc=0.643, acc_len=1.50, thru=15.3, I=22.6, T=27.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:46,909] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:53,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:19:54,967] loop.py:45 INFO TRAIN_STEP step=3608 loss=1.813800 Training: 20%|██ | 3608/18018 [7:49:41<31:46:23, 7.94s/step, loss=1.814, acc=0.620, acc_len=1.44, thru=15.9, I=18.2, T=28.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:19:55,401] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:01,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:20:01,952] loop.py:45 INFO TRAIN_STEP step=3609 loss=1.621757 Training: 20%|██ | 3609/18018 [7:49:48<30:37:31, 7.65s/step, loss=1.622, acc=0.648, acc_len=1.53, thru=16.2, I=21.1, T=31.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:02,431] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:08,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:20:09,418] loop.py:380 INFO TIMING step=3610: step=4.080s data=1.939s compute=4.050s [fwd=2.902s bwd=1.133s opt=0.003s] dispatch=3.379s [2026-10-18 21:20:09,419] loop.py:45 INFO TRAIN_STEP step=3610 loss=1.767112 Training: 20%|██ | 3610/18018 [7:49:55<30:24:14, 7.60s/step, loss=1.767, acc=0.624, acc_len=1.44, thru=16.9, I=19.6, T=31.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:09,538] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 21:20:16,000] loop.py:45 INFO TRAIN_STEP step=3611 loss=1.704722 Training: 20%|██ | 3611/18018 [7:50:02<29:10:51, 7.29s/step, loss=1.705, acc=0.634, acc_len=1.48, thru=17.0, I=20.7, T=33.6, wait=2.8s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:20:15,994] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3610: forward=2690.2ms backward=1071.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:25,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:20:26,755] loop.py:45 INFO TRAIN_STEP step=3612 loss=1.779503 Training: 20%|██ | 3612/18018 [7:50:12<33:20:07, 8.33s/step, loss=1.780, acc=0.627, acc_len=1.45, thru=13.2, I=14.0, T=25.7, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:27,259] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:32,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:20:33,721] loop.py:45 INFO TRAIN_STEP step=3613 loss=1.801960 Training: 20%|██ | 3613/18018 [7:50:19<31:41:57, 7.92s/step, loss=1.802, acc=0.622, acc_len=1.44, thru=16.4, I=21.1, T=30.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:33,873] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 21:20:41,469] loop.py:45 INFO TRAIN_STEP step=3614 loss=1.579859 Training: 20%|██ | 3614/18018 [7:50:27<31:29:04, 7.87s/step, loss=1.580, acc=0.660, acc_len=1.56, thru=16.1, I=15.6, T=26.8, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:49,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:20:50,894] loop.py:380 INFO TIMING step=3615: step=4.350s data=1.946s compute=4.314s [fwd=2.994s bwd=1.304s opt=0.003s] dispatch=5.066s [2026-10-18 21:20:50,894] loop.py:45 INFO TRAIN_STEP step=3615 loss=1.902116 Training: 20%|██ | 3615/18018 [7:50:36<33:21:02, 8.34s/step, loss=1.902, acc=0.603, acc_len=1.38, thru=14.8, I=16.4, T=29.4, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:51,082] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:57,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:20:57,786] loop.py:45 INFO TRAIN_STEP step=3616 loss=1.760677 Training: 20%|██ | 3616/18018 [7:50:43<31:36:53, 7.90s/step, loss=1.761, acc=0.626, acc_len=1.46, thru=17.2, I=18.4, T=33.2, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:20:57,782] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3615: forward=2742.4ms backward=1056.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:20:58,223] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 21:21:04,970] loop.py:45 INFO TRAIN_STEP step=3617 loss=1.709625 Training: 20%|██ | 3617/18018 [7:50:51<30:45:10, 7.69s/step, loss=1.710, acc=0.635, acc_len=1.49, thru=18.1, I=19.2, T=28.4, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:12,929] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:21:14,621] loop.py:45 INFO TRAIN_STEP step=3618 loss=1.664510 Training: 20%|██ | 3618/18018 [7:51:00<33:06:21, 8.28s/step, loss=1.665, acc=0.643, acc_len=1.51, thru=13.8, I=14.5, T=24.2, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:14,943] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:21,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:21:22,262] loop.py:45 INFO TRAIN_STEP step=3619 loss=1.610891 Training: 20%|██ | 3619/18018 [7:51:08<32:20:23, 8.09s/step, loss=1.611, acc=0.654, acc_len=1.54, thru=15.0, I=20.3, T=26.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:22,353] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:28,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:21:29,947] loop.py:380 INFO TIMING step=3620: step=4.211s data=1.903s compute=4.182s [fwd=2.988s bwd=1.179s opt=0.003s] dispatch=3.465s [2026-10-18 21:21:29,947] loop.py:45 INFO TRAIN_STEP step=3620 loss=1.629645 Training: 20%|██ | 3620/18018 [7:51:16<31:51:25, 7.97s/step, loss=1.630, acc=0.652, acc_len=1.55, thru=15.7, I=17.6, T=30.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:30,110] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:38,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:21:38,424] loop.py:45 INFO TRAIN_STEP step=3621 loss=1.705124 Training: 20%|██ | 3621/18018 [7:51:24<32:28:16, 8.12s/step, loss=1.705, acc=0.637, acc_len=1.48, thru=15.4, I=15.0, T=29.2, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:21:38,419] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3620: forward=3113.3ms backward=1219.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:38,559] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:45,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:21:48,335] loop.py:45 INFO TRAIN_STEP step=3622 loss=1.709645 Training: 20%|██ | 3622/18018 [7:51:34<34:37:05, 8.66s/step, loss=1.710, acc=0.640, acc_len=1.50, thru=13.1, I=16.7, T=26.9, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:21:48,787] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 21:21:56,174] loop.py:45 INFO TRAIN_STEP step=3623 loss=1.711691 Training: 20%|██ | 3623/18018 [7:51:42<33:37:50, 8.41s/step, loss=1.712, acc=0.637, acc_len=1.49, thru=16.2, I=16.5, T=27.9, wait=3.2s, pool=56, epoch=1/2][2026-10-18 21:22:04,649] loop.py:45 INFO TRAIN_STEP step=3624 loss=1.732860 Training: 20%|██ | 3624/18018 [7:51:50<33:42:25, 8.43s/step, loss=1.733, acc=0.636, acc_len=1.48, thru=14.5, I=14.4, T=34.2, wait=4.7s, pool=52, epoch=1/2][2026-10-18 21:22:14,387] loop.py:380 INFO TIMING step=3625: step=5.196s data=2.916s compute=5.164s [fwd=3.957s bwd=1.192s opt=0.003s] dispatch=4.533s [2026-10-18 21:22:14,387] loop.py:45 INFO TRAIN_STEP step=3625 loss=1.617656 Training: 20%|██ | 3625/18018 [7:52:00<35:16:32, 8.82s/step, loss=1.618, acc=0.653, acc_len=1.54, thru=13.8, I=14.8, T=24.6, wait=4.5s, pool=56, epoch=1/2][2026-10-18 21:22:22,832] loop.py:45 INFO TRAIN_STEP step=3626 loss=1.678076 Training: 20%|██ | 3626/18018 [7:52:08<34:49:11, 8.71s/step, loss=1.678, acc=0.638, acc_len=1.50, thru=12.1, I=14.3, T=30.9, wait=4.3s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:22:22,827] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3625: forward=2970.5ms backward=1133.3ms [2026-10-18 21:22:30,657] loop.py:45 INFO TRAIN_STEP step=3627 loss=1.595198 Training: 20%|██ | 3627/18018 [7:52:16<33:45:12, 8.44s/step, loss=1.595, acc=0.657, acc_len=1.55, thru=16.9, I=16.9, T=29.9, wait=3.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:22:37,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:22:40,061] loop.py:45 INFO TRAIN_STEP step=3628 loss=1.652679 Training: 20%|██ | 3628/18018 [7:52:26<34:54:20, 8.73s/step, loss=1.653, acc=0.646, acc_len=1.52, thru=14.7, I=15.7, T=28.4, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:22:40,371] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:22:47,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:22:49,427] loop.py:45 INFO TRAIN_STEP step=3629 loss=1.562657 Training: 20%|██ | 3629/18018 [7:52:35<35:39:34, 8.92s/step, loss=1.563, acc=0.659, acc_len=1.56, thru=14.1, I=17.4, T=27.9, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:22:49,478] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:22:54,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:22:55,835] loop.py:380 INFO TIMING step=3630: step=3.933s data=1.821s compute=3.893s [fwd=2.763s bwd=1.117s opt=0.003s] dispatch=2.466s [2026-10-18 21:22:55,835] loop.py:45 INFO TRAIN_STEP step=3630 loss=1.609925 Training: 20%|██ | 3630/18018 [7:52:41<32:38:36, 8.17s/step, loss=1.610, acc=0.655, acc_len=1.55, thru=17.2, I=25.4, T=32.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:22:56,156] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:23:06,063] loop.py:45 INFO TRAIN_STEP step=3631 loss=1.609745 Training: 20%|██ | 3631/18018 [7:52:52<35:06:51, 8.79s/step, loss=1.610, acc=0.658, acc_len=1.55, thru=12.5, I=12.6, T=31.8, wait=6.2s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:23:06,059] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3630: forward=2809.2ms backward=1175.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:12,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:23:13,786] loop.py:45 INFO TRAIN_STEP step=3632 loss=1.790139 Training: 20%|██ | 3632/18018 [7:52:59<33:50:12, 8.47s/step, loss=1.790, acc=0.625, acc_len=1.45, thru=17.5, I=19.2, T=29.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:13,913] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:19,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:23:20,947] loop.py:45 INFO TRAIN_STEP step=3633 loss=1.685444 Training: 20%|██ | 3633/18018 [7:53:07<32:16:03, 8.08s/step, loss=1.685, acc=0.636, acc_len=1.49, thru=16.8, I=20.3, T=32.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:21,054] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:25,173] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:23:26,745] loop.py:45 INFO TRAIN_STEP step=3634 loss=1.673817 Training: 20%|██ | 3634/18018 [7:53:12<29:32:07, 7.39s/step, loss=1.674, acc=0.641, acc_len=1.50, thru=20.4, I=29.3, T=32.4, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:27,185] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:32,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:23:33,372] loop.py:380 INFO TIMING step=3635: step=3.712s data=1.704s compute=3.680s [fwd=2.600s bwd=1.066s opt=0.003s] dispatch=2.905s [2026-10-18 21:23:33,373] loop.py:45 INFO TRAIN_STEP step=3635 loss=1.693679 Training: 20%|██ | 3635/18018 [7:53:19<28:37:03, 7.16s/step, loss=1.694, acc=0.637, acc_len=1.49, thru=17.8, I=23.2, T=34.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:33,410] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 21:23:39,965] loop.py:45 INFO TRAIN_STEP step=3636 loss=1.595577 Training: 20%|██ | 3636/18018 [7:53:26<27:55:56, 6.99s/step, loss=1.596, acc=0.657, acc_len=1.55, thru=17.3, I=17.6, T=34.6, wait=2.9s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:23:39,960] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3635: forward=2570.6ms backward=1082.6ms [2026-10-18 21:23:48,364] loop.py:45 INFO TRAIN_STEP step=3637 loss=1.712640 Training: 20%|██ | 3637/18018 [7:53:34<29:36:52, 7.41s/step, loss=1.713, acc=0.635, acc_len=1.48, thru=16.9, I=16.6, T=31.8, wait=4.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:56,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:23:57,603] loop.py:45 INFO TRAIN_STEP step=3638 loss=1.677173 Training: 20%|██ | 3638/18018 [7:53:43<31:47:52, 7.96s/step, loss=1.677, acc=0.644, acc_len=1.51, thru=15.7, I=15.6, T=30.5, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:23:57,685] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:02,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:24:03,636] loop.py:45 INFO TRAIN_STEP step=3639 loss=1.635186 Training: 20%|██ | 3639/18018 [7:53:49<29:29:13, 7.38s/step, loss=1.635, acc=0.648, acc_len=1.52, thru=18.9, I=23.4, T=34.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:03,827] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:24:09,809] loop.py:380 INFO TIMING step=3640: step=3.477s data=1.491s compute=3.447s [fwd=2.487s bwd=0.945s opt=0.003s] dispatch=2.688s [2026-10-18 21:24:09,810] loop.py:45 INFO TRAIN_STEP step=3640 loss=1.733893 Training: 20%|██ | 3640/18018 [7:53:55<28:02:15, 7.02s/step, loss=1.734, acc=0.632, acc_len=1.47, thru=18.2, I=24.6, T=36.8, wait=2.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:17,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 21:24:18,817] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3640: forward=3427.1ms backward=1517.5ms [2026-10-18 21:24:18,822] loop.py:45 INFO TRAIN_STEP step=3641 loss=1.655097 Training: 20%|██ | 3641/18018 [7:54:04<30:25:20, 7.62s/step, loss=1.655, acc=0.641, acc_len=1.50, thru=14.8, I=16.5, T=25.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:18,906] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:25,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:24:25,763] loop.py:45 INFO TRAIN_STEP step=3642 loss=1.631428 Training: 20%|██ | 3642/18018 [7:54:11<29:36:34, 7.41s/step, loss=1.631, acc=0.650, acc_len=1.53, thru=15.3, I=18.0, T=29.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:26,022] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:31,131] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:24:32,965] loop.py:45 INFO TRAIN_STEP step=3643 loss=1.771308 Training: 20%|██ | 3643/18018 [7:54:19<29:21:20, 7.35s/step, loss=1.771, acc=0.626, acc_len=1.45, thru=16.4, I=21.0, T=26.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:33,147] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:38,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:24:40,835] loop.py:45 INFO TRAIN_STEP step=3644 loss=1.678071 Training: 20%|██ | 3644/18018 [7:54:26<29:58:16, 7.51s/step, loss=1.678, acc=0.639, acc_len=1.49, thru=14.0, I=21.4, T=28.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:41,332] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 21:24:49,245] loop.py:380 INFO TIMING step=3645: step=3.936s data=1.822s compute=3.893s [fwd=2.752s bwd=1.127s opt=0.003s] dispatch=4.466s [2026-10-18 21:24:49,245] loop.py:45 INFO TRAIN_STEP step=3645 loss=1.540801 Training: 20%|██ | 3645/18018 [7:54:35<31:03:13, 7.78s/step, loss=1.541, acc=0.660, acc_len=1.56, thru=12.6, I=15.2, T=32.5, wait=4.5s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:57,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:24:58,843] loop.py:45 INFO TRAIN_STEP step=3646 loss=1.766810 Training: 20%|██ | 3646/18018 [7:54:44<33:13:45, 8.32s/step, loss=1.767, acc=0.622, acc_len=1.44, thru=16.6, I=15.7, T=28.4, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:24:58,837] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3645: forward=3181.6ms backward=1269.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:24:59,191] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:25:05,648] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:25:06,978] loop.py:45 INFO TRAIN_STEP step=3647 loss=1.661648 Training: 20%|██ | 3647/18018 [7:54:53<33:00:15, 8.27s/step, loss=1.662, acc=0.642, acc_len=1.51, thru=14.4, I=19.2, T=31.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:25:07,175] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:25:13,390] loop.py:45 INFO TRAIN_STEP step=3648 loss=1.646797 Training: 20%|██ | 3648/18018 [7:54:59<30:46:38, 7.71s/step, loss=1.647, acc=0.645, acc_len=1.52, thru=17.8, I=20.3, T=36.0, wait=2.8s, pool=56, epoch=1/2][2026-10-18 21:25:22,184] loop.py:45 INFO TRAIN_STEP step=3649 loss=1.729012 Training: 20%|██ | 3649/18018 [7:55:08<32:04:25, 8.04s/step, loss=1.729, acc=0.635, acc_len=1.47, thru=14.2, I=15.5, T=29.8, wait=4.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:25:31,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:25:32,290] loop.py:380 INFO TIMING step=3650: step=4.272s data=1.920s compute=4.232s [fwd=3.043s bwd=1.174s opt=0.003s] dispatch=5.824s [2026-10-18 21:25:32,290] loop.py:45 INFO TRAIN_STEP step=3650 loss=1.751119 Training: 20%|██ | 3650/18018 [7:55:18<34:33:00, 8.66s/step, loss=1.751, acc=0.632, acc_len=1.47, thru=14.5, I=15.2, T=30.0, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:25:32,362] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-18 21:25:39,445] loop.py:45 INFO TRAIN_STEP step=3651 loss=1.562562 Training: 20%|██ | 3651/18018 [7:55:25<32:45:03, 8.21s/step, loss=1.563, acc=0.661, acc_len=1.56, thru=11.8, I=15.6, T=32.6, wait=3.2s, pool=20, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:25:39,440] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3650: forward=2685.9ms backward=1177.1ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 21:25:39,865] eagle_store.py:517 WARNING batch_get_buffer missing keys (k116839-98ce8c42_hs, k116839-98ce8c42_ids, k116839-98ce8c42_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:25:47,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 21:25:39,863] eagle_store.py:517 WARNING batch_get_buffer missing keys (k116838-97183fde_hs, k116838-97183fde_ids, k116838-97183fde_lhs); sleeping 0.50s. [2026-10-18 21:25:49,824] loop.py:45 INFO TRAIN_STEP step=3652 loss=1.784110 Training: 20%|██ | 3652/18018 [7:55:35<35:20:54, 8.86s/step, loss=1.784, acc=0.624, acc_len=1.45, thru=17.2, I=15.8, T=27.3, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:25:49,936] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:25:57,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:25:57,584] loop.py:45 INFO TRAIN_STEP step=3653 loss=1.593832 Training: 20%|██ | 3653/18018 [7:55:43<34:01:54, 8.53s/step, loss=1.594, acc=0.657, acc_len=1.55, thru=14.2, I=16.6, T=31.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:25:57,815] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:04,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:26:04,650] loop.py:45 INFO TRAIN_STEP step=3654 loss=1.739392 Training: 20%|██ | 3654/18018 [7:55:50<32:16:41, 8.09s/step, loss=1.739, acc=0.631, acc_len=1.47, thru=18.5, I=18.9, T=30.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:04,771] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-18 21:26:12,050] loop.py:380 INFO TIMING step=3655: step=3.876s data=1.727s compute=3.829s [fwd=2.721s bwd=1.093s opt=0.003s] dispatch=3.514s [2026-10-18 21:26:12,050] loop.py:45 INFO TRAIN_STEP step=3655 loss=1.720597 Training: 20%|██ | 3655/18018 [7:55:58<31:26:50, 7.88s/step, loss=1.721, acc=0.629, acc_len=1.47, thru=17.0, I=17.2, T=33.0, wait=3.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:18,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:26:21,163] loop.py:45 INFO TRAIN_STEP step=3656 loss=1.622188 Training: 20%|██ | 3656/18018 [7:56:07<32:55:10, 8.25s/step, loss=1.622, acc=0.650, acc_len=1.53, thru=15.7, I=19.3, T=21.7, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:26:21,158] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3655: forward=4097.1ms backward=1754.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:21,436] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:28,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:26:29,848] loop.py:45 INFO TRAIN_STEP step=3657 loss=1.465051 Training: 20%|██ | 3657/18018 [7:56:15<33:26:14, 8.38s/step, loss=1.465, acc=0.676, acc_len=1.62, thru=15.2, I=17.7, T=28.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:30,084] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:35,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:26:35,837] loop.py:45 INFO TRAIN_STEP step=3658 loss=1.436570 Training: 20%|██ | 3658/18018 [7:56:21<30:34:07, 7.66s/step, loss=1.437, acc=0.682, acc_len=1.63, thru=18.5, I=22.2, T=35.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:36,289] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:42,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:26:43,588] loop.py:45 INFO TRAIN_STEP step=3659 loss=1.603310 Training: 20%|██ | 3659/18018 [7:56:29<30:40:27, 7.69s/step, loss=1.603, acc=0.649, acc_len=1.52, thru=16.5, I=19.4, T=29.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:43,622] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 21:26:49,920] loop.py:380 INFO TIMING step=3660: step=3.758s data=1.677s compute=3.717s [fwd=2.618s bwd=1.084s opt=0.003s] dispatch=2.563s [2026-10-18 21:26:49,920] loop.py:45 INFO TRAIN_STEP step=3660 loss=1.672640 Training: 20%|██ | 3660/18018 [7:56:36<29:02:35, 7.28s/step, loss=1.673, acc=0.642, acc_len=1.50, thru=17.7, I=22.1, T=34.1, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:55,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:26:56,622] loop.py:45 INFO TRAIN_STEP step=3661 loss=1.754198 Training: 20%|██ | 3661/18018 [7:56:42<28:21:01, 7.11s/step, loss=1.754, acc=0.628, acc_len=1.46, thru=18.8, I=20.0, T=33.4, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:26:56,618] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3660: forward=2700.0ms backward=1079.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:26:56,957] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:27:04,846] loop.py:45 INFO TRAIN_STEP step=3662 loss=1.656367 Training: 20%|██ | 3662/18018 [7:56:50<29:40:49, 7.44s/step, loss=1.656, acc=0.650, acc_len=1.53, thru=13.1, I=16.3, T=30.0, wait=3.9s, pool=52, epoch=1/2][2026-10-18 21:27:13,010] loop.py:45 INFO TRAIN_STEP step=3663 loss=1.672140 Training: 20%|██ | 3663/18018 [7:56:59<30:32:33, 7.66s/step, loss=1.672, acc=0.640, acc_len=1.50, thru=14.9, I=16.4, T=29.5, wait=3.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:20,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:27:21,122] loop.py:45 INFO TRAIN_STEP step=3664 loss=1.612633 Training: 20%|██ | 3664/18018 [7:57:07<31:04:54, 7.80s/step, loss=1.613, acc=0.647, acc_len=1.52, thru=15.7, I=16.2, T=27.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:21,560] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:27:28,231] loop.py:380 INFO TIMING step=3665: step=4.250s data=1.931s compute=4.220s [fwd=2.927s bwd=1.278s opt=0.003s] dispatch=2.850s [2026-10-18 21:27:28,231] loop.py:45 INFO TRAIN_STEP step=3665 loss=1.595892 Training: 20%|██ | 3665/18018 [7:57:14<30:15:28, 7.59s/step, loss=1.596, acc=0.652, acc_len=1.53, thru=15.0, I=16.8, T=30.1, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:35,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:27:37,033] loop.py:45 INFO TRAIN_STEP step=3666 loss=1.652911 Training: 20%|██ | 3666/18018 [7:57:23<31:42:21, 7.95s/step, loss=1.653, acc=0.643, acc_len=1.51, thru=16.3, I=15.4, T=30.6, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:27:37,028] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3665: forward=2907.1ms backward=1211.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:37,226] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:44,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:27:44,524] loop.py:45 INFO TRAIN_STEP step=3667 loss=1.718458 Training: 20%|██ | 3667/18018 [7:57:30<31:09:03, 7.81s/step, loss=1.718, acc=0.632, acc_len=1.47, thru=15.4, I=17.3, T=32.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:44,554] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:51,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:27:51,449] loop.py:45 INFO TRAIN_STEP step=3668 loss=1.629085 Training: 20%|██ | 3668/18018 [7:57:37<30:05:16, 7.55s/step, loss=1.629, acc=0.649, acc_len=1.53, thru=18.2, I=18.8, T=32.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:51,629] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:58,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:27:59,687] loop.py:45 INFO TRAIN_STEP step=3669 loss=1.548552 Training: 20%|██ | 3669/18018 [7:57:45<30:54:30, 7.75s/step, loss=1.549, acc=0.661, acc_len=1.56, thru=15.0, I=17.3, T=34.6, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:27:59,768] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:05,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:28:06,208] loop.py:380 INFO TIMING step=3670: step=4.211s data=1.939s compute=4.179s [fwd=2.908s bwd=1.256s opt=0.003s] dispatch=2.301s [2026-10-18 21:28:06,208] loop.py:45 INFO TRAIN_STEP step=3670 loss=1.674655 Training: 20%|██ | 3670/18018 [7:57:52<29:26:00, 7.39s/step, loss=1.675, acc=0.639, acc_len=1.49, thru=19.6, I=23.2, T=30.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:06,645] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:12,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:28:13,287] loop.py:45 INFO TRAIN_STEP step=3671 loss=1.604096 Training: 20%|██ | 3671/18018 [7:57:59<29:03:52, 7.29s/step, loss=1.604, acc=0.650, acc_len=1.53, thru=17.7, I=20.3, T=31.2, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:28:13,281] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3670: forward=2831.6ms backward=1226.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:13,767] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:20,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:28:21,755] loop.py:45 INFO TRAIN_STEP step=3672 loss=1.617241 Training: 20%|██ | 3672/18018 [7:58:07<30:27:55, 7.65s/step, loss=1.617, acc=0.653, acc_len=1.54, thru=14.0, I=17.5, T=30.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:22,135] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:28:27,536] loop.py:45 INFO TRAIN_STEP step=3673 loss=1.612814 Training: 20%|██ | 3673/18018 [7:58:13<28:14:02, 7.09s/step, loss=1.613, acc=0.648, acc_len=1.52, thru=19.4, I=18.5, T=33.3, wait=1.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:34,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:28:37,135] loop.py:45 INFO TRAIN_STEP step=3674 loss=1.860265 Training: 20%|██ | 3674/18018 [7:58:23<31:14:17, 7.84s/step, loss=1.860, acc=0.612, acc_len=1.41, thru=14.8, I=17.6, T=30.0, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:37,394] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:42,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:28:43,070] loop.py:380 INFO TIMING step=3675: step=3.857s data=1.757s compute=3.828s [fwd=2.714s bwd=1.099s opt=0.003s] dispatch=2.068s [2026-10-18 21:28:43,070] loop.py:45 INFO TRAIN_STEP step=3675 loss=1.544445 Training: 20%|██ | 3675/18018 [7:58:29<28:57:39, 7.27s/step, loss=1.544, acc=0.663, acc_len=1.57, thru=18.7, I=25.9, T=33.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:43,184] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:50,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:28:50,999] loop.py:45 INFO TRAIN_STEP step=3676 loss=1.728778 Training: 20%|██ | 3676/18018 [7:58:37<29:44:48, 7.47s/step, loss=1.729, acc=0.634, acc_len=1.48, thru=16.1, I=18.3, T=27.9, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:28:50,993] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3675: forward=3195.7ms backward=1330.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:51,159] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:57,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:28:58,296] loop.py:45 INFO TRAIN_STEP step=3677 loss=1.797685 Training: 20%|██ | 3677/18018 [7:58:44<29:32:25, 7.42s/step, loss=1.798, acc=0.625, acc_len=1.45, thru=16.2, I=20.2, T=31.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:28:58,604] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:29:05,738] loop.py:45 INFO TRAIN_STEP step=3678 loss=1.826939 Training: 20%|██ | 3678/18018 [7:58:51<29:34:15, 7.42s/step, loss=1.827, acc=0.621, acc_len=1.43, thru=14.0, I=19.0, T=30.7, wait=3.3s, pool=44, epoch=1/2][2026-10-18 21:29:15,123] loop.py:45 INFO TRAIN_STEP step=3679 loss=1.643385 Training: 20%|██ | 3679/18018 [7:59:01<31:54:40, 8.01s/step, loss=1.643, acc=0.643, acc_len=1.51, thru=15.0, I=15.5, T=29.6, wait=5.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:21,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:29:22,892] loop.py:380 INFO TIMING step=3680: step=4.296s data=1.986s compute=4.263s [fwd=3.062s bwd=1.185s opt=0.003s] dispatch=3.465s [2026-10-18 21:29:22,892] loop.py:45 INFO TRAIN_STEP step=3680 loss=1.586653 Training: 20%|██ | 3680/18018 [7:59:08<31:37:08, 7.94s/step, loss=1.587, acc=0.657, acc_len=1.55, thru=16.8, I=17.6, T=29.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:23,025] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:29,665] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:29:30,128] loop.py:45 INFO TRAIN_STEP step=3681 loss=1.739506 Training: 20%|██ | 3681/18018 [7:59:16<30:46:47, 7.73s/step, loss=1.740, acc=0.631, acc_len=1.47, thru=15.7, I=20.7, T=30.8, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:29:30,123] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3680: forward=2937.3ms backward=1167.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:30,197] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:34,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:29:36,255] loop.py:45 INFO TRAIN_STEP step=3682 loss=1.627167 Training: 20%|██ | 3682/18018 [7:59:22<28:51:47, 7.25s/step, loss=1.627, acc=0.647, acc_len=1.52, thru=20.7, I=22.4, T=30.9, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:36,469] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:42,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:29:43,024] loop.py:45 INFO TRAIN_STEP step=3683 loss=1.793989 Training: 20%|██ | 3683/18018 [7:59:29<28:17:15, 7.10s/step, loss=1.794, acc=0.624, acc_len=1.44, thru=18.8, I=21.3, T=32.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:43,201] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:49,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:29:50,160] loop.py:45 INFO TRAIN_STEP step=3684 loss=1.695705 Training: 20%|██ | 3684/18018 [7:59:36<28:19:30, 7.11s/step, loss=1.696, acc=0.640, acc_len=1.49, thru=16.9, I=19.5, T=27.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:50,597] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:56,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:29:58,349] loop.py:380 INFO TIMING step=3685: step=4.474s data=1.898s compute=4.434s [fwd=2.976s bwd=1.443s opt=0.003s] dispatch=3.704s [2026-10-18 21:29:58,349] loop.py:45 INFO TRAIN_STEP step=3685 loss=1.611819 Training: 20%|██ | 3685/18018 [7:59:44<29:36:21, 7.44s/step, loss=1.612, acc=0.654, acc_len=1.55, thru=16.2, I=19.1, T=28.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:29:58,609] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:03,325] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:30:04,737] loop.py:45 INFO TRAIN_STEP step=3686 loss=1.830084 Training: 20%|██ | 3686/18018 [7:59:50<28:21:09, 7.12s/step, loss=1.830, acc=0.620, acc_len=1.43, thru=18.6, I=23.6, T=31.5, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:30:04,732] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3685: forward=2849.8ms backward=1166.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:04,837] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:14,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:30:14,972] loop.py:45 INFO TRAIN_STEP step=3687 loss=1.674982 Training: 20%|██ | 3687/18018 [8:00:01<32:04:09, 8.06s/step, loss=1.675, acc=0.641, acc_len=1.51, thru=12.2, I=12.9, T=28.4, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:15,121] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:20,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:30:22,435] loop.py:45 INFO TRAIN_STEP step=3688 loss=1.723503 Training: 20%|██ | 3688/18018 [8:00:08<31:21:36, 7.88s/step, loss=1.724, acc=0.636, acc_len=1.49, thru=16.7, I=20.2, T=29.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:22,474] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:27,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:30:28,711] loop.py:45 INFO TRAIN_STEP step=3689 loss=1.764328 Training: 20%|██ | 3689/18018 [8:00:14<29:26:34, 7.40s/step, loss=1.764, acc=0.626, acc_len=1.45, thru=18.6, I=26.0, T=35.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:29,132] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:30:37,265] loop.py:380 INFO TIMING step=3690: step=5.105s data=2.488s compute=5.066s [fwd=3.519s bwd=1.531s opt=0.003s] dispatch=3.438s [2026-10-18 21:30:37,265] loop.py:45 INFO TRAIN_STEP step=3690 loss=1.660694 Training: 20%|██ | 3690/18018 [8:00:23<30:49:34, 7.75s/step, loss=1.661, acc=0.642, acc_len=1.50, thru=11.6, I=17.9, T=25.1, wait=3.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:44,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:30:46,575] loop.py:45 INFO TRAIN_STEP step=3691 loss=1.688830 Training: 20%|██ | 3691/18018 [8:00:32<32:41:14, 8.21s/step, loss=1.689, acc=0.640, acc_len=1.50, thru=15.8, I=16.3, T=27.4, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:30:46,570] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3690: forward=3256.0ms backward=1364.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:30:46,833] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:30:53,430] loop.py:45 INFO TRAIN_STEP step=3692 loss=1.664237 Training: 20%|██ | 3692/18018 [8:00:39<31:03:52, 7.81s/step, loss=1.664, acc=0.644, acc_len=1.51, thru=13.2, I=19.5, T=33.5, wait=3.0s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:01,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:31:04,314] loop.py:45 INFO TRAIN_STEP step=3693 loss=1.575239 Training: 20%|██ | 3693/18018 [8:00:50<34:44:15, 8.73s/step, loss=1.575, acc=0.658, acc_len=1.56, thru=14.6, I=17.0, T=27.7, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:04,430] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:10,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:31:10,753] loop.py:45 INFO TRAIN_STEP step=3694 loss=1.513338 Training: 21%|██ | 3694/18018 [8:00:56<31:59:58, 8.04s/step, loss=1.513, acc=0.668, acc_len=1.59, thru=15.7, I=23.7, T=33.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:11,239] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:16,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:31:17,616] loop.py:380 INFO TIMING step=3695: step=4.158s data=1.917s compute=4.127s [fwd=2.923s bwd=1.188s opt=0.003s] dispatch=2.695s [2026-10-18 21:31:17,616] loop.py:45 INFO TRAIN_STEP step=3695 loss=1.580214 Training: 21%|██ | 3695/18018 [8:01:03<30:35:15, 7.69s/step, loss=1.580, acc=0.656, acc_len=1.55, thru=18.6, I=21.2, T=30.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:18,013] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:25,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:31:26,176] loop.py:45 INFO TRAIN_STEP step=3696 loss=1.566625 Training: 21%|██ | 3696/18018 [8:01:12<31:37:43, 7.95s/step, loss=1.567, acc=0.656, acc_len=1.55, thru=14.3, I=18.0, T=28.3, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:31:26,171] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3695: forward=3080.0ms backward=1389.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:26,646] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:31,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:31:33,464] loop.py:45 INFO TRAIN_STEP step=3697 loss=1.567198 Training: 21%|██ | 3697/18018 [8:01:19<30:50:00, 7.75s/step, loss=1.567, acc=0.661, acc_len=1.56, thru=16.8, I=24.0, T=29.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:33,640] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-18 21:31:40,587] loop.py:45 INFO TRAIN_STEP step=3698 loss=1.602332 Training: 21%|██ | 3698/18018 [8:01:26<30:05:02, 7.56s/step, loss=1.602, acc=0.651, acc_len=1.54, thru=12.5, I=20.6, T=29.1, wait=2.7s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:50,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:31:50,854] loop.py:45 INFO TRAIN_STEP step=3699 loss=1.751122 Training: 21%|██ | 3699/18018 [8:01:36<33:18:27, 8.37s/step, loss=1.751, acc=0.635, acc_len=1.49, thru=16.8, I=14.4, T=29.1, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:50,918] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:57,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:31:57,321] loop.py:380 INFO TIMING step=3700: step=3.977s data=1.819s compute=3.951s [fwd=2.838s bwd=1.098s opt=0.003s] dispatch=2.480s [2026-10-18 21:31:57,321] loop.py:45 INFO TRAIN_STEP step=3700 loss=1.831383 Training: 21%|██ | 3700/18018 [8:01:43<31:01:43, 7.80s/step, loss=1.831, acc=0.615, acc_len=1.42, thru=18.8, I=18.8, T=32.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:31:57,606] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:04,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:32:06,237] loop.py:45 INFO TRAIN_STEP step=3701 loss=1.755500 Training: 21%|██ | 3701/18018 [8:01:52<32:21:25, 8.14s/step, loss=1.756, acc=0.628, acc_len=1.47, thru=14.8, I=18.0, T=28.0, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:32:06,232] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3700: forward=3193.6ms backward=1333.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:06,382] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:12,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:32:13,911] loop.py:45 INFO TRAIN_STEP step=3702 loss=1.675007 Training: 21%|██ | 3702/18018 [8:01:59<31:48:10, 8.00s/step, loss=1.675, acc=0.645, acc_len=1.52, thru=14.4, I=21.3, T=27.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:14,201] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:19,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:32:20,989] loop.py:45 INFO TRAIN_STEP step=3703 loss=1.658773 Training: 21%|██ | 3703/18018 [8:02:07<30:42:20, 7.72s/step, loss=1.659, acc=0.644, acc_len=1.51, thru=16.7, I=19.0, T=33.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:21,318] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:25,425] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:32:28,368] loop.py:45 INFO TRAIN_STEP step=3704 loss=1.757197 Training: 21%|██ | 3704/18018 [8:02:14<30:17:43, 7.62s/step, loss=1.757, acc=0.629, acc_len=1.46, thru=16.5, I=25.7, T=27.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:28,454] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:35,494] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:32:35,783] loop.py:380 INFO TIMING step=3705: step=3.950s data=1.785s compute=3.912s [fwd=2.702s bwd=1.195s opt=0.003s] dispatch=3.455s [2026-10-18 21:32:35,783] loop.py:45 INFO TRAIN_STEP step=3705 loss=1.597066 Training: 21%|██ | 3705/18018 [8:02:21<30:02:58, 7.56s/step, loss=1.597, acc=0.652, acc_len=1.54, thru=17.5, I=15.8, T=32.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:35,997] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 21:32:43,366] loop.py:45 INFO TRAIN_STEP step=3706 loss=1.587649 Training: 21%|██ | 3706/18018 [8:02:29<30:04:29, 7.56s/step, loss=1.588, acc=0.653, acc_len=1.54, thru=15.8, I=17.2, T=28.9, wait=3.1s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:32:43,361] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3705: forward=3093.1ms backward=1294.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:50,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:32:51,810] loop.py:45 INFO TRAIN_STEP step=3707 loss=1.671459 Training: 21%|██ | 3707/18018 [8:02:37<31:07:22, 7.83s/step, loss=1.671, acc=0.641, acc_len=1.50, thru=15.8, I=16.2, T=27.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:52,017] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:56,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:32:57,714] loop.py:45 INFO TRAIN_STEP step=3708 loss=1.690687 Training: 21%|██ | 3708/18018 [8:02:43<28:49:30, 7.25s/step, loss=1.691, acc=0.633, acc_len=1.48, thru=18.6, I=24.6, T=34.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:32:57,821] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:02,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:33:03,767] loop.py:45 INFO TRAIN_STEP step=3709 loss=1.434174 Training: 21%|██ | 3709/18018 [8:02:49<27:23:30, 6.89s/step, loss=1.434, acc=0.681, acc_len=1.64, thru=20.4, I=29.4, T=30.2, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:04,229] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:10,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:33:10,782] loop.py:380 INFO TIMING step=3710: step=4.223s data=1.935s compute=4.194s [fwd=2.956s bwd=1.223s opt=0.003s] dispatch=2.783s [2026-10-18 21:33:10,782] loop.py:45 INFO TRAIN_STEP step=3710 loss=1.604423 Training: 21%|██ | 3710/18018 [8:02:56<27:32:15, 6.93s/step, loss=1.604, acc=0.650, acc_len=1.53, thru=16.7, I=21.3, T=30.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:11,246] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:16,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:33:18,301] loop.py:45 INFO TRAIN_STEP step=3711 loss=1.703622 Training: 21%|██ | 3711/18018 [8:03:04<28:14:29, 7.11s/step, loss=1.704, acc=0.639, acc_len=1.49, thru=17.3, I=21.6, T=30.8, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:33:18,296] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3710: forward=2909.4ms backward=1206.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:18,424] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:23,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:33:25,482] loop.py:45 INFO TRAIN_STEP step=3712 loss=1.570059 Training: 21%|██ | 3712/18018 [8:03:11<28:19:42, 7.13s/step, loss=1.570, acc=0.657, acc_len=1.55, thru=17.0, I=24.1, T=27.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:25,893] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:30,192] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:33:32,204] loop.py:45 INFO TRAIN_STEP step=3713 loss=1.507968 Training: 21%|██ | 3713/18018 [8:03:18<27:50:30, 7.01s/step, loss=1.508, acc=0.665, acc_len=1.58, thru=16.8, I=26.9, T=29.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:32,711] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:37,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:33:38,873] loop.py:45 INFO TRAIN_STEP step=3714 loss=1.517248 Training: 21%|██ | 3714/18018 [8:03:24<27:26:04, 6.90s/step, loss=1.517, acc=0.665, acc_len=1.58, thru=16.6, I=26.0, T=33.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:39,141] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:47,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:33:48,835] loop.py:380 INFO TIMING step=3715: step=4.366s data=1.969s compute=4.335s [fwd=3.122s bwd=1.197s opt=0.003s] dispatch=5.588s [2026-10-18 21:33:48,835] loop.py:45 INFO TRAIN_STEP step=3715 loss=1.804717 Training: 21%|██ | 3715/18018 [8:03:34<31:04:32, 7.82s/step, loss=1.805, acc=0.624, acc_len=1.44, thru=13.2, I=14.1, T=29.3, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:49,291] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:54,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:33:55,788] loop.py:45 INFO TRAIN_STEP step=3716 loss=1.476666 Training: 21%|██ | 3716/18018 [8:03:41<30:02:30, 7.56s/step, loss=1.477, acc=0.675, acc_len=1.62, thru=18.1, I=23.4, T=33.0, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:33:55,782] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3715: forward=2715.8ms backward=1101.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:33:55,875] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:00,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:34:02,851] loop.py:45 INFO TRAIN_STEP step=3717 loss=1.809690 Training: 21%|██ | 3717/18018 [8:03:48<29:26:31, 7.41s/step, loss=1.810, acc=0.620, acc_len=1.43, thru=17.3, I=28.2, T=28.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:03,118] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:09,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:34:10,662] loop.py:45 INFO TRAIN_STEP step=3718 loss=1.655120 Training: 21%|██ | 3718/18018 [8:03:56<29:55:04, 7.53s/step, loss=1.655, acc=0.642, acc_len=1.50, thru=17.0, I=18.8, T=30.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:10,873] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:16,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:34:19,059] loop.py:45 INFO TRAIN_STEP step=3719 loss=1.568870 Training: 21%|██ | 3719/18018 [8:04:05<30:56:40, 7.79s/step, loss=1.569, acc=0.655, acc_len=1.55, thru=14.6, I=18.2, T=24.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:19,539] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 [2026-10-18 21:34:25,719] loop.py:380 INFO TIMING step=3720: step=4.299s data=1.933s compute=4.258s [fwd=2.927s bwd=1.316s opt=0.003s] dispatch=2.353s [2026-10-18 21:34:25,719] loop.py:45 INFO TRAIN_STEP step=3720 loss=1.610328 Training: 21%|██ | 3720/18018 [8:04:11<29:35:55, 7.45s/step, loss=1.610, acc=0.654, acc_len=1.54, thru=13.2, I=18.1, T=29.8, wait=2.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:30,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:34:32,730] loop.py:45 INFO TRAIN_STEP step=3721 loss=1.530266 Training: 21%|██ | 3721/18018 [8:04:18<29:04:04, 7.32s/step, loss=1.530, acc=0.665, acc_len=1.57, thru=18.2, I=21.8, T=28.6, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:34:32,724] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3720: forward=3136.9ms backward=1290.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:33,183] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 21:34:40,457] loop.py:45 INFO TRAIN_STEP step=3722 loss=1.618541 Training: 21%|██ | 3722/18018 [8:04:26<29:33:04, 7.44s/step, loss=1.619, acc=0.649, acc_len=1.53, thru=16.5, I=15.8, T=27.7, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:46,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:34:47,984] loop.py:45 INFO TRAIN_STEP step=3723 loss=1.970707 Training: 21%|██ | 3723/18018 [8:04:34<29:39:09, 7.47s/step, loss=1.971, acc=0.596, acc_len=1.36, thru=17.8, I=18.4, T=29.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:48,326] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:54,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:34:54,827] loop.py:45 INFO TRAIN_STEP step=3724 loss=1.683759 Training: 21%|██ | 3724/18018 [8:04:40<28:54:25, 7.28s/step, loss=1.684, acc=0.638, acc_len=1.49, thru=18.1, I=22.0, T=33.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:34:55,094] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:04,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:35:05,591] loop.py:380 INFO TIMING step=3725: step=4.387s data=2.036s compute=4.280s [fwd=2.841s bwd=1.425s opt=0.003s] dispatch=6.367s [2026-10-18 21:35:05,591] loop.py:45 INFO TRAIN_STEP step=3725 loss=1.757912 Training: 21%|██ | 3725/18018 [8:04:51<33:03:11, 8.33s/step, loss=1.758, acc=0.623, acc_len=1.44, thru=13.2, I=13.0, T=29.2, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:05,620] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:11,276] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:35:12,478] loop.py:45 INFO TRAIN_STEP step=3726 loss=1.724739 Training: 21%|██ | 3726/18018 [8:04:58<31:20:16, 7.89s/step, loss=1.725, acc=0.634, acc_len=1.48, thru=17.6, I=24.0, T=28.6, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:35:12,473] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3725: forward=3097.9ms backward=1324.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:12,783] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:35:18,918] loop.py:45 INFO TRAIN_STEP step=3727 loss=1.607597 Training: 21%|██ | 3727/18018 [8:05:05<29:36:23, 7.46s/step, loss=1.608, acc=0.651, acc_len=1.53, thru=16.4, I=20.5, T=30.6, wait=2.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:27,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:35:28,790] loop.py:45 INFO TRAIN_STEP step=3728 loss=1.725338 Training: 21%|██ | 3728/18018 [8:05:14<32:28:31, 8.18s/step, loss=1.725, acc=0.632, acc_len=1.47, thru=14.0, I=13.5, T=28.8, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:29,240] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:34,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:35:36,208] loop.py:45 INFO TRAIN_STEP step=3729 loss=1.518191 Training: 21%|██ | 3729/18018 [8:05:22<31:33:57, 7.95s/step, loss=1.518, acc=0.667, acc_len=1.59, thru=14.8, I=20.1, T=28.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:36,593] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:43,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:35:43,768] loop.py:380 INFO TIMING step=3730: step=3.772s data=1.673s compute=3.730s [fwd=2.668s bwd=1.048s opt=0.003s] dispatch=3.776s [2026-10-18 21:35:43,769] loop.py:45 INFO TRAIN_STEP step=3730 loss=1.730128 Training: 21%|██ | 3730/18018 [8:05:29<31:05:52, 7.84s/step, loss=1.730, acc=0.638, acc_len=1.50, thru=14.8, I=20.2, T=33.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:43,828] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:49,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:35:50,246] loop.py:45 INFO TRAIN_STEP step=3731 loss=1.634811 Training: 21%|██ | 3731/18018 [8:05:36<29:28:40, 7.43s/step, loss=1.635, acc=0.645, acc_len=1.51, thru=19.1, I=22.0, T=28.6, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:35:50,241] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3730: forward=3138.7ms backward=1288.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:50,595] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:35:58,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:36:00,699] loop.py:45 INFO TRAIN_STEP step=3732 loss=1.585310 Training: 21%|██ | 3732/18018 [8:05:46<33:04:39, 8.34s/step, loss=1.585, acc=0.656, acc_len=1.55, thru=12.1, I=15.6, T=27.0, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:00,729] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:05,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:36:07,103] loop.py:45 INFO TRAIN_STEP step=3733 loss=1.624911 Training: 21%|██ | 3733/18018 [8:05:53<30:46:33, 7.76s/step, loss=1.625, acc=0.650, acc_len=1.53, thru=17.6, I=26.6, T=28.7, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:07,224] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-18 21:36:15,332] loop.py:45 INFO TRAIN_STEP step=3734 loss=1.677400 Training: 21%|██ | 3734/18018 [8:06:01<31:20:30, 7.90s/step, loss=1.677, acc=0.641, acc_len=1.50, thru=15.3, I=17.4, T=33.0, wait=4.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:26,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:36:28,403] loop.py:380 INFO TIMING step=3735: step=5.107s data=2.440s compute=5.059s [fwd=3.618s bwd=1.429s opt=0.003s] dispatch=7.951s [2026-10-18 21:36:28,404] loop.py:45 INFO TRAIN_STEP step=3735 loss=1.828593 Training: 21%|██ | 3735/18018 [8:06:14<37:29:33, 9.45s/step, loss=1.829, acc=0.617, acc_len=1.43, thru=12.2, I=11.9, T=25.1, wait=8.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:28,704] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:33,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:36:35,380] loop.py:45 INFO TRAIN_STEP step=3736 loss=1.583283 Training: 21%|██ | 3736/18018 [8:06:21<34:32:37, 8.71s/step, loss=1.583, acc=0.655, acc_len=1.55, thru=14.6, I=23.7, T=28.0, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:36:35,375] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3735: forward=3403.1ms backward=1115.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:35,672] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:39,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:36:41,948] loop.py:45 INFO TRAIN_STEP step=3737 loss=1.525548 Training: 21%|██ | 3737/18018 [8:06:28<31:59:49, 8.07s/step, loss=1.526, acc=0.663, acc_len=1.57, thru=16.3, I=26.0, T=32.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:42,345] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:47,498] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:36:48,424] loop.py:45 INFO TRAIN_STEP step=3738 loss=1.743076 Training: 21%|██ | 3738/18018 [8:06:34<30:06:05, 7.59s/step, loss=1.743, acc=0.634, acc_len=1.47, thru=18.0, I=26.3, T=32.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:48,507] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:54,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:36:55,840] loop.py:45 INFO TRAIN_STEP step=3739 loss=1.788720 Training: 21%|██ | 3739/18018 [8:06:41<29:53:36, 7.54s/step, loss=1.789, acc=0.621, acc_len=1.44, thru=15.7, I=18.4, T=30.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:36:56,318] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:37:02,109] loop.py:380 INFO TIMING step=3740: step=3.727s data=1.557s compute=3.699s [fwd=2.591s bwd=1.093s opt=0.003s] dispatch=2.532s [2026-10-18 21:37:02,109] loop.py:45 INFO TRAIN_STEP step=3740 loss=1.677935 Training: 21%|██ | 3740/18018 [8:06:48<28:23:03, 7.16s/step, loss=1.678, acc=0.638, acc_len=1.49, thru=18.6, I=23.9, T=34.3, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:09,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:37:10,169] loop.py:45 INFO TRAIN_STEP step=3741 loss=1.708136 (TrainerActor pid=9310) [2026-10-18 21:37:10,164] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3740: forward=3333.3ms backward=1311.9ms Training: 21%|██ | 3741/18018 [8:06:56<29:27:23, 7.43s/step, loss=1.708, acc=0.636, acc_len=1.48, thru=16.1, I=16.5, T=27.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:10,330] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:17,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:37:20,477] loop.py:45 INFO TRAIN_STEP step=3742 loss=1.853142 Training: 21%|██ | 3742/18018 [8:07:06<32:52:47, 8.29s/step, loss=1.853, acc=0.614, acc_len=1.41, thru=12.4, I=17.9, T=24.4, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:20,976] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:26,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:37:27,967] loop.py:45 INFO TRAIN_STEP step=3743 loss=1.493014 Training: 21%|██ | 3743/18018 [8:07:14<31:55:39, 8.05s/step, loss=1.493, acc=0.670, acc_len=1.59, thru=18.3, I=22.0, T=27.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:28,151] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:38,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:37:38,935] loop.py:45 INFO TRAIN_STEP step=3744 loss=1.695041 Training: 21%|██ | 3744/18018 [8:07:25<35:23:28, 8.93s/step, loss=1.695, acc=0.636, acc_len=1.49, thru=11.2, I=12.0, T=26.1, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:39,224] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 21:37:45,397] loop.py:380 INFO TIMING step=3745: step=4.043s data=1.833s compute=4.001s [fwd=2.850s bwd=1.138s opt=0.003s] dispatch=2.410s [2026-10-18 21:37:45,397] loop.py:45 INFO TRAIN_STEP step=3745 loss=1.718639 Training: 21%|██ | 3745/18018 [8:07:31<32:27:42, 8.19s/step, loss=1.719, acc=0.632, acc_len=1.48, thru=17.3, I=20.0, T=31.7, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:52,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:37:52,513] loop.py:45 INFO TRAIN_STEP step=3746 loss=1.563030 Training: 21%|██ | 3746/18018 [8:07:38<31:10:55, 7.87s/step, loss=1.563, acc=0.662, acc_len=1.58, thru=17.2, I=17.3, T=30.2, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:37:52,508] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3745: forward=2910.4ms backward=1277.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:52,806] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:37:59,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:38:01,570] loop.py:45 INFO TRAIN_STEP step=3747 loss=1.736550 Training: 21%|██ | 3747/18018 [8:07:47<32:35:54, 8.22s/step, loss=1.737, acc=0.631, acc_len=1.47, thru=14.4, I=18.0, T=26.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:01,808] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:06,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:38:08,055] loop.py:45 INFO TRAIN_STEP step=3748 loss=1.520601 Training: 21%|██ | 3748/18018 [8:07:54<30:31:40, 7.70s/step, loss=1.521, acc=0.664, acc_len=1.58, thru=16.9, I=27.7, T=30.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:08,323] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:13,868] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:38:15,172] loop.py:45 INFO TRAIN_STEP step=3749 loss=1.655256 Training: 21%|██ | 3749/18018 [8:08:01<29:49:54, 7.53s/step, loss=1.655, acc=0.647, acc_len=1.52, thru=15.5, I=18.5, T=32.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:15,377] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:22,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:38:22,631] loop.py:380 INFO TIMING step=3750: step=4.152s data=1.999s compute=4.110s [fwd=2.926s bwd=1.171s opt=0.003s] dispatch=3.298s [2026-10-18 21:38:22,632] loop.py:45 INFO TRAIN_STEP step=3750 loss=1.718298 Training: 21%|██ | 3750/18018 [8:08:08<29:44:53, 7.51s/step, loss=1.718, acc=0.632, acc_len=1.47, thru=17.4, I=20.4, T=30.8, wait=3.3s, pool=64, epoch=1/2][2026-10-18 21:38:22,634] loop.py:419 INFO Saving checkpoint at step 3750... (TrainerActor pid=9310) [2026-10-18 21:38:26,392] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0003751 (TrainerActor pid=9310) [2026-10-18 21:38:26,403] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0003001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:27,141] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:27,666] inference_manager.py:408 INFO Pool capacity available after 5.6s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:33,107] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 21:38:26,393] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0003751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 21:38:27,625] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0003001 [2026-10-18 21:38:34,578] loop.py:45 INFO TRAIN_STEP step=3751 loss=1.644103 Training: 21%|██ | 3751/18018 [8:08:20<35:01:42, 8.84s/step, loss=1.644, acc=0.646, acc_len=1.51, thru=18.6, I=22.5, T=29.8, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:38:34,572] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3750: forward=2932.3ms backward=1300.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:34,634] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:41,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:38:43,212] loop.py:45 INFO TRAIN_STEP step=3752 loss=1.737025 Training: 21%|██ | 3752/18018 [8:08:29<34:46:48, 8.78s/step, loss=1.737, acc=0.635, acc_len=1.48, thru=15.0, I=16.8, T=31.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:43,441] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:49,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:38:50,166] loop.py:45 INFO TRAIN_STEP step=3753 loss=1.588915 Training: 21%|██ | 3753/18018 [8:08:36<32:36:47, 8.23s/step, loss=1.589, acc=0.651, acc_len=1.54, thru=17.1, I=20.3, T=28.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:50,237] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:57,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:38:58,487] loop.py:45 INFO TRAIN_STEP step=3754 loss=1.638794 Training: 21%|██ | 3754/18018 [8:08:44<32:43:07, 8.26s/step, loss=1.639, acc=0.652, acc_len=1.53, thru=14.9, I=15.5, T=28.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:38:58,747] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:03,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:39:05,092] loop.py:380 INFO TIMING step=3755: step=3.797s data=1.687s compute=3.752s [fwd=2.652s bwd=1.085s opt=0.003s] dispatch=2.797s [2026-10-18 21:39:05,092] loop.py:45 INFO TRAIN_STEP step=3755 loss=1.741143 Training: 21%|██ | 3755/18018 [8:08:51<30:44:57, 7.76s/step, loss=1.741, acc=0.628, acc_len=1.46, thru=18.3, I=20.8, T=33.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:05,449] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:12,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:39:13,419] loop.py:45 INFO TRAIN_STEP step=3756 loss=1.714123 Training: 21%|██ | 3756/18018 [8:08:59<31:25:16, 7.93s/step, loss=1.714, acc=0.637, acc_len=1.49, thru=14.7, I=18.0, T=30.9, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:39:13,413] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3755: forward=2853.7ms backward=1246.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:13,469] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:18,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:39:19,863] loop.py:45 INFO TRAIN_STEP step=3757 loss=1.693610 Training: 21%|██ | 3757/18018 [8:09:05<29:39:08, 7.49s/step, loss=1.694, acc=0.640, acc_len=1.49, thru=18.6, I=21.2, T=31.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:20,054] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:26,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:39:27,361] loop.py:45 INFO TRAIN_STEP step=3758 loss=1.682190 Training: 21%|██ | 3758/18018 [8:09:13<29:39:55, 7.49s/step, loss=1.682, acc=0.639, acc_len=1.50, thru=15.4, I=20.2, T=30.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:27,470] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:33,132] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:39:35,790] loop.py:45 INFO TRAIN_STEP step=3759 loss=1.743838 Training: 21%|██ | 3759/18018 [8:09:21<30:46:38, 7.77s/step, loss=1.744, acc=0.631, acc_len=1.46, thru=14.1, I=18.1, T=27.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:39:36,151] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 21:39:42,769] loop.py:380 INFO TIMING step=3760: step=4.297s data=1.942s compute=4.270s [fwd=3.054s bwd=1.202s opt=0.003s] dispatch=2.673s [2026-10-18 21:39:42,769] loop.py:45 INFO TRAIN_STEP step=3760 loss=1.648767 Training: 21%|██ | 3760/18018 [8:09:28<29:50:03, 7.53s/step, loss=1.649, acc=0.644, acc_len=1.51, thru=12.8, I=19.3, T=29.8, wait=2.7s, pool=52, epoch=1/2][2026-10-18 21:39:50,332] loop.py:45 INFO TRAIN_STEP step=3761 loss=1.478621 Training: 21%|██ | 3761/18018 [8:09:36<29:52:02, 7.54s/step, loss=1.479, acc=0.668, acc_len=1.58, thru=14.9, I=15.3, T=31.8, wait=3.5s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:39:50,326] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3760: forward=2817.4ms backward=1160.2ms [2026-10-18 21:39:58,765] loop.py:45 INFO TRAIN_STEP step=3762 loss=1.808035 Training: 21%|██ | 3762/18018 [8:09:44<30:55:39, 7.81s/step, loss=1.808, acc=0.624, acc_len=1.44, thru=17.8, I=14.9, T=33.1, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:05,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:40:07,115] loop.py:45 INFO TRAIN_STEP step=3763 loss=1.664380 Training: 21%|██ | 3763/18018 [8:09:53<31:33:57, 7.97s/step, loss=1.664, acc=0.642, acc_len=1.50, thru=15.4, I=15.8, T=32.5, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:07,506] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:14,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:40:15,256] loop.py:45 INFO TRAIN_STEP step=3764 loss=1.661617 Training: 21%|██ | 3764/18018 [8:10:01<31:45:59, 8.02s/step, loss=1.662, acc=0.642, acc_len=1.50, thru=15.5, I=18.5, T=29.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:15,375] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:22,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:40:23,229] loop.py:380 INFO TIMING step=3765: step=4.337s data=2.034s compute=4.303s [fwd=3.036s bwd=1.252s opt=0.003s] dispatch=3.625s [2026-10-18 21:40:23,229] loop.py:45 INFO TRAIN_STEP step=3765 loss=1.563040 Training: 21%|██ | 3765/18018 [8:10:09<31:42:18, 8.01s/step, loss=1.563, acc=0.657, acc_len=1.56, thru=15.4, I=16.7, T=29.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:23,459] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:29,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 21:40:31,486] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3765: forward=3119.7ms backward=1314.6ms [2026-10-18 21:40:31,491] loop.py:45 INFO TRAIN_STEP step=3766 loss=1.715236 Training: 21%|██ | 3766/18018 [8:10:17<32:00:16, 8.08s/step, loss=1.715, acc=0.636, acc_len=1.48, thru=15.4, I=18.4, T=28.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:31,584] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:37,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:40:38,876] loop.py:45 INFO TRAIN_STEP step=3767 loss=1.673356 Training: 21%|██ | 3767/18018 [8:10:24<31:10:15, 7.87s/step, loss=1.673, acc=0.643, acc_len=1.51, thru=15.1, I=21.5, T=32.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:39,188] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:40:47,619] loop.py:45 INFO TRAIN_STEP step=3768 loss=1.774276 Training: 21%|██ | 3768/18018 [8:10:33<32:11:49, 8.13s/step, loss=1.774, acc=0.633, acc_len=1.47, thru=13.3, I=17.0, T=31.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:47,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:40:48,118] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 21:40:54,854] loop.py:45 INFO TRAIN_STEP step=3769 loss=1.627738 Training: 21%|██ | 3769/18018 [8:10:40<31:07:46, 7.86s/step, loss=1.628, acc=0.647, acc_len=1.53, thru=17.4, I=18.9, T=29.8, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:00,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:41:03,318] loop.py:380 INFO TIMING step=3770: step=5.035s data=2.445s compute=5.008s [fwd=3.493s bwd=1.500s opt=0.003s] dispatch=3.419s [2026-10-18 21:41:03,318] loop.py:45 INFO TRAIN_STEP step=3770 loss=1.529098 Training: 21%|██ | 3770/18018 [8:10:49<31:50:21, 8.04s/step, loss=1.529, acc=0.662, acc_len=1.57, thru=15.7, I=19.9, T=25.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:03,776] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:10,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:41:10,386] loop.py:45 INFO TRAIN_STEP step=3771 loss=1.695769 Training: 21%|██ | 3771/18018 [8:10:56<30:40:29, 7.75s/step, loss=1.696, acc=0.634, acc_len=1.48, thru=14.2, I=20.8, T=32.4, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:41:10,382] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3770: forward=2810.8ms backward=1097.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:10,707] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:17,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:41:18,476] loop.py:45 INFO TRAIN_STEP step=3772 loss=1.777020 Training: 21%|██ | 3772/18018 [8:11:04<31:04:43, 7.85s/step, loss=1.777, acc=0.623, acc_len=1.44, thru=15.8, I=17.3, T=31.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:18,782] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:25,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:41:26,780] loop.py:45 INFO TRAIN_STEP step=3773 loss=1.675235 Training: 21%|██ | 3773/18018 [8:11:12<31:36:33, 7.99s/step, loss=1.675, acc=0.642, acc_len=1.51, thru=15.0, I=19.2, T=29.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:26,811] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:33,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:41:34,596] loop.py:45 INFO TRAIN_STEP step=3774 loss=1.865334 Training: 21%|██ | 3774/18018 [8:11:20<31:24:00, 7.94s/step, loss=1.865, acc=0.611, acc_len=1.41, thru=14.9, I=17.5, T=29.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:34,723] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:40,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:41:41,935] loop.py:380 INFO TIMING step=3775: step=4.401s data=2.164s compute=4.358s [fwd=3.055s bwd=1.287s opt=0.003s] dispatch=2.930s [2026-10-18 21:41:41,935] loop.py:45 INFO TRAIN_STEP step=3775 loss=1.705722 Training: 21%|██ | 3775/18018 [8:11:28<30:41:36, 7.76s/step, loss=1.706, acc=0.637, acc_len=1.49, thru=15.4, I=21.8, T=29.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:42,063] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:46,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:41:48,593] loop.py:45 INFO TRAIN_STEP step=3776 loss=1.621535 Training: 21%|██ | 3776/18018 [8:11:34<29:22:59, 7.43s/step, loss=1.622, acc=0.646, acc_len=1.52, thru=16.9, I=26.5, T=29.8, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:41:48,588] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3775: forward=2961.3ms backward=1275.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:48,973] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:55,888] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:41:56,267] loop.py:45 INFO TRAIN_STEP step=3777 loss=1.541445 Training: 21%|██ | 3777/18018 [8:11:42<29:40:30, 7.50s/step, loss=1.541, acc=0.658, acc_len=1.56, thru=14.0, I=17.8, T=31.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:41:56,402] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:01,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:42:04,049] loop.py:45 INFO TRAIN_STEP step=3778 loss=1.711830 Training: 21%|██ | 3778/18018 [8:11:50<30:00:14, 7.59s/step, loss=1.712, acc=0.639, acc_len=1.50, thru=18.1, I=19.4, T=30.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:04,495] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:08,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:42:10,089] loop.py:45 INFO TRAIN_STEP step=3779 loss=1.594764 Training: 21%|██ | 3779/18018 [8:11:56<28:10:07, 7.12s/step, loss=1.595, acc=0.651, acc_len=1.54, thru=18.4, I=28.4, T=32.8, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:10,555] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:16,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:42:17,834] loop.py:380 INFO TIMING step=3780: step=4.692s data=2.156s compute=4.664s [fwd=3.262s bwd=1.387s opt=0.003s] dispatch=3.045s [2026-10-18 21:42:17,834] loop.py:45 INFO TRAIN_STEP step=3780 loss=1.648983 Training: 21%|██ | 3780/18018 [8:12:03<28:54:32, 7.31s/step, loss=1.649, acc=0.643, acc_len=1.51, thru=14.4, I=18.6, T=27.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:17,888] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:24,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:42:25,094] loop.py:45 INFO TRAIN_STEP step=3781 loss=1.728848 Training: 21%|██ | 3781/18018 [8:12:11<28:50:54, 7.29s/step, loss=1.729, acc=0.631, acc_len=1.47, thru=17.0, I=18.7, T=31.8, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:42:25,090] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3780: forward=2800.2ms backward=1175.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:25,269] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:30,198] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:42:31,023] loop.py:45 INFO TRAIN_STEP step=3782 loss=1.631574 Training: 21%|██ | 3782/18018 [8:12:17<27:13:33, 6.88s/step, loss=1.632, acc=0.647, acc_len=1.52, thru=19.6, I=23.1, T=34.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:31,204] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:37,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:42:39,217] loop.py:45 INFO TRAIN_STEP step=3783 loss=1.666702 Training: 21%|██ | 3783/18018 [8:12:25<28:46:30, 7.28s/step, loss=1.667, acc=0.642, acc_len=1.50, thru=16.1, I=21.5, T=33.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:39,505] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:47,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:42:47,240] loop.py:45 INFO TRAIN_STEP step=3784 loss=1.773783 Training: 21%|██ | 3784/18018 [8:12:33<29:39:29, 7.50s/step, loss=1.774, acc=0.625, acc_len=1.45, thru=13.7, I=19.1, T=27.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:47,505] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:54,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:42:55,213] loop.py:380 INFO TIMING step=3785: step=4.458s data=2.087s compute=4.425s [fwd=3.097s bwd=1.313s opt=0.003s] dispatch=3.506s [2026-10-18 21:42:55,213] loop.py:45 INFO TRAIN_STEP step=3785 loss=1.673182 Training: 21%|██ | 3785/18018 [8:12:41<30:12:51, 7.64s/step, loss=1.673, acc=0.641, acc_len=1.50, thru=15.2, I=16.9, T=28.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:42:55,660] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:01,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:43:02,296] loop.py:45 INFO TRAIN_STEP step=3786 loss=1.590867 Training: 21%|██ | 3786/18018 [8:12:48<29:32:58, 7.47s/step, loss=1.591, acc=0.654, acc_len=1.55, thru=16.5, I=20.3, T=27.7, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:43:02,291] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3785: forward=3213.2ms backward=1352.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:02,631] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:11,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:43:11,684] loop.py:45 INFO TRAIN_STEP step=3787 loss=1.603106 Training: 21%|██ | 3787/18018 [8:12:57<31:49:08, 8.05s/step, loss=1.603, acc=0.651, acc_len=1.53, thru=14.1, I=13.7, T=31.3, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:12,042] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 21:43:18,080] loop.py:45 INFO TRAIN_STEP step=3788 loss=1.600429 Training: 21%|██ | 3788/18018 [8:13:04<29:51:24, 7.55s/step, loss=1.600, acc=0.652, acc_len=1.54, thru=17.6, I=18.2, T=31.3, wait=2.3s, pool=48, epoch=1/2][2026-10-18 21:43:25,827] loop.py:45 INFO TRAIN_STEP step=3789 loss=1.712365 Training: 21%|██ | 3789/18018 [8:13:11<30:04:51, 7.61s/step, loss=1.712, acc=0.637, acc_len=1.50, thru=15.0, I=13.5, T=28.4, wait=3.2s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:34,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:43:37,229] loop.py:380 INFO TIMING step=3790: step=4.602s data=2.159s compute=4.539s [fwd=3.108s bwd=1.415s opt=0.003s] dispatch=6.790s [2026-10-18 21:43:37,229] loop.py:45 INFO TRAIN_STEP step=3790 loss=1.664687 Training: 21%|██ | 3790/18018 [8:13:23<34:34:34, 8.75s/step, loss=1.665, acc=0.645, acc_len=1.51, thru=13.7, I=16.4, T=27.8, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:37,493] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:44,972] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:43:46,084] loop.py:45 INFO TRAIN_STEP step=3791 loss=1.672055 Training: 21%|██ | 3791/18018 [8:13:32<34:42:05, 8.78s/step, loss=1.672, acc=0.636, acc_len=1.47, thru=14.9, I=16.6, T=28.4, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:43:46,080] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3790: forward=3138.3ms backward=1319.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:46,481] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:51,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:43:52,506] loop.py:45 INFO TRAIN_STEP step=3792 loss=1.598229 Training: 21%|██ | 3792/18018 [8:13:38<31:53:54, 8.07s/step, loss=1.598, acc=0.654, acc_len=1.54, thru=17.5, I=19.2, T=31.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:43:52,963] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:00,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:44:00,373] loop.py:45 INFO TRAIN_STEP step=3793 loss=1.697862 Training: 21%|██ | 3793/18018 [8:13:46<31:39:16, 8.01s/step, loss=1.698, acc=0.637, acc_len=1.49, thru=15.4, I=16.2, T=32.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:00,647] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 21:44:07,687] loop.py:45 INFO TRAIN_STEP step=3794 loss=1.592494 Training: 21%|██ | 3794/18018 [8:13:53<30:49:33, 7.80s/step, loss=1.592, acc=0.652, acc_len=1.53, thru=17.6, I=19.2, T=33.2, wait=3.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:14,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:44:16,105] loop.py:380 INFO TIMING step=3795: step=4.513s data=2.185s compute=4.469s [fwd=3.066s bwd=1.388s opt=0.003s] dispatch=3.897s [2026-10-18 21:44:16,105] loop.py:45 INFO TRAIN_STEP step=3795 loss=1.510233 Training: 21%|██ | 3795/18018 [8:14:02<31:33:22, 7.99s/step, loss=1.510, acc=0.671, acc_len=1.60, thru=15.7, I=18.1, T=28.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:16,398] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:23,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:44:24,244] loop.py:45 INFO TRAIN_STEP step=3796 loss=1.706377 Training: 21%|██ | 3796/18018 [8:14:10<31:43:53, 8.03s/step, loss=1.706, acc=0.635, acc_len=1.47, thru=13.7, I=18.6, T=29.5, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:44:24,239] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3795: forward=2960.3ms backward=1323.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:24,453] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:30,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:44:31,250] loop.py:45 INFO TRAIN_STEP step=3797 loss=1.734181 Training: 21%|██ | 3797/18018 [8:14:17<30:30:49, 7.72s/step, loss=1.734, acc=0.631, acc_len=1.47, thru=17.8, I=19.1, T=33.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:31,484] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:38,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:44:40,009] loop.py:45 INFO TRAIN_STEP step=3798 loss=1.887264 Training: 21%|██ | 3798/18018 [8:14:26<31:44:21, 8.04s/step, loss=1.887, acc=0.609, acc_len=1.40, thru=14.5, I=17.4, T=24.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:40,312] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:47,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:44:48,863] loop.py:45 INFO TRAIN_STEP step=3799 loss=1.711050 Training: 21%|██ | 3799/18018 [8:14:34<32:42:25, 8.28s/step, loss=1.711, acc=0.637, acc_len=1.49, thru=15.0, I=16.8, T=25.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:44:49,035] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:44:56,731] loop.py:380 INFO TIMING step=3800: step=3.847s data=1.734s compute=3.799s [fwd=2.683s bwd=1.101s opt=0.003s] dispatch=4.011s [2026-10-18 21:44:56,731] loop.py:45 INFO TRAIN_STEP step=3800 loss=1.581313 Training: 21%|██ | 3800/18018 [8:14:42<32:12:49, 8.16s/step, loss=1.581, acc=0.653, acc_len=1.54, thru=11.4, I=17.3, T=33.3, wait=4.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:03,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:45:05,075] loop.py:45 INFO TRAIN_STEP step=3801 loss=1.652249 Training: 21%|██ | 3801/18018 [8:14:51<32:25:54, 8.21s/step, loss=1.652, acc=0.647, acc_len=1.52, thru=16.9, I=18.4, T=30.9, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:45:05,070] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3800: forward=2869.0ms backward=1219.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:05,575] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:10,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:45:11,663] loop.py:45 INFO TRAIN_STEP step=3802 loss=1.572715 Training: 21%|██ | 3802/18018 [8:14:57<30:30:22, 7.73s/step, loss=1.573, acc=0.660, acc_len=1.56, thru=17.3, I=24.8, T=32.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:12,075] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:18,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:45:19,076] loop.py:45 INFO TRAIN_STEP step=3803 loss=1.512395 Training: 21%|██ | 3803/18018 [8:15:05<30:08:12, 7.63s/step, loss=1.512, acc=0.668, acc_len=1.58, thru=16.7, I=21.2, T=30.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:19,115] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 21:45:26,671] loop.py:45 INFO TRAIN_STEP step=3804 loss=1.650032 Training: 21%|██ | 3804/18018 [8:15:12<30:05:37, 7.62s/step, loss=1.650, acc=0.643, acc_len=1.51, thru=15.2, I=18.6, T=27.4, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:37,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:45:39,052] loop.py:380 INFO TIMING step=3805: step=5.039s data=2.491s compute=5.010s [fwd=3.522s bwd=1.472s opt=0.003s] dispatch=7.329s [2026-10-18 21:45:39,052] loop.py:45 INFO TRAIN_STEP step=3805 loss=1.552041 Training: 21%|██ | 3805/18018 [8:15:25<35:43:31, 9.05s/step, loss=1.552, acc=0.663, acc_len=1.58, thru=10.8, I=10.8, T=25.4, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:39,161] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:44,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 21:45:45,714] loop.py:45 INFO TRAIN_STEP step=3806 loss=1.687381 Training: 21%|██ | 3806/18018 [8:15:31<32:53:43, 8.33s/step, loss=1.687, acc=0.643, acc_len=1.50, thru=17.5, I=22.3, T=32.9, wait=2.8s, pool=68, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:45:45,709] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3805: forward=2729.4ms backward=1116.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:45:45,915] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=4/64 [2026-10-18 21:45:54,695] loop.py:45 INFO TRAIN_STEP step=3807 loss=1.642203 Training: 21%|██ | 3807/18018 [8:15:40<33:39:28, 8.53s/step, loss=1.642, acc=0.647, acc_len=1.52, thru=13.2, I=13.3, T=32.3, wait=5.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:01,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:46:02,682] loop.py:45 INFO TRAIN_STEP step=3808 loss=1.640355 Training: 21%|██ | 3808/18018 [8:15:48<33:01:12, 8.37s/step, loss=1.640, acc=0.645, acc_len=1.52, thru=16.7, I=18.9, T=29.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:02,804] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:07,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:46:10,041] loop.py:45 INFO TRAIN_STEP step=3809 loss=1.654899 Training: 21%|██ | 3809/18018 [8:15:56<31:49:34, 8.06s/step, loss=1.655, acc=0.642, acc_len=1.51, thru=15.2, I=19.9, T=30.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:10,472] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:18,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:46:19,964] loop.py:380 INFO TIMING step=3810: step=4.413s data=2.228s compute=4.345s [fwd=3.073s bwd=1.256s opt=0.003s] dispatch=5.500s [2026-10-18 21:46:19,964] loop.py:45 INFO TRAIN_STEP step=3810 loss=1.764553 Training: 21%|██ | 3810/18018 [8:16:06<34:01:33, 8.62s/step, loss=1.765, acc=0.628, acc_len=1.46, thru=13.5, I=15.1, T=29.0, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:20,140] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:25,903] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:46:27,831] loop.py:45 INFO TRAIN_STEP step=3811 loss=1.710030 Training: 21%|██ | 3811/18018 [8:16:13<33:07:43, 8.39s/step, loss=1.710, acc=0.636, acc_len=1.48, thru=14.5, I=22.2, T=27.0, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:46:27,826] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3810: forward=2998.6ms backward=1701.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:27,930] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:33,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:46:35,278] loop.py:45 INFO TRAIN_STEP step=3812 loss=1.826908 Training: 21%|██ | 3812/18018 [8:16:21<32:00:17, 8.11s/step, loss=1.827, acc=0.619, acc_len=1.42, thru=14.8, I=19.9, T=31.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:35,700] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:40,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:46:41,815] loop.py:45 INFO TRAIN_STEP step=3813 loss=1.662358 Training: 21%|██ | 3813/18018 [8:16:27<30:08:28, 7.64s/step, loss=1.662, acc=0.642, acc_len=1.50, thru=17.9, I=25.6, T=32.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:41,903] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:48,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:46:49,561] loop.py:45 INFO TRAIN_STEP step=3814 loss=1.566041 Training: 21%|██ | 3814/18018 [8:16:35<30:15:57, 7.67s/step, loss=1.566, acc=0.657, acc_len=1.56, thru=14.8, I=17.8, T=28.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:49,711] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:56,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:46:57,628] loop.py:380 INFO TIMING step=3815: step=4.084s data=1.807s compute=4.053s [fwd=2.842s bwd=1.195s opt=0.003s] dispatch=3.972s [2026-10-18 21:46:57,628] loop.py:45 INFO TRAIN_STEP step=3815 loss=1.582093 Training: 21%|██ | 3815/18018 [8:16:43<30:43:50, 7.79s/step, loss=1.582, acc=0.654, acc_len=1.55, thru=14.4, I=18.3, T=31.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:46:57,813] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:05,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:47:07,139] loop.py:45 INFO TRAIN_STEP step=3816 loss=1.568496 Training: 21%|██ | 3816/18018 [8:16:53<32:45:57, 8.31s/step, loss=1.568, acc=0.657, acc_len=1.56, thru=13.7, I=16.2, T=33.7, wait=5.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:47:07,135] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3815: forward=2651.6ms backward=1088.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:07,198] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:13,114] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:47:15,098] loop.py:45 INFO TRAIN_STEP step=3817 loss=1.638954 Training: 21%|██ | 3817/18018 [8:17:01<32:21:22, 8.20s/step, loss=1.639, acc=0.646, acc_len=1.53, thru=14.2, I=19.1, T=30.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:15,136] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:21,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:47:21,996] loop.py:45 INFO TRAIN_STEP step=3818 loss=1.582593 Training: 21%|██ | 3818/18018 [8:17:08<30:48:35, 7.81s/step, loss=1.583, acc=0.656, acc_len=1.55, thru=16.3, I=21.4, T=28.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:22,148] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:27,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:47:29,177] loop.py:45 INFO TRAIN_STEP step=3819 loss=1.720886 Training: 21%|██ | 3819/18018 [8:17:15<30:03:42, 7.62s/step, loss=1.721, acc=0.635, acc_len=1.49, thru=17.9, I=21.2, T=30.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:29,649] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:35,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:47:37,457] loop.py:380 INFO TIMING step=3820: step=4.140s data=1.916s compute=4.107s [fwd=2.956s bwd=1.136s opt=0.003s] dispatch=4.130s [2026-10-18 21:47:37,457] loop.py:45 INFO TRAIN_STEP step=3820 loss=1.748462 Training: 21%|██ | 3820/18018 [8:17:23<30:50:18, 7.82s/step, loss=1.748, acc=0.629, acc_len=1.46, thru=13.7, I=21.4, T=30.9, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:37,733] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:47:43,362] loop.py:45 INFO TRAIN_STEP step=3821 loss=1.743620 Training: 21%|██ | 3821/18018 [8:17:29<28:34:04, 7.24s/step, loss=1.744, acc=0.626, acc_len=1.46, thru=19.4, I=22.2, T=33.6, wait=2.1s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:47:43,357] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3820: forward=2625.7ms backward=1131.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:50,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:47:51,427] loop.py:45 INFO TRAIN_STEP step=3822 loss=1.727512 Training: 21%|██ | 3822/18018 [8:17:37<29:32:24, 7.49s/step, loss=1.728, acc=0.633, acc_len=1.48, thru=18.0, I=18.3, T=30.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:51,575] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:58,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:47:59,074] loop.py:45 INFO TRAIN_STEP step=3823 loss=1.706748 Training: 21%|██ | 3823/18018 [8:17:45<29:43:22, 7.54s/step, loss=1.707, acc=0.634, acc_len=1.48, thru=15.8, I=18.8, T=27.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:47:59,193] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-18 21:48:05,802] loop.py:45 INFO TRAIN_STEP step=3824 loss=1.574763 Training: 21%|██ | 3824/18018 [8:17:51<28:45:31, 7.29s/step, loss=1.575, acc=0.656, acc_len=1.55, thru=18.1, I=17.7, T=32.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:05,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:06,298] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:10,218] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:48:12,587] loop.py:380 INFO TIMING step=3825: step=4.109s data=1.891s compute=4.080s [fwd=2.864s bwd=1.200s opt=0.003s] dispatch=2.669s [2026-10-18 21:48:12,588] loop.py:45 INFO TRAIN_STEP step=3825 loss=1.593905 Training: 21%|██ | 3825/18018 [8:17:58<28:09:34, 7.14s/step, loss=1.594, acc=0.656, acc_len=1.55, thru=19.2, I=22.4, T=31.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:12,749] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:19,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:48:19,184] loop.py:45 INFO TRAIN_STEP step=3826 loss=1.668035 Training: 21%|██ | 3826/18018 [8:18:05<27:30:32, 6.98s/step, loss=1.668, acc=0.640, acc_len=1.50, thru=15.8, I=22.7, T=31.2, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:48:19,179] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3825: forward=2830.2ms backward=1230.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:19,550] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:26,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:48:28,742] loop.py:45 INFO TRAIN_STEP step=3827 loss=1.609971 Training: 21%|██ | 3827/18018 [8:18:14<30:33:26, 7.75s/step, loss=1.610, acc=0.652, acc_len=1.54, thru=13.6, I=16.6, T=27.8, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:29,162] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 21:48:34,892] loop.py:45 INFO TRAIN_STEP step=3828 loss=1.595935 Training: 21%|██ | 3828/18018 [8:18:20<28:39:58, 7.27s/step, loss=1.596, acc=0.656, acc_len=1.54, thru=18.1, I=23.6, T=32.8, wait=2.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:41,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:48:42,098] loop.py:45 INFO TRAIN_STEP step=3829 loss=1.547967 Training: 21%|██▏ | 3829/18018 [8:18:28<28:34:51, 7.25s/step, loss=1.548, acc=0.666, acc_len=1.57, thru=17.8, I=18.0, T=31.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:42,364] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:47,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:48:49,251] loop.py:380 INFO TIMING step=3830: step=4.426s data=2.077s compute=4.400s [fwd=3.141s bwd=1.244s opt=0.003s] dispatch=2.718s [2026-10-18 21:48:49,252] loop.py:45 INFO TRAIN_STEP step=3830 loss=1.682482 Training: 21%|██▏ | 3830/18018 [8:18:35<28:27:48, 7.22s/step, loss=1.682, acc=0.634, acc_len=1.48, thru=19.0, I=20.4, T=28.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:49,419] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:54,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 21:48:56,983] loop.py:45 INFO TRAIN_STEP step=3831 loss=1.634547 Training: 21%|██▏ | 3831/18018 [8:18:43<29:03:56, 7.38s/step, loss=1.635, acc=0.646, acc_len=1.51, thru=15.4, I=20.4, T=28.7, wait=3.3s, pool=68, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:48:56,978] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3830: forward=3090.7ms backward=1315.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:48:57,137] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=12/64 [2026-10-18 21:49:03,005] loop.py:45 INFO TRAIN_STEP step=3832 loss=1.518828 Training: 21%|██▏ | 3832/18018 [8:18:49<27:27:45, 6.97s/step, loss=1.519, acc=0.664, acc_len=1.57, thru=16.8, I=25.5, T=31.9, wait=2.0s, pool=52, epoch=1/2][2026-10-18 21:49:10,466] loop.py:45 INFO TRAIN_STEP step=3833 loss=1.634734 Training: 21%|██▏ | 3833/18018 [8:18:56<28:02:47, 7.12s/step, loss=1.635, acc=0.650, acc_len=1.53, thru=15.0, I=16.4, T=33.9, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:49:17,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:49:19,547] loop.py:45 INFO TRAIN_STEP step=3834 loss=1.687057 Training: 21%|██▏ | 3834/18018 [8:19:05<30:21:39, 7.71s/step, loss=1.687, acc=0.643, acc_len=1.51, thru=15.1, I=16.6, T=24.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:49:19,870] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:49:26,942] loop.py:380 INFO TIMING step=3835: step=3.804s data=1.671s compute=3.775s [fwd=2.694s bwd=1.065s opt=0.003s] dispatch=3.581s [2026-10-18 21:49:26,942] loop.py:45 INFO TRAIN_STEP step=3835 loss=1.665772 Training: 21%|██▏ | 3835/18018 [8:19:13<29:59:33, 7.61s/step, loss=1.666, acc=0.644, acc_len=1.51, thru=12.2, I=16.4, T=33.7, wait=3.6s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:49:36,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:49:37,326] loop.py:45 INFO TRAIN_STEP step=3836 loss=1.680872 Training: 21%|██▏ | 3836/18018 [8:19:23<33:15:49, 8.44s/step, loss=1.681, acc=0.640, acc_len=1.49, thru=15.4, I=15.1, T=30.0, wait=6.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:49:37,321] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3835: forward=3016.3ms backward=1196.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:49:37,459] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:49:43,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:49:44,579] loop.py:45 INFO TRAIN_STEP step=3837 loss=1.598482 Training: 21%|██▏ | 3837/18018 [8:19:30<31:51:17, 8.09s/step, loss=1.598, acc=0.658, acc_len=1.55, thru=17.0, I=19.3, T=31.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:49:44,800] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:49:50,974] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:49:51,381] loop.py:45 INFO TRAIN_STEP step=3838 loss=1.591287 Training: 21%|██▏ | 3838/18018 [8:19:37<30:19:55, 7.70s/step, loss=1.591, acc=0.651, acc_len=1.53, thru=17.8, I=21.6, T=31.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:49:51,482] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-18 21:50:01,402] loop.py:45 INFO TRAIN_STEP step=3839 loss=1.736242 Training: 21%|██▏ | 3839/18018 [8:19:47<33:04:18, 8.40s/step, loss=1.736, acc=0.634, acc_len=1.47, thru=12.5, I=12.2, T=29.9, wait=5.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:06,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:50:08,055] loop.py:380 INFO TIMING step=3840: step=4.311s data=1.914s compute=4.280s [fwd=2.975s bwd=1.290s opt=0.003s] dispatch=2.333s [2026-10-18 21:50:08,055] loop.py:45 INFO TRAIN_STEP step=3840 loss=1.611870 Training: 21%|██▏ | 3840/18018 [8:19:54<31:00:43, 7.87s/step, loss=1.612, acc=0.647, acc_len=1.52, thru=18.5, I=20.2, T=29.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:08,203] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:12,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:50:14,044] loop.py:45 INFO TRAIN_STEP step=3841 loss=1.549173 Training: 21%|██▏ | 3841/18018 [8:20:00<28:46:46, 7.31s/step, loss=1.549, acc=0.659, acc_len=1.56, thru=19.2, I=26.8, T=34.2, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:50:14,038] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3840: forward=2653.2ms backward=1047.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:14,222] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:21,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:50:22,896] loop.py:45 INFO TRAIN_STEP step=3842 loss=1.698961 Training: 21%|██▏ | 3842/18018 [8:20:08<30:36:13, 7.77s/step, loss=1.699, acc=0.637, acc_len=1.49, thru=14.8, I=15.7, T=30.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:23,327] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:30,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:50:31,205] loop.py:45 INFO TRAIN_STEP step=3843 loss=1.673671 Training: 21%|██▏ | 3843/18018 [8:20:17<31:14:09, 7.93s/step, loss=1.674, acc=0.644, acc_len=1.51, thru=13.8, I=15.4, T=29.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:31,243] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:36,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:50:38,488] loop.py:45 INFO TRAIN_STEP step=3844 loss=1.501778 Training: 21%|██▏ | 3844/18018 [8:20:24<30:27:46, 7.74s/step, loss=1.502, acc=0.672, acc_len=1.60, thru=16.3, I=21.2, T=27.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:38,710] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 21:50:44,472] loop.py:380 INFO TIMING step=3845: step=3.790s data=1.662s compute=3.752s [fwd=2.591s bwd=1.146s opt=0.003s] dispatch=2.187s [2026-10-18 21:50:44,473] loop.py:45 INFO TRAIN_STEP step=3845 loss=1.620778 Training: 21%|██▏ | 3845/18018 [8:20:30<28:23:25, 7.21s/step, loss=1.621, acc=0.650, acc_len=1.53, thru=17.7, I=24.7, T=33.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:44,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:50:44,945] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 21:50:51,842] loop.py:45 INFO TRAIN_STEP step=3846 loss=1.682400 Training: 21%|██▏ | 3846/18018 [8:20:37<28:34:35, 7.26s/step, loss=1.682, acc=0.646, acc_len=1.51, thru=15.3, I=16.4, T=30.9, wait=3.2s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:50:51,837] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3845: forward=2889.0ms backward=1207.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:01,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:51:02,853] loop.py:45 INFO TRAIN_STEP step=3847 loss=1.826157 Training: 21%|██▏ | 3847/18018 [8:20:48<33:00:26, 8.39s/step, loss=1.826, acc=0.616, acc_len=1.42, thru=13.3, I=14.8, T=27.6, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:03,237] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:10,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:51:10,657] loop.py:45 INFO TRAIN_STEP step=3848 loss=1.697739 Training: 21%|██▏ | 3848/18018 [8:20:56<32:19:05, 8.21s/step, loss=1.698, acc=0.634, acc_len=1.48, thru=14.4, I=21.1, T=25.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:10,690] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 21:51:19,005] loop.py:45 INFO TRAIN_STEP step=3849 loss=1.714376 Training: 21%|██▏ | 3849/18018 [8:21:05<32:28:45, 8.25s/step, loss=1.714, acc=0.636, acc_len=1.49, thru=14.0, I=13.7, T=32.2, wait=4.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:30,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:51:32,299] loop.py:380 INFO TIMING step=3850: step=4.670s data=2.176s compute=4.646s [fwd=3.258s bwd=1.374s opt=0.003s] dispatch=8.612s [2026-10-18 21:51:32,300] loop.py:45 INFO TRAIN_STEP step=3850 loss=1.816110 Training: 21%|██▏ | 3850/18018 [8:21:18<38:25:46, 9.76s/step, loss=1.816, acc=0.622, acc_len=1.44, thru=10.4, I=10.1, T=27.4, wait=8.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:32,698] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:38,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:51:38,825] loop.py:45 INFO TRAIN_STEP step=3851 loss=1.579886 Training: 21%|██▏ | 3851/18018 [8:21:24<34:36:03, 8.79s/step, loss=1.580, acc=0.654, acc_len=1.55, thru=18.2, I=22.3, T=32.1, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:51:38,821] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3850: forward=2805.6ms backward=1139.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:39,097] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:44,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:51:46,449] loop.py:45 INFO TRAIN_STEP step=3852 loss=1.687184 Training: 21%|██▏ | 3852/18018 [8:21:32<33:13:07, 8.44s/step, loss=1.687, acc=0.639, acc_len=1.49, thru=16.1, I=20.0, T=31.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:46,754] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:51,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:51:52,880] loop.py:45 INFO TRAIN_STEP step=3853 loss=1.716131 Training: 21%|██▏ | 3853/18018 [8:21:38<30:50:34, 7.84s/step, loss=1.716, acc=0.635, acc_len=1.48, thru=17.6, I=25.6, T=34.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:53,343] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:51:59,938] loop.py:45 INFO TRAIN_STEP step=3854 loss=1.557490 Training: 21%|██▏ | 3854/18018 [8:21:46<29:55:02, 7.60s/step, loss=1.557, acc=0.659, acc_len=1.56, thru=17.8, I=20.4, T=31.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:51:59,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:00,370] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:06,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:52:07,048] loop.py:380 INFO TIMING step=3855: step=4.062s data=1.846s compute=4.013s [fwd=2.822s bwd=1.177s opt=0.003s] dispatch=3.041s [2026-10-18 21:52:07,048] loop.py:45 INFO TRAIN_STEP step=3855 loss=1.558383 Training: 21%|██▏ | 3855/18018 [8:21:53<29:20:02, 7.46s/step, loss=1.558, acc=0.660, acc_len=1.56, thru=17.9, I=18.0, T=31.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:07,383] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 21:52:16,191] loop.py:45 INFO TRAIN_STEP step=3856 loss=1.693321 Training: 21%|██▏ | 3856/18018 [8:22:02<31:19:15, 7.96s/step, loss=1.693, acc=0.640, acc_len=1.49, thru=12.5, I=14.0, T=31.6, wait=5.1s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:52:16,186] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3855: forward=2906.3ms backward=1101.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:24,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:52:26,411] loop.py:45 INFO TRAIN_STEP step=3857 loss=1.746763 Training: 21%|██▏ | 3857/18018 [8:22:12<33:59:13, 8.64s/step, loss=1.747, acc=0.631, acc_len=1.47, thru=14.9, I=16.1, T=27.2, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:26,541] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:31,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:52:33,434] loop.py:45 INFO TRAIN_STEP step=3858 loss=1.585642 Training: 21%|██▏ | 3858/18018 [8:22:19<32:04:31, 8.15s/step, loss=1.586, acc=0.654, acc_len=1.54, thru=15.3, I=24.4, T=26.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:33,940] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:40,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:52:40,899] loop.py:45 INFO TRAIN_STEP step=3859 loss=1.586923 Training: 21%|██▏ | 3859/18018 [8:22:26<31:15:31, 7.95s/step, loss=1.587, acc=0.654, acc_len=1.55, thru=14.4, I=20.2, T=29.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:40,930] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:47,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:52:48,194] loop.py:380 INFO TIMING step=3860: step=4.546s data=2.272s compute=4.478s [fwd=3.097s bwd=1.366s opt=0.003s] dispatch=2.741s [2026-10-18 21:52:48,194] loop.py:45 INFO TRAIN_STEP step=3860 loss=1.582278 Training: 21%|██▏ | 3860/18018 [8:22:34<30:29:12, 7.75s/step, loss=1.582, acc=0.659, acc_len=1.55, thru=16.3, I=16.5, T=28.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:48,460] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:54,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:52:55,768] loop.py:45 INFO TRAIN_STEP step=3861 loss=1.792470 Training: 21%|██▏ | 3861/18018 [8:22:41<30:16:36, 7.70s/step, loss=1.792, acc=0.623, acc_len=1.44, thru=17.0, I=18.3, T=29.8, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:52:55,763] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3860: forward=3055.7ms backward=1184.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:52:55,934] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:02,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:53:04,302] loop.py:45 INFO TRAIN_STEP step=3862 loss=1.736895 Training: 21%|██▏ | 3862/18018 [8:22:50<31:15:31, 7.95s/step, loss=1.737, acc=0.633, acc_len=1.47, thru=14.1, I=18.5, T=29.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:04,423] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:08,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:53:11,440] loop.py:45 INFO TRAIN_STEP step=3863 loss=1.867371 Training: 21%|██▏ | 3863/18018 [8:22:57<30:17:52, 7.71s/step, loss=1.867, acc=0.612, acc_len=1.41, thru=16.2, I=23.9, T=28.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:11,471] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:18,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:53:18,742] loop.py:45 INFO TRAIN_STEP step=3864 loss=1.577965 Training: 21%|██▏ | 3864/18018 [8:23:04<29:49:08, 7.58s/step, loss=1.578, acc=0.660, acc_len=1.56, thru=14.1, I=19.4, T=26.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:19,156] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:24,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:53:25,995] loop.py:380 INFO TIMING step=3865: step=4.598s data=2.240s compute=4.564s [fwd=3.155s bwd=1.394s opt=0.003s] dispatch=2.646s [2026-10-18 21:53:25,995] loop.py:45 INFO TRAIN_STEP step=3865 loss=1.552312 Training: 21%|██▏ | 3865/18018 [8:23:12<29:25:35, 7.48s/step, loss=1.552, acc=0.661, acc_len=1.56, thru=16.9, I=18.2, T=27.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:26,326] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:34,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:53:34,841] loop.py:45 INFO TRAIN_STEP step=3866 loss=1.643128 Training: 21%|██▏ | 3866/18018 [8:23:20<31:01:51, 7.89s/step, loss=1.643, acc=0.647, acc_len=1.52, thru=15.0, I=16.2, T=31.9, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:53:34,836] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3865: forward=2755.7ms backward=1200.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:35,079] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:44,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:53:44,958] loop.py:45 INFO TRAIN_STEP step=3867 loss=1.608880 Training: 21%|██▏ | 3867/18018 [8:23:31<33:38:58, 8.56s/step, loss=1.609, acc=0.654, acc_len=1.53, thru=13.0, I=13.2, T=30.1, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:45,115] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:51,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:53:52,140] loop.py:45 INFO TRAIN_STEP step=3868 loss=1.736119 Training: 21%|██▏ | 3868/18018 [8:23:38<32:01:19, 8.15s/step, loss=1.736, acc=0.624, acc_len=1.45, thru=17.6, I=18.5, T=29.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:52,646] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:53:59,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:54:01,353] loop.py:45 INFO TRAIN_STEP step=3869 loss=1.750099 Training: 21%|██▏ | 3869/18018 [8:23:47<33:16:34, 8.47s/step, loss=1.750, acc=0.627, acc_len=1.46, thru=13.6, I=18.1, T=27.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:01,471] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:09,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:54:10,001] loop.py:380 INFO TIMING step=3870: step=4.782s data=2.383s compute=4.756s [fwd=3.395s bwd=1.345s opt=0.003s] dispatch=3.856s [2026-10-18 21:54:10,001] loop.py:45 INFO TRAIN_STEP step=3870 loss=1.583925 Training: 21%|██▏ | 3870/18018 [8:23:56<33:29:14, 8.52s/step, loss=1.584, acc=0.657, acc_len=1.56, thru=15.0, I=13.5, T=26.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:10,284] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:16,538] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:54:17,088] loop.py:45 INFO TRAIN_STEP step=3871 loss=1.719898 Training: 21%|██▏ | 3871/18018 [8:24:03<31:47:38, 8.09s/step, loss=1.720, acc=0.635, acc_len=1.48, thru=17.7, I=19.2, T=30.3, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:54:17,083] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3870: forward=2950.9ms backward=1220.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:17,543] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 21:54:25,294] loop.py:45 INFO TRAIN_STEP step=3872 loss=1.699186 Training: 21%|██▏ | 3872/18018 [8:24:11<31:55:51, 8.13s/step, loss=1.699, acc=0.640, acc_len=1.50, thru=13.7, I=13.1, T=29.4, wait=3.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:35,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:54:36,776] loop.py:45 INFO TRAIN_STEP step=3873 loss=1.661248 Training: 21%|██▏ | 3873/18018 [8:24:22<35:52:53, 9.13s/step, loss=1.661, acc=0.645, acc_len=1.51, thru=13.3, I=13.6, T=25.5, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:37,028] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:41,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:54:43,303] loop.py:45 INFO TRAIN_STEP step=3874 loss=1.650376 Training: 22%|██▏ | 3874/18018 [8:24:29<32:48:37, 8.35s/step, loss=1.650, acc=0.644, acc_len=1.51, thru=16.6, I=23.4, T=33.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:43,714] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:49,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:54:49,954] loop.py:380 INFO TIMING step=3875: step=3.983s data=1.845s compute=3.955s [fwd=2.814s bwd=1.127s opt=0.003s] dispatch=2.659s [2026-10-18 21:54:49,954] loop.py:45 INFO TRAIN_STEP step=3875 loss=1.688641 Training: 22%|██▏ | 3875/18018 [8:24:36<30:48:11, 7.84s/step, loss=1.689, acc=0.636, acc_len=1.49, thru=18.2, I=23.8, T=32.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:50,325] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:56,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:54:56,986] loop.py:45 INFO TRAIN_STEP step=3876 loss=1.622307 Training: 22%|██▏ | 3876/18018 [8:24:43<29:50:58, 7.60s/step, loss=1.622, acc=0.651, acc_len=1.53, thru=19.2, I=20.3, T=33.0, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:54:56,980] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3875: forward=2743.2ms backward=1095.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:54:57,046] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:04,822] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:55:05,930] loop.py:45 INFO TRAIN_STEP step=3877 loss=1.736171 Training: 22%|██▏ | 3877/18018 [8:24:52<31:25:55, 8.00s/step, loss=1.736, acc=0.630, acc_len=1.47, thru=15.3, I=18.2, T=30.7, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:06,333] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:55:15,128] loop.py:45 INFO TRAIN_STEP step=3878 loss=1.667257 Training: 22%|██▏ | 3878/18018 [8:25:01<32:50:16, 8.36s/step, loss=1.667, acc=0.641, acc_len=1.51, thru=13.7, I=13.6, T=31.1, wait=5.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:21,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:55:21,870] loop.py:45 INFO TRAIN_STEP step=3879 loss=1.725952 Training: 22%|██▏ | 3879/18018 [8:25:07<30:55:43, 7.87s/step, loss=1.726, acc=0.634, acc_len=1.48, thru=19.1, I=18.4, T=31.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:21,956] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:27,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:55:28,597] loop.py:380 INFO TIMING step=3880: step=4.337s data=2.006s compute=4.304s [fwd=3.032s bwd=1.257s opt=0.003s] dispatch=2.382s [2026-10-18 21:55:28,597] loop.py:45 INFO TRAIN_STEP step=3880 loss=1.622456 Training: 22%|██▏ | 3880/18018 [8:25:14<29:34:29, 7.53s/step, loss=1.622, acc=0.652, acc_len=1.53, thru=17.3, I=19.8, T=29.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:29,050] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:55:36,817] loop.py:45 INFO TRAIN_STEP step=3881 loss=1.488670 Training: 22%|██▏ | 3881/18018 [8:25:22<30:23:07, 7.74s/step, loss=1.489, acc=0.670, acc_len=1.59, thru=14.1, I=14.8, T=33.1, wait=4.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:55:36,812] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3880: forward=2692.0ms backward=1133.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:42,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:55:44,799] loop.py:45 INFO TRAIN_STEP step=3882 loss=1.672216 Training: 22%|██▏ | 3882/18018 [8:25:30<30:40:09, 7.81s/step, loss=1.672, acc=0.641, acc_len=1.50, thru=17.0, I=20.8, T=28.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:45,068] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:52,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:55:52,588] loop.py:45 INFO TRAIN_STEP step=3883 loss=1.685221 Training: 22%|██▏ | 3883/18018 [8:25:38<30:38:31, 7.80s/step, loss=1.685, acc=0.640, acc_len=1.50, thru=17.0, I=16.0, T=29.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:52,926] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:55:59,912] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:56:00,117] loop.py:45 INFO TRAIN_STEP step=3884 loss=1.710639 Training: 22%|██▏ | 3884/18018 [8:25:46<30:18:59, 7.72s/step, loss=1.711, acc=0.636, acc_len=1.48, thru=15.8, I=17.4, T=31.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:56:00,414] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 21:56:08,831] loop.py:380 INFO TIMING step=3885: step=4.585s data=2.368s compute=4.558s [fwd=3.405s bwd=1.138s opt=0.003s] dispatch=4.119s [2026-10-18 21:56:08,831] loop.py:45 INFO TRAIN_STEP step=3885 loss=1.766312 Training: 22%|██▏ | 3885/18018 [8:25:54<31:29:02, 8.02s/step, loss=1.766, acc=0.628, acc_len=1.45, thru=12.4, I=13.2, T=27.9, wait=4.1s, pool=40, epoch=1/2][2026-10-18 21:56:17,059] loop.py:45 INFO TRAIN_STEP step=3886 loss=1.656089 Training: 22%|██▏ | 3886/18018 [8:26:03<31:43:33, 8.08s/step, loss=1.656, acc=0.647, acc_len=1.52, thru=16.5, I=16.0, T=28.9, wait=3.8s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:56:17,055] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3885: forward=3074.3ms backward=1297.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:56:23,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:56:24,859] loop.py:45 INFO TRAIN_STEP step=3887 loss=1.722703 Training: 22%|██▏ | 3887/18018 [8:26:10<31:23:23, 8.00s/step, loss=1.723, acc=0.634, acc_len=1.48, thru=15.7, I=17.0, T=31.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:56:25,344] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:56:32,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:56:33,209] loop.py:45 INFO TRAIN_STEP step=3888 loss=1.600730 Training: 22%|██▏ | 3888/18018 [8:26:19<31:48:28, 8.10s/step, loss=1.601, acc=0.653, acc_len=1.53, thru=13.4, I=18.8, T=26.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:56:33,412] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:56:39,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:56:39,526] loop.py:45 INFO TRAIN_STEP step=3889 loss=1.692542 Training: 22%|██▏ | 3889/18018 [8:26:25<29:42:06, 7.57s/step, loss=1.693, acc=0.643, acc_len=1.51, thru=17.7, I=20.9, T=35.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:56:39,563] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 21:56:45,749] loop.py:380 INFO TIMING step=3890: step=4.229s data=2.040s compute=4.187s [fwd=2.888s bwd=1.284s opt=0.003s] dispatch=1.983s [2026-10-18 21:56:45,749] loop.py:45 INFO TRAIN_STEP step=3890 loss=1.470527 Training: 22%|██▏ | 3890/18018 [8:26:31<28:06:53, 7.16s/step, loss=1.471, acc=0.671, acc_len=1.61, thru=19.0, I=22.1, T=30.3, wait=2.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:56:53,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:56:55,714] loop.py:45 INFO TRAIN_STEP step=3891 loss=1.700613 Training: 22%|██▏ | 3891/18018 [8:26:41<31:24:29, 8.00s/step, loss=1.701, acc=0.636, acc_len=1.48, thru=14.6, I=15.1, T=27.9, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:56:55,709] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3890: forward=3215.9ms backward=1331.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:56:56,206] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:02,019] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:57:02,510] loop.py:45 INFO TRAIN_STEP step=3892 loss=1.635188 Training: 22%|██▏ | 3892/18018 [8:26:48<29:59:14, 7.64s/step, loss=1.635, acc=0.646, acc_len=1.51, thru=16.6, I=22.1, T=34.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:02,545] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:09,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:57:10,405] loop.py:45 INFO TRAIN_STEP step=3893 loss=1.543965 Training: 22%|██▏ | 3893/18018 [8:26:56<30:16:58, 7.72s/step, loss=1.544, acc=0.661, acc_len=1.58, thru=16.4, I=17.8, T=35.7, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:10,441] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 21:57:19,033] loop.py:45 INFO TRAIN_STEP step=3894 loss=1.694725 Training: 22%|██▏ | 3894/18018 [8:27:05<31:21:00, 7.99s/step, loss=1.695, acc=0.638, acc_len=1.49, thru=13.0, I=14.5, T=31.5, wait=4.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:26,690] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:57:28,014] loop.py:380 INFO TIMING step=3895: step=4.562s data=2.124s compute=4.493s [fwd=2.985s bwd=1.491s opt=0.003s] dispatch=4.408s [2026-10-18 21:57:28,014] loop.py:45 INFO TRAIN_STEP step=3895 loss=1.671596 Training: 22%|██▏ | 3895/18018 [8:27:14<32:30:38, 8.29s/step, loss=1.672, acc=0.640, acc_len=1.50, thru=16.3, I=17.5, T=28.1, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:28,216] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:57:35,597] loop.py:45 INFO TRAIN_STEP step=3896 loss=1.750948 Training: 22%|██▏ | 3896/18018 [8:27:21<31:41:04, 8.08s/step, loss=1.751, acc=0.630, acc_len=1.47, thru=13.8, I=18.0, T=31.2, wait=3.5s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:57:35,593] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3895: forward=2828.8ms backward=1211.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:40,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:57:42,045] loop.py:45 INFO TRAIN_STEP step=3897 loss=1.672083 Training: 22%|██▏ | 3897/18018 [8:27:28<29:45:43, 7.59s/step, loss=1.672, acc=0.640, acc_len=1.50, thru=18.2, I=20.0, T=36.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:42,328] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:48,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:57:48,920] loop.py:45 INFO TRAIN_STEP step=3898 loss=1.547958 Training: 22%|██▏ | 3898/18018 [8:27:35<28:55:18, 7.37s/step, loss=1.548, acc=0.661, acc_len=1.57, thru=17.7, I=22.0, T=32.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:49,175] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:53,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:57:55,182] loop.py:45 INFO TRAIN_STEP step=3899 loss=1.508878 Training: 22%|██▏ | 3899/18018 [8:27:41<27:36:41, 7.04s/step, loss=1.509, acc=0.671, acc_len=1.59, thru=18.8, I=23.7, T=31.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:57:55,210] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 21:58:02,695] loop.py:380 INFO TIMING step=3900: step=3.761s data=1.561s compute=3.713s [fwd=2.628s bwd=1.070s opt=0.003s] dispatch=3.743s [2026-10-18 21:58:02,695] loop.py:45 INFO TRAIN_STEP step=3900 loss=1.614667 Training: 22%|██▏ | 3900/18018 [8:27:48<28:10:06, 7.18s/step, loss=1.615, acc=0.650, acc_len=1.53, thru=13.2, I=14.9, T=34.0, wait=3.7s, pool=48, epoch=1/2][2026-10-18 21:58:11,415] loop.py:45 INFO TRAIN_STEP step=3901 loss=1.699042 Training: 22%|██▏ | 3901/18018 [8:27:57<29:58:22, 7.64s/step, loss=1.699, acc=0.643, acc_len=1.51, thru=15.4, I=15.0, T=31.0, wait=4.6s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:58:11,411] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3900: forward=2825.9ms backward=1259.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:16,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:58:17,590] loop.py:45 INFO TRAIN_STEP step=3902 loss=1.649067 Training: 22%|██▏ | 3902/18018 [8:28:03<28:14:34, 7.20s/step, loss=1.649, acc=0.641, acc_len=1.50, thru=20.7, I=20.0, T=33.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:17,783] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:25,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:58:26,704] loop.py:45 INFO TRAIN_STEP step=3903 loss=1.564921 Training: 22%|██▏ | 3903/18018 [8:28:12<30:29:26, 7.78s/step, loss=1.565, acc=0.658, acc_len=1.56, thru=13.3, I=16.0, T=28.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:26,782] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 21:58:35,442] loop.py:45 INFO TRAIN_STEP step=3904 loss=1.599711 Training: 22%|██▏ | 3904/18018 [8:28:21<31:37:10, 8.07s/step, loss=1.600, acc=0.656, acc_len=1.55, thru=14.4, I=14.6, T=28.7, wait=4.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:42,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:58:43,776] loop.py:380 INFO TIMING step=3905: step=4.218s data=1.851s compute=4.176s [fwd=2.953s bwd=1.209s opt=0.003s] dispatch=4.105s [2026-10-18 21:58:43,776] loop.py:45 INFO TRAIN_STEP step=3905 loss=1.708774 Training: 22%|██▏ | 3905/18018 [8:28:29<31:56:00, 8.15s/step, loss=1.709, acc=0.632, acc_len=1.47, thru=16.8, I=19.3, T=30.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:43,810] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:48,907] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:58:50,287] loop.py:45 INFO TRAIN_STEP step=3906 loss=1.581507 Training: 22%|██▏ | 3906/18018 [8:28:36<30:00:30, 7.66s/step, loss=1.582, acc=0.655, acc_len=1.55, thru=18.2, I=23.8, T=31.8, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 21:58:50,282] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3905: forward=2797.2ms backward=1170.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:50,423] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:58,179] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:58:59,358] loop.py:45 INFO TRAIN_STEP step=3907 loss=1.643888 Training: 22%|██▏ | 3907/18018 [8:28:45<31:40:08, 8.08s/step, loss=1.644, acc=0.648, acc_len=1.52, thru=14.3, I=15.3, T=28.2, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:58:59,686] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 21:59:06,390] loop.py:45 INFO TRAIN_STEP step=3908 loss=1.670791 Training: 22%|██▏ | 3908/18018 [8:28:52<30:26:05, 7.77s/step, loss=1.671, acc=0.639, acc_len=1.50, thru=13.0, I=20.2, T=30.4, wait=2.8s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:13,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:59:15,393] loop.py:45 INFO TRAIN_STEP step=3909 loss=1.697038 Training: 22%|██▏ | 3909/18018 [8:29:01<31:53:21, 8.14s/step, loss=1.697, acc=0.638, acc_len=1.49, thru=17.6, I=17.4, T=30.9, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:15,810] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:22,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:59:24,365] loop.py:380 INFO TIMING step=3910: step=4.696s data=2.207s compute=4.661s [fwd=3.279s bwd=1.366s opt=0.003s] dispatch=4.267s [2026-10-18 21:59:24,365] loop.py:45 INFO TRAIN_STEP step=3910 loss=1.654853 Training: 22%|██▏ | 3910/18018 [8:29:10<32:52:11, 8.39s/step, loss=1.655, acc=0.647, acc_len=1.52, thru=15.0, I=17.9, T=27.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:24,808] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 21:59:31,578] loop.py:45 INFO TRAIN_STEP step=3911 loss=1.668677 Training: 22%|██▏ | 3911/18018 [8:29:17<31:28:59, 8.03s/step, loss=1.669, acc=0.642, acc_len=1.50, thru=16.5, I=18.9, T=31.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:31,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 21:59:31,573] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3910: forward=2873.6ms backward=1140.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:32,079] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:38,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:59:38,869] loop.py:45 INFO TRAIN_STEP step=3912 loss=1.602442 Training: 22%|██▏ | 3912/18018 [8:29:24<30:36:40, 7.81s/step, loss=1.602, acc=0.653, acc_len=1.53, thru=17.6, I=18.6, T=29.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:39,039] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:43,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 21:59:44,527] loop.py:45 INFO TRAIN_STEP step=3913 loss=1.689438 Training: 22%|██▏ | 3913/18018 [8:29:30<28:04:38, 7.17s/step, loss=1.689, acc=0.634, acc_len=1.48, thru=19.3, I=23.1, T=34.3, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 21:59:44,634] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 21:59:52,641] loop.py:45 INFO TRAIN_STEP step=3914 loss=1.480087 Training: 22%|██▏ | 3914/18018 [8:29:38<29:11:24, 7.45s/step, loss=1.480, acc=0.675, acc_len=1.61, thru=12.6, I=14.4, T=31.4, wait=4.0s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:03,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:00:05,329] loop.py:380 INFO TIMING step=3915: step=4.759s data=2.256s compute=4.710s [fwd=3.167s bwd=1.528s opt=0.003s] dispatch=7.917s [2026-10-18 22:00:05,329] loop.py:45 INFO TRAIN_STEP step=3915 loss=1.781590 Training: 22%|██▏ | 3915/18018 [8:29:51<35:20:28, 9.02s/step, loss=1.782, acc=0.627, acc_len=1.45, thru=13.2, I=16.2, T=26.9, wait=7.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:05,658] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 22:00:11,863] loop.py:45 INFO TRAIN_STEP step=3916 loss=1.615900 Training: 22%|██▏ | 3916/18018 [8:29:57<32:24:45, 8.27s/step, loss=1.616, acc=0.651, acc_len=1.53, thru=14.0, I=18.5, T=35.5, wait=2.9s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:00:11,859] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3915: forward=2561.0ms backward=1007.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:18,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:00:18,862] loop.py:45 INFO TRAIN_STEP step=3917 loss=1.669192 Training: 22%|██▏ | 3917/18018 [8:30:04<30:54:57, 7.89s/step, loss=1.669, acc=0.646, acc_len=1.52, thru=19.1, I=19.9, T=33.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:18,930] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:26,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:00:26,746] loop.py:45 INFO TRAIN_STEP step=3918 loss=1.902392 Training: 22%|██▏ | 3918/18018 [8:30:12<30:54:04, 7.89s/step, loss=1.902, acc=0.600, acc_len=1.38, thru=15.2, I=16.9, T=33.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:27,236] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:00:33,884] loop.py:45 INFO TRAIN_STEP step=3919 loss=1.746301 Training: 22%|██▏ | 3919/18018 [8:30:19<30:01:06, 7.66s/step, loss=1.746, acc=0.628, acc_len=1.46, thru=17.2, I=23.4, T=31.0, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:41,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:00:41,609] loop.py:380 INFO TIMING step=3920: step=3.967s data=1.717s compute=3.934s [fwd=2.691s bwd=1.229s opt=0.003s] dispatch=3.747s [2026-10-18 22:00:41,609] loop.py:45 INFO TRAIN_STEP step=3920 loss=1.485584 Training: 22%|██▏ | 3920/18018 [8:30:27<30:05:01, 7.68s/step, loss=1.486, acc=0.672, acc_len=1.60, thru=15.6, I=15.9, T=32.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:41,700] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:49,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 22:00:50,734] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3920: forward=3144.7ms backward=1313.3ms [2026-10-18 22:00:50,739] loop.py:45 INFO TRAIN_STEP step=3921 loss=1.611140 Training: 22%|██▏ | 3921/18018 [8:30:36<31:47:06, 8.12s/step, loss=1.611, acc=0.649, acc_len=1.52, thru=14.1, I=17.2, T=28.4, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:51,091] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:00:58,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:01:00,050] loop.py:45 INFO TRAIN_STEP step=3922 loss=1.636662 Training: 22%|██▏ | 3922/18018 [8:30:46<33:11:00, 8.47s/step, loss=1.637, acc=0.651, acc_len=1.53, thru=14.3, I=15.5, T=29.3, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:00,353] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:04,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:01:06,403] loop.py:45 INFO TRAIN_STEP step=3923 loss=1.682464 Training: 22%|██▏ | 3923/18018 [8:30:52<30:41:26, 7.84s/step, loss=1.682, acc=0.641, acc_len=1.50, thru=18.4, I=24.6, T=31.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:06,736] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:11,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:01:12,212] loop.py:45 INFO TRAIN_STEP step=3924 loss=1.531786 Training: 22%|██▏ | 3924/18018 [8:30:58<28:18:17, 7.23s/step, loss=1.532, acc=0.663, acc_len=1.56, thru=20.1, I=27.9, T=35.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:12,428] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:16,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:01:19,113] loop.py:380 INFO TIMING step=3925: step=4.118s data=1.929s compute=4.078s [fwd=2.895s bwd=1.167s opt=0.003s] dispatch=2.773s [2026-10-18 22:01:19,113] loop.py:45 INFO TRAIN_STEP step=3925 loss=1.534861 Training: 22%|██▏ | 3925/18018 [8:31:05<27:54:51, 7.13s/step, loss=1.535, acc=0.661, acc_len=1.56, thru=17.8, I=25.0, T=31.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:19,480] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:27,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:01:28,493] loop.py:45 INFO TRAIN_STEP step=3926 loss=1.530350 Training: 22%|██▏ | 3926/18018 [8:31:14<30:33:22, 7.81s/step, loss=1.530, acc=0.666, acc_len=1.58, thru=14.2, I=16.7, T=29.7, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:01:28,488] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3925: forward=3008.6ms backward=1253.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:28,680] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:34,733] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:01:36,027] loop.py:45 INFO TRAIN_STEP step=3927 loss=1.601889 Training: 22%|██▏ | 3927/18018 [8:31:22<30:14:02, 7.72s/step, loss=1.602, acc=0.653, acc_len=1.54, thru=14.9, I=20.5, T=30.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:36,245] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:41,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:01:42,615] loop.py:45 INFO TRAIN_STEP step=3928 loss=1.504882 Training: 22%|██▏ | 3928/18018 [8:31:28<28:53:44, 7.38s/step, loss=1.505, acc=0.667, acc_len=1.59, thru=17.9, I=21.5, T=29.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:43,000] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:49,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:01:50,563] loop.py:45 INFO TRAIN_STEP step=3929 loss=1.647641 Training: 22%|██▏ | 3929/18018 [8:31:36<29:33:35, 7.55s/step, loss=1.648, acc=0.640, acc_len=1.50, thru=15.4, I=18.8, T=29.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:50,694] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:56,888] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:01:58,224] loop.py:380 INFO TIMING step=3930: step=3.971s data=1.676s compute=3.946s [fwd=2.757s bwd=1.174s opt=0.003s] dispatch=3.677s [2026-10-18 22:01:58,224] loop.py:45 INFO TRAIN_STEP step=3930 loss=1.632298 Training: 22%|██▏ | 3930/18018 [8:31:44<29:41:04, 7.59s/step, loss=1.632, acc=0.648, acc_len=1.52, thru=15.1, I=19.4, T=32.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:01:58,397] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:03,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:02:05,122] loop.py:45 INFO TRAIN_STEP step=3931 loss=1.784411 Training: 22%|██▏ | 3931/18018 [8:31:51<28:52:21, 7.38s/step, loss=1.784, acc=0.622, acc_len=1.44, thru=17.4, I=22.4, T=32.0, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:02:05,117] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3930: forward=2812.0ms backward=1151.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:05,276] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:12,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:02:14,034] loop.py:45 INFO TRAIN_STEP step=3932 loss=1.644078 Training: 22%|██▏ | 3932/18018 [8:32:00<30:40:15, 7.84s/step, loss=1.644, acc=0.648, acc_len=1.52, thru=14.6, I=17.2, T=29.6, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:14,097] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:18,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:02:20,563] loop.py:45 INFO TRAIN_STEP step=3933 loss=1.710098 Training: 22%|██▏ | 3933/18018 [8:32:06<29:07:51, 7.45s/step, loss=1.710, acc=0.635, acc_len=1.49, thru=17.9, I=24.7, T=32.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:20,744] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:24,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:02:26,540] loop.py:45 INFO TRAIN_STEP step=3934 loss=1.652707 Training: 22%|██▏ | 3934/18018 [8:32:12<27:24:19, 7.01s/step, loss=1.653, acc=0.641, acc_len=1.49, thru=19.2, I=28.3, T=33.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:26,722] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:31,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:02:33,643] loop.py:380 INFO TIMING step=3935: step=4.598s data=2.230s compute=4.563s [fwd=3.168s bwd=1.380s opt=0.003s] dispatch=2.496s [2026-10-18 22:02:33,643] loop.py:45 INFO TRAIN_STEP step=3935 loss=1.481847 Training: 22%|██▏ | 3935/18018 [8:32:19<27:31:17, 7.04s/step, loss=1.482, acc=0.674, acc_len=1.61, thru=15.8, I=24.7, T=27.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:33,901] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:39,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:02:40,714] loop.py:45 INFO TRAIN_STEP step=3936 loss=1.637028 Training: 22%|██▏ | 3936/18018 [8:32:26<27:33:33, 7.05s/step, loss=1.637, acc=0.646, acc_len=1.52, thru=16.5, I=21.5, T=29.7, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:02:40,710] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3935: forward=2977.8ms backward=1282.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:40,931] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:48,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:02:48,486] loop.py:45 INFO TRAIN_STEP step=3937 loss=1.652357 Training: 22%|██▏ | 3937/18018 [8:32:34<28:24:38, 7.26s/step, loss=1.652, acc=0.644, acc_len=1.51, thru=16.1, I=20.1, T=26.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:48,541] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:57,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:02:57,067] loop.py:45 INFO TRAIN_STEP step=3938 loss=1.656882 Training: 22%|██▏ | 3938/18018 [8:32:43<29:57:05, 7.66s/step, loss=1.657, acc=0.644, acc_len=1.51, thru=14.9, I=13.6, T=30.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:02:57,523] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:04,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:03:05,417] loop.py:45 INFO TRAIN_STEP step=3939 loss=1.438555 Training: 22%|██▏ | 3939/18018 [8:32:51<30:45:48, 7.87s/step, loss=1.439, acc=0.680, acc_len=1.63, thru=15.8, I=16.6, T=33.5, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:05,747] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:03:13,129] loop.py:380 INFO TIMING step=3940: step=3.877s data=1.753s compute=3.839s [fwd=2.718s bwd=1.105s opt=0.003s] dispatch=3.826s [2026-10-18 22:03:13,129] loop.py:45 INFO TRAIN_STEP step=3940 loss=1.711507 Training: 22%|██▏ | 3940/18018 [8:32:59<30:34:51, 7.82s/step, loss=1.712, acc=0.635, acc_len=1.48, thru=15.1, I=16.9, T=33.0, wait=3.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:22,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:03:23,956] loop.py:45 INFO TRAIN_STEP step=3941 loss=1.659243 Training: 22%|██▏ | 3941/18018 [8:33:10<34:06:18, 8.72s/step, loss=1.659, acc=0.643, acc_len=1.51, thru=12.9, I=13.6, T=26.4, wait=6.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:03:23,951] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3940: forward=3341.6ms backward=1429.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:24,405] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:03:31,124] loop.py:45 INFO TRAIN_STEP step=3942 loss=1.515325 Training: 22%|██▏ | 3942/18018 [8:33:17<32:16:55, 8.26s/step, loss=1.515, acc=0.668, acc_len=1.58, thru=16.1, I=19.4, T=32.8, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:40,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:03:41,578] loop.py:45 INFO TRAIN_STEP step=3943 loss=1.659765 Training: 22%|██▏ | 3943/18018 [8:33:27<34:51:16, 8.91s/step, loss=1.660, acc=0.646, acc_len=1.51, thru=13.4, I=14.2, T=27.9, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:41,869] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:46,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:03:48,589] loop.py:45 INFO TRAIN_STEP step=3944 loss=1.671875 Training: 22%|██▏ | 3944/18018 [8:33:34<32:37:19, 8.34s/step, loss=1.672, acc=0.640, acc_len=1.50, thru=17.4, I=19.8, T=29.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:48,907] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:55,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:03:55,218] loop.py:380 INFO TIMING step=3945: step=3.434s data=1.381s compute=3.403s [fwd=2.426s bwd=0.963s opt=0.003s] dispatch=3.185s [2026-10-18 22:03:55,219] loop.py:45 INFO TRAIN_STEP step=3945 loss=1.630299 Training: 22%|██▏ | 3945/18018 [8:33:41<30:36:28, 7.83s/step, loss=1.630, acc=0.652, acc_len=1.53, thru=17.0, I=21.6, T=37.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:03:55,560] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:04:01,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:04:03,424] loop.py:45 INFO TRAIN_STEP step=3946 loss=1.623918 Training: 22%|██▏ | 3946/18018 [8:33:49<31:02:50, 7.94s/step, loss=1.624, acc=0.648, acc_len=1.52, thru=15.2, I=19.6, T=30.1, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:04:03,418] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3945: forward=2958.9ms backward=1242.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:04:03,569] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-18 22:04:10,370] loop.py:45 INFO TRAIN_STEP step=3947 loss=1.495203 Training: 22%|██▏ | 3947/18018 [8:33:56<29:52:24, 7.64s/step, loss=1.495, acc=0.668, acc_len=1.59, thru=14.8, I=19.4, T=31.2, wait=2.8s, pool=52, epoch=1/2][2026-10-18 22:04:19,271] loop.py:45 INFO TRAIN_STEP step=3948 loss=1.662224 Training: 22%|██▏ | 3948/18018 [8:34:05<31:20:53, 8.02s/step, loss=1.662, acc=0.646, acc_len=1.51, thru=12.6, I=11.8, T=26.5, wait=4.1s, pool=28, epoch=1/2][2026-10-18 22:04:27,803] loop.py:45 INFO TRAIN_STEP step=3949 loss=1.610367 Training: 22%|██▏ | 3949/18018 [8:34:13<31:56:44, 8.17s/step, loss=1.610, acc=0.651, acc_len=1.54, thru=17.3, I=17.7, T=28.0, wait=4.0s, pool=56, epoch=1/2][2026-10-18 22:04:36,872] loop.py:380 INFO TIMING step=3950: step=4.150s data=1.884s compute=4.112s [fwd=2.886s bwd=1.211s opt=0.003s] dispatch=4.908s [2026-10-18 22:04:36,872] loop.py:45 INFO TRAIN_STEP step=3950 loss=1.551821 Training: 22%|██▏ | 3950/18018 [8:34:22<32:59:34, 8.44s/step, loss=1.552, acc=0.660, acc_len=1.56, thru=13.6, I=12.5, T=30.8, wait=4.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:04:43,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:04:44,836] loop.py:45 INFO TRAIN_STEP step=3951 loss=1.548962 Training: 22%|██▏ | 3951/18018 [8:34:30<32:25:45, 8.30s/step, loss=1.549, acc=0.664, acc_len=1.57, thru=15.9, I=17.6, T=30.0, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:04:44,831] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3950: forward=2948.0ms backward=1263.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:04:44,886] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:04:50,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:04:51,041] loop.py:45 INFO TRAIN_STEP step=3952 loss=1.531248 Training: 22%|██▏ | 3952/18018 [8:34:37<29:58:18, 7.67s/step, loss=1.531, acc=0.664, acc_len=1.57, thru=19.5, I=21.8, T=37.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:04:51,228] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:04:57,749] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:04:59,110] loop.py:45 INFO TRAIN_STEP step=3953 loss=1.589347 Training: 22%|██▏ | 3953/18018 [8:34:45<30:26:23, 7.79s/step, loss=1.589, acc=0.659, acc_len=1.56, thru=17.2, I=19.5, T=30.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:04:59,276] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 22:05:07,642] loop.py:45 INFO TRAIN_STEP step=3954 loss=1.663826 Training: 22%|██▏ | 3954/18018 [8:34:53<31:18:01, 8.01s/step, loss=1.664, acc=0.645, acc_len=1.52, thru=11.3, I=14.7, T=32.0, wait=4.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:13,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:05:15,291] loop.py:380 INFO TIMING step=3955: step=4.020s data=1.706s compute=3.987s [fwd=2.848s bwd=1.124s opt=0.003s] dispatch=3.619s [2026-10-18 22:05:15,291] loop.py:45 INFO TRAIN_STEP step=3955 loss=1.618564 Training: 22%|██▏ | 3955/18018 [8:35:01<30:52:27, 7.90s/step, loss=1.619, acc=0.650, acc_len=1.53, thru=17.1, I=18.0, T=31.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:15,518] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:21,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:05:22,998] loop.py:45 INFO TRAIN_STEP step=3956 loss=1.629553 Training: 22%|██▏ | 3956/18018 [8:35:09<30:38:30, 7.84s/step, loss=1.630, acc=0.645, acc_len=1.51, thru=14.7, I=20.5, T=29.7, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:05:22,993] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3955: forward=2988.6ms backward=1255.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:23,409] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:27,917] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:05:29,824] loop.py:45 INFO TRAIN_STEP step=3957 loss=1.697950 Training: 22%|██▏ | 3957/18018 [8:35:15<29:26:44, 7.54s/step, loss=1.698, acc=0.635, acc_len=1.48, thru=17.6, I=24.1, T=31.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:29,939] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:35,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:05:36,952] loop.py:45 INFO TRAIN_STEP step=3958 loss=1.639361 Training: 22%|██▏ | 3958/18018 [8:35:23<28:57:37, 7.42s/step, loss=1.639, acc=0.646, acc_len=1.52, thru=15.8, I=23.6, T=28.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:37,165] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:41,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:05:42,736] loop.py:45 INFO TRAIN_STEP step=3959 loss=1.694528 Training: 22%|██▏ | 3959/18018 [8:35:28<27:03:00, 6.93s/step, loss=1.695, acc=0.642, acc_len=1.49, thru=19.7, I=25.3, T=33.3, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:42,977] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:05:49,455] loop.py:380 INFO TIMING step=3960: step=3.731s data=1.646s compute=3.689s [fwd=2.599s bwd=1.075s opt=0.003s] dispatch=2.976s [2026-10-18 22:05:49,455] loop.py:45 INFO TRAIN_STEP step=3960 loss=1.640158 Training: 22%|██▏ | 3960/18018 [8:35:35<26:48:05, 6.86s/step, loss=1.640, acc=0.642, acc_len=1.50, thru=17.1, I=19.3, T=34.3, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:55,897] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:05:57,205] loop.py:45 INFO TRAIN_STEP step=3961 loss=1.584829 Training: 22%|██▏ | 3961/18018 [8:35:43<27:50:28, 7.13s/step, loss=1.585, acc=0.658, acc_len=1.55, thru=17.3, I=18.4, T=34.0, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:05:57,199] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3960: forward=2646.4ms backward=1064.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:05:57,407] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:04,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:06:04,630] loop.py:45 INFO TRAIN_STEP step=3962 loss=1.509315 Training: 22%|██▏ | 3962/18018 [8:35:50<28:10:58, 7.22s/step, loss=1.509, acc=0.669, acc_len=1.59, thru=17.1, I=19.7, T=31.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:04,789] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:09,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:06:10,870] loop.py:45 INFO TRAIN_STEP step=3963 loss=1.602661 Training: 22%|██▏ | 3963/18018 [8:35:56<27:02:12, 6.93s/step, loss=1.603, acc=0.652, acc_len=1.54, thru=18.9, I=22.3, T=31.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:11,080] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:17,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:06:18,974] loop.py:45 INFO TRAIN_STEP step=3964 loss=1.656889 Training: 22%|██▏ | 3964/18018 [8:36:05<28:24:58, 7.28s/step, loss=1.657, acc=0.639, acc_len=1.50, thru=14.5, I=19.4, T=31.8, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:19,232] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:25,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:06:26,322] loop.py:380 INFO TIMING step=3965: step=4.099s data=1.712s compute=4.061s [fwd=2.854s bwd=1.191s opt=0.003s] dispatch=3.238s [2026-10-18 22:06:26,323] loop.py:45 INFO TRAIN_STEP step=3965 loss=1.732349 Training: 22%|██▏ | 3965/18018 [8:36:12<28:29:43, 7.30s/step, loss=1.732, acc=0.632, acc_len=1.48, thru=14.4, I=17.5, T=31.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:26,417] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:31,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:06:32,766] loop.py:45 INFO TRAIN_STEP step=3966 loss=1.649758 Training: 22%|██▏ | 3966/18018 [8:36:18<27:29:17, 7.04s/step, loss=1.650, acc=0.645, acc_len=1.51, thru=19.4, I=20.6, T=32.0, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:06:32,760] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3965: forward=2760.2ms backward=1198.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:33,111] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:37,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:06:39,079] loop.py:45 INFO TRAIN_STEP step=3967 loss=1.720410 Training: 22%|██▏ | 3967/18018 [8:36:25<26:37:57, 6.82s/step, loss=1.720, acc=0.632, acc_len=1.47, thru=18.6, I=24.4, T=33.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:39,268] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:06:46,866] loop.py:45 INFO TRAIN_STEP step=3968 loss=1.544367 Training: 22%|██▏ | 3968/18018 [8:36:32<27:45:30, 7.11s/step, loss=1.544, acc=0.665, acc_len=1.57, thru=15.3, I=18.0, T=28.8, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:53,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:06:54,136] loop.py:45 INFO TRAIN_STEP step=3969 loss=1.757645 Training: 22%|██▏ | 3969/18018 [8:36:40<27:56:35, 7.16s/step, loss=1.758, acc=0.626, acc_len=1.45, thru=18.3, I=17.5, T=31.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:06:54,642] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:07:01,177] loop.py:380 INFO TIMING step=3970: step=4.470s data=2.329s compute=4.435s [fwd=3.335s bwd=1.088s opt=0.003s] dispatch=2.559s [2026-10-18 22:07:01,177] loop.py:45 INFO TRAIN_STEP step=3970 loss=1.599350 Training: 22%|██▏ | 3970/18018 [8:36:47<27:47:53, 7.12s/step, loss=1.599, acc=0.657, acc_len=1.56, thru=16.5, I=17.6, T=28.6, wait=2.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:07,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:07:09,340] loop.py:45 INFO TRAIN_STEP step=3971 loss=1.704759 Training: 22%|██▏ | 3971/18018 [8:36:55<29:00:57, 7.44s/step, loss=1.705, acc=0.640, acc_len=1.50, thru=17.5, I=19.4, T=25.8, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:07:09,334] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3970: forward=3349.4ms backward=1558.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:09,571] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:07:16,964] loop.py:45 INFO TRAIN_STEP step=3972 loss=1.811191 Training: 22%|██▏ | 3972/18018 [8:37:03<29:13:55, 7.49s/step, loss=1.811, acc=0.619, acc_len=1.43, thru=12.9, I=17.5, T=31.8, wait=3.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:23,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:07:25,665] loop.py:45 INFO TRAIN_STEP step=3973 loss=1.491520 Training: 22%|██▏ | 3973/18018 [8:37:11<30:38:50, 7.86s/step, loss=1.492, acc=0.670, acc_len=1.59, thru=15.8, I=16.0, T=27.8, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:25,977] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:31,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:07:32,784] loop.py:45 INFO TRAIN_STEP step=3974 loss=1.634607 Training: 22%|██▏ | 3974/18018 [8:37:18<29:46:43, 7.63s/step, loss=1.635, acc=0.647, acc_len=1.53, thru=16.0, I=22.2, T=33.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:33,255] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:07:40,950] loop.py:380 INFO TIMING step=3975: step=4.230s data=1.848s compute=4.194s [fwd=2.907s bwd=1.273s opt=0.003s] dispatch=3.928s [2026-10-18 22:07:40,951] loop.py:45 INFO TRAIN_STEP step=3975 loss=1.598788 Training: 22%|██▏ | 3975/18018 [8:37:27<30:24:04, 7.79s/step, loss=1.599, acc=0.654, acc_len=1.54, thru=13.6, I=17.3, T=30.3, wait=3.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:49,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:07:50,324] loop.py:45 INFO TRAIN_STEP step=3976 loss=1.731675 (TrainerActor pid=9310) [2026-10-18 22:07:50,319] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3975: forward=3595.1ms backward=1222.8ms Training: 22%|██▏ | 3976/18018 [8:37:36<32:15:00, 8.27s/step, loss=1.732, acc=0.635, acc_len=1.48, thru=14.6, I=15.3, T=26.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:50,460] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:54,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:07:56,421] loop.py:45 INFO TRAIN_STEP step=3977 loss=1.656041 Training: 22%|██▏ | 3977/18018 [8:37:42<29:42:22, 7.62s/step, loss=1.656, acc=0.643, acc_len=1.50, thru=18.6, I=22.9, T=32.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:07:56,884] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:02,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:08:03,350] loop.py:45 INFO TRAIN_STEP step=3978 loss=1.629448 Training: 22%|██▏ | 3978/18018 [8:37:49<28:53:59, 7.41s/step, loss=1.629, acc=0.646, acc_len=1.52, thru=15.3, I=22.6, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:03,403] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:08,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:08:09,766] loop.py:45 INFO TRAIN_STEP step=3979 loss=1.674182 Training: 22%|██▏ | 3979/18018 [8:37:55<27:44:10, 7.11s/step, loss=1.674, acc=0.644, acc_len=1.51, thru=18.8, I=22.4, T=31.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:10,018] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:15,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:08:17,566] loop.py:380 INFO TIMING step=3980: step=4.216s data=1.893s compute=4.190s [fwd=2.930s bwd=1.247s opt=0.003s] dispatch=3.572s [2026-10-18 22:08:17,566] loop.py:45 INFO TRAIN_STEP step=3980 loss=1.656356 Training: 22%|██▏ | 3980/18018 [8:38:03<28:32:19, 7.32s/step, loss=1.656, acc=0.642, acc_len=1.50, thru=14.8, I=18.3, T=30.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:17,882] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:23,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:08:24,049] loop.py:45 INFO TRAIN_STEP step=3981 loss=1.637210 Training: 22%|██▏ | 3981/18018 [8:38:10<27:33:21, 7.07s/step, loss=1.637, acc=0.648, acc_len=1.52, thru=17.3, I=22.1, T=33.2, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:08:24,044] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3980: forward=2696.7ms backward=1115.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:24,466] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:30,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:08:31,544] loop.py:45 INFO TRAIN_STEP step=3982 loss=1.654677 Training: 22%|██▏ | 3982/18018 [8:38:17<28:03:20, 7.20s/step, loss=1.655, acc=0.646, acc_len=1.51, thru=17.6, I=18.9, T=30.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:31,929] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:36,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:08:37,945] loop.py:45 INFO TRAIN_STEP step=3983 loss=1.684708 Training: 22%|██▏ | 3983/18018 [8:38:24<27:07:27, 6.96s/step, loss=1.685, acc=0.638, acc_len=1.48, thru=18.6, I=22.6, T=31.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:38,275] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:44,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:08:46,000] loop.py:45 INFO TRAIN_STEP step=3984 loss=1.551334 Training: 22%|██▏ | 3984/18018 [8:38:32<28:24:26, 7.29s/step, loss=1.551, acc=0.663, acc_len=1.57, thru=14.1, I=19.5, T=29.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:46,346] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:51,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:08:53,724] loop.py:380 INFO TIMING step=3985: step=4.788s data=2.307s compute=4.755s [fwd=3.414s bwd=1.326s opt=0.003s] dispatch=2.925s [2026-10-18 22:08:53,724] loop.py:45 INFO TRAIN_STEP step=3985 loss=1.652848 Training: 22%|██▏ | 3985/18018 [8:38:39<28:54:59, 7.42s/step, loss=1.653, acc=0.643, acc_len=1.51, thru=15.4, I=20.3, T=26.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:53,778] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:08:59,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:09:00,732] loop.py:45 INFO TRAIN_STEP step=3986 loss=1.720875 Training: 22%|██▏ | 3986/18018 [8:38:46<28:26:07, 7.30s/step, loss=1.721, acc=0.635, acc_len=1.48, thru=15.6, I=24.0, T=26.8, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:09:00,727] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3985: forward=3495.3ms backward=1235.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:01,135] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:08,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:09:08,976] loop.py:45 INFO TRAIN_STEP step=3987 loss=1.623297 Training: 22%|██▏ | 3987/18018 [8:38:55<29:32:23, 7.58s/step, loss=1.623, acc=0.654, acc_len=1.55, thru=13.6, I=16.5, T=27.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:09,237] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:09:17,092] loop.py:45 INFO TRAIN_STEP step=3988 loss=1.655337 Training: 22%|██▏ | 3988/18018 [8:39:03<30:09:53, 7.74s/step, loss=1.655, acc=0.641, acc_len=1.50, thru=14.8, I=14.4, T=30.6, wait=3.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:25,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:09:26,431] loop.py:45 INFO TRAIN_STEP step=3989 loss=1.669945 Training: 22%|██▏ | 3989/18018 [8:39:12<32:01:59, 8.22s/step, loss=1.670, acc=0.639, acc_len=1.49, thru=15.9, I=14.9, T=26.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:26,909] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:32,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:09:33,451] loop.py:380 INFO TIMING step=3990: step=4.081s data=1.918s compute=4.033s [fwd=2.841s bwd=1.175s opt=0.003s] dispatch=2.931s [2026-10-18 22:09:33,452] loop.py:45 INFO TRAIN_STEP step=3990 loss=1.708537 Training: 22%|██▏ | 3990/18018 [8:39:19<30:37:49, 7.86s/step, loss=1.709, acc=0.636, acc_len=1.48, thru=16.1, I=19.1, T=31.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:33,732] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:09:41,863] loop.py:45 INFO TRAIN_STEP step=3991 loss=1.660946 Training: 22%|██▏ | 3991/18018 [8:39:27<31:16:16, 8.03s/step, loss=1.661, acc=0.647, acc_len=1.51, thru=11.2, I=14.3, T=31.4, wait=4.3s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:09:41,858] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3990: forward=2891.8ms backward=1137.4ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 22:09:41,927] eagle_store.py:517 WARNING batch_get_buffer missing keys (k127711-966721d0_hs, k127711-966721d0_ids, k127711-966721d0_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:49,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 22:09:41,910] eagle_store.py:517 WARNING batch_get_buffer missing keys (k127710-9017ed8f_hs, k127710-9017ed8f_ids, k127710-9017ed8f_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-18 22:09:51,653] loop.py:45 INFO TRAIN_STEP step=3992 loss=1.686450 Training: 22%|██▏ | 3992/18018 [8:39:37<33:19:56, 8.56s/step, loss=1.686, acc=0.642, acc_len=1.50, thru=16.5, I=17.3, T=30.9, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:51,991] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:57,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:09:59,453] loop.py:45 INFO TRAIN_STEP step=3993 loss=1.608671 Training: 22%|██▏ | 3993/18018 [8:39:45<32:26:41, 8.33s/step, loss=1.609, acc=0.648, acc_len=1.53, thru=14.6, I=21.2, T=25.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:09:59,955] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:05,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:10:06,676] loop.py:45 INFO TRAIN_STEP step=3994 loss=1.629883 Training: 22%|██▏ | 3994/18018 [8:39:52<31:08:59, 8.00s/step, loss=1.630, acc=0.647, acc_len=1.52, thru=15.6, I=20.9, T=29.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:07,088] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:10:15,500] loop.py:380 INFO TIMING step=3995: step=4.512s data=2.173s compute=4.470s [fwd=3.144s bwd=1.311s opt=0.003s] dispatch=4.305s [2026-10-18 22:10:15,501] loop.py:45 INFO TRAIN_STEP step=3995 loss=1.591437 Training: 22%|██▏ | 3995/18018 [8:40:01<32:06:58, 8.24s/step, loss=1.591, acc=0.659, acc_len=1.56, thru=12.8, I=14.9, T=28.4, wait=4.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:21,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:10:23,324] loop.py:45 INFO TRAIN_STEP step=3996 loss=1.594748 Training: 22%|██▏ | 3996/18018 [8:40:09<31:37:17, 8.12s/step, loss=1.595, acc=0.657, acc_len=1.55, thru=16.9, I=18.4, T=27.3, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:10:23,319] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=3995: forward=3155.7ms backward=1472.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:23,707] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:30,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:10:31,856] loop.py:45 INFO TRAIN_STEP step=3997 loss=1.625929 Training: 22%|██▏ | 3997/18018 [8:40:17<32:06:14, 8.24s/step, loss=1.626, acc=0.652, acc_len=1.53, thru=15.7, I=18.5, T=29.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:32,294] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:10:40,395] loop.py:45 INFO TRAIN_STEP step=3998 loss=1.545642 Training: 22%|██▏ | 3998/18018 [8:40:26<32:26:50, 8.33s/step, loss=1.546, acc=0.661, acc_len=1.56, thru=15.3, I=14.9, T=30.4, wait=4.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:46,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:10:47,858] loop.py:45 INFO TRAIN_STEP step=3999 loss=1.570323 Training: 22%|██▏ | 3999/18018 [8:40:33<31:25:54, 8.07s/step, loss=1.570, acc=0.660, acc_len=1.56, thru=17.4, I=18.9, T=29.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:47,944] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:51,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:10:53,495] loop.py:380 INFO TIMING step=4000: step=3.763s data=1.646s compute=3.725s [fwd=2.649s bwd=1.062s opt=0.003s] dispatch=1.862s [2026-10-18 22:10:53,495] loop.py:45 INFO TRAIN_STEP step=4000 loss=1.731447 Training: 22%|██▏ | 4000/18018 [8:40:39<28:35:05, 7.34s/step, loss=1.731, acc=0.633, acc_len=1.47, thru=19.7, I=25.6, T=34.0, wait=1.9s, pool=64, epoch=1/2][2026-10-18 22:10:53,500] loop.py:419 INFO Saving checkpoint at step 4000... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 22:10:56,898] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0004001 (TrainerActor pid=9310) [2026-10-18 22:10:56,905] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0003251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:57,028] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:10:58,033] inference_manager.py:408 INFO Pool capacity available after 6.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:03,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 22:10:56,896] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0004001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 22:10:57,979] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0003251 [2026-10-18 22:11:04,159] loop.py:45 INFO TRAIN_STEP step=4001 loss=1.648268 Training: 22%|██▏ | 4001/18018 [8:40:50<32:27:42, 8.34s/step, loss=1.648, acc=0.642, acc_len=1.50, thru=20.9, I=24.4, T=34.1, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:11:04,155] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4000: forward=2634.5ms backward=1067.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:04,662] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:11,971] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:11:14,069] loop.py:45 INFO TRAIN_STEP step=4002 loss=1.588800 Training: 22%|██▏ | 4002/18018 [8:41:00<34:17:55, 8.81s/step, loss=1.589, acc=0.652, acc_len=1.53, thru=13.6, I=16.7, T=31.2, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:14,491] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 22:11:20,520] loop.py:45 INFO TRAIN_STEP step=4003 loss=1.676701 Training: 22%|██▏ | 4003/18018 [8:41:06<31:32:18, 8.10s/step, loss=1.677, acc=0.642, acc_len=1.49, thru=17.3, I=22.7, T=32.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:20,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:21,015] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:28,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:11:29,257] loop.py:45 INFO TRAIN_STEP step=4004 loss=1.648449 Training: 22%|██▏ | 4004/18018 [8:41:15<32:16:48, 8.29s/step, loss=1.648, acc=0.649, acc_len=1.52, thru=15.3, I=17.8, T=30.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:29,749] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:11:36,601] loop.py:380 INFO TIMING step=4005: step=3.922s data=1.686s compute=3.876s [fwd=2.711s bwd=1.149s opt=0.003s] dispatch=3.413s [2026-10-18 22:11:36,602] loop.py:45 INFO TRAIN_STEP step=4005 loss=1.534873 Training: 22%|██▏ | 4005/18018 [8:41:22<31:10:16, 8.01s/step, loss=1.535, acc=0.661, acc_len=1.56, thru=16.8, I=20.8, T=32.6, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:42,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:11:43,226] loop.py:45 INFO TRAIN_STEP step=4006 loss=1.771523 Training: 22%|██▏ | 4006/18018 [8:41:29<29:33:20, 7.59s/step, loss=1.772, acc=0.623, acc_len=1.45, thru=18.9, I=20.0, T=33.5, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:11:43,222] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4005: forward=2656.1ms backward=1098.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:43,354] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-18 22:11:49,141] loop.py:45 INFO TRAIN_STEP step=4007 loss=1.498673 Training: 22%|██▏ | 4007/18018 [8:41:35<27:35:24, 7.09s/step, loss=1.499, acc=0.669, acc_len=1.59, thru=18.5, I=21.4, T=30.6, wait=1.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:57,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:11:57,386] loop.py:45 INFO TRAIN_STEP step=4008 loss=1.670791 Training: 22%|██▏ | 4008/18018 [8:41:43<28:56:19, 7.44s/step, loss=1.671, acc=0.644, acc_len=1.51, thru=17.3, I=15.0, T=32.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:11:57,758] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:12:04,631] loop.py:45 INFO TRAIN_STEP step=4009 loss=1.658920 Training: 22%|██▏ | 4009/18018 [8:41:50<28:42:50, 7.38s/step, loss=1.659, acc=0.647, acc_len=1.52, thru=16.0, I=19.7, T=32.8, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:13,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:12:15,488] loop.py:380 INFO TIMING step=4010: step=4.471s data=2.155s compute=4.392s [fwd=3.091s bwd=1.286s opt=0.003s] dispatch=6.377s [2026-10-18 22:12:15,488] loop.py:45 INFO TRAIN_STEP step=4010 loss=1.692246 Training: 22%|██▏ | 4010/18018 [8:42:01<32:46:22, 8.42s/step, loss=1.692, acc=0.637, acc_len=1.49, thru=13.7, I=13.1, T=28.6, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:15,582] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:20,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:12:21,399] loop.py:45 INFO TRAIN_STEP step=4011 loss=1.648414 Training: 22%|██▏ | 4011/18018 [8:42:07<29:50:22, 7.67s/step, loss=1.648, acc=0.643, acc_len=1.50, thru=18.4, I=23.7, T=36.3, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:12:21,393] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4010: forward=2498.1ms backward=973.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:21,491] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 [2026-10-18 22:12:29,109] loop.py:45 INFO TRAIN_STEP step=4012 loss=1.635556 Training: 22%|██▏ | 4012/18018 [8:42:15<29:53:07, 7.68s/step, loss=1.636, acc=0.654, acc_len=1.54, thru=16.5, I=19.3, T=29.0, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:34,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:12:36,248] loop.py:45 INFO TRAIN_STEP step=4013 loss=1.617082 Training: 22%|██▏ | 4013/18018 [8:42:22<29:15:06, 7.52s/step, loss=1.617, acc=0.653, acc_len=1.54, thru=16.6, I=19.5, T=33.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:36,454] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:42,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:12:43,053] loop.py:45 INFO TRAIN_STEP step=4014 loss=1.559920 Training: 22%|██▏ | 4014/18018 [8:42:29<28:24:46, 7.30s/step, loss=1.560, acc=0.659, acc_len=1.55, thru=17.8, I=25.8, T=31.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:43,425] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:49,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:12:50,169] loop.py:380 INFO TIMING step=4015: step=4.752s data=2.148s compute=4.715s [fwd=3.293s bwd=1.407s opt=0.003s] dispatch=2.354s [2026-10-18 22:12:50,169] loop.py:45 INFO TRAIN_STEP step=4015 loss=1.567566 Training: 22%|██▏ | 4015/18018 [8:42:36<28:11:34, 7.25s/step, loss=1.568, acc=0.655, acc_len=1.55, thru=16.6, I=18.3, T=26.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:12:50,563] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:12:56,678] loop.py:45 INFO TRAIN_STEP step=4016 loss=1.585193 (TrainerActor pid=9310) [2026-10-18 22:12:56,672] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4015: forward=2608.2ms backward=1046.0ms Training: 22%|██▏ | 4016/18018 [8:42:42<27:19:33, 7.03s/step, loss=1.585, acc=0.658, acc_len=1.55, thru=14.9, I=18.1, T=34.6, wait=2.8s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:04,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:13:05,491] loop.py:45 INFO TRAIN_STEP step=4017 loss=1.793085 Training: 22%|██▏ | 4017/18018 [8:42:51<29:24:40, 7.56s/step, loss=1.793, acc=0.620, acc_len=1.43, thru=17.3, I=16.0, T=35.2, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:05,869] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:11,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:13:13,521] loop.py:45 INFO TRAIN_STEP step=4018 loss=1.603033 Training: 22%|██▏ | 4018/18018 [8:42:59<29:57:23, 7.70s/step, loss=1.603, acc=0.649, acc_len=1.52, thru=15.3, I=19.7, T=29.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:13,941] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:20,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:13:21,232] loop.py:45 INFO TRAIN_STEP step=4019 loss=1.602552 Training: 22%|██▏ | 4019/18018 [8:43:07<29:57:45, 7.71s/step, loss=1.603, acc=0.655, acc_len=1.54, thru=14.1, I=20.6, T=30.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:21,558] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:27,258] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:13:28,007] loop.py:380 INFO TIMING step=4020: step=3.723s data=1.587s compute=3.692s [fwd=2.604s bwd=1.074s opt=0.003s] dispatch=3.041s [2026-10-18 22:13:28,007] loop.py:45 INFO TRAIN_STEP step=4020 loss=1.567675 Training: 22%|██▏ | 4020/18018 [8:43:14<28:52:26, 7.43s/step, loss=1.568, acc=0.653, acc_len=1.54, thru=18.1, I=21.2, T=34.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:28,262] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:33,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:13:35,098] loop.py:45 INFO TRAIN_STEP step=4021 loss=1.584276 Training: 22%|██▏ | 4021/18018 [8:43:21<28:28:51, 7.33s/step, loss=1.584, acc=0.657, acc_len=1.55, thru=18.9, I=21.8, T=34.0, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:13:35,092] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4020: forward=2624.4ms backward=1090.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:35,363] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:42,972] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:13:43,456] loop.py:45 INFO TRAIN_STEP step=4022 loss=1.505213 Training: 22%|██▏ | 4022/18018 [8:43:29<29:41:08, 7.64s/step, loss=1.505, acc=0.669, acc_len=1.59, thru=14.0, I=15.9, T=31.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:43,493] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:49,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:13:50,317] loop.py:45 INFO TRAIN_STEP step=4023 loss=1.522634 Training: 22%|██▏ | 4023/18018 [8:43:36<28:46:42, 7.40s/step, loss=1.523, acc=0.671, acc_len=1.59, thru=17.8, I=20.3, T=30.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:50,734] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:56,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:13:57,838] loop.py:45 INFO TRAIN_STEP step=4024 loss=1.689309 Training: 22%|██▏ | 4024/18018 [8:43:43<28:54:57, 7.44s/step, loss=1.689, acc=0.642, acc_len=1.50, thru=16.5, I=21.5, T=29.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:13:57,990] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-18 22:14:04,646] loop.py:380 INFO TIMING step=4025: step=4.099s data=1.850s compute=4.061s [fwd=2.795s bwd=1.252s opt=0.003s] dispatch=2.698s [2026-10-18 22:14:04,646] loop.py:45 INFO TRAIN_STEP step=4025 loss=1.627110 Training: 22%|██▏ | 4025/18018 [8:43:50<28:10:34, 7.25s/step, loss=1.627, acc=0.651, acc_len=1.53, thru=16.1, I=17.4, T=31.2, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:11,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:14:12,281] loop.py:45 INFO TRAIN_STEP step=4026 loss=1.694603 Training: 22%|██▏ | 4026/18018 [8:43:58<28:37:21, 7.36s/step, loss=1.695, acc=0.635, acc_len=1.48, thru=16.5, I=17.0, T=32.2, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:14:12,277] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4025: forward=2784.6ms backward=1127.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:12,627] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:18,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:14:21,180] loop.py:45 INFO TRAIN_STEP step=4027 loss=1.632372 Training: 22%|██▏ | 4027/18018 [8:44:07<30:24:43, 7.83s/step, loss=1.632, acc=0.646, acc_len=1.51, thru=14.7, I=20.0, T=29.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:21,408] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:27,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:14:28,189] loop.py:45 INFO TRAIN_STEP step=4028 loss=1.607196 Training: 22%|██▏ | 4028/18018 [8:44:14<29:27:30, 7.58s/step, loss=1.607, acc=0.651, acc_len=1.53, thru=16.1, I=21.6, T=32.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:28,463] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:33,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:14:35,196] loop.py:45 INFO TRAIN_STEP step=4029 loss=1.677922 Training: 22%|██▏ | 4029/18018 [8:44:21<28:47:15, 7.41s/step, loss=1.678, acc=0.641, acc_len=1.50, thru=18.2, I=20.0, T=30.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:35,576] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:40,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:14:41,264] loop.py:380 INFO TIMING step=4030: step=3.735s data=1.743s compute=3.697s [fwd=2.597s bwd=1.084s opt=0.003s] dispatch=2.325s [2026-10-18 22:14:41,264] loop.py:45 INFO TRAIN_STEP step=4030 loss=1.666002 Training: 22%|██▏ | 4030/18018 [8:44:27<27:13:24, 7.01s/step, loss=1.666, acc=0.637, acc_len=1.49, thru=18.9, I=25.1, T=34.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:41,561] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:46,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:14:48,224] loop.py:45 INFO TRAIN_STEP step=4031 loss=1.679310 Training: 22%|██▏ | 4031/18018 [8:44:34<27:10:10, 6.99s/step, loss=1.679, acc=0.641, acc_len=1.49, thru=18.0, I=23.1, T=28.9, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:14:48,219] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4030: forward=3163.3ms backward=1219.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:14:48,256] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 22:14:54,936] loop.py:45 INFO TRAIN_STEP step=4032 loss=1.534039 Training: 22%|██▏ | 4032/18018 [8:44:41<26:50:21, 6.91s/step, loss=1.534, acc=0.663, acc_len=1.57, thru=15.2, I=19.6, T=34.1, wait=2.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:04,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:15:05,448] loop.py:45 INFO TRAIN_STEP step=4033 loss=1.637943 Training: 22%|██▏ | 4033/18018 [8:44:51<31:02:14, 7.99s/step, loss=1.638, acc=0.643, acc_len=1.51, thru=14.4, I=14.4, T=29.7, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:05,551] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:14,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:15:15,199] loop.py:45 INFO TRAIN_STEP step=4034 loss=1.671402 Training: 22%|██▏ | 4034/18018 [8:45:01<33:05:08, 8.52s/step, loss=1.671, acc=0.640, acc_len=1.50, thru=13.3, I=13.9, T=27.3, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:15,416] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:20,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:15:22,861] loop.py:380 INFO TIMING step=4035: step=4.767s data=2.102s compute=4.738s [fwd=3.331s bwd=1.392s opt=0.003s] dispatch=2.886s [2026-10-18 22:15:22,861] loop.py:45 INFO TRAIN_STEP step=4035 loss=1.685081 Training: 22%|██▏ | 4035/18018 [8:45:08<32:05:06, 8.26s/step, loss=1.685, acc=0.643, acc_len=1.51, thru=16.5, I=19.4, T=26.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:23,355] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 22:15:29,114] loop.py:45 INFO TRAIN_STEP step=4036 loss=1.696545 Training: 22%|██▏ | 4036/18018 [8:45:15<29:44:50, 7.66s/step, loss=1.697, acc=0.636, acc_len=1.48, thru=16.0, I=22.0, T=36.7, wait=2.8s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:15:29,109] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4035: forward=2483.8ms backward=954.4ms [2026-10-18 22:15:36,601] loop.py:45 INFO TRAIN_STEP step=4037 loss=1.609466 Training: 22%|██▏ | 4037/18018 [8:45:22<29:32:37, 7.61s/step, loss=1.609, acc=0.659, acc_len=1.56, thru=15.0, I=15.6, T=31.3, wait=3.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:43,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:15:44,764] loop.py:45 INFO TRAIN_STEP step=4038 loss=1.713616 Training: 22%|██▏ | 4038/18018 [8:45:30<30:11:17, 7.77s/step, loss=1.714, acc=0.636, acc_len=1.48, thru=16.9, I=17.5, T=27.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:44,793] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:51,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:15:52,775] loop.py:45 INFO TRAIN_STEP step=4039 loss=1.650225 Training: 22%|██▏ | 4039/18018 [8:45:38<30:27:51, 7.85s/step, loss=1.650, acc=0.647, acc_len=1.52, thru=13.6, I=18.8, T=29.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:15:53,153] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:15:58,736] loop.py:380 INFO TIMING step=4040: step=3.551s data=1.527s compute=3.523s [fwd=2.551s bwd=0.956s opt=0.003s] dispatch=2.399s [2026-10-18 22:15:58,737] loop.py:45 INFO TRAIN_STEP step=4040 loss=1.612117 Training: 22%|██▏ | 4040/18018 [8:45:44<28:15:54, 7.28s/step, loss=1.612, acc=0.650, acc_len=1.52, thru=17.8, I=22.0, T=36.0, wait=2.4s, pool=52, epoch=1/2][2026-10-18 22:16:06,528] loop.py:45 INFO TRAIN_STEP step=4041 loss=1.642430 Training: 22%|██▏ | 4041/18018 [8:45:52<28:51:35, 7.43s/step, loss=1.642, acc=0.642, acc_len=1.50, thru=14.5, I=16.9, T=32.2, wait=3.8s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:16:06,523] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4040: forward=2840.5ms backward=1096.6ms [2026-10-18 22:16:18,676] loop.py:45 INFO TRAIN_STEP step=4042 loss=1.626784 Training: 22%|██▏ | 4042/18018 [8:46:04<34:21:10, 8.85s/step, loss=1.627, acc=0.650, acc_len=1.53, thru=14.1, I=13.2, T=28.8, wait=7.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:16:26,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:16:27,886] loop.py:45 INFO TRAIN_STEP step=4043 loss=1.685734 Training: 22%|██▏ | 4043/18018 [8:46:13<34:46:07, 8.96s/step, loss=1.686, acc=0.638, acc_len=1.49, thru=15.0, I=15.0, T=27.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:16:27,919] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:16:34,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:16:34,692] loop.py:45 INFO TRAIN_STEP step=4044 loss=1.635269 Training: 22%|██▏ | 4044/18018 [8:46:20<32:15:45, 8.31s/step, loss=1.635, acc=0.648, acc_len=1.52, thru=16.7, I=21.5, T=28.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:16:34,729] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:16:42,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:16:42,903] loop.py:380 INFO TIMING step=4045: step=4.238s data=1.959s compute=4.209s [fwd=2.996s bwd=1.199s opt=0.003s] dispatch=3.962s [2026-10-18 22:16:42,903] loop.py:45 INFO TRAIN_STEP step=4045 loss=1.743171 Training: 22%|██▏ | 4045/18018 [8:46:28<32:08:33, 8.28s/step, loss=1.743, acc=0.632, acc_len=1.47, thru=14.6, I=14.7, T=30.2, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:16:43,284] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:16:49,549] loop.py:45 INFO TRAIN_STEP step=4046 loss=1.651538 Training: 22%|██▏ | 4046/18018 [8:46:35<30:14:08, 7.79s/step, loss=1.652, acc=0.642, acc_len=1.50, thru=15.8, I=18.6, T=36.0, wait=3.1s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:16:49,544] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4045: forward=2523.9ms backward=977.6ms [2026-10-18 22:16:57,909] loop.py:45 INFO TRAIN_STEP step=4047 loss=1.788149 Training: 22%|██▏ | 4047/18018 [8:46:43<30:53:38, 7.96s/step, loss=1.788, acc=0.622, acc_len=1.43, thru=15.5, I=13.8, T=33.6, wait=4.5s, pool=44, epoch=1/2][2026-10-18 22:17:06,272] loop.py:45 INFO TRAIN_STEP step=4048 loss=1.583082 Training: 22%|██▏ | 4048/18018 [8:46:52<31:21:42, 8.08s/step, loss=1.583, acc=0.658, acc_len=1.56, thru=17.7, I=15.1, T=25.2, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:12,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:17:14,126] loop.py:45 INFO TRAIN_STEP step=4049 loss=1.665465 Training: 22%|██▏ | 4049/18018 [8:47:00<31:05:37, 8.01s/step, loss=1.665, acc=0.643, acc_len=1.51, thru=17.3, I=17.2, T=29.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:14,239] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-18 22:17:20,262] loop.py:380 INFO TIMING step=4050: step=3.881s data=1.750s compute=3.838s [fwd=2.704s bwd=1.118s opt=0.003s] dispatch=2.246s [2026-10-18 22:17:20,262] loop.py:45 INFO TRAIN_STEP step=4050 loss=1.738255 Training: 22%|██▏ | 4050/18018 [8:47:06<28:54:21, 7.45s/step, loss=1.738, acc=0.629, acc_len=1.47, thru=16.8, I=21.5, T=33.0, wait=2.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:28,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:17:28,547] loop.py:45 INFO TRAIN_STEP step=4051 loss=1.609733 Training: 22%|██▏ | 4051/18018 [8:47:14<29:52:34, 7.70s/step, loss=1.610, acc=0.653, acc_len=1.53, thru=16.9, I=16.9, T=30.0, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:17:28,542] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4050: forward=2990.3ms backward=1232.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:28,895] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:35,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:17:37,072] loop.py:45 INFO TRAIN_STEP step=4052 loss=1.639753 Training: 22%|██▏ | 4052/18018 [8:47:23<30:49:54, 7.95s/step, loss=1.640, acc=0.649, acc_len=1.53, thru=15.2, I=16.9, T=29.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:37,530] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:42,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:17:43,652] loop.py:45 INFO TRAIN_STEP step=4053 loss=1.658660 Training: 22%|██▏ | 4053/18018 [8:47:29<29:14:25, 7.54s/step, loss=1.659, acc=0.645, acc_len=1.51, thru=17.6, I=24.3, T=34.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:44,114] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:49,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:17:51,602] loop.py:45 INFO TRAIN_STEP step=4054 loss=1.698401 Training: 22%|██▏ | 4054/18018 [8:47:37<29:43:01, 7.66s/step, loss=1.698, acc=0.637, acc_len=1.48, thru=16.1, I=25.5, T=25.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:52,021] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:17:59,483] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:18:01,509] loop.py:380 INFO TIMING step=4055: step=4.547s data=2.128s compute=4.518s [fwd=3.231s bwd=1.271s opt=0.003s] dispatch=5.352s [2026-10-18 22:18:01,509] loop.py:45 INFO TRAIN_STEP step=4055 loss=1.686244 Training: 23%|██▎ | 4055/18018 [8:47:47<32:19:42, 8.34s/step, loss=1.686, acc=0.635, acc_len=1.48, thru=13.5, I=16.9, T=28.2, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:02,006] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:08,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:18:09,677] loop.py:45 INFO TRAIN_STEP step=4056 loss=1.487944 Training: 23%|██▎ | 4056/18018 [8:47:55<32:08:00, 8.29s/step, loss=1.488, acc=0.670, acc_len=1.59, thru=16.7, I=18.7, T=28.5, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:18:09,672] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4055: forward=3138.5ms backward=1318.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:09,821] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:17,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:18:17,823] loop.py:45 INFO TRAIN_STEP step=4057 loss=1.637806 Training: 23%|██▎ | 4057/18018 [8:48:03<31:58:08, 8.24s/step, loss=1.638, acc=0.649, acc_len=1.53, thru=16.5, I=18.8, T=32.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:17,998] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:23,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:18:25,972] loop.py:45 INFO TRAIN_STEP step=4058 loss=1.641017 Training: 23%|██▎ | 4058/18018 [8:48:12<31:51:27, 8.22s/step, loss=1.641, acc=0.645, acc_len=1.51, thru=15.6, I=19.9, T=28.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:26,245] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:33,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:18:33,272] loop.py:45 INFO TRAIN_STEP step=4059 loss=1.504540 Training: 23%|██▎ | 4059/18018 [8:48:19<30:47:15, 7.94s/step, loss=1.505, acc=0.665, acc_len=1.58, thru=14.6, I=20.4, T=32.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:33,607] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:39,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:18:40,501] loop.py:380 INFO TIMING step=4060: step=4.239s data=1.880s compute=4.204s [fwd=3.053s bwd=1.135s opt=0.003s] dispatch=2.981s [2026-10-18 22:18:40,501] loop.py:45 INFO TRAIN_STEP step=4060 loss=1.608006 Training: 23%|██▎ | 4060/18018 [8:48:26<29:57:30, 7.73s/step, loss=1.608, acc=0.650, acc_len=1.53, thru=17.3, I=19.4, T=30.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:40,588] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 22:18:49,326] loop.py:45 INFO TRAIN_STEP step=4061 loss=1.611556 Training: 23%|██▎ | 4061/18018 [8:48:35<31:14:02, 8.06s/step, loss=1.612, acc=0.649, acc_len=1.53, thru=12.8, I=15.0, T=25.2, wait=3.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:18:49,321] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4060: forward=3548.3ms backward=1478.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:59,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:18:59,400] loop.py:45 INFO TRAIN_STEP step=4062 loss=1.668300 Training: 23%|██▎ | 4062/18018 [8:48:45<33:34:36, 8.66s/step, loss=1.668, acc=0.645, acc_len=1.51, thru=14.6, I=13.5, T=26.4, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:18:59,661] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:08,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:19:08,835] loop.py:45 INFO TRAIN_STEP step=4063 loss=1.579374 Training: 23%|██▎ | 4063/18018 [8:48:54<34:28:33, 8.89s/step, loss=1.579, acc=0.661, acc_len=1.56, thru=13.6, I=13.8, T=23.8, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:09,095] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:16,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:19:17,481] loop.py:45 INFO TRAIN_STEP step=4064 loss=1.715475 Training: 23%|██▎ | 4064/18018 [8:49:03<34:11:07, 8.82s/step, loss=1.715, acc=0.638, acc_len=1.49, thru=14.1, I=16.1, T=30.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:17,928] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:23,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:19:24,515] loop.py:380 INFO TIMING step=4065: step=4.210s data=1.950s compute=4.176s [fwd=2.902s bwd=1.259s opt=0.003s] dispatch=2.815s [2026-10-18 22:19:24,515] loop.py:45 INFO TRAIN_STEP step=4065 loss=1.715745 Training: 23%|██▎ | 4065/18018 [8:49:10<32:06:22, 8.28s/step, loss=1.716, acc=0.634, acc_len=1.47, thru=18.1, I=20.9, T=30.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:24,763] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:29,418] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:19:31,824] loop.py:45 INFO TRAIN_STEP step=4066 loss=1.487551 Training: 23%|██▎ | 4066/18018 [8:49:17<30:58:21, 7.99s/step, loss=1.488, acc=0.669, acc_len=1.59, thru=16.4, I=25.5, T=27.4, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:19:31,819] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4065: forward=3271.2ms backward=1352.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:31,939] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:37,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:19:39,258] loop.py:45 INFO TRAIN_STEP step=4067 loss=1.601361 Training: 23%|██▎ | 4067/18018 [8:49:25<30:19:17, 7.82s/step, loss=1.601, acc=0.654, acc_len=1.54, thru=13.4, I=23.2, T=26.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:39,662] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:44,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:19:45,773] loop.py:45 INFO TRAIN_STEP step=4068 loss=1.576095 Training: 23%|██▎ | 4068/18018 [8:49:31<28:47:45, 7.43s/step, loss=1.576, acc=0.655, acc_len=1.54, thru=16.0, I=22.1, T=32.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:46,109] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:51,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:19:52,831] loop.py:45 INFO TRAIN_STEP step=4069 loss=1.546810 Training: 23%|██▎ | 4069/18018 [8:49:38<28:21:37, 7.32s/step, loss=1.547, acc=0.661, acc_len=1.56, thru=17.4, I=22.0, T=31.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:53,214] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:19:59,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:20:01,244] loop.py:380 INFO TIMING step=4070: step=4.438s data=2.099s compute=4.381s [fwd=3.028s bwd=1.338s opt=0.003s] dispatch=3.964s [2026-10-18 22:20:01,244] loop.py:45 INFO TRAIN_STEP step=4070 loss=1.716598 Training: 23%|██▎ | 4070/18018 [8:49:47<29:37:35, 7.65s/step, loss=1.717, acc=0.635, acc_len=1.48, thru=14.9, I=20.6, T=28.8, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:01,533] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:20:08,278] loop.py:45 INFO TRAIN_STEP step=4071 loss=1.532724 Training: 23%|██▎ | 4071/18018 [8:49:54<28:54:47, 7.46s/step, loss=1.533, acc=0.665, acc_len=1.57, thru=16.2, I=18.2, T=34.6, wait=3.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:20:08,273] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4070: forward=2656.5ms backward=996.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:19,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:20:19,973] loop.py:45 INFO TRAIN_STEP step=4072 loss=1.693341 Training: 23%|██▎ | 4072/18018 [8:50:06<33:49:57, 8.73s/step, loss=1.693, acc=0.639, acc_len=1.49, thru=12.1, I=11.9, T=29.8, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:20,077] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:25,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:20:27,081] loop.py:45 INFO TRAIN_STEP step=4073 loss=1.726717 Training: 23%|██▎ | 4073/18018 [8:50:13<31:56:27, 8.25s/step, loss=1.727, acc=0.635, acc_len=1.47, thru=16.6, I=21.4, T=28.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:27,213] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:31,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:20:33,327] loop.py:45 INFO TRAIN_STEP step=4074 loss=1.699230 Training: 23%|██▎ | 4074/18018 [8:50:19<29:36:58, 7.65s/step, loss=1.699, acc=0.636, acc_len=1.48, thru=18.0, I=25.4, T=29.6, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:33,439] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:38,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:20:40,111] loop.py:380 INFO TIMING step=4075: step=4.265s data=1.939s compute=4.234s [fwd=2.983s bwd=1.236s opt=0.003s] dispatch=2.508s [2026-10-18 22:20:40,111] loop.py:45 INFO TRAIN_STEP step=4075 loss=1.568363 Training: 23%|██▎ | 4075/18018 [8:50:26<28:36:34, 7.39s/step, loss=1.568, acc=0.659, acc_len=1.56, thru=17.6, I=23.7, T=30.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:40,177] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:46,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:20:46,381] loop.py:45 INFO TRAIN_STEP step=4076 loss=1.683094 Training: 23%|██▎ | 4076/18018 [8:50:32<27:18:37, 7.05s/step, loss=1.683, acc=0.638, acc_len=1.48, thru=18.5, I=21.4, T=35.2, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:20:46,376] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4075: forward=2545.4ms backward=1034.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:46,675] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:53,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:20:54,274] loop.py:45 INFO TRAIN_STEP step=4077 loss=1.650560 Training: 23%|██▎ | 4077/18018 [8:50:40<28:17:03, 7.30s/step, loss=1.651, acc=0.647, acc_len=1.51, thru=16.1, I=16.9, T=31.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:20:54,639] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:00,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:21:01,130] loop.py:45 INFO TRAIN_STEP step=4078 loss=1.536510 Training: 23%|██▎ | 4078/18018 [8:50:47<27:45:42, 7.17s/step, loss=1.537, acc=0.665, acc_len=1.58, thru=18.8, I=20.7, T=30.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:01,630] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:07,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:21:10,005] loop.py:45 INFO TRAIN_STEP step=4079 loss=1.586961 Training: 23%|██▎ | 4079/18018 [8:50:56<29:44:30, 7.68s/step, loss=1.587, acc=0.656, acc_len=1.54, thru=13.3, I=22.9, T=25.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:10,193] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:19,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:21:19,417] loop.py:380 INFO TIMING step=4080: step=4.502s data=2.180s compute=4.458s [fwd=3.138s bwd=1.305s opt=0.003s] dispatch=4.900s [2026-10-18 22:21:19,417] loop.py:45 INFO TRAIN_STEP step=4080 loss=1.508091 Training: 23%|██▎ | 4080/18018 [8:51:05<31:45:05, 8.20s/step, loss=1.508, acc=0.672, acc_len=1.60, thru=13.9, I=15.0, T=28.4, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:19,547] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:27,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:21:29,292] loop.py:45 INFO TRAIN_STEP step=4081 loss=1.811239 Training: 23%|██▎ | 4081/18018 [8:51:15<33:41:36, 8.70s/step, loss=1.811, acc=0.626, acc_len=1.46, thru=13.2, I=14.2, T=26.0, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:21:29,287] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4080: forward=3512.0ms backward=1367.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:29,327] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:36,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:21:37,090] loop.py:45 INFO TRAIN_STEP step=4082 loss=1.664046 Training: 23%|██▎ | 4082/18018 [8:51:23<32:38:23, 8.43s/step, loss=1.664, acc=0.644, acc_len=1.51, thru=14.8, I=18.6, T=30.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:37,442] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:44,141] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:21:45,584] loop.py:45 INFO TRAIN_STEP step=4083 loss=1.592065 Training: 23%|██▎ | 4083/18018 [8:51:31<32:42:29, 8.45s/step, loss=1.592, acc=0.657, acc_len=1.55, thru=14.7, I=17.5, T=26.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:45,650] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 22:21:52,507] loop.py:45 INFO TRAIN_STEP step=4084 loss=1.601436 Training: 23%|██▎ | 4084/18018 [8:51:38<30:55:54, 7.99s/step, loss=1.601, acc=0.652, acc_len=1.54, thru=15.7, I=17.7, T=33.2, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:58,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:21:59,484] loop.py:380 INFO TIMING step=4085: step=3.734s data=1.631s compute=3.695s [fwd=2.630s bwd=1.051s opt=0.003s] dispatch=3.234s [2026-10-18 22:21:59,484] loop.py:45 INFO TRAIN_STEP step=4085 loss=1.726189 Training: 23%|██▎ | 4085/18018 [8:51:45<29:45:09, 7.69s/step, loss=1.726, acc=0.629, acc_len=1.46, thru=19.1, I=19.4, T=34.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:21:59,795] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:05,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:22:06,864] loop.py:45 INFO TRAIN_STEP step=4086 loss=1.649610 Training: 23%|██▎ | 4086/18018 [8:51:52<29:23:34, 7.60s/step, loss=1.650, acc=0.644, acc_len=1.51, thru=16.1, I=19.0, T=28.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:22:06,859] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4085: forward=3189.4ms backward=1334.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:06,910] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:11,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:22:13,415] loop.py:45 INFO TRAIN_STEP step=4087 loss=1.514752 Training: 23%|██▎ | 4087/18018 [8:51:59<28:10:39, 7.28s/step, loss=1.515, acc=0.666, acc_len=1.58, thru=17.0, I=21.5, T=32.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:13,576] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:19,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:22:20,465] loop.py:45 INFO TRAIN_STEP step=4088 loss=1.609629 Training: 23%|██▎ | 4088/18018 [8:52:06<27:54:25, 7.21s/step, loss=1.610, acc=0.652, acc_len=1.53, thru=17.2, I=24.9, T=31.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:20,628] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:26,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:22:27,580] loop.py:45 INFO TRAIN_STEP step=4089 loss=1.556744 Training: 23%|██▎ | 4089/18018 [8:52:13<27:47:43, 7.18s/step, loss=1.557, acc=0.662, acc_len=1.57, thru=16.6, I=21.5, T=30.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:27,777] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:22:35,287] loop.py:380 INFO TIMING step=4090: step=3.841s data=1.694s compute=3.814s [fwd=2.796s bwd=1.003s opt=0.003s] dispatch=3.855s [2026-10-18 22:22:35,288] loop.py:45 INFO TRAIN_STEP step=4090 loss=1.683688 Training: 23%|██▎ | 4090/18018 [8:52:21<28:24:00, 7.34s/step, loss=1.684, acc=0.640, acc_len=1.49, thru=13.7, I=16.9, T=33.3, wait=3.9s, pool=44, epoch=1/2][2026-10-18 22:22:48,551] loop.py:45 INFO TRAIN_STEP step=4091 loss=1.693360 Training: 23%|██▎ | 4091/18018 [8:52:34<35:16:18, 9.12s/step, loss=1.693, acc=0.637, acc_len=1.48, thru=7.4, I=8.6, T=24.6, wait=8.1s, pool=8, epoch=1/2] (TrainerActor pid=9310) [2026-10-18 22:22:48,546] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4090: forward=3933.7ms backward=1204.0ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 22:22:49,062] eagle_store.py:517 WARNING batch_get_buffer missing keys (k130915-ace52148_hs, k130915-ace52148_ids, k130915-ace52148_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:57,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 22:22:49,059] eagle_store.py:517 WARNING batch_get_buffer missing keys (k130914-b9edd25f_hs, k130914-b9edd25f_ids, k130914-b9edd25f_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-18 22:22:59,203] loop.py:45 INFO TRAIN_STEP step=4092 loss=1.503080 Training: 23%|██▎ | 4092/18018 [8:52:45<37:02:56, 9.58s/step, loss=1.503, acc=0.673, acc_len=1.61, thru=17.5, I=18.1, T=32.6, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:22:59,642] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:07,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:23:08,550] loop.py:45 INFO TRAIN_STEP step=4093 loss=1.738944 Training: 23%|██▎ | 4093/18018 [8:52:54<36:46:53, 9.51s/step, loss=1.739, acc=0.636, acc_len=1.48, thru=14.4, I=14.8, T=29.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:08,625] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-18 22:23:16,118] loop.py:45 INFO TRAIN_STEP step=4094 loss=1.711066 Training: 23%|██▎ | 4094/18018 [8:53:02<34:31:23, 8.93s/step, loss=1.711, acc=0.634, acc_len=1.48, thru=15.5, I=15.7, T=31.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:16,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:16,586] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:23:23,977] loop.py:380 INFO TIMING step=4095: step=4.087s data=1.945s compute=4.041s [fwd=2.788s bwd=1.238s opt=0.003s] dispatch=3.764s [2026-10-18 22:23:23,977] loop.py:45 INFO TRAIN_STEP step=4095 loss=1.575265 Training: 23%|██▎ | 4095/18018 [8:53:10<33:17:09, 8.61s/step, loss=1.575, acc=0.658, acc_len=1.55, thru=14.9, I=15.2, T=31.3, wait=3.8s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:31,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:23:33,033] loop.py:45 INFO TRAIN_STEP step=4096 loss=1.701036 Training: 23%|██▎ | 4096/18018 [8:53:19<33:48:13, 8.74s/step, loss=1.701, acc=0.643, acc_len=1.50, thru=16.7, I=17.2, T=30.0, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:23:33,027] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4095: forward=2957.0ms backward=1262.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:33,277] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:39,567] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:23:39,960] loop.py:45 INFO TRAIN_STEP step=4097 loss=1.475465 Training: 23%|██▎ | 4097/18018 [8:53:26<31:41:50, 8.20s/step, loss=1.475, acc=0.668, acc_len=1.59, thru=16.0, I=19.4, T=32.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:40,083] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:46,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:23:47,011] loop.py:45 INFO TRAIN_STEP step=4098 loss=1.651809 Training: 23%|██▎ | 4098/18018 [8:53:33<30:21:57, 7.85s/step, loss=1.652, acc=0.647, acc_len=1.51, thru=17.9, I=18.9, T=32.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:47,266] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:55,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:23:56,117] loop.py:45 INFO TRAIN_STEP step=4099 loss=1.640707 Training: 23%|██▎ | 4099/18018 [8:53:42<31:49:05, 8.23s/step, loss=1.641, acc=0.646, acc_len=1.51, thru=13.8, I=14.7, T=30.0, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:23:56,301] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:24:03,542] loop.py:380 INFO TIMING step=4100: step=3.973s data=1.690s compute=3.945s [fwd=2.726s bwd=1.205s opt=0.003s] dispatch=3.441s [2026-10-18 22:24:03,542] loop.py:45 INFO TRAIN_STEP step=4100 loss=1.557369 Training: 23%|██▎ | 4100/18018 [8:53:49<30:52:49, 7.99s/step, loss=1.557, acc=0.660, acc_len=1.56, thru=13.6, I=17.3, T=32.2, wait=3.4s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:11,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:24:13,014] loop.py:45 INFO TRAIN_STEP step=4101 loss=1.725394 Training: 23%|██▎ | 4101/18018 [8:53:59<32:36:03, 8.43s/step, loss=1.725, acc=0.628, acc_len=1.45, thru=15.7, I=16.4, T=27.6, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:24:13,008] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4100: forward=3227.7ms backward=1354.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:13,182] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:19,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:24:20,522] loop.py:45 INFO TRAIN_STEP step=4102 loss=1.598801 Training: 23%|██▎ | 4102/18018 [8:54:06<31:31:28, 8.16s/step, loss=1.599, acc=0.655, acc_len=1.54, thru=14.8, I=20.4, T=30.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:20,784] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:25,499] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:24:26,937] loop.py:45 INFO TRAIN_STEP step=4103 loss=1.601631 Training: 23%|██▎ | 4103/18018 [8:54:13<29:30:15, 7.63s/step, loss=1.602, acc=0.654, acc_len=1.54, thru=18.1, I=22.5, T=31.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:27,019] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:32,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:24:33,349] loop.py:45 INFO TRAIN_STEP step=4104 loss=1.675984 Training: 23%|██▎ | 4104/18018 [8:54:19<28:05:05, 7.27s/step, loss=1.676, acc=0.638, acc_len=1.49, thru=18.4, I=24.3, T=34.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:33,795] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:24:40,461] loop.py:380 INFO TIMING step=4105: step=4.103s data=1.934s compute=4.058s [fwd=2.915s bwd=1.129s opt=0.003s] dispatch=3.001s [2026-10-18 22:24:40,461] loop.py:45 INFO TRAIN_STEP step=4105 loss=1.522123 Training: 23%|██▎ | 4105/18018 [8:54:26<27:54:30, 7.22s/step, loss=1.522, acc=0.666, acc_len=1.58, thru=16.9, I=20.8, T=31.2, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:47,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:24:49,524] loop.py:45 INFO TRAIN_STEP step=4106 loss=1.725888 Training: 23%|██▎ | 4106/18018 [8:54:35<30:02:19, 7.77s/step, loss=1.726, acc=0.640, acc_len=1.49, thru=15.3, I=17.2, T=30.7, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:24:49,520] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4105: forward=2938.8ms backward=1172.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:50,007] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:56,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:24:57,240] loop.py:45 INFO TRAIN_STEP step=4107 loss=1.646870 Training: 23%|██▎ | 4107/18018 [8:54:43<29:58:08, 7.76s/step, loss=1.647, acc=0.641, acc_len=1.50, thru=13.9, I=20.1, T=30.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:24:57,463] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:25:02,964] loop.py:45 INFO TRAIN_STEP step=4108 loss=1.617344 Training: 23%|██▎ | 4108/18018 [8:54:49<27:36:37, 7.15s/step, loss=1.617, acc=0.649, acc_len=1.52, thru=18.1, I=19.8, T=32.9, wait=1.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:08,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:25:10,710] loop.py:45 INFO TRAIN_STEP step=4109 loss=1.700017 Training: 23%|██▎ | 4109/18018 [8:54:56<28:18:30, 7.33s/step, loss=1.700, acc=0.634, acc_len=1.48, thru=18.1, I=20.0, T=30.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:10,910] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:25:17,141] loop.py:380 INFO TIMING step=4110: step=3.457s data=1.444s compute=3.430s [fwd=2.392s bwd=1.025s opt=0.003s] dispatch=2.963s [2026-10-18 22:25:17,141] loop.py:45 INFO TRAIN_STEP step=4110 loss=1.636054 Training: 23%|██▎ | 4110/18018 [8:55:03<27:15:56, 7.06s/step, loss=1.636, acc=0.648, acc_len=1.52, thru=16.4, I=23.3, T=37.0, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:26,009] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:25:26,811] loop.py:45 INFO TRAIN_STEP step=4111 loss=1.649721 Training: 23%|██▎ | 4111/18018 [8:55:12<30:17:32, 7.84s/step, loss=1.650, acc=0.649, acc_len=1.52, thru=14.5, I=13.9, T=29.9, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:25:26,806] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4110: forward=2950.2ms backward=1285.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:27,013] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:33,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:25:34,140] loop.py:45 INFO TRAIN_STEP step=4112 loss=1.638771 Training: 23%|██▎ | 4112/18018 [8:55:20<29:41:43, 7.69s/step, loss=1.639, acc=0.650, acc_len=1.53, thru=17.3, I=19.4, T=30.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:34,590] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:42,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:25:43,356] loop.py:45 INFO TRAIN_STEP step=4113 loss=1.699410 Training: 23%|██▎ | 4113/18018 [8:55:29<31:27:49, 8.15s/step, loss=1.699, acc=0.640, acc_len=1.50, thru=14.6, I=14.8, T=28.8, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:43,386] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:48,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:25:50,184] loop.py:45 INFO TRAIN_STEP step=4114 loss=1.583777 Training: 23%|██▎ | 4114/18018 [8:55:36<29:56:04, 7.75s/step, loss=1.584, acc=0.655, acc_len=1.54, thru=18.3, I=21.5, T=33.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:50,602] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:58,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:25:59,370] loop.py:380 INFO TIMING step=4115: step=3.862s data=1.705s compute=3.829s [fwd=2.675s bwd=1.139s opt=0.003s] dispatch=5.314s [2026-10-18 22:25:59,370] loop.py:45 INFO TRAIN_STEP step=4115 loss=1.570159 Training: 23%|██▎ | 4115/18018 [8:55:45<31:35:51, 8.18s/step, loss=1.570, acc=0.656, acc_len=1.55, thru=14.6, I=16.1, T=33.1, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:25:59,438] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:26:05,576] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:26:06,983] loop.py:45 INFO TRAIN_STEP step=4116 loss=1.671621 Training: 23%|██▎ | 4116/18018 [8:55:53<30:56:08, 8.01s/step, loss=1.672, acc=0.637, acc_len=1.49, thru=16.6, I=20.1, T=31.2, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:26:06,978] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4115: forward=2880.2ms backward=1176.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:26:07,090] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:26:12,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:26:13,372] loop.py:45 INFO TRAIN_STEP step=4117 loss=1.601607 Training: 23%|██▎ | 4117/18018 [8:55:59<29:03:17, 7.52s/step, loss=1.602, acc=0.656, acc_len=1.55, thru=17.7, I=23.1, T=31.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:26:13,498] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 22:26:20,264] loop.py:45 INFO TRAIN_STEP step=4118 loss=1.652272 Training: 23%|██▎ | 4118/18018 [8:56:06<28:19:11, 7.33s/step, loss=1.652, acc=0.647, acc_len=1.52, thru=16.2, I=19.8, T=30.2, wait=2.6s, pool=52, epoch=1/2][2026-10-18 22:26:29,979] loop.py:45 INFO TRAIN_STEP step=4119 loss=1.705848 Training: 23%|██▎ | 4119/18018 [8:56:16<31:04:32, 8.05s/step, loss=1.706, acc=0.640, acc_len=1.49, thru=13.9, I=14.5, T=27.3, wait=5.0s, pool=56, epoch=1/2][2026-10-18 22:26:38,905] loop.py:380 INFO TIMING step=4120: step=4.098s data=1.862s compute=4.026s [fwd=2.857s bwd=1.154s opt=0.003s] dispatch=4.817s [2026-10-18 22:26:38,905] loop.py:45 INFO TRAIN_STEP step=4120 loss=1.697135 Training: 23%|██▎ | 4120/18018 [8:56:24<32:05:14, 8.31s/step, loss=1.697, acc=0.641, acc_len=1.50, thru=13.3, I=12.8, T=31.2, wait=4.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:26:48,128] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:26:48,296] loop.py:45 INFO TRAIN_STEP step=4121 loss=1.656219 Training: 23%|██▎ | 4121/18018 [8:56:34<33:20:04, 8.64s/step, loss=1.656, acc=0.643, acc_len=1.51, thru=15.6, I=14.3, T=33.0, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:26:48,291] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4120: forward=2660.2ms backward=1155.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:26:48,632] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:26:56,678] loop.py:45 INFO TRAIN_STEP step=4122 loss=1.812954 Training: 23%|██▎ | 4122/18018 [8:56:42<33:02:22, 8.56s/step, loss=1.813, acc=0.620, acc_len=1.44, thru=14.8, I=14.8, T=30.1, wait=4.1s, pool=52, epoch=1/2][2026-10-18 22:27:05,563] loop.py:45 INFO TRAIN_STEP step=4123 loss=1.669723 Training: 23%|██▎ | 4123/18018 [8:56:51<33:24:48, 8.66s/step, loss=1.670, acc=0.642, acc_len=1.50, thru=15.9, I=14.4, T=33.3, wait=5.0s, pool=60, epoch=1/2][2026-10-18 22:27:11,550] loop.py:45 INFO TRAIN_STEP step=4124 loss=1.582455 Training: 23%|██▎ | 4124/18018 [8:56:57<30:19:20, 7.86s/step, loss=1.582, acc=0.655, acc_len=1.54, thru=19.5, I=19.1, T=32.8, wait=2.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:20,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:27:20,957] loop.py:380 INFO TIMING step=4125: step=4.501s data=2.185s compute=4.475s [fwd=3.260s bwd=1.200s opt=0.003s] dispatch=4.896s [2026-10-18 22:27:20,958] loop.py:45 INFO TRAIN_STEP step=4125 loss=1.686319 Training: 23%|██▎ | 4125/18018 [8:57:07<32:06:48, 8.32s/step, loss=1.686, acc=0.641, acc_len=1.49, thru=13.6, I=13.9, T=28.4, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:21,051] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:25,387] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:27:26,958] loop.py:45 INFO TRAIN_STEP step=4126 loss=1.570777 Training: 23%|██▎ | 4126/18018 [8:57:13<29:25:28, 7.63s/step, loss=1.571, acc=0.659, acc_len=1.55, thru=18.0, I=20.9, T=30.2, wait=1.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:27:26,953] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4125: forward=2939.7ms backward=1249.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:27,397] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:32,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:27:33,362] loop.py:45 INFO TRAIN_STEP step=4127 loss=1.672031 Training: 23%|██▎ | 4127/18018 [8:57:19<28:00:39, 7.26s/step, loss=1.672, acc=0.640, acc_len=1.50, thru=18.5, I=23.9, T=33.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:33,871] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:41,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:27:41,418] loop.py:45 INFO TRAIN_STEP step=4128 loss=1.616920 Training: 23%|██▎ | 4128/18018 [8:57:27<28:55:52, 7.50s/step, loss=1.617, acc=0.652, acc_len=1.53, thru=15.1, I=17.3, T=32.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:41,682] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:47,722] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:27:48,899] loop.py:45 INFO TRAIN_STEP step=4129 loss=1.675074 Training: 23%|██▎ | 4129/18018 [8:57:34<28:54:22, 7.49s/step, loss=1.675, acc=0.643, acc_len=1.50, thru=17.1, I=19.2, T=32.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:49,233] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:55,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:27:57,236] loop.py:380 INFO TIMING step=4130: step=5.185s data=2.552s compute=5.153s [fwd=3.594s bwd=1.543s opt=0.003s] dispatch=3.143s [2026-10-18 22:27:57,237] loop.py:45 INFO TRAIN_STEP step=4130 loss=1.643225 Training: 23%|██▎ | 4130/18018 [8:57:43<29:53:03, 7.75s/step, loss=1.643, acc=0.643, acc_len=1.51, thru=13.9, I=18.7, T=24.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:27:57,587] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:03,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 22:28:03,368] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4130: forward=2497.5ms backward=1001.8ms [2026-10-18 22:28:03,373] loop.py:45 INFO TRAIN_STEP step=4131 loss=1.569504 Training: 23%|██▎ | 4131/18018 [8:57:49<28:01:08, 7.26s/step, loss=1.570, acc=0.657, acc_len=1.55, thru=16.8, I=21.1, T=36.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:03,608] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:10,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:28:11,294] loop.py:45 INFO TRAIN_STEP step=4132 loss=1.757839 Training: 23%|██▎ | 4132/18018 [8:57:57<28:46:32, 7.46s/step, loss=1.758, acc=0.632, acc_len=1.47, thru=16.2, I=18.6, T=30.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:11,598] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:28:17,600] loop.py:45 INFO TRAIN_STEP step=4133 loss=1.548227 Training: 23%|██▎ | 4133/18018 [8:58:03<27:26:17, 7.11s/step, loss=1.548, acc=0.661, acc_len=1.55, thru=17.6, I=21.3, T=30.9, wait=2.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:26,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:28:27,030] loop.py:45 INFO TRAIN_STEP step=4134 loss=1.722267 Training: 23%|██▎ | 4134/18018 [8:58:13<30:06:51, 7.81s/step, loss=1.722, acc=0.626, acc_len=1.46, thru=14.8, I=13.8, T=31.8, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:27,518] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:32,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:28:34,516] loop.py:380 INFO TIMING step=4135: step=4.357s data=2.169s compute=4.317s [fwd=3.021s bwd=1.282s opt=0.003s] dispatch=3.121s [2026-10-18 22:28:34,516] loop.py:45 INFO TRAIN_STEP step=4135 loss=1.580672 Training: 23%|██▎ | 4135/18018 [8:58:20<29:44:34, 7.71s/step, loss=1.581, acc=0.657, acc_len=1.54, thru=17.6, I=22.6, T=29.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:34,550] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:39,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:28:40,111] loop.py:45 INFO TRAIN_STEP step=4136 loss=1.641910 Training: 23%|██▎ | 4136/18018 [8:58:26<27:17:19, 7.08s/step, loss=1.642, acc=0.645, acc_len=1.51, thru=19.6, I=24.5, T=33.4, wait=1.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:28:40,106] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4135: forward=2703.2ms backward=1085.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:40,460] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:28:46,261] loop.py:45 INFO TRAIN_STEP step=4137 loss=1.598756 Training: 23%|██▎ | 4137/18018 [8:58:32<26:12:57, 6.80s/step, loss=1.599, acc=0.653, acc_len=1.54, thru=18.7, I=20.8, T=33.5, wait=2.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:53,323] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:28:54,033] loop.py:45 INFO TRAIN_STEP step=4138 loss=1.535477 Training: 23%|██▎ | 4138/18018 [8:58:40<27:20:19, 7.09s/step, loss=1.535, acc=0.665, acc_len=1.58, thru=17.0, I=17.0, T=26.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:28:54,329] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:00,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:29:00,930] loop.py:45 INFO TRAIN_STEP step=4139 loss=1.674458 Training: 23%|██▎ | 4139/18018 [8:58:47<27:06:43, 7.03s/step, loss=1.674, acc=0.645, acc_len=1.51, thru=15.4, I=18.8, T=29.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:01,253] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:29:08,365] loop.py:380 INFO TIMING step=4140: step=4.328s data=1.987s compute=4.297s [fwd=3.008s bwd=1.274s opt=0.003s] dispatch=3.097s [2026-10-18 22:29:08,365] loop.py:45 INFO TRAIN_STEP step=4140 loss=1.746793 Training: 23%|██▎ | 4140/18018 [8:58:54<27:34:34, 7.15s/step, loss=1.747, acc=0.633, acc_len=1.47, thru=15.1, I=16.3, T=29.6, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:15,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:29:15,445] loop.py:45 INFO TRAIN_STEP step=4141 loss=1.700339 Training: 23%|██▎ | 4141/18018 [8:59:01<27:29:24, 7.13s/step, loss=1.700, acc=0.633, acc_len=1.48, thru=17.6, I=16.8, T=30.8, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:29:15,440] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4140: forward=2878.7ms backward=1220.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:15,685] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:21,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:29:21,686] loop.py:45 INFO TRAIN_STEP step=4142 loss=1.420325 Training: 23%|██▎ | 4142/18018 [8:59:07<26:27:25, 6.86s/step, loss=1.420, acc=0.679, acc_len=1.63, thru=18.9, I=19.7, T=32.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:22,134] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:28,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:29:29,751] loop.py:45 INFO TRAIN_STEP step=4143 loss=1.667015 Training: 23%|██▎ | 4143/18018 [8:59:15<27:50:32, 7.22s/step, loss=1.667, acc=0.647, acc_len=1.52, thru=14.9, I=18.9, T=28.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:30,234] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:35,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:29:37,589] loop.py:45 INFO TRAIN_STEP step=4144 loss=1.542820 Training: 23%|██▎ | 4144/18018 [8:59:23<28:33:15, 7.41s/step, loss=1.543, acc=0.662, acc_len=1.56, thru=14.5, I=20.1, T=26.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:37,652] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:43,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:29:44,180] loop.py:380 INFO TIMING step=4145: step=3.930s data=1.812s compute=3.885s [fwd=2.715s bwd=1.155s opt=0.003s] dispatch=2.650s [2026-10-18 22:29:44,180] loop.py:45 INFO TRAIN_STEP step=4145 loss=1.823996 Training: 23%|██▎ | 4145/18018 [8:59:30<27:36:10, 7.16s/step, loss=1.824, acc=0.613, acc_len=1.41, thru=16.9, I=20.7, T=32.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:44,495] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:49,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:29:50,289] loop.py:45 INFO TRAIN_STEP step=4146 loss=1.517597 Training: 23%|██▎ | 4146/18018 [8:59:36<26:22:57, 6.85s/step, loss=1.518, acc=0.665, acc_len=1.58, thru=18.9, I=22.8, T=32.4, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:29:50,283] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4145: forward=2799.7ms backward=1107.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:29:50,733] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:29:57,699] loop.py:45 INFO TRAIN_STEP step=4147 loss=1.673595 Training: 23%|██▎ | 4147/18018 [8:59:43<27:02:02, 7.02s/step, loss=1.674, acc=0.645, acc_len=1.51, thru=12.5, I=18.6, T=36.4, wait=3.9s, pool=24, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 22:29:57,765] eagle_store.py:517 WARNING batch_get_buffer missing keys (k132703-ab2d11ff_hs, k132703-ab2d11ff_ids, k132703-ab2d11ff_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:07,688] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 22:29:57,765] eagle_store.py:517 WARNING batch_get_buffer missing keys (k132702-baf28ff1_hs, k132702-baf28ff1_ids, k132702-baf28ff1_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-18 22:30:07,920] loop.py:45 INFO TRAIN_STEP step=4148 loss=1.577650 Training: 23%|██▎ | 4148/18018 [8:59:54<30:44:11, 7.98s/step, loss=1.578, acc=0.657, acc_len=1.55, thru=17.0, I=14.4, T=32.8, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:08,193] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:13,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:30:14,815] loop.py:45 INFO TRAIN_STEP step=4149 loss=1.526238 Training: 23%|██▎ | 4149/18018 [9:00:00<29:28:53, 7.65s/step, loss=1.526, acc=0.664, acc_len=1.58, thru=18.1, I=20.6, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:15,054] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:21,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:30:21,492] loop.py:380 INFO TIMING step=4150: step=3.856s data=1.731s compute=3.814s [fwd=2.677s bwd=1.123s opt=0.003s] dispatch=2.813s [2026-10-18 22:30:21,492] loop.py:45 INFO TRAIN_STEP step=4150 loss=1.582326 Training: 23%|██▎ | 4150/18018 [9:00:07<28:21:13, 7.36s/step, loss=1.582, acc=0.657, acc_len=1.55, thru=17.9, I=21.4, T=33.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:21,548] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:29,701] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:30:30,392] loop.py:45 INFO TRAIN_STEP step=4151 loss=1.716594 Training: 23%|██▎ | 4151/18018 [9:00:16<30:07:43, 7.82s/step, loss=1.717, acc=0.638, acc_len=1.48, thru=14.0, I=15.0, T=30.8, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:30:30,387] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4150: forward=2764.9ms backward=1322.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:30,705] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:37,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:30:38,004] loop.py:45 INFO TRAIN_STEP step=4152 loss=1.637732 Training: 23%|██▎ | 4152/18018 [9:00:24<29:53:02, 7.76s/step, loss=1.638, acc=0.649, acc_len=1.52, thru=15.6, I=17.9, T=29.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:38,333] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:43,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:30:45,322] loop.py:45 INFO TRAIN_STEP step=4153 loss=1.614097 Training: 23%|██▎ | 4153/18018 [9:00:31<29:22:27, 7.63s/step, loss=1.614, acc=0.648, acc_len=1.52, thru=16.9, I=20.9, T=28.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:45,686] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:50,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:30:52,219] loop.py:45 INFO TRAIN_STEP step=4154 loss=1.560838 Training: 23%|██▎ | 4154/18018 [9:00:38<28:31:41, 7.41s/step, loss=1.561, acc=0.659, acc_len=1.56, thru=15.7, I=24.3, T=31.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:52,373] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:30:59,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:30:59,742] loop.py:380 INFO TIMING step=4155: step=4.172s data=1.927s compute=4.144s [fwd=2.928s bwd=1.200s opt=0.003s] dispatch=3.343s [2026-10-18 22:30:59,742] loop.py:45 INFO TRAIN_STEP step=4155 loss=1.648095 Training: 23%|██▎ | 4155/18018 [9:00:45<28:39:29, 7.44s/step, loss=1.648, acc=0.643, acc_len=1.50, thru=16.1, I=19.1, T=30.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:00,100] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:04,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:31:06,210] loop.py:45 INFO TRAIN_STEP step=4156 loss=1.636084 Training: 23%|██▎ | 4156/18018 [9:00:52<27:32:00, 7.15s/step, loss=1.636, acc=0.645, acc_len=1.52, thru=17.7, I=23.7, T=32.5, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:31:06,204] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4155: forward=2767.1ms backward=1121.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:06,358] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-18 22:31:12,214] loop.py:45 INFO TRAIN_STEP step=4157 loss=1.557850 Training: 23%|██▎ | 4157/18018 [9:00:58<26:12:19, 6.81s/step, loss=1.558, acc=0.662, acc_len=1.56, thru=18.0, I=25.4, T=34.1, wait=2.2s, pool=48, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-18 22:31:12,291] eagle_store.py:517 WARNING batch_get_buffer missing keys (k133031-b3f822b3_hs, k133031-b3f822b3_ids, k133031-b3f822b3_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:20,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-18 22:31:12,290] eagle_store.py:517 WARNING batch_get_buffer missing keys (k133030-b48d8825_hs, k133030-b48d8825_ids, k133030-b48d8825_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-18 22:31:22,412] loop.py:45 INFO TRAIN_STEP step=4158 loss=1.544048 Training: 23%|██▎ | 4158/18018 [9:01:08<30:07:15, 7.82s/step, loss=1.544, acc=0.661, acc_len=1.56, thru=15.5, I=15.3, T=28.4, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:22,486] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:29,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:31:30,784] loop.py:45 INFO TRAIN_STEP step=4159 loss=1.664763 Training: 23%|██▎ | 4159/18018 [9:01:16<30:45:09, 7.99s/step, loss=1.665, acc=0.646, acc_len=1.52, thru=13.7, I=17.7, T=29.8, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:31,126] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:37,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:31:37,597] loop.py:380 INFO TIMING step=4160: step=3.602s data=1.588s compute=3.560s [fwd=2.492s bwd=1.053s opt=0.003s] dispatch=3.202s [2026-10-18 22:31:37,597] loop.py:45 INFO TRAIN_STEP step=4160 loss=1.614361 Training: 23%|██▎ | 4160/18018 [9:01:23<29:23:28, 7.64s/step, loss=1.614, acc=0.652, acc_len=1.54, thru=18.1, I=20.7, T=35.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:38,085] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:44,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:31:45,620] loop.py:45 INFO TRAIN_STEP step=4161 loss=1.631321 Training: 23%|██▎ | 4161/18018 [9:01:31<29:50:24, 7.75s/step, loss=1.631, acc=0.649, acc_len=1.52, thru=16.2, I=18.2, T=31.4, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:31:45,616] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4160: forward=2845.9ms backward=1180.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:45,955] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:53,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:31:53,691] loop.py:45 INFO TRAIN_STEP step=4162 loss=1.599741 Training: 23%|██▎ | 4162/18018 [9:01:39<30:12:19, 7.85s/step, loss=1.600, acc=0.657, acc_len=1.54, thru=16.5, I=17.6, T=30.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:53,860] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:31:58,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:32:00,347] loop.py:45 INFO TRAIN_STEP step=4163 loss=1.580330 Training: 23%|██▎ | 4163/18018 [9:01:46<28:49:30, 7.49s/step, loss=1.580, acc=0.659, acc_len=1.56, thru=16.8, I=21.5, T=30.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:00,832] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:06,565] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:32:06,707] loop.py:45 INFO TRAIN_STEP step=4164 loss=1.690175 Training: 23%|██▎ | 4164/18018 [9:01:52<27:31:04, 7.15s/step, loss=1.690, acc=0.638, acc_len=1.49, thru=18.5, I=24.3, T=36.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:07,068] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:13,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:32:14,794] loop.py:380 INFO TIMING step=4165: step=3.970s data=1.827s compute=3.937s [fwd=2.815s bwd=1.108s opt=0.003s] dispatch=4.109s [2026-10-18 22:32:14,795] loop.py:45 INFO TRAIN_STEP step=4165 loss=1.674765 Training: 23%|██▎ | 4165/18018 [9:02:00<28:36:03, 7.43s/step, loss=1.675, acc=0.642, acc_len=1.50, thru=15.5, I=16.7, T=32.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:14,902] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:20,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:32:21,733] loop.py:45 INFO TRAIN_STEP step=4166 loss=1.565292 Training: 23%|██▎ | 4166/18018 [9:02:07<28:01:36, 7.28s/step, loss=1.565, acc=0.659, acc_len=1.56, thru=17.8, I=23.0, T=31.3, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:32:21,729] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4165: forward=2873.9ms backward=1166.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:22,098] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:26,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:32:28,345] loop.py:45 INFO TRAIN_STEP step=4167 loss=1.562966 Training: 23%|██▎ | 4167/18018 [9:02:14<27:15:04, 7.08s/step, loss=1.563, acc=0.660, acc_len=1.55, thru=17.8, I=26.5, T=33.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:28,743] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:33,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:32:34,686] loop.py:45 INFO TRAIN_STEP step=4168 loss=1.588571 Training: 23%|██▎ | 4168/18018 [9:02:20<26:23:27, 6.86s/step, loss=1.589, acc=0.652, acc_len=1.54, thru=19.2, I=27.6, T=32.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:35,172] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:32:41,566] loop.py:45 INFO TRAIN_STEP step=4169 loss=1.536211 Training: 23%|██▎ | 4169/18018 [9:02:27<26:24:38, 6.87s/step, loss=1.536, acc=0.659, acc_len=1.56, thru=16.6, I=22.4, T=32.2, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:48,461] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:32:50,407] loop.py:380 INFO TIMING step=4170: step=4.706s data=2.263s compute=4.666s [fwd=3.307s bwd=1.343s opt=0.003s] dispatch=4.127s [2026-10-18 22:32:50,407] loop.py:45 INFO TRAIN_STEP step=4170 loss=1.644890 Training: 23%|██▎ | 4170/18018 [9:02:36<28:41:25, 7.46s/step, loss=1.645, acc=0.646, acc_len=1.52, thru=14.8, I=17.0, T=27.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:50,479] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:58,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:32:59,132] loop.py:45 INFO TRAIN_STEP step=4171 loss=1.709919 Training: 23%|██▎ | 4171/18018 [9:02:45<30:09:04, 7.84s/step, loss=1.710, acc=0.634, acc_len=1.48, thru=14.8, I=15.9, T=29.4, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:32:59,127] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4170: forward=2915.6ms backward=1395.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:32:59,167] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:04,794] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:33:06,244] loop.py:45 INFO TRAIN_STEP step=4172 loss=1.618608 Training: 23%|██▎ | 4172/18018 [9:02:52<29:18:37, 7.62s/step, loss=1.619, acc=0.647, acc_len=1.52, thru=16.7, I=22.0, T=33.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:06,319] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:12,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:33:13,456] loop.py:45 INFO TRAIN_STEP step=4173 loss=1.495207 Training: 23%|██▎ | 4173/18018 [9:02:59<28:50:13, 7.50s/step, loss=1.495, acc=0.668, acc_len=1.59, thru=16.8, I=20.8, T=31.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:13,634] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:19,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:33:21,607] loop.py:45 INFO TRAIN_STEP step=4174 loss=1.649320 Training: 23%|██▎ | 4174/18018 [9:03:07<29:35:14, 7.69s/step, loss=1.649, acc=0.648, acc_len=1.52, thru=14.8, I=19.0, T=27.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:21,700] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:26,600] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:33:29,586] loop.py:380 INFO TIMING step=4175: step=5.221s data=2.667s compute=5.192s [fwd=3.579s bwd=1.597s opt=0.003s] dispatch=2.748s [2026-10-18 22:33:29,586] loop.py:45 INFO TRAIN_STEP step=4175 loss=1.670405 Training: 23%|██▎ | 4175/18018 [9:03:15<29:54:53, 7.78s/step, loss=1.670, acc=0.640, acc_len=1.49, thru=13.3, I=25.0, T=24.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:29,635] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=48/64 [2026-10-18 22:33:36,329] loop.py:45 INFO TRAIN_STEP step=4176 loss=1.697185 Training: 23%|██▎ | 4176/18018 [9:03:22<28:42:52, 7.47s/step, loss=1.697, acc=0.638, acc_len=1.49, thru=13.3, I=20.4, T=32.6, wait=2.8s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:33:36,325] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4175: forward=2765.4ms backward=1111.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:43,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:33:43,887] loop.py:45 INFO TRAIN_STEP step=4177 loss=1.607201 Training: 23%|██▎ | 4177/18018 [9:03:29<28:49:06, 7.50s/step, loss=1.607, acc=0.653, acc_len=1.53, thru=17.3, I=17.5, T=31.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:44,073] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:50,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:33:50,830] loop.py:45 INFO TRAIN_STEP step=4178 loss=1.618362 Training: 23%|██▎ | 4178/18018 [9:03:36<28:10:33, 7.33s/step, loss=1.618, acc=0.646, acc_len=1.52, thru=17.1, I=20.4, T=29.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:51,305] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:57,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:33:58,112] loop.py:45 INFO TRAIN_STEP step=4179 loss=1.632270 Training: 23%|██▎ | 4179/18018 [9:03:44<28:07:09, 7.31s/step, loss=1.632, acc=0.646, acc_len=1.52, thru=16.6, I=18.5, T=33.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:33:58,250] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:03,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:34:04,904] loop.py:380 INFO TIMING step=4180: step=4.072s data=1.874s compute=4.036s [fwd=2.902s bwd=1.119s opt=0.003s] dispatch=2.712s [2026-10-18 22:34:04,905] loop.py:45 INFO TRAIN_STEP step=4180 loss=1.671665 Training: 23%|██▎ | 4180/18018 [9:03:50<27:31:05, 7.16s/step, loss=1.672, acc=0.639, acc_len=1.49, thru=18.2, I=23.3, T=31.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:05,104] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:11,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:34:11,373] loop.py:45 INFO TRAIN_STEP step=4181 loss=1.486661 Training: 23%|██▎ | 4181/18018 [9:03:57<26:43:05, 6.95s/step, loss=1.487, acc=0.674, acc_len=1.60, thru=17.7, I=21.6, T=29.8, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:34:11,368] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4180: forward=2940.7ms backward=1294.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:11,634] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:16,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:34:18,878] loop.py:45 INFO TRAIN_STEP step=4182 loss=1.620682 Training: 23%|██▎ | 4182/18018 [9:04:04<27:21:15, 7.12s/step, loss=1.621, acc=0.650, acc_len=1.52, thru=17.1, I=19.9, T=27.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:19,193] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 22:34:25,488] loop.py:45 INFO TRAIN_STEP step=4183 loss=1.461020 Training: 23%|██▎ | 4183/18018 [9:04:11<26:46:03, 6.97s/step, loss=1.461, acc=0.672, acc_len=1.60, thru=14.4, I=20.3, T=34.2, wait=2.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:34,746] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:34:36,100] loop.py:45 INFO TRAIN_STEP step=4184 loss=1.707741 Training: 23%|██▎ | 4184/18018 [9:04:22<30:58:11, 8.06s/step, loss=1.708, acc=0.634, acc_len=1.48, thru=14.1, I=15.1, T=28.3, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:36,264] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:34:42,711] loop.py:380 INFO TIMING step=4185: step=4.051s data=1.854s compute=4.015s [fwd=2.739s bwd=1.261s opt=0.003s] dispatch=2.551s [2026-10-18 22:34:42,711] loop.py:45 INFO TRAIN_STEP step=4185 loss=1.548534 Training: 23%|██▎ | 4185/18018 [9:04:28<29:17:48, 7.62s/step, loss=1.549, acc=0.656, acc_len=1.56, thru=17.7, I=19.5, T=31.6, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:51,131] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:34:51,220] loop.py:45 INFO TRAIN_STEP step=4186 loss=1.685154 Training: 23%|██▎ | 4186/18018 [9:04:37<30:18:58, 7.89s/step, loss=1.685, acc=0.636, acc_len=1.48, thru=16.9, I=15.2, T=29.4, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:34:51,215] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4185: forward=3033.0ms backward=1275.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:34:51,634] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:34:57,914] loop.py:45 INFO TRAIN_STEP step=4187 loss=1.568109 Training: 23%|██▎ | 4187/18018 [9:04:44<28:56:12, 7.53s/step, loss=1.568, acc=0.658, acc_len=1.55, thru=17.7, I=17.8, T=34.3, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:05,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:35:05,491] loop.py:45 INFO TRAIN_STEP step=4188 loss=1.550376 Training: 23%|██▎ | 4188/18018 [9:04:51<28:59:03, 7.54s/step, loss=1.550, acc=0.661, acc_len=1.56, thru=17.0, I=17.6, T=31.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:05,822] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:35:15,960] loop.py:45 INFO TRAIN_STEP step=4189 loss=1.545029 Training: 23%|██▎ | 4189/18018 [9:05:02<32:21:08, 8.42s/step, loss=1.545, acc=0.661, acc_len=1.56, thru=12.1, I=11.6, T=30.0, wait=6.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:21,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:35:22,869] loop.py:380 INFO TIMING step=4190: step=3.802s data=1.767s compute=3.759s [fwd=2.621s bwd=1.122s opt=0.003s] dispatch=3.098s [2026-10-18 22:35:22,869] loop.py:45 INFO TRAIN_STEP step=4190 loss=1.542610 Training: 23%|██▎ | 4190/18018 [9:05:08<30:36:25, 7.97s/step, loss=1.543, acc=0.664, acc_len=1.57, thru=18.4, I=18.4, T=33.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:23,225] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:30,646] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:35:31,270] loop.py:45 INFO TRAIN_STEP step=4191 loss=1.624701 Training: 23%|██▎ | 4191/18018 [9:05:17<31:06:16, 8.10s/step, loss=1.625, acc=0.651, acc_len=1.53, thru=14.1, I=17.0, T=29.8, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:35:31,265] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4190: forward=3047.8ms backward=1182.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:31,652] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:35:39,096] loop.py:45 INFO TRAIN_STEP step=4192 loss=1.711563 Training: 23%|██▎ | 4192/18018 [9:05:25<30:47:15, 8.02s/step, loss=1.712, acc=0.635, acc_len=1.48, thru=15.0, I=15.9, T=29.9, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:43,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:35:45,757] loop.py:45 INFO TRAIN_STEP step=4193 loss=1.638324 Training: 23%|██▎ | 4193/18018 [9:05:31<29:13:23, 7.61s/step, loss=1.638, acc=0.650, acc_len=1.52, thru=18.6, I=20.5, T=31.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:45,991] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:52,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:35:54,095] loop.py:45 INFO TRAIN_STEP step=4194 loss=1.554538 Training: 23%|██▎ | 4194/18018 [9:05:40<30:03:39, 7.83s/step, loss=1.555, acc=0.659, acc_len=1.56, thru=15.8, I=18.2, T=29.5, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:35:54,296] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:01,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:36:02,663] loop.py:380 INFO TIMING step=4195: step=4.706s data=2.334s compute=4.672s [fwd=3.247s bwd=1.409s opt=0.003s] dispatch=3.853s [2026-10-18 22:36:02,664] loop.py:45 INFO TRAIN_STEP step=4195 loss=1.596997 Training: 23%|██▎ | 4195/18018 [9:05:48<30:54:43, 8.05s/step, loss=1.597, acc=0.653, acc_len=1.54, thru=14.1, I=18.3, T=27.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:02,962] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:09,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:36:11,343] loop.py:45 INFO TRAIN_STEP step=4196 loss=1.584652 Training: 23%|██▎ | 4196/18018 [9:05:57<31:37:56, 8.24s/step, loss=1.585, acc=0.655, acc_len=1.55, thru=14.6, I=17.8, T=27.2, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:36:11,338] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4195: forward=3241.3ms backward=1408.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:11,720] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:16,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:36:18,427] loop.py:45 INFO TRAIN_STEP step=4197 loss=1.531452 Training: 23%|██▎ | 4197/18018 [9:06:04<30:18:06, 7.89s/step, loss=1.531, acc=0.665, acc_len=1.57, thru=14.9, I=26.0, T=31.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:18,462] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:23,428] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:36:24,768] loop.py:45 INFO TRAIN_STEP step=4198 loss=1.565182 Training: 23%|██▎ | 4198/18018 [9:06:10<28:30:40, 7.43s/step, loss=1.565, acc=0.658, acc_len=1.55, thru=17.8, I=26.0, T=31.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:24,937] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:30,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:36:31,622] loop.py:45 INFO TRAIN_STEP step=4199 loss=1.578589 Training: 23%|██▎ | 4199/18018 [9:06:17<27:51:04, 7.26s/step, loss=1.579, acc=0.656, acc_len=1.55, thru=17.1, I=23.2, T=30.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:31,837] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:36:37,637] loop.py:380 INFO TIMING step=4200: step=3.904s data=1.765s compute=3.868s [fwd=2.753s bwd=1.100s opt=0.003s] dispatch=2.101s [2026-10-18 22:36:37,638] loop.py:45 INFO TRAIN_STEP step=4200 loss=1.711600 Training: 23%|██▎ | 4200/18018 [9:06:23<26:25:22, 6.88s/step, loss=1.712, acc=0.631, acc_len=1.47, thru=18.5, I=23.1, T=32.8, wait=2.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:43,498] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:36:44,901] loop.py:45 INFO TRAIN_STEP step=4201 loss=1.659623 Training: 23%|██▎ | 4201/18018 [9:06:30<26:51:23, 7.00s/step, loss=1.660, acc=0.645, acc_len=1.51, thru=16.9, I=19.2, T=32.6, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:36:44,897] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4200: forward=2761.0ms backward=1112.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:45,013] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:49,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:36:50,477] loop.py:45 INFO TRAIN_STEP step=4202 loss=1.652374 Training: 23%|██▎ | 4202/18018 [9:06:36<25:13:05, 6.57s/step, loss=1.652, acc=0.644, acc_len=1.51, thru=21.3, I=26.4, T=34.4, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:50,984] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:36:58,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:36:59,796] loop.py:45 INFO TRAIN_STEP step=4203 loss=1.435447 Training: 23%|██▎ | 4203/18018 [9:06:45<28:22:45, 7.40s/step, loss=1.435, acc=0.677, acc_len=1.62, thru=14.5, I=16.1, T=28.6, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:00,210] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:06,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:37:08,522] loop.py:45 INFO TRAIN_STEP step=4204 loss=1.701571 Training: 23%|██▎ | 4204/18018 [9:06:54<29:54:29, 7.79s/step, loss=1.702, acc=0.636, acc_len=1.48, thru=13.4, I=17.6, T=25.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:08,893] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:14,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:37:15,328] loop.py:380 INFO TIMING step=4205: step=3.685s data=1.587s compute=3.650s [fwd=2.563s bwd=1.072s opt=0.003s] dispatch=3.113s [2026-10-18 22:37:15,328] loop.py:45 INFO TRAIN_STEP step=4205 loss=1.654628 Training: 23%|██▎ | 4205/18018 [9:07:01<28:46:02, 7.50s/step, loss=1.655, acc=0.643, acc_len=1.50, thru=16.3, I=23.1, T=34.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:15,668] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:37:23,296] loop.py:45 INFO TRAIN_STEP step=4206 loss=1.540220 Training: 23%|██▎ | 4206/18018 [9:07:09<29:18:30, 7.64s/step, loss=1.540, acc=0.666, acc_len=1.58, thru=15.3, I=18.8, T=33.6, wait=4.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:37:23,290] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4205: forward=2648.0ms backward=1113.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:30,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:37:31,358] loop.py:45 INFO TRAIN_STEP step=4207 loss=1.623093 Training: 23%|██▎ | 4207/18018 [9:07:17<29:47:38, 7.77s/step, loss=1.623, acc=0.648, acc_len=1.52, thru=15.8, I=16.8, T=30.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:31,439] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:36,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:37:38,416] loop.py:45 INFO TRAIN_STEP step=4208 loss=1.537017 Training: 23%|██▎ | 4208/18018 [9:07:24<28:58:31, 7.55s/step, loss=1.537, acc=0.662, acc_len=1.56, thru=16.7, I=19.4, T=32.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:38,854] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:42,903] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:37:44,835] loop.py:45 INFO TRAIN_STEP step=4209 loss=1.615650 Training: 23%|██▎ | 4209/18018 [9:07:30<27:40:02, 7.21s/step, loss=1.616, acc=0.653, acc_len=1.54, thru=18.6, I=28.6, T=32.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:44,927] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:51,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:37:52,774] loop.py:380 INFO TIMING step=4210: step=4.595s data=2.173s compute=4.566s [fwd=3.153s bwd=1.398s opt=0.003s] dispatch=3.334s [2026-10-18 22:37:52,774] loop.py:45 INFO TRAIN_STEP step=4210 loss=1.562675 Training: 23%|██▎ | 4210/18018 [9:07:38<28:30:02, 7.43s/step, loss=1.563, acc=0.659, acc_len=1.55, thru=13.4, I=21.7, T=27.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:53,053] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:37:57,745] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:37:59,996] loop.py:45 INFO TRAIN_STEP step=4211 loss=1.629209 Training: 23%|██▎ | 4211/18018 [9:07:46<28:15:40, 7.37s/step, loss=1.629, acc=0.650, acc_len=1.53, thru=15.7, I=26.2, T=28.7, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:37:59,991] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4210: forward=3151.0ms backward=1259.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:38:00,261] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 22:38:06,742] loop.py:45 INFO TRAIN_STEP step=4212 loss=1.524984 Training: 23%|██▎ | 4212/18018 [9:07:52<27:32:26, 7.18s/step, loss=1.525, acc=0.670, acc_len=1.59, thru=16.2, I=22.7, T=32.2, wait=2.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:38:12,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:38:13,522] loop.py:45 INFO TRAIN_STEP step=4213 loss=1.653359 Training: 23%|██▎ | 4213/18018 [9:07:59<27:04:37, 7.06s/step, loss=1.653, acc=0.646, acc_len=1.51, thru=18.6, I=21.1, T=30.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:38:13,607] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:38:19,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:38:21,312] loop.py:45 INFO TRAIN_STEP step=4214 loss=1.703813 Training: 23%|██▎ | 4214/18018 [9:08:07<27:54:45, 7.28s/step, loss=1.704, acc=0.638, acc_len=1.48, thru=14.1, I=17.9, T=30.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:38:21,744] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:38:29,935] loop.py:380 INFO TIMING step=4215: step=4.367s data=2.013s compute=4.334s [fwd=3.146s bwd=1.174s opt=0.003s] dispatch=4.246s [2026-10-18 22:38:29,935] loop.py:45 INFO TRAIN_STEP step=4215 loss=1.510262 Training: 23%|██▎ | 4215/18018 [9:08:16<29:27:22, 7.68s/step, loss=1.510, acc=0.663, acc_len=1.57, thru=15.2, I=14.9, T=29.3, wait=4.2s, pool=60, epoch=1/2][2026-10-18 22:38:37,083] loop.py:45 INFO TRAIN_STEP step=4216 loss=1.560212 Training: 23%|██▎ | 4216/18018 [9:08:23<28:50:25, 7.52s/step, loss=1.560, acc=0.663, acc_len=1.57, thru=15.8, I=14.9, T=34.6, wait=3.4s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:38:37,078] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4215: forward=2623.1ms backward=1035.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:38:44,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:38:45,849] loop.py:45 INFO TRAIN_STEP step=4217 loss=1.542685 Training: 23%|██▎ | 4217/18018 [9:08:31<30:16:09, 7.90s/step, loss=1.543, acc=0.663, acc_len=1.57, thru=17.4, I=18.9, T=31.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:38:45,887] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:38:52,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:38:53,667] loop.py:45 INFO TRAIN_STEP step=4218 loss=1.540198 Training: 23%|██▎ | 4218/18018 [9:08:39<30:10:36, 7.87s/step, loss=1.540, acc=0.662, acc_len=1.57, thru=14.0, I=18.6, T=26.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:38:54,151] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:00,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:39:03,464] loop.py:45 INFO TRAIN_STEP step=4219 loss=1.488917 Training: 23%|██▎ | 4219/18018 [9:08:49<32:23:10, 8.45s/step, loss=1.489, acc=0.674, acc_len=1.61, thru=13.8, I=18.6, T=24.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:03,925] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:08,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:39:10,818] loop.py:380 INFO TIMING step=4220: step=4.773s data=2.517s compute=4.743s [fwd=3.535s bwd=1.194s opt=0.003s] dispatch=2.572s [2026-10-18 22:39:10,818] loop.py:45 INFO TRAIN_STEP step=4220 loss=1.615459 Training: 23%|██▎ | 4220/18018 [9:08:56<31:07:30, 8.12s/step, loss=1.615, acc=0.649, acc_len=1.52, thru=13.6, I=24.6, T=26.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:10,882] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:15,457] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:39:18,114] loop.py:45 INFO TRAIN_STEP step=4221 loss=1.602832 Training: 23%|██▎ | 4221/18018 [9:09:04<30:10:28, 7.87s/step, loss=1.603, acc=0.654, acc_len=1.54, thru=14.8, I=24.6, T=25.3, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:39:18,109] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4220: forward=3630.9ms backward=1381.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:18,479] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:24,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:39:26,659] loop.py:45 INFO TRAIN_STEP step=4222 loss=1.604912 Training: 23%|██▎ | 4222/18018 [9:09:12<30:56:40, 8.07s/step, loss=1.605, acc=0.647, acc_len=1.52, thru=15.7, I=17.5, T=27.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:26,995] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:39:34,810] loop.py:45 INFO TRAIN_STEP step=4223 loss=1.718620 Training: 23%|██▎ | 4223/18018 [9:09:20<31:01:47, 8.10s/step, loss=1.719, acc=0.635, acc_len=1.47, thru=13.4, I=17.5, T=31.7, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:43,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:39:44,343] loop.py:45 INFO TRAIN_STEP step=4224 loss=1.673933 Training: 23%|██▎ | 4224/18018 [9:09:30<32:40:34, 8.53s/step, loss=1.674, acc=0.644, acc_len=1.51, thru=14.2, I=13.1, T=31.4, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:44,401] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 22:39:52,678] loop.py:380 INFO TIMING step=4225: step=4.634s data=2.258s compute=4.596s [fwd=3.152s bwd=1.430s opt=0.003s] dispatch=3.694s [2026-10-18 22:39:52,678] loop.py:45 INFO TRAIN_STEP step=4225 loss=1.520376 Training: 23%|██▎ | 4225/18018 [9:09:38<32:27:10, 8.47s/step, loss=1.520, acc=0.669, acc_len=1.59, thru=15.8, I=17.4, T=27.6, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:39:58,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:40:01,393] loop.py:45 INFO TRAIN_STEP step=4226 loss=1.585781 Training: 23%|██▎ | 4226/18018 [9:09:47<32:44:02, 8.54s/step, loss=1.586, acc=0.655, acc_len=1.53, thru=16.1, I=19.1, T=23.3, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:40:01,388] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4225: forward=3762.0ms backward=1673.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:40:01,597] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 22:40:07,827] loop.py:45 INFO TRAIN_STEP step=4227 loss=1.681456 Training: 23%|██▎ | 4227/18018 [9:09:53<30:18:13, 7.91s/step, loss=1.681, acc=0.637, acc_len=1.48, thru=14.1, I=19.5, T=33.6, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:40:16,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:40:18,146] loop.py:45 INFO TRAIN_STEP step=4228 loss=1.535256 Training: 23%|██▎ | 4228/18018 [9:10:04<33:04:23, 8.63s/step, loss=1.535, acc=0.662, acc_len=1.57, thru=13.2, I=16.0, T=27.9, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:40:18,234] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:40:23,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:40:24,673] loop.py:45 INFO TRAIN_STEP step=4229 loss=1.593947 Training: 23%|██▎ | 4229/18018 [9:10:10<30:38:50, 8.00s/step, loss=1.594, acc=0.655, acc_len=1.54, thru=17.6, I=24.7, T=32.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:40:24,883] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:40:31,233] loop.py:380 INFO TIMING step=4230: step=4.277s data=1.962s compute=4.242s [fwd=3.033s bwd=1.194s opt=0.003s] dispatch=2.273s [2026-10-18 22:40:31,234] loop.py:45 INFO TRAIN_STEP step=4230 loss=1.665067 Training: 23%|██▎ | 4230/18018 [9:10:17<28:59:27, 7.57s/step, loss=1.665, acc=0.640, acc_len=1.50, thru=17.2, I=23.7, T=29.9, wait=2.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:40:36,403] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:40:38,247] loop.py:45 INFO TRAIN_STEP step=4231 loss=1.658863 Training: 23%|██▎ | 4231/18018 [9:10:24<28:20:54, 7.40s/step, loss=1.659, acc=0.644, acc_len=1.51, thru=17.7, I=20.3, T=29.8, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:40:38,242] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4230: forward=2926.7ms backward=1309.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:40:38,418] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:40:45,638] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:40:46,669] loop.py:45 INFO TRAIN_STEP step=4232 loss=1.524709 Training: 23%|██▎ | 4232/18018 [9:10:32<29:31:00, 7.71s/step, loss=1.525, acc=0.664, acc_len=1.57, thru=15.5, I=17.4, T=27.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:40:47,149] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:40:54,327] loop.py:45 INFO TRAIN_STEP step=4233 loss=1.561355 Training: 23%|██▎ | 4233/18018 [9:10:40<29:27:42, 7.69s/step, loss=1.561, acc=0.662, acc_len=1.56, thru=14.1, I=15.9, T=30.9, wait=3.5s, pool=48, epoch=1/2][2026-10-18 22:41:03,384] loop.py:45 INFO TRAIN_STEP step=4234 loss=1.569004 Training: 23%|██▎ | 4234/18018 [9:10:49<31:01:30, 8.10s/step, loss=1.569, acc=0.658, acc_len=1.55, thru=14.6, I=15.6, T=34.4, wait=5.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:09,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:41:11,849] loop.py:380 INFO TIMING step=4235: step=4.326s data=1.920s compute=4.292s [fwd=2.936s bwd=1.340s opt=0.003s] dispatch=4.126s [2026-10-18 22:41:11,849] loop.py:45 INFO TRAIN_STEP step=4235 loss=1.673609 Training: 24%|██▎ | 4235/18018 [9:10:57<31:26:07, 8.21s/step, loss=1.674, acc=0.643, acc_len=1.50, thru=15.4, I=16.8, T=29.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:12,171] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:18,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:41:18,710] loop.py:45 INFO TRAIN_STEP step=4236 loss=1.626905 Training: 24%|██▎ | 4236/18018 [9:11:04<29:53:04, 7.81s/step, loss=1.627, acc=0.642, acc_len=1.50, thru=15.7, I=19.9, T=31.9, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:41:18,705] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4235: forward=2824.0ms backward=1150.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:19,004] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:26,179] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:41:27,574] loop.py:45 INFO TRAIN_STEP step=4237 loss=1.763541 Training: 24%|██▎ | 4237/18018 [9:11:13<31:05:51, 8.12s/step, loss=1.764, acc=0.631, acc_len=1.46, thru=15.8, I=19.9, T=33.9, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:27,696] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:35,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:41:36,801] loop.py:45 INFO TRAIN_STEP step=4238 loss=1.623479 Training: 24%|██▎ | 4238/18018 [9:11:22<32:21:43, 8.45s/step, loss=1.623, acc=0.648, acc_len=1.52, thru=14.7, I=13.5, T=30.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:36,957] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:42,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:41:44,089] loop.py:45 INFO TRAIN_STEP step=4239 loss=1.578898 Training: 24%|██▎ | 4239/18018 [9:11:30<31:01:03, 8.10s/step, loss=1.579, acc=0.656, acc_len=1.54, thru=16.0, I=20.9, T=29.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:44,116] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-18 22:41:49,302] loop.py:380 INFO TIMING step=4240: step=3.558s data=1.419s compute=3.530s [fwd=2.437s bwd=1.078s opt=0.003s] dispatch=1.648s [2026-10-18 22:41:49,303] loop.py:45 INFO TRAIN_STEP step=4240 loss=1.664270 Training: 24%|██▎ | 4240/18018 [9:11:35<27:41:56, 7.24s/step, loss=1.664, acc=0.640, acc_len=1.49, thru=19.2, I=26.0, T=36.0, wait=1.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:57,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:41:57,681] loop.py:45 INFO TRAIN_STEP step=4241 loss=1.542470 Training: 24%|██▎ | 4241/18018 [9:11:43<29:00:19, 7.58s/step, loss=1.542, acc=0.661, acc_len=1.56, thru=17.1, I=15.9, T=34.9, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:41:57,676] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4240: forward=2591.3ms backward=1030.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:41:57,805] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:42:03,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:42:04,160] loop.py:45 INFO TRAIN_STEP step=4242 loss=1.617807 Training: 24%|██▎ | 4242/18018 [9:11:50<27:44:24, 7.25s/step, loss=1.618, acc=0.652, acc_len=1.53, thru=18.8, I=23.2, T=32.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:42:04,466] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:42:10,872] loop.py:45 INFO TRAIN_STEP step=4243 loss=1.495026 Training: 24%|██▎ | 4243/18018 [9:11:56<27:07:18, 7.09s/step, loss=1.495, acc=0.668, acc_len=1.59, thru=17.5, I=20.9, T=29.7, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:42:18,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:42:20,123] loop.py:45 INFO TRAIN_STEP step=4244 loss=1.742503 Training: 24%|██▎ | 4244/18018 [9:12:06<29:36:09, 7.74s/step, loss=1.743, acc=0.632, acc_len=1.47, thru=14.1, I=15.7, T=29.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:42:20,552] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:42:27,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:42:29,115] loop.py:380 INFO TIMING step=4245: step=4.015s data=1.797s compute=3.986s [fwd=2.827s bwd=1.144s opt=0.003s] dispatch=4.968s [2026-10-18 22:42:29,115] loop.py:45 INFO TRAIN_STEP step=4245 loss=1.585799 Training: 24%|██▎ | 4245/18018 [9:12:15<31:02:41, 8.11s/step, loss=1.586, acc=0.656, acc_len=1.55, thru=14.9, I=17.9, T=31.9, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:42:29,153] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 22:42:36,093] loop.py:45 INFO TRAIN_STEP step=4246 loss=1.533431 Training: 24%|██▎ | 4246/18018 [9:12:22<29:44:05, 7.77s/step, loss=1.533, acc=0.662, acc_len=1.56, thru=17.0, I=20.1, T=31.9, wait=3.0s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:42:36,087] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4245: forward=2832.2ms backward=1128.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:42:45,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:42:45,283] loop.py:45 INFO TRAIN_STEP step=4247 loss=1.664393 Training: 24%|██▎ | 4247/18018 [9:12:31<31:21:35, 8.20s/step, loss=1.664, acc=0.646, acc_len=1.51, thru=14.8, I=15.3, T=26.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:42:45,627] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:42:53,776] loop.py:45 INFO TRAIN_STEP step=4248 loss=1.723994 Training: 24%|██▎ | 4248/18018 [9:12:39<31:41:43, 8.29s/step, loss=1.724, acc=0.633, acc_len=1.47, thru=13.9, I=16.3, T=26.4, wait=3.6s, pool=52, epoch=1/2][2026-10-18 22:43:03,605] loop.py:45 INFO TRAIN_STEP step=4249 loss=1.597729 Training: 24%|██▎ | 4249/18018 [9:12:49<33:27:51, 8.75s/step, loss=1.598, acc=0.650, acc_len=1.53, thru=12.4, I=10.8, T=29.7, wait=5.5s, pool=44, epoch=1/2][2026-10-18 22:43:12,337] loop.py:380 INFO TIMING step=4250: step=3.802s data=1.608s compute=3.764s [fwd=2.647s bwd=1.103s opt=0.003s] dispatch=4.920s [2026-10-18 22:43:12,338] loop.py:45 INFO TRAIN_STEP step=4250 loss=1.770281 Training: 24%|██▎ | 4250/18018 [9:12:58<33:26:40, 8.74s/step, loss=1.770, acc=0.625, acc_len=1.45, thru=16.2, I=15.2, T=33.7, wait=4.9s, pool=60, epoch=1/2][2026-10-18 22:43:12,344] loop.py:419 INFO Saving checkpoint at step 4250... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:13,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 22:43:15,825] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0004251 (TrainerActor pid=9310) [2026-10-18 22:43:15,832] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0003501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:17,271] inference_manager.py:408 INFO Pool capacity available after 4.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:25,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 22:43:15,824] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0004251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 22:43:16,795] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0003501 [2026-10-18 22:43:25,949] loop.py:45 INFO TRAIN_STEP step=4251 loss=1.716376 Training: 24%|██▎ | 4251/18018 [9:13:12<39:01:14, 10.20s/step, loss=1.716, acc=0.633, acc_len=1.47, thru=14.8, I=16.9, T=26.8, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:43:25,943] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4250: forward=3172.2ms backward=1546.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:26,388] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:33,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:43:33,761] loop.py:45 INFO TRAIN_STEP step=4252 loss=1.575647 Training: 24%|██▎ | 4252/18018 [9:13:19<36:16:35, 9.49s/step, loss=1.576, acc=0.656, acc_len=1.55, thru=14.9, I=15.7, T=34.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:34,267] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:39,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:43:40,276] loop.py:45 INFO TRAIN_STEP step=4253 loss=1.604497 Training: 24%|██▎ | 4253/18018 [9:13:26<32:51:45, 8.59s/step, loss=1.604, acc=0.656, acc_len=1.54, thru=19.5, I=23.4, T=31.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:40,771] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:45,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:43:46,541] loop.py:45 INFO TRAIN_STEP step=4254 loss=1.441425 Training: 24%|██▎ | 4254/18018 [9:13:32<30:11:25, 7.90s/step, loss=1.441, acc=0.679, acc_len=1.62, thru=19.1, I=24.5, T=32.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:46,755] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:53,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:43:53,389] loop.py:380 INFO TIMING step=4255: step=4.054s data=1.790s compute=4.026s [fwd=2.827s bwd=1.184s opt=0.003s] dispatch=2.783s [2026-10-18 22:43:53,389] loop.py:45 INFO TRAIN_STEP step=4255 loss=1.735167 Training: 24%|██▎ | 4255/18018 [9:13:39<28:59:03, 7.58s/step, loss=1.735, acc=0.631, acc_len=1.47, thru=17.5, I=21.0, T=31.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:53,759] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:43:59,416] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:44:00,325] loop.py:45 INFO TRAIN_STEP step=4256 loss=1.490907 Training: 24%|██▎ | 4256/18018 [9:13:46<28:14:31, 7.39s/step, loss=1.491, acc=0.672, acc_len=1.60, thru=18.1, I=20.4, T=27.8, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:44:00,320] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4255: forward=3078.1ms backward=1492.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:00,430] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:06,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:44:06,630] loop.py:45 INFO TRAIN_STEP step=4257 loss=1.540377 Training: 24%|██▎ | 4257/18018 [9:13:52<26:59:51, 7.06s/step, loss=1.540, acc=0.664, acc_len=1.57, thru=17.3, I=19.0, T=31.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:06,938] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:11,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:44:13,674] loop.py:45 INFO TRAIN_STEP step=4258 loss=1.543846 Training: 24%|██▎ | 4258/18018 [9:13:59<26:58:29, 7.06s/step, loss=1.544, acc=0.663, acc_len=1.57, thru=17.0, I=21.3, T=29.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:13,855] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:20,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:44:21,559] loop.py:45 INFO TRAIN_STEP step=4259 loss=1.659024 Training: 24%|██▎ | 4259/18018 [9:14:07<27:55:17, 7.31s/step, loss=1.659, acc=0.642, acc_len=1.50, thru=13.6, I=20.4, T=29.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:21,643] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:26,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:44:28,218] loop.py:380 INFO TIMING step=4260: step=4.397s data=2.159s compute=4.365s [fwd=2.972s bwd=1.378s opt=0.003s] dispatch=2.254s [2026-10-18 22:44:28,218] loop.py:45 INFO TRAIN_STEP step=4260 loss=1.632093 Training: 24%|██▎ | 4260/18018 [9:14:14<27:10:49, 7.11s/step, loss=1.632, acc=0.645, acc_len=1.51, thru=17.9, I=26.5, T=29.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:28,528] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:34,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:44:34,860] loop.py:45 INFO TRAIN_STEP step=4261 loss=1.551984 Training: 24%|██▎ | 4261/18018 [9:14:20<26:38:20, 6.97s/step, loss=1.552, acc=0.659, acc_len=1.55, thru=16.4, I=22.8, T=27.8, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:44:34,855] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4260: forward=3159.7ms backward=1385.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:35,156] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:39,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:44:42,359] loop.py:45 INFO TRAIN_STEP step=4262 loss=1.693183 Training: 24%|██▎ | 4262/18018 [9:14:28<27:14:35, 7.13s/step, loss=1.693, acc=0.642, acc_len=1.50, thru=15.6, I=21.1, T=28.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:42,472] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:48,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:44:49,864] loop.py:45 INFO TRAIN_STEP step=4263 loss=1.475493 Training: 24%|██▎ | 4263/18018 [9:14:35<27:40:12, 7.24s/step, loss=1.475, acc=0.672, acc_len=1.61, thru=13.7, I=23.0, T=32.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:44:49,898] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 22:44:56,012] loop.py:45 INFO TRAIN_STEP step=4264 loss=1.668810 Training: 24%|██▎ | 4264/18018 [9:14:42<26:24:52, 6.91s/step, loss=1.669, acc=0.641, acc_len=1.50, thru=19.4, I=22.8, T=37.2, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:03,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:45:03,141] loop.py:380 INFO TIMING step=4265: step=4.007s data=1.807s compute=3.980s [fwd=2.836s bwd=1.129s opt=0.003s] dispatch=3.112s [2026-10-18 22:45:03,141] loop.py:45 INFO TRAIN_STEP step=4265 loss=1.638678 Training: 24%|██▎ | 4265/18018 [9:14:49<26:39:23, 6.98s/step, loss=1.639, acc=0.647, acc_len=1.51, thru=18.6, I=19.7, T=31.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:03,512] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:09,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:45:10,315] loop.py:45 INFO TRAIN_STEP step=4266 loss=1.683155 Training: 24%|██▎ | 4266/18018 [9:14:56<26:52:49, 7.04s/step, loss=1.683, acc=0.645, acc_len=1.51, thru=18.2, I=18.5, T=32.0, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:45:10,311] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4265: forward=2788.8ms backward=1155.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:10,767] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:18,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:45:19,115] loop.py:45 INFO TRAIN_STEP step=4267 loss=1.482556 Training: 24%|██▎ | 4267/18018 [9:15:05<28:54:02, 7.57s/step, loss=1.483, acc=0.673, acc_len=1.60, thru=13.4, I=15.1, T=30.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:19,476] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:26,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:45:26,589] loop.py:45 INFO TRAIN_STEP step=4268 loss=1.655658 Training: 24%|██▎ | 4268/18018 [9:15:12<28:47:27, 7.54s/step, loss=1.656, acc=0.644, acc_len=1.51, thru=16.6, I=19.0, T=28.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:26,716] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:32,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:45:33,505] loop.py:45 INFO TRAIN_STEP step=4269 loss=1.649842 Training: 24%|██▎ | 4269/18018 [9:15:19<28:04:35, 7.35s/step, loss=1.650, acc=0.646, acc_len=1.51, thru=18.3, I=18.0, T=34.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:33,616] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:39,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:45:39,931] loop.py:380 INFO TIMING step=4270: step=3.940s data=1.709s compute=3.903s [fwd=2.732s bwd=1.155s opt=0.003s] dispatch=2.477s [2026-10-18 22:45:39,932] loop.py:45 INFO TRAIN_STEP step=4270 loss=1.611004 Training: 24%|██▎ | 4270/18018 [9:15:26<27:01:00, 7.07s/step, loss=1.611, acc=0.653, acc_len=1.54, thru=19.0, I=23.5, T=32.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:40,367] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:47,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:45:47,190] loop.py:45 INFO TRAIN_STEP step=4271 loss=1.582871 Training: 24%|██▎ | 4271/18018 [9:15:33<27:13:17, 7.13s/step, loss=1.583, acc=0.658, acc_len=1.56, thru=16.9, I=17.2, T=34.7, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:45:47,185] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4270: forward=2626.2ms backward=1019.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:45:47,620] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:45:54,842] loop.py:45 INFO TRAIN_STEP step=4272 loss=1.684173 Training: 24%|██▎ | 4272/18018 [9:15:40<27:49:23, 7.29s/step, loss=1.684, acc=0.635, acc_len=1.48, thru=16.2, I=17.4, T=28.7, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:46:01,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:46:02,841] loop.py:45 INFO TRAIN_STEP step=4273 loss=1.641584 Training: 24%|██▎ | 4273/18018 [9:15:48<28:38:09, 7.50s/step, loss=1.642, acc=0.646, acc_len=1.52, thru=16.4, I=17.1, T=29.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:46:03,022] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:46:10,561] loop.py:45 INFO TRAIN_STEP step=4274 loss=1.580541 Training: 24%|██▎ | 4274/18018 [9:15:56<28:53:05, 7.57s/step, loss=1.581, acc=0.659, acc_len=1.55, thru=12.5, I=15.7, T=27.0, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:46:16,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:46:18,540] loop.py:380 INFO TIMING step=4275: step=4.746s data=2.236s compute=4.710s [fwd=3.313s bwd=1.381s opt=0.003s] dispatch=3.225s [2026-10-18 22:46:18,540] loop.py:45 INFO TRAIN_STEP step=4275 loss=1.666312 Training: 24%|██▎ | 4275/18018 [9:16:04<29:21:26, 7.69s/step, loss=1.666, acc=0.644, acc_len=1.51, thru=18.3, I=17.3, T=27.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:46:18,867] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:46:25,631] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:46:26,876] loop.py:45 INFO TRAIN_STEP step=4276 loss=1.642136 (TrainerActor pid=9310) [2026-10-18 22:46:26,870] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4275: forward=2840.1ms backward=1355.9ms Training: 24%|██▎ | 4276/18018 [9:16:12<30:05:46, 7.88s/step, loss=1.642, acc=0.642, acc_len=1.50, thru=16.0, I=19.4, T=30.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:46:27,139] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:46:34,597] loop.py:45 INFO TRAIN_STEP step=4277 loss=1.677561 Training: 24%|██▎ | 4277/18018 [9:16:20<29:54:20, 7.83s/step, loss=1.678, acc=0.642, acc_len=1.50, thru=14.6, I=17.1, T=32.4, wait=3.8s, pool=44, epoch=1/2][2026-10-18 22:46:44,260] loop.py:45 INFO TRAIN_STEP step=4278 loss=1.729189 Training: 24%|██▎ | 4278/18018 [9:16:30<31:59:37, 8.38s/step, loss=1.729, acc=0.637, acc_len=1.48, thru=14.2, I=13.8, T=27.8, wait=5.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:46:52,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:46:53,110] loop.py:45 INFO TRAIN_STEP step=4279 loss=1.639417 Training: 24%|██▎ | 4279/18018 [9:16:39<32:31:37, 8.52s/step, loss=1.639, acc=0.647, acc_len=1.52, thru=16.3, I=17.1, T=30.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:46:53,592] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:47:00,615] loop.py:380 INFO TIMING step=4280: step=4.107s data=1.869s compute=4.075s [fwd=2.839s bwd=1.221s opt=0.003s] dispatch=3.390s [2026-10-18 22:47:00,615] loop.py:45 INFO TRAIN_STEP step=4280 loss=1.548242 Training: 24%|██▍ | 4280/18018 [9:16:46<31:21:30, 8.22s/step, loss=1.548, acc=0.665, acc_len=1.58, thru=16.5, I=20.1, T=31.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:00,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:01,088] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:05,339] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:47:07,253] loop.py:45 INFO TRAIN_STEP step=4281 loss=1.546417 Training: 24%|██▍ | 4281/18018 [9:16:53<29:33:04, 7.74s/step, loss=1.546, acc=0.662, acc_len=1.57, thru=17.7, I=21.8, T=32.2, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:47:07,248] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4280: forward=2747.9ms backward=1160.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:07,367] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-18 22:47:14,088] loop.py:45 INFO TRAIN_STEP step=4282 loss=1.596742 Training: 24%|██▍ | 4282/18018 [9:17:00<28:30:21, 7.47s/step, loss=1.597, acc=0.648, acc_len=1.52, thru=13.9, I=16.8, T=33.3, wait=3.0s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:22,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:47:24,637] loop.py:45 INFO TRAIN_STEP step=4283 loss=1.769876 Training: 24%|██▍ | 4283/18018 [9:17:10<32:01:46, 8.40s/step, loss=1.770, acc=0.628, acc_len=1.46, thru=14.9, I=18.3, T=26.8, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:24,897] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:30,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:47:32,212] loop.py:45 INFO TRAIN_STEP step=4284 loss=1.536079 Training: 24%|██▍ | 4284/18018 [9:17:18<31:05:07, 8.15s/step, loss=1.536, acc=0.661, acc_len=1.56, thru=14.9, I=22.8, T=29.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:32,704] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:47:40,075] loop.py:380 INFO TIMING step=4285: step=4.667s data=2.188s compute=4.631s [fwd=3.182s bwd=1.434s opt=0.003s] dispatch=3.188s [2026-10-18 22:47:40,076] loop.py:45 INFO TRAIN_STEP step=4285 loss=1.558863 Training: 24%|██▍ | 4285/18018 [9:17:26<30:45:31, 8.06s/step, loss=1.559, acc=0.658, acc_len=1.56, thru=13.3, I=16.9, T=27.4, wait=3.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:45,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:47:47,842] loop.py:45 INFO TRAIN_STEP step=4286 loss=1.655295 Training: 24%|██▍ | 4286/18018 [9:17:33<30:24:59, 7.97s/step, loss=1.655, acc=0.647, acc_len=1.52, thru=17.3, I=19.6, T=27.8, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:47:47,836] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4285: forward=3181.0ms backward=1371.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:47:48,193] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-18 22:47:54,199] loop.py:45 INFO TRAIN_STEP step=4287 loss=1.586765 Training: 24%|██▍ | 4287/18018 [9:17:40<28:33:49, 7.49s/step, loss=1.587, acc=0.650, acc_len=1.52, thru=14.5, I=21.2, T=34.0, wait=2.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:05,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:48:06,338] loop.py:45 INFO TRAIN_STEP step=4288 loss=1.635504 Training: 24%|██▍ | 4288/18018 [9:17:52<33:53:03, 8.88s/step, loss=1.636, acc=0.649, acc_len=1.53, thru=12.5, I=12.3, T=28.0, wait=7.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:06,536] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:12,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:48:13,777] loop.py:45 INFO TRAIN_STEP step=4289 loss=1.603818 Training: 24%|██▍ | 4289/18018 [9:17:59<32:13:35, 8.45s/step, loss=1.604, acc=0.653, acc_len=1.54, thru=15.7, I=20.4, T=30.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:14,159] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:18,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:48:20,345] loop.py:380 INFO TIMING step=4290: step=3.944s data=1.767s compute=3.915s [fwd=2.769s bwd=1.131s opt=0.003s] dispatch=2.616s [2026-10-18 22:48:20,346] loop.py:45 INFO TRAIN_STEP step=4290 loss=1.652166 Training: 24%|██▍ | 4290/18018 [9:18:06<30:04:24, 7.89s/step, loss=1.652, acc=0.644, acc_len=1.51, thru=18.6, I=24.6, T=32.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:20,491] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:25,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:48:26,488] loop.py:45 INFO TRAIN_STEP step=4291 loss=1.598015 Training: 24%|██▍ | 4291/18018 [9:18:12<28:04:28, 7.36s/step, loss=1.598, acc=0.652, acc_len=1.53, thru=19.2, I=25.4, T=32.9, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:48:26,483] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4290: forward=2730.0ms backward=1112.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:26,739] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:31,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 22:48:33,958] loop.py:45 INFO TRAIN_STEP step=4292 loss=1.553676 Training: 24%|██▍ | 4292/18018 [9:18:20<28:11:33, 7.39s/step, loss=1.554, acc=0.660, acc_len=1.56, thru=17.7, I=25.4, T=25.9, wait=2.5s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:34,363] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=4/64 [2026-10-18 22:48:40,591] loop.py:45 INFO TRAIN_STEP step=4293 loss=1.648149 Training: 24%|██▍ | 4293/18018 [9:18:26<27:19:20, 7.17s/step, loss=1.648, acc=0.647, acc_len=1.52, thru=13.6, I=20.9, T=35.7, wait=3.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:48,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:48:48,954] loop.py:45 INFO TRAIN_STEP step=4294 loss=1.578326 Training: 24%|██▍ | 4294/18018 [9:18:35<28:41:18, 7.53s/step, loss=1.578, acc=0.657, acc_len=1.55, thru=16.8, I=15.8, T=30.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:49,268] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:55,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:48:56,072] loop.py:380 INFO TIMING step=4295: step=4.073s data=1.715s compute=4.028s [fwd=2.804s bwd=1.209s opt=0.003s] dispatch=3.036s [2026-10-18 22:48:56,072] loop.py:45 INFO TRAIN_STEP step=4295 loss=1.663742 Training: 24%|██▍ | 4295/18018 [9:18:42<28:13:24, 7.40s/step, loss=1.664, acc=0.643, acc_len=1.50, thru=17.2, I=17.9, T=31.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:48:56,169] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 [2026-10-18 22:49:03,527] loop.py:45 INFO TRAIN_STEP step=4296 loss=1.552374 Training: 24%|██▍ | 4296/18018 [9:18:49<28:16:42, 7.42s/step, loss=1.552, acc=0.665, acc_len=1.57, thru=16.9, I=17.5, T=30.7, wait=3.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:49:03,522] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4295: forward=2925.1ms backward=1189.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:10,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:49:11,737] loop.py:45 INFO TRAIN_STEP step=4297 loss=1.696443 Training: 24%|██▍ | 4297/18018 [9:18:57<29:10:43, 7.66s/step, loss=1.696, acc=0.638, acc_len=1.49, thru=16.6, I=18.9, T=31.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:12,019] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:17,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:49:17,990] loop.py:45 INFO TRAIN_STEP step=4298 loss=1.696682 Training: 24%|██▍ | 4298/18018 [9:19:04<27:34:31, 7.24s/step, loss=1.697, acc=0.634, acc_len=1.48, thru=19.0, I=23.1, T=35.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:18,201] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:25,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:49:25,277] loop.py:45 INFO TRAIN_STEP step=4299 loss=1.587415 Training: 24%|██▍ | 4299/18018 [9:19:11<27:37:47, 7.25s/step, loss=1.587, acc=0.651, acc_len=1.53, thru=17.7, I=19.2, T=29.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:25,641] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:31,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:49:33,596] loop.py:380 INFO TIMING step=4300: step=4.091s data=1.799s compute=4.061s [fwd=2.878s bwd=1.168s opt=0.003s] dispatch=4.219s [2026-10-18 22:49:33,596] loop.py:45 INFO TRAIN_STEP step=4300 loss=1.636322 Training: 24%|██▍ | 4300/18018 [9:19:19<28:51:06, 7.57s/step, loss=1.636, acc=0.646, acc_len=1.51, thru=15.8, I=18.8, T=31.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:33,786] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:39,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 22:49:40,475] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4300: forward=2957.9ms backward=1308.4ms [2026-10-18 22:49:40,480] loop.py:45 INFO TRAIN_STEP step=4301 loss=1.496668 Training: 24%|██▍ | 4301/18018 [9:19:26<28:03:51, 7.37s/step, loss=1.497, acc=0.672, acc_len=1.61, thru=16.7, I=24.0, T=29.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:40,787] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:46,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:49:47,410] loop.py:45 INFO TRAIN_STEP step=4302 loss=1.719931 Training: 24%|██▍ | 4302/18018 [9:19:33<27:33:53, 7.23s/step, loss=1.720, acc=0.631, acc_len=1.47, thru=17.0, I=20.5, T=35.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:47,763] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:53,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:49:54,376] loop.py:45 INFO TRAIN_STEP step=4303 loss=1.627856 Training: 24%|██▍ | 4303/18018 [9:19:40<27:15:25, 7.15s/step, loss=1.628, acc=0.648, acc_len=1.52, thru=18.3, I=21.6, T=34.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:49:54,585] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:01,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:50:02,751] loop.py:45 INFO TRAIN_STEP step=4304 loss=1.705835 Training: 24%|██▍ | 4304/18018 [9:19:48<28:38:44, 7.52s/step, loss=1.706, acc=0.639, acc_len=1.49, thru=15.5, I=19.1, T=33.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:03,104] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:07,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:50:08,184] loop.py:380 INFO TIMING step=4305: step=3.281s data=1.347s compute=3.253s [fwd=2.328s bwd=0.910s opt=0.003s] dispatch=2.143s [2026-10-18 22:50:08,184] loop.py:45 INFO TRAIN_STEP step=4305 loss=1.705629 Training: 24%|██▍ | 4305/18018 [9:19:54<26:15:31, 6.89s/step, loss=1.706, acc=0.633, acc_len=1.47, thru=20.5, I=26.1, T=39.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:08,210] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:13,666] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:50:14,650] loop.py:45 INFO TRAIN_STEP step=4306 loss=1.618352 Training: 24%|██▍ | 4306/18018 [9:20:00<25:46:17, 6.77s/step, loss=1.618, acc=0.648, acc_len=1.52, thru=20.9, I=25.2, T=36.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:14,686] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-18 22:50:14,645] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4305: forward=2464.3ms backward=1026.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:19,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:50:20,794] loop.py:45 INFO TRAIN_STEP step=4307 loss=1.693096 Training: 24%|██▍ | 4307/18018 [9:20:06<25:03:22, 6.58s/step, loss=1.693, acc=0.640, acc_len=1.50, thru=20.2, I=24.0, T=33.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:20,906] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:27,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:50:28,087] loop.py:45 INFO TRAIN_STEP step=4308 loss=1.579450 Training: 24%|██▍ | 4308/18018 [9:20:14<25:52:11, 6.79s/step, loss=1.579, acc=0.655, acc_len=1.54, thru=17.3, I=20.4, T=29.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:28,134] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 22:50:36,129] loop.py:45 INFO TRAIN_STEP step=4309 loss=1.556477 Training: 24%|██▍ | 4309/18018 [9:20:22<27:17:39, 7.17s/step, loss=1.556, acc=0.657, acc_len=1.55, thru=14.2, I=13.8, T=30.1, wait=3.8s, pool=48, epoch=1/2][2026-10-18 22:50:45,655] loop.py:380 INFO TIMING step=4310: step=4.347s data=1.924s compute=4.314s [fwd=3.068s bwd=1.232s opt=0.003s] dispatch=5.171s [2026-10-18 22:50:45,655] loop.py:45 INFO TRAIN_STEP step=4310 loss=1.675755 Training: 24%|██▍ | 4310/18018 [9:20:31<29:59:18, 7.88s/step, loss=1.676, acc=0.642, acc_len=1.50, thru=13.7, I=13.5, T=29.4, wait=5.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:50:58,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:51:00,379] loop.py:45 INFO TRAIN_STEP step=4311 loss=1.717833 Training: 24%|██▍ | 4311/18018 [9:20:46<37:48:24, 9.93s/step, loss=1.718, acc=0.639, acc_len=1.49, thru=10.5, I=11.3, T=25.9, wait=9.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:51:00,374] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4310: forward=3794.7ms backward=1101.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:51:00,728] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:51:05,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:51:07,078] loop.py:45 INFO TRAIN_STEP step=4312 loss=1.583704 Training: 24%|██▍ | 4312/18018 [9:20:53<34:06:49, 8.96s/step, loss=1.584, acc=0.654, acc_len=1.54, thru=15.7, I=23.7, T=30.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:51:07,506] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:51:13,859] loop.py:45 INFO TRAIN_STEP step=4313 loss=1.538380 Training: 24%|██▍ | 4313/18018 [9:20:59<31:37:22, 8.31s/step, loss=1.538, acc=0.667, acc_len=1.58, thru=15.9, I=23.6, T=30.2, wait=2.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:51:23,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:51:24,019] loop.py:45 INFO TRAIN_STEP step=4314 loss=1.422328 Training: 24%|██▍ | 4314/18018 [9:21:10<33:44:24, 8.86s/step, loss=1.422, acc=0.683, acc_len=1.65, thru=14.8, I=13.7, T=31.9, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:51:24,109] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 22:51:31,861] loop.py:380 INFO TIMING step=4315: step=4.065s data=1.934s compute=4.025s [fwd=2.873s bwd=1.138s opt=0.003s] dispatch=3.765s [2026-10-18 22:51:31,861] loop.py:45 INFO TRAIN_STEP step=4315 loss=1.703874 Training: 24%|██▍ | 4315/18018 [9:21:17<32:34:19, 8.56s/step, loss=1.704, acc=0.640, acc_len=1.49, thru=15.4, I=17.1, T=31.5, wait=3.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:51:37,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:51:38,656] loop.py:45 INFO TRAIN_STEP step=4316 loss=1.624794 Training: 24%|██▍ | 4316/18018 [9:21:24<30:33:12, 8.03s/step, loss=1.625, acc=0.649, acc_len=1.53, thru=18.5, I=19.6, T=32.4, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:51:38,651] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4315: forward=2663.1ms backward=1221.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:51:38,752] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:51:44,079] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:51:45,039] loop.py:45 INFO TRAIN_STEP step=4317 loss=1.438367 Training: 24%|██▍ | 4317/18018 [9:21:31<28:40:36, 7.53s/step, loss=1.438, acc=0.678, acc_len=1.61, thru=19.3, I=23.1, T=35.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:51:45,089] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-18 22:51:52,177] loop.py:45 INFO TRAIN_STEP step=4318 loss=1.599586 Training: 24%|██▍ | 4318/18018 [9:21:38<28:13:24, 7.42s/step, loss=1.600, acc=0.656, acc_len=1.55, thru=16.8, I=21.1, T=31.2, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:01,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:52:01,460] loop.py:45 INFO TRAIN_STEP step=4319 loss=1.676567 Training: 24%|██▍ | 4319/18018 [9:21:47<30:20:52, 7.98s/step, loss=1.677, acc=0.639, acc_len=1.49, thru=15.3, I=13.8, T=32.4, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:01,772] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:52:08,244] loop.py:380 INFO TIMING step=4320: step=4.232s data=1.985s compute=4.208s [fwd=2.997s bwd=1.197s opt=0.003s] dispatch=2.542s [2026-10-18 22:52:08,244] loop.py:45 INFO TRAIN_STEP step=4320 loss=1.629646 Training: 24%|██▍ | 4320/18018 [9:21:54<28:59:08, 7.62s/step, loss=1.630, acc=0.647, acc_len=1.51, thru=18.0, I=18.7, T=30.2, wait=2.5s, pool=60, epoch=1/2][2026-10-18 22:52:15,479] loop.py:45 INFO TRAIN_STEP step=4321 loss=1.580394 Training: 24%|██▍ | 4321/18018 [9:22:01<28:32:47, 7.50s/step, loss=1.580, acc=0.659, acc_len=1.55, thru=15.6, I=15.2, T=31.7, wait=3.2s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:52:15,475] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4320: forward=2818.2ms backward=1178.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:22,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:52:23,515] loop.py:45 INFO TRAIN_STEP step=4322 loss=1.677419 Training: 24%|██▍ | 4322/18018 [9:22:09<29:09:19, 7.66s/step, loss=1.677, acc=0.642, acc_len=1.50, thru=16.9, I=17.5, T=31.0, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:23,707] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 22:52:29,829] loop.py:45 INFO TRAIN_STEP step=4323 loss=1.674128 Training: 24%|██▍ | 4323/18018 [9:22:15<27:36:37, 7.26s/step, loss=1.674, acc=0.637, acc_len=1.50, thru=16.0, I=23.1, T=33.8, wait=2.5s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:37,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:52:39,010] loop.py:45 INFO TRAIN_STEP step=4324 loss=1.585879 Training: 24%|██▍ | 4324/18018 [9:22:25<29:48:12, 7.83s/step, loss=1.586, acc=0.654, acc_len=1.54, thru=16.2, I=16.0, T=33.3, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:39,076] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:46,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:52:47,474] loop.py:380 INFO TIMING step=4325: step=4.365s data=2.072s compute=4.334s [fwd=3.019s bwd=1.300s opt=0.003s] dispatch=4.090s [2026-10-18 22:52:47,475] loop.py:45 INFO TRAIN_STEP step=4325 loss=1.422983 Training: 24%|██▍ | 4325/18018 [9:22:33<30:31:30, 8.03s/step, loss=1.423, acc=0.685, acc_len=1.64, thru=13.3, I=16.7, T=29.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:47,540] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:53,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:52:54,121] loop.py:45 INFO TRAIN_STEP step=4326 loss=1.729746 Training: 24%|██▍ | 4326/18018 [9:22:40<28:56:39, 7.61s/step, loss=1.730, acc=0.630, acc_len=1.46, thru=17.6, I=19.2, T=34.7, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:52:54,116] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4325: forward=2585.4ms backward=1050.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:52:54,437] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:53:00,564] loop.py:45 INFO TRAIN_STEP step=4327 loss=1.580050 Training: 24%|██▍ | 4327/18018 [9:22:46<27:36:36, 7.26s/step, loss=1.580, acc=0.655, acc_len=1.54, thru=17.2, I=17.4, T=32.9, wait=2.5s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:07,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:53:07,418] loop.py:45 INFO TRAIN_STEP step=4328 loss=1.868867 Training: 24%|██▍ | 4328/18018 [9:22:53<27:08:43, 7.14s/step, loss=1.869, acc=0.609, acc_len=1.40, thru=18.9, I=19.3, T=33.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:07,631] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:53:17,282] loop.py:45 INFO TRAIN_STEP step=4329 loss=1.772387 Training: 24%|██▍ | 4329/18018 [9:23:03<30:15:17, 7.96s/step, loss=1.772, acc=0.627, acc_len=1.45, thru=11.2, I=12.4, T=28.0, wait=5.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:25,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:53:27,081] loop.py:380 INFO TIMING step=4330: step=4.356s data=2.053s compute=4.288s [fwd=2.942s bwd=1.332s opt=0.003s] dispatch=5.431s [2026-10-18 22:53:27,081] loop.py:45 INFO TRAIN_STEP step=4330 loss=1.653387 Training: 24%|██▍ | 4330/18018 [9:23:13<32:21:15, 8.51s/step, loss=1.653, acc=0.643, acc_len=1.50, thru=15.6, I=14.8, T=29.4, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:27,376] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:34,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:53:35,050] loop.py:45 INFO TRAIN_STEP step=4331 loss=1.634554 Training: 24%|██▍ | 4331/18018 [9:23:21<31:44:05, 8.35s/step, loss=1.635, acc=0.645, acc_len=1.51, thru=13.7, I=18.2, T=26.4, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:53:35,045] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4330: forward=3544.7ms backward=1248.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:35,194] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:40,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:53:42,530] loop.py:45 INFO TRAIN_STEP step=4332 loss=1.525150 Training: 24%|██▍ | 4332/18018 [9:23:28<30:44:24, 8.09s/step, loss=1.525, acc=0.663, acc_len=1.56, thru=15.3, I=18.4, T=27.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:43,001] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 22:53:49,438] loop.py:45 INFO TRAIN_STEP step=4333 loss=1.513946 Training: 24%|██▍ | 4333/18018 [9:23:35<29:23:50, 7.73s/step, loss=1.514, acc=0.668, acc_len=1.58, thru=15.0, I=24.5, T=35.2, wait=3.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:56,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:53:57,806] loop.py:45 INFO TRAIN_STEP step=4334 loss=1.474671 Training: 24%|██▍ | 4334/18018 [9:23:43<30:07:04, 7.92s/step, loss=1.475, acc=0.670, acc_len=1.60, thru=15.0, I=15.8, T=32.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:53:57,953] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:03,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:54:04,830] loop.py:380 INFO TIMING step=4335: step=4.170s data=1.966s compute=4.127s [fwd=2.855s bwd=1.257s opt=0.003s] dispatch=2.845s [2026-10-18 22:54:04,830] loop.py:45 INFO TRAIN_STEP step=4335 loss=1.498441 Training: 24%|██▍ | 4335/18018 [9:23:50<29:05:25, 7.65s/step, loss=1.498, acc=0.669, acc_len=1.59, thru=16.1, I=22.4, T=30.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:05,146] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:10,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:54:11,037] loop.py:45 INFO TRAIN_STEP step=4336 loss=1.548256 Training: 24%|██▍ | 4336/18018 [9:23:57<27:26:24, 7.22s/step, loss=1.548, acc=0.661, acc_len=1.57, thru=19.0, I=25.9, T=32.5, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:54:11,032] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4335: forward=2751.6ms backward=1142.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:11,332] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:17,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:54:19,508] loop.py:45 INFO TRAIN_STEP step=4337 loss=1.531627 Training: 24%|██▍ | 4337/18018 [9:24:05<28:51:41, 7.59s/step, loss=1.532, acc=0.661, acc_len=1.56, thru=15.3, I=17.6, T=32.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:19,960] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:54:26,692] loop.py:45 INFO TRAIN_STEP step=4338 loss=1.649299 Training: 24%|██▍ | 4338/18018 [9:24:12<28:23:27, 7.47s/step, loss=1.649, acc=0.640, acc_len=1.50, thru=15.7, I=18.3, T=29.5, wait=2.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:32,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:54:34,000] loop.py:45 INFO TRAIN_STEP step=4339 loss=1.557994 Training: 24%|██▍ | 4339/18018 [9:24:20<28:12:25, 7.42s/step, loss=1.558, acc=0.661, acc_len=1.56, thru=20.0, I=18.7, T=29.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:34,105] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:38,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:54:40,674] loop.py:380 INFO TIMING step=4340: step=4.187s data=2.051s compute=4.156s [fwd=2.934s bwd=1.206s opt=0.003s] dispatch=2.475s [2026-10-18 22:54:40,674] loop.py:45 INFO TRAIN_STEP step=4340 loss=1.561962 Training: 24%|██▍ | 4340/18018 [9:24:26<27:20:56, 7.20s/step, loss=1.562, acc=0.656, acc_len=1.55, thru=16.8, I=25.0, T=30.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:40,900] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:50,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:54:50,705] loop.py:45 INFO TRAIN_STEP step=4341 loss=1.586547 Training: 24%|██▍ | 4341/18018 [9:24:36<30:34:32, 8.05s/step, loss=1.587, acc=0.659, acc_len=1.55, thru=13.1, I=13.5, T=28.1, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:54:50,701] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4340: forward=3286.6ms backward=1201.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:50,847] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:57,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:54:59,552] loop.py:45 INFO TRAIN_STEP step=4342 loss=1.802777 Training: 24%|██▍ | 4342/18018 [9:24:45<31:29:05, 8.29s/step, loss=1.803, acc=0.621, acc_len=1.43, thru=14.2, I=18.2, T=32.9, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:54:59,732] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:55:04,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:55:06,477] loop.py:45 INFO TRAIN_STEP step=4343 loss=1.667483 Training: 24%|██▍ | 4343/18018 [9:24:52<29:55:38, 7.88s/step, loss=1.667, acc=0.644, acc_len=1.51, thru=17.6, I=28.2, T=31.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:55:06,603] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-18 22:55:12,819] loop.py:45 INFO TRAIN_STEP step=4344 loss=1.576606 Training: 24%|██▍ | 4344/18018 [9:24:58<28:10:28, 7.42s/step, loss=1.577, acc=0.660, acc_len=1.56, thru=17.5, I=22.2, T=35.8, wait=2.8s, pool=56, epoch=1/2][2026-10-18 22:55:19,780] loop.py:380 INFO TIMING step=4345: step=3.820s data=1.674s compute=3.783s [fwd=2.697s bwd=1.071s opt=0.003s] dispatch=3.133s [2026-10-18 22:55:19,780] loop.py:45 INFO TRAIN_STEP step=4345 loss=1.661808 Training: 24%|██▍ | 4345/18018 [9:25:05<27:39:12, 7.28s/step, loss=1.662, acc=0.645, acc_len=1.51, thru=15.2, I=17.9, T=33.5, wait=3.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:55:28,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:55:29,116] loop.py:45 INFO TRAIN_STEP step=4346 loss=1.671930 Training: 24%|██▍ | 4346/18018 [9:25:15<29:59:36, 7.90s/step, loss=1.672, acc=0.639, acc_len=1.49, thru=16.2, I=14.8, T=29.0, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:55:29,111] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4345: forward=2998.4ms backward=1361.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:55:29,433] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:55:36,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:55:37,083] loop.py:45 INFO TRAIN_STEP step=4347 loss=1.641438 Training: 24%|██▍ | 4347/18018 [9:25:23<30:04:13, 7.92s/step, loss=1.641, acc=0.648, acc_len=1.51, thru=16.6, I=16.7, T=33.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:55:37,469] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:55:47,121] loop.py:45 INFO TRAIN_STEP step=4348 loss=1.665509 Training: 24%|██▍ | 4348/18018 [9:25:33<32:28:45, 8.55s/step, loss=1.666, acc=0.642, acc_len=1.50, thru=13.3, I=13.3, T=33.0, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:55:47,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:55:47,612] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 22:55:54,221] loop.py:45 INFO TRAIN_STEP step=4349 loss=1.588112 Training: 24%|██▍ | 4349/18018 [9:25:40<30:49:18, 8.12s/step, loss=1.588, acc=0.653, acc_len=1.54, thru=17.4, I=18.0, T=32.2, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:02,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:56:03,870] loop.py:380 INFO TIMING step=4350: step=4.092s data=1.764s compute=4.060s [fwd=2.855s bwd=1.190s opt=0.003s] dispatch=5.548s [2026-10-18 22:56:03,870] loop.py:45 INFO TRAIN_STEP step=4350 loss=1.856138 Training: 24%|██▍ | 4350/18018 [9:25:49<32:33:53, 8.58s/step, loss=1.856, acc=0.610, acc_len=1.40, thru=14.4, I=16.1, T=31.3, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:03,935] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:09,637] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 22:56:10,887] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4350: forward=2712.6ms backward=1170.2ms [2026-10-18 22:56:10,892] loop.py:45 INFO TRAIN_STEP step=4351 loss=1.565378 Training: 24%|██▍ | 4351/18018 [9:25:56<30:47:27, 8.11s/step, loss=1.565, acc=0.657, acc_len=1.56, thru=17.2, I=22.8, T=32.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:11,145] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:20,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:56:20,343] loop.py:45 INFO TRAIN_STEP step=4352 loss=1.619207 Training: 24%|██▍ | 4352/18018 [9:26:06<32:18:57, 8.51s/step, loss=1.619, acc=0.653, acc_len=1.52, thru=13.9, I=14.8, T=27.5, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:20,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:27,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:56:29,034] loop.py:45 INFO TRAIN_STEP step=4353 loss=1.682029 Training: 24%|██▍ | 4353/18018 [9:26:15<32:31:04, 8.57s/step, loss=1.682, acc=0.645, acc_len=1.51, thru=15.0, I=15.6, T=33.9, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:29,191] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:36,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:56:38,028] loop.py:45 INFO TRAIN_STEP step=4354 loss=1.627296 Training: 24%|██▍ | 4354/18018 [9:26:24<32:59:54, 8.69s/step, loss=1.627, acc=0.649, acc_len=1.52, thru=14.5, I=17.3, T=31.8, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:38,496] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:56:45,956] loop.py:380 INFO TIMING step=4355: step=4.934s data=2.464s compute=4.904s [fwd=3.345s bwd=1.544s opt=0.003s] dispatch=2.985s [2026-10-18 22:56:45,956] loop.py:45 INFO TRAIN_STEP step=4355 loss=1.591801 Training: 24%|██▍ | 4355/18018 [9:26:32<32:07:27, 8.46s/step, loss=1.592, acc=0.650, acc_len=1.53, thru=13.6, I=17.0, T=25.9, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:53,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:56:55,088] loop.py:45 INFO TRAIN_STEP step=4356 loss=1.561368 Training: 24%|██▍ | 4356/18018 [9:26:41<32:52:56, 8.66s/step, loss=1.561, acc=0.657, acc_len=1.55, thru=15.5, I=16.4, T=29.5, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:56:55,083] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4355: forward=2971.4ms backward=1301.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:56:55,559] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:00,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:57:02,410] loop.py:45 INFO TRAIN_STEP step=4357 loss=1.786818 Training: 24%|██▍ | 4357/18018 [9:26:48<31:21:15, 8.26s/step, loss=1.787, acc=0.627, acc_len=1.45, thru=16.2, I=24.9, T=32.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:02,442] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:08,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:57:09,260] loop.py:45 INFO TRAIN_STEP step=4358 loss=1.510222 Training: 24%|██▍ | 4358/18018 [9:26:55<29:44:34, 7.84s/step, loss=1.510, acc=0.670, acc_len=1.59, thru=17.1, I=25.3, T=30.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:09,502] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:14,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:57:15,913] loop.py:45 INFO TRAIN_STEP step=4359 loss=1.633050 Training: 24%|██▍ | 4359/18018 [9:27:01<28:23:27, 7.48s/step, loss=1.633, acc=0.651, acc_len=1.52, thru=17.6, I=21.6, T=32.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:16,309] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 22:57:22,979] loop.py:380 INFO TIMING step=4360: step=3.606s data=1.609s compute=3.564s [fwd=2.526s bwd=1.023s opt=0.003s] dispatch=3.451s [2026-10-18 22:57:22,979] loop.py:45 INFO TRAIN_STEP step=4360 loss=1.644150 Training: 24%|██▍ | 4360/18018 [9:27:09<27:54:54, 7.36s/step, loss=1.644, acc=0.644, acc_len=1.51, thru=16.3, I=20.4, T=35.5, wait=3.5s, pool=60, epoch=1/2][2026-10-18 22:57:30,563] loop.py:45 INFO TRAIN_STEP step=4361 loss=1.633375 Training: 24%|██▍ | 4361/18018 [9:27:16<28:10:06, 7.43s/step, loss=1.633, acc=0.643, acc_len=1.50, thru=15.7, I=15.5, T=35.6, wait=4.0s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:57:30,559] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4360: forward=2531.0ms backward=1022.9ms [2026-10-18 22:57:38,938] loop.py:45 INFO TRAIN_STEP step=4362 loss=1.649360 Training: 24%|██▍ | 4362/18018 [9:27:25<29:15:08, 7.71s/step, loss=1.649, acc=0.640, acc_len=1.49, thru=15.7, I=17.0, T=27.9, wait=3.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:47,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:57:47,675] loop.py:45 INFO TRAIN_STEP step=4363 loss=1.496862 Training: 24%|██▍ | 4363/18018 [9:27:33<30:24:53, 8.02s/step, loss=1.497, acc=0.674, acc_len=1.61, thru=15.7, I=14.6, T=27.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:47,726] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:53,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:57:53,990] loop.py:45 INFO TRAIN_STEP step=4364 loss=1.780819 Training: 24%|██▍ | 4364/18018 [9:27:40<28:28:26, 7.51s/step, loss=1.781, acc=0.625, acc_len=1.44, thru=18.5, I=19.0, T=32.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:57:54,099] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:01,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:58:02,408] loop.py:380 INFO TIMING step=4365: step=4.298s data=1.935s compute=4.268s [fwd=2.971s bwd=1.282s opt=0.003s] dispatch=4.110s [2026-10-18 22:58:02,408] loop.py:45 INFO TRAIN_STEP step=4365 loss=1.581564 Training: 24%|██▍ | 4365/18018 [9:27:48<29:30:26, 7.78s/step, loss=1.582, acc=0.656, acc_len=1.55, thru=15.4, I=17.2, T=29.8, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:02,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:08,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:58:09,407] loop.py:45 INFO TRAIN_STEP step=4366 loss=1.559245 Training: 24%|██▍ | 4366/18018 [9:27:55<28:37:00, 7.55s/step, loss=1.559, acc=0.658, acc_len=1.56, thru=16.7, I=20.4, T=31.2, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:58:09,401] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4365: forward=2916.8ms backward=1149.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:09,610] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:58:16,498] loop.py:45 INFO TRAIN_STEP step=4367 loss=1.464390 Training: 24%|██▍ | 4367/18018 [9:28:02<28:05:39, 7.41s/step, loss=1.464, acc=0.671, acc_len=1.60, thru=15.6, I=18.2, T=32.3, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:21,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:58:23,870] loop.py:45 INFO TRAIN_STEP step=4368 loss=1.652110 Training: 24%|██▍ | 4368/18018 [9:28:09<28:03:02, 7.40s/step, loss=1.652, acc=0.649, acc_len=1.51, thru=18.4, I=19.3, T=29.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:23,991] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=16/64 [2026-10-18 22:58:30,984] loop.py:45 INFO TRAIN_STEP step=4369 loss=1.530306 Training: 24%|██▍ | 4369/18018 [9:28:17<27:43:32, 7.31s/step, loss=1.530, acc=0.670, acc_len=1.60, thru=13.8, I=18.1, T=27.1, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:39,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:58:40,525] loop.py:380 INFO TIMING step=4370: step=4.540s data=2.039s compute=4.501s [fwd=3.126s bwd=1.359s opt=0.003s] dispatch=4.993s [2026-10-18 22:58:40,525] loop.py:45 INFO TRAIN_STEP step=4370 loss=1.680204 Training: 24%|██▍ | 4370/18018 [9:28:26<30:15:37, 7.98s/step, loss=1.680, acc=0.641, acc_len=1.50, thru=14.2, I=16.4, T=28.2, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:40,672] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (TrainerActor pid=9310) [2026-10-18 22:58:47,793] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4370: forward=2948.3ms backward=1216.6ms [2026-10-18 22:58:47,798] loop.py:45 INFO TRAIN_STEP step=4371 loss=1.551400 Training: 24%|██▍ | 4371/18018 [9:28:33<29:26:58, 7.77s/step, loss=1.551, acc=0.662, acc_len=1.56, thru=14.3, I=16.5, T=30.4, wait=3.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:58,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:58:59,373] loop.py:45 INFO TRAIN_STEP step=4372 loss=1.736479 Training: 24%|██▍ | 4372/18018 [9:28:45<33:46:39, 8.91s/step, loss=1.736, acc=0.639, acc_len=1.49, thru=12.4, I=14.2, T=27.4, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:58:59,879] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 22:59:07,889] loop.py:45 INFO TRAIN_STEP step=4373 loss=1.723910 Training: 24%|██▍ | 4373/18018 [9:28:53<33:19:34, 8.79s/step, loss=1.724, acc=0.635, acc_len=1.48, thru=12.4, I=17.6, T=32.0, wait=4.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:59:14,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:59:15,002] loop.py:45 INFO TRAIN_STEP step=4374 loss=1.582195 Training: 24%|██▍ | 4374/18018 [9:29:01<31:24:44, 8.29s/step, loss=1.582, acc=0.655, acc_len=1.54, thru=17.4, I=17.2, T=32.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:59:15,440] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:59:21,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:59:22,330] loop.py:380 INFO TIMING step=4375: step=4.700s data=2.378s compute=4.662s [fwd=3.063s bwd=1.584s opt=0.003s] dispatch=2.619s [2026-10-18 22:59:22,331] loop.py:45 INFO TRAIN_STEP step=4375 loss=1.644485 Training: 24%|██▍ | 4375/18018 [9:29:08<30:19:10, 8.00s/step, loss=1.644, acc=0.647, acc_len=1.52, thru=17.6, I=20.2, T=27.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:59:22,761] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:59:27,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:59:28,666] loop.py:45 INFO TRAIN_STEP step=4376 loss=1.793713 Training: 24%|██▍ | 4376/18018 [9:29:14<28:25:32, 7.50s/step, loss=1.794, acc=0.622, acc_len=1.43, thru=17.2, I=20.7, T=30.7, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 22:59:28,661] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4375: forward=2898.2ms backward=1224.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:59:28,828] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 22:59:37,574] loop.py:45 INFO TRAIN_STEP step=4377 loss=1.728588 Training: 24%|██▍ | 4377/18018 [9:29:23<30:01:15, 7.92s/step, loss=1.729, acc=0.633, acc_len=1.46, thru=13.7, I=12.9, T=33.1, wait=5.0s, pool=56, epoch=1/2][2026-10-18 22:59:44,758] loop.py:45 INFO TRAIN_STEP step=4378 loss=1.515401 Training: 24%|██▍ | 4378/18018 [9:29:30<29:10:44, 7.70s/step, loss=1.515, acc=0.666, acc_len=1.58, thru=18.8, I=15.7, T=32.8, wait=3.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:59:53,757] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 22:59:55,350] loop.py:45 INFO TRAIN_STEP step=4379 loss=1.532777 Training: 24%|██▍ | 4379/18018 [9:29:41<32:27:39, 8.57s/step, loss=1.533, acc=0.667, acc_len=1.58, thru=12.7, I=14.7, T=28.9, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 22:59:55,773] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:00:01,606] loop.py:380 INFO TIMING step=4380: step=3.647s data=1.638s compute=3.616s [fwd=2.603s bwd=0.998s opt=0.003s] dispatch=2.600s [2026-10-18 23:00:01,606] loop.py:45 INFO TRAIN_STEP step=4380 loss=1.537168 Training: 24%|██▍ | 4380/18018 [9:29:47<29:49:56, 7.87s/step, loss=1.537, acc=0.658, acc_len=1.55, thru=17.0, I=23.7, T=35.1, wait=2.6s, pool=56, epoch=1/2][2026-10-18 23:00:09,057] loop.py:45 INFO TRAIN_STEP step=4381 loss=1.657912 Training: 24%|██▍ | 4381/18018 [9:29:55<29:21:01, 7.75s/step, loss=1.658, acc=0.640, acc_len=1.49, thru=14.5, I=15.6, T=34.0, wait=3.7s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:00:09,052] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4380: forward=2649.9ms backward=1072.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:00:19,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:00:20,049] loop.py:45 INFO TRAIN_STEP step=4382 loss=1.540077 Training: 24%|██▍ | 4382/18018 [9:30:06<33:02:07, 8.72s/step, loss=1.540, acc=0.666, acc_len=1.58, thru=13.5, I=14.0, T=23.0, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:00:20,083] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-18 23:00:26,324] loop.py:45 INFO TRAIN_STEP step=4383 loss=1.623389 Training: 24%|██▍ | 4383/18018 [9:30:12<30:14:58, 7.99s/step, loss=1.623, acc=0.649, acc_len=1.52, thru=16.5, I=16.5, T=30.8, wait=2.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:00:33,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:00:35,173] loop.py:45 INFO TRAIN_STEP step=4384 loss=1.615245 Training: 24%|██▍ | 4384/18018 [9:30:21<31:13:46, 8.25s/step, loss=1.615, acc=0.653, acc_len=1.54, thru=15.9, I=17.3, T=31.0, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:00:35,299] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:00:40,544] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:00:41,473] loop.py:380 INFO TIMING step=4385: step=3.985s data=1.822s compute=3.951s [fwd=2.802s bwd=1.134s opt=0.003s] dispatch=2.305s [2026-10-18 23:00:41,474] loop.py:45 INFO TRAIN_STEP step=4385 loss=1.547156 Training: 24%|██▍ | 4385/18018 [9:30:27<29:00:53, 7.66s/step, loss=1.547, acc=0.660, acc_len=1.56, thru=19.0, I=23.3, T=32.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:00:41,548] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 23:00:48,269] loop.py:45 INFO TRAIN_STEP step=4386 loss=1.515901 Training: 24%|██▍ | 4386/18018 [9:30:34<28:01:45, 7.40s/step, loss=1.516, acc=0.667, acc_len=1.59, thru=14.5, I=21.4, T=29.2, wait=2.4s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:00:48,264] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4385: forward=3017.8ms backward=1304.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:00:56,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:00:57,521] loop.py:45 INFO TRAIN_STEP step=4387 loss=1.660059 Training: 24%|██▍ | 4387/18018 [9:30:43<30:07:45, 7.96s/step, loss=1.660, acc=0.642, acc_len=1.50, thru=17.3, I=14.8, T=30.9, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:00:57,953] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:04,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:01:06,425] loop.py:45 INFO TRAIN_STEP step=4388 loss=1.658742 Training: 24%|██▍ | 4388/18018 [9:30:52<31:12:12, 8.24s/step, loss=1.659, acc=0.640, acc_len=1.49, thru=14.4, I=20.7, T=30.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:06,760] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:11,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:01:13,267] loop.py:45 INFO TRAIN_STEP step=4389 loss=1.508173 Training: 24%|██▍ | 4389/18018 [9:30:59<29:36:40, 7.82s/step, loss=1.508, acc=0.671, acc_len=1.60, thru=17.4, I=28.5, T=29.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:13,327] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:18,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:01:19,464] loop.py:380 INFO TIMING step=4390: step=4.127s data=1.889s compute=4.085s [fwd=2.823s bwd=1.247s opt=0.003s] dispatch=2.058s [2026-10-18 23:01:19,464] loop.py:45 INFO TRAIN_STEP step=4390 loss=1.655120 Training: 24%|██▍ | 4390/18018 [9:31:05<27:45:44, 7.33s/step, loss=1.655, acc=0.646, acc_len=1.51, thru=17.6, I=23.3, T=31.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:19,564] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 23:01:25,768] loop.py:45 INFO TRAIN_STEP step=4391 loss=1.683832 Training: 24%|██▍ | 4391/18018 [9:31:11<26:35:34, 7.03s/step, loss=1.684, acc=0.642, acc_len=1.50, thru=17.9, I=19.4, T=33.6, wait=2.5s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:01:25,764] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4390: forward=2657.7ms backward=1104.8ms [2026-10-18 23:01:33,707] loop.py:45 INFO TRAIN_STEP step=4392 loss=1.678016 Training: 24%|██▍ | 4392/18018 [9:31:19<27:37:32, 7.30s/step, loss=1.678, acc=0.640, acc_len=1.49, thru=15.2, I=16.5, T=30.6, wait=3.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:42,044] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:01:42,577] loop.py:45 INFO TRAIN_STEP step=4393 loss=1.664814 Training: 24%|██▍ | 4393/18018 [9:31:28<29:24:33, 7.77s/step, loss=1.665, acc=0.644, acc_len=1.51, thru=16.9, I=14.9, T=27.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:43,048] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:47,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:01:49,473] loop.py:45 INFO TRAIN_STEP step=4394 loss=1.544088 Training: 24%|██▍ | 4394/18018 [9:31:35<28:24:49, 7.51s/step, loss=1.544, acc=0.666, acc_len=1.58, thru=16.5, I=20.4, T=31.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:01:49,732] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:01:57,236] loop.py:380 INFO TIMING step=4395: step=4.574s data=2.169s compute=4.540s [fwd=3.194s bwd=1.331s opt=0.003s] dispatch=3.180s [2026-10-18 23:01:57,236] loop.py:45 INFO TRAIN_STEP step=4395 loss=1.499841 Training: 24%|██▍ | 4395/18018 [9:31:43<28:42:01, 7.58s/step, loss=1.500, acc=0.670, acc_len=1.59, thru=11.6, I=15.9, T=28.0, wait=3.2s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:05,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:02:06,916] loop.py:45 INFO TRAIN_STEP step=4396 loss=1.628462 Training: 24%|██▍ | 4396/18018 [9:31:52<31:04:34, 8.21s/step, loss=1.628, acc=0.646, acc_len=1.52, thru=16.1, I=16.0, T=28.1, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:02:06,912] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4395: forward=3227.3ms backward=1264.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:07,310] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:14,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:02:15,628] loop.py:45 INFO TRAIN_STEP step=4397 loss=1.627869 Training: 24%|██▍ | 4397/18018 [9:32:01<31:38:38, 8.36s/step, loss=1.628, acc=0.647, acc_len=1.52, thru=13.4, I=17.1, T=28.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:15,983] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:23,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:02:24,328] loop.py:45 INFO TRAIN_STEP step=4398 loss=1.583743 Training: 24%|██▍ | 4398/18018 [9:32:10<32:01:25, 8.46s/step, loss=1.584, acc=0.655, acc_len=1.55, thru=13.4, I=17.3, T=27.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:24,704] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:31,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:02:32,706] loop.py:45 INFO TRAIN_STEP step=4399 loss=1.595981 Training: 24%|██▍ | 4399/18018 [9:32:18<31:55:22, 8.44s/step, loss=1.596, acc=0.651, acc_len=1.53, thru=14.1, I=17.6, T=27.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:32,860] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:37,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:02:39,277] loop.py:380 INFO TIMING step=4400: step=4.128s data=1.838s compute=4.100s [fwd=2.910s bwd=1.175s opt=0.003s] dispatch=2.432s [2026-10-18 23:02:39,278] loop.py:45 INFO TRAIN_STEP step=4400 loss=1.612484 Training: 24%|██▍ | 4400/18018 [9:32:25<29:48:05, 7.88s/step, loss=1.612, acc=0.649, acc_len=1.52, thru=18.8, I=22.8, T=31.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:39,605] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:46,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:02:47,275] loop.py:45 INFO TRAIN_STEP step=4401 loss=1.476228 (TrainerActor pid=9310) [2026-10-18 23:02:47,269] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4400: forward=3345.4ms backward=1253.2ms Training: 24%|██▍ | 4401/18018 [9:32:33<29:56:06, 7.91s/step, loss=1.476, acc=0.674, acc_len=1.60, thru=13.6, I=17.9, T=27.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:47,393] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:55,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:02:55,689] loop.py:45 INFO TRAIN_STEP step=4402 loss=1.440475 Training: 24%|██▍ | 4402/18018 [9:32:41<30:29:58, 8.06s/step, loss=1.440, acc=0.679, acc_len=1.62, thru=14.3, I=14.9, T=28.5, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:02:56,082] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 23:03:05,607] loop.py:45 INFO TRAIN_STEP step=4403 loss=1.662034 Training: 24%|██▍ | 4403/18018 [9:32:51<32:35:52, 8.62s/step, loss=1.662, acc=0.644, acc_len=1.51, thru=10.9, I=9.7, T=29.8, wait=5.6s, pool=40, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:16,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:03:18,403] loop.py:45 INFO TRAIN_STEP step=4404 loss=1.643739 Training: 24%|██▍ | 4404/18018 [9:33:04<37:20:13, 9.87s/step, loss=1.644, acc=0.647, acc_len=1.52, thru=10.8, I=12.2, T=26.5, wait=8.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:18,476] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:25,316] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:03:26,609] loop.py:380 INFO TIMING step=4405: step=4.862s data=2.425s compute=4.828s [fwd=3.426s bwd=1.390s opt=0.003s] dispatch=3.334s [2026-10-18 23:03:26,610] loop.py:45 INFO TRAIN_STEP step=4405 loss=1.597011 Training: 24%|██▍ | 4405/18018 [9:33:12<35:26:31, 9.37s/step, loss=1.597, acc=0.652, acc_len=1.54, thru=15.8, I=17.5, T=26.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:26,825] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:32,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:03:33,028] loop.py:45 INFO TRAIN_STEP step=4406 loss=1.607924 Training: 24%|██▍ | 4406/18018 [9:33:19<32:05:18, 8.49s/step, loss=1.608, acc=0.651, acc_len=1.52, thru=17.3, I=21.7, T=31.6, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:03:33,023] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4405: forward=2793.0ms backward=1217.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:33,056] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:39,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:03:39,843] loop.py:45 INFO TRAIN_STEP step=4407 loss=1.630887 Training: 24%|██▍ | 4407/18018 [9:33:25<30:11:22, 7.98s/step, loss=1.631, acc=0.647, acc_len=1.51, thru=18.1, I=22.4, T=32.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:40,184] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:46,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:03:48,614] loop.py:45 INFO TRAIN_STEP step=4408 loss=1.688866 Training: 24%|██▍ | 4408/18018 [9:33:34<31:04:54, 8.22s/step, loss=1.689, acc=0.642, acc_len=1.50, thru=14.5, I=16.1, T=27.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:48,933] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:54,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:03:56,883] loop.py:45 INFO TRAIN_STEP step=4409 loss=1.656305 Training: 24%|██▍ | 4409/18018 [9:33:42<31:07:52, 8.24s/step, loss=1.656, acc=0.646, acc_len=1.51, thru=13.2, I=20.5, T=26.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:03:57,217] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:04:02,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:04:04,364] loop.py:380 INFO TIMING step=4410: step=4.135s data=1.908s compute=4.103s [fwd=2.868s bwd=1.222s opt=0.003s] dispatch=3.336s [2026-10-18 23:04:04,364] loop.py:45 INFO TRAIN_STEP step=4410 loss=1.618911 Training: 24%|██▍ | 4410/18018 [9:33:50<30:16:24, 8.01s/step, loss=1.619, acc=0.653, acc_len=1.53, thru=14.1, I=22.8, T=31.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:04:04,500] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-18 23:04:11,619] loop.py:45 INFO TRAIN_STEP step=4411 loss=1.585141 Training: 24%|██▍ | 4411/18018 [9:33:57<29:25:06, 7.78s/step, loss=1.585, acc=0.658, acc_len=1.55, thru=14.7, I=18.4, T=32.1, wait=3.3s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:04:11,614] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4410: forward=2760.4ms backward=1170.8ms [2026-10-18 23:04:21,220] loop.py:45 INFO TRAIN_STEP step=4412 loss=1.559016 Training: 24%|██▍ | 4412/18018 [9:34:07<31:28:29, 8.33s/step, loss=1.559, acc=0.659, acc_len=1.55, thru=14.7, I=13.0, T=30.5, wait=5.4s, pool=60, epoch=1/2][2026-10-18 23:04:27,945] loop.py:45 INFO TRAIN_STEP step=4413 loss=1.570993 Training: 24%|██▍ | 4413/18018 [9:34:14<29:39:18, 7.85s/step, loss=1.571, acc=0.662, acc_len=1.56, thru=16.6, I=16.4, T=32.3, wait=2.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:04:36,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:04:37,558] loop.py:45 INFO TRAIN_STEP step=4414 loss=1.582460 Training: 24%|██▍ | 4414/18018 [9:34:23<31:39:25, 8.38s/step, loss=1.582, acc=0.654, acc_len=1.55, thru=15.8, I=15.2, T=28.4, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:04:37,974] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:04:47,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:04:48,629] loop.py:380 INFO TIMING step=4415: step=4.225s data=1.964s compute=4.176s [fwd=2.948s bwd=1.215s opt=0.003s] dispatch=6.835s [2026-10-18 23:04:48,629] loop.py:45 INFO TRAIN_STEP step=4415 loss=1.634302 Training: 25%|██▍ | 4415/18018 [9:34:34<34:42:27, 9.19s/step, loss=1.634, acc=0.647, acc_len=1.51, thru=12.1, I=13.5, T=30.3, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:04:49,062] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:04:56,357] loop.py:45 INFO TRAIN_STEP step=4416 loss=1.655361 Training: 25%|██▍ | 4416/18018 [9:34:42<33:03:08, 8.75s/step, loss=1.655, acc=0.644, acc_len=1.51, thru=13.8, I=17.3, T=29.8, wait=3.4s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:04:56,352] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4415: forward=2915.2ms backward=1332.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:04,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:05:06,807] loop.py:45 INFO TRAIN_STEP step=4417 loss=1.572318 Training: 25%|██▍ | 4417/18018 [9:34:52<34:58:43, 9.26s/step, loss=1.572, acc=0.657, acc_len=1.55, thru=12.9, I=17.7, T=25.9, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:07,135] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:12,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:05:13,573] loop.py:45 INFO TRAIN_STEP step=4418 loss=1.648998 Training: 25%|██▍ | 4418/18018 [9:34:59<32:09:06, 8.51s/step, loss=1.649, acc=0.646, acc_len=1.51, thru=14.6, I=25.3, T=32.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:14,061] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:21,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:05:22,459] loop.py:45 INFO TRAIN_STEP step=4419 loss=1.657090 Training: 25%|██▍ | 4419/18018 [9:35:08<32:34:30, 8.62s/step, loss=1.657, acc=0.644, acc_len=1.51, thru=15.2, I=16.4, T=28.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:22,696] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 23:05:29,288] loop.py:380 INFO TIMING step=4420: step=4.388s data=2.043s compute=4.354s [fwd=3.079s bwd=1.262s opt=0.003s] dispatch=2.431s [2026-10-18 23:05:29,288] loop.py:45 INFO TRAIN_STEP step=4420 loss=1.581581 Training: 25%|██▍ | 4420/18018 [9:35:15<30:32:36, 8.09s/step, loss=1.582, acc=0.655, acc_len=1.54, thru=16.5, I=19.2, T=29.2, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:37,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:05:38,526] loop.py:45 INFO TRAIN_STEP step=4421 loss=1.443908 Training: 25%|██▍ | 4421/18018 [9:35:24<31:50:26, 8.43s/step, loss=1.444, acc=0.683, acc_len=1.64, thru=14.3, I=14.5, T=27.3, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:05:38,521] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4420: forward=3259.4ms backward=1388.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:38,967] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:44,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:05:45,370] loop.py:45 INFO TRAIN_STEP step=4422 loss=1.734499 Training: 25%|██▍ | 4422/18018 [9:35:31<30:02:32, 7.95s/step, loss=1.734, acc=0.633, acc_len=1.47, thru=17.8, I=20.8, T=32.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:45,522] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:53,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:05:55,505] loop.py:45 INFO TRAIN_STEP step=4423 loss=1.529615 Training: 25%|██▍ | 4423/18018 [9:35:41<32:30:32, 8.61s/step, loss=1.530, acc=0.666, acc_len=1.58, thru=12.8, I=13.8, T=24.9, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:05:56,003] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:06:04,231] loop.py:45 INFO TRAIN_STEP step=4424 loss=1.570451 Training: 25%|██▍ | 4424/18018 [9:35:50<32:38:24, 8.64s/step, loss=1.570, acc=0.656, acc_len=1.55, thru=14.1, I=16.6, T=30.5, wait=4.5s, pool=52, epoch=1/2][2026-10-18 23:06:12,131] loop.py:380 INFO TIMING step=4425: step=4.109s data=1.781s compute=4.073s [fwd=2.905s bwd=1.154s opt=0.003s] dispatch=3.783s [2026-10-18 23:06:12,131] loop.py:45 INFO TRAIN_STEP step=4425 loss=1.605649 Training: 25%|██▍ | 4425/18018 [9:35:58<31:47:43, 8.42s/step, loss=1.606, acc=0.652, acc_len=1.53, thru=13.4, I=13.7, T=31.1, wait=3.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:06:18,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:06:20,251] loop.py:45 INFO TRAIN_STEP step=4426 loss=1.728970 Training: 25%|██▍ | 4426/18018 [9:36:06<31:27:10, 8.33s/step, loss=1.729, acc=0.633, acc_len=1.47, thru=18.0, I=18.5, T=27.9, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:06:20,246] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4425: forward=3134.3ms backward=1381.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:06:20,507] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:06:27,647] loop.py:45 INFO TRAIN_STEP step=4427 loss=1.502534 Training: 25%|██▍ | 4427/18018 [9:36:13<30:23:46, 8.05s/step, loss=1.503, acc=0.668, acc_len=1.58, thru=14.0, I=16.6, T=31.3, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:06:37,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:06:38,267] loop.py:45 INFO TRAIN_STEP step=4428 loss=1.625308 Training: 25%|██▍ | 4428/18018 [9:36:24<33:17:57, 8.82s/step, loss=1.625, acc=0.648, acc_len=1.52, thru=15.1, I=13.3, T=32.5, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:06:38,748] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:06:44,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:06:45,152] loop.py:45 INFO TRAIN_STEP step=4429 loss=1.428595 Training: 25%|██▍ | 4429/18018 [9:36:31<31:06:17, 8.24s/step, loss=1.429, acc=0.681, acc_len=1.64, thru=16.8, I=19.4, T=30.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:06:45,601] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:06:53,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:06:54,175] loop.py:380 INFO TIMING step=4430: step=4.287s data=2.052s compute=4.261s [fwd=3.020s bwd=1.227s opt=0.003s] dispatch=4.726s [2026-10-18 23:06:54,175] loop.py:45 INFO TRAIN_STEP step=4430 loss=1.575612 Training: 25%|██▍ | 4430/18018 [9:36:40<31:59:21, 8.48s/step, loss=1.576, acc=0.662, acc_len=1.56, thru=14.0, I=15.7, T=29.9, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:06:54,360] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:00,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:07:00,699] loop.py:45 INFO TRAIN_STEP step=4431 loss=1.605070 Training: 25%|██▍ | 4431/18018 [9:36:46<29:46:28, 7.89s/step, loss=1.605, acc=0.652, acc_len=1.53, thru=19.2, I=18.6, T=30.1, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:07:00,694] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4430: forward=2974.5ms backward=1229.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:00,730] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:06,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:07:07,512] loop.py:45 INFO TRAIN_STEP step=4432 loss=1.556997 Training: 25%|██▍ | 4432/18018 [9:36:53<28:33:20, 7.57s/step, loss=1.557, acc=0.660, acc_len=1.56, thru=16.4, I=18.6, T=31.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:07,676] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:13,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:07:14,829] loop.py:45 INFO TRAIN_STEP step=4433 loss=1.477252 Training: 25%|██▍ | 4433/18018 [9:37:00<28:16:12, 7.49s/step, loss=1.477, acc=0.671, acc_len=1.60, thru=16.6, I=20.6, T=27.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:14,944] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:21,614] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:07:22,513] loop.py:45 INFO TRAIN_STEP step=4434 loss=1.538881 Training: 25%|██▍ | 4434/18018 [9:37:08<28:29:13, 7.55s/step, loss=1.539, acc=0.659, acc_len=1.56, thru=15.9, I=20.5, T=31.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:22,642] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-18 23:07:31,343] loop.py:380 INFO TIMING step=4435: step=4.238s data=1.947s compute=4.209s [fwd=2.981s bwd=1.213s opt=0.003s] dispatch=4.581s [2026-10-18 23:07:31,343] loop.py:45 INFO TRAIN_STEP step=4435 loss=1.509324 Training: 25%|██▍ | 4435/18018 [9:37:17<29:55:56, 7.93s/step, loss=1.509, acc=0.668, acc_len=1.59, thru=13.7, I=16.8, T=30.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:31,325] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:31,827] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:36,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:07:38,005] loop.py:45 INFO TRAIN_STEP step=4436 loss=1.625006 Training: 25%|██▍ | 4436/18018 [9:37:24<28:29:34, 7.55s/step, loss=1.625, acc=0.650, acc_len=1.53, thru=17.8, I=21.6, T=29.8, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:07:38,000] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4435: forward=3012.9ms backward=1225.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:38,350] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:07:45,595] loop.py:45 INFO TRAIN_STEP step=4437 loss=1.541716 Training: 25%|██▍ | 4437/18018 [9:37:31<28:32:09, 7.56s/step, loss=1.542, acc=0.663, acc_len=1.56, thru=13.7, I=18.1, T=27.8, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:51,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:07:52,828] loop.py:45 INFO TRAIN_STEP step=4438 loss=1.662816 Training: 25%|██▍ | 4438/18018 [9:37:38<28:09:28, 7.46s/step, loss=1.663, acc=0.641, acc_len=1.50, thru=18.2, I=17.5, T=28.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:07:53,234] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:08:00,648] loop.py:45 INFO TRAIN_STEP step=4439 loss=1.576105 Training: 25%|██▍ | 4439/18018 [9:37:46<28:33:27, 7.57s/step, loss=1.576, acc=0.659, acc_len=1.55, thru=14.0, I=14.8, T=29.7, wait=3.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:06,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:08:08,227] loop.py:380 INFO TIMING step=4440: step=4.829s data=2.314s compute=4.794s [fwd=3.345s bwd=1.434s opt=0.003s] dispatch=2.740s [2026-10-18 23:08:08,228] loop.py:45 INFO TRAIN_STEP step=4440 loss=1.552269 Training: 25%|██▍ | 4440/18018 [9:37:54<28:33:51, 7.57s/step, loss=1.552, acc=0.662, acc_len=1.56, thru=17.7, I=18.6, T=26.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:08,530] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:15,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:08:16,666] loop.py:45 INFO TRAIN_STEP step=4441 loss=1.567444 Training: 25%|██▍ | 4441/18018 [9:38:02<29:32:25, 7.83s/step, loss=1.567, acc=0.658, acc_len=1.55, thru=15.7, I=17.5, T=24.5, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:08:16,661] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4440: forward=3504.0ms backward=1666.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:16,887] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:08:24,270] loop.py:45 INFO TRAIN_STEP step=4442 loss=1.587374 Training: 25%|██▍ | 4442/18018 [9:38:10<29:16:48, 7.76s/step, loss=1.587, acc=0.653, acc_len=1.53, thru=14.4, I=17.8, T=31.0, wait=3.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:30,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:08:31,498] loop.py:45 INFO TRAIN_STEP step=4443 loss=1.721271 Training: 25%|██▍ | 4443/18018 [9:38:17<28:40:13, 7.60s/step, loss=1.721, acc=0.634, acc_len=1.48, thru=17.2, I=19.5, T=31.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:31,528] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:35,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:08:38,335] loop.py:45 INFO TRAIN_STEP step=4444 loss=1.499525 Training: 25%|██▍ | 4444/18018 [9:38:24<27:48:18, 7.37s/step, loss=1.500, acc=0.666, acc_len=1.58, thru=17.0, I=25.6, T=28.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:38,374] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:45,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:08:46,197] loop.py:380 INFO TIMING step=4445: step=4.580s data=2.342s compute=4.548s [fwd=3.396s bwd=1.137s opt=0.003s] dispatch=3.270s [2026-10-18 23:08:46,198] loop.py:45 INFO TRAIN_STEP step=4445 loss=1.488682 Training: 25%|██▍ | 4445/18018 [9:38:32<28:21:17, 7.52s/step, loss=1.489, acc=0.671, acc_len=1.60, thru=16.4, I=17.1, T=27.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:08:46,256] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 23:08:53,822] loop.py:45 INFO TRAIN_STEP step=4446 loss=1.707406 Training: 25%|██▍ | 4446/18018 [9:38:39<28:28:02, 7.55s/step, loss=1.707, acc=0.637, acc_len=1.49, thru=15.3, I=16.4, T=31.4, wait=3.5s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:08:53,818] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4445: forward=2830.9ms backward=1190.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:02,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:09:02,459] loop.py:45 INFO TRAIN_STEP step=4447 loss=1.523350 Training: 25%|██▍ | 4447/18018 [9:38:48<29:41:42, 7.88s/step, loss=1.523, acc=0.666, acc_len=1.58, thru=16.3, I=17.1, T=31.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:02,643] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:11,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:09:11,765] loop.py:45 INFO TRAIN_STEP step=4448 loss=1.468427 Training: 25%|██▍ | 4448/18018 [9:38:57<31:18:27, 8.31s/step, loss=1.468, acc=0.675, acc_len=1.61, thru=13.8, I=13.8, T=30.0, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:12,142] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:18,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:09:18,340] loop.py:45 INFO TRAIN_STEP step=4449 loss=1.559891 Training: 25%|██▍ | 4449/18018 [9:39:04<29:20:53, 7.79s/step, loss=1.560, acc=0.660, acc_len=1.56, thru=17.4, I=19.0, T=34.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:18,751] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:24,773] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:09:26,333] loop.py:380 INFO TIMING step=4450: step=4.159s data=1.833s compute=4.122s [fwd=2.924s bwd=1.183s opt=0.003s] dispatch=3.825s [2026-10-18 23:09:26,333] loop.py:45 INFO TRAIN_STEP step=4450 loss=1.531232 Training: 25%|██▍ | 4450/18018 [9:39:12<29:34:45, 7.85s/step, loss=1.531, acc=0.662, acc_len=1.57, thru=16.9, I=20.1, T=30.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:26,782] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:33,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:09:33,765] loop.py:45 INFO TRAIN_STEP step=4451 loss=1.562435 Training: 25%|██▍ | 4451/18018 [9:39:19<29:06:32, 7.72s/step, loss=1.562, acc=0.662, acc_len=1.56, thru=15.7, I=19.5, T=29.2, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:09:33,760] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4450: forward=3027.2ms backward=1302.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:34,000] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:42,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:09:43,430] loop.py:45 INFO TRAIN_STEP step=4452 loss=1.558440 Training: 25%|██▍ | 4452/18018 [9:39:29<31:17:56, 8.31s/step, loss=1.558, acc=0.661, acc_len=1.56, thru=13.3, I=15.0, T=31.6, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:43,549] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:48,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:09:50,151] loop.py:45 INFO TRAIN_STEP step=4453 loss=1.672396 Training: 25%|██▍ | 4453/18018 [9:39:36<29:30:21, 7.83s/step, loss=1.672, acc=0.641, acc_len=1.49, thru=18.7, I=19.4, T=28.6, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:50,443] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:58,345] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:09:59,269] loop.py:45 INFO TRAIN_STEP step=4454 loss=1.642567 Training: 25%|██▍ | 4454/18018 [9:39:45<30:57:30, 8.22s/step, loss=1.643, acc=0.649, acc_len=1.53, thru=14.5, I=15.7, T=26.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:09:59,350] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:04,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:10:06,331] loop.py:380 INFO TIMING step=4455: step=4.486s data=2.298s compute=4.457s [fwd=3.359s bwd=1.083s opt=0.003s] dispatch=2.567s [2026-10-18 23:10:06,332] loop.py:45 INFO TRAIN_STEP step=4455 loss=1.666617 Training: 25%|██▍ | 4455/18018 [9:39:52<29:39:11, 7.87s/step, loss=1.667, acc=0.640, acc_len=1.49, thru=16.3, I=22.4, T=28.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:06,533] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:11,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:10:12,245] loop.py:45 INFO TRAIN_STEP step=4456 loss=1.639212 Training: 25%|██▍ | 4456/18018 [9:39:58<27:26:16, 7.28s/step, loss=1.639, acc=0.648, acc_len=1.52, thru=18.0, I=26.1, T=33.3, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:10:12,240] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4455: forward=2713.1ms backward=1085.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:12,639] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:17,265] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:10:17,968] loop.py:45 INFO TRAIN_STEP step=4457 loss=1.518514 Training: 25%|██▍ | 4457/18018 [9:40:04<25:40:17, 6.81s/step, loss=1.519, acc=0.669, acc_len=1.59, thru=20.8, I=26.7, T=34.7, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:18,269] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:24,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:10:25,085] loop.py:45 INFO TRAIN_STEP step=4458 loss=1.562521 Training: 25%|██▍ | 4458/18018 [9:40:11<26:00:39, 6.91s/step, loss=1.563, acc=0.658, acc_len=1.55, thru=17.3, I=20.2, T=28.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:25,235] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:31,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:10:32,959] loop.py:45 INFO TRAIN_STEP step=4459 loss=1.707351 Training: 25%|██▍ | 4459/18018 [9:40:19<27:06:11, 7.20s/step, loss=1.707, acc=0.637, acc_len=1.48, thru=15.4, I=18.0, T=31.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:33,437] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:10:39,662] loop.py:380 INFO TIMING step=4460: step=4.016s data=1.930s compute=3.987s [fwd=2.947s bwd=1.026s opt=0.003s] dispatch=2.679s [2026-10-18 23:10:39,662] loop.py:45 INFO TRAIN_STEP step=4460 loss=1.521621 Training: 25%|██▍ | 4460/18018 [9:40:25<26:32:40, 7.05s/step, loss=1.522, acc=0.663, acc_len=1.57, thru=16.7, I=23.2, T=31.9, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:47,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:10:48,764] loop.py:45 INFO TRAIN_STEP step=4461 loss=1.718155 Training: 25%|██▍ | 4461/18018 [9:40:34<28:51:46, 7.66s/step, loss=1.718, acc=0.635, acc_len=1.48, thru=14.4, I=15.5, T=32.4, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:10:48,759] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4460: forward=2742.4ms backward=1151.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:49,041] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:53,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:10:55,274] loop.py:45 INFO TRAIN_STEP step=4462 loss=1.613490 Training: 25%|██▍ | 4462/18018 [9:40:41<27:33:21, 7.32s/step, loss=1.613, acc=0.648, acc_len=1.53, thru=17.8, I=24.9, T=31.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:10:55,704] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:02,621] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:11:03,610] loop.py:45 INFO TRAIN_STEP step=4463 loss=1.505263 Training: 25%|██▍ | 4463/18018 [9:40:49<28:42:23, 7.62s/step, loss=1.505, acc=0.673, acc_len=1.60, thru=13.6, I=17.8, T=29.8, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:03,644] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:09,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:11:10,632] loop.py:45 INFO TRAIN_STEP step=4464 loss=1.611302 Training: 25%|██▍ | 4464/18018 [9:40:56<28:01:19, 7.44s/step, loss=1.611, acc=0.652, acc_len=1.53, thru=17.6, I=19.1, T=32.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:11,049] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:17,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:11:18,782] loop.py:380 INFO TIMING step=4465: step=3.942s data=1.796s compute=3.892s [fwd=2.725s bwd=1.152s opt=0.003s] dispatch=4.199s [2026-10-18 23:11:18,782] loop.py:45 INFO TRAIN_STEP step=4465 loss=1.719490 Training: 25%|██▍ | 4465/18018 [9:41:04<28:49:07, 7.65s/step, loss=1.719, acc=0.634, acc_len=1.48, thru=15.3, I=19.8, T=32.5, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:19,125] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:24,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:11:25,364] loop.py:45 INFO TRAIN_STEP step=4466 loss=1.567869 Training: 25%|██▍ | 4466/18018 [9:41:11<27:36:26, 7.33s/step, loss=1.568, acc=0.658, acc_len=1.55, thru=18.0, I=23.4, T=32.2, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:11:25,359] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4465: forward=2779.3ms backward=1154.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:25,568] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:30,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:11:31,402] loop.py:45 INFO TRAIN_STEP step=4467 loss=1.489131 Training: 25%|██▍ | 4467/18018 [9:41:17<26:08:31, 6.94s/step, loss=1.489, acc=0.667, acc_len=1.58, thru=19.7, I=26.1, T=34.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:31,760] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:38,631] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:11:39,521] loop.py:45 INFO TRAIN_STEP step=4468 loss=1.540394 Training: 25%|██▍ | 4468/18018 [9:41:25<27:27:42, 7.30s/step, loss=1.540, acc=0.661, acc_len=1.56, thru=15.3, I=18.8, T=32.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:39,643] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:46,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:11:47,504] loop.py:45 INFO TRAIN_STEP step=4469 loss=1.560502 Training: 25%|██▍ | 4469/18018 [9:41:33<28:14:19, 7.50s/step, loss=1.561, acc=0.663, acc_len=1.56, thru=17.1, I=20.9, T=30.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:47,942] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:55,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:11:55,418] loop.py:380 INFO TIMING step=4470: step=4.234s data=1.927s compute=4.206s [fwd=2.985s bwd=1.206s opt=0.003s] dispatch=3.669s [2026-10-18 23:11:55,418] loop.py:45 INFO TRAIN_STEP step=4470 loss=1.491824 Training: 25%|██▍ | 4470/18018 [9:41:41<28:41:52, 7.63s/step, loss=1.492, acc=0.677, acc_len=1.61, thru=14.8, I=14.7, T=30.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:11:55,865] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:00,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:12:01,675] loop.py:45 INFO TRAIN_STEP step=4471 loss=1.493576 Training: 25%|██▍ | 4471/18018 [9:41:47<27:09:21, 7.22s/step, loss=1.494, acc=0.667, acc_len=1.58, thru=18.4, I=22.1, T=33.2, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:12:01,670] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4470: forward=2668.5ms backward=1134.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:02,174] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:09,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:12:10,123] loop.py:45 INFO TRAIN_STEP step=4472 loss=1.615550 Training: 25%|██▍ | 4472/18018 [9:41:56<28:32:27, 7.59s/step, loss=1.616, acc=0.653, acc_len=1.54, thru=13.2, I=19.1, T=32.2, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:10,280] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:15,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:12:17,281] loop.py:45 INFO TRAIN_STEP step=4473 loss=1.505698 Training: 25%|██▍ | 4473/18018 [9:42:03<28:03:24, 7.46s/step, loss=1.506, acc=0.672, acc_len=1.60, thru=16.6, I=21.8, T=30.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:17,503] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:22,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:12:24,118] loop.py:45 INFO TRAIN_STEP step=4474 loss=1.673607 Training: 25%|██▍ | 4474/18018 [9:42:10<27:21:14, 7.27s/step, loss=1.674, acc=0.635, acc_len=1.48, thru=17.4, I=23.4, T=28.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:24,510] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:30,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:12:30,276] loop.py:380 INFO TIMING step=4475: step=3.772s data=1.673s compute=3.730s [fwd=2.641s bwd=1.075s opt=0.003s] dispatch=2.377s [2026-10-18 23:12:30,276] loop.py:45 INFO TRAIN_STEP step=4475 loss=1.616324 Training: 25%|██▍ | 4475/18018 [9:42:16<26:05:42, 6.94s/step, loss=1.616, acc=0.650, acc_len=1.53, thru=17.6, I=21.3, T=33.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:30,712] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:39,834] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:12:39,936] loop.py:45 INFO TRAIN_STEP step=4476 loss=1.543440 Training: 25%|██▍ | 4476/18018 [9:42:26<29:10:00, 7.75s/step, loss=1.543, acc=0.665, acc_len=1.58, thru=13.8, I=16.3, T=26.7, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:12:39,931] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4475: forward=3371.7ms backward=1384.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:40,336] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:46,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:12:48,497] loop.py:45 INFO TRAIN_STEP step=4477 loss=1.683997 Training: 25%|██▍ | 4477/18018 [9:42:34<30:04:33, 8.00s/step, loss=1.684, acc=0.638, acc_len=1.49, thru=14.3, I=17.6, T=27.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:12:48,790] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 23:12:55,927] loop.py:45 INFO TRAIN_STEP step=4478 loss=1.512629 Training: 25%|██▍ | 4478/18018 [9:42:42<29:26:07, 7.83s/step, loss=1.513, acc=0.669, acc_len=1.58, thru=13.9, I=22.7, T=31.6, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:04,940] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:13:07,489] loop.py:45 INFO TRAIN_STEP step=4479 loss=1.794783 Training: 25%|██▍ | 4479/18018 [9:42:53<33:38:51, 8.95s/step, loss=1.795, acc=0.625, acc_len=1.44, thru=12.4, I=14.6, T=22.8, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:07,957] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:15,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:13:15,807] loop.py:380 INFO TIMING step=4480: step=4.135s data=1.936s compute=4.105s [fwd=2.909s bwd=1.182s opt=0.003s] dispatch=4.175s [2026-10-18 23:13:15,807] loop.py:45 INFO TRAIN_STEP step=4480 loss=1.522149 Training: 25%|██▍ | 4480/18018 [9:43:01<32:56:15, 8.76s/step, loss=1.522, acc=0.670, acc_len=1.59, thru=16.3, I=18.1, T=31.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:16,145] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:24,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:13:24,358] loop.py:45 INFO TRAIN_STEP step=4481 loss=1.542069 Training: 25%|██▍ | 4481/18018 [9:43:10<32:41:56, 8.70s/step, loss=1.542, acc=0.661, acc_len=1.56, thru=14.6, I=16.6, T=31.9, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:13:24,353] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4480: forward=2827.3ms backward=1135.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:24,819] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:32,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:13:32,226] loop.py:45 INFO TRAIN_STEP step=4482 loss=1.509579 Training: 25%|██▍ | 4482/18018 [9:43:18<31:45:43, 8.45s/step, loss=1.510, acc=0.667, acc_len=1.58, thru=16.8, I=16.2, T=27.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:32,639] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:38,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:13:38,852] loop.py:45 INFO TRAIN_STEP step=4483 loss=1.537540 Training: 25%|██▍ | 4483/18018 [9:43:24<29:42:17, 7.90s/step, loss=1.538, acc=0.663, acc_len=1.56, thru=18.7, I=18.8, T=34.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:39,183] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:45,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:13:45,842] loop.py:45 INFO TRAIN_STEP step=4484 loss=1.641709 Training: 25%|██▍ | 4484/18018 [9:43:31<28:40:41, 7.63s/step, loss=1.642, acc=0.642, acc_len=1.50, thru=18.8, I=19.1, T=32.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:45,929] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:53,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:13:53,540] loop.py:380 INFO TIMING step=4485: step=3.911s data=1.727s compute=3.883s [fwd=2.747s bwd=1.122s opt=0.003s] dispatch=3.776s [2026-10-18 23:13:53,540] loop.py:45 INFO TRAIN_STEP step=4485 loss=1.607742 Training: 25%|██▍ | 4485/18018 [9:43:39<28:45:15, 7.65s/step, loss=1.608, acc=0.651, acc_len=1.53, thru=16.8, I=17.4, T=32.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:53,582] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:13:59,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:14:00,302] loop.py:45 INFO TRAIN_STEP step=4486 loss=1.628616 Training: 25%|██▍ | 4486/18018 [9:43:46<27:45:09, 7.38s/step, loss=1.629, acc=0.649, acc_len=1.52, thru=19.4, I=19.6, T=32.7, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:14:00,296] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4485: forward=2790.3ms backward=1077.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:14:00,337] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-18 23:14:06,714] loop.py:45 INFO TRAIN_STEP step=4487 loss=1.583453 Training: 25%|██▍ | 4487/18018 [9:43:52<26:39:16, 7.09s/step, loss=1.583, acc=0.656, acc_len=1.54, thru=17.6, I=19.3, T=34.2, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:14:12,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:14:13,404] loop.py:45 INFO TRAIN_STEP step=4488 loss=1.438188 Training: 25%|██▍ | 4488/18018 [9:43:59<26:11:53, 6.97s/step, loss=1.438, acc=0.679, acc_len=1.63, thru=19.2, I=21.6, T=30.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:14:13,859] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:14:22,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:14:24,103] loop.py:45 INFO TRAIN_STEP step=4489 loss=1.549679 Training: 25%|██▍ | 4489/18018 [9:44:10<30:24:02, 8.09s/step, loss=1.550, acc=0.663, acc_len=1.57, thru=12.8, I=13.6, T=26.7, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:14:24,200] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-18 23:14:30,785] loop.py:380 INFO TIMING step=4490: step=3.962s data=1.698s compute=3.933s [fwd=2.758s bwd=1.160s opt=0.003s] dispatch=2.711s [2026-10-18 23:14:30,785] loop.py:45 INFO TRAIN_STEP step=4490 loss=1.661773 Training: 25%|██▍ | 4490/18018 [9:44:16<28:48:35, 7.67s/step, loss=1.662, acc=0.642, acc_len=1.49, thru=16.2, I=20.0, T=32.3, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:14:37,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:14:38,852] loop.py:45 INFO TRAIN_STEP step=4491 loss=1.718145 Training: 25%|██▍ | 4491/18018 [9:44:24<29:15:38, 7.79s/step, loss=1.718, acc=0.633, acc_len=1.47, thru=15.8, I=16.9, T=33.9, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:14:38,846] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4490: forward=2695.5ms backward=1035.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:14:38,994] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:14:45,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:14:47,392] loop.py:45 INFO TRAIN_STEP step=4492 loss=1.653644 Training: 25%|██▍ | 4492/18018 [9:44:33<30:06:29, 8.01s/step, loss=1.654, acc=0.649, acc_len=1.51, thru=14.3, I=18.2, T=28.5, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:14:47,474] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 [2026-10-18 23:14:56,049] loop.py:45 INFO TRAIN_STEP step=4493 loss=1.508125 Training: 25%|██▍ | 4493/18018 [9:44:42<30:49:45, 8.21s/step, loss=1.508, acc=0.665, acc_len=1.58, thru=14.5, I=15.9, T=31.3, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:05,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:15:05,439] loop.py:45 INFO TRAIN_STEP step=4494 loss=1.808133 Training: 25%|██▍ | 4494/18018 [9:44:51<32:09:38, 8.56s/step, loss=1.808, acc=0.620, acc_len=1.43, thru=14.7, I=13.2, T=30.6, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:05,820] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:12,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:15:13,038] loop.py:380 INFO TIMING step=4495: step=3.953s data=1.783s compute=3.901s [fwd=2.701s bwd=1.185s opt=0.003s] dispatch=3.638s [2026-10-18 23:15:13,038] loop.py:45 INFO TRAIN_STEP step=4495 loss=1.464247 Training: 25%|██▍ | 4495/18018 [9:44:59<31:04:27, 8.27s/step, loss=1.464, acc=0.676, acc_len=1.62, thru=16.0, I=15.4, T=32.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:13,377] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 23:15:21,244] loop.py:45 INFO TRAIN_STEP step=4496 loss=1.412748 Training: 25%|██▍ | 4496/18018 [9:45:07<31:00:01, 8.25s/step, loss=1.413, acc=0.683, acc_len=1.63, thru=16.2, I=18.7, T=27.2, wait=3.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:15:21,240] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4495: forward=3265.8ms backward=1398.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:26,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:15:29,292] loop.py:45 INFO TRAIN_STEP step=4497 loss=1.488787 Training: 25%|██▍ | 4497/18018 [9:45:15<30:45:52, 8.19s/step, loss=1.489, acc=0.672, acc_len=1.59, thru=16.3, I=19.2, T=28.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:29,499] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:35,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:15:37,113] loop.py:45 INFO TRAIN_STEP step=4498 loss=1.582798 Training: 25%|██▍ | 4498/18018 [9:45:23<30:20:38, 8.08s/step, loss=1.583, acc=0.656, acc_len=1.55, thru=16.8, I=20.2, T=30.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:37,392] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:42,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:15:45,445] loop.py:45 INFO TRAIN_STEP step=4499 loss=1.602037 Training: 25%|██▍ | 4499/18018 [9:45:31<30:37:39, 8.16s/step, loss=1.602, acc=0.652, acc_len=1.53, thru=14.0, I=24.6, T=23.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:45,692] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:50,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:15:51,090] loop.py:380 INFO TIMING step=4500: step=3.911s data=1.820s compute=3.881s [fwd=2.665s bwd=1.201s opt=0.003s] dispatch=1.724s [2026-10-18 23:15:51,090] loop.py:45 INFO TRAIN_STEP step=4500 loss=1.474925 Training: 25%|██▍ | 4500/18018 [9:45:37<27:47:46, 7.40s/step, loss=1.475, acc=0.678, acc_len=1.61, thru=16.5, I=28.0, T=32.7, wait=1.7s, pool=64, epoch=1/2][2026-10-18 23:15:51,093] loop.py:419 INFO Saving checkpoint at step 4500... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 23:15:54,556] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0004501 (TrainerActor pid=9310) [2026-10-18 23:15:54,563] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0003751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:55,354] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:15:55,857] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:04,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 23:15:54,555] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0004501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 23:15:55,734] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0003751 [2026-10-18 23:16:05,498] loop.py:45 INFO TRAIN_STEP step=4501 loss=1.617811 (TrainerActor pid=9310) [2026-10-18 23:16:05,493] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4500: forward=3619.5ms backward=1578.2ms Training: 25%|██▍ | 4501/18018 [9:45:51<35:41:15, 9.50s/step, loss=1.618, acc=0.649, acc_len=1.52, thru=13.3, I=15.4, T=24.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:05,824] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:10,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:16:12,598] loop.py:45 INFO TRAIN_STEP step=4502 loss=1.555426 Training: 25%|██▍ | 4502/18018 [9:45:58<32:58:28, 8.78s/step, loss=1.555, acc=0.659, acc_len=1.56, thru=16.4, I=23.0, T=28.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:12,682] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-18 23:16:20,902] loop.py:45 INFO TRAIN_STEP step=4503 loss=1.608888 Training: 25%|██▍ | 4503/18018 [9:46:06<32:25:51, 8.64s/step, loss=1.609, acc=0.650, acc_len=1.53, thru=15.6, I=15.3, T=28.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:20,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:21,323] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 23:16:27,439] loop.py:45 INFO TRAIN_STEP step=4504 loss=1.453918 Training: 25%|██▍ | 4504/18018 [9:46:13<30:03:49, 8.01s/step, loss=1.454, acc=0.682, acc_len=1.63, thru=14.0, I=17.5, T=31.9, wait=2.5s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:35,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:16:36,313] loop.py:380 INFO TIMING step=4505: step=4.382s data=2.093s compute=4.330s [fwd=3.030s bwd=1.284s opt=0.003s] dispatch=4.484s [2026-10-18 23:16:36,313] loop.py:45 INFO TRAIN_STEP step=4505 loss=1.497431 Training: 25%|██▌ | 4505/18018 [9:46:22<31:02:17, 8.27s/step, loss=1.497, acc=0.672, acc_len=1.60, thru=17.6, I=16.7, T=29.2, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:36,406] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 23:16:42,610] loop.py:45 INFO TRAIN_STEP step=4506 loss=1.625775 Training: 25%|██▌ | 4506/18018 [9:46:28<28:48:47, 7.68s/step, loss=1.626, acc=0.646, acc_len=1.51, thru=14.9, I=17.8, T=35.8, wait=2.7s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:16:42,604] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4505: forward=2531.6ms backward=987.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:50,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:16:51,783] loop.py:45 INFO TRAIN_STEP step=4507 loss=1.528530 Training: 25%|██▌ | 4507/18018 [9:46:37<30:29:53, 8.13s/step, loss=1.529, acc=0.660, acc_len=1.56, thru=17.1, I=17.2, T=25.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:51,819] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:57,002] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:16:57,966] loop.py:45 INFO TRAIN_STEP step=4508 loss=1.583795 Training: 25%|██▌ | 4508/18018 [9:46:44<28:18:29, 7.54s/step, loss=1.584, acc=0.656, acc_len=1.55, thru=17.2, I=21.4, T=33.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:16:58,023] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:03,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:17:05,365] loop.py:45 INFO TRAIN_STEP step=4509 loss=1.652406 Training: 25%|██▌ | 4509/18018 [9:46:51<28:08:27, 7.50s/step, loss=1.652, acc=0.644, acc_len=1.51, thru=16.1, I=19.7, T=31.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:05,427] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-18 23:17:12,580] loop.py:380 INFO TIMING step=4510: step=4.446s data=1.974s compute=4.414s [fwd=3.012s bwd=1.388s opt=0.003s] dispatch=2.760s [2026-10-18 23:17:12,580] loop.py:45 INFO TRAIN_STEP step=4510 loss=1.741921 Training: 25%|██▌ | 4510/18018 [9:46:58<27:49:12, 7.41s/step, loss=1.742, acc=0.630, acc_len=1.45, thru=15.1, I=17.7, T=28.8, wait=2.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:19,708] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:17:21,160] loop.py:45 INFO TRAIN_STEP step=4511 loss=1.552953 Training: 25%|██▌ | 4511/18018 [9:47:07<29:07:42, 7.76s/step, loss=1.553, acc=0.664, acc_len=1.57, thru=16.6, I=16.6, T=27.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:21,217] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (TrainerActor pid=9310) [2026-10-18 23:17:21,155] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4510: forward=3157.2ms backward=1477.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:26,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:17:27,682] loop.py:45 INFO TRAIN_STEP step=4512 loss=1.573256 Training: 25%|██▌ | 4512/18018 [9:47:13<27:43:58, 7.39s/step, loss=1.573, acc=0.656, acc_len=1.54, thru=18.1, I=21.5, T=30.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:27,849] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:32,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:17:34,459] loop.py:45 INFO TRAIN_STEP step=4513 loss=1.560636 Training: 25%|██▌ | 4513/18018 [9:47:20<27:02:05, 7.21s/step, loss=1.561, acc=0.659, acc_len=1.56, thru=17.2, I=24.7, T=27.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:34,571] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-18 23:17:42,540] loop.py:45 INFO TRAIN_STEP step=4514 loss=1.542508 Training: 25%|██▌ | 4514/18018 [9:47:28<28:00:59, 7.47s/step, loss=1.543, acc=0.660, acc_len=1.57, thru=10.0, I=16.2, T=31.5, wait=4.0s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:51,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:17:51,783] loop.py:380 INFO TIMING step=4515: step=4.790s data=2.287s compute=4.754s [fwd=3.252s bwd=1.488s opt=0.003s] dispatch=4.445s [2026-10-18 23:17:51,784] loop.py:45 INFO TRAIN_STEP step=4515 loss=1.643190 Training: 25%|██▌ | 4515/18018 [9:47:37<30:00:42, 8.00s/step, loss=1.643, acc=0.646, acc_len=1.51, thru=17.5, I=13.5, T=26.7, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:17:51,973] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 23:18:00,020] loop.py:45 INFO TRAIN_STEP step=4516 loss=1.530421 Training: 25%|██▌ | 4516/18018 [9:47:46<30:16:35, 8.07s/step, loss=1.530, acc=0.666, acc_len=1.58, thru=15.6, I=15.6, T=33.7, wait=4.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:18:00,015] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4515: forward=2639.4ms backward=1115.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:10,485] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:18:10,741] loop.py:45 INFO TRAIN_STEP step=4517 loss=1.649684 Training: 25%|██▌ | 4517/18018 [9:47:56<33:15:14, 8.87s/step, loss=1.650, acc=0.646, acc_len=1.51, thru=12.0, I=12.0, T=26.9, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:10,988] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:18,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:18:19,142] loop.py:45 INFO TRAIN_STEP step=4518 loss=1.408366 Training: 25%|██▌ | 4518/18018 [9:48:05<32:43:38, 8.73s/step, loss=1.408, acc=0.685, acc_len=1.64, thru=15.2, I=16.5, T=25.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:19,479] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:27,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:18:29,139] loop.py:45 INFO TRAIN_STEP step=4519 loss=1.573840 Training: 25%|██▌ | 4519/18018 [9:48:15<34:08:58, 9.11s/step, loss=1.574, acc=0.657, acc_len=1.55, thru=13.2, I=15.4, T=26.7, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:29,592] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:36,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:18:36,831] loop.py:380 INFO TIMING step=4520: step=4.536s data=2.168s compute=4.496s [fwd=3.155s bwd=1.327s opt=0.003s] dispatch=3.149s [2026-10-18 23:18:36,831] loop.py:45 INFO TRAIN_STEP step=4520 loss=1.652913 Training: 25%|██▌ | 4520/18018 [9:48:22<32:33:26, 8.68s/step, loss=1.653, acc=0.644, acc_len=1.50, thru=13.9, I=19.1, T=28.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:37,025] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:42,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:18:44,077] loop.py:45 INFO TRAIN_STEP step=4521 loss=1.663346 (TrainerActor pid=9310) [2026-10-18 23:18:44,072] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4520: forward=2982.5ms backward=1223.2ms Training: 25%|██▌ | 4521/18018 [9:48:30<30:56:23, 8.25s/step, loss=1.663, acc=0.641, acc_len=1.50, thru=18.2, I=19.8, T=30.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:44,297] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:49,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:18:50,706] loop.py:45 INFO TRAIN_STEP step=4522 loss=1.519312 Training: 25%|██▌ | 4522/18018 [9:48:36<29:06:36, 7.77s/step, loss=1.519, acc=0.662, acc_len=1.56, thru=17.9, I=25.1, T=32.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:18:51,006] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:00,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:19:00,351] loop.py:45 INFO TRAIN_STEP step=4523 loss=1.559195 Training: 25%|██▌ | 4523/18018 [9:48:46<31:13:15, 8.33s/step, loss=1.559, acc=0.660, acc_len=1.56, thru=13.7, I=12.2, T=28.6, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:00,738] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 23:19:07,080] loop.py:45 INFO TRAIN_STEP step=4524 loss=1.619138 Training: 25%|██▌ | 4524/18018 [9:48:53<29:25:11, 7.85s/step, loss=1.619, acc=0.646, acc_len=1.52, thru=17.2, I=17.5, T=32.1, wait=2.7s, pool=60, epoch=1/2][2026-10-18 23:19:14,746] loop.py:380 INFO TIMING step=4525: step=4.081s data=1.795s compute=4.052s [fwd=2.904s bwd=1.134s opt=0.003s] dispatch=3.576s [2026-10-18 23:19:14,746] loop.py:45 INFO TRAIN_STEP step=4525 loss=1.614886 Training: 25%|██▌ | 4525/18018 [9:49:00<29:12:56, 7.79s/step, loss=1.615, acc=0.651, acc_len=1.52, thru=17.6, I=18.1, T=31.4, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:22,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:19:23,050] loop.py:45 INFO TRAIN_STEP step=4526 loss=1.610509 Training: 25%|██▌ | 4526/18018 [9:49:09<29:46:58, 7.95s/step, loss=1.611, acc=0.648, acc_len=1.52, thru=15.1, I=16.0, T=27.3, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:19:23,045] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4525: forward=3279.1ms backward=1367.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:23,519] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:27,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:19:30,772] loop.py:45 INFO TRAIN_STEP step=4527 loss=1.599682 Training: 25%|██▌ | 4527/18018 [9:49:16<29:31:41, 7.88s/step, loss=1.600, acc=0.652, acc_len=1.53, thru=15.0, I=20.8, T=25.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:30,887] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:35,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:19:37,170] loop.py:45 INFO TRAIN_STEP step=4528 loss=1.596343 Training: 25%|██▌ | 4528/18018 [9:49:23<27:51:44, 7.44s/step, loss=1.596, acc=0.655, acc_len=1.54, thru=14.7, I=24.0, T=33.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:37,330] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:44,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:19:45,718] loop.py:45 INFO TRAIN_STEP step=4529 loss=1.609136 Training: 25%|██▌ | 4529/18018 [9:49:31<29:06:32, 7.77s/step, loss=1.609, acc=0.648, acc_len=1.52, thru=15.3, I=18.6, T=29.9, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:46,051] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:19:53,041] loop.py:380 INFO TIMING step=4530: step=4.208s data=2.035s compute=4.170s [fwd=2.881s bwd=1.275s opt=0.003s] dispatch=3.106s [2026-10-18 23:19:53,041] loop.py:45 INFO TRAIN_STEP step=4530 loss=1.501721 Training: 25%|██▌ | 4530/18018 [9:49:39<28:36:28, 7.64s/step, loss=1.502, acc=0.668, acc_len=1.58, thru=15.3, I=19.8, T=30.4, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:19:59,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:20:01,348] loop.py:45 INFO TRAIN_STEP step=4531 loss=1.580499 Training: 25%|██▌ | 4531/18018 [9:49:47<29:21:30, 7.84s/step, loss=1.580, acc=0.655, acc_len=1.54, thru=16.4, I=18.8, T=31.7, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:20:01,343] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4530: forward=2856.1ms backward=1123.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:20:01,608] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:20:08,030] loop.py:45 INFO TRAIN_STEP step=4532 loss=1.489304 Training: 25%|██▌ | 4532/18018 [9:49:54<28:03:28, 7.49s/step, loss=1.489, acc=0.671, acc_len=1.59, thru=15.3, I=20.1, T=33.5, wait=2.9s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:20:14,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:20:16,480] loop.py:45 INFO TRAIN_STEP step=4533 loss=1.602216 Training: 25%|██▌ | 4533/18018 [9:50:02<29:08:03, 7.78s/step, loss=1.602, acc=0.652, acc_len=1.53, thru=16.8, I=17.9, T=31.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:20:16,895] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:20:23,206] loop.py:45 INFO TRAIN_STEP step=4534 loss=1.503878 Training: 25%|██▌ | 4534/18018 [9:50:09<27:57:02, 7.46s/step, loss=1.504, acc=0.669, acc_len=1.58, thru=16.1, I=19.4, T=33.3, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:20:29,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:20:31,130] loop.py:380 INFO TIMING step=4535: step=4.362s data=2.063s compute=4.323s [fwd=3.038s bwd=1.270s opt=0.003s] dispatch=3.553s [2026-10-18 23:20:31,130] loop.py:45 INFO TRAIN_STEP step=4535 loss=1.672325 Training: 25%|██▌ | 4535/18018 [9:50:17<28:28:05, 7.60s/step, loss=1.672, acc=0.636, acc_len=1.48, thru=18.0, I=18.5, T=29.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:20:31,496] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:20:37,182] loop.py:45 INFO TRAIN_STEP step=4536 loss=1.583205 Training: 25%|██▌ | 4536/18018 [9:50:23<26:43:27, 7.14s/step, loss=1.583, acc=0.653, acc_len=1.53, thru=16.8, I=19.8, T=36.5, wait=2.5s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:20:37,177] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4535: forward=2471.5ms backward=986.3ms [2026-10-18 23:20:44,912] loop.py:45 INFO TRAIN_STEP step=4537 loss=1.529028 Training: 25%|██▌ | 4537/18018 [9:50:30<27:23:25, 7.31s/step, loss=1.529, acc=0.666, acc_len=1.58, thru=16.0, I=17.1, T=30.9, wait=3.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:20:52,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:20:53,284] loop.py:45 INFO TRAIN_STEP step=4538 loss=1.628063 Training: 25%|██▌ | 4538/18018 [9:50:39<28:34:40, 7.63s/step, loss=1.628, acc=0.649, acc_len=1.52, thru=16.3, I=15.9, T=30.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:20:53,599] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:21:01,913] loop.py:45 INFO TRAIN_STEP step=4539 loss=1.652313 Training: 25%|██▌ | 4539/18018 [9:50:47<29:41:32, 7.93s/step, loss=1.652, acc=0.644, acc_len=1.51, thru=13.6, I=14.1, T=29.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:21:01,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:21:02,416] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:21:10,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:21:11,376] loop.py:380 INFO TIMING step=4540: step=4.541s data=2.028s compute=4.517s [fwd=3.234s bwd=1.269s opt=0.003s] dispatch=4.915s [2026-10-18 23:21:11,377] loop.py:45 INFO TRAIN_STEP step=4540 loss=1.525903 Training: 25%|██▌ | 4540/18018 [9:50:57<31:24:48, 8.39s/step, loss=1.526, acc=0.666, acc_len=1.57, thru=14.5, I=16.3, T=28.2, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:21:11,837] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:21:21,300] loop.py:45 INFO TRAIN_STEP step=4541 loss=1.506100 Training: 25%|██▌ | 4541/18018 [9:51:07<33:07:59, 8.85s/step, loss=1.506, acc=0.669, acc_len=1.59, thru=11.4, I=11.1, T=29.4, wait=5.6s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:21:21,295] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4540: forward=3094.4ms backward=1206.8ms [2026-10-18 23:21:31,934] loop.py:45 INFO TRAIN_STEP step=4542 loss=1.549726 Training: 25%|██▌ | 4542/18018 [9:51:18<35:08:10, 9.39s/step, loss=1.550, acc=0.661, acc_len=1.55, thru=10.9, I=11.1, T=29.2, wait=6.2s, pool=36, epoch=1/2][2026-10-18 23:21:39,782] loop.py:45 INFO TRAIN_STEP step=4543 loss=1.629803 Training: 25%|██▌ | 4543/18018 [9:51:25<33:24:26, 8.93s/step, loss=1.630, acc=0.652, acc_len=1.53, thru=18.5, I=16.9, T=30.4, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:21:45,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:21:47,702] loop.py:45 INFO TRAIN_STEP step=4544 loss=1.701703 Training: 25%|██▌ | 4544/18018 [9:51:33<32:16:20, 8.62s/step, loss=1.702, acc=0.638, acc_len=1.49, thru=16.9, I=19.8, T=27.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:21:48,197] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 23:21:54,886] loop.py:380 INFO TIMING step=4545: step=3.701s data=1.582s compute=3.673s [fwd=2.603s bwd=1.057s opt=0.003s] dispatch=3.474s [2026-10-18 23:21:54,886] loop.py:45 INFO TRAIN_STEP step=4545 loss=1.633676 Training: 25%|██▌ | 4545/18018 [9:51:40<30:39:19, 8.19s/step, loss=1.634, acc=0.648, acc_len=1.51, thru=13.8, I=19.4, T=34.6, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:04,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:22:05,337] loop.py:45 INFO TRAIN_STEP step=4546 loss=1.730451 Training: 25%|██▌ | 4546/18018 [9:51:51<33:11:23, 8.87s/step, loss=1.730, acc=0.633, acc_len=1.47, thru=14.9, I=15.0, T=28.5, wait=5.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:22:05,333] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4545: forward=3176.7ms backward=1253.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:05,547] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:22:11,589] loop.py:45 INFO TRAIN_STEP step=4547 loss=1.504999 Training: 25%|██▌ | 4547/18018 [9:51:57<30:15:05, 8.08s/step, loss=1.505, acc=0.670, acc_len=1.59, thru=15.4, I=18.7, T=32.6, wait=2.3s, pool=36, epoch=1/2][2026-10-18 23:22:22,239] loop.py:45 INFO TRAIN_STEP step=4548 loss=1.754021 Training: 25%|██▌ | 4548/18018 [9:52:08<33:07:56, 8.85s/step, loss=1.754, acc=0.634, acc_len=1.47, thru=13.2, I=11.9, T=31.8, wait=6.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:29,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:22:31,087] loop.py:45 INFO TRAIN_STEP step=4549 loss=1.532232 Training: 25%|██▌ | 4549/18018 [9:52:17<33:07:06, 8.85s/step, loss=1.532, acc=0.661, acc_len=1.57, thru=17.4, I=19.5, T=29.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:31,451] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:22:36,974] loop.py:380 INFO TIMING step=4550: step=3.472s data=1.400s compute=3.440s [fwd=2.452s bwd=0.974s opt=0.003s] dispatch=2.404s [2026-10-18 23:22:36,974] loop.py:45 INFO TRAIN_STEP step=4550 loss=1.621619 Training: 25%|██▌ | 4550/18018 [9:52:23<29:47:26, 7.96s/step, loss=1.622, acc=0.652, acc_len=1.53, thru=19.3, I=24.2, T=36.9, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:42,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:22:43,611] loop.py:45 INFO TRAIN_STEP step=4551 loss=1.646989 Training: 25%|██▌ | 4551/18018 [9:52:29<28:17:54, 7.56s/step, loss=1.647, acc=0.643, acc_len=1.49, thru=18.6, I=20.6, T=32.5, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:22:43,606] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4550: forward=2794.9ms backward=1096.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:43,815] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:50,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:22:51,929] loop.py:45 INFO TRAIN_STEP step=4552 loss=1.557140 Training: 25%|██▌ | 4552/18018 [9:52:38<29:08:26, 7.79s/step, loss=1.557, acc=0.656, acc_len=1.55, thru=15.0, I=19.0, T=28.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:52,123] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:58,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:22:59,679] loop.py:45 INFO TRAIN_STEP step=4553 loss=1.458587 Training: 25%|██▌ | 4553/18018 [9:52:45<29:05:29, 7.78s/step, loss=1.459, acc=0.675, acc_len=1.61, thru=14.4, I=20.5, T=27.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:22:59,886] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:06,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:23:08,679] loop.py:45 INFO TRAIN_STEP step=4554 loss=1.597399 Training: 25%|██▌ | 4554/18018 [9:52:54<30:27:40, 8.14s/step, loss=1.597, acc=0.651, acc_len=1.53, thru=14.6, I=15.2, T=30.3, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:08,950] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:15,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:23:15,925] loop.py:380 INFO TIMING step=4555: step=4.185s data=1.997s compute=4.133s [fwd=2.906s bwd=1.212s opt=0.003s] dispatch=3.052s [2026-10-18 23:23:15,925] loop.py:45 INFO TRAIN_STEP step=4555 loss=1.497369 Training: 25%|██▌ | 4555/18018 [9:53:02<29:27:02, 7.88s/step, loss=1.497, acc=0.670, acc_len=1.59, thru=15.9, I=18.8, T=30.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:16,057] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:21,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:23:22,859] loop.py:45 INFO TRAIN_STEP step=4556 loss=1.610634 Training: 25%|██▌ | 4556/18018 [9:53:08<28:23:34, 7.59s/step, loss=1.611, acc=0.649, acc_len=1.52, thru=18.0, I=19.5, T=28.9, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:23:22,855] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4555: forward=3080.1ms backward=1303.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:23,340] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:28,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:23:29,353] loop.py:45 INFO TRAIN_STEP step=4557 loss=1.552480 Training: 25%|██▌ | 4557/18018 [9:53:15<27:09:34, 7.26s/step, loss=1.552, acc=0.658, acc_len=1.55, thru=17.6, I=21.0, T=32.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:29,553] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:34,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:23:35,419] loop.py:45 INFO TRAIN_STEP step=4558 loss=1.667481 Training: 25%|██▌ | 4558/18018 [9:53:21<25:48:50, 6.90s/step, loss=1.667, acc=0.643, acc_len=1.50, thru=19.6, I=22.7, T=32.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:35,549] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:40,744] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:23:42,319] loop.py:45 INFO TRAIN_STEP step=4559 loss=1.590109 Training: 25%|██▌ | 4559/18018 [9:53:28<25:48:28, 6.90s/step, loss=1.590, acc=0.658, acc_len=1.55, thru=17.2, I=22.9, T=28.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:42,756] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:49,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:23:51,182] loop.py:380 INFO TIMING step=4560: step=5.826s data=3.266s compute=5.786s [fwd=4.247s bwd=1.524s opt=0.003s] dispatch=3.026s [2026-10-18 23:23:51,182] loop.py:45 INFO TRAIN_STEP step=4560 loss=1.443047 Training: 25%|██▌ | 4560/18018 [9:53:37<28:00:07, 7.49s/step, loss=1.443, acc=0.678, acc_len=1.61, thru=15.2, I=17.7, T=22.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:23:51,301] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 23:23:57,296] loop.py:45 INFO TRAIN_STEP step=4561 loss=1.562754 (TrainerActor pid=9310) [2026-10-18 23:23:57,290] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4560: forward=2748.0ms backward=1115.8ms Training: 25%|██▌ | 4561/18018 [9:53:43<26:27:29, 7.08s/step, loss=1.563, acc=0.657, acc_len=1.55, thru=17.2, I=19.7, T=32.7, wait=2.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:03,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:24:04,963] loop.py:45 INFO TRAIN_STEP step=4562 loss=1.647924 Training: 25%|██▌ | 4562/18018 [9:53:51<27:06:55, 7.25s/step, loss=1.648, acc=0.649, acc_len=1.52, thru=17.1, I=17.9, T=32.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:05,212] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:24:12,606] loop.py:45 INFO TRAIN_STEP step=4563 loss=1.714267 Training: 25%|██▌ | 4563/18018 [9:53:58<27:32:55, 7.37s/step, loss=1.714, acc=0.637, acc_len=1.48, thru=13.3, I=18.3, T=32.2, wait=3.7s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:18,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:24:20,541] loop.py:45 INFO TRAIN_STEP step=4564 loss=1.496574 Training: 25%|██▌ | 4564/18018 [9:54:06<28:10:48, 7.54s/step, loss=1.497, acc=0.672, acc_len=1.59, thru=17.5, I=19.0, T=31.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:20,577] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:27,485] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:24:27,922] loop.py:380 INFO TIMING step=4565: step=3.937s data=1.707s compute=3.902s [fwd=2.760s bwd=1.127s opt=0.003s] dispatch=3.434s [2026-10-18 23:24:27,922] loop.py:45 INFO TRAIN_STEP step=4565 loss=1.591005 Training: 25%|██▌ | 4565/18018 [9:54:14<28:00:00, 7.49s/step, loss=1.591, acc=0.655, acc_len=1.54, thru=14.6, I=18.7, T=32.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:28,005] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:32,415] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:24:33,694] loop.py:45 INFO TRAIN_STEP step=4566 loss=1.574953 Training: 25%|██▌ | 4566/18018 [9:54:19<26:04:00, 6.98s/step, loss=1.575, acc=0.655, acc_len=1.54, thru=20.1, I=21.6, T=33.0, wait=1.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:24:33,688] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4565: forward=2766.9ms backward=1067.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:33,924] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:40,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:24:41,456] loop.py:45 INFO TRAIN_STEP step=4567 loss=1.728559 Training: 25%|██▌ | 4567/18018 [9:54:27<26:56:55, 7.21s/step, loss=1.729, acc=0.629, acc_len=1.47, thru=15.9, I=20.0, T=28.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:41,764] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:49,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:24:50,921] loop.py:45 INFO TRAIN_STEP step=4568 loss=1.616774 Training: 25%|██▌ | 4568/18018 [9:54:37<29:28:08, 7.89s/step, loss=1.617, acc=0.650, acc_len=1.53, thru=14.0, I=15.8, T=27.6, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:24:51,093] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:24:59,280] loop.py:45 INFO TRAIN_STEP step=4569 loss=1.426788 Training: 25%|██▌ | 4569/18018 [9:54:45<29:59:45, 8.03s/step, loss=1.427, acc=0.683, acc_len=1.63, thru=12.1, I=13.3, T=31.0, wait=4.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:06,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:25:07,084] loop.py:380 INFO TIMING step=4570: step=4.070s data=1.814s compute=4.027s [fwd=2.865s bwd=1.147s opt=0.003s] dispatch=3.724s [2026-10-18 23:25:07,085] loop.py:45 INFO TRAIN_STEP step=4570 loss=1.556952 Training: 25%|██▌ | 4570/18018 [9:54:53<29:44:24, 7.96s/step, loss=1.557, acc=0.658, acc_len=1.55, thru=18.8, I=17.4, T=31.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:07,554] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:15,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:25:16,890] loop.py:45 INFO TRAIN_STEP step=4571 loss=1.627779 Training: 25%|██▌ | 4571/18018 [9:55:02<31:48:15, 8.51s/step, loss=1.628, acc=0.649, acc_len=1.52, thru=13.7, I=14.7, T=28.6, wait=5.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:25:16,885] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4570: forward=3160.7ms backward=1279.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:17,129] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:24,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:25:26,247] loop.py:45 INFO TRAIN_STEP step=4572 loss=1.697377 Training: 25%|██▌ | 4572/18018 [9:55:12<32:44:52, 8.77s/step, loss=1.697, acc=0.638, acc_len=1.49, thru=14.0, I=15.7, T=27.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:26,502] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:32,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:25:33,848] loop.py:45 INFO TRAIN_STEP step=4573 loss=1.516156 Training: 25%|██▌ | 4573/18018 [9:55:19<31:26:12, 8.42s/step, loss=1.516, acc=0.666, acc_len=1.58, thru=15.2, I=19.1, T=30.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:33,880] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 23:25:41,384] loop.py:45 INFO TRAIN_STEP step=4574 loss=1.655269 Training: 25%|██▌ | 4574/18018 [9:55:27<30:26:50, 8.15s/step, loss=1.655, acc=0.646, acc_len=1.51, thru=12.9, I=20.4, T=28.1, wait=3.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:50,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:25:52,011] loop.py:380 INFO TIMING step=4575: step=4.645s data=2.263s compute=4.600s [fwd=3.305s bwd=1.280s opt=0.003s] dispatch=5.971s [2026-10-18 23:25:52,011] loop.py:45 INFO TRAIN_STEP step=4575 loss=1.648653 Training: 25%|██▌ | 4575/18018 [9:55:38<33:12:55, 8.89s/step, loss=1.649, acc=0.645, acc_len=1.51, thru=14.6, I=13.9, T=27.6, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:52,176] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:59,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:25:59,520] loop.py:45 INFO TRAIN_STEP step=4576 loss=1.715911 Training: 25%|██▌ | 4576/18018 [9:55:45<31:39:36, 8.48s/step, loss=1.716, acc=0.636, acc_len=1.48, thru=15.4, I=17.4, T=32.6, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:25:59,515] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4575: forward=2781.8ms backward=1096.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:25:59,920] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:06,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:26:07,256] loop.py:45 INFO TRAIN_STEP step=4577 loss=1.601905 Training: 25%|██▌ | 4577/18018 [9:55:53<30:49:32, 8.26s/step, loss=1.602, acc=0.651, acc_len=1.54, thru=17.1, I=18.5, T=27.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:07,744] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:13,974] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:26:14,243] loop.py:45 INFO TRAIN_STEP step=4578 loss=1.570293 Training: 25%|██▌ | 4578/18018 [9:56:00<29:24:09, 7.88s/step, loss=1.570, acc=0.652, acc_len=1.54, thru=16.5, I=20.0, T=33.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:14,476] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:19,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:26:21,924] loop.py:45 INFO TRAIN_STEP step=4579 loss=1.651878 Training: 25%|██▌ | 4579/18018 [9:56:08<29:10:55, 7.82s/step, loss=1.652, acc=0.648, acc_len=1.51, thru=16.5, I=22.1, T=27.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:22,315] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:29,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:26:30,467] loop.py:380 INFO TIMING step=4580: step=3.997s data=1.767s compute=3.967s [fwd=2.850s bwd=1.101s opt=0.003s] dispatch=4.536s [2026-10-18 23:26:30,467] loop.py:45 INFO TRAIN_STEP step=4580 loss=1.543207 Training: 25%|██▌ | 4580/18018 [9:56:16<29:59:29, 8.03s/step, loss=1.543, acc=0.662, acc_len=1.57, thru=15.7, I=17.6, T=32.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:30,919] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:26:37,892] loop.py:45 INFO TRAIN_STEP step=4581 loss=1.654305 Training: 25%|██▌ | 4581/18018 [9:56:23<29:18:25, 7.85s/step, loss=1.654, acc=0.643, acc_len=1.50, thru=15.7, I=16.9, T=30.8, wait=3.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:26:37,887] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4580: forward=2945.3ms backward=1171.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:44,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:26:46,044] loop.py:45 INFO TRAIN_STEP step=4582 loss=1.592942 Training: 25%|██▌ | 4582/18018 [9:56:32<29:38:38, 7.94s/step, loss=1.593, acc=0.655, acc_len=1.54, thru=17.1, I=19.1, T=26.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:46,553] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 23:26:52,508] loop.py:45 INFO TRAIN_STEP step=4583 loss=1.543658 Training: 25%|██▌ | 4583/18018 [9:56:38<27:59:08, 7.50s/step, loss=1.544, acc=0.667, acc_len=1.57, thru=14.9, I=21.9, T=36.5, wait=2.9s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:26:59,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:27:00,329] loop.py:45 INFO TRAIN_STEP step=4584 loss=1.644432 Training: 25%|██▌ | 4584/18018 [9:56:46<28:20:40, 7.60s/step, loss=1.644, acc=0.648, acc_len=1.52, thru=17.8, I=17.7, T=35.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:00,701] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 23:27:06,298] loop.py:380 INFO TIMING step=4585: step=3.476s data=1.421s compute=3.434s [fwd=2.446s bwd=0.974s opt=0.003s] dispatch=2.482s [2026-10-18 23:27:06,298] loop.py:45 INFO TRAIN_STEP step=4585 loss=1.676260 Training: 25%|██▌ | 4585/18018 [9:56:52<26:31:08, 7.11s/step, loss=1.676, acc=0.634, acc_len=1.47, thru=18.4, I=22.8, T=36.8, wait=2.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:16,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:27:17,832] loop.py:45 INFO TRAIN_STEP step=4586 loss=1.703292 Training: 25%|██▌ | 4586/18018 [9:57:03<31:28:28, 8.44s/step, loss=1.703, acc=0.637, acc_len=1.48, thru=13.8, I=14.0, T=28.5, wait=7.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:27:17,827] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4585: forward=3193.3ms backward=1252.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:18,132] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:27:26,458] loop.py:45 INFO TRAIN_STEP step=4587 loss=1.464027 Training: 25%|██▌ | 4587/18018 [9:57:12<31:40:56, 8.49s/step, loss=1.464, acc=0.675, acc_len=1.60, thru=13.7, I=15.4, T=31.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:26,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:26,937] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:34,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:27:36,293] loop.py:45 INFO TRAIN_STEP step=4588 loss=1.551479 Training: 25%|██▌ | 4588/18018 [9:57:22<33:11:01, 8.90s/step, loss=1.551, acc=0.664, acc_len=1.57, thru=13.4, I=16.0, T=31.4, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:36,340] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 23:27:43,161] loop.py:45 INFO TRAIN_STEP step=4589 loss=1.484323 Training: 25%|██▌ | 4589/18018 [9:57:29<30:54:49, 8.29s/step, loss=1.484, acc=0.673, acc_len=1.60, thru=15.6, I=18.8, T=32.0, wait=2.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:49,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:27:51,082] loop.py:380 INFO TIMING step=4590: step=4.173s data=1.882s compute=4.144s [fwd=2.895s bwd=1.234s opt=0.003s] dispatch=3.738s [2026-10-18 23:27:51,083] loop.py:45 INFO TRAIN_STEP step=4590 loss=1.479443 Training: 25%|██▌ | 4590/18018 [9:57:37<30:30:11, 8.18s/step, loss=1.479, acc=0.674, acc_len=1.60, thru=18.5, I=19.5, T=30.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:51,287] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:56,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:27:57,892] loop.py:45 INFO TRAIN_STEP step=4591 loss=1.558327 Training: 25%|██▌ | 4591/18018 [9:57:43<28:58:05, 7.77s/step, loss=1.558, acc=0.655, acc_len=1.55, thru=15.3, I=21.3, T=30.7, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:27:57,887] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4590: forward=2921.0ms backward=1194.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:27:57,988] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:28:02,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:28:04,008] loop.py:45 INFO TRAIN_STEP step=4592 loss=1.611484 Training: 25%|██▌ | 4592/18018 [9:57:50<27:07:06, 7.27s/step, loss=1.611, acc=0.650, acc_len=1.53, thru=19.3, I=25.4, T=33.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:28:04,431] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:28:11,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:28:11,806] loop.py:45 INFO TRAIN_STEP step=4593 loss=1.572894 Training: 25%|██▌ | 4593/18018 [9:57:57<27:42:28, 7.43s/step, loss=1.573, acc=0.660, acc_len=1.55, thru=14.6, I=19.9, T=34.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:28:11,862] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-18 23:28:20,265] loop.py:45 INFO TRAIN_STEP step=4594 loss=1.701751 Training: 25%|██▌ | 4594/18018 [9:58:06<28:51:13, 7.74s/step, loss=1.702, acc=0.636, acc_len=1.48, thru=12.2, I=15.3, T=31.0, wait=4.3s, pool=32, epoch=1/2][2026-10-18 23:28:30,872] loop.py:380 INFO TIMING step=4595: step=3.897s data=1.591s compute=3.866s [fwd=2.725s bwd=1.126s opt=0.003s] dispatch=6.703s [2026-10-18 23:28:30,872] loop.py:45 INFO TRAIN_STEP step=4595 loss=1.652233 Training: 26%|██▌ | 4595/18018 [9:58:16<32:03:41, 8.60s/step, loss=1.652, acc=0.647, acc_len=1.51, thru=15.0, I=14.2, T=32.8, wait=6.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:28:40,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:28:40,866] loop.py:45 INFO TRAIN_STEP step=4596 loss=1.617835 (TrainerActor pid=9310) [2026-10-18 23:28:40,861] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4595: forward=3025.6ms backward=1198.1ms Training: 26%|██▌ | 4596/18018 [9:58:26<33:37:10, 9.02s/step, loss=1.618, acc=0.653, acc_len=1.53, thru=14.0, I=13.5, T=30.0, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:28:41,008] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:28:46,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:28:47,728] loop.py:45 INFO TRAIN_STEP step=4597 loss=1.522774 Training: 26%|██▌ | 4597/18018 [9:58:33<31:12:26, 8.37s/step, loss=1.523, acc=0.661, acc_len=1.56, thru=18.4, I=21.3, T=30.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:28:48,082] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:28:54,558] loop.py:45 INFO TRAIN_STEP step=4598 loss=1.599497 Training: 26%|██▌ | 4598/18018 [9:58:40<29:28:49, 7.91s/step, loss=1.599, acc=0.650, acc_len=1.53, thru=16.9, I=22.1, T=32.7, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:00,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:29:02,087] loop.py:45 INFO TRAIN_STEP step=4599 loss=1.728325 Training: 26%|██▌ | 4599/18018 [9:58:48<29:03:17, 7.79s/step, loss=1.728, acc=0.632, acc_len=1.46, thru=16.8, I=19.2, T=32.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:02,474] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:08,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:29:09,622] loop.py:380 INFO TIMING step=4600: step=3.544s data=1.444s compute=3.510s [fwd=2.428s bwd=1.067s opt=0.003s] dispatch=3.982s [2026-10-18 23:29:09,622] loop.py:45 INFO TRAIN_STEP step=4600 loss=1.561164 Training: 26%|██▌ | 4600/18018 [9:58:55<28:45:48, 7.72s/step, loss=1.561, acc=0.661, acc_len=1.56, thru=15.9, I=21.5, T=36.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:10,088] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:17,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:29:18,080] loop.py:45 INFO TRAIN_STEP step=4601 loss=1.502489 Training: 26%|██▌ | 4601/18018 [9:59:04<29:35:22, 7.94s/step, loss=1.502, acc=0.670, acc_len=1.59, thru=14.6, I=17.5, T=30.8, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:29:18,075] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4600: forward=2907.3ms backward=1184.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:18,270] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:23,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:29:25,319] loop.py:45 INFO TRAIN_STEP step=4602 loss=1.549176 Training: 26%|██▌ | 4602/18018 [9:59:11<28:48:21, 7.73s/step, loss=1.549, acc=0.660, acc_len=1.55, thru=17.1, I=20.0, T=25.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:25,638] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:32,733] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:29:35,330] loop.py:45 INFO TRAIN_STEP step=4603 loss=1.694883 Training: 26%|██▌ | 4603/18018 [9:59:21<31:21:06, 8.41s/step, loss=1.695, acc=0.637, acc_len=1.48, thru=13.2, I=17.8, T=20.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:35,751] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:42,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:29:43,688] loop.py:45 INFO TRAIN_STEP step=4604 loss=1.561736 Training: 26%|██▌ | 4604/18018 [9:59:29<31:17:19, 8.40s/step, loss=1.562, acc=0.660, acc_len=1.55, thru=16.1, I=20.3, T=28.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:44,134] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:50,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:29:50,739] loop.py:380 INFO TIMING step=4605: step=4.047s data=1.812s compute=4.017s [fwd=2.795s bwd=1.207s opt=0.003s] dispatch=2.993s [2026-10-18 23:29:50,739] loop.py:45 INFO TRAIN_STEP step=4605 loss=1.616589 Training: 26%|██▌ | 4605/18018 [9:59:36<29:46:50, 7.99s/step, loss=1.617, acc=0.651, acc_len=1.52, thru=15.6, I=19.7, T=31.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:29:50,859] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-18 23:29:58,404] loop.py:45 INFO TRAIN_STEP step=4606 loss=1.557860 Training: 26%|██▌ | 4606/18018 [9:59:44<29:24:45, 7.89s/step, loss=1.558, acc=0.657, acc_len=1.55, thru=15.2, I=16.2, T=35.7, wait=4.1s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:29:58,399] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4605: forward=2508.4ms backward=1034.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:09,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:30:09,382] loop.py:45 INFO TRAIN_STEP step=4607 loss=1.671289 Training: 26%|██▌ | 4607/18018 [9:59:55<32:51:23, 8.82s/step, loss=1.671, acc=0.643, acc_len=1.50, thru=13.0, I=13.3, T=27.0, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:09,668] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 23:30:18,543] loop.py:45 INFO TRAIN_STEP step=4608 loss=1.499161 Training: 26%|██▌ | 4608/18018 [10:00:04<33:14:05, 8.92s/step, loss=1.499, acc=0.669, acc_len=1.60, thru=13.6, I=13.6, T=26.2, wait=4.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:25,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:30:27,984] loop.py:45 INFO TRAIN_STEP step=4609 loss=1.669789 Training: 26%|██▌ | 4609/18018 [10:00:14<33:48:46, 9.08s/step, loss=1.670, acc=0.641, acc_len=1.50, thru=14.2, I=16.0, T=26.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:28,318] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:32,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:30:35,195] loop.py:380 INFO TIMING step=4610: step=4.298s data=2.103s compute=4.260s [fwd=2.978s bwd=1.267s opt=0.003s] dispatch=2.903s [2026-10-18 23:30:35,195] loop.py:45 INFO TRAIN_STEP step=4610 loss=1.523593 Training: 26%|██▌ | 4610/18018 [10:00:21<31:43:28, 8.52s/step, loss=1.524, acc=0.663, acc_len=1.57, thru=15.4, I=26.1, T=29.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:35,426] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:43,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:30:43,696] loop.py:45 INFO TRAIN_STEP step=4611 loss=1.586052 Training: 26%|██▌ | 4611/18018 [10:00:29<31:42:09, 8.51s/step, loss=1.586, acc=0.653, acc_len=1.54, thru=15.5, I=15.7, T=31.1, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:30:43,691] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4610: forward=2917.1ms backward=1150.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:44,111] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:49,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:30:50,810] loop.py:45 INFO TRAIN_STEP step=4612 loss=1.749876 Training: 26%|██▌ | 4612/18018 [10:00:36<30:08:14, 8.09s/step, loss=1.750, acc=0.632, acc_len=1.47, thru=17.3, I=19.9, T=31.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:50,901] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:56,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:30:58,314] loop.py:45 INFO TRAIN_STEP step=4613 loss=1.700116 Training: 26%|██▌ | 4613/18018 [10:00:44<29:28:34, 7.92s/step, loss=1.700, acc=0.638, acc_len=1.49, thru=16.5, I=21.7, T=29.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:30:58,366] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:31:05,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:31:05,979] loop.py:45 INFO TRAIN_STEP step=4614 loss=1.500583 Training: 26%|██▌ | 4614/18018 [10:00:52<29:11:43, 7.84s/step, loss=1.501, acc=0.671, acc_len=1.60, thru=14.0, I=16.8, T=29.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:31:06,177] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 23:31:14,232] loop.py:380 INFO TIMING step=4615: step=4.071s data=1.745s compute=4.039s [fwd=2.820s bwd=1.204s opt=0.003s] dispatch=4.171s [2026-10-18 23:31:14,232] loop.py:45 INFO TRAIN_STEP step=4615 loss=1.522607 Training: 26%|██▌ | 4615/18018 [10:01:00<29:39:00, 7.96s/step, loss=1.523, acc=0.664, acc_len=1.57, thru=12.9, I=15.1, T=31.4, wait=4.2s, pool=36, epoch=1/2][2026-10-18 23:31:24,282] loop.py:45 INFO TRAIN_STEP step=4616 loss=1.639974 Training: 26%|██▌ | 4616/18018 [10:01:10<31:58:42, 8.59s/step, loss=1.640, acc=0.643, acc_len=1.51, thru=15.1, I=13.0, T=29.6, wait=5.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:31:24,277] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4615: forward=2946.5ms backward=1320.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:31:30,387] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:31:32,417] loop.py:45 INFO TRAIN_STEP step=4617 loss=1.518204 Training: 26%|██▌ | 4617/18018 [10:01:18<31:28:00, 8.45s/step, loss=1.518, acc=0.668, acc_len=1.59, thru=16.5, I=19.0, T=25.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:31:32,909] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:31:38,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:31:39,275] loop.py:45 INFO TRAIN_STEP step=4618 loss=1.571671 Training: 26%|██▌ | 4618/18018 [10:01:25<29:41:11, 7.98s/step, loss=1.572, acc=0.656, acc_len=1.55, thru=15.2, I=23.7, T=32.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:31:39,634] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:31:46,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:31:47,234] loop.py:45 INFO TRAIN_STEP step=4619 loss=1.680487 Training: 26%|██▌ | 4619/18018 [10:01:33<29:39:57, 7.97s/step, loss=1.680, acc=0.643, acc_len=1.50, thru=15.2, I=19.3, T=30.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:31:47,512] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:31:54,915] loop.py:380 INFO TIMING step=4620: step=4.153s data=1.916s compute=4.124s [fwd=2.954s bwd=1.155s opt=0.003s] dispatch=3.518s [2026-10-18 23:31:54,916] loop.py:45 INFO TRAIN_STEP step=4620 loss=1.575041 Training: 26%|██▌ | 4620/18018 [10:01:41<29:20:25, 7.88s/step, loss=1.575, acc=0.656, acc_len=1.54, thru=13.8, I=17.7, T=30.8, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:01,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:32:03,396] loop.py:45 INFO TRAIN_STEP step=4621 loss=1.675409 Training: 26%|██▌ | 4621/18018 [10:01:49<30:00:09, 8.06s/step, loss=1.675, acc=0.640, acc_len=1.49, thru=16.5, I=17.9, T=27.2, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:32:03,391] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4620: forward=3314.9ms backward=1347.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:03,826] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:08,528] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:32:10,729] loop.py:45 INFO TRAIN_STEP step=4622 loss=1.606926 Training: 26%|██▌ | 4622/18018 [10:01:56<29:11:12, 7.84s/step, loss=1.607, acc=0.653, acc_len=1.53, thru=14.5, I=24.3, T=29.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:11,045] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:16,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:32:17,868] loop.py:45 INFO TRAIN_STEP step=4623 loss=1.469945 Training: 26%|██▌ | 4623/18018 [10:02:03<28:23:52, 7.63s/step, loss=1.470, acc=0.672, acc_len=1.60, thru=14.9, I=23.0, T=31.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:18,000] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:23,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:32:24,939] loop.py:45 INFO TRAIN_STEP step=4624 loss=1.546904 Training: 26%|██▌ | 4624/18018 [10:02:11<27:46:19, 7.46s/step, loss=1.547, acc=0.662, acc_len=1.57, thru=17.5, I=20.5, T=32.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:24,996] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-18 23:32:31,036] loop.py:380 INFO TIMING step=4625: step=3.836s data=1.673s compute=3.804s [fwd=2.711s bwd=1.078s opt=0.003s] dispatch=2.250s [2026-10-18 23:32:31,037] loop.py:45 INFO TRAIN_STEP step=4625 loss=1.619661 Training: 26%|██▌ | 4625/18018 [10:02:17<26:14:31, 7.05s/step, loss=1.620, acc=0.649, acc_len=1.52, thru=18.3, I=22.3, T=33.4, wait=2.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:38,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:32:38,532] loop.py:45 INFO TRAIN_STEP step=4626 loss=1.661606 Training: 26%|██▌ | 4626/18018 [10:02:24<26:44:00, 7.19s/step, loss=1.662, acc=0.645, acc_len=1.51, thru=17.3, I=17.3, T=33.3, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:32:38,527] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4625: forward=2769.3ms backward=1029.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:38,672] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-18 23:32:45,374] loop.py:45 INFO TRAIN_STEP step=4627 loss=1.630249 Training: 26%|██▌ | 4627/18018 [10:02:31<26:20:48, 7.08s/step, loss=1.630, acc=0.653, acc_len=1.53, thru=17.5, I=17.2, T=34.5, wait=3.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:52,387] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:32:53,423] loop.py:45 INFO TRAIN_STEP step=4628 loss=1.701046 Training: 26%|██▌ | 4628/18018 [10:02:39<27:25:22, 7.37s/step, loss=1.701, acc=0.636, acc_len=1.49, thru=18.6, I=18.5, T=32.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:32:53,895] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:33:01,004] loop.py:45 INFO TRAIN_STEP step=4629 loss=1.512827 Training: 26%|██▌ | 4629/18018 [10:02:47<27:39:04, 7.43s/step, loss=1.513, acc=0.667, acc_len=1.58, thru=15.7, I=18.1, T=31.7, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:09,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:33:09,658] loop.py:380 INFO TIMING step=4630: step=3.762s data=1.585s compute=3.720s [fwd=2.672s bwd=1.032s opt=0.003s] dispatch=4.885s [2026-10-18 23:33:09,658] loop.py:45 INFO TRAIN_STEP step=4630 loss=1.568563 Training: 26%|██▌ | 4630/18018 [10:02:55<29:00:40, 7.80s/step, loss=1.569, acc=0.656, acc_len=1.54, thru=15.1, I=14.8, T=34.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:09,747] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:16,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:33:17,414] loop.py:45 INFO TRAIN_STEP step=4631 loss=1.552291 Training: 26%|██▌ | 4631/18018 [10:03:03<28:57:32, 7.79s/step, loss=1.552, acc=0.662, acc_len=1.56, thru=18.0, I=18.3, T=26.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:33:17,409] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4630: forward=3413.6ms backward=1454.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:17,702] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:23,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:33:24,570] loop.py:45 INFO TRAIN_STEP step=4632 loss=1.650164 Training: 26%|██▌ | 4632/18018 [10:03:10<28:15:13, 7.60s/step, loss=1.650, acc=0.645, acc_len=1.51, thru=17.4, I=20.7, T=31.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:24,978] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:31,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:33:31,633] loop.py:45 INFO TRAIN_STEP step=4633 loss=1.556755 Training: 26%|██▌ | 4633/18018 [10:03:17<27:39:14, 7.44s/step, loss=1.557, acc=0.659, acc_len=1.55, thru=16.4, I=18.9, T=30.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:31,913] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:39,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:33:40,820] loop.py:45 INFO TRAIN_STEP step=4634 loss=1.618423 Training: 26%|██▌ | 4634/18018 [10:03:26<29:36:10, 7.96s/step, loss=1.618, acc=0.649, acc_len=1.53, thru=14.1, I=16.2, T=29.9, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:41,127] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:33:49,460] loop.py:380 INFO TIMING step=4635: step=5.215s data=2.620s compute=5.178s [fwd=3.539s bwd=1.626s opt=0.003s] dispatch=3.414s [2026-10-18 23:33:49,460] loop.py:45 INFO TRAIN_STEP step=4635 loss=1.590167 Training: 26%|██▌ | 4635/18018 [10:03:35<30:21:17, 8.17s/step, loss=1.590, acc=0.654, acc_len=1.53, thru=13.3, I=15.6, T=24.5, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:54,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:33:56,033] loop.py:45 INFO TRAIN_STEP step=4636 loss=1.641984 Training: 26%|██▌ | 4636/18018 [10:03:42<28:34:45, 7.69s/step, loss=1.642, acc=0.646, acc_len=1.52, thru=19.0, I=17.6, T=33.8, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:33:56,029] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4635: forward=2631.6ms backward=1104.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:33:56,169] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-18 23:34:04,917] loop.py:45 INFO TRAIN_STEP step=4637 loss=1.684646 Training: 26%|██▌ | 4637/18018 [10:03:51<29:54:35, 8.05s/step, loss=1.685, acc=0.639, acc_len=1.49, thru=12.8, I=13.9, T=31.0, wait=4.7s, pool=48, epoch=1/2][2026-10-18 23:34:12,817] loop.py:45 INFO TRAIN_STEP step=4638 loss=1.638229 Training: 26%|██▌ | 4638/18018 [10:03:58<29:44:39, 8.00s/step, loss=1.638, acc=0.645, acc_len=1.51, thru=17.2, I=16.1, T=31.6, wait=3.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:34:20,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:34:21,874] loop.py:45 INFO TRAIN_STEP step=4639 loss=1.717824 Training: 26%|██▌ | 4639/18018 [10:04:07<30:54:51, 8.32s/step, loss=1.718, acc=0.634, acc_len=1.47, thru=14.6, I=16.6, T=27.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:34:22,029] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:34:28,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:34:29,605] loop.py:380 INFO TIMING step=4640: step=4.024s data=1.715s compute=3.994s [fwd=2.891s bwd=1.089s opt=0.003s] dispatch=3.698s [2026-10-18 23:34:29,606] loop.py:45 INFO TRAIN_STEP step=4640 loss=1.571497 Training: 26%|██▌ | 4640/18018 [10:04:15<30:15:28, 8.14s/step, loss=1.571, acc=0.658, acc_len=1.55, thru=15.3, I=20.8, T=31.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:34:30,084] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:34:37,052] loop.py:45 INFO TRAIN_STEP step=4641 loss=1.579452 Training: 26%|██▌ | 4641/18018 [10:04:23<29:29:03, 7.93s/step, loss=1.579, acc=0.655, acc_len=1.55, thru=16.1, I=22.7, T=28.2, wait=2.9s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:34:37,047] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4640: forward=3180.9ms backward=1310.5ms [2026-10-18 23:34:45,524] loop.py:45 INFO TRAIN_STEP step=4642 loss=1.615181 Training: 26%|██▌ | 4642/18018 [10:04:31<30:04:43, 8.10s/step, loss=1.615, acc=0.652, acc_len=1.53, thru=11.6, I=13.3, T=29.8, wait=4.2s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:34:52,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:34:53,883] loop.py:45 INFO TRAIN_STEP step=4643 loss=1.554238 Training: 26%|██▌ | 4643/18018 [10:04:39<30:22:08, 8.17s/step, loss=1.554, acc=0.655, acc_len=1.55, thru=17.8, I=16.6, T=25.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:34:54,178] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:34:59,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:35:00,902] loop.py:45 INFO TRAIN_STEP step=4644 loss=1.616371 Training: 26%|██▌ | 4644/18018 [10:04:46<29:04:45, 7.83s/step, loss=1.616, acc=0.649, acc_len=1.53, thru=16.6, I=22.7, T=32.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:01,254] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:09,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:35:09,743] loop.py:380 INFO TIMING step=4645: step=3.912s data=1.756s compute=3.875s [fwd=2.685s bwd=1.175s opt=0.003s] dispatch=4.919s [2026-10-18 23:35:09,743] loop.py:45 INFO TRAIN_STEP step=4645 loss=1.516772 Training: 26%|██▌ | 4645/18018 [10:04:55<30:12:26, 8.13s/step, loss=1.517, acc=0.664, acc_len=1.58, thru=15.1, I=16.2, T=32.7, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:09,998] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 23:35:19,684] loop.py:45 INFO TRAIN_STEP step=4646 loss=1.613761 Training: 26%|██▌ | 4646/18018 [10:05:05<32:13:05, 8.67s/step, loss=1.614, acc=0.651, acc_len=1.52, thru=12.4, I=12.8, T=26.2, wait=5.0s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:35:19,680] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4645: forward=3563.5ms backward=1255.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:27,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:35:27,865] loop.py:45 INFO TRAIN_STEP step=4647 loss=1.545501 Training: 26%|██▌ | 4647/18018 [10:05:13<31:40:02, 8.53s/step, loss=1.546, acc=0.663, acc_len=1.56, thru=16.4, I=15.3, T=34.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:28,248] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 23:35:33,856] loop.py:45 INFO TRAIN_STEP step=4648 loss=1.590106 Training: 26%|██▌ | 4648/18018 [10:05:19<28:50:34, 7.77s/step, loss=1.590, acc=0.658, acc_len=1.55, thru=20.1, I=21.3, T=37.3, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:42,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:35:44,235] loop.py:45 INFO TRAIN_STEP step=4649 loss=1.599963 Training: 26%|██▌ | 4649/18018 [10:05:30<31:44:57, 8.55s/step, loss=1.600, acc=0.658, acc_len=1.55, thru=12.9, I=14.8, T=28.0, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:44,668] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:49,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:35:51,131] loop.py:380 INFO TIMING step=4650: step=3.724s data=1.659s compute=3.695s [fwd=2.638s bwd=1.043s opt=0.003s] dispatch=3.164s [2026-10-18 23:35:51,131] loop.py:45 INFO TRAIN_STEP step=4650 loss=1.614472 Training: 26%|██▌ | 4650/18018 [10:05:37<29:54:13, 8.05s/step, loss=1.614, acc=0.647, acc_len=1.51, thru=16.1, I=24.0, T=34.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:51,439] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:58,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:35:59,278] loop.py:45 INFO TRAIN_STEP step=4651 loss=1.510072 Training: 26%|██▌ | 4651/18018 [10:05:45<30:00:37, 8.08s/step, loss=1.510, acc=0.663, acc_len=1.57, thru=14.8, I=20.4, T=28.3, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:35:59,274] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4650: forward=3121.1ms backward=1356.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:35:59,551] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:04,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:36:05,807] loop.py:45 INFO TRAIN_STEP step=4652 loss=1.552178 Training: 26%|██▌ | 4652/18018 [10:05:51<28:16:39, 7.62s/step, loss=1.552, acc=0.654, acc_len=1.54, thru=19.1, I=21.5, T=32.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:05,841] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 23:36:14,976] loop.py:45 INFO TRAIN_STEP step=4653 loss=1.437058 Training: 26%|██▌ | 4653/18018 [10:06:01<30:00:12, 8.08s/step, loss=1.437, acc=0.677, acc_len=1.61, thru=12.3, I=11.4, T=32.2, wait=5.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:23,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:36:23,868] loop.py:45 INFO TRAIN_STEP step=4654 loss=1.476699 Training: 26%|██▌ | 4654/18018 [10:06:09<30:54:09, 8.32s/step, loss=1.477, acc=0.674, acc_len=1.61, thru=16.9, I=16.9, T=28.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:24,266] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:30,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:36:31,715] loop.py:380 INFO TIMING step=4655: step=4.306s data=1.917s compute=4.270s [fwd=3.000s bwd=1.256s opt=0.003s] dispatch=3.533s [2026-10-18 23:36:31,715] loop.py:45 INFO TRAIN_STEP step=4655 loss=1.630660 Training: 26%|██▌ | 4655/18018 [10:06:17<30:22:13, 8.18s/step, loss=1.631, acc=0.650, acc_len=1.53, thru=14.6, I=18.2, T=29.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:32,130] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:36:40,919] loop.py:45 INFO TRAIN_STEP step=4656 loss=1.586819 Training: 26%|██▌ | 4656/18018 [10:06:27<31:30:20, 8.49s/step, loss=1.587, acc=0.653, acc_len=1.54, thru=14.6, I=15.0, T=29.5, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:40,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 23:36:40,912] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4655: forward=2979.8ms backward=1306.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:41,377] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:46,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:36:48,068] loop.py:45 INFO TRAIN_STEP step=4657 loss=1.649470 Training: 26%|██▌ | 4657/18018 [10:06:34<30:00:41, 8.09s/step, loss=1.649, acc=0.644, acc_len=1.51, thru=18.1, I=19.3, T=31.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:48,463] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:54,929] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:36:55,858] loop.py:45 INFO TRAIN_STEP step=4658 loss=1.529937 Training: 26%|██▌ | 4658/18018 [10:06:41<29:40:43, 8.00s/step, loss=1.530, acc=0.667, acc_len=1.57, thru=14.9, I=18.0, T=29.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:36:55,947] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:02,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:37:04,597] loop.py:45 INFO TRAIN_STEP step=4659 loss=1.569504 Training: 26%|██▌ | 4659/18018 [10:06:50<30:30:15, 8.22s/step, loss=1.570, acc=0.657, acc_len=1.55, thru=13.6, I=17.9, T=25.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:04,630] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:10,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:37:11,077] loop.py:380 INFO TIMING step=4660: step=4.019s data=1.787s compute=3.995s [fwd=2.838s bwd=1.143s opt=0.003s] dispatch=2.450s [2026-10-18 23:37:11,078] loop.py:45 INFO TRAIN_STEP step=4660 loss=1.544443 Training: 26%|██▌ | 4660/18018 [10:06:57<28:33:54, 7.70s/step, loss=1.544, acc=0.664, acc_len=1.57, thru=16.4, I=22.2, T=31.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:11,276] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:17,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:37:18,589] loop.py:45 INFO TRAIN_STEP step=4661 loss=1.654155 Training: 26%|██▌ | 4661/18018 [10:07:04<28:21:11, 7.64s/step, loss=1.654, acc=0.646, acc_len=1.52, thru=15.1, I=17.1, T=31.9, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:37:18,584] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4660: forward=2777.4ms backward=1174.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:19,054] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:37:25,993] loop.py:45 INFO TRAIN_STEP step=4662 loss=1.768370 Training: 26%|██▌ | 4662/18018 [10:07:12<28:05:10, 7.57s/step, loss=1.768, acc=0.628, acc_len=1.45, thru=16.4, I=18.7, T=33.1, wait=3.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:30,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:37:33,207] loop.py:45 INFO TRAIN_STEP step=4663 loss=1.657440 Training: 26%|██▌ | 4663/18018 [10:07:19<27:41:22, 7.46s/step, loss=1.657, acc=0.644, acc_len=1.51, thru=19.0, I=21.4, T=29.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:33,465] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:41,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:37:41,391] loop.py:45 INFO TRAIN_STEP step=4664 loss=1.407238 Training: 26%|██▌ | 4664/18018 [10:07:27<28:29:20, 7.68s/step, loss=1.407, acc=0.685, acc_len=1.64, thru=16.2, I=19.4, T=33.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:41,518] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:48,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:37:50,726] loop.py:380 INFO TIMING step=4665: step=5.488s data=3.068s compute=5.445s [fwd=3.954s bwd=1.476s opt=0.003s] dispatch=3.836s [2026-10-18 23:37:50,726] loop.py:45 INFO TRAIN_STEP step=4665 loss=1.647375 Training: 26%|██▌ | 4665/18018 [10:07:36<30:19:34, 8.18s/step, loss=1.647, acc=0.642, acc_len=1.50, thru=13.6, I=15.2, T=23.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:51,020] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:56,484] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:37:56,829] loop.py:45 INFO TRAIN_STEP step=4666 loss=1.510563 Training: 26%|██▌ | 4666/18018 [10:07:42<28:01:02, 7.55s/step, loss=1.511, acc=0.666, acc_len=1.58, thru=16.0, I=20.7, T=34.3, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:37:56,823] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4665: forward=2595.8ms backward=1092.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:37:56,986] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:02,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:38:03,658] loop.py:45 INFO TRAIN_STEP step=4667 loss=1.738728 Training: 26%|██▌ | 4667/18018 [10:07:49<27:12:39, 7.34s/step, loss=1.739, acc=0.628, acc_len=1.45, thru=18.4, I=20.8, T=30.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:03,700] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:08,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:38:09,671] loop.py:45 INFO TRAIN_STEP step=4668 loss=1.576213 Training: 26%|██▌ | 4668/18018 [10:07:55<25:44:01, 6.94s/step, loss=1.576, acc=0.656, acc_len=1.55, thru=18.6, I=24.4, T=33.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:09,794] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:17,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:38:17,845] loop.py:45 INFO TRAIN_STEP step=4669 loss=1.597245 Training: 26%|██▌ | 4669/18018 [10:08:03<27:06:16, 7.31s/step, loss=1.597, acc=0.651, acc_len=1.52, thru=13.8, I=15.2, T=32.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:18,044] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 23:38:24,498] loop.py:380 INFO TIMING step=4670: step=3.780s data=1.657s compute=3.748s [fwd=2.705s bwd=1.028s opt=0.003s] dispatch=2.864s [2026-10-18 23:38:24,498] loop.py:45 INFO TRAIN_STEP step=4670 loss=1.573359 Training: 26%|██▌ | 4670/18018 [10:08:10<26:22:28, 7.11s/step, loss=1.573, acc=0.655, acc_len=1.54, thru=14.2, I=19.6, T=33.9, wait=2.9s, pool=24, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:35,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:38:35,811] loop.py:45 INFO TRAIN_STEP step=4671 loss=1.631709 Training: 26%|██▌ | 4671/18018 [10:08:21<31:02:34, 8.37s/step, loss=1.632, acc=0.655, acc_len=1.54, thru=15.3, I=15.6, T=27.3, wait=6.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:38:35,806] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4670: forward=3258.3ms backward=1377.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:35,971] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:44,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:38:45,093] loop.py:45 INFO TRAIN_STEP step=4672 loss=1.567780 Training: 26%|██▌ | 4672/18018 [10:08:31<32:03:02, 8.65s/step, loss=1.568, acc=0.659, acc_len=1.55, thru=13.7, I=14.4, T=27.5, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:45,407] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 23:38:52,419] loop.py:45 INFO TRAIN_STEP step=4673 loss=1.664498 Training: 26%|██▌ | 4673/18018 [10:08:38<30:34:48, 8.25s/step, loss=1.664, acc=0.640, acc_len=1.50, thru=12.2, I=13.7, T=29.9, wait=3.0s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:38:58,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:39:00,525] loop.py:45 INFO TRAIN_STEP step=4674 loss=1.656704 Training: 26%|██▌ | 4674/18018 [10:08:46<30:25:14, 8.21s/step, loss=1.657, acc=0.647, acc_len=1.52, thru=18.8, I=19.3, T=30.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:00,858] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:05,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:39:07,640] loop.py:380 INFO TIMING step=4675: step=4.422s data=2.058s compute=4.389s [fwd=3.062s bwd=1.311s opt=0.003s] dispatch=2.682s [2026-10-18 23:39:07,640] loop.py:45 INFO TRAIN_STEP step=4675 loss=1.642956 Training: 26%|██▌ | 4675/18018 [10:08:53<29:12:16, 7.88s/step, loss=1.643, acc=0.646, acc_len=1.51, thru=16.6, I=25.4, T=28.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:07,674] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:11,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:39:14,020] loop.py:45 INFO TRAIN_STEP step=4676 loss=1.570751 Training: 26%|██▌ | 4676/18018 [10:09:00<27:32:01, 7.43s/step, loss=1.571, acc=0.659, acc_len=1.56, thru=16.0, I=25.5, T=28.8, wait=1.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:39:14,016] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4675: forward=3077.5ms backward=1325.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:14,501] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:22,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:39:22,879] loop.py:45 INFO TRAIN_STEP step=4677 loss=1.530662 Training: 26%|██▌ | 4677/18018 [10:09:08<29:07:12, 7.86s/step, loss=1.531, acc=0.667, acc_len=1.58, thru=15.3, I=15.1, T=28.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:23,341] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:30,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:39:31,513] loop.py:45 INFO TRAIN_STEP step=4678 loss=1.382202 Training: 26%|██▌ | 4678/18018 [10:09:17<29:58:59, 8.09s/step, loss=1.382, acc=0.687, acc_len=1.65, thru=14.7, I=17.4, T=26.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:31,571] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:39,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:39:40,656] loop.py:45 INFO TRAIN_STEP step=4679 loss=1.642137 Training: 26%|██▌ | 4679/18018 [10:09:26<31:08:57, 8.41s/step, loss=1.642, acc=0.647, acc_len=1.51, thru=14.1, I=14.0, T=27.0, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:40,766] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:45,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:39:47,241] loop.py:380 INFO TIMING step=4680: step=4.127s data=1.941s compute=4.091s [fwd=2.881s bwd=1.195s opt=0.003s] dispatch=2.446s [2026-10-18 23:39:47,241] loop.py:45 INFO TRAIN_STEP step=4680 loss=1.625708 Training: 26%|██▌ | 4680/18018 [10:09:33<29:07:12, 7.86s/step, loss=1.626, acc=0.647, acc_len=1.52, thru=17.3, I=20.9, T=31.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:39:47,443] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:39:54,963] loop.py:45 INFO TRAIN_STEP step=4681 loss=1.598809 Training: 26%|██▌ | 4681/18018 [10:09:41<28:57:55, 7.82s/step, loss=1.599, acc=0.651, acc_len=1.53, thru=14.0, I=16.4, T=32.1, wait=3.7s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:39:54,959] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4680: forward=2803.5ms backward=1124.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:00,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:40:01,707] loop.py:45 INFO TRAIN_STEP step=4682 loss=1.639163 Training: 26%|██▌ | 4682/18018 [10:09:47<27:46:09, 7.50s/step, loss=1.639, acc=0.647, acc_len=1.51, thru=19.0, I=19.4, T=31.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:02,101] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:40:09,053] loop.py:45 INFO TRAIN_STEP step=4683 loss=1.451796 Training: 26%|██▌ | 4683/18018 [10:09:55<27:35:58, 7.45s/step, loss=1.452, acc=0.679, acc_len=1.62, thru=13.7, I=18.4, T=28.9, wait=2.9s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:15,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:40:17,717] loop.py:45 INFO TRAIN_STEP step=4684 loss=1.393279 Training: 26%|██▌ | 4684/18018 [10:10:03<28:56:47, 7.82s/step, loss=1.393, acc=0.682, acc_len=1.63, thru=16.9, I=18.0, T=26.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:17,896] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:24,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:40:24,881] loop.py:380 INFO TIMING step=4685: step=3.918s data=1.936s compute=3.888s [fwd=2.851s bwd=1.023s opt=0.003s] dispatch=3.236s [2026-10-18 23:40:24,881] loop.py:45 INFO TRAIN_STEP step=4685 loss=1.514548 Training: 26%|██▌ | 4685/18018 [10:10:10<28:13:10, 7.62s/step, loss=1.515, acc=0.670, acc_len=1.60, thru=14.0, I=20.5, T=32.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:25,280] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:33,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:40:34,838] loop.py:45 INFO TRAIN_STEP step=4686 loss=1.488208 Training: 26%|██▌ | 4686/18018 [10:10:20<30:48:51, 8.32s/step, loss=1.488, acc=0.673, acc_len=1.60, thru=13.4, I=15.9, T=28.8, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:40:34,833] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4685: forward=3023.7ms backward=1355.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:35,215] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:40,756] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:40:41,880] loop.py:45 INFO TRAIN_STEP step=4687 loss=1.545053 Training: 26%|██▌ | 4687/18018 [10:10:27<29:23:28, 7.94s/step, loss=1.545, acc=0.660, acc_len=1.56, thru=16.9, I=23.7, T=29.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:42,265] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:40:50,106] loop.py:45 INFO TRAIN_STEP step=4688 loss=1.563417 Training: 26%|██▌ | 4688/18018 [10:10:36<29:42:39, 8.02s/step, loss=1.563, acc=0.663, acc_len=1.57, thru=12.4, I=12.4, T=30.8, wait=4.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:56,027] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:40:57,041] loop.py:45 INFO TRAIN_STEP step=4689 loss=1.483049 Training: 26%|██▌ | 4689/18018 [10:10:43<28:29:59, 7.70s/step, loss=1.483, acc=0.672, acc_len=1.60, thru=17.8, I=19.1, T=32.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:40:57,537] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:41:03,586] loop.py:380 INFO TIMING step=4690: step=4.141s data=1.972s compute=4.102s [fwd=2.844s bwd=1.243s opt=0.003s] dispatch=2.394s [2026-10-18 23:41:03,586] loop.py:45 INFO TRAIN_STEP step=4690 loss=1.568568 Training: 26%|██▌ | 4690/18018 [10:10:49<27:13:03, 7.35s/step, loss=1.569, acc=0.662, acc_len=1.56, thru=16.9, I=19.1, T=30.9, wait=2.4s, pool=52, epoch=1/2][2026-10-18 23:41:11,426] loop.py:45 INFO TRAIN_STEP step=4691 loss=1.538509 (TrainerActor pid=9310) [2026-10-18 23:41:11,420] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4690: forward=3094.9ms backward=1138.5ms Training: 26%|██▌ | 4691/18018 [10:10:57<27:45:37, 7.50s/step, loss=1.539, acc=0.667, acc_len=1.58, thru=17.2, I=14.7, T=29.9, wait=3.5s, pool=52, epoch=1/2][2026-10-18 23:41:20,090] loop.py:45 INFO TRAIN_STEP step=4692 loss=1.542359 Training: 26%|██▌ | 4692/18018 [10:11:06<29:02:49, 7.85s/step, loss=1.542, acc=0.664, acc_len=1.58, thru=15.7, I=14.4, T=30.6, wait=4.5s, pool=56, epoch=1/2][2026-10-18 23:41:27,288] loop.py:45 INFO TRAIN_STEP step=4693 loss=1.563012 Training: 26%|██▌ | 4693/18018 [10:11:13<28:19:38, 7.65s/step, loss=1.563, acc=0.658, acc_len=1.55, thru=16.7, I=15.5, T=33.0, wait=3.3s, pool=52, epoch=1/2][2026-10-18 23:41:33,458] loop.py:45 INFO TRAIN_STEP step=4694 loss=1.620417 Training: 26%|██▌ | 4694/18018 [10:11:19<26:40:37, 7.21s/step, loss=1.620, acc=0.650, acc_len=1.52, thru=15.8, I=17.2, T=34.8, wait=2.5s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:41:41,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:41:43,144] loop.py:380 INFO TIMING step=4695: step=4.067s data=1.754s compute=4.029s [fwd=2.949s bwd=1.065s opt=0.003s] dispatch=5.610s [2026-10-18 23:41:43,144] loop.py:45 INFO TRAIN_STEP step=4695 loss=1.590208 Training: 26%|██▌ | 4695/18018 [10:11:29<29:25:31, 7.95s/step, loss=1.590, acc=0.658, acc_len=1.55, thru=16.5, I=17.9, T=31.5, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:41:43,555] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 23:41:50,460] loop.py:45 INFO TRAIN_STEP step=4696 loss=1.528849 Training: 26%|██▌ | 4696/18018 [10:11:36<28:43:05, 7.76s/step, loss=1.529, acc=0.664, acc_len=1.56, thru=11.7, I=20.5, T=29.5, wait=3.0s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:41:50,454] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4695: forward=3093.0ms backward=1207.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:00,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:42:00,596] loop.py:45 INFO TRAIN_STEP step=4697 loss=1.525197 Training: 26%|██▌ | 4697/18018 [10:11:46<31:21:19, 8.47s/step, loss=1.525, acc=0.667, acc_len=1.59, thru=16.1, I=13.7, T=32.2, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:00,783] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:06,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:42:07,781] loop.py:45 INFO TRAIN_STEP step=4698 loss=1.612948 Training: 26%|██▌ | 4698/18018 [10:11:53<29:55:16, 8.09s/step, loss=1.613, acc=0.649, acc_len=1.52, thru=18.6, I=20.1, T=33.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:08,020] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:15,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:42:15,189] loop.py:45 INFO TRAIN_STEP step=4699 loss=1.687456 Training: 26%|██▌ | 4699/18018 [10:12:01<29:09:54, 7.88s/step, loss=1.687, acc=0.637, acc_len=1.48, thru=17.1, I=19.1, T=30.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:15,563] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-18 23:42:21,785] loop.py:380 INFO TIMING step=4700: step=3.683s data=1.619s compute=3.651s [fwd=2.623s bwd=1.013s opt=0.003s] dispatch=2.904s [2026-10-18 23:42:21,785] loop.py:45 INFO TRAIN_STEP step=4700 loss=1.598754 Training: 26%|██▌ | 4700/18018 [10:12:07<27:44:02, 7.50s/step, loss=1.599, acc=0.653, acc_len=1.53, thru=18.4, I=19.7, T=34.8, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:27,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:42:29,626] loop.py:45 INFO TRAIN_STEP step=4701 loss=1.517568 Training: 26%|██▌ | 4701/18018 [10:12:15<28:06:52, 7.60s/step, loss=1.518, acc=0.665, acc_len=1.57, thru=16.5, I=20.0, T=31.2, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:42:29,622] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4700: forward=2846.1ms backward=1208.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:29,784] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:35,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:42:36,654] loop.py:45 INFO TRAIN_STEP step=4702 loss=1.796228 Training: 26%|██▌ | 4702/18018 [10:12:22<27:28:37, 7.43s/step, loss=1.796, acc=0.622, acc_len=1.43, thru=16.7, I=24.3, T=31.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:37,012] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:44,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:42:45,258] loop.py:45 INFO TRAIN_STEP step=4703 loss=1.631659 Training: 26%|██▌ | 4703/18018 [10:12:31<28:46:59, 7.78s/step, loss=1.632, acc=0.650, acc_len=1.52, thru=13.5, I=16.1, T=31.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:45,552] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:49,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:42:51,937] loop.py:45 INFO TRAIN_STEP step=4704 loss=1.627996 Training: 26%|██▌ | 4704/18018 [10:12:38<27:33:13, 7.45s/step, loss=1.628, acc=0.646, acc_len=1.51, thru=18.1, I=23.3, T=30.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:51,981] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:58,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:42:59,332] loop.py:380 INFO TIMING step=4705: step=4.566s data=1.732s compute=4.536s [fwd=3.300s bwd=1.222s opt=0.003s] dispatch=2.819s [2026-10-18 23:42:59,332] loop.py:45 INFO TRAIN_STEP step=4705 loss=1.519465 Training: 26%|██▌ | 4705/18018 [10:12:45<27:29:29, 7.43s/step, loss=1.519, acc=0.667, acc_len=1.58, thru=14.8, I=20.4, T=28.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:42:59,551] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 23:43:08,482] loop.py:45 INFO TRAIN_STEP step=4706 loss=1.573506 Training: 26%|██▌ | 4706/18018 [10:12:54<29:23:37, 7.95s/step, loss=1.574, acc=0.654, acc_len=1.55, thru=11.7, I=15.5, T=27.0, wait=4.4s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:43:08,476] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4705: forward=3305.2ms backward=1377.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:17,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:43:17,871] loop.py:45 INFO TRAIN_STEP step=4707 loss=1.580418 Training: 26%|██▌ | 4707/18018 [10:13:03<30:59:12, 8.38s/step, loss=1.580, acc=0.653, acc_len=1.53, thru=15.2, I=14.0, T=32.0, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:18,266] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:25,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:43:26,058] loop.py:45 INFO TRAIN_STEP step=4708 loss=1.529716 Training: 26%|██▌ | 4708/18018 [10:13:12<30:46:09, 8.32s/step, loss=1.530, acc=0.665, acc_len=1.58, thru=14.6, I=16.6, T=31.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:26,226] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:32,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:43:33,253] loop.py:45 INFO TRAIN_STEP step=4709 loss=1.635971 Training: 26%|██▌ | 4709/18018 [10:13:19<29:31:08, 7.98s/step, loss=1.636, acc=0.647, acc_len=1.52, thru=17.7, I=20.0, T=31.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:33,309] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-18 23:43:39,233] loop.py:380 INFO TIMING step=4710: step=4.040s data=1.987s compute=4.008s [fwd=2.964s bwd=1.030s opt=0.003s] dispatch=1.929s [2026-10-18 23:43:39,233] loop.py:45 INFO TRAIN_STEP step=4710 loss=1.614762 Training: 26%|██▌ | 4710/18018 [10:13:25<27:17:29, 7.38s/step, loss=1.615, acc=0.648, acc_len=1.53, thru=19.2, I=21.0, T=31.7, wait=1.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:45,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:43:48,100] loop.py:45 INFO TRAIN_STEP step=4711 loss=1.645218 Training: 26%|██▌ | 4711/18018 [10:13:34<28:56:18, 7.83s/step, loss=1.645, acc=0.644, acc_len=1.51, thru=15.4, I=16.9, T=26.7, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:43:48,095] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4710: forward=3331.5ms backward=1397.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:48,146] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:53,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:43:54,895] loop.py:45 INFO TRAIN_STEP step=4712 loss=1.677403 Training: 26%|██▌ | 4712/18018 [10:13:40<27:47:12, 7.52s/step, loss=1.677, acc=0.640, acc_len=1.49, thru=15.3, I=20.0, T=32.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:43:55,019] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:00,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:44:02,046] loop.py:45 INFO TRAIN_STEP step=4713 loss=1.565693 Training: 26%|██▌ | 4713/18018 [10:13:48<27:22:43, 7.41s/step, loss=1.566, acc=0.659, acc_len=1.55, thru=17.0, I=23.6, T=32.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:02,482] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:07,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:44:09,231] loop.py:45 INFO TRAIN_STEP step=4714 loss=1.599727 Training: 26%|██▌ | 4714/18018 [10:13:55<27:07:42, 7.34s/step, loss=1.600, acc=0.653, acc_len=1.54, thru=16.7, I=24.2, T=30.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:09,623] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:44:16,221] loop.py:380 INFO TIMING step=4715: step=4.096s data=1.883s compute=4.069s [fwd=2.896s bwd=1.157s opt=0.003s] dispatch=2.886s [2026-10-18 23:44:16,222] loop.py:45 INFO TRAIN_STEP step=4715 loss=1.521719 Training: 26%|██▌ | 4715/18018 [10:14:02<26:44:24, 7.24s/step, loss=1.522, acc=0.662, acc_len=1.57, thru=16.0, I=21.7, T=31.3, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:22,839] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:44:25,054] loop.py:45 INFO TRAIN_STEP step=4716 loss=1.582981 Training: 26%|██▌ | 4716/18018 [10:14:11<28:30:28, 7.72s/step, loss=1.583, acc=0.655, acc_len=1.54, thru=15.7, I=17.2, T=32.5, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:44:25,050] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4715: forward=2791.0ms backward=1092.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:25,358] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-18 23:44:31,887] loop.py:45 INFO TRAIN_STEP step=4717 loss=1.512939 Training: 26%|██▌ | 4717/18018 [10:14:17<27:31:33, 7.45s/step, loss=1.513, acc=0.665, acc_len=1.57, thru=15.0, I=27.3, T=32.0, wait=2.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:40,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:44:41,023] loop.py:45 INFO TRAIN_STEP step=4718 loss=1.529226 Training: 26%|██▌ | 4718/18018 [10:14:27<29:23:29, 7.96s/step, loss=1.529, acc=0.665, acc_len=1.58, thru=15.2, I=14.6, T=30.2, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:41,467] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:51,840] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:44:51,965] loop.py:45 INFO TRAIN_STEP step=4719 loss=1.624221 Training: 26%|██▌ | 4719/18018 [10:14:38<32:41:55, 8.85s/step, loss=1.624, acc=0.650, acc_len=1.53, thru=12.5, I=11.0, T=26.1, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:52,341] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:44:59,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:45:00,971] loop.py:380 INFO TIMING step=4720: step=4.346s data=1.909s compute=4.300s [fwd=2.910s bwd=1.376s opt=0.003s] dispatch=4.652s [2026-10-18 23:45:00,971] loop.py:45 INFO TRAIN_STEP step=4720 loss=1.560913 Training: 26%|██▌ | 4720/18018 [10:14:47<32:52:13, 8.90s/step, loss=1.561, acc=0.660, acc_len=1.55, thru=14.9, I=17.2, T=29.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:45:01,103] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:45:07,484] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:45:08,156] loop.py:45 INFO TRAIN_STEP step=4721 loss=1.597263 Training: 26%|██▌ | 4721/18018 [10:14:54<30:58:03, 8.38s/step, loss=1.597, acc=0.655, acc_len=1.54, thru=15.8, I=20.6, T=33.0, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:45:08,151] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4720: forward=2721.2ms backward=1105.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:45:08,488] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:45:12,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:45:14,865] loop.py:45 INFO TRAIN_STEP step=4722 loss=1.547443 Training: 26%|██▌ | 4722/18018 [10:15:00<29:06:38, 7.88s/step, loss=1.547, acc=0.661, acc_len=1.56, thru=18.5, I=23.4, T=29.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:45:14,977] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-18 23:45:21,645] loop.py:45 INFO TRAIN_STEP step=4723 loss=1.660613 Training: 26%|██▌ | 4723/18018 [10:15:07<27:53:11, 7.55s/step, loss=1.661, acc=0.644, acc_len=1.50, thru=15.7, I=20.4, T=30.9, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:45:29,395] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:45:30,165] loop.py:45 INFO TRAIN_STEP step=4724 loss=1.517683 Training: 26%|██▌ | 4724/18018 [10:15:16<28:57:32, 7.84s/step, loss=1.518, acc=0.665, acc_len=1.58, thru=16.1, I=15.0, T=30.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:45:30,398] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:45:35,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:45:36,959] loop.py:380 INFO TIMING step=4725: step=4.242s data=1.990s compute=4.216s [fwd=2.981s bwd=1.220s opt=0.003s] dispatch=2.541s [2026-10-18 23:45:36,959] loop.py:45 INFO TRAIN_STEP step=4725 loss=1.546873 Training: 26%|██▌ | 4725/18018 [10:15:23<27:47:43, 7.53s/step, loss=1.547, acc=0.662, acc_len=1.57, thru=18.1, I=21.7, T=30.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:45:37,324] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:45:44,923] loop.py:45 INFO TRAIN_STEP step=4726 loss=1.659086 Training: 26%|██▌ | 4726/18018 [10:15:31<28:16:38, 7.66s/step, loss=1.659, acc=0.642, acc_len=1.49, thru=14.1, I=19.7, T=30.1, wait=3.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:45:44,919] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4725: forward=2798.5ms backward=1395.4ms [2026-10-18 23:45:53,591] loop.py:45 INFO TRAIN_STEP step=4727 loss=1.579930 Training: 26%|██▌ | 4727/18018 [10:15:39<29:23:44, 7.96s/step, loss=1.580, acc=0.654, acc_len=1.53, thru=14.9, I=14.2, T=31.4, wait=4.6s, pool=60, epoch=1/2][2026-10-18 23:46:02,164] loop.py:45 INFO TRAIN_STEP step=4728 loss=1.594836 Training: 26%|██▌ | 4728/18018 [10:15:48<30:03:51, 8.14s/step, loss=1.595, acc=0.656, acc_len=1.54, thru=14.5, I=13.5, T=30.9, wait=4.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:11,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:46:13,517] loop.py:45 INFO TRAIN_STEP step=4729 loss=1.558504 Training: 26%|██▌ | 4729/18018 [10:15:59<33:36:59, 9.11s/step, loss=1.559, acc=0.657, acc_len=1.55, thru=13.5, I=14.8, T=23.1, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:13,941] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:21,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:46:22,640] loop.py:380 INFO TIMING step=4730: step=4.578s data=2.151s compute=4.528s [fwd=3.169s bwd=1.344s opt=0.003s] dispatch=4.536s [2026-10-18 23:46:22,640] loop.py:45 INFO TRAIN_STEP step=4730 loss=1.538466 Training: 26%|██▋ | 4730/18018 [10:16:08<33:38:05, 9.11s/step, loss=1.538, acc=0.665, acc_len=1.58, thru=14.7, I=15.8, T=28.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:22,696] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:29,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:46:30,857] loop.py:45 INFO TRAIN_STEP step=4731 loss=1.569106 Training: 26%|██▋ | 4731/18018 [10:16:16<32:38:24, 8.84s/step, loss=1.569, acc=0.662, acc_len=1.56, thru=14.9, I=18.2, T=30.0, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:46:30,852] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4730: forward=3013.3ms backward=1194.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:31,256] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:46:37,698] loop.py:45 INFO TRAIN_STEP step=4732 loss=1.516633 Training: 26%|██▋ | 4732/18018 [10:16:23<30:25:31, 8.24s/step, loss=1.517, acc=0.670, acc_len=1.59, thru=17.9, I=20.5, T=30.4, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:42,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:46:43,946] loop.py:45 INFO TRAIN_STEP step=4733 loss=1.496976 Training: 26%|██▋ | 4733/18018 [10:16:30<28:12:23, 7.64s/step, loss=1.497, acc=0.670, acc_len=1.59, thru=19.3, I=21.3, T=31.3, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:44,156] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:46:52,564] loop.py:45 INFO TRAIN_STEP step=4734 loss=1.683039 Training: 26%|██▋ | 4734/18018 [10:16:38<29:16:55, 7.94s/step, loss=1.683, acc=0.636, acc_len=1.48, thru=12.9, I=14.9, T=32.8, wait=4.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:58,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:46:59,431] loop.py:380 INFO TIMING step=4735: step=4.464s data=2.078s compute=4.428s [fwd=3.048s bwd=1.365s opt=0.003s] dispatch=2.397s [2026-10-18 23:46:59,431] loop.py:45 INFO TRAIN_STEP step=4735 loss=1.600364 Training: 26%|██▋ | 4735/18018 [10:16:45<28:05:48, 7.61s/step, loss=1.600, acc=0.655, acc_len=1.54, thru=18.2, I=18.5, T=28.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:46:59,711] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:05,576] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:47:06,349] loop.py:45 INFO TRAIN_STEP step=4736 loss=1.564211 Training: 26%|██▋ | 4736/18018 [10:16:52<27:19:27, 7.41s/step, loss=1.564, acc=0.662, acc_len=1.56, thru=17.6, I=20.3, T=31.5, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:47:06,344] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4735: forward=2861.7ms backward=1164.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:06,582] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:14,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:47:15,909] loop.py:45 INFO TRAIN_STEP step=4737 loss=1.505593 Training: 26%|██▋ | 4737/18018 [10:17:01<29:42:24, 8.05s/step, loss=1.506, acc=0.665, acc_len=1.57, thru=13.7, I=17.2, T=28.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:16,114] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:47:22,595] loop.py:45 INFO TRAIN_STEP step=4738 loss=1.471419 Training: 26%|██▋ | 4738/18018 [10:17:08<28:11:37, 7.64s/step, loss=1.471, acc=0.677, acc_len=1.61, thru=16.6, I=22.0, T=32.1, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:29,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:47:30,797] loop.py:45 INFO TRAIN_STEP step=4739 loss=1.614472 Training: 26%|██▋ | 4739/18018 [10:17:16<28:48:33, 7.81s/step, loss=1.614, acc=0.652, acc_len=1.53, thru=15.1, I=15.2, T=29.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:31,227] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:47:40,326] loop.py:380 INFO TIMING step=4740: step=4.054s data=1.829s compute=4.012s [fwd=2.764s bwd=1.234s opt=0.003s] dispatch=5.466s [2026-10-18 23:47:40,326] loop.py:45 INFO TRAIN_STEP step=4740 loss=1.632280 Training: 26%|██▋ | 4740/18018 [10:17:26<30:42:25, 8.33s/step, loss=1.632, acc=0.646, acc_len=1.51, thru=14.1, I=12.9, T=31.6, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:40,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:40,802] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:46,192] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:47:47,257] loop.py:45 INFO TRAIN_STEP step=4741 loss=1.619201 Training: 26%|██▋ | 4741/18018 [10:17:33<29:09:47, 7.91s/step, loss=1.619, acc=0.649, acc_len=1.52, thru=18.2, I=18.9, T=31.6, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:47:47,252] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4740: forward=2806.7ms backward=1199.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:47,704] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:51,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:47:53,477] loop.py:45 INFO TRAIN_STEP step=4742 loss=1.497235 Training: 26%|██▋ | 4742/18018 [10:17:39<27:17:39, 7.40s/step, loss=1.497, acc=0.669, acc_len=1.59, thru=19.0, I=26.9, T=32.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:53,626] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:58,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:47:59,182] loop.py:45 INFO TRAIN_STEP step=4743 loss=1.611244 Training: 26%|██▋ | 4743/18018 [10:17:45<25:24:58, 6.89s/step, loss=1.611, acc=0.654, acc_len=1.53, thru=20.7, I=27.1, T=35.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:47:59,269] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:04,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:48:05,359] loop.py:45 INFO TRAIN_STEP step=4744 loss=1.538886 Training: 26%|██▋ | 4744/18018 [10:17:51<24:37:25, 6.68s/step, loss=1.539, acc=0.663, acc_len=1.57, thru=20.5, I=22.4, T=32.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:05,790] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:12,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:48:14,446] loop.py:380 INFO TIMING step=4745: step=4.860s data=2.262s compute=4.826s [fwd=3.379s bwd=1.432s opt=0.003s] dispatch=4.217s [2026-10-18 23:48:14,447] loop.py:45 INFO TRAIN_STEP step=4745 loss=1.673634 Training: 26%|██▋ | 4745/18018 [10:18:00<27:17:07, 7.40s/step, loss=1.674, acc=0.641, acc_len=1.50, thru=14.2, I=18.5, T=26.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:14,640] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:22,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:48:24,073] loop.py:45 INFO TRAIN_STEP step=4746 loss=1.648073 Training: 26%|██▋ | 4746/18018 [10:18:10<29:44:39, 8.07s/step, loss=1.648, acc=0.644, acc_len=1.50, thru=13.6, I=15.9, T=30.1, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:48:24,068] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4745: forward=2928.6ms backward=1257.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:24,571] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:30,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:48:31,542] loop.py:45 INFO TRAIN_STEP step=4747 loss=1.557784 Training: 26%|██▋ | 4747/18018 [10:18:17<29:04:57, 7.89s/step, loss=1.558, acc=0.659, acc_len=1.56, thru=15.8, I=20.6, T=28.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:31,840] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:36,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:48:37,852] loop.py:45 INFO TRAIN_STEP step=4748 loss=1.746876 Training: 26%|██▋ | 4748/18018 [10:18:23<27:19:53, 7.41s/step, loss=1.747, acc=0.630, acc_len=1.46, thru=17.7, I=21.5, T=30.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:38,357] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:48:46,305] loop.py:45 INFO TRAIN_STEP step=4749 loss=1.548161 Training: 26%|██▋ | 4749/18018 [10:18:32<28:28:48, 7.73s/step, loss=1.548, acc=0.665, acc_len=1.58, thru=11.2, I=15.2, T=27.9, wait=3.9s, pool=44, epoch=1/2][2026-10-18 23:48:58,145] loop.py:380 INFO TIMING step=4750: step=4.386s data=1.968s compute=4.354s [fwd=3.042s bwd=1.299s opt=0.003s] dispatch=7.442s [2026-10-18 23:48:58,145] loop.py:45 INFO TRAIN_STEP step=4750 loss=1.675359 Training: 26%|██▋ | 4750/18018 [10:18:44<33:01:31, 8.96s/step, loss=1.675, acc=0.644, acc_len=1.50, thru=14.2, I=11.9, T=29.2, wait=7.4s, pool=56, epoch=1/2][2026-10-18 23:48:58,150] loop.py:419 INFO Saving checkpoint at step 4750... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:48:58,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 23:49:02,296] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0004751 (TrainerActor pid=9310) [2026-10-18 23:49:02,315] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0004001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:03,811] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:03,811] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:10,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-18 23:49:02,295] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0004751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-18 23:49:03,542] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0004001 [2026-10-18 23:49:10,798] loop.py:45 INFO TRAIN_STEP step=4751 loss=1.526213 Training: 26%|██▋ | 4751/18018 [10:18:56<37:06:12, 10.07s/step, loss=1.526, acc=0.663, acc_len=1.57, thru=18.3, I=19.6, T=32.6, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:49:10,793] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4750: forward=2717.1ms backward=1152.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:11,256] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 23:49:17,539] loop.py:45 INFO TRAIN_STEP step=4752 loss=1.615439 Training: 26%|██▋ | 4752/18018 [10:19:03<33:25:15, 9.07s/step, loss=1.615, acc=0.646, acc_len=1.51, thru=17.2, I=20.4, T=35.0, wait=3.1s, pool=48, epoch=1/2][2026-10-18 23:49:25,894] loop.py:45 INFO TRAIN_STEP step=4753 loss=1.721136 Training: 26%|██▋ | 4753/18018 [10:19:11<32:37:47, 8.86s/step, loss=1.721, acc=0.635, acc_len=1.48, thru=16.0, I=15.1, T=30.8, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:30,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:49:32,776] loop.py:45 INFO TRAIN_STEP step=4754 loss=1.609235 Training: 26%|██▋ | 4754/18018 [10:19:18<30:26:51, 8.26s/step, loss=1.609, acc=0.653, acc_len=1.53, thru=17.6, I=20.5, T=30.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:32,995] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:36,750] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-18 23:49:38,532] loop.py:380 INFO TIMING step=4755: step=3.632s data=1.618s compute=3.604s [fwd=2.576s bwd=1.014s opt=0.003s] dispatch=2.114s [2026-10-18 23:49:38,532] loop.py:45 INFO TRAIN_STEP step=4755 loss=1.644406 Training: 26%|██▋ | 4755/18018 [10:19:24<27:40:17, 7.51s/step, loss=1.644, acc=0.643, acc_len=1.50, thru=20.1, I=30.3, T=35.2, wait=2.1s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:38,760] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:44,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:49:44,401] loop.py:45 INFO TRAIN_STEP step=4756 loss=1.474408 Training: 26%|██▋ | 4756/18018 [10:19:30<25:51:24, 7.02s/step, loss=1.474, acc=0.674, acc_len=1.60, thru=21.7, I=30.1, T=36.7, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:49:44,396] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4755: forward=2413.5ms backward=1028.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:44,514] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:50,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:49:51,700] loop.py:45 INFO TRAIN_STEP step=4757 loss=1.524487 Training: 26%|██▋ | 4757/18018 [10:19:37<26:09:51, 7.10s/step, loss=1.524, acc=0.663, acc_len=1.57, thru=18.1, I=20.6, T=33.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:51,986] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:58,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:49:58,532] loop.py:45 INFO TRAIN_STEP step=4758 loss=1.661424 Training: 26%|██▋ | 4758/18018 [10:19:44<25:51:42, 7.02s/step, loss=1.661, acc=0.641, acc_len=1.50, thru=17.7, I=20.8, T=31.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:49:58,861] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:05,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:50:06,805] loop.py:45 INFO TRAIN_STEP step=4759 loss=1.525084 Training: 26%|██▋ | 4759/18018 [10:19:52<27:14:34, 7.40s/step, loss=1.525, acc=0.666, acc_len=1.58, thru=16.0, I=18.0, T=30.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:07,162] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:14,907] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:50:16,567] loop.py:380 INFO TIMING step=4760: step=4.313s data=1.939s compute=4.276s [fwd=3.040s bwd=1.221s opt=0.003s] dispatch=5.440s [2026-10-18 23:50:16,567] loop.py:45 INFO TRAIN_STEP step=4760 loss=1.596185 Training: 26%|██▋ | 4760/18018 [10:20:02<29:51:16, 8.11s/step, loss=1.596, acc=0.656, acc_len=1.54, thru=13.6, I=15.3, T=29.7, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:16,925] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:28,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:50:31,201] loop.py:45 INFO TRAIN_STEP step=4761 loss=1.547391 Training: 26%|██▋ | 4761/18018 [10:20:17<37:03:40, 10.06s/step, loss=1.547, acc=0.662, acc_len=1.57, thru=10.5, I=9.4, T=25.1, wait=9.5s, pool=64, epoch=1/2] (TrainerActor pid=9310) [2026-10-18 23:50:31,196] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4760: forward=3172.6ms backward=1491.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:31,701] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:36,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:50:38,095] loop.py:45 INFO TRAIN_STEP step=4762 loss=1.506958 Training: 26%|██▋ | 4762/18018 [10:20:24<33:33:33, 9.11s/step, loss=1.507, acc=0.668, acc_len=1.58, thru=14.4, I=24.9, T=31.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:38,315] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:50:44,945] loop.py:45 INFO TRAIN_STEP step=4763 loss=1.534711 Training: 26%|██▋ | 4763/18018 [10:20:31<31:03:11, 8.43s/step, loss=1.535, acc=0.659, acc_len=1.55, thru=17.3, I=23.5, T=31.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:44,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:45,412] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:53,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:50:53,752] loop.py:45 INFO TRAIN_STEP step=4764 loss=1.469822 Training: 26%|██▋ | 4764/18018 [10:20:39<31:27:59, 8.55s/step, loss=1.470, acc=0.676, acc_len=1.62, thru=14.9, I=15.1, T=30.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:50:54,039] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:00,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:51:01,544] loop.py:380 INFO TIMING step=4765: step=3.867s data=1.631s compute=3.836s [fwd=2.700s bwd=1.121s opt=0.003s] dispatch=3.913s [2026-10-18 23:51:01,544] loop.py:45 INFO TRAIN_STEP step=4765 loss=1.603398 Training: 26%|██▋ | 4765/18018 [10:20:47<30:37:39, 8.32s/step, loss=1.603, acc=0.654, acc_len=1.54, thru=17.1, I=18.4, T=33.1, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:01,827] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:09,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:51:10,799] loop.py:45 INFO TRAIN_STEP step=4766 loss=1.595357 Training: 26%|██▋ | 4766/18018 [10:20:56<31:39:41, 8.60s/step, loss=1.595, acc=0.656, acc_len=1.55, thru=14.3, I=16.1, T=29.5, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:51:10,794] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4765: forward=3025.9ms backward=1270.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:11,206] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:17,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:51:17,734] loop.py:45 INFO TRAIN_STEP step=4767 loss=1.623152 Training: 26%|██▋ | 4767/18018 [10:21:03<29:49:02, 8.10s/step, loss=1.623, acc=0.652, acc_len=1.53, thru=16.5, I=21.9, T=31.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:18,087] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:22,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:51:24,199] loop.py:45 INFO TRAIN_STEP step=4768 loss=1.535814 Training: 26%|██▋ | 4768/18018 [10:21:10<28:00:36, 7.61s/step, loss=1.536, acc=0.662, acc_len=1.57, thru=19.1, I=23.6, T=29.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:24,623] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:30,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:51:30,962] loop.py:45 INFO TRAIN_STEP step=4769 loss=1.548005 Training: 26%|██▋ | 4769/18018 [10:21:17<27:04:24, 7.36s/step, loss=1.548, acc=0.660, acc_len=1.56, thru=17.3, I=23.0, T=31.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:31,318] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:37,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:51:38,667] loop.py:380 INFO TIMING step=4770: step=4.023s data=1.776s compute=3.991s [fwd=2.850s bwd=1.127s opt=0.003s] dispatch=3.670s [2026-10-18 23:51:38,667] loop.py:45 INFO TRAIN_STEP step=4770 loss=1.558969 Training: 26%|██▋ | 4770/18018 [10:21:24<27:27:19, 7.46s/step, loss=1.559, acc=0.659, acc_len=1.56, thru=16.2, I=19.4, T=31.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:39,106] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:46,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:51:47,371] loop.py:45 INFO TRAIN_STEP step=4771 loss=1.518811 Training: 26%|██▋ | 4771/18018 [10:21:33<28:49:27, 7.83s/step, loss=1.519, acc=0.668, acc_len=1.59, thru=13.8, I=19.7, T=32.9, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:51:47,367] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4770: forward=2745.1ms backward=1097.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:47,581] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:55,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:51:57,168] loop.py:45 INFO TRAIN_STEP step=4772 loss=1.538172 Training: 26%|██▋ | 4772/18018 [10:21:43<30:59:24, 8.42s/step, loss=1.538, acc=0.663, acc_len=1.56, thru=14.2, I=14.1, T=29.1, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:51:57,373] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:02,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:52:04,544] loop.py:45 INFO TRAIN_STEP step=4773 loss=1.593143 Training: 26%|██▋ | 4773/18018 [10:21:50<29:49:55, 8.11s/step, loss=1.593, acc=0.650, acc_len=1.52, thru=15.7, I=23.3, T=27.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:04,846] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:11,033] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:52:11,207] loop.py:45 INFO TRAIN_STEP step=4774 loss=1.447710 Training: 26%|██▋ | 4774/18018 [10:21:57<28:14:02, 7.67s/step, loss=1.448, acc=0.675, acc_len=1.61, thru=16.2, I=19.5, T=32.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:11,535] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:16,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:52:18,563] loop.py:380 INFO TIMING step=4775: step=4.569s data=2.257s compute=4.536s [fwd=3.234s bwd=1.286s opt=0.003s] dispatch=2.778s [2026-10-18 23:52:18,563] loop.py:45 INFO TRAIN_STEP step=4775 loss=1.524757 Training: 27%|██▋ | 4775/18018 [10:22:04<27:52:59, 7.58s/step, loss=1.525, acc=0.663, acc_len=1.57, thru=17.8, I=21.5, T=28.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:18,619] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:24,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:52:24,354] loop.py:45 INFO TRAIN_STEP step=4776 loss=1.540098 Training: 27%|██▋ | 4776/18018 [10:22:10<25:54:18, 7.04s/step, loss=1.540, acc=0.663, acc_len=1.57, thru=18.1, I=23.5, T=33.1, wait=1.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:52:24,349] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4775: forward=2689.2ms backward=1123.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:24,510] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:30,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:52:32,017] loop.py:45 INFO TRAIN_STEP step=4777 loss=1.499737 Training: 27%|██▋ | 4777/18018 [10:22:18<26:35:14, 7.23s/step, loss=1.500, acc=0.672, acc_len=1.60, thru=17.3, I=18.9, T=26.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:32,496] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:37,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:52:38,082] loop.py:45 INFO TRAIN_STEP step=4778 loss=1.581551 Training: 27%|██▋ | 4778/18018 [10:22:24<25:18:06, 6.88s/step, loss=1.582, acc=0.653, acc_len=1.53, thru=18.1, I=22.5, T=32.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:38,240] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:43,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:52:45,507] loop.py:45 INFO TRAIN_STEP step=4779 loss=1.674742 Training: 27%|██▋ | 4779/18018 [10:22:34<29:01:37, 7.89s/step, loss=1.675, acc=0.640, acc_len=1.50, thru=12.2, I=23.2, T=28.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:48,843] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.5s): pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:48,843] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:54,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:52:54,926] loop.py:380 INFO TIMING step=4780: step=3.765s data=1.719s compute=3.730s [fwd=2.645s bwd=1.071s opt=0.003s] dispatch=2.812s [2026-10-18 23:52:54,926] loop.py:45 INFO TRAIN_STEP step=4780 loss=1.471458 Training: 27%|██▋ | 4780/18018 [10:22:41<27:35:04, 7.50s/step, loss=1.471, acc=0.672, acc_len=1.60, thru=21.0, I=24.3, T=34.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:52:55,115] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-18 23:53:01,350] loop.py:45 INFO TRAIN_STEP step=4781 loss=1.641721 Training: 27%|██▋ | 4781/18018 [10:22:47<26:23:31, 7.18s/step, loss=1.642, acc=0.648, acc_len=1.52, thru=18.6, I=21.8, T=29.3, wait=2.0s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:53:01,345] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4780: forward=2982.7ms backward=1346.1ms [2026-10-18 23:53:10,110] loop.py:45 INFO TRAIN_STEP step=4782 loss=1.486422 Training: 27%|██▋ | 4782/18018 [10:22:56<28:08:03, 7.65s/step, loss=1.486, acc=0.670, acc_len=1.59, thru=14.2, I=14.2, T=30.4, wait=4.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:53:17,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:53:19,095] loop.py:45 INFO TRAIN_STEP step=4783 loss=1.576043 Training: 27%|██▋ | 4783/18018 [10:23:05<29:36:17, 8.05s/step, loss=1.576, acc=0.655, acc_len=1.54, thru=16.4, I=17.0, T=29.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:53:19,147] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-18 23:53:25,319] loop.py:45 INFO TRAIN_STEP step=4784 loss=1.583737 Training: 27%|██▋ | 4784/18018 [10:23:11<27:35:03, 7.50s/step, loss=1.584, acc=0.650, acc_len=1.53, thru=17.6, I=22.9, T=32.1, wait=2.2s, pool=52, epoch=1/2][2026-10-18 23:53:34,114] loop.py:380 INFO TIMING step=4785: step=4.560s data=2.178s compute=4.472s [fwd=3.076s bwd=1.381s opt=0.003s] dispatch=4.226s [2026-10-18 23:53:34,114] loop.py:45 INFO TRAIN_STEP step=4785 loss=1.636812 Training: 27%|██▋ | 4785/18018 [10:23:20<29:00:16, 7.89s/step, loss=1.637, acc=0.646, acc_len=1.51, thru=14.4, I=14.0, T=28.1, wait=4.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:53:42,021] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:53:42,547] loop.py:45 INFO TRAIN_STEP step=4786 loss=1.589157 Training: 27%|██▋ | 4786/18018 [10:23:28<29:36:06, 8.05s/step, loss=1.589, acc=0.654, acc_len=1.54, thru=16.0, I=14.9, T=29.0, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:53:42,542] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4785: forward=3029.0ms backward=1338.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:53:43,028] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:53:51,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:53:52,810] loop.py:45 INFO TRAIN_STEP step=4787 loss=1.602516 Training: 27%|██▋ | 4787/18018 [10:23:38<32:02:06, 8.72s/step, loss=1.603, acc=0.655, acc_len=1.53, thru=14.4, I=14.2, T=30.5, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:53:53,218] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:53:58,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:53:59,807] loop.py:45 INFO TRAIN_STEP step=4788 loss=1.288205 Training: 27%|██▋ | 4788/18018 [10:23:45<30:08:14, 8.20s/step, loss=1.288, acc=0.702, acc_len=1.70, thru=16.0, I=23.5, T=27.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:54:00,069] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:54:06,234] loop.py:45 INFO TRAIN_STEP step=4789 loss=1.544994 Training: 27%|██▋ | 4789/18018 [10:23:52<28:10:55, 7.67s/step, loss=1.545, acc=0.662, acc_len=1.56, thru=15.1, I=23.9, T=35.5, wait=2.8s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:54:15,558] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:54:16,492] loop.py:380 INFO TIMING step=4790: step=4.491s data=2.037s compute=4.457s [fwd=3.129s bwd=1.313s opt=0.003s] dispatch=5.757s [2026-10-18 23:54:16,492] loop.py:45 INFO TRAIN_STEP step=4790 loss=1.621170 Training: 27%|██▋ | 4790/18018 [10:24:02<31:02:00, 8.45s/step, loss=1.621, acc=0.650, acc_len=1.52, thru=16.1, I=14.6, T=28.5, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:54:16,565] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:54:22,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:54:24,464] loop.py:45 INFO TRAIN_STEP step=4791 loss=1.634427 Training: 27%|██▋ | 4791/18018 [10:24:10<30:30:24, 8.30s/step, loss=1.634, acc=0.649, acc_len=1.52, thru=14.4, I=17.4, T=29.4, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:54:24,459] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4790: forward=3073.4ms backward=1232.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:54:24,785] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:54:30,882] loop.py:45 INFO TRAIN_STEP step=4792 loss=1.499173 Training: 27%|██▋ | 4792/18018 [10:24:16<28:25:47, 7.74s/step, loss=1.499, acc=0.672, acc_len=1.59, thru=16.5, I=21.9, T=33.9, wait=2.6s, pool=44, epoch=1/2][2026-10-18 23:54:37,961] loop.py:45 INFO TRAIN_STEP step=4793 loss=1.552966 Training: 27%|██▋ | 4793/18018 [10:24:24<27:41:51, 7.54s/step, loss=1.553, acc=0.657, acc_len=1.55, thru=16.9, I=17.0, T=34.9, wait=3.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:54:47,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:54:48,121] loop.py:45 INFO TRAIN_STEP step=4794 loss=1.559247 Training: 27%|██▋ | 4794/18018 [10:24:34<30:35:11, 8.33s/step, loss=1.559, acc=0.656, acc_len=1.55, thru=14.2, I=13.7, T=29.8, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:54:48,364] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:54:53,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:54:54,364] loop.py:380 INFO TIMING step=4795: step=3.978s data=1.824s compute=3.938s [fwd=2.768s bwd=1.155s opt=0.003s] dispatch=2.254s [2026-10-18 23:54:54,364] loop.py:45 INFO TRAIN_STEP step=4795 loss=1.538851 Training: 27%|██▋ | 4795/18018 [10:24:40<28:17:13, 7.70s/step, loss=1.539, acc=0.663, acc_len=1.57, thru=18.5, I=21.4, T=32.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:54:54,596] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:00,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:55:01,199] loop.py:45 INFO TRAIN_STEP step=4796 loss=1.544254 Training: 27%|██▋ | 4796/18018 [10:24:47<27:19:54, 7.44s/step, loss=1.544, acc=0.660, acc_len=1.57, thru=16.6, I=20.2, T=29.1, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:55:01,194] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4795: forward=3055.9ms backward=1287.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:01,536] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:06,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:55:07,058] loop.py:45 INFO TRAIN_STEP step=4797 loss=1.610656 Training: 27%|██▋ | 4797/18018 [10:24:53<25:35:00, 6.97s/step, loss=1.611, acc=0.647, acc_len=1.51, thru=19.3, I=22.7, T=33.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:07,195] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:12,919] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:55:14,337] loop.py:45 INFO TRAIN_STEP step=4798 loss=1.591389 Training: 27%|██▋ | 4798/18018 [10:25:00<25:55:38, 7.06s/step, loss=1.591, acc=0.653, acc_len=1.54, thru=17.3, I=21.6, T=31.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:14,432] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-18 23:55:20,339] loop.py:45 INFO TRAIN_STEP step=4799 loss=1.456598 Training: 27%|██▋ | 4799/18018 [10:25:06<24:45:28, 6.74s/step, loss=1.457, acc=0.679, acc_len=1.62, thru=17.0, I=18.5, T=36.4, wait=2.5s, pool=40, epoch=1/2][2026-10-18 23:55:28,709] loop.py:380 INFO TIMING step=4800: step=3.882s data=1.663s compute=3.845s [fwd=2.768s bwd=1.063s opt=0.003s] dispatch=4.479s [2026-10-18 23:55:28,709] loop.py:45 INFO TRAIN_STEP step=4800 loss=1.553669 Training: 27%|██▋ | 4800/18018 [10:25:14<26:32:55, 7.23s/step, loss=1.554, acc=0.663, acc_len=1.57, thru=15.9, I=15.7, T=33.0, wait=4.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:37,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:55:39,072] loop.py:45 INFO TRAIN_STEP step=4801 loss=1.631676 Training: 27%|██▋ | 4801/18018 [10:25:25<29:59:52, 8.17s/step, loss=1.632, acc=0.649, acc_len=1.52, thru=14.5, I=14.4, T=27.2, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:55:39,067] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4800: forward=3145.2ms backward=1502.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:39,287] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:45,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:55:46,125] loop.py:45 INFO TRAIN_STEP step=4802 loss=1.649483 Training: 27%|██▋ | 4802/18018 [10:25:32<28:45:58, 7.84s/step, loss=1.649, acc=0.645, acc_len=1.51, thru=16.5, I=20.0, T=29.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:46,420] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:52,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:55:52,657] loop.py:45 INFO TRAIN_STEP step=4803 loss=1.602069 Training: 27%|██▋ | 4803/18018 [10:25:38<27:19:29, 7.44s/step, loss=1.602, acc=0.656, acc_len=1.54, thru=18.3, I=19.4, T=33.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:52,936] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:55:58,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:56:00,100] loop.py:45 INFO TRAIN_STEP step=4804 loss=1.626673 Training: 27%|██▋ | 4804/18018 [10:25:46<27:19:20, 7.44s/step, loss=1.627, acc=0.647, acc_len=1.52, thru=18.0, I=20.6, T=32.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:00,267] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-18 23:56:07,729] loop.py:380 INFO TIMING step=4805: step=4.316s data=2.060s compute=4.279s [fwd=3.041s bwd=1.223s opt=0.003s] dispatch=3.304s [2026-10-18 23:56:07,730] loop.py:45 INFO TRAIN_STEP step=4805 loss=1.609082 Training: 27%|██▋ | 4805/18018 [10:25:53<27:31:36, 7.50s/step, loss=1.609, acc=0.649, acc_len=1.52, thru=14.2, I=19.1, T=29.7, wait=3.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:17,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:56:17,941] loop.py:45 INFO TRAIN_STEP step=4806 loss=1.792564 Training: 27%|██▋ | 4806/18018 [10:26:04<30:30:35, 8.31s/step, loss=1.793, acc=0.622, acc_len=1.44, thru=15.1, I=13.7, T=27.3, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:56:17,936] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4805: forward=3311.3ms backward=1317.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:17,975] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:23,733] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:56:24,663] loop.py:45 INFO TRAIN_STEP step=4807 loss=1.612315 Training: 27%|██▋ | 4807/18018 [10:26:10<28:45:24, 7.84s/step, loss=1.612, acc=0.650, acc_len=1.52, thru=17.3, I=19.0, T=31.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:24,749] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:32,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:56:34,238] loop.py:45 INFO TRAIN_STEP step=4808 loss=1.596621 Training: 27%|██▋ | 4808/18018 [10:26:20<30:39:56, 8.36s/step, loss=1.597, acc=0.655, acc_len=1.54, thru=13.5, I=15.9, T=31.3, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:34,343] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:37,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:56:40,543] loop.py:45 INFO TRAIN_STEP step=4809 loss=1.618312 Training: 27%|██▋ | 4809/18018 [10:26:26<28:24:25, 7.74s/step, loss=1.618, acc=0.650, acc_len=1.52, thru=18.9, I=27.6, T=27.5, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:40,938] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:47,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:56:49,703] loop.py:380 INFO TIMING step=4810: step=4.333s data=1.906s compute=4.296s [fwd=2.999s bwd=1.283s opt=0.003s] dispatch=4.815s [2026-10-18 23:56:49,703] loop.py:45 INFO TRAIN_STEP step=4810 loss=1.540655 Training: 27%|██▋ | 4810/18018 [10:26:35<29:57:54, 8.17s/step, loss=1.541, acc=0.661, acc_len=1.56, thru=14.6, I=18.2, T=29.5, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:56:49,963] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-18 23:56:55,855] loop.py:45 INFO TRAIN_STEP step=4811 loss=1.564478 (TrainerActor pid=9310) [2026-10-18 23:56:55,850] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4810: forward=2567.9ms backward=1108.9ms Training: 27%|██▋ | 4811/18018 [10:26:41<27:44:36, 7.56s/step, loss=1.564, acc=0.657, acc_len=1.55, thru=17.5, I=23.4, T=34.4, wait=2.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:03,218] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:57:03,361] loop.py:45 INFO TRAIN_STEP step=4812 loss=1.396278 Training: 27%|██▋ | 4812/18018 [10:26:49<27:40:44, 7.55s/step, loss=1.396, acc=0.683, acc_len=1.64, thru=17.2, I=17.3, T=32.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:03,720] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:10,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:57:10,658] loop.py:45 INFO TRAIN_STEP step=4813 loss=1.500643 Training: 27%|██▋ | 4813/18018 [10:26:56<27:24:18, 7.47s/step, loss=1.501, acc=0.671, acc_len=1.60, thru=17.0, I=18.4, T=34.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:11,113] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:18,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:57:18,659] loop.py:45 INFO TRAIN_STEP step=4814 loss=1.699367 Training: 27%|██▋ | 4814/18018 [10:27:04<27:59:01, 7.63s/step, loss=1.699, acc=0.639, acc_len=1.48, thru=15.6, I=16.7, T=33.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:18,984] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:24,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:57:24,988] loop.py:380 INFO TIMING step=4815: step=4.023s data=1.801s compute=3.985s [fwd=2.831s bwd=1.139s opt=0.003s] dispatch=2.297s [2026-10-18 23:57:24,988] loop.py:45 INFO TRAIN_STEP step=4815 loss=1.494328 Training: 27%|██▋ | 4815/18018 [10:27:11<26:33:04, 7.24s/step, loss=1.494, acc=0.669, acc_len=1.58, thru=19.6, I=19.9, T=31.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:25,204] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:32,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:57:33,223] loop.py:45 INFO TRAIN_STEP step=4816 loss=1.457905 Training: 27%|██▋ | 4816/18018 [10:27:19<27:38:40, 7.54s/step, loss=1.458, acc=0.674, acc_len=1.60, thru=17.1, I=16.6, T=30.9, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:57:33,219] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4815: forward=2819.7ms backward=1273.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:33,368] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:40,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:57:40,899] loop.py:45 INFO TRAIN_STEP step=4817 loss=1.596839 Training: 27%|██▋ | 4817/18018 [10:27:26<27:47:38, 7.58s/step, loss=1.597, acc=0.655, acc_len=1.54, thru=17.8, I=17.3, T=29.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:41,054] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:45,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:57:47,119] loop.py:45 INFO TRAIN_STEP step=4818 loss=1.652474 Training: 27%|██▋ | 4818/18018 [10:27:33<26:17:52, 7.17s/step, loss=1.652, acc=0.648, acc_len=1.52, thru=18.7, I=21.5, T=31.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:47,318] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:51,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:57:53,508] loop.py:45 INFO TRAIN_STEP step=4819 loss=1.538558 Training: 27%|██▋ | 4819/18018 [10:27:39<25:25:58, 6.94s/step, loss=1.539, acc=0.662, acc_len=1.57, thru=18.8, I=27.1, T=29.8, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:57:53,794] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:58:00,179] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:58:02,155] loop.py:380 INFO TIMING step=4820: step=5.330s data=2.603s compute=5.302s [fwd=3.699s bwd=1.588s opt=0.003s] dispatch=3.308s [2026-10-18 23:58:02,155] loop.py:45 INFO TRAIN_STEP step=4820 loss=1.520314 Training: 27%|██▋ | 4820/18018 [10:27:48<27:18:51, 7.45s/step, loss=1.520, acc=0.664, acc_len=1.57, thru=15.3, I=19.9, T=24.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:58:02,211] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:58:07,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:58:08,420] loop.py:45 INFO TRAIN_STEP step=4821 loss=1.520802 Training: 27%|██▋ | 4821/18018 [10:27:54<26:00:29, 7.09s/step, loss=1.521, acc=0.670, acc_len=1.59, thru=16.0, I=21.7, T=32.3, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:58:08,415] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4820: forward=2803.2ms backward=1111.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:58:08,529] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-18 23:58:15,158] loop.py:45 INFO TRAIN_STEP step=4822 loss=1.451872 Training: 27%|██▋ | 4822/18018 [10:28:01<25:36:46, 6.99s/step, loss=1.452, acc=0.675, acc_len=1.61, thru=16.6, I=18.0, T=30.8, wait=2.6s, pool=52, epoch=1/2][2026-10-18 23:58:24,865] loop.py:45 INFO TRAIN_STEP step=4823 loss=1.485180 Training: 27%|██▋ | 4823/18018 [10:28:10<28:35:55, 7.80s/step, loss=1.485, acc=0.673, acc_len=1.60, thru=14.2, I=13.7, T=28.7, wait=5.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:58:29,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:58:31,833] loop.py:45 INFO TRAIN_STEP step=4824 loss=1.346105 Training: 27%|██▋ | 4824/18018 [10:28:17<27:40:51, 7.55s/step, loss=1.346, acc=0.699, acc_len=1.68, thru=20.9, I=19.2, T=28.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:58:31,863] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 [2026-10-18 23:58:39,362] loop.py:380 INFO TIMING step=4825: step=4.785s data=2.221s compute=4.742s [fwd=3.329s bwd=1.398s opt=0.003s] dispatch=2.734s [2026-10-18 23:58:39,362] loop.py:45 INFO TRAIN_STEP step=4825 loss=1.599745 Training: 27%|██▋ | 4825/18018 [10:28:25<27:39:12, 7.55s/step, loss=1.600, acc=0.653, acc_len=1.53, thru=16.6, I=18.0, T=26.8, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:58:46,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:58:47,606] loop.py:45 INFO TRAIN_STEP step=4826 loss=1.507459 Training: 27%|██▋ | 4826/18018 [10:28:33<28:25:04, 7.76s/step, loss=1.507, acc=0.669, acc_len=1.58, thru=15.7, I=16.2, T=31.5, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:58:47,601] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4825: forward=2851.5ms backward=1165.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:58:47,636] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-18 23:58:55,439] loop.py:45 INFO TRAIN_STEP step=4827 loss=1.685969 Training: 27%|██▋ | 4827/18018 [10:28:41<28:30:01, 7.78s/step, loss=1.686, acc=0.639, acc_len=1.50, thru=14.1, I=16.0, T=31.4, wait=3.8s, pool=60, epoch=1/2][2026-10-18 23:59:02,987] loop.py:45 INFO TRAIN_STEP step=4828 loss=1.633905 Training: 27%|██▋ | 4828/18018 [10:28:49<28:14:52, 7.71s/step, loss=1.634, acc=0.648, acc_len=1.52, thru=13.8, I=16.1, T=34.9, wait=3.9s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:12,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:59:14,572] loop.py:45 INFO TRAIN_STEP step=4829 loss=1.694618 Training: 27%|██▋ | 4829/18018 [10:29:00<32:30:12, 8.87s/step, loss=1.695, acc=0.642, acc_len=1.49, thru=13.3, I=15.8, T=30.6, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:14,766] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:21,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:59:22,379] loop.py:380 INFO TIMING step=4830: step=4.583s data=2.125s compute=4.552s [fwd=3.214s bwd=1.321s opt=0.003s] dispatch=3.215s [2026-10-18 23:59:22,380] loop.py:45 INFO TRAIN_STEP step=4830 loss=1.540343 Training: 27%|██▋ | 4830/18018 [10:29:08<31:19:50, 8.55s/step, loss=1.540, acc=0.663, acc_len=1.57, thru=13.6, I=17.9, T=27.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:22,459] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-18 23:59:29,872] loop.py:45 INFO TRAIN_STEP step=4831 loss=1.398888 Training: 27%|██▋ | 4831/18018 [10:29:15<30:09:47, 8.23s/step, loss=1.399, acc=0.685, acc_len=1.64, thru=16.0, I=18.0, T=30.0, wait=3.2s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-18 23:59:29,867] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4830: forward=3035.8ms backward=1183.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:37,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:59:37,993] loop.py:45 INFO TRAIN_STEP step=4832 loss=1.548203 Training: 27%|██▋ | 4832/18018 [10:29:24<30:02:11, 8.20s/step, loss=1.548, acc=0.662, acc_len=1.56, thru=15.0, I=15.2, T=31.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:38,227] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:44,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:59:46,282] loop.py:45 INFO TRAIN_STEP step=4833 loss=1.646766 Training: 27%|██▋ | 4833/18018 [10:29:32<30:07:51, 8.23s/step, loss=1.647, acc=0.644, acc_len=1.51, thru=15.9, I=20.3, T=29.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:46,549] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:52,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-18 23:59:53,715] loop.py:45 INFO TRAIN_STEP step=4834 loss=1.511808 Training: 27%|██▋ | 4834/18018 [10:29:39<29:15:29, 7.99s/step, loss=1.512, acc=0.666, acc_len=1.57, thru=14.3, I=18.9, T=29.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-18 23:59:53,810] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-19 00:00:00,436] loop.py:380 INFO TIMING step=4835: step=3.606s data=1.578s compute=3.567s [fwd=2.542s bwd=1.010s opt=0.003s] dispatch=3.105s [2026-10-19 00:00:00,437] loop.py:45 INFO TRAIN_STEP step=4835 loss=1.614890 Training: 27%|██▋ | 4835/18018 [10:29:46<27:51:43, 7.61s/step, loss=1.615, acc=0.652, acc_len=1.53, thru=15.0, I=21.6, T=35.5, wait=3.1s, pool=44, epoch=1/2][2026-10-19 00:00:07,356] loop.py:45 INFO TRAIN_STEP step=4836 loss=1.554647 Training: 27%|██▋ | 4836/18018 [10:29:53<27:06:12, 7.40s/step, loss=1.555, acc=0.660, acc_len=1.56, thru=17.7, I=20.7, T=35.3, wait=3.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:00:07,351] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4835: forward=2543.1ms backward=1037.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:13,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:00:14,524] loop.py:45 INFO TRAIN_STEP step=4837 loss=1.572031 Training: 27%|██▋ | 4837/18018 [10:30:00<26:50:37, 7.33s/step, loss=1.572, acc=0.655, acc_len=1.54, thru=16.6, I=19.4, T=31.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:14,550] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:23,565] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:00:25,277] loop.py:45 INFO TRAIN_STEP step=4838 loss=1.554836 Training: 27%|██▋ | 4838/18018 [10:30:11<30:36:07, 8.36s/step, loss=1.555, acc=0.663, acc_len=1.57, thru=11.7, I=14.3, T=31.9, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:25,574] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:31,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:00:32,118] loop.py:45 INFO TRAIN_STEP step=4839 loss=1.399912 Training: 27%|██▋ | 4839/18018 [10:30:18<28:55:57, 7.90s/step, loss=1.400, acc=0.688, acc_len=1.65, thru=17.5, I=23.4, T=35.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:32,531] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:00:38,523] loop.py:380 INFO TIMING step=4840: step=3.682s data=1.563s compute=3.648s [fwd=2.554s bwd=1.079s opt=0.003s] dispatch=2.712s [2026-10-19 00:00:38,523] loop.py:45 INFO TRAIN_STEP step=4840 loss=1.627173 Training: 27%|██▋ | 4840/18018 [10:30:24<27:17:03, 7.45s/step, loss=1.627, acc=0.647, acc_len=1.52, thru=17.3, I=21.8, T=34.8, wait=2.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:45,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:00:47,839] loop.py:45 INFO TRAIN_STEP step=4841 loss=1.579198 Training: 27%|██▋ | 4841/18018 [10:30:33<29:19:38, 8.01s/step, loss=1.579, acc=0.654, acc_len=1.54, thru=16.1, I=16.3, T=29.4, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:00:47,833] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4840: forward=3030.8ms backward=1259.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:48,005] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:53,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:00:54,701] loop.py:45 INFO TRAIN_STEP step=4842 loss=1.482033 Training: 27%|██▋ | 4842/18018 [10:30:40<28:03:39, 7.67s/step, loss=1.482, acc=0.670, acc_len=1.59, thru=16.4, I=22.8, T=28.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:00:54,924] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:01:00,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:01:01,103] loop.py:45 INFO TRAIN_STEP step=4843 loss=1.512985 Training: 27%|██▋ | 4843/18018 [10:30:47<26:40:09, 7.29s/step, loss=1.513, acc=0.667, acc_len=1.57, thru=18.3, I=21.3, T=36.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:01:01,554] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:01:09,396] loop.py:45 INFO TRAIN_STEP step=4844 loss=1.562744 Training: 27%|██▋ | 4844/18018 [10:30:55<27:46:20, 7.59s/step, loss=1.563, acc=0.659, acc_len=1.55, thru=14.3, I=16.2, T=34.3, wait=4.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:01:16,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:01:17,872] loop.py:380 INFO TIMING step=4845: step=4.499s data=2.006s compute=4.462s [fwd=3.079s bwd=1.369s opt=0.003s] dispatch=3.967s [2026-10-19 00:01:17,872] loop.py:45 INFO TRAIN_STEP step=4845 loss=1.505070 Training: 27%|██▋ | 4845/18018 [10:31:03<28:44:41, 7.86s/step, loss=1.505, acc=0.669, acc_len=1.59, thru=16.6, I=18.7, T=28.5, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:01:17,962] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:01:25,136] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:01:26,036] loop.py:45 INFO TRAIN_STEP step=4846 loss=1.672271 Training: 27%|██▋ | 4846/18018 [10:31:12<29:04:57, 7.95s/step, loss=1.672, acc=0.642, acc_len=1.49, thru=14.2, I=19.7, T=28.5, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:01:26,031] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4845: forward=3204.7ms backward=1236.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:01:26,155] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 00:01:36,973] loop.py:45 INFO TRAIN_STEP step=4847 loss=1.700165 Training: 27%|██▋ | 4847/18018 [10:31:23<32:21:31, 8.84s/step, loss=1.700, acc=0.636, acc_len=1.48, thru=11.2, I=11.7, T=31.4, wait=6.8s, pool=56, epoch=1/2][2026-10-19 00:01:44,463] loop.py:45 INFO TRAIN_STEP step=4848 loss=1.645733 Training: 27%|██▋ | 4848/18018 [10:31:30<30:52:13, 8.44s/step, loss=1.646, acc=0.645, acc_len=1.50, thru=15.2, I=15.3, T=32.9, wait=3.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:01:55,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:01:56,566] loop.py:45 INFO TRAIN_STEP step=4849 loss=1.554620 Training: 27%|██▋ | 4849/18018 [10:31:42<34:53:22, 9.54s/step, loss=1.555, acc=0.662, acc_len=1.56, thru=14.1, I=13.1, T=30.0, wait=7.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:01:56,888] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:02:04,236] loop.py:380 INFO TIMING step=4850: step=3.918s data=1.731s compute=3.882s [fwd=2.772s bwd=1.096s opt=0.003s] dispatch=3.742s [2026-10-19 00:02:04,237] loop.py:45 INFO TRAIN_STEP step=4850 loss=1.556110 Training: 27%|██▋ | 4850/18018 [10:31:50<32:50:10, 8.98s/step, loss=1.556, acc=0.660, acc_len=1.56, thru=14.1, I=15.2, T=32.7, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:02:12,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:02:13,416] loop.py:45 INFO TRAIN_STEP step=4851 loss=1.563106 Training: 27%|██▋ | 4851/18018 [10:31:59<33:03:27, 9.04s/step, loss=1.563, acc=0.660, acc_len=1.56, thru=15.3, I=17.4, T=31.6, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:02:13,411] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4850: forward=2807.9ms backward=1190.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:02:13,762] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:02:21,222] loop.py:45 INFO TRAIN_STEP step=4852 loss=1.608700 Training: 27%|██▋ | 4852/18018 [10:32:07<31:42:05, 8.67s/step, loss=1.609, acc=0.653, acc_len=1.53, thru=13.1, I=14.8, T=30.1, wait=3.5s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:02:29,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:02:32,160] loop.py:45 INFO TRAIN_STEP step=4853 loss=1.715882 Training: 27%|██▋ | 4853/18018 [10:32:18<34:11:27, 9.35s/step, loss=1.716, acc=0.637, acc_len=1.48, thru=15.3, I=19.0, T=25.9, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:02:32,215] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 [2026-10-19 00:02:38,789] loop.py:45 INFO TRAIN_STEP step=4854 loss=1.467325 Training: 27%|██▋ | 4854/18018 [10:32:24<31:12:04, 8.53s/step, loss=1.467, acc=0.672, acc_len=1.60, thru=12.8, I=21.7, T=36.1, wait=3.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:02:45,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:02:47,889] loop.py:380 INFO TIMING step=4855: step=3.879s data=1.680s compute=3.854s [fwd=2.728s bwd=1.112s opt=0.003s] dispatch=5.213s [2026-10-19 00:02:47,890] loop.py:45 INFO TRAIN_STEP step=4855 loss=1.598358 Training: 27%|██▋ | 4855/18018 [10:32:33<31:49:19, 8.70s/step, loss=1.598, acc=0.650, acc_len=1.53, thru=16.7, I=17.1, T=33.0, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:02:48,244] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:02:52,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:02:54,529] loop.py:45 INFO TRAIN_STEP step=4856 loss=1.447625 Training: 27%|██▋ | 4856/18018 [10:32:40<29:33:20, 8.08s/step, loss=1.448, acc=0.679, acc_len=1.61, thru=18.0, I=28.3, T=29.6, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:02:54,524] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4855: forward=2968.5ms backward=1309.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:02:54,934] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:02:59,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:03:00,504] loop.py:45 INFO TRAIN_STEP step=4857 loss=1.523080 Training: 27%|██▋ | 4857/18018 [10:32:46<27:14:22, 7.45s/step, loss=1.523, acc=0.667, acc_len=1.58, thru=19.1, I=27.2, T=37.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:01,006] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:05,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:03:06,786] loop.py:45 INFO TRAIN_STEP step=4858 loss=1.490757 Training: 27%|██▋ | 4858/18018 [10:32:52<25:57:31, 7.10s/step, loss=1.491, acc=0.669, acc_len=1.59, thru=20.4, I=27.8, T=32.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:07,126] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:03:14,780] loop.py:45 INFO TRAIN_STEP step=4859 loss=1.594274 Training: 27%|██▋ | 4859/18018 [10:33:00<26:56:04, 7.37s/step, loss=1.594, acc=0.647, acc_len=1.52, thru=14.4, I=19.3, T=31.9, wait=4.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:19,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:03:21,701] loop.py:380 INFO TIMING step=4860: step=4.032s data=1.798s compute=3.982s [fwd=2.675s bwd=1.292s opt=0.003s] dispatch=2.879s [2026-10-19 00:03:21,701] loop.py:45 INFO TRAIN_STEP step=4860 loss=1.426561 Training: 27%|██▋ | 4860/18018 [10:33:07<26:26:29, 7.23s/step, loss=1.427, acc=0.677, acc_len=1.61, thru=17.0, I=20.9, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:21,795] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:28,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:03:28,760] loop.py:45 INFO TRAIN_STEP step=4861 loss=1.472813 Training: 27%|██▋ | 4861/18018 [10:33:14<26:14:55, 7.18s/step, loss=1.473, acc=0.672, acc_len=1.60, thru=16.5, I=22.5, T=30.7, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:03:28,755] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4860: forward=2844.6ms backward=1266.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:29,174] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:03:35,256] loop.py:45 INFO TRAIN_STEP step=4862 loss=1.629773 Training: 27%|██▋ | 4862/18018 [10:33:21<25:29:33, 6.98s/step, loss=1.630, acc=0.648, acc_len=1.51, thru=15.3, I=17.2, T=34.5, wait=2.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:42,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:03:44,969] loop.py:45 INFO TRAIN_STEP step=4863 loss=1.613751 Training: 27%|██▋ | 4863/18018 [10:33:31<28:29:27, 7.80s/step, loss=1.614, acc=0.647, acc_len=1.52, thru=16.0, I=18.0, T=30.4, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:45,367] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:51,395] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:03:52,732] loop.py:45 INFO TRAIN_STEP step=4864 loss=1.532615 Training: 27%|██▋ | 4864/18018 [10:33:38<28:27:15, 7.79s/step, loss=1.533, acc=0.663, acc_len=1.57, thru=13.3, I=21.8, T=31.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:03:52,905] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:03:58,574] loop.py:380 INFO TIMING step=4865: step=3.492s data=1.511s compute=3.465s [fwd=2.517s bwd=0.933s opt=0.003s] dispatch=2.339s [2026-10-19 00:03:58,574] loop.py:45 INFO TRAIN_STEP step=4865 loss=1.431365 Training: 27%|██▋ | 4865/18018 [10:33:44<26:19:08, 7.20s/step, loss=1.431, acc=0.683, acc_len=1.63, thru=16.9, I=20.2, T=36.7, wait=2.3s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:06,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:04:07,303] loop.py:45 INFO TRAIN_STEP step=4866 loss=1.444845 Training: 27%|██▋ | 4866/18018 [10:33:53<27:59:10, 7.66s/step, loss=1.445, acc=0.673, acc_len=1.60, thru=17.7, I=17.4, T=32.9, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:04:07,298] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4865: forward=2756.9ms backward=1076.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:07,762] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:15,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:04:15,555] loop.py:45 INFO TRAIN_STEP step=4867 loss=1.520750 Training: 27%|██▋ | 4867/18018 [10:34:01<28:38:06, 7.84s/step, loss=1.521, acc=0.665, acc_len=1.57, thru=15.2, I=17.8, T=32.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:15,873] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:21,587] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:04:22,551] loop.py:45 INFO TRAIN_STEP step=4868 loss=1.597432 Training: 27%|██▋ | 4868/18018 [10:34:08<27:42:27, 7.59s/step, loss=1.597, acc=0.652, acc_len=1.53, thru=18.6, I=18.6, T=29.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:22,600] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:28,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:04:30,281] loop.py:45 INFO TRAIN_STEP step=4869 loss=1.583479 Training: 27%|██▋ | 4869/18018 [10:34:16<27:51:48, 7.63s/step, loss=1.583, acc=0.655, acc_len=1.54, thru=15.8, I=19.9, T=27.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:30,743] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 00:04:39,064] loop.py:380 INFO TIMING step=4870: step=4.168s data=1.934s compute=4.142s [fwd=2.898s bwd=1.229s opt=0.003s] dispatch=4.607s [2026-10-19 00:04:39,064] loop.py:45 INFO TRAIN_STEP step=4870 loss=1.732903 Training: 27%|██▋ | 4870/18018 [10:34:25<29:07:33, 7.97s/step, loss=1.733, acc=0.631, acc_len=1.46, thru=15.4, I=15.0, T=30.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:39,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:39,562] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:45,265] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:04:46,491] loop.py:45 INFO TRAIN_STEP step=4871 loss=1.575330 Training: 27%|██▋ | 4871/18018 [10:34:32<28:31:28, 7.81s/step, loss=1.575, acc=0.657, acc_len=1.56, thru=16.6, I=19.7, T=28.6, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:04:46,487] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4870: forward=3056.4ms backward=1340.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:46,775] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:53,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:04:53,341] loop.py:45 INFO TRAIN_STEP step=4872 loss=1.507993 Training: 27%|██▋ | 4872/18018 [10:34:39<27:28:12, 7.52s/step, loss=1.508, acc=0.670, acc_len=1.60, thru=17.1, I=21.8, T=33.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:53,577] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:04:59,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:05:01,249] loop.py:45 INFO TRAIN_STEP step=4873 loss=1.563932 Training: 27%|██▋ | 4873/18018 [10:34:47<27:53:31, 7.64s/step, loss=1.564, acc=0.658, acc_len=1.55, thru=15.8, I=18.1, T=29.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:05:01,371] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-19 00:05:08,749] loop.py:45 INFO TRAIN_STEP step=4874 loss=1.566355 Training: 27%|██▋ | 4874/18018 [10:34:54<27:44:04, 7.60s/step, loss=1.566, acc=0.657, acc_len=1.55, thru=13.8, I=15.8, T=31.2, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:05:14,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:05:15,327] loop.py:380 INFO TIMING step=4875: step=4.054s data=1.849s compute=3.980s [fwd=2.748s bwd=1.217s opt=0.003s] dispatch=2.517s [2026-10-19 00:05:15,327] loop.py:45 INFO TRAIN_STEP step=4875 loss=1.593019 Training: 27%|██▋ | 4875/18018 [10:35:01<26:37:07, 7.29s/step, loss=1.593, acc=0.651, acc_len=1.53, thru=19.3, I=20.1, T=31.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:05:15,398] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:05:22,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:05:24,136] loop.py:45 INFO TRAIN_STEP step=4876 loss=1.537679 Training: 27%|██▋ | 4876/18018 [10:35:10<28:16:45, 7.75s/step, loss=1.538, acc=0.665, acc_len=1.57, thru=14.1, I=18.3, T=27.3, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:05:24,131] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4875: forward=3196.2ms backward=1433.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:05:24,260] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:05:29,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:05:32,134] loop.py:45 INFO TRAIN_STEP step=4877 loss=1.726952 Training: 27%|██▋ | 4877/18018 [10:35:18<28:33:10, 7.82s/step, loss=1.727, acc=0.632, acc_len=1.46, thru=13.3, I=20.3, T=22.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:05:32,278] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=16/64 [2026-10-19 00:05:38,788] loop.py:45 INFO TRAIN_STEP step=4878 loss=1.611659 Training: 27%|██▋ | 4878/18018 [10:35:24<27:16:20, 7.47s/step, loss=1.612, acc=0.649, acc_len=1.53, thru=13.7, I=16.4, T=29.4, wait=2.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:05:44,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:05:46,060] loop.py:45 INFO TRAIN_STEP step=4879 loss=1.476294 Training: 27%|██▋ | 4879/18018 [10:35:32<27:02:58, 7.41s/step, loss=1.476, acc=0.673, acc_len=1.61, thru=16.5, I=18.9, T=31.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:05:46,191] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-19 00:05:57,182] loop.py:380 INFO TIMING step=4880: step=4.531s data=2.096s compute=4.461s [fwd=3.048s bwd=1.397s opt=0.003s] dispatch=6.581s [2026-10-19 00:05:57,183] loop.py:45 INFO TRAIN_STEP step=4880 loss=1.544353 Training: 27%|██▋ | 4880/18018 [10:35:43<31:08:06, 8.53s/step, loss=1.544, acc=0.661, acc_len=1.56, thru=11.7, I=10.0, T=28.2, wait=6.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:03,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:06:05,318] loop.py:45 INFO TRAIN_STEP step=4881 loss=1.569797 Training: 27%|██▋ | 4881/18018 [10:35:51<30:40:36, 8.41s/step, loss=1.570, acc=0.661, acc_len=1.56, thru=17.5, I=18.2, T=28.7, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:06:05,313] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4880: forward=3138.9ms backward=1271.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:05,377] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:10,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:06:11,437] loop.py:45 INFO TRAIN_STEP step=4882 loss=1.566357 Training: 27%|██▋ | 4882/18018 [10:35:57<28:10:11, 7.72s/step, loss=1.566, acc=0.654, acc_len=1.54, thru=18.2, I=25.2, T=33.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:11,746] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:06:18,314] loop.py:45 INFO TRAIN_STEP step=4883 loss=1.595346 Training: 27%|██▋ | 4883/18018 [10:36:04<27:14:35, 7.47s/step, loss=1.595, acc=0.652, acc_len=1.52, thru=16.8, I=19.7, T=33.7, wait=3.1s, pool=52, epoch=1/2][2026-10-19 00:06:25,896] loop.py:45 INFO TRAIN_STEP step=4884 loss=1.571079 Training: 27%|██▋ | 4884/18018 [10:36:11<27:22:26, 7.50s/step, loss=1.571, acc=0.656, acc_len=1.54, thru=16.9, I=16.5, T=34.7, wait=3.9s, pool=52, epoch=1/2][2026-10-19 00:06:34,043] loop.py:380 INFO TIMING step=4885: step=4.510s data=2.181s compute=4.454s [fwd=3.074s bwd=1.364s opt=0.003s] dispatch=3.621s [2026-10-19 00:06:34,043] loop.py:45 INFO TRAIN_STEP step=4885 loss=1.562217 Training: 27%|██▋ | 4885/18018 [10:36:20<28:04:10, 7.69s/step, loss=1.562, acc=0.659, acc_len=1.55, thru=13.3, I=13.8, T=28.4, wait=3.6s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:42,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:06:43,944] loop.py:45 INFO TRAIN_STEP step=4886 loss=1.468521 Training: 27%|██▋ | 4886/18018 [10:36:30<30:29:00, 8.36s/step, loss=1.469, acc=0.672, acc_len=1.60, thru=16.8, I=18.0, T=30.6, wait=5.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:06:43,939] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4885: forward=2973.5ms backward=1173.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:44,360] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:50,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:06:51,276] loop.py:45 INFO TRAIN_STEP step=4887 loss=1.665599 Training: 27%|██▋ | 4887/18018 [10:36:37<29:21:28, 8.05s/step, loss=1.666, acc=0.647, acc_len=1.51, thru=16.7, I=21.3, T=30.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:51,685] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:56,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:06:59,059] loop.py:45 INFO TRAIN_STEP step=4888 loss=1.618744 Training: 27%|██▋ | 4888/18018 [10:36:45<29:04:05, 7.97s/step, loss=1.619, acc=0.650, acc_len=1.53, thru=15.6, I=20.7, T=27.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:06:59,321] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:04,122] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:07:06,148] loop.py:45 INFO TRAIN_STEP step=4889 loss=1.554464 Training: 27%|██▋ | 4889/18018 [10:36:52<28:05:57, 7.70s/step, loss=1.554, acc=0.655, acc_len=1.55, thru=15.3, I=26.5, T=31.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:06,634] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:12,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:07:13,108] loop.py:380 INFO TIMING step=4890: step=3.539s data=1.478s compute=3.513s [fwd=2.518s bwd=0.979s opt=0.003s] dispatch=3.413s [2026-10-19 00:07:13,108] loop.py:45 INFO TRAIN_STEP step=4890 loss=1.461904 Training: 27%|██▋ | 4890/18018 [10:36:59<27:17:04, 7.48s/step, loss=1.462, acc=0.676, acc_len=1.61, thru=16.0, I=23.5, T=36.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:13,526] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:19,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:07:20,145] loop.py:45 INFO TRAIN_STEP step=4891 loss=1.611783 (TrainerActor pid=9310) [2026-10-19 00:07:20,139] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4890: forward=3013.9ms backward=1234.3ms Training: 27%|██▋ | 4891/18018 [10:37:06<26:47:44, 7.35s/step, loss=1.612, acc=0.653, acc_len=1.53, thru=17.8, I=21.5, T=29.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:20,224] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:26,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:07:27,218] loop.py:45 INFO TRAIN_STEP step=4892 loss=1.584507 Training: 27%|██▋ | 4892/18018 [10:37:13<26:29:25, 7.27s/step, loss=1.585, acc=0.655, acc_len=1.55, thru=16.5, I=19.6, T=31.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:27,438] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:31,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:07:33,861] loop.py:45 INFO TRAIN_STEP step=4893 loss=1.707421 Training: 27%|██▋ | 4893/18018 [10:37:19<25:48:33, 7.08s/step, loss=1.707, acc=0.638, acc_len=1.49, thru=18.1, I=24.2, T=28.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:34,206] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:44,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:07:44,845] loop.py:45 INFO TRAIN_STEP step=4894 loss=1.589384 Training: 27%|██▋ | 4894/18018 [10:37:30<30:04:30, 8.25s/step, loss=1.589, acc=0.656, acc_len=1.55, thru=11.0, I=11.5, T=29.2, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:45,261] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:50,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:07:50,930] loop.py:380 INFO TIMING step=4895: step=3.845s data=1.636s compute=3.805s [fwd=2.615s bwd=1.176s opt=0.003s] dispatch=2.232s [2026-10-19 00:07:50,931] loop.py:45 INFO TRAIN_STEP step=4895 loss=1.513045 Training: 27%|██▋ | 4895/18018 [10:37:37<27:42:29, 7.60s/step, loss=1.513, acc=0.670, acc_len=1.59, thru=18.9, I=20.3, T=33.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:51,157] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:55,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:07:56,707] loop.py:45 INFO TRAIN_STEP step=4896 loss=1.542862 Training: 27%|██▋ | 4896/18018 [10:37:42<25:42:38, 7.05s/step, loss=1.543, acc=0.664, acc_len=1.57, thru=20.5, I=24.5, T=35.3, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:07:56,702] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4895: forward=2571.1ms backward=1004.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:07:57,021] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:08:04,808] loop.py:45 INFO TRAIN_STEP step=4897 loss=1.563587 Training: 27%|██▋ | 4897/18018 [10:37:50<26:51:05, 7.37s/step, loss=1.564, acc=0.662, acc_len=1.57, thru=14.2, I=18.9, T=31.9, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:11,402] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:08:12,813] loop.py:45 INFO TRAIN_STEP step=4898 loss=1.555266 Training: 27%|██▋ | 4898/18018 [10:37:58<27:33:01, 7.56s/step, loss=1.555, acc=0.661, acc_len=1.56, thru=15.6, I=17.5, T=27.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:12,917] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:19,555] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:08:20,099] loop.py:45 INFO TRAIN_STEP step=4899 loss=1.538389 Training: 27%|██▋ | 4899/18018 [10:38:06<27:14:52, 7.48s/step, loss=1.538, acc=0.661, acc_len=1.57, thru=16.1, I=19.8, T=32.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:20,558] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:08:26,973] loop.py:380 INFO TIMING step=4900: step=4.117s data=1.843s compute=4.075s [fwd=2.850s bwd=1.211s opt=0.003s] dispatch=2.747s [2026-10-19 00:08:26,974] loop.py:45 INFO TRAIN_STEP step=4900 loss=1.585376 Training: 27%|██▋ | 4900/18018 [10:38:13<26:35:20, 7.30s/step, loss=1.585, acc=0.656, acc_len=1.54, thru=18.3, I=19.7, T=31.1, wait=2.7s, pool=60, epoch=1/2][2026-10-19 00:08:34,563] loop.py:45 INFO TRAIN_STEP step=4901 loss=1.593214 Training: 27%|██▋ | 4901/18018 [10:38:20<26:54:10, 7.38s/step, loss=1.593, acc=0.652, acc_len=1.54, thru=16.6, I=15.7, T=29.9, wait=3.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:08:34,559] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4900: forward=2931.9ms backward=1282.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:42,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:08:42,648] loop.py:45 INFO TRAIN_STEP step=4902 loss=1.645962 Training: 27%|██▋ | 4902/18018 [10:38:28<27:40:04, 7.59s/step, loss=1.646, acc=0.648, acc_len=1.51, thru=16.4, I=16.0, T=26.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:43,049] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:48,373] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:08:49,531] loop.py:45 INFO TRAIN_STEP step=4903 loss=1.501090 Training: 27%|██▋ | 4903/18018 [10:38:35<26:53:22, 7.38s/step, loss=1.501, acc=0.664, acc_len=1.57, thru=16.8, I=18.6, T=29.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:49,881] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:55,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:08:57,117] loop.py:45 INFO TRAIN_STEP step=4904 loss=1.591793 Training: 27%|██▋ | 4904/18018 [10:38:43<27:06:41, 7.44s/step, loss=1.592, acc=0.653, acc_len=1.53, thru=16.1, I=20.7, T=29.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:08:57,481] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 00:09:03,936] loop.py:380 INFO TIMING step=4905: step=3.736s data=1.694s compute=3.704s [fwd=2.613s bwd=1.075s opt=0.003s] dispatch=3.075s [2026-10-19 00:09:03,936] loop.py:45 INFO TRAIN_STEP step=4905 loss=1.609656 Training: 27%|██▋ | 4905/18018 [10:38:50<26:25:37, 7.26s/step, loss=1.610, acc=0.646, acc_len=1.51, thru=15.4, I=17.6, T=34.3, wait=3.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:09,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:09:10,722] loop.py:45 INFO TRAIN_STEP step=4906 loss=1.602696 Training: 27%|██▋ | 4906/18018 [10:38:56<25:54:53, 7.12s/step, loss=1.603, acc=0.656, acc_len=1.54, thru=20.4, I=18.4, T=33.4, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:09:10,717] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4905: forward=2715.5ms backward=1075.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:11,001] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:17,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:09:18,138] loop.py:45 INFO TRAIN_STEP step=4907 loss=1.510871 Training: 27%|██▋ | 4907/18018 [10:39:04<26:14:26, 7.21s/step, loss=1.511, acc=0.667, acc_len=1.58, thru=17.6, I=19.6, T=33.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:18,397] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:24,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:09:26,306] loop.py:45 INFO TRAIN_STEP step=4908 loss=1.499497 Training: 27%|██▋ | 4908/18018 [10:39:12<27:17:25, 7.49s/step, loss=1.499, acc=0.670, acc_len=1.59, thru=16.1, I=18.3, T=29.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:26,445] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:33,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:09:34,924] loop.py:45 INFO TRAIN_STEP step=4909 loss=1.569422 Training: 27%|██▋ | 4909/18018 [10:39:21<28:30:54, 7.83s/step, loss=1.569, acc=0.659, acc_len=1.55, thru=13.2, I=15.6, T=23.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:35,401] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:43,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:09:43,671] loop.py:380 INFO TIMING step=4910: step=4.861s data=2.479s compute=4.825s [fwd=3.568s bwd=1.243s opt=0.003s] dispatch=3.877s [2026-10-19 00:09:43,671] loop.py:45 INFO TRAIN_STEP step=4910 loss=1.472820 Training: 27%|██▋ | 4910/18018 [10:39:29<29:30:54, 8.11s/step, loss=1.473, acc=0.675, acc_len=1.61, thru=13.7, I=15.4, T=26.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:44,021] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:09:51,128] loop.py:45 INFO TRAIN_STEP step=4911 loss=1.514272 Training: 27%|██▋ | 4911/18018 [10:39:37<28:48:07, 7.91s/step, loss=1.514, acc=0.667, acc_len=1.58, thru=14.5, I=15.5, T=32.3, wait=3.5s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:09:51,123] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4910: forward=2685.8ms backward=1212.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:58,213] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:09:58,657] loop.py:45 INFO TRAIN_STEP step=4912 loss=1.513432 Training: 27%|██▋ | 4912/18018 [10:39:44<28:23:07, 7.80s/step, loss=1.513, acc=0.661, acc_len=1.56, thru=18.4, I=16.2, T=29.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:09:58,731] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:05,856] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:10:08,779] loop.py:45 INFO TRAIN_STEP step=4913 loss=1.626678 Training: 27%|██▋ | 4913/18018 [10:39:54<30:55:15, 8.49s/step, loss=1.627, acc=0.645, acc_len=1.52, thru=12.6, I=17.3, T=25.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:08,877] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:15,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:10:16,262] loop.py:45 INFO TRAIN_STEP step=4914 loss=1.614171 Training: 27%|██▋ | 4914/18018 [10:40:02<29:48:51, 8.19s/step, loss=1.614, acc=0.649, acc_len=1.52, thru=17.4, I=20.3, T=28.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:16,582] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:24,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:10:24,947] loop.py:380 INFO TIMING step=4915: step=4.665s data=2.090s compute=4.623s [fwd=3.168s bwd=1.441s opt=0.003s] dispatch=4.011s [2026-10-19 00:10:24,948] loop.py:45 INFO TRAIN_STEP step=4915 loss=1.515257 Training: 27%|██▋ | 4915/18018 [10:40:11<30:21:04, 8.34s/step, loss=1.515, acc=0.665, acc_len=1.58, thru=13.4, I=15.7, T=27.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:25,388] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (TrainerActor pid=9310) [2026-10-19 00:10:34,991] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4915: forward=2954.4ms backward=1202.4ms [2026-10-19 00:10:34,995] loop.py:45 INFO TRAIN_STEP step=4916 loss=1.617550 Training: 27%|██▋ | 4916/18018 [10:40:21<32:12:57, 8.85s/step, loss=1.618, acc=0.654, acc_len=1.54, thru=12.5, I=11.9, T=30.4, wait=5.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:41,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:10:42,341] loop.py:45 INFO TRAIN_STEP step=4917 loss=1.505781 Training: 27%|██▋ | 4917/18018 [10:40:28<30:34:08, 8.40s/step, loss=1.506, acc=0.667, acc_len=1.58, thru=18.7, I=16.4, T=28.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:42,408] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:49,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:10:50,997] loop.py:45 INFO TRAIN_STEP step=4918 loss=1.502215 Training: 27%|██▋ | 4918/18018 [10:40:37<30:50:52, 8.48s/step, loss=1.502, acc=0.669, acc_len=1.59, thru=14.1, I=16.7, T=30.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:51,160] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:55,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:10:56,967] loop.py:45 INFO TRAIN_STEP step=4919 loss=1.420236 Training: 27%|██▋ | 4919/18018 [10:40:43<28:06:22, 7.72s/step, loss=1.420, acc=0.683, acc_len=1.64, thru=19.3, I=22.6, T=32.4, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:10:57,172] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-19 00:11:04,416] loop.py:380 INFO TIMING step=4920: step=4.938s data=2.474s compute=4.897s [fwd=3.371s bwd=1.512s opt=0.003s] dispatch=2.502s [2026-10-19 00:11:04,416] loop.py:45 INFO TRAIN_STEP step=4920 loss=1.624310 Training: 27%|██▋ | 4920/18018 [10:40:50<27:48:12, 7.64s/step, loss=1.624, acc=0.651, acc_len=1.52, thru=16.0, I=18.2, T=25.9, wait=2.5s, pool=60, epoch=1/2][2026-10-19 00:11:11,914] loop.py:45 INFO TRAIN_STEP step=4921 loss=1.651701 Training: 27%|██▋ | 4921/18018 [10:40:57<27:38:36, 7.60s/step, loss=1.652, acc=0.643, acc_len=1.51, thru=16.5, I=15.1, T=29.4, wait=3.1s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:11:11,909] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4920: forward=3017.2ms backward=1292.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:11:20,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:11:20,204] loop.py:45 INFO TRAIN_STEP step=4922 loss=1.600885 Training: 27%|██▋ | 4922/18018 [10:41:06<28:24:01, 7.81s/step, loss=1.601, acc=0.649, acc_len=1.52, thru=16.8, I=15.5, T=32.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:11:20,566] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:11:27,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:11:28,546] loop.py:45 INFO TRAIN_STEP step=4923 loss=1.587371 Training: 27%|██▋ | 4923/18018 [10:41:14<28:58:49, 7.97s/step, loss=1.587, acc=0.651, acc_len=1.53, thru=16.3, I=17.2, T=29.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:11:28,884] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:11:35,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:11:35,854] loop.py:45 INFO TRAIN_STEP step=4924 loss=1.537199 Training: 27%|██▋ | 4924/18018 [10:41:21<28:15:33, 7.77s/step, loss=1.537, acc=0.662, acc_len=1.56, thru=16.1, I=19.0, T=33.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:11:35,906] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-19 00:11:45,212] loop.py:380 INFO TIMING step=4925: step=4.179s data=1.656s compute=4.153s [fwd=2.897s bwd=1.241s opt=0.003s] dispatch=5.168s [2026-10-19 00:11:45,213] loop.py:45 INFO TRAIN_STEP step=4925 loss=1.601091 Training: 27%|██▋ | 4925/18018 [10:41:31<29:59:31, 8.25s/step, loss=1.601, acc=0.656, acc_len=1.54, thru=11.7, I=14.7, T=30.6, wait=5.2s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:11:55,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:11:56,957] loop.py:45 INFO TRAIN_STEP step=4926 loss=1.533350 Training: 27%|██▋ | 4926/18018 [10:41:43<33:48:10, 9.29s/step, loss=1.533, acc=0.661, acc_len=1.56, thru=14.1, I=13.5, T=30.3, wait=7.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:11:56,952] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4925: forward=2949.5ms backward=1222.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:11:56,988] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-19 00:12:02,961] loop.py:45 INFO TRAIN_STEP step=4927 loss=1.601179 Training: 27%|██▋ | 4927/18018 [10:41:49<30:12:42, 8.31s/step, loss=1.601, acc=0.650, acc_len=1.52, thru=17.6, I=22.6, T=32.6, wait=2.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:10,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:12:11,920] loop.py:45 INFO TRAIN_STEP step=4928 loss=1.489966 Training: 27%|██▋ | 4928/18018 [10:41:58<30:55:05, 8.50s/step, loss=1.490, acc=0.669, acc_len=1.59, thru=15.9, I=17.4, T=32.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:12,290] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:12:18,596] loop.py:45 INFO TRAIN_STEP step=4929 loss=1.528259 Training: 27%|██▋ | 4929/18018 [10:42:04<28:55:33, 7.96s/step, loss=1.528, acc=0.661, acc_len=1.56, thru=17.1, I=22.8, T=29.1, wait=2.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:26,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:12:27,755] loop.py:380 INFO TIMING step=4930: step=4.177s data=1.908s compute=4.143s [fwd=2.969s bwd=1.159s opt=0.003s] dispatch=4.969s [2026-10-19 00:12:27,755] loop.py:45 INFO TRAIN_STEP step=4930 loss=1.613628 Training: 27%|██▋ | 4930/18018 [10:42:13<30:13:57, 8.32s/step, loss=1.614, acc=0.647, acc_len=1.52, thru=15.5, I=15.2, T=30.6, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:27,878] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:36,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:12:37,195] loop.py:45 INFO TRAIN_STEP step=4931 loss=1.479522 Training: 27%|██▋ | 4931/18018 [10:42:23<31:27:26, 8.65s/step, loss=1.480, acc=0.677, acc_len=1.61, thru=13.8, I=13.9, T=26.5, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:12:37,190] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4930: forward=3333.8ms backward=1455.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:37,404] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:45,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:12:46,442] loop.py:45 INFO TRAIN_STEP step=4932 loss=1.572041 Training: 27%|██▋ | 4932/18018 [10:42:32<32:06:11, 8.83s/step, loss=1.572, acc=0.654, acc_len=1.53, thru=14.7, I=15.2, T=28.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:46,819] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:53,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:12:54,333] loop.py:45 INFO TRAIN_STEP step=4933 loss=1.676786 Training: 27%|██▋ | 4933/18018 [10:42:40<31:04:23, 8.55s/step, loss=1.677, acc=0.643, acc_len=1.50, thru=15.6, I=19.5, T=31.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:54,568] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:12:58,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:13:00,026] loop.py:45 INFO TRAIN_STEP step=4934 loss=1.469476 Training: 27%|██▋ | 4934/18018 [10:42:46<27:57:29, 7.69s/step, loss=1.469, acc=0.671, acc_len=1.59, thru=19.9, I=22.7, T=32.4, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:00,470] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:06,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:13:07,813] loop.py:380 INFO TIMING step=4935: step=3.858s data=1.738s compute=3.816s [fwd=2.678s bwd=1.123s opt=0.003s] dispatch=3.918s [2026-10-19 00:13:07,813] loop.py:45 INFO TRAIN_STEP step=4935 loss=1.638669 Training: 27%|██▋ | 4935/18018 [10:42:53<28:03:33, 7.72s/step, loss=1.639, acc=0.646, acc_len=1.52, thru=15.1, I=19.2, T=33.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:08,135] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:13:13,718] loop.py:45 INFO TRAIN_STEP step=4936 loss=1.400125 Training: 27%|██▋ | 4936/18018 [10:42:59<26:04:36, 7.18s/step, loss=1.400, acc=0.687, acc_len=1.64, thru=19.7, I=20.8, T=34.4, wait=2.2s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:13:13,713] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4935: forward=2541.3ms backward=1135.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:18,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:13:20,599] loop.py:45 INFO TRAIN_STEP step=4937 loss=1.489639 Training: 27%|██▋ | 4937/18018 [10:43:06<25:45:08, 7.09s/step, loss=1.490, acc=0.671, acc_len=1.59, thru=18.1, I=21.1, T=30.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:20,964] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:29,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:13:29,888] loop.py:45 INFO TRAIN_STEP step=4938 loss=1.562233 Training: 27%|██▋ | 4938/18018 [10:43:15<28:09:07, 7.75s/step, loss=1.562, acc=0.658, acc_len=1.55, thru=14.3, I=15.3, T=25.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:30,169] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:36,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:13:37,383] loop.py:45 INFO TRAIN_STEP step=4939 loss=1.527920 Training: 27%|██▋ | 4939/18018 [10:43:23<27:52:23, 7.67s/step, loss=1.528, acc=0.670, acc_len=1.59, thru=16.5, I=17.6, T=30.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:37,574] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:44,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:13:44,983] loop.py:380 INFO TIMING step=4940: step=4.020s data=1.864s compute=3.992s [fwd=2.835s bwd=1.143s opt=0.003s] dispatch=3.570s [2026-10-19 00:13:44,983] loop.py:45 INFO TRAIN_STEP step=4940 loss=1.548715 Training: 27%|██▋ | 4940/18018 [10:43:31<27:47:32, 7.65s/step, loss=1.549, acc=0.660, acc_len=1.55, thru=17.5, I=17.5, T=31.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:45,385] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:49,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:13:51,210] loop.py:45 INFO TRAIN_STEP step=4941 loss=1.548121 Training: 27%|██▋ | 4941/18018 [10:43:37<26:14:22, 7.22s/step, loss=1.548, acc=0.660, acc_len=1.56, thru=19.4, I=22.2, T=36.4, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:13:51,205] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4940: forward=2453.5ms backward=1004.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:51,294] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:58,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:13:58,511] loop.py:45 INFO TRAIN_STEP step=4942 loss=1.441893 Training: 27%|██▋ | 4942/18018 [10:43:44<26:19:13, 7.25s/step, loss=1.442, acc=0.676, acc_len=1.62, thru=16.9, I=19.7, T=34.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:13:58,648] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:06,947] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:14:08,410] loop.py:45 INFO TRAIN_STEP step=4943 loss=1.535575 Training: 27%|██▋ | 4943/18018 [10:43:54<29:12:36, 8.04s/step, loss=1.536, acc=0.668, acc_len=1.58, thru=13.1, I=15.6, T=25.8, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:08,464] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-19 00:14:16,866] loop.py:45 INFO TRAIN_STEP step=4944 loss=1.476300 Training: 27%|██▋ | 4944/18018 [10:44:02<29:39:19, 8.17s/step, loss=1.476, acc=0.670, acc_len=1.59, thru=12.9, I=14.9, T=29.8, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:22,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:14:24,486] loop.py:380 INFO TIMING step=4945: step=4.270s data=2.008s compute=4.238s [fwd=2.963s bwd=1.261s opt=0.003s] dispatch=3.342s [2026-10-19 00:14:24,486] loop.py:45 INFO TRAIN_STEP step=4945 loss=1.545816 Training: 27%|██▋ | 4945/18018 [10:44:10<29:03:31, 8.00s/step, loss=1.546, acc=0.659, acc_len=1.55, thru=17.2, I=18.0, T=30.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:24,945] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:30,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:14:31,765] loop.py:45 INFO TRAIN_STEP step=4946 loss=1.576259 Training: 27%|██▋ | 4946/18018 [10:44:17<28:16:18, 7.79s/step, loss=1.576, acc=0.651, acc_len=1.53, thru=15.6, I=20.9, T=28.5, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:14:31,760] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4945: forward=3070.6ms backward=1369.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:31,855] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:38,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:14:38,279] loop.py:45 INFO TRAIN_STEP step=4947 loss=1.510148 Training: 27%|██▋ | 4947/18018 [10:44:24<26:52:52, 7.40s/step, loss=1.510, acc=0.671, acc_len=1.59, thru=17.9, I=19.6, T=31.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:38,723] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:44,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:14:46,273] loop.py:45 INFO TRAIN_STEP step=4948 loss=1.585780 Training: 27%|██▋ | 4948/18018 [10:44:32<27:31:29, 7.58s/step, loss=1.586, acc=0.652, acc_len=1.53, thru=17.6, I=17.6, T=30.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:46,473] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:51,179] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:14:52,684] loop.py:45 INFO TRAIN_STEP step=4949 loss=1.528838 Training: 27%|██▋ | 4949/18018 [10:44:38<26:14:52, 7.23s/step, loss=1.529, acc=0.665, acc_len=1.57, thru=18.5, I=23.1, T=29.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:53,191] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:14:59,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:15:00,448] loop.py:380 INFO TIMING step=4950: step=4.851s data=2.365s compute=4.818s [fwd=3.448s bwd=1.355s opt=0.003s] dispatch=2.903s [2026-10-19 00:15:00,449] loop.py:45 INFO TRAIN_STEP step=4950 loss=1.669981 Training: 27%|██▋ | 4950/18018 [10:44:46<26:49:37, 7.39s/step, loss=1.670, acc=0.638, acc_len=1.49, thru=14.5, I=16.6, T=26.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:00,765] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:07,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:15:08,256] loop.py:45 INFO TRAIN_STEP step=4951 loss=1.502680 Training: 27%|██▋ | 4951/18018 [10:44:54<27:16:42, 7.52s/step, loss=1.503, acc=0.664, acc_len=1.58, thru=15.4, I=17.1, T=26.9, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:15:08,252] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4950: forward=3351.7ms backward=1360.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:08,559] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:13,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:15:15,447] loop.py:45 INFO TRAIN_STEP step=4952 loss=1.525431 Training: 27%|██▋ | 4952/18018 [10:45:01<26:55:21, 7.42s/step, loss=1.525, acc=0.665, acc_len=1.58, thru=16.0, I=19.6, T=28.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:15,666] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:20,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:15:21,308] loop.py:45 INFO TRAIN_STEP step=4953 loss=1.588156 Training: 27%|██▋ | 4953/18018 [10:45:07<25:13:35, 6.95s/step, loss=1.588, acc=0.653, acc_len=1.53, thru=18.4, I=25.2, T=34.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:21,512] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:27,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:15:28,103] loop.py:45 INFO TRAIN_STEP step=4954 loss=1.522356 Training: 27%|██▋ | 4954/18018 [10:45:14<25:03:21, 6.90s/step, loss=1.522, acc=0.666, acc_len=1.58, thru=18.5, I=21.6, T=29.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:28,199] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:35,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:15:36,115] loop.py:380 INFO TIMING step=4955: step=4.242s data=1.928s compute=4.188s [fwd=2.911s bwd=1.262s opt=0.003s] dispatch=3.759s [2026-10-19 00:15:36,115] loop.py:45 INFO TRAIN_STEP step=4955 loss=1.613863 Training: 28%|██▊ | 4955/18018 [10:45:22<26:15:27, 7.24s/step, loss=1.614, acc=0.654, acc_len=1.53, thru=14.5, I=16.8, T=30.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:15:36,258] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-19 00:15:42,455] loop.py:45 INFO TRAIN_STEP step=4956 loss=1.542452 Training: 28%|██▊ | 4956/18018 [10:45:28<25:16:43, 6.97s/step, loss=1.542, acc=0.660, acc_len=1.56, thru=18.2, I=18.9, T=34.4, wait=2.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:15:42,450] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4955: forward=2631.9ms backward=1043.7ms [2026-10-19 00:15:49,283] loop.py:45 INFO TRAIN_STEP step=4957 loss=1.529000 Training: 28%|██▊ | 4957/18018 [10:45:35<25:07:34, 6.93s/step, loss=1.529, acc=0.663, acc_len=1.57, thru=17.6, I=17.4, T=31.0, wait=2.7s, pool=56, epoch=1/2][2026-10-19 00:15:58,543] loop.py:45 INFO TRAIN_STEP step=4958 loss=1.606617 Training: 28%|██▊ | 4958/18018 [10:45:44<27:39:55, 7.63s/step, loss=1.607, acc=0.649, acc_len=1.52, thru=13.8, I=15.2, T=27.9, wait=4.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:07,039] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:16:08,332] loop.py:45 INFO TRAIN_STEP step=4959 loss=1.577534 Training: 28%|██▊ | 4959/18018 [10:45:54<30:01:07, 8.28s/step, loss=1.578, acc=0.656, acc_len=1.55, thru=15.0, I=14.7, T=30.9, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:08,549] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:13,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:16:14,547] loop.py:380 INFO TIMING step=4960: step=4.058s data=1.772s compute=4.028s [fwd=2.782s bwd=1.231s opt=0.003s] dispatch=2.147s [2026-10-19 00:16:14,548] loop.py:45 INFO TRAIN_STEP step=4960 loss=1.595799 Training: 28%|██▊ | 4960/18018 [10:46:00<27:46:31, 7.66s/step, loss=1.596, acc=0.653, acc_len=1.53, thru=19.2, I=22.4, T=31.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:15,001] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:16:21,019] loop.py:45 INFO TRAIN_STEP step=4961 loss=1.514596 Training: 28%|██▊ | 4961/18018 [10:46:07<26:28:49, 7.30s/step, loss=1.515, acc=0.664, acc_len=1.58, thru=15.3, I=20.6, T=33.5, wait=2.6s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:16:21,014] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4960: forward=2678.3ms backward=1088.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:28,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:16:31,056] loop.py:45 INFO TRAIN_STEP step=4962 loss=1.624373 Training: 28%|██▊ | 4962/18018 [10:46:17<29:27:18, 8.12s/step, loss=1.624, acc=0.651, acc_len=1.53, thru=16.1, I=19.3, T=24.3, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:31,546] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:36,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:16:37,714] loop.py:45 INFO TRAIN_STEP step=4963 loss=1.555603 Training: 28%|██▊ | 4963/18018 [10:46:23<27:51:38, 7.68s/step, loss=1.556, acc=0.659, acc_len=1.56, thru=13.7, I=23.5, T=31.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:38,125] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:16:45,493] loop.py:45 INFO TRAIN_STEP step=4964 loss=1.614622 Training: 28%|██▊ | 4964/18018 [10:46:31<27:57:47, 7.71s/step, loss=1.615, acc=0.647, acc_len=1.52, thru=13.7, I=17.4, T=31.8, wait=3.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:55,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:16:56,602] loop.py:380 INFO TIMING step=4965: step=4.499s data=2.125s compute=4.460s [fwd=3.169s bwd=1.275s opt=0.003s] dispatch=6.601s [2026-10-19 00:16:56,602] loop.py:45 INFO TRAIN_STEP step=4965 loss=1.527846 Training: 28%|██▊ | 4965/18018 [10:46:42<31:39:18, 8.73s/step, loss=1.528, acc=0.665, acc_len=1.58, thru=13.5, I=15.3, T=28.5, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:16:56,805] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:02,341] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:17:03,438] loop.py:45 INFO TRAIN_STEP step=4966 loss=1.458990 Training: 28%|██▊ | 4966/18018 [10:46:49<29:35:44, 8.16s/step, loss=1.459, acc=0.675, acc_len=1.61, thru=16.9, I=20.9, T=33.7, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:17:03,432] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4965: forward=2674.6ms backward=1077.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:03,850] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:08,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:17:10,093] loop.py:45 INFO TRAIN_STEP step=4967 loss=1.534034 Training: 28%|██▊ | 4967/18018 [10:46:56<27:57:06, 7.71s/step, loss=1.534, acc=0.664, acc_len=1.57, thru=18.8, I=25.1, T=31.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:10,212] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:15,773] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:17:16,697] loop.py:45 INFO TRAIN_STEP step=4968 loss=1.529405 Training: 28%|██▊ | 4968/18018 [10:47:02<26:44:49, 7.38s/step, loss=1.529, acc=0.664, acc_len=1.57, thru=18.3, I=22.8, T=34.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:16,789] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:22,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:17:23,619] loop.py:45 INFO TRAIN_STEP step=4969 loss=1.606729 Training: 28%|██▊ | 4969/18018 [10:47:09<26:15:00, 7.24s/step, loss=1.607, acc=0.650, acc_len=1.52, thru=18.2, I=21.2, T=30.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:23,690] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:30,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:17:31,190] loop.py:380 INFO TIMING step=4970: step=4.157s data=1.937s compute=4.102s [fwd=2.933s bwd=1.154s opt=0.003s] dispatch=3.402s [2026-10-19 00:17:31,190] loop.py:45 INFO TRAIN_STEP step=4970 loss=1.475141 Training: 28%|██▊ | 4970/18018 [10:47:17<26:36:17, 7.34s/step, loss=1.475, acc=0.680, acc_len=1.62, thru=16.0, I=19.2, T=30.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:31,507] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:37,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:17:37,992] loop.py:45 INFO TRAIN_STEP step=4971 loss=1.612647 Training: 28%|██▊ | 4971/18018 [10:47:24<26:00:57, 7.18s/step, loss=1.613, acc=0.651, acc_len=1.53, thru=18.5, I=19.3, T=33.3, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:17:37,987] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4970: forward=2629.2ms backward=1152.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:38,179] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:43,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:17:44,324] loop.py:45 INFO TRAIN_STEP step=4972 loss=1.561366 Training: 28%|██▊ | 4972/18018 [10:47:30<25:05:37, 6.92s/step, loss=1.561, acc=0.660, acc_len=1.56, thru=19.4, I=21.9, T=33.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:44,537] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:50,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:17:51,238] loop.py:45 INFO TRAIN_STEP step=4973 loss=1.520011 Training: 28%|██▊ | 4973/18018 [10:47:37<25:04:58, 6.92s/step, loss=1.520, acc=0.665, acc_len=1.57, thru=17.4, I=21.6, T=31.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:51,548] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:57,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:17:59,064] loop.py:45 INFO TRAIN_STEP step=4974 loss=1.603471 Training: 28%|██▊ | 4974/18018 [10:47:45<26:03:49, 7.19s/step, loss=1.603, acc=0.657, acc_len=1.54, thru=14.9, I=19.2, T=29.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:17:59,117] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-19 00:18:05,904] loop.py:380 INFO TIMING step=4975: step=3.975s data=1.801s compute=3.939s [fwd=2.768s bwd=1.157s opt=0.003s] dispatch=2.855s [2026-10-19 00:18:05,904] loop.py:45 INFO TRAIN_STEP step=4975 loss=1.468305 Training: 28%|██▊ | 4975/18018 [10:47:51<25:40:33, 7.09s/step, loss=1.468, acc=0.671, acc_len=1.59, thru=16.6, I=18.8, T=32.2, wait=2.9s, pool=56, epoch=1/2][2026-10-19 00:18:14,101] loop.py:45 INFO TRAIN_STEP step=4976 loss=1.606118 Training: 28%|██▊ | 4976/18018 [10:48:00<26:53:06, 7.42s/step, loss=1.606, acc=0.652, acc_len=1.53, thru=16.2, I=16.4, T=28.9, wait=3.8s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:18:14,097] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4975: forward=3099.4ms backward=1269.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:19,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:18:21,182] loop.py:45 INFO TRAIN_STEP step=4977 loss=1.491259 Training: 28%|██▊ | 4977/18018 [10:48:07<26:30:26, 7.32s/step, loss=1.491, acc=0.669, acc_len=1.59, thru=18.4, I=19.7, T=29.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:21,575] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 00:18:28,471] loop.py:45 INFO TRAIN_STEP step=4978 loss=1.518004 Training: 28%|██▊ | 4978/18018 [10:48:14<26:28:31, 7.31s/step, loss=1.518, acc=0.666, acc_len=1.58, thru=14.9, I=19.6, T=30.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:28,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:28,960] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:37,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:18:38,956] loop.py:45 INFO TRAIN_STEP step=4979 loss=1.618583 Training: 28%|██▊ | 4979/18018 [10:48:25<29:55:32, 8.26s/step, loss=1.619, acc=0.648, acc_len=1.52, thru=13.7, I=15.0, T=26.9, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:39,252] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:46,892] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:18:47,284] loop.py:380 INFO TIMING step=4980: step=4.426s data=2.111s compute=4.392s [fwd=3.076s bwd=1.302s opt=0.003s] dispatch=3.892s [2026-10-19 00:18:47,284] loop.py:45 INFO TRAIN_STEP step=4980 loss=1.552836 Training: 28%|██▊ | 4980/18018 [10:48:33<29:59:39, 8.28s/step, loss=1.553, acc=0.662, acc_len=1.56, thru=15.9, I=17.5, T=28.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:47,413] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:54,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:18:55,545] loop.py:45 INFO TRAIN_STEP step=4981 loss=1.526450 Training: 28%|██▊ | 4981/18018 [10:48:41<29:58:00, 8.27s/step, loss=1.526, acc=0.664, acc_len=1.56, thru=15.3, I=16.2, T=26.6, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:18:55,541] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4980: forward=3358.3ms backward=1409.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:18:55,974] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:02,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:19:04,191] loop.py:45 INFO TRAIN_STEP step=4982 loss=1.659024 Training: 28%|██▊ | 4982/18018 [10:48:50<30:22:05, 8.39s/step, loss=1.659, acc=0.643, acc_len=1.51, thru=14.0, I=19.3, T=30.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:04,632] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:09,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:19:11,727] loop.py:45 INFO TRAIN_STEP step=4983 loss=1.532591 Training: 28%|██▊ | 4983/18018 [10:48:57<29:26:39, 8.13s/step, loss=1.533, acc=0.665, acc_len=1.57, thru=16.5, I=25.7, T=30.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:11,763] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:16,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:19:17,785] loop.py:45 INFO TRAIN_STEP step=4984 loss=1.484290 Training: 28%|██▊ | 4984/18018 [10:49:03<27:11:23, 7.51s/step, loss=1.484, acc=0.669, acc_len=1.58, thru=18.6, I=26.4, T=31.9, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:18,035] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:19:26,091] loop.py:380 INFO TIMING step=4985: step=3.857s data=1.686s compute=3.810s [fwd=2.730s bwd=1.064s opt=0.003s] dispatch=4.438s [2026-10-19 00:19:26,091] loop.py:45 INFO TRAIN_STEP step=4985 loss=1.462448 Training: 28%|██▊ | 4985/18018 [10:49:12<28:03:18, 7.75s/step, loss=1.462, acc=0.673, acc_len=1.60, thru=13.4, I=16.8, T=33.2, wait=4.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:34,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:19:36,139] loop.py:45 INFO TRAIN_STEP step=4986 loss=1.395041 Training: 28%|██▊ | 4986/18018 [10:49:22<30:32:48, 8.44s/step, loss=1.395, acc=0.689, acc_len=1.65, thru=14.0, I=14.4, T=25.0, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:19:36,134] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4985: forward=3432.1ms backward=1628.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:36,236] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-19 00:19:44,436] loop.py:45 INFO TRAIN_STEP step=4987 loss=1.612341 Training: 28%|██▊ | 4987/18018 [10:49:30<30:23:14, 8.39s/step, loss=1.612, acc=0.654, acc_len=1.54, thru=13.0, I=15.3, T=29.5, wait=4.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:53,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:19:54,320] loop.py:45 INFO TRAIN_STEP step=4988 loss=1.546585 Training: 28%|██▊ | 4988/18018 [10:49:40<32:00:14, 8.84s/step, loss=1.547, acc=0.662, acc_len=1.57, thru=14.6, I=15.6, T=25.8, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:19:54,437] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:01,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:20:01,564] loop.py:45 INFO TRAIN_STEP step=4989 loss=1.604749 Training: 28%|██▊ | 4989/18018 [10:49:47<30:16:00, 8.36s/step, loss=1.605, acc=0.653, acc_len=1.53, thru=15.7, I=15.4, T=30.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:01,744] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:08,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:20:10,821] loop.py:380 INFO TIMING step=4990: step=5.490s data=2.661s compute=5.462s [fwd=3.893s bwd=1.554s opt=0.003s] dispatch=3.756s [2026-10-19 00:20:10,821] loop.py:45 INFO TRAIN_STEP step=4990 loss=1.557431 Training: 28%|██▊ | 4990/18018 [10:49:56<31:14:07, 8.63s/step, loss=1.557, acc=0.660, acc_len=1.56, thru=14.1, I=16.4, T=23.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:10,857] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:17,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:20:18,787] loop.py:45 INFO TRAIN_STEP step=4991 loss=1.454798 Training: 28%|██▊ | 4991/18018 [10:50:04<30:30:37, 8.43s/step, loss=1.455, acc=0.679, acc_len=1.62, thru=13.2, I=18.0, T=28.7, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:20:18,782] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4990: forward=3105.7ms backward=1314.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:19,187] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:25,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:20:26,149] loop.py:45 INFO TRAIN_STEP step=4992 loss=1.584144 Training: 28%|██▊ | 4992/18018 [10:50:12<29:20:43, 8.11s/step, loss=1.584, acc=0.655, acc_len=1.54, thru=16.8, I=18.8, T=28.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:26,253] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:31,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:20:32,466] loop.py:45 INFO TRAIN_STEP step=4993 loss=1.568396 Training: 28%|██▊ | 4993/18018 [10:50:18<27:23:48, 7.57s/step, loss=1.568, acc=0.661, acc_len=1.56, thru=18.6, I=20.7, T=35.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:32,901] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:20:40,097] loop.py:45 INFO TRAIN_STEP step=4994 loss=1.658369 Training: 28%|██▊ | 4994/18018 [10:50:26<27:27:27, 7.59s/step, loss=1.658, acc=0.641, acc_len=1.50, thru=15.4, I=16.5, T=30.8, wait=3.5s, pool=52, epoch=1/2][2026-10-19 00:20:46,723] loop.py:380 INFO TIMING step=4995: step=4.114s data=1.716s compute=4.078s [fwd=2.838s bwd=1.225s opt=0.003s] dispatch=2.504s [2026-10-19 00:20:46,723] loop.py:45 INFO TRAIN_STEP step=4995 loss=1.533298 Training: 28%|██▊ | 4995/18018 [10:50:32<26:24:35, 7.30s/step, loss=1.533, acc=0.667, acc_len=1.57, thru=17.4, I=15.7, T=31.1, wait=2.5s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:52,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:20:54,744] loop.py:45 INFO TRAIN_STEP step=4996 loss=1.468590 Training: 28%|██▊ | 4996/18018 [10:50:40<27:11:25, 7.52s/step, loss=1.469, acc=0.677, acc_len=1.61, thru=18.4, I=20.2, T=27.6, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:20:54,740] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=4995: forward=3187.5ms backward=1404.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:20:55,040] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:01,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:21:02,610] loop.py:45 INFO TRAIN_STEP step=4997 loss=1.424201 Training: 28%|██▊ | 4997/18018 [10:50:48<27:34:05, 7.62s/step, loss=1.424, acc=0.682, acc_len=1.63, thru=16.9, I=19.0, T=32.8, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:02,948] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:21:11,706] loop.py:45 INFO TRAIN_STEP step=4998 loss=1.416811 Training: 28%|██▊ | 4998/18018 [10:50:57<29:09:50, 8.06s/step, loss=1.417, acc=0.678, acc_len=1.62, thru=12.7, I=14.9, T=31.4, wait=5.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:18,131] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:21:19,929] loop.py:45 INFO TRAIN_STEP step=4999 loss=1.479139 Training: 28%|██▊ | 4999/18018 [10:51:06<29:20:12, 8.11s/step, loss=1.479, acc=0.675, acc_len=1.61, thru=16.6, I=17.0, T=28.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:20,146] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:26,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:21:27,717] loop.py:380 INFO TIMING step=5000: step=4.207s data=1.811s compute=4.170s [fwd=2.963s bwd=1.191s opt=0.003s] dispatch=3.570s [2026-10-19 00:21:27,717] loop.py:45 INFO TRAIN_STEP step=5000 loss=1.576925 Training: 28%|██▊ | 5000/18018 [10:51:13<28:58:58, 8.01s/step, loss=1.577, acc=0.659, acc_len=1.55, thru=14.0, I=19.3, T=30.4, wait=3.6s, pool=64, epoch=1/2][2026-10-19 00:21:27,722] loop.py:419 INFO Saving checkpoint at step 5000... (TrainerActor pid=9310) [2026-10-19 00:21:31,157] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0005001 (TrainerActor pid=9310) [2026-10-19 00:21:31,191] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0004251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:31,539] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:32,560] inference_manager.py:408 INFO Pool capacity available after 6.1s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:40,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 00:21:31,159] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0005001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 00:21:32,326] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0004251 [2026-10-19 00:21:40,992] loop.py:45 INFO TRAIN_STEP step=5001 loss=1.573655 Training: 28%|██▊ | 5001/18018 [10:51:27<34:41:11, 9.59s/step, loss=1.574, acc=0.655, acc_len=1.55, thru=15.2, I=16.4, T=30.8, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:21:40,987] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5000: forward=2886.4ms backward=1220.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:41,064] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:48,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:21:50,886] loop.py:45 INFO TRAIN_STEP step=5002 loss=1.672565 Training: 28%|██▊ | 5002/18018 [10:51:36<35:00:33, 9.68s/step, loss=1.673, acc=0.637, acc_len=1.48, thru=13.1, I=15.4, T=25.4, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:51,222] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:57,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:21:57,932] loop.py:45 INFO TRAIN_STEP step=5003 loss=1.570039 Training: 28%|██▊ | 5003/18018 [10:51:44<32:08:52, 8.89s/step, loss=1.570, acc=0.660, acc_len=1.56, thru=15.1, I=20.6, T=30.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:21:58,266] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:04,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:22:05,606] loop.py:45 INFO TRAIN_STEP step=5004 loss=1.699423 Training: 28%|██▊ | 5004/18018 [10:51:51<30:49:21, 8.53s/step, loss=1.699, acc=0.636, acc_len=1.48, thru=15.1, I=18.4, T=31.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:05,961] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:22:13,420] loop.py:380 INFO TIMING step=5005: step=4.581s data=2.266s compute=4.536s [fwd=3.199s bwd=1.321s opt=0.003s] dispatch=3.224s [2026-10-19 00:22:13,420] loop.py:45 INFO TRAIN_STEP step=5005 loss=1.662340 Training: 28%|██▊ | 5005/18018 [10:51:59<30:03:01, 8.31s/step, loss=1.662, acc=0.640, acc_len=1.49, thru=15.3, I=18.8, T=27.9, wait=3.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:19,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:22:21,701] loop.py:45 INFO TRAIN_STEP step=5006 loss=1.553782 Training: 28%|██▊ | 5006/18018 [10:52:07<30:00:31, 8.30s/step, loss=1.554, acc=0.657, acc_len=1.55, thru=15.9, I=18.0, T=26.0, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:22:21,696] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5005: forward=3434.5ms backward=1436.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:22,145] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:28,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:22:28,175] loop.py:45 INFO TRAIN_STEP step=5007 loss=1.665370 Training: 28%|██▊ | 5007/18018 [10:52:14<28:01:35, 7.75s/step, loss=1.665, acc=0.645, acc_len=1.51, thru=15.9, I=20.7, T=32.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:28,508] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:34,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:22:35,960] loop.py:45 INFO TRAIN_STEP step=5008 loss=1.652477 Training: 28%|██▊ | 5008/18018 [10:52:22<28:03:21, 7.76s/step, loss=1.652, acc=0.647, acc_len=1.51, thru=17.2, I=20.7, T=31.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:36,031] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:42,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:22:43,681] loop.py:45 INFO TRAIN_STEP step=5009 loss=1.567035 Training: 28%|██▊ | 5009/18018 [10:52:29<28:00:25, 7.75s/step, loss=1.567, acc=0.660, acc_len=1.55, thru=13.9, I=20.7, T=30.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:44,011] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:49,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:22:50,007] loop.py:380 INFO TIMING step=5010: step=3.968s data=1.796s compute=3.938s [fwd=2.759s bwd=1.163s opt=0.003s] dispatch=2.350s [2026-10-19 00:22:50,008] loop.py:45 INFO TRAIN_STEP step=5010 loss=1.570529 Training: 28%|██▊ | 5010/18018 [10:52:36<26:27:41, 7.32s/step, loss=1.571, acc=0.656, acc_len=1.54, thru=17.9, I=21.9, T=32.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:50,366] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:56,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:22:57,300] loop.py:45 INFO TRAIN_STEP step=5011 loss=1.580869 Training: 28%|██▊ | 5011/18018 [10:52:43<26:25:30, 7.31s/step, loss=1.581, acc=0.651, acc_len=1.54, thru=17.7, I=18.1, T=34.7, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:22:57,295] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5010: forward=2626.7ms backward=1016.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:22:57,431] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-19 00:23:04,007] loop.py:45 INFO TRAIN_STEP step=5012 loss=1.611603 Training: 28%|██▊ | 5012/18018 [10:52:50<25:46:01, 7.13s/step, loss=1.612, acc=0.652, acc_len=1.53, thru=17.9, I=20.4, T=32.4, wait=2.7s, pool=60, epoch=1/2][2026-10-19 00:23:11,427] loop.py:45 INFO TRAIN_STEP step=5013 loss=1.692344 Training: 28%|██▊ | 5013/18018 [10:52:57<26:04:34, 7.22s/step, loss=1.692, acc=0.640, acc_len=1.49, thru=16.6, I=16.1, T=32.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:11,387] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:11,889] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:18,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:23:19,495] loop.py:45 INFO TRAIN_STEP step=5014 loss=1.634060 Training: 28%|██▊ | 5014/18018 [10:53:05<26:59:50, 7.47s/step, loss=1.634, acc=0.649, acc_len=1.53, thru=16.0, I=17.8, T=33.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:19,887] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:27,119] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:23:27,229] loop.py:380 INFO TIMING step=5015: step=5.153s data=1.950s compute=5.126s [fwd=3.860s bwd=1.250s opt=0.003s] dispatch=2.571s [2026-10-19 00:23:27,229] loop.py:45 INFO TRAIN_STEP step=5015 loss=1.533022 Training: 28%|██▊ | 5015/18018 [10:53:13<27:16:34, 7.55s/step, loss=1.533, acc=0.663, acc_len=1.57, thru=16.7, I=21.6, T=24.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:27,620] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:35,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:23:36,305] loop.py:45 INFO TRAIN_STEP step=5016 loss=1.657500 Training: 28%|██▊ | 5016/18018 [10:53:22<28:55:30, 8.01s/step, loss=1.657, acc=0.646, acc_len=1.51, thru=13.7, I=13.9, T=26.6, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:23:36,300] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5015: forward=3420.3ms backward=1335.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:36,724] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:43,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:23:44,997] loop.py:45 INFO TRAIN_STEP step=5017 loss=1.530439 Training: 28%|██▊ | 5017/18018 [10:53:31<29:39:55, 8.21s/step, loss=1.530, acc=0.669, acc_len=1.58, thru=14.1, I=15.7, T=25.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:45,098] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:50,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:23:51,177] loop.py:45 INFO TRAIN_STEP step=5018 loss=1.607648 Training: 28%|██▊ | 5018/18018 [10:53:37<27:27:29, 7.60s/step, loss=1.608, acc=0.651, acc_len=1.53, thru=18.0, I=22.7, T=34.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:23:51,264] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-19 00:23:58,292] loop.py:45 INFO TRAIN_STEP step=5019 loss=1.555014 Training: 28%|██▊ | 5019/18018 [10:53:44<26:55:37, 7.46s/step, loss=1.555, acc=0.660, acc_len=1.56, thru=15.0, I=19.0, T=34.7, wait=3.4s, pool=48, epoch=1/2][2026-10-19 00:24:08,453] loop.py:380 INFO TIMING step=5020: step=5.126s data=2.806s compute=5.098s [fwd=3.928s bwd=1.156s opt=0.003s] dispatch=5.025s [2026-10-19 00:24:08,453] loop.py:45 INFO TRAIN_STEP step=5020 loss=1.525424 Training: 28%|██▊ | 5020/18018 [10:53:54<29:51:13, 8.27s/step, loss=1.525, acc=0.663, acc_len=1.56, thru=14.2, I=12.7, T=25.0, wait=5.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:24:15,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:24:18,496] loop.py:45 INFO TRAIN_STEP step=5021 loss=1.500661 Training: 28%|██▊ | 5021/18018 [10:54:04<31:46:16, 8.80s/step, loss=1.501, acc=0.667, acc_len=1.58, thru=13.8, I=16.1, T=27.3, wait=5.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:24:18,491] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5020: forward=3376.3ms backward=1257.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:24:18,984] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:24:24,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:24:24,897] loop.py:45 INFO TRAIN_STEP step=5022 loss=1.586104 Training: 28%|██▊ | 5022/18018 [10:54:10<29:10:20, 8.08s/step, loss=1.586, acc=0.654, acc_len=1.54, thru=16.1, I=24.9, T=31.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:24:25,163] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:24:30,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:24:32,549] loop.py:45 INFO TRAIN_STEP step=5023 loss=1.625968 Training: 28%|██▊ | 5023/18018 [10:54:18<28:42:11, 7.95s/step, loss=1.626, acc=0.652, acc_len=1.53, thru=16.1, I=19.7, T=31.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:24:32,817] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 00:24:39,302] loop.py:45 INFO TRAIN_STEP step=5024 loss=1.700284 Training: 28%|██▊ | 5024/18018 [10:54:25<27:24:08, 7.59s/step, loss=1.700, acc=0.641, acc_len=1.49, thru=16.3, I=19.3, T=31.9, wait=2.7s, pool=56, epoch=1/2][2026-10-19 00:24:49,232] loop.py:380 INFO TIMING step=5025: step=4.236s data=1.936s compute=4.185s [fwd=2.853s bwd=1.317s opt=0.003s] dispatch=5.686s [2026-10-19 00:24:49,232] loop.py:45 INFO TRAIN_STEP step=5025 loss=1.730787 Training: 28%|██▊ | 5025/18018 [10:54:35<29:55:55, 8.29s/step, loss=1.731, acc=0.631, acc_len=1.46, thru=13.4, I=13.8, T=30.2, wait=5.7s, pool=60, epoch=1/2][2026-10-19 00:24:56,394] loop.py:45 INFO TRAIN_STEP step=5026 loss=1.472946 Training: 28%|██▊ | 5026/18018 [10:54:42<28:42:23, 7.95s/step, loss=1.473, acc=0.673, acc_len=1.60, thru=16.2, I=15.9, T=30.7, wait=3.0s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:24:56,391] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5025: forward=2919.0ms backward=1198.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:03,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:25:05,156] loop.py:45 INFO TRAIN_STEP step=5027 loss=1.451816 Training: 28%|██▊ | 5027/18018 [10:54:51<29:34:40, 8.20s/step, loss=1.452, acc=0.680, acc_len=1.63, thru=16.4, I=16.6, T=32.0, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:05,310] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:12,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:25:12,805] loop.py:45 INFO TRAIN_STEP step=5028 loss=1.559149 Training: 28%|██▊ | 5028/18018 [10:54:58<28:58:58, 8.03s/step, loss=1.559, acc=0.658, acc_len=1.56, thru=16.0, I=20.1, T=30.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:13,051] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:21,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:25:22,753] loop.py:45 INFO TRAIN_STEP step=5029 loss=1.523937 Training: 28%|██▊ | 5029/18018 [10:55:08<31:03:21, 8.61s/step, loss=1.524, acc=0.670, acc_len=1.59, thru=13.2, I=15.2, T=29.2, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:23,196] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:29,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:25:29,726] loop.py:380 INFO TIMING step=5030: step=3.670s data=1.614s compute=3.641s [fwd=2.600s bwd=1.026s opt=0.003s] dispatch=3.293s [2026-10-19 00:25:29,726] loop.py:45 INFO TRAIN_STEP step=5030 loss=1.484752 Training: 28%|██▊ | 5030/18018 [10:55:15<29:17:03, 8.12s/step, loss=1.485, acc=0.674, acc_len=1.60, thru=16.5, I=21.5, T=34.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:30,114] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:35,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:25:37,441] loop.py:45 INFO TRAIN_STEP step=5031 loss=1.595410 Training: 28%|██▊ | 5031/18018 [10:55:23<28:50:37, 8.00s/step, loss=1.595, acc=0.654, acc_len=1.54, thru=17.5, I=23.1, T=30.1, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:25:37,437] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5030: forward=2900.9ms backward=1306.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:37,656] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:42,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:25:44,256] loop.py:45 INFO TRAIN_STEP step=5032 loss=1.549609 Training: 28%|██▊ | 5032/18018 [10:55:30<27:34:02, 7.64s/step, loss=1.550, acc=0.665, acc_len=1.57, thru=15.8, I=26.8, T=28.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:44,639] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:52,648] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:25:53,335] loop.py:45 INFO TRAIN_STEP step=5033 loss=1.490183 Training: 28%|██▊ | 5033/18018 [10:55:39<29:07:11, 8.07s/step, loss=1.490, acc=0.671, acc_len=1.59, thru=14.7, I=16.2, T=29.7, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:53,652] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:25:58,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:25:59,991] loop.py:45 INFO TRAIN_STEP step=5034 loss=1.623902 Training: 28%|██▊ | 5034/18018 [10:55:46<27:34:54, 7.65s/step, loss=1.624, acc=0.647, acc_len=1.52, thru=17.9, I=23.1, T=29.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:00,053] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:05,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:26:06,343] loop.py:380 INFO TIMING step=5035: step=4.165s data=1.980s compute=4.128s [fwd=2.851s bwd=1.263s opt=0.003s] dispatch=2.181s [2026-10-19 00:26:06,344] loop.py:45 INFO TRAIN_STEP step=5035 loss=1.399832 Training: 28%|██▊ | 5035/18018 [10:55:52<26:10:51, 7.26s/step, loss=1.400, acc=0.685, acc_len=1.64, thru=17.6, I=26.2, T=30.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:06,738] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:14,021] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:26:14,439] loop.py:45 INFO TRAIN_STEP step=5036 loss=1.610141 Training: 28%|██▊ | 5036/18018 [10:56:00<27:05:00, 7.51s/step, loss=1.610, acc=0.654, acc_len=1.53, thru=14.1, I=17.1, T=29.5, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:26:14,434] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5035: forward=3074.2ms backward=1219.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:14,536] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:20,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:26:21,057] loop.py:45 INFO TRAIN_STEP step=5037 loss=1.640610 Training: 28%|██▊ | 5037/18018 [10:56:07<26:07:00, 7.24s/step, loss=1.641, acc=0.645, acc_len=1.50, thru=17.0, I=18.9, T=33.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:21,163] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:26,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:26:28,042] loop.py:45 INFO TRAIN_STEP step=5038 loss=1.522398 Training: 28%|██▊ | 5038/18018 [10:56:14<25:50:09, 7.17s/step, loss=1.522, acc=0.664, acc_len=1.57, thru=17.0, I=21.3, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:28,079] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:36,237] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:26:36,667] loop.py:45 INFO TRAIN_STEP step=5039 loss=1.585796 Training: 28%|██▊ | 5039/18018 [10:56:22<27:24:40, 7.60s/step, loss=1.586, acc=0.655, acc_len=1.54, thru=14.9, I=16.7, T=28.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:36,752] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:45,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:26:46,971] loop.py:380 INFO TIMING step=5040: step=4.730s data=2.199s compute=4.681s [fwd=3.228s bwd=1.438s opt=0.003s] dispatch=5.564s [2026-10-19 00:26:46,971] loop.py:45 INFO TRAIN_STEP step=5040 loss=1.565440 Training: 28%|██▊ | 5040/18018 [10:56:33<30:19:42, 8.41s/step, loss=1.565, acc=0.657, acc_len=1.55, thru=12.9, I=12.9, T=27.1, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:26:47,445] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:26:54,409] loop.py:45 INFO TRAIN_STEP step=5041 loss=1.524151 Training: 28%|██▊ | 5041/18018 [10:56:40<29:16:23, 8.12s/step, loss=1.524, acc=0.666, acc_len=1.58, thru=15.3, I=20.0, T=29.3, wait=3.1s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:26:54,404] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5040: forward=3030.4ms backward=1282.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:00,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:27:03,082] loop.py:45 INFO TRAIN_STEP step=5042 loss=1.572631 Training: 28%|██▊ | 5042/18018 [10:56:49<29:52:06, 8.29s/step, loss=1.573, acc=0.657, acc_len=1.55, thru=15.0, I=16.2, T=27.0, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:03,429] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:09,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:27:10,580] loop.py:45 INFO TRAIN_STEP step=5043 loss=1.629370 Training: 28%|██▊ | 5043/18018 [10:56:56<29:00:47, 8.05s/step, loss=1.629, acc=0.649, acc_len=1.52, thru=13.7, I=20.2, T=32.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:10,892] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:16,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:27:17,934] loop.py:45 INFO TRAIN_STEP step=5044 loss=1.631094 Training: 28%|██▊ | 5044/18018 [10:57:04<28:15:27, 7.84s/step, loss=1.631, acc=0.647, acc_len=1.51, thru=16.9, I=19.7, T=33.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:18,177] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:27:25,860] loop.py:380 INFO TIMING step=5045: step=4.297s data=2.063s compute=4.267s [fwd=2.788s bwd=1.464s opt=0.003s] dispatch=3.620s [2026-10-19 00:27:25,860] loop.py:45 INFO TRAIN_STEP step=5045 loss=1.594726 Training: 28%|██▊ | 5045/18018 [10:57:11<28:20:57, 7.87s/step, loss=1.595, acc=0.653, acc_len=1.54, thru=12.0, I=16.5, T=29.8, wait=3.6s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:32,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:27:33,305] loop.py:45 INFO TRAIN_STEP step=5046 loss=1.543727 Training: 28%|██▊ | 5046/18018 [10:57:19<27:53:29, 7.74s/step, loss=1.544, acc=0.665, acc_len=1.57, thru=18.5, I=18.0, T=31.8, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:27:33,300] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5045: forward=2851.0ms backward=1115.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:33,407] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-19 00:27:39,299] loop.py:45 INFO TRAIN_STEP step=5047 loss=1.502754 Training: 28%|██▊ | 5047/18018 [10:57:25<26:00:01, 7.22s/step, loss=1.503, acc=0.668, acc_len=1.58, thru=19.0, I=22.0, T=32.1, wait=2.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:44,276] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:27:45,563] loop.py:45 INFO TRAIN_STEP step=5048 loss=1.454757 Training: 28%|██▊ | 5048/18018 [10:57:31<24:58:11, 6.93s/step, loss=1.455, acc=0.677, acc_len=1.61, thru=18.6, I=21.0, T=32.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:45,785] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:51,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:27:52,687] loop.py:45 INFO TRAIN_STEP step=5049 loss=1.507114 Training: 28%|██▊ | 5049/18018 [10:57:38<25:10:33, 6.99s/step, loss=1.507, acc=0.669, acc_len=1.58, thru=17.3, I=22.0, T=30.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:53,054] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:58,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:27:58,923] loop.py:380 INFO TIMING step=5050: step=4.132s data=1.940s compute=4.090s [fwd=2.866s bwd=1.208s opt=0.003s] dispatch=2.094s [2026-10-19 00:27:58,924] loop.py:45 INFO TRAIN_STEP step=5050 loss=1.613547 Training: 28%|██▊ | 5050/18018 [10:57:45<24:21:45, 6.76s/step, loss=1.614, acc=0.651, acc_len=1.53, thru=19.0, I=24.1, T=31.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:27:59,053] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 00:28:05,870] loop.py:45 INFO TRAIN_STEP step=5051 loss=1.618614 Training: 28%|██▊ | 5051/18018 [10:57:51<24:33:25, 6.82s/step, loss=1.619, acc=0.650, acc_len=1.52, thru=16.1, I=18.7, T=30.1, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:28:05,865] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5050: forward=2905.6ms backward=1285.5ms [2026-10-19 00:28:15,634] loop.py:45 INFO TRAIN_STEP step=5052 loss=1.572622 Training: 28%|██▊ | 5052/18018 [10:58:01<27:44:20, 7.70s/step, loss=1.573, acc=0.659, acc_len=1.55, thru=12.8, I=12.9, T=29.6, wait=5.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:28:24,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:28:25,150] loop.py:45 INFO TRAIN_STEP step=5053 loss=1.481935 Training: 28%|██▊ | 5053/18018 [10:58:11<29:41:52, 8.25s/step, loss=1.482, acc=0.676, acc_len=1.61, thru=14.5, I=15.7, T=26.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:28:25,252] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:28:30,659] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:28:31,588] loop.py:45 INFO TRAIN_STEP step=5054 loss=1.625241 Training: 28%|██▊ | 5054/18018 [10:58:17<27:44:31, 7.70s/step, loss=1.625, acc=0.648, acc_len=1.52, thru=17.9, I=20.3, T=32.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:28:31,664] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:28:38,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:28:38,927] loop.py:380 INFO TIMING step=5055: step=3.700s data=1.566s compute=3.655s [fwd=2.559s bwd=1.082s opt=0.003s] dispatch=3.628s [2026-10-19 00:28:38,927] loop.py:45 INFO TRAIN_STEP step=5055 loss=1.563321 Training: 28%|██▊ | 5055/18018 [10:58:25<27:20:34, 7.59s/step, loss=1.563, acc=0.658, acc_len=1.54, thru=16.2, I=19.3, T=34.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:28:39,404] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:28:45,293] loop.py:45 INFO TRAIN_STEP step=5056 loss=1.493350 Training: 28%|██▊ | 5056/18018 [10:58:31<26:00:54, 7.23s/step, loss=1.493, acc=0.671, acc_len=1.60, thru=18.6, I=19.1, T=35.3, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:28:45,288] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5055: forward=2520.1ms backward=1040.1ms [2026-10-19 00:28:56,562] loop.py:45 INFO TRAIN_STEP step=5057 loss=1.535950 Training: 28%|██▊ | 5057/18018 [10:58:42<30:23:01, 8.44s/step, loss=1.536, acc=0.663, acc_len=1.56, thru=11.1, I=11.8, T=28.8, wait=6.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:04,345] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:29:04,868] loop.py:45 INFO TRAIN_STEP step=5058 loss=1.422996 Training: 28%|██▊ | 5058/18018 [10:58:50<30:14:04, 8.40s/step, loss=1.423, acc=0.684, acc_len=1.65, thru=15.8, I=15.4, T=33.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:05,349] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:12,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:29:13,274] loop.py:45 INFO TRAIN_STEP step=5059 loss=1.636963 Training: 28%|██▊ | 5059/18018 [10:58:59<30:14:35, 8.40s/step, loss=1.637, acc=0.646, acc_len=1.51, thru=14.7, I=15.9, T=27.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:13,420] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:20,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:29:22,351] loop.py:380 INFO TIMING step=5060: step=4.487s data=2.040s compute=4.456s [fwd=3.151s bwd=1.291s opt=0.003s] dispatch=4.579s [2026-10-19 00:29:22,351] loop.py:45 INFO TRAIN_STEP step=5060 loss=1.636323 Training: 28%|██▊ | 5060/18018 [10:59:08<30:58:10, 8.60s/step, loss=1.636, acc=0.646, acc_len=1.51, thru=13.3, I=16.3, T=28.5, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:22,719] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:28,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:29:29,449] loop.py:45 INFO TRAIN_STEP step=5061 loss=1.518064 Training: 28%|██▊ | 5061/18018 [10:59:15<29:20:29, 8.15s/step, loss=1.518, acc=0.665, acc_len=1.57, thru=16.5, I=21.9, T=31.6, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:29:29,444] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5060: forward=2894.9ms backward=1106.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:29,537] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:34,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:29:36,078] loop.py:45 INFO TRAIN_STEP step=5062 loss=1.519497 Training: 28%|██▊ | 5062/18018 [10:59:22<27:41:38, 7.70s/step, loss=1.519, acc=0.666, acc_len=1.57, thru=17.4, I=23.3, T=34.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:36,273] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:40,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:29:43,065] loop.py:45 INFO TRAIN_STEP step=5063 loss=1.508155 Training: 28%|██▊ | 5063/18018 [10:59:29<26:55:36, 7.48s/step, loss=1.508, acc=0.670, acc_len=1.58, thru=18.6, I=25.4, T=28.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:43,280] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 00:29:49,466] loop.py:45 INFO TRAIN_STEP step=5064 loss=1.435013 Training: 28%|██▊ | 5064/18018 [10:59:35<25:45:27, 7.16s/step, loss=1.435, acc=0.679, acc_len=1.62, thru=16.1, I=22.1, T=33.1, wait=2.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:57,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:29:58,517] loop.py:380 INFO TIMING step=5065: step=4.506s data=2.091s compute=4.464s [fwd=3.173s bwd=1.277s opt=0.003s] dispatch=4.535s [2026-10-19 00:29:58,518] loop.py:45 INFO TRAIN_STEP step=5065 loss=1.535702 Training: 28%|██▊ | 5065/18018 [10:59:44<27:48:03, 7.73s/step, loss=1.536, acc=0.662, acc_len=1.56, thru=15.7, I=17.1, T=28.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:29:58,835] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:04,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:30:06,119] loop.py:45 INFO TRAIN_STEP step=5066 loss=1.699695 Training: 28%|██▊ | 5066/18018 [10:59:52<27:39:39, 7.69s/step, loss=1.700, acc=0.640, acc_len=1.49, thru=15.1, I=17.3, T=30.8, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:30:06,114] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5065: forward=2863.8ms backward=1228.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:06,495] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:11,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:30:12,920] loop.py:45 INFO TRAIN_STEP step=5067 loss=1.476581 Training: 28%|██▊ | 5067/18018 [10:59:59<26:42:03, 7.42s/step, loss=1.477, acc=0.670, acc_len=1.59, thru=17.9, I=21.8, T=31.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:13,140] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:30:19,084] loop.py:45 INFO TRAIN_STEP step=5068 loss=1.556096 Training: 28%|██▊ | 5068/18018 [11:00:05<25:20:32, 7.04s/step, loss=1.556, acc=0.660, acc_len=1.56, thru=19.0, I=24.2, T=33.4, wait=2.3s, pool=60, epoch=1/2][2026-10-19 00:30:26,090] loop.py:45 INFO TRAIN_STEP step=5069 loss=1.459896 Training: 28%|██▊ | 5069/18018 [11:00:12<25:18:03, 7.03s/step, loss=1.460, acc=0.679, acc_len=1.62, thru=17.1, I=17.1, T=29.6, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:30,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:30:32,999] loop.py:380 INFO TIMING step=5070: step=4.149s data=1.980s compute=4.112s [fwd=2.848s bwd=1.248s opt=0.003s] dispatch=2.748s [2026-10-19 00:30:33,000] loop.py:45 INFO TRAIN_STEP step=5070 loss=1.614491 Training: 28%|██▊ | 5070/18018 [11:00:19<25:09:38, 7.00s/step, loss=1.614, acc=0.649, acc_len=1.52, thru=17.4, I=21.3, T=30.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:33,388] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:39,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 00:30:41,015] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5070: forward=3081.2ms backward=1465.0ms [2026-10-19 00:30:41,020] loop.py:45 INFO TRAIN_STEP step=5071 loss=1.553434 Training: 28%|██▊ | 5071/18018 [11:00:27<26:15:59, 7.30s/step, loss=1.553, acc=0.664, acc_len=1.57, thru=16.8, I=21.6, T=27.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:41,232] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:45,756] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:30:47,786] loop.py:45 INFO TRAIN_STEP step=5072 loss=1.544841 Training: 28%|██▊ | 5072/18018 [11:00:33<25:41:03, 7.14s/step, loss=1.545, acc=0.661, acc_len=1.56, thru=17.0, I=26.0, T=27.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:30:48,274] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 00:30:54,961] loop.py:45 INFO TRAIN_STEP step=5073 loss=1.432943 Training: 28%|██▊ | 5073/18018 [11:00:41<25:43:04, 7.15s/step, loss=1.433, acc=0.682, acc_len=1.63, thru=13.5, I=18.0, T=33.5, wait=3.3s, pool=52, epoch=1/2][2026-10-19 00:31:07,257] loop.py:45 INFO TRAIN_STEP step=5074 loss=1.435395 Training: 28%|██▊ | 5074/18018 [11:00:53<31:15:44, 8.69s/step, loss=1.435, acc=0.680, acc_len=1.63, thru=10.8, I=9.1, T=31.2, wait=8.2s, pool=32, epoch=1/2] [2026-10-19 00:31:16,256] loop.py:380 INFO TIMING step=5075: step=4.066s data=1.786s compute=4.034s [fwd=2.926s bwd=1.092s opt=0.003s] dispatch=4.925s [2026-10-19 00:31:16,256] loop.py:45 INFO TRAIN_STEP step=5075 loss=1.481617 Training: 28%|██▊ | 5075/18018 [11:01:02<31:35:16, 8.79s/step, loss=1.482, acc=0.674, acc_len=1.60, thru=16.6, I=12.7, T=31.5, wait=4.9s, pool=52, epoch=1/2][2026-10-19 00:31:26,285] loop.py:45 INFO TRAIN_STEP step=5076 loss=1.569500 Training: 28%|██▊ | 5076/18018 [11:01:12<32:55:36, 9.16s/step, loss=1.569, acc=0.661, acc_len=1.56, thru=13.8, I=14.3, T=26.7, wait=5.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:31:26,280] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5075: forward=3332.3ms backward=1402.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:31:33,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:31:35,165] loop.py:45 INFO TRAIN_STEP step=5077 loss=1.652565 Training: 28%|██▊ | 5077/18018 [11:01:21<32:37:22, 9.08s/step, loss=1.653, acc=0.644, acc_len=1.50, thru=15.4, I=15.0, T=26.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:31:35,602] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 00:31:43,867] loop.py:45 INFO TRAIN_STEP step=5078 loss=1.468962 Training: 28%|██▊ | 5078/18018 [11:01:29<32:13:02, 8.96s/step, loss=1.469, acc=0.673, acc_len=1.60, thru=14.0, I=13.5, T=31.6, wait=4.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:31:52,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:31:52,601] loop.py:45 INFO TRAIN_STEP step=5079 loss=1.644183 Training: 28%|██▊ | 5079/18018 [11:01:38<31:58:08, 8.89s/step, loss=1.644, acc=0.645, acc_len=1.51, thru=16.0, I=15.7, T=28.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:31:52,709] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:31:57,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:31:59,510] loop.py:380 INFO TIMING step=5080: step=4.544s data=2.158s compute=4.511s [fwd=3.248s bwd=1.247s opt=0.003s] dispatch=2.356s [2026-10-19 00:31:59,510] loop.py:45 INFO TRAIN_STEP step=5080 loss=1.519587 Training: 28%|██▊ | 5080/18018 [11:01:45<29:49:28, 8.30s/step, loss=1.520, acc=0.669, acc_len=1.58, thru=18.1, I=20.0, T=28.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:31:59,986] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:03,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:32:05,291] loop.py:45 INFO TRAIN_STEP step=5081 loss=1.497482 Training: 28%|██▊ | 5081/18018 [11:01:51<27:06:30, 7.54s/step, loss=1.497, acc=0.667, acc_len=1.58, thru=18.9, I=27.7, T=34.4, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:32:05,286] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5080: forward=2601.6ms backward=1072.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:05,459] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:11,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:32:12,450] loop.py:45 INFO TRAIN_STEP step=5082 loss=1.455601 Training: 28%|██▊ | 5082/18018 [11:01:58<26:41:26, 7.43s/step, loss=1.456, acc=0.676, acc_len=1.61, thru=17.8, I=22.9, T=30.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:12,940] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:20,834] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:32:21,386] loop.py:45 INFO TRAIN_STEP step=5083 loss=1.579533 Training: 28%|██▊ | 5083/18018 [11:02:07<28:18:55, 7.88s/step, loss=1.580, acc=0.652, acc_len=1.53, thru=13.2, I=16.6, T=28.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:21,840] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:28,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:32:30,421] loop.py:45 INFO TRAIN_STEP step=5084 loss=1.643572 Training: 28%|██▊ | 5084/18018 [11:02:16<29:33:27, 8.23s/step, loss=1.644, acc=0.647, acc_len=1.51, thru=14.0, I=18.5, T=27.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:30,632] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:37,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:32:38,426] loop.py:380 INFO TIMING step=5085: step=4.441s data=2.092s compute=4.415s [fwd=3.105s bwd=1.296s opt=0.003s] dispatch=3.555s [2026-10-19 00:32:38,426] loop.py:45 INFO TRAIN_STEP step=5085 loss=1.567975 Training: 28%|██▊ | 5085/18018 [11:02:24<29:18:58, 8.16s/step, loss=1.568, acc=0.661, acc_len=1.56, thru=13.7, I=18.1, T=28.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:38,634] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:44,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:32:45,272] loop.py:45 INFO TRAIN_STEP step=5086 loss=1.429022 Training: 28%|██▊ | 5086/18018 [11:02:31<27:53:47, 7.77s/step, loss=1.429, acc=0.680, acc_len=1.62, thru=17.6, I=18.3, T=31.7, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:32:45,268] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5085: forward=2831.9ms backward=1155.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:45,362] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-19 00:32:52,103] loop.py:45 INFO TRAIN_STEP step=5087 loss=1.546607 Training: 28%|██▊ | 5087/18018 [11:02:38<26:53:15, 7.49s/step, loss=1.547, acc=0.660, acc_len=1.56, thru=17.3, I=19.9, T=29.4, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:58,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:32:59,548] loop.py:45 INFO TRAIN_STEP step=5088 loss=1.420982 Training: 28%|██▊ | 5088/18018 [11:02:45<26:50:32, 7.47s/step, loss=1.421, acc=0.685, acc_len=1.64, thru=16.0, I=17.7, T=34.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:32:59,576] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:04,647] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:33:06,373] loop.py:45 INFO TRAIN_STEP step=5089 loss=1.629226 Training: 28%|██▊ | 5089/18018 [11:02:52<26:08:36, 7.28s/step, loss=1.629, acc=0.645, acc_len=1.51, thru=18.1, I=25.1, T=33.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:06,662] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:12,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:33:13,380] loop.py:380 INFO TIMING step=5090: step=4.125s data=1.897s compute=4.094s [fwd=2.846s bwd=1.233s opt=0.003s] dispatch=2.871s [2026-10-19 00:33:13,380] loop.py:45 INFO TRAIN_STEP step=5090 loss=1.679552 Training: 28%|██▊ | 5090/18018 [11:02:59<25:50:45, 7.20s/step, loss=1.680, acc=0.640, acc_len=1.48, thru=17.5, I=23.8, T=31.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:13,779] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:19,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:33:21,165] loop.py:45 INFO TRAIN_STEP step=5091 loss=1.602547 Training: 28%|██▊ | 5091/18018 [11:03:07<26:28:44, 7.37s/step, loss=1.603, acc=0.654, acc_len=1.53, thru=15.5, I=17.4, T=29.2, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:33:21,160] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5090: forward=3111.6ms backward=1220.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:21,236] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-19 00:33:27,909] loop.py:45 INFO TRAIN_STEP step=5092 loss=1.603862 Training: 28%|██▊ | 5092/18018 [11:03:13<25:47:42, 7.18s/step, loss=1.604, acc=0.652, acc_len=1.53, thru=15.2, I=16.4, T=35.2, wait=3.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:35,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:33:36,926] loop.py:45 INFO TRAIN_STEP step=5093 loss=1.547482 Training: 28%|██▊ | 5093/18018 [11:03:23<27:46:12, 7.73s/step, loss=1.547, acc=0.659, acc_len=1.56, thru=17.7, I=17.1, T=29.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:37,237] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:44,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:33:45,106] loop.py:45 INFO TRAIN_STEP step=5094 loss=1.614252 Training: 28%|██▊ | 5094/18018 [11:03:31<28:14:44, 7.87s/step, loss=1.614, acc=0.650, acc_len=1.52, thru=14.3, I=16.1, T=30.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:45,447] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:50,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:33:51,802] loop.py:380 INFO TIMING step=5095: step=3.728s data=1.570s compute=3.698s [fwd=2.630s bwd=1.054s opt=0.003s] dispatch=2.960s [2026-10-19 00:33:51,802] loop.py:45 INFO TRAIN_STEP step=5095 loss=1.690651 Training: 28%|██▊ | 5095/18018 [11:03:37<26:58:53, 7.52s/step, loss=1.691, acc=0.637, acc_len=1.48, thru=17.8, I=21.1, T=34.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:52,279] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:57,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:33:58,544] loop.py:45 INFO TRAIN_STEP step=5096 loss=1.541390 Training: 28%|██▊ | 5096/18018 [11:03:44<26:08:40, 7.28s/step, loss=1.541, acc=0.658, acc_len=1.55, thru=18.2, I=23.7, T=32.5, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:33:58,538] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5095: forward=2716.1ms backward=1166.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:33:59,034] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:04,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:34:05,529] loop.py:45 INFO TRAIN_STEP step=5097 loss=1.456941 Training: 28%|██▊ | 5097/18018 [11:03:51<25:49:27, 7.20s/step, loss=1.457, acc=0.675, acc_len=1.60, thru=17.2, I=21.9, T=30.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:05,625] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:09,745] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:34:11,481] loop.py:45 INFO TRAIN_STEP step=5098 loss=1.524626 Training: 28%|██▊ | 5098/18018 [11:03:57<24:29:09, 6.82s/step, loss=1.525, acc=0.663, acc_len=1.56, thru=19.2, I=24.7, T=33.3, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:11,760] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:17,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:34:17,983] loop.py:45 INFO TRAIN_STEP step=5099 loss=1.498425 Training: 28%|██▊ | 5099/18018 [11:04:04<24:07:57, 6.72s/step, loss=1.498, acc=0.671, acc_len=1.59, thru=18.1, I=23.7, T=31.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:18,284] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:24,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:34:26,240] loop.py:380 INFO TIMING step=5100: step=4.913s data=2.430s compute=4.881s [fwd=3.402s bwd=1.464s opt=0.003s] dispatch=3.337s [2026-10-19 00:34:26,241] loop.py:45 INFO TRAIN_STEP step=5100 loss=1.506229 Training: 28%|██▊ | 5100/18018 [11:04:12<25:46:58, 7.19s/step, loss=1.506, acc=0.668, acc_len=1.59, thru=14.9, I=19.6, T=26.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:26,359] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:32,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:34:32,954] loop.py:45 INFO TRAIN_STEP step=5101 loss=1.303388 Training: 28%|██▊ | 5101/18018 [11:04:19<25:16:21, 7.04s/step, loss=1.303, acc=0.699, acc_len=1.69, thru=16.5, I=21.6, T=32.1, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:34:32,950] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5100: forward=2789.4ms backward=1150.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:33,144] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:38,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:34:39,522] loop.py:45 INFO TRAIN_STEP step=5102 loss=1.555261 Training: 28%|██▊ | 5102/18018 [11:04:25<24:45:40, 6.90s/step, loss=1.555, acc=0.660, acc_len=1.56, thru=18.8, I=22.3, T=30.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:39,892] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:44,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:34:46,179] loop.py:45 INFO TRAIN_STEP step=5103 loss=1.518553 Training: 28%|██▊ | 5103/18018 [11:04:32<24:29:45, 6.83s/step, loss=1.519, acc=0.667, acc_len=1.58, thru=17.6, I=23.3, T=36.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:46,471] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:51,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:34:52,830] loop.py:45 INFO TRAIN_STEP step=5104 loss=1.591054 Training: 28%|██▊ | 5104/18018 [11:04:38<24:18:12, 6.78s/step, loss=1.591, acc=0.651, acc_len=1.52, thru=19.2, I=27.4, T=30.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:34:53,017] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 00:34:59,568] loop.py:380 INFO TIMING step=5105: step=3.812s data=1.650s compute=3.780s [fwd=2.651s bwd=1.117s opt=0.003s] dispatch=2.915s [2026-10-19 00:34:59,568] loop.py:45 INFO TRAIN_STEP step=5105 loss=1.533182 Training: 28%|██▊ | 5105/18018 [11:04:45<24:15:31, 6.76s/step, loss=1.533, acc=0.662, acc_len=1.57, thru=16.8, I=22.7, T=33.6, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:08,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:35:09,315] loop.py:45 INFO TRAIN_STEP step=5106 loss=1.628567 Training: 28%|██▊ | 5106/18018 [11:04:55<27:28:14, 7.66s/step, loss=1.629, acc=0.650, acc_len=1.53, thru=13.8, I=13.2, T=30.1, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:35:09,310] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5105: forward=2935.0ms backward=1176.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:09,532] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:14,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:35:16,347] loop.py:45 INFO TRAIN_STEP step=5107 loss=1.658431 Training: 28%|██▊ | 5107/18018 [11:05:02<26:47:27, 7.47s/step, loss=1.658, acc=0.645, acc_len=1.50, thru=17.8, I=21.3, T=30.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:16,836] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 00:35:24,337] loop.py:45 INFO TRAIN_STEP step=5108 loss=1.575443 Training: 28%|██▊ | 5108/18018 [11:05:10<27:21:04, 7.63s/step, loss=1.575, acc=0.657, acc_len=1.55, thru=14.6, I=18.6, T=29.7, wait=3.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:31,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:35:31,532] loop.py:45 INFO TRAIN_STEP step=5109 loss=1.377875 Training: 28%|██▊ | 5109/18018 [11:05:17<26:52:55, 7.50s/step, loss=1.378, acc=0.688, acc_len=1.65, thru=18.1, I=16.2, T=32.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:31,845] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:37,097] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:35:38,350] loop.py:380 INFO TIMING step=5110: step=4.135s data=1.912s compute=4.101s [fwd=2.910s bwd=1.177s opt=0.003s] dispatch=2.675s [2026-10-19 00:35:38,351] loop.py:45 INFO TRAIN_STEP step=5110 loss=1.599642 Training: 28%|██▊ | 5110/18018 [11:05:24<26:09:07, 7.29s/step, loss=1.600, acc=0.650, acc_len=1.53, thru=19.0, I=21.5, T=31.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:38,613] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:44,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:35:45,551] loop.py:45 INFO TRAIN_STEP step=5111 loss=1.582975 Training: 28%|██▊ | 5111/18018 [11:05:31<26:02:57, 7.27s/step, loss=1.583, acc=0.658, acc_len=1.54, thru=17.7, I=20.4, T=31.9, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:35:45,546] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5110: forward=2729.1ms backward=1220.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:45,973] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:51,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:35:52,333] loop.py:45 INFO TRAIN_STEP step=5112 loss=1.468546 Training: 28%|██▊ | 5112/18018 [11:05:38<25:31:32, 7.12s/step, loss=1.469, acc=0.674, acc_len=1.61, thru=18.2, I=21.2, T=30.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:52,588] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:35:58,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:35:59,846] loop.py:45 INFO TRAIN_STEP step=5113 loss=1.582479 Training: 28%|██▊ | 5113/18018 [11:05:45<25:56:48, 7.24s/step, loss=1.582, acc=0.659, acc_len=1.55, thru=16.8, I=21.6, T=31.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:00,100] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:36:08,217] loop.py:45 INFO TRAIN_STEP step=5114 loss=1.532896 Training: 28%|██▊ | 5114/18018 [11:05:54<27:09:48, 7.58s/step, loss=1.533, acc=0.661, acc_len=1.57, thru=11.8, I=17.6, T=30.5, wait=4.2s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:14,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:36:17,166] loop.py:380 INFO TIMING step=5115: step=4.233s data=2.003s compute=4.186s [fwd=3.039s bwd=1.132s opt=0.003s] dispatch=4.705s [2026-10-19 00:36:17,166] loop.py:45 INFO TRAIN_STEP step=5115 loss=1.553625 Training: 28%|██▊ | 5115/18018 [11:06:03<28:38:04, 7.99s/step, loss=1.554, acc=0.662, acc_len=1.56, thru=17.5, I=16.2, T=30.2, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:17,232] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:23,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:36:24,450] loop.py:45 INFO TRAIN_STEP step=5116 loss=1.511456 (TrainerActor pid=9310) [2026-10-19 00:36:24,445] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5115: forward=3804.1ms backward=1165.8ms Training: 28%|██▊ | 5116/18018 [11:06:10<27:52:29, 7.78s/step, loss=1.511, acc=0.667, acc_len=1.58, thru=14.3, I=20.4, T=25.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:24,698] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:31,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:36:31,838] loop.py:45 INFO TRAIN_STEP step=5117 loss=1.534412 Training: 28%|██▊ | 5117/18018 [11:06:17<27:27:17, 7.66s/step, loss=1.534, acc=0.665, acc_len=1.57, thru=16.5, I=17.1, T=28.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:32,221] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:37,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:36:38,490] loop.py:45 INFO TRAIN_STEP step=5118 loss=1.642430 Training: 28%|██▊ | 5118/18018 [11:06:24<26:21:57, 7.36s/step, loss=1.642, acc=0.648, acc_len=1.51, thru=17.3, I=21.0, T=31.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:38,697] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:45,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:36:46,392] loop.py:45 INFO TRAIN_STEP step=5119 loss=1.662651 Training: 28%|██▊ | 5119/18018 [11:06:32<26:56:55, 7.52s/step, loss=1.663, acc=0.642, acc_len=1.49, thru=14.9, I=19.2, T=27.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:46,855] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:51,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:36:53,882] loop.py:380 INFO TIMING step=5120: step=4.487s data=2.193s compute=4.450s [fwd=3.134s bwd=1.303s opt=0.003s] dispatch=2.994s [2026-10-19 00:36:53,882] loop.py:45 INFO TRAIN_STEP step=5120 loss=1.478667 Training: 28%|██▊ | 5120/18018 [11:06:39<26:54:53, 7.51s/step, loss=1.479, acc=0.674, acc_len=1.61, thru=14.4, I=18.9, T=28.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:36:53,991] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 00:37:03,243] loop.py:45 INFO TRAIN_STEP step=5121 loss=1.589384 Training: 28%|██▊ | 5121/18018 [11:06:49<28:53:59, 8.07s/step, loss=1.589, acc=0.654, acc_len=1.54, thru=11.7, I=11.7, T=26.4, wait=4.5s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:37:03,238] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5120: forward=3355.1ms backward=1443.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:09,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:37:10,870] loop.py:45 INFO TRAIN_STEP step=5122 loss=1.463085 Training: 28%|██▊ | 5122/18018 [11:06:56<28:25:23, 7.93s/step, loss=1.463, acc=0.674, acc_len=1.60, thru=17.2, I=18.0, T=30.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:11,111] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:15,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:37:17,917] loop.py:45 INFO TRAIN_STEP step=5123 loss=1.524346 Training: 28%|██▊ | 5123/18018 [11:07:04<27:28:00, 7.67s/step, loss=1.524, acc=0.665, acc_len=1.57, thru=16.0, I=26.4, T=30.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:18,352] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:25,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:37:26,441] loop.py:45 INFO TRAIN_STEP step=5124 loss=1.597673 Training: 28%|██▊ | 5124/18018 [11:07:12<28:22:58, 7.92s/step, loss=1.598, acc=0.652, acc_len=1.52, thru=15.8, I=17.8, T=31.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:26,753] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:32,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:37:33,239] loop.py:380 INFO TIMING step=5125: step=4.518s data=2.110s compute=4.489s [fwd=3.196s bwd=1.281s opt=0.003s] dispatch=2.273s [2026-10-19 00:37:33,239] loop.py:45 INFO TRAIN_STEP step=5125 loss=1.409485 Training: 28%|██▊ | 5125/18018 [11:07:19<27:10:18, 7.59s/step, loss=1.409, acc=0.683, acc_len=1.63, thru=17.6, I=21.8, T=28.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:33,563] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:40,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:37:41,971] loop.py:45 INFO TRAIN_STEP step=5126 loss=1.454290 Training: 28%|██▊ | 5126/18018 [11:07:28<28:24:05, 7.93s/step, loss=1.454, acc=0.678, acc_len=1.61, thru=13.3, I=15.6, T=27.1, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:37:41,966] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5125: forward=3180.6ms backward=1466.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:42,251] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:48,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:37:50,318] loop.py:45 INFO TRAIN_STEP step=5127 loss=1.508039 Training: 28%|██▊ | 5127/18018 [11:07:36<28:50:48, 8.06s/step, loss=1.508, acc=0.668, acc_len=1.59, thru=13.8, I=17.9, T=25.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:50,673] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:56,978] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:37:58,195] loop.py:45 INFO TRAIN_STEP step=5128 loss=1.549848 Training: 28%|██▊ | 5128/18018 [11:07:44<28:38:59, 8.00s/step, loss=1.550, acc=0.660, acc_len=1.55, thru=14.0, I=17.5, T=31.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:37:58,484] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:04,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:38:05,074] loop.py:45 INFO TRAIN_STEP step=5129 loss=1.679571 Training: 28%|██▊ | 5129/18018 [11:07:51<27:26:30, 7.66s/step, loss=1.680, acc=0.640, acc_len=1.49, thru=17.2, I=20.7, T=27.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:05,547] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:10,968] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:38:12,664] loop.py:380 INFO TIMING step=5130: step=4.468s data=2.278s compute=4.439s [fwd=3.368s bwd=1.056s opt=0.003s] dispatch=3.114s [2026-10-19 00:38:12,664] loop.py:45 INFO TRAIN_STEP step=5130 loss=1.516425 Training: 28%|██▊ | 5130/18018 [11:07:58<27:21:43, 7.64s/step, loss=1.516, acc=0.669, acc_len=1.59, thru=16.7, I=18.8, T=28.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:12,983] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:20,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:38:20,855] loop.py:45 INFO TRAIN_STEP step=5131 loss=1.600363 Training: 28%|██▊ | 5131/18018 [11:08:06<27:56:41, 7.81s/step, loss=1.600, acc=0.652, acc_len=1.52, thru=13.3, I=18.2, T=32.5, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:38:20,850] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5130: forward=2779.7ms backward=1100.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:21,356] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:29,033] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:38:31,363] loop.py:45 INFO TRAIN_STEP step=5132 loss=1.573770 Training: 28%|██▊ | 5132/18018 [11:08:17<30:50:45, 8.62s/step, loss=1.574, acc=0.657, acc_len=1.54, thru=11.8, I=15.9, T=25.2, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:31,551] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 00:38:38,833] loop.py:45 INFO TRAIN_STEP step=5133 loss=1.447121 Training: 28%|██▊ | 5133/18018 [11:08:24<29:36:45, 8.27s/step, loss=1.447, acc=0.677, acc_len=1.61, thru=10.2, I=16.1, T=31.6, wait=3.4s, pool=32, epoch=1/2][2026-10-19 00:38:48,998] loop.py:45 INFO TRAIN_STEP step=5134 loss=1.449499 Training: 28%|██▊ | 5134/18018 [11:08:35<31:38:19, 8.84s/step, loss=1.449, acc=0.681, acc_len=1.63, thru=14.4, I=12.6, T=28.9, wait=5.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:55,659] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:38:56,164] loop.py:380 INFO TIMING step=5135: step=4.435s data=2.092s compute=4.393s [fwd=3.164s bwd=1.214s opt=0.003s] dispatch=2.722s [2026-10-19 00:38:56,164] loop.py:45 INFO TRAIN_STEP step=5135 loss=1.515480 Training: 28%|██▊ | 5135/18018 [11:08:42<29:50:16, 8.34s/step, loss=1.515, acc=0.667, acc_len=1.58, thru=17.3, I=16.9, T=28.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:38:56,664] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:39:04,709] loop.py:45 INFO TRAIN_STEP step=5136 loss=1.536991 Training: 29%|██▊ | 5136/18018 [11:08:50<30:03:31, 8.40s/step, loss=1.537, acc=0.665, acc_len=1.57, thru=13.6, I=15.1, T=29.3, wait=4.2s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:39:04,704] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5135: forward=3025.7ms backward=1296.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:10,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:39:12,189] loop.py:45 INFO TRAIN_STEP step=5137 loss=1.527110 Training: 29%|██▊ | 5137/18018 [11:08:58<29:04:13, 8.12s/step, loss=1.527, acc=0.663, acc_len=1.56, thru=17.1, I=18.0, T=28.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:12,393] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:39:21,252] loop.py:45 INFO TRAIN_STEP step=5138 loss=1.655118 Training: 29%|██▊ | 5138/18018 [11:09:07<30:04:32, 8.41s/step, loss=1.655, acc=0.643, acc_len=1.50, thru=13.6, I=12.9, T=30.8, wait=4.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:26,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:39:27,870] loop.py:45 INFO TRAIN_STEP step=5139 loss=1.405309 Training: 29%|██▊ | 5139/18018 [11:09:13<28:09:08, 7.87s/step, loss=1.405, acc=0.684, acc_len=1.64, thru=18.6, I=20.7, T=32.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:28,046] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:33,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:39:34,635] loop.py:380 INFO TIMING step=5140: step=4.781s data=2.264s compute=4.746s [fwd=3.342s bwd=1.389s opt=0.003s] dispatch=1.975s [2026-10-19 00:39:34,635] loop.py:45 INFO TRAIN_STEP step=5140 loss=1.415087 Training: 29%|██▊ | 5140/18018 [11:09:20<26:58:03, 7.54s/step, loss=1.415, acc=0.680, acc_len=1.62, thru=17.5, I=24.9, T=26.8, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:34,906] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:40,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:39:40,926] loop.py:45 INFO TRAIN_STEP step=5141 loss=1.585703 Training: 29%|██▊ | 5141/18018 [11:09:27<25:37:36, 7.16s/step, loss=1.586, acc=0.658, acc_len=1.54, thru=18.0, I=22.3, T=34.1, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:39:40,921] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5140: forward=2618.2ms backward=1098.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:41,096] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:46,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:39:48,017] loop.py:45 INFO TRAIN_STEP step=5142 loss=1.428993 Training: 29%|██▊ | 5142/18018 [11:09:34<25:32:42, 7.14s/step, loss=1.429, acc=0.682, acc_len=1.63, thru=18.0, I=22.0, T=27.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:48,306] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:54,544] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:39:55,913] loop.py:45 INFO TRAIN_STEP step=5143 loss=1.413983 Training: 29%|██▊ | 5143/18018 [11:09:41<26:21:07, 7.37s/step, loss=1.414, acc=0.687, acc_len=1.64, thru=14.4, I=20.0, T=27.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:39:56,066] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:02,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:40:02,768] loop.py:45 INFO TRAIN_STEP step=5144 loss=1.671560 Training: 29%|██▊ | 5144/18018 [11:09:48<25:47:55, 7.21s/step, loss=1.672, acc=0.641, acc_len=1.49, thru=17.4, I=20.2, T=30.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:03,118] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:40:10,829] loop.py:380 INFO TIMING step=5145: step=4.199s data=1.925s compute=4.167s [fwd=2.969s bwd=1.183s opt=0.003s] dispatch=3.852s [2026-10-19 00:40:10,829] loop.py:45 INFO TRAIN_STEP step=5145 loss=1.572778 Training: 29%|██▊ | 5145/18018 [11:09:56<26:42:14, 7.47s/step, loss=1.573, acc=0.658, acc_len=1.55, thru=13.9, I=16.9, T=30.5, wait=3.9s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:20,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:40:20,703] loop.py:45 INFO TRAIN_STEP step=5146 loss=1.508962 Training: 29%|██▊ | 5146/18018 [11:10:06<29:16:56, 8.19s/step, loss=1.509, acc=0.666, acc_len=1.58, thru=15.4, I=13.5, T=30.1, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:40:20,699] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5145: forward=2879.8ms backward=1341.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:21,110] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:26,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:40:27,588] loop.py:45 INFO TRAIN_STEP step=5147 loss=1.469183 Training: 29%|██▊ | 5147/18018 [11:10:13<27:52:55, 7.80s/step, loss=1.469, acc=0.674, acc_len=1.60, thru=18.0, I=20.2, T=32.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:27,673] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:32,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:40:33,758] loop.py:45 INFO TRAIN_STEP step=5148 loss=1.473755 Training: 29%|██▊ | 5148/18018 [11:10:19<26:08:03, 7.31s/step, loss=1.474, acc=0.676, acc_len=1.61, thru=19.7, I=24.0, T=31.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:33,915] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:40,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:40:42,227] loop.py:45 INFO TRAIN_STEP step=5149 loss=1.530093 Training: 29%|██▊ | 5149/18018 [11:10:28<27:22:22, 7.66s/step, loss=1.530, acc=0.665, acc_len=1.57, thru=15.4, I=21.0, T=27.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:40:42,447] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 00:40:48,027] loop.py:380 INFO TIMING step=5150: step=3.699s data=1.616s compute=3.668s [fwd=2.628s bwd=1.025s opt=0.003s] dispatch=2.092s [2026-10-19 00:40:48,027] loop.py:45 INFO TRAIN_STEP step=5150 loss=1.485394 Training: 29%|██▊ | 5150/18018 [11:10:34<25:22:52, 7.10s/step, loss=1.485, acc=0.668, acc_len=1.59, thru=18.8, I=22.5, T=34.6, wait=2.1s, pool=60, epoch=1/2][2026-10-19 00:40:57,509] loop.py:45 INFO TRAIN_STEP step=5151 loss=1.531259 Training: 29%|██▊ | 5151/18018 [11:10:43<27:55:49, 7.81s/step, loss=1.531, acc=0.662, acc_len=1.56, thru=13.0, I=12.9, T=25.1, wait=4.4s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:40:57,505] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5150: forward=3753.6ms backward=1309.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:03,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:41:04,832] loop.py:45 INFO TRAIN_STEP step=5152 loss=1.523270 Training: 29%|██▊ | 5152/18018 [11:10:50<27:24:12, 7.67s/step, loss=1.523, acc=0.663, acc_len=1.57, thru=17.6, I=17.9, T=29.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:04,976] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:10,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:41:12,065] loop.py:45 INFO TRAIN_STEP step=5153 loss=1.647649 Training: 29%|██▊ | 5153/18018 [11:10:58<26:56:06, 7.54s/step, loss=1.648, acc=0.645, acc_len=1.50, thru=15.8, I=22.3, T=28.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:12,445] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 00:41:21,568] loop.py:45 INFO TRAIN_STEP step=5154 loss=1.478590 Training: 29%|██▊ | 5154/18018 [11:11:07<29:02:16, 8.13s/step, loss=1.479, acc=0.672, acc_len=1.60, thru=13.6, I=13.3, T=28.9, wait=5.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:27,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:41:28,495] loop.py:380 INFO TIMING step=5155: step=4.193s data=1.815s compute=4.166s [fwd=2.941s bwd=1.210s opt=0.003s] dispatch=2.726s [2026-10-19 00:41:28,496] loop.py:45 INFO TRAIN_STEP step=5155 loss=1.468444 Training: 29%|██▊ | 5155/18018 [11:11:14<27:45:04, 7.77s/step, loss=1.468, acc=0.673, acc_len=1.60, thru=18.5, I=18.2, T=30.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:28,767] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:33,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:41:35,022] loop.py:45 INFO TRAIN_STEP step=5156 loss=1.515399 Training: 29%|██▊ | 5156/18018 [11:11:21<26:25:05, 7.39s/step, loss=1.515, acc=0.669, acc_len=1.58, thru=17.0, I=20.2, T=30.3, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:41:35,017] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5155: forward=2943.8ms backward=1240.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:35,315] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:41,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:41:41,907] loop.py:45 INFO TRAIN_STEP step=5157 loss=1.539047 Training: 29%|██▊ | 5157/18018 [11:11:27<25:52:18, 7.24s/step, loss=1.539, acc=0.669, acc_len=1.59, thru=17.2, I=19.4, T=31.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:42,061] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 00:41:49,359] loop.py:45 INFO TRAIN_STEP step=5158 loss=1.327006 Training: 29%|██▊ | 5158/18018 [11:11:35<26:05:37, 7.30s/step, loss=1.327, acc=0.700, acc_len=1.69, thru=15.3, I=16.3, T=34.6, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:59,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:41:59,708] loop.py:45 INFO TRAIN_STEP step=5159 loss=1.526189 Training: 29%|██▊ | 5159/18018 [11:11:45<29:21:18, 8.22s/step, loss=1.526, acc=0.666, acc_len=1.57, thru=13.9, I=13.8, T=31.2, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:41:59,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:42:06,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:42:07,185] loop.py:380 INFO TIMING step=5160: step=3.745s data=1.656s compute=3.705s [fwd=2.590s bwd=1.101s opt=0.003s] dispatch=3.723s [2026-10-19 00:42:07,185] loop.py:45 INFO TRAIN_STEP step=5160 loss=1.568568 Training: 29%|██▊ | 5160/18018 [11:11:53<28:33:30, 8.00s/step, loss=1.569, acc=0.655, acc_len=1.54, thru=15.7, I=16.5, T=34.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:42:07,479] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:42:14,955] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:42:15,831] loop.py:45 INFO TRAIN_STEP step=5161 loss=1.619038 Training: 29%|██▊ | 5161/18018 [11:12:01<29:15:13, 8.19s/step, loss=1.619, acc=0.650, acc_len=1.52, thru=15.2, I=18.7, T=30.1, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:42:15,826] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5160: forward=2988.0ms backward=1212.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:42:16,013] inference_manager.py:408 INFO Pool capacity available after 1.1s, pool_size=24/64 [2026-10-19 00:42:22,608] loop.py:45 INFO TRAIN_STEP step=5162 loss=1.418897 Training: 29%|██▊ | 5162/18018 [11:12:08<27:44:11, 7.77s/step, loss=1.419, acc=0.681, acc_len=1.63, thru=17.6, I=20.9, T=35.4, wait=3.1s, pool=60, epoch=1/2][2026-10-19 00:42:30,253] loop.py:45 INFO TRAIN_STEP step=5163 loss=1.470003 Training: 29%|██▊ | 5163/18018 [11:12:16<27:36:21, 7.73s/step, loss=1.470, acc=0.676, acc_len=1.61, thru=16.0, I=16.6, T=29.7, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:42:39,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:42:40,215] loop.py:45 INFO TRAIN_STEP step=5164 loss=1.449971 Training: 29%|██▊ | 5164/18018 [11:12:26<29:59:24, 8.40s/step, loss=1.450, acc=0.682, acc_len=1.63, thru=13.3, I=13.8, T=30.4, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:42:40,350] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 00:42:46,758] loop.py:380 INFO TIMING step=5165: step=4.032s data=1.850s compute=4.002s [fwd=2.835s bwd=1.153s opt=0.003s] dispatch=2.502s [2026-10-19 00:42:46,758] loop.py:45 INFO TRAIN_STEP step=5165 loss=1.497701 Training: 29%|██▊ | 5165/18018 [11:12:32<28:00:03, 7.84s/step, loss=1.498, acc=0.670, acc_len=1.58, thru=18.1, I=20.9, T=31.7, wait=2.5s, pool=56, epoch=1/2][2026-10-19 00:42:57,469] loop.py:45 INFO TRAIN_STEP step=5166 loss=1.563160 Training: 29%|██▊ | 5166/18018 [11:12:43<31:04:11, 8.70s/step, loss=1.563, acc=0.660, acc_len=1.55, thru=12.7, I=11.5, T=27.0, wait=6.0s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:42:57,464] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5165: forward=3355.0ms backward=1339.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:03,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:43:04,870] loop.py:45 INFO TRAIN_STEP step=5167 loss=1.517316 Training: 29%|██▊ | 5167/18018 [11:12:50<29:40:22, 8.31s/step, loss=1.517, acc=0.668, acc_len=1.58, thru=17.7, I=16.4, T=30.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:05,211] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:43:12,429] loop.py:45 INFO TRAIN_STEP step=5168 loss=1.542123 Training: 29%|██▊ | 5168/18018 [11:12:58<28:51:55, 8.09s/step, loss=1.542, acc=0.664, acc_len=1.57, thru=14.5, I=20.9, T=29.7, wait=3.2s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:20,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:43:21,155] loop.py:45 INFO TRAIN_STEP step=5169 loss=1.530151 Training: 29%|██▊ | 5169/18018 [11:13:07<29:32:39, 8.28s/step, loss=1.530, acc=0.668, acc_len=1.58, thru=16.2, I=15.1, T=32.5, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:21,580] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:43:29,132] loop.py:380 INFO TIMING step=5170: step=3.774s data=1.575s compute=3.742s [fwd=2.624s bwd=1.102s opt=0.003s] dispatch=4.195s [2026-10-19 00:43:29,132] loop.py:45 INFO TRAIN_STEP step=5170 loss=1.448803 Training: 29%|██▊ | 5170/18018 [11:13:15<29:13:20, 8.19s/step, loss=1.449, acc=0.682, acc_len=1.63, thru=14.0, I=17.4, T=33.9, wait=4.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:36,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:43:37,561] loop.py:45 INFO TRAIN_STEP step=5171 loss=1.598769 (TrainerActor pid=9310) [2026-10-19 00:43:37,557] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5170: forward=2652.6ms backward=1161.6ms Training: 29%|██▊ | 5171/18018 [11:13:23<29:28:38, 8.26s/step, loss=1.599, acc=0.651, acc_len=1.52, thru=16.1, I=16.8, T=33.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:38,032] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:44,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:43:45,188] loop.py:45 INFO TRAIN_STEP step=5172 loss=1.452641 Training: 29%|██▊ | 5172/18018 [11:13:31<28:47:44, 8.07s/step, loss=1.453, acc=0.672, acc_len=1.59, thru=15.6, I=19.0, T=31.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:45,597] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:50,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:43:51,374] loop.py:45 INFO TRAIN_STEP step=5173 loss=1.448308 Training: 29%|██▊ | 5173/18018 [11:13:37<26:46:37, 7.50s/step, loss=1.448, acc=0.678, acc_len=1.62, thru=18.6, I=23.6, T=35.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:43:51,784] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:43:59,376] loop.py:45 INFO TRAIN_STEP step=5174 loss=1.573336 Training: 29%|██▊ | 5174/18018 [11:13:45<27:18:26, 7.65s/step, loss=1.573, acc=0.660, acc_len=1.56, thru=14.4, I=14.9, T=34.3, wait=4.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:05,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:44:07,252] loop.py:380 INFO TIMING step=5175: step=4.154s data=1.864s compute=4.115s [fwd=2.928s bwd=1.172s opt=0.003s] dispatch=3.714s [2026-10-19 00:44:07,252] loop.py:45 INFO TRAIN_STEP step=5175 loss=1.583446 Training: 29%|██▊ | 5175/18018 [11:13:53<27:32:40, 7.72s/step, loss=1.583, acc=0.657, acc_len=1.54, thru=17.4, I=18.2, T=30.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:07,432] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:13,767] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:44:15,183] loop.py:45 INFO TRAIN_STEP step=5176 loss=1.552425 Training: 29%|██▊ | 5176/18018 [11:14:01<27:46:06, 7.78s/step, loss=1.552, acc=0.663, acc_len=1.57, thru=14.2, I=19.5, T=33.0, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:44:15,178] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5175: forward=2782.0ms backward=1049.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:15,280] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:23,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:44:23,676] loop.py:45 INFO TRAIN_STEP step=5177 loss=1.578628 Training: 29%|██▊ | 5177/18018 [11:14:09<28:31:26, 8.00s/step, loss=1.579, acc=0.658, acc_len=1.55, thru=13.6, I=16.8, T=31.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:23,925] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:28,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:44:30,158] loop.py:45 INFO TRAIN_STEP step=5178 loss=1.332928 Training: 29%|██▊ | 5178/18018 [11:14:16<26:53:59, 7.54s/step, loss=1.333, acc=0.697, acc_len=1.68, thru=18.8, I=21.8, T=31.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:30,570] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:35,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:44:37,791] loop.py:45 INFO TRAIN_STEP step=5179 loss=1.573300 Training: 29%|██▊ | 5179/18018 [11:14:23<26:59:42, 7.57s/step, loss=1.573, acc=0.659, acc_len=1.55, thru=15.3, I=22.0, T=29.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:38,293] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:45,518] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:44:46,136] loop.py:380 INFO TIMING step=5180: step=3.896s data=1.678s compute=3.870s [fwd=2.800s bwd=1.056s opt=0.003s] dispatch=4.440s [2026-10-19 00:44:46,136] loop.py:45 INFO TRAIN_STEP step=5180 loss=1.611908 Training: 29%|██▊ | 5180/18018 [11:14:32<27:49:26, 7.80s/step, loss=1.612, acc=0.653, acc_len=1.53, thru=16.3, I=15.5, T=32.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:44:46,525] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:44:52,114] loop.py:45 INFO TRAIN_STEP step=5181 loss=1.592950 Training: 29%|██▉ | 5181/18018 [11:14:38<25:52:10, 7.25s/step, loss=1.593, acc=0.656, acc_len=1.54, thru=17.4, I=20.8, T=33.4, wait=2.1s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:44:52,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5180: forward=2698.1ms backward=1090.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:00,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:45:01,352] loop.py:45 INFO TRAIN_STEP step=5182 loss=1.533728 Training: 29%|██▉ | 5182/18018 [11:14:47<27:59:16, 7.85s/step, loss=1.534, acc=0.662, acc_len=1.56, thru=16.7, I=17.2, T=30.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:01,550] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:07,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:45:08,238] loop.py:45 INFO TRAIN_STEP step=5183 loss=1.585553 Training: 29%|██▉ | 5183/18018 [11:14:54<26:57:18, 7.56s/step, loss=1.586, acc=0.654, acc_len=1.53, thru=17.4, I=22.2, T=32.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:08,576] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:15,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:45:16,051] loop.py:45 INFO TRAIN_STEP step=5184 loss=1.677671 Training: 29%|██▉ | 5184/18018 [11:15:02<27:13:24, 7.64s/step, loss=1.678, acc=0.639, acc_len=1.49, thru=14.5, I=16.3, T=28.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:16,344] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:21,928] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:45:23,235] loop.py:380 INFO TIMING step=5185: step=4.303s data=1.997s compute=4.262s [fwd=2.987s bwd=1.260s opt=0.003s] dispatch=2.872s [2026-10-19 00:45:23,236] loop.py:45 INFO TRAIN_STEP step=5185 loss=1.727294 Training: 29%|██▉ | 5185/18018 [11:15:09<26:44:25, 7.50s/step, loss=1.727, acc=0.633, acc_len=1.47, thru=17.0, I=19.5, T=29.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:23,435] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:30,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:45:31,615] loop.py:45 INFO TRAIN_STEP step=5186 loss=1.530454 Training: 29%|██▉ | 5186/18018 [11:15:17<27:40:39, 7.76s/step, loss=1.530, acc=0.667, acc_len=1.58, thru=13.7, I=16.7, T=30.8, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:45:31,611] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5185: forward=2963.3ms backward=1133.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:31,828] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:38,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:45:38,734] loop.py:45 INFO TRAIN_STEP step=5187 loss=1.499920 Training: 29%|██▉ | 5187/18018 [11:15:24<26:58:54, 7.57s/step, loss=1.500, acc=0.669, acc_len=1.59, thru=17.2, I=18.8, T=32.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:39,225] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:46,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:45:47,441] loop.py:45 INFO TRAIN_STEP step=5188 loss=1.562948 Training: 29%|██▉ | 5188/18018 [11:15:33<28:11:45, 7.91s/step, loss=1.563, acc=0.661, acc_len=1.55, thru=14.7, I=17.1, T=31.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:47,833] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:53,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:45:53,622] loop.py:45 INFO TRAIN_STEP step=5189 loss=1.643299 Training: 29%|██▉ | 5189/18018 [11:15:39<26:20:33, 7.39s/step, loss=1.643, acc=0.644, acc_len=1.50, thru=19.1, I=22.2, T=32.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:53,832] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:45:59,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:46:00,593] loop.py:380 INFO TIMING step=5190: step=3.993s data=1.785s compute=3.964s [fwd=2.805s bwd=1.144s opt=0.003s] dispatch=2.969s [2026-10-19 00:46:00,593] loop.py:45 INFO TRAIN_STEP step=5190 loss=1.534910 Training: 29%|██▉ | 5190/18018 [11:15:46<25:53:24, 7.27s/step, loss=1.535, acc=0.665, acc_len=1.57, thru=17.7, I=19.7, T=32.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:01,055] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:06,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:46:07,093] loop.py:45 INFO TRAIN_STEP step=5191 loss=1.573365 Training: 29%|██▉ | 5191/18018 [11:15:53<25:04:11, 7.04s/step, loss=1.573, acc=0.656, acc_len=1.54, thru=18.6, I=22.6, T=35.5, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:46:07,087] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5190: forward=2545.4ms backward=1010.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:07,482] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:46:13,689] loop.py:45 INFO TRAIN_STEP step=5192 loss=1.447284 Training: 29%|██▉ | 5192/18018 [11:15:59<24:36:02, 6.90s/step, loss=1.447, acc=0.677, acc_len=1.62, thru=18.1, I=24.5, T=30.3, wait=2.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:20,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:46:21,513] loop.py:45 INFO TRAIN_STEP step=5193 loss=1.570334 Training: 29%|██▉ | 5193/18018 [11:16:07<25:34:44, 7.18s/step, loss=1.570, acc=0.656, acc_len=1.55, thru=15.1, I=15.6, T=28.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:21,598] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:26,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:46:27,575] loop.py:45 INFO TRAIN_STEP step=5194 loss=1.499553 Training: 29%|██▉ | 5194/18018 [11:16:13<24:22:54, 6.84s/step, loss=1.500, acc=0.668, acc_len=1.58, thru=17.8, I=21.6, T=31.9, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:27,663] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:33,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:46:34,285] loop.py:380 INFO TIMING step=5195: step=3.916s data=1.779s compute=3.886s [fwd=2.810s bwd=1.061s opt=0.003s] dispatch=2.785s [2026-10-19 00:46:34,285] loop.py:45 INFO TRAIN_STEP step=5195 loss=1.584319 Training: 29%|██▉ | 5195/18018 [11:16:20<24:14:09, 6.80s/step, loss=1.584, acc=0.657, acc_len=1.55, thru=17.6, I=21.9, T=32.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:34,515] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:39,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:46:40,494] loop.py:45 INFO TRAIN_STEP step=5196 loss=1.538415 Training: 29%|██▉ | 5196/18018 [11:16:26<23:36:00, 6.63s/step, loss=1.538, acc=0.662, acc_len=1.57, thru=18.8, I=25.3, T=32.4, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:46:40,490] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5195: forward=2721.5ms backward=1187.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:40,665] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:46:46,638] loop.py:45 INFO TRAIN_STEP step=5197 loss=1.635711 Training: 29%|██▉ | 5197/18018 [11:16:32<23:04:52, 6.48s/step, loss=1.636, acc=0.650, acc_len=1.52, thru=19.0, I=21.9, T=33.0, wait=2.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:52,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:46:54,439] loop.py:45 INFO TRAIN_STEP step=5198 loss=1.425260 Training: 29%|██▉ | 5198/18018 [11:16:40<24:29:25, 6.88s/step, loss=1.425, acc=0.679, acc_len=1.62, thru=18.6, I=19.7, T=27.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:46:54,546] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-19 00:47:01,854] loop.py:45 INFO TRAIN_STEP step=5199 loss=1.554712 Training: 29%|██▉ | 5199/18018 [11:16:47<25:03:39, 7.04s/step, loss=1.555, acc=0.658, acc_len=1.55, thru=13.7, I=18.8, T=31.5, wait=3.3s, pool=48, epoch=1/2][2026-10-19 00:47:11,828] loop.py:380 INFO TIMING step=5200: step=4.324s data=1.952s compute=4.293s [fwd=3.079s bwd=1.199s opt=0.003s] dispatch=5.642s [2026-10-19 00:47:11,828] loop.py:45 INFO TRAIN_STEP step=5200 loss=1.575409 Training: 29%|██▉ | 5200/18018 [11:16:57<28:11:53, 7.92s/step, loss=1.575, acc=0.658, acc_len=1.55, thru=14.6, I=14.1, T=29.6, wait=5.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:18,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:47:19,903] loop.py:45 INFO TRAIN_STEP step=5201 loss=1.528233 Training: 29%|██▉ | 5201/18018 [11:17:05<28:21:33, 7.97s/step, loss=1.528, acc=0.665, acc_len=1.57, thru=16.3, I=15.6, T=29.5, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:47:19,898] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5200: forward=3061.7ms backward=1218.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:20,280] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:26,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:47:27,467] loop.py:45 INFO TRAIN_STEP step=5202 loss=1.530774 Training: 29%|██▉ | 5202/18018 [11:17:13<27:55:45, 7.85s/step, loss=1.531, acc=0.666, acc_len=1.58, thru=15.6, I=20.8, T=25.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:27,805] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:33,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:47:34,784] loop.py:45 INFO TRAIN_STEP step=5203 loss=1.512641 Training: 29%|██▉ | 5203/18018 [11:17:20<27:21:52, 7.69s/step, loss=1.513, acc=0.663, acc_len=1.56, thru=15.2, I=18.7, T=30.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:35,028] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:47:41,902] loop.py:45 INFO TRAIN_STEP step=5204 loss=1.633330 Training: 29%|██▉ | 5204/18018 [11:17:27<26:45:04, 7.52s/step, loss=1.633, acc=0.647, acc_len=1.51, thru=16.3, I=19.5, T=30.1, wait=2.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:49,845] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:47:50,126] loop.py:380 INFO TIMING step=5205: step=3.838s data=1.650s compute=3.800s [fwd=2.737s bwd=1.047s opt=0.003s] dispatch=4.378s [2026-10-19 00:47:50,126] loop.py:45 INFO TRAIN_STEP step=5205 loss=1.470456 Training: 29%|██▉ | 5205/18018 [11:17:36<27:30:35, 7.73s/step, loss=1.470, acc=0.673, acc_len=1.60, thru=16.8, I=16.1, T=33.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:50,347] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:57,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:47:58,793] loop.py:45 INFO TRAIN_STEP step=5206 loss=1.567659 Training: 29%|██▉ | 5206/18018 [11:17:44<28:30:25, 8.01s/step, loss=1.568, acc=0.656, acc_len=1.54, thru=15.3, I=17.9, T=29.6, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:47:58,788] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5205: forward=3065.3ms backward=1215.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:47:59,165] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 00:48:06,994] loop.py:45 INFO TRAIN_STEP step=5207 loss=1.586145 Training: 29%|██▉ | 5207/18018 [11:17:53<28:42:24, 8.07s/step, loss=1.586, acc=0.655, acc_len=1.53, thru=13.6, I=16.9, T=28.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:06,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:07,452] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:14,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:48:15,515] loop.py:45 INFO TRAIN_STEP step=5208 loss=1.604405 Training: 29%|██▉ | 5208/18018 [11:18:01<29:11:33, 8.20s/step, loss=1.604, acc=0.652, acc_len=1.53, thru=15.6, I=16.7, T=29.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:15,905] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:24,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:48:25,459] loop.py:45 INFO TRAIN_STEP step=5209 loss=1.529134 Training: 29%|██▉ | 5209/18018 [11:18:11<31:02:44, 8.73s/step, loss=1.529, acc=0.662, acc_len=1.56, thru=13.4, I=13.2, T=29.9, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:25,488] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:37,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:48:38,009] loop.py:380 INFO TIMING step=5210: step=4.913s data=2.200s compute=4.877s [fwd=3.264s bwd=1.597s opt=0.003s] dispatch=7.627s [2026-10-19 00:48:38,009] loop.py:45 INFO TRAIN_STEP step=5210 loss=1.640549 Training: 29%|██▉ | 5210/18018 [11:18:24<35:07:39, 9.87s/step, loss=1.641, acc=0.647, acc_len=1.52, thru=9.7, I=10.6, T=26.1, wait=7.6s, pool=64, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:38,086] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:43,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:48:44,092] loop.py:45 INFO TRAIN_STEP step=5211 loss=1.327723 Training: 29%|██▉ | 5211/18018 [11:18:30<31:04:34, 8.74s/step, loss=1.328, acc=0.699, acc_len=1.69, thru=18.6, I=15.3, T=36.0, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:48:44,087] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5210: forward=2543.3ms backward=977.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:44,516] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:51,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:48:52,169] loop.py:45 INFO TRAIN_STEP step=5212 loss=1.681283 Training: 29%|██▉ | 5212/18018 [11:18:38<30:22:17, 8.54s/step, loss=1.681, acc=0.640, acc_len=1.49, thru=16.7, I=19.3, T=30.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:52,426] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:58,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:48:59,823] loop.py:45 INFO TRAIN_STEP step=5213 loss=1.624826 Training: 29%|██▉ | 5213/18018 [11:18:45<29:25:39, 8.27s/step, loss=1.625, acc=0.652, acc_len=1.52, thru=16.1, I=18.3, T=27.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:48:59,985] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:07,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:49:08,765] loop.py:45 INFO TRAIN_STEP step=5214 loss=1.550036 Training: 29%|██▉ | 5214/18018 [11:18:54<30:08:16, 8.47s/step, loss=1.550, acc=0.666, acc_len=1.57, thru=13.8, I=14.2, T=31.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:08,813] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:15,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:49:16,032] loop.py:380 INFO TIMING step=5215: step=4.277s data=1.956s compute=4.248s [fwd=2.972s bwd=1.260s opt=0.003s] dispatch=2.981s [2026-10-19 00:49:16,032] loop.py:45 INFO TRAIN_STEP step=5215 loss=1.557443 Training: 29%|██▉ | 5215/18018 [11:19:02<28:50:55, 8.11s/step, loss=1.557, acc=0.661, acc_len=1.55, thru=16.1, I=16.9, T=29.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:16,313] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:21,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:49:22,413] loop.py:45 INFO TRAIN_STEP step=5216 loss=1.520818 Training: 29%|██▉ | 5216/18018 [11:19:08<26:59:57, 7.59s/step, loss=1.521, acc=0.671, acc_len=1.58, thru=19.3, I=19.5, T=31.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:22,479] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (TrainerActor pid=9310) [2026-10-19 00:49:22,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5215: forward=2766.2ms backward=1218.6ms [2026-10-19 00:49:29,694] loop.py:45 INFO TRAIN_STEP step=5217 loss=1.630490 Training: 29%|██▉ | 5217/18018 [11:19:15<26:39:47, 7.50s/step, loss=1.630, acc=0.646, acc_len=1.51, thru=16.8, I=17.2, T=33.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:29,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:30,170] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:37,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:49:37,904] loop.py:45 INFO TRAIN_STEP step=5218 loss=1.493618 Training: 29%|██▉ | 5218/18018 [11:19:23<27:25:42, 7.71s/step, loss=1.494, acc=0.668, acc_len=1.59, thru=17.3, I=17.4, T=29.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:38,399] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:45,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:49:46,391] loop.py:45 INFO TRAIN_STEP step=5219 loss=1.446494 Training: 29%|██▉ | 5219/18018 [11:19:32<28:14:43, 7.94s/step, loss=1.446, acc=0.675, acc_len=1.61, thru=16.0, I=18.1, T=33.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:46,771] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:52,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:49:53,330] loop.py:380 INFO TIMING step=5220: step=4.411s data=1.957s compute=4.383s [fwd=3.093s bwd=1.274s opt=0.003s] dispatch=2.518s [2026-10-19 00:49:53,331] loop.py:45 INFO TRAIN_STEP step=5220 loss=1.546315 Training: 29%|██▉ | 5220/18018 [11:19:39<27:10:16, 7.64s/step, loss=1.546, acc=0.662, acc_len=1.57, thru=18.0, I=24.5, T=29.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:53,561] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:57,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:49:59,828] loop.py:45 INFO TRAIN_STEP step=5221 loss=1.592702 Training: 29%|██▉ | 5221/18018 [11:19:45<25:56:51, 7.30s/step, loss=1.593, acc=0.650, acc_len=1.52, thru=17.8, I=24.7, T=31.3, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:49:59,822] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5220: forward=2909.8ms backward=1142.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:49:59,990] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:05,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:50:06,601] loop.py:45 INFO TRAIN_STEP step=5222 loss=1.505613 Training: 29%|██▉ | 5222/18018 [11:19:52<25:23:00, 7.14s/step, loss=1.506, acc=0.665, acc_len=1.57, thru=16.7, I=22.7, T=33.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:07,074] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:50:12,997] loop.py:45 INFO TRAIN_STEP step=5223 loss=1.560956 Training: 29%|██▉ | 5223/18018 [11:19:59<24:35:02, 6.92s/step, loss=1.561, acc=0.662, acc_len=1.56, thru=19.6, I=21.1, T=36.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:12,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:13,462] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:17,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:50:19,400] loop.py:45 INFO TRAIN_STEP step=5224 loss=1.504662 Training: 29%|██▉ | 5224/18018 [11:20:05<24:02:14, 6.76s/step, loss=1.505, acc=0.666, acc_len=1.57, thru=19.7, I=25.7, T=30.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:19,631] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 00:50:25,964] loop.py:380 INFO TIMING step=5225: step=3.984s data=1.803s compute=3.956s [fwd=2.785s bwd=1.155s opt=0.003s] dispatch=2.568s [2026-10-19 00:50:25,964] loop.py:45 INFO TRAIN_STEP step=5225 loss=1.439346 Training: 29%|██▉ | 5225/18018 [11:20:12<23:49:27, 6.70s/step, loss=1.439, acc=0.675, acc_len=1.61, thru=14.7, I=21.8, T=32.1, wait=2.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:36,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:50:38,446] loop.py:45 INFO TRAIN_STEP step=5226 loss=1.470884 Training: 29%|██▉ | 5226/18018 [11:20:24<29:58:39, 8.44s/step, loss=1.471, acc=0.673, acc_len=1.60, thru=14.1, I=14.2, T=24.9, wait=7.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:50:38,441] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5225: forward=3543.4ms backward=1528.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:38,859] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:45,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:50:45,789] loop.py:45 INFO TRAIN_STEP step=5227 loss=1.469540 Training: 29%|██▉ | 5227/18018 [11:20:31<28:48:35, 8.11s/step, loss=1.470, acc=0.672, acc_len=1.60, thru=14.1, I=18.8, T=27.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:45,869] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:51,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:50:52,147] loop.py:45 INFO TRAIN_STEP step=5228 loss=1.603207 Training: 29%|██▉ | 5228/18018 [11:20:38<26:56:33, 7.58s/step, loss=1.603, acc=0.650, acc_len=1.53, thru=18.4, I=19.4, T=33.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:52,517] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:50:57,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:51:00,012] loop.py:45 INFO TRAIN_STEP step=5229 loss=1.546255 Training: 29%|██▉ | 5229/18018 [11:20:46<27:14:23, 7.67s/step, loss=1.546, acc=0.660, acc_len=1.56, thru=16.1, I=24.9, T=25.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:00,491] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:06,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:51:06,939] loop.py:380 INFO TIMING step=5230: step=3.785s data=1.661s compute=3.753s [fwd=2.600s bwd=1.141s opt=0.003s] dispatch=3.133s [2026-10-19 00:51:06,939] loop.py:45 INFO TRAIN_STEP step=5230 loss=1.568625 Training: 29%|██▉ | 5230/18018 [11:20:53<26:26:54, 7.45s/step, loss=1.569, acc=0.656, acc_len=1.54, thru=14.2, I=22.4, T=33.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:07,209] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:51:15,081] loop.py:45 INFO TRAIN_STEP step=5231 loss=1.562057 Training: 29%|██▉ | 5231/18018 [11:21:01<27:11:23, 7.65s/step, loss=1.562, acc=0.656, acc_len=1.54, thru=15.4, I=18.6, T=28.0, wait=3.6s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:51:15,076] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5230: forward=3178.6ms backward=1329.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:21,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:51:23,049] loop.py:45 INFO TRAIN_STEP step=5232 loss=1.638145 Training: 29%|██▉ | 5232/18018 [11:21:09<27:31:20, 7.75s/step, loss=1.638, acc=0.647, acc_len=1.52, thru=15.2, I=16.6, T=28.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:23,085] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:28,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:51:29,764] loop.py:45 INFO TRAIN_STEP step=5233 loss=1.467706 Training: 29%|██▉ | 5233/18018 [11:21:15<26:24:58, 7.44s/step, loss=1.468, acc=0.674, acc_len=1.60, thru=17.2, I=24.8, T=35.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:30,134] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:36,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:51:36,925] loop.py:45 INFO TRAIN_STEP step=5234 loss=1.636818 Training: 29%|██▉ | 5234/18018 [11:21:23<26:07:13, 7.36s/step, loss=1.637, acc=0.650, acc_len=1.52, thru=17.1, I=22.3, T=31.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:37,120] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:44,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:51:44,458] loop.py:380 INFO TIMING step=5235: step=3.707s data=1.591s compute=3.676s [fwd=2.639s bwd=1.022s opt=0.003s] dispatch=3.816s [2026-10-19 00:51:44,459] loop.py:45 INFO TRAIN_STEP step=5235 loss=1.621952 Training: 29%|██▉ | 5235/18018 [11:21:30<26:18:22, 7.41s/step, loss=1.622, acc=0.652, acc_len=1.53, thru=17.4, I=18.7, T=34.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:44,627] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:51,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:51:53,002] loop.py:45 INFO TRAIN_STEP step=5236 loss=1.523639 Training: 29%|██▉ | 5236/18018 [11:21:39<27:30:52, 7.75s/step, loss=1.524, acc=0.672, acc_len=1.59, thru=16.2, I=18.4, T=29.7, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:51:52,997] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5235: forward=2945.9ms backward=1296.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:51:53,320] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:51:59,200] loop.py:45 INFO TRAIN_STEP step=5237 loss=1.429784 Training: 29%|██▉ | 5237/18018 [11:21:45<25:51:27, 7.28s/step, loss=1.430, acc=0.681, acc_len=1.62, thru=18.8, I=21.7, T=35.4, wait=2.6s, pool=60, epoch=1/2][2026-10-19 00:52:07,849] loop.py:45 INFO TRAIN_STEP step=5238 loss=1.633185 Training: 29%|██▉ | 5238/18018 [11:21:53<27:18:49, 7.69s/step, loss=1.633, acc=0.648, acc_len=1.52, thru=12.2, I=12.1, T=32.9, wait=4.7s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:18,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:52:19,428] loop.py:45 INFO TRAIN_STEP step=5239 loss=1.564136 Training: 29%|██▉ | 5239/18018 [11:22:05<31:26:40, 8.86s/step, loss=1.564, acc=0.662, acc_len=1.56, thru=13.0, I=13.5, T=27.7, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:19,925] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:27,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:52:28,367] loop.py:380 INFO TIMING step=5240: step=4.976s data=2.215s compute=4.937s [fwd=3.290s bwd=1.631s opt=0.003s] dispatch=3.955s [2026-10-19 00:52:28,368] loop.py:45 INFO TRAIN_STEP step=5240 loss=1.674089 Training: 29%|██▉ | 5240/18018 [11:22:14<31:31:48, 8.88s/step, loss=1.674, acc=0.643, acc_len=1.50, thru=13.2, I=14.8, T=25.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:28,538] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:52:36,753] loop.py:45 INFO TRAIN_STEP step=5241 loss=1.627178 Training: 29%|██▉ | 5241/18018 [11:22:22<30:59:51, 8.73s/step, loss=1.627, acc=0.648, acc_len=1.52, thru=12.9, I=14.1, T=29.4, wait=4.0s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:52:36,748] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5240: forward=3087.6ms backward=1220.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:42,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:52:43,640] loop.py:45 INFO TRAIN_STEP step=5242 loss=1.480540 Training: 29%|██▉ | 5242/18018 [11:22:29<29:01:51, 8.18s/step, loss=1.481, acc=0.673, acc_len=1.60, thru=17.5, I=20.2, T=33.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:43,721] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:48,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:52:49,828] loop.py:45 INFO TRAIN_STEP step=5243 loss=1.513290 Training: 29%|██▉ | 5243/18018 [11:22:35<26:54:30, 7.58s/step, loss=1.513, acc=0.666, acc_len=1.58, thru=19.3, I=27.2, T=35.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:49,862] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:54,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:52:57,434] loop.py:45 INFO TRAIN_STEP step=5244 loss=1.475166 Training: 29%|██▉ | 5244/18018 [11:22:43<26:55:38, 7.59s/step, loss=1.475, acc=0.673, acc_len=1.60, thru=16.1, I=26.6, T=25.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:52:57,918] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:02,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:53:04,372] loop.py:380 INFO TIMING step=5245: step=3.892s data=1.719s compute=3.858s [fwd=2.659s bwd=1.185s opt=0.003s] dispatch=3.038s [2026-10-19 00:53:04,372] loop.py:45 INFO TRAIN_STEP step=5245 loss=1.585654 Training: 29%|██▉ | 5245/18018 [11:22:50<26:14:00, 7.39s/step, loss=1.586, acc=0.649, acc_len=1.52, thru=13.2, I=23.6, T=32.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:04,515] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:09,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:53:11,140] loop.py:45 INFO TRAIN_STEP step=5246 loss=1.484975 Training: 29%|██▉ | 5246/18018 [11:22:57<25:34:04, 7.21s/step, loss=1.485, acc=0.676, acc_len=1.61, thru=17.7, I=26.3, T=28.9, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:53:11,135] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5245: forward=3148.9ms backward=1231.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:11,317] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:16,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:53:18,593] loop.py:45 INFO TRAIN_STEP step=5247 loss=1.434572 Training: 29%|██▉ | 5247/18018 [11:23:04<25:49:31, 7.28s/step, loss=1.435, acc=0.677, acc_len=1.62, thru=14.4, I=20.1, T=28.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:18,821] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:23,903] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:53:25,471] loop.py:45 INFO TRAIN_STEP step=5248 loss=1.543345 Training: 29%|██▉ | 5248/18018 [11:23:11<25:23:42, 7.16s/step, loss=1.543, acc=0.658, acc_len=1.54, thru=15.7, I=23.1, T=32.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:25,916] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:31,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:53:31,884] loop.py:45 INFO TRAIN_STEP step=5249 loss=1.580992 Training: 29%|██▉ | 5249/18018 [11:23:17<24:35:58, 6.94s/step, loss=1.581, acc=0.656, acc_len=1.54, thru=18.3, I=25.9, T=30.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:32,093] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:38,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:53:39,303] loop.py:380 INFO TIMING step=5250: step=3.895s data=1.728s compute=3.869s [fwd=2.742s bwd=1.113s opt=0.003s] dispatch=3.515s [2026-10-19 00:53:39,304] loop.py:45 INFO TRAIN_STEP step=5250 loss=1.570975 Training: 29%|██▉ | 5250/18018 [11:23:25<25:06:53, 7.08s/step, loss=1.571, acc=0.657, acc_len=1.55, thru=15.3, I=16.7, T=32.9, wait=3.5s, pool=64, epoch=1/2][2026-10-19 00:53:39,308] loop.py:419 INFO Saving checkpoint at step 5250... (TrainerActor pid=9310) [2026-10-19 00:53:42,743] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0005251 (TrainerActor pid=9310) [2026-10-19 00:53:42,779] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0004501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:43,470] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:43,972] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=0/64 [2026-10-19 00:53:50,673] loop.py:45 INFO TRAIN_STEP step=5251 loss=1.489064 Training: 29%|██▉ | 5251/18018 [11:23:36<29:40:22, 8.37s/step, loss=1.489, acc=0.670, acc_len=1.59, thru=17.3, I=22.8, T=28.7, wait=2.5s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:53:50,668] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5250: forward=2967.5ms backward=1444.4ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 00:53:42,744] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0005251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 00:53:43,730] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0004501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:53:59,885] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:54:01,138] loop.py:45 INFO TRAIN_STEP step=5252 loss=1.516999 Training: 29%|██▉ | 5252/18018 [11:23:47<31:54:10, 9.00s/step, loss=1.517, acc=0.666, acc_len=1.57, thru=16.0, I=13.2, T=29.3, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:01,392] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:09,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:54:11,086] loop.py:45 INFO TRAIN_STEP step=5253 loss=1.480906 Training: 29%|██▉ | 5253/18018 [11:23:57<32:54:50, 9.28s/step, loss=1.481, acc=0.673, acc_len=1.60, thru=13.2, I=13.5, T=28.8, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:11,380] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:16,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:54:17,154] loop.py:45 INFO TRAIN_STEP step=5254 loss=1.642345 Training: 29%|██▉ | 5254/18018 [11:24:03<29:29:28, 8.32s/step, loss=1.642, acc=0.647, acc_len=1.51, thru=19.3, I=20.7, T=31.8, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:17,485] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:23,173] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:54:24,851] loop.py:380 INFO TIMING step=5255: step=3.835s data=1.648s compute=3.805s [fwd=2.702s bwd=1.090s opt=0.003s] dispatch=3.853s [2026-10-19 00:54:24,851] loop.py:45 INFO TRAIN_STEP step=5255 loss=1.482389 Training: 29%|██▉ | 5255/18018 [11:24:10<28:49:37, 8.13s/step, loss=1.482, acc=0.670, acc_len=1.59, thru=16.1, I=18.5, T=33.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:25,177] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:29,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:54:31,096] loop.py:45 INFO TRAIN_STEP step=5256 loss=1.462472 Training: 29%|██▉ | 5256/18018 [11:24:17<26:49:18, 7.57s/step, loss=1.462, acc=0.674, acc_len=1.60, thru=19.7, I=26.6, T=30.4, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:54:31,091] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5255: forward=3041.3ms backward=1126.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:31,241] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:36,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:54:38,492] loop.py:45 INFO TRAIN_STEP step=5257 loss=1.540744 Training: 29%|██▉ | 5257/18018 [11:24:24<26:38:19, 7.52s/step, loss=1.541, acc=0.659, acc_len=1.55, thru=14.9, I=20.3, T=28.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:38,741] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:42,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:54:44,965] loop.py:45 INFO TRAIN_STEP step=5258 loss=1.587731 Training: 29%|██▉ | 5258/18018 [11:24:31<25:31:38, 7.20s/step, loss=1.588, acc=0.654, acc_len=1.54, thru=17.1, I=25.3, T=29.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:45,011] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:51,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:54:52,395] loop.py:45 INFO TRAIN_STEP step=5259 loss=1.630012 Training: 29%|██▉ | 5259/18018 [11:24:38<25:46:04, 7.27s/step, loss=1.630, acc=0.647, acc_len=1.51, thru=14.9, I=21.1, T=29.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:54:52,842] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:03,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:55:04,555] loop.py:380 INFO TIMING step=5260: step=4.952s data=2.561s compute=4.912s [fwd=3.626s bwd=1.271s opt=0.003s] dispatch=7.199s [2026-10-19 00:55:04,555] loop.py:45 INFO TRAIN_STEP step=5260 loss=1.546004 Training: 29%|██▉ | 5260/18018 [11:24:50<30:57:56, 8.74s/step, loss=1.546, acc=0.664, acc_len=1.57, thru=11.2, I=10.9, T=25.8, wait=7.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:04,665] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-19 00:55:09,816] loop.py:45 INFO TRAIN_STEP step=5261 loss=1.563173 Training: 29%|██▉ | 5261/18018 [11:24:55<27:16:00, 7.69s/step, loss=1.563, acc=0.655, acc_len=1.54, thru=19.0, I=23.7, T=35.7, wait=1.7s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:55:09,810] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5260: forward=2515.6ms backward=1030.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:17,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:55:17,613] loop.py:45 INFO TRAIN_STEP step=5262 loss=1.540739 Training: 29%|██▉ | 5262/18018 [11:25:03<27:22:14, 7.72s/step, loss=1.541, acc=0.662, acc_len=1.57, thru=16.0, I=15.6, T=31.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:18,032] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 00:55:24,304] loop.py:45 INFO TRAIN_STEP step=5263 loss=1.433873 Training: 29%|██▉ | 5263/18018 [11:25:10<26:16:20, 7.42s/step, loss=1.434, acc=0.682, acc_len=1.63, thru=18.2, I=19.2, T=35.5, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:28,947] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:55:30,860] loop.py:45 INFO TRAIN_STEP step=5264 loss=1.484476 Training: 29%|██▉ | 5264/18018 [11:25:16<25:21:27, 7.16s/step, loss=1.484, acc=0.676, acc_len=1.61, thru=19.3, I=23.2, T=30.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:30,972] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:38,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:55:39,519] loop.py:380 INFO TIMING step=5265: step=4.095s data=1.836s compute=4.069s [fwd=2.894s bwd=1.159s opt=0.003s] dispatch=4.554s [2026-10-19 00:55:39,519] loop.py:45 INFO TRAIN_STEP step=5265 loss=1.512415 Training: 29%|██▉ | 5265/18018 [11:25:25<26:56:56, 7.61s/step, loss=1.512, acc=0.666, acc_len=1.58, thru=15.0, I=15.6, T=31.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:39,986] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:55:49,704] loop.py:45 INFO TRAIN_STEP step=5266 loss=1.397120 Training: 29%|██▉ | 5266/18018 [11:25:35<29:41:13, 8.38s/step, loss=1.397, acc=0.693, acc_len=1.66, thru=11.9, I=11.4, T=29.8, wait=5.9s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:55:49,699] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5265: forward=3047.8ms backward=1202.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:55,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:55:56,846] loop.py:45 INFO TRAIN_STEP step=5267 loss=1.450022 Training: 29%|██▉ | 5267/18018 [11:25:42<28:21:58, 8.01s/step, loss=1.450, acc=0.678, acc_len=1.62, thru=19.4, I=18.6, T=33.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:55:57,344] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:03,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:56:04,613] loop.py:45 INFO TRAIN_STEP step=5268 loss=1.670235 Training: 29%|██▉ | 5268/18018 [11:25:50<28:06:30, 7.94s/step, loss=1.670, acc=0.642, acc_len=1.50, thru=15.2, I=20.1, T=31.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:04,780] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:56:11,135] loop.py:45 INFO TRAIN_STEP step=5269 loss=1.641623 Training: 29%|██▉ | 5269/18018 [11:25:57<26:36:18, 7.51s/step, loss=1.642, acc=0.644, acc_len=1.50, thru=17.8, I=19.6, T=33.0, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:16,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:56:18,618] loop.py:380 INFO TIMING step=5270: step=4.375s data=2.069s compute=4.337s [fwd=3.034s bwd=1.287s opt=0.003s] dispatch=3.098s [2026-10-19 00:56:18,618] loop.py:45 INFO TRAIN_STEP step=5270 loss=1.561989 Training: 29%|██▉ | 5270/18018 [11:26:04<26:34:14, 7.50s/step, loss=1.562, acc=0.662, acc_len=1.56, thru=17.2, I=20.2, T=29.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:18,941] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 00:56:26,107] loop.py:45 INFO TRAIN_STEP step=5271 loss=1.619448 Training: 29%|██▉ | 5271/18018 [11:26:12<26:33:08, 7.50s/step, loss=1.619, acc=0.653, acc_len=1.53, thru=12.9, I=17.5, T=30.7, wait=3.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:56:26,101] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5270: forward=2916.5ms backward=1204.5ms [2026-10-19 00:56:35,568] loop.py:45 INFO TRAIN_STEP step=5272 loss=1.580424 Training: 29%|██▉ | 5272/18018 [11:26:21<28:38:11, 8.09s/step, loss=1.580, acc=0.655, acc_len=1.55, thru=12.9, I=13.2, T=30.6, wait=5.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:42,484] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:56:44,769] loop.py:45 INFO TRAIN_STEP step=5273 loss=1.463311 Training: 29%|██▉ | 5273/18018 [11:26:30<29:49:00, 8.42s/step, loss=1.463, acc=0.677, acc_len=1.61, thru=16.4, I=18.0, T=25.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:45,004] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:50,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:56:52,048] loop.py:45 INFO TRAIN_STEP step=5274 loss=1.507236 Training: 29%|██▉ | 5274/18018 [11:26:38<28:35:55, 8.08s/step, loss=1.507, acc=0.672, acc_len=1.60, thru=14.5, I=23.1, T=29.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:52,394] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:57,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:56:58,613] loop.py:380 INFO TIMING step=5275: step=4.064s data=1.808s compute=4.035s [fwd=2.834s bwd=1.187s opt=0.003s] dispatch=2.491s [2026-10-19 00:56:58,614] loop.py:45 INFO TRAIN_STEP step=5275 loss=1.414947 Training: 29%|██▉ | 5275/18018 [11:26:44<26:59:26, 7.63s/step, loss=1.415, acc=0.684, acc_len=1.64, thru=19.1, I=24.0, T=31.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:56:58,774] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:57:04,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:57:05,361] loop.py:45 INFO TRAIN_STEP step=5276 loss=1.515152 Training: 29%|██▉ | 5276/18018 [11:26:51<26:03:13, 7.36s/step, loss=1.515, acc=0.662, acc_len=1.57, thru=17.6, I=21.2, T=32.6, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:57:05,356] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5275: forward=2693.4ms backward=1187.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:57:05,437] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-19 00:57:13,372] loop.py:45 INFO TRAIN_STEP step=5277 loss=1.516430 Training: 29%|██▉ | 5277/18018 [11:26:59<26:44:34, 7.56s/step, loss=1.516, acc=0.668, acc_len=1.58, thru=13.7, I=14.4, T=30.7, wait=3.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:57:22,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:57:22,509] loop.py:45 INFO TRAIN_STEP step=5278 loss=1.489108 Training: 29%|██▉ | 5278/18018 [11:27:08<28:25:12, 8.03s/step, loss=1.489, acc=0.673, acc_len=1.59, thru=15.2, I=14.9, T=32.0, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:57:22,746] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:57:28,140] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:57:29,319] loop.py:45 INFO TRAIN_STEP step=5279 loss=1.544018 Training: 29%|██▉ | 5279/18018 [11:27:15<27:07:22, 7.66s/step, loss=1.544, acc=0.660, acc_len=1.56, thru=18.7, I=20.0, T=32.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:57:29,651] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 00:57:36,321] loop.py:380 INFO TIMING step=5280: step=4.259s data=1.930s compute=4.231s [fwd=3.041s bwd=1.175s opt=0.003s] dispatch=2.733s [2026-10-19 00:57:36,321] loop.py:45 INFO TRAIN_STEP step=5280 loss=1.511467 Training: 29%|██▉ | 5280/18018 [11:27:22<26:24:50, 7.47s/step, loss=1.511, acc=0.663, acc_len=1.57, thru=14.3, I=16.3, T=30.1, wait=2.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:57:44,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:57:44,227] loop.py:45 INFO TRAIN_STEP step=5281 loss=1.583902 Training: 29%|██▉ | 5281/18018 [11:27:30<26:52:52, 7.60s/step, loss=1.584, acc=0.652, acc_len=1.53, thru=17.6, I=16.6, T=30.8, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:57:44,223] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5280: forward=2915.4ms backward=1195.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:57:44,593] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 00:57:51,512] loop.py:45 INFO TRAIN_STEP step=5282 loss=1.572820 Training: 29%|██▉ | 5282/18018 [11:27:37<26:33:09, 7.51s/step, loss=1.573, acc=0.662, acc_len=1.57, thru=15.6, I=17.1, T=30.2, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:00,119] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:58:02,419] loop.py:45 INFO TRAIN_STEP step=5283 loss=1.489280 Training: 29%|██▉ | 5283/18018 [11:27:48<30:09:17, 8.52s/step, loss=1.489, acc=0.670, acc_len=1.60, thru=14.0, I=14.6, T=29.3, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:02,639] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:07,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:58:08,086] loop.py:45 INFO TRAIN_STEP step=5284 loss=1.502880 Training: 29%|██▉ | 5284/18018 [11:27:54<27:07:21, 7.67s/step, loss=1.503, acc=0.670, acc_len=1.59, thru=19.6, I=28.7, T=34.2, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:08,439] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:18,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:58:18,926] loop.py:380 INFO TIMING step=5285: step=4.701s data=2.101s compute=4.622s [fwd=3.261s bwd=1.345s opt=0.003s] dispatch=6.127s [2026-10-19 00:58:18,927] loop.py:45 INFO TRAIN_STEP step=5285 loss=1.577035 Training: 29%|██▉ | 5285/18018 [11:28:05<30:29:04, 8.62s/step, loss=1.577, acc=0.660, acc_len=1.55, thru=12.8, I=12.0, T=27.2, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:19,126] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:25,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:58:28,592] loop.py:45 INFO TRAIN_STEP step=5286 loss=1.677474 Training: 29%|██▉ | 5286/18018 [11:28:14<31:35:42, 8.93s/step, loss=1.677, acc=0.640, acc_len=1.49, thru=13.5, I=17.1, T=25.2, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:58:28,587] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5285: forward=3457.2ms backward=1539.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:28,854] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:34,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:58:36,065] loop.py:45 INFO TRAIN_STEP step=5287 loss=1.430673 Training: 29%|██▉ | 5287/18018 [11:28:22<30:02:27, 8.49s/step, loss=1.431, acc=0.680, acc_len=1.62, thru=17.8, I=20.3, T=31.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:36,531] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 00:58:44,630] loop.py:45 INFO TRAIN_STEP step=5288 loss=1.536175 Training: 29%|██▉ | 5288/18018 [11:28:30<30:06:55, 8.52s/step, loss=1.536, acc=0.664, acc_len=1.56, thru=15.3, I=20.1, T=30.8, wait=4.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:50,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:58:51,927] loop.py:45 INFO TRAIN_STEP step=5289 loss=1.586300 Training: 29%|██▉ | 5289/18018 [11:28:38<28:48:56, 8.15s/step, loss=1.586, acc=0.657, acc_len=1.54, thru=16.2, I=17.6, T=30.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:58:51,979] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-19 00:58:59,346] loop.py:380 INFO TIMING step=5290: step=4.571s data=2.207s compute=4.533s [fwd=3.145s bwd=1.372s opt=0.003s] dispatch=2.840s [2026-10-19 00:58:59,346] loop.py:45 INFO TRAIN_STEP step=5290 loss=1.537665 Training: 29%|██▉ | 5290/18018 [11:28:45<28:02:28, 7.93s/step, loss=1.538, acc=0.663, acc_len=1.56, thru=13.8, I=16.6, T=28.0, wait=2.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:08,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:59:09,250] loop.py:45 INFO TRAIN_STEP step=5291 loss=1.458630 Training: 29%|██▉ | 5291/18018 [11:28:55<30:07:52, 8.52s/step, loss=1.459, acc=0.676, acc_len=1.61, thru=14.7, I=14.6, T=29.2, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:59:09,245] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5290: forward=2983.0ms backward=1317.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:09,710] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:15,979] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:59:16,438] loop.py:45 INFO TRAIN_STEP step=5292 loss=1.473575 Training: 29%|██▉ | 5292/18018 [11:29:02<28:42:46, 8.12s/step, loss=1.474, acc=0.676, acc_len=1.61, thru=18.0, I=19.0, T=27.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:16,493] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-19 00:59:23,867] loop.py:45 INFO TRAIN_STEP step=5293 loss=1.551216 Training: 29%|██▉ | 5293/18018 [11:29:09<27:58:26, 7.91s/step, loss=1.551, acc=0.662, acc_len=1.57, thru=15.0, I=15.9, T=29.1, wait=3.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:31,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:59:33,429] loop.py:45 INFO TRAIN_STEP step=5294 loss=1.514170 Training: 29%|██▉ | 5294/18018 [11:29:19<29:43:06, 8.41s/step, loss=1.514, acc=0.667, acc_len=1.58, thru=15.0, I=16.7, T=26.9, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:33,494] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:39,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:59:41,133] loop.py:380 INFO TIMING step=5295: step=4.141s data=1.898s compute=4.104s [fwd=2.882s bwd=1.206s opt=0.003s] dispatch=3.554s [2026-10-19 00:59:41,133] loop.py:45 INFO TRAIN_STEP step=5295 loss=1.468839 Training: 29%|██▉ | 5295/18018 [11:29:27<28:58:05, 8.20s/step, loss=1.469, acc=0.677, acc_len=1.62, thru=14.1, I=18.0, T=30.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:41,351] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:46,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 00:59:47,192] loop.py:45 INFO TRAIN_STEP step=5296 loss=1.488933 Training: 29%|██▉ | 5296/18018 [11:29:33<26:42:00, 7.56s/step, loss=1.489, acc=0.671, acc_len=1.58, thru=19.3, I=21.5, T=34.6, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 00:59:47,187] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5295: forward=2611.1ms backward=1052.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 00:59:47,661] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 00:59:54,883] loop.py:45 INFO TRAIN_STEP step=5297 loss=1.548109 Training: 29%|██▉ | 5297/18018 [11:29:40<26:50:42, 7.60s/step, loss=1.548, acc=0.664, acc_len=1.57, thru=14.9, I=20.4, T=29.2, wait=3.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:01,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:00:03,605] loop.py:45 INFO TRAIN_STEP step=5298 loss=1.469472 Training: 29%|██▉ | 5298/18018 [11:29:49<28:02:01, 7.93s/step, loss=1.469, acc=0.677, acc_len=1.61, thru=17.5, I=16.3, T=29.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:03,711] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:09,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:00:09,881] loop.py:45 INFO TRAIN_STEP step=5299 loss=1.445822 Training: 29%|██▉ | 5299/18018 [11:29:55<26:16:31, 7.44s/step, loss=1.446, acc=0.676, acc_len=1.61, thru=17.0, I=22.1, T=32.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:09,987] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-19 01:00:15,839] loop.py:380 INFO TIMING step=5300: step=3.689s data=1.652s compute=3.646s [fwd=2.582s bwd=1.049s opt=0.003s] dispatch=2.258s [2026-10-19 01:00:15,839] loop.py:45 INFO TRAIN_STEP step=5300 loss=1.675596 Training: 29%|██▉ | 5300/18018 [11:30:01<24:42:14, 6.99s/step, loss=1.676, acc=0.634, acc_len=1.47, thru=18.8, I=20.2, T=34.7, wait=2.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:21,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:00:23,614] loop.py:45 INFO TRAIN_STEP step=5301 loss=1.505428 Training: 29%|██▉ | 5301/18018 [11:30:09<25:32:00, 7.23s/step, loss=1.505, acc=0.666, acc_len=1.57, thru=16.8, I=20.1, T=26.8, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:00:23,608] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5300: forward=3288.9ms backward=1433.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:23,989] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:30,377] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:00:30,659] loop.py:45 INFO TRAIN_STEP step=5302 loss=1.483410 Training: 29%|██▉ | 5302/18018 [11:30:16<25:20:05, 7.17s/step, loss=1.483, acc=0.673, acc_len=1.60, thru=14.4, I=19.0, T=31.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:30,878] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:36,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:00:38,496] loop.py:45 INFO TRAIN_STEP step=5303 loss=1.566538 Training: 29%|██▉ | 5303/18018 [11:30:24<26:02:18, 7.37s/step, loss=1.567, acc=0.657, acc_len=1.55, thru=16.8, I=20.5, T=27.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:38,934] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:44,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:00:45,055] loop.py:45 INFO TRAIN_STEP step=5304 loss=1.403237 Training: 29%|██▉ | 5304/18018 [11:30:31<25:10:23, 7.13s/step, loss=1.403, acc=0.688, acc_len=1.65, thru=15.6, I=23.5, T=33.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:45,398] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:54,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:00:56,200] loop.py:380 INFO TIMING step=5305: step=5.307s data=2.595s compute=5.245s [fwd=3.704s bwd=1.525s opt=0.003s] dispatch=5.829s [2026-10-19 01:00:56,200] loop.py:45 INFO TRAIN_STEP step=5305 loss=1.571411 Training: 29%|██▉ | 5305/18018 [11:30:42<29:25:36, 8.33s/step, loss=1.571, acc=0.659, acc_len=1.55, thru=13.1, I=13.4, T=24.1, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:00:56,609] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:03,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:01:04,411] loop.py:45 INFO TRAIN_STEP step=5306 loss=1.632219 Training: 29%|██▉ | 5306/18018 [11:30:50<29:17:38, 8.30s/step, loss=1.632, acc=0.646, acc_len=1.51, thru=16.4, I=17.1, T=28.1, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:01:04,406] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5305: forward=3209.4ms backward=1290.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:04,894] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:01:10,840] loop.py:45 INFO TRAIN_STEP step=5307 loss=1.451117 Training: 29%|██▉ | 5307/18018 [11:30:56<27:19:00, 7.74s/step, loss=1.451, acc=0.678, acc_len=1.61, thru=17.7, I=19.9, T=33.6, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:20,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:01:20,652] loop.py:45 INFO TRAIN_STEP step=5308 loss=1.473407 Training: 29%|██▉ | 5308/18018 [11:31:06<29:30:39, 8.36s/step, loss=1.473, acc=0.676, acc_len=1.61, thru=14.1, I=13.4, T=32.7, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:20,731] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:26,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:01:26,572] loop.py:45 INFO TRAIN_STEP step=5309 loss=1.685992 Training: 29%|██▉ | 5309/18018 [11:31:12<26:55:40, 7.63s/step, loss=1.686, acc=0.641, acc_len=1.49, thru=20.1, I=22.2, T=32.9, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:26,754] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:31,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:01:33,485] loop.py:380 INFO TIMING step=5310: step=4.157s data=1.815s compute=4.127s [fwd=2.890s bwd=1.222s opt=0.003s] dispatch=2.746s [2026-10-19 01:01:33,485] loop.py:45 INFO TRAIN_STEP step=5310 loss=1.423823 Training: 29%|██▉ | 5310/18018 [11:31:19<26:10:10, 7.41s/step, loss=1.424, acc=0.680, acc_len=1.63, thru=18.1, I=21.5, T=30.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:33,956] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:40,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:01:42,012] loop.py:45 INFO TRAIN_STEP step=5311 loss=1.548935 Training: 29%|██▉ | 5311/18018 [11:31:28<27:20:41, 7.75s/step, loss=1.549, acc=0.662, acc_len=1.56, thru=15.9, I=17.9, T=28.8, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:01:42,007] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5310: forward=3076.2ms backward=1323.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:42,391] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:47,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:01:48,962] loop.py:45 INFO TRAIN_STEP step=5312 loss=1.565106 Training: 29%|██▉ | 5312/18018 [11:31:35<26:29:56, 7.51s/step, loss=1.565, acc=0.661, acc_len=1.56, thru=17.1, I=22.4, T=30.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:49,141] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:54,979] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:01:55,267] loop.py:45 INFO TRAIN_STEP step=5313 loss=1.556478 Training: 29%|██▉ | 5313/18018 [11:31:41<25:13:27, 7.15s/step, loss=1.556, acc=0.660, acc_len=1.56, thru=17.7, I=24.1, T=31.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:01:55,481] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:02:02,077] loop.py:45 INFO TRAIN_STEP step=5314 loss=1.572906 Training: 29%|██▉ | 5314/18018 [11:31:48<24:51:51, 7.05s/step, loss=1.573, acc=0.656, acc_len=1.55, thru=14.6, I=16.5, T=31.2, wait=2.7s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:10,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:02:11,840] loop.py:380 INFO TIMING step=5315: step=4.440s data=1.991s compute=4.392s [fwd=3.000s bwd=1.377s opt=0.003s] dispatch=5.314s [2026-10-19 01:02:11,841] loop.py:45 INFO TRAIN_STEP step=5315 loss=1.499381 Training: 29%|██▉ | 5315/18018 [11:31:57<27:44:22, 7.86s/step, loss=1.499, acc=0.667, acc_len=1.58, thru=16.4, I=16.6, T=28.8, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:12,041] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:02:18,423] loop.py:45 INFO TRAIN_STEP step=5316 loss=1.587303 Training: 30%|██▉ | 5316/18018 [11:32:04<26:22:54, 7.48s/step, loss=1.587, acc=0.649, acc_len=1.52, thru=15.1, I=16.5, T=32.2, wait=2.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:02:18,419] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5315: forward=2753.5ms backward=1176.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:25,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:02:26,385] loop.py:45 INFO TRAIN_STEP step=5317 loss=1.410379 Training: 30%|██▉ | 5317/18018 [11:32:12<26:53:47, 7.62s/step, loss=1.410, acc=0.686, acc_len=1.64, thru=16.7, I=17.7, T=27.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:26,456] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:34,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:02:35,474] loop.py:45 INFO TRAIN_STEP step=5318 loss=1.414581 Training: 30%|██▉ | 5318/18018 [11:32:21<28:26:39, 8.06s/step, loss=1.415, acc=0.682, acc_len=1.63, thru=14.5, I=14.8, T=26.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:35,810] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:40,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:02:41,167] loop.py:45 INFO TRAIN_STEP step=5319 loss=1.502130 Training: 30%|██▉ | 5319/18018 [11:32:27<25:55:53, 7.35s/step, loss=1.502, acc=0.669, acc_len=1.58, thru=20.1, I=20.8, T=37.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:41,578] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:46,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:02:47,115] loop.py:380 INFO TIMING step=5320: step=3.655s data=1.594s compute=3.628s [fwd=2.561s bwd=1.052s opt=0.003s] dispatch=2.285s [2026-10-19 01:02:47,115] loop.py:45 INFO TRAIN_STEP step=5320 loss=1.502436 Training: 30%|██▉ | 5320/18018 [11:32:33<24:26:50, 6.93s/step, loss=1.502, acc=0.665, acc_len=1.58, thru=21.7, I=24.3, T=35.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:47,207] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:51,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:02:53,422] loop.py:45 INFO TRAIN_STEP step=5321 loss=1.598878 Training: 30%|██▉ | 5321/18018 [11:32:39<23:46:59, 6.74s/step, loss=1.599, acc=0.653, acc_len=1.54, thru=19.6, I=24.0, T=31.5, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:02:53,417] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5320: forward=2784.5ms backward=1231.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:02:53,913] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:00,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:03:01,656] loop.py:45 INFO TRAIN_STEP step=5322 loss=1.544851 Training: 30%|██▉ | 5322/18018 [11:32:47<25:21:39, 7.19s/step, loss=1.545, acc=0.665, acc_len=1.57, thru=13.8, I=19.4, T=33.1, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:01,713] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 01:03:08,358] loop.py:45 INFO TRAIN_STEP step=5323 loss=1.494377 Training: 30%|██▉ | 5323/18018 [11:32:54<24:50:39, 7.05s/step, loss=1.494, acc=0.670, acc_len=1.59, thru=17.3, I=22.0, T=35.2, wait=3.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:19,092] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:03:19,558] loop.py:45 INFO TRAIN_STEP step=5324 loss=1.511255 Training: 30%|██▉ | 5324/18018 [11:33:05<29:13:59, 8.29s/step, loss=1.511, acc=0.665, acc_len=1.57, thru=13.4, I=12.4, T=28.7, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:19,609] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:24,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:03:25,611] loop.py:380 INFO TIMING step=5325: step=3.891s data=1.732s compute=3.866s [fwd=2.778s bwd=1.073s opt=0.003s] dispatch=2.152s [2026-10-19 01:03:25,612] loop.py:45 INFO TRAIN_STEP step=5325 loss=1.458773 Training: 30%|██▉ | 5325/18018 [11:33:11<26:51:49, 7.62s/step, loss=1.459, acc=0.676, acc_len=1.61, thru=18.7, I=20.5, T=32.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:25,825] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:03:33,452] loop.py:45 INFO TRAIN_STEP step=5326 loss=1.436826 Training: 30%|██▉ | 5326/18018 [11:33:19<27:05:46, 7.69s/step, loss=1.437, acc=0.676, acc_len=1.61, thru=15.8, I=16.1, T=29.7, wait=3.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:03:33,445] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5325: forward=3000.9ms backward=1248.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:38,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:03:39,759] loop.py:45 INFO TRAIN_STEP step=5327 loss=1.392448 Training: 30%|██▉ | 5327/18018 [11:33:25<25:38:05, 7.27s/step, loss=1.392, acc=0.686, acc_len=1.64, thru=18.9, I=19.7, T=32.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:40,022] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:46,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:03:48,391] loop.py:45 INFO TRAIN_STEP step=5328 loss=1.637325 Training: 30%|██▉ | 5328/18018 [11:33:34<27:04:26, 7.68s/step, loss=1.637, acc=0.649, acc_len=1.52, thru=13.6, I=17.8, T=26.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:48,606] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:53,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:03:55,266] loop.py:45 INFO TRAIN_STEP step=5329 loss=1.579377 Training: 30%|██▉ | 5329/18018 [11:33:41<26:13:05, 7.44s/step, loss=1.579, acc=0.654, acc_len=1.54, thru=15.9, I=21.9, T=31.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:03:55,643] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:02,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:04:03,018] loop.py:380 INFO TIMING step=5330: step=4.409s data=1.995s compute=4.357s [fwd=3.174s bwd=1.167s opt=0.003s] dispatch=3.334s [2026-10-19 01:04:03,018] loop.py:45 INFO TRAIN_STEP step=5330 loss=1.511841 Training: 30%|██▉ | 5330/18018 [11:33:49<26:32:57, 7.53s/step, loss=1.512, acc=0.671, acc_len=1.59, thru=14.4, I=18.1, T=29.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:03,318] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:12,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:04:13,579] loop.py:45 INFO TRAIN_STEP step=5331 loss=1.538161 Training: 30%|██▉ | 5331/18018 [11:33:59<29:44:50, 8.44s/step, loss=1.538, acc=0.661, acc_len=1.56, thru=12.5, I=13.2, T=26.0, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:04:13,574] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5330: forward=3697.5ms backward=1173.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:13,993] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:04:21,047] loop.py:45 INFO TRAIN_STEP step=5332 loss=1.635587 Training: 30%|██▉ | 5332/18018 [11:34:07<28:42:57, 8.15s/step, loss=1.636, acc=0.651, acc_len=1.52, thru=15.0, I=14.5, T=33.2, wait=3.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:30,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:04:30,849] loop.py:45 INFO TRAIN_STEP step=5333 loss=1.654523 Training: 30%|██▉ | 5333/18018 [11:34:16<30:27:40, 8.64s/step, loss=1.655, acc=0.646, acc_len=1.51, thru=14.7, I=15.6, T=30.1, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:31,300] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:36,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:04:37,764] loop.py:45 INFO TRAIN_STEP step=5334 loss=1.551617 Training: 30%|██▉ | 5334/18018 [11:34:23<28:37:56, 8.13s/step, loss=1.552, acc=0.661, acc_len=1.57, thru=18.6, I=20.9, T=32.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:37,796] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:43,854] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:04:44,282] loop.py:380 INFO TIMING step=5335: step=3.793s data=1.659s compute=3.763s [fwd=2.647s bwd=1.103s opt=0.003s] dispatch=2.714s [2026-10-19 01:04:44,282] loop.py:45 INFO TRAIN_STEP step=5335 loss=1.604228 Training: 30%|██▉ | 5335/18018 [11:34:30<26:55:40, 7.64s/step, loss=1.604, acc=0.655, acc_len=1.53, thru=18.5, I=20.4, T=33.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:44,372] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:49,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:04:51,682] loop.py:45 INFO TRAIN_STEP step=5336 loss=1.674586 Training: 30%|██▉ | 5336/18018 [11:34:37<26:40:05, 7.57s/step, loss=1.675, acc=0.645, acc_len=1.50, thru=18.3, I=21.9, T=30.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:51,709] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-19 01:04:51,676] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5335: forward=2981.7ms backward=1184.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:57,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:04:59,571] loop.py:45 INFO TRAIN_STEP step=5337 loss=1.594552 Training: 30%|██▉ | 5337/18018 [11:34:45<27:00:19, 7.67s/step, loss=1.595, acc=0.656, acc_len=1.54, thru=13.3, I=20.1, T=30.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:04:59,937] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:07,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:05:08,906] loop.py:45 INFO TRAIN_STEP step=5338 loss=1.534536 Training: 30%|██▉ | 5338/18018 [11:34:54<28:45:49, 8.17s/step, loss=1.535, acc=0.665, acc_len=1.57, thru=14.3, I=16.4, T=30.3, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:09,357] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:05:14,616] loop.py:45 INFO TRAIN_STEP step=5339 loss=1.417889 Training: 30%|██▉ | 5339/18018 [11:35:00<26:10:09, 7.43s/step, loss=1.418, acc=0.681, acc_len=1.63, thru=19.3, I=22.1, T=37.2, wait=2.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:22,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:05:24,877] loop.py:380 INFO TIMING step=5340: step=5.046s data=2.515s compute=5.021s [fwd=3.691s bwd=1.316s opt=0.003s] dispatch=5.203s [2026-10-19 01:05:24,877] loop.py:45 INFO TRAIN_STEP step=5340 loss=1.577761 Training: 30%|██▉ | 5340/18018 [11:35:10<29:09:26, 8.28s/step, loss=1.578, acc=0.660, acc_len=1.55, thru=13.5, I=15.7, T=25.4, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:25,166] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:28,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:05:30,799] loop.py:45 INFO TRAIN_STEP step=5341 loss=1.423191 Training: 30%|██▉ | 5341/18018 [11:35:16<26:39:36, 7.57s/step, loss=1.423, acc=0.684, acc_len=1.64, thru=16.6, I=27.6, T=32.4, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:05:30,794] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5340: forward=2689.4ms backward=1210.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:30,954] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:35,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:05:37,222] loop.py:45 INFO TRAIN_STEP step=5342 loss=1.491807 Training: 30%|██▉ | 5342/18018 [11:35:23<25:26:52, 7.23s/step, loss=1.492, acc=0.669, acc_len=1.59, thru=18.4, I=28.0, T=32.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:37,594] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:42,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:05:43,700] loop.py:45 INFO TRAIN_STEP step=5343 loss=1.604620 Training: 30%|██▉ | 5343/18018 [11:35:29<24:39:23, 7.00s/step, loss=1.605, acc=0.653, acc_len=1.54, thru=18.6, I=25.9, T=32.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:44,014] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:51,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:05:52,475] loop.py:45 INFO TRAIN_STEP step=5344 loss=1.643516 Training: 30%|██▉ | 5344/18018 [11:35:38<26:31:29, 7.53s/step, loss=1.644, acc=0.644, acc_len=1.51, thru=13.2, I=16.8, T=29.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:05:52,662] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:06:00,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:06:01,169] loop.py:380 INFO TIMING step=5345: step=4.305s data=1.907s compute=4.261s [fwd=3.001s bwd=1.245s opt=0.003s] dispatch=4.379s [2026-10-19 01:06:01,170] loop.py:45 INFO TRAIN_STEP step=5345 loss=1.504134 Training: 30%|██▉ | 5345/18018 [11:35:47<27:44:55, 7.88s/step, loss=1.504, acc=0.668, acc_len=1.59, thru=14.2, I=16.8, T=29.7, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:06:01,367] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:06:07,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:06:08,353] loop.py:45 INFO TRAIN_STEP step=5346 loss=1.492144 Training: 30%|██▉ | 5346/18018 [11:35:54<27:00:30, 7.67s/step, loss=1.492, acc=0.669, acc_len=1.58, thru=17.6, I=20.2, T=29.8, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:06:08,348] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5345: forward=2873.9ms backward=1357.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:06:08,584] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:06:18,373] loop.py:45 INFO TRAIN_STEP step=5347 loss=1.610083 Training: 30%|██▉ | 5347/18018 [11:36:04<29:28:56, 8.38s/step, loss=1.610, acc=0.654, acc_len=1.53, thru=11.4, I=10.9, T=29.1, wait=5.6s, pool=48, epoch=1/2][2026-10-19 01:06:26,256] loop.py:45 INFO TRAIN_STEP step=5348 loss=1.621655 Training: 30%|██▉ | 5348/18018 [11:36:12<28:57:39, 8.23s/step, loss=1.622, acc=0.645, acc_len=1.51, thru=17.6, I=14.5, T=28.9, wait=3.4s, pool=52, epoch=1/2][2026-10-19 01:06:33,929] loop.py:45 INFO TRAIN_STEP step=5349 loss=1.621161 Training: 30%|██▉ | 5349/18018 [11:36:20<28:22:31, 8.06s/step, loss=1.621, acc=0.650, acc_len=1.52, thru=17.7, I=18.6, T=31.0, wait=3.5s, pool=60, epoch=1/2][2026-10-19 01:06:42,950] loop.py:380 INFO TIMING step=5350: step=4.333s data=1.976s compute=4.292s [fwd=3.073s bwd=1.205s opt=0.003s] dispatch=4.675s [2026-10-19 01:06:42,950] loop.py:45 INFO TRAIN_STEP step=5350 loss=1.568181 Training: 30%|██▉ | 5350/18018 [11:36:29<29:22:41, 8.35s/step, loss=1.568, acc=0.659, acc_len=1.56, thru=14.3, I=14.4, T=29.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:06:42,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:06:43,455] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:06:51,933] loop.py:45 INFO TRAIN_STEP step=5351 loss=1.554435 Training: 30%|██▉ | 5351/18018 [11:36:38<30:02:48, 8.54s/step, loss=1.554, acc=0.659, acc_len=1.56, thru=11.0, I=11.8, T=34.9, wait=5.3s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:06:51,928] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5350: forward=2644.3ms backward=981.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:01,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:07:01,349] loop.py:45 INFO TRAIN_STEP step=5352 loss=1.592644 Training: 30%|██▉ | 5352/18018 [11:36:47<30:58:16, 8.80s/step, loss=1.593, acc=0.656, acc_len=1.55, thru=17.1, I=14.6, T=32.4, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:01,564] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:08,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:07:09,602] loop.py:45 INFO TRAIN_STEP step=5353 loss=1.634101 Training: 30%|██▉ | 5353/18018 [11:36:55<30:23:16, 8.64s/step, loss=1.634, acc=0.647, acc_len=1.51, thru=15.4, I=17.9, T=28.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:09,636] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:16,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:07:17,409] loop.py:45 INFO TRAIN_STEP step=5354 loss=1.310774 Training: 30%|██▉ | 5354/18018 [11:37:03<29:30:36, 8.39s/step, loss=1.311, acc=0.698, acc_len=1.68, thru=15.3, I=19.0, T=29.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:17,730] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:23,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:07:24,620] loop.py:380 INFO TIMING step=5355: step=4.238s data=1.986s compute=4.204s [fwd=3.040s bwd=1.150s opt=0.003s] dispatch=2.961s [2026-10-19 01:07:24,620] loop.py:45 INFO TRAIN_STEP step=5355 loss=1.567290 Training: 30%|██▉ | 5355/18018 [11:37:10<28:15:44, 8.03s/step, loss=1.567, acc=0.659, acc_len=1.55, thru=17.7, I=21.2, T=30.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:25,124] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:07:32,161] loop.py:45 INFO TRAIN_STEP step=5356 loss=1.496528 Training: 30%|██▉ | 5356/18018 [11:37:18<27:44:20, 7.89s/step, loss=1.497, acc=0.670, acc_len=1.59, thru=15.5, I=17.6, T=31.7, wait=3.5s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:07:32,157] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5355: forward=2822.8ms backward=1165.4ms [2026-10-19 01:07:39,594] loop.py:45 INFO TRAIN_STEP step=5357 loss=1.595930 Training: 30%|██▉ | 5357/18018 [11:37:25<27:15:33, 7.75s/step, loss=1.596, acc=0.653, acc_len=1.53, thru=16.7, I=17.2, T=32.1, wait=3.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:46,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:07:48,442] loop.py:45 INFO TRAIN_STEP step=5358 loss=1.644952 Training: 30%|██▉ | 5358/18018 [11:37:34<28:24:56, 8.08s/step, loss=1.645, acc=0.644, acc_len=1.50, thru=16.4, I=16.4, T=28.8, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:48,535] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:54,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:07:54,696] loop.py:45 INFO TRAIN_STEP step=5359 loss=1.384114 Training: 30%|██▉ | 5359/18018 [11:37:40<26:29:11, 7.53s/step, loss=1.384, acc=0.689, acc_len=1.65, thru=16.9, I=20.9, T=35.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:07:54,977] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:00,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:08:01,421] loop.py:380 INFO TIMING step=5360: step=3.789s data=1.703s compute=3.751s [fwd=2.665s bwd=1.072s opt=0.003s] dispatch=2.927s [2026-10-19 01:08:01,421] loop.py:45 INFO TRAIN_STEP step=5360 loss=1.463792 Training: 30%|██▉ | 5360/18018 [11:37:47<25:37:55, 7.29s/step, loss=1.464, acc=0.673, acc_len=1.60, thru=18.7, I=20.6, T=33.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:01,900] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:14,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:08:14,518] loop.py:45 INFO TRAIN_STEP step=5361 loss=1.677599 Training: 30%|██▉ | 5361/18018 [11:38:00<31:45:19, 9.03s/step, loss=1.678, acc=0.641, acc_len=1.50, thru=11.9, I=10.1, T=25.8, wait=8.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:08:14,513] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5360: forward=3097.8ms backward=1798.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:14,623] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:19,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:08:20,734] loop.py:45 INFO TRAIN_STEP step=5362 loss=1.575717 Training: 30%|██▉ | 5362/18018 [11:38:06<28:46:55, 8.19s/step, loss=1.576, acc=0.658, acc_len=1.55, thru=17.2, I=18.5, T=34.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:20,984] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:08:26,770] loop.py:45 INFO TRAIN_STEP step=5363 loss=1.510290 Training: 30%|██▉ | 5363/18018 [11:38:12<26:30:42, 7.54s/step, loss=1.510, acc=0.669, acc_len=1.58, thru=19.8, I=27.2, T=33.3, wait=2.2s, pool=56, epoch=1/2][2026-10-19 01:08:35,228] loop.py:45 INFO TRAIN_STEP step=5364 loss=1.551944 Training: 30%|██▉ | 5364/18018 [11:38:21<27:28:24, 7.82s/step, loss=1.552, acc=0.661, acc_len=1.55, thru=15.3, I=17.5, T=30.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:35,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:35,659] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:08:42,806] loop.py:380 INFO TIMING step=5365: step=4.206s data=2.021s compute=4.146s [fwd=2.858s bwd=1.273s opt=0.003s] dispatch=3.365s [2026-10-19 01:08:42,806] loop.py:45 INFO TRAIN_STEP step=5365 loss=1.628640 Training: 30%|██▉ | 5365/18018 [11:38:28<27:13:19, 7.75s/step, loss=1.629, acc=0.652, acc_len=1.53, thru=13.5, I=15.0, T=30.4, wait=3.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:50,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:08:52,483] loop.py:45 INFO TRAIN_STEP step=5366 loss=1.540358 Training: 30%|██▉ | 5366/18018 [11:38:38<29:15:18, 8.32s/step, loss=1.540, acc=0.664, acc_len=1.57, thru=15.3, I=16.3, T=28.5, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:08:52,478] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5365: forward=3181.3ms backward=1261.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:52,936] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:58,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:08:58,969] loop.py:45 INFO TRAIN_STEP step=5367 loss=1.569079 Training: 30%|██▉ | 5367/18018 [11:38:45<27:18:57, 7.77s/step, loss=1.569, acc=0.659, acc_len=1.55, thru=17.5, I=23.6, T=34.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:08:59,307] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:09:05,201] loop.py:45 INFO TRAIN_STEP step=5368 loss=1.466136 Training: 30%|██▉ | 5368/18018 [11:38:51<25:41:21, 7.31s/step, loss=1.466, acc=0.673, acc_len=1.59, thru=19.3, I=20.9, T=32.2, wait=2.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:09:10,733] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:09:12,774] loop.py:45 INFO TRAIN_STEP step=5369 loss=1.584064 Training: 30%|██▉ | 5369/18018 [11:38:58<25:57:43, 7.39s/step, loss=1.584, acc=0.652, acc_len=1.53, thru=17.6, I=19.9, T=27.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:09:13,251] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:09:20,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:09:21,612] loop.py:380 INFO TIMING step=5370: step=4.594s data=1.990s compute=4.540s [fwd=3.221s bwd=1.304s opt=0.003s] dispatch=4.236s [2026-10-19 01:09:21,612] loop.py:45 INFO TRAIN_STEP step=5370 loss=1.479382 Training: 30%|██▉ | 5370/18018 [11:39:07<27:29:23, 7.82s/step, loss=1.479, acc=0.676, acc_len=1.61, thru=15.3, I=16.3, T=27.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:09:22,115] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:09:27,988] loop.py:45 INFO TRAIN_STEP step=5371 loss=1.641621 Training: 30%|██▉ | 5371/18018 [11:39:14<25:57:37, 7.39s/step, loss=1.642, acc=0.643, acc_len=1.50, thru=16.4, I=19.4, T=33.8, wait=2.6s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:09:27,983] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5370: forward=2638.1ms backward=1092.3ms [2026-10-19 01:09:36,767] loop.py:45 INFO TRAIN_STEP step=5372 loss=1.485005 Training: 30%|██▉ | 5372/18018 [11:39:22<27:25:24, 7.81s/step, loss=1.485, acc=0.671, acc_len=1.60, thru=15.6, I=17.6, T=24.1, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:09:42,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:09:45,343] loop.py:45 INFO TRAIN_STEP step=5373 loss=1.558472 Training: 30%|██▉ | 5373/18018 [11:39:31<28:13:58, 8.04s/step, loss=1.558, acc=0.661, acc_len=1.56, thru=14.4, I=18.4, T=27.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:09:45,657] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:09:51,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:09:51,416] loop.py:45 INFO TRAIN_STEP step=5374 loss=1.551384 Training: 30%|██▉ | 5374/18018 [11:39:37<26:09:32, 7.45s/step, loss=1.551, acc=0.662, acc_len=1.56, thru=16.9, I=24.2, T=35.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:09:51,575] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:09:57,898] loop.py:380 INFO TIMING step=5375: step=3.985s data=1.837s compute=3.951s [fwd=2.790s bwd=1.147s opt=0.003s] dispatch=2.488s [2026-10-19 01:09:57,899] loop.py:45 INFO TRAIN_STEP step=5375 loss=1.488096 Training: 30%|██▉ | 5375/18018 [11:39:43<25:08:16, 7.16s/step, loss=1.488, acc=0.670, acc_len=1.60, thru=19.1, I=20.5, T=32.1, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:05,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:10:05,925] loop.py:45 INFO TRAIN_STEP step=5376 loss=1.374980 Training: 30%|██▉ | 5376/18018 [11:39:52<26:03:10, 7.42s/step, loss=1.375, acc=0.692, acc_len=1.66, thru=17.4, I=17.3, T=32.9, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:10:05,920] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5375: forward=2663.1ms backward=1171.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:06,165] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:11,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:10:11,658] loop.py:45 INFO TRAIN_STEP step=5377 loss=1.376163 Training: 30%|██▉ | 5377/18018 [11:39:57<24:16:23, 6.91s/step, loss=1.376, acc=0.691, acc_len=1.65, thru=20.7, I=21.7, T=37.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:12,102] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:10:17,744] loop.py:45 INFO TRAIN_STEP step=5378 loss=1.605364 Training: 30%|██▉ | 5378/18018 [11:40:03<23:23:59, 6.66s/step, loss=1.605, acc=0.652, acc_len=1.53, thru=18.9, I=21.5, T=35.0, wait=2.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:27,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:10:27,860] loop.py:45 INFO TRAIN_STEP step=5379 loss=1.543659 Training: 30%|██▉ | 5379/18018 [11:40:13<27:02:05, 7.70s/step, loss=1.544, acc=0.663, acc_len=1.57, thru=15.3, I=14.9, T=31.5, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:28,244] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:35,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:10:35,528] loop.py:380 INFO TIMING step=5380: step=3.974s data=1.712s compute=3.933s [fwd=2.723s bwd=1.195s opt=0.003s] dispatch=3.686s [2026-10-19 01:10:35,528] loop.py:45 INFO TRAIN_STEP step=5380 loss=1.647381 Training: 30%|██▉ | 5380/18018 [11:40:21<26:59:52, 7.69s/step, loss=1.647, acc=0.645, acc_len=1.51, thru=17.5, I=17.5, T=32.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:35,881] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:40,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:10:42,815] loop.py:45 INFO TRAIN_STEP step=5381 loss=1.512647 Training: 30%|██▉ | 5381/18018 [11:40:28<26:34:14, 7.57s/step, loss=1.513, acc=0.666, acc_len=1.57, thru=17.9, I=21.2, T=31.2, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:10:42,810] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5380: forward=2889.1ms backward=1170.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:42,860] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:48,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:10:49,573] loop.py:45 INFO TRAIN_STEP step=5382 loss=1.514234 Training: 30%|██▉ | 5382/18018 [11:40:35<25:42:52, 7.33s/step, loss=1.514, acc=0.668, acc_len=1.58, thru=17.7, I=22.9, T=33.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:10:49,776] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:10:55,738] loop.py:45 INFO TRAIN_STEP step=5383 loss=1.591828 Training: 30%|██▉ | 5383/18018 [11:40:41<24:29:18, 6.98s/step, loss=1.592, acc=0.653, acc_len=1.53, thru=19.0, I=20.9, T=30.8, wait=2.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:03,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:11:05,049] loop.py:45 INFO TRAIN_STEP step=5384 loss=1.517129 Training: 30%|██▉ | 5384/18018 [11:40:51<26:56:43, 7.68s/step, loss=1.517, acc=0.669, acc_len=1.59, thru=14.6, I=16.7, T=30.0, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:05,078] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 01:11:11,016] loop.py:380 INFO TIMING step=5385: step=3.540s data=1.532s compute=3.511s [fwd=2.519s bwd=0.977s opt=0.003s] dispatch=2.417s [2026-10-19 01:11:11,016] loop.py:45 INFO TRAIN_STEP step=5385 loss=1.501929 Training: 30%|██▉ | 5385/18018 [11:40:57<25:08:29, 7.16s/step, loss=1.502, acc=0.665, acc_len=1.57, thru=17.0, I=21.2, T=36.2, wait=2.4s, pool=48, epoch=1/2][2026-10-19 01:11:20,705] loop.py:45 INFO TRAIN_STEP step=5386 loss=1.606899 Training: 30%|██▉ | 5386/18018 [11:41:06<27:47:56, 7.92s/step, loss=1.607, acc=0.652, acc_len=1.53, thru=13.7, I=15.4, T=32.0, wait=5.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:11:20,699] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5385: forward=2877.7ms backward=1077.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:31,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:11:32,734] loop.py:45 INFO TRAIN_STEP step=5387 loss=1.558062 Training: 30%|██▉ | 5387/18018 [11:41:18<32:07:04, 9.15s/step, loss=1.558, acc=0.657, acc_len=1.55, thru=11.8, I=13.7, T=29.4, wait=7.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:32,817] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:37,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:11:38,689] loop.py:45 INFO TRAIN_STEP step=5388 loss=1.461511 Training: 30%|██▉ | 5388/18018 [11:41:24<28:45:02, 8.19s/step, loss=1.462, acc=0.676, acc_len=1.61, thru=19.0, I=22.9, T=32.6, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:39,171] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:44,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:11:45,875] loop.py:45 INFO TRAIN_STEP step=5389 loss=1.489652 Training: 30%|██▉ | 5389/18018 [11:41:31<27:41:09, 7.89s/step, loss=1.490, acc=0.671, acc_len=1.59, thru=15.7, I=23.6, T=30.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:46,025] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:53,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:11:54,790] loop.py:380 INFO TIMING step=5390: step=3.805s data=1.676s compute=3.759s [fwd=2.689s bwd=1.055s opt=0.003s] dispatch=5.099s [2026-10-19 01:11:54,790] loop.py:45 INFO TRAIN_STEP step=5390 loss=1.600397 Training: 30%|██▉ | 5390/18018 [11:41:40<28:45:27, 8.20s/step, loss=1.600, acc=0.652, acc_len=1.53, thru=14.6, I=15.3, T=33.6, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:11:54,965] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:00,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:12:01,080] loop.py:45 INFO TRAIN_STEP step=5391 loss=1.660885 Training: 30%|██▉ | 5391/18018 [11:41:47<26:44:49, 7.63s/step, loss=1.661, acc=0.642, acc_len=1.50, thru=19.3, I=24.2, T=34.5, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:12:01,076] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5390: forward=2579.2ms backward=1088.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:01,523] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:09,978] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:12:10,837] loop.py:45 INFO TRAIN_STEP step=5392 loss=1.656330 Training: 30%|██▉ | 5392/18018 [11:41:56<28:59:23, 8.27s/step, loss=1.656, acc=0.643, acc_len=1.50, thru=13.7, I=15.4, T=27.8, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:10,982] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:16,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:12:17,246] loop.py:45 INFO TRAIN_STEP step=5393 loss=1.471046 Training: 30%|██▉ | 5393/18018 [11:42:03<27:02:03, 7.71s/step, loss=1.471, acc=0.674, acc_len=1.60, thru=17.8, I=20.7, T=31.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:17,296] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:21,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:12:24,149] loop.py:45 INFO TRAIN_STEP step=5394 loss=1.639370 Training: 30%|██▉ | 5394/18018 [11:42:10<26:10:58, 7.47s/step, loss=1.639, acc=0.645, acc_len=1.50, thru=16.9, I=22.9, T=29.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:24,475] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:30,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:12:30,690] loop.py:380 INFO TIMING step=5395: step=3.968s data=1.787s compute=3.937s [fwd=2.761s bwd=1.161s opt=0.003s] dispatch=2.566s [2026-10-19 01:12:30,690] loop.py:45 INFO TRAIN_STEP step=5395 loss=1.511000 Training: 30%|██▉ | 5395/18018 [11:42:16<25:12:24, 7.19s/step, loss=1.511, acc=0.665, acc_len=1.58, thru=16.7, I=23.5, T=32.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:12:31,183] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:12:37,418] loop.py:45 INFO TRAIN_STEP step=5396 loss=1.581296 Training: 30%|██▉ | 5396/18018 [11:42:23<24:43:12, 7.05s/step, loss=1.581, acc=0.655, acc_len=1.54, thru=16.4, I=20.1, T=35.6, wait=3.1s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:12:37,413] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5395: forward=2443.0ms backward=1098.4ms [2026-10-19 01:12:45,768] loop.py:45 INFO TRAIN_STEP step=5397 loss=1.478773 Training: 30%|██▉ | 5397/18018 [11:42:31<26:05:16, 7.44s/step, loss=1.479, acc=0.673, acc_len=1.60, thru=17.0, I=15.0, T=31.7, wait=4.3s, pool=60, epoch=1/2][2026-10-19 01:12:53,422] loop.py:45 INFO TRAIN_STEP step=5398 loss=1.513993 Training: 30%|██▉ | 5398/18018 [11:42:39<26:18:34, 7.51s/step, loss=1.514, acc=0.664, acc_len=1.57, thru=15.4, I=17.1, T=32.3, wait=3.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:03,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:13:04,396] loop.py:45 INFO TRAIN_STEP step=5399 loss=1.585943 Training: 30%|██▉ | 5399/18018 [11:42:50<29:57:15, 8.55s/step, loss=1.586, acc=0.652, acc_len=1.53, thru=14.4, I=13.7, T=27.7, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:04,568] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:09,972] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:13:11,342] loop.py:380 INFO TIMING step=5400: step=4.228s data=1.874s compute=4.199s [fwd=2.972s bwd=1.212s opt=0.003s] dispatch=2.706s [2026-10-19 01:13:11,342] loop.py:45 INFO TRAIN_STEP step=5400 loss=1.481050 Training: 30%|██▉ | 5400/18018 [11:42:57<28:16:05, 8.07s/step, loss=1.481, acc=0.673, acc_len=1.60, thru=15.7, I=20.7, T=30.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:11,493] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:17,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:13:18,666] loop.py:45 INFO TRAIN_STEP step=5401 loss=1.475337 Training: 30%|██▉ | 5401/18018 [11:43:04<27:29:18, 7.84s/step, loss=1.475, acc=0.671, acc_len=1.59, thru=16.2, I=20.7, T=30.6, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:13:18,660] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5400: forward=2888.6ms backward=1246.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:18,726] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:23,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:13:25,188] loop.py:45 INFO TRAIN_STEP step=5402 loss=1.424387 Training: 30%|██▉ | 5402/18018 [11:43:11<26:05:45, 7.45s/step, loss=1.424, acc=0.684, acc_len=1.63, thru=17.7, I=23.5, T=28.4, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:25,265] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:31,122] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:13:32,480] loop.py:45 INFO TRAIN_STEP step=5403 loss=1.603897 Training: 30%|██▉ | 5403/18018 [11:43:18<25:55:56, 7.40s/step, loss=1.604, acc=0.653, acc_len=1.53, thru=16.2, I=22.3, T=30.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:32,636] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:39,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:13:40,655] loop.py:45 INFO TRAIN_STEP step=5404 loss=1.560205 Training: 30%|██▉ | 5404/18018 [11:43:26<26:44:32, 7.63s/step, loss=1.560, acc=0.657, acc_len=1.54, thru=14.0, I=19.8, T=29.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:40,973] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:46,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:13:47,941] loop.py:380 INFO TIMING step=5405: step=4.704s data=2.319s compute=4.664s [fwd=3.245s bwd=1.403s opt=0.003s] dispatch=2.574s [2026-10-19 01:13:47,941] loop.py:45 INFO TRAIN_STEP step=5405 loss=1.666560 Training: 30%|██▉ | 5405/18018 [11:43:34<26:22:45, 7.53s/step, loss=1.667, acc=0.643, acc_len=1.50, thru=16.6, I=21.4, T=27.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:13:48,237] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:13:55,545] loop.py:45 INFO TRAIN_STEP step=5406 loss=1.612819 Training: 30%|███ | 5406/18018 [11:43:41<26:27:11, 7.55s/step, loss=1.613, acc=0.648, acc_len=1.52, thru=13.3, I=15.0, T=30.6, wait=3.4s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:13:55,540] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5405: forward=2944.3ms backward=1184.5ms [2026-10-19 01:14:06,297] loop.py:45 INFO TRAIN_STEP step=5407 loss=1.588422 Training: 30%|███ | 5407/18018 [11:43:52<29:49:03, 8.51s/step, loss=1.588, acc=0.663, acc_len=1.57, thru=13.7, I=13.0, T=30.8, wait=6.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:13,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:14:14,426] loop.py:45 INFO TRAIN_STEP step=5408 loss=1.640805 Training: 30%|███ | 5408/18018 [11:44:00<29:24:38, 8.40s/step, loss=1.641, acc=0.649, acc_len=1.52, thru=17.5, I=18.5, T=34.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:14,579] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:14:21,182] loop.py:45 INFO TRAIN_STEP step=5409 loss=1.535229 Training: 30%|███ | 5409/18018 [11:44:07<27:41:04, 7.90s/step, loss=1.535, acc=0.667, acc_len=1.58, thru=17.3, I=21.0, T=29.4, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:27,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:14:28,096] loop.py:380 INFO TIMING step=5410: step=4.151s data=1.871s compute=4.120s [fwd=2.880s bwd=1.224s opt=0.003s] dispatch=2.755s [2026-10-19 01:14:28,096] loop.py:45 INFO TRAIN_STEP step=5410 loss=1.543371 Training: 30%|███ | 5410/18018 [11:44:14<26:38:39, 7.61s/step, loss=1.543, acc=0.659, acc_len=1.54, thru=16.6, I=17.6, T=30.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:28,238] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:36,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:14:37,891] loop.py:45 INFO TRAIN_STEP step=5411 loss=1.571145 Training: 30%|███ | 5411/18018 [11:44:23<28:56:17, 8.26s/step, loss=1.571, acc=0.652, acc_len=1.54, thru=13.3, I=14.6, T=30.0, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:14:37,886] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5410: forward=3008.6ms backward=1208.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:38,021] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:44,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:14:46,436] loop.py:45 INFO TRAIN_STEP step=5412 loss=1.429579 Training: 30%|███ | 5412/18018 [11:44:32<29:13:51, 8.35s/step, loss=1.430, acc=0.683, acc_len=1.63, thru=13.6, I=18.2, T=31.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:46,736] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:51,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:14:52,621] loop.py:45 INFO TRAIN_STEP step=5413 loss=1.617180 Training: 30%|███ | 5413/18018 [11:44:38<26:57:33, 7.70s/step, loss=1.617, acc=0.651, acc_len=1.52, thru=18.8, I=25.8, T=35.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:53,000] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:14:58,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:14:59,754] loop.py:45 INFO TRAIN_STEP step=5414 loss=1.647705 Training: 30%|███ | 5414/18018 [11:44:45<26:21:44, 7.53s/step, loss=1.648, acc=0.645, acc_len=1.51, thru=18.1, I=24.1, T=29.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:15:00,137] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:15:06,408] loop.py:380 INFO TIMING step=5415: step=3.857s data=1.694s compute=3.821s [fwd=2.693s bwd=1.114s opt=0.003s] dispatch=2.787s [2026-10-19 01:15:06,408] loop.py:45 INFO TRAIN_STEP step=5415 loss=1.637131 Training: 30%|███ | 5415/18018 [11:44:52<25:26:32, 7.27s/step, loss=1.637, acc=0.645, acc_len=1.51, thru=15.0, I=20.6, T=33.2, wait=2.8s, pool=40, epoch=1/2][2026-10-19 01:15:15,148] loop.py:45 INFO TRAIN_STEP step=5416 loss=1.523098 (TrainerActor pid=9310) [2026-10-19 01:15:15,144] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5415: forward=3075.0ms backward=1312.7ms Training: 30%|███ | 5416/18018 [11:45:01<26:59:06, 7.71s/step, loss=1.523, acc=0.664, acc_len=1.57, thru=14.6, I=14.8, T=28.8, wait=4.3s, pool=40, epoch=1/2][2026-10-19 01:15:24,571] loop.py:45 INFO TRAIN_STEP step=5417 loss=1.698107 Training: 30%|███ | 5417/18018 [11:45:10<28:46:52, 8.22s/step, loss=1.698, acc=0.637, acc_len=1.48, thru=16.1, I=13.3, T=29.8, wait=5.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:15:33,544] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:15:33,986] loop.py:45 INFO TRAIN_STEP step=5418 loss=1.484967 Training: 30%|███ | 5418/18018 [11:45:20<30:01:57, 8.58s/step, loss=1.485, acc=0.671, acc_len=1.59, thru=16.1, I=14.3, T=30.3, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:15:34,058] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:15:39,416] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:15:41,453] loop.py:45 INFO TRAIN_STEP step=5419 loss=1.699331 Training: 30%|███ | 5419/18018 [11:45:27<28:51:28, 8.25s/step, loss=1.699, acc=0.635, acc_len=1.48, thru=17.5, I=20.3, T=30.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:15:41,938] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:15:49,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:15:50,396] loop.py:380 INFO TIMING step=5420: step=3.949s data=1.782s compute=3.911s [fwd=2.726s bwd=1.169s opt=0.003s] dispatch=4.986s [2026-10-19 01:15:50,397] loop.py:45 INFO TRAIN_STEP step=5420 loss=1.551817 Training: 30%|███ | 5420/18018 [11:45:36<29:35:20, 8.46s/step, loss=1.552, acc=0.664, acc_len=1.57, thru=15.1, I=17.2, T=32.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:15:50,708] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:15:57,282] loop.py:45 INFO TRAIN_STEP step=5421 loss=1.484407 Training: 30%|███ | 5421/18018 [11:45:43<27:56:18, 7.98s/step, loss=1.484, acc=0.675, acc_len=1.60, thru=16.6, I=18.4, T=33.3, wait=3.0s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:15:57,278] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5420: forward=2730.7ms backward=1067.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:16:05,568] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:16:06,566] loop.py:45 INFO TRAIN_STEP step=5422 loss=1.416614 Training: 30%|███ | 5422/18018 [11:45:52<29:18:11, 8.37s/step, loss=1.417, acc=0.687, acc_len=1.64, thru=15.3, I=16.0, T=31.4, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:16:07,074] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:16:12,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:16:13,588] loop.py:45 INFO TRAIN_STEP step=5423 loss=1.549254 Training: 30%|███ | 5423/18018 [11:45:59<27:52:42, 7.97s/step, loss=1.549, acc=0.662, acc_len=1.56, thru=18.5, I=22.2, T=35.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:16:14,046] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:16:22,213] loop.py:45 INFO TRAIN_STEP step=5424 loss=1.548357 Training: 30%|███ | 5424/18018 [11:46:08<28:34:14, 8.17s/step, loss=1.548, acc=0.661, acc_len=1.56, thru=13.5, I=16.4, T=28.9, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:16:29,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:16:31,836] loop.py:380 INFO TIMING step=5425: step=5.043s data=2.281s compute=4.998s [fwd=3.505s bwd=1.477s opt=0.003s] dispatch=4.565s [2026-10-19 01:16:31,836] loop.py:45 INFO TRAIN_STEP step=5425 loss=1.490704 Training: 30%|███ | 5425/18018 [11:46:17<30:05:33, 8.60s/step, loss=1.491, acc=0.669, acc_len=1.58, thru=13.8, I=16.7, T=25.4, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:16:31,878] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 01:16:40,043] loop.py:45 INFO TRAIN_STEP step=5426 loss=1.498683 Training: 30%|███ | 5426/18018 [11:46:26<29:40:20, 8.48s/step, loss=1.499, acc=0.669, acc_len=1.58, thru=13.8, I=13.8, T=33.0, wait=4.3s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:16:40,037] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5425: forward=2701.2ms backward=1119.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:16:45,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:16:47,478] loop.py:45 INFO TRAIN_STEP step=5427 loss=1.506336 Training: 30%|███ | 5427/18018 [11:46:33<28:34:16, 8.17s/step, loss=1.506, acc=0.668, acc_len=1.59, thru=18.6, I=19.0, T=29.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:16:47,508] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-19 01:16:55,501] loop.py:45 INFO TRAIN_STEP step=5428 loss=1.524160 Training: 30%|███ | 5428/18018 [11:46:41<28:24:57, 8.13s/step, loss=1.524, acc=0.664, acc_len=1.57, thru=12.6, I=17.9, T=33.4, wait=4.2s, pool=44, epoch=1/2][2026-10-19 01:17:04,411] loop.py:45 INFO TRAIN_STEP step=5429 loss=1.534227 Training: 30%|███ | 5429/18018 [11:46:50<29:14:08, 8.36s/step, loss=1.534, acc=0.662, acc_len=1.56, thru=15.2, I=13.7, T=32.4, wait=5.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:10,108] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:17:11,879] loop.py:380 INFO TIMING step=5430: step=4.061s data=1.847s compute=4.010s [fwd=2.809s bwd=1.186s opt=0.003s] dispatch=3.399s [2026-10-19 01:17:11,879] loop.py:45 INFO TRAIN_STEP step=5430 loss=1.631913 Training: 30%|███ | 5430/18018 [11:46:57<28:17:58, 8.09s/step, loss=1.632, acc=0.646, acc_len=1.51, thru=16.8, I=19.0, T=31.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:12,117] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:18,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:17:18,874] loop.py:45 INFO TRAIN_STEP step=5431 loss=1.501858 Training: 30%|███ | 5431/18018 [11:47:04<27:08:38, 7.76s/step, loss=1.502, acc=0.674, acc_len=1.60, thru=16.3, I=21.3, T=34.7, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:17:18,868] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5430: forward=2581.9ms backward=1058.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:19,378] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:24,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:17:25,596] loop.py:45 INFO TRAIN_STEP step=5432 loss=1.560363 Training: 30%|███ | 5432/18018 [11:47:11<26:03:03, 7.45s/step, loss=1.560, acc=0.658, acc_len=1.54, thru=18.4, I=21.2, T=35.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:26,002] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:17:33,197] loop.py:45 INFO TRAIN_STEP step=5433 loss=1.600360 Training: 30%|███ | 5433/18018 [11:47:19<26:12:14, 7.50s/step, loss=1.600, acc=0.653, acc_len=1.54, thru=17.4, I=20.4, T=34.2, wait=3.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:41,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:17:42,410] loop.py:45 INFO TRAIN_STEP step=5434 loss=1.474415 Training: 30%|███ | 5434/18018 [11:47:28<28:00:09, 8.01s/step, loss=1.474, acc=0.671, acc_len=1.59, thru=14.5, I=15.5, T=31.6, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:42,736] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:50,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:17:51,071] loop.py:380 INFO TIMING step=5435: step=5.043s data=2.427s compute=5.006s [fwd=3.445s bwd=1.546s opt=0.003s] dispatch=3.609s [2026-10-19 01:17:51,071] loop.py:45 INFO TRAIN_STEP step=5435 loss=1.579914 Training: 30%|███ | 5435/18018 [11:47:37<28:40:54, 8.21s/step, loss=1.580, acc=0.656, acc_len=1.54, thru=13.4, I=20.4, T=25.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:51,528] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:17:59,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:18:00,354] loop.py:45 INFO TRAIN_STEP step=5436 loss=1.562212 Training: 30%|███ | 5436/18018 [11:47:46<29:48:36, 8.53s/step, loss=1.562, acc=0.663, acc_len=1.56, thru=13.4, I=13.7, T=28.9, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:18:00,348] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5435: forward=3125.9ms backward=1251.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:00,675] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:06,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:18:07,660] loop.py:45 INFO TRAIN_STEP step=5437 loss=1.567020 Training: 30%|███ | 5437/18018 [11:47:53<28:31:27, 8.16s/step, loss=1.567, acc=0.657, acc_len=1.55, thru=16.2, I=19.5, T=32.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:07,941] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:13,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:18:15,706] loop.py:45 INFO TRAIN_STEP step=5438 loss=1.494888 Training: 30%|███ | 5438/18018 [11:48:01<28:23:59, 8.13s/step, loss=1.495, acc=0.671, acc_len=1.59, thru=15.3, I=19.7, T=24.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:16,064] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:21,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:18:22,160] loop.py:45 INFO TRAIN_STEP step=5439 loss=1.603951 Training: 30%|███ | 5439/18018 [11:48:08<26:38:40, 7.63s/step, loss=1.604, acc=0.655, acc_len=1.54, thru=15.8, I=20.4, T=33.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:22,190] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:26,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:18:28,653] loop.py:380 INFO TIMING step=5440: step=4.180s data=1.882s compute=4.140s [fwd=2.844s bwd=1.279s opt=0.003s] dispatch=2.303s [2026-10-19 01:18:28,653] loop.py:45 INFO TRAIN_STEP step=5440 loss=1.661476 Training: 30%|███ | 5440/18018 [11:48:14<25:27:14, 7.29s/step, loss=1.661, acc=0.645, acc_len=1.51, thru=19.3, I=23.2, T=30.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:28,776] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-19 01:18:35,815] loop.py:45 INFO TRAIN_STEP step=5441 loss=1.476354 Training: 30%|███ | 5441/18018 [11:48:21<25:19:23, 7.25s/step, loss=1.476, acc=0.674, acc_len=1.59, thru=14.0, I=17.3, T=29.9, wait=2.9s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:18:35,810] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5440: forward=2979.8ms backward=1252.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:44,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:18:44,188] loop.py:45 INFO TRAIN_STEP step=5442 loss=1.624190 Training: 30%|███ | 5442/18018 [11:48:30<26:29:59, 7.59s/step, loss=1.624, acc=0.647, acc_len=1.51, thru=17.6, I=16.4, T=31.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:44,623] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:51,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:18:53,022] loop.py:45 INFO TRAIN_STEP step=5443 loss=1.522328 Training: 30%|███ | 5443/18018 [11:48:39<27:48:21, 7.96s/step, loss=1.522, acc=0.663, acc_len=1.57, thru=14.9, I=15.2, T=33.1, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:18:53,449] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:01,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:19:02,096] loop.py:45 INFO TRAIN_STEP step=5444 loss=1.577187 Training: 30%|███ | 5444/18018 [11:48:48<28:58:20, 8.29s/step, loss=1.577, acc=0.655, acc_len=1.54, thru=14.8, I=14.2, T=32.9, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:02,404] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:07,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:19:08,742] loop.py:380 INFO TIMING step=5445: step=3.670s data=1.597s compute=3.640s [fwd=2.632s bwd=0.994s opt=0.003s] dispatch=2.966s [2026-10-19 01:19:08,743] loop.py:45 INFO TRAIN_STEP step=5445 loss=1.501192 Training: 30%|███ | 5445/18018 [11:48:54<27:14:28, 7.80s/step, loss=1.501, acc=0.665, acc_len=1.57, thru=18.8, I=22.2, T=34.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:09,215] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:15,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 01:19:16,283] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5445: forward=2794.9ms backward=1096.3ms [2026-10-19 01:19:16,288] loop.py:45 INFO TRAIN_STEP step=5446 loss=1.545836 Training: 30%|███ | 5446/18018 [11:49:02<26:58:19, 7.72s/step, loss=1.546, acc=0.656, acc_len=1.55, thru=15.5, I=22.5, T=32.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:16,790] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:22,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:19:23,599] loop.py:45 INFO TRAIN_STEP step=5447 loss=1.553738 Training: 30%|███ | 5447/18018 [11:49:09<26:32:25, 7.60s/step, loss=1.554, acc=0.660, acc_len=1.55, thru=16.4, I=20.5, T=32.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:23,632] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:30,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:19:30,972] loop.py:45 INFO TRAIN_STEP step=5448 loss=1.485760 Training: 30%|███ | 5448/18018 [11:49:17<26:17:57, 7.53s/step, loss=1.486, acc=0.670, acc_len=1.59, thru=16.6, I=19.1, T=32.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:31,286] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:38,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:19:38,691] loop.py:45 INFO TRAIN_STEP step=5449 loss=1.627522 Training: 30%|███ | 5449/18018 [11:49:24<26:29:37, 7.59s/step, loss=1.628, acc=0.646, acc_len=1.51, thru=15.7, I=18.1, T=32.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:39,005] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:19:46,424] loop.py:380 INFO TIMING step=5450: step=3.958s data=1.772s compute=3.922s [fwd=2.820s bwd=1.087s opt=0.003s] dispatch=3.764s [2026-10-19 01:19:46,424] loop.py:45 INFO TRAIN_STEP step=5450 loss=1.471046 Training: 30%|███ | 5450/18018 [11:49:32<26:38:29, 7.63s/step, loss=1.471, acc=0.676, acc_len=1.61, thru=14.4, I=17.9, T=32.3, wait=3.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:53,308] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:19:55,263] loop.py:45 INFO TRAIN_STEP step=5451 loss=1.694247 Training: 30%|███ | 5451/18018 [11:49:41<27:54:20, 7.99s/step, loss=1.694, acc=0.640, acc_len=1.49, thru=16.4, I=18.7, T=29.3, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:19:55,258] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5450: forward=3087.4ms backward=1223.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:19:55,340] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:01,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:20:01,755] loop.py:45 INFO TRAIN_STEP step=5452 loss=1.600783 Training: 30%|███ | 5452/18018 [11:49:47<26:19:52, 7.54s/step, loss=1.601, acc=0.655, acc_len=1.53, thru=16.9, I=23.0, T=35.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:02,101] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:20:11,637] loop.py:45 INFO TRAIN_STEP step=5453 loss=1.519967 Training: 30%|███ | 5453/18018 [11:49:57<28:46:30, 8.24s/step, loss=1.520, acc=0.663, acc_len=1.57, thru=12.2, I=11.7, T=26.9, wait=5.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:18,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:20:19,881] loop.py:45 INFO TRAIN_STEP step=5454 loss=1.458107 Training: 30%|███ | 5454/18018 [11:50:05<28:46:27, 8.24s/step, loss=1.458, acc=0.673, acc_len=1.59, thru=16.0, I=16.4, T=31.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:20,332] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:25,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:20:26,536] loop.py:380 INFO TIMING step=5455: step=4.551s data=2.173s compute=4.511s [fwd=3.160s bwd=1.336s opt=0.003s] dispatch=2.093s [2026-10-19 01:20:26,536] loop.py:45 INFO TRAIN_STEP step=5455 loss=1.456508 Training: 30%|███ | 5455/18018 [11:50:12<27:06:20, 7.77s/step, loss=1.457, acc=0.673, acc_len=1.61, thru=16.9, I=22.6, T=28.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:26,960] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:20:34,723] loop.py:45 INFO TRAIN_STEP step=5456 loss=1.507784 Training: 30%|███ | 5456/18018 [11:50:20<27:32:37, 7.89s/step, loss=1.508, acc=0.669, acc_len=1.59, thru=13.3, I=14.7, T=27.4, wait=3.5s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:20:34,718] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5455: forward=3228.7ms backward=1398.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:43,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:20:44,732] loop.py:45 INFO TRAIN_STEP step=5457 loss=1.631336 Training: 30%|███ | 5457/18018 [11:50:30<29:45:19, 8.53s/step, loss=1.631, acc=0.646, acc_len=1.51, thru=15.2, I=15.1, T=27.5, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:44,783] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:50,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:20:50,843] loop.py:45 INFO TRAIN_STEP step=5458 loss=1.681993 Training: 30%|███ | 5458/18018 [11:50:36<27:13:24, 7.80s/step, loss=1.682, acc=0.643, acc_len=1.51, thru=18.2, I=19.7, T=32.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:51,163] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:57,860] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:20:59,455] loop.py:45 INFO TRAIN_STEP step=5459 loss=1.544004 Training: 30%|███ | 5459/18018 [11:50:45<28:04:08, 8.05s/step, loss=1.544, acc=0.662, acc_len=1.57, thru=15.1, I=16.2, T=30.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:20:59,875] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:05,415] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:21:07,028] loop.py:380 INFO TIMING step=5460: step=4.081s data=1.840s compute=4.041s [fwd=2.850s bwd=1.176s opt=0.003s] dispatch=3.483s [2026-10-19 01:21:07,028] loop.py:45 INFO TRAIN_STEP step=5460 loss=1.486361 Training: 30%|███ | 5460/18018 [11:50:53<27:34:22, 7.90s/step, loss=1.486, acc=0.678, acc_len=1.61, thru=14.9, I=21.6, T=31.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:07,430] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:16,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 01:21:17,776] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5460: forward=3433.6ms backward=1178.7ms [2026-10-19 01:21:17,781] loop.py:45 INFO TRAIN_STEP step=5461 loss=1.549732 Training: 30%|███ | 5461/18018 [11:51:03<30:32:58, 8.76s/step, loss=1.550, acc=0.664, acc_len=1.56, thru=11.8, I=13.8, T=27.5, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:18,089] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:21:25,700] loop.py:45 INFO TRAIN_STEP step=5462 loss=1.529350 Training: 30%|███ | 5462/18018 [11:51:11<29:40:12, 8.51s/step, loss=1.529, acc=0.667, acc_len=1.57, thru=13.7, I=15.4, T=29.6, wait=3.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:33,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:21:35,087] loop.py:45 INFO TRAIN_STEP step=5463 loss=1.446152 Training: 30%|███ | 5463/18018 [11:51:21<30:35:21, 8.77s/step, loss=1.446, acc=0.679, acc_len=1.61, thru=15.7, I=16.5, T=31.1, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:35,325] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:21:44,314] loop.py:45 INFO TRAIN_STEP step=5464 loss=1.595390 Training: 30%|███ | 5464/18018 [11:51:30<31:03:36, 8.91s/step, loss=1.595, acc=0.649, acc_len=1.52, thru=14.2, I=13.6, T=30.0, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:44,306] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:44,807] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:50,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:21:52,823] loop.py:380 INFO TIMING step=5465: step=4.719s data=2.061s compute=4.684s [fwd=3.450s bwd=1.220s opt=0.003s] dispatch=3.783s [2026-10-19 01:21:52,823] loop.py:45 INFO TRAIN_STEP step=5465 loss=1.550000 Training: 30%|███ | 5465/18018 [11:51:38<30:38:36, 8.79s/step, loss=1.550, acc=0.659, acc_len=1.56, thru=16.3, I=19.1, T=27.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:21:53,116] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:00,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:22:00,413] loop.py:45 INFO TRAIN_STEP step=5466 loss=1.439913 Training: 30%|███ | 5466/18018 [11:51:46<29:23:16, 8.43s/step, loss=1.440, acc=0.682, acc_len=1.63, thru=13.6, I=18.2, T=35.4, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:22:00,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5465: forward=2545.6ms backward=1021.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:00,555] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:05,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:22:07,812] loop.py:45 INFO TRAIN_STEP step=5467 loss=1.488530 Training: 30%|███ | 5467/18018 [11:51:53<28:18:29, 8.12s/step, loss=1.489, acc=0.668, acc_len=1.58, thru=18.0, I=22.1, T=29.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:07,922] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:14,171] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:22:14,887] loop.py:45 INFO TRAIN_STEP step=5468 loss=1.519542 Training: 30%|███ | 5468/18018 [11:52:00<27:12:58, 7.81s/step, loss=1.520, acc=0.668, acc_len=1.58, thru=16.1, I=21.8, T=30.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:15,176] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:20,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:22:22,366] loop.py:45 INFO TRAIN_STEP step=5469 loss=1.509232 Training: 30%|███ | 5469/18018 [11:52:08<26:52:12, 7.71s/step, loss=1.509, acc=0.669, acc_len=1.58, thru=16.0, I=21.4, T=26.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:22,631] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 01:22:29,038] loop.py:380 INFO TIMING step=5470: step=3.897s data=1.647s compute=3.865s [fwd=2.760s bwd=1.090s opt=0.003s] dispatch=2.764s [2026-10-19 01:22:29,038] loop.py:45 INFO TRAIN_STEP step=5470 loss=1.527238 Training: 30%|███ | 5470/18018 [11:52:15<25:47:08, 7.40s/step, loss=1.527, acc=0.664, acc_len=1.57, thru=16.0, I=26.1, T=32.8, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:36,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:22:37,454] loop.py:45 INFO TRAIN_STEP step=5471 loss=1.516040 Training: 30%|███ | 5471/18018 [11:52:23<26:50:42, 7.70s/step, loss=1.516, acc=0.667, acc_len=1.57, thru=15.0, I=15.6, T=29.8, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:22:37,448] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5470: forward=2946.2ms backward=1294.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:37,705] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:42,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:22:44,341] loop.py:45 INFO TRAIN_STEP step=5472 loss=1.555671 Training: 30%|███ | 5472/18018 [11:52:30<25:59:31, 7.46s/step, loss=1.556, acc=0.662, acc_len=1.57, thru=18.2, I=22.9, T=29.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:48,343] inference_manager.py:398 INFO Still waiting for pool capacity (waited 6.2s): pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:22:48,343] inference_manager.py:408 INFO Pool capacity available after 6.2s, pool_size=16/64 [2026-10-19 01:22:55,046] loop.py:45 INFO TRAIN_STEP step=5473 loss=1.585255 Training: 30%|███ | 5473/18018 [11:52:41<29:22:54, 8.43s/step, loss=1.585, acc=0.655, acc_len=1.54, thru=15.6, I=18.1, T=32.6, wait=6.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:05,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:23:06,805] loop.py:45 INFO TRAIN_STEP step=5474 loss=1.657458 Training: 30%|███ | 5474/18018 [11:52:52<32:51:28, 9.43s/step, loss=1.657, acc=0.649, acc_len=1.52, thru=14.2, I=13.5, T=27.0, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:07,217] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:12,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:23:13,561] loop.py:380 INFO TIMING step=5475: step=3.955s data=1.762s compute=3.925s [fwd=2.755s bwd=1.156s opt=0.003s] dispatch=2.793s [2026-10-19 01:23:13,561] loop.py:45 INFO TRAIN_STEP step=5475 loss=1.500732 Training: 30%|███ | 5475/18018 [11:52:59<30:03:41, 8.63s/step, loss=1.501, acc=0.671, acc_len=1.59, thru=16.6, I=22.1, T=32.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:13,909] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:19,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:23:21,380] loop.py:45 INFO TRAIN_STEP step=5476 loss=1.495936 Training: 30%|███ | 5476/18018 [11:53:07<29:12:42, 8.38s/step, loss=1.496, acc=0.669, acc_len=1.58, thru=15.3, I=19.6, T=29.0, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:23:21,375] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5475: forward=3020.4ms backward=1346.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:21,748] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:23:27,583] loop.py:45 INFO TRAIN_STEP step=5477 loss=1.374954 Training: 30%|███ | 5477/18018 [11:53:13<26:55:48, 7.73s/step, loss=1.375, acc=0.686, acc_len=1.64, thru=17.9, I=20.4, T=30.5, wait=2.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:33,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:23:34,383] loop.py:45 INFO TRAIN_STEP step=5478 loss=1.519143 Training: 30%|███ | 5478/18018 [11:53:20<25:57:18, 7.45s/step, loss=1.519, acc=0.666, acc_len=1.57, thru=18.7, I=18.2, T=34.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:34,888] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:43,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:23:44,496] loop.py:45 INFO TRAIN_STEP step=5479 loss=1.703182 Training: 30%|███ | 5479/18018 [11:53:30<28:44:10, 8.25s/step, loss=1.703, acc=0.640, acc_len=1.49, thru=13.3, I=14.4, T=31.9, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:44,586] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-19 01:23:53,293] loop.py:380 INFO TIMING step=5480: step=3.664s data=1.608s compute=3.635s [fwd=2.534s bwd=1.087s opt=0.003s] dispatch=5.122s [2026-10-19 01:23:53,293] loop.py:45 INFO TRAIN_STEP step=5480 loss=1.511975 Training: 30%|███ | 5480/18018 [11:53:39<29:18:20, 8.41s/step, loss=1.512, acc=0.666, acc_len=1.57, thru=13.2, I=15.8, T=34.9, wait=5.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:23:59,892] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:24:00,862] loop.py:45 INFO TRAIN_STEP step=5481 loss=1.430006 Training: 30%|███ | 5481/18018 [11:53:46<28:25:04, 8.16s/step, loss=1.430, acc=0.683, acc_len=1.64, thru=17.2, I=18.6, T=28.3, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:24:00,856] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5480: forward=3239.7ms backward=1243.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:00,909] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 01:24:09,674] loop.py:45 INFO TRAIN_STEP step=5482 loss=1.562186 Training: 30%|███ | 5482/18018 [11:53:55<29:05:52, 8.36s/step, loss=1.562, acc=0.661, acc_len=1.56, thru=12.7, I=14.4, T=30.7, wait=4.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:17,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:24:18,241] loop.py:45 INFO TRAIN_STEP step=5483 loss=1.482531 Training: 30%|███ | 5483/18018 [11:54:04<29:18:51, 8.42s/step, loss=1.483, acc=0.673, acc_len=1.61, thru=16.0, I=16.4, T=28.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:18,287] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:23,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:24:25,339] loop.py:45 INFO TRAIN_STEP step=5484 loss=1.425506 Training: 30%|███ | 5484/18018 [11:54:11<27:55:56, 8.02s/step, loss=1.426, acc=0.681, acc_len=1.63, thru=17.9, I=20.7, T=29.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:25,609] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:31,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:24:33,254] loop.py:380 INFO TIMING step=5485: step=4.096s data=1.945s compute=4.051s [fwd=2.885s bwd=1.152s opt=0.003s] dispatch=3.811s [2026-10-19 01:24:33,254] loop.py:45 INFO TRAIN_STEP step=5485 loss=1.518326 Training: 30%|███ | 5485/18018 [11:54:19<27:49:03, 7.99s/step, loss=1.518, acc=0.664, acc_len=1.57, thru=16.8, I=19.2, T=31.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:33,376] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:41,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:24:42,115] loop.py:45 INFO TRAIN_STEP step=5486 loss=1.573010 Training: 30%|███ | 5486/18018 [11:54:28<28:43:30, 8.25s/step, loss=1.573, acc=0.656, acc_len=1.55, thru=14.7, I=15.6, T=31.8, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:24:42,110] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5485: forward=2887.1ms backward=1089.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:42,186] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-19 01:24:50,182] loop.py:45 INFO TRAIN_STEP step=5487 loss=1.477109 Training: 30%|███ | 5487/18018 [11:54:36<28:31:55, 8.20s/step, loss=1.477, acc=0.668, acc_len=1.58, thru=15.3, I=17.6, T=26.9, wait=3.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:24:57,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:24:59,930] loop.py:45 INFO TRAIN_STEP step=5488 loss=1.677438 Training: 30%|███ | 5488/18018 [11:54:46<30:08:54, 8.66s/step, loss=1.677, acc=0.639, acc_len=1.49, thru=14.1, I=15.6, T=23.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:00,086] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:06,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:25:08,316] loop.py:45 INFO TRAIN_STEP step=5489 loss=1.623080 Training: 30%|███ | 5489/18018 [11:54:54<29:51:27, 8.58s/step, loss=1.623, acc=0.646, acc_len=1.51, thru=15.6, I=20.5, T=29.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:08,712] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:14,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:25:15,208] loop.py:380 INFO TIMING step=5490: step=3.754s data=1.662s compute=3.720s [fwd=2.616s bwd=1.091s opt=0.003s] dispatch=3.127s [2026-10-19 01:25:15,208] loop.py:45 INFO TRAIN_STEP step=5490 loss=1.526527 Training: 30%|███ | 5490/18018 [11:55:01<28:05:40, 8.07s/step, loss=1.527, acc=0.665, acc_len=1.58, thru=16.6, I=24.9, T=34.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:15,267] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:20,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:25:20,973] loop.py:45 INFO TRAIN_STEP step=5491 loss=1.602236 Training: 30%|███ | 5491/18018 [11:55:07<25:40:54, 7.38s/step, loss=1.602, acc=0.652, acc_len=1.53, thru=20.8, I=23.8, T=33.4, wait=1.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:25:20,968] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5490: forward=2671.3ms backward=1118.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:21,375] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:32,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:25:34,285] loop.py:45 INFO TRAIN_STEP step=5492 loss=1.625826 Training: 30%|███ | 5492/18018 [11:55:20<31:52:21, 9.16s/step, loss=1.626, acc=0.653, acc_len=1.52, thru=10.6, I=10.8, T=30.0, wait=9.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:34,363] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:39,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:25:40,701] loop.py:45 INFO TRAIN_STEP step=5493 loss=1.521062 Training: 30%|███ | 5493/18018 [11:55:26<29:00:09, 8.34s/step, loss=1.521, acc=0.668, acc_len=1.58, thru=17.7, I=24.7, T=27.9, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:41,200] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:45,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:25:47,695] loop.py:45 INFO TRAIN_STEP step=5494 loss=1.530632 Training: 30%|███ | 5494/18018 [11:55:33<27:36:11, 7.93s/step, loss=1.531, acc=0.661, acc_len=1.56, thru=16.4, I=25.8, T=31.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:25:47,940] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 01:25:55,017] loop.py:380 INFO TIMING step=5495: step=4.091s data=1.917s compute=4.059s [fwd=2.842s bwd=1.202s opt=0.003s] dispatch=3.220s [2026-10-19 01:25:55,017] loop.py:45 INFO TRAIN_STEP step=5495 loss=1.532953 Training: 30%|███ | 5495/18018 [11:55:41<26:57:34, 7.75s/step, loss=1.533, acc=0.664, acc_len=1.57, thru=9.7, I=21.4, T=31.3, wait=3.2s, pool=16, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:06,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:26:07,827] loop.py:45 INFO TRAIN_STEP step=5496 loss=1.542527 Training: 31%|███ | 5496/18018 [11:55:53<32:14:20, 9.27s/step, loss=1.543, acc=0.668, acc_len=1.58, thru=13.2, I=15.3, T=31.4, wait=8.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:26:07,821] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5495: forward=2906.6ms backward=1125.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:07,915] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:16,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:26:17,277] loop.py:45 INFO TRAIN_STEP step=5497 loss=1.579147 Training: 31%|███ | 5497/18018 [11:56:03<32:25:35, 9.32s/step, loss=1.579, acc=0.657, acc_len=1.55, thru=13.7, I=14.8, T=31.8, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:17,317] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:25,660] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:26:27,330] loop.py:45 INFO TRAIN_STEP step=5498 loss=1.532143 Training: 31%|███ | 5498/18018 [11:56:13<33:11:05, 9.54s/step, loss=1.532, acc=0.664, acc_len=1.57, thru=13.3, I=14.0, T=30.1, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:27,673] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:26:34,412] loop.py:45 INFO TRAIN_STEP step=5499 loss=1.492390 Training: 31%|███ | 5499/18018 [11:56:20<30:36:47, 8.80s/step, loss=1.492, acc=0.671, acc_len=1.59, thru=14.8, I=18.6, T=30.6, wait=2.9s, pool=52, epoch=1/2][2026-10-19 01:26:42,195] loop.py:380 INFO TIMING step=5500: step=4.493s data=2.140s compute=4.461s [fwd=3.153s bwd=1.293s opt=0.003s] dispatch=3.282s [2026-10-19 01:26:42,196] loop.py:45 INFO TRAIN_STEP step=5500 loss=1.427986 Training: 31%|███ | 5500/18018 [11:56:28<29:32:53, 8.50s/step, loss=1.428, acc=0.677, acc_len=1.61, thru=15.0, I=14.3, T=28.5, wait=3.3s, pool=44, epoch=1/2][2026-10-19 01:26:42,199] loop.py:419 INFO Saving checkpoint at step 5500... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:42,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 01:26:45,667] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0005501 (TrainerActor pid=9310) [2026-10-19 01:26:45,674] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0004751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:47,304] inference_manager.py:408 INFO Pool capacity available after 4.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:52,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 01:26:45,668] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0005501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 01:26:46,837] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0004751 [2026-10-19 01:26:54,416] loop.py:45 INFO TRAIN_STEP step=5501 loss=1.551735 Training: 31%|███ | 5501/18018 [11:56:40<33:25:49, 9.61s/step, loss=1.552, acc=0.662, acc_len=1.56, thru=18.0, I=23.6, T=29.0, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:26:54,411] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5500: forward=3004.7ms backward=1326.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:26:54,757] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:03,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:27:04,586] loop.py:45 INFO TRAIN_STEP step=5502 loss=1.485828 Training: 31%|███ | 5502/18018 [11:56:50<34:00:17, 9.78s/step, loss=1.486, acc=0.672, acc_len=1.59, thru=13.0, I=13.7, T=27.1, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:04,940] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:27:11,468] loop.py:45 INFO TRAIN_STEP step=5503 loss=1.412336 Training: 31%|███ | 5503/18018 [11:56:57<30:58:45, 8.91s/step, loss=1.412, acc=0.683, acc_len=1.63, thru=16.2, I=19.8, T=29.3, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:17,680] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:27:18,868] loop.py:45 INFO TRAIN_STEP step=5504 loss=1.456188 Training: 31%|███ | 5504/18018 [11:57:04<29:24:06, 8.46s/step, loss=1.456, acc=0.677, acc_len=1.61, thru=17.8, I=17.9, T=32.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:19,191] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:25,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:27:26,649] loop.py:380 INFO TIMING step=5505: step=4.067s data=1.856s compute=4.023s [fwd=2.794s bwd=1.214s opt=0.003s] dispatch=3.704s [2026-10-19 01:27:26,649] loop.py:45 INFO TRAIN_STEP step=5505 loss=1.603129 Training: 31%|███ | 5505/18018 [11:57:12<28:41:29, 8.25s/step, loss=1.603, acc=0.649, acc_len=1.53, thru=15.3, I=20.3, T=31.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:26,981] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:32,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:27:33,098] loop.py:45 INFO TRAIN_STEP step=5506 loss=1.581073 Training: 31%|███ | 5506/18018 [11:57:19<26:48:24, 7.71s/step, loss=1.581, acc=0.653, acc_len=1.53, thru=18.5, I=20.7, T=34.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:27:33,093] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5505: forward=2629.5ms backward=1015.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:33,501] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:27:41,296] loop.py:45 INFO TRAIN_STEP step=5507 loss=1.572354 Training: 31%|███ | 5507/18018 [11:57:27<27:18:47, 7.86s/step, loss=1.572, acc=0.660, acc_len=1.56, thru=13.0, I=17.8, T=33.1, wait=4.3s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:49,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:27:51,003] loop.py:45 INFO TRAIN_STEP step=5508 loss=1.573445 Training: 31%|███ | 5508/18018 [11:57:37<29:14:07, 8.41s/step, loss=1.573, acc=0.657, acc_len=1.55, thru=17.2, I=15.9, T=31.0, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:51,322] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:56,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:27:57,480] loop.py:45 INFO TRAIN_STEP step=5509 loss=1.433324 Training: 31%|███ | 5509/18018 [11:57:43<27:12:53, 7.83s/step, loss=1.433, acc=0.679, acc_len=1.62, thru=18.1, I=23.3, T=36.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:27:57,890] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:28:04,396] loop.py:380 INFO TIMING step=5510: step=3.486s data=1.394s compute=3.452s [fwd=2.425s bwd=1.012s opt=0.003s] dispatch=3.421s [2026-10-19 01:28:04,397] loop.py:45 INFO TRAIN_STEP step=5510 loss=1.518890 Training: 31%|███ | 5510/18018 [11:57:50<26:15:33, 7.56s/step, loss=1.519, acc=0.667, acc_len=1.58, thru=17.1, I=20.2, T=36.7, wait=3.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:12,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:28:12,906] loop.py:45 INFO TRAIN_STEP step=5511 loss=1.407091 Training: 31%|███ | 5511/18018 [11:57:58<27:14:54, 7.84s/step, loss=1.407, acc=0.685, acc_len=1.63, thru=16.7, I=15.4, T=34.1, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:28:12,901] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5510: forward=2600.4ms backward=1109.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:13,115] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:18,906] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:28:20,480] loop.py:45 INFO TRAIN_STEP step=5512 loss=1.571659 Training: 31%|███ | 5512/18018 [11:58:06<26:57:46, 7.76s/step, loss=1.572, acc=0.661, acc_len=1.55, thru=17.2, I=19.8, T=30.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:20,909] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:26,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:28:28,597] loop.py:45 INFO TRAIN_STEP step=5513 loss=1.433604 Training: 31%|███ | 5513/18018 [11:58:14<27:19:55, 7.87s/step, loss=1.434, acc=0.679, acc_len=1.61, thru=14.0, I=19.9, T=30.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:28,741] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:35,395] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:28:36,154] loop.py:45 INFO TRAIN_STEP step=5514 loss=1.625953 Training: 31%|███ | 5514/18018 [11:58:22<27:00:27, 7.78s/step, loss=1.626, acc=0.645, acc_len=1.51, thru=14.5, I=19.4, T=32.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:36,401] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:28:47,587] loop.py:380 INFO TIMING step=5515: step=4.643s data=2.010s compute=4.608s [fwd=3.300s bwd=1.293s opt=0.003s] dispatch=6.779s [2026-10-19 01:28:47,587] loop.py:45 INFO TRAIN_STEP step=5515 loss=1.688490 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:47,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 Training: 31%|███ | 5515/18018 [11:58:33<30:48:50, 8.87s/step, loss=1.688, acc=0.638, acc_len=1.49, thru=11.9, I=10.8, T=27.6, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:48,004] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:59,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:28:59,603] loop.py:45 INFO TRAIN_STEP step=5516 loss=1.604998 Training: 31%|███ | 5516/18018 [11:58:45<34:05:12, 9.82s/step, loss=1.605, acc=0.649, acc_len=1.51, thru=11.3, I=10.9, T=24.9, wait=6.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:28:59,598] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5515: forward=3481.1ms backward=1597.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:28:59,814] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:07,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:29:09,942] loop.py:45 INFO TRAIN_STEP step=5517 loss=1.604107 Training: 31%|███ | 5517/18018 [11:58:56<34:37:46, 9.97s/step, loss=1.604, acc=0.653, acc_len=1.52, thru=12.9, I=15.4, T=29.4, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:10,332] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:15,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:29:16,489] loop.py:45 INFO TRAIN_STEP step=5518 loss=1.400351 Training: 31%|███ | 5518/18018 [11:59:02<31:03:36, 8.95s/step, loss=1.400, acc=0.684, acc_len=1.64, thru=17.4, I=26.6, T=30.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:16,808] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:21,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:29:23,480] loop.py:45 INFO TRAIN_STEP step=5519 loss=1.464420 Training: 31%|███ | 5519/18018 [11:59:09<29:01:19, 8.36s/step, loss=1.464, acc=0.672, acc_len=1.59, thru=17.2, I=25.1, T=27.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:23,799] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:29,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:29:31,135] loop.py:380 INFO TIMING step=5520: step=4.529s data=2.223s compute=4.502s [fwd=3.149s bwd=1.337s opt=0.003s] dispatch=3.116s [2026-10-19 01:29:31,135] loop.py:45 INFO TRAIN_STEP step=5520 loss=1.429833 Training: 31%|███ | 5520/18018 [11:59:17<28:17:12, 8.15s/step, loss=1.430, acc=0.684, acc_len=1.62, thru=14.4, I=21.8, T=28.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:31,472] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:36,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:29:37,739] loop.py:45 INFO TRAIN_STEP step=5521 loss=1.536301 Training: 31%|███ | 5521/18018 [11:59:23<26:40:38, 7.68s/step, loss=1.536, acc=0.666, acc_len=1.58, thru=17.0, I=23.0, T=33.8, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:29:37,734] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5520: forward=2680.1ms backward=1059.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:37,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:44,929] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:29:45,705] loop.py:45 INFO TRAIN_STEP step=5522 loss=1.598795 Training: 31%|███ | 5522/18018 [11:59:31<26:58:01, 7.77s/step, loss=1.599, acc=0.651, acc_len=1.52, thru=15.0, I=16.8, T=30.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:45,935] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:50,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:29:52,462] loop.py:45 INFO TRAIN_STEP step=5523 loss=1.552530 Training: 31%|███ | 5523/18018 [11:59:38<25:54:28, 7.46s/step, loss=1.553, acc=0.662, acc_len=1.56, thru=17.0, I=24.0, T=29.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:29:52,953] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 01:30:00,050] loop.py:45 INFO TRAIN_STEP step=5524 loss=1.401563 Training: 31%|███ | 5524/18018 [11:59:46<26:02:09, 7.50s/step, loss=1.402, acc=0.685, acc_len=1.64, thru=12.1, I=18.3, T=31.1, wait=3.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:09,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:30:10,252] loop.py:380 INFO TIMING step=5525: step=4.315s data=1.876s compute=4.284s [fwd=3.021s bwd=1.248s opt=0.003s] dispatch=5.879s [2026-10-19 01:30:10,252] loop.py:45 INFO TRAIN_STEP step=5525 loss=1.573605 Training: 31%|███ | 5525/18018 [11:59:56<28:50:46, 8.31s/step, loss=1.574, acc=0.657, acc_len=1.55, thru=14.6, I=13.6, T=29.7, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:10,307] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:18,499] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:30:20,575] loop.py:45 INFO TRAIN_STEP step=5526 loss=1.607625 Training: 31%|███ | 5526/18018 [12:00:06<30:56:05, 8.91s/step, loss=1.608, acc=0.655, acc_len=1.54, thru=12.4, I=14.5, T=27.0, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:30:20,570] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5525: forward=3380.4ms backward=1300.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:21,011] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:26,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:30:27,936] loop.py:45 INFO TRAIN_STEP step=5527 loss=1.599602 Training: 31%|███ | 5527/18018 [12:00:14<29:18:55, 8.45s/step, loss=1.600, acc=0.653, acc_len=1.54, thru=14.5, I=22.4, T=29.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:28,406] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:37,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:30:37,546] loop.py:45 INFO TRAIN_STEP step=5528 loss=1.556565 Training: 31%|███ | 5528/18018 [12:00:23<30:31:24, 8.80s/step, loss=1.557, acc=0.663, acc_len=1.56, thru=14.0, I=12.5, T=30.8, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:37,619] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:43,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:30:44,700] loop.py:45 INFO TRAIN_STEP step=5529 loss=1.512826 Training: 31%|███ | 5529/18018 [12:00:30<28:48:32, 8.30s/step, loss=1.513, acc=0.666, acc_len=1.58, thru=17.2, I=19.4, T=30.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:44,955] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:49,459] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:30:51,171] loop.py:380 INFO TIMING step=5530: step=3.855s data=1.723s compute=3.826s [fwd=2.699s bwd=1.113s opt=0.003s] dispatch=2.607s [2026-10-19 01:30:51,171] loop.py:45 INFO TRAIN_STEP step=5530 loss=1.612967 Training: 31%|███ | 5530/18018 [12:00:37<26:54:02, 7.75s/step, loss=1.613, acc=0.650, acc_len=1.52, thru=18.1, I=25.2, T=33.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:51,474] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:56,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:30:57,307] loop.py:45 INFO TRAIN_STEP step=5531 loss=1.451411 Training: 31%|███ | 5531/18018 [12:00:43<25:12:39, 7.27s/step, loss=1.451, acc=0.677, acc_len=1.61, thru=19.5, I=28.0, T=34.6, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:30:57,302] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5530: forward=2612.2ms backward=1035.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:30:57,788] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:03,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:31:04,008] loop.py:45 INFO TRAIN_STEP step=5532 loss=1.469219 Training: 31%|███ | 5532/18018 [12:00:50<24:37:21, 7.10s/step, loss=1.469, acc=0.671, acc_len=1.59, thru=18.9, I=24.0, T=32.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:04,149] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:11,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:31:12,519] loop.py:45 INFO TRAIN_STEP step=5533 loss=1.632592 Training: 31%|███ | 5533/18018 [12:00:58<26:05:11, 7.52s/step, loss=1.633, acc=0.644, acc_len=1.51, thru=14.5, I=16.3, T=27.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:12,672] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:18,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:31:20,537] loop.py:45 INFO TRAIN_STEP step=5534 loss=1.602194 Training: 31%|███ | 5534/18018 [12:01:06<26:36:06, 7.67s/step, loss=1.602, acc=0.652, acc_len=1.53, thru=15.4, I=19.4, T=33.5, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:20,630] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:27,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:31:28,476] loop.py:380 INFO TIMING step=5535: step=3.892s data=1.695s compute=3.861s [fwd=2.670s bwd=1.176s opt=0.003s] dispatch=4.036s [2026-10-19 01:31:28,476] loop.py:45 INFO TRAIN_STEP step=5535 loss=1.469264 Training: 31%|███ | 5535/18018 [12:01:14<26:52:43, 7.75s/step, loss=1.469, acc=0.677, acc_len=1.62, thru=13.8, I=18.8, T=32.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:28,699] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:31:35,066] loop.py:45 INFO TRAIN_STEP step=5536 loss=1.493948 Training: 31%|███ | 5536/18018 [12:01:21<25:39:58, 7.40s/step, loss=1.494, acc=0.664, acc_len=1.57, thru=19.5, I=22.3, T=31.0, wait=2.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:31:35,061] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5535: forward=2831.0ms backward=1240.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:41,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:31:44,744] loop.py:45 INFO TRAIN_STEP step=5537 loss=1.494516 Training: 31%|███ | 5537/18018 [12:01:30<28:01:50, 8.09s/step, loss=1.495, acc=0.666, acc_len=1.57, thru=13.8, I=17.9, T=25.5, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:45,179] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:49,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:31:51,561] loop.py:45 INFO TRAIN_STEP step=5538 loss=1.566252 Training: 31%|███ | 5538/18018 [12:01:37<26:42:41, 7.71s/step, loss=1.566, acc=0.655, acc_len=1.54, thru=13.3, I=28.6, T=30.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:51,958] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:55,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:31:57,938] loop.py:45 INFO TRAIN_STEP step=5539 loss=1.478292 Training: 31%|███ | 5539/18018 [12:01:44<25:19:37, 7.31s/step, loss=1.478, acc=0.677, acc_len=1.61, thru=18.0, I=30.2, T=30.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:31:58,261] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:01,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:32:04,443] loop.py:380 INFO TIMING step=5540: step=4.396s data=2.111s compute=4.363s [fwd=3.045s bwd=1.302s opt=0.003s] dispatch=2.099s [2026-10-19 01:32:04,444] loop.py:45 INFO TRAIN_STEP step=5540 loss=1.418487 Training: 31%|███ | 5540/18018 [12:01:50<24:29:26, 7.07s/step, loss=1.418, acc=0.684, acc_len=1.63, thru=17.0, I=30.9, T=29.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:04,891] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 01:32:11,711] loop.py:45 INFO TRAIN_STEP step=5541 loss=1.412342 Training: 31%|███ | 5541/18018 [12:01:57<24:41:56, 7.13s/step, loss=1.412, acc=0.679, acc_len=1.62, thru=12.2, I=17.8, T=32.2, wait=3.3s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:32:11,707] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5540: forward=2762.0ms backward=1162.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:20,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:32:20,881] loop.py:45 INFO TRAIN_STEP step=5542 loss=1.588431 Training: 31%|███ | 5542/18018 [12:02:06<26:49:26, 7.74s/step, loss=1.588, acc=0.654, acc_len=1.53, thru=15.5, I=16.1, T=28.0, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:21,228] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:26,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:32:27,749] loop.py:45 INFO TRAIN_STEP step=5543 loss=1.535393 Training: 31%|███ | 5543/18018 [12:02:13<25:54:54, 7.48s/step, loss=1.535, acc=0.664, acc_len=1.56, thru=16.9, I=18.4, T=31.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:27,917] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:36,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:32:36,873] loop.py:45 INFO TRAIN_STEP step=5544 loss=1.527126 Training: 31%|███ | 5544/18018 [12:02:22<27:37:15, 7.97s/step, loss=1.527, acc=0.664, acc_len=1.57, thru=13.2, I=14.3, T=31.3, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:37,271] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:44,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:32:44,910] loop.py:380 INFO TIMING step=5545: step=5.161s data=2.516s compute=5.128s [fwd=3.567s bwd=1.545s opt=0.003s] dispatch=2.868s [2026-10-19 01:32:44,910] loop.py:45 INFO TRAIN_STEP step=5545 loss=1.610692 Training: 31%|███ | 5545/18018 [12:02:30<27:41:17, 7.99s/step, loss=1.611, acc=0.649, acc_len=1.52, thru=15.9, I=17.4, T=24.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:45,357] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:53,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:32:55,811] loop.py:45 INFO TRAIN_STEP step=5546 loss=1.588140 Training: 31%|███ | 5546/18018 [12:02:41<30:42:32, 8.86s/step, loss=1.588, acc=0.657, acc_len=1.55, thru=12.5, I=14.6, T=28.1, wait=6.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:32:55,806] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5545: forward=3230.4ms backward=1272.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:32:56,098] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:01,171] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:33:02,199] loop.py:45 INFO TRAIN_STEP step=5547 loss=1.662929 Training: 31%|███ | 5547/18018 [12:02:48<28:08:00, 8.12s/step, loss=1.663, acc=0.647, acc_len=1.51, thru=17.2, I=25.4, T=32.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:02,679] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:33:09,379] loop.py:45 INFO TRAIN_STEP step=5548 loss=1.618245 Training: 31%|███ | 5548/18018 [12:02:55<27:09:06, 7.84s/step, loss=1.618, acc=0.645, acc_len=1.50, thru=15.5, I=17.5, T=27.5, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:15,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:33:17,879] loop.py:45 INFO TRAIN_STEP step=5549 loss=1.695234 Training: 31%|███ | 5549/18018 [12:03:03<27:50:17, 8.04s/step, loss=1.695, acc=0.636, acc_len=1.48, thru=16.6, I=18.9, T=25.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:18,200] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:23,416] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:33:24,924] loop.py:380 INFO TIMING step=5550: step=4.032s data=1.687s compute=3.991s [fwd=2.764s bwd=1.212s opt=0.003s] dispatch=3.004s [2026-10-19 01:33:24,924] loop.py:45 INFO TRAIN_STEP step=5550 loss=1.508968 Training: 31%|███ | 5550/18018 [12:03:11<26:48:25, 7.74s/step, loss=1.509, acc=0.670, acc_len=1.59, thru=14.6, I=22.9, T=31.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:25,431] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:31,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:33:32,527] loop.py:45 INFO TRAIN_STEP step=5551 loss=1.483061 Training: 31%|███ | 5551/18018 [12:03:18<26:39:43, 7.70s/step, loss=1.483, acc=0.671, acc_len=1.59, thru=15.2, I=21.2, T=34.2, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:33:32,521] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5550: forward=2625.9ms backward=1070.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:32,876] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:38,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:33:38,981] loop.py:45 INFO TRAIN_STEP step=5552 loss=1.520326 Training: 31%|███ | 5552/18018 [12:03:25<25:21:47, 7.32s/step, loss=1.520, acc=0.664, acc_len=1.57, thru=19.6, I=20.1, T=30.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:39,380] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:47,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:33:48,782] loop.py:45 INFO TRAIN_STEP step=5553 loss=1.472283 Training: 31%|███ | 5553/18018 [12:03:34<27:56:01, 8.07s/step, loss=1.472, acc=0.673, acc_len=1.60, thru=13.6, I=16.3, T=30.0, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:33:49,195] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:33:56,574] loop.py:45 INFO TRAIN_STEP step=5554 loss=1.492106 Training: 31%|███ | 5554/18018 [12:03:42<27:38:56, 7.99s/step, loss=1.492, acc=0.673, acc_len=1.60, thru=14.1, I=18.6, T=27.7, wait=3.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:05,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:34:05,936] loop.py:380 INFO TIMING step=5555: step=4.574s data=2.157s compute=4.509s [fwd=3.212s bwd=1.282s opt=0.003s] dispatch=4.777s [2026-10-19 01:34:05,937] loop.py:45 INFO TRAIN_STEP step=5555 loss=1.518465 Training: 31%|███ | 5555/18018 [12:03:52<29:04:27, 8.40s/step, loss=1.518, acc=0.670, acc_len=1.59, thru=16.0, I=13.0, T=28.0, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:06,274] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:10,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:34:13,494] loop.py:45 INFO TRAIN_STEP step=5556 loss=1.381794 Training: 31%|███ | 5556/18018 [12:03:59<28:11:48, 8.15s/step, loss=1.382, acc=0.688, acc_len=1.65, thru=18.4, I=20.9, T=26.3, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:34:13,488] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5555: forward=3336.7ms backward=1485.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:13,827] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:18,880] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:34:20,718] loop.py:45 INFO TRAIN_STEP step=5557 loss=1.400670 Training: 31%|███ | 5557/18018 [12:04:06<27:14:28, 7.87s/step, loss=1.401, acc=0.684, acc_len=1.63, thru=13.7, I=24.7, T=27.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:20,893] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:28,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:34:29,620] loop.py:45 INFO TRAIN_STEP step=5558 loss=1.422513 Training: 31%|███ | 5558/18018 [12:04:15<28:18:38, 8.18s/step, loss=1.423, acc=0.685, acc_len=1.64, thru=14.7, I=15.2, T=26.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:29,729] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 [2026-10-19 01:34:35,116] loop.py:45 INFO TRAIN_STEP step=5559 loss=1.532639 Training: 31%|███ | 5559/18018 [12:04:21<25:31:18, 7.37s/step, loss=1.533, acc=0.664, acc_len=1.56, thru=18.7, I=23.2, T=33.3, wait=1.6s, pool=60, epoch=1/2][2026-10-19 01:34:43,579] loop.py:380 INFO TIMING step=5560: step=4.528s data=2.168s compute=4.498s [fwd=3.222s bwd=1.260s opt=0.003s] dispatch=3.925s [2026-10-19 01:34:43,580] loop.py:45 INFO TRAIN_STEP step=5560 loss=1.477188 Training: 31%|███ | 5560/18018 [12:04:29<26:38:55, 7.70s/step, loss=1.477, acc=0.673, acc_len=1.59, thru=14.2, I=15.4, T=28.3, wait=3.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:53,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:34:56,635] loop.py:45 INFO TRAIN_STEP step=5561 loss=1.744535 Training: 31%|███ | 5561/18018 [12:04:42<32:12:19, 9.31s/step, loss=1.745, acc=0.630, acc_len=1.46, thru=12.5, I=12.0, T=23.8, wait=7.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:34:56,631] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5560: forward=3669.1ms backward=1632.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:34:56,965] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:02,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:35:03,588] loop.py:45 INFO TRAIN_STEP step=5562 loss=1.481590 Training: 31%|███ | 5562/18018 [12:04:49<29:45:34, 8.60s/step, loss=1.482, acc=0.673, acc_len=1.60, thru=14.0, I=21.9, T=31.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:03,959] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:35:10,780] loop.py:45 INFO TRAIN_STEP step=5563 loss=1.594275 Training: 31%|███ | 5563/18018 [12:04:56<28:17:46, 8.18s/step, loss=1.594, acc=0.656, acc_len=1.55, thru=16.3, I=18.5, T=33.3, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:15,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:35:17,626] loop.py:45 INFO TRAIN_STEP step=5564 loss=1.583740 Training: 31%|███ | 5564/18018 [12:05:03<26:54:33, 7.78s/step, loss=1.584, acc=0.657, acc_len=1.54, thru=19.0, I=21.2, T=28.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:17,654] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:22,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:35:23,836] loop.py:380 INFO TIMING step=5565: step=3.890s data=1.723s compute=3.862s [fwd=2.724s bwd=1.123s opt=0.003s] dispatch=2.310s [2026-10-19 01:35:23,836] loop.py:45 INFO TRAIN_STEP step=5565 loss=1.468580 Training: 31%|███ | 5565/18018 [12:05:09<25:16:42, 7.31s/step, loss=1.469, acc=0.672, acc_len=1.59, thru=17.5, I=24.2, T=32.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:24,230] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:30,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:35:30,753] loop.py:45 INFO TRAIN_STEP step=5566 loss=1.542063 Training: 31%|███ | 5566/18018 [12:05:16<24:52:20, 7.19s/step, loss=1.542, acc=0.660, acc_len=1.56, thru=17.9, I=23.1, T=33.8, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:35:30,747] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5565: forward=2651.1ms backward=1082.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:31,086] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:42,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:35:43,770] loop.py:45 INFO TRAIN_STEP step=5567 loss=1.494452 Training: 31%|███ | 5567/18018 [12:05:29<30:54:54, 8.94s/step, loss=1.494, acc=0.674, acc_len=1.60, thru=11.2, I=11.5, T=28.5, wait=8.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:43,886] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:50,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:35:51,313] loop.py:45 INFO TRAIN_STEP step=5568 loss=1.477240 Training: 31%|███ | 5568/18018 [12:05:37<29:27:48, 8.52s/step, loss=1.477, acc=0.674, acc_len=1.60, thru=14.9, I=19.2, T=30.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:35:51,738] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:35:59,054] loop.py:45 INFO TRAIN_STEP step=5569 loss=1.539783 Training: 31%|███ | 5569/18018 [12:05:45<28:39:10, 8.29s/step, loss=1.540, acc=0.666, acc_len=1.58, thru=14.4, I=14.6, T=29.9, wait=3.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:05,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:36:05,889] loop.py:380 INFO TIMING step=5570: step=3.830s data=1.684s compute=3.803s [fwd=2.648s bwd=1.140s opt=0.003s] dispatch=2.997s [2026-10-19 01:36:05,889] loop.py:45 INFO TRAIN_STEP step=5570 loss=1.525823 Training: 31%|███ | 5570/18018 [12:05:51<27:08:53, 7.85s/step, loss=1.526, acc=0.664, acc_len=1.56, thru=19.0, I=18.9, T=33.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:06,250] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:12,955] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:36:13,662] loop.py:45 INFO TRAIN_STEP step=5571 loss=1.592579 Training: 31%|███ | 5571/18018 [12:05:59<27:03:56, 7.83s/step, loss=1.593, acc=0.652, acc_len=1.54, thru=16.0, I=18.6, T=27.1, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:36:13,657] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5570: forward=3277.7ms backward=1382.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:13,959] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:19,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:36:21,724] loop.py:45 INFO TRAIN_STEP step=5572 loss=1.533242 Training: 31%|███ | 5572/18018 [12:06:07<27:18:15, 7.90s/step, loss=1.533, acc=0.661, acc_len=1.56, thru=15.3, I=20.8, T=29.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:22,078] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:25,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:36:27,728] loop.py:45 INFO TRAIN_STEP step=5573 loss=1.538863 Training: 31%|███ | 5573/18018 [12:06:13<25:20:25, 7.33s/step, loss=1.539, acc=0.660, acc_len=1.55, thru=19.1, I=30.8, T=31.7, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:27,809] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:34,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:36:34,468] loop.py:45 INFO TRAIN_STEP step=5574 loss=1.505423 Training: 31%|███ | 5574/18018 [12:06:20<24:43:28, 7.15s/step, loss=1.505, acc=0.668, acc_len=1.57, thru=16.8, I=21.4, T=29.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:34,584] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:40,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:36:40,951] loop.py:380 INFO TIMING step=5575: step=4.168s data=1.934s compute=4.136s [fwd=2.972s bwd=1.150s opt=0.003s] dispatch=2.305s [2026-10-19 01:36:40,951] loop.py:45 INFO TRAIN_STEP step=5575 loss=1.524230 Training: 31%|███ | 5575/18018 [12:06:27<24:01:37, 6.95s/step, loss=1.524, acc=0.662, acc_len=1.56, thru=18.1, I=19.6, T=30.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:41,109] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:46,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:36:47,719] loop.py:45 INFO TRAIN_STEP step=5576 loss=1.549357 Training: 31%|███ | 5576/18018 [12:06:33<23:50:12, 6.90s/step, loss=1.549, acc=0.657, acc_len=1.55, thru=17.6, I=22.1, T=30.5, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:36:47,715] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5575: forward=2947.9ms backward=1205.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:36:48,226] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:36:55,023] loop.py:45 INFO TRAIN_STEP step=5577 loss=1.578195 Training: 31%|███ | 5577/18018 [12:06:41<24:15:16, 7.02s/step, loss=1.578, acc=0.655, acc_len=1.54, thru=14.1, I=18.0, T=29.4, wait=2.9s, pool=44, epoch=1/2][2026-10-19 01:37:03,238] loop.py:45 INFO TRAIN_STEP step=5578 loss=1.557534 Training: 31%|███ | 5578/18018 [12:06:49<25:29:31, 7.38s/step, loss=1.558, acc=0.657, acc_len=1.55, thru=16.7, I=14.5, T=33.0, wait=4.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:07,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:37:09,694] loop.py:45 INFO TRAIN_STEP step=5579 loss=1.446192 Training: 31%|███ | 5579/18018 [12:06:55<24:32:18, 7.10s/step, loss=1.446, acc=0.674, acc_len=1.60, thru=19.4, I=22.9, T=30.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:09,825] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:14,866] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:37:17,555] loop.py:380 INFO TIMING step=5580: step=5.492s data=3.327s compute=5.457s [fwd=4.277s bwd=1.166s opt=0.003s] dispatch=2.358s [2026-10-19 01:37:17,556] loop.py:45 INFO TRAIN_STEP step=5580 loss=1.555285 Training: 31%|███ | 5580/18018 [12:07:03<25:19:24, 7.33s/step, loss=1.555, acc=0.657, acc_len=1.55, thru=13.8, I=23.6, T=23.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:17,885] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:24,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:37:24,765] loop.py:45 INFO TRAIN_STEP step=5581 loss=1.512100 Training: 31%|███ | 5581/18018 [12:07:10<25:11:49, 7.29s/step, loss=1.512, acc=0.662, acc_len=1.56, thru=13.3, I=20.0, T=34.4, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:37:24,759] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5580: forward=2590.2ms backward=1089.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:25,056] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:31,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:37:33,255] loop.py:45 INFO TRAIN_STEP step=5582 loss=1.594507 Training: 31%|███ | 5582/18018 [12:07:19<26:26:00, 7.65s/step, loss=1.595, acc=0.656, acc_len=1.53, thru=15.2, I=19.4, T=28.8, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:33,341] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 01:37:39,507] loop.py:45 INFO TRAIN_STEP step=5583 loss=1.603298 Training: 31%|███ | 5583/18018 [12:07:25<24:58:45, 7.23s/step, loss=1.603, acc=0.652, acc_len=1.53, thru=17.4, I=23.5, T=33.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:39,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:37:40,008] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:37:46,503] loop.py:45 INFO TRAIN_STEP step=5584 loss=1.592798 Training: 31%|███ | 5584/18018 [12:07:32<24:44:03, 7.16s/step, loss=1.593, acc=0.652, acc_len=1.53, thru=17.3, I=20.7, T=30.8, wait=2.8s, pool=60, epoch=1/2][2026-10-19 01:37:53,747] loop.py:380 INFO TIMING step=5585: step=4.282s data=1.902s compute=4.254s [fwd=2.936s bwd=1.303s opt=0.003s] dispatch=2.953s [2026-10-19 01:37:53,747] loop.py:45 INFO TRAIN_STEP step=5585 loss=1.542678 Training: 31%|███ | 5585/18018 [12:07:39<24:49:15, 7.19s/step, loss=1.543, acc=0.661, acc_len=1.56, thru=15.6, I=16.9, T=29.9, wait=3.0s, pool=52, epoch=1/2][2026-10-19 01:38:02,305] loop.py:45 INFO TRAIN_STEP step=5586 loss=1.413216 Training: 31%|███ | 5586/18018 [12:07:48<26:14:10, 7.60s/step, loss=1.413, acc=0.686, acc_len=1.65, thru=14.7, I=14.1, T=27.3, wait=3.8s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:38:02,300] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5585: forward=3169.8ms backward=1464.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:08,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:38:10,863] loop.py:45 INFO TRAIN_STEP step=5587 loss=1.537738 Training: 31%|███ | 5587/18018 [12:07:56<27:13:53, 7.89s/step, loss=1.538, acc=0.666, acc_len=1.57, thru=15.8, I=16.7, T=25.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:11,083] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:20,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:38:20,896] loop.py:45 INFO TRAIN_STEP step=5588 loss=1.559876 Training: 31%|███ | 5588/18018 [12:08:06<29:27:12, 8.53s/step, loss=1.560, acc=0.656, acc_len=1.55, thru=13.1, I=13.6, T=31.6, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:21,403] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:38:28,349] loop.py:45 INFO TRAIN_STEP step=5589 loss=1.522881 Training: 31%|███ | 5589/18018 [12:08:14<28:19:58, 8.21s/step, loss=1.523, acc=0.666, acc_len=1.57, thru=16.1, I=18.0, T=32.9, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:35,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:38:37,735] loop.py:380 INFO TIMING step=5590: step=5.225s data=2.471s compute=5.187s [fwd=3.711s bwd=1.461s opt=0.003s] dispatch=4.151s [2026-10-19 01:38:37,735] loop.py:45 INFO TRAIN_STEP step=5590 loss=1.474139 Training: 31%|███ | 5590/18018 [12:08:23<29:33:12, 8.56s/step, loss=1.474, acc=0.676, acc_len=1.61, thru=15.1, I=18.2, T=24.5, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:38,137] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 01:38:44,950] loop.py:45 INFO TRAIN_STEP step=5591 loss=1.445978 Training: 31%|███ | 5591/18018 [12:08:31<28:09:25, 8.16s/step, loss=1.446, acc=0.678, acc_len=1.62, thru=14.1, I=20.0, T=30.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:44,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 01:38:44,945] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5590: forward=2893.8ms backward=1209.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:45,371] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:52,071] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:38:53,738] loop.py:45 INFO TRAIN_STEP step=5592 loss=1.477021 Training: 31%|███ | 5592/18018 [12:08:39<28:48:26, 8.35s/step, loss=1.477, acc=0.673, acc_len=1.60, thru=14.6, I=17.8, T=32.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:38:54,080] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:39:00,407] loop.py:45 INFO TRAIN_STEP step=5593 loss=1.515134 Training: 31%|███ | 5593/18018 [12:08:46<27:04:07, 7.84s/step, loss=1.515, acc=0.665, acc_len=1.57, thru=14.4, I=17.7, T=34.7, wait=3.0s, pool=36, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 01:39:00,475] eagle_store.py:517 WARNING batch_get_buffer missing keys (k178979-aadd380d_hs, k178979-aadd380d_ids, k178979-aadd380d_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:07,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:39:08,551] loop.py:45 INFO TRAIN_STEP step=5594 loss=1.470479 Training: 31%|███ | 5594/18018 [12:08:54<27:22:49, 7.93s/step, loss=1.470, acc=0.675, acc_len=1.60, thru=19.4, I=16.7, T=29.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:08,808] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:39:15,229] loop.py:380 INFO TIMING step=5595: step=4.108s data=1.867s compute=4.080s [fwd=2.923s bwd=1.142s opt=0.003s] dispatch=2.559s [2026-10-19 01:39:15,229] loop.py:45 INFO TRAIN_STEP step=5595 loss=1.528677 Training: 31%|███ | 5595/18018 [12:09:01<26:04:33, 7.56s/step, loss=1.529, acc=0.665, acc_len=1.57, thru=15.9, I=17.7, T=31.2, wait=2.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:22,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:39:24,847] loop.py:45 INFO TRAIN_STEP step=5596 loss=1.400523 Training: 31%|███ | 5596/18018 [12:09:10<28:12:26, 8.17s/step, loss=1.401, acc=0.688, acc_len=1.65, thru=15.1, I=16.7, T=24.9, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:39:24,842] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5595: forward=3553.1ms backward=1538.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:25,344] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:32,131] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:39:32,402] loop.py:45 INFO TRAIN_STEP step=5597 loss=1.477810 Training: 31%|███ | 5597/18018 [12:09:18<27:33:49, 7.99s/step, loss=1.478, acc=0.672, acc_len=1.59, thru=14.3, I=20.5, T=28.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:32,633] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:40,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:39:40,650] loop.py:45 INFO TRAIN_STEP step=5598 loss=1.466827 Training: 31%|███ | 5598/18018 [12:09:26<27:49:46, 8.07s/step, loss=1.467, acc=0.674, acc_len=1.60, thru=14.2, I=14.3, T=27.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:41,087] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:47,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:39:49,871] loop.py:45 INFO TRAIN_STEP step=5599 loss=1.468227 Training: 31%|███ | 5599/18018 [12:09:35<29:01:27, 8.41s/step, loss=1.468, acc=0.673, acc_len=1.60, thru=14.4, I=17.3, T=26.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:39:50,060] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 01:39:57,014] loop.py:380 INFO TIMING step=5600: step=3.717s data=1.601s compute=3.686s [fwd=2.626s bwd=1.045s opt=0.003s] dispatch=3.415s [2026-10-19 01:39:57,014] loop.py:45 INFO TRAIN_STEP step=5600 loss=1.563104 Training: 31%|███ | 5600/18018 [12:09:43<27:42:18, 8.03s/step, loss=1.563, acc=0.660, acc_len=1.55, thru=13.3, I=17.7, T=34.4, wait=3.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:06,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:40:06,862] loop.py:45 INFO TRAIN_STEP step=5601 loss=1.429931 Training: 31%|███ | 5601/18018 [12:09:52<29:35:03, 8.58s/step, loss=1.430, acc=0.679, acc_len=1.62, thru=14.4, I=15.3, T=30.3, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:40:06,857] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5600: forward=2971.1ms backward=1205.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:07,234] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:12,485] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:40:12,786] loop.py:45 INFO TRAIN_STEP step=5602 loss=1.423741 Training: 31%|███ | 5602/18018 [12:09:58<26:50:10, 7.78s/step, loss=1.424, acc=0.680, acc_len=1.62, thru=19.6, I=20.6, T=38.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:12,987] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:19,760] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:40:20,725] loop.py:45 INFO TRAIN_STEP step=5603 loss=1.544752 Training: 31%|███ | 5603/18018 [12:10:06<26:59:51, 7.83s/step, loss=1.545, acc=0.666, acc_len=1.57, thru=15.9, I=17.8, T=30.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:20,790] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:27,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:40:28,920] loop.py:45 INFO TRAIN_STEP step=5604 loss=1.449399 Training: 31%|███ | 5604/18018 [12:10:15<27:22:29, 7.94s/step, loss=1.449, acc=0.674, acc_len=1.60, thru=14.5, I=17.2, T=29.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:28,994] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:33,583] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:40:35,559] loop.py:380 INFO TIMING step=5605: step=3.873s data=1.761s compute=3.845s [fwd=2.727s bwd=1.104s opt=0.003s] dispatch=2.755s [2026-10-19 01:40:35,559] loop.py:45 INFO TRAIN_STEP step=5605 loss=1.476653 Training: 31%|███ | 5605/18018 [12:10:21<26:01:42, 7.55s/step, loss=1.477, acc=0.671, acc_len=1.59, thru=18.1, I=26.6, T=33.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:35,601] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:41,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:40:42,000] loop.py:45 INFO TRAIN_STEP step=5606 loss=1.417115 Training: 31%|███ | 5606/18018 [12:10:28<24:52:43, 7.22s/step, loss=1.417, acc=0.679, acc_len=1.62, thru=18.5, I=26.1, T=35.4, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:40:41,995] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5605: forward=2568.9ms backward=998.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:42,044] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:47,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:40:49,335] loop.py:45 INFO TRAIN_STEP step=5607 loss=1.522682 Training: 31%|███ | 5607/18018 [12:10:35<24:59:59, 7.25s/step, loss=1.523, acc=0.665, acc_len=1.57, thru=16.8, I=21.3, T=28.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:49,637] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:55,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:40:56,293] loop.py:45 INFO TRAIN_STEP step=5608 loss=1.564540 Training: 31%|███ | 5608/18018 [12:10:42<24:41:45, 7.16s/step, loss=1.565, acc=0.656, acc_len=1.54, thru=18.1, I=23.1, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:40:56,403] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 01:41:05,256] loop.py:45 INFO TRAIN_STEP step=5609 loss=1.472785 Training: 31%|███ | 5609/18018 [12:10:51<26:33:11, 7.70s/step, loss=1.473, acc=0.676, acc_len=1.60, thru=9.3, I=14.3, T=30.0, wait=4.7s, pool=24, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:41:13,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:41:14,552] loop.py:380 INFO TIMING step=5610: step=3.668s data=1.639s compute=3.613s [fwd=2.487s bwd=1.111s opt=0.003s] dispatch=5.618s [2026-10-19 01:41:14,552] loop.py:45 INFO TRAIN_STEP step=5610 loss=1.485713 Training: 31%|███ | 5610/18018 [12:11:00<28:11:50, 8.18s/step, loss=1.486, acc=0.671, acc_len=1.59, thru=17.4, I=16.0, T=34.9, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:41:14,941] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:41:20,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:41:21,755] loop.py:45 INFO TRAIN_STEP step=5611 loss=1.421648 Training: 31%|███ | 5611/18018 [12:11:07<27:11:08, 7.89s/step, loss=1.422, acc=0.679, acc_len=1.61, thru=17.4, I=22.4, T=28.8, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:41:21,750] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5610: forward=3000.0ms backward=1394.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:41:21,980] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:41:27,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:41:29,412] loop.py:45 INFO TRAIN_STEP step=5612 loss=1.593700 Training: 31%|███ | 5612/18018 [12:11:15<26:56:39, 7.82s/step, loss=1.594, acc=0.655, acc_len=1.53, thru=16.6, I=21.2, T=29.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:41:29,800] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:41:36,991] loop.py:45 INFO TRAIN_STEP step=5613 loss=1.538450 Training: 31%|███ | 5613/18018 [12:11:23<26:41:55, 7.75s/step, loss=1.538, acc=0.658, acc_len=1.55, thru=12.6, I=18.8, T=31.1, wait=3.4s, pool=36, epoch=1/2][2026-10-19 01:41:47,159] loop.py:45 INFO TRAIN_STEP step=5614 loss=1.681343 Training: 31%|███ | 5614/18018 [12:11:33<29:11:33, 8.47s/step, loss=1.681, acc=0.639, acc_len=1.49, thru=15.5, I=13.9, T=27.2, wait=5.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:41:54,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:41:55,625] loop.py:380 INFO TIMING step=5615: step=4.294s data=2.017s compute=4.231s [fwd=2.928s bwd=1.288s opt=0.003s] dispatch=4.161s [2026-10-19 01:41:55,625] loop.py:45 INFO TRAIN_STEP step=5615 loss=1.525068 Training: 31%|███ | 5615/18018 [12:11:41<29:10:52, 8.47s/step, loss=1.525, acc=0.664, acc_len=1.57, thru=15.8, I=18.5, T=29.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:41:56,061] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:42:01,511] loop.py:45 INFO TRAIN_STEP step=5616 loss=1.615751 Training: 31%|███ | 5616/18018 [12:11:47<26:30:39, 7.70s/step, loss=1.616, acc=0.648, acc_len=1.52, thru=17.3, I=27.0, T=37.5, wait=2.5s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:42:01,506] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5615: forward=2440.4ms backward=932.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:10,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:42:11,088] loop.py:45 INFO TRAIN_STEP step=5617 loss=1.401327 Training: 31%|███ | 5617/18018 [12:11:57<28:27:06, 8.26s/step, loss=1.401, acc=0.688, acc_len=1.64, thru=15.2, I=13.3, T=31.1, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:11,260] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:18,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:42:18,321] loop.py:45 INFO TRAIN_STEP step=5618 loss=1.604490 Training: 31%|███ | 5618/18018 [12:12:04<27:23:23, 7.95s/step, loss=1.604, acc=0.653, acc_len=1.54, thru=16.7, I=19.6, T=33.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:18,688] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:42:26,118] loop.py:45 INFO TRAIN_STEP step=5619 loss=1.469797 Training: 31%|███ | 5619/18018 [12:12:12<27:13:34, 7.90s/step, loss=1.470, acc=0.679, acc_len=1.62, thru=13.6, I=15.1, T=32.4, wait=3.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:33,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:42:33,450] loop.py:380 INFO TIMING step=5620: step=3.865s data=1.754s compute=3.825s [fwd=2.710s bwd=1.100s opt=0.003s] dispatch=3.458s [2026-10-19 01:42:33,451] loop.py:45 INFO TRAIN_STEP step=5620 loss=1.496030 Training: 31%|███ | 5620/18018 [12:12:19<26:38:04, 7.73s/step, loss=1.496, acc=0.670, acc_len=1.59, thru=17.6, I=17.0, T=33.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:33,757] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:41,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:42:41,399] loop.py:45 INFO TRAIN_STEP step=5621 loss=1.667711 Training: 31%|███ | 5621/18018 [12:12:27<26:51:08, 7.80s/step, loss=1.668, acc=0.642, acc_len=1.50, thru=17.1, I=16.9, T=30.6, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:42:41,394] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5620: forward=2952.5ms backward=1182.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:41,617] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:42:47,998] loop.py:45 INFO TRAIN_STEP step=5622 loss=1.459635 Training: 31%|███ | 5622/18018 [12:12:34<25:36:44, 7.44s/step, loss=1.460, acc=0.674, acc_len=1.60, thru=16.0, I=17.2, T=32.5, wait=2.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:59,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:42:59,359] loop.py:45 INFO TRAIN_STEP step=5623 loss=1.495665 Training: 31%|███ | 5623/18018 [12:12:45<29:39:43, 8.62s/step, loss=1.496, acc=0.671, acc_len=1.60, thru=13.4, I=12.5, T=26.7, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:42:59,708] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:05,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:43:05,298] loop.py:45 INFO TRAIN_STEP step=5624 loss=1.481896 Training: 31%|███ | 5624/18018 [12:12:51<26:53:45, 7.81s/step, loss=1.482, acc=0.672, acc_len=1.59, thru=18.6, I=18.5, T=33.4, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:05,637] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:43:12,609] loop.py:380 INFO TIMING step=5625: step=4.062s data=1.881s compute=4.031s [fwd=2.870s bwd=1.146s opt=0.003s] dispatch=3.240s [2026-10-19 01:43:12,610] loop.py:45 INFO TRAIN_STEP step=5625 loss=1.386636 Training: 31%|███ | 5625/18018 [12:12:58<26:22:34, 7.66s/step, loss=1.387, acc=0.690, acc_len=1.66, thru=16.8, I=17.0, T=31.5, wait=3.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:20,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:43:21,867] loop.py:45 INFO TRAIN_STEP step=5626 loss=1.388857 Training: 31%|███ | 5626/18018 [12:13:07<28:01:23, 8.14s/step, loss=1.389, acc=0.692, acc_len=1.66, thru=15.1, I=17.2, T=30.2, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:43:21,862] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5625: forward=2904.1ms backward=1263.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:21,943] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-19 01:43:29,242] loop.py:45 INFO TRAIN_STEP step=5627 loss=1.504844 Training: 31%|███ | 5627/18018 [12:13:15<27:13:53, 7.91s/step, loss=1.505, acc=0.671, acc_len=1.59, thru=16.1, I=18.9, T=32.6, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:36,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:43:36,681] loop.py:45 INFO TRAIN_STEP step=5628 loss=1.589071 Training: 31%|███ | 5628/18018 [12:13:22<26:44:16, 7.77s/step, loss=1.589, acc=0.652, acc_len=1.52, thru=18.6, I=19.5, T=30.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:37,142] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:43,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:43:44,238] loop.py:45 INFO TRAIN_STEP step=5629 loss=1.511640 Training: 31%|███ | 5629/18018 [12:13:30<26:31:10, 7.71s/step, loss=1.512, acc=0.666, acc_len=1.57, thru=15.7, I=16.7, T=31.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:44,316] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:51,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:43:51,597] loop.py:380 INFO TIMING step=5630: step=3.950s data=1.699s compute=3.914s [fwd=2.737s bwd=1.162s opt=0.003s] dispatch=3.397s [2026-10-19 01:43:51,597] loop.py:45 INFO TRAIN_STEP step=5630 loss=1.488109 Training: 31%|███ | 5630/18018 [12:13:37<26:09:28, 7.60s/step, loss=1.488, acc=0.674, acc_len=1.59, thru=17.4, I=17.8, T=32.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:43:51,855] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:44:01,210] loop.py:45 INFO TRAIN_STEP step=5631 loss=1.482391 Training: 31%|███▏ | 5631/18018 [12:13:47<28:14:03, 8.21s/step, loss=1.482, acc=0.672, acc_len=1.59, thru=9.7, I=14.3, T=29.8, wait=5.3s, pool=24, epoch=1/2] (TrainerActor pid=9310) [2026-10-19 01:44:01,205] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5630: forward=2912.7ms backward=1166.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:10,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:44:12,255] loop.py:45 INFO TRAIN_STEP step=5632 loss=1.647347 Training: 31%|███▏ | 5632/18018 [12:13:58<31:09:42, 9.06s/step, loss=1.647, acc=0.644, acc_len=1.50, thru=16.9, I=15.2, T=28.2, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:12,527] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:44:19,207] loop.py:45 INFO TRAIN_STEP step=5633 loss=1.477167 Training: 31%|███▏ | 5633/18018 [12:14:05<28:59:04, 8.43s/step, loss=1.477, acc=0.671, acc_len=1.59, thru=16.0, I=21.0, T=30.4, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:26,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:44:27,765] loop.py:45 INFO TRAIN_STEP step=5634 loss=1.613731 Training: 31%|███▏ | 5634/18018 [12:14:13<29:07:16, 8.47s/step, loss=1.614, acc=0.651, acc_len=1.53, thru=16.1, I=16.0, T=32.1, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:27,966] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:34,568] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:44:34,626] loop.py:380 INFO TIMING step=5635: step=3.979s data=1.737s compute=3.941s [fwd=2.791s bwd=1.136s opt=0.003s] dispatch=2.872s [2026-10-19 01:44:34,626] loop.py:45 INFO TRAIN_STEP step=5635 loss=1.546788 Training: 31%|███▏ | 5635/18018 [12:14:20<27:27:36, 7.98s/step, loss=1.547, acc=0.660, acc_len=1.55, thru=18.0, I=18.4, T=32.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:35,070] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:40,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:44:42,186] loop.py:45 INFO TRAIN_STEP step=5636 loss=1.460034 Training: 31%|███▏ | 5636/18018 [12:14:28<27:01:26, 7.86s/step, loss=1.460, acc=0.675, acc_len=1.60, thru=17.4, I=20.3, T=27.4, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:44:42,181] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5635: forward=3297.5ms backward=1329.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:42,323] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:49,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:44:49,892] loop.py:45 INFO TRAIN_STEP step=5637 loss=1.491985 Training: 31%|███▏ | 5637/18018 [12:14:35<26:51:57, 7.81s/step, loss=1.492, acc=0.670, acc_len=1.58, thru=14.4, I=19.5, T=27.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:44:50,063] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:44:59,960] loop.py:45 INFO TRAIN_STEP step=5638 loss=1.600134 Training: 31%|███▏ | 5638/18018 [12:14:46<29:11:21, 8.49s/step, loss=1.600, acc=0.655, acc_len=1.54, thru=10.5, I=11.6, T=27.2, wait=5.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:09,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:45:10,769] loop.py:45 INFO TRAIN_STEP step=5639 loss=1.647615 Training: 31%|███▏ | 5639/18018 [12:14:56<31:34:57, 9.18s/step, loss=1.648, acc=0.646, acc_len=1.51, thru=15.1, I=14.1, T=29.2, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:10,903] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:18,207] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:45:19,799] loop.py:380 INFO TIMING step=5640: step=4.088s data=1.840s compute=4.054s [fwd=2.892s bwd=1.147s opt=0.003s] dispatch=4.932s [2026-10-19 01:45:19,799] loop.py:45 INFO TRAIN_STEP step=5640 loss=1.496793 Training: 31%|███▏ | 5640/18018 [12:15:05<31:25:08, 9.14s/step, loss=1.497, acc=0.670, acc_len=1.59, thru=14.4, I=16.6, T=31.3, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:20,222] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:24,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:45:26,653] loop.py:45 INFO TRAIN_STEP step=5641 loss=1.610051 Training: 31%|███▏ | 5641/18018 [12:15:12<29:03:46, 8.45s/step, loss=1.610, acc=0.648, acc_len=1.52, thru=16.0, I=25.2, T=30.1, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:45:26,647] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5640: forward=2855.5ms backward=1351.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:26,708] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 [2026-10-19 01:45:33,984] loop.py:45 INFO TRAIN_STEP step=5642 loss=1.538333 Training: 31%|███▏ | 5642/18018 [12:15:20<27:54:05, 8.12s/step, loss=1.538, acc=0.662, acc_len=1.56, thru=14.4, I=16.2, T=29.9, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:39,892] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:45:42,633] loop.py:45 INFO TRAIN_STEP step=5643 loss=1.490832 Training: 31%|███▏ | 5643/18018 [12:15:28<28:26:57, 8.28s/step, loss=1.491, acc=0.672, acc_len=1.59, thru=16.0, I=19.0, T=25.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:42,913] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:48,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:45:48,718] loop.py:45 INFO TRAIN_STEP step=5644 loss=1.426037 Training: 31%|███▏ | 5644/18018 [12:15:34<26:11:11, 7.62s/step, loss=1.426, acc=0.680, acc_len=1.62, thru=16.3, I=23.7, T=33.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:48,892] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:55,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:45:55,422] loop.py:380 INFO TIMING step=5645: step=4.164s data=1.905s compute=4.132s [fwd=2.893s bwd=1.224s opt=0.003s] dispatch=2.531s [2026-10-19 01:45:55,422] loop.py:45 INFO TRAIN_STEP step=5645 loss=1.498632 Training: 31%|███▏ | 5645/18018 [12:15:41<25:14:30, 7.34s/step, loss=1.499, acc=0.672, acc_len=1.60, thru=19.4, I=20.3, T=30.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:45:55,714] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:01,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:46:02,316] loop.py:45 INFO TRAIN_STEP step=5646 loss=1.585709 Training: 31%|███▏ | 5646/18018 [12:15:48<24:46:38, 7.21s/step, loss=1.586, acc=0.653, acc_len=1.53, thru=18.0, I=19.7, T=30.4, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:46:02,311] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5645: forward=2944.2ms backward=1214.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:02,774] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:09,674] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:46:10,205] loop.py:45 INFO TRAIN_STEP step=5647 loss=1.629169 Training: 31%|███▏ | 5647/18018 [12:15:56<25:28:30, 7.41s/step, loss=1.629, acc=0.645, acc_len=1.51, thru=15.0, I=17.9, T=32.1, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:10,678] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:17,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:46:19,115] loop.py:45 INFO TRAIN_STEP step=5648 loss=1.527083 Training: 31%|███▏ | 5648/18018 [12:16:05<27:00:55, 7.86s/step, loss=1.527, acc=0.662, acc_len=1.56, thru=14.2, I=18.1, T=30.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:19,618] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:24,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:46:26,439] loop.py:45 INFO TRAIN_STEP step=5649 loss=1.457725 Training: 31%|███▏ | 5649/18018 [12:16:12<26:27:36, 7.70s/step, loss=1.458, acc=0.679, acc_len=1.61, thru=16.0, I=24.2, T=29.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:26,514] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:36,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:46:37,019] loop.py:380 INFO TIMING step=5650: step=4.390s data=2.134s compute=4.334s [fwd=2.982s bwd=1.336s opt=0.003s] dispatch=6.179s [2026-10-19 01:46:37,019] loop.py:45 INFO TRAIN_STEP step=5650 loss=1.471920 Training: 31%|███▏ | 5650/18018 [12:16:23<29:25:18, 8.56s/step, loss=1.472, acc=0.669, acc_len=1.59, thru=12.0, I=12.7, T=29.2, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:37,516] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:46:44,527] loop.py:45 INFO TRAIN_STEP step=5651 loss=1.596444 Training: 31%|███▏ | 5651/18018 [12:16:30<28:19:51, 8.25s/step, loss=1.596, acc=0.654, acc_len=1.53, thru=15.3, I=16.0, T=30.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:44,457] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 01:46:44,523] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5650: forward=2870.8ms backward=1306.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:44,958] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:50,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:46:51,195] loop.py:45 INFO TRAIN_STEP step=5652 loss=1.469563 Training: 31%|███▏ | 5652/18018 [12:16:37<26:42:13, 7.77s/step, loss=1.470, acc=0.674, acc_len=1.60, thru=18.8, I=20.3, T=32.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:51,462] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:46:57,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:46:59,903] loop.py:45 INFO TRAIN_STEP step=5653 loss=1.491113 Training: 31%|███▏ | 5653/18018 [12:16:45<27:39:51, 8.05s/step, loss=1.491, acc=0.671, acc_len=1.59, thru=14.1, I=20.3, T=28.9, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:00,151] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:05,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:47:06,637] loop.py:45 INFO TRAIN_STEP step=5654 loss=1.416049 Training: 31%|███▏ | 5654/18018 [12:16:52<26:18:07, 7.66s/step, loss=1.416, acc=0.683, acc_len=1.62, thru=16.4, I=28.1, T=32.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:07,051] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:17,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:47:19,491] loop.py:380 INFO TIMING step=5655: step=4.688s data=2.340s compute=4.636s [fwd=3.162s bwd=1.459s opt=0.003s] dispatch=8.154s [2026-10-19 01:47:19,491] loop.py:45 INFO TRAIN_STEP step=5655 loss=1.462047 Training: 31%|███▏ | 5655/18018 [12:17:05<31:39:01, 9.22s/step, loss=1.462, acc=0.676, acc_len=1.61, thru=8.9, I=10.3, T=27.3, wait=8.2s, pool=64, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:19,931] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:25,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:47:26,232] loop.py:45 INFO TRAIN_STEP step=5656 loss=1.432365 Training: 31%|███▏ | 5656/18018 [12:17:12<29:05:57, 8.47s/step, loss=1.432, acc=0.682, acc_len=1.63, thru=16.8, I=24.4, T=32.2, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:47:26,227] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5655: forward=2816.8ms backward=1111.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:26,528] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:32,415] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:47:33,627] loop.py:45 INFO TRAIN_STEP step=5657 loss=1.477666 Training: 31%|███▏ | 5657/18018 [12:17:19<27:59:07, 8.15s/step, loss=1.478, acc=0.671, acc_len=1.59, thru=16.6, I=21.5, T=32.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:33,924] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:39,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:47:41,442] loop.py:45 INFO TRAIN_STEP step=5658 loss=1.472539 Training: 31%|███▏ | 5658/18018 [12:17:27<27:38:11, 8.05s/step, loss=1.473, acc=0.673, acc_len=1.60, thru=15.4, I=21.8, T=29.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:41,630] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:46,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:47:47,097] loop.py:45 INFO TRAIN_STEP step=5659 loss=1.366407 Training: 31%|███▏ | 5659/18018 [12:17:33<25:10:06, 7.33s/step, loss=1.366, acc=0.689, acc_len=1.64, thru=19.7, I=23.1, T=33.1, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:47,372] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:52,947] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:47:53,932] loop.py:380 INFO TIMING step=5660: step=4.133s data=1.945s compute=4.105s [fwd=2.895s bwd=1.195s opt=0.003s] dispatch=2.692s [2026-10-19 01:47:53,932] loop.py:45 INFO TRAIN_STEP step=5660 loss=1.466226 Training: 31%|███▏ | 5660/18018 [12:17:40<24:39:26, 7.18s/step, loss=1.466, acc=0.678, acc_len=1.61, thru=17.8, I=21.6, T=31.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:47:53,967] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 01:48:01,746] loop.py:45 INFO TRAIN_STEP step=5661 loss=1.493801 Training: 31%|███▏ | 5661/18018 [12:17:47<25:18:33, 7.37s/step, loss=1.494, acc=0.674, acc_len=1.59, thru=14.1, I=16.8, T=31.2, wait=3.7s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:48:01,739] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5660: forward=2924.0ms backward=1104.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:48:08,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:48:09,102] loop.py:45 INFO TRAIN_STEP step=5662 loss=1.460235 Training: 31%|███▏ | 5662/18018 [12:17:55<25:16:56, 7.37s/step, loss=1.460, acc=0.672, acc_len=1.60, thru=16.8, I=17.1, T=28.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:48:09,500] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:48:16,258] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:48:17,061] loop.py:45 INFO TRAIN_STEP step=5663 loss=1.487545 Training: 31%|███▏ | 5663/18018 [12:18:03<25:53:29, 7.54s/step, loss=1.488, acc=0.669, acc_len=1.59, thru=16.4, I=18.4, T=28.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:48:17,262] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:48:24,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:48:24,368] loop.py:45 INFO TRAIN_STEP step=5664 loss=1.624822 Training: 31%|███▏ | 5664/18018 [12:18:10<25:38:46, 7.47s/step, loss=1.625, acc=0.647, acc_len=1.51, thru=15.7, I=18.6, T=29.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:48:24,532] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:48:31,788] loop.py:380 INFO TIMING step=5665: step=4.377s data=2.042s compute=4.337s [fwd=3.059s bwd=1.262s opt=0.003s] dispatch=3.033s [2026-10-19 01:48:31,789] loop.py:45 INFO TRAIN_STEP step=5665 loss=1.445151 Training: 31%|███▏ | 5665/18018 [12:18:17<25:35:16, 7.46s/step, loss=1.445, acc=0.677, acc_len=1.61, thru=14.5, I=14.3, T=29.2, wait=3.0s, pool=40, epoch=1/2][2026-10-19 01:48:42,207] loop.py:45 INFO TRAIN_STEP step=5666 loss=1.494696 Training: 31%|███▏ | 5666/18018 [12:18:28<28:38:03, 8.35s/step, loss=1.495, acc=0.670, acc_len=1.59, thru=13.3, I=14.4, T=28.8, wait=6.0s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:48:42,201] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5665: forward=3144.5ms backward=1249.9ms [2026-10-19 01:48:52,163] loop.py:45 INFO TRAIN_STEP step=5667 loss=1.725423 Training: 31%|███▏ | 5667/18018 [12:18:38<30:17:19, 8.83s/step, loss=1.725, acc=0.631, acc_len=1.46, thru=12.9, I=11.9, T=31.1, wait=5.8s, pool=48, epoch=1/2][2026-10-19 01:49:00,551] loop.py:45 INFO TRAIN_STEP step=5668 loss=1.505867 Training: 31%|███▏ | 5668/18018 [12:18:46<29:50:12, 8.70s/step, loss=1.506, acc=0.668, acc_len=1.59, thru=17.2, I=15.1, T=27.6, wait=3.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:06,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:49:07,129] loop.py:45 INFO TRAIN_STEP step=5669 loss=1.505501 Training: 31%|███▏ | 5669/18018 [12:18:53<27:39:09, 8.06s/step, loss=1.506, acc=0.669, acc_len=1.59, thru=17.9, I=18.4, T=33.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:07,367] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:12,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:49:13,592] loop.py:380 INFO TIMING step=5670: step=4.203s data=1.893s compute=4.168s [fwd=2.912s bwd=1.241s opt=0.003s] dispatch=2.250s [2026-10-19 01:49:13,592] loop.py:45 INFO TRAIN_STEP step=5670 loss=1.344273 Training: 31%|███▏ | 5670/18018 [12:18:59<26:00:15, 7.58s/step, loss=1.344, acc=0.696, acc_len=1.67, thru=18.7, I=22.6, T=30.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:13,767] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:49:19,943] loop.py:45 INFO TRAIN_STEP step=5671 loss=1.528057 Training: 31%|███▏ | 5671/18018 [12:19:06<24:44:08, 7.21s/step, loss=1.528, acc=0.665, acc_len=1.57, thru=17.7, I=22.6, T=32.7, wait=2.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:49:19,938] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5670: forward=2741.2ms backward=1118.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:25,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:49:27,238] loop.py:45 INFO TRAIN_STEP step=5672 loss=1.344959 Training: 31%|███▏ | 5672/18018 [12:19:13<24:49:13, 7.24s/step, loss=1.345, acc=0.694, acc_len=1.67, thru=16.5, I=19.6, T=33.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:27,544] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:33,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:49:34,138] loop.py:45 INFO TRAIN_STEP step=5673 loss=1.458254 Training: 31%|███▏ | 5673/18018 [12:19:20<24:28:16, 7.14s/step, loss=1.458, acc=0.675, acc_len=1.61, thru=17.4, I=22.5, T=33.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:34,272] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:39,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:49:41,663] loop.py:45 INFO TRAIN_STEP step=5674 loss=1.461696 Training: 31%|███▏ | 5674/18018 [12:19:27<24:52:04, 7.25s/step, loss=1.462, acc=0.674, acc_len=1.60, thru=18.4, I=20.3, T=33.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:49:41,881] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:49:47,676] loop.py:380 INFO TIMING step=5675: step=3.710s data=1.495s compute=3.680s [fwd=2.579s bwd=1.085s opt=0.003s] dispatch=2.295s [2026-10-19 01:49:47,676] loop.py:45 INFO TRAIN_STEP step=5675 loss=1.434176 Training: 31%|███▏ | 5675/18018 [12:19:33<23:35:30, 6.88s/step, loss=1.434, acc=0.677, acc_len=1.60, thru=18.9, I=26.6, T=34.5, wait=2.3s, pool=52, epoch=1/2][2026-10-19 01:49:55,027] loop.py:45 INFO TRAIN_STEP step=5676 loss=1.605120 Training: 32%|███▏ | 5676/18018 [12:19:41<24:04:31, 7.02s/step, loss=1.605, acc=0.649, acc_len=1.52, thru=16.1, I=17.9, T=33.9, wait=3.6s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:49:55,022] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5675: forward=2623.0ms backward=1101.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:50:03,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:50:04,316] loop.py:45 INFO TRAIN_STEP step=5677 loss=1.465270 Training: 32%|███▏ | 5677/18018 [12:19:50<26:24:11, 7.70s/step, loss=1.465, acc=0.673, acc_len=1.60, thru=15.4, I=15.4, T=28.9, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:50:04,556] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:50:13,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:50:14,426] loop.py:45 INFO TRAIN_STEP step=5678 loss=1.582798 Training: 32%|███▏ | 5678/18018 [12:20:00<28:52:38, 8.42s/step, loss=1.583, acc=0.657, acc_len=1.55, thru=13.0, I=13.4, T=27.6, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:50:14,566] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-19 01:50:22,096] loop.py:45 INFO TRAIN_STEP step=5679 loss=1.538495 Training: 32%|███▏ | 5679/18018 [12:20:08<28:05:57, 8.20s/step, loss=1.538, acc=0.665, acc_len=1.56, thru=13.9, I=16.0, T=31.8, wait=3.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:50:30,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:50:31,093] loop.py:380 INFO TIMING step=5680: step=4.130s data=1.844s compute=4.102s [fwd=2.910s bwd=1.177s opt=0.003s] dispatch=4.855s [2026-10-19 01:50:31,093] loop.py:45 INFO TRAIN_STEP step=5680 loss=1.459160 Training: 32%|███▏ | 5680/18018 [12:20:17<28:55:05, 8.44s/step, loss=1.459, acc=0.678, acc_len=1.61, thru=15.6, I=14.7, T=31.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:50:31,600] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:50:38,525] loop.py:45 INFO TRAIN_STEP step=5681 loss=1.542618 (TrainerActor pid=9310) [2026-10-19 01:50:38,520] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5680: forward=2886.3ms backward=1237.9ms Training: 32%|███▏ | 5681/18018 [12:20:24<27:52:48, 8.14s/step, loss=1.543, acc=0.661, acc_len=1.56, thru=15.6, I=16.0, T=30.7, wait=3.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:50:45,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:50:45,461] loop.py:45 INFO TRAIN_STEP step=5682 loss=1.537132 Training: 32%|███▏ | 5682/18018 [12:20:31<26:38:39, 7.78s/step, loss=1.537, acc=0.661, acc_len=1.56, thru=18.5, I=18.2, T=33.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:50:45,658] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:50:52,303] loop.py:45 INFO TRAIN_STEP step=5683 loss=1.504993 Training: 32%|███▏ | 5683/18018 [12:20:38<25:41:14, 7.50s/step, loss=1.505, acc=0.667, acc_len=1.58, thru=15.7, I=19.5, T=33.9, wait=3.1s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:01,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:51:02,018] loop.py:45 INFO TRAIN_STEP step=5684 loss=1.402607 Training: 32%|███▏ | 5684/18018 [12:20:48<27:57:38, 8.16s/step, loss=1.403, acc=0.688, acc_len=1.64, thru=16.5, I=15.1, T=30.8, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:02,066] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 01:51:08,983] loop.py:380 INFO TIMING step=5685: step=3.870s data=1.727s compute=3.833s [fwd=2.752s bwd=1.067s opt=0.003s] dispatch=3.085s [2026-10-19 01:51:08,983] loop.py:45 INFO TRAIN_STEP step=5685 loss=1.567401 Training: 32%|███▏ | 5685/18018 [12:20:55<26:43:51, 7.80s/step, loss=1.567, acc=0.651, acc_len=1.53, thru=15.2, I=21.5, T=33.1, wait=3.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:16,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:51:18,050] loop.py:45 INFO TRAIN_STEP step=5686 loss=1.449719 Training: 32%|███▏ | 5686/18018 [12:21:04<28:01:40, 8.18s/step, loss=1.450, acc=0.678, acc_len=1.61, thru=16.6, I=15.4, T=29.5, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:51:18,044] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5685: forward=3045.6ms backward=1248.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:18,160] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:24,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:51:26,081] loop.py:45 INFO TRAIN_STEP step=5687 loss=1.611874 Training: 32%|███▏ | 5687/18018 [12:21:12<27:52:12, 8.14s/step, loss=1.612, acc=0.653, acc_len=1.53, thru=13.3, I=20.0, T=29.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:26,566] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 01:51:32,437] loop.py:45 INFO TRAIN_STEP step=5688 loss=1.554362 Training: 32%|███▏ | 5688/18018 [12:21:18<26:02:13, 7.60s/step, loss=1.554, acc=0.659, acc_len=1.55, thru=16.5, I=23.3, T=36.7, wait=2.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:40,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:51:41,545] loop.py:45 INFO TRAIN_STEP step=5689 loss=1.630916 Training: 32%|███▏ | 5689/18018 [12:21:27<27:34:56, 8.05s/step, loss=1.631, acc=0.644, acc_len=1.51, thru=15.7, I=15.7, T=32.2, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:41,891] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:49,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:51:50,565] loop.py:380 INFO TIMING step=5690: step=4.503s data=2.240s compute=4.453s [fwd=3.122s bwd=1.316s opt=0.003s] dispatch=4.508s [2026-10-19 01:51:50,566] loop.py:45 INFO TRAIN_STEP step=5690 loss=1.375920 Training: 32%|███▏ | 5690/18018 [12:21:36<28:34:27, 8.34s/step, loss=1.376, acc=0.692, acc_len=1.66, thru=14.8, I=14.6, T=28.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:50,599] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:51:58,345] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:52:00,133] loop.py:45 INFO TRAIN_STEP step=5691 loss=1.612657 Training: 32%|███▏ | 5691/18018 [12:21:46<29:49:44, 8.71s/step, loss=1.613, acc=0.647, acc_len=1.51, thru=13.4, I=15.5, T=23.8, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:52:00,128] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5690: forward=3982.3ms backward=1327.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:00,356] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:09,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:52:10,775] loop.py:45 INFO TRAIN_STEP step=5692 loss=1.453404 Training: 32%|███▏ | 5692/18018 [12:21:56<31:48:34, 9.29s/step, loss=1.453, acc=0.676, acc_len=1.61, thru=11.8, I=14.3, T=22.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:11,173] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:17,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:52:19,355] loop.py:45 INFO TRAIN_STEP step=5693 loss=1.533207 Training: 32%|███▏ | 5693/18018 [12:22:05<31:04:30, 9.08s/step, loss=1.533, acc=0.665, acc_len=1.58, thru=15.7, I=18.8, T=26.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:19,408] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:24,693] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:52:25,618] loop.py:45 INFO TRAIN_STEP step=5694 loss=1.461445 Training: 32%|███▏ | 5694/18018 [12:22:11<28:11:09, 8.23s/step, loss=1.461, acc=0.675, acc_len=1.61, thru=17.0, I=20.7, T=34.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:25,697] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:32,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:52:34,211] loop.py:380 INFO TIMING step=5695: step=4.386s data=2.147s compute=4.356s [fwd=3.116s bwd=1.226s opt=0.003s] dispatch=4.195s [2026-10-19 01:52:34,211] loop.py:45 INFO TRAIN_STEP step=5695 loss=1.465270 Training: 32%|███▏ | 5695/18018 [12:22:20<28:33:03, 8.34s/step, loss=1.465, acc=0.675, acc_len=1.61, thru=14.9, I=19.9, T=29.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:34,582] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:42,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:52:42,164] loop.py:45 INFO TRAIN_STEP step=5696 loss=1.382904 Training: 32%|███▏ | 5696/18018 [12:22:28<28:09:00, 8.22s/step, loss=1.383, acc=0.691, acc_len=1.65, thru=16.9, I=17.4, T=30.2, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:52:42,159] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5695: forward=2936.6ms backward=1269.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:52:42,519] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:52:48,597] loop.py:45 INFO TRAIN_STEP step=5697 loss=1.469118 Training: 32%|███▏ | 5697/18018 [12:22:34<26:18:24, 7.69s/step, loss=1.469, acc=0.674, acc_len=1.60, thru=16.3, I=16.4, T=33.3, wait=2.6s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:00,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:53:02,647] loop.py:45 INFO TRAIN_STEP step=5698 loss=1.553538 Training: 32%|███▏ | 5698/18018 [12:22:48<32:50:26, 9.60s/step, loss=1.554, acc=0.670, acc_len=1.59, thru=14.8, I=12.5, T=25.8, wait=9.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:02,684] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:09,648] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:53:11,035] loop.py:45 INFO TRAIN_STEP step=5699 loss=1.471666 Training: 32%|███▏ | 5699/18018 [12:22:57<31:35:45, 9.23s/step, loss=1.472, acc=0.671, acc_len=1.60, thru=15.4, I=18.1, T=29.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:11,167] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:18,892] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:53:19,589] loop.py:380 INFO TIMING step=5700: step=4.036s data=1.821s compute=3.999s [fwd=2.836s bwd=1.148s opt=0.003s] dispatch=4.509s [2026-10-19 01:53:19,589] loop.py:45 INFO TRAIN_STEP step=5700 loss=1.552844 Training: 32%|███▏ | 5700/18018 [12:23:05<30:53:48, 9.03s/step, loss=1.553, acc=0.661, acc_len=1.56, thru=13.3, I=17.2, T=31.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:19,899] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 01:53:26,682] loop.py:45 INFO TRAIN_STEP step=5701 loss=1.607143 Training: 32%|███▏ | 5701/18018 [12:23:12<28:54:16, 8.45s/step, loss=1.607, acc=0.647, acc_len=1.51, thru=17.9, I=18.3, T=32.7, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:53:26,677] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5700: forward=2732.8ms backward=1128.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:26,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:27,165] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:32,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:53:34,881] loop.py:45 INFO TRAIN_STEP step=5702 loss=1.447958 Training: 32%|███▏ | 5702/18018 [12:23:20<28:38:56, 8.37s/step, loss=1.448, acc=0.682, acc_len=1.62, thru=16.6, I=19.1, T=28.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:35,290] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 01:53:43,297] loop.py:45 INFO TRAIN_STEP step=5703 loss=1.465340 Training: 32%|███▏ | 5703/18018 [12:23:29<28:41:13, 8.39s/step, loss=1.465, acc=0.672, acc_len=1.60, thru=13.0, I=12.6, T=30.6, wait=4.2s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:51,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:53:54,558] loop.py:45 INFO TRAIN_STEP step=5704 loss=1.586394 Training: 32%|███▏ | 5704/18018 [12:23:40<31:38:08, 9.25s/step, loss=1.586, acc=0.659, acc_len=1.55, thru=13.1, I=16.6, T=24.5, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:53:54,988] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 01:54:03,392] loop.py:380 INFO TIMING step=5705: step=4.241s data=1.957s compute=4.194s [fwd=2.946s bwd=1.233s opt=0.003s] dispatch=4.585s [2026-10-19 01:54:03,393] loop.py:45 INFO TRAIN_STEP step=5705 loss=1.539637 Training: 32%|███▏ | 5705/18018 [12:23:49<31:12:37, 9.13s/step, loss=1.540, acc=0.664, acc_len=1.57, thru=13.8, I=13.7, T=30.2, wait=4.6s, pool=52, epoch=1/2][2026-10-19 01:54:11,002] loop.py:45 INFO TRAIN_STEP step=5706 loss=1.541871 Training: 32%|███▏ | 5706/18018 [12:23:57<29:39:00, 8.67s/step, loss=1.542, acc=0.658, acc_len=1.54, thru=17.2, I=16.9, T=31.8, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:54:10,997] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5705: forward=2818.4ms backward=1150.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:10,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:11,469] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:54:18,878] loop.py:45 INFO TRAIN_STEP step=5707 loss=1.433428 Training: 32%|███▏ | 5707/18018 [12:24:04<28:50:01, 8.43s/step, loss=1.433, acc=0.683, acc_len=1.63, thru=15.3, I=15.6, T=26.9, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:25,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:54:26,269] loop.py:45 INFO TRAIN_STEP step=5708 loss=1.686138 Training: 32%|███▏ | 5708/18018 [12:24:12<27:45:59, 8.12s/step, loss=1.686, acc=0.638, acc_len=1.48, thru=18.1, I=16.1, T=29.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:26,302] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 01:54:32,465] loop.py:45 INFO TRAIN_STEP step=5709 loss=1.471219 Training: 32%|███▏ | 5709/18018 [12:24:18<25:47:19, 7.54s/step, loss=1.471, acc=0.674, acc_len=1.60, thru=15.7, I=16.9, T=30.4, wait=2.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:38,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:54:39,982] loop.py:380 INFO TIMING step=5710: step=4.375s data=2.244s compute=4.324s [fwd=3.064s bwd=1.245s opt=0.003s] dispatch=3.133s [2026-10-19 01:54:39,982] loop.py:45 INFO TRAIN_STEP step=5710 loss=1.418238 Training: 32%|███▏ | 5710/18018 [12:24:26<25:45:38, 7.53s/step, loss=1.418, acc=0.681, acc_len=1.62, thru=18.4, I=17.8, T=29.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:40,011] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:45,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:54:47,426] loop.py:45 INFO TRAIN_STEP step=5711 loss=1.308720 Training: 32%|███▏ | 5711/18018 [12:24:33<25:39:55, 7.51s/step, loss=1.309, acc=0.702, acc_len=1.69, thru=15.6, I=19.9, T=28.2, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:54:47,421] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5710: forward=3419.3ms backward=1082.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:47,478] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:54,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:54:55,616] loop.py:45 INFO TRAIN_STEP step=5712 loss=1.424522 Training: 32%|███▏ | 5712/18018 [12:24:41<26:21:53, 7.71s/step, loss=1.425, acc=0.683, acc_len=1.63, thru=13.7, I=16.2, T=24.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:54:56,026] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:00,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:55:02,712] loop.py:45 INFO TRAIN_STEP step=5713 loss=1.469867 Training: 32%|███▏ | 5713/18018 [12:24:48<25:43:48, 7.53s/step, loss=1.470, acc=0.672, acc_len=1.60, thru=15.0, I=23.2, T=29.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:02,988] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:09,916] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:55:10,119] loop.py:45 INFO TRAIN_STEP step=5714 loss=1.345127 Training: 32%|███▏ | 5714/18018 [12:24:56<25:36:10, 7.49s/step, loss=1.345, acc=0.695, acc_len=1.67, thru=14.8, I=19.4, T=34.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:10,419] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:17,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:55:18,817] loop.py:380 INFO TIMING step=5715: step=4.391s data=2.071s compute=4.359s [fwd=3.067s bwd=1.277s opt=0.003s] dispatch=4.297s [2026-10-19 01:55:18,817] loop.py:45 INFO TRAIN_STEP step=5715 loss=1.417340 Training: 32%|███▏ | 5715/18018 [12:25:04<26:50:22, 7.85s/step, loss=1.417, acc=0.684, acc_len=1.64, thru=15.1, I=17.6, T=29.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:18,968] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:25,897] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:55:26,435] loop.py:45 INFO TRAIN_STEP step=5716 loss=1.471936 (TrainerActor pid=9310) [2026-10-19 01:55:26,431] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5715: forward=3040.6ms backward=1251.9ms Training: 32%|███▏ | 5716/18018 [12:25:12<26:35:37, 7.78s/step, loss=1.472, acc=0.669, acc_len=1.58, thru=14.6, I=17.5, T=29.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:26,903] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:31,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:55:32,703] loop.py:45 INFO TRAIN_STEP step=5717 loss=1.551018 Training: 32%|███▏ | 5717/18018 [12:25:18<25:02:19, 7.33s/step, loss=1.551, acc=0.662, acc_len=1.57, thru=19.3, I=23.1, T=30.8, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:33,142] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:40,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:55:40,778] loop.py:45 INFO TRAIN_STEP step=5718 loss=1.507368 Training: 32%|███▏ | 5718/18018 [12:25:26<25:48:09, 7.55s/step, loss=1.507, acc=0.669, acc_len=1.58, thru=14.2, I=17.7, T=32.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:41,153] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:47,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:55:48,224] loop.py:45 INFO TRAIN_STEP step=5719 loss=1.455710 Training: 32%|███▏ | 5719/18018 [12:25:34<25:41:42, 7.52s/step, loss=1.456, acc=0.671, acc_len=1.59, thru=18.2, I=19.7, T=29.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:48,308] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:54,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:55:56,892] loop.py:380 INFO TIMING step=5720: step=4.453s data=2.048s compute=4.425s [fwd=3.155s bwd=1.255s opt=0.003s] dispatch=4.205s [2026-10-19 01:55:56,893] loop.py:45 INFO TRAIN_STEP step=5720 loss=1.330805 Training: 32%|███▏ | 5720/18018 [12:25:42<26:51:56, 7.86s/step, loss=1.331, acc=0.697, acc_len=1.68, thru=14.0, I=18.7, T=28.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:55:57,275] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:02,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:56:03,937] loop.py:45 INFO TRAIN_STEP step=5721 loss=1.443279 Training: 32%|███▏ | 5721/18018 [12:25:50<26:01:32, 7.62s/step, loss=1.443, acc=0.677, acc_len=1.61, thru=15.1, I=25.0, T=31.2, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:56:03,932] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5720: forward=2820.9ms backward=1236.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:03,972] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:10,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:56:11,154] loop.py:45 INFO TRAIN_STEP step=5722 loss=1.434292 Training: 32%|███▏ | 5722/18018 [12:25:57<25:36:38, 7.50s/step, loss=1.434, acc=0.677, acc_len=1.61, thru=16.9, I=22.4, T=27.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:11,608] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:16,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:56:18,122] loop.py:45 INFO TRAIN_STEP step=5723 loss=1.521585 Training: 32%|███▏ | 5723/18018 [12:26:04<25:03:52, 7.34s/step, loss=1.522, acc=0.666, acc_len=1.57, thru=15.3, I=19.3, T=33.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:18,296] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:24,308] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:56:24,838] loop.py:45 INFO TRAIN_STEP step=5724 loss=1.392416 Training: 32%|███▏ | 5724/18018 [12:26:10<24:25:42, 7.15s/step, loss=1.392, acc=0.688, acc_len=1.65, thru=18.3, I=22.8, T=30.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:25,314] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:33,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:56:34,272] loop.py:380 INFO TIMING step=5725: step=4.364s data=1.997s compute=4.322s [fwd=3.077s bwd=1.230s opt=0.003s] dispatch=5.057s [2026-10-19 01:56:34,272] loop.py:45 INFO TRAIN_STEP step=5725 loss=1.590039 Training: 32%|███▏ | 5725/18018 [12:26:20<26:45:36, 7.84s/step, loss=1.590, acc=0.655, acc_len=1.53, thru=13.2, I=15.8, T=29.3, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:34,511] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:56:41,638] loop.py:45 INFO TRAIN_STEP step=5726 loss=1.644376 Training: 32%|███▏ | 5726/18018 [12:26:27<26:16:29, 7.70s/step, loss=1.644, acc=0.647, acc_len=1.51, thru=13.8, I=15.8, T=29.0, wait=2.9s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:56:41,633] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5725: forward=3041.9ms backward=1322.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:47,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:56:49,246] loop.py:45 INFO TRAIN_STEP step=5727 loss=1.736780 Training: 32%|███▏ | 5727/18018 [12:26:35<26:10:54, 7.67s/step, loss=1.737, acc=0.629, acc_len=1.46, thru=18.7, I=18.6, T=32.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:49,736] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:55,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:56:56,475] loop.py:45 INFO TRAIN_STEP step=5728 loss=1.449538 Training: 32%|███▏ | 5728/18018 [12:26:42<25:43:49, 7.54s/step, loss=1.450, acc=0.677, acc_len=1.61, thru=16.5, I=21.2, T=30.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:56:56,945] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:03,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:57:03,725] loop.py:45 INFO TRAIN_STEP step=5729 loss=1.675850 Training: 32%|███▏ | 5729/18018 [12:26:49<25:26:00, 7.45s/step, loss=1.676, acc=0.642, acc_len=1.49, thru=15.8, I=18.9, T=30.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:04,157] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:09,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:57:11,075] loop.py:380 INFO TIMING step=5730: step=3.956s data=1.779s compute=3.930s [fwd=2.808s bwd=1.108s opt=0.003s] dispatch=3.385s [2026-10-19 01:57:11,075] loop.py:45 INFO TRAIN_STEP step=5730 loss=1.571114 Training: 32%|███▏ | 5730/18018 [12:26:57<25:19:40, 7.42s/step, loss=1.571, acc=0.653, acc_len=1.53, thru=16.5, I=21.4, T=32.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:11,573] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:17,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:57:18,331] loop.py:45 INFO TRAIN_STEP step=5731 loss=1.504827 Training: 32%|███▏ | 5731/18018 [12:27:04<25:09:35, 7.37s/step, loss=1.505, acc=0.667, acc_len=1.58, thru=15.8, I=23.5, T=34.2, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:57:18,324] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5730: forward=2577.5ms backward=1109.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:18,616] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:24,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:57:25,729] loop.py:45 INFO TRAIN_STEP step=5732 loss=1.374609 Training: 32%|███▏ | 5732/18018 [12:27:11<25:11:06, 7.38s/step, loss=1.375, acc=0.688, acc_len=1.65, thru=16.8, I=21.2, T=27.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:26,108] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 01:57:37,010] loop.py:45 INFO TRAIN_STEP step=5733 loss=1.593613 Training: 32%|███▏ | 5733/18018 [12:27:23<29:10:29, 8.55s/step, loss=1.594, acc=0.655, acc_len=1.53, thru=10.9, I=11.0, T=27.6, wait=6.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:43,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-19 01:57:44,118] loop.py:45 INFO TRAIN_STEP step=5734 loss=1.543978 Training: 32%|███▏ | 5734/18018 [12:27:30<27:41:59, 8.12s/step, loss=1.544, acc=0.659, acc_len=1.55, thru=18.1, I=17.4, T=32.8, wait=3.2s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:44,516] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:49,665] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:57:50,153] loop.py:380 INFO TIMING step=5735: step=3.512s data=1.510s compute=3.475s [fwd=2.457s bwd=1.002s opt=0.003s] dispatch=2.512s [2026-10-19 01:57:50,153] loop.py:45 INFO TRAIN_STEP step=5735 loss=1.444383 Training: 32%|███▏ | 5735/18018 [12:27:36<25:33:56, 7.49s/step, loss=1.444, acc=0.677, acc_len=1.61, thru=19.6, I=22.6, T=36.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:50,185] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:55,209] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:57:56,962] loop.py:45 INFO TRAIN_STEP step=5736 loss=1.451130 Training: 32%|███▏ | 5736/18018 [12:27:43<24:51:49, 7.29s/step, loss=1.451, acc=0.679, acc_len=1.61, thru=19.3, I=22.4, T=29.9, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:57:56,957] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5735: forward=2931.9ms backward=1298.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:57:57,220] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:02,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:58:04,632] loop.py:45 INFO TRAIN_STEP step=5737 loss=1.441250 Training: 32%|███▏ | 5737/18018 [12:27:50<25:15:06, 7.40s/step, loss=1.441, acc=0.675, acc_len=1.61, thru=14.6, I=21.5, T=33.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:04,744] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:11,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:58:12,546] loop.py:45 INFO TRAIN_STEP step=5738 loss=1.540932 Training: 32%|███▏ | 5738/18018 [12:27:58<25:46:20, 7.56s/step, loss=1.541, acc=0.663, acc_len=1.56, thru=14.2, I=20.2, T=27.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:12,591] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:17,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:58:19,475] loop.py:45 INFO TRAIN_STEP step=5739 loss=1.580229 Training: 32%|███▏ | 5739/18018 [12:28:05<25:07:47, 7.37s/step, loss=1.580, acc=0.656, acc_len=1.54, thru=17.6, I=20.1, T=31.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:19,923] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:25,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:58:27,197] loop.py:380 INFO TIMING step=5740: step=4.047s data=1.742s compute=4.014s [fwd=2.778s bwd=1.221s opt=0.003s] dispatch=3.665s [2026-10-19 01:58:27,197] loop.py:45 INFO TRAIN_STEP step=5740 loss=1.406717 Training: 32%|███▏ | 5740/18018 [12:28:13<25:29:27, 7.47s/step, loss=1.407, acc=0.680, acc_len=1.63, thru=14.9, I=23.9, T=31.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:27,703] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 01:58:35,890] loop.py:45 INFO TRAIN_STEP step=5741 loss=1.436563 Training: 32%|███▏ | 5741/18018 [12:28:21<26:44:02, 7.84s/step, loss=1.437, acc=0.681, acc_len=1.63, thru=13.7, I=13.7, T=30.0, wait=4.4s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:58:35,886] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5740: forward=3009.0ms backward=1202.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:44,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:58:46,416] loop.py:45 INFO TRAIN_STEP step=5742 loss=1.560392 Training: 32%|███▏ | 5742/18018 [12:28:32<29:28:54, 8.65s/step, loss=1.560, acc=0.658, acc_len=1.55, thru=13.8, I=16.2, T=27.4, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:46,819] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:52,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:58:52,928] loop.py:45 INFO TRAIN_STEP step=5743 loss=1.576485 Training: 32%|███▏ | 5743/18018 [12:28:39<27:17:42, 8.01s/step, loss=1.576, acc=0.654, acc_len=1.54, thru=18.1, I=22.6, T=35.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:58:53,376] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 01:59:00,951] loop.py:45 INFO TRAIN_STEP step=5744 loss=1.443209 Training: 32%|███▏ | 5744/18018 [12:28:47<27:18:42, 8.01s/step, loss=1.443, acc=0.677, acc_len=1.61, thru=13.7, I=14.8, T=31.3, wait=3.9s, pool=44, epoch=1/2][2026-10-19 01:59:11,316] loop.py:380 INFO TIMING step=5745: step=4.021s data=1.825s compute=3.992s [fwd=2.839s bwd=1.138s opt=0.003s] dispatch=6.335s [2026-10-19 01:59:11,316] loop.py:45 INFO TRAIN_STEP step=5745 loss=1.611936 Training: 32%|███▏ | 5745/18018 [12:28:57<29:43:03, 8.72s/step, loss=1.612, acc=0.653, acc_len=1.53, thru=13.2, I=13.4, T=31.8, wait=6.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:59:18,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:59:20,398] loop.py:45 INFO TRAIN_STEP step=5746 loss=1.521669 Training: 32%|███▏ | 5746/18018 [12:29:06<30:05:23, 8.83s/step, loss=1.522, acc=0.663, acc_len=1.57, thru=14.9, I=17.8, T=29.2, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 01:59:20,392] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5745: forward=3052.2ms backward=1284.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:59:20,776] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:59:26,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:59:27,804] loop.py:45 INFO TRAIN_STEP step=5747 loss=1.473709 Training: 32%|███▏ | 5747/18018 [12:29:13<28:37:59, 8.40s/step, loss=1.474, acc=0.669, acc_len=1.58, thru=14.4, I=23.6, T=28.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:59:28,236] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 01:59:36,521] loop.py:45 INFO TRAIN_STEP step=5748 loss=1.579565 Training: 32%|███▏ | 5748/18018 [12:29:22<28:57:16, 8.50s/step, loss=1.580, acc=0.651, acc_len=1.53, thru=15.0, I=14.0, T=33.0, wait=4.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:59:42,736] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:59:44,337] loop.py:45 INFO TRAIN_STEP step=5749 loss=1.406906 Training: 32%|███▏ | 5749/18018 [12:29:30<28:15:27, 8.29s/step, loss=1.407, acc=0.683, acc_len=1.62, thru=16.2, I=18.6, T=31.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:59:44,746] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:59:50,568] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 01:59:51,575] loop.py:380 INFO TIMING step=5750: step=3.904s data=1.745s compute=3.867s [fwd=2.746s bwd=1.106s opt=0.003s] dispatch=3.326s [2026-10-19 01:59:51,575] loop.py:45 INFO TRAIN_STEP step=5750 loss=1.594642 Training: 32%|███▏ | 5750/18018 [12:29:37<27:10:43, 7.98s/step, loss=1.595, acc=0.655, acc_len=1.53, thru=16.2, I=21.9, T=32.8, wait=3.3s, pool=64, epoch=1/2][2026-10-19 01:59:51,578] loop.py:419 INFO Saving checkpoint at step 5750... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 01:59:54,993] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0005751 (TrainerActor pid=9310) [2026-10-19 01:59:55,001] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0005001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:59:55,607] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 01:59:56,611] inference_manager.py:408 INFO Pool capacity available after 6.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:03,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 01:59:54,992] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0005751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 01:59:56,239] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0005001 [2026-10-19 02:00:04,263] loop.py:45 INFO TRAIN_STEP step=5751 loss=1.524441 Training: 32%|███▏ | 5751/18018 [12:29:50<31:59:38, 9.39s/step, loss=1.524, acc=0.670, acc_len=1.58, thru=16.7, I=19.9, T=30.7, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:00:04,259] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5750: forward=2921.0ms backward=1205.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:04,643] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:10,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:00:10,689] loop.py:45 INFO TRAIN_STEP step=5752 loss=1.675341 Training: 32%|███▏ | 5752/18018 [12:29:56<28:57:45, 8.50s/step, loss=1.675, acc=0.635, acc_len=1.47, thru=18.4, I=22.4, T=34.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:10,818] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-19 02:00:16,818] loop.py:45 INFO TRAIN_STEP step=5753 loss=1.462446 Training: 32%|███▏ | 5753/18018 [12:30:02<26:32:07, 7.79s/step, loss=1.462, acc=0.677, acc_len=1.61, thru=19.4, I=19.9, T=31.4, wait=2.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:24,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:00:26,260] loop.py:45 INFO TRAIN_STEP step=5754 loss=1.444707 Training: 32%|███▏ | 5754/18018 [12:30:12<28:13:33, 8.29s/step, loss=1.445, acc=0.679, acc_len=1.62, thru=14.5, I=16.8, T=30.4, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:26,550] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:31,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:00:32,608] loop.py:380 INFO TIMING step=5755: step=3.714s data=1.665s compute=3.683s [fwd=2.631s bwd=1.037s opt=0.003s] dispatch=2.622s [2026-10-19 02:00:32,608] loop.py:45 INFO TRAIN_STEP step=5755 loss=1.223229 Training: 32%|███▏ | 5755/18018 [12:30:18<26:14:37, 7.70s/step, loss=1.223, acc=0.715, acc_len=1.75, thru=17.8, I=27.2, T=34.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:32,811] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:37,712] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:00:39,315] loop.py:45 INFO TRAIN_STEP step=5756 loss=1.451853 Training: 32%|███▏ | 5756/18018 [12:30:25<25:13:17, 7.40s/step, loss=1.452, acc=0.676, acc_len=1.60, thru=18.5, I=26.1, T=31.8, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:00:39,309] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5755: forward=2653.5ms backward=1325.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:39,728] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:44,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:00:45,211] loop.py:45 INFO TRAIN_STEP step=5757 loss=1.466967 Training: 32%|███▏ | 5757/18018 [12:30:31<23:40:43, 6.95s/step, loss=1.467, acc=0.676, acc_len=1.61, thru=19.7, I=27.8, T=35.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:45,283] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:50,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:00:50,989] loop.py:45 INFO TRAIN_STEP step=5758 loss=1.477633 Training: 32%|███▏ | 5758/18018 [12:30:37<22:28:40, 6.60s/step, loss=1.478, acc=0.675, acc_len=1.60, thru=21.7, I=24.8, T=35.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:51,293] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:56,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:00:57,691] loop.py:45 INFO TRAIN_STEP step=5759 loss=1.433525 Training: 32%|███▏ | 5759/18018 [12:30:43<22:34:42, 6.63s/step, loss=1.434, acc=0.680, acc_len=1.63, thru=19.6, I=22.8, T=31.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:00:57,838] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-19 02:01:03,959] loop.py:380 INFO TIMING step=5760: step=4.125s data=1.932s compute=4.092s [fwd=2.876s bwd=1.202s opt=0.003s] dispatch=2.134s [2026-10-19 02:01:03,959] loop.py:45 INFO TRAIN_STEP step=5760 loss=1.399128 Training: 32%|███▏ | 5760/18018 [12:30:50<22:12:17, 6.52s/step, loss=1.399, acc=0.683, acc_len=1.63, thru=18.0, I=24.8, T=31.0, wait=2.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:12,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:01:12,680] loop.py:45 INFO TRAIN_STEP step=5761 loss=1.521634 Training: 32%|███▏ | 5761/18018 [12:30:58<24:27:00, 7.18s/step, loss=1.522, acc=0.662, acc_len=1.56, thru=15.8, I=13.8, T=28.2, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:01:12,676] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5760: forward=3260.3ms backward=1230.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:12,929] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:19,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:01:19,256] loop.py:45 INFO TRAIN_STEP step=5762 loss=1.500781 Training: 32%|███▏ | 5762/18018 [12:31:05<23:49:44, 7.00s/step, loss=1.501, acc=0.671, acc_len=1.59, thru=17.9, I=17.0, T=33.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:19,720] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:24,538] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:01:26,398] loop.py:45 INFO TRAIN_STEP step=5763 loss=1.574167 Training: 32%|███▏ | 5763/18018 [12:31:12<23:58:33, 7.04s/step, loss=1.574, acc=0.656, acc_len=1.54, thru=16.9, I=21.5, T=28.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:26,557] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:32,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:01:33,467] loop.py:45 INFO TRAIN_STEP step=5764 loss=1.621328 Training: 32%|███▏ | 5764/18018 [12:31:19<23:59:59, 7.05s/step, loss=1.621, acc=0.649, acc_len=1.52, thru=15.7, I=22.4, T=29.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:33,943] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:01:39,983] loop.py:380 INFO TIMING step=5765: step=4.146s data=1.980s compute=4.108s [fwd=2.926s bwd=1.167s opt=0.003s] dispatch=2.359s [2026-10-19 02:01:39,984] loop.py:45 INFO TRAIN_STEP step=5765 loss=1.509282 Training: 32%|███▏ | 5765/18018 [12:31:26<23:27:00, 6.89s/step, loss=1.509, acc=0.667, acc_len=1.58, thru=16.1, I=19.4, T=30.9, wait=2.4s, pool=48, epoch=1/2][2026-10-19 02:01:49,046] loop.py:45 INFO TRAIN_STEP step=5766 loss=1.510651 Training: 32%|███▏ | 5766/18018 [12:31:35<25:39:57, 7.54s/step, loss=1.511, acc=0.669, acc_len=1.59, thru=15.0, I=14.1, T=32.9, wait=5.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:01:49,042] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5765: forward=2761.2ms backward=1078.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:55,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:01:57,530] loop.py:45 INFO TRAIN_STEP step=5767 loss=1.572706 Training: 32%|███▏ | 5767/18018 [12:31:43<26:37:43, 7.82s/step, loss=1.573, acc=0.654, acc_len=1.54, thru=15.7, I=19.6, T=28.0, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:01:57,618] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:02:03,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:02:04,807] loop.py:45 INFO TRAIN_STEP step=5768 loss=1.443280 Training: 32%|███▏ | 5768/18018 [12:31:50<26:04:02, 7.66s/step, loss=1.443, acc=0.675, acc_len=1.60, thru=14.8, I=22.5, T=27.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:02:05,270] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:02:10,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:02:11,505] loop.py:45 INFO TRAIN_STEP step=5769 loss=1.536421 Training: 32%|███▏ | 5769/18018 [12:31:57<25:04:57, 7.37s/step, loss=1.536, acc=0.660, acc_len=1.55, thru=17.7, I=21.2, T=30.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:02:11,935] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:02:18,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:02:20,010] loop.py:380 INFO TIMING step=5770: step=4.105s data=1.904s compute=4.057s [fwd=2.881s bwd=1.162s opt=0.003s] dispatch=4.389s [2026-10-19 02:02:20,010] loop.py:45 INFO TRAIN_STEP step=5770 loss=1.686882 Training: 32%|███▏ | 5770/18018 [12:32:06<26:14:14, 7.71s/step, loss=1.687, acc=0.640, acc_len=1.49, thru=15.3, I=17.9, T=31.2, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:02:20,075] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-19 02:02:26,729] loop.py:45 INFO TRAIN_STEP step=5771 loss=1.546971 Training: 32%|███▏ | 5771/18018 [12:32:12<25:13:19, 7.41s/step, loss=1.547, acc=0.662, acc_len=1.56, thru=16.9, I=20.6, T=31.8, wait=2.7s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:02:26,724] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5770: forward=2812.0ms backward=1163.2ms [2026-10-19 02:02:34,982] loop.py:45 INFO TRAIN_STEP step=5772 loss=1.582405 Training: 32%|███▏ | 5772/18018 [12:32:21<26:04:23, 7.66s/step, loss=1.582, acc=0.658, acc_len=1.54, thru=14.7, I=14.5, T=31.8, wait=4.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:02:42,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:02:43,101] loop.py:45 INFO TRAIN_STEP step=5773 loss=1.514694 Training: 32%|███▏ | 5773/18018 [12:32:29<26:32:06, 7.80s/step, loss=1.515, acc=0.665, acc_len=1.58, thru=17.3, I=17.6, T=28.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:02:43,565] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:02:50,391] loop.py:45 INFO TRAIN_STEP step=5774 loss=1.478971 Training: 32%|███▏ | 5774/18018 [12:32:36<26:00:43, 7.65s/step, loss=1.479, acc=0.674, acc_len=1.60, thru=16.3, I=18.8, T=32.5, wait=3.3s, pool=56, epoch=1/2][2026-10-19 02:02:58,829] loop.py:380 INFO TIMING step=5775: step=4.347s data=1.934s compute=4.319s [fwd=3.079s bwd=1.224s opt=0.003s] dispatch=4.081s [2026-10-19 02:02:58,830] loop.py:45 INFO TRAIN_STEP step=5775 loss=1.655022 Training: 32%|███▏ | 5775/18018 [12:32:44<26:48:58, 7.89s/step, loss=1.655, acc=0.635, acc_len=1.48, thru=14.6, I=15.5, T=29.4, wait=4.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:06,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:03:07,247] loop.py:45 INFO TRAIN_STEP step=5776 loss=1.668520 Training: 32%|███▏ | 5776/18018 [12:32:53<27:21:29, 8.05s/step, loss=1.669, acc=0.639, acc_len=1.48, thru=16.1, I=17.6, T=30.1, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:03:07,242] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5775: forward=3005.7ms backward=1191.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:07,342] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:15,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:03:16,099] loop.py:45 INFO TRAIN_STEP step=5777 loss=1.558333 Training: 32%|███▏ | 5777/18018 [12:33:02<28:10:36, 8.29s/step, loss=1.558, acc=0.658, acc_len=1.55, thru=13.9, I=13.6, T=33.1, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:16,342] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:20,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:03:22,557] loop.py:45 INFO TRAIN_STEP step=5778 loss=1.452289 Training: 32%|███▏ | 5778/18018 [12:33:08<26:18:33, 7.74s/step, loss=1.452, acc=0.675, acc_len=1.60, thru=18.4, I=22.5, T=31.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:22,760] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:03:28,735] loop.py:45 INFO TRAIN_STEP step=5779 loss=1.436721 Training: 32%|███▏ | 5779/18018 [12:33:14<24:42:59, 7.27s/step, loss=1.437, acc=0.675, acc_len=1.60, thru=15.3, I=24.2, T=33.2, wait=2.3s, pool=32, epoch=1/2][2026-10-19 02:03:38,626] loop.py:380 INFO TIMING step=5780: step=4.335s data=1.894s compute=4.303s [fwd=3.041s bwd=1.247s opt=0.003s] dispatch=5.547s [2026-10-19 02:03:38,626] loop.py:45 INFO TRAIN_STEP step=5780 loss=1.536800 Training: 32%|███▏ | 5780/18018 [12:33:24<27:23:13, 8.06s/step, loss=1.537, acc=0.665, acc_len=1.57, thru=14.9, I=12.6, T=29.5, wait=5.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:46,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:03:47,370] loop.py:45 INFO TRAIN_STEP step=5781 loss=1.545459 Training: 32%|███▏ | 5781/18018 [12:33:33<28:05:18, 8.26s/step, loss=1.545, acc=0.659, acc_len=1.55, thru=16.9, I=16.8, T=32.8, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:03:47,365] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5780: forward=2782.4ms backward=1074.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:47,698] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:54,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:03:55,728] loop.py:45 INFO TRAIN_STEP step=5782 loss=1.440461 Training: 32%|███▏ | 5782/18018 [12:33:41<28:10:57, 8.29s/step, loss=1.440, acc=0.678, acc_len=1.61, thru=14.8, I=17.1, T=28.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:03:55,889] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:04:02,823] loop.py:45 INFO TRAIN_STEP step=5783 loss=1.514101 Training: 32%|███▏ | 5783/18018 [12:33:48<26:57:38, 7.93s/step, loss=1.514, acc=0.664, acc_len=1.57, thru=16.6, I=18.9, T=30.5, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:07,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:04:09,204] loop.py:45 INFO TRAIN_STEP step=5784 loss=1.655108 Training: 32%|███▏ | 5784/18018 [12:33:55<25:22:27, 7.47s/step, loss=1.655, acc=0.642, acc_len=1.50, thru=19.5, I=19.8, T=32.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:09,672] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:14,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:04:16,024] loop.py:380 INFO TIMING step=5785: step=3.769s data=1.590s compute=3.739s [fwd=2.598s bwd=1.126s opt=0.003s] dispatch=3.041s [2026-10-19 02:04:16,024] loop.py:45 INFO TRAIN_STEP step=5785 loss=1.436648 Training: 32%|███▏ | 5785/18018 [12:34:02<24:42:42, 7.27s/step, loss=1.437, acc=0.678, acc_len=1.61, thru=17.7, I=24.8, T=34.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:16,052] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:24,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:04:24,989] loop.py:45 INFO TRAIN_STEP step=5786 loss=1.618662 Training: 32%|███▏ | 5786/18018 [12:34:11<26:26:18, 7.78s/step, loss=1.619, acc=0.649, acc_len=1.52, thru=14.3, I=14.8, T=27.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:25,023] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-19 02:04:24,984] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5785: forward=3246.1ms backward=1349.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:32,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:04:32,786] loop.py:45 INFO TRAIN_STEP step=5787 loss=1.547732 Training: 32%|███▏ | 5787/18018 [12:34:18<26:26:56, 7.78s/step, loss=1.548, acc=0.659, acc_len=1.56, thru=16.3, I=15.7, T=30.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:33,282] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:38,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:04:39,801] loop.py:45 INFO TRAIN_STEP step=5788 loss=1.449653 Training: 32%|███▏ | 5788/18018 [12:34:25<25:39:54, 7.55s/step, loss=1.450, acc=0.676, acc_len=1.60, thru=18.0, I=19.1, T=29.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:39,975] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:46,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:04:46,876] loop.py:45 INFO TRAIN_STEP step=5789 loss=1.595717 Training: 32%|███▏ | 5789/18018 [12:34:32<25:10:24, 7.41s/step, loss=1.596, acc=0.652, acc_len=1.52, thru=16.7, I=20.3, T=34.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:46,949] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:54,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:04:56,415] loop.py:380 INFO TIMING step=5790: step=4.816s data=2.352s compute=4.779s [fwd=3.280s bwd=1.483s opt=0.003s] dispatch=4.712s [2026-10-19 02:04:56,415] loop.py:45 INFO TRAIN_STEP step=5790 loss=1.493797 Training: 32%|███▏ | 5790/18018 [12:34:42<27:20:19, 8.05s/step, loss=1.494, acc=0.667, acc_len=1.57, thru=13.2, I=16.7, T=26.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:04:56,651] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:02,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:05:02,574] loop.py:45 INFO TRAIN_STEP step=5791 loss=1.454700 Training: 32%|███▏ | 5791/18018 [12:34:48<25:24:39, 7.48s/step, loss=1.455, acc=0.672, acc_len=1.59, thru=18.5, I=21.7, T=36.0, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:05:02,570] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5790: forward=2500.3ms backward=1008.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:02,864] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:08,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:05:09,335] loop.py:45 INFO TRAIN_STEP step=5792 loss=1.438270 Training: 32%|███▏ | 5792/18018 [12:34:55<24:40:30, 7.27s/step, loss=1.438, acc=0.679, acc_len=1.62, thru=19.6, I=20.0, T=33.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:09,364] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:16,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:05:17,408] loop.py:45 INFO TRAIN_STEP step=5793 loss=1.526219 Training: 32%|███▏ | 5793/18018 [12:35:03<25:29:42, 7.51s/step, loss=1.526, acc=0.663, acc_len=1.57, thru=15.8, I=17.3, T=25.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:17,505] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:24,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:05:25,148] loop.py:45 INFO TRAIN_STEP step=5794 loss=1.587198 Training: 32%|███▏ | 5794/18018 [12:35:11<25:43:48, 7.58s/step, loss=1.587, acc=0.654, acc_len=1.54, thru=15.5, I=16.6, T=29.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:25,414] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:30,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:05:32,044] loop.py:380 INFO TIMING step=5795: step=3.957s data=1.815s compute=3.926s [fwd=2.787s bwd=1.124s opt=0.003s] dispatch=2.929s [2026-10-19 02:05:32,044] loop.py:45 INFO TRAIN_STEP step=5795 loss=1.510653 Training: 32%|███▏ | 5795/18018 [12:35:18<25:02:06, 7.37s/step, loss=1.511, acc=0.669, acc_len=1.58, thru=18.4, I=20.9, T=32.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:32,205] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:39,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:05:39,721] loop.py:45 INFO TRAIN_STEP step=5796 loss=1.558300 Training: 32%|███▏ | 5796/18018 [12:35:25<25:20:24, 7.46s/step, loss=1.558, acc=0.659, acc_len=1.56, thru=15.6, I=19.8, T=29.5, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:05:39,715] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5795: forward=2918.7ms backward=1364.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:39,869] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:05:46,990] loop.py:45 INFO TRAIN_STEP step=5797 loss=1.381049 Training: 32%|███▏ | 5797/18018 [12:35:33<25:08:31, 7.41s/step, loss=1.381, acc=0.693, acc_len=1.66, thru=15.5, I=19.5, T=31.0, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:54,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:05:56,480] loop.py:45 INFO TRAIN_STEP step=5798 loss=1.385499 Training: 32%|███▏ | 5798/18018 [12:35:42<27:15:33, 8.03s/step, loss=1.385, acc=0.685, acc_len=1.64, thru=13.9, I=15.7, T=27.3, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:05:56,631] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:02,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:06:04,254] loop.py:45 INFO TRAIN_STEP step=5799 loss=1.427451 Training: 32%|███▏ | 5799/18018 [12:35:50<26:59:51, 7.95s/step, loss=1.427, acc=0.679, acc_len=1.62, thru=16.8, I=22.4, T=25.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:04,760] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:11,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:06:13,036] loop.py:380 INFO TIMING step=5800: step=4.464s data=2.295s compute=4.427s [fwd=3.304s bwd=1.108s opt=0.003s] dispatch=4.307s [2026-10-19 02:06:13,036] loop.py:45 INFO TRAIN_STEP step=5800 loss=1.389645 Training: 32%|███▏ | 5800/18018 [12:35:59<27:50:12, 8.20s/step, loss=1.390, acc=0.686, acc_len=1.64, thru=15.5, I=20.5, T=28.7, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:13,085] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:20,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:06:20,671] loop.py:45 INFO TRAIN_STEP step=5801 loss=1.560715 Training: 32%|███▏ | 5801/18018 [12:36:06<27:15:28, 8.03s/step, loss=1.561, acc=0.659, acc_len=1.55, thru=13.8, I=18.6, T=28.9, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:06:20,667] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5800: forward=3256.6ms backward=1133.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:20,805] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:27,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:06:28,835] loop.py:45 INFO TRAIN_STEP step=5802 loss=1.603614 Training: 32%|███▏ | 5802/18018 [12:36:14<27:23:28, 8.07s/step, loss=1.604, acc=0.652, acc_len=1.53, thru=17.3, I=18.4, T=30.0, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:29,140] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:35,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:06:36,647] loop.py:45 INFO TRAIN_STEP step=5803 loss=1.484156 Training: 32%|███▏ | 5803/18018 [12:36:22<27:07:18, 7.99s/step, loss=1.484, acc=0.672, acc_len=1.59, thru=15.7, I=18.2, T=27.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:36,964] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:42,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:06:43,186] loop.py:45 INFO TRAIN_STEP step=5804 loss=1.503903 Training: 32%|███▏ | 5804/18018 [12:36:29<25:38:23, 7.56s/step, loss=1.504, acc=0.671, acc_len=1.59, thru=16.9, I=19.7, T=33.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:43,420] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:49,880] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:06:50,632] loop.py:380 INFO TIMING step=5805: step=4.078s data=1.812s compute=4.041s [fwd=2.901s bwd=1.126s opt=0.003s] dispatch=3.360s [2026-10-19 02:06:50,632] loop.py:45 INFO TRAIN_STEP step=5805 loss=1.536388 Training: 32%|███▏ | 5805/18018 [12:36:36<25:31:32, 7.52s/step, loss=1.536, acc=0.664, acc_len=1.57, thru=17.8, I=19.7, T=31.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:50,886] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:06:58,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:06:59,891] loop.py:45 INFO TRAIN_STEP step=5806 loss=1.555097 Training: 32%|███▏ | 5806/18018 [12:36:45<27:17:11, 8.04s/step, loss=1.555, acc=0.658, acc_len=1.55, thru=14.2, I=17.5, T=24.5, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:06:59,886] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5805: forward=3677.1ms backward=1497.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:07:00,119] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:07:05,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:07:06,511] loop.py:45 INFO TRAIN_STEP step=5807 loss=1.651240 Training: 32%|███▏ | 5807/18018 [12:36:52<25:50:15, 7.62s/step, loss=1.651, acc=0.641, acc_len=1.49, thru=16.0, I=20.4, T=28.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:07:06,745] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:07:15,066] loop.py:45 INFO TRAIN_STEP step=5808 loss=1.551097 Training: 32%|███▏ | 5808/18018 [12:37:01<26:47:17, 7.90s/step, loss=1.551, acc=0.662, acc_len=1.55, thru=12.4, I=12.8, T=30.7, wait=4.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:07:21,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:07:22,355] loop.py:45 INFO TRAIN_STEP step=5809 loss=1.458910 Training: 32%|███▏ | 5809/18018 [12:37:08<26:10:04, 7.72s/step, loss=1.459, acc=0.677, acc_len=1.61, thru=17.9, I=17.3, T=29.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:07:22,429] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-19 02:07:29,427] loop.py:380 INFO TIMING step=5810: step=3.946s data=1.785s compute=3.908s [fwd=2.771s bwd=1.122s opt=0.003s] dispatch=3.116s [2026-10-19 02:07:29,427] loop.py:45 INFO TRAIN_STEP step=5810 loss=1.580462 Training: 32%|███▏ | 5810/18018 [12:37:15<25:30:33, 7.52s/step, loss=1.580, acc=0.655, acc_len=1.54, thru=14.8, I=16.8, T=32.4, wait=3.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:07:37,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:07:38,086] loop.py:45 INFO TRAIN_STEP step=5811 loss=1.619960 Training: 32%|███▏ | 5811/18018 [12:37:24<26:39:57, 7.86s/step, loss=1.620, acc=0.653, acc_len=1.52, thru=16.8, I=14.5, T=29.3, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:07:38,079] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5810: forward=2905.4ms backward=1393.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:07:38,187] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 [2026-10-19 02:07:47,187] loop.py:45 INFO TRAIN_STEP step=5812 loss=1.513660 Training: 32%|███▏ | 5812/18018 [12:37:33<27:55:08, 8.23s/step, loss=1.514, acc=0.663, acc_len=1.56, thru=13.1, I=14.0, T=26.6, wait=4.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:07:55,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:07:56,611] loop.py:45 INFO TRAIN_STEP step=5813 loss=1.622266 Training: 32%|███▏ | 5813/18018 [12:37:42<29:07:43, 8.59s/step, loss=1.622, acc=0.649, acc_len=1.51, thru=15.2, I=16.4, T=29.6, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:07:56,814] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:03,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:08:04,060] loop.py:45 INFO TRAIN_STEP step=5814 loss=1.477939 Training: 32%|███▏ | 5814/18018 [12:37:50<27:57:40, 8.25s/step, loss=1.478, acc=0.670, acc_len=1.59, thru=15.2, I=17.8, T=30.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:04,365] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:08:12,900] loop.py:380 INFO TIMING step=5815: step=4.261s data=1.876s compute=4.227s [fwd=3.025s bwd=1.187s opt=0.003s] dispatch=4.572s [2026-10-19 02:08:12,900] loop.py:45 INFO TRAIN_STEP step=5815 loss=1.418159 Training: 32%|███▏ | 5815/18018 [12:37:58<28:33:52, 8.43s/step, loss=1.418, acc=0.680, acc_len=1.62, thru=15.1, I=17.8, T=30.0, wait=4.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:21,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:08:22,533] loop.py:45 INFO TRAIN_STEP step=5816 loss=1.449927 Training: 32%|███▏ | 5816/18018 [12:38:08<29:47:15, 8.79s/step, loss=1.450, acc=0.675, acc_len=1.60, thru=14.6, I=12.7, T=27.1, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:08:22,527] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5815: forward=3348.3ms backward=1315.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:22,917] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:08:30,292] loop.py:45 INFO TRAIN_STEP step=5817 loss=1.620072 Training: 32%|███▏ | 5817/18018 [12:38:16<28:44:10, 8.48s/step, loss=1.620, acc=0.647, acc_len=1.51, thru=15.6, I=17.8, T=31.1, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:37,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:08:38,592] loop.py:45 INFO TRAIN_STEP step=5818 loss=1.469737 Training: 32%|███▏ | 5818/18018 [12:38:24<28:33:17, 8.43s/step, loss=1.470, acc=0.673, acc_len=1.59, thru=14.9, I=16.4, T=29.1, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:38,874] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:45,179] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:08:47,402] loop.py:45 INFO TRAIN_STEP step=5819 loss=1.471464 Training: 32%|███▏ | 5819/18018 [12:38:33<28:56:33, 8.54s/step, loss=1.471, acc=0.673, acc_len=1.60, thru=15.0, I=18.4, T=27.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:47,695] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:54,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:08:56,209] loop.py:380 INFO TIMING step=5820: step=4.473s data=2.031s compute=4.441s [fwd=3.220s bwd=1.207s opt=0.003s] dispatch=4.324s [2026-10-19 02:08:56,209] loop.py:45 INFO TRAIN_STEP step=5820 loss=1.568501 Training: 32%|███▏ | 5820/18018 [12:38:42<29:12:31, 8.62s/step, loss=1.569, acc=0.655, acc_len=1.53, thru=15.0, I=17.3, T=28.6, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:08:56,447] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:09:03,915] loop.py:45 INFO TRAIN_STEP step=5821 loss=1.598536 Training: 32%|███▏ | 5821/18018 [12:38:50<28:16:42, 8.35s/step, loss=1.599, acc=0.650, acc_len=1.53, thru=14.1, I=17.3, T=28.7, wait=3.2s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:09:03,911] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5820: forward=3145.3ms backward=1264.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:11,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:09:13,663] loop.py:45 INFO TRAIN_STEP step=5822 loss=1.453627 Training: 32%|███▏ | 5822/18018 [12:38:59<29:41:59, 8.77s/step, loss=1.454, acc=0.676, acc_len=1.61, thru=15.3, I=17.8, T=28.7, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:14,135] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:19,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:09:21,954] loop.py:45 INFO TRAIN_STEP step=5823 loss=1.498268 Training: 32%|███▏ | 5823/18018 [12:39:08<29:12:53, 8.62s/step, loss=1.498, acc=0.671, acc_len=1.60, thru=16.4, I=20.7, T=30.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:21,989] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:28,883] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:09:29,508] loop.py:45 INFO TRAIN_STEP step=5824 loss=1.440541 Training: 32%|███▏ | 5824/18018 [12:39:15<28:07:30, 8.30s/step, loss=1.441, acc=0.679, acc_len=1.61, thru=14.8, I=19.0, T=28.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:29,889] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:09:35,953] loop.py:380 INFO TIMING step=5825: step=3.979s data=1.814s compute=3.954s [fwd=2.827s bwd=1.113s opt=0.003s] dispatch=2.456s [2026-10-19 02:09:35,953] loop.py:45 INFO TRAIN_STEP step=5825 loss=1.425198 Training: 32%|███▏ | 5825/18018 [12:39:22<26:13:57, 7.75s/step, loss=1.425, acc=0.684, acc_len=1.64, thru=16.7, I=18.3, T=32.2, wait=2.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:43,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:09:45,034] loop.py:45 INFO TRAIN_STEP step=5826 loss=1.563340 Training: 32%|███▏ | 5826/18018 [12:39:31<27:35:23, 8.15s/step, loss=1.563, acc=0.659, acc_len=1.55, thru=15.7, I=16.5, T=30.3, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:09:45,029] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5825: forward=2946.0ms backward=1224.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:45,238] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:52,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:09:53,713] loop.py:45 INFO TRAIN_STEP step=5827 loss=1.586087 Training: 32%|███▏ | 5827/18018 [12:39:39<28:07:29, 8.31s/step, loss=1.586, acc=0.658, acc_len=1.55, thru=13.2, I=17.8, T=26.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:09:54,128] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:00,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:10:00,496] loop.py:45 INFO TRAIN_STEP step=5828 loss=1.402180 Training: 32%|███▏ | 5828/18018 [12:39:46<26:34:36, 7.85s/step, loss=1.402, acc=0.680, acc_len=1.62, thru=18.3, I=19.5, T=34.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:00,834] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:08,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:10:11,129] loop.py:45 INFO TRAIN_STEP step=5829 loss=1.636523 Training: 32%|███▏ | 5829/18018 [12:39:57<29:24:12, 8.68s/step, loss=1.637, acc=0.644, acc_len=1.51, thru=12.4, I=15.2, T=24.4, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:11,353] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 02:10:18,364] loop.py:380 INFO TIMING step=5830: step=4.330s data=2.014s compute=4.303s [fwd=3.114s bwd=1.174s opt=0.003s] dispatch=2.896s [2026-10-19 02:10:18,364] loop.py:45 INFO TRAIN_STEP step=5830 loss=1.558001 Training: 32%|███▏ | 5830/18018 [12:40:04<27:55:42, 8.25s/step, loss=1.558, acc=0.656, acc_len=1.55, thru=12.2, I=16.2, T=29.6, wait=2.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:24,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:10:26,326] loop.py:45 INFO TRAIN_STEP step=5831 loss=1.571365 Training: 32%|███▏ | 5831/18018 [12:40:12<27:38:14, 8.16s/step, loss=1.571, acc=0.658, acc_len=1.55, thru=17.6, I=17.6, T=31.1, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:10:26,322] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5830: forward=2938.7ms backward=1136.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:26,611] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:33,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:10:33,539] loop.py:45 INFO TRAIN_STEP step=5832 loss=1.593643 Training: 32%|███▏ | 5832/18018 [12:40:19<26:40:04, 7.88s/step, loss=1.594, acc=0.652, acc_len=1.53, thru=15.2, I=17.8, T=33.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:33,855] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:10:40,884] loop.py:45 INFO TRAIN_STEP step=5833 loss=1.586341 Training: 32%|███▏ | 5833/18018 [12:40:26<26:07:24, 7.72s/step, loss=1.586, acc=0.653, acc_len=1.53, thru=15.9, I=15.6, T=28.2, wait=2.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:48,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:10:48,923] loop.py:45 INFO TRAIN_STEP step=5834 loss=1.535949 Training: 32%|███▏ | 5834/18018 [12:40:35<26:26:55, 7.81s/step, loss=1.536, acc=0.665, acc_len=1.57, thru=17.4, I=16.0, T=29.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:49,266] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:54,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:10:55,908] loop.py:380 INFO TIMING step=5835: step=4.120s data=1.871s compute=4.082s [fwd=2.867s bwd=1.200s opt=0.003s] dispatch=2.854s [2026-10-19 02:10:55,909] loop.py:45 INFO TRAIN_STEP step=5835 loss=1.492027 Training: 32%|███▏ | 5835/18018 [12:40:41<25:36:10, 7.57s/step, loss=1.492, acc=0.667, acc_len=1.58, thru=17.4, I=20.9, T=31.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:10:56,015] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:00,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:11:02,212] loop.py:45 INFO TRAIN_STEP step=5836 loss=1.482718 Training: 32%|███▏ | 5836/18018 [12:40:48<24:19:18, 7.19s/step, loss=1.483, acc=0.674, acc_len=1.60, thru=18.6, I=24.9, T=28.8, wait=1.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:11:02,207] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5835: forward=3221.3ms backward=1178.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:02,246] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:06,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:11:08,861] loop.py:45 INFO TRAIN_STEP step=5837 loss=1.436563 Training: 32%|███▏ | 5837/18018 [12:40:54<23:46:21, 7.03s/step, loss=1.437, acc=0.679, acc_len=1.61, thru=17.2, I=25.9, T=29.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:08,966] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:15,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:11:16,420] loop.py:45 INFO TRAIN_STEP step=5838 loss=1.453743 Training: 32%|███▏ | 5838/18018 [12:41:02<24:18:34, 7.19s/step, loss=1.454, acc=0.672, acc_len=1.60, thru=13.2, I=19.3, T=29.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:16,763] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:21,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:11:23,139] loop.py:45 INFO TRAIN_STEP step=5839 loss=1.616415 Training: 32%|███▏ | 5839/18018 [12:41:09<23:50:05, 7.05s/step, loss=1.616, acc=0.648, acc_len=1.51, thru=17.8, I=20.9, T=32.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:23,633] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:11:31,393] loop.py:380 INFO TIMING step=5840: step=3.951s data=1.768s compute=3.918s [fwd=2.802s bwd=1.101s opt=0.003s] dispatch=4.294s [2026-10-19 02:11:31,393] loop.py:45 INFO TRAIN_STEP step=5840 loss=1.498630 Training: 32%|███▏ | 5840/18018 [12:41:17<25:03:40, 7.41s/step, loss=1.499, acc=0.671, acc_len=1.60, thru=11.7, I=15.6, T=32.4, wait=4.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:40,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:11:41,578] loop.py:45 INFO TRAIN_STEP step=5841 loss=1.388065 Training: 32%|███▏ | 5841/18018 [12:41:27<27:52:38, 8.24s/step, loss=1.388, acc=0.686, acc_len=1.64, thru=15.3, I=14.2, T=28.0, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:11:41,572] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5840: forward=3228.3ms backward=1289.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:41,841] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:47,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:11:47,903] loop.py:45 INFO TRAIN_STEP step=5842 loss=1.521594 Training: 32%|███▏ | 5842/18018 [12:41:33<25:55:41, 7.67s/step, loss=1.522, acc=0.663, acc_len=1.57, thru=17.4, I=20.0, T=32.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:48,231] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:52,822] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:11:54,933] loop.py:45 INFO TRAIN_STEP step=5843 loss=1.377147 Training: 32%|███▏ | 5843/18018 [12:41:41<25:16:56, 7.48s/step, loss=1.377, acc=0.690, acc_len=1.65, thru=17.8, I=21.9, T=28.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:11:55,340] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:01,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:12:02,586] loop.py:45 INFO TRAIN_STEP step=5844 loss=1.600753 Training: 32%|███▏ | 5844/18018 [12:41:48<25:27:45, 7.53s/step, loss=1.601, acc=0.653, acc_len=1.53, thru=14.1, I=21.3, T=32.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:02,869] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:12:08,941] loop.py:380 INFO TIMING step=5845: step=3.694s data=1.594s compute=3.661s [fwd=2.615s bwd=1.031s opt=0.003s] dispatch=2.647s [2026-10-19 02:12:08,941] loop.py:45 INFO TRAIN_STEP step=5845 loss=1.493124 Training: 32%|███▏ | 5845/18018 [12:41:55<24:16:00, 7.18s/step, loss=1.493, acc=0.669, acc_len=1.58, thru=15.6, I=21.3, T=34.6, wait=2.6s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:16,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:12:18,600] loop.py:45 INFO TRAIN_STEP step=5846 loss=1.563455 Training: 32%|███▏ | 5846/18018 [12:42:04<26:46:56, 7.92s/step, loss=1.563, acc=0.661, acc_len=1.56, thru=15.7, I=17.2, T=30.2, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:12:18,594] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5845: forward=2910.3ms backward=1281.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:18,963] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:25,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:12:26,265] loop.py:45 INFO TRAIN_STEP step=5847 loss=1.589284 Training: 32%|███▏ | 5847/18018 [12:42:12<26:31:15, 7.84s/step, loss=1.589, acc=0.654, acc_len=1.54, thru=15.3, I=20.0, T=33.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:26,772] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:32,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:12:32,773] loop.py:45 INFO TRAIN_STEP step=5848 loss=1.493748 Training: 32%|███▏ | 5848/18018 [12:42:18<25:09:38, 7.44s/step, loss=1.494, acc=0.670, acc_len=1.59, thru=19.4, I=20.7, T=33.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:32,819] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:38,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:12:39,130] loop.py:45 INFO TRAIN_STEP step=5849 loss=1.644105 Training: 32%|███▏ | 5849/18018 [12:42:25<24:03:27, 7.12s/step, loss=1.644, acc=0.647, acc_len=1.51, thru=19.4, I=21.3, T=30.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:39,415] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:44,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:12:45,833] loop.py:380 INFO TIMING step=5850: step=3.896s data=1.662s compute=3.855s [fwd=2.742s bwd=1.099s opt=0.003s] dispatch=2.798s [2026-10-19 02:12:45,834] loop.py:45 INFO TRAIN_STEP step=5850 loss=1.605987 Training: 32%|███▏ | 5850/18018 [12:42:31<23:38:13, 6.99s/step, loss=1.606, acc=0.648, acc_len=1.52, thru=16.9, I=20.2, T=32.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:45,900] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-19 02:12:52,592] loop.py:45 INFO TRAIN_STEP step=5851 loss=1.474139 Training: 32%|███▏ | 5851/18018 [12:42:38<23:23:48, 6.92s/step, loss=1.474, acc=0.669, acc_len=1.58, thru=16.6, I=19.6, T=32.6, wait=2.8s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:12:52,586] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5850: forward=2733.6ms backward=1157.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:12:58,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:13:00,346] loop.py:45 INFO TRAIN_STEP step=5852 loss=1.522736 Training: 32%|███▏ | 5852/18018 [12:42:46<24:14:14, 7.17s/step, loss=1.523, acc=0.663, acc_len=1.56, thru=17.8, I=19.0, T=31.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:00,839] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:06,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:13:07,375] loop.py:45 INFO TRAIN_STEP step=5853 loss=1.601846 Training: 32%|███▏ | 5853/18018 [12:42:53<24:05:23, 7.13s/step, loss=1.602, acc=0.652, acc_len=1.53, thru=16.4, I=22.2, T=31.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:07,722] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:12,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:13:14,324] loop.py:45 INFO TRAIN_STEP step=5854 loss=1.468297 Training: 32%|███▏ | 5854/18018 [12:43:00<23:54:18, 7.07s/step, loss=1.468, acc=0.675, acc_len=1.61, thru=17.6, I=22.4, T=29.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:14,742] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:13:24,973] loop.py:380 INFO TIMING step=5855: step=4.815s data=2.191s compute=4.767s [fwd=3.108s bwd=1.644s opt=0.003s] dispatch=5.827s [2026-10-19 02:13:24,974] loop.py:45 INFO TRAIN_STEP step=5855 loss=1.553616 Training: 32%|███▏ | 5855/18018 [12:43:11<27:31:51, 8.15s/step, loss=1.554, acc=0.661, acc_len=1.56, thru=11.0, I=11.8, T=26.6, wait=5.8s, pool=52, epoch=1/2][2026-10-19 02:13:34,091] loop.py:45 INFO TRAIN_STEP step=5856 loss=1.511984 Training: 33%|███▎ | 5856/18018 [12:43:20<28:30:25, 8.44s/step, loss=1.512, acc=0.668, acc_len=1.58, thru=15.7, I=13.9, T=27.2, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:34,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 02:13:34,087] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5855: forward=3143.8ms backward=1492.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:34,536] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:13:43,020] loop.py:45 INFO TRAIN_STEP step=5857 loss=1.527367 Training: 33%|███▎ | 5857/18018 [12:43:29<29:00:14, 8.59s/step, loss=1.527, acc=0.662, acc_len=1.56, thru=13.8, I=13.6, T=27.9, wait=4.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:50,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:13:51,055] loop.py:45 INFO TRAIN_STEP step=5858 loss=1.392243 Training: 33%|███▎ | 5858/18018 [12:43:37<28:26:33, 8.42s/step, loss=1.392, acc=0.689, acc_len=1.65, thru=17.1, I=15.8, T=29.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:51,189] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:56,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:13:58,290] loop.py:45 INFO TRAIN_STEP step=5859 loss=1.543142 Training: 33%|███▎ | 5859/18018 [12:43:44<27:14:23, 8.07s/step, loss=1.543, acc=0.663, acc_len=1.56, thru=17.0, I=20.0, T=29.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:13:58,327] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:04,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:14:05,714] loop.py:380 INFO TIMING step=5860: step=4.266s data=2.006s compute=4.225s [fwd=2.974s bwd=1.236s opt=0.003s] dispatch=3.148s [2026-10-19 02:14:05,714] loop.py:45 INFO TRAIN_STEP step=5860 loss=1.439993 Training: 33%|███▎ | 5860/18018 [12:43:51<26:35:08, 7.87s/step, loss=1.440, acc=0.679, acc_len=1.61, thru=15.2, I=20.0, T=30.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:05,842] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 02:14:13,343] loop.py:45 INFO TRAIN_STEP step=5861 loss=1.507385 Training: 33%|███▎ | 5861/18018 [12:43:59<26:20:10, 7.80s/step, loss=1.507, acc=0.667, acc_len=1.57, thru=16.7, I=17.5, T=30.2, wait=3.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:14:13,338] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5860: forward=2898.3ms backward=1300.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:20,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:14:22,776] loop.py:45 INFO TRAIN_STEP step=5862 loss=1.395078 Training: 33%|███▎ | 5862/18018 [12:44:08<27:59:36, 8.29s/step, loss=1.395, acc=0.684, acc_len=1.64, thru=14.6, I=16.9, T=26.4, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:23,038] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:29,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:14:29,863] loop.py:45 INFO TRAIN_STEP step=5863 loss=1.565120 Training: 33%|███▎ | 5863/18018 [12:44:15<26:46:09, 7.93s/step, loss=1.565, acc=0.660, acc_len=1.56, thru=14.8, I=19.1, T=31.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:30,104] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:35,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:14:37,572] loop.py:45 INFO TRAIN_STEP step=5864 loss=1.428864 Training: 33%|███▎ | 5864/18018 [12:44:23<26:32:51, 7.86s/step, loss=1.429, acc=0.683, acc_len=1.63, thru=17.9, I=20.1, T=28.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:37,810] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:14:45,480] loop.py:380 INFO TIMING step=5865: step=4.021s data=1.845s compute=3.983s [fwd=2.815s bwd=1.153s opt=0.003s] dispatch=3.876s [2026-10-19 02:14:45,481] loop.py:45 INFO TRAIN_STEP step=5865 loss=1.596024 Training: 33%|███▎ | 5865/18018 [12:44:31<26:35:24, 7.88s/step, loss=1.596, acc=0.655, acc_len=1.53, thru=13.4, I=16.2, T=31.8, wait=3.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:52,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:14:53,921] loop.py:45 INFO TRAIN_STEP step=5866 loss=1.424811 Training: 33%|███▎ | 5866/18018 [12:44:40<27:09:33, 8.05s/step, loss=1.425, acc=0.683, acc_len=1.63, thru=16.1, I=17.0, T=31.5, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:14:53,916] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5865: forward=2823.6ms backward=1201.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:54,045] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:14:59,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:15:00,157] loop.py:45 INFO TRAIN_STEP step=5867 loss=1.702001 Training: 33%|███▎ | 5867/18018 [12:44:46<25:19:22, 7.50s/step, loss=1.702, acc=0.636, acc_len=1.47, thru=19.0, I=21.1, T=33.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:15:00,483] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:15:07,718] loop.py:45 INFO TRAIN_STEP step=5868 loss=1.493386 Training: 33%|███▎ | 5868/18018 [12:44:53<25:22:50, 7.52s/step, loss=1.493, acc=0.669, acc_len=1.59, thru=14.9, I=16.3, T=31.0, wait=3.4s, pool=44, epoch=1/2][2026-10-19 02:15:16,716] loop.py:45 INFO TRAIN_STEP step=5869 loss=1.542390 Training: 33%|███▎ | 5869/18018 [12:45:02<26:52:27, 7.96s/step, loss=1.542, acc=0.662, acc_len=1.57, thru=14.6, I=16.6, T=29.7, wait=4.7s, pool=44, epoch=1/2][2026-10-19 02:15:25,025] loop.py:380 INFO TIMING step=5870: step=4.168s data=1.909s compute=4.129s [fwd=2.944s bwd=1.170s opt=0.003s] dispatch=4.132s [2026-10-19 02:15:25,025] loop.py:45 INFO TRAIN_STEP step=5870 loss=1.555500 Training: 33%|███▎ | 5870/18018 [12:45:11<27:13:19, 8.07s/step, loss=1.556, acc=0.656, acc_len=1.54, thru=15.1, I=13.7, T=30.7, wait=4.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:15:31,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:15:33,371] loop.py:45 INFO TRAIN_STEP step=5871 loss=1.535806 Training: 33%|███▎ | 5871/18018 [12:45:19<27:30:09, 8.15s/step, loss=1.536, acc=0.668, acc_len=1.59, thru=17.0, I=18.4, T=31.8, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:15:33,366] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5870: forward=2709.1ms backward=1248.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:15:33,460] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 02:15:39,668] loop.py:45 INFO TRAIN_STEP step=5872 loss=1.342850 Training: 33%|███▎ | 5872/18018 [12:45:25<25:37:26, 7.59s/step, loss=1.343, acc=0.698, acc_len=1.68, thru=16.6, I=21.9, T=35.1, wait=2.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:15:47,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:15:47,302] loop.py:45 INFO TRAIN_STEP step=5873 loss=1.466573 Training: 33%|███▎ | 5873/18018 [12:45:33<25:39:36, 7.61s/step, loss=1.467, acc=0.673, acc_len=1.60, thru=17.7, I=18.7, T=32.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:15:47,515] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:15:54,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:15:56,288] loop.py:45 INFO TRAIN_STEP step=5874 loss=1.545825 Training: 33%|███▎ | 5874/18018 [12:45:42<27:03:23, 8.02s/step, loss=1.546, acc=0.666, acc_len=1.57, thru=13.7, I=15.6, T=31.9, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:15:56,474] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:16:05,707] loop.py:380 INFO TIMING step=5875: step=3.861s data=1.686s compute=3.820s [fwd=2.739s bwd=1.067s opt=0.003s] dispatch=5.549s [2026-10-19 02:16:05,708] loop.py:45 INFO TRAIN_STEP step=5875 loss=1.691917 Training: 33%|███▎ | 5875/18018 [12:45:51<28:28:02, 8.44s/step, loss=1.692, acc=0.642, acc_len=1.50, thru=12.6, I=11.7, T=33.2, wait=5.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:11,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:16:12,208] loop.py:45 INFO TRAIN_STEP step=5876 loss=1.640432 Training: 33%|███▎ | 5876/18018 [12:45:58<26:30:12, 7.86s/step, loss=1.640, acc=0.643, acc_len=1.50, thru=19.8, I=19.1, T=31.5, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:16:12,203] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5875: forward=2770.8ms backward=1244.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:12,610] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:17,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:16:18,412] loop.py:45 INFO TRAIN_STEP step=5877 loss=1.500065 Training: 33%|███▎ | 5877/18018 [12:46:04<24:49:39, 7.36s/step, loss=1.500, acc=0.667, acc_len=1.58, thru=19.4, I=21.7, T=34.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:18,854] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:25,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:16:25,975] loop.py:45 INFO TRAIN_STEP step=5878 loss=1.512781 Training: 33%|███▎ | 5878/18018 [12:46:12<25:01:46, 7.42s/step, loss=1.513, acc=0.667, acc_len=1.58, thru=17.3, I=20.3, T=36.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:26,392] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:33,168] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:16:34,003] loop.py:45 INFO TRAIN_STEP step=5879 loss=1.508983 Training: 33%|███▎ | 5879/18018 [12:46:20<25:38:38, 7.61s/step, loss=1.509, acc=0.668, acc_len=1.59, thru=16.5, I=18.7, T=30.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:34,171] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:40,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:16:41,174] loop.py:380 INFO TIMING step=5880: step=3.981s data=1.824s compute=3.946s [fwd=2.775s bwd=1.156s opt=0.003s] dispatch=3.177s [2026-10-19 02:16:41,174] loop.py:45 INFO TRAIN_STEP step=5880 loss=1.624453 Training: 33%|███▎ | 5880/18018 [12:46:27<25:11:53, 7.47s/step, loss=1.624, acc=0.646, acc_len=1.51, thru=18.6, I=21.1, T=32.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:41,657] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:16:49,572] loop.py:45 INFO TRAIN_STEP step=5881 loss=1.425783 Training: 33%|███▎ | 5881/18018 [12:46:35<26:08:06, 7.75s/step, loss=1.426, acc=0.679, acc_len=1.62, thru=12.1, I=16.3, T=35.9, wait=4.8s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:16:49,568] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5880: forward=2512.9ms backward=1007.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:57,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:16:59,513] loop.py:45 INFO TRAIN_STEP step=5882 loss=1.662011 Training: 33%|███▎ | 5882/18018 [12:46:45<28:20:35, 8.41s/step, loss=1.662, acc=0.639, acc_len=1.49, thru=16.0, I=16.7, T=31.4, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:16:59,790] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:04,906] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:17:06,348] loop.py:45 INFO TRAIN_STEP step=5883 loss=1.510597 Training: 33%|███▎ | 5883/18018 [12:46:52<26:45:09, 7.94s/step, loss=1.511, acc=0.667, acc_len=1.57, thru=16.4, I=24.6, T=30.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:06,424] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:12,757] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:17:13,536] loop.py:45 INFO TRAIN_STEP step=5884 loss=1.409669 Training: 33%|███▎ | 5884/18018 [12:46:59<25:59:37, 7.71s/step, loss=1.410, acc=0.685, acc_len=1.63, thru=17.4, I=20.9, T=28.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:13,766] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:17:22,103] loop.py:380 INFO TIMING step=5885: step=4.189s data=1.904s compute=4.145s [fwd=3.008s bwd=1.122s opt=0.003s] dispatch=4.368s [2026-10-19 02:17:22,104] loop.py:45 INFO TRAIN_STEP step=5885 loss=1.423607 Training: 33%|███▎ | 5885/18018 [12:47:08<26:51:16, 7.97s/step, loss=1.424, acc=0.682, acc_len=1.62, thru=14.9, I=17.9, T=30.6, wait=4.4s, pool=60, epoch=1/2][2026-10-19 02:17:30,510] loop.py:45 INFO TRAIN_STEP step=5886 loss=1.544586 Training: 33%|███▎ | 5886/18018 [12:47:16<27:17:46, 8.10s/step, loss=1.545, acc=0.661, acc_len=1.55, thru=14.4, I=16.3, T=29.8, wait=4.1s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:17:30,506] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5885: forward=3058.2ms backward=1185.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:37,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:17:39,243] loop.py:45 INFO TRAIN_STEP step=5887 loss=1.584475 Training: 33%|███▎ | 5887/18018 [12:47:25<27:55:59, 8.29s/step, loss=1.584, acc=0.659, acc_len=1.55, thru=15.0, I=16.6, T=30.5, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:39,288] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:44,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:17:45,653] loop.py:45 INFO TRAIN_STEP step=5888 loss=1.439162 Training: 33%|███▎ | 5888/18018 [12:47:31<26:02:00, 7.73s/step, loss=1.439, acc=0.677, acc_len=1.61, thru=17.4, I=25.1, T=33.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:45,715] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:55,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:17:55,319] loop.py:45 INFO TRAIN_STEP step=5889 loss=1.448732 Training: 33%|███▎ | 5889/18018 [12:47:41<27:59:21, 8.31s/step, loss=1.449, acc=0.677, acc_len=1.61, thru=13.3, I=12.7, T=26.8, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:17:55,602] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:04,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:18:04,778] loop.py:380 INFO TIMING step=5890: step=5.312s data=2.588s compute=5.278s [fwd=3.599s bwd=1.662s opt=0.003s] dispatch=4.139s [2026-10-19 02:18:04,778] loop.py:45 INFO TRAIN_STEP step=5890 loss=1.496859 Training: 33%|███▎ | 5890/18018 [12:47:50<29:09:07, 8.65s/step, loss=1.497, acc=0.671, acc_len=1.59, thru=13.7, I=16.9, T=24.1, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:05,217] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:10,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:18:12,160] loop.py:45 INFO TRAIN_STEP step=5891 loss=1.501578 Training: 33%|███▎ | 5891/18018 [12:47:58<27:51:55, 8.27s/step, loss=1.502, acc=0.671, acc_len=1.59, thru=15.5, I=16.4, T=30.4, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:18:12,156] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5890: forward=2969.7ms backward=1191.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:12,548] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:18,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:18:19,267] loop.py:45 INFO TRAIN_STEP step=5892 loss=1.415336 Training: 33%|███▎ | 5892/18018 [12:48:05<26:41:02, 7.92s/step, loss=1.415, acc=0.681, acc_len=1.63, thru=17.7, I=24.2, T=31.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:19,693] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:26,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:18:26,970] loop.py:45 INFO TRAIN_STEP step=5893 loss=1.479726 Training: 33%|███▎ | 5893/18018 [12:48:13<26:27:39, 7.86s/step, loss=1.480, acc=0.673, acc_len=1.61, thru=14.9, I=17.2, T=33.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:27,315] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:33,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:18:34,803] loop.py:45 INFO TRAIN_STEP step=5894 loss=1.379402 Training: 33%|███▎ | 5894/18018 [12:48:20<26:26:15, 7.85s/step, loss=1.379, acc=0.686, acc_len=1.65, thru=16.9, I=21.5, T=28.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:35,096] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:41,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:18:42,545] loop.py:380 INFO TIMING step=5895: step=4.358s data=1.958s compute=4.310s [fwd=2.921s bwd=1.374s opt=0.003s] dispatch=3.373s [2026-10-19 02:18:42,546] loop.py:45 INFO TRAIN_STEP step=5895 loss=1.546544 Training: 33%|███▎ | 5895/18018 [12:48:28<26:19:26, 7.82s/step, loss=1.547, acc=0.660, acc_len=1.55, thru=15.0, I=21.2, T=29.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:42,624] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-19 02:18:49,543] loop.py:45 INFO TRAIN_STEP step=5896 loss=1.696786 Training: 33%|███▎ | 5896/18018 [12:48:35<25:29:40, 7.57s/step, loss=1.697, acc=0.637, acc_len=1.48, thru=14.2, I=15.7, T=30.7, wait=2.8s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:18:49,537] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5895: forward=2956.0ms backward=1163.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:57,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:18:59,187] loop.py:45 INFO TRAIN_STEP step=5897 loss=1.629452 Training: 33%|███▎ | 5897/18018 [12:48:45<27:35:14, 8.19s/step, loss=1.629, acc=0.649, acc_len=1.51, thru=15.8, I=18.0, T=29.2, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:18:59,371] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:04,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:19:05,230] loop.py:45 INFO TRAIN_STEP step=5898 loss=1.413130 Training: 33%|███▎ | 5898/18018 [12:48:51<25:24:39, 7.55s/step, loss=1.413, acc=0.682, acc_len=1.63, thru=18.9, I=27.7, T=35.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:05,613] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:19:13,519] loop.py:45 INFO TRAIN_STEP step=5899 loss=1.511214 Training: 33%|███▎ | 5899/18018 [12:48:59<26:09:34, 7.77s/step, loss=1.511, acc=0.662, acc_len=1.56, thru=13.0, I=15.5, T=30.7, wait=4.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:19,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:19:22,989] loop.py:380 INFO TIMING step=5900: step=4.917s data=2.421s compute=4.870s [fwd=3.393s bwd=1.462s opt=0.003s] dispatch=4.542s [2026-10-19 02:19:22,989] loop.py:45 INFO TRAIN_STEP step=5900 loss=1.564753 Training: 33%|███▎ | 5900/18018 [12:49:09<27:52:23, 8.28s/step, loss=1.565, acc=0.659, acc_len=1.56, thru=14.6, I=19.8, T=26.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:23,257] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:28,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:19:29,145] loop.py:45 INFO TRAIN_STEP step=5901 loss=1.437632 Training: 33%|███▎ | 5901/18018 [12:49:15<25:43:25, 7.64s/step, loss=1.438, acc=0.682, acc_len=1.63, thru=15.2, I=27.6, T=34.9, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:19:29,140] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5900: forward=2558.6ms backward=1071.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:29,212] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-19 02:19:34,830] loop.py:45 INFO TRAIN_STEP step=5902 loss=1.496289 Training: 33%|███▎ | 5902/18018 [12:49:20<23:44:44, 7.06s/step, loss=1.496, acc=0.674, acc_len=1.60, thru=21.0, I=22.6, T=35.0, wait=2.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:40,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:19:41,880] loop.py:45 INFO TRAIN_STEP step=5903 loss=1.551859 Training: 33%|███▎ | 5903/18018 [12:49:27<23:44:17, 7.05s/step, loss=1.552, acc=0.658, acc_len=1.55, thru=17.6, I=19.7, T=29.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:41,969] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:47,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:19:48,624] loop.py:45 INFO TRAIN_STEP step=5904 loss=1.574587 Training: 33%|███▎ | 5904/18018 [12:49:34<23:25:26, 6.96s/step, loss=1.575, acc=0.654, acc_len=1.53, thru=17.0, I=21.2, T=31.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:48,973] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:55,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:19:58,000] loop.py:380 INFO TIMING step=5905: step=5.027s data=2.454s compute=4.977s [fwd=3.528s bwd=1.434s opt=0.003s] dispatch=4.339s [2026-10-19 02:19:58,001] loop.py:45 INFO TRAIN_STEP step=5905 loss=1.626664 Training: 33%|███▎ | 5905/18018 [12:49:44<25:51:38, 7.69s/step, loss=1.627, acc=0.653, acc_len=1.53, thru=14.5, I=16.5, T=25.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:19:58,377] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 02:20:04,854] loop.py:45 INFO TRAIN_STEP step=5906 loss=1.588877 Training: 33%|███▎ | 5906/18018 [12:49:50<25:01:01, 7.44s/step, loss=1.589, acc=0.648, acc_len=1.52, thru=13.4, I=19.5, T=29.9, wait=2.6s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:20:04,849] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5905: forward=2955.3ms backward=1274.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:20:14,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:20:15,664] loop.py:45 INFO TRAIN_STEP step=5907 loss=1.408798 Training: 33%|███▎ | 5907/18018 [12:50:01<28:25:13, 8.45s/step, loss=1.409, acc=0.687, acc_len=1.64, thru=14.2, I=14.4, T=26.8, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:20:15,697] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 02:20:22,591] loop.py:45 INFO TRAIN_STEP step=5908 loss=1.515462 Training: 33%|███▎ | 5908/18018 [12:50:08<26:53:01, 7.99s/step, loss=1.515, acc=0.665, acc_len=1.58, thru=15.5, I=17.9, T=32.3, wait=3.0s, pool=60, epoch=1/2][2026-10-19 02:20:31,544] loop.py:45 INFO TRAIN_STEP step=5909 loss=1.466176 Training: 33%|███▎ | 5909/18018 [12:50:17<27:51:09, 8.28s/step, loss=1.466, acc=0.673, acc_len=1.60, thru=14.6, I=14.2, T=32.8, wait=5.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:20:38,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:20:41,321] loop.py:380 INFO TIMING step=5910: step=5.006s data=2.698s compute=4.959s [fwd=3.639s bwd=1.305s opt=0.003s] dispatch=4.760s [2026-10-19 02:20:41,321] loop.py:45 INFO TRAIN_STEP step=5910 loss=1.597400 Training: 33%|███▎ | 5910/18018 [12:50:27<29:21:24, 8.73s/step, loss=1.597, acc=0.652, acc_len=1.53, thru=13.9, I=17.1, T=25.6, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:20:41,377] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 [2026-10-19 02:20:48,410] loop.py:45 INFO TRAIN_STEP step=5911 loss=1.518571 Training: 33%|███▎ | 5911/18018 [12:50:34<27:42:19, 8.24s/step, loss=1.519, acc=0.665, acc_len=1.57, thru=14.1, I=20.6, T=29.9, wait=2.8s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:20:48,406] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5910: forward=3031.5ms backward=1195.4ms [2026-10-19 02:20:55,849] loop.py:45 INFO TRAIN_STEP step=5912 loss=1.427303 Training: 33%|███▎ | 5912/18018 [12:50:41<26:53:36, 8.00s/step, loss=1.427, acc=0.685, acc_len=1.64, thru=14.7, I=15.2, T=28.4, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:03,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:21:05,632] loop.py:45 INFO TRAIN_STEP step=5913 loss=1.516906 Training: 33%|███▎ | 5913/18018 [12:50:51<28:41:32, 8.53s/step, loss=1.517, acc=0.666, acc_len=1.57, thru=14.1, I=16.1, T=23.8, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:05,827] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:12,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:21:13,567] loop.py:45 INFO TRAIN_STEP step=5914 loss=1.598045 Training: 33%|███▎ | 5914/18018 [12:50:59<28:05:17, 8.35s/step, loss=1.598, acc=0.651, acc_len=1.52, thru=13.5, I=17.8, T=24.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:13,671] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:18,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:21:19,812] loop.py:380 INFO TIMING step=5915: step=3.857s data=1.693s compute=3.822s [fwd=2.675s bwd=1.133s opt=0.003s] dispatch=2.378s [2026-10-19 02:21:19,813] loop.py:45 INFO TRAIN_STEP step=5915 loss=1.388427 Training: 33%|███▎ | 5915/18018 [12:51:05<25:57:32, 7.72s/step, loss=1.388, acc=0.689, acc_len=1.65, thru=16.4, I=21.3, T=33.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:19,884] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:25,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:21:25,863] loop.py:45 INFO TRAIN_STEP step=5916 loss=1.539608 Training: 33%|███▎ | 5916/18018 [12:51:11<24:16:19, 7.22s/step, loss=1.540, acc=0.664, acc_len=1.56, thru=20.1, I=25.2, T=31.0, wait=1.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:21:25,859] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5915: forward=2913.4ms backward=1167.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:26,289] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:32,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:21:34,183] loop.py:45 INFO TRAIN_STEP step=5917 loss=1.528643 Training: 33%|███▎ | 5917/18018 [12:51:20<25:22:45, 7.55s/step, loss=1.529, acc=0.669, acc_len=1.58, thru=14.8, I=19.2, T=31.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:34,360] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:39,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:21:40,488] loop.py:45 INFO TRAIN_STEP step=5918 loss=1.487817 Training: 33%|███▎ | 5918/18018 [12:51:26<24:07:10, 7.18s/step, loss=1.488, acc=0.673, acc_len=1.59, thru=18.9, I=26.6, T=33.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:40,793] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:45,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:21:46,558] loop.py:45 INFO TRAIN_STEP step=5919 loss=1.464394 Training: 33%|███▎ | 5919/18018 [12:51:32<23:00:08, 6.84s/step, loss=1.464, acc=0.675, acc_len=1.60, thru=19.7, I=25.9, T=33.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:47,049] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:53,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:21:54,804] loop.py:380 INFO TIMING step=5920: step=4.926s data=2.593s compute=4.888s [fwd=3.459s bwd=1.414s opt=0.003s] dispatch=3.311s [2026-10-19 02:21:54,805] loop.py:45 INFO TRAIN_STEP step=5920 loss=1.446754 Training: 33%|███▎ | 5920/18018 [12:51:40<24:24:56, 7.27s/step, loss=1.447, acc=0.674, acc_len=1.60, thru=14.8, I=18.2, T=26.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:55,109] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:21:59,753] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:22:01,272] loop.py:45 INFO TRAIN_STEP step=5921 loss=1.551671 Training: 33%|███▎ | 5921/18018 [12:51:47<23:36:34, 7.03s/step, loss=1.552, acc=0.657, acc_len=1.55, thru=18.6, I=22.8, T=30.0, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:22:01,267] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5920: forward=3155.3ms backward=1065.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:22:01,766] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:22:08,439] loop.py:45 INFO TRAIN_STEP step=5922 loss=1.607027 Training: 33%|███▎ | 5922/18018 [12:51:54<23:45:14, 7.07s/step, loss=1.607, acc=0.652, acc_len=1.53, thru=14.4, I=18.2, T=32.6, wait=3.2s, pool=44, epoch=1/2][2026-10-19 02:22:16,957] loop.py:45 INFO TRAIN_STEP step=5923 loss=1.442830 Training: 33%|███▎ | 5923/18018 [12:52:03<25:12:21, 7.50s/step, loss=1.443, acc=0.681, acc_len=1.63, thru=16.2, I=15.0, T=32.8, wait=4.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:22:23,494] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:22:25,314] loop.py:45 INFO TRAIN_STEP step=5924 loss=1.451199 Training: 33%|███▎ | 5924/18018 [12:52:11<26:03:59, 7.76s/step, loss=1.451, acc=0.680, acc_len=1.62, thru=17.1, I=19.5, T=31.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:22:25,509] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:22:30,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:22:31,725] loop.py:380 INFO TIMING step=5925: step=3.949s data=1.788s compute=3.915s [fwd=2.742s bwd=1.158s opt=0.003s] dispatch=2.452s [2026-10-19 02:22:31,726] loop.py:45 INFO TRAIN_STEP step=5925 loss=1.562537 Training: 33%|███▎ | 5925/18018 [12:52:17<24:42:22, 7.35s/step, loss=1.563, acc=0.660, acc_len=1.55, thru=18.0, I=24.4, T=32.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:22:31,819] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:22:37,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:22:38,012] loop.py:45 INFO TRAIN_STEP step=5926 loss=1.375253 Training: 33%|███▎ | 5926/18018 [12:52:24<23:37:39, 7.03s/step, loss=1.375, acc=0.689, acc_len=1.65, thru=18.9, I=22.2, T=34.1, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:22:38,007] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5925: forward=2648.0ms backward=1068.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:22:38,168] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:22:44,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:22:45,060] loop.py:45 INFO TRAIN_STEP step=5927 loss=1.515755 Training: 33%|███▎ | 5927/18018 [12:52:31<23:38:22, 7.04s/step, loss=1.516, acc=0.669, acc_len=1.58, thru=18.3, I=21.2, T=30.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:22:45,230] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:22:51,702] loop.py:45 INFO TRAIN_STEP step=5928 loss=1.449528 Training: 33%|███▎ | 5928/18018 [12:52:37<23:14:14, 6.92s/step, loss=1.450, acc=0.679, acc_len=1.62, thru=15.8, I=18.9, T=35.1, wait=3.0s, pool=44, epoch=1/2][2026-10-19 02:23:00,823] loop.py:45 INFO TRAIN_STEP step=5929 loss=1.542722 Training: 33%|███▎ | 5929/18018 [12:52:46<25:27:07, 7.58s/step, loss=1.543, acc=0.663, acc_len=1.56, thru=14.9, I=14.1, T=30.2, wait=4.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:05,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:23:06,888] loop.py:380 INFO TIMING step=5930: step=3.723s data=1.606s compute=3.696s [fwd=2.612s bwd=1.070s opt=0.003s] dispatch=2.333s [2026-10-19 02:23:06,888] loop.py:45 INFO TRAIN_STEP step=5930 loss=1.391807 Training: 33%|███▎ | 5930/18018 [12:52:52<23:55:27, 7.13s/step, loss=1.392, acc=0.687, acc_len=1.64, thru=18.9, I=22.1, T=34.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:07,370] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:23:14,267] loop.py:45 INFO TRAIN_STEP step=5931 loss=1.339487 Training: 33%|███▎ | 5931/18018 [12:53:00<24:10:42, 7.20s/step, loss=1.339, acc=0.698, acc_len=1.69, thru=13.6, I=19.1, T=34.2, wait=3.6s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:23:14,262] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5930: forward=2638.1ms backward=1066.3ms [2026-10-19 02:23:23,270] loop.py:45 INFO TRAIN_STEP step=5932 loss=1.607179 Training: 33%|███▎ | 5932/18018 [12:53:09<25:59:29, 7.74s/step, loss=1.607, acc=0.655, acc_len=1.54, thru=14.6, I=16.0, T=30.0, wait=4.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:34,753] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:23:36,747] loop.py:45 INFO TRAIN_STEP step=5933 loss=1.397631 Training: 33%|███▎ | 5933/18018 [12:53:22<31:46:00, 9.46s/step, loss=1.398, acc=0.686, acc_len=1.64, thru=13.6, I=12.1, T=29.2, wait=9.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:36,780] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:42,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:23:43,270] loop.py:45 INFO TRAIN_STEP step=5934 loss=1.450621 Training: 33%|███▎ | 5934/18018 [12:53:29<28:48:12, 8.58s/step, loss=1.451, acc=0.675, acc_len=1.61, thru=17.4, I=23.9, T=32.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:43,440] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:49,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:23:50,104] loop.py:380 INFO TIMING step=5935: step=3.760s data=1.659s compute=3.725s [fwd=2.663s bwd=1.048s opt=0.003s] dispatch=3.064s [2026-10-19 02:23:50,104] loop.py:45 INFO TRAIN_STEP step=5935 loss=1.513785 Training: 33%|███▎ | 5935/18018 [12:53:36<27:02:22, 8.06s/step, loss=1.514, acc=0.665, acc_len=1.57, thru=17.6, I=20.6, T=34.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:50,503] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:56,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:23:57,828] loop.py:45 INFO TRAIN_STEP step=5936 loss=1.478844 Training: 33%|███▎ | 5936/18018 [12:53:43<26:42:18, 7.96s/step, loss=1.479, acc=0.676, acc_len=1.61, thru=16.2, I=21.1, T=29.8, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:23:57,823] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5935: forward=2914.9ms backward=1333.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:23:58,044] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:24:04,866] loop.py:45 INFO TRAIN_STEP step=5937 loss=1.598585 Training: 33%|███▎ | 5937/18018 [12:53:50<25:46:27, 7.68s/step, loss=1.599, acc=0.649, acc_len=1.52, thru=15.8, I=18.9, T=33.8, wait=3.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:12,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:24:14,039] loop.py:45 INFO TRAIN_STEP step=5938 loss=1.484650 Training: 33%|███▎ | 5938/18018 [12:54:00<27:16:39, 8.13s/step, loss=1.485, acc=0.667, acc_len=1.57, thru=15.7, I=16.5, T=28.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:14,235] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:21,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:24:21,742] loop.py:45 INFO TRAIN_STEP step=5939 loss=1.451786 Training: 33%|███▎ | 5939/18018 [12:54:07<26:50:39, 8.00s/step, loss=1.452, acc=0.675, acc_len=1.61, thru=15.0, I=16.6, T=31.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:21,990] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:26,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:24:28,329] loop.py:380 INFO TIMING step=5940: step=4.175s data=1.692s compute=4.123s [fwd=2.796s bwd=1.312s opt=0.003s] dispatch=2.403s [2026-10-19 02:24:28,329] loop.py:45 INFO TRAIN_STEP step=5940 loss=1.435191 Training: 33%|███▎ | 5940/18018 [12:54:14<25:25:17, 7.58s/step, loss=1.435, acc=0.682, acc_len=1.63, thru=19.4, I=23.1, T=30.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:28,586] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:34,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:24:34,836] loop.py:45 INFO TRAIN_STEP step=5941 loss=1.384675 Training: 33%|███▎ | 5941/18018 [12:54:20<24:20:27, 7.26s/step, loss=1.385, acc=0.688, acc_len=1.64, thru=16.9, I=23.1, T=35.2, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:24:34,831] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5940: forward=2574.2ms backward=1024.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:35,108] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:41,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:24:41,760] loop.py:45 INFO TRAIN_STEP step=5942 loss=1.427176 Training: 33%|███▎ | 5942/18018 [12:54:27<24:00:21, 7.16s/step, loss=1.427, acc=0.682, acc_len=1.62, thru=18.4, I=18.8, T=31.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:41,811] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:48,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:24:51,099] loop.py:45 INFO TRAIN_STEP step=5943 loss=1.543089 Training: 33%|███▎ | 5943/18018 [12:54:37<26:11:55, 7.81s/step, loss=1.543, acc=0.663, acc_len=1.57, thru=13.5, I=17.9, T=26.2, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:51,129] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:55,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:24:57,271] loop.py:45 INFO TRAIN_STEP step=5944 loss=1.390472 Training: 33%|███▎ | 5944/18018 [12:54:43<24:32:58, 7.32s/step, loss=1.390, acc=0.688, acc_len=1.64, thru=16.1, I=24.8, T=30.9, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:24:57,504] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:25:04,019] loop.py:380 INFO TIMING step=5945: step=4.022s data=1.753s compute=3.981s [fwd=2.759s bwd=1.207s opt=0.003s] dispatch=2.716s [2026-10-19 02:25:04,019] loop.py:45 INFO TRAIN_STEP step=5945 loss=1.515483 Training: 33%|███▎ | 5945/18018 [12:54:50<23:58:09, 7.15s/step, loss=1.515, acc=0.666, acc_len=1.57, thru=15.8, I=18.6, T=31.8, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:10,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:25:11,748] loop.py:45 INFO TRAIN_STEP step=5946 loss=1.438203 Training: 33%|███▎ | 5946/18018 [12:54:57<24:33:21, 7.32s/step, loss=1.438, acc=0.679, acc_len=1.61, thru=16.9, I=18.4, T=30.3, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:25:11,743] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5945: forward=2925.7ms backward=1242.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:12,058] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:19,638] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:25:20,025] loop.py:45 INFO TRAIN_STEP step=5947 loss=1.495284 Training: 33%|███▎ | 5947/18018 [12:55:06<25:30:40, 7.61s/step, loss=1.495, acc=0.670, acc_len=1.59, thru=13.2, I=16.3, T=33.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:20,155] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:25,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:25:26,411] loop.py:45 INFO TRAIN_STEP step=5948 loss=1.414541 Training: 33%|███▎ | 5948/18018 [12:55:12<24:16:52, 7.24s/step, loss=1.415, acc=0.682, acc_len=1.63, thru=18.8, I=21.7, T=31.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:26,679] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:35,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:25:36,711] loop.py:45 INFO TRAIN_STEP step=5949 loss=1.516247 Training: 33%|███▎ | 5949/18018 [12:55:22<27:21:12, 8.16s/step, loss=1.516, acc=0.667, acc_len=1.57, thru=12.6, I=14.4, T=25.6, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:36,739] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:42,926] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:25:43,936] loop.py:380 INFO TIMING step=5950: step=4.390s data=1.997s compute=4.366s [fwd=3.017s bwd=1.335s opt=0.003s] dispatch=2.826s [2026-10-19 02:25:43,937] loop.py:45 INFO TRAIN_STEP step=5950 loss=1.513561 Training: 33%|███▎ | 5950/18018 [12:55:30<26:24:39, 7.88s/step, loss=1.514, acc=0.664, acc_len=1.57, thru=15.6, I=17.6, T=29.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:44,439] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:51,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:25:51,943] loop.py:45 INFO TRAIN_STEP step=5951 loss=1.426577 Training: 33%|███▎ | 5951/18018 [12:55:38<26:32:15, 7.92s/step, loss=1.427, acc=0.676, acc_len=1.61, thru=17.0, I=17.9, T=30.6, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:25:51,938] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5950: forward=3035.2ms backward=1099.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:52,360] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:58,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:25:59,560] loop.py:45 INFO TRAIN_STEP step=5952 loss=1.462345 Training: 33%|███▎ | 5952/18018 [12:55:45<26:14:10, 7.83s/step, loss=1.462, acc=0.672, acc_len=1.59, thru=16.9, I=19.4, T=30.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:25:59,681] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:06,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:26:08,526] loop.py:45 INFO TRAIN_STEP step=5953 loss=1.492721 Training: 33%|███▎ | 5953/18018 [12:55:54<27:22:36, 8.17s/step, loss=1.493, acc=0.668, acc_len=1.58, thru=14.5, I=16.5, T=30.5, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:08,718] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:14,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:26:15,254] loop.py:45 INFO TRAIN_STEP step=5954 loss=1.528208 Training: 33%|███▎ | 5954/18018 [12:56:01<25:55:33, 7.74s/step, loss=1.528, acc=0.666, acc_len=1.56, thru=17.7, I=23.9, T=31.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:15,735] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:22,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:26:22,560] loop.py:380 INFO TIMING step=5955: step=4.158s data=1.814s compute=4.124s [fwd=2.843s bwd=1.266s opt=0.003s] dispatch=3.139s [2026-10-19 02:26:22,560] loop.py:45 INFO TRAIN_STEP step=5955 loss=1.515385 Training: 33%|███▎ | 5955/18018 [12:56:08<25:29:27, 7.61s/step, loss=1.515, acc=0.666, acc_len=1.57, thru=17.8, I=18.4, T=30.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:22,853] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:27,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:26:29,091] loop.py:45 INFO TRAIN_STEP step=5956 loss=1.547107 Training: 33%|███▎ | 5956/18018 [12:56:15<24:24:25, 7.28s/step, loss=1.547, acc=0.660, acc_len=1.55, thru=18.3, I=20.3, T=31.0, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:26:29,086] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5955: forward=2895.2ms backward=1181.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:29,471] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:37,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:26:38,005] loop.py:45 INFO TRAIN_STEP step=5957 loss=1.479652 Training: 33%|███▎ | 5957/18018 [12:56:24<26:02:33, 7.77s/step, loss=1.480, acc=0.672, acc_len=1.59, thru=15.0, I=15.3, T=32.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:38,055] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:43,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:26:44,192] loop.py:45 INFO TRAIN_STEP step=5958 loss=1.669781 Training: 33%|███▎ | 5958/18018 [12:56:30<24:26:50, 7.30s/step, loss=1.670, acc=0.640, acc_len=1.49, thru=19.7, I=21.9, T=34.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:44,348] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:50,854] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:26:51,257] loop.py:45 INFO TRAIN_STEP step=5959 loss=1.545028 Training: 33%|███▎ | 5959/18018 [12:56:37<24:12:43, 7.23s/step, loss=1.545, acc=0.657, acc_len=1.55, thru=18.2, I=20.1, T=34.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:51,368] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:58,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:26:59,086] loop.py:380 INFO TIMING step=5960: step=4.275s data=2.019s compute=4.223s [fwd=2.958s bwd=1.250s opt=0.003s] dispatch=3.545s [2026-10-19 02:26:59,086] loop.py:45 INFO TRAIN_STEP step=5960 loss=1.452372 Training: 33%|███▎ | 5960/18018 [12:56:45<24:48:42, 7.41s/step, loss=1.452, acc=0.675, acc_len=1.60, thru=17.7, I=17.9, T=29.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:26:59,309] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:03,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:27:06,882] loop.py:45 INFO TRAIN_STEP step=5961 loss=1.537683 Training: 33%|███▎ | 5961/18018 [12:56:52<25:12:05, 7.52s/step, loss=1.538, acc=0.664, acc_len=1.57, thru=15.8, I=21.5, T=25.0, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:27:06,877] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5960: forward=3661.2ms backward=1405.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:06,991] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:13,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:27:14,018] loop.py:45 INFO TRAIN_STEP step=5962 loss=1.504094 Training: 33%|███▎ | 5962/18018 [12:57:00<24:48:36, 7.41s/step, loss=1.504, acc=0.669, acc_len=1.59, thru=18.2, I=19.5, T=32.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:14,275] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:27:20,796] loop.py:45 INFO TRAIN_STEP step=5963 loss=1.524530 Training: 33%|███▎ | 5963/18018 [12:57:06<24:10:22, 7.22s/step, loss=1.525, acc=0.661, acc_len=1.56, thru=18.0, I=19.6, T=29.2, wait=2.4s, pool=56, epoch=1/2][2026-10-19 02:27:27,193] loop.py:45 INFO TRAIN_STEP step=5964 loss=1.429935 Training: 33%|███▎ | 5964/18018 [12:57:13<23:20:48, 6.97s/step, loss=1.430, acc=0.678, acc_len=1.61, thru=16.2, I=18.1, T=32.5, wait=2.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:35,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:27:36,217] loop.py:380 INFO TIMING step=5965: step=3.864s data=1.640s compute=3.831s [fwd=2.699s bwd=1.116s opt=0.003s] dispatch=5.149s [2026-10-19 02:27:36,217] loop.py:45 INFO TRAIN_STEP step=5965 loss=1.446368 Training: 33%|███▎ | 5965/18018 [12:57:22<25:24:16, 7.59s/step, loss=1.446, acc=0.678, acc_len=1.61, thru=15.7, I=15.4, T=33.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:36,551] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:44,441] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:27:45,641] loop.py:45 INFO TRAIN_STEP step=5966 loss=1.457791 Training: 33%|███▎ | 5966/18018 [12:57:31<27:14:47, 8.14s/step, loss=1.458, acc=0.676, acc_len=1.61, thru=14.1, I=15.0, T=30.0, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:27:45,636] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5965: forward=3069.8ms backward=1147.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:45,951] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:51,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:27:52,060] loop.py:45 INFO TRAIN_STEP step=5967 loss=1.389304 Training: 33%|███▎ | 5967/18018 [12:57:38<25:31:04, 7.62s/step, loss=1.389, acc=0.692, acc_len=1.66, thru=18.1, I=21.5, T=30.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:27:52,339] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:27:59,072] loop.py:45 INFO TRAIN_STEP step=5968 loss=1.542746 Training: 33%|███▎ | 5968/18018 [12:57:45<24:53:58, 7.44s/step, loss=1.543, acc=0.658, acc_len=1.56, thru=16.4, I=17.9, T=30.7, wait=2.8s, pool=60, epoch=1/2][2026-10-19 02:28:06,559] loop.py:45 INFO TRAIN_STEP step=5969 loss=1.409639 Training: 33%|███▎ | 5969/18018 [12:57:52<24:56:59, 7.45s/step, loss=1.410, acc=0.683, acc_len=1.62, thru=14.5, I=13.8, T=31.3, wait=3.4s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:14,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:28:16,709] loop.py:380 INFO TIMING step=5970: step=4.655s data=2.148s compute=4.628s [fwd=3.354s bwd=1.259s opt=0.003s] dispatch=5.483s [2026-10-19 02:28:16,710] loop.py:45 INFO TRAIN_STEP step=5970 loss=1.507286 Training: 33%|███▎ | 5970/18018 [12:58:02<27:39:11, 8.26s/step, loss=1.507, acc=0.673, acc_len=1.60, thru=15.7, I=17.7, T=27.5, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:16,864] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:20,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:28:22,629] loop.py:45 INFO TRAIN_STEP step=5971 loss=1.488299 Training: 33%|███▎ | 5971/18018 [12:58:08<25:17:50, 7.56s/step, loss=1.488, acc=0.670, acc_len=1.58, thru=18.1, I=29.2, T=32.7, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:28:22,624] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5970: forward=2714.0ms backward=1157.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:22,982] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:28,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:28:30,091] loop.py:45 INFO TRAIN_STEP step=5972 loss=1.452024 Training: 33%|███▎ | 5972/18018 [12:58:16<25:11:48, 7.53s/step, loss=1.452, acc=0.673, acc_len=1.59, thru=14.1, I=23.9, T=27.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:30,264] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:35,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:28:36,587] loop.py:45 INFO TRAIN_STEP step=5973 loss=1.506534 Training: 33%|███▎ | 5973/18018 [12:58:22<24:09:22, 7.22s/step, loss=1.507, acc=0.667, acc_len=1.58, thru=16.6, I=23.2, T=32.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:36,989] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:28:43,897] loop.py:45 INFO TRAIN_STEP step=5974 loss=1.569167 Training: 33%|███▎ | 5974/18018 [12:58:29<24:14:50, 7.25s/step, loss=1.569, acc=0.659, acc_len=1.55, thru=14.1, I=19.9, T=28.5, wait=2.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:50,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:28:52,460] loop.py:380 INFO TIMING step=5975: step=4.378s data=1.925s compute=4.341s [fwd=3.109s bwd=1.217s opt=0.003s] dispatch=4.174s [2026-10-19 02:28:52,460] loop.py:45 INFO TRAIN_STEP step=5975 loss=1.561063 Training: 33%|███▎ | 5975/18018 [12:58:38<25:33:52, 7.64s/step, loss=1.561, acc=0.659, acc_len=1.55, thru=16.6, I=15.9, T=29.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:52,967] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:57,401] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:28:59,051] loop.py:45 INFO TRAIN_STEP step=5976 loss=1.633398 Training: 33%|███▎ | 5976/18018 [12:58:45<24:30:22, 7.33s/step, loss=1.633, acc=0.645, acc_len=1.50, thru=18.1, I=26.0, T=32.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:28:59,046] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5975: forward=2735.8ms backward=1130.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:28:59,417] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:29:04,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:29:05,308] loop.py:45 INFO TRAIN_STEP step=5977 loss=1.497407 Training: 33%|███▎ | 5977/18018 [12:58:51<23:25:53, 7.01s/step, loss=1.497, acc=0.666, acc_len=1.57, thru=18.9, I=25.4, T=38.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:29:05,443] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:29:10,637] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:29:12,637] loop.py:45 INFO TRAIN_STEP step=5978 loss=1.340075 Training: 33%|███▎ | 5978/18018 [12:58:58<23:45:20, 7.10s/step, loss=1.340, acc=0.697, acc_len=1.67, thru=17.8, I=23.6, T=29.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:29:12,670] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 02:29:21,991] loop.py:45 INFO TRAIN_STEP step=5979 loss=1.367139 Training: 33%|███▎ | 5979/18018 [12:59:08<26:00:41, 7.78s/step, loss=1.367, acc=0.693, acc_len=1.66, thru=13.3, I=12.3, T=31.1, wait=5.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:29:27,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:29:29,739] loop.py:380 INFO TIMING step=5980: step=5.262s data=2.576s compute=5.227s [fwd=3.714s bwd=1.497s opt=0.003s] dispatch=2.476s [2026-10-19 02:29:29,740] loop.py:45 INFO TRAIN_STEP step=5980 loss=1.456806 Training: 33%|███▎ | 5980/18018 [12:59:15<25:58:48, 7.77s/step, loss=1.457, acc=0.677, acc_len=1.61, thru=17.4, I=19.7, T=24.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:29:29,811] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 [2026-10-19 02:29:38,290] loop.py:45 INFO TRAIN_STEP step=5981 loss=1.571071 Training: 33%|███▎ | 5981/18018 [12:59:24<26:45:40, 8.00s/step, loss=1.571, acc=0.658, acc_len=1.55, thru=13.2, I=14.5, T=31.7, wait=4.5s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:29:38,285] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5980: forward=2846.5ms backward=1146.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:29:49,722] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:29:49,818] loop.py:45 INFO TRAIN_STEP step=5982 loss=1.491369 Training: 33%|███▎ | 5982/18018 [12:59:35<30:17:28, 9.06s/step, loss=1.491, acc=0.669, acc_len=1.59, thru=14.2, I=11.6, T=30.6, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:29:50,225] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:29:56,256] loop.py:45 INFO TRAIN_STEP step=5983 loss=1.563992 Training: 33%|███▎ | 5983/18018 [12:59:42<27:39:43, 8.27s/step, loss=1.564, acc=0.657, acc_len=1.54, thru=18.1, I=20.4, T=31.9, wait=2.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:05,666] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:30:06,670] loop.py:45 INFO TRAIN_STEP step=5984 loss=1.484649 Training: 33%|███▎ | 5984/18018 [12:59:52<29:48:11, 8.92s/step, loss=1.485, acc=0.669, acc_len=1.58, thru=13.7, I=13.4, T=30.3, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:07,175] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:12,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:30:13,371] loop.py:380 INFO TIMING step=5985: step=3.927s data=1.783s compute=3.894s [fwd=2.742s bwd=1.138s opt=0.003s] dispatch=2.764s [2026-10-19 02:30:13,371] loop.py:45 INFO TRAIN_STEP step=5985 loss=1.513106 Training: 33%|███▎ | 5985/18018 [12:59:59<27:34:53, 8.25s/step, loss=1.513, acc=0.665, acc_len=1.57, thru=18.3, I=20.7, T=32.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:13,635] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:30:20,681] loop.py:45 INFO TRAIN_STEP step=5986 loss=1.454455 Training: 33%|███▎ | 5986/18018 [13:00:06<26:38:05, 7.97s/step, loss=1.454, acc=0.678, acc_len=1.61, thru=15.3, I=21.1, T=31.3, wait=3.2s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:30:20,677] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5985: forward=2832.9ms backward=1211.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:29,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:30:31,966] loop.py:45 INFO TRAIN_STEP step=5987 loss=1.550400 Training: 33%|███▎ | 5987/18018 [13:00:18<29:57:17, 8.96s/step, loss=1.550, acc=0.661, acc_len=1.55, thru=12.0, I=14.7, T=26.7, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:32,147] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:37,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:30:39,907] loop.py:45 INFO TRAIN_STEP step=5988 loss=1.578154 Training: 33%|███▎ | 5988/18018 [13:00:25<28:55:38, 8.66s/step, loss=1.578, acc=0.651, acc_len=1.53, thru=16.5, I=18.0, T=27.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:40,342] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:46,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:30:46,775] loop.py:45 INFO TRAIN_STEP step=5989 loss=1.398757 Training: 33%|███▎ | 5989/18018 [13:00:32<27:07:57, 8.12s/step, loss=1.399, acc=0.682, acc_len=1.63, thru=15.0, I=22.9, T=29.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:30:47,130] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:30:54,160] loop.py:380 INFO TIMING step=5990: step=3.953s data=1.725s compute=3.917s [fwd=2.770s bwd=1.132s opt=0.003s] dispatch=3.423s [2026-10-19 02:30:54,160] loop.py:45 INFO TRAIN_STEP step=5990 loss=1.434291 Training: 33%|███▎ | 5990/18018 [13:00:40<26:23:42, 7.90s/step, loss=1.434, acc=0.676, acc_len=1.60, thru=16.0, I=16.1, T=32.4, wait=3.4s, pool=60, epoch=1/2][2026-10-19 02:31:02,185] loop.py:45 INFO TRAIN_STEP step=5991 loss=1.478056 Training: 33%|███▎ | 5991/18018 [13:00:48<26:30:53, 7.94s/step, loss=1.478, acc=0.677, acc_len=1.61, thru=15.3, I=14.8, T=28.8, wait=3.6s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:31:02,180] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5990: forward=3120.1ms backward=1285.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:09,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:31:09,692] loop.py:45 INFO TRAIN_STEP step=5992 loss=1.524491 Training: 33%|███▎ | 5992/18018 [13:00:55<26:05:08, 7.81s/step, loss=1.524, acc=0.666, acc_len=1.57, thru=15.6, I=15.8, T=31.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:09,868] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:17,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:31:19,267] loop.py:45 INFO TRAIN_STEP step=5993 loss=1.544223 Training: 33%|███▎ | 5993/18018 [13:01:05<27:51:09, 8.34s/step, loss=1.544, acc=0.661, acc_len=1.56, thru=13.3, I=16.4, T=27.7, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:19,594] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:27,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:31:28,736] loop.py:45 INFO TRAIN_STEP step=5994 loss=1.576763 Training: 33%|███▎ | 5994/18018 [13:01:14<28:59:03, 8.68s/step, loss=1.577, acc=0.661, acc_len=1.56, thru=14.0, I=15.7, T=25.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:28,955] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:33,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:31:35,057] loop.py:380 INFO TIMING step=5995: step=3.993s data=1.806s compute=3.958s [fwd=2.836s bwd=1.108s opt=0.003s] dispatch=2.317s [2026-10-19 02:31:35,057] loop.py:45 INFO TRAIN_STEP step=5995 loss=1.424650 Training: 33%|███▎ | 5995/18018 [13:01:21<26:37:12, 7.97s/step, loss=1.425, acc=0.681, acc_len=1.63, thru=18.8, I=20.5, T=32.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:35,293] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:31:42,391] loop.py:45 INFO TRAIN_STEP step=5996 loss=1.422450 Training: 33%|███▎ | 5996/18018 [13:01:28<25:58:54, 7.78s/step, loss=1.422, acc=0.681, acc_len=1.62, thru=16.5, I=19.1, T=27.2, wait=2.6s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:31:42,386] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=5995: forward=3251.2ms backward=1405.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:47,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:31:50,393] loop.py:45 INFO TRAIN_STEP step=5997 loss=1.460171 Training: 33%|███▎ | 5997/18018 [13:01:36<26:11:53, 7.85s/step, loss=1.460, acc=0.674, acc_len=1.61, thru=16.2, I=19.3, T=25.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:50,714] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:57,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:31:57,509] loop.py:45 INFO TRAIN_STEP step=5998 loss=1.514365 Training: 33%|███▎ | 5998/18018 [13:01:43<25:27:46, 7.63s/step, loss=1.514, acc=0.666, acc_len=1.57, thru=13.6, I=22.0, T=32.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:31:57,540] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-19 02:32:06,553] loop.py:45 INFO TRAIN_STEP step=5999 loss=1.507887 Training: 33%|███▎ | 5999/18018 [13:01:52<26:52:53, 8.05s/step, loss=1.508, acc=0.662, acc_len=1.56, thru=13.4, I=13.6, T=31.1, wait=4.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:13,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:32:15,487] loop.py:380 INFO TIMING step=6000: step=4.483s data=2.051s compute=4.453s [fwd=3.163s bwd=1.275s opt=0.003s] dispatch=4.444s [2026-10-19 02:32:15,487] loop.py:45 INFO TRAIN_STEP step=6000 loss=1.541616 Training: 33%|███▎ | 6000/18018 [13:02:01<27:45:47, 8.32s/step, loss=1.542, acc=0.658, acc_len=1.55, thru=16.1, I=17.6, T=28.6, wait=4.4s, pool=64, epoch=1/2][2026-10-19 02:32:15,489] loop.py:419 INFO Saving checkpoint at step 6000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:18,500] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 02:32:18,893] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0006001 (TrainerActor pid=9310) [2026-10-19 02:32:18,900] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0005251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:20,024] inference_manager.py:408 INFO Pool capacity available after 6.6s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:27,091] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 02:32:18,892] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0006001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 02:32:19,987] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0005251 [2026-10-19 02:32:27,876] loop.py:45 INFO TRAIN_STEP step=6001 loss=1.582478 Training: 33%|███▎ | 6001/18018 [13:02:13<31:50:31, 9.54s/step, loss=1.582, acc=0.652, acc_len=1.53, thru=16.3, I=19.4, T=27.6, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:32:27,871] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6000: forward=3232.8ms backward=1368.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:28,098] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:36,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:32:36,970] loop.py:45 INFO TRAIN_STEP step=6002 loss=1.533184 Training: 33%|███▎ | 6002/18018 [13:02:23<31:23:32, 9.41s/step, loss=1.533, acc=0.665, acc_len=1.57, thru=13.6, I=13.9, T=28.9, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:37,048] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:42,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:32:43,697] loop.py:45 INFO TRAIN_STEP step=6003 loss=1.500780 Training: 33%|███▎ | 6003/18018 [13:02:29<28:42:35, 8.60s/step, loss=1.501, acc=0.670, acc_len=1.59, thru=17.2, I=20.3, T=28.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:43,890] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:32:50,714] loop.py:45 INFO TRAIN_STEP step=6004 loss=1.638461 Training: 33%|███▎ | 6004/18018 [13:02:36<27:07:07, 8.13s/step, loss=1.638, acc=0.648, acc_len=1.51, thru=14.6, I=16.3, T=30.4, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:32:59,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:32:59,803] loop.py:380 INFO TIMING step=6005: step=4.052s data=1.835s compute=4.005s [fwd=2.833s bwd=1.158s opt=0.003s] dispatch=5.029s [2026-10-19 02:32:59,804] loop.py:45 INFO TRAIN_STEP step=6005 loss=1.497298 Training: 33%|███▎ | 6005/18018 [13:02:45<28:04:55, 8.42s/step, loss=1.497, acc=0.672, acc_len=1.60, thru=15.9, I=15.3, T=31.6, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:33:00,011] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:33:04,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:33:06,050] loop.py:45 INFO TRAIN_STEP step=6006 loss=1.312577 Training: 33%|███▎ | 6006/18018 [13:02:52<25:54:22, 7.76s/step, loss=1.313, acc=0.701, acc_len=1.69, thru=19.0, I=22.7, T=31.7, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:33:06,046] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6005: forward=2795.0ms backward=1179.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:33:06,527] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:33:11,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:33:13,477] loop.py:45 INFO TRAIN_STEP step=6007 loss=1.489996 Training: 33%|███▎ | 6007/18018 [13:02:59<25:34:09, 7.66s/step, loss=1.490, acc=0.667, acc_len=1.57, thru=16.4, I=21.9, T=28.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:33:13,591] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:33:19,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:33:20,775] loop.py:45 INFO TRAIN_STEP step=6008 loss=1.564979 Training: 33%|███▎ | 6008/18018 [13:03:06<25:11:54, 7.55s/step, loss=1.565, acc=0.655, acc_len=1.54, thru=14.8, I=19.2, T=34.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:33:21,210] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:33:28,483] loop.py:45 INFO TRAIN_STEP step=6009 loss=1.500971 Training: 33%|███▎ | 6009/18018 [13:03:14<25:21:14, 7.60s/step, loss=1.501, acc=0.670, acc_len=1.58, thru=14.2, I=22.6, T=30.5, wait=3.5s, pool=48, epoch=1/2][2026-10-19 02:33:37,347] loop.py:380 INFO TIMING step=6010: step=4.028s data=1.825s compute=3.997s [fwd=2.881s bwd=1.102s opt=0.003s] dispatch=4.825s [2026-10-19 02:33:37,347] loop.py:45 INFO TRAIN_STEP step=6010 loss=1.360742 Training: 33%|███▎ | 6010/18018 [13:03:23<26:36:51, 7.98s/step, loss=1.361, acc=0.693, acc_len=1.66, thru=13.8, I=12.2, T=31.8, wait=4.8s, pool=48, epoch=1/2][2026-10-19 02:33:47,558] loop.py:45 INFO TRAIN_STEP step=6011 loss=1.520638 Training: 33%|███▎ | 6011/18018 [13:03:33<28:50:44, 8.65s/step, loss=1.521, acc=0.663, acc_len=1.57, thru=12.5, I=12.8, T=29.5, wait=5.9s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:33:47,552] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6010: forward=2988.0ms backward=1285.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:33:53,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:33:55,521] loop.py:45 INFO TRAIN_STEP step=6012 loss=1.653991 Training: 33%|███▎ | 6012/18018 [13:03:41<28:09:31, 8.44s/step, loss=1.654, acc=0.642, acc_len=1.50, thru=16.5, I=18.9, T=27.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:33:55,840] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:34:01,854] loop.py:45 INFO TRAIN_STEP step=6013 loss=1.557783 Training: 33%|███▎ | 6013/18018 [13:03:47<26:02:36, 7.81s/step, loss=1.558, acc=0.661, acc_len=1.55, thru=15.8, I=19.1, T=31.9, wait=2.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:09,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:34:10,340] loop.py:45 INFO TRAIN_STEP step=6014 loss=1.526510 Training: 33%|███▎ | 6014/18018 [13:03:56<26:43:03, 8.01s/step, loss=1.527, acc=0.667, acc_len=1.58, thru=17.2, I=18.2, T=30.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:10,790] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:14,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:34:17,209] loop.py:380 INFO TIMING step=6015: step=4.168s data=1.972s compute=4.138s [fwd=2.886s bwd=1.236s opt=0.003s] dispatch=2.693s [2026-10-19 02:34:17,209] loop.py:45 INFO TRAIN_STEP step=6015 loss=1.546987 Training: 33%|███▎ | 6015/18018 [13:04:03<25:34:22, 7.67s/step, loss=1.547, acc=0.665, acc_len=1.57, thru=16.5, I=26.0, T=30.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:17,605] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:22,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:34:24,344] loop.py:45 INFO TRAIN_STEP step=6016 loss=1.486930 Training: 33%|███▎ | 6016/18018 [13:04:10<25:02:02, 7.51s/step, loss=1.487, acc=0.669, acc_len=1.58, thru=14.1, I=23.2, T=31.8, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:34:24,339] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6015: forward=2847.6ms backward=1132.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:24,480] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:30,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:34:31,495] loop.py:45 INFO TRAIN_STEP step=6017 loss=1.575171 Training: 33%|███▎ | 6017/18018 [13:04:17<24:40:33, 7.40s/step, loss=1.575, acc=0.660, acc_len=1.56, thru=16.5, I=21.8, T=30.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:31,677] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:35,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:34:37,494] loop.py:45 INFO TRAIN_STEP step=6018 loss=1.642272 Training: 33%|███▎ | 6018/18018 [13:04:23<23:16:02, 6.98s/step, loss=1.642, acc=0.647, acc_len=1.51, thru=19.9, I=23.9, T=33.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:37,956] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:43,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:34:44,432] loop.py:45 INFO TRAIN_STEP step=6019 loss=1.565371 Training: 33%|███▎ | 6019/18018 [13:04:30<23:13:35, 6.97s/step, loss=1.565, acc=0.657, acc_len=1.54, thru=16.4, I=22.1, T=32.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:44,803] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:49,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:34:51,065] loop.py:380 INFO TIMING step=6020: step=4.393s data=2.256s compute=4.358s [fwd=3.217s bwd=1.126s opt=0.003s] dispatch=2.230s [2026-10-19 02:34:51,065] loop.py:45 INFO TRAIN_STEP step=6020 loss=1.469520 Training: 33%|███▎ | 6020/18018 [13:04:37<22:53:19, 6.87s/step, loss=1.470, acc=0.673, acc_len=1.59, thru=18.2, I=22.3, T=29.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:34:51,326] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:34:58,076] loop.py:45 INFO TRAIN_STEP step=6021 loss=1.336326 Training: 33%|███▎ | 6021/18018 [13:04:44<23:01:44, 6.91s/step, loss=1.336, acc=0.697, acc_len=1.68, thru=15.6, I=19.8, T=31.7, wait=3.0s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:34:58,071] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6020: forward=2843.3ms backward=1140.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:06,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:35:08,125] loop.py:45 INFO TRAIN_STEP step=6022 loss=1.470332 Training: 33%|███▎ | 6022/18018 [13:04:54<26:09:57, 7.85s/step, loss=1.470, acc=0.675, acc_len=1.60, thru=14.2, I=14.3, T=27.3, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:08,497] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:12,943] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:35:14,807] loop.py:45 INFO TRAIN_STEP step=6023 loss=1.622274 Training: 33%|███▎ | 6023/18018 [13:05:00<24:59:39, 7.50s/step, loss=1.622, acc=0.648, acc_len=1.51, thru=17.7, I=23.2, T=32.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:14,959] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:35:21,409] loop.py:45 INFO TRAIN_STEP step=6024 loss=1.491786 Training: 33%|███▎ | 6024/18018 [13:05:07<24:05:26, 7.23s/step, loss=1.492, acc=0.670, acc_len=1.59, thru=17.4, I=21.5, T=30.3, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:28,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:35:29,766] loop.py:380 INFO TIMING step=6025: step=4.424s data=2.042s compute=4.387s [fwd=3.102s bwd=1.269s opt=0.003s] dispatch=3.924s [2026-10-19 02:35:29,766] loop.py:45 INFO TRAIN_STEP step=6025 loss=1.384052 Training: 33%|███▎ | 6025/18018 [13:05:15<25:12:59, 7.57s/step, loss=1.384, acc=0.692, acc_len=1.66, thru=16.0, I=15.6, T=28.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:30,058] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:34,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:35:37,197] loop.py:45 INFO TRAIN_STEP step=6026 loss=1.544826 Training: 33%|███▎ | 6026/18018 [13:05:23<25:04:25, 7.53s/step, loss=1.545, acc=0.657, acc_len=1.54, thru=15.9, I=22.7, T=25.4, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:35:37,192] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6025: forward=3743.2ms backward=1231.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:37,500] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:45,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:35:45,821] loop.py:45 INFO TRAIN_STEP step=6027 loss=1.418412 Training: 33%|███▎ | 6027/18018 [13:05:31<26:10:12, 7.86s/step, loss=1.418, acc=0.683, acc_len=1.63, thru=15.4, I=16.7, T=24.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:46,026] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:52,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:35:54,275] loop.py:45 INFO TRAIN_STEP step=6028 loss=1.466576 Training: 33%|███▎ | 6028/18018 [13:05:40<26:45:44, 8.04s/step, loss=1.467, acc=0.673, acc_len=1.60, thru=14.6, I=15.1, T=29.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:35:54,444] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:01,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:36:02,004] loop.py:45 INFO TRAIN_STEP step=6029 loss=1.645425 Training: 33%|███▎ | 6029/18018 [13:05:48<26:27:15, 7.94s/step, loss=1.645, acc=0.646, acc_len=1.51, thru=14.6, I=17.5, T=33.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:02,243] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:08,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:36:09,259] loop.py:380 INFO TIMING step=6030: step=4.074s data=1.984s compute=4.042s [fwd=2.969s bwd=1.059s opt=0.003s] dispatch=3.172s [2026-10-19 02:36:09,259] loop.py:45 INFO TRAIN_STEP step=6030 loss=1.411351 Training: 33%|███▎ | 6030/18018 [13:05:55<25:45:50, 7.74s/step, loss=1.411, acc=0.682, acc_len=1.62, thru=16.4, I=20.8, T=31.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:09,549] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:14,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:36:16,118] loop.py:45 INFO TRAIN_STEP step=6031 loss=1.431599 Training: 33%|███▎ | 6031/18018 [13:06:02<24:53:02, 7.47s/step, loss=1.432, acc=0.675, acc_len=1.60, thru=16.6, I=22.7, T=35.6, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:36:16,113] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6030: forward=2550.2ms backward=1001.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:16,619] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:22,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:36:23,751] loop.py:45 INFO TRAIN_STEP step=6032 loss=1.436312 Training: 33%|███▎ | 6032/18018 [13:06:09<25:02:33, 7.52s/step, loss=1.436, acc=0.678, acc_len=1.61, thru=15.9, I=22.3, T=33.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:23,810] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:28,207] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:36:30,412] loop.py:45 INFO TRAIN_STEP step=6033 loss=1.423011 Training: 33%|███▎ | 6033/18018 [13:06:16<24:10:56, 7.26s/step, loss=1.423, acc=0.685, acc_len=1.63, thru=19.2, I=25.1, T=27.2, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:30,731] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:35,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:36:36,733] loop.py:45 INFO TRAIN_STEP step=6034 loss=1.497744 Training: 33%|███▎ | 6034/18018 [13:06:22<23:14:20, 6.98s/step, loss=1.498, acc=0.671, acc_len=1.59, thru=17.2, I=28.2, T=35.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:36,766] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:42,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:36:42,388] loop.py:380 INFO TIMING step=6035: step=3.712s data=1.591s compute=3.688s [fwd=2.544s bwd=1.131s opt=0.003s] dispatch=1.932s [2026-10-19 02:36:42,388] loop.py:45 INFO TRAIN_STEP step=6035 loss=1.374688 Training: 33%|███▎ | 6035/18018 [13:06:28<21:54:45, 6.58s/step, loss=1.375, acc=0.689, acc_len=1.65, thru=20.9, I=25.7, T=34.5, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:42,535] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:47,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:36:48,771] loop.py:45 INFO TRAIN_STEP step=6036 loss=1.451615 Training: 33%|███▎ | 6036/18018 [13:06:34<21:42:35, 6.52s/step, loss=1.452, acc=0.678, acc_len=1.61, thru=18.7, I=21.4, T=35.9, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:36:48,767] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6035: forward=2540.2ms backward=969.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:48,835] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:54,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:36:55,001] loop.py:45 INFO TRAIN_STEP step=6037 loss=1.463389 Training: 34%|███▎ | 6037/18018 [13:06:41<21:25:03, 6.44s/step, loss=1.463, acc=0.674, acc_len=1.61, thru=20.1, I=24.4, T=35.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:36:55,134] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:00,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:37:02,269] loop.py:45 INFO TRAIN_STEP step=6038 loss=1.565509 Training: 34%|███▎ | 6038/18018 [13:06:48<22:14:47, 6.69s/step, loss=1.566, acc=0.658, acc_len=1.55, thru=17.6, I=21.5, T=30.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:02,451] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:08,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:37:09,057] loop.py:45 INFO TRAIN_STEP step=6039 loss=1.419323 Training: 34%|███▎ | 6039/18018 [13:06:55<22:20:43, 6.72s/step, loss=1.419, acc=0.682, acc_len=1.63, thru=17.2, I=22.0, T=28.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:09,382] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:37:15,794] loop.py:380 INFO TIMING step=6040: step=4.000s data=1.688s compute=3.971s [fwd=2.756s bwd=1.200s opt=0.003s] dispatch=2.728s [2026-10-19 02:37:15,794] loop.py:45 INFO TRAIN_STEP step=6040 loss=1.400230 Training: 34%|███▎ | 6040/18018 [13:07:01<22:21:54, 6.72s/step, loss=1.400, acc=0.689, acc_len=1.65, thru=16.9, I=19.4, T=32.0, wait=2.7s, pool=44, epoch=1/2][2026-10-19 02:37:24,041] loop.py:45 INFO TRAIN_STEP step=6041 loss=1.514921 Training: 34%|███▎ | 6041/18018 [13:07:10<23:53:04, 7.18s/step, loss=1.515, acc=0.667, acc_len=1.58, thru=17.2, I=14.8, T=33.2, wait=4.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:37:24,036] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6040: forward=2715.5ms backward=1093.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:31,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:37:33,426] loop.py:45 INFO TRAIN_STEP step=6042 loss=1.346703 Training: 34%|███▎ | 6042/18018 [13:07:19<26:05:14, 7.84s/step, loss=1.347, acc=0.694, acc_len=1.66, thru=15.2, I=17.9, T=26.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:33,819] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:39,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:37:39,692] loop.py:45 INFO TRAIN_STEP step=6043 loss=1.558455 Training: 34%|███▎ | 6043/18018 [13:07:25<24:30:39, 7.37s/step, loss=1.558, acc=0.659, acc_len=1.54, thru=18.7, I=22.4, T=35.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:39,824] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:45,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:37:46,721] loop.py:45 INFO TRAIN_STEP step=6044 loss=1.400587 Training: 34%|███▎ | 6044/18018 [13:07:32<24:10:10, 7.27s/step, loss=1.401, acc=0.684, acc_len=1.63, thru=17.6, I=20.2, T=28.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:47,121] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:53,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:37:56,007] loop.py:380 INFO TIMING step=6045: step=4.751s data=2.279s compute=4.726s [fwd=3.332s bwd=1.379s opt=0.003s] dispatch=4.526s [2026-10-19 02:37:56,007] loop.py:45 INFO TRAIN_STEP step=6045 loss=1.422548 Training: 34%|███▎ | 6045/18018 [13:07:42<26:11:00, 7.87s/step, loss=1.423, acc=0.682, acc_len=1.62, thru=14.4, I=19.0, T=26.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:37:56,381] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:04,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:38:05,034] loop.py:45 INFO TRAIN_STEP step=6046 loss=1.338302 Training: 34%|███▎ | 6046/18018 [13:07:51<27:20:01, 8.22s/step, loss=1.338, acc=0.695, acc_len=1.67, thru=14.8, I=14.9, T=34.3, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:05,087] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (TrainerActor pid=9310) [2026-10-19 02:38:05,029] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6045: forward=2649.9ms backward=1038.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:11,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:38:13,672] loop.py:45 INFO TRAIN_STEP step=6047 loss=1.551005 Training: 34%|███▎ | 6047/18018 [13:07:59<27:44:58, 8.35s/step, loss=1.551, acc=0.662, acc_len=1.57, thru=16.1, I=20.2, T=27.8, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:13,897] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:19,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:38:20,380] loop.py:45 INFO TRAIN_STEP step=6048 loss=1.413775 Training: 34%|███▎ | 6048/18018 [13:08:06<26:06:43, 7.85s/step, loss=1.414, acc=0.683, acc_len=1.63, thru=15.4, I=22.3, T=30.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:20,428] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:27,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:38:28,546] loop.py:45 INFO TRAIN_STEP step=6049 loss=1.581054 Training: 34%|███▎ | 6049/18018 [13:08:14<26:25:23, 7.95s/step, loss=1.581, acc=0.655, acc_len=1.54, thru=14.8, I=16.0, T=32.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:28,603] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 02:38:35,237] loop.py:380 INFO TIMING step=6050: step=3.743s data=1.626s compute=3.697s [fwd=2.550s bwd=1.132s opt=0.003s] dispatch=2.938s [2026-10-19 02:38:35,237] loop.py:45 INFO TRAIN_STEP step=6050 loss=1.411287 Training: 34%|███▎ | 6050/18018 [13:08:21<25:09:57, 7.57s/step, loss=1.411, acc=0.682, acc_len=1.62, thru=17.1, I=18.1, T=34.2, wait=2.9s, pool=52, epoch=1/2][2026-10-19 02:38:41,244] loop.py:45 INFO TRAIN_STEP step=6051 loss=1.426538 Training: 34%|███▎ | 6051/18018 [13:08:27<23:36:26, 7.10s/step, loss=1.427, acc=0.679, acc_len=1.63, thru=16.8, I=18.6, T=36.4, wait=2.5s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:38:41,240] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6050: forward=2480.1ms backward=986.3ms [2026-10-19 02:38:52,083] loop.py:45 INFO TRAIN_STEP step=6052 loss=1.552222 Training: 34%|███▎ | 6052/18018 [13:08:38<27:19:44, 8.22s/step, loss=1.552, acc=0.660, acc_len=1.55, thru=15.0, I=16.2, T=31.0, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:52,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:52,576] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:38:59,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:39:01,927] loop.py:45 INFO TRAIN_STEP step=6053 loss=1.429673 Training: 34%|███▎ | 6053/18018 [13:08:48<28:56:41, 8.71s/step, loss=1.430, acc=0.682, acc_len=1.63, thru=13.4, I=15.1, T=31.0, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:02,411] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:06,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:39:08,150] loop.py:45 INFO TRAIN_STEP step=6054 loss=1.432688 Training: 34%|███▎ | 6054/18018 [13:08:54<26:27:49, 7.96s/step, loss=1.433, acc=0.679, acc_len=1.62, thru=18.7, I=31.0, T=34.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:08,361] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:14,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:39:16,378] loop.py:380 INFO TIMING step=6055: step=4.148s data=1.854s compute=4.117s [fwd=2.870s bwd=1.232s opt=0.003s] dispatch=4.072s [2026-10-19 02:39:16,379] loop.py:45 INFO TRAIN_STEP step=6055 loss=1.502453 Training: 34%|███▎ | 6055/18018 [13:09:02<26:43:38, 8.04s/step, loss=1.502, acc=0.671, acc_len=1.59, thru=16.0, I=19.3, T=30.9, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:16,630] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:21,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:39:22,437] loop.py:45 INFO TRAIN_STEP step=6056 loss=1.366271 Training: 34%|███▎ | 6056/18018 [13:09:08<24:44:51, 7.45s/step, loss=1.366, acc=0.689, acc_len=1.65, thru=19.7, I=23.8, T=35.6, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:39:22,433] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6055: forward=2481.3ms backward=1070.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:22,470] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:27,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:39:28,748] loop.py:45 INFO TRAIN_STEP step=6057 loss=1.423952 Training: 34%|███▎ | 6057/18018 [13:09:14<23:36:45, 7.11s/step, loss=1.424, acc=0.682, acc_len=1.63, thru=19.7, I=22.6, T=32.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:29,001] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:35,538] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:39:35,933] loop.py:45 INFO TRAIN_STEP step=6058 loss=1.492630 Training: 34%|███▎ | 6058/18018 [13:09:22<23:41:16, 7.13s/step, loss=1.493, acc=0.664, acc_len=1.57, thru=16.9, I=18.6, T=31.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:36,046] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:42,659] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:39:44,625] loop.py:45 INFO TRAIN_STEP step=6059 loss=1.435138 Training: 34%|███▎ | 6059/18018 [13:09:30<25:14:29, 7.60s/step, loss=1.435, acc=0.676, acc_len=1.60, thru=14.8, I=18.3, T=26.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:44,672] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:51,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:39:52,566] loop.py:380 INFO TIMING step=6060: step=4.610s data=2.207s compute=4.573s [fwd=3.182s bwd=1.376s opt=0.003s] dispatch=3.322s [2026-10-19 02:39:52,567] loop.py:45 INFO TRAIN_STEP step=6060 loss=1.562550 Training: 34%|███▎ | 6060/18018 [13:09:38<25:34:58, 7.70s/step, loss=1.563, acc=0.659, acc_len=1.55, thru=13.3, I=17.1, T=27.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:52,703] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:57,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:39:58,454] loop.py:45 INFO TRAIN_STEP step=6061 loss=1.331267 Training: 34%|███▎ | 6061/18018 [13:09:44<23:46:22, 7.16s/step, loss=1.331, acc=0.699, acc_len=1.68, thru=18.2, I=22.2, T=34.9, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:39:58,449] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6060: forward=2593.9ms backward=1026.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:39:58,648] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:04,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:40:05,986] loop.py:45 INFO TRAIN_STEP step=6062 loss=1.550670 Training: 34%|███▎ | 6062/18018 [13:09:52<24:08:31, 7.27s/step, loss=1.551, acc=0.660, acc_len=1.56, thru=17.7, I=21.8, T=29.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:06,368] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:11,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:40:12,805] loop.py:45 INFO TRAIN_STEP step=6063 loss=1.376374 Training: 34%|███▎ | 6063/18018 [13:09:58<23:41:39, 7.14s/step, loss=1.376, acc=0.689, acc_len=1.65, thru=16.3, I=23.6, T=30.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:12,843] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-19 02:40:21,353] loop.py:45 INFO TRAIN_STEP step=6064 loss=1.544876 Training: 34%|███▎ | 6064/18018 [13:10:07<25:05:47, 7.56s/step, loss=1.545, acc=0.659, acc_len=1.55, thru=14.6, I=14.5, T=30.3, wait=4.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:26,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:40:29,424] loop.py:380 INFO TIMING step=6065: step=4.734s data=2.099s compute=4.698s [fwd=3.142s bwd=1.541s opt=0.003s] dispatch=3.327s [2026-10-19 02:40:29,424] loop.py:45 INFO TRAIN_STEP step=6065 loss=1.492602 Training: 34%|███▎ | 6065/18018 [13:10:15<25:36:28, 7.71s/step, loss=1.493, acc=0.669, acc_len=1.58, thru=16.8, I=20.0, T=27.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:29,516] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:35,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:40:36,154] loop.py:45 INFO TRAIN_STEP step=6066 loss=1.338126 Training: 34%|███▎ | 6066/18018 [13:10:22<24:37:36, 7.42s/step, loss=1.338, acc=0.695, acc_len=1.67, thru=15.8, I=22.6, T=28.1, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:40:36,149] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6065: forward=3176.4ms backward=1327.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:36,376] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:40:43,679] loop.py:45 INFO TRAIN_STEP step=6067 loss=1.444867 Training: 34%|███▎ | 6067/18018 [13:10:29<24:43:48, 7.45s/step, loss=1.445, acc=0.679, acc_len=1.61, thru=14.1, I=15.4, T=28.6, wait=3.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:53,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:40:54,100] loop.py:45 INFO TRAIN_STEP step=6068 loss=1.443477 Training: 34%|███▎ | 6068/18018 [13:10:40<27:41:19, 8.34s/step, loss=1.443, acc=0.682, acc_len=1.62, thru=15.0, I=14.4, T=29.8, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:40:54,526] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:00,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:41:02,034] loop.py:45 INFO TRAIN_STEP step=6069 loss=1.564046 Training: 34%|███▎ | 6069/18018 [13:10:48<27:16:40, 8.22s/step, loss=1.564, acc=0.663, acc_len=1.56, thru=15.4, I=20.5, T=28.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:02,529] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:08,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:41:09,472] loop.py:380 INFO TIMING step=6070: step=4.185s data=1.916s compute=4.151s [fwd=2.894s bwd=1.241s opt=0.003s] dispatch=3.246s [2026-10-19 02:41:09,473] loop.py:45 INFO TRAIN_STEP step=6070 loss=1.412398 Training: 34%|███▎ | 6070/18018 [13:10:55<26:29:59, 7.98s/step, loss=1.412, acc=0.678, acc_len=1.62, thru=16.0, I=22.2, T=30.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:09,902] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:16,044] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:41:16,339] loop.py:45 INFO TRAIN_STEP step=6071 loss=1.534681 Training: 34%|███▎ | 6071/18018 [13:11:02<25:23:06, 7.65s/step, loss=1.535, acc=0.663, acc_len=1.56, thru=16.8, I=20.8, T=33.4, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:41:16,335] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6070: forward=2713.8ms backward=1079.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:16,546] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:41:25,310] loop.py:45 INFO TRAIN_STEP step=6072 loss=1.566329 Training: 34%|███▎ | 6072/18018 [13:11:11<26:41:50, 8.05s/step, loss=1.566, acc=0.661, acc_len=1.55, thru=14.5, I=13.4, T=30.6, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:25,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:25,806] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:35,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:41:35,880] loop.py:45 INFO TRAIN_STEP step=6073 loss=1.554006 Training: 34%|███▎ | 6073/18018 [13:11:21<29:12:40, 8.80s/step, loss=1.554, acc=0.667, acc_len=1.57, thru=12.6, I=14.2, T=24.8, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:36,235] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:43,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:41:43,701] loop.py:45 INFO TRAIN_STEP step=6074 loss=1.438742 Training: 34%|███▎ | 6074/18018 [13:11:29<28:13:45, 8.51s/step, loss=1.439, acc=0.678, acc_len=1.61, thru=13.8, I=14.0, T=29.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:43,753] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:49,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:41:50,315] loop.py:380 INFO TIMING step=6075: step=4.141s data=1.872s compute=4.116s [fwd=2.887s bwd=1.214s opt=0.003s] dispatch=2.464s [2026-10-19 02:41:50,315] loop.py:45 INFO TRAIN_STEP step=6075 loss=1.684448 Training: 34%|███▎ | 6075/18018 [13:11:36<26:20:26, 7.94s/step, loss=1.684, acc=0.641, acc_len=1.48, thru=18.2, I=18.7, T=30.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:50,741] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:57,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:41:58,256] loop.py:45 INFO TRAIN_STEP step=6076 loss=1.528915 Training: 34%|███▎ | 6076/18018 [13:11:44<26:20:32, 7.94s/step, loss=1.529, acc=0.667, acc_len=1.57, thru=15.8, I=17.9, T=32.1, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:41:58,251] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6075: forward=2837.2ms backward=1103.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:41:58,586] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:03,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:42:07,128] loop.py:45 INFO TRAIN_STEP step=6077 loss=1.541288 Training: 34%|███▎ | 6077/18018 [13:11:53<27:15:52, 8.22s/step, loss=1.541, acc=0.662, acc_len=1.57, thru=15.0, I=20.7, T=25.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:07,398] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:13,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:42:14,062] loop.py:45 INFO TRAIN_STEP step=6078 loss=1.362800 Training: 34%|███▎ | 6078/18018 [13:12:00<25:58:57, 7.83s/step, loss=1.363, acc=0.690, acc_len=1.66, thru=19.2, I=20.6, T=34.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:14,215] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:42:20,912] loop.py:45 INFO TRAIN_STEP step=6079 loss=1.551116 Training: 34%|███▎ | 6079/18018 [13:12:06<25:00:02, 7.54s/step, loss=1.551, acc=0.659, acc_len=1.55, thru=18.1, I=22.9, T=30.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:20,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:21,400] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:28,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:42:30,781] loop.py:380 INFO TIMING step=6080: step=4.692s data=2.301s compute=4.652s [fwd=3.204s bwd=1.432s opt=0.003s] dispatch=5.170s [2026-10-19 02:42:30,781] loop.py:45 INFO TRAIN_STEP step=6080 loss=1.615657 Training: 34%|███▎ | 6080/18018 [13:12:16<27:19:03, 8.24s/step, loss=1.616, acc=0.649, acc_len=1.52, thru=13.3, I=15.2, T=27.3, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:31,021] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 02:42:37,747] loop.py:45 INFO TRAIN_STEP step=6081 loss=1.397857 Training: 34%|███▎ | 6081/18018 [13:12:23<26:03:06, 7.86s/step, loss=1.398, acc=0.683, acc_len=1.63, thru=14.7, I=19.3, T=29.5, wait=2.6s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:42:37,742] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6080: forward=2869.0ms backward=1429.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:43,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:42:45,645] loop.py:45 INFO TRAIN_STEP step=6082 loss=1.471140 Training: 34%|███▍ | 6082/18018 [13:12:31<26:05:28, 7.87s/step, loss=1.471, acc=0.676, acc_len=1.60, thru=18.5, I=18.6, T=28.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:45,882] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:53,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:42:53,898] loop.py:45 INFO TRAIN_STEP step=6083 loss=1.443999 Training: 34%|███▍ | 6083/18018 [13:12:39<26:28:06, 7.98s/step, loss=1.444, acc=0.678, acc_len=1.62, thru=16.0, I=14.9, T=27.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:54,223] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:42:58,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:43:00,994] loop.py:45 INFO TRAIN_STEP step=6084 loss=1.543103 Training: 34%|███▍ | 6084/18018 [13:12:47<25:34:57, 7.72s/step, loss=1.543, acc=0.663, acc_len=1.56, thru=16.8, I=20.9, T=32.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:01,380] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:07,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:43:07,526] loop.py:380 INFO TIMING step=6085: step=3.683s data=1.662s compute=3.642s [fwd=2.552s bwd=1.076s opt=0.003s] dispatch=2.842s [2026-10-19 02:43:07,526] loop.py:45 INFO TRAIN_STEP step=6085 loss=1.413480 Training: 34%|███▍ | 6085/18018 [13:12:53<24:24:09, 7.36s/step, loss=1.413, acc=0.682, acc_len=1.63, thru=17.2, I=24.7, T=34.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:07,769] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:43:17,967] loop.py:45 INFO TRAIN_STEP step=6086 loss=1.425333 Training: 34%|███▍ | 6086/18018 [13:13:04<27:27:39, 8.29s/step, loss=1.425, acc=0.678, acc_len=1.61, thru=12.8, I=11.8, T=28.1, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:17,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 02:43:17,963] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6085: forward=3264.4ms backward=1254.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:18,461] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:23,928] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:43:25,520] loop.py:45 INFO TRAIN_STEP step=6087 loss=1.426488 Training: 34%|███▍ | 6087/18018 [13:13:11<26:43:55, 8.07s/step, loss=1.426, acc=0.683, acc_len=1.63, thru=17.6, I=20.2, T=31.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:25,942] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:32,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:43:33,077] loop.py:45 INFO TRAIN_STEP step=6088 loss=1.442397 Training: 34%|███▍ | 6088/18018 [13:13:19<26:13:30, 7.91s/step, loss=1.442, acc=0.677, acc_len=1.61, thru=15.0, I=20.6, T=29.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:33,409] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:38,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:43:39,480] loop.py:45 INFO TRAIN_STEP step=6089 loss=1.350684 Training: 34%|███▍ | 6089/18018 [13:13:25<24:43:16, 7.46s/step, loss=1.351, acc=0.693, acc_len=1.67, thru=17.9, I=20.9, T=32.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:39,858] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:45,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:43:46,212] loop.py:380 INFO TIMING step=6090: step=4.203s data=1.992s compute=4.179s [fwd=2.965s bwd=1.200s opt=0.003s] dispatch=2.517s [2026-10-19 02:43:46,212] loop.py:45 INFO TRAIN_STEP step=6090 loss=1.398822 Training: 34%|███▍ | 6090/18018 [13:13:32<23:59:35, 7.24s/step, loss=1.399, acc=0.683, acc_len=1.63, thru=18.2, I=21.6, T=30.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:46,481] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:53,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:43:53,436] loop.py:45 INFO TRAIN_STEP step=6091 loss=1.548915 Training: 34%|███▍ | 6091/18018 [13:13:39<23:58:31, 7.24s/step, loss=1.549, acc=0.656, acc_len=1.55, thru=17.5, I=20.4, T=28.6, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:43:53,431] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6090: forward=3000.2ms backward=1436.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:43:53,678] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:02,237] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:44:03,406] loop.py:45 INFO TRAIN_STEP step=6092 loss=1.473768 Training: 34%|███▍ | 6092/18018 [13:13:49<26:41:25, 8.06s/step, loss=1.474, acc=0.673, acc_len=1.60, thru=13.4, I=13.6, T=28.7, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:03,748] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:07,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:44:09,668] loop.py:45 INFO TRAIN_STEP step=6093 loss=1.483602 Training: 34%|███▍ | 6093/18018 [13:13:55<24:54:10, 7.52s/step, loss=1.484, acc=0.668, acc_len=1.58, thru=18.5, I=22.3, T=32.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:09,701] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 02:44:18,137] loop.py:45 INFO TRAIN_STEP step=6094 loss=1.522254 Training: 34%|███▍ | 6094/18018 [13:14:04<25:50:45, 7.80s/step, loss=1.522, acc=0.665, acc_len=1.57, thru=14.3, I=13.2, T=30.1, wait=4.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:24,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:44:25,988] loop.py:380 INFO TIMING step=6095: step=4.358s data=1.848s compute=4.318s [fwd=2.960s bwd=1.343s opt=0.003s] dispatch=3.484s [2026-10-19 02:44:25,988] loop.py:45 INFO TRAIN_STEP step=6095 loss=1.550406 Training: 34%|███▍ | 6095/18018 [13:14:12<25:53:27, 7.82s/step, loss=1.550, acc=0.658, acc_len=1.55, thru=16.0, I=17.5, T=29.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:26,220] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:32,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:44:32,873] loop.py:45 INFO TRAIN_STEP step=6096 loss=1.582188 Training: 34%|███▍ | 6096/18018 [13:14:18<24:57:52, 7.54s/step, loss=1.582, acc=0.655, acc_len=1.53, thru=17.3, I=21.7, T=33.1, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:44:32,868] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6095: forward=2734.4ms backward=1087.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:33,183] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:36,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:44:38,741] loop.py:45 INFO TRAIN_STEP step=6097 loss=1.436603 Training: 34%|███▍ | 6097/18018 [13:14:24<23:18:10, 7.04s/step, loss=1.437, acc=0.676, acc_len=1.61, thru=20.0, I=28.1, T=34.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:39,126] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:44,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:44:45,956] loop.py:45 INFO TRAIN_STEP step=6098 loss=1.400503 Training: 34%|███▍ | 6098/18018 [13:14:32<23:28:39, 7.09s/step, loss=1.401, acc=0.686, acc_len=1.64, thru=16.0, I=26.3, T=29.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:46,302] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:51,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:44:53,354] loop.py:45 INFO TRAIN_STEP step=6099 loss=1.438064 Training: 34%|███▍ | 6099/18018 [13:14:39<23:46:47, 7.18s/step, loss=1.438, acc=0.680, acc_len=1.62, thru=16.6, I=22.3, T=30.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:44:53,472] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:00,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:45:01,491] loop.py:380 INFO TIMING step=6100: step=3.832s data=1.635s compute=3.806s [fwd=2.722s bwd=1.069s opt=0.003s] dispatch=4.295s [2026-10-19 02:45:01,491] loop.py:45 INFO TRAIN_STEP step=6100 loss=1.499830 Training: 34%|███▍ | 6100/18018 [13:14:47<24:43:35, 7.47s/step, loss=1.500, acc=0.669, acc_len=1.58, thru=14.1, I=18.3, T=33.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:01,754] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:07,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 02:45:08,741] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6100: forward=3076.6ms backward=1332.6ms [2026-10-19 02:45:08,745] loop.py:45 INFO TRAIN_STEP step=6101 loss=1.515271 Training: 34%|███▍ | 6101/18018 [13:14:54<24:30:40, 7.40s/step, loss=1.515, acc=0.665, acc_len=1.57, thru=18.0, I=24.1, T=28.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:09,125] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:14,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:45:16,884] loop.py:45 INFO TRAIN_STEP step=6102 loss=1.467625 Training: 34%|███▍ | 6102/18018 [13:15:02<25:14:18, 7.62s/step, loss=1.468, acc=0.675, acc_len=1.61, thru=13.6, I=24.6, T=25.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:17,121] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 02:45:22,800] loop.py:45 INFO TRAIN_STEP step=6103 loss=1.528763 Training: 34%|███▍ | 6103/18018 [13:15:08<23:32:26, 7.11s/step, loss=1.529, acc=0.665, acc_len=1.56, thru=15.0, I=25.3, T=34.2, wait=2.2s, pool=44, epoch=1/2][2026-10-19 02:45:30,913] loop.py:45 INFO TRAIN_STEP step=6104 loss=1.562666 Training: 34%|███▍ | 6104/18018 [13:15:16<24:31:52, 7.41s/step, loss=1.563, acc=0.661, acc_len=1.55, thru=14.9, I=15.9, T=29.3, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:41,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:45:44,091] loop.py:380 INFO TIMING step=6105: step=4.699s data=2.006s compute=4.666s [fwd=3.321s bwd=1.330s opt=0.003s] dispatch=8.469s [2026-10-19 02:45:44,091] loop.py:45 INFO TRAIN_STEP step=6105 loss=1.392383 Training: 34%|███▍ | 6105/18018 [13:15:30<30:15:09, 9.14s/step, loss=1.392, acc=0.687, acc_len=1.64, thru=14.2, I=12.9, T=27.2, wait=8.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:44,339] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:52,528] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:45:52,740] loop.py:45 INFO TRAIN_STEP step=6106 loss=1.584615 Training: 34%|███▍ | 6106/18018 [13:15:38<29:45:39, 8.99s/step, loss=1.585, acc=0.657, acc_len=1.54, thru=15.2, I=15.9, T=28.7, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:45:52,734] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6105: forward=3094.6ms backward=1302.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:45:53,031] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:45:59,236] loop.py:45 INFO TRAIN_STEP step=6107 loss=1.544413 Training: 34%|███▍ | 6107/18018 [13:15:45<27:16:30, 8.24s/step, loss=1.544, acc=0.660, acc_len=1.55, thru=16.9, I=16.5, T=32.7, wait=2.6s, pool=56, epoch=1/2][2026-10-19 02:46:06,642] loop.py:45 INFO TRAIN_STEP step=6108 loss=1.466283 Training: 34%|███▍ | 6108/18018 [13:15:52<26:26:34, 7.99s/step, loss=1.466, acc=0.673, acc_len=1.59, thru=17.3, I=17.2, T=29.2, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:13,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:46:16,073] loop.py:45 INFO TRAIN_STEP step=6109 loss=1.568808 Training: 34%|███▍ | 6109/18018 [13:16:02<27:52:05, 8.42s/step, loss=1.569, acc=0.661, acc_len=1.56, thru=15.3, I=16.5, T=28.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:16,135] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:20,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:46:22,366] loop.py:380 INFO TIMING step=6110: step=4.133s data=1.911s compute=4.105s [fwd=2.909s bwd=1.181s opt=0.003s] dispatch=2.151s [2026-10-19 02:46:22,366] loop.py:45 INFO TRAIN_STEP step=6110 loss=1.578418 Training: 34%|███▍ | 6110/18018 [13:16:08<25:45:02, 7.78s/step, loss=1.578, acc=0.656, acc_len=1.55, thru=16.6, I=29.5, T=31.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:22,616] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:28,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:46:29,262] loop.py:45 INFO TRAIN_STEP step=6111 loss=1.483835 Training: 34%|███▍ | 6111/18018 [13:16:15<24:51:57, 7.52s/step, loss=1.484, acc=0.671, acc_len=1.58, thru=16.4, I=24.5, T=31.5, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:46:29,257] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6110: forward=2876.0ms backward=1147.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:29,511] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:35,340] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:46:35,631] loop.py:45 INFO TRAIN_STEP step=6112 loss=1.535486 Training: 34%|███▍ | 6112/18018 [13:16:21<23:43:32, 7.17s/step, loss=1.535, acc=0.664, acc_len=1.58, thru=18.1, I=21.1, T=32.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:35,843] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:40,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:46:42,107] loop.py:45 INFO TRAIN_STEP step=6113 loss=1.500504 Training: 34%|███▍ | 6113/18018 [13:16:28<23:01:51, 6.96s/step, loss=1.501, acc=0.670, acc_len=1.59, thru=19.9, I=21.7, T=33.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:42,410] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:49,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:46:51,839] loop.py:45 INFO TRAIN_STEP step=6114 loss=1.443020 Training: 34%|███▍ | 6114/18018 [13:16:37<25:46:30, 7.79s/step, loss=1.443, acc=0.678, acc_len=1.62, thru=13.6, I=16.3, T=26.6, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:51,948] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:56,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:46:58,481] loop.py:380 INFO TIMING step=6115: step=4.051s data=1.896s compute=4.023s [fwd=2.842s bwd=1.166s opt=0.003s] dispatch=2.581s [2026-10-19 02:46:58,482] loop.py:45 INFO TRAIN_STEP step=6115 loss=1.576595 Training: 34%|███▍ | 6115/18018 [13:16:44<24:37:42, 7.45s/step, loss=1.577, acc=0.652, acc_len=1.52, thru=16.7, I=23.4, T=31.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:46:58,511] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:04,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:47:05,748] loop.py:45 INFO TRAIN_STEP step=6116 loss=1.548151 Training: 34%|███▍ | 6116/18018 [13:16:51<24:26:48, 7.39s/step, loss=1.548, acc=0.658, acc_len=1.54, thru=16.2, I=22.4, T=32.1, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:47:05,742] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6115: forward=2837.5ms backward=1111.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:06,174] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:47:13,628] loop.py:45 INFO TRAIN_STEP step=6117 loss=1.529052 Training: 34%|███▍ | 6117/18018 [13:16:59<24:55:28, 7.54s/step, loss=1.529, acc=0.666, acc_len=1.57, thru=14.2, I=17.2, T=33.2, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:13,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:14,061] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:20,425] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:47:21,243] loop.py:45 INFO TRAIN_STEP step=6118 loss=1.517326 Training: 34%|███▍ | 6118/18018 [13:17:07<24:59:56, 7.56s/step, loss=1.517, acc=0.667, acc_len=1.58, thru=17.7, I=18.1, T=29.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:21,432] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:27,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:47:29,562] loop.py:45 INFO TRAIN_STEP step=6119 loss=1.527868 Training: 34%|███▍ | 6119/18018 [13:17:15<25:44:44, 7.79s/step, loss=1.528, acc=0.667, acc_len=1.57, thru=14.6, I=17.0, T=26.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:29,813] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:37,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:47:38,518] loop.py:380 INFO TIMING step=6120: step=4.780s data=2.207s compute=4.750s [fwd=3.305s bwd=1.430s opt=0.003s] dispatch=4.167s [2026-10-19 02:47:38,519] loop.py:45 INFO TRAIN_STEP step=6120 loss=1.590347 Training: 34%|███▍ | 6120/18018 [13:17:24<26:54:08, 8.14s/step, loss=1.590, acc=0.654, acc_len=1.54, thru=14.7, I=15.0, T=26.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:38,747] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:46,171] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:47:47,414] loop.py:45 INFO TRAIN_STEP step=6121 loss=1.355229 Training: 34%|███▍ | 6121/18018 [13:17:33<27:38:56, 8.37s/step, loss=1.355, acc=0.695, acc_len=1.67, thru=14.0, I=15.8, T=27.7, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:47:47,409] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6120: forward=3223.7ms backward=1346.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:47,683] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:53,529] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:47:54,420] loop.py:45 INFO TRAIN_STEP step=6122 loss=1.492308 Training: 34%|███▍ | 6122/18018 [13:17:40<26:17:50, 7.96s/step, loss=1.492, acc=0.673, acc_len=1.59, thru=17.3, I=20.0, T=28.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:47:54,535] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:02,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:48:04,282] loop.py:45 INFO TRAIN_STEP step=6123 loss=1.526825 Training: 34%|███▍ | 6123/18018 [13:17:50<28:10:53, 8.53s/step, loss=1.527, acc=0.664, acc_len=1.57, thru=13.2, I=14.7, T=26.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:04,687] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:11,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:48:12,658] loop.py:45 INFO TRAIN_STEP step=6124 loss=1.513495 Training: 34%|███▍ | 6124/18018 [13:17:58<28:01:37, 8.48s/step, loss=1.513, acc=0.668, acc_len=1.57, thru=13.2, I=16.2, T=30.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:12,875] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:22,757] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:48:23,330] loop.py:380 INFO TIMING step=6125: step=5.106s data=2.500s compute=5.073s [fwd=3.630s bwd=1.429s opt=0.003s] dispatch=5.558s [2026-10-19 02:48:23,330] loop.py:45 INFO TRAIN_STEP step=6125 loss=1.444376 Training: 34%|███▍ | 6125/18018 [13:18:09<30:11:37, 9.14s/step, loss=1.444, acc=0.679, acc_len=1.62, thru=12.8, I=12.2, T=25.1, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:23,762] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:31,425] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:48:31,830] loop.py:45 INFO TRAIN_STEP step=6126 loss=1.503060 Training: 34%|███▍ | 6126/18018 [13:18:17<29:33:28, 8.95s/step, loss=1.503, acc=0.668, acc_len=1.58, thru=15.8, I=17.5, T=32.2, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:48:31,825] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6125: forward=2675.9ms backward=1243.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:31,929] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:37,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:48:37,990] loop.py:45 INFO TRAIN_STEP step=6127 loss=1.468050 Training: 34%|███▍ | 6127/18018 [13:18:24<26:47:31, 8.11s/step, loss=1.468, acc=0.676, acc_len=1.60, thru=19.5, I=18.2, T=30.2, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:38,442] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:44,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:48:46,939] loop.py:45 INFO TRAIN_STEP step=6128 loss=1.402479 Training: 34%|███▍ | 6128/18018 [13:18:33<27:37:22, 8.36s/step, loss=1.402, acc=0.689, acc_len=1.64, thru=15.3, I=18.6, T=25.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:47,002] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:53,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:48:54,577] loop.py:45 INFO TRAIN_STEP step=6129 loss=1.446927 Training: 34%|███▍ | 6129/18018 [13:18:40<26:53:54, 8.14s/step, loss=1.447, acc=0.679, acc_len=1.62, thru=16.9, I=19.5, T=28.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:48:54,990] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:49:02,842] loop.py:380 INFO TIMING step=6130: step=4.173s data=1.775s compute=4.120s [fwd=2.954s bwd=1.153s opt=0.003s] dispatch=4.085s [2026-10-19 02:49:02,842] loop.py:45 INFO TRAIN_STEP step=6130 loss=1.470475 Training: 34%|███▍ | 6130/18018 [13:18:48<27:01:03, 8.18s/step, loss=1.470, acc=0.674, acc_len=1.61, thru=13.7, I=15.2, T=30.7, wait=4.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:08,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:49:10,924] loop.py:45 INFO TRAIN_STEP step=6131 loss=1.553402 Training: 34%|███▍ | 6131/18018 [13:18:57<26:54:53, 8.15s/step, loss=1.553, acc=0.660, acc_len=1.55, thru=17.6, I=20.1, T=29.4, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:49:10,919] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6130: forward=2953.3ms backward=1358.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:11,350] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:18,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:49:18,785] loop.py:45 INFO TRAIN_STEP step=6132 loss=1.390149 Training: 34%|███▍ | 6132/18018 [13:19:04<26:37:39, 8.06s/step, loss=1.390, acc=0.687, acc_len=1.64, thru=13.3, I=18.3, T=32.0, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:18,819] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:27,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:49:28,562] loop.py:45 INFO TRAIN_STEP step=6133 loss=1.414977 Training: 34%|███▍ | 6133/18018 [13:19:14<28:19:07, 8.58s/step, loss=1.415, acc=0.683, acc_len=1.63, thru=13.2, I=15.4, T=25.8, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:28,679] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:34,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:49:36,001] loop.py:45 INFO TRAIN_STEP step=6134 loss=1.592793 Training: 34%|███▍ | 6134/18018 [13:19:22<27:11:19, 8.24s/step, loss=1.593, acc=0.655, acc_len=1.53, thru=14.5, I=18.7, T=27.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:36,110] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:44,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:49:44,382] loop.py:380 INFO TIMING step=6135: step=4.726s data=2.262s compute=4.683s [fwd=3.219s bwd=1.449s opt=0.003s] dispatch=3.646s [2026-10-19 02:49:44,382] loop.py:45 INFO TRAIN_STEP step=6135 loss=1.449657 Training: 34%|███▍ | 6135/18018 [13:19:30<27:19:46, 8.28s/step, loss=1.450, acc=0.677, acc_len=1.61, thru=15.5, I=15.4, T=27.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:44,689] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:52,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:49:53,329] loop.py:45 INFO TRAIN_STEP step=6136 loss=1.437380 Training: 34%|███▍ | 6136/18018 [13:19:39<27:59:16, 8.48s/step, loss=1.437, acc=0.678, acc_len=1.62, thru=14.8, I=16.5, T=25.5, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:49:53,324] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6135: forward=3546.8ms backward=1416.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:53,429] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:49:58,529] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:50:00,784] loop.py:45 INFO TRAIN_STEP step=6137 loss=1.364334 Training: 34%|███▍ | 6137/18018 [13:19:46<26:58:16, 8.17s/step, loss=1.364, acc=0.693, acc_len=1.66, thru=17.4, I=18.8, T=27.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:01,041] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:07,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:50:08,269] loop.py:45 INFO TRAIN_STEP step=6138 loss=1.362131 Training: 34%|███▍ | 6138/18018 [13:19:54<26:17:27, 7.97s/step, loss=1.362, acc=0.690, acc_len=1.66, thru=13.6, I=18.8, T=29.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:08,520] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:50:16,913] loop.py:45 INFO TRAIN_STEP step=6139 loss=1.445950 Training: 34%|███▍ | 6139/18018 [13:20:02<26:57:24, 8.17s/step, loss=1.446, acc=0.679, acc_len=1.61, thru=12.2, I=14.0, T=31.0, wait=4.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:25,237] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:50:26,503] loop.py:380 INFO TIMING step=6140: step=5.359s data=2.087s compute=5.331s [fwd=3.982s bwd=1.333s opt=0.003s] dispatch=4.222s [2026-10-19 02:50:26,504] loop.py:45 INFO TRAIN_STEP step=6140 loss=1.568316 Training: 34%|███▍ | 6140/18018 [13:20:12<28:21:37, 8.60s/step, loss=1.568, acc=0.659, acc_len=1.55, thru=15.7, I=16.7, T=23.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:26,746] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:32,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:50:33,948] loop.py:45 INFO TRAIN_STEP step=6141 loss=1.547426 Training: 34%|███▍ | 6141/18018 [13:20:20<27:13:06, 8.25s/step, loss=1.547, acc=0.662, acc_len=1.56, thru=16.2, I=17.9, T=31.6, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:50:33,943] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6140: forward=2886.7ms backward=1112.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:34,417] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:39,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:50:40,248] loop.py:45 INFO TRAIN_STEP step=6142 loss=1.536248 Training: 34%|███▍ | 6142/18018 [13:20:26<25:17:14, 7.67s/step, loss=1.536, acc=0.660, acc_len=1.56, thru=18.9, I=23.7, T=34.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:40,278] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:44,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:50:47,103] loop.py:45 INFO TRAIN_STEP step=6143 loss=1.428357 Training: 34%|███▍ | 6143/18018 [13:20:33<24:29:06, 7.42s/step, loss=1.428, acc=0.682, acc_len=1.63, thru=17.5, I=26.5, T=29.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:47,255] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:51,825] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:50:53,250] loop.py:45 INFO TRAIN_STEP step=6144 loss=1.476581 Training: 34%|███▍ | 6144/18018 [13:20:39<23:13:12, 7.04s/step, loss=1.477, acc=0.669, acc_len=1.59, thru=17.4, I=28.5, T=32.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:50:53,347] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-19 02:51:01,533] loop.py:380 INFO TIMING step=6145: step=4.696s data=2.193s compute=4.663s [fwd=3.247s bwd=1.402s opt=0.003s] dispatch=3.576s [2026-10-19 02:51:01,533] loop.py:45 INFO TRAIN_STEP step=6145 loss=1.562928 Training: 34%|███▍ | 6145/18018 [13:20:47<24:26:51, 7.41s/step, loss=1.563, acc=0.661, acc_len=1.56, thru=10.7, I=14.5, T=27.3, wait=3.6s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:12,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:51:14,842] loop.py:45 INFO TRAIN_STEP step=6146 loss=1.602856 Training: 34%|███▍ | 6146/18018 [13:21:00<30:16:49, 9.18s/step, loss=1.603, acc=0.652, acc_len=1.53, thru=12.4, I=12.9, T=24.8, wait=8.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:51:14,837] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6145: forward=3651.2ms backward=1432.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:14,964] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-19 02:51:24,125] loop.py:45 INFO TRAIN_STEP step=6147 loss=1.446190 Training: 34%|███▍ | 6147/18018 [13:21:10<30:22:28, 9.21s/step, loss=1.446, acc=0.679, acc_len=1.62, thru=13.1, I=15.8, T=29.0, wait=4.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:30,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:51:32,431] loop.py:45 INFO TRAIN_STEP step=6148 loss=1.535426 Training: 34%|███▍ | 6148/18018 [13:21:18<29:28:38, 8.94s/step, loss=1.535, acc=0.661, acc_len=1.55, thru=15.8, I=17.5, T=25.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:32,630] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:38,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:51:39,098] loop.py:45 INFO TRAIN_STEP step=6149 loss=1.498579 Training: 34%|███▍ | 6149/18018 [13:21:25<27:13:38, 8.26s/step, loss=1.499, acc=0.667, acc_len=1.58, thru=14.8, I=20.0, T=30.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:39,360] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:46,525] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:51:46,744] loop.py:380 INFO TIMING step=6150: step=4.495s data=2.082s compute=4.450s [fwd=3.038s bwd=1.398s opt=0.003s] dispatch=3.140s [2026-10-19 02:51:46,744] loop.py:45 INFO TRAIN_STEP step=6150 loss=1.334248 Training: 34%|███▍ | 6150/18018 [13:21:32<26:37:01, 8.07s/step, loss=1.334, acc=0.698, acc_len=1.68, thru=16.5, I=17.8, T=28.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:47,026] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:52,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:51:53,269] loop.py:45 INFO TRAIN_STEP step=6151 loss=1.593419 Training: 34%|███▍ | 6151/18018 [13:21:39<25:05:01, 7.61s/step, loss=1.593, acc=0.651, acc_len=1.52, thru=18.1, I=19.7, T=34.9, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:51:53,264] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6150: forward=2611.4ms backward=1007.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:53,297] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:51:58,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:52:00,538] loop.py:45 INFO TRAIN_STEP step=6152 loss=1.546600 Training: 34%|███▍ | 6152/18018 [13:21:46<24:44:48, 7.51s/step, loss=1.547, acc=0.661, acc_len=1.56, thru=16.9, I=21.8, T=31.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:52:00,625] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 02:52:08,078] loop.py:45 INFO TRAIN_STEP step=6153 loss=1.464672 Training: 34%|███▍ | 6153/18018 [13:21:54<24:46:36, 7.52s/step, loss=1.465, acc=0.677, acc_len=1.61, thru=14.4, I=18.3, T=31.2, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:52:15,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:52:16,021] loop.py:45 INFO TRAIN_STEP step=6154 loss=1.458250 Training: 34%|███▍ | 6154/18018 [13:22:02<25:11:38, 7.64s/step, loss=1.458, acc=0.678, acc_len=1.61, thru=16.9, I=18.0, T=32.2, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:52:16,255] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:52:26,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:52:26,870] loop.py:380 INFO TIMING step=6155: step=5.322s data=2.900s compute=5.291s [fwd=3.831s bwd=1.445s opt=0.003s] dispatch=5.517s [2026-10-19 02:52:26,870] loop.py:45 INFO TRAIN_STEP step=6155 loss=1.473059 Training: 34%|███▍ | 6155/18018 [13:22:12<28:21:31, 8.61s/step, loss=1.473, acc=0.676, acc_len=1.60, thru=12.5, I=11.4, T=24.1, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:52:27,022] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:52:35,758] loop.py:45 INFO TRAIN_STEP step=6156 loss=1.547814 Training: 34%|███▍ | 6156/18018 [13:22:21<28:38:08, 8.69s/step, loss=1.548, acc=0.664, acc_len=1.57, thru=14.1, I=15.7, T=29.0, wait=4.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:52:35,753] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6155: forward=3168.5ms backward=1203.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:52:43,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:52:44,391] loop.py:45 INFO TRAIN_STEP step=6157 loss=1.386361 Training: 34%|███▍ | 6157/18018 [13:22:30<28:34:40, 8.67s/step, loss=1.386, acc=0.687, acc_len=1.64, thru=15.4, I=15.5, T=28.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:52:44,433] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:52:54,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:52:56,257] loop.py:45 INFO TRAIN_STEP step=6158 loss=1.503038 Training: 34%|███▍ | 6158/18018 [13:22:42<31:43:48, 9.63s/step, loss=1.503, acc=0.670, acc_len=1.59, thru=16.1, I=14.8, T=26.2, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:52:56,289] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-19 02:53:04,785] loop.py:45 INFO TRAIN_STEP step=6159 loss=1.535529 Training: 34%|███▍ | 6159/18018 [13:22:50<30:38:06, 9.30s/step, loss=1.536, acc=0.662, acc_len=1.55, thru=14.2, I=16.9, T=32.5, wait=4.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:53:15,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:53:15,837] loop.py:380 INFO TIMING step=6160: step=5.347s data=2.672s compute=5.289s [fwd=3.749s bwd=1.525s opt=0.003s] dispatch=5.696s [2026-10-19 02:53:15,837] loop.py:45 INFO TRAIN_STEP step=6160 loss=1.527956 Training: 34%|███▍ | 6160/18018 [13:23:01<32:21:56, 9.83s/step, loss=1.528, acc=0.663, acc_len=1.57, thru=13.5, I=12.3, T=23.9, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:53:16,077] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:53:23,528] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:53:23,953] loop.py:45 INFO TRAIN_STEP step=6161 loss=1.522199 Training: 34%|███▍ | 6161/18018 [13:23:10<30:40:23, 9.31s/step, loss=1.522, acc=0.665, acc_len=1.57, thru=14.5, I=13.8, T=28.6, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:53:23,947] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6160: forward=3120.5ms backward=1311.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:53:24,040] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 [2026-10-19 02:53:31,184] loop.py:45 INFO TRAIN_STEP step=6162 loss=1.408366 Training: 34%|███▍ | 6162/18018 [13:23:17<28:36:47, 8.69s/step, loss=1.408, acc=0.685, acc_len=1.63, thru=15.4, I=15.3, T=31.5, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:53:40,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:53:40,986] loop.py:45 INFO TRAIN_STEP step=6163 loss=1.629414 Training: 34%|███▍ | 6163/18018 [13:23:27<29:42:41, 9.02s/step, loss=1.629, acc=0.651, acc_len=1.52, thru=14.9, I=16.2, T=28.0, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:53:41,479] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:53:49,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:53:50,401] loop.py:45 INFO TRAIN_STEP step=6164 loss=1.587687 Training: 34%|███▍ | 6164/18018 [13:23:36<30:05:47, 9.14s/step, loss=1.588, acc=0.653, acc_len=1.52, thru=13.3, I=14.5, T=27.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:53:50,859] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:53:56,738] loop.py:380 INFO TIMING step=6165: step=3.742s data=1.626s compute=3.707s [fwd=2.633s bwd=1.060s opt=0.003s] dispatch=2.585s [2026-10-19 02:53:56,738] loop.py:45 INFO TRAIN_STEP step=6165 loss=1.497816 Training: 34%|███▍ | 6165/18018 [13:23:42<27:19:32, 8.30s/step, loss=1.498, acc=0.668, acc_len=1.57, thru=17.5, I=20.7, T=34.2, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:03,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:54:05,004] loop.py:45 INFO TRAIN_STEP step=6166 loss=1.652519 Training: 34%|███▍ | 6166/18018 [13:23:51<27:17:27, 8.29s/step, loss=1.653, acc=0.646, acc_len=1.51, thru=16.4, I=19.1, T=26.7, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:54:04,998] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6165: forward=3312.0ms backward=1432.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:05,443] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:54:13,567] loop.py:45 INFO TRAIN_STEP step=6167 loss=1.405918 Training: 34%|███▍ | 6167/18018 [13:23:59<27:33:25, 8.37s/step, loss=1.406, acc=0.685, acc_len=1.64, thru=14.3, I=13.3, T=30.6, wait=4.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:21,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:54:23,494] loop.py:45 INFO TRAIN_STEP step=6168 loss=1.515689 Training: 34%|███▍ | 6168/18018 [13:24:09<29:05:41, 8.84s/step, loss=1.516, acc=0.668, acc_len=1.58, thru=14.2, I=16.7, T=28.8, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:23,742] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:28,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:54:29,745] loop.py:45 INFO TRAIN_STEP step=6169 loss=1.551879 Training: 34%|███▍ | 6169/18018 [13:24:15<26:31:53, 8.06s/step, loss=1.552, acc=0.660, acc_len=1.55, thru=17.4, I=25.2, T=32.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:30,185] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:36,706] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:54:36,952] loop.py:380 INFO TIMING step=6170: step=4.546s data=2.148s compute=4.511s [fwd=3.132s bwd=1.365s opt=0.003s] dispatch=2.653s [2026-10-19 02:54:36,952] loop.py:45 INFO TRAIN_STEP step=6170 loss=1.299550 Training: 34%|███▍ | 6170/18018 [13:24:23<25:41:20, 7.81s/step, loss=1.300, acc=0.701, acc_len=1.69, thru=17.9, I=21.6, T=28.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:37,206] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:41,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:54:44,421] loop.py:45 INFO TRAIN_STEP step=6171 loss=1.479956 Training: 34%|███▍ | 6171/18018 [13:24:30<25:21:16, 7.70s/step, loss=1.480, acc=0.668, acc_len=1.58, thru=16.6, I=21.4, T=29.2, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:54:44,415] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6170: forward=2983.2ms backward=1344.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:44,494] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 [2026-10-19 02:54:51,244] loop.py:45 INFO TRAIN_STEP step=6172 loss=1.510070 Training: 34%|███▍ | 6172/18018 [13:24:37<24:28:49, 7.44s/step, loss=1.510, acc=0.667, acc_len=1.58, thru=15.6, I=21.2, T=33.2, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:56,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:54:57,632] loop.py:45 INFO TRAIN_STEP step=6173 loss=1.456679 Training: 34%|███▍ | 6173/18018 [13:24:43<23:26:26, 7.12s/step, loss=1.457, acc=0.675, acc_len=1.60, thru=18.6, I=20.7, T=32.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:54:57,853] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:02,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:55:03,511] loop.py:45 INFO TRAIN_STEP step=6174 loss=1.404533 Training: 34%|███▍ | 6174/18018 [13:24:49<22:12:41, 6.75s/step, loss=1.405, acc=0.684, acc_len=1.63, thru=20.1, I=23.5, T=36.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:03,802] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:55:10,554] loop.py:380 INFO TIMING step=6175: step=3.570s data=1.573s compute=3.539s [fwd=2.534s bwd=0.991s opt=0.003s] dispatch=3.461s [2026-10-19 02:55:10,554] loop.py:45 INFO TRAIN_STEP step=6175 loss=1.616916 Training: 34%|███▍ | 6175/18018 [13:24:56<22:29:48, 6.84s/step, loss=1.617, acc=0.645, acc_len=1.51, thru=16.8, I=19.1, T=35.9, wait=3.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:16,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:55:18,248] loop.py:45 INFO TRAIN_STEP step=6176 loss=1.418531 Training: 34%|███▍ | 6176/18018 [13:25:04<23:20:16, 7.09s/step, loss=1.419, acc=0.683, acc_len=1.63, thru=17.9, I=19.2, T=29.3, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:55:18,244] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6175: forward=3058.1ms backward=1258.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:18,316] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:23,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:55:24,518] loop.py:45 INFO TRAIN_STEP step=6177 loss=1.404183 Training: 34%|███▍ | 6177/18018 [13:25:10<22:31:21, 6.85s/step, loss=1.404, acc=0.688, acc_len=1.65, thru=17.7, I=26.4, T=32.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:24,769] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 02:55:32,171] loop.py:45 INFO TRAIN_STEP step=6178 loss=1.538987 Training: 34%|███▍ | 6178/18018 [13:25:18<23:19:01, 7.09s/step, loss=1.539, acc=0.666, acc_len=1.57, thru=14.7, I=14.9, T=31.7, wait=3.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:40,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:55:40,827] loop.py:45 INFO TRAIN_STEP step=6179 loss=1.471188 Training: 34%|███▍ | 6179/18018 [13:25:26<24:51:26, 7.56s/step, loss=1.471, acc=0.674, acc_len=1.60, thru=16.8, I=17.4, T=28.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:41,229] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 02:55:49,213] loop.py:380 INFO TIMING step=6180: step=4.539s data=2.120s compute=4.506s [fwd=3.113s bwd=1.379s opt=0.003s] dispatch=3.839s [2026-10-19 02:55:49,213] loop.py:45 INFO TRAIN_STEP step=6180 loss=1.566934 Training: 34%|███▍ | 6180/18018 [13:25:35<25:40:19, 7.81s/step, loss=1.567, acc=0.656, acc_len=1.55, thru=12.5, I=12.4, T=28.2, wait=3.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:57,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:55:58,796] loop.py:45 INFO TRAIN_STEP step=6181 loss=1.655766 Training: 34%|███▍ | 6181/18018 [13:25:44<27:25:17, 8.34s/step, loss=1.656, acc=0.643, acc_len=1.49, thru=16.4, I=16.0, T=30.8, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:55:58,791] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6180: forward=2940.1ms backward=1161.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:55:58,825] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 02:56:04,948] loop.py:45 INFO TRAIN_STEP step=6182 loss=1.370604 Training: 34%|███▍ | 6182/18018 [13:25:51<25:15:43, 7.68s/step, loss=1.371, acc=0.690, acc_len=1.66, thru=18.1, I=23.5, T=30.3, wait=1.9s, pool=52, epoch=1/2][2026-10-19 02:56:15,058] loop.py:45 INFO TRAIN_STEP step=6183 loss=1.570449 Training: 34%|███▍ | 6183/18018 [13:26:01<27:39:12, 8.41s/step, loss=1.570, acc=0.659, acc_len=1.55, thru=12.5, I=11.0, T=28.7, wait=5.6s, pool=48, epoch=1/2][2026-10-19 02:56:24,468] loop.py:45 INFO TRAIN_STEP step=6184 loss=1.586641 Training: 34%|███▍ | 6184/18018 [13:26:10<28:38:04, 8.71s/step, loss=1.587, acc=0.653, acc_len=1.53, thru=15.2, I=14.5, T=26.9, wait=4.6s, pool=60, epoch=1/2][2026-10-19 02:56:31,120] loop.py:380 INFO TIMING step=6185: step=4.159s data=1.811s compute=4.131s [fwd=2.886s bwd=1.230s opt=0.003s] dispatch=2.484s [2026-10-19 02:56:31,120] loop.py:45 INFO TRAIN_STEP step=6185 loss=1.409766 Training: 34%|███▍ | 6185/18018 [13:26:17<26:36:13, 8.09s/step, loss=1.410, acc=0.682, acc_len=1.62, thru=16.6, I=18.3, T=30.8, wait=2.5s, pool=56, epoch=1/2][2026-10-19 02:56:39,264] loop.py:45 INFO TRAIN_STEP step=6186 loss=1.508843 Training: 34%|███▍ | 6186/18018 [13:26:25<26:39:00, 8.11s/step, loss=1.509, acc=0.670, acc_len=1.58, thru=16.0, I=16.8, T=30.7, wait=4.0s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:56:39,259] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6185: forward=2965.3ms backward=1165.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:56:46,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:56:47,856] loop.py:45 INFO TRAIN_STEP step=6187 loss=1.453898 Training: 34%|███▍ | 6187/18018 [13:26:33<27:07:30, 8.25s/step, loss=1.454, acc=0.678, acc_len=1.61, thru=15.1, I=15.6, T=27.5, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:56:48,329] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:56:55,654] loop.py:45 INFO TRAIN_STEP step=6188 loss=1.507624 Training: 34%|███▍ | 6188/18018 [13:26:41<26:40:20, 8.12s/step, loss=1.508, acc=0.671, acc_len=1.59, thru=13.3, I=16.7, T=32.3, wait=3.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:57:03,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:57:04,288] loop.py:45 INFO TRAIN_STEP step=6189 loss=1.333678 Training: 34%|███▍ | 6189/18018 [13:26:50<27:10:47, 8.27s/step, loss=1.334, acc=0.699, acc_len=1.69, thru=16.0, I=16.3, T=30.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:57:04,631] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:57:11,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:57:12,768] loop.py:380 INFO TIMING step=6190: step=4.207s data=1.968s compute=4.166s [fwd=2.963s bwd=1.188s opt=0.003s] dispatch=4.264s [2026-10-19 02:57:12,768] loop.py:45 INFO TRAIN_STEP step=6190 loss=1.485164 Training: 34%|███▍ | 6190/18018 [13:26:58<27:22:55, 8.33s/step, loss=1.485, acc=0.673, acc_len=1.59, thru=14.0, I=16.9, T=30.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:57:13,207] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:57:19,531] loop.py:45 INFO TRAIN_STEP step=6191 loss=1.420984 Training: 34%|███▍ | 6191/18018 [13:27:05<25:49:55, 7.86s/step, loss=1.421, acc=0.684, acc_len=1.63, thru=16.0, I=18.9, T=30.4, wait=2.5s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:57:19,526] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6190: forward=2953.9ms backward=1210.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:57:26,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:57:28,956] loop.py:45 INFO TRAIN_STEP step=6192 loss=1.377191 Training: 34%|███▍ | 6192/18018 [13:27:15<27:22:10, 8.33s/step, loss=1.377, acc=0.694, acc_len=1.66, thru=16.1, I=17.9, T=27.2, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:57:28,986] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 02:57:35,656] loop.py:45 INFO TRAIN_STEP step=6193 loss=1.396238 Training: 34%|███▍ | 6193/18018 [13:27:21<25:45:29, 7.84s/step, loss=1.396, acc=0.685, acc_len=1.63, thru=14.2, I=17.5, T=31.8, wait=2.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:57:44,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:57:46,057] loop.py:45 INFO TRAIN_STEP step=6194 loss=1.490162 Training: 34%|███▍ | 6194/18018 [13:27:32<28:16:43, 8.61s/step, loss=1.490, acc=0.667, acc_len=1.58, thru=15.1, I=16.5, T=29.7, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:57:46,350] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:57:53,462] loop.py:380 INFO TIMING step=6195: step=4.917s data=2.367s compute=4.890s [fwd=3.329s bwd=1.546s opt=0.003s] dispatch=2.478s [2026-10-19 02:57:53,462] loop.py:45 INFO TRAIN_STEP step=6195 loss=1.477080 Training: 34%|███▍ | 6195/18018 [13:27:39<27:05:20, 8.25s/step, loss=1.477, acc=0.671, acc_len=1.59, thru=15.7, I=20.0, T=26.0, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:00,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:58:01,146] loop.py:45 INFO TRAIN_STEP step=6196 loss=1.484535 Training: 34%|███▍ | 6196/18018 [13:27:47<26:31:51, 8.08s/step, loss=1.485, acc=0.671, acc_len=1.59, thru=16.0, I=15.6, T=31.7, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:58:01,141] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6195: forward=2875.7ms backward=1101.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:01,292] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:05,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:58:06,761] loop.py:45 INFO TRAIN_STEP step=6197 loss=1.421373 Training: 34%|███▍ | 6197/18018 [13:27:52<24:06:06, 7.34s/step, loss=1.421, acc=0.680, acc_len=1.61, thru=20.6, I=24.5, T=37.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:06,793] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 02:58:12,978] loop.py:45 INFO TRAIN_STEP step=6198 loss=1.488945 Training: 34%|███▍ | 6198/18018 [13:27:59<22:59:34, 7.00s/step, loss=1.489, acc=0.666, acc_len=1.57, thru=17.4, I=20.1, T=33.7, wait=2.4s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:20,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:58:21,983] loop.py:45 INFO TRAIN_STEP step=6199 loss=1.530518 Training: 34%|███▍ | 6199/18018 [13:28:08<24:57:44, 7.60s/step, loss=1.531, acc=0.664, acc_len=1.57, thru=16.9, I=17.2, T=26.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:22,483] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 02:58:29,448] loop.py:380 INFO TIMING step=6200: step=4.701s data=2.184s compute=4.660s [fwd=3.227s bwd=1.419s opt=0.003s] dispatch=2.755s [2026-10-19 02:58:29,449] loop.py:45 INFO TRAIN_STEP step=6200 loss=1.536206 Training: 34%|███▍ | 6200/18018 [13:28:15<24:49:28, 7.56s/step, loss=1.536, acc=0.662, acc_len=1.56, thru=14.3, I=18.6, T=27.2, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:36,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:58:38,834] loop.py:45 INFO TRAIN_STEP step=6201 loss=1.496137 Training: 34%|███▍ | 6201/18018 [13:28:24<26:37:14, 8.11s/step, loss=1.496, acc=0.668, acc_len=1.57, thru=14.4, I=17.3, T=26.0, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:58:38,829] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6200: forward=3175.4ms backward=1682.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:38,990] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:44,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:58:46,114] loop.py:45 INFO TRAIN_STEP step=6202 loss=1.577435 Training: 34%|███▍ | 6202/18018 [13:28:32<25:48:00, 7.86s/step, loss=1.577, acc=0.655, acc_len=1.54, thru=14.4, I=21.1, T=31.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:58:46,270] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:58:53,884] loop.py:45 INFO TRAIN_STEP step=6203 loss=1.470623 Training: 34%|███▍ | 6203/18018 [13:28:39<25:42:26, 7.83s/step, loss=1.471, acc=0.675, acc_len=1.60, thru=12.7, I=14.8, T=31.4, wait=3.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:02,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:59:02,857] loop.py:45 INFO TRAIN_STEP step=6204 loss=1.516906 Training: 34%|███▍ | 6204/18018 [13:28:48<26:49:44, 8.18s/step, loss=1.517, acc=0.664, acc_len=1.57, thru=15.3, I=15.7, T=28.1, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:03,051] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:08,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:59:10,021] loop.py:380 INFO TIMING step=6205: step=3.901s data=1.772s compute=3.864s [fwd=2.718s bwd=1.132s opt=0.003s] dispatch=3.253s [2026-10-19 02:59:10,021] loop.py:45 INFO TRAIN_STEP step=6205 loss=1.526851 Training: 34%|███▍ | 6205/18018 [13:28:56<25:49:53, 7.87s/step, loss=1.527, acc=0.662, acc_len=1.56, thru=17.1, I=19.3, T=32.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:10,063] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:15,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:59:16,895] loop.py:45 INFO TRAIN_STEP step=6206 loss=1.474308 Training: 34%|███▍ | 6206/18018 [13:29:02<24:50:44, 7.57s/step, loss=1.474, acc=0.674, acc_len=1.60, thru=17.5, I=23.0, T=27.9, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:59:16,890] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6205: forward=3163.5ms backward=1380.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:17,027] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:21,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:59:24,088] loop.py:45 INFO TRAIN_STEP step=6207 loss=1.535279 Training: 34%|███▍ | 6207/18018 [13:29:10<24:28:18, 7.46s/step, loss=1.535, acc=0.664, acc_len=1.57, thru=15.7, I=22.4, T=27.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:24,512] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 02:59:31,742] loop.py:45 INFO TRAIN_STEP step=6208 loss=1.536744 Training: 34%|███▍ | 6208/18018 [13:29:17<24:39:39, 7.52s/step, loss=1.537, acc=0.663, acc_len=1.56, thru=12.6, I=16.5, T=29.2, wait=3.3s, pool=52, epoch=1/2][2026-10-19 02:59:39,343] loop.py:45 INFO TRAIN_STEP step=6209 loss=1.414897 Training: 34%|███▍ | 6209/18018 [13:29:25<24:44:40, 7.54s/step, loss=1.415, acc=0.682, acc_len=1.62, thru=17.0, I=16.7, T=30.0, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:46,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 02:59:48,180] loop.py:380 INFO TIMING step=6210: step=4.558s data=2.133s compute=4.521s [fwd=3.224s bwd=1.281s opt=0.003s] dispatch=4.265s [2026-10-19 02:59:48,180] loop.py:45 INFO TRAIN_STEP step=6210 loss=1.562843 Training: 34%|███▍ | 6210/18018 [13:29:34<26:00:38, 7.93s/step, loss=1.563, acc=0.658, acc_len=1.55, thru=14.7, I=15.5, T=28.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 02:59:48,393] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 02:59:55,417] loop.py:45 INFO TRAIN_STEP step=6211 loss=1.620315 Training: 34%|███▍ | 6211/18018 [13:29:41<25:19:45, 7.72s/step, loss=1.620, acc=0.648, acc_len=1.51, thru=14.8, I=17.8, T=32.8, wait=3.3s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 02:59:55,413] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6210: forward=2688.4ms backward=1162.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:02,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:00:03,319] loop.py:45 INFO TRAIN_STEP step=6212 loss=1.440837 Training: 34%|███▍ | 6212/18018 [13:29:49<25:30:10, 7.78s/step, loss=1.441, acc=0.681, acc_len=1.63, thru=16.9, I=16.9, T=30.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:03,590] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:09,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:00:10,867] loop.py:45 INFO TRAIN_STEP step=6213 loss=1.526031 Training: 34%|███▍ | 6213/18018 [13:29:56<25:16:30, 7.71s/step, loss=1.526, acc=0.665, acc_len=1.57, thru=15.4, I=18.9, T=27.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:11,166] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:16,968] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:00:19,353] loop.py:45 INFO TRAIN_STEP step=6214 loss=1.431444 Training: 34%|███▍ | 6214/18018 [13:30:05<26:02:21, 7.94s/step, loss=1.431, acc=0.677, acc_len=1.61, thru=13.8, I=19.5, T=22.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:19,496] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=16/64 [2026-10-19 03:00:26,568] loop.py:380 INFO TIMING step=6215: step=4.470s data=2.063s compute=4.441s [fwd=3.099s bwd=1.327s opt=0.003s] dispatch=2.733s [2026-10-19 03:00:26,568] loop.py:45 INFO TRAIN_STEP step=6215 loss=1.572562 Training: 34%|███▍ | 6215/18018 [13:30:12<25:19:14, 7.72s/step, loss=1.573, acc=0.658, acc_len=1.55, thru=12.9, I=17.2, T=28.6, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:33,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 03:00:35,307] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6215: forward=2722.6ms backward=1281.1ms [2026-10-19 03:00:35,312] loop.py:45 INFO TRAIN_STEP step=6216 loss=1.622827 Training: 34%|███▍ | 6216/18018 [13:30:21<26:19:27, 8.03s/step, loss=1.623, acc=0.647, acc_len=1.50, thru=15.2, I=16.2, T=31.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:35,376] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:41,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:00:43,100] loop.py:45 INFO TRAIN_STEP step=6217 loss=1.579042 Training: 35%|███▍ | 6217/18018 [13:30:29<26:04:54, 7.96s/step, loss=1.579, acc=0.655, acc_len=1.54, thru=15.4, I=20.3, T=28.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:43,330] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:50,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:00:52,201] loop.py:45 INFO TRAIN_STEP step=6218 loss=1.489253 Training: 35%|███▍ | 6218/18018 [13:30:38<27:12:22, 8.30s/step, loss=1.489, acc=0.675, acc_len=1.60, thru=14.4, I=16.3, T=23.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:52,467] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:00:59,659] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:01:00,540] loop.py:45 INFO TRAIN_STEP step=6219 loss=1.469149 Training: 35%|███▍ | 6219/18018 [13:30:46<27:14:29, 8.31s/step, loss=1.469, acc=0.670, acc_len=1.58, thru=15.9, I=17.5, T=32.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:00,678] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-19 03:01:08,351] loop.py:380 INFO TIMING step=6220: step=5.030s data=1.996s compute=4.994s [fwd=3.079s bwd=1.899s opt=0.003s] dispatch=2.772s [2026-10-19 03:01:08,352] loop.py:45 INFO TRAIN_STEP step=6220 loss=1.414990 Training: 35%|███▍ | 6220/18018 [13:30:54<26:44:47, 8.16s/step, loss=1.415, acc=0.684, acc_len=1.63, thru=15.2, I=15.9, T=25.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:08,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:08,856] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:01:17,040] loop.py:45 INFO TRAIN_STEP step=6221 loss=1.533419 Training: 35%|███▍ | 6221/18018 [13:31:03<27:15:48, 8.32s/step, loss=1.533, acc=0.666, acc_len=1.57, thru=14.5, I=14.1, T=28.1, wait=4.1s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:01:17,035] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6220: forward=3243.6ms backward=1264.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:23,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:01:26,392] loop.py:45 INFO TRAIN_STEP step=6222 loss=1.393283 Training: 35%|███▍ | 6222/18018 [13:31:12<28:16:38, 8.63s/step, loss=1.393, acc=0.687, acc_len=1.64, thru=15.3, I=20.6, T=25.6, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:26,545] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:31,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:01:33,062] loop.py:45 INFO TRAIN_STEP step=6223 loss=1.413804 Training: 35%|███▍ | 6223/18018 [13:31:19<26:20:48, 8.04s/step, loss=1.414, acc=0.680, acc_len=1.62, thru=15.0, I=26.6, T=28.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:33,428] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:38,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:01:39,841] loop.py:45 INFO TRAIN_STEP step=6224 loss=1.644698 Training: 35%|███▍ | 6224/18018 [13:31:25<25:06:12, 7.66s/step, loss=1.645, acc=0.642, acc_len=1.49, thru=16.3, I=23.2, T=32.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:40,344] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:47,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:01:48,772] loop.py:380 INFO TIMING step=6225: step=4.798s data=2.304s compute=4.756s [fwd=3.328s bwd=1.412s opt=0.003s] dispatch=4.124s [2026-10-19 03:01:48,772] loop.py:45 INFO TRAIN_STEP step=6225 loss=1.431274 Training: 35%|███▍ | 6225/18018 [13:31:34<26:20:58, 8.04s/step, loss=1.431, acc=0.681, acc_len=1.62, thru=13.7, I=16.0, T=26.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:48,921] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:57,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:01:59,342] loop.py:45 INFO TRAIN_STEP step=6226 loss=1.561182 Training: 35%|███▍ | 6226/18018 [13:31:45<28:49:40, 8.80s/step, loss=1.561, acc=0.656, acc_len=1.54, thru=12.5, I=13.0, T=28.7, wait=6.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:01:59,337] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6225: forward=3128.2ms backward=1270.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:01:59,695] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:06,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:02:06,905] loop.py:45 INFO TRAIN_STEP step=6227 loss=1.530137 Training: 35%|███▍ | 6227/18018 [13:31:52<27:36:34, 8.43s/step, loss=1.530, acc=0.664, acc_len=1.57, thru=14.8, I=17.4, T=30.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:07,066] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:02:14,559] loop.py:45 INFO TRAIN_STEP step=6228 loss=1.410897 Training: 35%|███▍ | 6228/18018 [13:32:00<26:50:42, 8.20s/step, loss=1.411, acc=0.682, acc_len=1.63, thru=12.7, I=13.7, T=31.4, wait=3.6s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:22,009] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:02:22,641] loop.py:45 INFO TRAIN_STEP step=6229 loss=1.479307 Training: 35%|███▍ | 6229/18018 [13:32:08<26:43:47, 8.16s/step, loss=1.479, acc=0.671, acc_len=1.59, thru=17.9, I=17.0, T=29.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:23,014] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:28,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:02:30,435] loop.py:380 INFO TIMING step=6230: step=4.164s data=1.850s compute=4.131s [fwd=2.889s bwd=1.226s opt=0.003s] dispatch=3.620s [2026-10-19 03:02:30,435] loop.py:45 INFO TRAIN_STEP step=6230 loss=1.584432 Training: 35%|███▍ | 6230/18018 [13:32:16<26:21:59, 8.05s/step, loss=1.584, acc=0.658, acc_len=1.54, thru=16.5, I=20.8, T=30.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:30,792] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:37,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:02:38,384] loop.py:45 INFO TRAIN_STEP step=6231 loss=1.477234 Training: 35%|███▍ | 6231/18018 [13:32:24<26:15:42, 8.02s/step, loss=1.477, acc=0.673, acc_len=1.59, thru=14.1, I=20.8, T=29.5, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:02:38,378] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6230: forward=3018.2ms backward=1267.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:38,809] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:44,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:02:46,622] loop.py:45 INFO TRAIN_STEP step=6232 loss=1.601865 Training: 35%|███▍ | 6232/18018 [13:32:32<26:28:28, 8.09s/step, loss=1.602, acc=0.652, acc_len=1.53, thru=14.9, I=17.9, T=26.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:46,679] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:55,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:02:56,954] loop.py:45 INFO TRAIN_STEP step=6233 loss=1.452684 Training: 35%|███▍ | 6233/18018 [13:32:43<28:40:39, 8.76s/step, loss=1.453, acc=0.681, acc_len=1.62, thru=11.7, I=14.0, T=26.8, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:02:57,101] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:01,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:03:02,944] loop.py:45 INFO TRAIN_STEP step=6234 loss=1.422806 Training: 35%|███▍ | 6234/18018 [13:32:49<25:57:10, 7.93s/step, loss=1.423, acc=0.683, acc_len=1.63, thru=18.0, I=21.8, T=31.2, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:03,435] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:09,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:03:09,992] loop.py:380 INFO TIMING step=6235: step=4.424s data=2.139s compute=4.386s [fwd=3.105s bwd=1.265s opt=0.003s] dispatch=2.615s [2026-10-19 03:03:09,992] loop.py:45 INFO TRAIN_STEP step=6235 loss=1.491921 Training: 35%|███▍ | 6235/18018 [13:32:56<25:05:09, 7.66s/step, loss=1.492, acc=0.675, acc_len=1.59, thru=16.8, I=20.7, T=28.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:10,251] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:16,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:03:18,259] loop.py:45 INFO TRAIN_STEP step=6236 loss=1.536666 Training: 35%|███▍ | 6236/18018 [13:33:04<25:40:37, 7.85s/step, loss=1.537, acc=0.661, acc_len=1.56, thru=15.3, I=17.9, T=27.3, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:03:18,254] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6235: forward=3257.3ms backward=1385.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:18,528] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:03:24,451] loop.py:45 INFO TRAIN_STEP step=6237 loss=1.390117 Training: 35%|███▍ | 6237/18018 [13:33:10<24:03:00, 7.35s/step, loss=1.390, acc=0.691, acc_len=1.66, thru=16.4, I=21.4, T=33.5, wait=2.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:34,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:03:35,593] loop.py:45 INFO TRAIN_STEP step=6238 loss=1.476242 Training: 35%|███▍ | 6238/18018 [13:33:21<27:46:23, 8.49s/step, loss=1.476, acc=0.668, acc_len=1.58, thru=15.5, I=14.0, T=29.4, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:35,628] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-19 03:03:42,389] loop.py:45 INFO TRAIN_STEP step=6239 loss=1.445629 Training: 35%|███▍ | 6239/18018 [13:33:28<26:06:31, 7.98s/step, loss=1.446, acc=0.674, acc_len=1.60, thru=15.7, I=16.6, T=31.3, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:49,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:03:51,695] loop.py:380 INFO TIMING step=6240: step=4.439s data=2.023s compute=4.400s [fwd=3.061s bwd=1.324s opt=0.003s] dispatch=4.857s [2026-10-19 03:03:51,696] loop.py:45 INFO TRAIN_STEP step=6240 loss=1.382601 Training: 35%|███▍ | 6240/18018 [13:33:37<27:24:26, 8.38s/step, loss=1.383, acc=0.685, acc_len=1.63, thru=16.1, I=18.4, T=28.8, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:03:52,188] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:00,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:04:01,172] loop.py:45 INFO TRAIN_STEP step=6241 loss=1.490189 Training: 35%|███▍ | 6241/18018 [13:33:47<28:29:04, 8.71s/step, loss=1.490, acc=0.669, acc_len=1.59, thru=14.2, I=14.8, T=34.1, wait=5.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:04:01,167] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6240: forward=2587.4ms backward=1120.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:01,287] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 03:04:07,732] loop.py:45 INFO TRAIN_STEP step=6242 loss=1.540677 Training: 35%|███▍ | 6242/18018 [13:33:53<26:22:38, 8.06s/step, loss=1.541, acc=0.659, acc_len=1.55, thru=19.0, I=18.2, T=31.0, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:15,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:04:15,778] loop.py:45 INFO TRAIN_STEP step=6243 loss=1.438392 Training: 35%|███▍ | 6243/18018 [13:34:01<26:21:25, 8.06s/step, loss=1.438, acc=0.680, acc_len=1.62, thru=15.7, I=16.0, T=27.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:16,095] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:21,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:04:23,557] loop.py:45 INFO TRAIN_STEP step=6244 loss=1.571154 Training: 35%|███▍ | 6244/18018 [13:34:09<26:04:53, 7.97s/step, loss=1.571, acc=0.656, acc_len=1.53, thru=17.2, I=17.4, T=28.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:23,931] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:29,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:04:30,226] loop.py:380 INFO TIMING step=6245: step=4.064s data=1.868s compute=4.033s [fwd=2.848s bwd=1.169s opt=0.003s] dispatch=2.595s [2026-10-19 03:04:30,226] loop.py:45 INFO TRAIN_STEP step=6245 loss=1.506658 Training: 35%|███▍ | 6245/18018 [13:34:16<24:47:52, 7.58s/step, loss=1.507, acc=0.665, acc_len=1.57, thru=17.1, I=24.4, T=31.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:30,613] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (TrainerActor pid=9310) [2026-10-19 03:04:38,629] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6245: forward=2683.2ms backward=1195.1ms [2026-10-19 03:04:38,634] loop.py:45 INFO TRAIN_STEP step=6246 loss=1.614068 Training: 35%|███▍ | 6246/18018 [13:34:24<25:36:12, 7.83s/step, loss=1.614, acc=0.649, acc_len=1.52, thru=12.4, I=13.3, T=32.4, wait=4.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:43,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:04:45,996] loop.py:45 INFO TRAIN_STEP step=6247 loss=1.590445 Training: 35%|███▍ | 6247/18018 [13:34:32<25:08:32, 7.69s/step, loss=1.590, acc=0.659, acc_len=1.54, thru=17.8, I=21.1, T=29.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:46,299] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:51,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:04:52,045] loop.py:45 INFO TRAIN_STEP step=6248 loss=1.487317 Training: 35%|███▍ | 6248/18018 [13:34:38<23:31:54, 7.20s/step, loss=1.487, acc=0.673, acc_len=1.60, thru=18.2, I=27.0, T=33.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:04:52,281] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 03:04:58,126] loop.py:45 INFO TRAIN_STEP step=6249 loss=1.486170 Training: 35%|███▍ | 6249/18018 [13:34:44<22:26:08, 6.86s/step, loss=1.486, acc=0.669, acc_len=1.58, thru=18.9, I=22.5, T=32.7, wait=2.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:05:04,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:05:05,533] loop.py:380 INFO TIMING step=6250: step=4.020s data=1.831s compute=3.985s [fwd=2.857s bwd=1.113s opt=0.003s] dispatch=3.378s [2026-10-19 03:05:05,533] loop.py:45 INFO TRAIN_STEP step=6250 loss=1.597989 Training: 35%|███▍ | 6250/18018 [13:34:51<22:58:06, 7.03s/step, loss=1.598, acc=0.648, acc_len=1.52, thru=17.5, I=19.5, T=31.8, wait=3.4s, pool=64, epoch=1/2][2026-10-19 03:05:05,538] loop.py:419 INFO Saving checkpoint at step 6250... (TrainerActor pid=9310) [2026-10-19 03:05:09,703] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0006251 (TrainerActor pid=9310) [2026-10-19 03:05:09,711] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0005501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:05:09,919] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:05:10,924] inference_manager.py:408 INFO Pool capacity available after 6.0s, pool_size=0/64 [2026-10-19 03:05:16,858] loop.py:45 INFO TRAIN_STEP step=6251 loss=1.497311 Training: 35%|███▍ | 6251/18018 [13:35:02<27:10:52, 8.32s/step, loss=1.497, acc=0.667, acc_len=1.57, thru=18.2, I=27.1, T=32.5, wait=2.2s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:05:16,853] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6250: forward=2736.8ms backward=1163.8ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 03:05:09,704] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0006251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 03:05:10,718] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0005501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:05:25,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:05:27,862] loop.py:45 INFO TRAIN_STEP step=6252 loss=1.293486 Training: 35%|███▍ | 6252/18018 [13:35:13<29:48:47, 9.12s/step, loss=1.293, acc=0.703, acc_len=1.69, thru=14.1, I=14.5, T=23.3, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:05:28,040] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:05:35,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:05:36,354] loop.py:45 INFO TRAIN_STEP step=6253 loss=1.405759 Training: 35%|███▍ | 6253/18018 [13:35:22<29:11:42, 8.93s/step, loss=1.406, acc=0.684, acc_len=1.64, thru=15.4, I=15.6, T=23.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:05:36,819] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:05:43,302] loop.py:45 INFO TRAIN_STEP step=6254 loss=1.609100 Training: 35%|███▍ | 6254/18018 [13:35:29<27:14:36, 8.34s/step, loss=1.609, acc=0.649, acc_len=1.52, thru=16.5, I=17.4, T=29.4, wait=2.6s, pool=52, epoch=1/2][2026-10-19 03:05:52,008] loop.py:380 INFO TIMING step=6255: step=4.573s data=2.194s compute=4.546s [fwd=3.180s bwd=1.352s opt=0.003s] dispatch=4.124s [2026-10-19 03:05:52,008] loop.py:45 INFO TRAIN_STEP step=6255 loss=1.457947 Training: 35%|███▍ | 6255/18018 [13:35:38<27:36:11, 8.45s/step, loss=1.458, acc=0.674, acc_len=1.60, thru=14.4, I=12.9, T=28.0, wait=4.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:05:57,753] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:05:59,681] loop.py:45 INFO TRAIN_STEP step=6256 loss=1.595663 Training: 35%|███▍ | 6256/18018 [13:35:45<26:50:37, 8.22s/step, loss=1.596, acc=0.655, acc_len=1.54, thru=17.0, I=19.6, T=31.7, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:05:59,675] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6255: forward=2766.2ms backward=1220.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:05:59,785] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-19 03:06:07,145] loop.py:45 INFO TRAIN_STEP step=6257 loss=1.491307 Training: 35%|███▍ | 6257/18018 [13:35:53<26:06:08, 7.99s/step, loss=1.491, acc=0.670, acc_len=1.59, thru=14.2, I=19.1, T=34.2, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:06:17,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:06:19,839] loop.py:45 INFO TRAIN_STEP step=6258 loss=1.480393 Training: 35%|███▍ | 6258/18018 [13:36:05<30:42:34, 9.40s/step, loss=1.480, acc=0.671, acc_len=1.58, thru=11.4, I=12.2, T=26.2, wait=7.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:06:20,297] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:06:26,831] loop.py:45 INFO TRAIN_STEP step=6259 loss=1.423928 Training: 35%|███▍ | 6259/18018 [13:36:12<28:20:46, 8.68s/step, loss=1.424, acc=0.679, acc_len=1.61, thru=14.0, I=19.4, T=33.0, wait=3.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:06:33,441] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:06:35,582] loop.py:380 INFO TIMING step=6260: step=4.954s data=2.465s compute=4.924s [fwd=3.454s bwd=1.455s opt=0.003s] dispatch=3.788s [2026-10-19 03:06:35,582] loop.py:45 INFO TRAIN_STEP step=6260 loss=1.487660 Training: 35%|███▍ | 6260/18018 [13:36:21<28:25:03, 8.70s/step, loss=1.488, acc=0.669, acc_len=1.58, thru=16.6, I=18.7, T=25.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:06:35,957] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:06:41,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:06:41,725] loop.py:45 INFO TRAIN_STEP step=6261 loss=1.581642 Training: 35%|███▍ | 6261/18018 [13:36:27<25:54:29, 7.93s/step, loss=1.582, acc=0.655, acc_len=1.54, thru=16.5, I=23.6, T=33.8, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:06:41,720] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6260: forward=2675.3ms backward=1062.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:06:42,072] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:06:48,002] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:06:48,265] loop.py:45 INFO TRAIN_STEP step=6262 loss=1.367256 Training: 35%|███▍ | 6262/18018 [13:36:34<24:32:28, 7.52s/step, loss=1.367, acc=0.690, acc_len=1.65, thru=18.4, I=20.8, T=33.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:06:48,504] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:06:58,009] loop.py:45 INFO TRAIN_STEP step=6263 loss=1.646127 Training: 35%|███▍ | 6263/18018 [13:36:44<26:43:19, 8.18s/step, loss=1.646, acc=0.646, acc_len=1.50, thru=12.2, I=12.2, T=27.9, wait=5.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:05,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:07:07,609] loop.py:45 INFO TRAIN_STEP step=6264 loss=1.713046 Training: 35%|███▍ | 6264/18018 [13:36:53<28:06:27, 8.61s/step, loss=1.713, acc=0.636, acc_len=1.48, thru=14.7, I=16.7, T=26.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:08,058] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:12,744] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:07:14,835] loop.py:380 INFO TIMING step=6265: step=4.463s data=2.035s compute=4.426s [fwd=3.067s bwd=1.343s opt=0.003s] dispatch=2.753s [2026-10-19 03:07:14,835] loop.py:45 INFO TRAIN_STEP step=6265 loss=1.622163 Training: 35%|███▍ | 6265/18018 [13:37:00<26:45:00, 8.19s/step, loss=1.622, acc=0.649, acc_len=1.51, thru=14.4, I=22.7, T=28.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:15,257] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:19,314] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:07:21,392] loop.py:45 INFO TRAIN_STEP step=6266 loss=1.543181 Training: 35%|███▍ | 6266/18018 [13:37:07<25:08:45, 7.70s/step, loss=1.543, acc=0.659, acc_len=1.54, thru=17.2, I=28.4, T=29.4, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:07:21,387] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6265: forward=3024.6ms backward=1283.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:21,837] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:28,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:07:29,000] loop.py:45 INFO TRAIN_STEP step=6267 loss=1.476352 Training: 35%|███▍ | 6267/18018 [13:37:15<25:02:56, 7.67s/step, loss=1.476, acc=0.674, acc_len=1.60, thru=14.1, I=20.4, T=26.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:29,365] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:34,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:07:37,767] loop.py:45 INFO TRAIN_STEP step=6268 loss=1.560904 Training: 35%|███▍ | 6268/18018 [13:37:23<26:07:12, 8.00s/step, loss=1.561, acc=0.660, acc_len=1.55, thru=15.1, I=22.4, T=21.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:37,957] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:43,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:07:44,952] loop.py:45 INFO TRAIN_STEP step=6269 loss=1.416280 Training: 35%|███▍ | 6269/18018 [13:37:31<25:19:03, 7.76s/step, loss=1.416, acc=0.680, acc_len=1.62, thru=18.3, I=22.1, T=25.4, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:45,170] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:50,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:07:51,913] loop.py:380 INFO TIMING step=6270: step=4.277s data=2.068s compute=4.241s [fwd=2.997s bwd=1.229s opt=0.003s] dispatch=2.673s [2026-10-19 03:07:51,914] loop.py:45 INFO TRAIN_STEP step=6270 loss=1.355078 Training: 35%|███▍ | 6270/18018 [13:37:37<24:32:01, 7.52s/step, loss=1.355, acc=0.695, acc_len=1.67, thru=15.3, I=19.4, T=29.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:07:51,959] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:00,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:08:00,395] loop.py:45 INFO TRAIN_STEP step=6271 loss=1.411929 Training: 35%|███▍ | 6271/18018 [13:37:46<25:28:26, 7.81s/step, loss=1.412, acc=0.683, acc_len=1.63, thru=13.3, I=14.4, T=29.2, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:08:00,390] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6270: forward=2968.6ms backward=1325.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:00,823] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:06,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:08:07,693] loop.py:45 INFO TRAIN_STEP step=6272 loss=1.426930 Training: 35%|███▍ | 6272/18018 [13:37:53<24:58:28, 7.65s/step, loss=1.427, acc=0.682, acc_len=1.62, thru=18.3, I=18.7, T=26.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:07,956] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:14,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:08:15,405] loop.py:45 INFO TRAIN_STEP step=6273 loss=1.636131 Training: 35%|███▍ | 6273/18018 [13:38:01<25:01:44, 7.67s/step, loss=1.636, acc=0.648, acc_len=1.52, thru=15.0, I=16.5, T=29.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:15,827] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:21,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:08:23,811] loop.py:45 INFO TRAIN_STEP step=6274 loss=1.569564 Training: 35%|███▍ | 6274/18018 [13:38:09<25:44:51, 7.89s/step, loss=1.570, acc=0.657, acc_len=1.55, thru=14.0, I=20.5, T=26.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:23,979] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:31,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:08:31,624] loop.py:380 INFO TIMING step=6275: step=4.158s data=1.918s compute=4.127s [fwd=2.935s bwd=1.177s opt=0.003s] dispatch=3.645s [2026-10-19 03:08:31,624] loop.py:45 INFO TRAIN_STEP step=6275 loss=1.432832 Training: 35%|███▍ | 6275/18018 [13:38:17<25:40:05, 7.87s/step, loss=1.433, acc=0.680, acc_len=1.62, thru=14.0, I=16.6, T=30.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:31,979] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:38,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:08:38,978] loop.py:45 INFO TRAIN_STEP step=6276 loss=1.515294 Training: 35%|███▍ | 6276/18018 [13:38:25<25:09:37, 7.71s/step, loss=1.515, acc=0.664, acc_len=1.57, thru=17.2, I=17.4, T=31.6, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:08:38,973] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6275: forward=2815.7ms backward=1190.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:39,337] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:44,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:08:45,424] loop.py:45 INFO TRAIN_STEP step=6277 loss=1.478501 Training: 35%|███▍ | 6277/18018 [13:38:31<23:55:05, 7.33s/step, loss=1.479, acc=0.672, acc_len=1.59, thru=18.8, I=21.2, T=35.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:45,705] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:51,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:08:52,015] loop.py:45 INFO TRAIN_STEP step=6278 loss=1.413883 Training: 35%|███▍ | 6278/18018 [13:38:38<23:11:17, 7.11s/step, loss=1.414, acc=0.681, acc_len=1.62, thru=19.3, I=21.7, T=32.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:52,220] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:08:58,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:09:00,860] loop.py:45 INFO TRAIN_STEP step=6279 loss=1.376784 Training: 35%|███▍ | 6279/18018 [13:38:46<24:52:53, 7.63s/step, loss=1.377, acc=0.689, acc_len=1.65, thru=14.8, I=18.6, T=26.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:01,304] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:07,171] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:09:08,511] loop.py:380 INFO TIMING step=6280: step=4.447s data=2.103s compute=4.412s [fwd=3.031s bwd=1.366s opt=0.003s] dispatch=3.196s [2026-10-19 03:09:08,511] loop.py:45 INFO TRAIN_STEP step=6280 loss=1.476579 Training: 35%|███▍ | 6280/18018 [13:38:54<24:54:08, 7.64s/step, loss=1.477, acc=0.674, acc_len=1.60, thru=17.8, I=22.1, T=28.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:08,680] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:11,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:09:14,360] loop.py:45 INFO TRAIN_STEP step=6281 loss=1.538875 Training: 35%|███▍ | 6281/18018 [13:39:00<23:09:03, 7.10s/step, loss=1.539, acc=0.662, acc_len=1.55, thru=18.9, I=28.3, T=31.1, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:14,415] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (TrainerActor pid=9310) [2026-10-19 03:09:14,355] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6280: forward=2800.2ms backward=1263.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:19,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:09:20,787] loop.py:45 INFO TRAIN_STEP step=6282 loss=1.490123 Training: 35%|███▍ | 6282/18018 [13:39:06<22:29:21, 6.90s/step, loss=1.490, acc=0.666, acc_len=1.57, thru=17.0, I=24.6, T=33.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:20,988] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:25,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:09:26,489] loop.py:45 INFO TRAIN_STEP step=6283 loss=1.492719 Training: 35%|███▍ | 6283/18018 [13:39:12<21:18:56, 6.54s/step, loss=1.493, acc=0.671, acc_len=1.59, thru=21.1, I=24.3, T=34.7, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:26,570] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:33,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:09:34,941] loop.py:45 INFO TRAIN_STEP step=6284 loss=1.588321 Training: 35%|███▍ | 6284/18018 [13:39:21<23:11:04, 7.11s/step, loss=1.588, acc=0.654, acc_len=1.53, thru=14.9, I=19.4, T=27.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:35,289] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:09:42,046] loop.py:380 INFO TIMING step=6285: step=4.110s data=1.914s compute=4.081s [fwd=2.951s bwd=1.116s opt=0.003s] dispatch=2.986s [2026-10-19 03:09:42,047] loop.py:45 INFO TRAIN_STEP step=6285 loss=1.372849 Training: 35%|███▍ | 6285/18018 [13:39:28<23:10:32, 7.11s/step, loss=1.373, acc=0.688, acc_len=1.64, thru=14.9, I=17.7, T=31.1, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:49,002] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:09:50,945] loop.py:45 INFO TRAIN_STEP step=6286 loss=1.666824 Training: 35%|███▍ | 6286/18018 [13:39:37<24:55:17, 7.65s/step, loss=1.667, acc=0.646, acc_len=1.50, thru=14.9, I=16.6, T=25.4, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:09:50,940] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6285: forward=3457.6ms backward=1521.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:51,018] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:56,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:09:57,432] loop.py:45 INFO TRAIN_STEP step=6287 loss=1.540924 Training: 35%|███▍ | 6287/18018 [13:39:43<23:47:04, 7.30s/step, loss=1.541, acc=0.663, acc_len=1.56, thru=16.0, I=20.6, T=29.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:09:57,567] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:10:03,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:10:04,707] loop.py:45 INFO TRAIN_STEP step=6288 loss=1.411295 Training: 35%|███▍ | 6288/18018 [13:39:50<23:45:44, 7.29s/step, loss=1.411, acc=0.683, acc_len=1.63, thru=15.7, I=18.3, T=27.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:10:04,784] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-19 03:10:13,195] loop.py:45 INFO TRAIN_STEP step=6289 loss=1.435780 Training: 35%|███▍ | 6289/18018 [13:39:59<24:55:30, 7.65s/step, loss=1.436, acc=0.677, acc_len=1.61, thru=13.2, I=14.9, T=33.1, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:10:18,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:10:20,943] loop.py:380 INFO TIMING step=6290: step=4.638s data=2.328s compute=4.590s [fwd=3.142s bwd=1.433s opt=0.003s] dispatch=3.102s [2026-10-19 03:10:20,943] loop.py:45 INFO TRAIN_STEP step=6290 loss=1.622561 Training: 35%|███▍ | 6290/18018 [13:40:07<25:01:06, 7.68s/step, loss=1.623, acc=0.654, acc_len=1.53, thru=17.1, I=20.6, T=27.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:10:21,287] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:10:27,158] loop.py:45 INFO TRAIN_STEP step=6291 loss=1.445801 Training: 35%|███▍ | 6291/18018 [13:40:13<23:35:08, 7.24s/step, loss=1.446, acc=0.677, acc_len=1.61, thru=17.2, I=22.2, T=32.0, wait=2.2s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:10:27,153] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6290: forward=2811.7ms backward=1144.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:10:34,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:10:35,844] loop.py:45 INFO TRAIN_STEP step=6292 loss=1.385560 Training: 35%|███▍ | 6292/18018 [13:40:21<24:59:53, 7.67s/step, loss=1.386, acc=0.687, acc_len=1.64, thru=16.2, I=17.0, T=28.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:10:36,224] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:10:44,735] loop.py:45 INFO TRAIN_STEP step=6293 loss=1.386452 Training: 35%|███▍ | 6293/18018 [13:40:30<26:11:02, 8.04s/step, loss=1.386, acc=0.686, acc_len=1.65, thru=13.2, I=14.4, T=28.0, wait=4.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:10:54,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:10:55,812] loop.py:45 INFO TRAIN_STEP step=6294 loss=1.599070 Training: 35%|███▍ | 6294/18018 [13:40:41<29:08:52, 8.95s/step, loss=1.599, acc=0.653, acc_len=1.53, thru=12.9, I=14.7, T=28.2, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:10:56,303] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:02,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:11:03,510] loop.py:380 INFO TIMING step=6295: step=4.674s data=2.244s compute=4.628s [fwd=3.204s bwd=1.408s opt=0.003s] dispatch=3.015s [2026-10-19 03:11:03,510] loop.py:45 INFO TRAIN_STEP step=6295 loss=1.414869 Training: 35%|███▍ | 6295/18018 [13:40:49<27:55:25, 8.58s/step, loss=1.415, acc=0.679, acc_len=1.61, thru=15.2, I=19.8, T=27.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:03,893] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:11:10,270] loop.py:45 INFO TRAIN_STEP step=6296 loss=1.440085 Training: 35%|███▍ | 6296/18018 [13:40:56<26:08:48, 8.03s/step, loss=1.440, acc=0.683, acc_len=1.63, thru=15.7, I=19.2, T=28.7, wait=2.3s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:11:10,264] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6295: forward=3108.2ms backward=1305.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:18,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:11:20,004] loop.py:45 INFO TRAIN_STEP step=6297 loss=1.372943 Training: 35%|███▍ | 6297/18018 [13:41:06<27:48:35, 8.54s/step, loss=1.373, acc=0.689, acc_len=1.65, thru=14.6, I=15.3, T=30.6, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:20,283] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:28,326] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:11:28,510] loop.py:45 INFO TRAIN_STEP step=6298 loss=1.388468 Training: 35%|███▍ | 6298/18018 [13:41:14<27:46:16, 8.53s/step, loss=1.388, acc=0.689, acc_len=1.65, thru=13.7, I=16.0, T=31.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:28,828] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:34,760] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:11:35,828] loop.py:45 INFO TRAIN_STEP step=6299 loss=1.491830 Training: 35%|███▍ | 6299/18018 [13:41:21<26:35:12, 8.17s/step, loss=1.492, acc=0.670, acc_len=1.59, thru=17.1, I=20.1, T=28.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:36,269] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:42,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:11:43,555] loop.py:380 INFO TIMING step=6300: step=5.081s data=2.160s compute=5.050s [fwd=3.357s bwd=1.678s opt=0.003s] dispatch=2.636s [2026-10-19 03:11:43,556] loop.py:45 INFO TRAIN_STEP step=6300 loss=1.449665 Training: 35%|███▍ | 6300/18018 [13:41:29<26:09:18, 8.04s/step, loss=1.450, acc=0.677, acc_len=1.61, thru=15.0, I=19.7, T=25.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:43,808] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:50,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:11:51,814] loop.py:45 INFO TRAIN_STEP step=6301 loss=1.653684 Training: 35%|███▍ | 6301/18018 [13:41:37<26:22:06, 8.10s/step, loss=1.654, acc=0.643, acc_len=1.49, thru=14.0, I=17.9, T=27.6, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:11:51,808] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6300: forward=3207.2ms backward=1377.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:11:52,227] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:11:58,243] loop.py:45 INFO TRAIN_STEP step=6302 loss=1.482028 Training: 35%|███▍ | 6302/18018 [13:41:44<24:44:03, 7.60s/step, loss=1.482, acc=0.672, acc_len=1.59, thru=17.6, I=19.5, T=34.2, wait=2.7s, pool=60, epoch=1/2][2026-10-19 03:12:05,869] loop.py:45 INFO TRAIN_STEP step=6303 loss=1.470450 Training: 35%|███▍ | 6303/18018 [13:41:51<24:45:25, 7.61s/step, loss=1.470, acc=0.677, acc_len=1.61, thru=18.4, I=16.8, T=30.3, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:12,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:12:12,921] loop.py:45 INFO TRAIN_STEP step=6304 loss=1.515709 Training: 35%|███▍ | 6304/18018 [13:41:59<24:12:45, 7.44s/step, loss=1.516, acc=0.666, acc_len=1.57, thru=17.1, I=17.4, T=34.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:13,288] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:19,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:12:21,841] loop.py:380 INFO TIMING step=6305: step=4.108s data=1.678s compute=4.075s [fwd=2.867s bwd=1.193s opt=0.003s] dispatch=4.802s [2026-10-19 03:12:21,841] loop.py:45 INFO TRAIN_STEP step=6305 loss=1.449680 Training: 35%|███▍ | 6305/18018 [13:42:07<25:39:07, 7.88s/step, loss=1.450, acc=0.678, acc_len=1.61, thru=14.1, I=18.7, T=31.2, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:22,334] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:12:28,245] loop.py:45 INFO TRAIN_STEP step=6306 loss=1.468174 Training: 35%|███▍ | 6306/18018 [13:42:14<24:12:26, 7.44s/step, loss=1.468, acc=0.674, acc_len=1.60, thru=17.3, I=22.3, T=35.1, wait=2.8s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:12:28,239] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6305: forward=2608.6ms backward=988.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:34,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:12:36,983] loop.py:45 INFO TRAIN_STEP step=6307 loss=1.471188 Training: 35%|███▌ | 6307/18018 [13:42:23<25:28:20, 7.83s/step, loss=1.471, acc=0.676, acc_len=1.60, thru=17.1, I=18.5, T=29.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:37,016] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:44,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:12:45,624] loop.py:45 INFO TRAIN_STEP step=6308 loss=1.316959 Training: 35%|███▌ | 6308/18018 [13:42:31<26:15:39, 8.07s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=14.9, I=17.2, T=25.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:46,004] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:51,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:12:53,127] loop.py:45 INFO TRAIN_STEP step=6309 loss=1.554889 Training: 35%|███▌ | 6309/18018 [13:42:39<25:42:03, 7.90s/step, loss=1.555, acc=0.659, acc_len=1.54, thru=16.5, I=22.5, T=27.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:53,330] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:12:59,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:13:00,029] loop.py:380 INFO TIMING step=6310: step=4.172s data=1.979s compute=4.135s [fwd=2.899s bwd=1.221s opt=0.003s] dispatch=2.720s [2026-10-19 03:13:00,029] loop.py:45 INFO TRAIN_STEP step=6310 loss=1.472208 Training: 35%|███▌ | 6310/18018 [13:42:46<24:43:25, 7.60s/step, loss=1.472, acc=0.669, acc_len=1.58, thru=16.1, I=20.2, T=30.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:00,404] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:05,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:13:07,606] loop.py:45 INFO TRAIN_STEP step=6311 loss=1.432239 Training: 35%|███▌ | 6311/18018 [13:42:53<24:41:47, 7.59s/step, loss=1.432, acc=0.678, acc_len=1.62, thru=17.1, I=22.5, T=27.5, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:13:07,601] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6310: forward=3265.8ms backward=1340.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:08,055] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:14,428] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:13:15,004] loop.py:45 INFO TRAIN_STEP step=6312 loss=1.470006 Training: 35%|███▌ | 6312/18018 [13:43:01<24:30:07, 7.54s/step, loss=1.470, acc=0.676, acc_len=1.60, thru=13.6, I=19.2, T=30.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:15,434] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:20,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:13:21,431] loop.py:45 INFO TRAIN_STEP step=6313 loss=1.393603 Training: 35%|███▌ | 6313/18018 [13:43:07<23:25:14, 7.20s/step, loss=1.394, acc=0.688, acc_len=1.65, thru=18.7, I=22.0, T=28.7, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:21,553] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:26,979] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:13:28,111] loop.py:45 INFO TRAIN_STEP step=6314 loss=1.433299 Training: 35%|███▌ | 6314/18018 [13:43:14<22:54:32, 7.05s/step, loss=1.433, acc=0.680, acc_len=1.62, thru=16.7, I=20.1, T=33.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:28,489] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:32,517] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:13:34,358] loop.py:380 INFO TIMING step=6315: step=3.790s data=1.472s compute=3.756s [fwd=2.530s bwd=1.212s opt=0.003s] dispatch=2.446s [2026-10-19 03:13:34,358] loop.py:45 INFO TRAIN_STEP step=6315 loss=1.484130 Training: 35%|███▌ | 6315/18018 [13:43:20<22:07:30, 6.81s/step, loss=1.484, acc=0.671, acc_len=1.59, thru=19.3, I=27.9, T=33.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:34,526] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:42,032] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:13:42,838] loop.py:45 INFO TRAIN_STEP step=6316 loss=1.534049 Training: 35%|███▌ | 6316/18018 [13:43:28<23:45:25, 7.31s/step, loss=1.534, acc=0.664, acc_len=1.56, thru=15.4, I=16.6, T=31.3, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:13:42,832] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6315: forward=2941.4ms backward=1109.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:43,037] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:48,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:13:49,364] loop.py:45 INFO TRAIN_STEP step=6317 loss=1.302346 Training: 35%|███▌ | 6317/18018 [13:43:35<22:59:33, 7.07s/step, loss=1.302, acc=0.703, acc_len=1.70, thru=18.1, I=20.2, T=33.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:49,661] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:57,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:13:58,053] loop.py:45 INFO TRAIN_STEP step=6318 loss=1.557518 Training: 35%|███▌ | 6318/18018 [13:43:44<24:33:51, 7.56s/step, loss=1.558, acc=0.656, acc_len=1.54, thru=14.6, I=17.5, T=31.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:13:58,313] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:14:05,652] loop.py:45 INFO TRAIN_STEP step=6319 loss=1.350887 Training: 35%|███▌ | 6319/18018 [13:43:51<24:36:09, 7.57s/step, loss=1.351, acc=0.693, acc_len=1.66, thru=13.7, I=14.3, T=32.4, wait=3.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:11,943] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:14:13,640] loop.py:380 INFO TIMING step=6320: step=4.542s data=2.009s compute=4.507s [fwd=3.216s bwd=1.276s opt=0.003s] dispatch=3.435s [2026-10-19 03:14:13,640] loop.py:45 INFO TRAIN_STEP step=6320 loss=1.399332 Training: 35%|███▌ | 6320/18018 [13:43:59<25:00:21, 7.70s/step, loss=1.399, acc=0.684, acc_len=1.63, thru=16.2, I=18.0, T=28.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:13,952] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:20,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:14:21,771] loop.py:45 INFO TRAIN_STEP step=6321 loss=1.487305 Training: 35%|███▌ | 6321/18018 [13:44:07<25:25:40, 7.83s/step, loss=1.487, acc=0.667, acc_len=1.57, thru=13.4, I=18.9, T=32.3, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:14:21,766] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6320: forward=2789.7ms backward=1100.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:21,892] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-19 03:14:28,868] loop.py:45 INFO TRAIN_STEP step=6322 loss=1.533976 Training: 35%|███▌ | 6322/18018 [13:44:14<24:43:04, 7.61s/step, loss=1.534, acc=0.664, acc_len=1.56, thru=17.1, I=20.8, T=34.6, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:34,915] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:14:36,419] loop.py:45 INFO TRAIN_STEP step=6323 loss=1.490782 Training: 35%|███▌ | 6323/18018 [13:44:22<24:39:34, 7.59s/step, loss=1.491, acc=0.668, acc_len=1.58, thru=16.6, I=18.7, T=29.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:36,926] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:43,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:14:44,016] loop.py:45 INFO TRAIN_STEP step=6324 loss=1.469856 Training: 35%|███▌ | 6324/18018 [13:44:30<24:39:48, 7.59s/step, loss=1.470, acc=0.671, acc_len=1.58, thru=15.5, I=20.1, T=27.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:44,342] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:48,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:14:50,370] loop.py:380 INFO TIMING step=6325: step=3.670s data=1.576s compute=3.634s [fwd=2.550s bwd=1.070s opt=0.003s] dispatch=2.672s [2026-10-19 03:14:50,370] loop.py:45 INFO TRAIN_STEP step=6325 loss=1.436241 Training: 35%|███▌ | 6325/18018 [13:44:36<23:27:05, 7.22s/step, loss=1.436, acc=0.681, acc_len=1.62, thru=18.9, I=22.8, T=34.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:50,873] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:56,323] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:14:57,450] loop.py:45 INFO TRAIN_STEP step=6326 loss=1.385320 (TrainerActor pid=9310) [2026-10-19 03:14:57,445] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6325: forward=2667.7ms backward=1111.9ms Training: 35%|███▌ | 6326/18018 [13:44:43<23:18:57, 7.18s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=16.9, I=23.9, T=33.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:14:57,833] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:02,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:15:03,125] loop.py:45 INFO TRAIN_STEP step=6327 loss=1.451994 Training: 35%|███▌ | 6327/18018 [13:44:49<21:50:51, 6.73s/step, loss=1.452, acc=0.678, acc_len=1.61, thru=20.8, I=26.6, T=36.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:03,479] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:10,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:15:11,846] loop.py:45 INFO TRAIN_STEP step=6328 loss=1.583604 Training: 35%|███▌ | 6328/18018 [13:44:57<23:47:20, 7.33s/step, loss=1.584, acc=0.653, acc_len=1.53, thru=14.8, I=17.3, T=27.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:12,264] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:15:19,633] loop.py:45 INFO TRAIN_STEP step=6329 loss=1.236790 Training: 35%|███▌ | 6329/18018 [13:45:05<24:14:09, 7.46s/step, loss=1.237, acc=0.714, acc_len=1.74, thru=15.9, I=18.5, T=28.5, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:27,693] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:15:28,754] loop.py:380 INFO TIMING step=6330: step=4.280s data=1.731s compute=4.250s [fwd=2.929s bwd=1.305s opt=0.003s] dispatch=4.831s [2026-10-19 03:15:28,755] loop.py:45 INFO TRAIN_STEP step=6330 loss=1.438493 Training: 35%|███▌ | 6330/18018 [13:45:14<25:50:45, 7.96s/step, loss=1.438, acc=0.678, acc_len=1.61, thru=14.7, I=15.2, T=29.9, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:29,200] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:15:35,746] loop.py:45 INFO TRAIN_STEP step=6331 loss=1.516801 Training: 35%|███▌ | 6331/18018 [13:45:21<24:54:02, 7.67s/step, loss=1.517, acc=0.664, acc_len=1.57, thru=13.4, I=20.7, T=32.5, wait=3.0s, pool=28, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:15:35,741] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6330: forward=2731.0ms backward=1154.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:49,140] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:15:49,297] loop.py:45 INFO TRAIN_STEP step=6332 loss=1.634311 Training: 35%|███▌ | 6332/18018 [13:45:35<30:37:27, 9.43s/step, loss=1.634, acc=0.647, acc_len=1.51, thru=11.8, I=11.7, T=29.2, wait=9.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:49,642] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:54,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:15:56,069] loop.py:45 INFO TRAIN_STEP step=6333 loss=1.539454 Training: 35%|███▌ | 6333/18018 [13:45:42<28:01:44, 8.64s/step, loss=1.539, acc=0.657, acc_len=1.54, thru=18.0, I=21.1, T=31.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:15:56,326] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:02,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:16:02,718] loop.py:45 INFO TRAIN_STEP step=6334 loss=1.524712 Training: 35%|███▌ | 6334/18018 [13:45:48<26:05:39, 8.04s/step, loss=1.525, acc=0.668, acc_len=1.58, thru=17.2, I=21.0, T=32.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:02,757] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-19 03:16:10,315] loop.py:380 INFO TIMING step=6335: step=4.409s data=2.186s compute=4.369s [fwd=3.081s bwd=1.273s opt=0.003s] dispatch=3.178s [2026-10-19 03:16:10,316] loop.py:45 INFO TRAIN_STEP step=6335 loss=1.627638 Training: 35%|███▌ | 6335/18018 [13:45:56<25:39:41, 7.91s/step, loss=1.628, acc=0.642, acc_len=1.49, thru=16.4, I=16.9, T=29.0, wait=3.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:16,033] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:16:17,820] loop.py:45 INFO TRAIN_STEP step=6336 loss=1.393273 Training: 35%|███▌ | 6336/18018 [13:46:03<25:16:02, 7.79s/step, loss=1.393, acc=0.686, acc_len=1.64, thru=17.8, I=20.0, T=30.4, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:16:17,815] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6335: forward=2786.7ms backward=1343.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:18,046] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:25,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:16:27,663] loop.py:45 INFO TRAIN_STEP step=6337 loss=1.424114 Training: 35%|███▌ | 6337/18018 [13:46:13<27:16:02, 8.40s/step, loss=1.424, acc=0.681, acc_len=1.62, thru=13.3, I=16.2, T=26.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:27,797] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:32,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:16:34,177] loop.py:45 INFO TRAIN_STEP step=6338 loss=1.507296 Training: 35%|███▌ | 6338/18018 [13:46:20<25:25:25, 7.84s/step, loss=1.507, acc=0.664, acc_len=1.57, thru=16.9, I=25.9, T=31.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:34,451] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:41,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:16:41,740] loop.py:45 INFO TRAIN_STEP step=6339 loss=1.590970 Training: 35%|███▌ | 6339/18018 [13:46:27<25:09:25, 7.75s/step, loss=1.591, acc=0.656, acc_len=1.54, thru=14.0, I=17.7, T=32.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:42,023] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:47,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:16:49,215] loop.py:380 INFO TIMING step=6340: step=4.388s data=1.979s compute=4.346s [fwd=2.989s bwd=1.341s opt=0.003s] dispatch=3.078s [2026-10-19 03:16:49,215] loop.py:45 INFO TRAIN_STEP step=6340 loss=1.535044 Training: 35%|███▌ | 6340/18018 [13:46:35<24:52:54, 7.67s/step, loss=1.535, acc=0.660, acc_len=1.55, thru=17.7, I=19.9, T=29.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:49,675] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:55,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:16:57,142] loop.py:45 INFO TRAIN_STEP step=6341 loss=1.498844 Training: 35%|███▌ | 6341/18018 [13:46:43<25:07:43, 7.75s/step, loss=1.499, acc=0.668, acc_len=1.58, thru=13.2, I=18.9, T=27.9, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:16:57,137] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6340: forward=3120.0ms backward=1416.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:16:57,383] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:02,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:17:03,416] loop.py:45 INFO TRAIN_STEP step=6342 loss=1.447998 Training: 35%|███▌ | 6342/18018 [13:46:49<23:41:39, 7.31s/step, loss=1.448, acc=0.677, acc_len=1.61, thru=17.2, I=20.8, T=33.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:03,547] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:12,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:17:13,444] loop.py:45 INFO TRAIN_STEP step=6343 loss=1.476959 Training: 35%|███▌ | 6343/18018 [13:46:59<26:20:23, 8.12s/step, loss=1.477, acc=0.675, acc_len=1.61, thru=13.0, I=14.0, T=29.3, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:13,822] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:18,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:17:20,289] loop.py:45 INFO TRAIN_STEP step=6344 loss=1.416616 Training: 35%|███▌ | 6344/18018 [13:47:06<25:05:51, 7.74s/step, loss=1.417, acc=0.688, acc_len=1.64, thru=17.4, I=22.5, T=29.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:20,375] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:25,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:17:27,181] loop.py:380 INFO TIMING step=6345: step=4.686s data=2.212s compute=4.658s [fwd=3.067s bwd=1.575s opt=0.003s] dispatch=2.195s [2026-10-19 03:17:27,181] loop.py:45 INFO TRAIN_STEP step=6345 loss=1.514128 Training: 35%|███▌ | 6345/18018 [13:47:13<24:16:13, 7.49s/step, loss=1.514, acc=0.668, acc_len=1.58, thru=15.7, I=24.4, T=27.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:27,479] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:17:33,695] loop.py:45 INFO TRAIN_STEP step=6346 loss=1.513744 Training: 35%|███▌ | 6346/18018 [13:47:19<23:19:22, 7.19s/step, loss=1.514, acc=0.668, acc_len=1.57, thru=15.4, I=20.5, T=34.8, wait=2.8s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:17:33,690] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6345: forward=2583.0ms backward=1045.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:40,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:17:42,303] loop.py:45 INFO TRAIN_STEP step=6347 loss=1.653352 Training: 35%|███▌ | 6347/18018 [13:47:28<24:41:54, 7.62s/step, loss=1.653, acc=0.645, acc_len=1.50, thru=15.6, I=16.7, T=28.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:42,362] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:48,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:17:49,777] loop.py:45 INFO TRAIN_STEP step=6348 loss=1.461443 Training: 35%|███▌ | 6348/18018 [13:47:35<24:33:23, 7.58s/step, loss=1.461, acc=0.673, acc_len=1.60, thru=15.7, I=20.6, T=29.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:49,888] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:57,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:17:59,305] loop.py:45 INFO TRAIN_STEP step=6349 loss=1.571448 Training: 35%|███▌ | 6349/18018 [13:47:45<26:27:06, 8.16s/step, loss=1.571, acc=0.656, acc_len=1.54, thru=13.6, I=16.3, T=25.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:17:59,589] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:04,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:18:05,548] loop.py:380 INFO TIMING step=6350: step=4.010s data=1.849s compute=3.979s [fwd=2.871s bwd=1.093s opt=0.003s] dispatch=2.223s [2026-10-19 03:18:05,548] loop.py:45 INFO TRAIN_STEP step=6350 loss=1.418582 Training: 35%|███▌ | 6350/18018 [13:47:51<24:35:08, 7.59s/step, loss=1.419, acc=0.681, acc_len=1.62, thru=16.8, I=22.3, T=31.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:05,843] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:12,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:18:12,554] loop.py:45 INFO TRAIN_STEP step=6351 loss=1.507127 Training: 35%|███▌ | 6351/18018 [13:47:58<24:01:10, 7.41s/step, loss=1.507, acc=0.667, acc_len=1.57, thru=16.5, I=19.7, T=28.4, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:18:12,548] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6350: forward=3100.8ms backward=1353.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:12,742] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:19,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:18:21,099] loop.py:45 INFO TRAIN_STEP step=6352 loss=1.566083 Training: 35%|███▌ | 6352/18018 [13:48:07<25:07:11, 7.75s/step, loss=1.566, acc=0.655, acc_len=1.54, thru=15.7, I=19.7, T=25.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:21,137] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 03:18:28,127] loop.py:45 INFO TRAIN_STEP step=6353 loss=1.517152 Training: 35%|███▌ | 6353/18018 [13:48:14<24:24:54, 7.53s/step, loss=1.517, acc=0.663, acc_len=1.56, thru=12.1, I=18.1, T=31.2, wait=2.9s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:35,954] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:18:37,831] loop.py:45 INFO TRAIN_STEP step=6354 loss=1.608081 Training: 35%|███▌ | 6354/18018 [13:48:23<26:31:05, 8.18s/step, loss=1.608, acc=0.650, acc_len=1.52, thru=16.4, I=16.8, T=29.7, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:37,975] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:42,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:18:44,660] loop.py:380 INFO TIMING step=6355: step=4.722s data=2.354s compute=4.696s [fwd=3.289s bwd=1.391s opt=0.003s] dispatch=2.098s [2026-10-19 03:18:44,660] loop.py:45 INFO TRAIN_STEP step=6355 loss=1.453879 Training: 35%|███▌ | 6355/18018 [13:48:30<25:11:53, 7.78s/step, loss=1.454, acc=0.675, acc_len=1.60, thru=15.6, I=26.6, T=27.1, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:45,154] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 03:18:52,539] loop.py:45 INFO TRAIN_STEP step=6356 loss=1.453886 Training: 35%|███▌ | 6356/18018 [13:48:38<25:17:42, 7.81s/step, loss=1.454, acc=0.671, acc_len=1.59, thru=15.2, I=17.8, T=31.1, wait=3.8s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:18:52,534] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6355: forward=2916.9ms backward=1142.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:18:59,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:19:00,552] loop.py:45 INFO TRAIN_STEP step=6357 loss=1.343603 Training: 35%|███▌ | 6357/18018 [13:48:46<25:29:31, 7.87s/step, loss=1.344, acc=0.695, acc_len=1.67, thru=16.4, I=17.4, T=33.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:00,907] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:08,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:19:09,051] loop.py:45 INFO TRAIN_STEP step=6358 loss=1.545010 Training: 35%|███▌ | 6358/18018 [13:48:55<26:06:03, 8.06s/step, loss=1.545, acc=0.663, acc_len=1.56, thru=14.9, I=17.3, T=22.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:09,122] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:14,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:19:16,486] loop.py:45 INFO TRAIN_STEP step=6359 loss=1.531922 Training: 35%|███▌ | 6359/18018 [13:49:02<25:29:40, 7.87s/step, loss=1.532, acc=0.666, acc_len=1.58, thru=15.0, I=17.5, T=28.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:16,988] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:25,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:19:25,375] loop.py:380 INFO TIMING step=6360: step=4.468s data=2.095s compute=4.436s [fwd=3.133s bwd=1.288s opt=0.003s] dispatch=4.409s [2026-10-19 03:19:25,375] loop.py:45 INFO TRAIN_STEP step=6360 loss=1.496587 Training: 35%|███▌ | 6360/18018 [13:49:11<26:28:37, 8.18s/step, loss=1.497, acc=0.669, acc_len=1.58, thru=15.3, I=13.3, T=28.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:25,670] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:34,022] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:19:35,127] loop.py:45 INFO TRAIN_STEP step=6361 loss=1.570520 Training: 35%|███▌ | 6361/18018 [13:49:21<28:00:20, 8.65s/step, loss=1.571, acc=0.656, acc_len=1.54, thru=13.2, I=14.7, T=25.2, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:19:35,122] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6360: forward=3480.6ms backward=1530.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:35,531] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:41,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:19:42,075] loop.py:45 INFO TRAIN_STEP step=6362 loss=1.454064 Training: 35%|███▌ | 6362/18018 [13:49:28<26:21:12, 8.14s/step, loss=1.454, acc=0.677, acc_len=1.60, thru=16.9, I=18.9, T=28.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:42,582] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:19:51,319] loop.py:45 INFO TRAIN_STEP step=6363 loss=1.587849 Training: 35%|███▌ | 6363/18018 [13:49:37<27:25:22, 8.47s/step, loss=1.588, acc=0.654, acc_len=1.53, thru=13.7, I=15.6, T=30.7, wait=5.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:57,856] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:19:59,274] loop.py:45 INFO TRAIN_STEP step=6364 loss=1.493300 Training: 35%|███▌ | 6364/18018 [13:49:45<26:55:17, 8.32s/step, loss=1.493, acc=0.673, acc_len=1.60, thru=15.9, I=17.4, T=30.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:19:59,381] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:04,107] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:20:06,700] loop.py:380 INFO TIMING step=6365: step=4.750s data=1.827s compute=4.718s [fwd=3.540s bwd=1.164s opt=0.003s] dispatch=2.665s [2026-10-19 03:20:06,700] loop.py:45 INFO TRAIN_STEP step=6365 loss=1.392647 Training: 35%|███▌ | 6365/18018 [13:49:52<26:03:15, 8.05s/step, loss=1.393, acc=0.686, acc_len=1.64, thru=16.2, I=24.7, T=26.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:07,129] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:11,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:20:12,935] loop.py:45 INFO TRAIN_STEP step=6366 loss=1.415849 Training: 35%|███▌ | 6366/18018 [13:49:59<24:17:27, 7.50s/step, loss=1.416, acc=0.681, acc_len=1.62, thru=16.1, I=25.0, T=35.5, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:20:12,930] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6365: forward=2494.7ms backward=1059.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:13,150] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:18,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:20:19,933] loop.py:45 INFO TRAIN_STEP step=6367 loss=1.406191 Training: 35%|███▌ | 6367/18018 [13:50:06<23:47:40, 7.35s/step, loss=1.406, acc=0.684, acc_len=1.64, thru=18.0, I=23.8, T=28.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:20,367] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:28,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:20:29,742] loop.py:45 INFO TRAIN_STEP step=6368 loss=1.552549 Training: 35%|███▌ | 6368/18018 [13:50:15<26:10:48, 8.09s/step, loss=1.553, acc=0.658, acc_len=1.55, thru=13.7, I=14.1, T=30.6, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:29,977] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:37,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:20:38,909] loop.py:45 INFO TRAIN_STEP step=6369 loss=1.560827 Training: 35%|███▌ | 6369/18018 [13:50:24<27:13:22, 8.41s/step, loss=1.561, acc=0.659, acc_len=1.55, thru=14.3, I=17.2, T=25.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:39,139] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:43,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:20:44,716] loop.py:380 INFO TIMING step=6370: step=3.867s data=1.728s compute=3.839s [fwd=2.718s bwd=1.107s opt=0.003s] dispatch=1.929s [2026-10-19 03:20:44,716] loop.py:45 INFO TRAIN_STEP step=6370 loss=1.336613 Training: 35%|███▌ | 6370/18018 [13:50:30<24:41:21, 7.63s/step, loss=1.337, acc=0.697, acc_len=1.68, thru=18.6, I=25.1, T=33.1, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:20:45,138] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:20:52,094] loop.py:45 INFO TRAIN_STEP step=6371 loss=1.460307 Training: 35%|███▌ | 6371/18018 [13:50:38<24:26:35, 7.56s/step, loss=1.460, acc=0.676, acc_len=1.60, thru=15.2, I=21.2, T=31.8, wait=3.3s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:20:52,089] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6370: forward=2852.8ms backward=1127.6ms [2026-10-19 03:21:01,292] loop.py:45 INFO TRAIN_STEP step=6372 loss=1.300996 Training: 35%|███▌ | 6372/18018 [13:50:47<26:02:10, 8.05s/step, loss=1.301, acc=0.702, acc_len=1.70, thru=14.9, I=14.5, T=25.4, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:07,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:21:08,976] loop.py:45 INFO TRAIN_STEP step=6373 loss=1.463084 Training: 35%|███▌ | 6373/18018 [13:50:55<25:40:43, 7.94s/step, loss=1.463, acc=0.675, acc_len=1.60, thru=16.1, I=17.3, T=28.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:09,005] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:14,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:21:15,178] loop.py:45 INFO TRAIN_STEP step=6374 loss=1.438718 Training: 35%|███▌ | 6374/18018 [13:51:01<23:59:30, 7.42s/step, loss=1.439, acc=0.679, acc_len=1.61, thru=18.2, I=22.7, T=32.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:15,559] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:23,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:21:25,049] loop.py:380 INFO TIMING step=6375: step=4.994s data=2.464s compute=4.966s [fwd=3.478s bwd=1.473s opt=0.003s] dispatch=4.867s [2026-10-19 03:21:25,049] loop.py:45 INFO TRAIN_STEP step=6375 loss=1.585919 Training: 35%|███▌ | 6375/18018 [13:51:11<26:22:16, 8.15s/step, loss=1.586, acc=0.655, acc_len=1.54, thru=13.5, I=15.2, T=25.6, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:25,513] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:35,979] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:21:36,040] loop.py:45 INFO TRAIN_STEP step=6376 loss=1.525818 Training: 35%|███▌ | 6376/18018 [13:51:22<29:07:09, 9.00s/step, loss=1.526, acc=0.664, acc_len=1.56, thru=11.4, I=11.4, T=30.0, wait=6.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:21:36,035] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6375: forward=2885.0ms backward=1321.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:36,480] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:42,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:21:43,685] loop.py:45 INFO TRAIN_STEP step=6377 loss=1.501386 Training: 35%|███▌ | 6377/18018 [13:51:29<27:48:03, 8.60s/step, loss=1.501, acc=0.668, acc_len=1.58, thru=16.8, I=18.6, T=28.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:44,090] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:48,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:21:51,073] loop.py:45 INFO TRAIN_STEP step=6378 loss=1.458567 Training: 35%|███▌ | 6378/18018 [13:51:37<26:37:30, 8.23s/step, loss=1.459, acc=0.676, acc_len=1.61, thru=16.4, I=24.1, T=24.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:51,448] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:56,674] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:21:58,971] loop.py:45 INFO TRAIN_STEP step=6379 loss=1.456487 Training: 35%|███▌ | 6379/18018 [13:51:45<26:17:39, 8.13s/step, loss=1.456, acc=0.675, acc_len=1.60, thru=17.0, I=22.6, T=26.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:21:59,191] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:04,722] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:22:06,317] loop.py:380 INFO TIMING step=6380: step=4.745s data=1.876s compute=4.708s [fwd=3.269s bwd=1.425s opt=0.003s] dispatch=2.592s [2026-10-19 03:22:06,317] loop.py:45 INFO TRAIN_STEP step=6380 loss=1.419282 Training: 35%|███▌ | 6380/18018 [13:51:52<25:31:44, 7.90s/step, loss=1.419, acc=0.685, acc_len=1.64, thru=13.8, I=21.6, T=27.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:06,735] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:13,767] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:22:15,620] loop.py:45 INFO TRAIN_STEP step=6381 loss=1.444062 Training: 35%|███▌ | 6381/18018 [13:52:01<26:53:28, 8.32s/step, loss=1.444, acc=0.678, acc_len=1.61, thru=14.4, I=16.8, T=23.1, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:22:15,616] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6380: forward=4108.0ms backward=1377.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:15,788] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:22,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:22:25,540] loop.py:45 INFO TRAIN_STEP step=6382 loss=1.381717 Training: 35%|███▌ | 6382/18018 [13:52:11<28:26:26, 8.80s/step, loss=1.382, acc=0.688, acc_len=1.65, thru=13.1, I=19.1, T=23.5, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:25,896] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:30,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:22:31,807] loop.py:45 INFO TRAIN_STEP step=6383 loss=1.498980 Training: 35%|███▌ | 6383/18018 [13:52:17<25:58:56, 8.04s/step, loss=1.499, acc=0.670, acc_len=1.59, thru=15.2, I=27.9, T=34.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:32,108] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:22:38,136] loop.py:45 INFO TRAIN_STEP step=6384 loss=1.534817 Training: 35%|███▌ | 6384/18018 [13:52:24<24:19:26, 7.53s/step, loss=1.535, acc=0.666, acc_len=1.57, thru=17.2, I=23.1, T=31.8, wait=2.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:45,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:22:46,300] loop.py:380 INFO TIMING step=6385: step=4.003s data=1.821s compute=3.957s [fwd=2.754s bwd=1.188s opt=0.003s] dispatch=4.151s [2026-10-19 03:22:46,300] loop.py:45 INFO TRAIN_STEP step=6385 loss=1.541319 Training: 35%|███▌ | 6385/18018 [13:52:32<24:56:25, 7.72s/step, loss=1.541, acc=0.662, acc_len=1.56, thru=17.5, I=17.5, T=32.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:46,338] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 03:22:52,629] loop.py:45 INFO TRAIN_STEP step=6386 loss=1.346222 Training: 35%|███▌ | 6386/18018 [13:52:38<23:35:20, 7.30s/step, loss=1.346, acc=0.696, acc_len=1.68, thru=19.0, I=21.1, T=31.1, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:22:52,624] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6385: forward=2814.5ms backward=1255.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:52,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:22:53,118] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:02,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:23:03,325] loop.py:45 INFO TRAIN_STEP step=6387 loss=1.577040 Training: 35%|███▌ | 6387/18018 [13:52:49<26:52:41, 8.32s/step, loss=1.577, acc=0.653, acc_len=1.53, thru=12.1, I=12.7, T=30.1, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:03,802] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 03:23:12,049] loop.py:45 INFO TRAIN_STEP step=6388 loss=1.579654 Training: 35%|███▌ | 6388/18018 [13:52:58<27:16:05, 8.44s/step, loss=1.580, acc=0.654, acc_len=1.54, thru=13.3, I=17.7, T=30.1, wait=4.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:19,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:23:20,223] loop.py:45 INFO TRAIN_STEP step=6389 loss=1.459925 Training: 35%|███▌ | 6389/18018 [13:53:06<27:00:33, 8.36s/step, loss=1.460, acc=0.678, acc_len=1.60, thru=15.8, I=16.9, T=31.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:20,264] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:25,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:23:26,723] loop.py:380 INFO TIMING step=6390: step=3.947s data=1.845s compute=3.906s [fwd=2.756s bwd=1.135s opt=0.003s] dispatch=2.543s [2026-10-19 03:23:26,724] loop.py:45 INFO TRAIN_STEP step=6390 loss=1.455914 Training: 35%|███▌ | 6390/18018 [13:53:12<25:12:14, 7.80s/step, loss=1.456, acc=0.676, acc_len=1.61, thru=18.5, I=22.6, T=32.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:27,000] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:23:33,589] loop.py:45 INFO TRAIN_STEP step=6391 loss=1.466789 Training: 35%|███▌ | 6391/18018 [13:53:19<24:17:29, 7.52s/step, loss=1.467, acc=0.673, acc_len=1.59, thru=14.7, I=19.7, T=30.1, wait=2.6s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:23:33,585] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6390: forward=2980.3ms backward=1221.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:41,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:23:42,134] loop.py:45 INFO TRAIN_STEP step=6392 loss=1.528773 Training: 35%|███▌ | 6392/18018 [13:53:28<25:16:49, 7.83s/step, loss=1.529, acc=0.660, acc_len=1.55, thru=17.8, I=15.5, T=28.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:42,495] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:50,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:23:51,436] loop.py:45 INFO TRAIN_STEP step=6393 loss=1.512746 Training: 35%|███▌ | 6393/18018 [13:53:37<26:42:30, 8.27s/step, loss=1.513, acc=0.665, acc_len=1.57, thru=14.2, I=15.8, T=29.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:23:51,514] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-19 03:23:58,921] loop.py:45 INFO TRAIN_STEP step=6394 loss=1.586231 Training: 35%|███▌ | 6394/18018 [13:53:45<25:56:37, 8.03s/step, loss=1.586, acc=0.650, acc_len=1.52, thru=13.4, I=14.8, T=33.7, wait=3.7s, pool=44, epoch=1/2][2026-10-19 03:24:09,248] loop.py:380 INFO TIMING step=6395: step=4.295s data=2.060s compute=4.246s [fwd=2.953s bwd=1.279s opt=0.003s] dispatch=6.022s [2026-10-19 03:24:09,248] loop.py:45 INFO TRAIN_STEP step=6395 loss=1.579868 Training: 35%|███▌ | 6395/18018 [13:53:55<28:09:46, 8.72s/step, loss=1.580, acc=0.656, acc_len=1.53, thru=12.1, I=12.4, T=29.8, wait=6.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:16,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:24:17,238] loop.py:45 INFO TRAIN_STEP step=6396 loss=1.485255 Training: 35%|███▌ | 6396/18018 [13:54:03<27:27:03, 8.50s/step, loss=1.485, acc=0.672, acc_len=1.60, thru=18.6, I=16.1, T=32.7, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:24:17,234] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6395: forward=2747.3ms backward=1116.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:17,443] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:23,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:24:23,988] loop.py:45 INFO TRAIN_STEP step=6397 loss=1.492717 Training: 36%|███▌ | 6397/18018 [13:54:10<25:44:56, 7.98s/step, loss=1.493, acc=0.668, acc_len=1.58, thru=18.7, I=21.0, T=30.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:24,252] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:31,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:24:33,096] loop.py:45 INFO TRAIN_STEP step=6398 loss=1.465249 Training: 36%|███▌ | 6398/18018 [13:54:19<26:50:33, 8.32s/step, loss=1.465, acc=0.672, acc_len=1.59, thru=13.4, I=19.0, T=25.2, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:33,243] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:38,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:24:39,391] loop.py:45 INFO TRAIN_STEP step=6399 loss=1.573382 Training: 36%|███▌ | 6399/18018 [13:54:25<24:53:06, 7.71s/step, loss=1.573, acc=0.654, acc_len=1.53, thru=16.5, I=24.2, T=32.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:39,696] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:45,325] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:24:47,598] loop.py:380 INFO TIMING step=6400: step=4.730s data=2.516s compute=4.690s [fwd=3.476s bwd=1.199s opt=0.003s] dispatch=3.467s [2026-10-19 03:24:47,599] loop.py:45 INFO TRAIN_STEP step=6400 loss=1.494736 Training: 36%|███▌ | 6400/18018 [13:54:33<25:21:50, 7.86s/step, loss=1.495, acc=0.670, acc_len=1.59, thru=14.5, I=22.0, T=27.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:24:47,844] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:24:55,557] loop.py:45 INFO TRAIN_STEP step=6401 loss=1.466708 Training: 36%|███▌ | 6401/18018 [13:54:41<25:27:32, 7.89s/step, loss=1.467, acc=0.673, acc_len=1.60, thru=15.0, I=17.6, T=29.5, wait=3.6s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:24:55,552] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6400: forward=3134.9ms backward=1160.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:03,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:25:03,932] loop.py:45 INFO TRAIN_STEP step=6402 loss=1.529591 Training: 36%|███▌ | 6402/18018 [13:54:50<25:55:23, 8.03s/step, loss=1.530, acc=0.663, acc_len=1.56, thru=16.2, I=16.1, T=28.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:04,343] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:10,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:25:11,385] loop.py:45 INFO TRAIN_STEP step=6403 loss=1.474286 Training: 36%|███▌ | 6403/18018 [13:54:57<25:21:30, 7.86s/step, loss=1.474, acc=0.672, acc_len=1.59, thru=17.3, I=19.6, T=30.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:11,784] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 03:25:17,984] loop.py:45 INFO TRAIN_STEP step=6404 loss=1.382164 Training: 36%|███▌ | 6404/18018 [13:55:04<24:08:12, 7.48s/step, loss=1.382, acc=0.689, acc_len=1.65, thru=17.6, I=22.1, T=30.7, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:23,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:25:24,801] loop.py:380 INFO TIMING step=6405: step=4.283s data=1.934s compute=4.239s [fwd=2.929s bwd=1.296s opt=0.003s] dispatch=2.524s [2026-10-19 03:25:24,801] loop.py:45 INFO TRAIN_STEP step=6405 loss=1.506406 Training: 36%|███▌ | 6405/18018 [13:55:10<23:29:34, 7.28s/step, loss=1.506, acc=0.666, acc_len=1.58, thru=16.9, I=19.6, T=29.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:25,226] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:30,600] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:25:32,517] loop.py:45 INFO TRAIN_STEP step=6406 loss=1.465245 Training: 36%|███▌ | 6406/18018 [13:55:18<23:54:36, 7.41s/step, loss=1.465, acc=0.679, acc_len=1.62, thru=15.2, I=22.0, T=27.9, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:25:32,512] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6405: forward=3214.0ms backward=1320.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:32,628] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:37,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:25:38,795] loop.py:45 INFO TRAIN_STEP step=6407 loss=1.357602 Training: 36%|███▌ | 6407/18018 [13:55:24<22:48:35, 7.07s/step, loss=1.358, acc=0.691, acc_len=1.66, thru=17.5, I=23.2, T=30.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:39,248] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:47,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:25:48,472] loop.py:45 INFO TRAIN_STEP step=6408 loss=1.537356 Training: 36%|███▌ | 6408/18018 [13:55:34<25:19:42, 7.85s/step, loss=1.537, acc=0.660, acc_len=1.55, thru=13.9, I=13.5, T=29.6, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:25:48,574] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-19 03:25:56,800] loop.py:45 INFO TRAIN_STEP step=6409 loss=1.458476 Training: 36%|███▌ | 6409/18018 [13:55:42<25:47:00, 8.00s/step, loss=1.458, acc=0.678, acc_len=1.61, thru=13.7, I=16.0, T=31.2, wait=4.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:04,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:26:05,965] loop.py:380 INFO TIMING step=6410: step=4.266s data=1.974s compute=4.237s [fwd=3.017s bwd=1.205s opt=0.003s] dispatch=4.889s [2026-10-19 03:26:05,965] loop.py:45 INFO TRAIN_STEP step=6410 loss=1.394827 Training: 36%|███▌ | 6410/18018 [13:55:52<26:54:49, 8.35s/step, loss=1.395, acc=0.683, acc_len=1.63, thru=15.3, I=16.9, T=30.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:06,003] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:12,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:26:13,930] loop.py:45 INFO TRAIN_STEP step=6411 loss=1.440378 Training: 36%|███▌ | 6411/18018 [13:56:00<26:32:32, 8.23s/step, loss=1.440, acc=0.682, acc_len=1.62, thru=14.1, I=19.7, T=30.2, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:26:13,924] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6410: forward=2875.8ms backward=1298.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:14,120] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:21,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:26:22,396] loop.py:45 INFO TRAIN_STEP step=6412 loss=1.517641 Training: 36%|███▌ | 6412/18018 [13:56:08<26:45:59, 8.30s/step, loss=1.518, acc=0.666, acc_len=1.57, thru=13.6, I=18.1, T=31.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:22,473] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-19 03:26:29,850] loop.py:45 INFO TRAIN_STEP step=6413 loss=1.382970 Training: 36%|███▌ | 6413/18018 [13:56:15<25:56:32, 8.05s/step, loss=1.383, acc=0.689, acc_len=1.65, thru=16.0, I=16.9, T=32.0, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:38,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:26:40,327] loop.py:45 INFO TRAIN_STEP step=6414 loss=1.535926 Training: 36%|███▌ | 6414/18018 [13:56:26<28:17:19, 8.78s/step, loss=1.536, acc=0.664, acc_len=1.56, thru=12.9, I=15.5, T=28.5, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:40,685] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:45,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:26:46,633] loop.py:380 INFO TIMING step=6415: step=4.175s data=1.918s compute=4.144s [fwd=2.893s bwd=1.235s opt=0.003s] dispatch=2.123s [2026-10-19 03:26:46,634] loop.py:45 INFO TRAIN_STEP step=6415 loss=1.416673 Training: 36%|███▌ | 6415/18018 [13:56:32<25:53:55, 8.04s/step, loss=1.417, acc=0.680, acc_len=1.63, thru=17.1, I=26.2, T=30.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:46,801] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:51,415] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:26:54,245] loop.py:45 INFO TRAIN_STEP step=6416 loss=1.408945 Training: 36%|███▌ | 6416/18018 [13:56:40<25:29:14, 7.91s/step, loss=1.409, acc=0.684, acc_len=1.63, thru=16.3, I=25.0, T=26.4, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:26:54,240] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6415: forward=3378.2ms backward=1429.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:54,439] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:26:59,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:27:00,656] loop.py:45 INFO TRAIN_STEP step=6417 loss=1.421416 Training: 36%|███▌ | 6417/18018 [13:56:46<24:02:08, 7.46s/step, loss=1.421, acc=0.686, acc_len=1.63, thru=14.9, I=24.3, T=30.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:01,119] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:06,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:27:07,729] loop.py:45 INFO TRAIN_STEP step=6418 loss=1.412008 Training: 36%|███▌ | 6418/18018 [13:56:53<23:39:39, 7.34s/step, loss=1.412, acc=0.682, acc_len=1.63, thru=17.6, I=22.3, T=33.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:07,894] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:15,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:27:15,548] loop.py:45 INFO TRAIN_STEP step=6419 loss=1.418197 Training: 36%|███▌ | 6419/18018 [13:57:01<24:07:05, 7.49s/step, loss=1.418, acc=0.682, acc_len=1.62, thru=14.9, I=16.1, T=31.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:15,944] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:24,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:27:26,179] loop.py:380 INFO TIMING step=6420: step=4.454s data=1.891s compute=4.394s [fwd=3.064s bwd=1.314s opt=0.003s] dispatch=6.169s [2026-10-19 03:27:26,179] loop.py:45 INFO TRAIN_STEP step=6420 loss=1.535918 Training: 36%|███▌ | 6420/18018 [13:57:12<27:09:25, 8.43s/step, loss=1.536, acc=0.664, acc_len=1.57, thru=12.3, I=14.3, T=28.7, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:26,209] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:32,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:27:33,424] loop.py:45 INFO TRAIN_STEP step=6421 loss=1.462653 Training: 36%|███▌ | 6421/18018 [13:57:19<26:00:39, 8.07s/step, loss=1.463, acc=0.675, acc_len=1.60, thru=16.2, I=21.2, T=30.7, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:27:33,418] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6420: forward=2958.8ms backward=1157.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:33,796] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:27:40,322] loop.py:45 INFO TRAIN_STEP step=6422 loss=1.470982 Training: 36%|███▌ | 6422/18018 [13:57:26<24:52:13, 7.72s/step, loss=1.471, acc=0.673, acc_len=1.60, thru=14.3, I=16.4, T=32.8, wait=3.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:48,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:27:49,488] loop.py:45 INFO TRAIN_STEP step=6423 loss=1.418502 Training: 36%|███▌ | 6423/18018 [13:57:35<26:15:55, 8.15s/step, loss=1.419, acc=0.682, acc_len=1.63, thru=17.2, I=16.0, T=33.5, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:49,649] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:56,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:27:57,073] loop.py:45 INFO TRAIN_STEP step=6424 loss=1.443909 Training: 36%|███▌ | 6424/18018 [13:57:43<25:42:45, 7.98s/step, loss=1.444, acc=0.678, acc_len=1.62, thru=15.5, I=18.4, T=32.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:27:57,306] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:02,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:28:04,913] loop.py:380 INFO TIMING step=6425: step=4.331s data=1.995s compute=4.294s [fwd=3.061s bwd=1.218s opt=0.003s] dispatch=3.499s [2026-10-19 03:28:04,913] loop.py:45 INFO TRAIN_STEP step=6425 loss=1.391153 Training: 36%|███▌ | 6425/18018 [13:57:50<25:34:20, 7.94s/step, loss=1.391, acc=0.688, acc_len=1.64, thru=15.7, I=19.8, T=29.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:05,086] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:10,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:28:12,366] loop.py:45 INFO TRAIN_STEP step=6426 loss=1.392960 Training: 36%|███▌ | 6426/18018 [13:57:58<25:05:43, 7.79s/step, loss=1.393, acc=0.686, acc_len=1.63, thru=13.6, I=23.8, T=27.9, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:28:12,361] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6425: forward=3171.1ms backward=1363.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:12,858] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:19,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:28:21,103] loop.py:45 INFO TRAIN_STEP step=6427 loss=1.495868 Training: 36%|███▌ | 6427/18018 [13:58:07<26:00:28, 8.08s/step, loss=1.496, acc=0.666, acc_len=1.57, thru=15.5, I=18.1, T=25.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:21,279] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:26,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:28:27,472] loop.py:45 INFO TRAIN_STEP step=6428 loss=1.346014 Training: 36%|███▌ | 6428/18018 [13:58:13<24:21:17, 7.56s/step, loss=1.346, acc=0.693, acc_len=1.67, thru=16.6, I=20.8, T=34.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:27,880] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:33,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:28:34,376] loop.py:45 INFO TRAIN_STEP step=6429 loss=1.505553 Training: 36%|███▌ | 6429/18018 [13:58:20<23:42:53, 7.37s/step, loss=1.506, acc=0.670, acc_len=1.58, thru=18.2, I=21.9, T=33.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:34,703] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 03:28:42,007] loop.py:380 INFO TIMING step=6430: step=4.331s data=2.056s compute=4.297s [fwd=3.049s bwd=1.233s opt=0.003s] dispatch=3.290s [2026-10-19 03:28:42,007] loop.py:45 INFO TRAIN_STEP step=6430 loss=1.408456 Training: 36%|███▌ | 6430/18018 [13:58:28<23:57:55, 7.45s/step, loss=1.408, acc=0.685, acc_len=1.63, thru=14.3, I=15.9, T=29.6, wait=3.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:48,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:28:50,395] loop.py:45 INFO TRAIN_STEP step=6431 loss=1.439437 Training: 36%|███▌ | 6431/18018 [13:58:36<24:52:27, 7.73s/step, loss=1.439, acc=0.677, acc_len=1.61, thru=16.9, I=17.9, T=28.1, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:28:50,390] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6430: forward=3135.0ms backward=1361.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:28:50,424] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 03:28:56,522] loop.py:45 INFO TRAIN_STEP step=6432 loss=1.431578 Training: 36%|███▌ | 6432/18018 [13:58:42<23:19:31, 7.25s/step, loss=1.432, acc=0.682, acc_len=1.63, thru=17.6, I=22.9, T=34.4, wait=2.4s, pool=60, epoch=1/2][2026-10-19 03:29:05,067] loop.py:45 INFO TRAIN_STEP step=6433 loss=1.484509 Training: 36%|███▌ | 6433/18018 [13:58:51<24:34:42, 7.64s/step, loss=1.485, acc=0.673, acc_len=1.60, thru=15.2, I=14.9, T=31.6, wait=4.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:10,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:29:12,599] loop.py:45 INFO TRAIN_STEP step=6434 loss=1.552127 Training: 36%|███▌ | 6434/18018 [13:58:58<24:28:19, 7.61s/step, loss=1.552, acc=0.657, acc_len=1.53, thru=18.0, I=20.9, T=29.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:13,000] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:19,615] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:29:21,137] loop.py:380 INFO TIMING step=6435: step=4.609s data=2.288s compute=4.574s [fwd=3.138s bwd=1.421s opt=0.003s] dispatch=3.920s [2026-10-19 03:29:21,137] loop.py:45 INFO TRAIN_STEP step=6435 loss=1.438689 Training: 36%|███▌ | 6435/18018 [13:59:07<25:22:12, 7.89s/step, loss=1.439, acc=0.677, acc_len=1.61, thru=15.7, I=17.8, T=27.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:21,625] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:27,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:29:28,412] loop.py:45 INFO TRAIN_STEP step=6436 loss=1.375692 Training: 36%|███▌ | 6436/18018 [13:59:14<24:46:47, 7.70s/step, loss=1.376, acc=0.687, acc_len=1.64, thru=16.4, I=21.3, T=27.7, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:29:28,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6435: forward=3288.5ms backward=1288.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:28,700] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:33,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:29:34,976] loop.py:45 INFO TRAIN_STEP step=6437 loss=1.480842 Training: 36%|███▌ | 6437/18018 [13:59:21<23:40:52, 7.36s/step, loss=1.481, acc=0.668, acc_len=1.58, thru=16.2, I=23.1, T=29.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:35,076] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 03:29:41,455] loop.py:45 INFO TRAIN_STEP step=6438 loss=1.439248 Training: 36%|███▌ | 6438/18018 [13:59:27<22:49:32, 7.10s/step, loss=1.439, acc=0.675, acc_len=1.60, thru=15.1, I=21.0, T=31.3, wait=2.4s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:50,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:29:51,411] loop.py:45 INFO TRAIN_STEP step=6439 loss=1.470029 Training: 36%|███▌ | 6439/18018 [13:59:37<25:35:05, 7.95s/step, loss=1.470, acc=0.675, acc_len=1.61, thru=15.7, I=16.2, T=26.9, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:29:51,607] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:29:57,578] loop.py:380 INFO TIMING step=6440: step=4.001s data=1.640s compute=3.968s [fwd=2.673s bwd=1.281s opt=0.003s] dispatch=2.155s [2026-10-19 03:29:57,578] loop.py:45 INFO TRAIN_STEP step=6440 loss=1.489493 Training: 36%|███▌ | 6440/18018 [13:59:43<23:51:25, 7.42s/step, loss=1.489, acc=0.670, acc_len=1.59, thru=16.1, I=17.9, T=32.0, wait=2.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:03,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:30:05,928] loop.py:45 INFO TRAIN_STEP step=6441 loss=1.461301 Training: 36%|███▌ | 6441/18018 [13:59:52<24:45:15, 7.70s/step, loss=1.461, acc=0.678, acc_len=1.62, thru=17.9, I=20.3, T=26.0, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:30:05,923] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6440: forward=3374.9ms backward=1499.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:06,333] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:30:13,820] loop.py:45 INFO TRAIN_STEP step=6442 loss=1.499257 Training: 36%|███▌ | 6442/18018 [13:59:59<24:56:26, 7.76s/step, loss=1.499, acc=0.666, acc_len=1.58, thru=15.0, I=17.8, T=29.1, wait=3.5s, pool=48, epoch=1/2][2026-10-19 03:30:22,991] loop.py:45 INFO TRAIN_STEP step=6443 loss=1.514818 Training: 36%|███▌ | 6443/18018 [14:00:09<26:18:02, 8.18s/step, loss=1.515, acc=0.663, acc_len=1.56, thru=14.4, I=13.0, T=30.4, wait=4.9s, pool=56, epoch=1/2][2026-10-19 03:30:29,908] loop.py:45 INFO TRAIN_STEP step=6444 loss=1.458820 Training: 36%|███▌ | 6444/18018 [14:00:15<25:04:45, 7.80s/step, loss=1.459, acc=0.674, acc_len=1.59, thru=19.1, I=16.2, T=30.8, wait=2.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:37,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:30:38,575] loop.py:380 INFO TIMING step=6445: step=4.158s data=1.903s compute=4.122s [fwd=2.905s bwd=1.202s opt=0.003s] dispatch=4.502s [2026-10-19 03:30:38,575] loop.py:45 INFO TRAIN_STEP step=6445 loss=1.495579 Training: 36%|███▌ | 6445/18018 [14:00:24<25:54:56, 8.06s/step, loss=1.496, acc=0.670, acc_len=1.58, thru=15.2, I=16.2, T=30.8, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:38,648] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:43,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:30:44,741] loop.py:45 INFO TRAIN_STEP step=6446 loss=1.527914 Training: 36%|███▌ | 6446/18018 [14:00:30<24:05:00, 7.49s/step, loss=1.528, acc=0.663, acc_len=1.55, thru=19.2, I=23.4, T=32.1, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:30:44,734] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6445: forward=2675.0ms backward=1257.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:45,159] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:51,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:30:51,324] loop.py:45 INFO TRAIN_STEP step=6447 loss=1.507394 Training: 36%|███▌ | 6447/18018 [14:00:37<23:12:16, 7.22s/step, loss=1.507, acc=0.665, acc_len=1.57, thru=18.3, I=20.5, T=31.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:51,752] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:57,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:30:58,307] loop.py:45 INFO TRAIN_STEP step=6448 loss=1.467527 Training: 36%|███▌ | 6448/18018 [14:00:44<22:58:31, 7.15s/step, loss=1.468, acc=0.669, acc_len=1.58, thru=17.7, I=20.2, T=32.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:30:58,453] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:04,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:31:06,761] loop.py:45 INFO TRAIN_STEP step=6449 loss=1.437427 Training: 36%|███▌ | 6449/18018 [14:00:52<24:14:00, 7.54s/step, loss=1.437, acc=0.681, acc_len=1.62, thru=14.2, I=18.1, T=29.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:06,911] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:13,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:31:14,986] loop.py:380 INFO TIMING step=6450: step=4.644s data=2.224s compute=4.617s [fwd=3.284s bwd=1.319s opt=0.003s] dispatch=3.571s [2026-10-19 03:31:14,987] loop.py:45 INFO TRAIN_STEP step=6450 loss=1.623760 Training: 36%|███▌ | 6450/18018 [14:01:01<24:53:27, 7.75s/step, loss=1.624, acc=0.646, acc_len=1.51, thru=15.9, I=16.8, T=27.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:15,309] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:21,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:31:24,631] loop.py:45 INFO TRAIN_STEP step=6451 loss=1.322315 Training: 36%|███▌ | 6451/18018 [14:01:10<26:43:06, 8.32s/step, loss=1.322, acc=0.699, acc_len=1.68, thru=13.7, I=17.1, T=22.4, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:31:24,627] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6450: forward=3995.5ms backward=1661.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:24,990] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:29,947] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:31:31,486] loop.py:45 INFO TRAIN_STEP step=6452 loss=1.435317 Training: 36%|███▌ | 6452/18018 [14:01:17<25:18:22, 7.88s/step, loss=1.435, acc=0.678, acc_len=1.61, thru=13.7, I=23.7, T=30.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:31,957] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:31:38,206] loop.py:45 INFO TRAIN_STEP step=6453 loss=1.438689 Training: 36%|███▌ | 6453/18018 [14:01:24<24:11:25, 7.53s/step, loss=1.439, acc=0.677, acc_len=1.60, thru=17.2, I=24.2, T=33.7, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:46,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:31:48,368] loop.py:45 INFO TRAIN_STEP step=6454 loss=1.463892 Training: 36%|███▌ | 6454/18018 [14:01:34<26:43:27, 8.32s/step, loss=1.464, acc=0.681, acc_len=1.62, thru=13.7, I=14.6, T=27.6, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:48,465] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:54,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:31:54,711] loop.py:380 INFO TIMING step=6455: step=4.486s data=2.141s compute=4.452s [fwd=3.143s bwd=1.294s opt=0.003s] dispatch=1.848s [2026-10-19 03:31:54,712] loop.py:45 INFO TRAIN_STEP step=6455 loss=1.349028 Training: 36%|███▌ | 6455/18018 [14:01:40<24:49:03, 7.73s/step, loss=1.349, acc=0.695, acc_len=1.67, thru=17.4, I=21.7, T=28.5, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:31:54,949] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:32:02,041] loop.py:45 INFO TRAIN_STEP step=6456 loss=1.401074 Training: 36%|███▌ | 6456/18018 [14:01:48<24:25:53, 7.61s/step, loss=1.401, acc=0.684, acc_len=1.63, thru=15.3, I=16.0, T=29.6, wait=3.0s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:32:02,037] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6455: forward=3051.4ms backward=1227.9ms [2026-10-19 03:32:10,183] loop.py:45 INFO TRAIN_STEP step=6457 loss=1.411325 Training: 36%|███▌ | 6457/18018 [14:01:56<24:56:52, 7.77s/step, loss=1.411, acc=0.679, acc_len=1.61, thru=16.8, I=15.5, T=28.8, wait=3.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:15,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:32:17,071] loop.py:45 INFO TRAIN_STEP step=6458 loss=1.420420 Training: 36%|███▌ | 6458/18018 [14:02:03<24:05:46, 7.50s/step, loss=1.420, acc=0.677, acc_len=1.61, thru=17.8, I=19.9, T=31.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:17,212] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:23,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:32:24,145] loop.py:45 INFO TRAIN_STEP step=6459 loss=1.258420 Training: 36%|███▌ | 6459/18018 [14:02:10<23:40:45, 7.37s/step, loss=1.258, acc=0.711, acc_len=1.72, thru=16.9, I=23.5, T=27.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:24,213] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:31,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:32:32,977] loop.py:380 INFO TIMING step=6460: step=4.195s data=1.936s compute=4.158s [fwd=2.988s bwd=1.155s opt=0.003s] dispatch=4.628s [2026-10-19 03:32:32,977] loop.py:45 INFO TRAIN_STEP step=6460 loss=1.635207 Training: 36%|███▌ | 6460/18018 [14:02:19<25:04:51, 7.81s/step, loss=1.635, acc=0.647, acc_len=1.51, thru=13.8, I=15.4, T=30.5, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:33,436] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:38,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:32:40,021] loop.py:45 INFO TRAIN_STEP step=6461 loss=1.482724 Training: 36%|███▌ | 6461/18018 [14:02:26<24:20:19, 7.58s/step, loss=1.483, acc=0.669, acc_len=1.59, thru=16.5, I=23.5, T=30.2, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:32:40,016] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6460: forward=2941.9ms backward=1247.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:40,204] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:46,888] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:32:47,884] loop.py:45 INFO TRAIN_STEP step=6462 loss=1.412985 Training: 36%|███▌ | 6462/18018 [14:02:33<24:36:36, 7.67s/step, loss=1.413, acc=0.683, acc_len=1.63, thru=14.5, I=19.1, T=30.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:32:47,921] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 03:32:54,729] loop.py:45 INFO TRAIN_STEP step=6463 loss=1.477208 Training: 36%|███▌ | 6463/18018 [14:02:40<23:48:50, 7.42s/step, loss=1.477, acc=0.671, acc_len=1.59, thru=17.5, I=16.9, T=35.1, wait=3.2s, pool=48, epoch=1/2][2026-10-19 03:33:03,418] loop.py:45 INFO TRAIN_STEP step=6464 loss=1.539826 Training: 36%|███▌ | 6464/18018 [14:02:49<25:02:03, 7.80s/step, loss=1.540, acc=0.661, acc_len=1.55, thru=16.3, I=15.2, T=32.5, wait=4.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:08,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:33:10,670] loop.py:380 INFO TIMING step=6465: step=4.155s data=1.922s compute=4.117s [fwd=2.927s bwd=1.175s opt=0.003s] dispatch=3.088s [2026-10-19 03:33:10,670] loop.py:45 INFO TRAIN_STEP step=6465 loss=1.487442 Training: 36%|███▌ | 6465/18018 [14:02:56<24:30:23, 7.64s/step, loss=1.487, acc=0.670, acc_len=1.58, thru=19.8, I=19.9, T=30.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:10,830] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:16,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:33:17,560] loop.py:45 INFO TRAIN_STEP step=6466 loss=1.548037 Training: 36%|███▌ | 6466/18018 [14:03:03<23:47:03, 7.41s/step, loss=1.548, acc=0.661, acc_len=1.55, thru=17.7, I=22.6, T=28.6, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:33:17,555] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6465: forward=3057.3ms backward=1378.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:17,754] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:25,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:33:26,505] loop.py:45 INFO TRAIN_STEP step=6467 loss=1.459924 Training: 36%|███▌ | 6467/18018 [14:03:12<25:15:36, 7.87s/step, loss=1.460, acc=0.673, acc_len=1.60, thru=13.6, I=17.2, T=29.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:26,543] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:32,155] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:33:34,736] loop.py:45 INFO TRAIN_STEP step=6468 loss=1.375288 Training: 36%|███▌ | 6468/18018 [14:03:20<25:36:08, 7.98s/step, loss=1.375, acc=0.688, acc_len=1.64, thru=13.6, I=20.6, T=22.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:35,181] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:41,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:33:42,545] loop.py:45 INFO TRAIN_STEP step=6469 loss=1.509015 Training: 36%|███▌ | 6469/18018 [14:03:28<25:26:08, 7.93s/step, loss=1.509, acc=0.667, acc_len=1.57, thru=17.4, I=20.8, T=29.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:42,863] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:48,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:33:50,235] loop.py:380 INFO TIMING step=6470: step=4.718s data=2.194s compute=4.684s [fwd=3.266s bwd=1.403s opt=0.003s] dispatch=2.960s [2026-10-19 03:33:50,235] loop.py:45 INFO TRAIN_STEP step=6470 loss=1.423330 Training: 36%|███▌ | 6470/18018 [14:03:36<25:12:03, 7.86s/step, loss=1.423, acc=0.679, acc_len=1.62, thru=15.4, I=17.6, T=27.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:33:50,679] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:33:57,258] loop.py:45 INFO TRAIN_STEP step=6471 loss=1.469383 Training: 36%|███▌ | 6471/18018 [14:03:43<24:23:53, 7.61s/step, loss=1.469, acc=0.676, acc_len=1.61, thru=14.4, I=19.5, T=33.9, wait=3.2s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:33:57,254] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6470: forward=2607.2ms backward=1114.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:08,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:34:10,696] loop.py:45 INFO TRAIN_STEP step=6472 loss=1.565456 Training: 36%|███▌ | 6472/18018 [14:03:56<30:00:26, 9.36s/step, loss=1.565, acc=0.657, acc_len=1.54, thru=12.2, I=12.8, T=26.2, wait=8.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:11,144] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:17,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:34:18,176] loop.py:45 INFO TRAIN_STEP step=6473 loss=1.534361 Training: 36%|███▌ | 6473/18018 [14:04:04<28:11:58, 8.79s/step, loss=1.534, acc=0.663, acc_len=1.56, thru=14.9, I=20.4, T=32.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:18,550] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:25,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:34:25,398] loop.py:45 INFO TRAIN_STEP step=6474 loss=1.534948 Training: 36%|███▌ | 6474/18018 [14:04:11<26:41:12, 8.32s/step, loss=1.535, acc=0.663, acc_len=1.56, thru=16.9, I=19.5, T=31.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:25,574] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:34:32,702] loop.py:380 INFO TIMING step=6475: step=4.003s data=1.718s compute=3.973s [fwd=2.789s bwd=1.169s opt=0.003s] dispatch=3.290s [2026-10-19 03:34:32,702] loop.py:45 INFO TRAIN_STEP step=6475 loss=1.346229 Training: 36%|███▌ | 6475/18018 [14:04:18<25:42:12, 8.02s/step, loss=1.346, acc=0.696, acc_len=1.67, thru=15.4, I=15.4, T=32.0, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:38,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:34:40,705] loop.py:45 INFO TRAIN_STEP step=6476 loss=1.497087 Training: 36%|███▌ | 6476/18018 [14:04:26<25:41:21, 8.01s/step, loss=1.497, acc=0.670, acc_len=1.58, thru=18.3, I=18.6, T=27.3, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:34:40,700] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6475: forward=3222.6ms backward=1419.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:41,113] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:34:47,210] loop.py:45 INFO TRAIN_STEP step=6477 loss=1.431044 Training: 36%|███▌ | 6477/18018 [14:04:33<24:14:23, 7.56s/step, loss=1.431, acc=0.680, acc_len=1.61, thru=15.7, I=21.7, T=33.3, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:56,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:34:57,110] loop.py:45 INFO TRAIN_STEP step=6478 loss=1.488240 Training: 36%|███▌ | 6478/18018 [14:04:43<26:28:58, 8.26s/step, loss=1.488, acc=0.673, acc_len=1.59, thru=14.2, I=15.7, T=27.0, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:34:57,481] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 03:35:05,469] loop.py:45 INFO TRAIN_STEP step=6479 loss=1.548856 Training: 36%|███▌ | 6479/18018 [14:04:51<26:34:32, 8.29s/step, loss=1.549, acc=0.659, acc_len=1.55, thru=14.0, I=15.3, T=29.9, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:14,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:35:14,688] loop.py:380 INFO TIMING step=6480: step=4.354s data=2.129s compute=4.218s [fwd=2.830s bwd=1.373s opt=0.003s] dispatch=4.854s [2026-10-19 03:35:14,688] loop.py:45 INFO TRAIN_STEP step=6480 loss=1.568545 Training: 36%|███▌ | 6480/18018 [14:05:00<27:27:50, 8.57s/step, loss=1.569, acc=0.660, acc_len=1.55, thru=14.0, I=14.6, T=29.4, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:15,139] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:20,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:35:21,333] loop.py:45 INFO TRAIN_STEP step=6481 loss=1.403865 Training: 36%|███▌ | 6481/18018 [14:05:07<25:36:50, 7.99s/step, loss=1.404, acc=0.686, acc_len=1.64, thru=18.1, I=20.3, T=32.4, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:35:21,328] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6480: forward=2691.4ms backward=1195.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:21,704] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 03:35:29,107] loop.py:45 INFO TRAIN_STEP step=6482 loss=1.503237 Training: 36%|███▌ | 6482/18018 [14:05:15<25:24:06, 7.93s/step, loss=1.503, acc=0.666, acc_len=1.58, thru=13.0, I=16.9, T=30.7, wait=3.6s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:35,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:35:36,915] loop.py:45 INFO TRAIN_STEP step=6483 loss=1.492156 Training: 36%|███▌ | 6483/18018 [14:05:22<25:16:57, 7.89s/step, loss=1.492, acc=0.668, acc_len=1.58, thru=17.4, I=17.5, T=34.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:37,034] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:42,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:35:44,174] loop.py:45 INFO TRAIN_STEP step=6484 loss=1.497308 Training: 36%|███▌ | 6484/18018 [14:05:30<24:40:26, 7.70s/step, loss=1.497, acc=0.669, acc_len=1.58, thru=17.6, I=24.1, T=30.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:44,680] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:50,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:35:50,825] loop.py:380 INFO TIMING step=6485: step=3.920s data=1.860s compute=3.893s [fwd=2.695s bwd=1.185s opt=0.003s] dispatch=2.722s [2026-10-19 03:35:50,825] loop.py:45 INFO TRAIN_STEP step=6485 loss=1.409916 Training: 36%|███▌ | 6485/18018 [14:05:36<23:39:40, 7.39s/step, loss=1.410, acc=0.683, acc_len=1.63, thru=17.3, I=21.9, T=32.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:51,237] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:57,736] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:35:58,158] loop.py:45 INFO TRAIN_STEP step=6486 loss=1.456813 Training: 36%|███▌ | 6486/18018 [14:05:44<23:36:34, 7.37s/step, loss=1.457, acc=0.678, acc_len=1.61, thru=18.1, I=17.8, T=31.8, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:35:58,152] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6485: forward=2791.2ms backward=1181.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:35:58,252] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:06,136] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:36:07,731] loop.py:45 INFO TRAIN_STEP step=6487 loss=1.470230 Training: 36%|███▌ | 6487/18018 [14:05:53<25:43:25, 8.03s/step, loss=1.470, acc=0.674, acc_len=1.60, thru=13.2, I=16.5, T=29.7, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:08,154] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:16,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:36:16,185] loop.py:45 INFO TRAIN_STEP step=6488 loss=1.428398 Training: 36%|███▌ | 6488/18018 [14:06:02<26:07:39, 8.16s/step, loss=1.428, acc=0.680, acc_len=1.61, thru=15.9, I=17.4, T=33.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:16,519] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:21,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:36:23,776] loop.py:45 INFO TRAIN_STEP step=6489 loss=1.325129 Training: 36%|███▌ | 6489/18018 [14:06:09<25:34:52, 7.99s/step, loss=1.325, acc=0.697, acc_len=1.68, thru=17.0, I=23.3, T=25.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:24,218] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 03:36:31,206] loop.py:380 INFO TIMING step=6490: step=3.845s data=1.613s compute=3.811s [fwd=2.688s bwd=1.108s opt=0.003s] dispatch=3.576s [2026-10-19 03:36:31,206] loop.py:45 INFO TRAIN_STEP step=6490 loss=1.615166 Training: 36%|███▌ | 6490/18018 [14:06:17<25:02:35, 7.82s/step, loss=1.615, acc=0.650, acc_len=1.52, thru=17.8, I=18.5, T=33.3, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:36,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:36:38,087] loop.py:45 INFO TRAIN_STEP step=6491 loss=1.433052 Training: 36%|███▌ | 6491/18018 [14:06:24<24:08:23, 7.54s/step, loss=1.433, acc=0.678, acc_len=1.62, thru=19.6, I=20.2, T=27.6, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:36:38,082] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6490: forward=3126.1ms backward=1455.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:38,188] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-19 03:36:45,592] loop.py:45 INFO TRAIN_STEP step=6492 loss=1.311405 Training: 36%|███▌ | 6492/18018 [14:06:31<24:06:14, 7.53s/step, loss=1.311, acc=0.697, acc_len=1.69, thru=14.2, I=18.1, T=27.1, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:53,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:36:55,096] loop.py:45 INFO TRAIN_STEP step=6493 loss=1.504882 Training: 36%|███▌ | 6493/18018 [14:06:41<25:59:58, 8.12s/step, loss=1.505, acc=0.671, acc_len=1.58, thru=14.7, I=16.6, T=28.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:36:55,547] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:01,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:37:03,997] loop.py:45 INFO TRAIN_STEP step=6494 loss=1.418725 Training: 36%|███▌ | 6494/18018 [14:06:50<26:44:46, 8.36s/step, loss=1.419, acc=0.680, acc_len=1.62, thru=15.2, I=20.8, T=25.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:04,251] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:10,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:37:10,550] loop.py:380 INFO TIMING step=6495: step=4.040s data=1.870s compute=3.998s [fwd=2.717s bwd=1.267s opt=0.003s] dispatch=2.503s [2026-10-19 03:37:10,550] loop.py:45 INFO TRAIN_STEP step=6495 loss=1.430065 Training: 36%|███▌ | 6495/18018 [14:06:56<25:00:47, 7.81s/step, loss=1.430, acc=0.675, acc_len=1.60, thru=15.2, I=19.8, T=31.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:10,735] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:17,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:37:20,623] loop.py:45 INFO TRAIN_STEP step=6496 loss=1.600596 Training: 36%|███▌ | 6496/18018 [14:07:06<27:10:36, 8.49s/step, loss=1.601, acc=0.652, acc_len=1.53, thru=12.9, I=17.8, T=24.3, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:37:20,617] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6495: forward=3603.7ms backward=1611.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:20,750] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:26,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:37:27,620] loop.py:45 INFO TRAIN_STEP step=6497 loss=1.455129 Training: 36%|███▌ | 6497/18018 [14:07:13<25:44:36, 8.04s/step, loss=1.455, acc=0.674, acc_len=1.59, thru=14.0, I=23.4, T=33.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:27,770] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:34,979] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:37:35,072] loop.py:45 INFO TRAIN_STEP step=6498 loss=1.399485 Training: 36%|███▌ | 6498/18018 [14:07:21<25:10:10, 7.87s/step, loss=1.399, acc=0.690, acc_len=1.65, thru=17.8, I=20.6, T=33.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:35,483] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:37:41,821] loop.py:45 INFO TRAIN_STEP step=6499 loss=1.484089 Training: 36%|███▌ | 6499/18018 [14:07:27<24:05:43, 7.53s/step, loss=1.484, acc=0.668, acc_len=1.58, thru=17.7, I=18.6, T=31.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:41,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:42,315] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:47,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:37:50,057] loop.py:380 INFO TIMING step=6500: step=4.837s data=2.323s compute=4.804s [fwd=3.336s bwd=1.453s opt=0.003s] dispatch=3.392s [2026-10-19 03:37:50,057] loop.py:45 INFO TRAIN_STEP step=6500 loss=1.516834 Training: 36%|███▌ | 6500/18018 [14:07:36<24:46:16, 7.74s/step, loss=1.517, acc=0.668, acc_len=1.58, thru=16.2, I=21.2, T=26.5, wait=3.4s, pool=64, epoch=1/2][2026-10-19 03:37:50,059] loop.py:419 INFO Saving checkpoint at step 6500... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:52,560] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 03:37:53,488] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0006501 (TrainerActor pid=9310) [2026-10-19 03:37:53,497] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0005751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:55,077] inference_manager.py:408 INFO Pool capacity available after 7.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:37:59,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 03:37:53,489] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0006501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 03:37:54,637] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0005751 [2026-10-19 03:38:01,621] loop.py:45 INFO TRAIN_STEP step=6501 loss=1.470436 Training: 36%|███▌ | 6501/18018 [14:07:47<28:26:21, 8.89s/step, loss=1.470, acc=0.676, acc_len=1.60, thru=19.6, I=28.3, T=30.5, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:38:01,616] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6500: forward=2844.6ms backward=1284.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:01,875] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:07,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:38:10,221] loop.py:45 INFO TRAIN_STEP step=6502 loss=1.510051 Training: 36%|███▌ | 6502/18018 [14:07:56<28:09:33, 8.80s/step, loss=1.510, acc=0.669, acc_len=1.57, thru=15.3, I=19.8, T=25.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:10,285] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 [2026-10-19 03:38:16,008] loop.py:45 INFO TRAIN_STEP step=6503 loss=1.407584 Training: 36%|███▌ | 6503/18018 [14:08:02<25:15:37, 7.90s/step, loss=1.408, acc=0.687, acc_len=1.64, thru=16.5, I=22.0, T=32.4, wait=1.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:24,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:38:25,018] loop.py:45 INFO TRAIN_STEP step=6504 loss=1.371552 Training: 36%|███▌ | 6504/18018 [14:08:11<26:19:40, 8.23s/step, loss=1.372, acc=0.690, acc_len=1.65, thru=15.8, I=15.0, T=26.0, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:25,429] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:29,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:38:32,697] loop.py:380 INFO TIMING step=6505: step=5.073s data=2.591s compute=5.043s [fwd=3.461s bwd=1.567s opt=0.003s] dispatch=2.596s [2026-10-19 03:38:32,697] loop.py:45 INFO TRAIN_STEP step=6505 loss=1.511349 Training: 36%|███▌ | 6505/18018 [14:08:18<25:47:43, 8.07s/step, loss=1.511, acc=0.670, acc_len=1.58, thru=15.2, I=18.6, T=25.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:32,890] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:37,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:38:39,268] loop.py:45 INFO TRAIN_STEP step=6506 loss=1.483744 Training: 36%|███▌ | 6506/18018 [14:08:25<24:21:23, 7.62s/step, loss=1.484, acc=0.669, acc_len=1.58, thru=14.5, I=23.8, T=29.4, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:38:39,263] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6505: forward=3205.7ms backward=1108.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:39,695] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:48,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:38:48,547] loop.py:45 INFO TRAIN_STEP step=6507 loss=1.561441 Training: 36%|███▌ | 6507/18018 [14:08:34<25:56:59, 8.12s/step, loss=1.561, acc=0.656, acc_len=1.54, thru=14.5, I=15.7, T=26.5, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:48,650] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:55,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:38:57,597] loop.py:45 INFO TRAIN_STEP step=6508 loss=1.498142 Training: 36%|███▌ | 6508/18018 [14:08:43<26:50:36, 8.40s/step, loss=1.498, acc=0.670, acc_len=1.59, thru=14.3, I=17.5, T=28.8, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:38:57,798] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:39:06,762] loop.py:45 INFO TRAIN_STEP step=6509 loss=1.384482 Training: 36%|███▌ | 6509/18018 [14:08:52<27:34:42, 8.63s/step, loss=1.384, acc=0.687, acc_len=1.64, thru=13.8, I=13.4, T=26.9, wait=4.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:15,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:39:17,875] loop.py:380 INFO TIMING step=6510: step=5.455s data=2.846s compute=5.404s [fwd=3.988s bwd=1.400s opt=0.003s] dispatch=5.650s [2026-10-19 03:39:17,876] loop.py:45 INFO TRAIN_STEP step=6510 loss=1.448232 Training: 36%|███▌ | 6510/18018 [14:09:03<29:57:41, 9.37s/step, loss=1.448, acc=0.681, acc_len=1.62, thru=10.5, I=13.9, T=23.5, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:17,985] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 [2026-10-19 03:39:25,882] loop.py:45 INFO TRAIN_STEP step=6511 loss=1.519379 Training: 36%|███▌ | 6511/18018 [14:09:11<28:38:55, 8.96s/step, loss=1.519, acc=0.665, acc_len=1.57, thru=13.2, I=16.0, T=30.6, wait=3.8s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:39:25,877] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6510: forward=2980.2ms backward=1165.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:35,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:39:36,246] loop.py:45 INFO TRAIN_STEP step=6512 loss=1.465216 Training: 36%|███▌ | 6512/18018 [14:09:22<29:59:20, 9.38s/step, loss=1.465, acc=0.675, acc_len=1.60, thru=15.9, I=13.3, T=28.7, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:36,728] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:42,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:39:44,603] loop.py:45 INFO TRAIN_STEP step=6513 loss=1.528968 Training: 36%|███▌ | 6513/18018 [14:09:30<29:00:14, 9.08s/step, loss=1.529, acc=0.662, acc_len=1.56, thru=14.8, I=19.4, T=25.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:44,839] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:50,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:39:51,084] loop.py:45 INFO TRAIN_STEP step=6514 loss=1.515470 Training: 36%|███▌ | 6514/18018 [14:09:37<26:30:54, 8.30s/step, loss=1.515, acc=0.665, acc_len=1.57, thru=16.9, I=21.2, T=30.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:51,156] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:55,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:39:57,200] loop.py:380 INFO TIMING step=6515: step=4.215s data=1.908s compute=4.185s [fwd=3.022s bwd=1.147s opt=0.003s] dispatch=1.891s [2026-10-19 03:39:57,200] loop.py:45 INFO TRAIN_STEP step=6515 loss=1.480859 Training: 36%|███▌ | 6515/18018 [14:09:43<24:25:10, 7.64s/step, loss=1.481, acc=0.666, acc_len=1.57, thru=18.9, I=24.3, T=30.4, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:39:57,628] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:03,401] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 03:40:04,768] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6515: forward=3104.8ms backward=1356.7ms [2026-10-19 03:40:04,773] loop.py:45 INFO TRAIN_STEP step=6516 loss=1.476182 Training: 36%|███▌ | 6516/18018 [14:09:50<24:21:05, 7.62s/step, loss=1.476, acc=0.673, acc_len=1.60, thru=13.7, I=20.4, T=28.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:04,916] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-19 03:40:12,925] loop.py:45 INFO TRAIN_STEP step=6517 loss=1.543632 Training: 36%|███▌ | 6517/18018 [14:09:59<24:51:26, 7.78s/step, loss=1.544, acc=0.663, acc_len=1.57, thru=12.9, I=13.9, T=30.7, wait=4.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:19,760] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:40:22,078] loop.py:45 INFO TRAIN_STEP step=6518 loss=1.445355 Training: 36%|███▌ | 6518/18018 [14:10:08<26:10:14, 8.19s/step, loss=1.445, acc=0.677, acc_len=1.61, thru=15.6, I=18.9, T=24.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:22,277] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:28,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:40:29,464] loop.py:45 INFO TRAIN_STEP step=6519 loss=1.477572 Training: 36%|███▌ | 6519/18018 [14:10:15<25:23:48, 7.95s/step, loss=1.478, acc=0.671, acc_len=1.59, thru=13.6, I=18.7, T=29.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:29,608] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:34,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:40:36,377] loop.py:380 INFO TIMING step=6520: step=4.143s data=1.862s compute=4.111s [fwd=2.890s bwd=1.206s opt=0.003s] dispatch=2.759s [2026-10-19 03:40:36,377] loop.py:45 INFO TRAIN_STEP step=6520 loss=1.464800 Training: 36%|███▌ | 6520/18018 [14:10:22<24:23:54, 7.64s/step, loss=1.465, acc=0.675, acc_len=1.60, thru=17.2, I=22.8, T=30.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:36,744] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:42,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:40:45,129] loop.py:45 INFO TRAIN_STEP step=6521 loss=1.351973 Training: 36%|███▌ | 6521/18018 [14:10:31<25:27:41, 7.97s/step, loss=1.352, acc=0.694, acc_len=1.67, thru=15.3, I=22.0, T=29.1, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:40:45,124] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6520: forward=3085.3ms backward=1252.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:45,585] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:51,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:40:52,599] loop.py:45 INFO TRAIN_STEP step=6522 loss=1.489161 Training: 36%|███▌ | 6522/18018 [14:10:38<24:58:51, 7.82s/step, loss=1.489, acc=0.668, acc_len=1.58, thru=18.2, I=23.6, T=25.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:40:52,824] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:00,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:41:01,229] loop.py:45 INFO TRAIN_STEP step=6523 loss=1.574180 Training: 36%|███▌ | 6523/18018 [14:10:47<25:45:02, 8.06s/step, loss=1.574, acc=0.660, acc_len=1.55, thru=13.3, I=13.8, T=28.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:01,413] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:07,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:41:07,873] loop.py:45 INFO TRAIN_STEP step=6524 loss=1.471565 Training: 36%|███▌ | 6524/18018 [14:10:53<24:23:11, 7.64s/step, loss=1.472, acc=0.675, acc_len=1.59, thru=18.8, I=17.4, T=31.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:08,066] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:12,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:41:14,179] loop.py:380 INFO TIMING step=6525: step=4.006s data=1.777s compute=3.977s [fwd=2.771s bwd=1.191s opt=0.003s] dispatch=2.291s [2026-10-19 03:41:14,179] loop.py:45 INFO TRAIN_STEP step=6525 loss=1.459776 Training: 36%|███▌ | 6525/18018 [14:11:00<23:06:34, 7.24s/step, loss=1.460, acc=0.675, acc_len=1.60, thru=18.1, I=25.1, T=32.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:14,393] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:41:22,022] loop.py:45 INFO TRAIN_STEP step=6526 loss=1.372955 Training: 36%|███▌ | 6526/18018 [14:11:08<23:41:04, 7.42s/step, loss=1.373, acc=0.691, acc_len=1.65, thru=12.9, I=16.1, T=34.9, wait=4.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:41:22,017] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6525: forward=2543.1ms backward=1067.3ms [2026-10-19 03:41:31,620] loop.py:45 INFO TRAIN_STEP step=6527 loss=1.472470 Training: 36%|███▌ | 6527/18018 [14:11:17<25:46:09, 8.07s/step, loss=1.472, acc=0.675, acc_len=1.60, thru=13.4, I=13.5, T=29.7, wait=5.3s, pool=48, epoch=1/2][2026-10-19 03:41:41,760] loop.py:45 INFO TRAIN_STEP step=6528 loss=1.404681 Training: 36%|███▌ | 6528/18018 [14:11:27<27:44:49, 8.69s/step, loss=1.405, acc=0.685, acc_len=1.64, thru=13.3, I=12.6, T=32.4, wait=6.2s, pool=52, epoch=1/2][2026-10-19 03:41:48,816] loop.py:45 INFO TRAIN_STEP step=6529 loss=1.446173 Training: 36%|███▌ | 6529/18018 [14:11:34<26:10:32, 8.20s/step, loss=1.446, acc=0.679, acc_len=1.61, thru=18.6, I=17.7, T=33.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:48,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:49,313] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:54,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:41:55,769] loop.py:380 INFO TIMING step=6530: step=4.030s data=1.831s compute=4.000s [fwd=2.810s bwd=1.175s opt=0.003s] dispatch=2.915s [2026-10-19 03:41:55,769] loop.py:45 INFO TRAIN_STEP step=6530 loss=1.439041 Training: 36%|███▌ | 6530/18018 [14:11:41<24:58:42, 7.83s/step, loss=1.439, acc=0.679, acc_len=1.62, thru=17.4, I=19.4, T=31.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:41:55,976] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 03:42:04,750] loop.py:45 INFO TRAIN_STEP step=6531 loss=1.523882 Training: 36%|███▌ | 6531/18018 [14:11:50<26:04:51, 8.17s/step, loss=1.524, acc=0.667, acc_len=1.57, thru=12.3, I=16.3, T=29.6, wait=4.6s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:42:04,745] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6530: forward=3046.0ms backward=1221.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:16,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:42:18,359] loop.py:45 INFO TRAIN_STEP step=6532 loss=1.444116 Training: 36%|███▋ | 6532/18018 [14:12:04<31:16:57, 9.80s/step, loss=1.444, acc=0.681, acc_len=1.62, thru=12.6, I=13.2, T=29.1, wait=9.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:18,497] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:25,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:42:26,622] loop.py:45 INFO TRAIN_STEP step=6533 loss=1.516552 Training: 36%|███▋ | 6533/18018 [14:12:12<29:48:07, 9.34s/step, loss=1.517, acc=0.666, acc_len=1.57, thru=13.7, I=16.9, T=30.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:26,970] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:34,274] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:42:35,377] loop.py:45 INFO TRAIN_STEP step=6534 loss=1.656090 Training: 36%|███▋ | 6534/18018 [14:12:21<29:14:23, 9.17s/step, loss=1.656, acc=0.642, acc_len=1.49, thru=14.6, I=17.1, T=24.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:35,783] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:43,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:42:44,012] loop.py:380 INFO TIMING step=6535: step=4.202s data=1.923s compute=4.169s [fwd=3.044s bwd=1.110s opt=0.003s] dispatch=4.423s [2026-10-19 03:42:44,013] loop.py:45 INFO TRAIN_STEP step=6535 loss=1.610847 Training: 36%|███▋ | 6535/18018 [14:12:30<28:43:40, 9.01s/step, loss=1.611, acc=0.652, acc_len=1.52, thru=13.8, I=16.4, T=30.5, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:44,116] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:49,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:42:51,172] loop.py:45 INFO TRAIN_STEP step=6536 loss=1.350276 Training: 36%|███▋ | 6536/18018 [14:12:37<26:57:29, 8.45s/step, loss=1.350, acc=0.694, acc_len=1.66, thru=16.4, I=21.2, T=30.7, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:42:51,166] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6535: forward=2921.2ms backward=1189.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:42:51,262] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 03:42:59,365] loop.py:45 INFO TRAIN_STEP step=6537 loss=1.475569 Training: 36%|███▋ | 6537/18018 [14:12:45<26:42:23, 8.37s/step, loss=1.476, acc=0.676, acc_len=1.60, thru=14.3, I=15.3, T=33.4, wait=4.4s, pool=52, epoch=1/2][2026-10-19 03:43:10,525] loop.py:45 INFO TRAIN_STEP step=6538 loss=1.517758 Training: 36%|███▋ | 6538/18018 [14:12:56<29:22:10, 9.21s/step, loss=1.518, acc=0.668, acc_len=1.58, thru=10.9, I=10.1, T=28.8, wait=6.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:17,518] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:43:17,853] loop.py:45 INFO TRAIN_STEP step=6539 loss=1.566113 Training: 36%|███▋ | 6539/18018 [14:13:03<27:34:08, 8.65s/step, loss=1.566, acc=0.663, acc_len=1.55, thru=18.9, I=15.6, T=32.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:18,020] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:24,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:43:26,440] loop.py:380 INFO TIMING step=6540: step=4.619s data=2.272s compute=4.589s [fwd=3.185s bwd=1.389s opt=0.003s] dispatch=3.958s [2026-10-19 03:43:26,440] loop.py:45 INFO TRAIN_STEP step=6540 loss=1.463535 Training: 36%|███▋ | 6540/18018 [14:13:12<27:30:38, 8.63s/step, loss=1.464, acc=0.675, acc_len=1.61, thru=14.1, I=16.2, T=27.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:26,490] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:32,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:43:33,548] loop.py:45 INFO TRAIN_STEP step=6541 loss=1.427757 Training: 36%|███▋ | 6541/18018 [14:13:19<26:03:10, 8.17s/step, loss=1.428, acc=0.683, acc_len=1.63, thru=15.3, I=19.5, T=32.1, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:43:33,543] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6540: forward=2767.8ms backward=1146.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:34,018] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:39,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:43:41,863] loop.py:45 INFO TRAIN_STEP step=6542 loss=1.472129 Training: 36%|███▋ | 6542/18018 [14:13:27<26:11:09, 8.21s/step, loss=1.472, acc=0.676, acc_len=1.60, thru=13.7, I=19.5, T=27.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:42,240] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:46,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:43:48,889] loop.py:45 INFO TRAIN_STEP step=6543 loss=1.384525 Training: 36%|███▋ | 6543/18018 [14:13:34<25:02:59, 7.86s/step, loss=1.385, acc=0.683, acc_len=1.62, thru=15.4, I=26.6, T=29.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:43:49,003] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 [2026-10-19 03:43:55,824] loop.py:45 INFO TRAIN_STEP step=6544 loss=1.413738 Training: 36%|███▋ | 6544/18018 [14:13:41<24:09:46, 7.58s/step, loss=1.414, acc=0.683, acc_len=1.63, thru=14.8, I=20.8, T=31.8, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:02,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:44:04,887] loop.py:380 INFO TIMING step=6545: step=5.141s data=2.622s compute=5.095s [fwd=3.583s bwd=1.496s opt=0.003s] dispatch=3.912s [2026-10-19 03:44:04,887] loop.py:45 INFO TRAIN_STEP step=6545 loss=1.555456 Training: 36%|███▋ | 6545/18018 [14:13:50<25:34:35, 8.03s/step, loss=1.555, acc=0.661, acc_len=1.55, thru=15.0, I=18.0, T=24.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:05,339] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:11,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:44:11,580] loop.py:45 INFO TRAIN_STEP step=6546 loss=1.487150 Training: 36%|███▋ | 6546/18018 [14:13:57<24:18:10, 7.63s/step, loss=1.487, acc=0.669, acc_len=1.58, thru=14.2, I=21.3, T=30.5, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:44:11,575] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6545: forward=2908.7ms backward=1244.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:11,617] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:17,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:44:18,771] loop.py:45 INFO TRAIN_STEP step=6547 loss=1.400584 Training: 36%|███▋ | 6547/18018 [14:14:04<23:52:58, 7.50s/step, loss=1.401, acc=0.683, acc_len=1.63, thru=17.0, I=20.4, T=29.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:19,074] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:44:26,020] loop.py:45 INFO TRAIN_STEP step=6548 loss=1.461875 Training: 36%|███▋ | 6548/18018 [14:14:12<23:38:49, 7.42s/step, loss=1.462, acc=0.674, acc_len=1.61, thru=14.2, I=24.8, T=26.0, wait=2.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:32,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:44:34,583] loop.py:45 INFO TRAIN_STEP step=6549 loss=1.642807 Training: 36%|███▋ | 6549/18018 [14:14:20<24:43:59, 7.76s/step, loss=1.643, acc=0.645, acc_len=1.50, thru=16.6, I=15.2, T=30.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:34,669] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 03:44:41,635] loop.py:380 INFO TIMING step=6550: step=4.199s data=1.898s compute=4.163s [fwd=2.877s bwd=1.271s opt=0.003s] dispatch=2.845s [2026-10-19 03:44:41,635] loop.py:45 INFO TRAIN_STEP step=6550 loss=1.492246 Training: 36%|███▋ | 6550/18018 [14:14:27<24:03:27, 7.55s/step, loss=1.492, acc=0.669, acc_len=1.57, thru=14.9, I=22.2, T=30.5, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:46,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:44:48,754] loop.py:45 INFO TRAIN_STEP step=6551 loss=1.448067 Training: 36%|███▋ | 6551/18018 [14:14:34<23:38:09, 7.42s/step, loss=1.448, acc=0.677, acc_len=1.61, thru=17.3, I=20.5, T=31.5, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:44:48,750] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6550: forward=2856.1ms backward=1161.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:44:48,784] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 [2026-10-19 03:44:55,738] loop.py:45 INFO TRAIN_STEP step=6552 loss=1.279225 Training: 36%|███▋ | 6552/18018 [14:14:41<23:13:09, 7.29s/step, loss=1.279, acc=0.704, acc_len=1.70, thru=14.0, I=19.5, T=31.8, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:45:02,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:45:03,351] loop.py:45 INFO TRAIN_STEP step=6553 loss=1.449351 Training: 36%|███▋ | 6553/18018 [14:14:49<23:31:28, 7.39s/step, loss=1.449, acc=0.679, acc_len=1.62, thru=16.9, I=17.8, T=32.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:45:03,637] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 03:45:11,112] loop.py:45 INFO TRAIN_STEP step=6554 loss=1.527287 Training: 36%|███▋ | 6554/18018 [14:14:57<23:52:41, 7.50s/step, loss=1.527, acc=0.667, acc_len=1.57, thru=14.3, I=14.7, T=31.8, wait=3.7s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:45:19,279] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:45:21,757] loop.py:380 INFO TIMING step=6555: step=4.576s data=2.237s compute=4.490s [fwd=3.056s bwd=1.419s opt=0.003s] dispatch=6.061s [2026-10-19 03:45:21,758] loop.py:45 INFO TRAIN_STEP step=6555 loss=1.535195 Training: 36%|███▋ | 6555/18018 [14:15:07<26:53:03, 8.44s/step, loss=1.535, acc=0.666, acc_len=1.56, thru=15.5, I=18.0, T=28.0, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:45:21,810] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:45:25,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:45:28,187] loop.py:45 INFO TRAIN_STEP step=6556 loss=1.507671 Training: 36%|███▋ | 6556/18018 [14:15:14<24:57:23, 7.84s/step, loss=1.508, acc=0.668, acc_len=1.58, thru=17.0, I=31.2, T=29.8, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:45:28,182] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6555: forward=2923.0ms backward=1320.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:45:28,486] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:45:36,327] loop.py:45 INFO TRAIN_STEP step=6557 loss=1.510475 Training: 36%|███▋ | 6557/18018 [14:15:22<25:14:33, 7.93s/step, loss=1.510, acc=0.666, acc_len=1.58, thru=13.8, I=14.3, T=31.3, wait=4.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:45:45,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:45:45,954] loop.py:45 INFO TRAIN_STEP step=6558 loss=1.358975 Training: 36%|███▋ | 6558/18018 [14:15:32<26:51:50, 8.44s/step, loss=1.359, acc=0.691, acc_len=1.66, thru=15.6, I=14.6, T=28.9, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:45:46,240] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:45:53,944] loop.py:45 INFO TRAIN_STEP step=6559 loss=1.463923 Training: 36%|███▋ | 6559/18018 [14:15:40<26:26:05, 8.30s/step, loss=1.464, acc=0.676, acc_len=1.61, thru=15.8, I=16.6, T=30.4, wait=3.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:01,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:46:03,912] loop.py:380 INFO TIMING step=6560: step=5.366s data=2.566s compute=5.324s [fwd=3.598s bwd=1.710s opt=0.003s] dispatch=4.589s [2026-10-19 03:46:03,912] loop.py:45 INFO TRAIN_STEP step=6560 loss=1.509063 Training: 36%|███▋ | 6560/18018 [14:15:49<28:01:05, 8.80s/step, loss=1.509, acc=0.668, acc_len=1.58, thru=13.5, I=14.6, T=23.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:04,278] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:11,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:46:12,569] loop.py:45 INFO TRAIN_STEP step=6561 loss=1.477839 Training: 36%|███▋ | 6561/18018 [14:15:58<27:52:37, 8.76s/step, loss=1.478, acc=0.672, acc_len=1.59, thru=15.4, I=15.2, T=25.2, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:46:12,565] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6560: forward=3516.9ms backward=1506.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:12,761] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:18,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:46:19,127] loop.py:45 INFO TRAIN_STEP step=6562 loss=1.424536 Training: 36%|███▋ | 6562/18018 [14:16:05<25:46:14, 8.10s/step, loss=1.425, acc=0.679, acc_len=1.62, thru=17.5, I=18.8, T=32.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:19,213] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:24,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:46:25,534] loop.py:45 INFO TRAIN_STEP step=6563 loss=1.543764 Training: 36%|███▋ | 6563/18018 [14:16:11<24:09:28, 7.59s/step, loss=1.544, acc=0.661, acc_len=1.55, thru=18.6, I=22.1, T=35.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:25,572] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:32,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:46:32,490] loop.py:45 INFO TRAIN_STEP step=6564 loss=1.422729 Training: 36%|███▋ | 6564/18018 [14:16:18<23:32:47, 7.40s/step, loss=1.423, acc=0.682, acc_len=1.62, thru=17.8, I=19.2, T=33.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:32,701] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:39,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:46:41,934] loop.py:380 INFO TIMING step=6565: step=4.425s data=2.143s compute=4.388s [fwd=3.029s bwd=1.344s opt=0.003s] dispatch=5.009s [2026-10-19 03:46:41,934] loop.py:45 INFO TRAIN_STEP step=6565 loss=1.432630 Training: 36%|███▋ | 6565/18018 [14:16:28<25:29:35, 8.01s/step, loss=1.433, acc=0.682, acc_len=1.63, thru=13.9, I=17.1, T=28.9, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:41,980] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:47,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:46:49,252] loop.py:45 INFO TRAIN_STEP step=6566 loss=1.380051 Training: 36%|███▋ | 6566/18018 [14:16:35<24:49:37, 7.80s/step, loss=1.380, acc=0.687, acc_len=1.65, thru=16.0, I=23.3, T=30.7, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:46:49,248] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6565: forward=2849.8ms backward=1272.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:46:49,645] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:46:55,911] loop.py:45 INFO TRAIN_STEP step=6567 loss=1.566188 Training: 36%|███▋ | 6567/18018 [14:16:41<23:44:07, 7.46s/step, loss=1.566, acc=0.659, acc_len=1.55, thru=16.5, I=23.5, T=30.4, wait=2.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:04,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:47:06,912] loop.py:45 INFO TRAIN_STEP step=6568 loss=1.629624 Training: 36%|███▋ | 6568/18018 [14:16:52<27:06:30, 8.52s/step, loss=1.630, acc=0.653, acc_len=1.53, thru=13.4, I=15.2, T=26.6, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:07,325] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:13,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:47:14,993] loop.py:45 INFO TRAIN_STEP step=6569 loss=1.594862 Training: 36%|███▋ | 6569/18018 [14:17:01<26:40:56, 8.39s/step, loss=1.595, acc=0.650, acc_len=1.52, thru=16.7, I=19.5, T=30.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:15,392] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:20,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:47:22,201] loop.py:380 INFO TIMING step=6570: step=4.088s data=1.871s compute=4.051s [fwd=2.824s bwd=1.212s opt=0.003s] dispatch=3.112s [2026-10-19 03:47:22,201] loop.py:45 INFO TRAIN_STEP step=6570 loss=1.467126 Training: 36%|███▋ | 6570/18018 [14:17:08<25:33:14, 8.04s/step, loss=1.467, acc=0.670, acc_len=1.59, thru=16.6, I=21.8, T=31.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:22,390] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:28,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:47:29,048] loop.py:45 INFO TRAIN_STEP step=6571 loss=1.451855 Training: 36%|███▋ | 6571/18018 [14:17:15<24:25:03, 7.68s/step, loss=1.452, acc=0.673, acc_len=1.60, thru=16.1, I=22.3, T=29.3, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:47:29,044] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6570: forward=3034.7ms backward=1294.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:29,248] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:35,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:47:35,645] loop.py:45 INFO TRAIN_STEP step=6572 loss=1.463556 Training: 36%|███▋ | 6572/18018 [14:17:21<23:22:56, 7.35s/step, loss=1.464, acc=0.672, acc_len=1.59, thru=18.4, I=18.7, T=32.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:47:35,955] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:47:42,435] loop.py:45 INFO TRAIN_STEP step=6573 loss=1.419596 Training: 36%|███▋ | 6573/18018 [14:17:28<22:50:29, 7.18s/step, loss=1.420, acc=0.681, acc_len=1.62, thru=15.7, I=16.2, T=31.5, wait=2.7s, pool=40, epoch=1/2][2026-10-19 03:47:51,398] loop.py:45 INFO TRAIN_STEP step=6574 loss=1.402306 Training: 36%|███▋ | 6574/18018 [14:17:37<24:32:17, 7.72s/step, loss=1.402, acc=0.684, acc_len=1.63, thru=15.1, I=15.2, T=26.2, wait=4.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:01,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:48:04,562] loop.py:380 INFO TIMING step=6575: step=5.250s data=2.450s compute=5.201s [fwd=3.744s bwd=1.442s opt=0.003s] dispatch=7.903s [2026-10-19 03:48:04,562] loop.py:45 INFO TRAIN_STEP step=6575 loss=1.508577 Training: 36%|███▋ | 6575/18018 [14:17:50<29:43:38, 9.35s/step, loss=1.509, acc=0.666, acc_len=1.57, thru=11.8, I=13.4, T=24.4, wait=7.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:04,843] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:12,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:48:12,655] loop.py:45 INFO TRAIN_STEP step=6576 loss=1.455321 Training: 36%|███▋ | 6576/18018 [14:17:58<28:31:27, 8.97s/step, loss=1.455, acc=0.677, acc_len=1.62, thru=16.4, I=17.6, T=31.4, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:48:12,651] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6575: forward=2882.0ms backward=1145.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:12,770] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:18,033] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:48:19,211] loop.py:45 INFO TRAIN_STEP step=6577 loss=1.440424 Training: 37%|███▋ | 6577/18018 [14:18:05<26:12:55, 8.25s/step, loss=1.440, acc=0.675, acc_len=1.60, thru=18.5, I=20.5, T=30.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:19,543] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:26,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:48:27,972] loop.py:45 INFO TRAIN_STEP step=6578 loss=1.471745 Training: 37%|███▋ | 6578/18018 [14:18:14<26:41:53, 8.40s/step, loss=1.472, acc=0.672, acc_len=1.59, thru=13.3, I=16.5, T=29.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:28,451] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:34,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:48:36,674] loop.py:45 INFO TRAIN_STEP step=6579 loss=1.453838 Training: 37%|███▋ | 6579/18018 [14:18:22<26:58:59, 8.49s/step, loss=1.454, acc=0.678, acc_len=1.61, thru=14.0, I=19.5, T=25.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:36,987] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:42,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:48:43,944] loop.py:380 INFO TIMING step=6580: step=4.378s data=2.009s compute=4.339s [fwd=3.071s bwd=1.253s opt=0.003s] dispatch=2.884s [2026-10-19 03:48:43,944] loop.py:45 INFO TRAIN_STEP step=6580 loss=1.602981 Training: 37%|███▋ | 6580/18018 [14:18:30<25:49:01, 8.13s/step, loss=1.603, acc=0.653, acc_len=1.53, thru=14.3, I=21.8, T=29.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:44,086] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:49,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:48:50,320] loop.py:45 INFO TRAIN_STEP step=6581 loss=1.501719 Training: 37%|███▋ | 6581/18018 [14:18:36<24:08:48, 7.60s/step, loss=1.502, acc=0.666, acc_len=1.56, thru=18.0, I=22.8, T=35.5, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:48:50,315] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6580: forward=2535.4ms backward=1011.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:50,415] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:55,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:48:56,999] loop.py:45 INFO TRAIN_STEP step=6582 loss=1.442760 Training: 37%|███▋ | 6582/18018 [14:18:43<23:15:59, 7.32s/step, loss=1.443, acc=0.675, acc_len=1.61, thru=18.9, I=22.6, T=33.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:48:57,318] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:03,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:49:04,083] loop.py:45 INFO TRAIN_STEP step=6583 loss=1.613208 Training: 37%|███▋ | 6583/18018 [14:18:50<23:02:08, 7.25s/step, loss=1.613, acc=0.648, acc_len=1.52, thru=17.5, I=21.8, T=31.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:04,207] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:08,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:49:11,384] loop.py:45 INFO TRAIN_STEP step=6584 loss=1.489369 Training: 37%|███▋ | 6584/18018 [14:18:57<23:04:51, 7.27s/step, loss=1.489, acc=0.668, acc_len=1.58, thru=16.8, I=24.5, T=26.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:11,723] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:18,416] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:49:19,228] loop.py:380 INFO TIMING step=6585: step=4.520s data=2.134s compute=4.476s [fwd=3.166s bwd=1.297s opt=0.003s] dispatch=3.314s [2026-10-19 03:49:19,229] loop.py:45 INFO TRAIN_STEP step=6585 loss=1.394256 Training: 37%|███▋ | 6585/18018 [14:19:05<23:37:40, 7.44s/step, loss=1.394, acc=0.689, acc_len=1.65, thru=17.1, I=20.6, T=28.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:19,421] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:29,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:49:29,794] loop.py:45 INFO TRAIN_STEP step=6586 loss=1.677145 Training: 37%|███▋ | 6586/18018 [14:19:15<26:36:22, 8.38s/step, loss=1.677, acc=0.641, acc_len=1.49, thru=12.9, I=11.6, T=26.3, wait=5.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:49:29,790] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6585: forward=3504.9ms backward=1309.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:29,878] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-19 03:49:38,746] loop.py:45 INFO TRAIN_STEP step=6587 loss=1.499664 Training: 37%|███▋ | 6587/18018 [14:19:24<27:08:53, 8.55s/step, loss=1.500, acc=0.670, acc_len=1.59, thru=13.7, I=13.9, T=27.3, wait=4.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:45,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:49:46,978] loop.py:45 INFO TRAIN_STEP step=6588 loss=1.556924 Training: 37%|███▋ | 6588/18018 [14:19:33<26:50:27, 8.45s/step, loss=1.557, acc=0.654, acc_len=1.53, thru=16.5, I=18.4, T=28.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:47,053] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:54,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:49:55,371] loop.py:45 INFO TRAIN_STEP step=6589 loss=1.524260 Training: 37%|███▋ | 6589/18018 [14:19:41<26:46:57, 8.44s/step, loss=1.524, acc=0.665, acc_len=1.57, thru=14.1, I=20.3, T=29.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:49:55,683] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:01,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:50:03,564] loop.py:380 INFO TIMING step=6590: step=4.322s data=1.847s compute=4.295s [fwd=2.942s bwd=1.338s opt=0.003s] dispatch=3.862s [2026-10-19 03:50:03,564] loop.py:45 INFO TRAIN_STEP step=6590 loss=1.402680 Training: 37%|███▋ | 6590/18018 [14:19:49<26:32:51, 8.36s/step, loss=1.403, acc=0.685, acc_len=1.64, thru=14.3, I=17.3, T=29.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:03,674] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:07,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:50:10,206] loop.py:45 INFO TRAIN_STEP step=6591 loss=1.282359 Training: 37%|███▋ | 6591/18018 [14:19:56<24:54:25, 7.85s/step, loss=1.282, acc=0.705, acc_len=1.70, thru=16.9, I=26.7, T=29.6, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:50:10,201] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6590: forward=2943.4ms backward=1322.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:10,350] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:16,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:50:17,636] loop.py:45 INFO TRAIN_STEP step=6592 loss=1.414677 Training: 37%|███▋ | 6592/18018 [14:20:03<24:30:28, 7.72s/step, loss=1.415, acc=0.682, acc_len=1.62, thru=13.7, I=18.5, T=31.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:18,065] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:22,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:50:24,700] loop.py:45 INFO TRAIN_STEP step=6593 loss=1.484591 Training: 37%|███▋ | 6593/18018 [14:20:10<23:52:44, 7.52s/step, loss=1.485, acc=0.668, acc_len=1.57, thru=17.7, I=23.4, T=29.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:25,161] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:29,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:50:32,130] loop.py:45 INFO TRAIN_STEP step=6594 loss=1.567292 Training: 37%|███▋ | 6594/18018 [14:20:18<23:47:22, 7.50s/step, loss=1.567, acc=0.657, acc_len=1.54, thru=14.6, I=24.0, T=25.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:32,367] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:38,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:50:39,148] loop.py:380 INFO TIMING step=6595: step=4.405s data=1.950s compute=4.376s [fwd=2.991s bwd=1.371s opt=0.003s] dispatch=2.602s [2026-10-19 03:50:39,148] loop.py:45 INFO TRAIN_STEP step=6595 loss=1.637512 Training: 37%|███▋ | 6595/18018 [14:20:25<23:19:47, 7.35s/step, loss=1.638, acc=0.643, acc_len=1.49, thru=14.8, I=19.7, T=29.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:39,499] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:50:46,495] loop.py:45 INFO TRAIN_STEP step=6596 loss=1.498418 Training: 37%|███▋ | 6596/18018 [14:20:32<23:19:21, 7.35s/step, loss=1.498, acc=0.669, acc_len=1.58, thru=17.0, I=17.5, T=28.4, wait=2.8s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:50:46,490] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6595: forward=3052.3ms backward=1413.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:53,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:50:54,700] loop.py:45 INFO TRAIN_STEP step=6597 loss=1.504705 Training: 37%|███▋ | 6597/18018 [14:20:40<24:08:07, 7.61s/step, loss=1.505, acc=0.670, acc_len=1.58, thru=15.6, I=17.4, T=29.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:50:54,821] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:02,930] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:51:03,343] loop.py:45 INFO TRAIN_STEP step=6598 loss=1.436954 Training: 37%|███▋ | 6598/18018 [14:20:49<25:07:06, 7.92s/step, loss=1.437, acc=0.679, acc_len=1.61, thru=15.1, I=15.5, T=31.8, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:03,437] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:09,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:51:10,242] loop.py:45 INFO TRAIN_STEP step=6599 loss=1.527728 Training: 37%|███▋ | 6599/18018 [14:20:56<24:08:42, 7.61s/step, loss=1.528, acc=0.664, acc_len=1.56, thru=16.8, I=18.7, T=29.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:10,443] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:16,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:51:18,686] loop.py:380 INFO TIMING step=6600: step=4.533s data=2.156s compute=4.503s [fwd=3.165s bwd=1.323s opt=0.003s] dispatch=3.902s [2026-10-19 03:51:18,686] loop.py:45 INFO TRAIN_STEP step=6600 loss=1.550675 Training: 37%|███▋ | 6600/18018 [14:21:04<24:56:08, 7.86s/step, loss=1.551, acc=0.659, acc_len=1.55, thru=15.4, I=17.0, T=28.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:18,991] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:23,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:51:26,302] loop.py:45 INFO TRAIN_STEP step=6601 loss=1.402246 Training: 37%|███▋ | 6601/18018 [14:21:12<24:41:53, 7.79s/step, loss=1.402, acc=0.685, acc_len=1.63, thru=15.3, I=26.2, T=26.7, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:51:26,297] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6600: forward=3194.3ms backward=1544.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:26,697] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 03:51:35,262] loop.py:45 INFO TRAIN_STEP step=6602 loss=1.505468 Training: 37%|███▋ | 6602/18018 [14:21:21<25:48:38, 8.14s/step, loss=1.505, acc=0.667, acc_len=1.58, thru=14.9, I=15.5, T=34.4, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:35,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:35,716] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:41,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:51:42,864] loop.py:45 INFO TRAIN_STEP step=6603 loss=1.476763 Training: 37%|███▋ | 6603/18018 [14:21:28<25:17:49, 7.98s/step, loss=1.477, acc=0.676, acc_len=1.60, thru=17.4, I=19.2, T=32.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:43,030] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:48,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:51:49,393] loop.py:45 INFO TRAIN_STEP step=6604 loss=1.447458 Training: 37%|███▋ | 6604/18018 [14:21:35<23:55:07, 7.54s/step, loss=1.447, acc=0.679, acc_len=1.61, thru=18.7, I=22.5, T=34.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:49,560] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:54,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:51:56,435] loop.py:380 INFO TIMING step=6605: step=4.710s data=2.310s compute=4.679s [fwd=3.436s bwd=1.229s opt=0.003s] dispatch=2.323s [2026-10-19 03:51:56,436] loop.py:45 INFO TRAIN_STEP step=6605 loss=1.392444 Training: 37%|███▋ | 6605/18018 [14:21:42<23:26:22, 7.39s/step, loss=1.392, acc=0.688, acc_len=1.64, thru=18.0, I=22.8, T=27.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:51:56,490] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:00,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:52:03,149] loop.py:45 INFO TRAIN_STEP step=6606 loss=1.388606 Training: 37%|███▋ | 6606/18018 [14:21:49<22:47:24, 7.19s/step, loss=1.389, acc=0.686, acc_len=1.63, thru=16.5, I=26.3, T=27.0, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:52:03,144] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6605: forward=3232.0ms backward=1465.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:03,354] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:52:10,134] loop.py:45 INFO TRAIN_STEP step=6607 loss=1.542787 Training: 37%|███▋ | 6607/18018 [14:21:56<22:35:32, 7.13s/step, loss=1.543, acc=0.662, acc_len=1.55, thru=13.0, I=19.6, T=29.5, wait=2.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:17,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:52:19,808] loop.py:45 INFO TRAIN_STEP step=6608 loss=1.436793 Training: 37%|███▋ | 6608/18018 [14:22:05<25:00:43, 7.89s/step, loss=1.437, acc=0.677, acc_len=1.61, thru=15.8, I=15.2, T=29.2, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:20,259] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:27,755] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:52:29,122] loop.py:45 INFO TRAIN_STEP step=6609 loss=1.599526 Training: 37%|███▋ | 6609/18018 [14:22:15<26:21:47, 8.32s/step, loss=1.600, acc=0.650, acc_len=1.52, thru=14.4, I=16.2, T=29.3, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:29,272] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:36,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:52:36,240] loop.py:380 INFO TIMING step=6610: step=4.397s data=2.039s compute=4.366s [fwd=3.076s bwd=1.274s opt=0.003s] dispatch=2.712s [2026-10-19 03:52:36,240] loop.py:45 INFO TRAIN_STEP step=6610 loss=1.337785 Training: 37%|███▋ | 6610/18018 [14:22:22<25:13:11, 7.96s/step, loss=1.338, acc=0.694, acc_len=1.66, thru=16.8, I=20.6, T=29.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:36,526] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:43,209] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:52:45,284] loop.py:45 INFO TRAIN_STEP step=6611 loss=1.461754 Training: 37%|███▋ | 6611/18018 [14:22:31<26:14:53, 8.28s/step, loss=1.462, acc=0.672, acc_len=1.59, thru=14.1, I=17.0, T=26.2, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:52:45,280] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6610: forward=3347.4ms backward=1484.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:45,724] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:53,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:52:53,413] loop.py:45 INFO TRAIN_STEP step=6612 loss=1.485149 Training: 37%|███▋ | 6612/18018 [14:22:39<26:06:00, 8.24s/step, loss=1.485, acc=0.665, acc_len=1.57, thru=16.6, I=20.1, T=24.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:52:53,585] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:00,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:53:02,683] loop.py:45 INFO TRAIN_STEP step=6613 loss=1.384027 Training: 37%|███▋ | 6613/18018 [14:22:48<27:04:42, 8.55s/step, loss=1.384, acc=0.690, acc_len=1.65, thru=13.7, I=14.8, T=25.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:03,026] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:53:10,405] loop.py:45 INFO TRAIN_STEP step=6614 loss=1.486765 Training: 37%|███▋ | 6614/18018 [14:22:56<26:17:26, 8.30s/step, loss=1.487, acc=0.671, acc_len=1.59, thru=12.5, I=20.1, T=27.9, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:17,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:53:19,531] loop.py:380 INFO TIMING step=6615: step=4.431s data=2.085s compute=4.397s [fwd=3.165s bwd=1.217s opt=0.003s] dispatch=4.686s [2026-10-19 03:53:19,531] loop.py:45 INFO TRAIN_STEP step=6615 loss=1.563414 Training: 37%|███▋ | 6615/18018 [14:23:05<27:04:31, 8.55s/step, loss=1.563, acc=0.661, acc_len=1.55, thru=15.1, I=16.0, T=28.9, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:19,611] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:26,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:53:28,154] loop.py:45 INFO TRAIN_STEP step=6616 loss=1.455150 Training: 37%|███▋ | 6616/18018 [14:23:14<27:08:45, 8.57s/step, loss=1.455, acc=0.672, acc_len=1.60, thru=15.1, I=18.6, T=30.3, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:53:28,149] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6615: forward=2935.6ms backward=1233.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:28,386] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:36,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:53:37,156] loop.py:45 INFO TRAIN_STEP step=6617 loss=1.558012 Training: 37%|███▋ | 6617/18018 [14:23:23<27:32:57, 8.70s/step, loss=1.558, acc=0.660, acc_len=1.55, thru=14.6, I=14.8, T=22.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:37,448] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:43,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:53:44,765] loop.py:45 INFO TRAIN_STEP step=6618 loss=1.467800 Training: 37%|███▋ | 6618/18018 [14:23:30<26:30:49, 8.37s/step, loss=1.468, acc=0.675, acc_len=1.60, thru=17.1, I=17.5, T=25.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:44,843] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:52,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:53:53,899] loop.py:45 INFO TRAIN_STEP step=6619 loss=1.478485 Training: 37%|███▋ | 6619/18018 [14:23:39<27:14:05, 8.60s/step, loss=1.478, acc=0.674, acc_len=1.60, thru=14.2, I=16.1, T=28.4, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:53:54,219] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:00,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:54:00,877] loop.py:380 INFO TIMING step=6620: step=4.283s data=2.045s compute=4.249s [fwd=3.009s bwd=1.225s opt=0.003s] dispatch=2.684s [2026-10-19 03:54:00,878] loop.py:45 INFO TRAIN_STEP step=6620 loss=1.501372 Training: 37%|███▋ | 6620/18018 [14:23:46<25:41:25, 8.11s/step, loss=1.501, acc=0.665, acc_len=1.56, thru=16.4, I=19.2, T=29.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:01,350] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:08,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:54:09,892] loop.py:45 INFO TRAIN_STEP step=6621 loss=1.430579 Training: 37%|███▋ | 6621/18018 [14:23:55<26:32:32, 8.38s/step, loss=1.431, acc=0.683, acc_len=1.63, thru=14.1, I=18.1, T=25.1, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:54:09,888] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6620: forward=3490.1ms backward=1560.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:10,029] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:14,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:54:16,602] loop.py:45 INFO TRAIN_STEP step=6622 loss=1.441669 Training: 37%|███▋ | 6622/18018 [14:24:02<24:56:57, 7.88s/step, loss=1.442, acc=0.678, acc_len=1.61, thru=15.1, I=23.1, T=29.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:16,954] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 03:54:22,416] loop.py:45 INFO TRAIN_STEP step=6623 loss=1.367007 Training: 37%|███▋ | 6623/18018 [14:24:08<22:59:04, 7.26s/step, loss=1.367, acc=0.689, acc_len=1.65, thru=18.2, I=25.7, T=35.0, wait=2.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:28,457] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:54:28,898] loop.py:45 INFO TRAIN_STEP step=6624 loss=1.555085 Training: 37%|███▋ | 6624/18018 [14:24:14<22:14:39, 7.03s/step, loss=1.555, acc=0.659, acc_len=1.55, thru=19.2, I=20.2, T=34.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:28,976] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:34,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:54:35,715] loop.py:380 INFO TIMING step=6625: step=4.698s data=2.317s compute=4.659s [fwd=3.381s bwd=1.263s opt=0.003s] dispatch=2.108s [2026-10-19 03:54:35,715] loop.py:45 INFO TRAIN_STEP step=6625 loss=1.491029 Training: 37%|███▋ | 6625/18018 [14:24:21<22:02:28, 6.96s/step, loss=1.491, acc=0.671, acc_len=1.59, thru=18.2, I=21.4, T=27.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:35,863] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:40,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:54:42,520] loop.py:45 INFO TRAIN_STEP step=6626 loss=1.449803 Training: 37%|███▋ | 6626/18018 [14:24:28<21:53:12, 6.92s/step, loss=1.450, acc=0.676, acc_len=1.60, thru=15.8, I=22.9, T=26.8, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:54:42,516] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6625: forward=3239.0ms backward=1478.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:42,821] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:48,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:54:49,532] loop.py:45 INFO TRAIN_STEP step=6627 loss=1.525478 Training: 37%|███▋ | 6627/18018 [14:24:35<21:58:27, 6.94s/step, loss=1.525, acc=0.664, acc_len=1.57, thru=15.7, I=25.6, T=30.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:49,984] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:56,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:54:57,307] loop.py:45 INFO TRAIN_STEP step=6628 loss=1.477063 Training: 37%|███▋ | 6628/18018 [14:24:43<22:45:46, 7.19s/step, loss=1.477, acc=0.670, acc_len=1.59, thru=14.7, I=17.9, T=27.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:54:57,539] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:02,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:55:04,667] loop.py:45 INFO TRAIN_STEP step=6629 loss=1.345749 Training: 37%|███▋ | 6629/18018 [14:24:50<22:54:56, 7.24s/step, loss=1.346, acc=0.695, acc_len=1.67, thru=16.3, I=19.0, T=31.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:05,096] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:12,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:55:14,565] loop.py:380 INFO TIMING step=6630: step=4.342s data=2.116s compute=4.298s [fwd=2.949s bwd=1.334s opt=0.003s] dispatch=5.548s [2026-10-19 03:55:14,565] loop.py:45 INFO TRAIN_STEP step=6630 loss=1.503151 Training: 37%|███▋ | 6630/18018 [14:25:00<25:25:58, 8.04s/step, loss=1.503, acc=0.671, acc_len=1.58, thru=13.5, I=17.5, T=29.5, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:15,042] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:19,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:55:21,664] loop.py:45 INFO TRAIN_STEP step=6631 loss=1.459639 Training: 37%|███▋ | 6631/18018 [14:25:07<24:32:20, 7.76s/step, loss=1.460, acc=0.678, acc_len=1.61, thru=16.8, I=28.4, T=33.0, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:55:21,659] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6630: forward=2666.4ms backward=1165.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:21,892] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:55:30,682] loop.py:45 INFO TRAIN_STEP step=6632 loss=1.458694 Training: 37%|███▋ | 6632/18018 [14:25:16<25:44:00, 8.14s/step, loss=1.459, acc=0.675, acc_len=1.60, thru=11.4, I=11.0, T=28.2, wait=4.5s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:38,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:55:38,243] loop.py:45 INFO TRAIN_STEP step=6633 loss=1.440636 Training: 37%|███▋ | 6633/18018 [14:25:24<25:10:57, 7.96s/step, loss=1.441, acc=0.678, acc_len=1.61, thru=19.3, I=15.4, T=29.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:38,661] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:43,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:55:45,728] loop.py:45 INFO TRAIN_STEP step=6634 loss=1.382067 Training: 37%|███▋ | 6634/18018 [14:25:31<24:43:48, 7.82s/step, loss=1.382, acc=0.686, acc_len=1.64, thru=16.6, I=20.7, T=27.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:46,007] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:53,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:55:53,723] loop.py:380 INFO TIMING step=6635: step=4.659s data=2.281s compute=4.629s [fwd=3.358s bwd=1.256s opt=0.003s] dispatch=3.327s [2026-10-19 03:55:53,724] loop.py:45 INFO TRAIN_STEP step=6635 loss=1.491815 Training: 37%|███▋ | 6635/18018 [14:25:39<24:53:39, 7.87s/step, loss=1.492, acc=0.672, acc_len=1.59, thru=13.9, I=16.0, T=27.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:55:53,851] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:00,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:56:03,114] loop.py:45 INFO TRAIN_STEP step=6636 loss=1.639623 Training: 37%|███▋ | 6636/18018 [14:25:49<26:19:48, 8.33s/step, loss=1.640, acc=0.643, acc_len=1.50, thru=13.7, I=19.2, T=24.4, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:56:03,110] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6635: forward=3649.8ms backward=1536.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:03,415] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 03:56:09,551] loop.py:45 INFO TRAIN_STEP step=6637 loss=1.404007 Training: 37%|███▋ | 6637/18018 [14:25:55<24:31:57, 7.76s/step, loss=1.404, acc=0.682, acc_len=1.62, thru=12.9, I=19.3, T=33.8, wait=2.6s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:18,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:56:19,363] loop.py:45 INFO TRAIN_STEP step=6638 loss=1.398593 Training: 37%|███▋ | 6638/18018 [14:26:05<26:28:37, 8.38s/step, loss=1.399, acc=0.688, acc_len=1.64, thru=15.6, I=15.8, T=26.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:19,790] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:25,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:56:26,697] loop.py:45 INFO TRAIN_STEP step=6639 loss=1.343940 Training: 37%|███▋ | 6639/18018 [14:26:12<25:29:10, 8.06s/step, loss=1.344, acc=0.691, acc_len=1.67, thru=15.2, I=17.1, T=26.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:26,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:33,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:56:34,173] loop.py:380 INFO TIMING step=6640: step=4.364s data=1.885s compute=4.331s [fwd=3.077s bwd=1.239s opt=0.003s] dispatch=3.103s [2026-10-19 03:56:34,173] loop.py:45 INFO TRAIN_STEP step=6640 loss=1.237530 Training: 37%|███▋ | 6640/18018 [14:26:20<24:55:42, 7.89s/step, loss=1.238, acc=0.711, acc_len=1.72, thru=16.9, I=17.1, T=29.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:34,637] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:42,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:56:42,648] loop.py:45 INFO TRAIN_STEP step=6641 loss=1.542309 Training: 37%|███▋ | 6641/18018 [14:26:28<25:29:01, 8.06s/step, loss=1.542, acc=0.663, acc_len=1.56, thru=14.8, I=17.7, T=27.7, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:56:42,644] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6640: forward=3449.1ms backward=1128.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:42,788] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:56:49,625] loop.py:45 INFO TRAIN_STEP step=6642 loss=1.506224 Training: 37%|███▋ | 6642/18018 [14:26:35<24:27:04, 7.74s/step, loss=1.506, acc=0.665, acc_len=1.57, thru=17.2, I=16.4, T=31.1, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:56,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:56:58,037] loop.py:45 INFO TRAIN_STEP step=6643 loss=1.463524 Training: 37%|███▋ | 6643/18018 [14:26:44<25:05:14, 7.94s/step, loss=1.464, acc=0.674, acc_len=1.60, thru=16.7, I=18.1, T=31.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:56:58,376] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:57:03,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:57:04,684] loop.py:45 INFO TRAIN_STEP step=6644 loss=1.414537 Training: 37%|███▋ | 6644/18018 [14:26:50<23:51:35, 7.55s/step, loss=1.415, acc=0.683, acc_len=1.62, thru=17.4, I=24.1, T=32.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:57:04,959] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:57:09,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:57:11,346] loop.py:380 INFO TIMING step=6645: step=3.742s data=1.582s compute=3.707s [fwd=2.637s bwd=1.055s opt=0.003s] dispatch=2.911s [2026-10-19 03:57:11,347] loop.py:45 INFO TRAIN_STEP step=6645 loss=1.471335 Training: 37%|███▋ | 6645/18018 [14:26:57<23:00:58, 7.29s/step, loss=1.471, acc=0.672, acc_len=1.60, thru=18.1, I=23.6, T=34.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:57:11,404] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-19 03:57:21,823] loop.py:45 INFO TRAIN_STEP step=6646 loss=1.601447 Training: 37%|███▋ | 6646/18018 [14:27:07<26:02:11, 8.24s/step, loss=1.601, acc=0.654, acc_len=1.52, thru=7.0, I=9.7, T=25.9, wait=5.5s, pool=20, epoch=1/2] (TrainerActor pid=9310) [2026-10-19 03:57:21,818] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6645: forward=3351.6ms backward=1534.7ms (TrainerActor pid=9310) [2026-10-19 03:57:22,310] eagle_store.py:517 WARNING batch_get_buffer missing keys (k212664-8fa0190b_hs, k212664-8fa0190b_ids, k212664-8fa0190b_lhs); sleeping 0.50s. [2026-10-19 03:57:32,095] loop.py:45 INFO TRAIN_STEP step=6647 loss=1.534756 Training: 37%|███▋ | 6647/18018 [14:27:18<27:57:28, 8.85s/step, loss=1.535, acc=0.665, acc_len=1.56, thru=16.3, I=13.1, T=31.6, wait=6.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:57:41,098] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9371) [2026-10-19 03:57:22,514] eagle_store.py:517 WARNING batch_get_buffer missing keys (k212666-b1335e8b_hs, k212666-b1335e8b_ids, k212666-b1335e8b_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 03:57:42,738] loop.py:45 INFO TRAIN_STEP step=6648 loss=1.487585 Training: 37%|███▋ | 6648/18018 [14:27:28<29:39:09, 9.39s/step, loss=1.488, acc=0.669, acc_len=1.59, thru=13.2, I=14.5, T=25.2, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:57:43,108] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:57:50,268] loop.py:45 INFO TRAIN_STEP step=6649 loss=1.578776 Training: 37%|███▋ | 6649/18018 [14:27:36<27:53:25, 8.83s/step, loss=1.579, acc=0.657, acc_len=1.54, thru=13.7, I=19.4, T=30.5, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:57:56,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:57:58,281] loop.py:380 INFO TIMING step=6650: step=4.254s data=1.869s compute=4.219s [fwd=2.921s bwd=1.284s opt=0.003s] dispatch=3.750s [2026-10-19 03:57:58,281] loop.py:45 INFO TRAIN_STEP step=6650 loss=1.453616 Training: 37%|███▋ | 6650/18018 [14:27:44<27:06:46, 8.59s/step, loss=1.454, acc=0.681, acc_len=1.62, thru=17.1, I=17.4, T=30.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:57:58,316] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:04,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:58:05,437] loop.py:45 INFO TRAIN_STEP step=6651 loss=1.553961 Training: 37%|███▋ | 6651/18018 [14:27:51<25:45:19, 8.16s/step, loss=1.554, acc=0.658, acc_len=1.54, thru=15.2, I=21.5, T=26.7, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:58:05,433] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6650: forward=3260.4ms backward=1484.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:05,469] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:10,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:58:12,456] loop.py:45 INFO TRAIN_STEP step=6652 loss=1.515184 Training: 37%|███▋ | 6652/18018 [14:27:58<24:40:28, 7.82s/step, loss=1.515, acc=0.663, acc_len=1.56, thru=17.9, I=21.3, T=29.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:12,748] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:20,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:58:21,257] loop.py:45 INFO TRAIN_STEP step=6653 loss=1.541480 Training: 37%|███▋ | 6653/18018 [14:28:07<25:36:24, 8.11s/step, loss=1.541, acc=0.662, acc_len=1.56, thru=15.1, I=13.6, T=28.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:21,413] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 03:58:29,063] loop.py:45 INFO TRAIN_STEP step=6654 loss=1.605172 Training: 37%|███▋ | 6654/18018 [14:28:15<25:18:48, 8.02s/step, loss=1.605, acc=0.652, acc_len=1.52, thru=15.7, I=17.3, T=30.2, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:35,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:58:38,462] loop.py:380 INFO TIMING step=6655: step=5.402s data=2.657s compute=5.362s [fwd=3.728s bwd=1.618s opt=0.003s] dispatch=3.989s [2026-10-19 03:58:38,463] loop.py:45 INFO TRAIN_STEP step=6655 loss=1.448027 Training: 37%|███▋ | 6655/18018 [14:28:24<26:37:16, 8.43s/step, loss=1.448, acc=0.680, acc_len=1.61, thru=14.4, I=18.1, T=23.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:38,784] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 [2026-10-19 03:58:45,586] loop.py:45 INFO TRAIN_STEP step=6656 loss=1.394579 (TrainerActor pid=9310) [2026-10-19 03:58:45,581] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6655: forward=2941.8ms backward=1324.8ms Training: 37%|███▋ | 6656/18018 [14:28:31<25:22:28, 8.04s/step, loss=1.395, acc=0.686, acc_len=1.63, thru=17.6, I=18.6, T=29.7, wait=2.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:51,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:58:53,349] loop.py:45 INFO TRAIN_STEP step=6657 loss=1.593800 Training: 37%|███▋ | 6657/18018 [14:28:39<25:06:52, 7.96s/step, loss=1.594, acc=0.648, acc_len=1.51, thru=17.4, I=18.7, T=28.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:53,565] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 03:58:59,900] loop.py:45 INFO TRAIN_STEP step=6658 loss=1.458735 Training: 37%|███▋ | 6658/18018 [14:28:45<23:46:43, 7.54s/step, loss=1.459, acc=0.675, acc_len=1.60, thru=16.4, I=21.4, T=34.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:58:59,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:00,407] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:06,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:59:07,378] loop.py:45 INFO TRAIN_STEP step=6659 loss=1.459007 Training: 37%|███▋ | 6659/18018 [14:28:53<23:43:16, 7.52s/step, loss=1.459, acc=0.673, acc_len=1.59, thru=18.2, I=19.4, T=30.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:07,543] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:14,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:59:15,619] loop.py:380 INFO TIMING step=6660: step=4.471s data=2.048s compute=4.440s [fwd=3.163s bwd=1.262s opt=0.003s] dispatch=3.762s [2026-10-19 03:59:15,619] loop.py:45 INFO TRAIN_STEP step=6660 loss=1.420871 Training: 37%|███▋ | 6660/18018 [14:29:01<24:24:10, 7.73s/step, loss=1.421, acc=0.680, acc_len=1.61, thru=14.6, I=15.9, T=28.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:15,791] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:23,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:59:23,819] loop.py:45 INFO TRAIN_STEP step=6661 loss=1.495659 Training: 37%|███▋ | 6661/18018 [14:29:09<24:50:27, 7.87s/step, loss=1.496, acc=0.666, acc_len=1.57, thru=15.9, I=16.9, T=31.8, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 03:59:23,814] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6660: forward=2887.0ms backward=1087.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:24,324] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 03:59:32,423] loop.py:45 INFO TRAIN_STEP step=6662 loss=1.565470 Training: 37%|███▋ | 6662/18018 [14:29:18<25:31:50, 8.09s/step, loss=1.565, acc=0.660, acc_len=1.55, thru=13.3, I=14.0, T=30.2, wait=4.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:39,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:59:40,209] loop.py:45 INFO TRAIN_STEP step=6663 loss=1.470460 Training: 37%|███▋ | 6663/18018 [14:29:26<25:14:07, 8.00s/step, loss=1.470, acc=0.672, acc_len=1.59, thru=17.8, I=16.5, T=28.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:40,598] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 03:59:47,614] loop.py:45 INFO TRAIN_STEP step=6664 loss=1.478099 Training: 37%|███▋ | 6664/18018 [14:29:33<24:40:11, 7.82s/step, loss=1.478, acc=0.673, acc_len=1.60, thru=16.8, I=20.1, T=31.8, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:55,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 03:59:55,706] loop.py:380 INFO TIMING step=6665: step=4.379s data=1.999s compute=4.344s [fwd=2.977s bwd=1.352s opt=0.003s] dispatch=3.704s [2026-10-19 03:59:55,707] loop.py:45 INFO TRAIN_STEP step=6665 loss=1.322195 Training: 37%|███▋ | 6665/18018 [14:29:41<24:55:27, 7.90s/step, loss=1.322, acc=0.697, acc_len=1.68, thru=15.8, I=15.5, T=29.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 03:59:55,784] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-19 04:00:02,743] loop.py:45 INFO TRAIN_STEP step=6666 loss=1.442566 Training: 37%|███▋ | 6666/18018 [14:29:48<24:06:20, 7.64s/step, loss=1.443, acc=0.680, acc_len=1.62, thru=13.8, I=18.6, T=32.5, wait=3.1s, pool=28, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:00:02,739] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6665: forward=2787.7ms backward=1100.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:15,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:00:16,024] loop.py:45 INFO TRAIN_STEP step=6667 loss=1.523636 Training: 37%|███▋ | 6667/18018 [14:30:02<29:25:51, 9.33s/step, loss=1.524, acc=0.664, acc_len=1.56, thru=13.7, I=12.1, T=26.4, wait=8.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:16,321] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:21,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:00:23,144] loop.py:45 INFO TRAIN_STEP step=6668 loss=1.519847 Training: 37%|███▋ | 6668/18018 [14:30:09<27:20:02, 8.67s/step, loss=1.520, acc=0.665, acc_len=1.57, thru=16.0, I=19.9, T=29.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:23,643] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:00:31,248] loop.py:45 INFO TRAIN_STEP step=6669 loss=1.393157 Training: 37%|███▋ | 6669/18018 [14:30:17<26:47:50, 8.50s/step, loss=1.393, acc=0.685, acc_len=1.64, thru=14.2, I=16.1, T=30.8, wait=3.9s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:38,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:00:40,549] loop.py:380 INFO TIMING step=6670: step=5.145s data=2.514s compute=5.111s [fwd=3.678s bwd=1.418s opt=0.003s] dispatch=4.148s [2026-10-19 04:00:40,549] loop.py:45 INFO TRAIN_STEP step=6670 loss=1.450305 Training: 37%|███▋ | 6670/18018 [14:30:26<27:33:12, 8.74s/step, loss=1.450, acc=0.677, acc_len=1.60, thru=16.2, I=16.8, T=24.9, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:40,820] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:45,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:00:47,791] loop.py:45 INFO TRAIN_STEP step=6671 loss=1.578543 Training: 37%|███▋ | 6671/18018 [14:30:33<26:08:02, 8.29s/step, loss=1.579, acc=0.655, acc_len=1.54, thru=15.6, I=24.9, T=27.2, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:00:47,787] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6670: forward=3214.4ms backward=1428.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:48,018] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:55,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:00:56,373] loop.py:45 INFO TRAIN_STEP step=6672 loss=1.529369 Training: 37%|███▋ | 6672/18018 [14:30:42<26:24:30, 8.38s/step, loss=1.529, acc=0.665, acc_len=1.57, thru=15.3, I=16.6, T=30.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:00:56,530] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:01:04,209] loop.py:45 INFO TRAIN_STEP step=6673 loss=1.379487 Training: 37%|███▋ | 6673/18018 [14:30:50<25:53:30, 8.22s/step, loss=1.379, acc=0.689, acc_len=1.65, thru=13.8, I=15.7, T=31.3, wait=3.7s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:01:11,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:01:13,663] loop.py:45 INFO TRAIN_STEP step=6674 loss=1.511162 Training: 37%|███▋ | 6674/18018 [14:30:59<27:03:31, 8.59s/step, loss=1.511, acc=0.667, acc_len=1.57, thru=16.8, I=17.3, T=28.3, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:01:13,981] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:01:20,755] loop.py:380 INFO TIMING step=6675: step=4.505s data=2.098s compute=4.474s [fwd=3.086s bwd=1.373s opt=0.003s] dispatch=2.576s [2026-10-19 04:01:20,755] loop.py:45 INFO TRAIN_STEP step=6675 loss=1.520746 Training: 37%|███▋ | 6675/18018 [14:31:06<25:38:27, 8.14s/step, loss=1.521, acc=0.664, acc_len=1.56, thru=12.9, I=28.4, T=28.4, wait=2.6s, pool=32, epoch=1/2][2026-10-19 04:01:31,255] loop.py:45 INFO TRAIN_STEP step=6676 loss=1.484747 Training: 37%|███▋ | 6676/18018 [14:31:17<27:52:15, 8.85s/step, loss=1.485, acc=0.671, acc_len=1.59, thru=17.7, I=12.4, T=32.0, wait=6.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:01:31,251] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6675: forward=2806.9ms backward=1117.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:01:38,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:01:39,216] loop.py:45 INFO TRAIN_STEP step=6677 loss=1.363503 Training: 37%|███▋ | 6677/18018 [14:31:25<27:01:58, 8.58s/step, loss=1.364, acc=0.690, acc_len=1.65, thru=15.8, I=15.0, T=31.5, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:01:39,452] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:01:45,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:01:46,725] loop.py:45 INFO TRAIN_STEP step=6678 loss=1.482595 Training: 37%|███▋ | 6678/18018 [14:31:32<26:00:54, 8.26s/step, loss=1.483, acc=0.670, acc_len=1.58, thru=16.6, I=19.2, T=31.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:01:47,223] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:01:53,155] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:01:53,724] loop.py:45 INFO TRAIN_STEP step=6679 loss=1.384134 Training: 37%|███▋ | 6679/18018 [14:31:39<24:49:30, 7.88s/step, loss=1.384, acc=0.688, acc_len=1.65, thru=16.5, I=21.2, T=27.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:01:54,158] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:01:59,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:02:00,099] loop.py:380 INFO TIMING step=6680: step=3.753s data=1.630s compute=3.711s [fwd=2.671s bwd=1.025s opt=0.003s] dispatch=2.614s [2026-10-19 04:02:00,100] loop.py:45 INFO TRAIN_STEP step=6680 loss=1.546209 Training: 37%|███▋ | 6680/18018 [14:31:46<23:23:50, 7.43s/step, loss=1.546, acc=0.656, acc_len=1.55, thru=19.0, I=20.2, T=34.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:00,170] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:06,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:02:07,611] loop.py:45 INFO TRAIN_STEP step=6681 loss=1.551757 Training: 37%|███▋ | 6681/18018 [14:31:53<23:28:33, 7.45s/step, loss=1.552, acc=0.661, acc_len=1.55, thru=16.8, I=19.2, T=31.2, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:02:07,606] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6680: forward=2910.5ms backward=1144.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:07,871] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:15,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:02:15,970] loop.py:45 INFO TRAIN_STEP step=6682 loss=1.547625 Training: 37%|███▋ | 6682/18018 [14:32:02<24:19:34, 7.73s/step, loss=1.548, acc=0.664, acc_len=1.57, thru=15.6, I=16.4, T=33.2, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:16,215] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:21,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:02:22,691] loop.py:45 INFO TRAIN_STEP step=6683 loss=1.314827 Training: 37%|███▋ | 6683/18018 [14:32:08<23:22:28, 7.42s/step, loss=1.315, acc=0.698, acc_len=1.68, thru=19.2, I=21.4, T=34.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:23,125] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:29,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:02:29,676] loop.py:45 INFO TRAIN_STEP step=6684 loss=1.394626 Training: 37%|███▋ | 6684/18018 [14:32:15<22:57:31, 7.29s/step, loss=1.395, acc=0.684, acc_len=1.64, thru=17.3, I=20.6, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:30,094] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 04:02:37,893] loop.py:380 INFO TIMING step=6685: step=3.676s data=1.630s compute=3.629s [fwd=2.554s bwd=1.061s opt=0.003s] dispatch=4.533s [2026-10-19 04:02:37,893] loop.py:45 INFO TRAIN_STEP step=6685 loss=1.459290 Training: 37%|███▋ | 6685/18018 [14:32:23<23:49:50, 7.57s/step, loss=1.459, acc=0.678, acc_len=1.61, thru=15.9, I=16.0, T=34.8, wait=4.5s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:45,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:02:46,146] loop.py:45 INFO TRAIN_STEP step=6686 loss=1.567852 Training: 37%|███▋ | 6686/18018 [14:32:32<24:28:28, 7.78s/step, loss=1.568, acc=0.656, acc_len=1.54, thru=17.3, I=18.1, T=30.3, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:02:46,141] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6685: forward=2952.6ms backward=1212.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:46,452] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:53,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:02:54,885] loop.py:45 INFO TRAIN_STEP step=6687 loss=1.535904 Training: 37%|███▋ | 6687/18018 [14:32:40<25:22:52, 8.06s/step, loss=1.536, acc=0.665, acc_len=1.57, thru=14.0, I=17.4, T=25.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:02:55,034] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:01,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:03:03,187] loop.py:45 INFO TRAIN_STEP step=6688 loss=1.423732 Training: 37%|███▋ | 6688/18018 [14:32:49<25:36:20, 8.14s/step, loss=1.424, acc=0.683, acc_len=1.63, thru=13.8, I=15.9, T=26.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:03,396] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:10,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:03:12,113] loop.py:45 INFO TRAIN_STEP step=6689 loss=1.486039 Training: 37%|███▋ | 6689/18018 [14:32:58<26:20:51, 8.37s/step, loss=1.486, acc=0.672, acc_len=1.59, thru=14.7, I=17.2, T=28.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:12,162] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:19,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:03:20,626] loop.py:380 INFO TIMING step=6690: step=5.047s data=2.233s compute=5.018s [fwd=3.351s bwd=1.652s opt=0.003s] dispatch=3.456s [2026-10-19 04:03:20,626] loop.py:45 INFO TRAIN_STEP step=6690 loss=1.390136 Training: 37%|███▋ | 6690/18018 [14:33:06<26:28:36, 8.41s/step, loss=1.390, acc=0.689, acc_len=1.64, thru=13.2, I=18.9, T=25.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:20,853] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:27,583] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:03:28,380] loop.py:45 INFO TRAIN_STEP step=6691 loss=1.435401 Training: 37%|███▋ | 6691/18018 [14:33:14<25:51:15, 8.22s/step, loss=1.435, acc=0.682, acc_len=1.62, thru=16.5, I=19.2, T=30.0, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:03:28,375] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6690: forward=3023.4ms backward=1195.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:28,590] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:03:37,076] loop.py:45 INFO TRAIN_STEP step=6692 loss=1.425637 Training: 37%|███▋ | 6692/18018 [14:33:23<26:18:04, 8.36s/step, loss=1.426, acc=0.682, acc_len=1.62, thru=13.4, I=17.3, T=26.8, wait=3.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:43,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:03:45,267] loop.py:45 INFO TRAIN_STEP step=6693 loss=1.500615 Training: 37%|███▋ | 6693/18018 [14:33:31<26:08:24, 8.31s/step, loss=1.501, acc=0.667, acc_len=1.57, thru=14.9, I=16.8, T=26.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:45,708] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:53,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:03:53,926] loop.py:45 INFO TRAIN_STEP step=6694 loss=1.383869 Training: 37%|███▋ | 6694/18018 [14:33:40<26:28:12, 8.42s/step, loss=1.384, acc=0.685, acc_len=1.64, thru=15.6, I=18.1, T=28.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:54,092] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:03:58,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:04:00,634] loop.py:380 INFO TIMING step=6695: step=4.076s data=1.837s compute=4.036s [fwd=2.925s bwd=1.096s opt=0.003s] dispatch=2.621s [2026-10-19 04:04:00,635] loop.py:45 INFO TRAIN_STEP step=6695 loss=1.486101 Training: 37%|███▋ | 6695/18018 [14:33:46<24:51:19, 7.90s/step, loss=1.486, acc=0.670, acc_len=1.59, thru=17.6, I=22.2, T=31.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:04:00,885] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:04:08,669] loop.py:45 INFO TRAIN_STEP step=6696 loss=1.439247 Training: 37%|███▋ | 6696/18018 [14:33:54<24:58:52, 7.94s/step, loss=1.439, acc=0.680, acc_len=1.61, thru=14.4, I=14.8, T=32.5, wait=4.1s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:04:08,665] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6695: forward=2761.6ms backward=1122.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:04:15,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:04:17,358] loop.py:45 INFO TRAIN_STEP step=6697 loss=1.340366 Training: 37%|███▋ | 6697/18018 [14:34:03<25:40:44, 8.17s/step, loss=1.340, acc=0.700, acc_len=1.69, thru=16.2, I=20.8, T=27.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:04:17,549] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:04:27,032] loop.py:45 INFO TRAIN_STEP step=6698 loss=1.489776 Training: 37%|███▋ | 6698/18018 [14:34:13<27:05:58, 8.62s/step, loss=1.490, acc=0.669, acc_len=1.59, thru=12.7, I=12.1, T=25.3, wait=4.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:04:32,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:04:34,857] loop.py:45 INFO TRAIN_STEP step=6699 loss=1.592701 Training: 37%|███▋ | 6699/18018 [14:34:20<26:21:00, 8.38s/step, loss=1.593, acc=0.654, acc_len=1.53, thru=17.5, I=16.3, T=25.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:04:35,005] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:04:41,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:04:41,939] loop.py:380 INFO TIMING step=6700: step=4.367s data=2.047s compute=4.334s [fwd=3.040s bwd=1.278s opt=0.003s] dispatch=2.706s [2026-10-19 04:04:41,939] loop.py:45 INFO TRAIN_STEP step=6700 loss=1.474859 Training: 37%|███▋ | 6700/18018 [14:34:28<25:07:24, 7.99s/step, loss=1.475, acc=0.673, acc_len=1.60, thru=14.8, I=17.4, T=29.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:04:42,095] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:04:47,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:04:49,206] loop.py:45 INFO TRAIN_STEP step=6701 loss=1.428819 Training: 37%|███▋ | 6701/18018 [14:34:35<24:26:13, 7.77s/step, loss=1.429, acc=0.683, acc_len=1.63, thru=17.1, I=18.9, T=29.7, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:04:49,201] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6700: forward=3051.0ms backward=1212.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:04:49,418] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:04:55,087] loop.py:45 INFO TRAIN_STEP step=6702 loss=1.474915 Training: 37%|███▋ | 6702/18018 [14:34:41<22:39:05, 7.21s/step, loss=1.475, acc=0.671, acc_len=1.59, thru=18.3, I=23.6, T=33.5, wait=2.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:05,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:05:06,388] loop.py:45 INFO TRAIN_STEP step=6703 loss=1.535119 Training: 37%|███▋ | 6703/18018 [14:34:52<26:30:32, 8.43s/step, loss=1.535, acc=0.662, acc_len=1.57, thru=13.9, I=13.0, T=26.6, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:06,483] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:12,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:05:13,427] loop.py:45 INFO TRAIN_STEP step=6704 loss=1.401539 Training: 37%|███▋ | 6704/18018 [14:34:59<25:11:27, 8.02s/step, loss=1.402, acc=0.686, acc_len=1.64, thru=18.2, I=20.4, T=28.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:13,842] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:19,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:05:20,915] loop.py:380 INFO TIMING step=6705: step=4.631s data=2.224s compute=4.603s [fwd=3.224s bwd=1.364s opt=0.003s] dispatch=2.849s [2026-10-19 04:05:20,916] loop.py:45 INFO TRAIN_STEP step=6705 loss=1.522657 Training: 37%|███▋ | 6705/18018 [14:35:06<24:41:27, 7.86s/step, loss=1.523, acc=0.665, acc_len=1.57, thru=16.1, I=20.7, T=27.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:21,411] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:28,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:05:28,420] loop.py:45 INFO TRAIN_STEP step=6706 loss=1.502767 Training: 37%|███▋ | 6706/18018 [14:35:14<24:21:33, 7.75s/step, loss=1.503, acc=0.668, acc_len=1.58, thru=16.0, I=17.1, T=30.7, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:05:28,416] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6705: forward=2900.5ms backward=1202.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:28,688] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:34,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:05:35,634] loop.py:45 INFO TRAIN_STEP step=6707 loss=1.374747 Training: 37%|███▋ | 6707/18018 [14:35:21<23:50:51, 7.59s/step, loss=1.375, acc=0.692, acc_len=1.66, thru=16.8, I=19.2, T=30.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:35,742] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 04:05:43,199] loop.py:45 INFO TRAIN_STEP step=6708 loss=1.474046 Training: 37%|███▋ | 6708/18018 [14:35:29<23:49:26, 7.58s/step, loss=1.474, acc=0.674, acc_len=1.60, thru=14.3, I=17.9, T=30.8, wait=3.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:51,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:05:51,477] loop.py:45 INFO TRAIN_STEP step=6709 loss=1.421650 Training: 37%|███▋ | 6709/18018 [14:35:37<24:28:34, 7.79s/step, loss=1.422, acc=0.677, acc_len=1.61, thru=17.4, I=15.7, T=31.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:51,651] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:56,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:05:58,501] loop.py:380 INFO TIMING step=6710: step=4.005s data=1.754s compute=3.971s [fwd=2.811s bwd=1.146s opt=0.003s] dispatch=3.010s [2026-10-19 04:05:58,501] loop.py:45 INFO TRAIN_STEP step=6710 loss=1.562390 Training: 37%|███▋ | 6710/18018 [14:35:44<23:45:03, 7.56s/step, loss=1.562, acc=0.657, acc_len=1.55, thru=18.8, I=20.9, T=32.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:05:58,556] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 [2026-10-19 04:06:04,958] loop.py:45 INFO TRAIN_STEP step=6711 loss=1.467851 Training: 37%|███▋ | 6711/18018 [14:35:51<22:42:19, 7.23s/step, loss=1.468, acc=0.675, acc_len=1.60, thru=16.8, I=21.3, T=33.9, wait=2.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:06:04,954] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6710: forward=2653.0ms backward=1069.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:13,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:06:16,046] loop.py:45 INFO TRAIN_STEP step=6712 loss=1.476514 Training: 37%|███▋ | 6712/18018 [14:36:02<26:20:30, 8.39s/step, loss=1.477, acc=0.677, acc_len=1.60, thru=11.6, I=15.9, T=20.4, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:16,319] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:22,136] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:06:23,434] loop.py:45 INFO TRAIN_STEP step=6713 loss=1.351128 Training: 37%|███▋ | 6713/18018 [14:36:09<25:23:52, 8.09s/step, loss=1.351, acc=0.692, acc_len=1.66, thru=18.0, I=24.2, T=28.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:23,645] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:30,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:06:31,418] loop.py:45 INFO TRAIN_STEP step=6714 loss=1.570464 Training: 37%|███▋ | 6714/18018 [14:36:17<25:17:50, 8.06s/step, loss=1.570, acc=0.660, acc_len=1.55, thru=13.6, I=14.5, T=33.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:31,753] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:06:38,749] loop.py:380 INFO TIMING step=6715: step=4.202s data=1.957s compute=4.172s [fwd=2.966s bwd=1.192s opt=0.003s] dispatch=3.117s [2026-10-19 04:06:38,749] loop.py:45 INFO TRAIN_STEP step=6715 loss=1.494579 Training: 37%|███▋ | 6715/18018 [14:36:24<24:36:42, 7.84s/step, loss=1.495, acc=0.670, acc_len=1.59, thru=16.9, I=17.1, T=30.5, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:42,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:06:45,375] loop.py:45 INFO TRAIN_STEP step=6716 loss=1.338538 Training: 37%|███▋ | 6716/18018 [14:36:31<23:28:05, 7.48s/step, loss=1.339, acc=0.700, acc_len=1.68, thru=19.0, I=21.7, T=28.9, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:06:45,370] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6715: forward=3055.2ms backward=1321.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:45,478] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:50,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:06:52,171] loop.py:45 INFO TRAIN_STEP step=6717 loss=1.391173 Training: 37%|███▋ | 6717/18018 [14:36:38<22:49:35, 7.27s/step, loss=1.391, acc=0.685, acc_len=1.64, thru=15.3, I=23.5, T=29.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:52,422] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:57,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:06:58,319] loop.py:45 INFO TRAIN_STEP step=6718 loss=1.426148 Training: 37%|███▋ | 6718/18018 [14:36:44<21:45:52, 6.93s/step, loss=1.426, acc=0.680, acc_len=1.62, thru=18.5, I=22.5, T=33.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:06:58,568] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:05,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:07:05,706] loop.py:45 INFO TRAIN_STEP step=6719 loss=1.493531 Training: 37%|███▋ | 6719/18018 [14:36:51<22:11:20, 7.07s/step, loss=1.494, acc=0.672, acc_len=1.59, thru=15.6, I=17.3, T=30.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:06,096] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:11,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:07:13,088] loop.py:380 INFO TIMING step=6720: step=4.412s data=2.081s compute=4.386s [fwd=3.143s bwd=1.229s opt=0.003s] dispatch=2.961s [2026-10-19 04:07:13,088] loop.py:45 INFO TRAIN_STEP step=6720 loss=1.436718 Training: 37%|███▋ | 6720/18018 [14:36:59<22:28:52, 7.16s/step, loss=1.437, acc=0.680, acc_len=1.62, thru=17.6, I=19.9, T=29.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:13,285] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:19,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:07:20,035] loop.py:45 INFO TRAIN_STEP step=6721 loss=1.399368 Training: 37%|███▋ | 6721/18018 [14:37:06<22:16:32, 7.10s/step, loss=1.399, acc=0.685, acc_len=1.64, thru=15.7, I=18.9, T=33.9, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:07:20,030] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6720: forward=2643.6ms backward=1076.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:20,415] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 04:07:28,799] loop.py:45 INFO TRAIN_STEP step=6722 loss=1.472317 Training: 37%|███▋ | 6722/18018 [14:37:14<23:50:25, 7.60s/step, loss=1.472, acc=0.675, acc_len=1.60, thru=16.1, I=16.2, T=26.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:28,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:29,303] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:37,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:07:37,846] loop.py:45 INFO TRAIN_STEP step=6723 loss=1.417769 Training: 37%|███▋ | 6723/18018 [14:37:23<25:12:09, 8.03s/step, loss=1.418, acc=0.684, acc_len=1.63, thru=15.2, I=17.6, T=30.4, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:38,337] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:07:46,188] loop.py:45 INFO TRAIN_STEP step=6724 loss=1.417207 Training: 37%|███▋ | 6724/18018 [14:37:32<25:29:29, 8.13s/step, loss=1.417, acc=0.681, acc_len=1.62, thru=13.0, I=13.4, T=30.2, wait=4.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:54,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:07:57,682] loop.py:380 INFO TIMING step=6725: step=5.475s data=2.666s compute=5.443s [fwd=3.862s bwd=1.566s opt=0.003s] dispatch=6.010s [2026-10-19 04:07:57,682] loop.py:45 INFO TRAIN_STEP step=6725 loss=1.398163 Training: 37%|███▋ | 6725/18018 [14:37:43<28:39:30, 9.14s/step, loss=1.398, acc=0.688, acc_len=1.65, thru=12.8, I=15.3, T=23.4, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:07:57,899] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:03,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:08:03,531] loop.py:45 INFO TRAIN_STEP step=6726 loss=1.354654 Training: 37%|███▋ | 6726/18018 [14:37:49<25:33:56, 8.15s/step, loss=1.355, acc=0.687, acc_len=1.65, thru=16.1, I=22.7, T=34.6, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:08:03,527] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6725: forward=2625.7ms backward=1038.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:03,832] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:10,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:08:10,665] loop.py:45 INFO TRAIN_STEP step=6727 loss=1.519539 Training: 37%|███▋ | 6727/18018 [14:37:56<24:36:25, 7.85s/step, loss=1.520, acc=0.662, acc_len=1.56, thru=18.0, I=20.0, T=35.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:11,012] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:08:18,442] loop.py:45 INFO TRAIN_STEP step=6728 loss=1.570171 Training: 37%|███▋ | 6728/18018 [14:38:04<24:32:17, 7.82s/step, loss=1.570, acc=0.657, acc_len=1.54, thru=14.4, I=14.4, T=33.9, wait=4.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:29,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:08:30,179] loop.py:45 INFO TRAIN_STEP step=6729 loss=1.513991 Training: 37%|███▋ | 6729/18018 [14:38:16<28:13:05, 9.00s/step, loss=1.514, acc=0.665, acc_len=1.56, thru=12.5, I=12.7, T=29.8, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:30,684] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:35,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:08:37,651] loop.py:380 INFO TIMING step=6730: step=4.530s data=2.141s compute=4.470s [fwd=3.038s bwd=1.416s opt=0.003s] dispatch=2.931s [2026-10-19 04:08:37,651] loop.py:45 INFO TRAIN_STEP step=6730 loss=1.494738 Training: 37%|███▋ | 6730/18018 [14:38:23<26:46:48, 8.54s/step, loss=1.495, acc=0.671, acc_len=1.59, thru=16.5, I=23.4, T=28.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:37,856] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:42,840] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:08:44,505] loop.py:45 INFO TRAIN_STEP step=6731 loss=1.446241 (TrainerActor pid=9310) [2026-10-19 04:08:44,500] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6730: forward=2749.6ms backward=1215.5ms Training: 37%|███▋ | 6731/18018 [14:38:30<25:11:21, 8.03s/step, loss=1.446, acc=0.677, acc_len=1.60, thru=15.4, I=25.5, T=31.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:44,851] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:53,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:08:53,585] loop.py:45 INFO TRAIN_STEP step=6732 loss=1.544648 Training: 37%|███▋ | 6732/18018 [14:38:39<26:10:18, 8.35s/step, loss=1.545, acc=0.659, acc_len=1.55, thru=14.7, I=15.7, T=29.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:08:54,067] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:09:00,384] loop.py:45 INFO TRAIN_STEP step=6733 loss=1.637784 Training: 37%|███▋ | 6733/18018 [14:38:46<24:42:41, 7.88s/step, loss=1.638, acc=0.641, acc_len=1.49, thru=15.8, I=17.7, T=33.4, wait=3.0s, pool=44, epoch=1/2][2026-10-19 04:09:08,781] loop.py:45 INFO TRAIN_STEP step=6734 loss=1.510806 Training: 37%|███▋ | 6734/18018 [14:38:54<25:11:39, 8.04s/step, loss=1.511, acc=0.669, acc_len=1.58, thru=15.3, I=15.7, T=33.5, wait=4.6s, pool=48, epoch=1/2][2026-10-19 04:09:16,506] loop.py:380 INFO TIMING step=6735: step=4.323s data=1.648s compute=4.293s [fwd=3.125s bwd=1.153s opt=0.003s] dispatch=3.392s [2026-10-19 04:09:16,506] loop.py:45 INFO TRAIN_STEP step=6735 loss=1.529734 Training: 37%|███▋ | 6735/18018 [14:39:02<24:53:45, 7.94s/step, loss=1.530, acc=0.664, acc_len=1.56, thru=16.9, I=17.0, T=29.6, wait=3.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:09:23,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:09:24,564] loop.py:45 INFO TRAIN_STEP step=6736 loss=1.445612 Training: 37%|███▋ | 6736/18018 [14:39:10<25:00:13, 7.98s/step, loss=1.446, acc=0.677, acc_len=1.61, thru=16.4, I=17.0, T=24.9, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:09:24,559] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6735: forward=3710.1ms backward=1381.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:09:24,951] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:09:30,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:09:30,919] loop.py:45 INFO TRAIN_STEP step=6737 loss=1.546666 Training: 37%|███▋ | 6737/18018 [14:39:17<23:28:28, 7.49s/step, loss=1.547, acc=0.662, acc_len=1.57, thru=17.4, I=21.3, T=34.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:09:31,083] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:09:39,956] loop.py:45 INFO TRAIN_STEP step=6738 loss=1.485030 Training: 37%|███▋ | 6738/18018 [14:39:26<24:55:33, 7.96s/step, loss=1.485, acc=0.670, acc_len=1.59, thru=11.7, I=13.7, T=31.5, wait=5.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:09:47,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:09:48,466] loop.py:45 INFO TRAIN_STEP step=6739 loss=1.380636 Training: 37%|███▋ | 6739/18018 [14:39:34<25:26:47, 8.12s/step, loss=1.381, acc=0.688, acc_len=1.65, thru=16.6, I=16.0, T=29.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:09:48,676] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:09:54,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:09:56,593] loop.py:380 INFO TIMING step=6740: step=4.916s data=2.509s compute=4.871s [fwd=3.360s bwd=1.495s opt=0.003s] dispatch=3.199s [2026-10-19 04:09:56,593] loop.py:45 INFO TRAIN_STEP step=6740 loss=1.356170 Training: 37%|███▋ | 6740/18018 [14:39:42<25:26:44, 8.12s/step, loss=1.356, acc=0.692, acc_len=1.66, thru=14.6, I=20.4, T=26.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:09:56,910] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:03,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:10:03,553] loop.py:45 INFO TRAIN_STEP step=6741 loss=1.352474 Training: 37%|███▋ | 6741/18018 [14:39:49<24:21:12, 7.77s/step, loss=1.352, acc=0.694, acc_len=1.67, thru=15.6, I=21.5, T=35.6, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:10:03,548] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6740: forward=2481.3ms backward=1054.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:03,656] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:10,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:10:10,998] loop.py:45 INFO TRAIN_STEP step=6742 loss=1.499894 Training: 37%|███▋ | 6742/18018 [14:39:57<24:02:27, 7.68s/step, loss=1.500, acc=0.668, acc_len=1.59, thru=18.1, I=20.3, T=34.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:11,259] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:19,002] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:10:19,845] loop.py:45 INFO TRAIN_STEP step=6743 loss=1.492731 Training: 37%|███▋ | 6743/18018 [14:40:05<25:08:25, 8.03s/step, loss=1.493, acc=0.668, acc_len=1.57, thru=14.4, I=16.3, T=32.7, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:20,007] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:27,712] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:10:27,837] loop.py:45 INFO TRAIN_STEP step=6744 loss=1.432876 Training: 37%|███▋ | 6744/18018 [14:40:13<25:06:11, 8.02s/step, loss=1.433, acc=0.679, acc_len=1.61, thru=16.6, I=17.0, T=32.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:28,215] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:34,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:10:35,551] loop.py:380 INFO TIMING step=6745: step=4.181s data=1.793s compute=4.154s [fwd=2.870s bwd=1.269s opt=0.003s] dispatch=3.525s [2026-10-19 04:10:35,551] loop.py:45 INFO TRAIN_STEP step=6745 loss=1.520635 Training: 37%|███▋ | 6745/18018 [14:40:21<24:49:09, 7.93s/step, loss=1.521, acc=0.663, acc_len=1.57, thru=18.1, I=19.7, T=30.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:35,588] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:42,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:10:43,702] loop.py:45 INFO TRAIN_STEP step=6746 loss=1.346870 Training: 37%|███▋ | 6746/18018 [14:40:29<25:01:35, 7.99s/step, loss=1.347, acc=0.697, acc_len=1.68, thru=13.9, I=19.0, T=30.0, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:10:43,697] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6745: forward=2998.6ms backward=1222.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:10:43,755] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-19 04:10:52,305] loop.py:45 INFO TRAIN_STEP step=6747 loss=1.458106 Training: 37%|███▋ | 6747/18018 [14:40:38<25:36:04, 8.18s/step, loss=1.458, acc=0.680, acc_len=1.61, thru=13.6, I=15.9, T=25.7, wait=3.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:00,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:11:03,159] loop.py:45 INFO TRAIN_STEP step=6748 loss=1.641561 Training: 37%|███▋ | 6748/18018 [14:40:49<28:06:40, 8.98s/step, loss=1.642, acc=0.645, acc_len=1.51, thru=13.7, I=14.8, T=28.3, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:03,357] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:11:09,075] loop.py:45 INFO TRAIN_STEP step=6749 loss=1.334186 Training: 37%|███▋ | 6749/18018 [14:40:55<25:13:46, 8.06s/step, loss=1.334, acc=0.697, acc_len=1.68, thru=18.1, I=25.6, T=36.1, wait=2.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:17,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:11:17,963] loop.py:380 INFO TIMING step=6750: step=4.861s data=2.569s compute=4.828s [fwd=3.600s bwd=1.214s opt=0.003s] dispatch=4.019s [2026-10-19 04:11:17,963] loop.py:45 INFO TRAIN_STEP step=6750 loss=1.422430 Training: 37%|███▋ | 6750/18018 [14:41:04<26:00:15, 8.31s/step, loss=1.422, acc=0.682, acc_len=1.62, thru=15.4, I=14.7, T=26.3, wait=4.0s, pool=64, epoch=1/2][2026-10-19 04:11:17,965] loop.py:419 INFO Saving checkpoint at step 6750... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 04:11:21,355] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0006751 (TrainerActor pid=9310) [2026-10-19 04:11:21,362] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0006001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:22,841] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:22,842] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:28,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 04:11:21,354] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0006751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 04:11:22,633] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0006001 [2026-10-19 04:11:29,934] loop.py:45 INFO TRAIN_STEP step=6751 loss=1.480803 Training: 37%|███▋ | 6751/18018 [14:41:16<29:26:33, 9.41s/step, loss=1.481, acc=0.673, acc_len=1.60, thru=18.1, I=24.9, T=33.4, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:11:29,929] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6750: forward=2695.8ms backward=1091.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:30,311] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:36,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:11:37,699] loop.py:45 INFO TRAIN_STEP step=6752 loss=1.451676 Training: 37%|███▋ | 6752/18018 [14:41:23<27:53:58, 8.92s/step, loss=1.452, acc=0.681, acc_len=1.62, thru=14.6, I=19.7, T=30.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:37,881] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:45,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:11:46,175] loop.py:45 INFO TRAIN_STEP step=6753 loss=1.561198 Training: 37%|███▋ | 6753/18018 [14:41:32<27:28:58, 8.78s/step, loss=1.561, acc=0.655, acc_len=1.53, thru=14.9, I=17.2, T=32.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:46,462] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:51,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:11:54,003] loop.py:45 INFO TRAIN_STEP step=6754 loss=1.514262 Training: 37%|███▋ | 6754/18018 [14:41:40<26:35:07, 8.50s/step, loss=1.514, acc=0.665, acc_len=1.56, thru=15.7, I=21.6, T=25.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:54,231] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:11:59,983] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:12:00,714] loop.py:380 INFO TIMING step=6755: step=3.504s data=1.391s compute=3.470s [fwd=2.453s bwd=1.002s opt=0.003s] dispatch=3.197s [2026-10-19 04:12:00,715] loop.py:45 INFO TRAIN_STEP step=6755 loss=1.509641 Training: 37%|███▋ | 6755/18018 [14:41:46<24:54:27, 7.96s/step, loss=1.510, acc=0.669, acc_len=1.58, thru=14.9, I=21.6, T=36.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:00,987] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:12:07,015] loop.py:45 INFO TRAIN_STEP step=6756 loss=1.465470 Training: 37%|███▋ | 6756/18018 [14:41:53<23:20:45, 7.46s/step, loss=1.465, acc=0.670, acc_len=1.59, thru=19.2, I=24.0, T=32.4, wait=2.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:12:07,010] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6755: forward=2742.2ms backward=1154.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:14,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:12:15,300] loop.py:45 INFO TRAIN_STEP step=6757 loss=1.405014 Training: 38%|███▊ | 6757/18018 [14:42:01<24:06:52, 7.71s/step, loss=1.405, acc=0.682, acc_len=1.63, thru=15.6, I=15.3, T=31.9, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:15,427] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:21,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:12:22,505] loop.py:45 INFO TRAIN_STEP step=6758 loss=1.516447 Training: 38%|███▊ | 6758/18018 [14:42:08<23:38:22, 7.56s/step, loss=1.516, acc=0.667, acc_len=1.57, thru=17.8, I=19.5, T=27.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:22,533] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:29,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:12:31,359] loop.py:45 INFO TRAIN_STEP step=6759 loss=1.448699 Training: 38%|███▊ | 6759/18018 [14:42:17<24:51:13, 7.95s/step, loss=1.449, acc=0.678, acc_len=1.61, thru=14.5, I=15.9, T=26.8, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:31,614] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:38,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:12:39,189] loop.py:380 INFO TIMING step=6760: step=3.921s data=1.764s compute=3.892s [fwd=2.796s bwd=1.081s opt=0.003s] dispatch=3.899s [2026-10-19 04:12:39,189] loop.py:45 INFO TRAIN_STEP step=6760 loss=1.420256 Training: 38%|███▊ | 6760/18018 [14:42:25<24:44:35, 7.91s/step, loss=1.420, acc=0.680, acc_len=1.62, thru=13.9, I=18.6, T=32.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:39,438] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:46,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:12:47,852] loop.py:45 INFO TRAIN_STEP step=6761 loss=1.469702 Training: 38%|███▊ | 6761/18018 [14:42:33<25:26:43, 8.14s/step, loss=1.470, acc=0.675, acc_len=1.61, thru=14.5, I=17.2, T=35.3, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:12:47,847] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6760: forward=2542.7ms backward=1032.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:47,999] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:52,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:12:54,592] loop.py:45 INFO TRAIN_STEP step=6762 loss=1.545675 Training: 38%|███▊ | 6762/18018 [14:42:40<24:07:55, 7.72s/step, loss=1.546, acc=0.655, acc_len=1.53, thru=18.1, I=24.6, T=28.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:54,816] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:12:59,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:13:01,764] loop.py:45 INFO TRAIN_STEP step=6763 loss=1.336041 Training: 38%|███▊ | 6763/18018 [14:42:47<23:37:05, 7.55s/step, loss=1.336, acc=0.696, acc_len=1.67, thru=14.9, I=24.5, T=25.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:02,045] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:06,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:13:08,453] loop.py:45 INFO TRAIN_STEP step=6764 loss=1.347198 Training: 38%|███▊ | 6764/18018 [14:42:54<22:48:10, 7.29s/step, loss=1.347, acc=0.694, acc_len=1.67, thru=16.3, I=24.3, T=32.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:08,743] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:17,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:13:18,012] loop.py:380 INFO TIMING step=6765: step=4.960s data=2.584s compute=4.929s [fwd=3.695s bwd=1.218s opt=0.003s] dispatch=4.590s [2026-10-19 04:13:18,012] loop.py:45 INFO TRAIN_STEP step=6765 loss=1.609114 Training: 38%|███▊ | 6765/18018 [14:43:04<24:55:32, 7.97s/step, loss=1.609, acc=0.655, acc_len=1.53, thru=13.8, I=14.5, T=25.8, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:18,153] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:25,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:13:26,999] loop.py:45 INFO TRAIN_STEP step=6766 loss=1.545030 Training: 38%|███▊ | 6766/18018 [14:43:13<25:52:19, 8.28s/step, loss=1.545, acc=0.661, acc_len=1.55, thru=14.2, I=15.9, T=27.8, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:13:26,994] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6765: forward=3221.4ms backward=1336.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:27,231] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:13:32,417] loop.py:45 INFO TRAIN_STEP step=6767 loss=1.438057 Training: 38%|███▊ | 6767/18018 [14:43:18<23:11:13, 7.42s/step, loss=1.438, acc=0.675, acc_len=1.60, thru=19.3, I=26.4, T=36.0, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:32,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:32,913] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:37,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:13:38,551] loop.py:45 INFO TRAIN_STEP step=6768 loss=1.363901 Training: 38%|███▊ | 6768/18018 [14:43:24<21:58:52, 7.03s/step, loss=1.364, acc=0.692, acc_len=1.66, thru=20.4, I=22.8, T=34.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:38,671] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:45,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:13:46,779] loop.py:45 INFO TRAIN_STEP step=6769 loss=1.425794 Training: 38%|███▊ | 6769/18018 [14:43:32<23:05:57, 7.39s/step, loss=1.426, acc=0.677, acc_len=1.61, thru=15.7, I=18.4, T=30.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:13:46,908] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-19 04:13:54,592] loop.py:380 INFO TIMING step=6770: step=4.165s data=1.694s compute=4.123s [fwd=2.920s bwd=1.187s opt=0.003s] dispatch=3.639s [2026-10-19 04:13:54,592] loop.py:45 INFO TRAIN_STEP step=6770 loss=1.525059 Training: 38%|███▊ | 6770/18018 [14:43:40<23:29:40, 7.52s/step, loss=1.525, acc=0.663, acc_len=1.56, thru=13.3, I=17.1, T=30.7, wait=3.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:14:02,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:14:03,052] loop.py:45 INFO TRAIN_STEP step=6771 loss=1.460551 Training: 38%|███▊ | 6771/18018 [14:43:49<24:22:21, 7.80s/step, loss=1.461, acc=0.674, acc_len=1.60, thru=16.5, I=16.2, T=30.1, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:14:03,048] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6770: forward=2998.4ms backward=1199.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:14:03,334] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 04:14:11,125] loop.py:45 INFO TRAIN_STEP step=6772 loss=1.462981 Training: 38%|███▊ | 6772/18018 [14:43:57<24:37:16, 7.88s/step, loss=1.463, acc=0.674, acc_len=1.60, thru=15.5, I=14.2, T=29.1, wait=3.7s, pool=56, epoch=1/2][2026-10-19 04:14:20,197] loop.py:45 INFO TRAIN_STEP step=6773 loss=1.524103 Training: 38%|███▊ | 6773/18018 [14:44:06<25:44:08, 8.24s/step, loss=1.524, acc=0.666, acc_len=1.58, thru=15.6, I=14.7, T=27.8, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:14:20,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:14:20,650] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 04:14:28,275] loop.py:45 INFO TRAIN_STEP step=6774 loss=1.468164 Training: 38%|███▊ | 6774/18018 [14:44:14<25:34:56, 8.19s/step, loss=1.468, acc=0.675, acc_len=1.60, thru=14.5, I=13.7, T=30.0, wait=3.8s, pool=52, epoch=1/2][2026-10-19 04:14:36,846] loop.py:380 INFO TIMING step=6775: step=3.687s data=1.480s compute=3.656s [fwd=2.630s bwd=1.011s opt=0.003s] dispatch=4.876s [2026-10-19 04:14:36,847] loop.py:45 INFO TRAIN_STEP step=6775 loss=1.483325 Training: 38%|███▊ | 6775/18018 [14:44:22<25:56:26, 8.31s/step, loss=1.483, acc=0.672, acc_len=1.59, thru=14.4, I=13.5, T=34.7, wait=4.9s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:14:45,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:14:47,664] loop.py:45 INFO TRAIN_STEP step=6776 loss=1.448310 Training: 38%|███▊ | 6776/18018 [14:44:33<28:17:24, 9.06s/step, loss=1.448, acc=0.676, acc_len=1.61, thru=12.8, I=16.0, T=27.0, wait=6.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:14:47,660] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6775: forward=3295.4ms backward=1388.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:14:47,743] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:14:53,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:14:53,818] loop.py:45 INFO TRAIN_STEP step=6777 loss=1.391863 Training: 38%|███▊ | 6777/18018 [14:44:39<25:33:57, 8.19s/step, loss=1.392, acc=0.687, acc_len=1.65, thru=18.0, I=22.5, T=34.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:14:53,871] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:00,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:15:01,881] loop.py:45 INFO TRAIN_STEP step=6778 loss=1.508294 Training: 38%|███▊ | 6778/18018 [14:44:47<25:26:41, 8.15s/step, loss=1.508, acc=0.666, acc_len=1.56, thru=17.0, I=21.5, T=29.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:02,216] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:07,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:15:09,396] loop.py:45 INFO TRAIN_STEP step=6779 loss=1.485661 Training: 38%|███▊ | 6779/18018 [14:44:55<24:50:53, 7.96s/step, loss=1.486, acc=0.669, acc_len=1.58, thru=15.1, I=23.4, T=29.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:09,733] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:16,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:15:16,961] loop.py:380 INFO TIMING step=6780: step=4.120s data=1.916s compute=4.085s [fwd=2.901s bwd=1.169s opt=0.003s] dispatch=3.438s [2026-10-19 04:15:16,961] loop.py:45 INFO TRAIN_STEP step=6780 loss=1.463449 Training: 38%|███▊ | 6780/18018 [14:45:03<24:28:32, 7.84s/step, loss=1.463, acc=0.678, acc_len=1.61, thru=14.6, I=19.8, T=31.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:17,459] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:15:23,973] loop.py:45 INFO TRAIN_STEP step=6781 loss=1.404565 Training: 38%|███▊ | 6781/18018 [14:45:10<23:41:56, 7.59s/step, loss=1.405, acc=0.687, acc_len=1.64, thru=16.5, I=17.2, T=31.9, wait=3.0s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:15:23,969] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6780: forward=2841.1ms backward=1123.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:29,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:15:30,959] loop.py:45 INFO TRAIN_STEP step=6782 loss=1.452974 Training: 38%|███▊ | 6782/18018 [14:45:17<23:07:49, 7.41s/step, loss=1.453, acc=0.673, acc_len=1.60, thru=19.1, I=18.8, T=32.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:31,170] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:38,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:15:38,656] loop.py:45 INFO TRAIN_STEP step=6783 loss=1.314265 Training: 38%|███▊ | 6783/18018 [14:45:24<23:23:43, 7.50s/step, loss=1.314, acc=0.698, acc_len=1.68, thru=15.7, I=19.8, T=31.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:39,129] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:44,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:15:46,741] loop.py:45 INFO TRAIN_STEP step=6784 loss=1.467869 Training: 38%|███▊ | 6784/18018 [14:45:32<23:56:40, 7.67s/step, loss=1.468, acc=0.671, acc_len=1.59, thru=15.2, I=19.1, T=30.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:15:46,816] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 [2026-10-19 04:15:53,045] loop.py:380 INFO TIMING step=6785: step=3.761s data=1.706s compute=3.723s [fwd=2.627s bwd=1.081s opt=0.003s] dispatch=2.534s [2026-10-19 04:15:53,045] loop.py:45 INFO TRAIN_STEP step=6785 loss=1.385863 Training: 38%|███▊ | 6785/18018 [14:45:39<22:39:29, 7.26s/step, loss=1.386, acc=0.686, acc_len=1.64, thru=15.9, I=22.3, T=34.0, wait=2.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:01,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:16:03,168] loop.py:45 INFO TRAIN_STEP step=6786 loss=1.405443 Training: 38%|███▊ | 6786/18018 [14:45:49<25:20:09, 8.12s/step, loss=1.405, acc=0.684, acc_len=1.63, thru=14.7, I=15.3, T=27.9, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:16:03,164] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6785: forward=3112.5ms backward=1426.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:03,518] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:08,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:16:09,840] loop.py:45 INFO TRAIN_STEP step=6787 loss=1.407219 Training: 38%|███▊ | 6787/18018 [14:45:55<23:58:46, 7.69s/step, loss=1.407, acc=0.683, acc_len=1.63, thru=16.7, I=23.2, T=33.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:10,133] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:15,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:16:16,199] loop.py:45 INFO TRAIN_STEP step=6788 loss=1.430323 Training: 38%|███▊ | 6788/18018 [14:46:02<22:44:01, 7.29s/step, loss=1.430, acc=0.681, acc_len=1.62, thru=19.3, I=23.5, T=32.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:16,525] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:25,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:16:26,100] loop.py:45 INFO TRAIN_STEP step=6789 loss=1.500784 Training: 38%|███▊ | 6789/18018 [14:46:12<25:10:43, 8.07s/step, loss=1.501, acc=0.667, acc_len=1.58, thru=13.4, I=14.2, T=26.4, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:26,243] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:30,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:16:32,723] loop.py:380 INFO TIMING step=6790: step=4.154s data=1.886s compute=4.122s [fwd=2.908s bwd=1.198s opt=0.003s] dispatch=2.457s [2026-10-19 04:16:32,723] loop.py:45 INFO TRAIN_STEP step=6790 loss=1.368623 Training: 38%|███▊ | 6790/18018 [14:46:18<23:49:13, 7.64s/step, loss=1.369, acc=0.689, acc_len=1.65, thru=18.2, I=22.8, T=30.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:32,832] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:39,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:16:39,742] loop.py:45 INFO TRAIN_STEP step=6791 loss=1.311114 Training: 38%|███▊ | 6791/18018 [14:46:25<23:14:21, 7.45s/step, loss=1.311, acc=0.697, acc_len=1.68, thru=15.7, I=18.7, T=28.7, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:16:39,738] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6790: forward=3042.2ms backward=1368.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:40,125] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:44,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:16:46,452] loop.py:45 INFO TRAIN_STEP step=6792 loss=1.409250 Training: 38%|███▊ | 6792/18018 [14:46:32<22:32:35, 7.23s/step, loss=1.409, acc=0.685, acc_len=1.64, thru=18.5, I=20.7, T=32.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:46,541] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:52,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:16:53,326] loop.py:45 INFO TRAIN_STEP step=6793 loss=1.494775 Training: 38%|███▊ | 6793/18018 [14:46:39<22:12:24, 7.12s/step, loss=1.495, acc=0.672, acc_len=1.58, thru=17.1, I=21.7, T=31.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:16:53,794] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:00,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:17:01,414] loop.py:45 INFO TRAIN_STEP step=6794 loss=1.541046 Training: 38%|███▊ | 6794/18018 [14:46:47<23:06:28, 7.41s/step, loss=1.541, acc=0.655, acc_len=1.53, thru=16.6, I=18.1, T=31.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:01,906] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:09,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:17:10,400] loop.py:380 INFO TIMING step=6795: step=4.293s data=1.785s compute=4.240s [fwd=2.897s bwd=1.328s opt=0.003s] dispatch=4.685s [2026-10-19 04:17:10,400] loop.py:45 INFO TRAIN_STEP step=6795 loss=1.587430 Training: 38%|███▊ | 6795/18018 [14:46:56<24:34:50, 7.88s/step, loss=1.587, acc=0.654, acc_len=1.53, thru=15.1, I=17.4, T=29.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:10,571] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:16,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:17:17,616] loop.py:45 INFO TRAIN_STEP step=6796 loss=1.479937 Training: 38%|███▊ | 6796/18018 [14:47:03<23:57:05, 7.68s/step, loss=1.480, acc=0.671, acc_len=1.58, thru=16.4, I=22.4, T=31.9, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:17:17,612] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6795: forward=2862.6ms backward=1099.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:18,100] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:22,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:17:24,807] loop.py:45 INFO TRAIN_STEP step=6797 loss=1.496405 Training: 38%|███▊ | 6797/18018 [14:47:10<23:29:27, 7.54s/step, loss=1.496, acc=0.670, acc_len=1.58, thru=16.2, I=24.6, T=31.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:24,840] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:30,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:17:31,554] loop.py:45 INFO TRAIN_STEP step=6798 loss=1.466471 Training: 38%|███▊ | 6798/18018 [14:47:17<22:44:53, 7.30s/step, loss=1.466, acc=0.675, acc_len=1.60, thru=17.4, I=23.5, T=27.2, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:31,895] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:17:38,648] loop.py:45 INFO TRAIN_STEP step=6799 loss=1.589882 Training: 38%|███▊ | 6799/18018 [14:47:24<22:33:36, 7.24s/step, loss=1.590, acc=0.653, acc_len=1.53, thru=16.0, I=18.0, T=30.3, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:44,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:17:44,955] loop.py:380 INFO TIMING step=6800: step=4.036s data=1.847s compute=4.007s [fwd=2.849s bwd=1.143s opt=0.003s] dispatch=2.257s [2026-10-19 04:17:44,955] loop.py:45 INFO TRAIN_STEP step=6800 loss=1.413843 Training: 38%|███▊ | 6800/18018 [14:47:31<21:40:57, 6.96s/step, loss=1.414, acc=0.681, acc_len=1.63, thru=18.0, I=18.6, T=31.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:45,126] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:51,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:17:52,019] loop.py:45 INFO TRAIN_STEP step=6801 loss=1.543557 Training: 38%|███▊ | 6801/18018 [14:47:38<21:46:43, 6.99s/step, loss=1.544, acc=0.661, acc_len=1.55, thru=16.8, I=18.9, T=32.4, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:17:52,015] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6800: forward=2775.1ms backward=1136.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:52,300] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:57,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:17:59,174] loop.py:45 INFO TRAIN_STEP step=6802 loss=1.416855 Training: 38%|███▊ | 6802/18018 [14:47:45<21:55:52, 7.04s/step, loss=1.417, acc=0.683, acc_len=1.63, thru=17.5, I=22.8, T=31.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:17:59,633] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:18:04,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:18:07,623] loop.py:45 INFO TRAIN_STEP step=6803 loss=1.351022 Training: 38%|███▊ | 6803/18018 [14:47:53<23:14:48, 7.46s/step, loss=1.351, acc=0.696, acc_len=1.67, thru=16.0, I=20.7, T=26.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:18:07,896] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 04:18:13,611] loop.py:45 INFO TRAIN_STEP step=6804 loss=1.441171 Training: 38%|███▊ | 6804/18018 [14:47:59<21:51:58, 7.02s/step, loss=1.441, acc=0.679, acc_len=1.61, thru=16.0, I=22.2, T=33.7, wait=2.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:18:19,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:18:22,068] loop.py:380 INFO TIMING step=6805: step=4.974s data=2.237s compute=4.938s [fwd=3.369s bwd=1.553s opt=0.003s] dispatch=3.475s [2026-10-19 04:18:22,068] loop.py:45 INFO TRAIN_STEP step=6805 loss=1.460371 Training: 38%|███▊ | 6805/18018 [14:48:08<23:12:29, 7.45s/step, loss=1.460, acc=0.676, acc_len=1.60, thru=16.6, I=20.4, T=25.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:18:22,305] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:18:28,840] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:18:29,425] loop.py:45 INFO TRAIN_STEP step=6806 loss=1.559585 Training: 38%|███▊ | 6806/18018 [14:48:15<23:07:06, 7.42s/step, loss=1.560, acc=0.655, acc_len=1.53, thru=18.0, I=19.8, T=27.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:18:29,420] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6805: forward=3172.3ms backward=1395.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:18:29,844] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:18:37,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:18:39,397] loop.py:45 INFO TRAIN_STEP step=6807 loss=1.576198 Training: 38%|███▊ | 6807/18018 [14:48:25<25:29:56, 8.19s/step, loss=1.576, acc=0.653, acc_len=1.53, thru=13.4, I=16.1, T=25.9, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:18:39,649] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:18:47,293] loop.py:45 INFO TRAIN_STEP step=6808 loss=1.508412 Training: 38%|███▊ | 6808/18018 [14:48:33<25:13:19, 8.10s/step, loss=1.508, acc=0.663, acc_len=1.57, thru=16.2, I=15.3, T=30.4, wait=3.7s, pool=60, epoch=1/2][2026-10-19 04:18:54,081] loop.py:45 INFO TRAIN_STEP step=6809 loss=1.434211 Training: 38%|███▊ | 6809/18018 [14:48:40<23:59:41, 7.71s/step, loss=1.434, acc=0.677, acc_len=1.60, thru=15.2, I=16.9, T=33.0, wait=2.9s, pool=32, epoch=1/2][2026-10-19 04:19:03,464] loop.py:380 INFO TIMING step=6810: step=4.526s data=2.094s compute=4.446s [fwd=3.140s bwd=1.291s opt=0.003s] dispatch=4.848s [2026-10-19 04:19:03,464] loop.py:45 INFO TRAIN_STEP step=6810 loss=1.606076 Training: 38%|███▊ | 6810/18018 [14:48:49<25:33:30, 8.21s/step, loss=1.606, acc=0.650, acc_len=1.52, thru=16.2, I=15.5, T=28.3, wait=4.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:11,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:19:13,112] loop.py:45 INFO TRAIN_STEP step=6811 loss=1.521279 Training: 38%|███▊ | 6811/18018 [14:48:59<26:53:52, 8.64s/step, loss=1.521, acc=0.663, acc_len=1.56, thru=14.2, I=13.5, T=25.0, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:19:13,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6810: forward=3824.3ms backward=1252.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:13,511] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:19,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:19:21,434] loop.py:45 INFO TRAIN_STEP step=6812 loss=1.473593 Training: 38%|███▊ | 6812/18018 [14:49:07<26:36:05, 8.55s/step, loss=1.474, acc=0.672, acc_len=1.59, thru=13.7, I=19.9, T=29.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:21,522] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:26,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:19:28,020] loop.py:45 INFO TRAIN_STEP step=6813 loss=1.453321 Training: 38%|███▊ | 6813/18018 [14:49:14<24:46:10, 7.96s/step, loss=1.453, acc=0.680, acc_len=1.62, thru=16.8, I=25.0, T=30.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:28,387] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:33,916] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:19:34,919] loop.py:45 INFO TRAIN_STEP step=6814 loss=1.485372 Training: 38%|███▊ | 6814/18018 [14:49:21<23:46:43, 7.64s/step, loss=1.485, acc=0.670, acc_len=1.58, thru=16.9, I=22.5, T=33.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:35,426] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:19:42,587] loop.py:380 INFO TIMING step=6815: step=4.185s data=1.855s compute=4.141s [fwd=2.861s bwd=1.266s opt=0.003s] dispatch=3.473s [2026-10-19 04:19:42,587] loop.py:45 INFO TRAIN_STEP step=6815 loss=1.419747 Training: 38%|███▊ | 6815/18018 [14:49:28<23:48:14, 7.65s/step, loss=1.420, acc=0.680, acc_len=1.61, thru=14.2, I=22.7, T=30.6, wait=3.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:52,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:19:54,872] loop.py:45 INFO TRAIN_STEP step=6816 loss=1.316057 Training: 38%|███▊ | 6816/18018 [14:49:40<28:07:31, 9.04s/step, loss=1.316, acc=0.699, acc_len=1.69, thru=14.0, I=14.6, T=29.8, wait=8.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:19:54,867] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6815: forward=3027.3ms backward=1232.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:55,001] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:19:59,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:20:00,626] loop.py:45 INFO TRAIN_STEP step=6817 loss=1.383495 Training: 38%|███▊ | 6817/18018 [14:49:46<25:03:31, 8.05s/step, loss=1.383, acc=0.687, acc_len=1.64, thru=19.2, I=25.7, T=36.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:00,789] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:11,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:20:11,677] loop.py:45 INFO TRAIN_STEP step=6818 loss=1.562940 Training: 38%|███▊ | 6818/18018 [14:49:57<27:51:14, 8.95s/step, loss=1.563, acc=0.655, acc_len=1.53, thru=12.1, I=11.7, T=28.5, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:11,890] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:16,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:20:18,103] loop.py:45 INFO TRAIN_STEP step=6819 loss=1.446349 Training: 38%|███▊ | 6819/18018 [14:50:04<25:29:35, 8.19s/step, loss=1.446, acc=0.679, acc_len=1.62, thru=17.6, I=21.5, T=31.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:18,183] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:24,027] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:20:25,982] loop.py:380 INFO TIMING step=6820: step=4.630s data=2.244s compute=4.595s [fwd=3.182s bwd=1.398s opt=0.003s] dispatch=3.238s [2026-10-19 04:20:25,982] loop.py:45 INFO TRAIN_STEP step=6820 loss=1.530999 Training: 38%|███▊ | 6820/18018 [14:50:12<25:11:44, 8.10s/step, loss=1.531, acc=0.658, acc_len=1.55, thru=13.4, I=21.4, T=27.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:26,058] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:33,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 04:20:34,588] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6820: forward=2911.3ms backward=1400.0ms [2026-10-19 04:20:34,592] loop.py:45 INFO TRAIN_STEP step=6821 loss=1.477687 Training: 38%|███▊ | 6821/18018 [14:50:20<25:40:08, 8.25s/step, loss=1.478, acc=0.675, acc_len=1.60, thru=15.0, I=17.3, T=29.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:34,990] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:42,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:20:42,833] loop.py:45 INFO TRAIN_STEP step=6822 loss=1.482653 Training: 38%|███▊ | 6822/18018 [14:50:28<25:39:18, 8.25s/step, loss=1.483, acc=0.670, acc_len=1.59, thru=14.3, I=17.2, T=28.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:43,156] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 04:20:49,440] loop.py:45 INFO TRAIN_STEP step=6823 loss=1.465730 Training: 38%|███▊ | 6823/18018 [14:50:35<24:07:09, 7.76s/step, loss=1.466, acc=0.675, acc_len=1.60, thru=18.4, I=18.3, T=34.1, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:55,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:20:56,538] loop.py:45 INFO TRAIN_STEP step=6824 loss=1.535934 Training: 38%|███▊ | 6824/18018 [14:50:42<23:30:12, 7.56s/step, loss=1.536, acc=0.660, acc_len=1.55, thru=19.1, I=19.6, T=32.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:20:57,035] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:03,648] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:21:04,168] loop.py:380 INFO TIMING step=6825: step=4.011s data=1.843s compute=3.974s [fwd=2.854s bwd=1.105s opt=0.003s] dispatch=3.609s [2026-10-19 04:21:04,168] loop.py:45 INFO TRAIN_STEP step=6825 loss=1.370013 Training: 38%|███▊ | 6825/18018 [14:50:50<23:34:23, 7.58s/step, loss=1.370, acc=0.695, acc_len=1.67, thru=15.5, I=17.9, T=31.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:04,651] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:21:10,319] loop.py:45 INFO TRAIN_STEP step=6826 loss=1.414309 Training: 38%|███▊ | 6826/18018 [14:50:56<22:13:52, 7.15s/step, loss=1.414, acc=0.682, acc_len=1.63, thru=16.6, I=18.5, T=35.1, wait=2.5s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:21:10,315] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6825: forward=2599.6ms backward=1001.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:19,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:21:20,798] loop.py:45 INFO TRAIN_STEP step=6827 loss=1.397346 Training: 38%|███▊ | 6827/18018 [14:51:06<25:19:58, 8.15s/step, loss=1.397, acc=0.688, acc_len=1.65, thru=14.7, I=16.1, T=24.0, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:21,265] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:27,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:21:29,691] loop.py:45 INFO TRAIN_STEP step=6828 loss=1.427342 Training: 38%|███▊ | 6828/18018 [14:51:15<26:01:35, 8.37s/step, loss=1.427, acc=0.685, acc_len=1.63, thru=15.2, I=17.5, T=24.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:29,839] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:37,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:21:38,971] loop.py:45 INFO TRAIN_STEP step=6829 loss=1.536333 Training: 38%|███▊ | 6829/18018 [14:51:25<26:52:03, 8.64s/step, loss=1.536, acc=0.660, acc_len=1.55, thru=14.0, I=14.5, T=30.0, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:39,363] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:47,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:21:48,290] loop.py:380 INFO TIMING step=6830: step=4.560s data=2.191s compute=4.525s [fwd=3.174s bwd=1.337s opt=0.003s] dispatch=4.751s [2026-10-19 04:21:48,291] loop.py:45 INFO TRAIN_STEP step=6830 loss=1.325904 Training: 38%|███▊ | 6830/18018 [14:51:34<27:29:44, 8.85s/step, loss=1.326, acc=0.699, acc_len=1.68, thru=14.3, I=15.1, T=28.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:48,763] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:53,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:21:56,084] loop.py:45 INFO TRAIN_STEP step=6831 loss=1.462839 Training: 38%|███▊ | 6831/18018 [14:51:42<26:30:33, 8.53s/step, loss=1.463, acc=0.674, acc_len=1.60, thru=15.3, I=23.7, T=25.6, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:21:56,078] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6830: forward=3408.8ms backward=1541.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:21:56,311] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:02,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:22:03,491] loop.py:45 INFO TRAIN_STEP step=6832 loss=1.388688 Training: 38%|███▊ | 6832/18018 [14:51:49<25:27:35, 8.19s/step, loss=1.389, acc=0.686, acc_len=1.64, thru=17.8, I=19.4, T=27.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:03,744] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:22:12,973] loop.py:45 INFO TRAIN_STEP step=6833 loss=1.452244 Training: 38%|███▊ | 6833/18018 [14:51:59<26:39:33, 8.58s/step, loss=1.452, acc=0.673, acc_len=1.60, thru=9.1, I=12.2, T=31.0, wait=5.3s, pool=20, epoch=1/2] (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 04:22:13,189] eagle_store.py:517 WARNING batch_get_buffer missing keys (k218655-b2aac650_hs, k218655-b2aac650_ids, k218655-b2aac650_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:22,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 04:22:13,187] eagle_store.py:517 WARNING batch_get_buffer missing keys (k218654-9f2da570_hs, k218654-9f2da570_ids, k218654-9f2da570_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 04:22:23,753] loop.py:45 INFO TRAIN_STEP step=6834 loss=1.523602 Training: 38%|███▊ | 6834/18018 [14:52:09<28:42:21, 9.24s/step, loss=1.524, acc=0.661, acc_len=1.55, thru=15.7, I=17.1, T=28.7, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:23,904] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:28,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:22:30,079] loop.py:380 INFO TIMING step=6835: step=4.232s data=1.986s compute=4.192s [fwd=2.932s bwd=1.244s opt=0.003s] dispatch=2.086s [2026-10-19 04:22:30,079] loop.py:45 INFO TRAIN_STEP step=6835 loss=1.541956 Training: 38%|███▊ | 6835/18018 [14:52:16<25:59:13, 8.37s/step, loss=1.542, acc=0.657, acc_len=1.55, thru=18.4, I=24.3, T=30.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:30,564] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:37,744] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:22:38,862] loop.py:45 INFO TRAIN_STEP step=6836 loss=1.424042 Training: 38%|███▊ | 6836/18018 [14:52:24<26:22:27, 8.49s/step, loss=1.424, acc=0.679, acc_len=1.61, thru=15.4, I=17.5, T=27.9, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:22:38,858] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6835: forward=3206.6ms backward=1333.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:39,253] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:22:46,462] loop.py:45 INFO TRAIN_STEP step=6837 loss=1.447065 Training: 38%|███▊ | 6837/18018 [14:52:34<27:13:22, 8.77s/step, loss=1.447, acc=0.677, acc_len=1.61, thru=10.7, I=16.7, T=28.8, wait=3.1s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:58,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:22:59,728] loop.py:45 INFO TRAIN_STEP step=6838 loss=1.543766 Training: 38%|███▊ | 6838/18018 [14:52:45<29:43:55, 9.57s/step, loss=1.544, acc=0.665, acc_len=1.57, thru=12.2, I=14.2, T=28.7, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:22:59,930] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:23:07,305] loop.py:45 INFO TRAIN_STEP step=6839 loss=1.500844 Training: 38%|███▊ | 6839/18018 [14:52:53<27:52:06, 8.97s/step, loss=1.501, acc=0.666, acc_len=1.57, thru=15.6, I=17.3, T=33.6, wait=3.8s, pool=60, epoch=1/2][2026-10-19 04:23:15,485] loop.py:380 INFO TIMING step=6840: step=4.208s data=1.880s compute=4.180s [fwd=3.001s bwd=1.165s opt=0.003s] dispatch=3.966s [2026-10-19 04:23:15,486] loop.py:45 INFO TRAIN_STEP step=6840 loss=1.347243 Training: 38%|███▊ | 6840/18018 [14:53:01<27:07:56, 8.74s/step, loss=1.347, acc=0.694, acc_len=1.66, thru=16.5, I=16.8, T=30.4, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:23:21,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:23:23,416] loop.py:45 INFO TRAIN_STEP step=6841 loss=1.469450 Training: 38%|███▊ | 6841/18018 [14:53:09<26:22:23, 8.49s/step, loss=1.469, acc=0.673, acc_len=1.59, thru=15.9, I=17.6, T=31.1, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:23:23,412] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6840: forward=2955.2ms backward=1110.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:23:23,851] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:23:30,155] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:23:30,842] loop.py:45 INFO TRAIN_STEP step=6842 loss=1.424989 Training: 38%|███▊ | 6842/18018 [14:53:16<25:22:32, 8.17s/step, loss=1.425, acc=0.679, acc_len=1.61, thru=15.1, I=19.9, T=27.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:23:31,159] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:23:35,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:23:38,278] loop.py:45 INFO TRAIN_STEP step=6843 loss=1.438562 Training: 38%|███▊ | 6843/18018 [14:53:24<24:41:08, 7.95s/step, loss=1.439, acc=0.679, acc_len=1.61, thru=15.5, I=20.0, T=25.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:23:38,498] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:23:44,306] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:23:44,996] loop.py:45 INFO TRAIN_STEP step=6844 loss=1.422775 Training: 38%|███▊ | 6844/18018 [14:53:31<23:32:07, 7.58s/step, loss=1.423, acc=0.681, acc_len=1.62, thru=15.9, I=23.4, T=35.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:23:45,313] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:23:53,130] loop.py:380 INFO TIMING step=6845: step=4.392s data=1.931s compute=4.358s [fwd=2.994s bwd=1.349s opt=0.003s] dispatch=3.732s [2026-10-19 04:23:53,130] loop.py:45 INFO TRAIN_STEP step=6845 loss=1.539175 Training: 38%|███▊ | 6845/18018 [14:53:39<24:02:56, 7.75s/step, loss=1.539, acc=0.660, acc_len=1.56, thru=14.5, I=15.8, T=29.1, wait=3.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:23:59,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:24:00,441] loop.py:45 INFO TRAIN_STEP step=6846 loss=1.408981 Training: 38%|███▊ | 6846/18018 [14:53:46<23:38:05, 7.62s/step, loss=1.409, acc=0.685, acc_len=1.64, thru=18.4, I=18.2, T=31.7, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:24:00,437] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6845: forward=2811.7ms backward=1183.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:00,896] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:04,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:24:08,041] loop.py:45 INFO TRAIN_STEP step=6847 loss=1.539541 Training: 38%|███▊ | 6847/18018 [14:53:54<23:37:11, 7.61s/step, loss=1.540, acc=0.657, acc_len=1.54, thru=15.7, I=26.8, T=26.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:08,519] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:15,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:24:16,524] loop.py:45 INFO TRAIN_STEP step=6848 loss=1.369082 Training: 38%|███▊ | 6848/18018 [14:54:02<24:25:44, 7.87s/step, loss=1.369, acc=0.687, acc_len=1.64, thru=16.0, I=18.2, T=28.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:16,557] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:23,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:24:25,296] loop.py:45 INFO TRAIN_STEP step=6849 loss=1.405414 Training: 38%|███▊ | 6849/18018 [14:54:11<25:15:41, 8.14s/step, loss=1.405, acc=0.685, acc_len=1.64, thru=13.7, I=18.5, T=28.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:25,726] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:31,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:24:32,849] loop.py:380 INFO TIMING step=6850: step=5.217s data=2.674s compute=5.190s [fwd=3.679s bwd=1.496s opt=0.003s] dispatch=2.327s [2026-10-19 04:24:32,849] loop.py:45 INFO TRAIN_STEP step=6850 loss=1.497396 Training: 38%|███▊ | 6850/18018 [14:54:18<24:42:37, 7.97s/step, loss=1.497, acc=0.667, acc_len=1.58, thru=14.5, I=23.9, T=24.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:33,018] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:40,840] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:24:41,619] loop.py:45 INFO TRAIN_STEP step=6851 loss=1.355167 Training: 38%|███▊ | 6851/18018 [14:54:27<25:27:29, 8.21s/step, loss=1.355, acc=0.695, acc_len=1.66, thru=14.9, I=15.5, T=31.3, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:24:41,614] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6850: forward=2915.0ms backward=1135.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:41,848] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:46,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:24:48,345] loop.py:45 INFO TRAIN_STEP step=6852 loss=1.461394 Training: 38%|███▊ | 6852/18018 [14:54:34<24:04:45, 7.76s/step, loss=1.461, acc=0.679, acc_len=1.61, thru=18.1, I=21.4, T=30.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:48,508] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:54,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:24:56,201] loop.py:45 INFO TRAIN_STEP step=6853 loss=1.492614 Training: 38%|███▊ | 6853/18018 [14:54:42<24:09:42, 7.79s/step, loss=1.493, acc=0.667, acc_len=1.58, thru=15.2, I=20.8, T=26.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:24:56,647] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:01,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:25:03,570] loop.py:45 INFO TRAIN_STEP step=6854 loss=1.337413 Training: 38%|███▊ | 6854/18018 [14:54:49<23:46:02, 7.66s/step, loss=1.337, acc=0.696, acc_len=1.67, thru=16.7, I=23.5, T=28.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:03,968] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:25:11,236] loop.py:380 INFO TIMING step=6855: step=4.606s data=2.216s compute=4.579s [fwd=3.242s bwd=1.322s opt=0.003s] dispatch=3.052s [2026-10-19 04:25:11,237] loop.py:45 INFO TRAIN_STEP step=6855 loss=1.467151 Training: 38%|███▊ | 6855/18018 [14:54:57<23:45:55, 7.66s/step, loss=1.467, acc=0.676, acc_len=1.60, thru=13.8, I=20.2, T=27.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:11,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:11,732] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:18,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:25:21,010] loop.py:45 INFO TRAIN_STEP step=6856 loss=1.462175 Training: 38%|███▊ | 6856/18018 [14:55:07<25:43:37, 8.30s/step, loss=1.462, acc=0.674, acc_len=1.60, thru=13.5, I=16.8, T=21.9, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:25:21,005] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6855: forward=3939.7ms backward=1833.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:21,470] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:27,705] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:25:29,591] loop.py:45 INFO TRAIN_STEP step=6857 loss=1.552953 Training: 38%|███▊ | 6857/18018 [14:55:15<25:59:23, 8.38s/step, loss=1.553, acc=0.661, acc_len=1.55, thru=15.8, I=20.2, T=29.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:29,726] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:34,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:25:36,082] loop.py:45 INFO TRAIN_STEP step=6858 loss=1.396767 Training: 38%|███▊ | 6858/18018 [14:55:22<24:13:39, 7.82s/step, loss=1.397, acc=0.684, acc_len=1.63, thru=17.2, I=24.4, T=30.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:36,434] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:25:44,579] loop.py:45 INFO TRAIN_STEP step=6859 loss=1.366819 Training: 38%|███▊ | 6859/18018 [14:55:30<24:51:29, 8.02s/step, loss=1.367, acc=0.691, acc_len=1.65, thru=12.4, I=15.3, T=28.9, wait=4.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:50,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:25:52,062] loop.py:380 INFO TIMING step=6860: step=4.086s data=1.935s compute=4.054s [fwd=2.802s bwd=1.237s opt=0.003s] dispatch=3.387s [2026-10-19 04:25:52,062] loop.py:45 INFO TRAIN_STEP step=6860 loss=1.428826 Training: 38%|███▊ | 6860/18018 [14:55:38<24:21:22, 7.86s/step, loss=1.429, acc=0.679, acc_len=1.62, thru=17.1, I=17.9, T=31.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:52,556] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:58,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:25:58,989] loop.py:45 INFO TRAIN_STEP step=6861 loss=1.321787 Training: 38%|███▊ | 6861/18018 [14:55:45<23:29:12, 7.58s/step, loss=1.322, acc=0.700, acc_len=1.69, thru=16.0, I=21.3, T=34.6, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:25:58,984] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6860: forward=2582.9ms backward=1077.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:25:59,140] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:05,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:26:05,365] loop.py:45 INFO TRAIN_STEP step=6862 loss=1.586048 Training: 38%|███▊ | 6862/18018 [14:55:51<22:22:10, 7.22s/step, loss=1.586, acc=0.654, acc_len=1.53, thru=19.6, I=22.1, T=29.8, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:05,543] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:12,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:26:13,053] loop.py:45 INFO TRAIN_STEP step=6863 loss=1.560736 Training: 38%|███▊ | 6863/18018 [14:55:59<22:48:08, 7.36s/step, loss=1.561, acc=0.658, acc_len=1.54, thru=16.4, I=17.5, T=30.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:13,521] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:19,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:26:21,789] loop.py:45 INFO TRAIN_STEP step=6864 loss=1.512915 Training: 38%|███▊ | 6864/18018 [14:56:07<24:04:47, 7.77s/step, loss=1.513, acc=0.666, acc_len=1.57, thru=13.7, I=19.4, T=29.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:22,236] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:31,079] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:26:32,025] loop.py:380 INFO TIMING step=6865: step=4.805s data=2.379s compute=4.751s [fwd=3.343s bwd=1.392s opt=0.003s] dispatch=5.423s [2026-10-19 04:26:32,025] loop.py:45 INFO TRAIN_STEP step=6865 loss=1.467435 Training: 38%|███▊ | 6865/18018 [14:56:18<26:22:12, 8.51s/step, loss=1.467, acc=0.674, acc_len=1.59, thru=13.1, I=14.1, T=26.6, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:32,103] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:38,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:26:38,732] loop.py:45 INFO TRAIN_STEP step=6866 loss=1.463118 Training: 38%|███▊ | 6866/18018 [14:56:24<24:41:27, 7.97s/step, loss=1.463, acc=0.670, acc_len=1.58, thru=17.4, I=18.8, T=31.6, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:26:38,728] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6865: forward=2823.5ms backward=1172.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:39,119] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:45,518] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:26:47,028] loop.py:45 INFO TRAIN_STEP step=6867 loss=1.452568 Training: 38%|███▊ | 6867/18018 [14:56:33<24:59:24, 8.07s/step, loss=1.453, acc=0.673, acc_len=1.59, thru=15.1, I=17.8, T=29.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:26:47,533] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:26:56,915] loop.py:45 INFO TRAIN_STEP step=6868 loss=1.350554 Training: 38%|███▊ | 6868/18018 [14:56:42<26:40:34, 8.61s/step, loss=1.351, acc=0.693, acc_len=1.66, thru=11.1, I=10.9, T=28.8, wait=5.4s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:07,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:27:08,750] loop.py:45 INFO TRAIN_STEP step=6869 loss=1.652710 Training: 38%|███▊ | 6869/18018 [14:56:54<29:40:07, 9.58s/step, loss=1.653, acc=0.642, acc_len=1.49, thru=14.8, I=13.1, T=28.4, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:09,240] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:14,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:27:15,775] loop.py:380 INFO TIMING step=6870: step=4.057s data=1.859s compute=4.027s [fwd=2.824s bwd=1.189s opt=0.003s] dispatch=2.960s [2026-10-19 04:27:15,775] loop.py:45 INFO TRAIN_STEP step=6870 loss=1.586146 Training: 38%|███▊ | 6870/18018 [14:57:01<27:17:33, 8.81s/step, loss=1.586, acc=0.653, acc_len=1.52, thru=17.3, I=20.5, T=31.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:15,925] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:21,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:27:22,617] loop.py:45 INFO TRAIN_STEP step=6871 loss=1.462128 Training: 38%|███▊ | 6871/18018 [14:57:08<25:27:37, 8.22s/step, loss=1.462, acc=0.677, acc_len=1.61, thru=17.2, I=19.7, T=33.6, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:27:22,613] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6870: forward=2641.3ms backward=1105.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:22,791] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:28,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:27:30,581] loop.py:45 INFO TRAIN_STEP step=6872 loss=1.389154 Training: 38%|███▊ | 6872/18018 [14:57:16<25:12:56, 8.14s/step, loss=1.389, acc=0.688, acc_len=1.64, thru=15.7, I=21.9, T=29.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:30,820] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:35,306] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:27:37,405] loop.py:45 INFO TRAIN_STEP step=6873 loss=1.379629 Training: 38%|███▊ | 6873/18018 [14:57:23<23:59:14, 7.75s/step, loss=1.380, acc=0.687, acc_len=1.64, thru=16.4, I=29.0, T=29.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:37,820] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:27:44,970] loop.py:45 INFO TRAIN_STEP step=6874 loss=1.420302 Training: 38%|███▊ | 6874/18018 [14:57:31<23:48:51, 7.69s/step, loss=1.420, acc=0.682, acc_len=1.63, thru=13.2, I=18.3, T=29.4, wait=3.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:51,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:27:52,914] loop.py:380 INFO TIMING step=6875: step=5.273s data=2.459s compute=5.245s [fwd=3.472s bwd=1.758s opt=0.003s] dispatch=2.662s [2026-10-19 04:27:52,914] loop.py:45 INFO TRAIN_STEP step=6875 loss=1.385757 Training: 38%|███▊ | 6875/18018 [14:57:38<24:02:44, 7.77s/step, loss=1.386, acc=0.687, acc_len=1.63, thru=17.1, I=18.0, T=24.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:27:52,946] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:00,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:28:01,363] loop.py:45 INFO TRAIN_STEP step=6876 loss=1.381812 Training: 38%|███▊ | 6876/18018 [14:57:47<24:40:38, 7.97s/step, loss=1.382, acc=0.689, acc_len=1.65, thru=13.3, I=14.7, T=28.5, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:28:01,358] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6875: forward=3040.1ms backward=1384.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:01,597] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:07,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:28:09,255] loop.py:45 INFO TRAIN_STEP step=6877 loss=1.367489 Training: 38%|███▊ | 6877/18018 [14:57:55<24:35:58, 7.95s/step, loss=1.367, acc=0.690, acc_len=1.65, thru=16.1, I=20.0, T=26.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:09,482] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:16,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:28:18,635] loop.py:45 INFO TRAIN_STEP step=6878 loss=1.531825 Training: 38%|███▊ | 6878/18018 [14:58:04<25:55:35, 8.38s/step, loss=1.532, acc=0.665, acc_len=1.56, thru=14.0, I=16.1, T=25.7, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:18,710] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:24,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:28:26,733] loop.py:45 INFO TRAIN_STEP step=6879 loss=1.404939 Training: 38%|███▊ | 6879/18018 [14:58:12<25:39:48, 8.29s/step, loss=1.405, acc=0.684, acc_len=1.64, thru=16.0, I=21.6, T=23.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:27,129] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:32,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:28:33,369] loop.py:380 INFO TIMING step=6880: step=3.799s data=1.603s compute=3.758s [fwd=2.652s bwd=1.090s opt=0.003s] dispatch=2.827s [2026-10-19 04:28:33,369] loop.py:45 INFO TRAIN_STEP step=6880 loss=1.536151 Training: 38%|███▊ | 6880/18018 [14:58:19<24:07:22, 7.80s/step, loss=1.536, acc=0.664, acc_len=1.56, thru=14.6, I=21.2, T=33.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:33,496] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:41,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:28:42,231] loop.py:45 INFO TRAIN_STEP step=6881 loss=1.503300 Training: 38%|███▊ | 6881/18018 [14:58:28<25:06:28, 8.12s/step, loss=1.503, acc=0.670, acc_len=1.58, thru=13.5, I=17.0, T=28.4, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:28:42,226] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6880: forward=3112.3ms backward=1357.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:42,322] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:50,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:28:51,373] loop.py:45 INFO TRAIN_STEP step=6882 loss=1.423410 Training: 38%|███▊ | 6882/18018 [14:58:37<26:03:26, 8.42s/step, loss=1.423, acc=0.680, acc_len=1.61, thru=14.2, I=16.8, T=29.8, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:51,773] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:59,071] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:28:59,263] loop.py:45 INFO TRAIN_STEP step=6883 loss=1.454361 Training: 38%|███▊ | 6883/18018 [14:58:45<25:33:32, 8.26s/step, loss=1.454, acc=0.678, acc_len=1.61, thru=15.6, I=18.8, T=33.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:28:59,573] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:04,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:29:06,314] loop.py:45 INFO TRAIN_STEP step=6884 loss=1.490060 Training: 38%|███▊ | 6884/18018 [14:58:52<24:26:01, 7.90s/step, loss=1.490, acc=0.669, acc_len=1.58, thru=16.7, I=19.8, T=32.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:06,508] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:12,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:29:12,743] loop.py:380 INFO TIMING step=6885: step=3.645s data=1.635s compute=3.617s [fwd=2.571s bwd=1.030s opt=0.003s] dispatch=2.774s [2026-10-19 04:29:12,743] loop.py:45 INFO TRAIN_STEP step=6885 loss=1.423631 Training: 38%|███▊ | 6885/18018 [14:58:58<23:03:53, 7.46s/step, loss=1.424, acc=0.682, acc_len=1.62, thru=18.6, I=24.1, T=35.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:13,021] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:19,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:29:20,073] loop.py:45 INFO TRAIN_STEP step=6886 loss=1.520742 Training: 38%|███▊ | 6886/18018 [14:59:06<22:56:36, 7.42s/step, loss=1.521, acc=0.662, acc_len=1.55, thru=17.1, I=19.0, T=27.3, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:29:20,068] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6885: forward=3197.3ms backward=1441.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:20,273] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:25,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:29:27,814] loop.py:45 INFO TRAIN_STEP step=6887 loss=1.460917 Training: 38%|███▊ | 6887/18018 [14:59:13<23:14:22, 7.52s/step, loss=1.461, acc=0.675, acc_len=1.60, thru=16.1, I=19.4, T=26.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:27,993] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:33,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:29:34,124] loop.py:45 INFO TRAIN_STEP step=6888 loss=1.366624 Training: 38%|███▊ | 6888/18018 [14:59:20<22:07:05, 7.15s/step, loss=1.367, acc=0.690, acc_len=1.65, thru=16.9, I=20.5, T=30.8, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:34,377] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:41,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:29:42,293] loop.py:45 INFO TRAIN_STEP step=6889 loss=1.537778 Training: 38%|███▊ | 6889/18018 [14:59:28<23:03:29, 7.46s/step, loss=1.538, acc=0.660, acc_len=1.55, thru=15.7, I=16.0, T=31.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:42,342] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:47,324] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:29:49,387] loop.py:380 INFO TIMING step=6890: step=4.537s data=2.186s compute=4.507s [fwd=3.168s bwd=1.324s opt=0.003s] dispatch=2.548s [2026-10-19 04:29:49,387] loop.py:45 INFO TRAIN_STEP step=6890 loss=1.440907 Training: 38%|███▊ | 6890/18018 [14:59:35<22:43:01, 7.35s/step, loss=1.441, acc=0.677, acc_len=1.61, thru=16.7, I=21.3, T=28.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:29:49,837] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:29:56,765] loop.py:45 INFO TRAIN_STEP step=6891 loss=1.408562 Training: 38%|███▊ | 6891/18018 [14:59:42<22:44:38, 7.36s/step, loss=1.409, acc=0.684, acc_len=1.63, thru=10.5, I=18.7, T=32.2, wait=3.4s, pool=20, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:29:56,761] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6890: forward=2641.8ms backward=1285.8ms (TrainerActor pid=9370) [2026-10-19 04:29:56,827] eagle_store.py:517 WARNING batch_get_buffer missing keys (k220519-80847da7_hs, k220519-80847da7_ids, k220519-80847da7_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:05,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9368) [2026-10-19 04:29:56,865] eagle_store.py:517 WARNING batch_get_buffer missing keys (k220517-ad6fb848_hs, k220517-ad6fb848_ids, k220517-ad6fb848_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 04:30:05,694] loop.py:45 INFO TRAIN_STEP step=6892 loss=1.526868 Training: 38%|███▊ | 6892/18018 [14:59:51<24:11:45, 7.83s/step, loss=1.527, acc=0.666, acc_len=1.57, thru=18.6, I=15.8, T=31.6, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:06,048] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:11,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:30:12,783] loop.py:45 INFO TRAIN_STEP step=6893 loss=1.460266 Training: 38%|███▊ | 6893/18018 [14:59:58<23:30:26, 7.61s/step, loss=1.460, acc=0.674, acc_len=1.60, thru=17.3, I=20.8, T=30.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:12,972] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:18,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:30:19,385] loop.py:45 INFO TRAIN_STEP step=6894 loss=1.535688 Training: 38%|███▊ | 6894/18018 [15:00:05<22:34:33, 7.31s/step, loss=1.536, acc=0.660, acc_len=1.55, thru=18.6, I=22.5, T=29.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:19,544] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:26,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:30:27,599] loop.py:380 INFO TIMING step=6895: step=4.278s data=1.859s compute=4.252s [fwd=2.949s bwd=1.288s opt=0.003s] dispatch=3.925s [2026-10-19 04:30:27,599] loop.py:45 INFO TRAIN_STEP step=6895 loss=1.493717 Training: 38%|███▊ | 6895/18018 [15:00:13<23:24:54, 7.58s/step, loss=1.494, acc=0.668, acc_len=1.58, thru=14.3, I=18.7, T=29.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:28,080] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:33,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:30:34,795] loop.py:45 INFO TRAIN_STEP step=6896 loss=1.292377 Training: 38%|███▊ | 6896/18018 [15:00:20<23:03:22, 7.46s/step, loss=1.292, acc=0.705, acc_len=1.71, thru=17.1, I=23.2, T=27.6, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:30:34,791] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6895: forward=3015.7ms backward=1555.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:35,288] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:40,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:30:42,504] loop.py:45 INFO TRAIN_STEP step=6897 loss=1.404551 Training: 38%|███▊ | 6897/18018 [15:00:28<23:17:00, 7.54s/step, loss=1.405, acc=0.685, acc_len=1.64, thru=14.9, I=18.6, T=32.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:42,880] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:50,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:30:51,802] loop.py:45 INFO TRAIN_STEP step=6898 loss=1.453040 Training: 38%|███▊ | 6898/18018 [15:00:37<24:54:46, 8.07s/step, loss=1.453, acc=0.677, acc_len=1.61, thru=14.3, I=16.4, T=30.9, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:30:52,204] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:31:01,333] loop.py:45 INFO TRAIN_STEP step=6899 loss=1.441560 Training: 38%|███▊ | 6899/18018 [15:00:47<26:16:18, 8.51s/step, loss=1.442, acc=0.674, acc_len=1.61, thru=13.6, I=13.9, T=28.5, wait=5.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:06,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:31:09,422] loop.py:380 INFO TIMING step=6900: step=4.457s data=2.123s compute=4.428s [fwd=3.101s bwd=1.312s opt=0.003s] dispatch=3.621s [2026-10-19 04:31:09,422] loop.py:45 INFO TRAIN_STEP step=6900 loss=1.491445 Training: 38%|███▊ | 6900/18018 [15:00:55<25:52:48, 8.38s/step, loss=1.491, acc=0.671, acc_len=1.59, thru=16.6, I=18.9, T=28.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:09,496] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:17,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:31:18,988] loop.py:45 INFO TRAIN_STEP step=6901 loss=1.322391 Training: 38%|███▊ | 6901/18018 [15:01:05<26:58:39, 8.74s/step, loss=1.322, acc=0.698, acc_len=1.68, thru=13.5, I=15.5, T=26.0, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:31:18,983] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6900: forward=3480.0ms backward=1399.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:19,196] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:24,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:31:25,242] loop.py:45 INFO TRAIN_STEP step=6902 loss=1.396598 Training: 38%|███▊ | 6902/18018 [15:01:11<24:40:34, 7.99s/step, loss=1.397, acc=0.686, acc_len=1.64, thru=18.2, I=20.4, T=31.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:25,277] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:31,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:31:31,424] loop.py:45 INFO TRAIN_STEP step=6903 loss=1.443933 Training: 38%|███▊ | 6903/18018 [15:01:17<22:59:47, 7.45s/step, loss=1.444, acc=0.677, acc_len=1.61, thru=18.6, I=22.1, T=30.2, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:31,527] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:38,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:31:39,046] loop.py:45 INFO TRAIN_STEP step=6904 loss=1.395260 Training: 38%|███▊ | 6904/18018 [15:01:25<23:09:18, 7.50s/step, loss=1.395, acc=0.687, acc_len=1.64, thru=16.0, I=17.1, T=32.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:39,161] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 04:31:45,886] loop.py:380 INFO TIMING step=6905: step=3.573s data=1.452s compute=3.537s [fwd=2.494s bwd=1.028s opt=0.003s] dispatch=3.258s [2026-10-19 04:31:45,886] loop.py:45 INFO TRAIN_STEP step=6905 loss=1.489528 Training: 38%|███▊ | 6905/18018 [15:01:31<22:32:32, 7.30s/step, loss=1.490, acc=0.672, acc_len=1.59, thru=16.8, I=21.9, T=35.8, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:51,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:31:53,343] loop.py:45 INFO TRAIN_STEP step=6906 loss=1.408809 Training: 38%|███▊ | 6906/18018 [15:01:39<22:40:55, 7.35s/step, loss=1.409, acc=0.684, acc_len=1.63, thru=16.5, I=19.8, T=30.5, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:31:53,338] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6905: forward=2908.8ms backward=1253.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:31:53,810] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:31:59,876] loop.py:45 INFO TRAIN_STEP step=6907 loss=1.501197 Training: 38%|███▊ | 6907/18018 [15:01:45<21:55:31, 7.10s/step, loss=1.501, acc=0.673, acc_len=1.59, thru=15.7, I=21.0, T=30.0, wait=2.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:07,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:32:08,774] loop.py:45 INFO TRAIN_STEP step=6908 loss=1.465429 Training: 38%|███▊ | 6908/18018 [15:01:54<23:35:02, 7.64s/step, loss=1.465, acc=0.673, acc_len=1.60, thru=15.8, I=16.8, T=30.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:09,123] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:14,736] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:32:16,004] loop.py:45 INFO TRAIN_STEP step=6909 loss=1.487776 Training: 38%|███▊ | 6909/18018 [15:02:02<23:12:02, 7.52s/step, loss=1.488, acc=0.668, acc_len=1.58, thru=15.7, I=22.1, T=32.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:16,244] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:22,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:32:23,277] loop.py:380 INFO TIMING step=6910: step=4.799s data=2.304s compute=4.758s [fwd=3.251s bwd=1.492s opt=0.003s] dispatch=2.465s [2026-10-19 04:32:23,278] loop.py:45 INFO TRAIN_STEP step=6910 loss=1.445891 Training: 38%|███▊ | 6910/18018 [15:02:09<22:58:25, 7.45s/step, loss=1.446, acc=0.673, acc_len=1.60, thru=16.5, I=20.5, T=26.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:23,501] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:29,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:32:30,673] loop.py:45 INFO TRAIN_STEP step=6911 loss=1.449562 Training: 38%|███▊ | 6911/18018 [15:02:16<22:55:26, 7.43s/step, loss=1.450, acc=0.674, acc_len=1.59, thru=16.1, I=21.0, T=30.9, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:32:30,669] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6910: forward=2880.8ms backward=1215.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:31,159] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:36,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:32:37,197] loop.py:45 INFO TRAIN_STEP step=6912 loss=1.301305 Training: 38%|███▊ | 6912/18018 [15:02:23<22:05:03, 7.16s/step, loss=1.301, acc=0.699, acc_len=1.68, thru=17.6, I=25.8, T=29.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:37,269] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-19 04:32:43,980] loop.py:45 INFO TRAIN_STEP step=6913 loss=1.514578 Training: 38%|███▊ | 6913/18018 [15:02:30<21:43:57, 7.05s/step, loss=1.515, acc=0.666, acc_len=1.57, thru=15.9, I=17.5, T=33.3, wait=2.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:49,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:32:51,081] loop.py:45 INFO TRAIN_STEP step=6914 loss=1.564277 Training: 38%|███▊ | 6914/18018 [15:02:37<21:47:05, 7.06s/step, loss=1.564, acc=0.656, acc_len=1.54, thru=18.4, I=20.6, T=29.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:51,231] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:32:59,471] loop.py:380 INFO TIMING step=6915: step=4.499s data=2.171s compute=4.453s [fwd=3.143s bwd=1.295s opt=0.003s] dispatch=3.880s [2026-10-19 04:32:59,471] loop.py:45 INFO TRAIN_STEP step=6915 loss=1.535830 Training: 38%|███▊ | 6915/18018 [15:02:45<23:00:31, 7.46s/step, loss=1.536, acc=0.662, acc_len=1.56, thru=13.9, I=15.8, T=28.5, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:59,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:32:59,944] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:07,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:33:08,121] loop.py:45 INFO TRAIN_STEP step=6916 loss=1.590087 Training: 38%|███▊ | 6916/18018 [15:02:54<24:06:27, 7.82s/step, loss=1.590, acc=0.657, acc_len=1.54, thru=15.1, I=15.8, T=27.9, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:33:08,117] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6915: forward=3198.4ms backward=1333.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:08,610] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:33:18,118] loop.py:45 INFO TRAIN_STEP step=6917 loss=1.558175 Training: 38%|███▊ | 6917/18018 [15:03:04<26:07:17, 8.47s/step, loss=1.558, acc=0.658, acc_len=1.55, thru=13.5, I=13.8, T=25.4, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:18,066] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:18,567] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:23,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:33:25,189] loop.py:45 INFO TRAIN_STEP step=6918 loss=1.328653 Training: 38%|███▊ | 6918/18018 [15:03:11<24:49:26, 8.05s/step, loss=1.329, acc=0.696, acc_len=1.67, thru=15.4, I=18.6, T=26.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:25,594] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:33:33,022] loop.py:45 INFO TRAIN_STEP step=6919 loss=1.494511 Training: 38%|███▊ | 6919/18018 [15:03:19<24:37:15, 7.99s/step, loss=1.495, acc=0.670, acc_len=1.58, thru=12.7, I=18.7, T=31.6, wait=3.8s, pool=52, epoch=1/2][2026-10-19 04:33:41,391] loop.py:380 INFO TIMING step=6920: step=4.268s data=1.927s compute=4.226s [fwd=3.012s bwd=1.198s opt=0.003s] dispatch=4.092s [2026-10-19 04:33:41,391] loop.py:45 INFO TRAIN_STEP step=6920 loss=1.432988 Training: 38%|███▊ | 6920/18018 [15:03:27<24:58:25, 8.10s/step, loss=1.433, acc=0.682, acc_len=1.62, thru=14.0, I=14.1, T=30.0, wait=4.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:47,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:33:49,506] loop.py:45 INFO TRAIN_STEP step=6921 loss=1.438671 Training: 38%|███▊ | 6921/18018 [15:03:35<24:59:04, 8.11s/step, loss=1.439, acc=0.681, acc_len=1.62, thru=16.7, I=20.1, T=25.0, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:33:49,501] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6920: forward=3517.3ms backward=1551.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:49,565] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:53,928] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:33:55,659] loop.py:45 INFO TRAIN_STEP step=6922 loss=1.532277 Training: 38%|███▊ | 6922/18018 [15:03:41<23:10:33, 7.52s/step, loss=1.532, acc=0.661, acc_len=1.56, thru=17.3, I=26.7, T=29.6, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:33:55,938] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:02,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:34:02,988] loop.py:45 INFO TRAIN_STEP step=6923 loss=1.380683 Training: 38%|███▊ | 6923/18018 [15:03:49<22:59:59, 7.46s/step, loss=1.381, acc=0.690, acc_len=1.65, thru=15.2, I=20.1, T=31.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:03,153] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:08,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:34:10,704] loop.py:45 INFO TRAIN_STEP step=6924 loss=1.449781 Training: 38%|███▊ | 6924/18018 [15:03:56<23:13:54, 7.54s/step, loss=1.450, acc=0.674, acc_len=1.61, thru=15.6, I=20.2, T=26.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:11,055] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:18,258] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:34:19,475] loop.py:380 INFO TIMING step=6925: step=4.260s data=2.060s compute=4.220s [fwd=2.998s bwd=1.207s opt=0.003s] dispatch=4.501s [2026-10-19 04:34:19,475] loop.py:45 INFO TRAIN_STEP step=6925 loss=1.456338 Training: 38%|███▊ | 6925/18018 [15:04:05<24:22:06, 7.91s/step, loss=1.456, acc=0.677, acc_len=1.61, thru=15.2, I=18.3, T=30.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:19,770] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:26,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:34:28,603] loop.py:45 INFO TRAIN_STEP step=6926 loss=1.598755 Training: 38%|███▊ | 6926/18018 [15:04:14<25:29:32, 8.27s/step, loss=1.599, acc=0.653, acc_len=1.53, thru=14.5, I=17.6, T=22.9, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:34:28,598] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6925: forward=3916.0ms backward=1620.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:28,910] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:33,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:34:35,487] loop.py:45 INFO TRAIN_STEP step=6927 loss=1.484880 Training: 38%|███▊ | 6927/18018 [15:04:21<24:12:20, 7.86s/step, loss=1.485, acc=0.668, acc_len=1.58, thru=15.7, I=22.2, T=31.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:35,858] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:43,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:34:44,118] loop.py:45 INFO TRAIN_STEP step=6928 loss=1.425113 Training: 38%|███▊ | 6928/18018 [15:04:30<24:55:08, 8.09s/step, loss=1.425, acc=0.677, acc_len=1.61, thru=15.5, I=16.4, T=28.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:44,493] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:50,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:34:53,071] loop.py:45 INFO TRAIN_STEP step=6929 loss=1.453327 Training: 38%|███▊ | 6929/18018 [15:04:39<25:42:53, 8.35s/step, loss=1.453, acc=0.679, acc_len=1.62, thru=14.6, I=16.8, T=28.0, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:53,445] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:34:58,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:35:00,162] loop.py:380 INFO TIMING step=6930: step=4.572s data=2.250s compute=4.546s [fwd=3.236s bwd=1.296s opt=0.003s] dispatch=2.510s [2026-10-19 04:35:00,162] loop.py:45 INFO TRAIN_STEP step=6930 loss=1.454084 Training: 38%|███▊ | 6930/18018 [15:04:46<24:33:03, 7.97s/step, loss=1.454, acc=0.672, acc_len=1.59, thru=15.4, I=26.0, T=28.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:00,411] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:04,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:35:07,071] loop.py:45 INFO TRAIN_STEP step=6931 loss=1.451212 Training: 38%|███▊ | 6931/18018 [15:04:53<23:34:00, 7.65s/step, loss=1.451, acc=0.676, acc_len=1.61, thru=15.8, I=27.2, T=29.3, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:35:07,067] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6930: forward=3021.7ms backward=1289.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:07,330] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:12,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:35:13,553] loop.py:45 INFO TRAIN_STEP step=6932 loss=1.463190 Training: 38%|███▊ | 6932/18018 [15:04:59<22:29:08, 7.30s/step, loss=1.463, acc=0.675, acc_len=1.60, thru=17.2, I=28.1, T=31.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:14,048] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:20,749] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:35:22,424] loop.py:45 INFO TRAIN_STEP step=6933 loss=1.508682 Training: 38%|███▊ | 6933/18018 [15:05:08<23:56:01, 7.77s/step, loss=1.509, acc=0.668, acc_len=1.58, thru=15.3, I=16.4, T=27.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:22,765] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:27,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:35:29,712] loop.py:45 INFO TRAIN_STEP step=6934 loss=1.344446 Training: 38%|███▊ | 6934/18018 [15:05:15<23:28:54, 7.63s/step, loss=1.344, acc=0.695, acc_len=1.67, thru=16.4, I=23.3, T=30.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:30,110] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:36,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:35:36,489] loop.py:380 INFO TIMING step=6935: step=4.088s data=1.942s compute=4.054s [fwd=2.817s bwd=1.223s opt=0.003s] dispatch=2.680s [2026-10-19 04:35:36,489] loop.py:45 INFO TRAIN_STEP step=6935 loss=1.452900 Training: 38%|███▊ | 6935/18018 [15:05:22<22:41:40, 7.37s/step, loss=1.453, acc=0.678, acc_len=1.61, thru=16.8, I=22.9, T=31.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:36,824] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:43,128] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:35:44,619] loop.py:45 INFO TRAIN_STEP step=6936 loss=1.465286 Training: 38%|███▊ | 6936/18018 [15:05:30<23:23:40, 7.60s/step, loss=1.465, acc=0.673, acc_len=1.60, thru=15.6, I=18.8, T=28.8, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:35:44,615] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6935: forward=3101.6ms backward=1294.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:44,650] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:50,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:35:51,027] loop.py:45 INFO TRAIN_STEP step=6937 loss=1.432909 Training: 39%|███▊ | 6937/18018 [15:05:37<22:17:20, 7.24s/step, loss=1.433, acc=0.675, acc_len=1.61, thru=18.7, I=21.8, T=30.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:51,478] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:58,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:35:59,008] loop.py:45 INFO TRAIN_STEP step=6938 loss=1.546526 Training: 39%|███▊ | 6938/18018 [15:05:45<22:58:15, 7.46s/step, loss=1.547, acc=0.661, acc_len=1.56, thru=15.4, I=17.1, T=32.1, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:35:59,364] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:07,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:36:08,335] loop.py:45 INFO TRAIN_STEP step=6939 loss=1.407212 Training: 39%|███▊ | 6939/18018 [15:05:54<24:41:30, 8.02s/step, loss=1.407, acc=0.679, acc_len=1.62, thru=14.2, I=14.7, T=30.2, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:08,794] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:14,079] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:36:14,733] loop.py:380 INFO TIMING step=6940: step=3.610s data=1.510s compute=3.569s [fwd=2.550s bwd=1.005s opt=0.003s] dispatch=2.777s [2026-10-19 04:36:14,733] loop.py:45 INFO TRAIN_STEP step=6940 loss=1.405713 Training: 39%|███▊ | 6940/18018 [15:06:00<23:11:11, 7.53s/step, loss=1.406, acc=0.688, acc_len=1.64, thru=18.0, I=20.0, T=35.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:15,084] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:20,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:36:22,884] loop.py:45 INFO TRAIN_STEP step=6941 loss=1.432758 Training: 39%|███▊ | 6941/18018 [15:06:08<23:45:14, 7.72s/step, loss=1.433, acc=0.680, acc_len=1.62, thru=16.3, I=21.4, T=29.2, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:36:22,879] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6940: forward=2937.5ms backward=1384.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:22,912] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:29,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:36:30,962] loop.py:45 INFO TRAIN_STEP step=6942 loss=1.462592 Training: 39%|███▊ | 6942/18018 [15:06:17<24:04:51, 7.83s/step, loss=1.463, acc=0.675, acc_len=1.60, thru=15.9, I=19.0, T=28.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:31,392] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:37,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:36:39,000] loop.py:45 INFO TRAIN_STEP step=6943 loss=1.477562 Training: 39%|███▊ | 6943/18018 [15:06:25<24:16:32, 7.89s/step, loss=1.478, acc=0.672, acc_len=1.59, thru=13.7, I=20.3, T=29.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:39,327] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:36:47,044] loop.py:45 INFO TRAIN_STEP step=6944 loss=1.328912 Training: 39%|███▊ | 6944/18018 [15:06:33<24:24:49, 7.94s/step, loss=1.329, acc=0.699, acc_len=1.69, thru=11.3, I=14.6, T=29.9, wait=3.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:54,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:36:56,598] loop.py:380 INFO TIMING step=6945: step=4.307s data=1.982s compute=4.269s [fwd=3.037s bwd=1.218s opt=0.003s] dispatch=5.238s [2026-10-19 04:36:56,598] loop.py:45 INFO TRAIN_STEP step=6945 loss=1.492103 Training: 39%|███▊ | 6945/18018 [15:06:42<25:54:19, 8.42s/step, loss=1.492, acc=0.673, acc_len=1.60, thru=16.1, I=16.9, T=29.7, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:36:57,041] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:03,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 04:37:03,593] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6945: forward=2727.9ms backward=1069.3ms [2026-10-19 04:37:03,597] loop.py:45 INFO TRAIN_STEP step=6946 loss=1.457289 Training: 39%|███▊ | 6946/18018 [15:06:49<24:35:22, 8.00s/step, loss=1.457, acc=0.675, acc_len=1.60, thru=15.2, I=21.0, T=33.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:03,884] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 04:37:10,795] loop.py:45 INFO TRAIN_STEP step=6947 loss=1.403244 Training: 39%|███▊ | 6947/18018 [15:06:56<23:51:13, 7.76s/step, loss=1.403, acc=0.683, acc_len=1.62, thru=15.8, I=18.0, T=31.0, wait=3.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:17,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:37:18,278] loop.py:45 INFO TRAIN_STEP step=6948 loss=1.453229 Training: 39%|███▊ | 6948/18018 [15:07:04<23:35:46, 7.67s/step, loss=1.453, acc=0.676, acc_len=1.61, thru=17.7, I=18.9, T=31.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:18,696] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:25,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:37:30,124] loop.py:45 INFO TRAIN_STEP step=6949 loss=1.387408 Training: 39%|███▊ | 6949/18018 [15:07:16<27:26:31, 8.93s/step, loss=1.387, acc=0.688, acc_len=1.65, thru=11.2, I=18.3, T=17.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:30,306] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:30,307] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=0/64 [2026-10-19 04:37:37,988] loop.py:380 INFO TIMING step=6950: step=4.820s data=2.496s compute=4.786s [fwd=3.417s bwd=1.354s opt=0.003s] dispatch=3.036s [2026-10-19 04:37:37,988] loop.py:45 INFO TRAIN_STEP step=6950 loss=1.284477 Training: 39%|███▊ | 6950/18018 [15:07:24<26:27:38, 8.61s/step, loss=1.284, acc=0.709, acc_len=1.72, thru=15.1, I=17.5, T=26.6, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:46,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:37:47,450] loop.py:45 INFO TRAIN_STEP step=6951 loss=1.517703 Training: 39%|███▊ | 6951/18018 [15:07:33<27:15:00, 8.86s/step, loss=1.518, acc=0.667, acc_len=1.57, thru=15.2, I=15.1, T=25.3, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:37:47,445] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6950: forward=3772.4ms backward=1241.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:47,759] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:55,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:37:57,011] loop.py:45 INFO TRAIN_STEP step=6952 loss=1.357236 Training: 39%|███▊ | 6952/18018 [15:07:43<27:53:16, 9.07s/step, loss=1.357, acc=0.690, acc_len=1.65, thru=13.8, I=15.4, T=23.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:37:57,297] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:38:02,776] loop.py:45 INFO TRAIN_STEP step=6953 loss=1.478309 Training: 39%|███▊ | 6953/18018 [15:07:48<24:50:06, 8.08s/step, loss=1.478, acc=0.670, acc_len=1.59, thru=16.3, I=22.2, T=33.3, wait=1.9s, pool=48, epoch=1/2][2026-10-19 04:38:10,939] loop.py:45 INFO TRAIN_STEP step=6954 loss=1.551636 Training: 39%|███▊ | 6954/18018 [15:07:57<24:54:35, 8.11s/step, loss=1.552, acc=0.660, acc_len=1.55, thru=14.9, I=15.7, T=29.8, wait=3.9s, pool=56, epoch=1/2][2026-10-19 04:38:18,327] loop.py:380 INFO TIMING step=6955: step=4.424s data=2.186s compute=4.379s [fwd=3.089s bwd=1.275s opt=0.003s] dispatch=2.955s [2026-10-19 04:38:18,327] loop.py:45 INFO TRAIN_STEP step=6955 loss=1.375957 Training: 39%|███▊ | 6955/18018 [15:08:04<24:14:50, 7.89s/step, loss=1.376, acc=0.687, acc_len=1.64, thru=16.1, I=16.2, T=28.9, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:38:25,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:38:27,745] loop.py:45 INFO TRAIN_STEP step=6956 loss=1.575898 Training: 39%|███▊ | 6956/18018 [15:08:13<25:39:08, 8.35s/step, loss=1.576, acc=0.653, acc_len=1.52, thru=14.1, I=16.7, T=28.7, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:38:27,740] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6955: forward=3121.6ms backward=1276.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:38:27,826] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 04:38:36,565] loop.py:45 INFO TRAIN_STEP step=6957 loss=1.453220 Training: 39%|███▊ | 6957/18018 [15:08:22<26:05:11, 8.49s/step, loss=1.453, acc=0.675, acc_len=1.60, thru=14.2, I=14.0, T=30.5, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:38:42,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:38:43,724] loop.py:45 INFO TRAIN_STEP step=6958 loss=1.436503 Training: 39%|███▊ | 6958/18018 [15:08:29<24:51:20, 8.09s/step, loss=1.437, acc=0.683, acc_len=1.62, thru=16.8, I=18.2, T=31.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:38:44,073] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:38:50,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:38:50,648] loop.py:45 INFO TRAIN_STEP step=6959 loss=1.486738 Training: 39%|███▊ | 6959/18018 [15:08:36<23:46:40, 7.74s/step, loss=1.487, acc=0.672, acc_len=1.59, thru=16.9, I=20.2, T=31.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:38:51,056] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:38:57,730] loop.py:380 INFO TIMING step=6960: step=4.555s data=2.004s compute=4.526s [fwd=3.122s bwd=1.389s opt=0.003s] dispatch=2.520s [2026-10-19 04:38:57,731] loop.py:45 INFO TRAIN_STEP step=6960 loss=1.351613 Training: 39%|███▊ | 6960/18018 [15:08:43<23:10:15, 7.54s/step, loss=1.352, acc=0.694, acc_len=1.66, thru=17.2, I=18.1, T=28.1, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:39:06,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:39:07,547] loop.py:45 INFO TRAIN_STEP step=6961 loss=1.546900 Training: 39%|███▊ | 6961/18018 [15:08:53<25:15:44, 8.23s/step, loss=1.547, acc=0.660, acc_len=1.55, thru=13.8, I=16.7, T=27.7, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:39:07,542] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6960: forward=3156.7ms backward=1412.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:39:07,577] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-19 04:39:14,103] loop.py:45 INFO TRAIN_STEP step=6962 loss=1.523949 Training: 39%|███▊ | 6962/18018 [15:09:00<23:43:31, 7.73s/step, loss=1.524, acc=0.659, acc_len=1.55, thru=15.9, I=23.8, T=30.0, wait=2.3s, pool=44, epoch=1/2][2026-10-19 04:39:22,136] loop.py:45 INFO TRAIN_STEP step=6963 loss=1.368953 Training: 39%|███▊ | 6963/18018 [15:09:08<24:00:11, 7.82s/step, loss=1.369, acc=0.692, acc_len=1.65, thru=15.6, I=14.5, T=30.7, wait=3.9s, pool=60, epoch=1/2][2026-10-19 04:39:29,881] loop.py:45 INFO TRAIN_STEP step=6964 loss=1.639537 Training: 39%|███▊ | 6964/18018 [15:09:15<23:56:06, 7.80s/step, loss=1.640, acc=0.641, acc_len=1.49, thru=16.6, I=18.2, T=27.5, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:39:36,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:39:38,701] loop.py:380 INFO TIMING step=6965: step=4.666s data=2.119s compute=4.621s [fwd=3.253s bwd=1.353s opt=0.003s] dispatch=4.145s [2026-10-19 04:39:38,701] loop.py:45 INFO TRAIN_STEP step=6965 loss=1.518873 Training: 39%|███▊ | 6965/18018 [15:09:24<24:52:45, 8.10s/step, loss=1.519, acc=0.666, acc_len=1.56, thru=14.7, I=16.4, T=27.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:39:38,950] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:39:44,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:39:45,011] loop.py:45 INFO TRAIN_STEP step=6966 loss=1.424206 Training: 39%|███▊ | 6966/18018 [15:09:31<23:13:32, 7.57s/step, loss=1.424, acc=0.683, acc_len=1.62, thru=16.6, I=21.7, T=33.2, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:39:45,007] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6965: forward=2635.3ms backward=1168.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:39:45,125] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-19 04:39:52,823] loop.py:45 INFO TRAIN_STEP step=6967 loss=1.479172 Training: 39%|███▊ | 6967/18018 [15:09:38<23:26:58, 7.64s/step, loss=1.479, acc=0.671, acc_len=1.59, thru=16.7, I=15.9, T=29.6, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:00,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:40:01,240] loop.py:45 INFO TRAIN_STEP step=6968 loss=1.497526 Training: 39%|███▊ | 6968/18018 [15:09:47<24:09:51, 7.87s/step, loss=1.498, acc=0.670, acc_len=1.58, thru=15.7, I=15.8, T=27.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:01,656] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:09,544] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:40:10,637] loop.py:45 INFO TRAIN_STEP step=6969 loss=1.323720 Training: 39%|███▊ | 6969/18018 [15:09:56<25:33:59, 8.33s/step, loss=1.324, acc=0.701, acc_len=1.69, thru=13.2, I=16.4, T=25.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:11,055] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:15,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:40:16,901] loop.py:380 INFO TIMING step=6970: step=3.916s data=1.737s compute=3.888s [fwd=2.734s bwd=1.140s opt=0.003s] dispatch=2.337s [2026-10-19 04:40:16,902] loop.py:45 INFO TRAIN_STEP step=6970 loss=1.386852 Training: 39%|███▊ | 6970/18018 [15:10:02<23:39:35, 7.71s/step, loss=1.387, acc=0.684, acc_len=1.63, thru=16.4, I=21.8, T=32.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:17,185] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:40:23,494] loop.py:45 INFO TRAIN_STEP step=6971 loss=1.524124 (TrainerActor pid=9310) [2026-10-19 04:40:23,489] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6970: forward=2634.5ms backward=1197.1ms Training: 39%|███▊ | 6971/18018 [15:10:09<22:37:50, 7.37s/step, loss=1.524, acc=0.662, acc_len=1.55, thru=14.8, I=18.7, T=33.0, wait=2.7s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:34,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:40:35,470] loop.py:45 INFO TRAIN_STEP step=6972 loss=1.381345 Training: 39%|███▊ | 6972/18018 [15:10:21<26:51:45, 8.75s/step, loss=1.381, acc=0.690, acc_len=1.66, thru=14.1, I=14.1, T=28.1, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:35,499] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:40,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:40:42,068] loop.py:45 INFO TRAIN_STEP step=6973 loss=1.378204 Training: 39%|███▊ | 6973/18018 [15:10:28<24:52:29, 8.11s/step, loss=1.378, acc=0.689, acc_len=1.65, thru=17.4, I=21.0, T=34.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:42,497] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:47,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:40:48,248] loop.py:45 INFO TRAIN_STEP step=6974 loss=1.535503 Training: 39%|███▊ | 6974/18018 [15:10:34<23:06:00, 7.53s/step, loss=1.536, acc=0.661, acc_len=1.55, thru=19.6, I=25.1, T=34.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:48,437] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:55,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:40:56,219] loop.py:380 INFO TIMING step=6975: step=4.374s data=2.077s compute=4.336s [fwd=3.062s bwd=1.259s opt=0.003s] dispatch=3.587s [2026-10-19 04:40:56,219] loop.py:45 INFO TRAIN_STEP step=6975 loss=1.359630 Training: 39%|███▊ | 6975/18018 [15:10:42<23:30:13, 7.66s/step, loss=1.360, acc=0.693, acc_len=1.66, thru=15.5, I=17.0, T=29.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:40:56,703] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:02,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:41:03,725] loop.py:45 INFO TRAIN_STEP step=6976 loss=1.456726 Training: 39%|███▊ | 6976/18018 [15:10:49<23:21:25, 7.62s/step, loss=1.457, acc=0.674, acc_len=1.60, thru=16.0, I=20.5, T=33.1, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:41:03,720] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6975: forward=2725.9ms backward=1066.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:04,076] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:11,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:41:12,746] loop.py:45 INFO TRAIN_STEP step=6977 loss=1.526065 Training: 39%|███▊ | 6977/18018 [15:10:58<24:39:01, 8.04s/step, loss=1.526, acc=0.664, acc_len=1.58, thru=14.8, I=16.9, T=30.9, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:12,780] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:20,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:41:20,605] loop.py:45 INFO TRAIN_STEP step=6978 loss=1.433672 Training: 39%|███▊ | 6978/18018 [15:11:06<24:28:59, 7.98s/step, loss=1.434, acc=0.678, acc_len=1.62, thru=15.7, I=18.5, T=30.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:20,899] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:28,182] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:41:28,713] loop.py:45 INFO TRAIN_STEP step=6979 loss=1.420234 Training: 39%|███▊ | 6979/18018 [15:11:14<24:35:38, 8.02s/step, loss=1.420, acc=0.679, acc_len=1.62, thru=16.3, I=17.3, T=29.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:29,191] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:34,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:41:35,800] loop.py:380 INFO TIMING step=6980: step=4.186s data=1.901s compute=4.154s [fwd=2.969s bwd=1.170s opt=0.003s] dispatch=2.891s [2026-10-19 04:41:35,800] loop.py:45 INFO TRAIN_STEP step=6980 loss=1.511932 Training: 39%|███▊ | 6980/18018 [15:11:21<23:43:58, 7.74s/step, loss=1.512, acc=0.662, acc_len=1.56, thru=16.3, I=19.5, T=30.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:35,859] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:42,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:41:43,552] loop.py:45 INFO TRAIN_STEP step=6981 loss=1.497543 Training: 39%|███▊ | 6981/18018 [15:11:29<23:44:36, 7.74s/step, loss=1.498, acc=0.673, acc_len=1.60, thru=15.3, I=18.5, T=32.4, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:41:43,547] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6980: forward=2711.7ms backward=1188.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:43,923] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:41:50,971] loop.py:45 INFO TRAIN_STEP step=6982 loss=1.341570 Training: 39%|███▉ | 6982/18018 [15:11:37<23:26:23, 7.65s/step, loss=1.342, acc=0.693, acc_len=1.66, thru=15.6, I=18.4, T=34.6, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:57,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:41:59,344] loop.py:45 INFO TRAIN_STEP step=6983 loss=1.480694 Training: 39%|███▉ | 6983/18018 [15:11:45<24:06:22, 7.86s/step, loss=1.481, acc=0.668, acc_len=1.59, thru=15.8, I=18.4, T=30.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:41:59,645] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:06,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:42:06,980] loop.py:45 INFO TRAIN_STEP step=6984 loss=1.401793 Training: 39%|███▉ | 6984/18018 [15:11:53<23:53:47, 7.80s/step, loss=1.402, acc=0.685, acc_len=1.63, thru=14.4, I=20.9, T=28.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:07,181] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:13,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:42:14,327] loop.py:380 INFO TIMING step=6985: step=3.937s data=1.730s compute=3.908s [fwd=2.783s bwd=1.110s opt=0.003s] dispatch=3.399s [2026-10-19 04:42:14,327] loop.py:45 INFO TRAIN_STEP step=6985 loss=1.343301 Training: 39%|███▉ | 6985/18018 [15:12:00<23:28:48, 7.66s/step, loss=1.343, acc=0.693, acc_len=1.66, thru=16.8, I=19.7, T=32.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:14,766] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:42:21,601] loop.py:45 INFO TRAIN_STEP step=6986 loss=1.410221 Training: 39%|███▉ | 6986/18018 [15:12:07<23:07:13, 7.54s/step, loss=1.410, acc=0.688, acc_len=1.64, thru=14.1, I=15.9, T=35.1, wait=3.6s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:42:21,596] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6985: forward=2522.2ms backward=1066.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:30,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:42:30,834] loop.py:45 INFO TRAIN_STEP step=6987 loss=1.368715 Training: 39%|███▉ | 6987/18018 [15:12:16<24:40:21, 8.05s/step, loss=1.369, acc=0.692, acc_len=1.66, thru=17.9, I=17.1, T=31.0, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:31,113] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:39,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:42:40,444] loop.py:45 INFO TRAIN_STEP step=6988 loss=1.492813 Training: 39%|███▉ | 6988/18018 [15:12:26<26:06:06, 8.52s/step, loss=1.493, acc=0.668, acc_len=1.57, thru=13.7, I=13.6, T=29.9, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:40,503] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:48,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:42:49,963] loop.py:45 INFO TRAIN_STEP step=6989 loss=1.544388 Training: 39%|███▉ | 6989/18018 [15:12:36<27:00:57, 8.82s/step, loss=1.544, acc=0.665, acc_len=1.56, thru=13.2, I=15.6, T=28.2, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:50,367] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:56,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:42:56,658] loop.py:380 INFO TIMING step=6990: step=4.123s data=1.928s compute=4.091s [fwd=2.864s bwd=1.212s opt=0.003s] dispatch=2.563s [2026-10-19 04:42:56,658] loop.py:45 INFO TRAIN_STEP step=6990 loss=1.468463 Training: 39%|███▉ | 6990/18018 [15:12:42<25:03:45, 8.18s/step, loss=1.468, acc=0.671, acc_len=1.58, thru=17.1, I=20.9, T=31.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:42:57,028] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:04,879] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:43:05,816] loop.py:45 INFO TRAIN_STEP step=6991 loss=1.448791 Training: 39%|███▉ | 6991/18018 [15:12:51<25:57:34, 8.48s/step, loss=1.449, acc=0.678, acc_len=1.61, thru=14.5, I=16.3, T=26.5, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:43:05,812] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6990: forward=3216.0ms backward=1558.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:05,884] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:11,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:43:12,694] loop.py:45 INFO TRAIN_STEP step=6992 loss=1.423866 Training: 39%|███▉ | 6992/18018 [15:12:58<24:29:17, 8.00s/step, loss=1.424, acc=0.681, acc_len=1.62, thru=17.3, I=19.3, T=32.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:12,723] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:18,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:43:19,054] loop.py:45 INFO TRAIN_STEP step=6993 loss=1.381616 Training: 39%|███▉ | 6993/18018 [15:13:05<22:58:59, 7.50s/step, loss=1.382, acc=0.683, acc_len=1.63, thru=19.3, I=22.0, T=32.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:19,253] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:24,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:43:25,229] loop.py:45 INFO TRAIN_STEP step=6994 loss=1.470032 Training: 39%|███▉ | 6994/18018 [15:13:11<21:45:35, 7.11s/step, loss=1.470, acc=0.673, acc_len=1.59, thru=19.4, I=23.6, T=35.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:25,644] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:32,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:43:33,513] loop.py:380 INFO TIMING step=6995: step=3.796s data=1.673s compute=3.755s [fwd=2.691s bwd=1.048s opt=0.003s] dispatch=4.480s [2026-10-19 04:43:33,514] loop.py:45 INFO TRAIN_STEP step=6995 loss=1.469194 Training: 39%|███▉ | 6995/18018 [15:13:19<22:50:29, 7.46s/step, loss=1.469, acc=0.673, acc_len=1.59, thru=14.3, I=18.6, T=33.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:33,848] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:43:40,927] loop.py:45 INFO TRAIN_STEP step=6996 loss=1.451161 Training: 39%|███▉ | 6996/18018 [15:13:27<22:47:49, 7.45s/step, loss=1.451, acc=0.677, acc_len=1.60, thru=15.0, I=18.0, T=34.0, wait=3.6s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:43:40,922] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=6995: forward=2653.6ms backward=1061.5ms [2026-10-19 04:43:48,328] loop.py:45 INFO TRAIN_STEP step=6997 loss=1.474443 Training: 39%|███▉ | 6997/18018 [15:13:34<22:45:06, 7.43s/step, loss=1.474, acc=0.675, acc_len=1.61, thru=17.6, I=16.2, T=30.7, wait=3.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:57,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:43:57,708] loop.py:45 INFO TRAIN_STEP step=6998 loss=1.517705 Training: 39%|███▉ | 6998/18018 [15:13:43<24:32:16, 8.02s/step, loss=1.518, acc=0.663, acc_len=1.56, thru=14.9, I=15.1, T=29.2, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:43:58,168] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:03,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:44:05,330] loop.py:45 INFO TRAIN_STEP step=6999 loss=1.395914 Training: 39%|███▉ | 6999/18018 [15:13:51<24:10:31, 7.90s/step, loss=1.396, acc=0.683, acc_len=1.62, thru=16.4, I=20.6, T=26.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:05,734] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:13,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:44:14,796] loop.py:380 INFO TIMING step=7000: step=4.557s data=2.293s compute=4.508s [fwd=3.093s bwd=1.400s opt=0.003s] dispatch=4.901s [2026-10-19 04:44:14,797] loop.py:45 INFO TRAIN_STEP step=7000 loss=1.452732 Training: 39%|███▉ | 7000/18018 [15:14:00<25:36:53, 8.37s/step, loss=1.453, acc=0.677, acc_len=1.61, thru=14.1, I=16.7, T=28.1, wait=4.9s, pool=64, epoch=1/2][2026-10-19 04:44:14,802] loop.py:419 INFO Saving checkpoint at step 7000... (TrainerActor pid=9310) [2026-10-19 04:44:18,162] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0007001 (TrainerActor pid=9310) [2026-10-19 04:44:18,170] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0006251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:18,401] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:19,408] inference_manager.py:408 INFO Pool capacity available after 6.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:26,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 04:44:18,163] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0007001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 04:44:19,256] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0006251 [2026-10-19 04:44:26,359] loop.py:45 INFO TRAIN_STEP step=7001 loss=1.362295 Training: 39%|███▉ | 7001/18018 [15:14:12<28:32:36, 9.33s/step, loss=1.362, acc=0.689, acc_len=1.64, thru=18.4, I=20.9, T=25.0, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:44:26,355] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7000: forward=3532.3ms backward=1520.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:26,617] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:33,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:44:34,449] loop.py:45 INFO TRAIN_STEP step=7002 loss=1.389190 Training: 39%|███▉ | 7002/18018 [15:14:20<27:24:12, 8.96s/step, loss=1.389, acc=0.686, acc_len=1.64, thru=14.4, I=15.9, T=29.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:34,699] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:44:41,614] loop.py:45 INFO TRAIN_STEP step=7003 loss=1.423433 Training: 39%|███▉ | 7003/18018 [15:14:27<25:45:26, 8.42s/step, loss=1.423, acc=0.684, acc_len=1.64, thru=15.7, I=16.3, T=28.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:41,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:42,094] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:46,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:44:48,475] loop.py:45 INFO TRAIN_STEP step=7004 loss=1.560410 Training: 39%|███▉ | 7004/18018 [15:14:34<24:19:35, 7.95s/step, loss=1.560, acc=0.660, acc_len=1.55, thru=18.0, I=19.8, T=31.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:48,825] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:55,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:44:56,313] loop.py:380 INFO TIMING step=7005: step=3.630s data=1.605s compute=3.601s [fwd=2.608s bwd=0.978s opt=0.003s] dispatch=4.200s [2026-10-19 04:44:56,313] loop.py:45 INFO TRAIN_STEP step=7005 loss=1.403122 Training: 39%|███▉ | 7005/18018 [15:14:42<24:13:14, 7.92s/step, loss=1.403, acc=0.684, acc_len=1.63, thru=14.6, I=20.0, T=35.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:44:56,508] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:02,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:45:03,102] loop.py:45 INFO TRAIN_STEP step=7006 loss=1.266826 Training: 39%|███▉ | 7006/18018 [15:14:49<23:10:56, 7.58s/step, loss=1.267, acc=0.707, acc_len=1.71, thru=18.6, I=22.2, T=32.1, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:45:03,098] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7005: forward=2778.9ms backward=1146.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:03,360] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:10,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:45:11,087] loop.py:45 INFO TRAIN_STEP step=7007 loss=1.439069 Training: 39%|███▉ | 7007/18018 [15:14:57<23:33:10, 7.70s/step, loss=1.439, acc=0.677, acc_len=1.61, thru=16.4, I=19.5, T=30.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:11,457] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:17,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:45:18,396] loop.py:45 INFO TRAIN_STEP step=7008 loss=1.447772 Training: 39%|███▉ | 7008/18018 [15:15:04<23:11:28, 7.58s/step, loss=1.448, acc=0.677, acc_len=1.60, thru=16.9, I=19.0, T=29.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:18,866] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:45:26,378] loop.py:45 INFO TRAIN_STEP step=7009 loss=1.346303 Training: 39%|███▉ | 7009/18018 [15:15:12<23:33:21, 7.70s/step, loss=1.346, acc=0.693, acc_len=1.66, thru=15.3, I=16.9, T=33.8, wait=4.2s, pool=60, epoch=1/2][2026-10-19 04:45:33,838] loop.py:380 INFO TIMING step=7010: step=3.780s data=1.625s compute=3.750s [fwd=2.683s bwd=1.051s opt=0.003s] dispatch=3.670s [2026-10-19 04:45:33,838] loop.py:45 INFO TRAIN_STEP step=7010 loss=1.428036 Training: 39%|███▉ | 7010/18018 [15:15:19<23:19:58, 7.63s/step, loss=1.428, acc=0.679, acc_len=1.62, thru=16.7, I=15.3, T=33.9, wait=3.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:41,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:45:42,762] loop.py:45 INFO TRAIN_STEP step=7011 loss=1.420131 Training: 39%|███▉ | 7011/18018 [15:15:28<24:30:59, 8.02s/step, loss=1.420, acc=0.686, acc_len=1.63, thru=17.1, I=19.2, T=29.4, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:45:42,758] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7010: forward=3090.6ms backward=1212.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:42,906] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:51,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:45:51,942] loop.py:45 INFO TRAIN_STEP step=7012 loss=1.349222 Training: 39%|███▉ | 7012/18018 [15:15:38<25:34:38, 8.37s/step, loss=1.349, acc=0.692, acc_len=1.66, thru=14.3, I=12.7, T=26.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:52,335] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:56,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:45:59,143] loop.py:45 INFO TRAIN_STEP step=7013 loss=1.480310 Training: 39%|███▉ | 7013/18018 [15:15:45<24:30:33, 8.02s/step, loss=1.480, acc=0.674, acc_len=1.59, thru=16.5, I=21.6, T=26.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:45:59,351] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:07,314] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:46:07,489] loop.py:45 INFO TRAIN_STEP step=7014 loss=1.353565 Training: 39%|███▉ | 7014/18018 [15:15:53<24:48:21, 8.12s/step, loss=1.354, acc=0.692, acc_len=1.66, thru=15.7, I=15.5, T=32.8, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:07,816] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:13,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:46:14,358] loop.py:380 INFO TIMING step=7015: step=4.437s data=1.996s compute=4.406s [fwd=3.079s bwd=1.311s opt=0.003s] dispatch=2.425s [2026-10-19 04:46:14,359] loop.py:45 INFO TRAIN_STEP step=7015 loss=1.440570 Training: 39%|███▉ | 7015/18018 [15:16:00<23:39:41, 7.74s/step, loss=1.441, acc=0.680, acc_len=1.61, thru=18.6, I=20.3, T=28.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:14,562] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:46:20,604] loop.py:45 INFO TRAIN_STEP step=7016 loss=1.470832 Training: 39%|███▉ | 7016/18018 [15:16:06<22:17:16, 7.29s/step, loss=1.471, acc=0.672, acc_len=1.59, thru=17.6, I=21.2, T=35.3, wait=2.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:46:20,599] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7015: forward=2561.7ms backward=1011.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:27,265] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:46:27,880] loop.py:45 INFO TRAIN_STEP step=7017 loss=1.348639 Training: 39%|███▉ | 7017/18018 [15:16:13<22:16:11, 7.29s/step, loss=1.349, acc=0.694, acc_len=1.66, thru=17.2, I=17.9, T=32.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:28,268] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:34,854] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:46:35,368] loop.py:45 INFO TRAIN_STEP step=7018 loss=1.521722 Training: 39%|███▉ | 7018/18018 [15:16:21<22:27:04, 7.35s/step, loss=1.522, acc=0.663, acc_len=1.56, thru=16.7, I=18.0, T=28.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:35,860] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:41,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:46:43,428] loop.py:45 INFO TRAIN_STEP step=7019 loss=1.478220 Training: 39%|███▉ | 7019/18018 [15:16:29<23:06:15, 7.56s/step, loss=1.478, acc=0.672, acc_len=1.58, thru=15.6, I=19.1, T=29.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:43,665] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:48,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:46:50,868] loop.py:380 INFO TIMING step=7020: step=5.099s data=2.564s compute=5.070s [fwd=3.402s bwd=1.652s opt=0.003s] dispatch=2.331s [2026-10-19 04:46:50,868] loop.py:45 INFO TRAIN_STEP step=7020 loss=1.388105 Training: 39%|███▉ | 7020/18018 [15:16:36<22:59:19, 7.52s/step, loss=1.388, acc=0.689, acc_len=1.65, thru=15.8, I=26.8, T=25.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:46:51,269] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 04:46:57,250] loop.py:45 INFO TRAIN_STEP step=7021 loss=1.461696 Training: 39%|███▉ | 7021/18018 [15:16:43<21:56:19, 7.18s/step, loss=1.462, acc=0.672, acc_len=1.59, thru=15.4, I=22.5, T=33.6, wait=2.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:46:57,245] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7020: forward=2656.8ms backward=1111.9ms [2026-10-19 04:47:03,864] loop.py:45 INFO TRAIN_STEP step=7022 loss=1.388646 Training: 39%|███▉ | 7022/18018 [15:16:49<21:25:04, 7.01s/step, loss=1.389, acc=0.684, acc_len=1.63, thru=19.2, I=18.3, T=32.1, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:09,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:47:12,302] loop.py:45 INFO TRAIN_STEP step=7023 loss=1.499843 Training: 39%|███▉ | 7023/18018 [15:16:58<22:43:20, 7.44s/step, loss=1.500, acc=0.669, acc_len=1.57, thru=16.5, I=18.8, T=25.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:12,334] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:17,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:47:18,363] loop.py:45 INFO TRAIN_STEP step=7024 loss=1.528543 Training: 39%|███▉ | 7024/18018 [15:17:04<21:27:28, 7.03s/step, loss=1.529, acc=0.661, acc_len=1.56, thru=16.5, I=23.0, T=31.6, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:18,586] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:47:24,553] loop.py:380 INFO TIMING step=7025: step=3.704s data=1.658s compute=3.676s [fwd=2.621s bwd=1.040s opt=0.003s] dispatch=2.475s [2026-10-19 04:47:24,553] loop.py:45 INFO TRAIN_STEP step=7025 loss=1.468630 Training: 39%|███▉ | 7025/18018 [15:17:10<20:41:28, 6.78s/step, loss=1.469, acc=0.672, acc_len=1.60, thru=18.4, I=21.5, T=34.6, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:31,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:47:33,054] loop.py:45 INFO TRAIN_STEP step=7026 loss=1.471853 Training: 39%|███▉ | 7026/18018 [15:17:19<22:16:02, 7.29s/step, loss=1.472, acc=0.672, acc_len=1.59, thru=14.9, I=17.0, T=29.5, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:47:33,049] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7025: forward=3037.1ms backward=1238.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:33,124] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:38,638] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:47:39,849] loop.py:45 INFO TRAIN_STEP step=7027 loss=1.439388 Training: 39%|███▉ | 7027/18018 [15:17:25<21:48:31, 7.14s/step, loss=1.439, acc=0.679, acc_len=1.61, thru=16.6, I=21.6, T=30.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:40,146] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:44,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:47:46,552] loop.py:45 INFO TRAIN_STEP step=7028 loss=1.498603 Training: 39%|███▉ | 7028/18018 [15:17:32<21:24:18, 7.01s/step, loss=1.499, acc=0.667, acc_len=1.58, thru=17.3, I=22.7, T=30.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:46,676] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:51,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:47:54,336] loop.py:45 INFO TRAIN_STEP step=7029 loss=1.354741 Training: 39%|███▉ | 7029/18018 [15:17:40<22:06:31, 7.24s/step, loss=1.355, acc=0.690, acc_len=1.65, thru=14.4, I=23.0, T=24.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:47:54,512] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:48:02,245] loop.py:380 INFO TIMING step=7030: step=4.109s data=1.937s compute=4.079s [fwd=2.903s bwd=1.161s opt=0.003s] dispatch=3.791s [2026-10-19 04:48:02,245] loop.py:45 INFO TRAIN_STEP step=7030 loss=1.513959 Training: 39%|███▉ | 7030/18018 [15:17:48<22:42:54, 7.44s/step, loss=1.514, acc=0.665, acc_len=1.57, thru=16.6, I=17.7, T=31.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:02,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:02,693] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:09,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:48:11,665] loop.py:45 INFO TRAIN_STEP step=7031 loss=1.417966 Training: 39%|███▉ | 7031/18018 [15:17:57<24:31:37, 8.04s/step, loss=1.418, acc=0.685, acc_len=1.63, thru=14.1, I=15.2, T=28.0, wait=4.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:48:11,661] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7030: forward=3258.4ms backward=1267.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:11,851] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:48:18,835] loop.py:45 INFO TRAIN_STEP step=7032 loss=1.328797 Training: 39%|███▉ | 7032/18018 [15:18:04<23:43:45, 7.78s/step, loss=1.329, acc=0.699, acc_len=1.68, thru=14.2, I=16.2, T=31.6, wait=3.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:25,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:48:26,927] loop.py:45 INFO TRAIN_STEP step=7033 loss=1.443341 Training: 39%|███▉ | 7033/18018 [15:18:13<24:01:06, 7.87s/step, loss=1.443, acc=0.676, acc_len=1.61, thru=17.8, I=18.5, T=28.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:27,176] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:33,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:48:34,750] loop.py:45 INFO TRAIN_STEP step=7034 loss=1.453160 Training: 39%|███▉ | 7034/18018 [15:18:20<23:58:09, 7.86s/step, loss=1.453, acc=0.676, acc_len=1.61, thru=14.0, I=19.2, T=30.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:35,185] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:41,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:48:43,298] loop.py:380 INFO TIMING step=7035: step=4.487s data=2.141s compute=4.453s [fwd=3.200s bwd=1.238s opt=0.003s] dispatch=4.052s [2026-10-19 04:48:43,298] loop.py:45 INFO TRAIN_STEP step=7035 loss=1.517838 Training: 39%|███▉ | 7035/18018 [15:18:29<24:36:10, 8.06s/step, loss=1.518, acc=0.667, acc_len=1.58, thru=14.1, I=21.0, T=28.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:43,724] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:51,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:48:51,420] loop.py:45 INFO TRAIN_STEP step=7036 loss=1.400480 Training: 39%|███▉ | 7036/18018 [15:18:37<24:39:05, 8.08s/step, loss=1.400, acc=0.686, acc_len=1.63, thru=16.6, I=16.3, T=29.4, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:48:51,415] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7035: forward=3014.4ms backward=1298.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:51,838] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:48:58,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:48:59,980] loop.py:45 INFO TRAIN_STEP step=7037 loss=1.319342 Training: 39%|███▉ | 7037/18018 [15:18:46<25:05:16, 8.22s/step, loss=1.319, acc=0.696, acc_len=1.67, thru=16.1, I=18.2, T=26.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:49:00,398] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:49:06,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:49:07,865] loop.py:45 INFO TRAIN_STEP step=7038 loss=1.540342 Training: 39%|███▉ | 7038/18018 [15:18:53<24:46:36, 8.12s/step, loss=1.540, acc=0.657, acc_len=1.54, thru=14.2, I=21.2, T=29.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:49:08,038] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:49:14,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:49:14,879] loop.py:45 INFO TRAIN_STEP step=7039 loss=1.379188 Training: 39%|███▉ | 7039/18018 [15:19:00<23:45:25, 7.79s/step, loss=1.379, acc=0.691, acc_len=1.66, thru=16.0, I=21.0, T=28.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:49:15,282] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:49:20,325] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:49:22,355] loop.py:380 INFO TIMING step=7040: step=4.341s data=2.113s compute=4.305s [fwd=3.047s bwd=1.243s opt=0.003s] dispatch=3.127s [2026-10-19 04:49:22,355] loop.py:45 INFO TRAIN_STEP step=7040 loss=1.528945 Training: 39%|███▉ | 7040/18018 [15:19:08<23:28:11, 7.70s/step, loss=1.529, acc=0.660, acc_len=1.55, thru=15.9, I=19.4, T=29.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:49:22,843] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 04:49:30,184] loop.py:45 INFO TRAIN_STEP step=7041 loss=1.439601 Training: 39%|███▉ | 7041/18018 [15:19:16<23:35:11, 7.74s/step, loss=1.440, acc=0.677, acc_len=1.61, thru=12.2, I=15.9, T=29.7, wait=3.5s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:49:30,179] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7040: forward=3076.7ms backward=1190.7ms [2026-10-19 04:49:41,084] loop.py:45 INFO TRAIN_STEP step=7042 loss=1.301495 Training: 39%|███▉ | 7042/18018 [15:19:27<26:28:54, 8.69s/step, loss=1.301, acc=0.701, acc_len=1.69, thru=12.3, I=13.0, T=26.0, wait=6.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:49:49,032] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:49:49,865] loop.py:45 INFO TRAIN_STEP step=7043 loss=1.457317 Training: 39%|███▉ | 7043/18018 [15:19:35<26:33:58, 8.71s/step, loss=1.457, acc=0.677, acc_len=1.60, thru=15.9, I=16.5, T=30.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:49:50,037] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:49:56,669] loop.py:45 INFO TRAIN_STEP step=7044 loss=1.518390 Training: 39%|███▉ | 7044/18018 [15:19:42<24:48:53, 8.14s/step, loss=1.518, acc=0.665, acc_len=1.57, thru=16.0, I=17.6, T=35.2, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:03,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:50:04,888] loop.py:380 INFO TIMING step=7045: step=4.597s data=2.180s compute=4.565s [fwd=3.220s bwd=1.330s opt=0.003s] dispatch=3.613s [2026-10-19 04:50:04,888] loop.py:45 INFO TRAIN_STEP step=7045 loss=1.559342 Training: 39%|███▉ | 7045/18018 [15:19:50<24:53:06, 8.16s/step, loss=1.559, acc=0.662, acc_len=1.56, thru=16.9, I=20.2, T=27.8, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:05,072] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:09,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:50:10,650] loop.py:45 INFO TRAIN_STEP step=7046 loss=1.353037 Training: 39%|███▉ | 7046/18018 [15:19:56<22:41:15, 7.44s/step, loss=1.353, acc=0.691, acc_len=1.66, thru=18.6, I=25.7, T=36.0, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:50:10,645] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7045: forward=2496.9ms backward=1014.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:10,686] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 04:50:18,652] loop.py:45 INFO TRAIN_STEP step=7047 loss=1.441473 Training: 39%|███▉ | 7047/18018 [15:20:04<23:11:38, 7.61s/step, loss=1.441, acc=0.678, acc_len=1.61, thru=15.7, I=17.3, T=31.0, wait=3.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:25,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:50:27,405] loop.py:45 INFO TRAIN_STEP step=7048 loss=1.467915 Training: 39%|███▉ | 7048/18018 [15:20:13<24:14:11, 7.95s/step, loss=1.468, acc=0.673, acc_len=1.59, thru=15.8, I=18.2, T=20.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:27,540] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 [2026-10-19 04:50:34,410] loop.py:45 INFO TRAIN_STEP step=7049 loss=1.431193 Training: 39%|███▉ | 7049/18018 [15:20:20<23:22:02, 7.67s/step, loss=1.431, acc=0.684, acc_len=1.63, thru=13.2, I=15.1, T=30.5, wait=2.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:40,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:50:41,438] loop.py:380 INFO TIMING step=7050: step=4.749s data=2.253s compute=4.724s [fwd=3.356s bwd=1.354s opt=0.003s] dispatch=2.270s [2026-10-19 04:50:41,439] loop.py:45 INFO TRAIN_STEP step=7050 loss=1.467920 Training: 39%|███▉ | 7050/18018 [15:20:27<22:46:45, 7.48s/step, loss=1.468, acc=0.673, acc_len=1.59, thru=17.2, I=18.1, T=27.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:41,487] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-19 04:50:49,258] loop.py:45 INFO TRAIN_STEP step=7051 loss=1.552022 Training: 39%|███▉ | 7051/18018 [15:20:35<23:05:26, 7.58s/step, loss=1.552, acc=0.664, acc_len=1.57, thru=12.5, I=13.5, T=27.4, wait=3.1s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:50:49,252] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7050: forward=3132.9ms backward=1482.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:56,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:50:56,995] loop.py:45 INFO TRAIN_STEP step=7052 loss=1.460709 Training: 39%|███▉ | 7052/18018 [15:20:43<23:14:01, 7.63s/step, loss=1.461, acc=0.673, acc_len=1.59, thru=18.3, I=16.3, T=32.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:50:57,118] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 04:51:03,536] loop.py:45 INFO TRAIN_STEP step=7053 loss=1.365829 Training: 39%|███▉ | 7053/18018 [15:20:49<22:14:20, 7.30s/step, loss=1.366, acc=0.691, acc_len=1.66, thru=17.7, I=23.0, T=28.1, wait=2.0s, pool=56, epoch=1/2][2026-10-19 04:51:12,873] loop.py:45 INFO TRAIN_STEP step=7054 loss=1.450073 Training: 39%|███▉ | 7054/18018 [15:20:58<24:05:39, 7.91s/step, loss=1.450, acc=0.677, acc_len=1.61, thru=13.8, I=13.1, T=28.3, wait=4.8s, pool=60, epoch=1/2][2026-10-19 04:51:23,045] loop.py:380 INFO TIMING step=7055: step=4.386s data=2.106s compute=4.351s [fwd=3.186s bwd=1.151s opt=0.003s] dispatch=5.779s [2026-10-19 04:51:23,046] loop.py:45 INFO TRAIN_STEP step=7055 loss=1.531983 Training: 39%|███▉ | 7055/18018 [15:21:09<26:09:32, 8.59s/step, loss=1.532, acc=0.664, acc_len=1.56, thru=12.4, I=12.5, T=29.2, wait=5.8s, pool=52, epoch=1/2][2026-10-19 04:51:30,120] loop.py:45 INFO TRAIN_STEP step=7056 loss=1.484553 Training: 39%|███▉ | 7056/18018 [15:21:16<24:46:15, 8.13s/step, loss=1.485, acc=0.673, acc_len=1.59, thru=14.9, I=16.8, T=32.0, wait=3.1s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:51:30,115] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7055: forward=2809.8ms backward=1142.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:51:40,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:51:41,564] loop.py:45 INFO TRAIN_STEP step=7057 loss=1.501175 Training: 39%|███▉ | 7057/18018 [15:21:27<27:47:29, 9.13s/step, loss=1.501, acc=0.666, acc_len=1.57, thru=12.2, I=13.8, T=27.4, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:51:41,839] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:51:46,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:51:48,084] loop.py:45 INFO TRAIN_STEP step=7058 loss=1.328055 Training: 39%|███▉ | 7058/18018 [15:21:34<25:24:29, 8.35s/step, loss=1.328, acc=0.692, acc_len=1.66, thru=17.9, I=22.0, T=29.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:51:48,378] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:51:55,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:51:56,244] loop.py:45 INFO TRAIN_STEP step=7059 loss=1.453861 Training: 39%|███▉ | 7059/18018 [15:21:42<25:14:07, 8.29s/step, loss=1.454, acc=0.674, acc_len=1.60, thru=14.9, I=16.4, T=24.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:51:56,355] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:01,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:52:04,450] loop.py:380 INFO TIMING step=7060: step=5.061s data=2.769s compute=5.017s [fwd=3.708s bwd=1.294s opt=0.003s] dispatch=3.137s [2026-10-19 04:52:04,450] loop.py:45 INFO TRAIN_STEP step=7060 loss=1.460552 Training: 39%|███▉ | 7060/18018 [15:21:50<25:09:22, 8.26s/step, loss=1.461, acc=0.679, acc_len=1.62, thru=15.6, I=18.7, T=25.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:04,831] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 04:52:12,348] loop.py:45 INFO TRAIN_STEP step=7061 loss=1.554733 Training: 39%|███▉ | 7061/18018 [15:21:58<24:49:19, 8.16s/step, loss=1.555, acc=0.661, acc_len=1.55, thru=16.0, I=22.6, T=25.7, wait=2.9s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:52:12,343] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7060: forward=3653.9ms backward=1287.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:19,722] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:52:21,318] loop.py:45 INFO TRAIN_STEP step=7062 loss=1.490204 Training: 39%|███▉ | 7062/18018 [15:22:07<25:33:42, 8.40s/step, loss=1.490, acc=0.676, acc_len=1.60, thru=14.5, I=14.8, T=28.0, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:21,733] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:28,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:52:30,904] loop.py:45 INFO TRAIN_STEP step=7063 loss=1.514387 Training: 39%|███▉ | 7063/18018 [15:22:16<26:38:33, 8.76s/step, loss=1.514, acc=0.667, acc_len=1.57, thru=14.0, I=17.6, T=25.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:31,365] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:36,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:52:38,206] loop.py:45 INFO TRAIN_STEP step=7064 loss=1.524250 Training: 39%|███▉ | 7064/18018 [15:22:24<25:18:47, 8.32s/step, loss=1.524, acc=0.664, acc_len=1.57, thru=14.2, I=20.7, T=30.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:38,421] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:43,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:52:44,231] loop.py:380 INFO TIMING step=7065: step=3.889s data=1.729s compute=3.859s [fwd=2.728s bwd=1.116s opt=0.003s] dispatch=2.128s [2026-10-19 04:52:44,231] loop.py:45 INFO TRAIN_STEP step=7065 loss=1.368163 Training: 39%|███▉ | 7065/18018 [15:22:30<23:13:02, 7.63s/step, loss=1.368, acc=0.690, acc_len=1.66, thru=19.1, I=23.6, T=32.9, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:44,678] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:50,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:52:51,577] loop.py:45 INFO TRAIN_STEP step=7066 loss=1.401896 Training: 39%|███▉ | 7066/18018 [15:22:37<22:57:19, 7.55s/step, loss=1.402, acc=0.683, acc_len=1.63, thru=16.9, I=20.2, T=31.6, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:52:51,572] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7065: forward=2828.6ms backward=1180.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:52:51,937] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:00,983] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:53:01,922] loop.py:45 INFO TRAIN_STEP step=7067 loss=1.350036 Training: 39%|███▉ | 7067/18018 [15:22:48<25:30:35, 8.39s/step, loss=1.350, acc=0.694, acc_len=1.66, thru=12.8, I=14.2, T=28.5, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:02,006] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:09,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:53:11,222] loop.py:45 INFO TRAIN_STEP step=7068 loss=1.443989 Training: 39%|███▉ | 7068/18018 [15:22:57<26:20:28, 8.66s/step, loss=1.444, acc=0.675, acc_len=1.61, thru=13.9, I=16.0, T=29.0, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:11,294] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:16,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:53:18,124] loop.py:45 INFO TRAIN_STEP step=7069 loss=1.565207 Training: 39%|███▉ | 7069/18018 [15:23:04<24:44:01, 8.13s/step, loss=1.565, acc=0.658, acc_len=1.54, thru=17.5, I=23.0, T=29.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:18,552] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:24,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:53:26,390] loop.py:380 INFO TIMING step=7070: step=4.340s data=1.915s compute=4.310s [fwd=2.942s bwd=1.353s opt=0.003s] dispatch=3.916s [2026-10-19 04:53:26,390] loop.py:45 INFO TRAIN_STEP step=7070 loss=1.373021 Training: 39%|███▉ | 7070/18018 [15:23:12<24:51:10, 8.17s/step, loss=1.373, acc=0.687, acc_len=1.64, thru=16.3, I=19.4, T=29.5, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:26,745] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:36,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:53:36,268] loop.py:45 INFO TRAIN_STEP step=7071 loss=1.534927 Training: 39%|███▉ | 7071/18018 [15:23:22<26:24:19, 8.68s/step, loss=1.535, acc=0.662, acc_len=1.55, thru=13.4, I=13.1, T=28.4, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:53:36,263] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7070: forward=3082.4ms backward=1364.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:36,698] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 04:53:42,759] loop.py:45 INFO TRAIN_STEP step=7072 loss=1.398857 Training: 39%|███▉ | 7072/18018 [15:23:28<24:24:10, 8.03s/step, loss=1.399, acc=0.685, acc_len=1.64, thru=15.2, I=17.0, T=31.0, wait=2.4s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:53,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:53:54,644] loop.py:45 INFO TRAIN_STEP step=7073 loss=1.498904 Training: 39%|███▉ | 7073/18018 [15:23:40<27:55:18, 9.18s/step, loss=1.499, acc=0.671, acc_len=1.59, thru=16.9, I=14.5, T=26.5, wait=7.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:53:54,999] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:01,183] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:54:01,494] loop.py:45 INFO TRAIN_STEP step=7074 loss=1.300371 Training: 39%|███▉ | 7074/18018 [15:23:47<25:47:29, 8.48s/step, loss=1.300, acc=0.701, acc_len=1.69, thru=15.7, I=18.2, T=33.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:01,686] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 04:54:11,423] loop.py:380 INFO TIMING step=7075: step=4.192s data=1.903s compute=4.161s [fwd=2.922s bwd=1.225s opt=0.003s] dispatch=5.728s [2026-10-19 04:54:11,424] loop.py:45 INFO TRAIN_STEP step=7075 loss=1.504416 Training: 39%|███▉ | 7075/18018 [15:23:57<27:06:19, 8.92s/step, loss=1.504, acc=0.668, acc_len=1.58, thru=11.9, I=11.6, T=30.5, wait=5.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:19,324] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:54:20,849] loop.py:45 INFO TRAIN_STEP step=7076 loss=1.460271 Training: 39%|███▉ | 7076/18018 [15:24:06<27:34:03, 9.07s/step, loss=1.460, acc=0.674, acc_len=1.60, thru=15.3, I=15.4, T=24.6, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:54:20,844] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7075: forward=3532.0ms backward=1632.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:21,337] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 04:54:28,910] loop.py:45 INFO TRAIN_STEP step=7077 loss=1.423395 Training: 39%|███▉ | 7077/18018 [15:24:14<26:38:44, 8.77s/step, loss=1.423, acc=0.676, acc_len=1.60, thru=13.8, I=17.4, T=27.5, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:35,757] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:54:37,202] loop.py:45 INFO TRAIN_STEP step=7078 loss=1.490723 Training: 39%|███▉ | 7078/18018 [15:24:23<26:12:38, 8.63s/step, loss=1.491, acc=0.671, acc_len=1.58, thru=15.9, I=15.7, T=28.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:37,276] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:42,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:54:44,246] loop.py:45 INFO TRAIN_STEP step=7079 loss=1.463814 Training: 39%|███▉ | 7079/18018 [15:24:30<24:45:53, 8.15s/step, loss=1.464, acc=0.676, acc_len=1.60, thru=16.1, I=21.0, T=31.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:44,740] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:52,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:54:52,792] loop.py:380 INFO TIMING step=7080: step=4.211s data=1.980s compute=4.179s [fwd=2.953s bwd=1.211s opt=0.003s] dispatch=4.326s [2026-10-19 04:54:52,792] loop.py:45 INFO TRAIN_STEP step=7080 loss=1.370625 Training: 39%|███▉ | 7080/18018 [15:24:38<25:07:29, 8.27s/step, loss=1.371, acc=0.689, acc_len=1.65, thru=15.9, I=16.8, T=30.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:53,022] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:54:59,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:55:00,462] loop.py:45 INFO TRAIN_STEP step=7081 loss=1.487420 Training: 39%|███▉ | 7081/18018 [15:24:46<24:34:37, 8.09s/step, loss=1.487, acc=0.669, acc_len=1.58, thru=16.6, I=22.4, T=29.1, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:55:00,457] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7080: forward=3098.5ms backward=1242.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:00,637] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:07,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:55:07,756] loop.py:45 INFO TRAIN_STEP step=7082 loss=1.490856 Training: 39%|███▉ | 7082/18018 [15:24:53<23:50:51, 7.85s/step, loss=1.491, acc=0.668, acc_len=1.58, thru=15.7, I=18.3, T=30.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:07,891] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:15,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:55:16,619] loop.py:45 INFO TRAIN_STEP step=7083 loss=1.552148 Training: 39%|███▉ | 7083/18018 [15:25:02<24:46:04, 8.15s/step, loss=1.552, acc=0.663, acc_len=1.56, thru=14.0, I=15.6, T=30.0, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:17,067] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:25,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:55:26,027] loop.py:45 INFO TRAIN_STEP step=7084 loss=1.486866 Training: 39%|███▉ | 7084/18018 [15:25:12<25:54:30, 8.53s/step, loss=1.487, acc=0.669, acc_len=1.58, thru=14.3, I=14.6, T=30.8, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:26,259] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:32,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:55:32,785] loop.py:380 INFO TIMING step=7085: step=4.444s data=2.069s compute=4.410s [fwd=3.087s bwd=1.308s opt=0.003s] dispatch=2.306s [2026-10-19 04:55:32,785] loop.py:45 INFO TRAIN_STEP step=7085 loss=1.424548 Training: 39%|███▉ | 7085/18018 [15:25:18<24:17:29, 8.00s/step, loss=1.425, acc=0.681, acc_len=1.62, thru=18.1, I=20.9, T=28.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:33,250] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:39,492] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:55:40,339] loop.py:45 INFO TRAIN_STEP step=7086 loss=1.437565 Training: 39%|███▉ | 7086/18018 [15:25:26<23:53:04, 7.87s/step, loss=1.438, acc=0.679, acc_len=1.61, thru=14.9, I=17.3, T=28.9, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:55:40,335] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7085: forward=3126.0ms backward=1252.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:40,500] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:45,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:55:47,366] loop.py:45 INFO TRAIN_STEP step=7087 loss=1.418422 Training: 39%|███▉ | 7087/18018 [15:25:33<23:07:12, 7.61s/step, loss=1.418, acc=0.679, acc_len=1.62, thru=18.3, I=21.6, T=27.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:47,777] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:54,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:55:55,204] loop.py:45 INFO TRAIN_STEP step=7088 loss=1.431970 Training: 39%|███▉ | 7088/18018 [15:25:41<23:19:12, 7.68s/step, loss=1.432, acc=0.677, acc_len=1.61, thru=14.6, I=16.2, T=31.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:55:55,350] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:05,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:56:05,872] loop.py:45 INFO TRAIN_STEP step=7089 loss=1.425445 Training: 39%|███▉ | 7089/18018 [15:25:51<26:02:19, 8.58s/step, loss=1.425, acc=0.681, acc_len=1.62, thru=12.0, I=12.2, T=30.2, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:05,977] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:12,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:56:13,124] loop.py:380 INFO TIMING step=7090: step=4.834s data=2.334s compute=4.806s [fwd=3.382s bwd=1.409s opt=0.003s] dispatch=2.409s [2026-10-19 04:56:13,124] loop.py:45 INFO TRAIN_STEP step=7090 loss=1.466855 Training: 39%|███▉ | 7090/18018 [15:25:59<24:49:49, 8.18s/step, loss=1.467, acc=0.671, acc_len=1.59, thru=16.8, I=19.1, T=26.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:13,365] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:17,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:56:20,848] loop.py:45 INFO TRAIN_STEP step=7091 loss=1.487533 Training: 39%|███▉ | 7091/18018 [15:26:06<24:24:41, 8.04s/step, loss=1.488, acc=0.672, acc_len=1.59, thru=15.2, I=21.8, T=26.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:56:20,844] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7090: forward=3325.9ms backward=1555.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:21,297] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 [2026-10-19 04:56:27,715] loop.py:45 INFO TRAIN_STEP step=7092 loss=1.437312 Training: 39%|███▉ | 7092/18018 [15:26:13<23:20:17, 7.69s/step, loss=1.437, acc=0.680, acc_len=1.61, thru=11.7, I=21.3, T=33.1, wait=3.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:37,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:56:38,247] loop.py:45 INFO TRAIN_STEP step=7093 loss=1.352098 Training: 39%|███▉ | 7093/18018 [15:26:24<25:55:26, 8.54s/step, loss=1.352, acc=0.695, acc_len=1.67, thru=14.4, I=14.3, T=29.9, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:38,401] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:42,712] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:56:44,289] loop.py:45 INFO TRAIN_STEP step=7094 loss=1.434393 Training: 39%|███▉ | 7094/18018 [15:26:30<23:38:47, 7.79s/step, loss=1.434, acc=0.684, acc_len=1.63, thru=18.6, I=23.3, T=29.8, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:44,727] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:49,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:56:51,318] loop.py:380 INFO TIMING step=7095: step=4.408s data=2.093s compute=4.378s [fwd=3.014s bwd=1.349s opt=0.003s] dispatch=2.612s [2026-10-19 04:56:51,318] loop.py:45 INFO TRAIN_STEP step=7095 loss=1.486873 Training: 39%|███▉ | 7095/18018 [15:26:37<22:57:00, 7.56s/step, loss=1.487, acc=0.673, acc_len=1.59, thru=16.3, I=24.6, T=29.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:51,669] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:56,155] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:56:58,213] loop.py:45 INFO TRAIN_STEP step=7096 loss=1.424590 Training: 39%|███▉ | 7096/18018 [15:26:44<22:20:11, 7.36s/step, loss=1.425, acc=0.681, acc_len=1.62, thru=16.2, I=23.7, T=33.0, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:56:58,208] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7095: forward=2723.0ms backward=1112.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:56:58,669] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:03,525] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:57:05,203] loop.py:45 INFO TRAIN_STEP step=7097 loss=1.448798 Training: 39%|███▉ | 7097/18018 [15:26:51<21:59:54, 7.25s/step, loss=1.449, acc=0.674, acc_len=1.60, thru=16.3, I=23.9, T=28.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:05,541] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:11,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:57:12,018] loop.py:45 INFO TRAIN_STEP step=7098 loss=1.468388 Training: 39%|███▉ | 7098/18018 [15:26:58<21:35:49, 7.12s/step, loss=1.468, acc=0.674, acc_len=1.59, thru=16.3, I=20.7, T=30.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:12,317] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:21,097] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:57:21,658] loop.py:45 INFO TRAIN_STEP step=7099 loss=1.463156 Training: 39%|███▉ | 7099/18018 [15:27:07<23:53:16, 7.88s/step, loss=1.463, acc=0.675, acc_len=1.61, thru=14.0, I=13.7, T=26.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:22,102] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:28,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:57:29,790] loop.py:380 INFO TIMING step=7100: step=4.934s data=2.385s compute=4.897s [fwd=3.409s bwd=1.472s opt=0.003s] dispatch=3.191s [2026-10-19 04:57:29,791] loop.py:45 INFO TRAIN_STEP step=7100 loss=1.518587 Training: 39%|███▉ | 7100/18018 [15:27:15<24:07:17, 7.95s/step, loss=1.519, acc=0.666, acc_len=1.57, thru=15.3, I=20.1, T=25.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:30,059] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:35,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:57:37,330] loop.py:45 INFO TRAIN_STEP step=7101 loss=1.511200 Training: 39%|███▉ | 7101/18018 [15:27:23<23:44:29, 7.83s/step, loss=1.511, acc=0.664, acc_len=1.57, thru=14.2, I=17.3, T=29.3, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:57:37,325] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7100: forward=2938.4ms backward=1372.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:37,724] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:45,428] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:57:46,902] loop.py:45 INFO TRAIN_STEP step=7102 loss=1.624321 Training: 39%|███▉ | 7102/18018 [15:27:32<25:19:36, 8.35s/step, loss=1.624, acc=0.651, acc_len=1.52, thru=13.9, I=15.6, T=27.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:46,953] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:52,141] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:57:53,362] loop.py:45 INFO TRAIN_STEP step=7103 loss=1.260496 Training: 39%|███▉ | 7103/18018 [15:27:39<23:36:05, 7.78s/step, loss=1.260, acc=0.711, acc_len=1.72, thru=16.5, I=22.2, T=29.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:57:53,650] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:00,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:58:01,128] loop.py:45 INFO TRAIN_STEP step=7104 loss=1.523435 Training: 39%|███▉ | 7104/18018 [15:27:47<23:34:55, 7.78s/step, loss=1.523, acc=0.666, acc_len=1.57, thru=15.0, I=17.9, T=29.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:01,495] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:07,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:58:09,300] loop.py:380 INFO TIMING step=7105: step=4.335s data=2.141s compute=4.277s [fwd=3.028s bwd=1.234s opt=0.003s] dispatch=3.827s [2026-10-19 04:58:09,300] loop.py:45 INFO TRAIN_STEP step=7105 loss=1.461851 Training: 39%|███▉ | 7105/18018 [15:27:55<23:56:13, 7.90s/step, loss=1.462, acc=0.676, acc_len=1.60, thru=15.4, I=17.7, T=29.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:09,624] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:16,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:58:17,081] loop.py:45 INFO TRAIN_STEP step=7106 loss=1.340777 Training: 39%|███▉ | 7106/18018 [15:28:03<23:49:50, 7.86s/step, loss=1.341, acc=0.698, acc_len=1.68, thru=15.4, I=21.0, T=32.2, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:58:17,077] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7105: forward=2819.2ms backward=1113.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:17,526] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 04:58:23,458] loop.py:45 INFO TRAIN_STEP step=7107 loss=1.502028 Training: 39%|███▉ | 7107/18018 [15:28:09<22:28:36, 7.42s/step, loss=1.502, acc=0.669, acc_len=1.58, thru=16.9, I=18.6, T=34.4, wait=2.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:30,450] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:58:30,995] loop.py:45 INFO TRAIN_STEP step=7108 loss=1.381628 Training: 39%|███▉ | 7108/18018 [15:28:17<22:35:05, 7.45s/step, loss=1.382, acc=0.689, acc_len=1.64, thru=17.4, I=17.4, T=34.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:31,457] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:38,032] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:58:39,108] loop.py:45 INFO TRAIN_STEP step=7109 loss=1.483423 Training: 39%|███▉ | 7109/18018 [15:28:25<23:11:02, 7.65s/step, loss=1.483, acc=0.670, acc_len=1.58, thru=16.6, I=19.9, T=27.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:39,543] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:44,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:58:46,232] loop.py:380 INFO TIMING step=7110: step=4.278s data=2.040s compute=4.244s [fwd=2.967s bwd=1.261s opt=0.003s] dispatch=2.837s [2026-10-19 04:58:46,233] loop.py:45 INFO TRAIN_STEP step=7110 loss=1.337026 Training: 39%|███▉ | 7110/18018 [15:28:32<22:42:19, 7.49s/step, loss=1.337, acc=0.695, acc_len=1.67, thru=16.2, I=20.5, T=29.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:46,321] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:51,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:58:52,569] loop.py:45 INFO TRAIN_STEP step=7111 loss=1.576752 Training: 39%|███▉ | 7111/18018 [15:28:38<21:39:01, 7.15s/step, loss=1.577, acc=0.654, acc_len=1.53, thru=18.2, I=23.3, T=33.3, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 04:58:52,565] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7110: forward=2702.8ms backward=1091.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:52,955] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:58,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:58:58,769] loop.py:45 INFO TRAIN_STEP step=7112 loss=1.482358 Training: 39%|███▉ | 7112/18018 [15:28:44<20:47:22, 6.86s/step, loss=1.482, acc=0.670, acc_len=1.58, thru=19.7, I=24.2, T=35.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:58:58,886] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:04,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:59:06,442] loop.py:45 INFO TRAIN_STEP step=7113 loss=1.508486 Training: 39%|███▉ | 7113/18018 [15:28:52<21:31:22, 7.11s/step, loss=1.508, acc=0.664, acc_len=1.56, thru=16.6, I=20.5, T=29.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:06,792] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:14,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:59:15,836] loop.py:45 INFO TRAIN_STEP step=7114 loss=1.405419 Training: 39%|███▉ | 7114/18018 [15:29:01<23:35:58, 7.79s/step, loss=1.405, acc=0.689, acc_len=1.65, thru=14.2, I=16.7, T=29.9, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:16,335] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:21,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:59:23,178] loop.py:380 INFO TIMING step=7115: step=4.320s data=2.062s compute=4.292s [fwd=3.100s bwd=1.176s opt=0.003s] dispatch=3.014s [2026-10-19 04:59:23,178] loop.py:45 INFO TRAIN_STEP step=7115 loss=1.371435 Training: 39%|███▉ | 7115/18018 [15:29:09<23:11:25, 7.66s/step, loss=1.371, acc=0.689, acc_len=1.65, thru=15.8, I=23.2, T=29.6, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:23,307] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:29,293] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:59:29,916] loop.py:45 INFO TRAIN_STEP step=7116 loss=1.490256 (TrainerActor pid=9310) [2026-10-19 04:59:29,911] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7115: forward=2634.4ms backward=1033.6ms Training: 39%|███▉ | 7116/18018 [15:29:16<22:21:16, 7.38s/step, loss=1.490, acc=0.671, acc_len=1.58, thru=17.2, I=22.0, T=34.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:30,301] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 04:59:37,573] loop.py:45 INFO TRAIN_STEP step=7117 loss=1.549407 Training: 39%|███▉ | 7117/18018 [15:29:23<22:36:13, 7.46s/step, loss=1.549, acc=0.661, acc_len=1.56, thru=16.1, I=18.0, T=29.4, wait=3.3s, pool=60, epoch=1/2][2026-10-19 04:59:47,639] loop.py:45 INFO TRAIN_STEP step=7118 loss=1.623486 Training: 40%|███▉ | 7118/18018 [15:29:33<24:57:41, 8.24s/step, loss=1.623, acc=0.649, acc_len=1.51, thru=13.2, I=12.5, T=26.3, wait=5.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:56,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 04:59:57,500] loop.py:45 INFO TRAIN_STEP step=7119 loss=1.466428 Training: 40%|███▉ | 7119/18018 [15:29:43<26:25:43, 8.73s/step, loss=1.466, acc=0.672, acc_len=1.58, thru=14.1, I=13.7, T=29.6, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 04:59:57,757] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:02,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:00:04,355] loop.py:380 INFO TIMING step=7120: step=4.395s data=2.111s compute=4.363s [fwd=3.066s bwd=1.281s opt=0.003s] dispatch=2.449s [2026-10-19 05:00:04,355] loop.py:45 INFO TRAIN_STEP step=7120 loss=1.572393 Training: 40%|███▉ | 7120/18018 [15:29:50<24:43:22, 8.17s/step, loss=1.572, acc=0.654, acc_len=1.54, thru=17.7, I=21.7, T=29.1, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:04,521] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:09,528] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:00:11,098] loop.py:45 INFO TRAIN_STEP step=7121 loss=1.321581 Training: 40%|███▉ | 7121/18018 [15:29:57<23:25:38, 7.74s/step, loss=1.322, acc=0.697, acc_len=1.68, thru=16.1, I=24.0, T=32.4, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:00:11,093] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7120: forward=2791.9ms backward=1120.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:11,533] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:18,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:00:20,159] loop.py:45 INFO TRAIN_STEP step=7122 loss=1.560101 Training: 40%|███▉ | 7122/18018 [15:30:06<24:37:31, 8.14s/step, loss=1.560, acc=0.659, acc_len=1.55, thru=14.8, I=17.6, T=30.0, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:20,370] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:00:29,910] loop.py:45 INFO TRAIN_STEP step=7123 loss=1.514839 Training: 40%|███▉ | 7123/18018 [15:30:15<26:05:27, 8.62s/step, loss=1.515, acc=0.669, acc_len=1.57, thru=11.3, I=11.7, T=25.7, wait=4.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:38,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:00:39,343] loop.py:45 INFO TRAIN_STEP step=7124 loss=1.486388 Training: 40%|███▉ | 7124/18018 [15:30:25<26:49:31, 8.86s/step, loss=1.486, acc=0.670, acc_len=1.59, thru=15.8, I=16.3, T=25.7, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:39,637] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:46,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:00:47,774] loop.py:380 INFO TIMING step=7125: step=4.371s data=2.047s compute=4.344s [fwd=3.140s bwd=1.187s opt=0.003s] dispatch=4.049s [2026-10-19 05:00:47,774] loop.py:45 INFO TRAIN_STEP step=7125 loss=1.511797 Training: 40%|███▉ | 7125/18018 [15:30:33<26:25:39, 8.73s/step, loss=1.512, acc=0.663, acc_len=1.56, thru=14.5, I=16.3, T=29.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:47,838] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:55,071] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:00:56,189] loop.py:45 INFO TRAIN_STEP step=7126 loss=1.388742 (TrainerActor pid=9310) [2026-10-19 05:00:56,185] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7125: forward=2850.0ms backward=1123.7ms Training: 40%|███▉ | 7126/18018 [15:30:42<26:08:13, 8.64s/step, loss=1.389, acc=0.687, acc_len=1.63, thru=14.1, I=14.9, T=31.8, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:00:56,580] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:03,669] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:01:04,573] loop.py:45 INFO TRAIN_STEP step=7127 loss=1.402884 Training: 40%|███▉ | 7127/18018 [15:30:50<25:54:14, 8.56s/step, loss=1.403, acc=0.687, acc_len=1.65, thru=14.2, I=17.1, T=33.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:04,679] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:09,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:01:11,252] loop.py:45 INFO TRAIN_STEP step=7128 loss=1.403597 Training: 40%|███▉ | 7128/18018 [15:30:57<24:11:27, 8.00s/step, loss=1.404, acc=0.684, acc_len=1.63, thru=18.5, I=23.3, T=31.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:11,569] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:01:17,710] loop.py:45 INFO TRAIN_STEP step=7129 loss=1.442490 Training: 40%|███▉ | 7129/18018 [15:31:03<22:47:36, 7.54s/step, loss=1.442, acc=0.674, acc_len=1.59, thru=16.9, I=20.7, T=35.9, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:24,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:01:25,902] loop.py:380 INFO TIMING step=7130: step=3.820s data=1.593s compute=3.775s [fwd=2.676s bwd=1.083s opt=0.003s] dispatch=4.362s [2026-10-19 05:01:25,902] loop.py:45 INFO TRAIN_STEP step=7130 loss=1.523946 Training: 40%|███▉ | 7130/18018 [15:31:11<23:23:11, 7.73s/step, loss=1.524, acc=0.664, acc_len=1.56, thru=17.3, I=18.5, T=33.5, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:26,238] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:32,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:01:33,184] loop.py:45 INFO TRAIN_STEP step=7131 loss=1.436710 Training: 40%|███▉ | 7131/18018 [15:31:19<22:58:27, 7.60s/step, loss=1.437, acc=0.679, acc_len=1.61, thru=16.2, I=20.1, T=29.4, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:01:33,179] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7130: forward=2994.1ms backward=1320.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:33,267] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:39,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:01:40,377] loop.py:45 INFO TRAIN_STEP step=7132 loss=1.420100 Training: 40%|███▉ | 7132/18018 [15:31:26<22:36:19, 7.48s/step, loss=1.420, acc=0.680, acc_len=1.62, thru=14.9, I=17.8, T=29.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:40,771] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:47,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:01:49,121] loop.py:45 INFO TRAIN_STEP step=7133 loss=1.418286 Training: 40%|███▉ | 7133/18018 [15:31:35<23:45:13, 7.86s/step, loss=1.418, acc=0.679, acc_len=1.61, thru=13.8, I=19.1, T=27.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:49,593] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:55,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:01:56,650] loop.py:45 INFO TRAIN_STEP step=7134 loss=1.460265 Training: 40%|███▉ | 7134/18018 [15:31:42<23:27:24, 7.76s/step, loss=1.460, acc=0.673, acc_len=1.59, thru=14.8, I=20.0, T=32.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:01:56,728] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:02,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:02:03,383] loop.py:380 INFO TIMING step=7135: step=3.802s data=1.506s compute=3.771s [fwd=2.755s bwd=1.002s opt=0.003s] dispatch=2.920s [2026-10-19 05:02:03,383] loop.py:45 INFO TRAIN_STEP step=7135 loss=1.443764 Training: 40%|███▉ | 7135/18018 [15:31:49<22:31:30, 7.45s/step, loss=1.444, acc=0.680, acc_len=1.62, thru=18.0, I=22.9, T=33.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:03,424] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:09,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:02:09,480] loop.py:45 INFO TRAIN_STEP step=7136 loss=1.481791 Training: 40%|███▉ | 7136/18018 [15:31:55<21:17:40, 7.04s/step, loss=1.482, acc=0.670, acc_len=1.58, thru=20.0, I=22.7, T=35.3, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:02:09,476] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7135: forward=2579.8ms backward=1003.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:09,549] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:15,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:02:17,436] loop.py:45 INFO TRAIN_STEP step=7137 loss=1.498084 Training: 40%|███▉ | 7137/18018 [15:32:03<22:07:07, 7.32s/step, loss=1.498, acc=0.666, acc_len=1.57, thru=15.8, I=19.1, T=27.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:17,582] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:22,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:02:23,410] loop.py:45 INFO TRAIN_STEP step=7138 loss=1.264865 Training: 40%|███▉ | 7138/18018 [15:32:09<20:53:54, 6.91s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=17.8, I=22.6, T=33.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:23,687] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:30,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:02:31,093] loop.py:45 INFO TRAIN_STEP step=7139 loss=1.418373 Training: 40%|███▉ | 7139/18018 [15:32:17<21:35:23, 7.14s/step, loss=1.418, acc=0.678, acc_len=1.61, thru=17.3, I=19.6, T=32.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:31,290] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:37,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:02:37,402] loop.py:380 INFO TIMING step=7140: step=4.133s data=1.872s compute=4.107s [fwd=2.898s bwd=1.197s opt=0.003s] dispatch=2.168s [2026-10-19 05:02:37,402] loop.py:45 INFO TRAIN_STEP step=7140 loss=1.337230 Training: 40%|███▉ | 7140/18018 [15:32:23<20:49:57, 6.89s/step, loss=1.337, acc=0.694, acc_len=1.66, thru=19.2, I=18.9, T=31.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:37,772] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:44,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:02:45,230] loop.py:45 INFO TRAIN_STEP step=7141 loss=1.410902 Training: 40%|███▉ | 7141/18018 [15:32:31<21:40:38, 7.17s/step, loss=1.411, acc=0.683, acc_len=1.63, thru=16.1, I=16.2, T=30.7, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:02:45,226] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7140: forward=2914.8ms backward=1206.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:45,499] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:51,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:02:53,638] loop.py:45 INFO TRAIN_STEP step=7142 loss=1.576981 Training: 40%|███▉ | 7142/18018 [15:32:39<22:47:37, 7.54s/step, loss=1.577, acc=0.653, acc_len=1.53, thru=15.8, I=19.4, T=30.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:02:53,839] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:00,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:03:00,917] loop.py:45 INFO TRAIN_STEP step=7143 loss=1.458001 Training: 40%|███▉ | 7143/18018 [15:32:47<22:32:53, 7.46s/step, loss=1.458, acc=0.673, acc_len=1.59, thru=15.8, I=19.6, T=25.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:01,337] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:08,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:03:09,419] loop.py:45 INFO TRAIN_STEP step=7144 loss=1.466458 Training: 40%|███▉ | 7144/18018 [15:32:55<23:29:15, 7.78s/step, loss=1.466, acc=0.670, acc_len=1.58, thru=15.4, I=16.7, T=28.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:09,483] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:15,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:03:17,237] loop.py:380 INFO TIMING step=7145: step=4.276s data=1.931s compute=4.235s [fwd=2.963s bwd=1.258s opt=0.003s] dispatch=3.533s [2026-10-19 05:03:17,237] loop.py:45 INFO TRAIN_STEP step=7145 loss=1.464898 Training: 40%|███▉ | 7145/18018 [15:33:03<23:31:22, 7.79s/step, loss=1.465, acc=0.673, acc_len=1.59, thru=15.6, I=21.5, T=29.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:17,549] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 05:03:24,215] loop.py:45 INFO TRAIN_STEP step=7146 loss=1.448567 Training: 40%|███▉ | 7146/18018 [15:33:10<22:47:19, 7.55s/step, loss=1.449, acc=0.671, acc_len=1.59, thru=12.5, I=22.4, T=33.5, wait=3.2s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:03:24,211] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7145: forward=2700.9ms backward=1071.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:32,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:03:34,572] loop.py:45 INFO TRAIN_STEP step=7147 loss=1.427986 Training: 40%|███▉ | 7147/18018 [15:33:20<25:19:59, 8.39s/step, loss=1.428, acc=0.678, acc_len=1.62, thru=15.9, I=16.8, T=27.2, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:34,677] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:41,002] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:03:41,285] loop.py:45 INFO TRAIN_STEP step=7148 loss=1.531903 Training: 40%|███▉ | 7148/18018 [15:33:27<23:48:37, 7.89s/step, loss=1.532, acc=0.663, acc_len=1.57, thru=16.2, I=20.8, T=32.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:41,505] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:46,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:03:47,648] loop.py:45 INFO TRAIN_STEP step=7149 loss=1.440629 Training: 40%|███▉ | 7149/18018 [15:33:33<22:25:50, 7.43s/step, loss=1.441, acc=0.675, acc_len=1.60, thru=18.0, I=21.9, T=31.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:47,775] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:54,907] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:03:55,791] loop.py:380 INFO TIMING step=7150: step=4.049s data=1.810s compute=4.001s [fwd=2.700s bwd=1.287s opt=0.003s] dispatch=4.084s [2026-10-19 05:03:55,791] loop.py:45 INFO TRAIN_STEP step=7150 loss=1.460105 Training: 40%|███▉ | 7150/18018 [15:33:41<23:04:32, 7.64s/step, loss=1.460, acc=0.672, acc_len=1.59, thru=13.8, I=17.4, T=31.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:03:55,921] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:00,749] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:04:02,109] loop.py:45 INFO TRAIN_STEP step=7151 loss=1.483441 Training: 40%|███▉ | 7151/18018 [15:33:48<21:52:20, 7.25s/step, loss=1.483, acc=0.674, acc_len=1.60, thru=18.5, I=23.2, T=33.5, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:04:02,104] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7150: forward=2654.6ms backward=1124.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:02,269] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:08,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:04:08,889] loop.py:45 INFO TRAIN_STEP step=7152 loss=1.488228 Training: 40%|███▉ | 7152/18018 [15:33:54<21:26:55, 7.11s/step, loss=1.488, acc=0.672, acc_len=1.59, thru=17.8, I=22.5, T=31.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:09,036] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:14,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:04:16,732] loop.py:45 INFO TRAIN_STEP step=7153 loss=1.485926 Training: 40%|███▉ | 7153/18018 [15:34:02<22:06:43, 7.33s/step, loss=1.486, acc=0.666, acc_len=1.57, thru=15.9, I=21.9, T=27.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:16,910] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:21,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:04:23,855] loop.py:45 INFO TRAIN_STEP step=7154 loss=1.284015 Training: 40%|███▉ | 7154/18018 [15:34:09<21:55:38, 7.27s/step, loss=1.284, acc=0.703, acc_len=1.70, thru=15.2, I=27.2, T=25.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:24,002] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:29,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:04:30,919] loop.py:380 INFO TIMING step=7155: step=4.681s data=2.276s compute=4.648s [fwd=3.256s bwd=1.379s opt=0.003s] dispatch=2.373s [2026-10-19 05:04:30,919] loop.py:45 INFO TRAIN_STEP step=7155 loss=1.487391 Training: 40%|███▉ | 7155/18018 [15:34:17<21:44:32, 7.21s/step, loss=1.487, acc=0.668, acc_len=1.58, thru=14.3, I=22.4, T=27.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:31,245] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:04:37,589] loop.py:45 INFO TRAIN_STEP step=7156 loss=1.586475 Training: 40%|███▉ | 7156/18018 [15:34:23<21:15:17, 7.04s/step, loss=1.586, acc=0.654, acc_len=1.53, thru=15.1, I=22.1, T=32.5, wait=2.7s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:04:37,584] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7155: forward=2762.8ms backward=1130.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:46,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:04:46,442] loop.py:45 INFO TRAIN_STEP step=7157 loss=1.471262 Training: 40%|███▉ | 7157/18018 [15:34:32<22:53:26, 7.59s/step, loss=1.471, acc=0.678, acc_len=1.62, thru=15.9, I=15.9, T=30.9, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:46,570] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:51,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:04:53,710] loop.py:45 INFO TRAIN_STEP step=7158 loss=1.546015 Training: 40%|███▉ | 7158/18018 [15:34:39<22:35:52, 7.49s/step, loss=1.546, acc=0.660, acc_len=1.56, thru=17.8, I=21.7, T=26.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:53,975] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:04:59,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:05:00,977] loop.py:45 INFO TRAIN_STEP step=7159 loss=1.468243 Training: 40%|███▉ | 7159/18018 [15:34:47<22:23:34, 7.42s/step, loss=1.468, acc=0.673, acc_len=1.60, thru=15.3, I=22.3, T=33.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:01,465] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:08,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:05:09,207] loop.py:380 INFO TIMING step=7160: step=5.218s data=2.051s compute=5.189s [fwd=3.873s bwd=1.301s opt=0.003s] dispatch=3.004s [2026-10-19 05:05:09,207] loop.py:45 INFO TRAIN_STEP step=7160 loss=1.329838 Training: 40%|███▉ | 7160/18018 [15:34:55<23:07:14, 7.67s/step, loss=1.330, acc=0.695, acc_len=1.67, thru=14.1, I=19.3, T=24.5, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:09,583] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:18,241] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:05:19,494] loop.py:45 INFO TRAIN_STEP step=7161 loss=1.550420 Training: 40%|███▉ | 7161/18018 [15:35:05<25:29:37, 8.45s/step, loss=1.550, acc=0.660, acc_len=1.55, thru=12.9, I=13.3, T=24.1, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:05:19,490] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7160: forward=3745.8ms backward=1521.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:19,749] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:27,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:05:28,309] loop.py:45 INFO TRAIN_STEP step=7162 loss=1.509061 Training: 40%|███▉ | 7162/18018 [15:35:14<25:48:53, 8.56s/step, loss=1.509, acc=0.670, acc_len=1.58, thru=15.0, I=15.2, T=31.8, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:28,528] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:35,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:05:36,350] loop.py:45 INFO TRAIN_STEP step=7163 loss=1.472623 Training: 40%|███▉ | 7163/18018 [15:35:22<25:20:40, 8.41s/step, loss=1.473, acc=0.671, acc_len=1.59, thru=14.7, I=19.1, T=31.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:36,575] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:41,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:05:43,489] loop.py:45 INFO TRAIN_STEP step=7164 loss=1.351973 Training: 40%|███▉ | 7164/18018 [15:35:29<24:11:47, 8.03s/step, loss=1.352, acc=0.691, acc_len=1.66, thru=17.6, I=24.5, T=27.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:43,646] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:52,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:05:53,389] loop.py:380 INFO TIMING step=7165: step=4.383s data=1.947s compute=4.354s [fwd=2.986s bwd=1.353s opt=0.003s] dispatch=5.507s [2026-10-19 05:05:53,389] loop.py:45 INFO TRAIN_STEP step=7165 loss=1.329441 Training: 40%|███▉ | 7165/18018 [15:35:39<25:53:14, 8.59s/step, loss=1.329, acc=0.697, acc_len=1.68, thru=13.1, I=13.6, T=29.2, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:53,869] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:05:59,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:06:00,689] loop.py:45 INFO TRAIN_STEP step=7166 loss=1.481131 Training: 40%|███▉ | 7166/18018 [15:35:46<24:43:19, 8.20s/step, loss=1.481, acc=0.667, acc_len=1.57, thru=16.9, I=20.5, T=29.9, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:06:00,685] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7165: forward=3133.8ms backward=1094.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:01,113] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:05,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:06:07,402] loop.py:45 INFO TRAIN_STEP step=7167 loss=1.343490 Training: 40%|███▉ | 7167/18018 [15:35:53<23:22:31, 7.76s/step, loss=1.343, acc=0.695, acc_len=1.67, thru=16.6, I=22.8, T=27.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:07,577] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 05:06:16,233] loop.py:45 INFO TRAIN_STEP step=7168 loss=1.479041 Training: 40%|███▉ | 7168/18018 [15:36:02<24:20:46, 8.08s/step, loss=1.479, acc=0.673, acc_len=1.59, thru=12.5, I=12.4, T=31.5, wait=4.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:22,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:06:24,039] loop.py:45 INFO TRAIN_STEP step=7169 loss=1.577856 Training: 40%|███▉ | 7169/18018 [15:36:10<24:05:50, 8.00s/step, loss=1.578, acc=0.658, acc_len=1.54, thru=17.4, I=18.9, T=30.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:24,273] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:30,047] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:06:30,541] loop.py:380 INFO TIMING step=7170: step=3.903s data=1.721s compute=3.872s [fwd=2.685s bwd=1.173s opt=0.003s] dispatch=2.589s [2026-10-19 05:06:30,541] loop.py:45 INFO TRAIN_STEP step=7170 loss=1.470890 Training: 40%|███▉ | 7170/18018 [15:36:16<22:44:42, 7.55s/step, loss=1.471, acc=0.672, acc_len=1.59, thru=17.3, I=22.1, T=32.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:30,578] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:37,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:06:37,868] loop.py:45 INFO TRAIN_STEP step=7171 loss=1.304557 Training: 40%|███▉ | 7171/18018 [15:36:23<22:32:34, 7.48s/step, loss=1.305, acc=0.700, acc_len=1.69, thru=17.1, I=18.6, T=28.8, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:06:37,864] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7170: forward=3136.6ms backward=1267.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:38,206] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:46,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:06:46,599] loop.py:45 INFO TRAIN_STEP step=7172 loss=1.461016 Training: 40%|███▉ | 7172/18018 [15:36:32<23:40:10, 7.86s/step, loss=1.461, acc=0.678, acc_len=1.61, thru=14.3, I=14.5, T=27.6, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:46,803] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:50,659] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:06:55,594] loop.py:45 INFO TRAIN_STEP step=7173 loss=1.490739 Training: 40%|███▉ | 7173/18018 [15:36:41<24:41:48, 8.20s/step, loss=1.491, acc=0.667, acc_len=1.58, thru=14.3, I=27.8, T=19.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:55,700] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:06:55,700] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=24/64 [2026-10-19 05:07:02,365] loop.py:45 INFO TRAIN_STEP step=7174 loss=1.330400 Training: 40%|███▉ | 7174/18018 [15:36:48<23:24:12, 7.77s/step, loss=1.330, acc=0.694, acc_len=1.66, thru=16.9, I=18.4, T=35.6, wait=3.2s, pool=48, epoch=1/2][2026-10-19 05:07:10,488] loop.py:380 INFO TIMING step=7175: step=4.239s data=1.696s compute=4.211s [fwd=2.856s bwd=1.341s opt=0.003s] dispatch=3.877s [2026-10-19 05:07:10,488] loop.py:45 INFO TRAIN_STEP step=7175 loss=1.523446 Training: 40%|███▉ | 7175/18018 [15:36:56<23:43:19, 7.88s/step, loss=1.523, acc=0.662, acc_len=1.56, thru=16.3, I=16.6, T=30.2, wait=3.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:18,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 05:07:19,220] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7175: forward=2858.2ms backward=1119.0ms [2026-10-19 05:07:19,224] loop.py:45 INFO TRAIN_STEP step=7176 loss=1.493371 Training: 40%|███▉ | 7176/18018 [15:37:05<24:29:42, 8.13s/step, loss=1.493, acc=0.672, acc_len=1.59, thru=14.9, I=14.9, T=31.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:19,498] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:25,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:07:26,632] loop.py:45 INFO TRAIN_STEP step=7177 loss=1.504270 Training: 40%|███▉ | 7177/18018 [15:37:12<23:50:22, 7.92s/step, loss=1.504, acc=0.664, acc_len=1.57, thru=17.0, I=20.6, T=33.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:26,752] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:31,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:07:33,388] loop.py:45 INFO TRAIN_STEP step=7178 loss=1.526768 Training: 40%|███▉ | 7178/18018 [15:37:19<22:47:19, 7.57s/step, loss=1.527, acc=0.661, acc_len=1.56, thru=17.3, I=24.4, T=28.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:33,872] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:38,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:07:41,188] loop.py:45 INFO TRAIN_STEP step=7179 loss=1.515614 Training: 40%|███▉ | 7179/18018 [15:37:27<22:59:47, 7.64s/step, loss=1.516, acc=0.664, acc_len=1.56, thru=13.7, I=23.0, T=25.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:41,479] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:49,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:07:50,793] loop.py:380 INFO TIMING step=7180: step=4.209s data=1.945s compute=4.175s [fwd=2.942s bwd=1.218s opt=0.003s] dispatch=5.385s [2026-10-19 05:07:50,793] loop.py:45 INFO TRAIN_STEP step=7180 loss=1.460819 Training: 40%|███▉ | 7180/18018 [15:37:36<24:46:14, 8.23s/step, loss=1.461, acc=0.676, acc_len=1.60, thru=13.8, I=15.5, T=30.4, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:07:50,962] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:07:57,133] loop.py:45 INFO TRAIN_STEP step=7181 loss=1.379668 Training: 40%|███▉ | 7181/18018 [15:37:43<23:03:39, 7.66s/step, loss=1.380, acc=0.691, acc_len=1.65, thru=17.1, I=19.2, T=31.5, wait=2.3s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:07:57,130] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7180: forward=2843.1ms backward=1179.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:04,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:08:05,672] loop.py:45 INFO TRAIN_STEP step=7182 loss=1.375229 Training: 40%|███▉ | 7182/18018 [15:37:51<23:51:11, 7.92s/step, loss=1.375, acc=0.690, acc_len=1.65, thru=16.7, I=17.3, T=28.5, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:05,987] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:10,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:08:11,670] loop.py:45 INFO TRAIN_STEP step=7183 loss=1.494730 Training: 40%|███▉ | 7183/18018 [15:37:57<22:06:44, 7.35s/step, loss=1.495, acc=0.671, acc_len=1.59, thru=18.5, I=22.0, T=35.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:11,866] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:16,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:08:18,296] loop.py:45 INFO TRAIN_STEP step=7184 loss=1.450393 Training: 40%|███▉ | 7184/18018 [15:38:04<21:27:34, 7.13s/step, loss=1.450, acc=0.676, acc_len=1.60, thru=19.3, I=24.5, T=29.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:18,469] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:08:26,510] loop.py:380 INFO TIMING step=7185: step=4.474s data=2.081s compute=4.445s [fwd=3.062s bwd=1.366s opt=0.003s] dispatch=3.730s [2026-10-19 05:08:26,511] loop.py:45 INFO TRAIN_STEP step=7185 loss=1.502545 Training: 40%|███▉ | 7185/18018 [15:38:12<22:26:14, 7.46s/step, loss=1.503, acc=0.667, acc_len=1.58, thru=11.4, I=13.8, T=28.6, wait=3.7s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:34,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:08:36,661] loop.py:45 INFO TRAIN_STEP step=7186 loss=1.383099 Training: 40%|███▉ | 7186/18018 [15:38:22<24:51:50, 8.26s/step, loss=1.383, acc=0.687, acc_len=1.64, thru=15.6, I=19.1, T=30.2, wait=5.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:08:36,656] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7185: forward=2897.6ms backward=1285.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:36,688] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:42,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:08:44,233] loop.py:45 INFO TRAIN_STEP step=7187 loss=1.496409 Training: 40%|███▉ | 7187/18018 [15:38:30<24:14:17, 8.06s/step, loss=1.496, acc=0.671, acc_len=1.58, thru=14.4, I=22.2, T=31.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:44,632] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:51,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:08:51,069] loop.py:45 INFO TRAIN_STEP step=7188 loss=1.487687 Training: 40%|███▉ | 7188/18018 [15:38:37<23:08:00, 7.69s/step, loss=1.488, acc=0.664, acc_len=1.57, thru=17.7, I=21.5, T=30.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:08:51,548] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:00,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:09:00,515] loop.py:45 INFO TRAIN_STEP step=7189 loss=1.508545 Training: 40%|███▉ | 7189/18018 [15:38:46<24:42:57, 8.22s/step, loss=1.509, acc=0.663, acc_len=1.56, thru=14.4, I=14.3, T=30.5, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:00,916] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:08,631] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:09:10,162] loop.py:380 INFO TIMING step=7190: step=4.266s data=1.997s compute=4.218s [fwd=2.974s bwd=1.230s opt=0.003s] dispatch=5.374s [2026-10-19 05:09:10,162] loop.py:45 INFO TRAIN_STEP step=7190 loss=1.650210 Training: 40%|███▉ | 7190/18018 [15:38:56<26:00:18, 8.65s/step, loss=1.650, acc=0.644, acc_len=1.50, thru=13.5, I=17.0, T=30.0, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:10,644] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:16,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:09:17,110] loop.py:45 INFO TRAIN_STEP step=7191 loss=1.389444 Training: 40%|███▉ | 7191/18018 [15:39:03<24:28:16, 8.14s/step, loss=1.389, acc=0.688, acc_len=1.63, thru=17.1, I=21.8, T=33.2, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:09:17,105] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7190: forward=2700.1ms backward=1110.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:17,244] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:22,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:09:24,386] loop.py:45 INFO TRAIN_STEP step=7192 loss=1.407955 Training: 40%|███▉ | 7192/18018 [15:39:10<23:41:38, 7.88s/step, loss=1.408, acc=0.681, acc_len=1.62, thru=18.5, I=22.2, T=26.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:24,588] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 05:09:31,251] loop.py:45 INFO TRAIN_STEP step=7193 loss=1.318575 Training: 40%|███▉ | 7193/18018 [15:39:17<22:46:34, 7.57s/step, loss=1.319, acc=0.700, acc_len=1.69, thru=15.2, I=21.9, T=29.0, wait=2.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:37,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:09:38,161] loop.py:45 INFO TRAIN_STEP step=7194 loss=1.293624 Training: 40%|███▉ | 7194/18018 [15:39:24<22:10:21, 7.37s/step, loss=1.294, acc=0.705, acc_len=1.71, thru=17.8, I=16.7, T=30.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:38,544] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:42,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:09:44,305] loop.py:380 INFO TIMING step=7195: step=3.970s data=1.889s compute=3.940s [fwd=2.799s bwd=1.126s opt=0.003s] dispatch=2.167s [2026-10-19 05:09:44,306] loop.py:45 INFO TRAIN_STEP step=7195 loss=1.259759 Training: 40%|███▉ | 7195/18018 [15:39:30<21:03:42, 7.01s/step, loss=1.260, acc=0.710, acc_len=1.72, thru=18.9, I=25.1, T=32.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:44,805] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:49,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 05:09:51,423] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7195: forward=2811.5ms backward=1164.6ms [2026-10-19 05:09:51,429] loop.py:45 INFO TRAIN_STEP step=7196 loss=1.459804 Training: 40%|███▉ | 7196/18018 [15:39:37<21:09:58, 7.04s/step, loss=1.460, acc=0.673, acc_len=1.60, thru=17.0, I=28.0, T=31.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:51,554] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:56,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:09:58,082] loop.py:45 INFO TRAIN_STEP step=7197 loss=1.489177 Training: 40%|███▉ | 7197/18018 [15:39:44<20:48:53, 6.92s/step, loss=1.489, acc=0.670, acc_len=1.58, thru=16.9, I=26.7, T=32.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:09:58,403] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:05,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:10:05,372] loop.py:45 INFO TRAIN_STEP step=7198 loss=1.457912 Training: 40%|███▉ | 7198/18018 [15:39:51<21:08:28, 7.03s/step, loss=1.458, acc=0.676, acc_len=1.60, thru=15.6, I=19.5, T=29.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:05,793] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:12,674] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:10:14,090] loop.py:45 INFO TRAIN_STEP step=7199 loss=1.411602 Training: 40%|███▉ | 7199/18018 [15:40:00<22:39:36, 7.54s/step, loss=1.412, acc=0.686, acc_len=1.63, thru=15.5, I=16.6, T=26.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:14,193] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:21,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:10:22,182] loop.py:380 INFO TIMING step=7200: step=4.697s data=2.186s compute=4.669s [fwd=3.232s bwd=1.422s opt=0.003s] dispatch=3.386s [2026-10-19 05:10:22,183] loop.py:45 INFO TRAIN_STEP step=7200 loss=1.492937 Training: 40%|███▉ | 7200/18018 [15:40:08<23:09:20, 7.71s/step, loss=1.493, acc=0.674, acc_len=1.60, thru=15.2, I=18.2, T=27.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:22,235] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 05:10:30,256] loop.py:45 INFO TRAIN_STEP step=7201 loss=1.372223 Training: 40%|███▉ | 7201/18018 [15:40:16<23:28:59, 7.82s/step, loss=1.372, acc=0.689, acc_len=1.65, thru=15.1, I=16.0, T=29.6, wait=3.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:10:30,252] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7200: forward=3064.7ms backward=1204.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:36,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:10:38,205] loop.py:45 INFO TRAIN_STEP step=7202 loss=1.486912 Training: 40%|███▉ | 7202/18018 [15:40:24<23:36:10, 7.86s/step, loss=1.487, acc=0.670, acc_len=1.59, thru=17.3, I=19.3, T=29.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:38,644] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 05:10:45,183] loop.py:45 INFO TRAIN_STEP step=7203 loss=1.413333 Training: 40%|███▉ | 7203/18018 [15:40:31<22:48:33, 7.59s/step, loss=1.413, acc=0.683, acc_len=1.63, thru=15.7, I=20.6, T=29.7, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:51,999] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:10:53,407] loop.py:45 INFO TRAIN_STEP step=7204 loss=1.447860 Training: 40%|███▉ | 7204/18018 [15:40:39<23:22:34, 7.78s/step, loss=1.448, acc=0.677, acc_len=1.61, thru=17.1, I=18.7, T=27.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:10:53,519] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-19 05:10:59,973] loop.py:380 INFO TIMING step=7205: step=4.011s data=1.767s compute=3.987s [fwd=2.840s bwd=1.132s opt=0.003s] dispatch=2.545s [2026-10-19 05:10:59,973] loop.py:45 INFO TRAIN_STEP step=7205 loss=1.485046 Training: 40%|███▉ | 7205/18018 [15:40:46<22:16:45, 7.42s/step, loss=1.485, acc=0.671, acc_len=1.59, thru=16.2, I=18.4, T=31.9, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:04,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:11:06,482] loop.py:45 INFO TRAIN_STEP step=7206 loss=1.407322 Training: 40%|███▉ | 7206/18018 [15:40:52<21:27:28, 7.14s/step, loss=1.407, acc=0.680, acc_len=1.62, thru=19.3, I=22.4, T=33.3, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:11:06,477] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7205: forward=2747.8ms backward=1050.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:06,648] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:12,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:11:14,037] loop.py:45 INFO TRAIN_STEP step=7207 loss=1.380921 Training: 40%|███▉ | 7207/18018 [15:41:00<21:49:33, 7.27s/step, loss=1.381, acc=0.687, acc_len=1.64, thru=14.5, I=21.2, T=30.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:14,258] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:20,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:11:21,238] loop.py:45 INFO TRAIN_STEP step=7208 loss=1.478975 Training: 40%|████ | 7208/18018 [15:41:07<21:45:42, 7.25s/step, loss=1.479, acc=0.670, acc_len=1.59, thru=16.9, I=20.3, T=32.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:21,578] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:27,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:11:29,178] loop.py:45 INFO TRAIN_STEP step=7209 loss=1.528980 Training: 40%|████ | 7209/18018 [15:41:15<22:23:07, 7.46s/step, loss=1.529, acc=0.661, acc_len=1.56, thru=15.6, I=19.6, T=31.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:29,306] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:34,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:11:36,134] loop.py:380 INFO TIMING step=7210: step=3.958s data=1.668s compute=3.923s [fwd=2.641s bwd=1.267s opt=0.003s] dispatch=2.988s [2026-10-19 05:11:36,134] loop.py:45 INFO TRAIN_STEP step=7210 loss=1.447034 Training: 40%|████ | 7210/18018 [15:41:22<21:55:59, 7.31s/step, loss=1.447, acc=0.677, acc_len=1.61, thru=17.7, I=24.1, T=32.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:36,288] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:42,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:11:44,653] loop.py:45 INFO TRAIN_STEP step=7211 loss=1.518842 Training: 40%|████ | 7211/18018 [15:41:30<23:01:21, 7.67s/step, loss=1.519, acc=0.666, acc_len=1.57, thru=13.3, I=19.6, T=29.9, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:11:44,648] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7210: forward=3010.7ms backward=1221.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:44,681] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:49,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:11:50,263] loop.py:45 INFO TRAIN_STEP step=7212 loss=1.518216 Training: 40%|████ | 7212/18018 [15:41:36<21:09:55, 7.05s/step, loss=1.518, acc=0.663, acc_len=1.56, thru=19.8, I=24.7, T=34.2, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:11:50,713] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:11:56,336] loop.py:45 INFO TRAIN_STEP step=7213 loss=1.458220 Training: 40%|████ | 7213/18018 [15:41:42<20:16:58, 6.76s/step, loss=1.458, acc=0.674, acc_len=1.60, thru=18.7, I=22.3, T=36.3, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:02,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:12:03,723] loop.py:45 INFO TRAIN_STEP step=7214 loss=1.422587 Training: 40%|████ | 7214/18018 [15:41:49<20:51:01, 6.95s/step, loss=1.423, acc=0.682, acc_len=1.62, thru=18.0, I=18.3, T=29.7, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:03,941] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:08,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:12:10,081] loop.py:380 INFO TIMING step=7215: step=4.412s data=2.033s compute=4.377s [fwd=3.029s bwd=1.333s opt=0.003s] dispatch=1.935s [2026-10-19 05:12:10,082] loop.py:45 INFO TRAIN_STEP step=7215 loss=1.277119 Training: 40%|████ | 7215/18018 [15:41:56<20:18:57, 6.77s/step, loss=1.277, acc=0.706, acc_len=1.70, thru=18.0, I=22.2, T=29.0, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:10,286] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:17,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:12:18,138] loop.py:45 INFO TRAIN_STEP step=7216 loss=1.475486 Training: 40%|████ | 7216/18018 [15:42:04<21:28:18, 7.16s/step, loss=1.475, acc=0.668, acc_len=1.57, thru=14.6, I=17.3, T=28.7, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:12:18,134] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7215: forward=3118.4ms backward=1303.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:18,596] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:25,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:12:26,227] loop.py:45 INFO TRAIN_STEP step=7217 loss=1.387843 Training: 40%|████ | 7217/18018 [15:42:12<22:18:38, 7.44s/step, loss=1.388, acc=0.689, acc_len=1.65, thru=16.0, I=17.8, T=30.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:26,540] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:31,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:12:33,918] loop.py:45 INFO TRAIN_STEP step=7218 loss=1.472796 Training: 40%|████ | 7218/18018 [15:42:20<22:32:13, 7.51s/step, loss=1.473, acc=0.679, acc_len=1.61, thru=14.9, I=22.5, T=30.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:34,332] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:39,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:12:40,494] loop.py:45 INFO TRAIN_STEP step=7219 loss=1.396062 Training: 40%|████ | 7219/18018 [15:42:26<21:41:33, 7.23s/step, loss=1.396, acc=0.684, acc_len=1.62, thru=16.5, I=27.1, T=31.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:40,563] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:45,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:12:46,986] loop.py:380 INFO TIMING step=7220: step=4.469s data=2.143s compute=4.430s [fwd=3.128s bwd=1.286s opt=0.003s] dispatch=2.013s [2026-10-19 05:12:46,986] loop.py:45 INFO TRAIN_STEP step=7220 loss=1.459664 Training: 40%|████ | 7220/18018 [15:42:33<21:01:27, 7.01s/step, loss=1.460, acc=0.676, acc_len=1.60, thru=18.4, I=25.0, T=28.6, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:47,322] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:52,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:12:53,630] loop.py:45 INFO TRAIN_STEP step=7221 loss=1.460170 Training: 40%|████ | 7221/18018 [15:42:39<20:41:38, 6.90s/step, loss=1.460, acc=0.674, acc_len=1.61, thru=16.4, I=21.1, T=29.2, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:12:53,626] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7220: forward=3050.0ms backward=1287.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:53,843] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:12:59,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:13:01,158] loop.py:45 INFO TRAIN_STEP step=7222 loss=1.462112 Training: 40%|████ | 7222/18018 [15:42:47<21:15:24, 7.09s/step, loss=1.462, acc=0.676, acc_len=1.60, thru=15.0, I=20.9, T=27.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:01,586] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:08,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:13:10,148] loop.py:45 INFO TRAIN_STEP step=7223 loss=1.417212 Training: 40%|████ | 7223/18018 [15:42:56<22:58:02, 7.66s/step, loss=1.417, acc=0.684, acc_len=1.64, thru=14.9, I=18.8, T=29.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:10,248] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:18,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:13:18,667] loop.py:45 INFO TRAIN_STEP step=7224 loss=1.446930 Training: 40%|████ | 7224/18018 [15:43:04<23:44:14, 7.92s/step, loss=1.447, acc=0.679, acc_len=1.61, thru=13.3, I=14.4, T=31.2, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:19,086] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:25,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:13:27,296] loop.py:380 INFO TIMING step=7225: step=4.498s data=2.145s compute=4.465s [fwd=3.121s bwd=1.329s opt=0.003s] dispatch=4.122s [2026-10-19 05:13:27,296] loop.py:45 INFO TRAIN_STEP step=7225 loss=1.426571 Training: 40%|████ | 7225/18018 [15:43:13<24:22:35, 8.13s/step, loss=1.427, acc=0.681, acc_len=1.62, thru=15.6, I=20.3, T=28.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:27,582] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:34,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:13:35,953] loop.py:45 INFO TRAIN_STEP step=7226 loss=1.451280 Training: 40%|████ | 7226/18018 [15:43:22<24:50:48, 8.29s/step, loss=1.451, acc=0.679, acc_len=1.62, thru=15.3, I=17.5, T=31.9, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:13:35,949] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7225: forward=2800.8ms backward=1169.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:36,273] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:41,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:13:42,909] loop.py:45 INFO TRAIN_STEP step=7227 loss=1.339228 Training: 40%|████ | 7227/18018 [15:43:28<23:38:39, 7.89s/step, loss=1.339, acc=0.695, acc_len=1.66, thru=18.0, I=22.4, T=33.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:43,076] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:49,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:13:50,547] loop.py:45 INFO TRAIN_STEP step=7228 loss=1.448861 Training: 40%|████ | 7228/18018 [15:43:36<23:25:06, 7.81s/step, loss=1.449, acc=0.679, acc_len=1.62, thru=15.4, I=19.9, T=25.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:50,872] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:13:57,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:14:00,236] loop.py:45 INFO TRAIN_STEP step=7229 loss=1.440022 Training: 40%|████ | 7229/18018 [15:43:46<25:06:16, 8.38s/step, loss=1.440, acc=0.675, acc_len=1.61, thru=13.7, I=17.2, T=27.8, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:00,315] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 [2026-10-19 05:14:06,023] loop.py:380 INFO TIMING step=7230: step=3.364s data=1.395s compute=3.329s [fwd=2.364s bwd=0.950s opt=0.003s] dispatch=2.413s [2026-10-19 05:14:06,024] loop.py:45 INFO TRAIN_STEP step=7230 loss=1.444778 Training: 40%|████ | 7230/18018 [15:43:52<22:46:23, 7.60s/step, loss=1.445, acc=0.680, acc_len=1.62, thru=16.6, I=21.5, T=38.1, wait=2.4s, pool=44, epoch=1/2][2026-10-19 05:14:15,615] loop.py:45 INFO TRAIN_STEP step=7231 loss=1.455975 Training: 40%|████ | 7231/18018 [15:44:01<24:33:48, 8.20s/step, loss=1.456, acc=0.675, acc_len=1.60, thru=13.6, I=13.5, T=32.1, wait=5.6s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:14:15,610] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7230: forward=2806.9ms backward=1101.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:25,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:14:26,304] loop.py:45 INFO TRAIN_STEP step=7232 loss=1.390197 Training: 40%|████ | 7232/18018 [15:44:12<26:48:01, 8.95s/step, loss=1.390, acc=0.689, acc_len=1.65, thru=15.8, I=14.4, T=32.8, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:26,444] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:33,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:14:34,876] loop.py:45 INFO TRAIN_STEP step=7233 loss=1.462630 Training: 40%|████ | 7233/18018 [15:44:20<26:27:44, 8.83s/step, loss=1.463, acc=0.674, acc_len=1.60, thru=14.5, I=18.7, T=28.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:35,035] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:40,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:14:40,539] loop.py:45 INFO TRAIN_STEP step=7234 loss=1.471854 Training: 40%|████ | 7234/18018 [15:44:26<23:36:40, 7.88s/step, loss=1.472, acc=0.671, acc_len=1.59, thru=19.2, I=23.0, T=32.3, wait=1.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:40,789] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:47,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:14:47,873] loop.py:380 INFO TIMING step=7235: step=3.709s data=1.519s compute=3.664s [fwd=2.572s bwd=1.077s opt=0.003s] dispatch=3.614s [2026-10-19 05:14:47,874] loop.py:45 INFO TRAIN_STEP step=7235 loss=1.441452 Training: 40%|████ | 7235/18018 [15:44:33<23:06:54, 7.72s/step, loss=1.441, acc=0.679, acc_len=1.61, thru=17.4, I=19.6, T=34.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:48,110] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:55,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:14:56,126] loop.py:45 INFO TRAIN_STEP step=7236 loss=1.443015 Training: 40%|████ | 7236/18018 [15:44:42<23:35:42, 7.88s/step, loss=1.443, acc=0.676, acc_len=1.60, thru=15.5, I=18.2, T=27.9, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:14:56,121] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7235: forward=3142.8ms backward=1395.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:14:56,191] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:03,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:15:03,900] loop.py:45 INFO TRAIN_STEP step=7237 loss=1.481658 Training: 40%|████ | 7237/18018 [15:44:49<23:29:58, 7.85s/step, loss=1.482, acc=0.674, acc_len=1.60, thru=16.6, I=17.5, T=26.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:04,246] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:15:11,487] loop.py:45 INFO TRAIN_STEP step=7238 loss=1.437513 Training: 40%|████ | 7238/18018 [15:44:57<23:15:44, 7.77s/step, loss=1.438, acc=0.679, acc_len=1.62, thru=13.3, I=14.0, T=31.6, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:17,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:15:19,699] loop.py:45 INFO TRAIN_STEP step=7239 loss=1.475021 Training: 40%|████ | 7239/18018 [15:45:05<23:39:34, 7.90s/step, loss=1.475, acc=0.674, acc_len=1.60, thru=16.6, I=19.8, T=26.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:20,152] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:24,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:15:26,741] loop.py:380 INFO TIMING step=7240: step=4.558s data=2.071s compute=4.526s [fwd=3.171s bwd=1.340s opt=0.003s] dispatch=2.475s [2026-10-19 05:15:26,741] loop.py:45 INFO TRAIN_STEP step=7240 loss=1.407806 Training: 40%|████ | 7240/18018 [15:45:12<22:53:03, 7.64s/step, loss=1.408, acc=0.685, acc_len=1.62, thru=15.7, I=29.0, T=28.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:26,971] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:32,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:15:33,324] loop.py:45 INFO TRAIN_STEP step=7241 loss=1.306647 Training: 40%|████ | 7241/18018 [15:45:19<21:55:51, 7.33s/step, loss=1.307, acc=0.699, acc_len=1.68, thru=15.4, I=25.0, T=36.2, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:15:33,319] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7240: forward=2508.4ms backward=987.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:33,603] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:38,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:15:40,151] loop.py:45 INFO TRAIN_STEP step=7242 loss=1.577379 Training: 40%|████ | 7242/18018 [15:45:26<21:28:51, 7.18s/step, loss=1.577, acc=0.653, acc_len=1.52, thru=18.6, I=25.8, T=32.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:40,271] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:49,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:15:49,808] loop.py:45 INFO TRAIN_STEP step=7243 loss=1.475553 Training: 40%|████ | 7243/18018 [15:45:35<23:42:21, 7.92s/step, loss=1.476, acc=0.675, acc_len=1.60, thru=13.4, I=12.9, T=27.5, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:50,222] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:57,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:15:57,821] loop.py:45 INFO TRAIN_STEP step=7244 loss=1.499496 Training: 40%|████ | 7244/18018 [15:45:43<23:47:16, 7.95s/step, loss=1.499, acc=0.670, acc_len=1.59, thru=14.9, I=16.3, T=29.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:15:58,217] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:04,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:16:05,234] loop.py:380 INFO TIMING step=7245: step=4.455s data=2.034s compute=4.422s [fwd=3.065s bwd=1.343s opt=0.003s] dispatch=2.948s [2026-10-19 05:16:05,235] loop.py:45 INFO TRAIN_STEP step=7245 loss=1.314322 Training: 40%|████ | 7245/18018 [15:45:51<23:18:17, 7.79s/step, loss=1.314, acc=0.702, acc_len=1.69, thru=17.8, I=19.6, T=28.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:05,553] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:10,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:16:11,590] loop.py:45 INFO TRAIN_STEP step=7246 loss=1.389430 Training: 40%|████ | 7246/18018 [15:45:57<22:00:58, 7.36s/step, loss=1.389, acc=0.689, acc_len=1.65, thru=17.2, I=21.6, T=28.9, wait=1.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:16:11,586] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7245: forward=3042.1ms backward=1335.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:11,637] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 05:16:19,279] loop.py:45 INFO TRAIN_STEP step=7247 loss=1.561325 Training: 40%|████ | 7247/18018 [15:46:05<22:18:41, 7.46s/step, loss=1.561, acc=0.656, acc_len=1.54, thru=13.1, I=16.7, T=32.0, wait=3.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:29,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:16:32,512] loop.py:45 INFO TRAIN_STEP step=7248 loss=1.501756 Training: 40%|████ | 7248/18018 [15:46:18<27:29:33, 9.19s/step, loss=1.502, acc=0.671, acc_len=1.59, thru=13.6, I=14.6, T=23.3, wait=7.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:32,955] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:38,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:16:39,029] loop.py:45 INFO TRAIN_STEP step=7249 loss=1.401801 Training: 40%|████ | 7249/18018 [15:46:25<25:05:27, 8.39s/step, loss=1.402, acc=0.684, acc_len=1.63, thru=15.3, I=23.3, T=34.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:39,368] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:46,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:16:46,996] loop.py:380 INFO TIMING step=7250: step=3.996s data=1.771s compute=3.965s [fwd=2.831s bwd=1.119s opt=0.003s] dispatch=3.962s [2026-10-19 05:16:46,996] loop.py:45 INFO TRAIN_STEP step=7250 loss=1.546167 Training: 40%|████ | 7250/18018 [15:46:33<24:42:44, 8.26s/step, loss=1.546, acc=0.662, acc_len=1.56, thru=15.8, I=16.4, T=32.0, wait=4.0s, pool=64, epoch=1/2][2026-10-19 05:16:47,000] loop.py:419 INFO Saving checkpoint at step 7250... (TrainerActor pid=9310) [2026-10-19 05:16:49,847] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0007251 (TrainerActor pid=9310) [2026-10-19 05:16:49,854] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0006501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:51,285] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:51,285] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:58,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 05:16:49,848] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0007251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 05:16:50,839] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0006501 [2026-10-19 05:16:58,952] loop.py:45 INFO TRAIN_STEP step=7251 loss=1.417848 Training: 40%|████ | 7251/18018 [15:46:45<28:01:30, 9.37s/step, loss=1.418, acc=0.679, acc_len=1.61, thru=16.7, I=17.9, T=26.3, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:16:58,948] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7250: forward=3326.1ms backward=1491.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:16:59,202] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:05,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:17:06,441] loop.py:45 INFO TRAIN_STEP step=7252 loss=1.443139 Training: 40%|████ | 7252/18018 [15:46:52<26:20:00, 8.81s/step, loss=1.443, acc=0.678, acc_len=1.61, thru=16.5, I=18.1, T=32.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:06,731] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:13,192] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:17:15,647] loop.py:45 INFO TRAIN_STEP step=7253 loss=1.462085 Training: 40%|████ | 7253/18018 [15:47:01<26:41:29, 8.93s/step, loss=1.462, acc=0.677, acc_len=1.60, thru=14.4, I=18.1, T=26.3, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:15,731] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:21,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:17:23,610] loop.py:45 INFO TRAIN_STEP step=7254 loss=1.533503 Training: 40%|████ | 7254/18018 [15:47:09<25:49:26, 8.64s/step, loss=1.534, acc=0.664, acc_len=1.56, thru=16.3, I=20.6, T=25.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:24,016] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 05:17:31,090] loop.py:380 INFO TIMING step=7255: step=3.580s data=1.517s compute=3.549s [fwd=2.560s bwd=0.974s opt=0.003s] dispatch=3.890s [2026-10-19 05:17:31,090] loop.py:45 INFO TRAIN_STEP step=7255 loss=1.454733 Training: 40%|████ | 7255/18018 [15:47:17<24:47:01, 8.29s/step, loss=1.455, acc=0.675, acc_len=1.59, thru=11.0, I=17.1, T=35.8, wait=3.9s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:38,745] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:17:41,200] loop.py:45 INFO TRAIN_STEP step=7256 loss=1.550848 Training: 40%|████ | 7256/18018 [15:47:27<26:24:56, 8.84s/step, loss=1.551, acc=0.663, acc_len=1.56, thru=15.6, I=18.6, T=29.1, wait=5.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:17:41,196] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7255: forward=3124.7ms backward=1232.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:41,270] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:47,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:17:47,913] loop.py:45 INFO TRAIN_STEP step=7257 loss=1.411986 Training: 40%|████ | 7257/18018 [15:47:33<24:30:32, 8.20s/step, loss=1.412, acc=0.680, acc_len=1.61, thru=15.6, I=23.0, T=30.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:48,162] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:53,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:17:54,737] loop.py:45 INFO TRAIN_STEP step=7258 loss=1.501965 Training: 40%|████ | 7258/18018 [15:47:40<23:16:23, 7.79s/step, loss=1.502, acc=0.668, acc_len=1.58, thru=18.9, I=21.8, T=29.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:54,877] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:17:59,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:18:00,615] loop.py:45 INFO TRAIN_STEP step=7259 loss=1.456326 Training: 40%|████ | 7259/18018 [15:47:46<21:33:27, 7.21s/step, loss=1.456, acc=0.676, acc_len=1.60, thru=18.3, I=23.4, T=33.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:00,899] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:06,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:18:08,899] loop.py:380 INFO TIMING step=7260: step=4.451s data=2.150s compute=4.414s [fwd=3.133s bwd=1.266s opt=0.003s] dispatch=3.826s [2026-10-19 05:18:08,900] loop.py:45 INFO TRAIN_STEP step=7260 loss=1.581642 Training: 40%|████ | 7260/18018 [15:47:54<22:31:06, 7.54s/step, loss=1.582, acc=0.653, acc_len=1.53, thru=15.2, I=20.8, T=28.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:09,005] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:14,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:18:16,030] loop.py:45 INFO TRAIN_STEP step=7261 loss=1.440444 Training: 40%|████ | 7261/18018 [15:48:02<22:09:11, 7.41s/step, loss=1.440, acc=0.675, acc_len=1.60, thru=15.1, I=21.2, T=33.0, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:18:16,025] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7260: forward=2739.6ms backward=1088.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:16,476] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:22,930] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:18:23,244] loop.py:45 INFO TRAIN_STEP step=7262 loss=1.367499 Training: 40%|████ | 7262/18018 [15:48:09<21:58:14, 7.35s/step, loss=1.367, acc=0.691, acc_len=1.65, thru=17.5, I=19.3, T=33.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:23,433] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:18:32,246] loop.py:45 INFO TRAIN_STEP step=7263 loss=1.304423 Training: 40%|████ | 7263/18018 [15:48:18<23:26:42, 7.85s/step, loss=1.304, acc=0.697, acc_len=1.68, thru=14.5, I=14.0, T=32.3, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:32,198] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:32,700] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:39,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:18:40,817] loop.py:45 INFO TRAIN_STEP step=7264 loss=1.552334 Training: 40%|████ | 7264/18018 [15:48:26<24:05:30, 8.06s/step, loss=1.552, acc=0.655, acc_len=1.53, thru=16.7, I=17.8, T=27.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:40,961] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:18:47,678] loop.py:380 INFO TIMING step=7265: step=4.194s data=1.878s compute=4.159s [fwd=2.968s bwd=1.177s opt=0.003s] dispatch=2.659s [2026-10-19 05:18:47,678] loop.py:45 INFO TRAIN_STEP step=7265 loss=1.517223 Training: 40%|████ | 7265/18018 [15:48:33<23:00:39, 7.70s/step, loss=1.517, acc=0.662, acc_len=1.56, thru=14.3, I=16.4, T=30.5, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:54,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:18:55,179] loop.py:45 INFO TRAIN_STEP step=7266 loss=1.465189 Training: 40%|████ | 7266/18018 [15:48:41<22:49:35, 7.64s/step, loss=1.465, acc=0.673, acc_len=1.60, thru=16.6, I=17.3, T=31.3, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:18:55,174] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7265: forward=2937.4ms backward=1107.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:18:55,617] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:02,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:19:03,371] loop.py:45 INFO TRAIN_STEP step=7267 loss=1.371182 Training: 40%|████ | 7267/18018 [15:48:49<23:19:03, 7.81s/step, loss=1.371, acc=0.686, acc_len=1.63, thru=15.9, I=18.4, T=25.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:03,570] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:09,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:19:11,668] loop.py:45 INFO TRAIN_STEP step=7268 loss=1.438480 Training: 40%|████ | 7268/18018 [15:48:57<23:45:08, 7.95s/step, loss=1.438, acc=0.682, acc_len=1.62, thru=14.1, I=17.0, T=26.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:12,129] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:17,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:19:18,288] loop.py:45 INFO TRAIN_STEP step=7269 loss=1.320028 Training: 40%|████ | 7269/18018 [15:49:04<22:33:27, 7.55s/step, loss=1.320, acc=0.697, acc_len=1.68, thru=17.2, I=22.8, T=31.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:18,495] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:19:27,598] loop.py:380 INFO TIMING step=7270: step=4.061s data=1.862s compute=4.032s [fwd=2.869s bwd=1.148s opt=0.003s] dispatch=5.234s [2026-10-19 05:19:27,598] loop.py:45 INFO TRAIN_STEP step=7270 loss=1.421701 Training: 40%|████ | 7270/18018 [15:49:13<24:07:39, 8.08s/step, loss=1.422, acc=0.680, acc_len=1.61, thru=11.4, I=12.7, T=31.5, wait=5.2s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:34,276] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:19:35,464] loop.py:45 INFO TRAIN_STEP step=7271 loss=1.440975 Training: 40%|████ | 7271/18018 [15:49:21<23:55:47, 8.02s/step, loss=1.441, acc=0.681, acc_len=1.61, thru=17.3, I=16.8, T=33.8, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:19:35,459] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7270: forward=2691.4ms backward=1042.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:35,785] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:43,581] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:19:44,100] loop.py:45 INFO TRAIN_STEP step=7272 loss=1.518622 Training: 40%|████ | 7272/18018 [15:49:30<24:28:57, 8.20s/step, loss=1.519, acc=0.666, acc_len=1.58, thru=13.7, I=15.8, T=29.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:44,588] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:51,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:19:51,911] loop.py:45 INFO TRAIN_STEP step=7273 loss=1.475660 Training: 40%|████ | 7273/18018 [15:49:37<24:07:56, 8.09s/step, loss=1.476, acc=0.672, acc_len=1.59, thru=16.1, I=17.7, T=33.1, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:52,403] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:58,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:19:58,658] loop.py:45 INFO TRAIN_STEP step=7274 loss=1.445967 Training: 40%|████ | 7274/18018 [15:49:44<22:55:55, 7.68s/step, loss=1.446, acc=0.680, acc_len=1.61, thru=18.5, I=21.8, T=36.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:19:59,047] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:20:05,844] loop.py:380 INFO TIMING step=7275: step=4.447s data=2.144s compute=4.410s [fwd=3.064s bwd=1.330s opt=0.003s] dispatch=2.729s [2026-10-19 05:20:05,844] loop.py:45 INFO TRAIN_STEP step=7275 loss=1.527629 Training: 40%|████ | 7275/18018 [15:49:51<22:29:01, 7.53s/step, loss=1.528, acc=0.660, acc_len=1.55, thru=14.9, I=19.2, T=28.8, wait=2.7s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:13,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:20:14,988] loop.py:45 INFO TRAIN_STEP step=7276 loss=1.410832 Training: 40%|████ | 7276/18018 [15:50:01<23:55:21, 8.02s/step, loss=1.411, acc=0.684, acc_len=1.63, thru=16.4, I=16.9, T=28.3, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:20:14,984] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7275: forward=3153.4ms backward=1297.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:15,097] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:19,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:20:21,008] loop.py:45 INFO TRAIN_STEP step=7277 loss=1.509478 Training: 40%|████ | 7277/18018 [15:50:07<22:07:51, 7.42s/step, loss=1.509, acc=0.669, acc_len=1.58, thru=18.4, I=25.6, T=30.1, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:21,423] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:27,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:20:28,197] loop.py:45 INFO TRAIN_STEP step=7278 loss=1.371153 Training: 40%|████ | 7278/18018 [15:50:14<21:55:31, 7.35s/step, loss=1.371, acc=0.690, acc_len=1.65, thru=15.6, I=20.2, T=30.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:28,557] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:34,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:20:35,496] loop.py:45 INFO TRAIN_STEP step=7279 loss=1.322151 Training: 40%|████ | 7279/18018 [15:50:21<21:52:45, 7.33s/step, loss=1.322, acc=0.696, acc_len=1.67, thru=17.1, I=19.9, T=27.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:35,642] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:41,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:20:43,095] loop.py:380 INFO TIMING step=7280: step=4.942s data=2.587s compute=4.908s [fwd=3.515s bwd=1.379s opt=0.003s] dispatch=2.646s [2026-10-19 05:20:43,096] loop.py:45 INFO TRAIN_STEP step=7280 loss=1.478181 Training: 40%|████ | 7280/18018 [15:50:29<22:06:50, 7.41s/step, loss=1.478, acc=0.671, acc_len=1.58, thru=16.7, I=20.6, T=25.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:43,308] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:54,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:20:54,873] loop.py:45 INFO TRAIN_STEP step=7281 loss=1.476681 Training: 40%|████ | 7281/18018 [15:50:40<26:00:59, 8.72s/step, loss=1.477, acc=0.674, acc_len=1.60, thru=10.7, I=10.6, T=28.0, wait=7.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:20:54,929] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (TrainerActor pid=9310) [2026-10-19 05:20:54,869] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7280: forward=3218.9ms backward=1312.9ms [2026-10-19 05:21:02,211] loop.py:45 INFO TRAIN_STEP step=7282 loss=1.337221 Training: 40%|████ | 7282/18018 [15:50:48<24:46:24, 8.31s/step, loss=1.337, acc=0.694, acc_len=1.66, thru=15.3, I=17.2, T=33.2, wait=3.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:10,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:21:10,793] loop.py:45 INFO TRAIN_STEP step=7283 loss=1.513959 Training: 40%|████ | 7283/18018 [15:50:56<25:01:07, 8.39s/step, loss=1.514, acc=0.665, acc_len=1.57, thru=16.5, I=16.4, T=31.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:11,133] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:16,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:21:19,086] loop.py:45 INFO TRAIN_STEP step=7284 loss=1.468954 Training: 40%|████ | 7284/18018 [15:51:05<24:55:45, 8.36s/step, loss=1.469, acc=0.671, acc_len=1.59, thru=15.4, I=20.7, T=28.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:19,340] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 05:21:26,137] loop.py:380 INFO TIMING step=7285: step=4.004s data=1.824s compute=3.970s [fwd=2.852s bwd=1.103s opt=0.003s] dispatch=3.038s [2026-10-19 05:21:26,137] loop.py:45 INFO TRAIN_STEP step=7285 loss=1.430349 Training: 40%|████ | 7285/18018 [15:51:12<23:45:13, 7.97s/step, loss=1.430, acc=0.675, acc_len=1.60, thru=15.3, I=19.6, T=32.0, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:33,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:21:34,634] loop.py:45 INFO TRAIN_STEP step=7286 loss=1.334899 Training: 40%|████ | 7286/18018 [15:51:20<24:13:29, 8.13s/step, loss=1.335, acc=0.696, acc_len=1.67, thru=16.3, I=16.1, T=23.3, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:21:34,629] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7285: forward=3974.9ms backward=1465.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:34,784] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:42,638] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:21:43,592] loop.py:45 INFO TRAIN_STEP step=7287 loss=1.386219 Training: 40%|████ | 7287/18018 [15:51:29<24:58:05, 8.38s/step, loss=1.386, acc=0.691, acc_len=1.65, thru=14.5, I=17.7, T=21.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:43,644] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:49,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:21:50,509] loop.py:45 INFO TRAIN_STEP step=7288 loss=1.441077 Training: 40%|████ | 7288/18018 [15:51:36<23:39:40, 7.94s/step, loss=1.441, acc=0.679, acc_len=1.61, thru=15.3, I=15.8, T=30.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:21:50,557] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:00,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:22:00,286] loop.py:45 INFO TRAIN_STEP step=7289 loss=1.421812 Training: 40%|████ | 7289/18018 [15:51:46<25:18:02, 8.49s/step, loss=1.422, acc=0.678, acc_len=1.61, thru=13.2, I=13.4, T=26.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:00,702] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:09,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:22:10,556] loop.py:380 INFO TIMING step=7290: step=4.406s data=1.960s compute=4.380s [fwd=3.023s bwd=1.342s opt=0.003s] dispatch=5.857s [2026-10-19 05:22:10,556] loop.py:45 INFO TRAIN_STEP step=7290 loss=1.364790 Training: 40%|████ | 7290/18018 [15:51:56<26:53:25, 9.02s/step, loss=1.365, acc=0.691, acc_len=1.64, thru=13.0, I=13.8, T=29.0, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:10,966] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:19,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:22:20,137] loop.py:45 INFO TRAIN_STEP step=7291 loss=1.393282 Training: 40%|████ | 7291/18018 [15:52:06<27:23:14, 9.19s/step, loss=1.393, acc=0.688, acc_len=1.64, thru=14.0, I=14.7, T=32.0, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:22:20,132] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7290: forward=2819.0ms backward=1130.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:20,600] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:27,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:22:29,270] loop.py:45 INFO TRAIN_STEP step=7292 loss=1.419928 Training: 40%|████ | 7292/18018 [15:52:15<27:19:57, 9.17s/step, loss=1.420, acc=0.681, acc_len=1.63, thru=14.1, I=17.4, T=25.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:29,393] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:34,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:22:36,898] loop.py:45 INFO TRAIN_STEP step=7293 loss=1.542718 Training: 40%|████ | 7293/18018 [15:52:22<25:56:57, 8.71s/step, loss=1.543, acc=0.661, acc_len=1.56, thru=15.4, I=22.4, T=28.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:37,395] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:44,173] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:22:45,440] loop.py:45 INFO TRAIN_STEP step=7294 loss=1.491822 Training: 40%|████ | 7294/18018 [15:52:31<25:47:42, 8.66s/step, loss=1.492, acc=0.670, acc_len=1.59, thru=15.9, I=18.5, T=31.1, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:45,679] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:51,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:22:51,911] loop.py:380 INFO TIMING step=7295: step=3.846s data=1.757s compute=3.814s [fwd=2.714s bwd=1.085s opt=0.003s] dispatch=2.616s [2026-10-19 05:22:51,911] loop.py:45 INFO TRAIN_STEP step=7295 loss=1.469128 Training: 40%|████ | 7295/18018 [15:52:37<23:50:19, 8.00s/step, loss=1.469, acc=0.674, acc_len=1.60, thru=18.3, I=21.9, T=33.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:52,064] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:22:57,749] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:23:00,566] loop.py:45 INFO TRAIN_STEP step=7296 loss=1.449637 Training: 40%|████ | 7296/18018 [15:52:46<24:25:02, 8.20s/step, loss=1.450, acc=0.671, acc_len=1.59, thru=15.1, I=22.4, T=26.0, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:23:00,562] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7295: forward=3365.3ms backward=1520.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:00,775] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:07,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:23:08,513] loop.py:45 INFO TRAIN_STEP step=7297 loss=1.556342 Training: 40%|████ | 7297/18018 [15:52:54<24:11:25, 8.12s/step, loss=1.556, acc=0.656, acc_len=1.54, thru=16.6, I=19.6, T=27.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:08,803] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:23:15,743] loop.py:45 INFO TRAIN_STEP step=7298 loss=1.401364 Training: 41%|████ | 7298/18018 [15:53:01<23:23:25, 7.85s/step, loss=1.401, acc=0.685, acc_len=1.63, thru=15.3, I=18.6, T=28.9, wait=2.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:24,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:23:26,029] loop.py:45 INFO TRAIN_STEP step=7299 loss=1.418166 Training: 41%|████ | 7299/18018 [15:53:12<25:33:32, 8.58s/step, loss=1.418, acc=0.682, acc_len=1.62, thru=13.7, I=14.9, T=24.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:26,057] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:31,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:23:32,525] loop.py:380 INFO TIMING step=7300: step=4.281s data=1.897s compute=4.249s [fwd=2.918s bwd=1.315s opt=0.003s] dispatch=2.208s [2026-10-19 05:23:32,525] loop.py:45 INFO TRAIN_STEP step=7300 loss=1.285029 Training: 41%|████ | 7300/18018 [15:53:18<23:41:34, 7.96s/step, loss=1.285, acc=0.702, acc_len=1.69, thru=16.5, I=21.8, T=29.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:32,928] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:37,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:23:39,122] loop.py:45 INFO TRAIN_STEP step=7301 loss=1.432070 Training: 41%|████ | 7301/18018 [15:53:25<22:28:28, 7.55s/step, loss=1.432, acc=0.681, acc_len=1.63, thru=16.1, I=22.3, T=32.9, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:23:39,118] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7300: forward=2728.9ms backward=1105.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:39,512] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:47,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:23:47,778] loop.py:45 INFO TRAIN_STEP step=7302 loss=1.461756 Training: 41%|████ | 7302/18018 [15:53:33<23:27:38, 7.88s/step, loss=1.462, acc=0.671, acc_len=1.59, thru=15.5, I=15.8, T=26.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:47,806] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:53,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:23:56,123] loop.py:45 INFO TRAIN_STEP step=7303 loss=1.468042 Training: 41%|████ | 7303/18018 [15:53:42<23:52:28, 8.02s/step, loss=1.468, acc=0.675, acc_len=1.60, thru=15.8, I=18.5, T=24.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:23:56,183] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:02,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:24:03,282] loop.py:45 INFO TRAIN_STEP step=7304 loss=1.414908 Training: 41%|████ | 7304/18018 [15:53:49<23:06:00, 7.76s/step, loss=1.415, acc=0.685, acc_len=1.63, thru=13.8, I=20.4, T=31.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:03,524] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:09,587] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:24:10,405] loop.py:380 INFO TIMING step=7305: step=3.962s data=1.842s compute=3.907s [fwd=2.760s bwd=1.132s opt=0.003s] dispatch=3.152s [2026-10-19 05:24:10,405] loop.py:45 INFO TRAIN_STEP step=7305 loss=1.543389 Training: 41%|████ | 7305/18018 [15:53:56<22:31:43, 7.57s/step, loss=1.543, acc=0.666, acc_len=1.57, thru=16.6, I=20.2, T=32.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:10,593] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:15,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:24:17,313] loop.py:45 INFO TRAIN_STEP step=7306 loss=1.299835 Training: 41%|████ | 7306/18018 [15:54:03<21:56:11, 7.37s/step, loss=1.300, acc=0.703, acc_len=1.69, thru=17.9, I=23.6, T=28.9, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:24:17,309] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7305: forward=3042.6ms backward=1329.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:17,524] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:25,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:24:27,347] loop.py:45 INFO TRAIN_STEP step=7307 loss=1.485434 Training: 41%|████ | 7307/18018 [15:54:13<24:18:35, 8.17s/step, loss=1.485, acc=0.671, acc_len=1.59, thru=13.0, I=14.7, T=29.7, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:27,435] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:33,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:24:33,692] loop.py:45 INFO TRAIN_STEP step=7308 loss=1.445019 Training: 41%|████ | 7308/18018 [15:54:19<22:40:40, 7.62s/step, loss=1.445, acc=0.682, acc_len=1.62, thru=17.7, I=21.9, T=32.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:34,105] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:39,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:24:40,062] loop.py:45 INFO TRAIN_STEP step=7309 loss=1.351756 Training: 41%|████ | 7309/18018 [15:54:26<21:33:27, 7.25s/step, loss=1.352, acc=0.695, acc_len=1.66, thru=18.6, I=21.3, T=32.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:40,355] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:46,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:24:47,710] loop.py:380 INFO TIMING step=7310: step=4.572s data=2.206s compute=4.527s [fwd=3.207s bwd=1.305s opt=0.003s] dispatch=3.066s [2026-10-19 05:24:47,710] loop.py:45 INFO TRAIN_STEP step=7310 loss=1.357403 Training: 41%|████ | 7310/18018 [15:54:33<21:54:50, 7.37s/step, loss=1.357, acc=0.692, acc_len=1.66, thru=16.1, I=18.4, T=28.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:47,798] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:54,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:24:54,749] loop.py:45 INFO TRAIN_STEP step=7311 loss=1.488853 Training: 41%|████ | 7311/18018 [15:54:40<21:37:05, 7.27s/step, loss=1.489, acc=0.671, acc_len=1.59, thru=18.2, I=20.1, T=29.9, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:24:54,745] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7310: forward=3060.0ms backward=1175.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:24:55,088] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:00,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:25:02,249] loop.py:45 INFO TRAIN_STEP step=7312 loss=1.413937 Training: 41%|████ | 7312/18018 [15:54:48<21:49:22, 7.34s/step, loss=1.414, acc=0.683, acc_len=1.62, thru=15.8, I=21.0, T=27.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:02,287] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-19 05:25:08,465] loop.py:45 INFO TRAIN_STEP step=7313 loss=1.404211 Training: 41%|████ | 7313/18018 [15:54:54<20:49:04, 7.00s/step, loss=1.404, acc=0.682, acc_len=1.62, thru=16.4, I=22.2, T=31.5, wait=2.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:18,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:25:18,819] loop.py:45 INFO TRAIN_STEP step=7314 loss=1.373920 Training: 41%|████ | 7314/18018 [15:55:04<23:48:28, 8.01s/step, loss=1.374, acc=0.691, acc_len=1.65, thru=13.7, I=14.5, T=27.7, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:19,253] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:25,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:25:27,942] loop.py:380 INFO TIMING step=7315: step=4.782s data=2.254s compute=4.736s [fwd=3.328s bwd=1.392s opt=0.003s] dispatch=4.332s [2026-10-19 05:25:27,943] loop.py:45 INFO TRAIN_STEP step=7315 loss=1.456824 Training: 41%|████ | 7315/18018 [15:55:14<24:48:01, 8.34s/step, loss=1.457, acc=0.679, acc_len=1.61, thru=14.2, I=16.2, T=26.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:28,320] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:34,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:25:36,266] loop.py:45 INFO TRAIN_STEP step=7316 loss=1.337431 Training: 41%|████ | 7316/18018 [15:55:22<24:47:00, 8.34s/step, loss=1.337, acc=0.693, acc_len=1.66, thru=16.1, I=18.7, T=26.4, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:25:36,261] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7315: forward=3388.0ms backward=1402.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:36,368] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:42,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:25:43,759] loop.py:45 INFO TRAIN_STEP step=7317 loss=1.419004 Training: 41%|████ | 7317/18018 [15:55:29<24:01:44, 8.08s/step, loss=1.419, acc=0.679, acc_len=1.61, thru=16.4, I=21.1, T=31.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:43,977] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:25:51,683] loop.py:45 INFO TRAIN_STEP step=7318 loss=1.396773 Training: 41%|████ | 7318/18018 [15:55:37<23:53:09, 8.04s/step, loss=1.397, acc=0.684, acc_len=1.63, thru=16.1, I=16.8, T=32.6, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:25:59,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:26:01,536] loop.py:45 INFO TRAIN_STEP step=7319 loss=1.359527 Training: 41%|████ | 7319/18018 [15:55:47<25:30:04, 8.58s/step, loss=1.360, acc=0.692, acc_len=1.66, thru=13.5, I=18.2, T=26.2, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:01,691] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:06,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:26:07,341] loop.py:380 INFO TIMING step=7320: step=3.736s data=1.662s compute=3.704s [fwd=2.624s bwd=1.066s opt=0.003s] dispatch=2.058s [2026-10-19 05:26:07,341] loop.py:45 INFO TRAIN_STEP step=7320 loss=1.263199 Training: 41%|████ | 7320/18018 [15:55:53<23:01:22, 7.75s/step, loss=1.263, acc=0.706, acc_len=1.71, thru=18.4, I=26.9, T=34.3, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:07,474] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:15,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:26:16,501] loop.py:45 INFO TRAIN_STEP step=7321 loss=1.494846 Training: 41%|████ | 7321/18018 [15:56:02<24:16:52, 8.17s/step, loss=1.495, acc=0.670, acc_len=1.58, thru=14.1, I=16.1, T=26.7, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:26:16,496] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7320: forward=3227.7ms backward=1499.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:16,732] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:26:24,007] loop.py:45 INFO TRAIN_STEP step=7322 loss=1.380724 Training: 41%|████ | 7322/18018 [15:56:10<23:41:05, 7.97s/step, loss=1.381, acc=0.688, acc_len=1.64, thru=14.6, I=18.7, T=27.9, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:31,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:26:32,219] loop.py:45 INFO TRAIN_STEP step=7323 loss=1.475367 Training: 41%|████ | 7323/18018 [15:56:18<23:53:48, 8.04s/step, loss=1.475, acc=0.671, acc_len=1.58, thru=15.6, I=15.0, T=30.1, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:32,593] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:37,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:26:39,322] loop.py:45 INFO TRAIN_STEP step=7324 loss=1.336745 Training: 41%|████ | 7324/18018 [15:56:25<23:03:17, 7.76s/step, loss=1.337, acc=0.693, acc_len=1.66, thru=17.1, I=21.3, T=29.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:39,771] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:45,621] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:26:46,378] loop.py:380 INFO TIMING step=7325: step=3.765s data=1.675s compute=3.729s [fwd=2.632s bwd=1.082s opt=0.003s] dispatch=3.284s [2026-10-19 05:26:46,378] loop.py:45 INFO TRAIN_STEP step=7325 loss=1.384597 Training: 41%|████ | 7325/18018 [15:56:32<22:25:36, 7.55s/step, loss=1.385, acc=0.686, acc_len=1.63, thru=16.6, I=22.8, T=34.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:46,627] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:52,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:26:53,372] loop.py:45 INFO TRAIN_STEP step=7326 loss=1.498672 Training: 41%|████ | 7326/18018 [15:56:39<21:55:42, 7.38s/step, loss=1.499, acc=0.670, acc_len=1.59, thru=18.0, I=20.3, T=30.4, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:26:53,367] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7325: forward=2875.6ms backward=1287.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:53,796] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:26:59,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:27:00,071] loop.py:45 INFO TRAIN_STEP step=7327 loss=1.295653 Training: 41%|████ | 7327/18018 [15:56:46<21:19:01, 7.18s/step, loss=1.296, acc=0.704, acc_len=1.70, thru=18.9, I=21.9, T=33.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:00,467] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:05,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:27:07,014] loop.py:45 INFO TRAIN_STEP step=7328 loss=1.385683 Training: 41%|████ | 7328/18018 [15:56:53<21:06:20, 7.11s/step, loss=1.386, acc=0.688, acc_len=1.64, thru=17.4, I=21.1, T=32.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:07,362] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:12,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:27:15,075] loop.py:45 INFO TRAIN_STEP step=7329 loss=1.431822 Training: 41%|████ | 7329/18018 [15:57:01<21:57:05, 7.39s/step, loss=1.432, acc=0.680, acc_len=1.61, thru=15.6, I=23.5, T=28.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:15,390] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:22,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:27:23,675] loop.py:380 INFO TIMING step=7330: step=4.227s data=1.994s compute=4.193s [fwd=2.976s bwd=1.202s opt=0.003s] dispatch=4.364s [2026-10-19 05:27:23,675] loop.py:45 INFO TRAIN_STEP step=7330 loss=1.492149 Training: 41%|████ | 7330/18018 [15:57:09<23:01:26, 7.76s/step, loss=1.492, acc=0.669, acc_len=1.58, thru=15.4, I=17.7, T=30.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:24,136] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:31,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:27:32,207] loop.py:45 INFO TRAIN_STEP step=7331 loss=1.470777 (TrainerActor pid=9310) [2026-10-19 05:27:32,202] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7330: forward=2834.2ms backward=1145.0ms Training: 41%|████ | 7331/18018 [15:57:18<23:42:55, 7.99s/step, loss=1.471, acc=0.670, acc_len=1.58, thru=14.1, I=16.6, T=31.8, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:32,423] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:38,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:27:39,086] loop.py:45 INFO TRAIN_STEP step=7332 loss=1.423618 Training: 41%|████ | 7332/18018 [15:57:25<22:43:28, 7.66s/step, loss=1.424, acc=0.680, acc_len=1.62, thru=19.1, I=19.0, T=30.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:39,121] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 05:27:45,237] loop.py:45 INFO TRAIN_STEP step=7333 loss=1.464728 Training: 41%|████ | 7333/18018 [15:57:31<21:22:59, 7.20s/step, loss=1.465, acc=0.673, acc_len=1.59, thru=15.3, I=21.7, T=33.9, wait=2.4s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:53,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:27:55,769] loop.py:45 INFO TRAIN_STEP step=7334 loss=1.335846 Training: 41%|████ | 7334/18018 [15:57:41<24:20:39, 8.20s/step, loss=1.336, acc=0.694, acc_len=1.66, thru=14.9, I=18.7, T=28.0, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:27:55,884] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:02,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:28:04,012] loop.py:380 INFO TIMING step=7335: step=4.637s data=2.336s compute=4.599s [fwd=3.188s bwd=1.397s opt=0.003s] dispatch=3.596s [2026-10-19 05:28:04,012] loop.py:45 INFO TRAIN_STEP step=7335 loss=1.367534 Training: 41%|████ | 7335/18018 [15:57:50<24:22:37, 8.21s/step, loss=1.368, acc=0.692, acc_len=1.66, thru=15.8, I=18.3, T=27.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:04,394] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:28:12,758] loop.py:45 INFO TRAIN_STEP step=7336 loss=1.448707 Training: 41%|████ | 7336/18018 [15:57:58<24:50:48, 8.37s/step, loss=1.449, acc=0.676, acc_len=1.60, thru=14.8, I=16.1, T=31.0, wait=4.6s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:28:12,753] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7335: forward=2908.7ms backward=1171.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:19,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:28:19,746] loop.py:45 INFO TRAIN_STEP step=7337 loss=1.373087 Training: 41%|████ | 7337/18018 [15:58:05<23:36:39, 7.96s/step, loss=1.373, acc=0.690, acc_len=1.65, thru=18.1, I=17.3, T=34.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:20,061] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:24,916] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:28:25,843] loop.py:45 INFO TRAIN_STEP step=7338 loss=1.374612 Training: 41%|████ | 7338/18018 [15:58:11<21:57:16, 7.40s/step, loss=1.375, acc=0.687, acc_len=1.64, thru=19.8, I=22.3, T=36.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:25,940] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-19 05:28:32,580] loop.py:45 INFO TRAIN_STEP step=7339 loss=1.424357 Training: 41%|████ | 7339/18018 [15:58:18<21:21:39, 7.20s/step, loss=1.424, acc=0.680, acc_len=1.63, thru=17.8, I=20.2, T=31.4, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:41,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:28:42,815] loop.py:380 INFO TIMING step=7340: step=4.718s data=2.313s compute=4.669s [fwd=3.269s bwd=1.385s opt=0.003s] dispatch=5.506s [2026-10-19 05:28:42,815] loop.py:45 INFO TRAIN_STEP step=7340 loss=1.531015 Training: 41%|████ | 7340/18018 [15:58:28<24:03:31, 8.11s/step, loss=1.531, acc=0.663, acc_len=1.56, thru=13.3, I=15.9, T=27.1, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:43,036] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:49,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:28:50,261] loop.py:45 INFO TRAIN_STEP step=7341 loss=1.351258 Training: 41%|████ | 7341/18018 [15:58:36<23:27:51, 7.91s/step, loss=1.351, acc=0.691, acc_len=1.66, thru=15.1, I=20.1, T=26.6, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:28:50,256] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7340: forward=3271.0ms backward=1491.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:50,538] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:56,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:28:57,275] loop.py:45 INFO TRAIN_STEP step=7342 loss=1.438868 Training: 41%|████ | 7342/18018 [15:58:43<22:39:51, 7.64s/step, loss=1.439, acc=0.675, acc_len=1.59, thru=16.6, I=19.6, T=30.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:28:57,417] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 05:29:03,554] loop.py:45 INFO TRAIN_STEP step=7343 loss=1.576285 Training: 41%|████ | 7343/18018 [15:58:49<21:26:57, 7.23s/step, loss=1.576, acc=0.656, acc_len=1.54, thru=17.3, I=19.3, T=32.4, wait=2.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:11,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:29:12,121] loop.py:45 INFO TRAIN_STEP step=7344 loss=1.414397 Training: 41%|████ | 7344/18018 [15:58:58<22:37:55, 7.63s/step, loss=1.414, acc=0.685, acc_len=1.64, thru=16.3, I=16.5, T=33.2, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:12,472] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:17,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:29:20,174] loop.py:380 INFO TIMING step=7345: step=4.471s data=2.034s compute=4.441s [fwd=3.212s bwd=1.217s opt=0.003s] dispatch=3.573s [2026-10-19 05:29:20,174] loop.py:45 INFO TRAIN_STEP step=7345 loss=1.425939 Training: 41%|████ | 7345/18018 [15:59:06<23:00:14, 7.76s/step, loss=1.426, acc=0.684, acc_len=1.63, thru=15.4, I=21.0, T=28.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:20,628] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:24,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 05:29:27,074] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7345: forward=2973.3ms backward=1226.2ms [2026-10-19 05:29:27,079] loop.py:45 INFO TRAIN_STEP step=7346 loss=1.217982 Training: 41%|████ | 7346/18018 [15:59:13<22:14:27, 7.50s/step, loss=1.218, acc=0.715, acc_len=1.74, thru=14.7, I=28.0, T=30.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:27,449] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:35,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:29:36,537] loop.py:45 INFO TRAIN_STEP step=7347 loss=1.451068 Training: 41%|████ | 7347/18018 [15:59:22<23:58:40, 8.09s/step, loss=1.451, acc=0.679, acc_len=1.61, thru=14.1, I=16.2, T=29.4, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:36,704] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:41,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:29:43,731] loop.py:45 INFO TRAIN_STEP step=7348 loss=1.496056 Training: 41%|████ | 7348/18018 [15:59:29<23:10:47, 7.82s/step, loss=1.496, acc=0.665, acc_len=1.56, thru=16.3, I=22.6, T=33.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:43,841] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 [2026-10-19 05:29:49,622] loop.py:45 INFO TRAIN_STEP step=7349 loss=1.415162 Training: 41%|████ | 7349/18018 [15:59:35<21:27:47, 7.24s/step, loss=1.415, acc=0.683, acc_len=1.63, thru=18.5, I=27.8, T=33.6, wait=2.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:56,906] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=68/64 [2026-10-19 05:29:57,627] loop.py:380 INFO TIMING step=7350: step=4.099s data=1.846s compute=4.068s [fwd=2.833s bwd=1.222s opt=0.003s] dispatch=3.897s [2026-10-19 05:29:57,627] loop.py:45 INFO TRAIN_STEP step=7350 loss=1.395257 Training: 41%|████ | 7350/18018 [15:59:43<22:08:19, 7.47s/step, loss=1.395, acc=0.685, acc_len=1.64, thru=16.9, I=17.1, T=31.2, wait=3.9s, pool=68, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:29:57,911] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=4/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:02,639] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:30:04,979] loop.py:45 INFO TRAIN_STEP step=7351 loss=1.498539 Training: 41%|████ | 7351/18018 [15:59:51<22:01:51, 7.44s/step, loss=1.499, acc=0.668, acc_len=1.58, thru=16.4, I=23.6, T=26.6, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:30:04,974] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7350: forward=3266.3ms backward=1495.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:05,157] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:09,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:30:10,643] loop.py:45 INFO TRAIN_STEP step=7352 loss=1.414528 Training: 41%|████ | 7352/18018 [15:59:56<20:27:21, 6.90s/step, loss=1.415, acc=0.682, acc_len=1.62, thru=18.4, I=28.2, T=37.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:10,879] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:15,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:30:17,693] loop.py:45 INFO TRAIN_STEP step=7353 loss=1.348619 Training: 41%|████ | 7353/18018 [16:00:03<20:34:52, 6.95s/step, loss=1.349, acc=0.693, acc_len=1.65, thru=18.3, I=27.1, T=27.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:17,741] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:23,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:30:24,027] loop.py:45 INFO TRAIN_STEP step=7354 loss=1.489505 Training: 41%|████ | 7354/18018 [16:00:10<20:02:09, 6.76s/step, loss=1.490, acc=0.664, acc_len=1.57, thru=17.0, I=22.3, T=31.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:24,117] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:29,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:30:30,787] loop.py:380 INFO TIMING step=7355: step=4.126s data=1.860s compute=4.096s [fwd=2.824s bwd=1.258s opt=0.003s] dispatch=2.623s [2026-10-19 05:30:30,787] loop.py:45 INFO TRAIN_STEP step=7355 loss=1.469510 Training: 41%|████ | 7355/18018 [16:00:16<20:01:40, 6.76s/step, loss=1.470, acc=0.675, acc_len=1.60, thru=18.7, I=20.7, T=31.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:31,290] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:37,341] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:30:38,700] loop.py:45 INFO TRAIN_STEP step=7356 loss=1.383275 Training: 41%|████ | 7356/18018 [16:00:24<21:03:07, 7.11s/step, loss=1.383, acc=0.688, acc_len=1.65, thru=14.5, I=20.1, T=27.2, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:30:38,696] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7355: forward=3325.0ms backward=1340.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:30:38,864] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:30:44,435] loop.py:45 INFO TRAIN_STEP step=7357 loss=1.502149 Training: 41%|████ | 7357/18018 [16:00:30<19:49:39, 6.70s/step, loss=1.502, acc=0.669, acc_len=1.58, thru=17.7, I=19.8, T=34.9, wait=2.1s, pool=56, epoch=1/2][2026-10-19 05:30:51,050] loop.py:45 INFO TRAIN_STEP step=7358 loss=1.461811 Training: 41%|████ | 7358/18018 [16:00:37<19:45:18, 6.67s/step, loss=1.462, acc=0.675, acc_len=1.59, thru=17.6, I=19.6, T=33.3, wait=2.8s, pool=48, epoch=1/2][2026-10-19 05:30:59,304] loop.py:45 INFO TRAIN_STEP step=7359 loss=1.497465 Training: 41%|████ | 7359/18018 [16:00:45<21:09:32, 7.15s/step, loss=1.497, acc=0.667, acc_len=1.57, thru=15.4, I=15.7, T=30.8, wait=4.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:06,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:31:07,760] loop.py:380 INFO TIMING step=7360: step=4.341s data=1.947s compute=4.301s [fwd=3.032s bwd=1.255s opt=0.003s] dispatch=4.107s [2026-10-19 05:31:07,760] loop.py:45 INFO TRAIN_STEP step=7360 loss=1.515778 Training: 41%|████ | 7360/18018 [16:00:53<22:19:17, 7.54s/step, loss=1.516, acc=0.665, acc_len=1.56, thru=16.5, I=16.8, T=29.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:07,884] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:15,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:31:16,506] loop.py:45 INFO TRAIN_STEP step=7361 loss=1.537614 Training: 41%|████ | 7361/18018 [16:01:02<23:23:25, 7.90s/step, loss=1.538, acc=0.663, acc_len=1.56, thru=13.7, I=15.8, T=33.4, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:31:16,501] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7360: forward=2721.9ms backward=1069.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:16,686] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:31:23,485] loop.py:45 INFO TRAIN_STEP step=7362 loss=1.371881 Training: 41%|████ | 7362/18018 [16:01:09<22:34:00, 7.62s/step, loss=1.372, acc=0.689, acc_len=1.65, thru=14.5, I=17.5, T=32.9, wait=3.1s, pool=40, epoch=1/2][2026-10-19 05:31:31,629] loop.py:45 INFO TRAIN_STEP step=7363 loss=1.428792 Training: 41%|████ | 7363/18018 [16:01:17<23:01:40, 7.78s/step, loss=1.429, acc=0.683, acc_len=1.62, thru=16.4, I=15.0, T=34.2, wait=4.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:39,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:31:40,923] loop.py:45 INFO TRAIN_STEP step=7364 loss=1.443426 Training: 41%|████ | 7364/18018 [16:01:27<24:22:14, 8.23s/step, loss=1.443, acc=0.679, acc_len=1.62, thru=15.6, I=15.5, T=32.0, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:41,144] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:46,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:31:47,662] loop.py:380 INFO TIMING step=7365: step=4.576s data=1.707s compute=4.546s [fwd=3.411s bwd=1.120s opt=0.003s] dispatch=2.154s [2026-10-19 05:31:47,662] loop.py:45 INFO TRAIN_STEP step=7365 loss=1.274334 Training: 41%|████ | 7365/18018 [16:01:33<23:02:19, 7.79s/step, loss=1.274, acc=0.704, acc_len=1.69, thru=18.0, I=21.7, T=28.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:48,162] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:54,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:31:54,883] loop.py:45 INFO TRAIN_STEP step=7366 loss=1.502727 Training: 41%|████ | 7366/18018 [16:01:40<22:32:08, 7.62s/step, loss=1.503, acc=0.666, acc_len=1.57, thru=16.5, I=18.5, T=29.6, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:31:54,879] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7365: forward=3078.1ms backward=1199.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:31:55,325] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:32:03,485] loop.py:45 INFO TRAIN_STEP step=7367 loss=1.410091 Training: 41%|████ | 7367/18018 [16:01:49<23:24:44, 7.91s/step, loss=1.410, acc=0.682, acc_len=1.63, thru=15.5, I=17.4, T=27.8, wait=4.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:09,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:32:11,090] loop.py:45 INFO TRAIN_STEP step=7368 loss=1.494939 Training: 41%|████ | 7368/18018 [16:01:57<23:08:02, 7.82s/step, loss=1.495, acc=0.669, acc_len=1.58, thru=15.7, I=16.3, T=29.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:11,356] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:32:17,739] loop.py:45 INFO TRAIN_STEP step=7369 loss=1.444835 Training: 41%|████ | 7369/18018 [16:02:03<22:05:34, 7.47s/step, loss=1.445, acc=0.673, acc_len=1.60, thru=17.3, I=21.9, T=34.4, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:26,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:32:27,492] loop.py:380 INFO TIMING step=7370: step=4.285s data=1.951s compute=4.259s [fwd=3.099s bwd=1.146s opt=0.003s] dispatch=5.456s [2026-10-19 05:32:27,492] loop.py:45 INFO TRAIN_STEP step=7370 loss=1.576817 Training: 41%|████ | 7370/18018 [16:02:13<24:06:59, 8.15s/step, loss=1.577, acc=0.654, acc_len=1.53, thru=14.0, I=13.8, T=29.9, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:27,920] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:33,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:32:34,896] loop.py:45 INFO TRAIN_STEP step=7371 loss=1.535686 Training: 41%|████ | 7371/18018 [16:02:20<23:27:02, 7.93s/step, loss=1.536, acc=0.662, acc_len=1.55, thru=16.3, I=20.3, T=30.2, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:32:34,891] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7370: forward=2899.9ms backward=1269.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:35,267] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:41,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:32:42,425] loop.py:45 INFO TRAIN_STEP step=7372 loss=1.353326 Training: 41%|████ | 7372/18018 [16:02:28<23:05:35, 7.81s/step, loss=1.353, acc=0.693, acc_len=1.66, thru=14.7, I=20.2, T=31.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:42,880] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:49,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:32:51,190] loop.py:45 INFO TRAIN_STEP step=7373 loss=1.385061 Training: 41%|████ | 7373/18018 [16:02:37<23:56:17, 8.10s/step, loss=1.385, acc=0.687, acc_len=1.64, thru=13.6, I=20.2, T=29.1, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:51,339] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:57,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:32:57,902] loop.py:45 INFO TRAIN_STEP step=7374 loss=1.489156 Training: 41%|████ | 7374/18018 [16:02:43<22:42:28, 7.68s/step, loss=1.489, acc=0.676, acc_len=1.60, thru=17.7, I=22.4, T=35.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:32:58,247] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:03,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:33:04,676] loop.py:380 INFO TIMING step=7375: step=3.869s data=1.727s compute=3.838s [fwd=2.698s bwd=1.126s opt=0.003s] dispatch=2.895s [2026-10-19 05:33:04,676] loop.py:45 INFO TRAIN_STEP step=7375 loss=1.486698 Training: 41%|████ | 7375/18018 [16:02:50<21:54:12, 7.41s/step, loss=1.487, acc=0.672, acc_len=1.59, thru=18.5, I=24.0, T=33.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:04,749] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:10,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:33:12,067] loop.py:45 INFO TRAIN_STEP step=7376 loss=1.360684 Training: 41%|████ | 7376/18018 [16:02:58<21:53:07, 7.40s/step, loss=1.361, acc=0.690, acc_len=1.65, thru=16.7, I=21.3, T=31.7, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:33:12,062] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7375: forward=2799.7ms backward=1191.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:12,250] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:17,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:33:19,256] loop.py:45 INFO TRAIN_STEP step=7377 loss=1.390699 Training: 41%|████ | 7377/18018 [16:03:05<21:41:30, 7.34s/step, loss=1.391, acc=0.686, acc_len=1.63, thru=16.1, I=20.9, T=30.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:19,658] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:27,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:33:27,349] loop.py:45 INFO TRAIN_STEP step=7378 loss=1.554459 Training: 41%|████ | 7378/18018 [16:03:13<22:21:29, 7.56s/step, loss=1.554, acc=0.659, acc_len=1.54, thru=13.8, I=15.9, T=33.6, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:27,705] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:33:35,968] loop.py:45 INFO TRAIN_STEP step=7379 loss=1.412099 Training: 41%|████ | 7379/18018 [16:03:22<23:17:28, 7.88s/step, loss=1.412, acc=0.679, acc_len=1.62, thru=14.9, I=14.8, T=29.5, wait=4.3s, pool=60, epoch=1/2][2026-10-19 05:33:43,422] loop.py:380 INFO TIMING step=7380: step=3.837s data=1.697s compute=3.810s [fwd=2.748s bwd=1.048s opt=0.003s] dispatch=3.608s [2026-10-19 05:33:43,422] loop.py:45 INFO TRAIN_STEP step=7380 loss=1.504139 Training: 41%|████ | 7380/18018 [16:03:29<22:54:46, 7.75s/step, loss=1.504, acc=0.671, acc_len=1.58, thru=15.7, I=14.4, T=33.4, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:49,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:33:50,860] loop.py:45 INFO TRAIN_STEP step=7381 loss=1.439309 Training: 41%|████ | 7381/18018 [16:03:36<22:37:47, 7.66s/step, loss=1.439, acc=0.676, acc_len=1.60, thru=18.5, I=20.5, T=29.8, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:33:50,856] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7380: forward=2933.2ms backward=1307.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:51,267] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:57,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:33:58,890] loop.py:45 INFO TRAIN_STEP step=7382 loss=1.476870 Training: 41%|████ | 7382/18018 [16:03:44<22:57:16, 7.77s/step, loss=1.477, acc=0.670, acc_len=1.58, thru=14.6, I=21.3, T=26.6, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:33:59,193] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:05,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:34:07,234] loop.py:45 INFO TRAIN_STEP step=7383 loss=1.441349 Training: 41%|████ | 7383/18018 [16:03:53<23:27:46, 7.94s/step, loss=1.441, acc=0.682, acc_len=1.62, thru=13.9, I=17.0, T=26.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:07,268] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:11,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:34:13,836] loop.py:45 INFO TRAIN_STEP step=7384 loss=1.406435 Training: 41%|████ | 7384/18018 [16:03:59<22:16:25, 7.54s/step, loss=1.406, acc=0.685, acc_len=1.64, thru=15.7, I=25.0, T=32.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:14,219] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:18,868] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:34:20,201] loop.py:380 INFO TIMING step=7385: step=4.040s data=1.877s compute=4.016s [fwd=2.762s bwd=1.239s opt=0.003s] dispatch=2.315s [2026-10-19 05:34:20,202] loop.py:45 INFO TRAIN_STEP step=7385 loss=1.466367 Training: 41%|████ | 7385/18018 [16:04:06<21:13:47, 7.19s/step, loss=1.466, acc=0.672, acc_len=1.59, thru=18.1, I=28.8, T=31.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:20,394] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:25,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:34:26,170] loop.py:45 INFO TRAIN_STEP step=7386 loss=1.446779 Training: 41%|████ | 7386/18018 [16:04:12<20:08:48, 6.82s/step, loss=1.447, acc=0.678, acc_len=1.61, thru=19.4, I=23.6, T=36.5, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:34:26,165] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7385: forward=2445.5ms backward=1010.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:26,566] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:31,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:34:33,057] loop.py:45 INFO TRAIN_STEP step=7387 loss=1.514914 Training: 41%|████ | 7387/18018 [16:04:19<20:12:06, 6.84s/step, loss=1.515, acc=0.667, acc_len=1.57, thru=18.9, I=25.2, T=28.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:34:33,561] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 05:34:41,313] loop.py:45 INFO TRAIN_STEP step=7388 loss=1.491976 Training: 41%|████ | 7388/18018 [16:04:27<21:27:12, 7.27s/step, loss=1.492, acc=0.670, acc_len=1.58, thru=16.0, I=17.4, T=29.8, wait=4.0s, pool=60, epoch=1/2][2026-10-19 05:34:49,715] loop.py:45 INFO TRAIN_STEP step=7389 loss=1.354894 Training: 41%|████ | 7389/18018 [16:04:35<22:27:31, 7.61s/step, loss=1.355, acc=0.694, acc_len=1.67, thru=14.7, I=12.5, T=30.6, wait=4.2s, pool=48, epoch=1/2][2026-10-19 05:34:59,242] loop.py:380 INFO TIMING step=7390: step=4.286s data=1.979s compute=4.239s [fwd=2.979s bwd=1.244s opt=0.003s] dispatch=5.231s [2026-10-19 05:34:59,242] loop.py:45 INFO TRAIN_STEP step=7390 loss=1.476507 Training: 41%|████ | 7390/18018 [16:04:45<24:09:31, 8.18s/step, loss=1.477, acc=0.674, acc_len=1.60, thru=15.6, I=16.0, T=29.9, wait=5.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:05,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:35:06,912] loop.py:45 INFO TRAIN_STEP step=7391 loss=1.570912 Training: 41%|████ | 7391/18018 [16:04:52<23:42:00, 8.03s/step, loss=1.571, acc=0.655, acc_len=1.53, thru=16.4, I=17.5, T=29.2, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:35:06,907] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7390: forward=3087.1ms backward=1234.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:07,229] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:12,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:35:13,165] loop.py:45 INFO TRAIN_STEP step=7392 loss=1.278448 Training: 41%|████ | 7392/18018 [16:04:59<22:07:34, 7.50s/step, loss=1.278, acc=0.705, acc_len=1.70, thru=18.1, I=21.1, T=33.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:13,600] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:19,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:35:20,919] loop.py:45 INFO TRAIN_STEP step=7393 loss=1.417676 Training: 41%|████ | 7393/18018 [16:05:07<22:21:12, 7.57s/step, loss=1.418, acc=0.682, acc_len=1.62, thru=17.6, I=19.8, T=29.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:20,952] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:27,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:35:27,848] loop.py:45 INFO TRAIN_STEP step=7394 loss=1.473550 Training: 41%|████ | 7394/18018 [16:05:13<21:46:42, 7.38s/step, loss=1.474, acc=0.672, acc_len=1.59, thru=16.8, I=17.8, T=29.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:28,236] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:35:37,286] loop.py:380 INFO TIMING step=7395: step=4.130s data=1.888s compute=4.078s [fwd=2.823s bwd=1.240s opt=0.003s] dispatch=5.299s [2026-10-19 05:35:37,286] loop.py:45 INFO TRAIN_STEP step=7395 loss=1.501597 Training: 41%|████ | 7395/18018 [16:05:23<23:35:52, 8.00s/step, loss=1.502, acc=0.672, acc_len=1.59, thru=13.7, I=13.9, T=31.0, wait=5.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:44,647] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:35:44,769] loop.py:45 INFO TRAIN_STEP step=7396 loss=1.526989 Training: 41%|████ | 7396/18018 [16:05:30<23:08:26, 7.84s/step, loss=1.527, acc=0.664, acc_len=1.57, thru=18.4, I=17.4, T=33.1, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:35:44,765] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7395: forward=2703.7ms backward=1104.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:45,150] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:51,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:35:53,892] loop.py:45 INFO TRAIN_STEP step=7397 loss=1.405160 Training: 41%|████ | 7397/18018 [16:05:39<24:16:26, 8.23s/step, loss=1.405, acc=0.684, acc_len=1.63, thru=14.2, I=17.9, T=24.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:35:54,087] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 05:36:00,713] loop.py:45 INFO TRAIN_STEP step=7398 loss=1.521239 Training: 41%|████ | 7398/18018 [16:05:46<23:01:39, 7.81s/step, loss=1.521, acc=0.664, acc_len=1.57, thru=13.8, I=15.6, T=32.4, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:10,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:36:12,346] loop.py:45 INFO TRAIN_STEP step=7399 loss=1.500777 Training: 41%|████ | 7399/18018 [16:05:58<26:24:34, 8.95s/step, loss=1.501, acc=0.670, acc_len=1.58, thru=12.1, I=13.2, T=25.7, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:12,797] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:18,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:36:19,705] loop.py:380 INFO TIMING step=7400: step=4.419s data=2.291s compute=4.387s [fwd=3.323s bwd=1.050s opt=0.003s] dispatch=2.931s [2026-10-19 05:36:19,706] loop.py:45 INFO TRAIN_STEP step=7400 loss=1.304776 Training: 41%|████ | 7400/18018 [16:06:05<24:59:42, 8.47s/step, loss=1.305, acc=0.707, acc_len=1.71, thru=14.1, I=20.2, T=29.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:20,162] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:25,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:36:27,675] loop.py:45 INFO TRAIN_STEP step=7401 loss=1.522605 Training: 41%|████ | 7401/18018 [16:06:13<24:32:52, 8.32s/step, loss=1.523, acc=0.663, acc_len=1.56, thru=14.6, I=23.1, T=28.0, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:36:27,670] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7400: forward=3426.9ms backward=1088.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:27,944] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 05:36:33,804] loop.py:45 INFO TRAIN_STEP step=7402 loss=1.475676 Training: 41%|████ | 7402/18018 [16:06:19<22:36:11, 7.67s/step, loss=1.476, acc=0.671, acc_len=1.58, thru=15.4, I=19.8, T=37.4, wait=2.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:44,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:36:45,229] loop.py:45 INFO TRAIN_STEP step=7403 loss=1.388252 Training: 41%|████ | 7403/18018 [16:06:31<25:55:42, 8.79s/step, loss=1.388, acc=0.682, acc_len=1.63, thru=14.1, I=13.3, T=32.0, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:45,490] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:51,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:36:52,377] loop.py:45 INFO TRAIN_STEP step=7404 loss=1.555598 Training: 41%|████ | 7404/18018 [16:06:38<24:28:14, 8.30s/step, loss=1.556, acc=0.661, acc_len=1.56, thru=17.3, I=19.8, T=34.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:52,669] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:58,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:36:58,852] loop.py:380 INFO TIMING step=7405: step=3.888s data=1.783s compute=3.856s [fwd=2.707s bwd=1.135s opt=0.003s] dispatch=2.578s [2026-10-19 05:36:58,853] loop.py:45 INFO TRAIN_STEP step=7405 loss=1.402631 Training: 41%|████ | 7405/18018 [16:06:44<22:51:14, 7.75s/step, loss=1.403, acc=0.683, acc_len=1.62, thru=19.6, I=23.4, T=32.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:36:58,986] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-19 05:37:07,827] loop.py:45 INFO TRAIN_STEP step=7406 loss=1.417985 Training: 41%|████ | 7406/18018 [16:06:53<23:55:53, 8.12s/step, loss=1.418, acc=0.679, acc_len=1.62, thru=13.1, I=14.1, T=31.5, wait=4.9s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:37:07,822] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7405: forward=2898.9ms backward=1109.1ms [2026-10-19 05:37:14,533] loop.py:45 INFO TRAIN_STEP step=7407 loss=1.519893 Training: 41%|████ | 7407/18018 [16:07:00<22:41:01, 7.70s/step, loss=1.520, acc=0.669, acc_len=1.58, thru=18.5, I=16.3, T=35.5, wait=3.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:37:22,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:37:23,977] loop.py:45 INFO TRAIN_STEP step=7408 loss=1.358507 Training: 41%|████ | 7408/18018 [16:07:10<24:13:29, 8.22s/step, loss=1.359, acc=0.690, acc_len=1.65, thru=15.0, I=16.9, T=27.3, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:37:24,313] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:37:29,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:37:31,295] loop.py:45 INFO TRAIN_STEP step=7409 loss=1.494748 Training: 41%|████ | 7409/18018 [16:07:17<23:25:37, 7.95s/step, loss=1.495, acc=0.664, acc_len=1.57, thru=16.5, I=23.1, T=28.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:37:31,644] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:37:37,978] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:37:38,476] loop.py:380 INFO TIMING step=7410: step=4.015s data=1.823s compute=3.975s [fwd=2.853s bwd=1.107s opt=0.003s] dispatch=3.155s [2026-10-19 05:37:38,476] loop.py:45 INFO TRAIN_STEP step=7410 loss=1.498120 Training: 41%|████ | 7410/18018 [16:07:24<22:44:41, 7.72s/step, loss=1.498, acc=0.668, acc_len=1.57, thru=14.9, I=20.5, T=31.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:37:38,983] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:37:44,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:37:45,102] loop.py:45 INFO TRAIN_STEP step=7411 loss=1.379258 Training: 41%|████ | 7411/18018 [16:07:31<21:46:37, 7.39s/step, loss=1.379, acc=0.687, acc_len=1.64, thru=18.9, I=22.1, T=30.4, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:37:45,098] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7410: forward=2870.1ms backward=1287.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:37:45,276] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:37:54,093] loop.py:45 INFO TRAIN_STEP step=7412 loss=1.359687 Training: 41%|████ | 7412/18018 [16:07:40<23:11:11, 7.87s/step, loss=1.360, acc=0.691, acc_len=1.66, thru=13.6, I=13.6, T=29.6, wait=4.7s, pool=56, epoch=1/2][2026-10-19 05:38:00,456] loop.py:45 INFO TRAIN_STEP step=7413 loss=1.610889 Training: 41%|████ | 7413/18018 [16:07:46<21:51:15, 7.42s/step, loss=1.611, acc=0.649, acc_len=1.51, thru=16.6, I=15.4, T=34.6, wait=2.7s, pool=44, epoch=1/2][2026-10-19 05:38:10,451] loop.py:45 INFO TRAIN_STEP step=7414 loss=1.376866 Training: 41%|████ | 7414/18018 [16:07:56<24:07:42, 8.19s/step, loss=1.377, acc=0.690, acc_len=1.65, thru=12.2, I=12.5, T=31.6, wait=5.9s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:38:19,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:38:20,773] loop.py:380 INFO TIMING step=7415: step=4.418s data=1.984s compute=4.376s [fwd=3.137s bwd=1.224s opt=0.003s] dispatch=5.894s [2026-10-19 05:38:20,773] loop.py:45 INFO TRAIN_STEP step=7415 loss=1.488589 Training: 41%|████ | 7415/18018 [16:08:06<26:00:32, 8.83s/step, loss=1.489, acc=0.672, acc_len=1.59, thru=15.2, I=16.8, T=29.0, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:38:21,035] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:38:29,271] loop.py:45 INFO TRAIN_STEP step=7416 loss=1.430604 Training: 41%|████ | 7416/18018 [16:08:15<25:42:41, 8.73s/step, loss=1.431, acc=0.684, acc_len=1.63, thru=14.1, I=14.0, T=34.8, wait=4.8s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:38:29,265] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7415: forward=2619.1ms backward=1007.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:38:37,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:38:38,971] loop.py:45 INFO TRAIN_STEP step=7417 loss=1.256070 Training: 41%|████ | 7417/18018 [16:08:25<26:33:57, 9.02s/step, loss=1.256, acc=0.715, acc_len=1.73, thru=15.3, I=17.3, T=26.4, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:38:39,314] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:38:45,956] loop.py:45 INFO TRAIN_STEP step=7418 loss=1.464260 Training: 41%|████ | 7418/18018 [16:08:32<24:45:55, 8.41s/step, loss=1.464, acc=0.673, acc_len=1.60, thru=13.6, I=16.4, T=34.1, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:38:56,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:38:58,942] loop.py:45 INFO TRAIN_STEP step=7419 loss=1.433630 Training: 41%|████ | 7419/18018 [16:08:45<28:48:15, 9.78s/step, loss=1.434, acc=0.681, acc_len=1.61, thru=12.8, I=12.8, T=26.3, wait=8.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:38:58,975] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:04,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:39:05,012] loop.py:380 INFO TIMING step=7420: step=3.602s data=1.553s compute=3.571s [fwd=2.504s bwd=1.053s opt=0.003s] dispatch=2.458s [2026-10-19 05:39:05,012] loop.py:45 INFO TRAIN_STEP step=7420 loss=1.410299 Training: 41%|████ | 7420/18018 [16:08:51<25:31:18, 8.67s/step, loss=1.410, acc=0.683, acc_len=1.63, thru=17.0, I=24.8, T=35.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:05,327] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:11,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:39:13,178] loop.py:45 INFO TRAIN_STEP step=7421 loss=1.287551 Training: 41%|████ | 7421/18018 [16:08:59<25:04:25, 8.52s/step, loss=1.288, acc=0.704, acc_len=1.70, thru=16.2, I=22.3, T=30.8, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:39:13,173] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7420: forward=2904.8ms backward=1204.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:13,338] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:18,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:39:19,377] loop.py:45 INFO TRAIN_STEP step=7422 loss=1.412475 Training: 41%|████ | 7422/18018 [16:09:05<23:01:24, 7.82s/step, loss=1.412, acc=0.682, acc_len=1.62, thru=18.9, I=26.0, T=35.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:19,436] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 05:39:26,632] loop.py:45 INFO TRAIN_STEP step=7423 loss=1.391262 Training: 41%|████ | 7423/18018 [16:09:12<22:31:12, 7.65s/step, loss=1.391, acc=0.684, acc_len=1.63, thru=16.2, I=20.5, T=29.4, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:33,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:39:35,536] loop.py:45 INFO TRAIN_STEP step=7424 loss=1.389239 Training: 41%|████ | 7424/18018 [16:09:21<23:37:25, 8.03s/step, loss=1.389, acc=0.686, acc_len=1.64, thru=15.4, I=16.5, T=29.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:35,706] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:40,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:39:41,593] loop.py:380 INFO TIMING step=7425: step=3.688s data=1.565s compute=3.656s [fwd=2.553s bwd=1.088s opt=0.003s] dispatch=2.361s [2026-10-19 05:39:41,593] loop.py:45 INFO TRAIN_STEP step=7425 loss=1.333818 Training: 41%|████ | 7425/18018 [16:09:27<21:53:00, 7.44s/step, loss=1.334, acc=0.699, acc_len=1.67, thru=18.7, I=25.4, T=34.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:41,693] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:48,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:39:48,498] loop.py:45 INFO TRAIN_STEP step=7426 loss=1.400592 Training: 41%|████ | 7426/18018 [16:09:34<21:24:35, 7.28s/step, loss=1.401, acc=0.686, acc_len=1.63, thru=18.2, I=20.7, T=34.5, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:39:48,494] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7425: forward=2675.7ms backward=990.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:48,823] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:54,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:39:55,526] loop.py:45 INFO TRAIN_STEP step=7427 loss=1.371620 Training: 41%|████ | 7427/18018 [16:09:41<21:11:19, 7.20s/step, loss=1.372, acc=0.695, acc_len=1.66, thru=20.2, I=19.2, T=33.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:39:55,717] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:00,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:40:02,150] loop.py:45 INFO TRAIN_STEP step=7428 loss=1.412100 Training: 41%|████ | 7428/18018 [16:09:48<20:40:37, 7.03s/step, loss=1.412, acc=0.684, acc_len=1.63, thru=19.2, I=22.5, T=29.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:02,444] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:08,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:40:09,687] loop.py:45 INFO TRAIN_STEP step=7429 loss=1.498279 Training: 41%|████ | 7429/18018 [16:09:55<21:07:19, 7.18s/step, loss=1.498, acc=0.669, acc_len=1.58, thru=14.4, I=19.0, T=27.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:09,912] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:14,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:40:15,952] loop.py:380 INFO TIMING step=7430: step=4.218s data=1.938s compute=4.183s [fwd=2.865s bwd=1.302s opt=0.003s] dispatch=2.039s [2026-10-19 05:40:15,953] loop.py:45 INFO TRAIN_STEP step=7430 loss=1.540616 Training: 41%|████ | 7430/18018 [16:10:02<20:18:44, 6.91s/step, loss=1.541, acc=0.656, acc_len=1.53, thru=17.9, I=22.6, T=30.3, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:16,313] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:40:23,443] loop.py:45 INFO TRAIN_STEP step=7431 loss=1.430510 Training: 41%|████ | 7431/18018 [16:10:09<20:49:30, 7.08s/step, loss=1.431, acc=0.680, acc_len=1.62, thru=12.9, I=13.8, T=31.5, wait=3.4s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:40:23,438] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7430: forward=2896.5ms backward=1115.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:32,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:40:33,075] loop.py:45 INFO TRAIN_STEP step=7432 loss=1.426534 Training: 41%|████ | 7432/18018 [16:10:19<23:04:24, 7.85s/step, loss=1.427, acc=0.683, acc_len=1.63, thru=16.1, I=16.9, T=32.2, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:33,572] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:40,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:40:40,282] loop.py:45 INFO TRAIN_STEP step=7433 loss=1.358132 Training: 41%|████▏ | 7433/18018 [16:10:26<22:30:25, 7.65s/step, loss=1.358, acc=0.694, acc_len=1.67, thru=17.7, I=22.3, T=32.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:40,644] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:47,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:40:48,348] loop.py:45 INFO TRAIN_STEP step=7434 loss=1.431300 Training: 41%|████▏ | 7434/18018 [16:10:34<22:52:12, 7.78s/step, loss=1.431, acc=0.677, acc_len=1.61, thru=15.3, I=17.3, T=31.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:40:48,635] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:40:55,315] loop.py:380 INFO TIMING step=7435: step=4.502s data=2.233s compute=4.474s [fwd=3.101s bwd=1.358s opt=0.003s] dispatch=2.455s [2026-10-19 05:40:55,315] loop.py:45 INFO TRAIN_STEP step=7435 loss=1.497001 Training: 41%|████▏ | 7435/18018 [16:10:41<22:09:12, 7.54s/step, loss=1.497, acc=0.668, acc_len=1.57, thru=16.6, I=20.1, T=28.4, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:01,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:41:03,183] loop.py:45 INFO TRAIN_STEP step=7436 loss=1.429939 Training: 41%|████▏ | 7436/18018 [16:10:49<22:26:23, 7.63s/step, loss=1.430, acc=0.679, acc_len=1.62, thru=16.1, I=19.9, T=30.7, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:41:03,178] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7435: forward=2899.2ms backward=1204.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:03,681] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:10,517] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:41:10,547] loop.py:45 INFO TRAIN_STEP step=7437 loss=1.449054 Training: 41%|████▏ | 7437/18018 [16:10:56<22:12:00, 7.55s/step, loss=1.449, acc=0.673, acc_len=1.58, thru=14.4, I=19.5, T=31.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:11,019] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:18,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:41:20,782] loop.py:45 INFO TRAIN_STEP step=7438 loss=1.411446 Training: 41%|████▏ | 7438/18018 [16:11:06<24:33:48, 8.36s/step, loss=1.411, acc=0.686, acc_len=1.63, thru=13.1, I=16.0, T=26.5, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:20,812] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 05:41:27,344] loop.py:45 INFO TRAIN_STEP step=7439 loss=1.415399 Training: 41%|████▏ | 7439/18018 [16:11:13<22:58:41, 7.82s/step, loss=1.415, acc=0.679, acc_len=1.61, thru=15.7, I=20.7, T=31.7, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:35,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:41:37,217] loop.py:380 INFO TIMING step=7440: step=4.263s data=1.900s compute=4.199s [fwd=3.005s bwd=1.178s opt=0.003s] dispatch=5.601s [2026-10-19 05:41:37,218] loop.py:45 INFO TRAIN_STEP step=7440 loss=1.356588 Training: 41%|████▏ | 7440/18018 [16:11:23<24:47:07, 8.44s/step, loss=1.357, acc=0.696, acc_len=1.67, thru=14.4, I=15.6, T=30.0, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:37,688] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:43,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:41:45,215] loop.py:45 INFO TRAIN_STEP step=7441 loss=1.392129 Training: 41%|████▏ | 7441/18018 [16:11:31<24:23:49, 8.30s/step, loss=1.392, acc=0.688, acc_len=1.64, thru=14.1, I=20.9, T=32.1, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:41:45,211] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7440: forward=2771.6ms backward=1172.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:45,705] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:41:52,942] loop.py:45 INFO TRAIN_STEP step=7442 loss=1.440053 Training: 41%|████▏ | 7442/18018 [16:11:39<23:53:13, 8.13s/step, loss=1.440, acc=0.677, acc_len=1.60, thru=14.0, I=18.4, T=33.9, wait=3.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:41:59,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:42:00,017] loop.py:45 INFO TRAIN_STEP step=7443 loss=1.576803 Training: 41%|████▏ | 7443/18018 [16:11:46<22:57:15, 7.81s/step, loss=1.577, acc=0.654, acc_len=1.53, thru=18.0, I=18.5, T=34.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:00,113] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:04,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:42:06,525] loop.py:45 INFO TRAIN_STEP step=7444 loss=1.478932 Training: 41%|████▏ | 7444/18018 [16:11:52<21:48:09, 7.42s/step, loss=1.479, acc=0.673, acc_len=1.59, thru=19.6, I=24.6, T=30.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:06,694] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:12,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:42:13,422] loop.py:380 INFO TIMING step=7445: step=3.854s data=1.705s compute=3.824s [fwd=2.722s bwd=1.087s opt=0.003s] dispatch=3.033s [2026-10-19 05:42:13,422] loop.py:45 INFO TRAIN_STEP step=7445 loss=1.463524 Training: 41%|████▏ | 7445/18018 [16:11:59<21:20:07, 7.26s/step, loss=1.464, acc=0.671, acc_len=1.58, thru=15.2, I=20.6, T=33.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:13,856] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:19,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:42:19,997] loop.py:45 INFO TRAIN_STEP step=7446 loss=1.515086 Training: 41%|████▏ | 7446/18018 [16:12:06<20:43:35, 7.06s/step, loss=1.515, acc=0.665, acc_len=1.56, thru=18.9, I=21.6, T=30.5, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:42:19,993] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7445: forward=2939.9ms backward=1213.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:20,341] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:26,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:42:28,403] loop.py:45 INFO TRAIN_STEP step=7447 loss=1.537253 Training: 41%|████▏ | 7447/18018 [16:12:14<21:54:46, 7.46s/step, loss=1.537, acc=0.661, acc_len=1.56, thru=15.5, I=16.8, T=30.5, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:28,714] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:33,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:42:35,280] loop.py:45 INFO TRAIN_STEP step=7448 loss=1.494339 Training: 41%|████▏ | 7448/18018 [16:12:21<21:23:44, 7.29s/step, loss=1.494, acc=0.666, acc_len=1.58, thru=16.0, I=23.1, T=31.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:35,425] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-19 05:42:41,194] loop.py:45 INFO TRAIN_STEP step=7449 loss=1.461084 Training: 41%|████▏ | 7449/18018 [16:12:27<20:10:57, 6.87s/step, loss=1.461, acc=0.673, acc_len=1.59, thru=19.2, I=22.2, T=35.4, wait=2.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:48,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:42:49,482] loop.py:380 INFO TIMING step=7450: step=4.633s data=2.314s compute=4.602s [fwd=3.292s bwd=1.295s opt=0.003s] dispatch=3.646s [2026-10-19 05:42:49,482] loop.py:45 INFO TRAIN_STEP step=7450 loss=1.573032 Training: 41%|████▏ | 7450/18018 [16:12:35<21:25:31, 7.30s/step, loss=1.573, acc=0.654, acc_len=1.53, thru=17.1, I=18.2, T=27.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:42:49,575] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-19 05:42:55,924] loop.py:45 INFO TRAIN_STEP step=7451 loss=1.426203 Training: 41%|████▏ | 7451/18018 [16:12:42<20:40:13, 7.04s/step, loss=1.426, acc=0.676, acc_len=1.61, thru=16.1, I=21.5, T=31.1, wait=2.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:42:55,919] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7450: forward=2896.0ms backward=1172.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:02,165] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:43:03,420] loop.py:45 INFO TRAIN_STEP step=7452 loss=1.354730 Training: 41%|████▏ | 7452/18018 [16:12:49<21:04:00, 7.18s/step, loss=1.355, acc=0.693, acc_len=1.66, thru=17.1, I=17.5, T=34.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:03,674] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:09,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:43:09,593] loop.py:45 INFO TRAIN_STEP step=7453 loss=1.654726 Training: 41%|████▏ | 7453/18018 [16:12:55<20:10:52, 6.88s/step, loss=1.655, acc=0.640, acc_len=1.49, thru=19.7, I=21.7, T=32.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:09,948] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:16,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:43:17,802] loop.py:45 INFO TRAIN_STEP step=7454 loss=1.455833 Training: 41%|████▏ | 7454/18018 [16:13:03<21:21:10, 7.28s/step, loss=1.456, acc=0.675, acc_len=1.60, thru=16.3, I=19.7, T=29.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:18,035] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:22,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:43:23,327] loop.py:380 INFO TIMING step=7455: step=3.495s data=1.480s compute=3.460s [fwd=2.455s bwd=0.990s opt=0.003s] dispatch=2.020s [2026-10-19 05:43:23,327] loop.py:45 INFO TRAIN_STEP step=7455 loss=1.407754 Training: 41%|████▏ | 7455/18018 [16:13:09<19:48:25, 6.75s/step, loss=1.408, acc=0.682, acc_len=1.63, thru=19.9, I=25.0, T=36.6, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:23,433] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (TrainerActor pid=9310) [2026-10-19 05:43:30,913] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7455: forward=2809.4ms backward=1191.3ms [2026-10-19 05:43:30,918] loop.py:45 INFO TRAIN_STEP step=7456 loss=1.320220 Training: 41%|████▏ | 7456/18018 [16:13:17<20:32:48, 7.00s/step, loss=1.320, acc=0.698, acc_len=1.68, thru=16.5, I=17.8, T=31.5, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:38,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:43:39,107] loop.py:45 INFO TRAIN_STEP step=7457 loss=1.467010 Training: 41%|████▏ | 7457/18018 [16:13:25<21:35:15, 7.36s/step, loss=1.467, acc=0.676, acc_len=1.60, thru=15.2, I=16.2, T=32.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:39,372] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:43:46,486] loop.py:45 INFO TRAIN_STEP step=7458 loss=1.382052 Training: 41%|████▏ | 7458/18018 [16:13:32<21:36:07, 7.36s/step, loss=1.382, acc=0.688, acc_len=1.64, thru=16.0, I=20.1, T=28.3, wait=2.8s, pool=56, epoch=1/2][2026-10-19 05:43:53,531] loop.py:45 INFO TRAIN_STEP step=7459 loss=1.474539 Training: 41%|████▏ | 7459/18018 [16:13:39<21:19:12, 7.27s/step, loss=1.475, acc=0.673, acc_len=1.59, thru=14.3, I=14.1, T=30.0, wait=2.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:43:59,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:44:01,305] loop.py:380 INFO TIMING step=7460: step=4.096s data=1.845s compute=4.071s [fwd=2.888s bwd=1.169s opt=0.003s] dispatch=3.668s [2026-10-19 05:44:01,305] loop.py:45 INFO TRAIN_STEP step=7460 loss=1.373582 Training: 41%|████▏ | 7460/18018 [16:13:47<21:45:39, 7.42s/step, loss=1.374, acc=0.689, acc_len=1.65, thru=18.5, I=18.3, T=31.2, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:44:01,702] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:44:08,487] loop.py:45 INFO TRAIN_STEP step=7461 loss=1.380162 Training: 41%|████▏ | 7461/18018 [16:13:54<21:33:09, 7.35s/step, loss=1.380, acc=0.686, acc_len=1.64, thru=13.6, I=15.5, T=32.8, wait=3.3s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:44:08,482] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7460: forward=2729.6ms backward=1123.9ms [2026-10-19 05:44:18,006] loop.py:45 INFO TRAIN_STEP step=7462 loss=1.418665 Training: 41%|████▏ | 7462/18018 [16:14:04<23:27:22, 8.00s/step, loss=1.419, acc=0.684, acc_len=1.63, thru=15.7, I=17.9, T=31.7, wait=5.5s, pool=52, epoch=1/2][2026-10-19 05:44:29,677] loop.py:45 INFO TRAIN_STEP step=7463 loss=1.474919 Training: 41%|████▏ | 7463/18018 [16:14:15<26:41:06, 9.10s/step, loss=1.475, acc=0.671, acc_len=1.58, thru=14.1, I=10.0, T=31.3, wait=7.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:44:35,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:44:36,540] loop.py:45 INFO TRAIN_STEP step=7464 loss=1.448771 Training: 41%|████▏ | 7464/18018 [16:14:22<24:42:50, 8.43s/step, loss=1.449, acc=0.680, acc_len=1.61, thru=17.6, I=17.8, T=33.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:44:36,832] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:44:42,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:44:43,004] loop.py:380 INFO TIMING step=7465: step=4.241s data=1.929s compute=4.211s [fwd=2.967s bwd=1.230s opt=0.003s] dispatch=2.212s [2026-10-19 05:44:43,004] loop.py:45 INFO TRAIN_STEP step=7465 loss=1.380825 Training: 41%|████▏ | 7465/18018 [16:14:29<22:58:57, 7.84s/step, loss=1.381, acc=0.690, acc_len=1.65, thru=18.6, I=22.1, T=30.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:44:43,437] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:44:49,635] loop.py:45 INFO TRAIN_STEP step=7466 loss=1.457994 Training: 41%|████▏ | 7466/18018 [16:14:35<21:55:03, 7.48s/step, loss=1.458, acc=0.672, acc_len=1.59, thru=15.3, I=18.4, T=33.5, wait=2.8s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:44:49,630] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7465: forward=2656.3ms backward=1104.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:44:58,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:44:58,616] loop.py:45 INFO TRAIN_STEP step=7467 loss=1.379528 Training: 41%|████▏ | 7467/18018 [16:14:44<23:14:12, 7.93s/step, loss=1.380, acc=0.692, acc_len=1.66, thru=15.3, I=13.9, T=31.3, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:44:58,875] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:45:05,659] loop.py:45 INFO TRAIN_STEP step=7468 loss=1.482671 Training: 41%|████▏ | 7468/18018 [16:14:51<22:27:23, 7.66s/step, loss=1.483, acc=0.669, acc_len=1.58, thru=18.6, I=18.0, T=31.1, wait=2.9s, pool=60, epoch=1/2][2026-10-19 05:45:13,864] loop.py:45 INFO TRAIN_STEP step=7469 loss=1.508511 Training: 41%|████▏ | 7469/18018 [16:14:59<22:55:44, 7.82s/step, loss=1.509, acc=0.663, acc_len=1.57, thru=13.6, I=15.3, T=31.3, wait=4.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:45:21,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:45:22,946] loop.py:380 INFO TIMING step=7470: step=3.699s data=1.481s compute=3.675s [fwd=2.652s bwd=1.010s opt=0.003s] dispatch=5.375s [2026-10-19 05:45:22,946] loop.py:45 INFO TRAIN_STEP step=7470 loss=1.445315 Training: 41%|████▏ | 7470/18018 [16:15:09<24:02:00, 8.20s/step, loss=1.445, acc=0.679, acc_len=1.61, thru=16.2, I=16.3, T=34.6, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:45:23,257] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:45:30,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:45:32,099] loop.py:45 INFO TRAIN_STEP step=7471 loss=1.465224 Training: 41%|████▏ | 7471/18018 [16:15:18<24:51:59, 8.49s/step, loss=1.465, acc=0.673, acc_len=1.59, thru=14.5, I=17.3, T=30.5, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:45:32,095] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7470: forward=2995.7ms backward=1145.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:45:32,434] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:45:38,736] loop.py:45 INFO TRAIN_STEP step=7472 loss=1.521828 Training: 41%|████▏ | 7472/18018 [16:15:24<23:14:12, 7.93s/step, loss=1.522, acc=0.662, acc_len=1.56, thru=16.6, I=22.3, T=37.0, wait=3.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:45:44,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:45:45,725] loop.py:45 INFO TRAIN_STEP step=7473 loss=1.292609 Training: 41%|████▏ | 7473/18018 [16:15:31<22:24:25, 7.65s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=18.9, I=21.8, T=35.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:45:46,139] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:45:51,948] loop.py:45 INFO TRAIN_STEP step=7474 loss=1.246210 Training: 41%|████▏ | 7474/18018 [16:15:38<21:09:04, 7.22s/step, loss=1.246, acc=0.708, acc_len=1.71, thru=16.2, I=25.4, T=33.6, wait=2.4s, pool=28, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:03,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:46:04,842] loop.py:380 INFO TIMING step=7475: step=5.424s data=2.730s compute=5.394s [fwd=3.728s bwd=1.652s opt=0.003s] dispatch=7.460s [2026-10-19 05:46:04,843] loop.py:45 INFO TRAIN_STEP step=7475 loss=1.417260 Training: 41%|████▏ | 7475/18018 [16:15:50<26:08:00, 8.92s/step, loss=1.417, acc=0.682, acc_len=1.62, thru=14.0, I=13.1, T=23.6, wait=7.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:05,137] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 05:46:11,968] loop.py:45 INFO TRAIN_STEP step=7476 loss=1.527932 (TrainerActor pid=9310) [2026-10-19 05:46:11,964] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7475: forward=3266.5ms backward=1463.0ms Training: 41%|████▏ | 7476/18018 [16:15:58<24:33:02, 8.38s/step, loss=1.528, acc=0.664, acc_len=1.56, thru=13.5, I=15.7, T=26.7, wait=2.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:19,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:46:20,984] loop.py:45 INFO TRAIN_STEP step=7477 loss=1.435667 Training: 41%|████▏ | 7477/18018 [16:16:07<25:06:07, 8.57s/step, loss=1.436, acc=0.676, acc_len=1.60, thru=16.4, I=15.5, T=27.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:21,391] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:28,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:46:28,661] loop.py:45 INFO TRAIN_STEP step=7478 loss=1.433032 Training: 42%|████▏ | 7478/18018 [16:16:14<24:18:55, 8.31s/step, loss=1.433, acc=0.679, acc_len=1.62, thru=15.5, I=18.2, T=32.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:29,042] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:46:34,301] loop.py:45 INFO TRAIN_STEP step=7479 loss=1.470803 Training: 42%|████▏ | 7479/18018 [16:16:20<21:58:15, 7.51s/step, loss=1.471, acc=0.672, acc_len=1.60, thru=19.8, I=21.0, T=35.4, wait=2.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:40,395] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:46:41,500] loop.py:380 INFO TIMING step=7480: step=3.928s data=1.734s compute=3.888s [fwd=2.740s bwd=1.134s opt=0.003s] dispatch=3.262s [2026-10-19 05:46:41,501] loop.py:45 INFO TRAIN_STEP step=7480 loss=1.546747 Training: 42%|████▏ | 7480/18018 [16:16:27<21:42:01, 7.41s/step, loss=1.547, acc=0.661, acc_len=1.55, thru=17.9, I=19.6, T=32.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:41,903] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:47,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:46:48,713] loop.py:45 INFO TRAIN_STEP step=7481 loss=1.500044 Training: 42%|████▏ | 7481/18018 [16:16:34<21:31:15, 7.35s/step, loss=1.500, acc=0.669, acc_len=1.58, thru=17.2, I=24.8, T=29.8, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:46:48,709] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7480: forward=3011.7ms backward=1237.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:49,183] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:55,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:46:56,588] loop.py:45 INFO TRAIN_STEP step=7482 loss=1.368366 Training: 42%|████▏ | 7482/18018 [16:16:42<21:58:47, 7.51s/step, loss=1.368, acc=0.691, acc_len=1.66, thru=14.2, I=16.5, T=31.2, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:46:56,863] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:47:01,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:47:03,292] loop.py:45 INFO TRAIN_STEP step=7483 loss=1.428219 Training: 42%|████▏ | 7483/18018 [16:16:49<21:16:06, 7.27s/step, loss=1.428, acc=0.677, acc_len=1.61, thru=17.9, I=22.7, T=29.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:47:03,441] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 05:47:09,659] loop.py:45 INFO TRAIN_STEP step=7484 loss=1.547592 Training: 42%|████▏ | 7484/18018 [16:16:55<20:28:39, 7.00s/step, loss=1.548, acc=0.659, acc_len=1.55, thru=13.1, I=21.2, T=34.8, wait=2.7s, pool=24, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:47:19,947] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:47:21,338] loop.py:380 INFO TIMING step=7485: step=4.252s data=1.807s compute=4.217s [fwd=3.027s bwd=1.177s opt=0.003s] dispatch=7.417s [2026-10-19 05:47:21,338] loop.py:45 INFO TRAIN_STEP step=7485 loss=1.486961 Training: 42%|████▏ | 7485/18018 [16:17:07<24:35:01, 8.40s/step, loss=1.487, acc=0.671, acc_len=1.58, thru=15.7, I=15.6, T=30.1, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:47:21,462] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:47:25,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:47:26,693] loop.py:45 INFO TRAIN_STEP step=7486 loss=1.330235 Training: 42%|████▏ | 7486/18018 [16:17:12<21:54:27, 7.49s/step, loss=1.330, acc=0.698, acc_len=1.67, thru=20.6, I=23.3, T=36.6, wait=1.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:47:26,688] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7485: forward=2469.6ms backward=979.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:47:26,800] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:47:35,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:47:35,503] loop.py:45 INFO TRAIN_STEP step=7487 loss=1.303833 Training: 42%|████▏ | 7487/18018 [16:17:21<23:03:46, 7.88s/step, loss=1.304, acc=0.700, acc_len=1.69, thru=13.6, I=16.8, T=30.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:47:35,947] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:47:42,583] loop.py:45 INFO TRAIN_STEP step=7488 loss=1.454696 Training: 42%|████▏ | 7488/18018 [16:17:28<22:21:31, 7.64s/step, loss=1.455, acc=0.674, acc_len=1.60, thru=14.7, I=15.5, T=36.0, wait=3.5s, pool=52, epoch=1/2][2026-10-19 05:47:50,639] loop.py:45 INFO TRAIN_STEP step=7489 loss=1.482848 Training: 42%|████▏ | 7489/18018 [16:17:36<22:42:53, 7.77s/step, loss=1.483, acc=0.675, acc_len=1.60, thru=12.9, I=17.2, T=32.6, wait=4.1s, pool=24, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:47:59,583] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:48:01,957] loop.py:380 INFO TIMING step=7490: step=4.452s data=2.003s compute=4.418s [fwd=3.175s bwd=1.228s opt=0.003s] dispatch=6.858s [2026-10-19 05:48:01,957] loop.py:45 INFO TRAIN_STEP step=7490 loss=1.471462 Training: 42%|████▏ | 7490/18018 [16:17:48<25:49:50, 8.83s/step, loss=1.471, acc=0.674, acc_len=1.60, thru=16.0, I=17.0, T=28.8, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:02,115] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:08,352] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:48:08,843] loop.py:45 INFO TRAIN_STEP step=7491 loss=1.530742 Training: 42%|████▏ | 7491/18018 [16:17:54<24:07:04, 8.25s/step, loss=1.531, acc=0.660, acc_len=1.55, thru=15.0, I=20.4, T=30.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:08,870] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-19 05:48:08,838] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7490: forward=2871.9ms backward=1294.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:16,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:48:17,814] loop.py:45 INFO TRAIN_STEP step=7492 loss=1.572968 Training: 42%|████▏ | 7492/18018 [16:18:03<24:45:00, 8.46s/step, loss=1.573, acc=0.655, acc_len=1.54, thru=14.5, I=18.0, T=26.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:18,299] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:23,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:48:25,050] loop.py:45 INFO TRAIN_STEP step=7493 loss=1.438162 Training: 42%|████▏ | 7493/18018 [16:18:11<23:40:19, 8.10s/step, loss=1.438, acc=0.681, acc_len=1.61, thru=16.6, I=23.0, T=30.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:25,422] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:32,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:48:32,444] loop.py:45 INFO TRAIN_STEP step=7494 loss=1.472066 Training: 42%|████▏ | 7494/18018 [16:18:18<23:03:10, 7.89s/step, loss=1.472, acc=0.674, acc_len=1.60, thru=16.9, I=20.7, T=27.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:32,559] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:38,461] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:48:38,661] loop.py:380 INFO TIMING step=7495: step=3.727s data=1.645s compute=3.695s [fwd=2.636s bwd=1.044s opt=0.003s] dispatch=2.480s [2026-10-19 05:48:38,661] loop.py:45 INFO TRAIN_STEP step=7495 loss=1.420539 Training: 42%|████▏ | 7495/18018 [16:18:24<21:35:15, 7.39s/step, loss=1.421, acc=0.680, acc_len=1.62, thru=18.6, I=18.2, T=34.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:38,964] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:45,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:48:45,856] loop.py:45 INFO TRAIN_STEP step=7496 loss=1.416650 Training: 42%|████▏ | 7496/18018 [16:18:31<21:24:58, 7.33s/step, loss=1.417, acc=0.684, acc_len=1.63, thru=18.1, I=18.9, T=32.2, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:48:45,852] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7495: forward=2734.3ms backward=1199.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:46,342] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:50,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:48:52,062] loop.py:45 INFO TRAIN_STEP step=7497 loss=1.421558 Training: 42%|████▏ | 7497/18018 [16:18:38<20:26:00, 6.99s/step, loss=1.422, acc=0.680, acc_len=1.61, thru=19.5, I=22.5, T=34.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:52,303] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:48:59,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:49:00,790] loop.py:45 INFO TRAIN_STEP step=7498 loss=1.307691 Training: 42%|████▏ | 7498/18018 [16:18:46<21:57:12, 7.51s/step, loss=1.308, acc=0.701, acc_len=1.68, thru=15.1, I=17.6, T=28.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:00,889] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:06,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:49:09,625] loop.py:45 INFO TRAIN_STEP step=7499 loss=1.351373 Training: 42%|████▏ | 7499/18018 [16:18:55<23:06:36, 7.91s/step, loss=1.351, acc=0.691, acc_len=1.65, thru=13.6, I=20.8, T=25.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:09,802] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:13,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:49:15,015] loop.py:380 INFO TIMING step=7500: step=3.538s data=1.503s compute=3.496s [fwd=2.418s bwd=1.063s opt=0.003s] dispatch=1.843s [2026-10-19 05:49:15,015] loop.py:45 INFO TRAIN_STEP step=7500 loss=1.445388 Training: 42%|████▏ | 7500/18018 [16:19:01<20:54:00, 7.15s/step, loss=1.445, acc=0.678, acc_len=1.61, thru=17.6, I=28.0, T=36.2, wait=1.8s, pool=64, epoch=1/2][2026-10-19 05:49:15,019] loop.py:419 INFO Saving checkpoint at step 7500... (TrainerActor pid=9310) [2026-10-19 05:49:18,502] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0007501 (TrainerActor pid=9310) [2026-10-19 05:49:18,511] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0006751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:18,960] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:19,966] inference_manager.py:408 INFO Pool capacity available after 6.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:25,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 05:49:18,504] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0007501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 05:49:19,676] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0006751 [2026-10-19 05:49:27,512] loop.py:45 INFO TRAIN_STEP step=7501 loss=1.414920 Training: 42%|████▏ | 7501/18018 [16:19:13<25:34:46, 8.76s/step, loss=1.415, acc=0.682, acc_len=1.63, thru=17.0, I=23.1, T=25.8, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:49:27,507] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7500: forward=3414.5ms backward=1506.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:27,997] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:32,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:49:34,833] loop.py:45 INFO TRAIN_STEP step=7502 loss=1.413499 Training: 42%|████▏ | 7502/18018 [16:19:20<24:19:12, 8.33s/step, loss=1.413, acc=0.678, acc_len=1.61, thru=13.8, I=24.9, T=28.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:34,877] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 [2026-10-19 05:49:43,048] loop.py:45 INFO TRAIN_STEP step=7503 loss=1.296657 Training: 42%|████▏ | 7503/18018 [16:19:29<24:13:17, 8.29s/step, loss=1.297, acc=0.698, acc_len=1.69, thru=15.2, I=18.4, T=27.4, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:47,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:49:49,891] loop.py:45 INFO TRAIN_STEP step=7504 loss=1.524681 Training: 42%|████▏ | 7504/18018 [16:19:35<22:56:52, 7.86s/step, loss=1.525, acc=0.666, acc_len=1.57, thru=16.8, I=20.2, T=29.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:50,347] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:56,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:49:57,124] loop.py:380 INFO TIMING step=7505: step=3.742s data=1.693s compute=3.715s [fwd=2.656s bwd=1.045s opt=0.003s] dispatch=3.482s [2026-10-19 05:49:57,124] loop.py:45 INFO TRAIN_STEP step=7505 loss=1.312838 Training: 42%|████▏ | 7505/18018 [16:19:43<22:24:03, 7.67s/step, loss=1.313, acc=0.699, acc_len=1.68, thru=15.7, I=23.1, T=34.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:49:57,178] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:02,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:50:04,310] loop.py:45 INFO TRAIN_STEP step=7506 loss=1.513514 Training: 42%|████▏ | 7506/18018 [16:19:50<21:58:25, 7.53s/step, loss=1.514, acc=0.666, acc_len=1.57, thru=17.1, I=22.1, T=29.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:50:04,306] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7505: forward=3045.6ms backward=1318.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:04,472] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:09,860] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:50:11,826] loop.py:45 INFO TRAIN_STEP step=7507 loss=1.436172 Training: 42%|████▏ | 7507/18018 [16:19:57<21:57:50, 7.52s/step, loss=1.436, acc=0.675, acc_len=1.60, thru=15.6, I=23.1, T=29.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:11,877] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:16,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:50:17,919] loop.py:45 INFO TRAIN_STEP step=7508 loss=1.362138 Training: 42%|████▏ | 7508/18018 [16:20:04<20:42:36, 7.09s/step, loss=1.362, acc=0.685, acc_len=1.63, thru=19.0, I=24.6, T=31.6, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:17,981] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 05:50:25,084] loop.py:45 INFO TRAIN_STEP step=7509 loss=1.418322 Training: 42%|████▏ | 7509/18018 [16:20:11<20:46:04, 7.11s/step, loss=1.418, acc=0.677, acc_len=1.61, thru=17.1, I=17.1, T=33.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:25,019] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:25,521] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:29,917] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:50:31,515] loop.py:380 INFO TIMING step=7510: step=4.144s data=1.835s compute=4.114s [fwd=2.854s bwd=1.245s opt=0.003s] dispatch=2.279s [2026-10-19 05:50:31,515] loop.py:45 INFO TRAIN_STEP step=7510 loss=1.332299 Training: 42%|████▏ | 7510/18018 [16:20:17<20:10:04, 6.91s/step, loss=1.332, acc=0.697, acc_len=1.67, thru=19.6, I=23.0, T=30.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:31,933] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:38,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:50:39,375] loop.py:45 INFO TRAIN_STEP step=7511 loss=1.483295 Training: 42%|████▏ | 7511/18018 [16:20:25<20:59:55, 7.19s/step, loss=1.483, acc=0.672, acc_len=1.59, thru=14.5, I=18.4, T=26.7, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:50:39,371] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7510: forward=3515.9ms backward=1237.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:39,771] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:46,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:50:47,736] loop.py:45 INFO TRAIN_STEP step=7512 loss=1.451260 Training: 42%|████▏ | 7512/18018 [16:20:33<22:01:05, 7.54s/step, loss=1.451, acc=0.674, acc_len=1.60, thru=15.7, I=17.2, T=30.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:48,052] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:53,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:50:54,568] loop.py:45 INFO TRAIN_STEP step=7513 loss=1.408815 Training: 42%|████▏ | 7513/18018 [16:20:40<21:23:36, 7.33s/step, loss=1.409, acc=0.680, acc_len=1.62, thru=17.9, I=21.5, T=31.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:50:54,689] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 05:51:01,374] loop.py:45 INFO TRAIN_STEP step=7514 loss=1.360459 Training: 42%|████▏ | 7514/18018 [16:20:47<20:55:43, 7.17s/step, loss=1.360, acc=0.690, acc_len=1.65, thru=18.0, I=18.6, T=28.9, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:07,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:51:08,966] loop.py:380 INFO TIMING step=7515: step=4.420s data=2.005s compute=4.386s [fwd=3.062s bwd=1.310s opt=0.003s] dispatch=3.164s [2026-10-19 05:51:08,967] loop.py:45 INFO TRAIN_STEP step=7515 loss=1.424226 Training: 42%|████▏ | 7515/18018 [16:20:55<21:17:44, 7.30s/step, loss=1.424, acc=0.676, acc_len=1.60, thru=18.6, I=17.7, T=29.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:09,396] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:15,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:51:16,878] loop.py:45 INFO TRAIN_STEP step=7516 loss=1.374007 Training: 42%|████▏ | 7516/18018 [16:21:02<21:49:40, 7.48s/step, loss=1.374, acc=0.687, acc_len=1.64, thru=14.5, I=21.3, T=30.2, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:51:16,873] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7515: forward=2880.7ms backward=1301.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:16,982] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:22,750] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:51:24,271] loop.py:45 INFO TRAIN_STEP step=7517 loss=1.448291 Training: 42%|████▏ | 7517/18018 [16:21:10<21:44:48, 7.46s/step, loss=1.448, acc=0.673, acc_len=1.59, thru=16.4, I=22.6, T=31.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:24,756] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:30,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:51:31,943] loop.py:45 INFO TRAIN_STEP step=7518 loss=1.437485 Training: 42%|████▏ | 7518/18018 [16:21:18<21:56:06, 7.52s/step, loss=1.437, acc=0.678, acc_len=1.61, thru=15.3, I=20.8, T=30.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:32,225] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:36,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:51:38,447] loop.py:45 INFO TRAIN_STEP step=7519 loss=1.373098 Training: 42%|████▏ | 7519/18018 [16:21:24<21:02:44, 7.22s/step, loss=1.373, acc=0.691, acc_len=1.66, thru=17.7, I=23.5, T=27.9, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:38,485] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:45,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:51:46,890] loop.py:380 INFO TIMING step=7520: step=5.337s data=2.992s compute=5.306s [fwd=4.148s bwd=1.144s opt=0.003s] dispatch=3.096s [2026-10-19 05:51:46,891] loop.py:45 INFO TRAIN_STEP step=7520 loss=1.455301 Training: 42%|████▏ | 7520/18018 [16:21:32<22:06:56, 7.58s/step, loss=1.455, acc=0.679, acc_len=1.61, thru=15.3, I=16.9, T=24.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:47,204] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:53,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:51:54,767] loop.py:45 INFO TRAIN_STEP step=7521 loss=1.386310 Training: 42%|████▏ | 7521/18018 [16:21:40<22:22:09, 7.67s/step, loss=1.386, acc=0.684, acc_len=1.63, thru=14.9, I=19.6, T=24.9, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:51:54,763] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7520: forward=3588.8ms backward=1505.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:51:54,921] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:01,182] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:52:01,273] loop.py:45 INFO TRAIN_STEP step=7522 loss=1.465873 Training: 42%|████▏ | 7522/18018 [16:21:47<21:20:51, 7.32s/step, loss=1.466, acc=0.673, acc_len=1.59, thru=17.1, I=19.6, T=31.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:01,684] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:52:08,318] loop.py:45 INFO TRAIN_STEP step=7523 loss=1.324836 Training: 42%|████▏ | 7523/18018 [16:21:54<21:06:11, 7.24s/step, loss=1.325, acc=0.700, acc_len=1.69, thru=15.6, I=17.7, T=32.9, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:14,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:52:15,233] loop.py:45 INFO TRAIN_STEP step=7524 loss=1.365300 Training: 42%|████▏ | 7524/18018 [16:22:01<20:49:06, 7.14s/step, loss=1.365, acc=0.689, acc_len=1.65, thru=18.2, I=18.4, T=36.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:15,315] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 [2026-10-19 05:52:22,953] loop.py:380 INFO TIMING step=7525: step=3.692s data=1.530s compute=3.663s [fwd=2.613s bwd=1.035s opt=0.003s] dispatch=4.019s [2026-10-19 05:52:22,953] loop.py:45 INFO TRAIN_STEP step=7525 loss=1.447744 Training: 42%|████▏ | 7525/18018 [16:22:09<21:19:37, 7.32s/step, loss=1.448, acc=0.674, acc_len=1.60, thru=15.8, I=16.7, T=34.7, wait=4.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:29,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:52:31,194] loop.py:45 INFO TRAIN_STEP step=7526 loss=1.550475 Training: 42%|████▏ | 7526/18018 [16:22:17<22:07:45, 7.59s/step, loss=1.550, acc=0.661, acc_len=1.55, thru=16.2, I=18.0, T=31.2, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:52:31,189] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7525: forward=2914.7ms backward=1138.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:31,582] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:37,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:52:38,869] loop.py:45 INFO TRAIN_STEP step=7527 loss=1.528025 Training: 42%|████▏ | 7527/18018 [16:22:24<22:11:49, 7.62s/step, loss=1.528, acc=0.662, acc_len=1.56, thru=14.6, I=19.4, T=27.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:38,919] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 05:52:47,042] loop.py:45 INFO TRAIN_STEP step=7528 loss=1.424921 Training: 42%|████▏ | 7528/18018 [16:22:34<23:42:12, 8.13s/step, loss=1.425, acc=0.680, acc_len=1.62, thru=13.8, I=19.5, T=24.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:48,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:52:48,716] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:52:55,576] loop.py:45 INFO TRAIN_STEP step=7529 loss=1.312598 Training: 42%|████▏ | 7529/18018 [16:22:41<23:01:47, 7.90s/step, loss=1.313, acc=0.698, acc_len=1.68, thru=16.3, I=18.6, T=32.9, wait=3.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:01,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:53:02,713] loop.py:380 INFO TIMING step=7530: step=4.126s data=1.882s compute=4.096s [fwd=2.954s bwd=1.128s opt=0.003s] dispatch=2.999s [2026-10-19 05:53:02,713] loop.py:45 INFO TRAIN_STEP step=7530 loss=1.371345 Training: 42%|████▏ | 7530/18018 [16:22:48<22:21:15, 7.67s/step, loss=1.371, acc=0.694, acc_len=1.67, thru=17.5, I=20.1, T=31.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:03,172] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:09,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:53:10,228] loop.py:45 INFO TRAIN_STEP step=7531 loss=1.383712 Training: 42%|████▏ | 7531/18018 [16:22:56<22:12:55, 7.63s/step, loss=1.384, acc=0.694, acc_len=1.66, thru=16.5, I=20.9, T=30.6, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:53:10,224] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7530: forward=2902.2ms backward=1236.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:10,448] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:53:17,308] loop.py:45 INFO TRAIN_STEP step=7532 loss=1.389053 Training: 42%|████▏ | 7532/18018 [16:23:03<21:44:12, 7.46s/step, loss=1.389, acc=0.684, acc_len=1.63, thru=16.8, I=19.4, T=35.6, wait=3.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:23,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:53:25,369] loop.py:45 INFO TRAIN_STEP step=7533 loss=1.576147 Training: 42%|████▏ | 7533/18018 [16:23:11<22:15:27, 7.64s/step, loss=1.576, acc=0.655, acc_len=1.53, thru=17.2, I=18.7, T=33.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:25,700] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:31,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:53:32,309] loop.py:45 INFO TRAIN_STEP step=7534 loss=1.373453 Training: 42%|████▏ | 7534/18018 [16:23:18<21:38:25, 7.43s/step, loss=1.373, acc=0.690, acc_len=1.64, thru=16.1, I=22.2, T=35.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:32,728] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:37,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:53:38,824] loop.py:380 INFO TIMING step=7535: step=3.835s data=1.706s compute=3.786s [fwd=2.653s bwd=1.119s opt=0.003s] dispatch=2.672s [2026-10-19 05:53:38,824] loop.py:45 INFO TRAIN_STEP step=7535 loss=1.425115 Training: 42%|████▏ | 7535/18018 [16:23:24<20:50:24, 7.16s/step, loss=1.425, acc=0.682, acc_len=1.63, thru=19.4, I=24.6, T=33.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:39,028] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:44,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:53:45,729] loop.py:45 INFO TRAIN_STEP step=7536 loss=1.330708 Training: 42%|████▏ | 7536/18018 [16:23:31<20:37:01, 7.08s/step, loss=1.331, acc=0.696, acc_len=1.67, thru=17.2, I=26.3, T=30.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:53:45,724] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7535: forward=2870.0ms backward=1253.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:45,812] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:50,706] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:53:52,150] loop.py:45 INFO TRAIN_STEP step=7537 loss=1.359993 Training: 42%|████▏ | 7537/18018 [16:23:38<20:02:26, 6.88s/step, loss=1.360, acc=0.691, acc_len=1.65, thru=17.9, I=24.3, T=33.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:52,227] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:58,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:53:59,505] loop.py:45 INFO TRAIN_STEP step=7538 loss=1.360992 Training: 42%|████▏ | 7538/18018 [16:23:45<20:26:54, 7.02s/step, loss=1.361, acc=0.690, acc_len=1.65, thru=16.7, I=20.8, T=33.3, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:53:59,853] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:04,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:54:05,281] loop.py:45 INFO TRAIN_STEP step=7539 loss=1.262148 Training: 42%|████▏ | 7539/18018 [16:23:51<19:21:21, 6.65s/step, loss=1.262, acc=0.709, acc_len=1.71, thru=20.3, I=23.3, T=35.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:05,424] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:13,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:54:13,718] loop.py:380 INFO TIMING step=7540: step=4.407s data=2.026s compute=4.363s [fwd=3.073s bwd=1.275s opt=0.003s] dispatch=4.022s [2026-10-19 05:54:13,718] loop.py:45 INFO TRAIN_STEP step=7540 loss=1.582936 Training: 42%|████▏ | 7540/18018 [16:23:59<20:54:57, 7.19s/step, loss=1.583, acc=0.651, acc_len=1.52, thru=14.4, I=15.8, T=29.0, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:14,213] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:20,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:54:20,487] loop.py:45 INFO TRAIN_STEP step=7541 loss=1.471105 Training: 42%|████▏ | 7541/18018 [16:24:06<20:32:52, 7.06s/step, loss=1.471, acc=0.671, acc_len=1.58, thru=17.7, I=18.2, T=32.4, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:54:20,483] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7540: forward=2773.6ms backward=1135.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:20,895] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:30,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:54:31,158] loop.py:45 INFO TRAIN_STEP step=7542 loss=1.453865 Training: 42%|████▏ | 7542/18018 [16:24:17<23:42:00, 8.14s/step, loss=1.454, acc=0.673, acc_len=1.59, thru=12.2, I=14.1, T=29.4, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:31,649] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:38,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:54:40,257] loop.py:45 INFO TRAIN_STEP step=7543 loss=1.500307 Training: 42%|████▏ | 7543/18018 [16:24:26<24:31:48, 8.43s/step, loss=1.500, acc=0.667, acc_len=1.57, thru=14.9, I=17.0, T=28.8, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:40,608] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 05:54:47,127] loop.py:45 INFO TRAIN_STEP step=7544 loss=1.406366 Training: 42%|████▏ | 7544/18018 [16:24:33<23:09:54, 7.96s/step, loss=1.406, acc=0.686, acc_len=1.64, thru=14.5, I=20.0, T=37.1, wait=3.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:53,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:54:54,525] loop.py:380 INFO TIMING step=7545: step=3.815s data=1.702s compute=3.787s [fwd=2.735s bwd=1.037s opt=0.003s] dispatch=3.576s [2026-10-19 05:54:54,525] loop.py:45 INFO TRAIN_STEP step=7545 loss=1.252776 Training: 42%|████▏ | 7545/18018 [16:24:40<22:40:21, 7.79s/step, loss=1.253, acc=0.711, acc_len=1.73, thru=19.9, I=17.8, T=33.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:54,591] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:54:59,044] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:55:00,014] loop.py:45 INFO TRAIN_STEP step=7546 loss=1.372520 Training: 42%|████▏ | 7546/18018 [16:24:46<20:39:35, 7.10s/step, loss=1.373, acc=0.688, acc_len=1.64, thru=20.8, I=24.4, T=33.3, wait=1.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:55:00,010] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7545: forward=2677.3ms backward=1121.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:00,067] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:04,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:55:07,061] loop.py:45 INFO TRAIN_STEP step=7547 loss=1.322958 Training: 42%|████▏ | 7547/18018 [16:24:53<20:36:26, 7.08s/step, loss=1.323, acc=0.702, acc_len=1.69, thru=16.1, I=22.1, T=27.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:07,465] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:12,345] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:55:13,425] loop.py:45 INFO TRAIN_STEP step=7548 loss=1.310055 Training: 42%|████▏ | 7548/18018 [16:24:59<19:58:35, 6.87s/step, loss=1.310, acc=0.698, acc_len=1.68, thru=17.0, I=25.2, T=29.0, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:13,855] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:18,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:55:20,043] loop.py:45 INFO TRAIN_STEP step=7549 loss=1.303319 Training: 42%|████▏ | 7549/18018 [16:25:06<19:45:22, 6.79s/step, loss=1.303, acc=0.697, acc_len=1.68, thru=17.3, I=22.8, T=30.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:20,297] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:25,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:55:25,798] loop.py:380 INFO TIMING step=7550: step=3.708s data=1.553s compute=3.676s [fwd=2.595s bwd=1.067s opt=0.003s] dispatch=2.038s [2026-10-19 05:55:25,798] loop.py:45 INFO TRAIN_STEP step=7550 loss=1.445932 Training: 42%|████▏ | 7550/18018 [16:25:11<18:51:00, 6.48s/step, loss=1.446, acc=0.678, acc_len=1.61, thru=19.5, I=24.9, T=34.5, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:26,098] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:33,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:55:33,293] loop.py:45 INFO TRAIN_STEP step=7551 loss=1.476706 Training: 42%|████▏ | 7551/18018 [16:25:19<19:43:50, 6.79s/step, loss=1.477, acc=0.669, acc_len=1.58, thru=17.6, I=19.3, T=31.8, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:55:33,289] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7550: forward=2865.8ms backward=1120.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:33,561] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:39,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:55:40,322] loop.py:45 INFO TRAIN_STEP step=7552 loss=1.420893 Training: 42%|████▏ | 7552/18018 [16:25:26<19:56:20, 6.86s/step, loss=1.421, acc=0.683, acc_len=1.63, thru=17.7, I=19.0, T=32.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:55:40,353] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 05:55:47,175] loop.py:45 INFO TRAIN_STEP step=7553 loss=1.317281 Training: 42%|████▏ | 7553/18018 [16:25:33<19:56:01, 6.86s/step, loss=1.317, acc=0.700, acc_len=1.69, thru=16.1, I=22.2, T=30.3, wait=2.6s, pool=52, epoch=1/2][2026-10-19 05:55:56,874] loop.py:45 INFO TRAIN_STEP step=7554 loss=1.313287 Training: 42%|████▏ | 7554/18018 [16:25:42<22:24:39, 7.71s/step, loss=1.313, acc=0.702, acc_len=1.70, thru=10.0, I=10.8, T=28.3, wait=5.2s, pool=12, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 05:55:57,133] eagle_store.py:517 WARNING batch_get_buffer missing keys (k241735-aca9c571_hs, k241735-aca9c571_ids, k241735-aca9c571_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:06,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 05:55:58,044] eagle_store.py:517 WARNING batch_get_buffer missing keys (k241738-9a0ea926_hs, k241738-9a0ea926_ids, k241738-9a0ea926_lhs); sleeping 0.50s. [repeated 6x across cluster] [2026-10-19 05:56:07,611] loop.py:380 INFO TIMING step=7555: step=4.414s data=2.046s compute=4.360s [fwd=3.017s bwd=1.328s opt=0.003s] dispatch=6.312s [2026-10-19 05:56:07,611] loop.py:45 INFO TRAIN_STEP step=7555 loss=1.372187 Training: 42%|████▏ | 7555/18018 [16:25:53<25:02:52, 8.62s/step, loss=1.372, acc=0.695, acc_len=1.67, thru=17.6, I=13.6, T=29.0, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:07,898] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:56:16,395] loop.py:45 INFO TRAIN_STEP step=7556 loss=1.493134 Training: 42%|████▏ | 7556/18018 [16:26:02<25:11:13, 8.67s/step, loss=1.493, acc=0.671, acc_len=1.58, thru=14.2, I=15.4, T=28.5, wait=4.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:56:16,390] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7555: forward=3144.8ms backward=1293.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:21,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:56:23,085] loop.py:45 INFO TRAIN_STEP step=7557 loss=1.417833 Training: 42%|████▏ | 7557/18018 [16:26:09<23:27:42, 8.07s/step, loss=1.418, acc=0.687, acc_len=1.63, thru=17.5, I=20.5, T=28.6, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:23,444] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:30,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:56:31,844] loop.py:45 INFO TRAIN_STEP step=7558 loss=1.497204 Training: 42%|████▏ | 7558/18018 [16:26:17<24:03:27, 8.28s/step, loss=1.497, acc=0.667, acc_len=1.57, thru=15.2, I=15.2, T=32.7, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:32,001] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:37,814] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:56:38,072] loop.py:45 INFO TRAIN_STEP step=7559 loss=1.470407 Training: 42%|████▏ | 7559/18018 [16:26:24<22:16:02, 7.66s/step, loss=1.470, acc=0.671, acc_len=1.59, thru=20.0, I=20.2, T=33.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:38,315] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:56:44,968] loop.py:380 INFO TIMING step=7560: step=3.847s data=1.662s compute=3.817s [fwd=2.670s bwd=1.132s opt=0.003s] dispatch=3.040s [2026-10-19 05:56:44,968] loop.py:45 INFO TRAIN_STEP step=7560 loss=1.372455 Training: 42%|████▏ | 7560/18018 [16:26:31<21:35:43, 7.43s/step, loss=1.372, acc=0.692, acc_len=1.65, thru=17.3, I=18.2, T=33.3, wait=3.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:51,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:56:52,421] loop.py:45 INFO TRAIN_STEP step=7561 loss=1.335477 Training: 42%|████▏ | 7561/18018 [16:26:38<21:36:30, 7.44s/step, loss=1.335, acc=0.699, acc_len=1.68, thru=17.7, I=18.2, T=34.7, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:56:52,417] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7560: forward=2651.8ms backward=991.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:52,837] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:56:58,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:56:59,863] loop.py:45 INFO TRAIN_STEP step=7562 loss=1.383419 Training: 42%|████▏ | 7562/18018 [16:26:45<21:36:40, 7.44s/step, loss=1.383, acc=0.691, acc_len=1.66, thru=16.7, I=20.9, T=27.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:00,075] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:06,276] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:57:06,745] loop.py:45 INFO TRAIN_STEP step=7563 loss=1.488112 Training: 42%|████▏ | 7563/18018 [16:26:52<21:07:13, 7.27s/step, loss=1.488, acc=0.667, acc_len=1.58, thru=17.0, I=19.5, T=30.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:06,792] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:11,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:57:13,193] loop.py:45 INFO TRAIN_STEP step=7564 loss=1.401658 Training: 42%|████▏ | 7564/18018 [16:26:59<20:24:02, 7.03s/step, loss=1.402, acc=0.684, acc_len=1.63, thru=18.5, I=21.1, T=30.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:13,223] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 [2026-10-19 05:57:20,688] loop.py:380 INFO TIMING step=7565: step=3.551s data=1.498s compute=3.518s [fwd=2.466s bwd=1.038s opt=0.003s] dispatch=3.934s [2026-10-19 05:57:20,688] loop.py:45 INFO TRAIN_STEP step=7565 loss=1.552117 Training: 42%|████▏ | 7565/18018 [16:27:06<20:48:27, 7.17s/step, loss=1.552, acc=0.658, acc_len=1.54, thru=14.8, I=19.4, T=36.0, wait=3.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:30,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:57:31,518] loop.py:45 INFO TRAIN_STEP step=7566 loss=1.508602 Training: 42%|████▏ | 7566/18018 [16:27:17<23:59:57, 8.27s/step, loss=1.509, acc=0.667, acc_len=1.58, thru=15.6, I=13.5, T=29.1, wait=6.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:57:31,514] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7565: forward=3043.5ms backward=1287.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:31,595] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-19 05:57:38,345] loop.py:45 INFO TRAIN_STEP step=7567 loss=1.465348 Training: 42%|████▏ | 7567/18018 [16:27:24<22:44:27, 7.83s/step, loss=1.465, acc=0.674, acc_len=1.59, thru=15.5, I=19.7, T=35.1, wait=3.2s, pool=52, epoch=1/2][2026-10-19 05:57:47,329] loop.py:45 INFO TRAIN_STEP step=7568 loss=1.320872 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:47,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 Training: 42%|████▏ | 7568/18018 [16:27:33<23:44:25, 8.18s/step, loss=1.321, acc=0.697, acc_len=1.67, thru=15.8, I=14.4, T=29.9, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:47,746] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:52,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:57:53,502] loop.py:45 INFO TRAIN_STEP step=7569 loss=1.350284 Training: 42%|████▏ | 7569/18018 [16:27:39<21:59:31, 7.58s/step, loss=1.350, acc=0.696, acc_len=1.67, thru=18.9, I=20.5, T=32.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:57:53,552] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:01,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:58:02,455] loop.py:380 INFO TIMING step=7570: step=4.448s data=2.059s compute=4.419s [fwd=3.150s bwd=1.255s opt=0.003s] dispatch=4.497s [2026-10-19 05:58:02,455] loop.py:45 INFO TRAIN_STEP step=7570 loss=1.517593 Training: 42%|████▏ | 7570/18018 [16:27:48<23:11:18, 7.99s/step, loss=1.518, acc=0.669, acc_len=1.59, thru=14.1, I=17.1, T=28.8, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:02,595] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:09,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 05:58:11,080] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7570: forward=3010.2ms backward=1236.6ms [2026-10-19 05:58:11,084] loop.py:45 INFO TRAIN_STEP step=7571 loss=1.499533 Training: 42%|████▏ | 7571/18018 [16:27:57<23:44:39, 8.18s/step, loss=1.500, acc=0.670, acc_len=1.59, thru=15.1, I=16.5, T=29.7, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:11,547] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:16,916] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:58:18,430] loop.py:45 INFO TRAIN_STEP step=7572 loss=1.532738 Training: 42%|████▏ | 7572/18018 [16:28:04<23:00:39, 7.93s/step, loss=1.533, acc=0.662, acc_len=1.56, thru=14.8, I=24.0, T=32.7, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:18,931] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:23,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:58:24,669] loop.py:45 INFO TRAIN_STEP step=7573 loss=1.304867 Training: 42%|████▏ | 7573/18018 [16:28:10<21:32:23, 7.42s/step, loss=1.305, acc=0.699, acc_len=1.68, thru=18.8, I=26.4, T=34.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:24,901] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:31,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:58:33,637] loop.py:45 INFO TRAIN_STEP step=7574 loss=1.550225 Training: 42%|████▏ | 7574/18018 [16:28:19<22:52:47, 7.89s/step, loss=1.550, acc=0.660, acc_len=1.55, thru=14.7, I=19.2, T=25.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:33,721] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 [2026-10-19 05:58:41,552] loop.py:380 INFO TIMING step=7575: step=4.097s data=1.915s compute=4.057s [fwd=2.827s bwd=1.216s opt=0.003s] dispatch=3.808s [2026-10-19 05:58:41,552] loop.py:45 INFO TRAIN_STEP step=7575 loss=1.408663 Training: 42%|████▏ | 7575/18018 [16:28:27<22:54:13, 7.90s/step, loss=1.409, acc=0.682, acc_len=1.62, thru=14.3, I=20.4, T=31.2, wait=3.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:49,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:58:52,182] loop.py:45 INFO TRAIN_STEP step=7576 loss=1.460400 Training: 42%|████▏ | 7576/18018 [16:28:38<25:16:48, 8.72s/step, loss=1.460, acc=0.676, acc_len=1.60, thru=14.8, I=14.6, T=26.5, wait=5.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:58:52,177] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7575: forward=3424.2ms backward=1350.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:52,649] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:56,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:58:57,983] loop.py:45 INFO TRAIN_STEP step=7577 loss=1.306104 Training: 42%|████▏ | 7577/18018 [16:28:44<22:44:34, 7.84s/step, loss=1.306, acc=0.703, acc_len=1.69, thru=17.3, I=27.8, T=35.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:58:58,348] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:03,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:59:05,122] loop.py:45 INFO TRAIN_STEP step=7578 loss=1.593630 Training: 42%|████▏ | 7578/18018 [16:28:51<22:07:44, 7.63s/step, loss=1.594, acc=0.653, acc_len=1.52, thru=17.2, I=23.1, T=34.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:05,347] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:11,983] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:59:12,631] loop.py:45 INFO TRAIN_STEP step=7579 loss=1.365937 Training: 42%|████▏ | 7579/18018 [16:28:58<22:01:16, 7.59s/step, loss=1.366, acc=0.689, acc_len=1.65, thru=16.0, I=20.4, T=33.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:12,988] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 05:59:20,043] loop.py:380 INFO TIMING step=7580: step=3.860s data=1.773s compute=3.819s [fwd=2.699s bwd=1.106s opt=0.003s] dispatch=3.542s [2026-10-19 05:59:20,044] loop.py:45 INFO TRAIN_STEP step=7580 loss=1.442740 Training: 42%|████▏ | 7580/18018 [16:29:06<21:51:40, 7.54s/step, loss=1.443, acc=0.677, acc_len=1.61, thru=14.2, I=17.3, T=33.2, wait=3.5s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:28,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:59:28,575] loop.py:45 INFO TRAIN_STEP step=7581 loss=1.529146 Training: 42%|████▏ | 7581/18018 [16:29:14<22:43:17, 7.84s/step, loss=1.529, acc=0.665, acc_len=1.56, thru=17.4, I=16.6, T=31.5, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 05:59:28,571] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7580: forward=2836.9ms backward=1179.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:28,813] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 05:59:39,948] loop.py:45 INFO TRAIN_STEP step=7582 loss=1.580465 Training: 42%|████▏ | 7582/18018 [16:29:26<25:47:44, 8.90s/step, loss=1.580, acc=0.656, acc_len=1.53, thru=10.5, I=9.8, T=26.8, wait=6.6s, pool=36, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:47,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:59:49,529] loop.py:45 INFO TRAIN_STEP step=7583 loss=1.366534 Training: 42%|████▏ | 7583/18018 [16:29:35<26:23:01, 9.10s/step, loss=1.367, acc=0.690, acc_len=1.65, thru=16.4, I=17.4, T=25.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:49,796] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:57,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 05:59:58,526] loop.py:45 INFO TRAIN_STEP step=7584 loss=1.404571 Training: 42%|████▏ | 7584/18018 [16:29:44<26:17:23, 9.07s/step, loss=1.405, acc=0.680, acc_len=1.62, thru=14.7, I=16.5, T=25.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 05:59:58,608] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:00:05,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:00:05,764] loop.py:380 INFO TIMING step=7585: step=4.315s data=1.972s compute=4.274s [fwd=2.999s bwd=1.259s opt=0.003s] dispatch=2.915s [2026-10-19 06:00:05,764] loop.py:45 INFO TRAIN_STEP step=7585 loss=1.503164 Training: 42%|████▏ | 7585/18018 [16:29:51<24:41:42, 8.52s/step, loss=1.503, acc=0.665, acc_len=1.56, thru=15.7, I=18.0, T=29.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:00:06,081] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:00:11,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:00:12,947] loop.py:45 INFO TRAIN_STEP step=7586 loss=1.489370 Training: 42%|████▏ | 7586/18018 [16:29:59<23:31:38, 8.12s/step, loss=1.489, acc=0.671, acc_len=1.59, thru=16.3, I=19.3, T=29.8, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:00:12,942] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7585: forward=3024.9ms backward=1213.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:00:13,340] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:00:18,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:00:20,363] loop.py:45 INFO TRAIN_STEP step=7587 loss=1.295930 Training: 42%|████▏ | 7587/18018 [16:30:06<22:54:52, 7.91s/step, loss=1.296, acc=0.699, acc_len=1.69, thru=14.6, I=21.1, T=32.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:00:20,569] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:00:27,762] loop.py:45 INFO TRAIN_STEP step=7588 loss=1.402062 Training: 42%|████▏ | 7588/18018 [16:30:13<22:28:13, 7.76s/step, loss=1.402, acc=0.685, acc_len=1.63, thru=15.6, I=20.4, T=32.6, wait=3.5s, pool=60, epoch=1/2][2026-10-19 06:00:38,600] loop.py:45 INFO TRAIN_STEP step=7589 loss=1.491356 Training: 42%|████▏ | 7589/18018 [16:30:24<25:08:42, 8.68s/step, loss=1.491, acc=0.671, acc_len=1.58, thru=13.2, I=10.8, T=26.9, wait=6.1s, pool=60, epoch=1/2][2026-10-19 06:00:44,851] loop.py:380 INFO TIMING step=7590: step=3.904s data=1.787s compute=3.860s [fwd=2.737s bwd=1.107s opt=0.003s] dispatch=2.339s [2026-10-19 06:00:44,851] loop.py:45 INFO TRAIN_STEP step=7590 loss=1.390463 Training: 42%|████▏ | 7590/18018 [16:30:30<23:01:57, 7.95s/step, loss=1.390, acc=0.685, acc_len=1.63, thru=17.8, I=15.7, T=32.8, wait=2.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:00:51,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:00:52,670] loop.py:45 INFO TRAIN_STEP step=7591 loss=1.420676 Training: 42%|████▏ | 7591/18018 [16:30:38<22:54:56, 7.91s/step, loss=1.421, acc=0.683, acc_len=1.63, thru=18.6, I=19.3, T=30.3, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:00:52,665] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7590: forward=2985.7ms backward=1188.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:00:53,021] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:00:57,528] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:00:59,877] loop.py:45 INFO TRAIN_STEP step=7592 loss=1.537028 Training: 42%|████▏ | 7592/18018 [16:30:45<22:18:03, 7.70s/step, loss=1.537, acc=0.663, acc_len=1.56, thru=14.8, I=25.3, T=25.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:00,045] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:05,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:01:06,827] loop.py:45 INFO TRAIN_STEP step=7593 loss=1.353488 Training: 42%|████▏ | 7593/18018 [16:30:52<21:38:55, 7.48s/step, loss=1.353, acc=0.692, acc_len=1.66, thru=15.0, I=24.7, T=29.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:07,033] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:01:13,842] loop.py:45 INFO TRAIN_STEP step=7594 loss=1.445933 Training: 42%|████▏ | 7594/18018 [16:30:59<21:14:39, 7.34s/step, loss=1.446, acc=0.677, acc_len=1.61, thru=15.1, I=16.1, T=33.0, wait=3.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:20,377] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:01:21,806] loop.py:380 INFO TIMING step=7595: step=4.161s data=1.905s compute=4.119s [fwd=2.888s bwd=1.216s opt=0.003s] dispatch=3.796s [2026-10-19 06:01:21,806] loop.py:45 INFO TRAIN_STEP step=7595 loss=1.324194 Training: 42%|████▏ | 7595/18018 [16:31:07<21:47:14, 7.53s/step, loss=1.324, acc=0.700, acc_len=1.69, thru=18.2, I=18.9, T=30.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:21,889] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-19 06:01:28,263] loop.py:45 INFO TRAIN_STEP step=7596 loss=1.509352 Training: 42%|████▏ | 7596/18018 [16:31:14<20:51:22, 7.20s/step, loss=1.509, acc=0.666, acc_len=1.57, thru=17.4, I=21.8, T=30.9, wait=2.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:01:28,259] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7595: forward=2858.3ms backward=1232.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:35,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:01:37,116] loop.py:45 INFO TRAIN_STEP step=7597 loss=1.285912 Training: 42%|████▏ | 7597/18018 [16:31:23<22:17:20, 7.70s/step, loss=1.286, acc=0.706, acc_len=1.71, thru=15.7, I=16.1, T=30.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:37,315] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:42,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:01:43,795] loop.py:45 INFO TRAIN_STEP step=7598 loss=1.480026 Training: 42%|████▏ | 7598/18018 [16:31:29<21:24:00, 7.39s/step, loss=1.480, acc=0.672, acc_len=1.59, thru=17.9, I=24.0, T=30.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:44,231] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:49,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:01:51,281] loop.py:45 INFO TRAIN_STEP step=7599 loss=1.503874 Training: 42%|████▏ | 7599/18018 [16:31:37<21:28:42, 7.42s/step, loss=1.504, acc=0.666, acc_len=1.56, thru=14.2, I=20.6, T=32.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:51,357] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:57,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:01:58,079] loop.py:380 INFO TIMING step=7600: step=4.536s data=2.256s compute=4.506s [fwd=3.190s bwd=1.300s opt=0.003s] dispatch=2.252s [2026-10-19 06:01:58,079] loop.py:45 INFO TRAIN_STEP step=7600 loss=1.549326 Training: 42%|████▏ | 7600/18018 [16:31:44<20:56:00, 7.23s/step, loss=1.549, acc=0.658, acc_len=1.55, thru=17.9, I=21.4, T=28.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:01:58,319] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:03,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:02:05,171] loop.py:45 INFO TRAIN_STEP step=7601 loss=1.423824 Training: 42%|████▏ | 7601/18018 [16:31:51<20:48:33, 7.19s/step, loss=1.424, acc=0.682, acc_len=1.62, thru=18.5, I=20.3, T=29.5, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:02:05,167] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7600: forward=3030.2ms backward=1261.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:05,356] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:10,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:02:11,389] loop.py:45 INFO TRAIN_STEP step=7602 loss=1.352289 Training: 42%|████▏ | 7602/18018 [16:31:57<19:57:48, 6.90s/step, loss=1.352, acc=0.691, acc_len=1.65, thru=18.0, I=26.5, T=32.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:11,574] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:02:18,609] loop.py:45 INFO TRAIN_STEP step=7603 loss=1.434102 Training: 42%|████▏ | 7603/18018 [16:32:04<20:14:12, 6.99s/step, loss=1.434, acc=0.676, acc_len=1.61, thru=16.4, I=18.6, T=33.3, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:25,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:02:26,610] loop.py:45 INFO TRAIN_STEP step=7604 loss=1.476471 Training: 42%|████▏ | 7604/18018 [16:32:12<21:06:35, 7.30s/step, loss=1.476, acc=0.672, acc_len=1.59, thru=17.1, I=17.7, T=29.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:26,991] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:33,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:02:35,302] loop.py:380 INFO TIMING step=7605: step=4.566s data=2.146s compute=4.521s [fwd=3.156s bwd=1.350s opt=0.003s] dispatch=4.116s [2026-10-19 06:02:35,303] loop.py:45 INFO TRAIN_STEP step=7605 loss=1.384555 Training: 42%|████▏ | 7605/18018 [16:32:21<22:19:07, 7.72s/step, loss=1.385, acc=0.688, acc_len=1.64, thru=14.3, I=18.1, T=28.0, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:35,527] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:39,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:02:41,890] loop.py:45 INFO TRAIN_STEP step=7606 loss=1.328287 Training: 42%|████▏ | 7606/18018 [16:32:27<21:20:09, 7.38s/step, loss=1.328, acc=0.696, acc_len=1.67, thru=17.8, I=26.3, T=30.6, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:02:41,885] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7605: forward=2835.5ms backward=1289.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:42,294] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 06:02:50,042] loop.py:45 INFO TRAIN_STEP step=7607 loss=1.430985 Training: 42%|████▏ | 7607/18018 [16:32:36<22:00:26, 7.61s/step, loss=1.431, acc=0.680, acc_len=1.62, thru=15.0, I=15.8, T=31.0, wait=4.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:57,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:02:57,479] loop.py:45 INFO TRAIN_STEP step=7608 loss=1.437490 Training: 42%|████▏ | 7608/18018 [16:32:43<21:51:18, 7.56s/step, loss=1.437, acc=0.678, acc_len=1.61, thru=19.0, I=16.8, T=32.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:02:57,728] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:03,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:03:04,458] loop.py:45 INFO TRAIN_STEP step=7609 loss=1.333789 Training: 42%|████▏ | 7609/18018 [16:32:50<21:20:57, 7.38s/step, loss=1.334, acc=0.694, acc_len=1.66, thru=18.1, I=19.0, T=29.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:04,930] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:09,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:03:10,831] loop.py:380 INFO TIMING step=7610: step=3.888s data=1.710s compute=3.858s [fwd=2.812s bwd=1.031s opt=0.003s] dispatch=2.477s [2026-10-19 06:03:10,831] loop.py:45 INFO TRAIN_STEP step=7610 loss=1.428677 Training: 42%|████▏ | 7610/18018 [16:32:56<20:28:21, 7.08s/step, loss=1.429, acc=0.680, acc_len=1.62, thru=18.0, I=20.5, T=32.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:11,135] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:17,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:03:18,665] loop.py:45 INFO TRAIN_STEP step=7611 loss=1.518200 Training: 42%|████▏ | 7611/18018 [16:33:04<21:07:26, 7.31s/step, loss=1.518, acc=0.660, acc_len=1.55, thru=14.9, I=18.1, T=31.1, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:03:18,661] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7610: forward=2872.2ms backward=1187.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:18,859] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:25,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:03:25,518] loop.py:45 INFO TRAIN_STEP step=7612 loss=1.485766 Training: 42%|████▏ | 7612/18018 [16:33:11<20:43:39, 7.17s/step, loss=1.486, acc=0.670, acc_len=1.58, thru=17.2, I=20.2, T=29.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:25,784] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:32,066] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:03:33,066] loop.py:45 INFO TRAIN_STEP step=7613 loss=1.386057 Training: 42%|████▏ | 7613/18018 [16:33:19<21:03:12, 7.28s/step, loss=1.386, acc=0.688, acc_len=1.64, thru=15.3, I=18.2, T=32.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:33,100] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:38,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:03:40,156] loop.py:45 INFO TRAIN_STEP step=7614 loss=1.310431 Training: 42%|████▏ | 7614/18018 [16:33:26<20:52:48, 7.22s/step, loss=1.310, acc=0.699, acc_len=1.68, thru=18.5, I=22.0, T=29.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:40,650] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:45,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:03:46,939] loop.py:380 INFO TIMING step=7615: step=3.760s data=1.628s compute=3.732s [fwd=2.630s bwd=1.087s opt=0.003s] dispatch=3.016s [2026-10-19 06:03:46,939] loop.py:45 INFO TRAIN_STEP step=7615 loss=1.441065 Training: 42%|████▏ | 7615/18018 [16:33:33<20:29:41, 7.09s/step, loss=1.441, acc=0.674, acc_len=1.60, thru=16.4, I=24.3, T=34.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:47,229] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:54,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:03:54,299] loop.py:45 INFO TRAIN_STEP step=7616 loss=1.372159 Training: 42%|████▏ | 7616/18018 [16:33:40<20:43:30, 7.17s/step, loss=1.372, acc=0.691, acc_len=1.66, thru=16.1, I=20.7, T=33.1, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:03:54,294] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7615: forward=2636.2ms backward=1184.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:03:54,550] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:02,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:04:03,150] loop.py:45 INFO TRAIN_STEP step=7617 loss=1.481224 Training: 42%|████▏ | 7617/18018 [16:33:49<22:10:39, 7.68s/step, loss=1.481, acc=0.672, acc_len=1.59, thru=14.6, I=15.7, T=31.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:03,263] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 06:04:09,092] loop.py:45 INFO TRAIN_STEP step=7618 loss=1.453788 Training: 42%|████▏ | 7618/18018 [16:33:55<20:40:25, 7.16s/step, loss=1.454, acc=0.676, acc_len=1.60, thru=19.1, I=23.1, T=33.2, wait=2.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:15,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:04:16,482] loop.py:45 INFO TRAIN_STEP step=7619 loss=1.445835 Training: 42%|████▏ | 7619/18018 [16:34:02<20:52:32, 7.23s/step, loss=1.446, acc=0.675, acc_len=1.61, thru=15.8, I=18.0, T=29.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:16,709] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:23,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:04:23,679] loop.py:380 INFO TIMING step=7620: step=4.451s data=2.028s compute=4.414s [fwd=3.096s bwd=1.302s opt=0.003s] dispatch=2.737s [2026-10-19 06:04:23,679] loop.py:45 INFO TRAIN_STEP step=7620 loss=1.446209 Training: 42%|████▏ | 7620/18018 [16:34:09<20:50:46, 7.22s/step, loss=1.446, acc=0.678, acc_len=1.61, thru=16.4, I=19.6, T=28.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:23,837] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:27,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:04:29,824] loop.py:45 INFO TRAIN_STEP step=7621 loss=1.349965 Training: 42%|████▏ | 7621/18018 [16:34:15<19:55:00, 6.90s/step, loss=1.350, acc=0.693, acc_len=1.66, thru=18.1, I=21.7, T=30.6, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:04:29,820] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7620: forward=2999.0ms backward=1134.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:29,891] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:34,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:04:36,033] loop.py:45 INFO TRAIN_STEP step=7622 loss=1.487022 Training: 42%|████▏ | 7622/18018 [16:34:22<19:19:04, 6.69s/step, loss=1.487, acc=0.668, acc_len=1.58, thru=18.6, I=27.7, T=34.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:36,185] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:04:43,605] loop.py:45 INFO TRAIN_STEP step=7623 loss=1.475389 Training: 42%|████▏ | 7623/18018 [16:34:29<20:04:50, 6.95s/step, loss=1.475, acc=0.671, acc_len=1.57, thru=14.2, I=16.2, T=31.4, wait=3.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:52,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:04:54,166] loop.py:45 INFO TRAIN_STEP step=7624 loss=1.546576 Training: 42%|████▏ | 7624/18018 [16:34:40<23:12:13, 8.04s/step, loss=1.547, acc=0.658, acc_len=1.55, thru=12.9, I=14.7, T=26.6, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:54,281] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:04:59,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:05:01,156] loop.py:380 INFO TIMING step=7625: step=4.246s data=1.849s compute=4.218s [fwd=2.927s bwd=1.276s opt=0.003s] dispatch=2.733s [2026-10-19 06:05:01,156] loop.py:45 INFO TRAIN_STEP step=7625 loss=1.472229 Training: 42%|████▏ | 7625/18018 [16:34:47<22:17:40, 7.72s/step, loss=1.472, acc=0.668, acc_len=1.58, thru=16.5, I=21.0, T=30.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:01,330] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:07,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:05:09,603] loop.py:45 INFO TRAIN_STEP step=7626 loss=1.411792 Training: 42%|████▏ | 7626/18018 [16:34:55<22:55:13, 7.94s/step, loss=1.412, acc=0.683, acc_len=1.62, thru=13.5, I=18.8, T=29.5, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:05:09,599] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7625: forward=3054.2ms backward=1230.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:09,972] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:16,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:05:18,729] loop.py:45 INFO TRAIN_STEP step=7627 loss=1.405977 Training: 42%|████▏ | 7627/18018 [16:35:04<23:56:32, 8.29s/step, loss=1.406, acc=0.682, acc_len=1.62, thru=14.6, I=18.7, T=28.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:19,217] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 06:05:25,335] loop.py:45 INFO TRAIN_STEP step=7628 loss=1.328668 Training: 42%|████▏ | 7628/18018 [16:35:11<22:28:44, 7.79s/step, loss=1.329, acc=0.694, acc_len=1.67, thru=17.0, I=25.0, T=34.4, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:35,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:05:36,143] loop.py:45 INFO TRAIN_STEP step=7629 loss=1.421732 Training: 42%|████▏ | 7629/18018 [16:35:22<25:05:26, 8.69s/step, loss=1.422, acc=0.683, acc_len=1.63, thru=13.3, I=12.6, T=30.0, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:36,569] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:41,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:05:43,449] loop.py:380 INFO TIMING step=7630: step=4.284s data=1.980s compute=4.247s [fwd=2.989s bwd=1.244s opt=0.003s] dispatch=3.013s [2026-10-19 06:05:43,449] loop.py:45 INFO TRAIN_STEP step=7630 loss=1.367515 Training: 42%|████▏ | 7630/18018 [16:35:29<23:53:11, 8.28s/step, loss=1.368, acc=0.690, acc_len=1.65, thru=16.4, I=19.8, T=29.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:43,821] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:50,568] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:05:50,754] loop.py:45 INFO TRAIN_STEP step=7631 loss=1.407959 Training: 42%|████▏ | 7631/18018 [16:35:36<23:02:27, 7.99s/step, loss=1.408, acc=0.684, acc_len=1.63, thru=15.8, I=17.9, T=32.9, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:05:50,750] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7630: forward=2696.0ms backward=1147.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:51,070] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:57,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:05:58,229] loop.py:45 INFO TRAIN_STEP step=7632 loss=1.514877 Training: 42%|████▏ | 7632/18018 [16:35:44<22:35:53, 7.83s/step, loss=1.515, acc=0.665, acc_len=1.57, thru=17.7, I=18.5, T=32.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:05:58,497] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:06:04,056] loop.py:45 INFO TRAIN_STEP step=7633 loss=1.340120 Training: 42%|████▏ | 7633/18018 [16:35:50<20:51:35, 7.23s/step, loss=1.340, acc=0.693, acc_len=1.67, thru=19.2, I=24.0, T=36.7, wait=2.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:11,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:06:12,486] loop.py:45 INFO TRAIN_STEP step=7634 loss=1.421688 Training: 42%|████▏ | 7634/18018 [16:35:58<21:53:47, 7.59s/step, loss=1.422, acc=0.682, acc_len=1.63, thru=15.2, I=16.6, T=32.3, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:12,647] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:19,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:06:20,828] loop.py:380 INFO TIMING step=7635: step=4.356s data=1.965s compute=4.326s [fwd=3.052s bwd=1.259s opt=0.003s] dispatch=3.976s [2026-10-19 06:06:20,828] loop.py:45 INFO TRAIN_STEP step=7635 loss=1.370589 Training: 42%|████▏ | 7635/18018 [16:36:06<22:32:39, 7.82s/step, loss=1.371, acc=0.690, acc_len=1.65, thru=14.9, I=19.1, T=29.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:20,880] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-19 06:06:28,994] loop.py:45 INFO TRAIN_STEP step=7636 loss=1.489877 Training: 42%|████▏ | 7636/18018 [16:36:15<22:50:32, 7.92s/step, loss=1.490, acc=0.669, acc_len=1.58, thru=12.9, I=16.5, T=26.9, wait=3.4s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:06:28,989] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7635: forward=3343.5ms backward=1350.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:34,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:06:36,748] loop.py:45 INFO TRAIN_STEP step=7637 loss=1.489935 Training: 42%|████▏ | 7637/18018 [16:36:22<22:41:50, 7.87s/step, loss=1.490, acc=0.670, acc_len=1.58, thru=17.7, I=18.8, T=28.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:36,828] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:42,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:06:42,837] loop.py:45 INFO TRAIN_STEP step=7638 loss=1.515180 Training: 42%|████▏ | 7638/18018 [16:36:28<21:09:08, 7.34s/step, loss=1.515, acc=0.667, acc_len=1.57, thru=18.3, I=23.1, T=34.2, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:43,235] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:49,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:06:50,284] loop.py:45 INFO TRAIN_STEP step=7639 loss=1.479382 Training: 42%|████▏ | 7639/18018 [16:36:36<21:14:54, 7.37s/step, loss=1.479, acc=0.673, acc_len=1.59, thru=16.9, I=19.7, T=32.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:06:50,322] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 [2026-10-19 06:06:58,002] loop.py:380 INFO TIMING step=7640: step=3.995s data=1.753s compute=3.966s [fwd=2.802s bwd=1.149s opt=0.003s] dispatch=3.713s [2026-10-19 06:06:58,002] loop.py:45 INFO TRAIN_STEP step=7640 loss=1.321310 Training: 42%|████▏ | 7640/18018 [16:36:44<21:32:37, 7.47s/step, loss=1.321, acc=0.698, acc_len=1.68, thru=14.3, I=14.4, T=32.0, wait=3.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:07:05,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:07:08,690] loop.py:45 INFO TRAIN_STEP step=7641 loss=1.451329 Training: 42%|████▏ | 7641/18018 [16:36:54<24:19:29, 8.44s/step, loss=1.451, acc=0.676, acc_len=1.60, thru=14.1, I=20.5, T=22.1, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:07:08,685] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7640: forward=4366.9ms backward=1386.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:07:08,876] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 [2026-10-19 06:07:15,868] loop.py:45 INFO TRAIN_STEP step=7642 loss=1.555408 Training: 42%|████▏ | 7642/18018 [16:37:01<23:13:48, 8.06s/step, loss=1.555, acc=0.655, acc_len=1.53, thru=15.4, I=20.1, T=27.4, wait=2.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:07:22,092] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:07:23,708] loop.py:45 INFO TRAIN_STEP step=7643 loss=1.351743 Training: 42%|████▏ | 7643/18018 [16:37:09<23:02:20, 7.99s/step, loss=1.352, acc=0.694, acc_len=1.66, thru=16.4, I=17.2, T=30.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:07:24,110] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:07:30,553] loop.py:45 INFO TRAIN_STEP step=7644 loss=1.469627 Training: 42%|████▏ | 7644/18018 [16:37:16<22:02:31, 7.65s/step, loss=1.470, acc=0.670, acc_len=1.58, thru=14.3, I=18.4, T=35.2, wait=3.2s, pool=40, epoch=1/2][2026-10-19 06:07:40,847] loop.py:380 INFO TIMING step=7645: step=4.739s data=2.281s compute=4.709s [fwd=3.455s bwd=1.239s opt=0.003s] dispatch=5.546s [2026-10-19 06:07:40,847] loop.py:45 INFO TRAIN_STEP step=7645 loss=1.388785 Training: 42%|████▏ | 7645/18018 [16:37:26<24:19:36, 8.44s/step, loss=1.389, acc=0.693, acc_len=1.66, thru=12.9, I=13.5, T=27.0, wait=5.5s, pool=52, epoch=1/2][2026-10-19 06:07:49,354] loop.py:45 INFO TRAIN_STEP step=7646 loss=1.488148 Training: 42%|████▏ | 7646/18018 [16:37:35<24:22:47, 8.46s/step, loss=1.488, acc=0.671, acc_len=1.58, thru=15.1, I=14.2, T=32.0, wait=4.5s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:07:49,349] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7645: forward=2844.6ms backward=1108.8ms [2026-10-19 06:07:58,119] loop.py:45 INFO TRAIN_STEP step=7647 loss=1.445159 Training: 42%|████▏ | 7647/18018 [16:37:44<24:38:23, 8.55s/step, loss=1.445, acc=0.675, acc_len=1.60, thru=15.3, I=14.0, T=31.0, wait=4.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:05,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:08:06,790] loop.py:45 INFO TRAIN_STEP step=7648 loss=1.468806 Training: 42%|████▏ | 7648/18018 [16:37:52<24:44:24, 8.59s/step, loss=1.469, acc=0.675, acc_len=1.60, thru=15.6, I=16.6, T=28.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:07,114] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:11,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:08:13,260] loop.py:45 INFO TRAIN_STEP step=7649 loss=1.352337 Training: 42%|████▏ | 7649/18018 [16:37:59<22:54:24, 7.95s/step, loss=1.352, acc=0.693, acc_len=1.66, thru=18.0, I=22.5, T=33.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:13,370] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:18,868] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:08:19,103] loop.py:380 INFO TIMING step=7650: step=3.566s data=1.508s compute=3.532s [fwd=2.482s bwd=1.036s opt=0.003s] dispatch=2.265s [2026-10-19 06:08:19,103] loop.py:45 INFO TRAIN_STEP step=7650 loss=1.378586 Training: 42%|████▏ | 7650/18018 [16:38:05<21:04:57, 7.32s/step, loss=1.379, acc=0.686, acc_len=1.63, thru=20.4, I=22.5, T=35.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:19,369] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:08:25,607] loop.py:45 INFO TRAIN_STEP step=7651 loss=1.413265 Training: 42%|████▏ | 7651/18018 [16:38:11<20:22:22, 7.07s/step, loss=1.413, acc=0.682, acc_len=1.63, thru=15.4, I=19.1, T=37.3, wait=3.1s, pool=24, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:08:25,602] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7650: forward=2451.1ms backward=943.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:33,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:08:35,571] loop.py:45 INFO TRAIN_STEP step=7652 loss=1.480691 Training: 42%|████▏ | 7652/18018 [16:38:21<22:52:07, 7.94s/step, loss=1.481, acc=0.673, acc_len=1.59, thru=17.1, I=16.8, T=29.9, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:36,006] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:41,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:08:43,078] loop.py:45 INFO TRAIN_STEP step=7653 loss=1.413679 Training: 42%|████▏ | 7653/18018 [16:38:29<22:29:16, 7.81s/step, loss=1.414, acc=0.684, acc_len=1.62, thru=16.1, I=23.3, T=30.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:43,268] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:47,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:08:49,747] loop.py:45 INFO TRAIN_STEP step=7654 loss=1.492148 Training: 42%|████▏ | 7654/18018 [16:38:35<21:30:01, 7.47s/step, loss=1.492, acc=0.669, acc_len=1.58, thru=16.8, I=26.1, T=29.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:49,935] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:55,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:08:55,460] loop.py:380 INFO TIMING step=7655: step=3.453s data=1.450s compute=3.425s [fwd=2.426s bwd=0.985s opt=0.003s] dispatch=2.252s [2026-10-19 06:08:55,460] loop.py:45 INFO TRAIN_STEP step=7655 loss=1.287025 Training: 42%|████▏ | 7655/18018 [16:38:41<19:58:57, 6.94s/step, loss=1.287, acc=0.701, acc_len=1.69, thru=19.8, I=25.5, T=37.1, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:08:55,907] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:09:03,492] loop.py:45 INFO TRAIN_STEP step=7656 loss=1.526083 Training: 42%|████▏ | 7656/18018 [16:38:49<20:55:27, 7.27s/step, loss=1.526, acc=0.663, acc_len=1.55, thru=15.6, I=16.5, T=33.3, wait=4.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:09:03,488] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7655: forward=2581.9ms backward=1183.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:11,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:09:13,101] loop.py:45 INFO TRAIN_STEP step=7657 loss=1.566368 Training: 42%|████▏ | 7657/18018 [16:38:59<22:56:23, 7.97s/step, loss=1.566, acc=0.656, acc_len=1.53, thru=14.6, I=16.2, T=25.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:13,475] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:18,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:09:21,261] loop.py:45 INFO TRAIN_STEP step=7658 loss=1.316871 Training: 43%|████▎ | 7658/18018 [16:39:07<23:06:05, 8.03s/step, loss=1.317, acc=0.701, acc_len=1.69, thru=13.5, I=23.0, T=23.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:21,699] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:28,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:09:29,014] loop.py:45 INFO TRAIN_STEP step=7659 loss=1.350685 Training: 43%|████▎ | 7659/18018 [16:39:15<22:51:45, 7.95s/step, loss=1.351, acc=0.696, acc_len=1.68, thru=17.5, I=19.1, T=27.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:29,370] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:09:36,843] loop.py:380 INFO TIMING step=7660: step=4.239s data=1.953s compute=4.209s [fwd=2.986s bwd=1.210s opt=0.003s] dispatch=3.580s [2026-10-19 06:09:36,843] loop.py:45 INFO TRAIN_STEP step=7660 loss=1.388359 Training: 43%|████▎ | 7660/18018 [16:39:22<22:45:42, 7.91s/step, loss=1.388, acc=0.686, acc_len=1.64, thru=14.8, I=17.1, T=30.2, wait=3.6s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:09:45,447] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7660: forward=2797.0ms backward=1229.1ms [2026-10-19 06:09:45,452] loop.py:45 INFO TRAIN_STEP step=7661 loss=1.382840 Training: 43%|████▎ | 7661/18018 [16:39:31<23:21:33, 8.12s/step, loss=1.383, acc=0.692, acc_len=1.66, thru=14.6, I=13.2, T=31.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:45,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:45,901] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:52,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:09:53,538] loop.py:45 INFO TRAIN_STEP step=7662 loss=1.399433 Training: 43%|████▎ | 7662/18018 [16:39:39<23:19:49, 8.11s/step, loss=1.399, acc=0.687, acc_len=1.64, thru=17.1, I=18.4, T=28.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:53,598] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:09:59,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:10:01,532] loop.py:45 INFO TRAIN_STEP step=7663 loss=1.492506 Training: 43%|████▎ | 7663/18018 [16:39:47<23:13:34, 8.07s/step, loss=1.493, acc=0.672, acc_len=1.59, thru=14.4, I=18.9, T=28.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:01,906] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:06,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:10:07,688] loop.py:45 INFO TRAIN_STEP step=7664 loss=1.400051 Training: 43%|████▎ | 7664/18018 [16:39:53<21:34:12, 7.50s/step, loss=1.400, acc=0.686, acc_len=1.64, thru=17.5, I=25.7, T=32.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:07,851] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:15,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:10:16,245] loop.py:380 INFO TIMING step=7665: step=4.047s data=1.887s compute=4.011s [fwd=2.861s bwd=1.136s opt=0.003s] dispatch=4.498s [2026-10-19 06:10:16,246] loop.py:45 INFO TRAIN_STEP step=7665 loss=1.582484 Training: 43%|████▎ | 7665/18018 [16:40:02<22:28:47, 7.82s/step, loss=1.582, acc=0.651, acc_len=1.52, thru=14.0, I=15.7, T=31.6, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:16,651] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:25,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:10:25,483] loop.py:45 INFO TRAIN_STEP step=7666 loss=1.344468 (TrainerActor pid=9310) [2026-10-19 06:10:25,478] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7665: forward=3065.0ms backward=1192.2ms Training: 43%|████▎ | 7666/18018 [16:40:11<23:42:15, 8.24s/step, loss=1.344, acc=0.696, acc_len=1.67, thru=14.5, I=16.1, T=29.7, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:25,573] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:31,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:10:32,235] loop.py:45 INFO TRAIN_STEP step=7667 loss=1.390277 Training: 43%|████▎ | 7667/18018 [16:40:18<22:24:55, 7.80s/step, loss=1.390, acc=0.684, acc_len=1.62, thru=16.1, I=18.1, T=36.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:32,424] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:10:41,240] loop.py:45 INFO TRAIN_STEP step=7668 loss=1.358315 Training: 43%|████▎ | 7668/18018 [16:40:27<23:27:16, 8.16s/step, loss=1.358, acc=0.691, acc_len=1.66, thru=13.0, I=14.7, T=28.8, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:48,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:10:50,107] loop.py:45 INFO TRAIN_STEP step=7669 loss=1.424090 Training: 43%|████▎ | 7669/18018 [16:40:36<24:03:47, 8.37s/step, loss=1.424, acc=0.680, acc_len=1.62, thru=14.9, I=15.8, T=28.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:50,418] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:56,712] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:10:57,567] loop.py:380 INFO TIMING step=7670: step=3.860s data=1.669s compute=3.827s [fwd=2.765s bwd=1.048s opt=0.003s] dispatch=3.591s [2026-10-19 06:10:57,567] loop.py:45 INFO TRAIN_STEP step=7670 loss=1.386178 Training: 43%|████▎ | 7670/18018 [16:40:43<23:16:38, 8.10s/step, loss=1.386, acc=0.687, acc_len=1.64, thru=15.1, I=18.5, T=33.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:10:57,729] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:02,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:11:03,541] loop.py:45 INFO TRAIN_STEP step=7671 loss=1.325383 Training: 43%|████▎ | 7671/18018 [16:40:49<21:26:38, 7.46s/step, loss=1.325, acc=0.692, acc_len=1.66, thru=20.2, I=22.8, T=33.7, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:11:03,536] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7670: forward=2669.8ms backward=1080.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:03,777] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:11:10,414] loop.py:45 INFO TRAIN_STEP step=7672 loss=1.554789 Training: 43%|████▎ | 7672/18018 [16:40:56<20:55:55, 7.28s/step, loss=1.555, acc=0.657, acc_len=1.54, thru=16.9, I=19.9, T=32.9, wait=3.0s, pool=60, epoch=1/2][2026-10-19 06:11:18,217] loop.py:45 INFO TRAIN_STEP step=7673 loss=1.332082 Training: 43%|████▎ | 7673/18018 [16:41:04<21:22:43, 7.44s/step, loss=1.332, acc=0.699, acc_len=1.68, thru=15.5, I=15.2, T=31.5, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:25,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:11:26,283] loop.py:45 INFO TRAIN_STEP step=7674 loss=1.486125 Training: 43%|████▎ | 7674/18018 [16:41:12<21:55:01, 7.63s/step, loss=1.486, acc=0.671, acc_len=1.59, thru=16.6, I=17.4, T=29.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:26,342] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:32,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:11:34,097] loop.py:380 INFO TIMING step=7675: step=4.648s data=2.176s compute=4.599s [fwd=3.288s bwd=1.296s opt=0.003s] dispatch=3.159s [2026-10-19 06:11:34,097] loop.py:45 INFO TRAIN_STEP step=7675 loss=1.499630 Training: 43%|████▎ | 7675/18018 [16:41:20<22:04:36, 7.68s/step, loss=1.500, acc=0.673, acc_len=1.59, thru=14.1, I=16.3, T=27.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:34,306] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:39,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:11:41,178] loop.py:45 INFO TRAIN_STEP step=7676 loss=1.362909 Training: 43%|████▎ | 7676/18018 [16:41:27<21:33:16, 7.50s/step, loss=1.363, acc=0.692, acc_len=1.66, thru=15.7, I=24.5, T=28.9, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:11:41,173] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7675: forward=3179.0ms backward=1203.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:41,376] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:49,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:11:50,128] loop.py:45 INFO TRAIN_STEP step=7677 loss=1.304780 Training: 43%|████▎ | 7677/18018 [16:41:36<22:47:52, 7.94s/step, loss=1.305, acc=0.704, acc_len=1.70, thru=14.6, I=14.4, T=30.7, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:50,633] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:55,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:11:57,087] loop.py:45 INFO TRAIN_STEP step=7678 loss=1.471737 Training: 43%|████▎ | 7678/18018 [16:41:43<21:57:17, 7.64s/step, loss=1.472, acc=0.670, acc_len=1.58, thru=18.1, I=22.5, T=30.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:11:57,255] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:04,274] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:12:04,603] loop.py:45 INFO TRAIN_STEP step=7679 loss=1.358340 Training: 43%|████▎ | 7679/18018 [16:41:50<21:50:32, 7.61s/step, loss=1.358, acc=0.691, acc_len=1.65, thru=15.1, I=18.4, T=32.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:04,776] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:09,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:12:11,072] loop.py:380 INFO TIMING step=7680: step=4.217s data=1.928s compute=4.183s [fwd=2.914s bwd=1.255s opt=0.003s] dispatch=2.243s [2026-10-19 06:12:11,072] loop.py:45 INFO TRAIN_STEP step=7680 loss=1.405441 Training: 43%|████▎ | 7680/18018 [16:41:57<20:51:41, 7.26s/step, loss=1.405, acc=0.685, acc_len=1.63, thru=19.1, I=20.5, T=30.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:11,381] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:16,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:12:17,287] loop.py:45 INFO TRAIN_STEP step=7681 loss=1.464610 Training: 43%|████▎ | 7681/18018 [16:42:03<19:57:20, 6.95s/step, loss=1.465, acc=0.676, acc_len=1.60, thru=18.5, I=23.3, T=33.8, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:12:17,283] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7680: forward=2698.0ms backward=1054.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:17,612] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:22,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:12:23,552] loop.py:45 INFO TRAIN_STEP step=7682 loss=1.379753 Training: 43%|████▎ | 7682/18018 [16:42:09<19:21:48, 6.74s/step, loss=1.380, acc=0.688, acc_len=1.64, thru=19.2, I=24.5, T=34.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:23,724] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:32,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:12:33,857] loop.py:45 INFO TRAIN_STEP step=7683 loss=1.362744 Training: 43%|████▎ | 7683/18018 [16:42:19<22:25:35, 7.81s/step, loss=1.363, acc=0.694, acc_len=1.66, thru=12.5, I=13.7, T=26.5, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:34,294] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:39,659] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:12:41,952] loop.py:45 INFO TRAIN_STEP step=7684 loss=1.391672 Training: 43%|████▎ | 7684/18018 [16:42:28<22:40:15, 7.90s/step, loss=1.392, acc=0.686, acc_len=1.63, thru=14.4, I=18.6, T=24.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:42,172] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:46,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:12:47,687] loop.py:380 INFO TIMING step=7685: step=3.765s data=1.652s compute=3.734s [fwd=2.656s bwd=1.063s opt=0.003s] dispatch=1.959s [2026-10-19 06:12:47,688] loop.py:45 INFO TRAIN_STEP step=7685 loss=1.421598 Training: 43%|████▎ | 7685/18018 [16:42:33<20:48:22, 7.25s/step, loss=1.422, acc=0.678, acc_len=1.60, thru=17.4, I=24.1, T=34.0, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:47,957] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:53,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:12:55,021] loop.py:45 INFO TRAIN_STEP step=7686 loss=1.407463 Training: 43%|████▎ | 7686/18018 [16:42:41<20:52:39, 7.27s/step, loss=1.407, acc=0.681, acc_len=1.62, thru=16.6, I=22.7, T=28.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:12:55,016] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7685: forward=3158.1ms backward=1366.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:12:55,271] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:13:04,130] loop.py:45 INFO TRAIN_STEP step=7687 loss=1.475319 Training: 43%|████▎ | 7687/18018 [16:42:50<22:27:09, 7.82s/step, loss=1.475, acc=0.672, acc_len=1.58, thru=14.0, I=12.2, T=30.7, wait=4.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:11,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:13:13,675] loop.py:45 INFO TRAIN_STEP step=7688 loss=1.518499 Training: 43%|████▎ | 7688/18018 [16:42:59<23:55:56, 8.34s/step, loss=1.518, acc=0.662, acc_len=1.56, thru=14.4, I=16.9, T=28.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:13,705] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:20,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:13:22,002] loop.py:45 INFO TRAIN_STEP step=7689 loss=1.427935 Training: 43%|████▎ | 7689/18018 [16:43:08<23:55:04, 8.34s/step, loss=1.428, acc=0.682, acc_len=1.63, thru=15.5, I=19.4, T=28.9, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:22,480] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:28,450] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:13:28,639] loop.py:380 INFO TIMING step=7690: step=3.900s data=1.712s compute=3.869s [fwd=2.752s bwd=1.100s opt=0.003s] dispatch=2.730s [2026-10-19 06:13:28,639] loop.py:45 INFO TRAIN_STEP step=7690 loss=1.512134 Training: 43%|████▎ | 7690/18018 [16:43:14<22:27:18, 7.83s/step, loss=1.512, acc=0.664, acc_len=1.56, thru=16.2, I=19.8, T=32.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:28,952] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:35,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:13:36,880] loop.py:45 INFO TRAIN_STEP step=7691 loss=1.349135 Training: 43%|████▎ | 7691/18018 [16:43:22<22:48:27, 7.95s/step, loss=1.349, acc=0.694, acc_len=1.66, thru=15.2, I=17.2, T=27.1, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:13:36,876] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7690: forward=3184.4ms backward=1467.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:37,238] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:42,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:13:44,703] loop.py:45 INFO TRAIN_STEP step=7692 loss=1.402684 Training: 43%|████▎ | 7692/18018 [16:43:30<22:41:50, 7.91s/step, loss=1.403, acc=0.685, acc_len=1.63, thru=13.7, I=19.8, T=26.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:44,993] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 06:13:51,482] loop.py:45 INFO TRAIN_STEP step=7693 loss=1.360548 Training: 43%|████▎ | 7693/18018 [16:43:37<21:43:03, 7.57s/step, loss=1.361, acc=0.695, acc_len=1.67, thru=14.5, I=18.4, T=32.2, wait=2.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:13:58,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:14:00,716] loop.py:45 INFO TRAIN_STEP step=7694 loss=1.516551 Training: 43%|████▎ | 7694/18018 [16:43:46<23:08:48, 8.07s/step, loss=1.517, acc=0.665, acc_len=1.57, thru=15.8, I=18.1, T=29.2, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:00,905] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:07,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:14:07,712] loop.py:380 INFO TIMING step=7695: step=4.072s data=1.863s compute=4.042s [fwd=2.882s bwd=1.145s opt=0.003s] dispatch=2.912s [2026-10-19 06:14:07,712] loop.py:45 INFO TRAIN_STEP step=7695 loss=1.403950 Training: 43%|████▎ | 7695/18018 [16:43:53<22:13:08, 7.75s/step, loss=1.404, acc=0.686, acc_len=1.64, thru=16.1, I=20.4, T=31.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:07,891] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:15,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:14:16,606] loop.py:45 INFO TRAIN_STEP step=7696 loss=1.395241 Training: 43%|████▎ | 7696/18018 [16:44:02<23:12:04, 8.09s/step, loss=1.395, acc=0.689, acc_len=1.65, thru=14.5, I=17.0, T=29.5, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:14:16,601] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7695: forward=2877.7ms backward=1283.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:16,850] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:14:24,528] loop.py:45 INFO TRAIN_STEP step=7697 loss=1.476133 Training: 43%|████▎ | 7697/18018 [16:44:10<23:03:10, 8.04s/step, loss=1.476, acc=0.670, acc_len=1.58, thru=14.2, I=17.3, T=28.8, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:31,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:14:31,706] loop.py:45 INFO TRAIN_STEP step=7698 loss=1.386703 Training: 43%|████▎ | 7698/18018 [16:44:17<22:18:33, 7.78s/step, loss=1.387, acc=0.686, acc_len=1.64, thru=17.3, I=17.3, T=30.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:31,864] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:36,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:14:38,194] loop.py:45 INFO TRAIN_STEP step=7699 loss=1.311617 Training: 43%|████▎ | 7699/18018 [16:44:24<21:11:39, 7.39s/step, loss=1.312, acc=0.698, acc_len=1.68, thru=18.6, I=21.5, T=30.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:38,681] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:45,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:14:45,317] loop.py:380 INFO TIMING step=7700: step=3.916s data=1.811s compute=3.879s [fwd=2.757s bwd=1.108s opt=0.003s] dispatch=3.198s [2026-10-19 06:14:45,317] loop.py:45 INFO TRAIN_STEP step=7700 loss=1.406413 Training: 43%|████▎ | 7700/18018 [16:44:31<20:57:30, 7.31s/step, loss=1.406, acc=0.685, acc_len=1.63, thru=16.2, I=22.6, T=32.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:45,730] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:53,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:14:54,451] loop.py:45 INFO TRAIN_STEP step=7701 loss=1.581791 Training: 43%|████▎ | 7701/18018 [16:44:40<22:31:24, 7.86s/step, loss=1.582, acc=0.654, acc_len=1.53, thru=14.1, I=15.1, T=27.5, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:14:54,446] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7700: forward=3099.8ms backward=1478.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:14:54,697] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:01,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:15:01,927] loop.py:45 INFO TRAIN_STEP step=7702 loss=1.353806 Training: 43%|████▎ | 7702/18018 [16:44:48<22:11:31, 7.74s/step, loss=1.354, acc=0.693, acc_len=1.66, thru=15.7, I=18.1, T=30.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:02,092] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:06,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:15:08,344] loop.py:45 INFO TRAIN_STEP step=7703 loss=1.308467 Training: 43%|████▎ | 7703/18018 [16:44:54<21:02:54, 7.35s/step, loss=1.308, acc=0.697, acc_len=1.67, thru=18.5, I=22.0, T=30.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:08,472] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:15,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:15:15,713] loop.py:45 INFO TRAIN_STEP step=7704 loss=1.428600 Training: 43%|████▎ | 7704/18018 [16:45:01<21:03:54, 7.35s/step, loss=1.429, acc=0.683, acc_len=1.62, thru=15.7, I=18.9, T=30.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:16,024] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:22,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:15:23,426] loop.py:380 INFO TIMING step=7705: step=4.734s data=2.239s compute=4.690s [fwd=3.246s bwd=1.428s opt=0.003s] dispatch=2.970s [2026-10-19 06:15:23,426] loop.py:45 INFO TRAIN_STEP step=7705 loss=1.460657 Training: 43%|████▎ | 7705/18018 [16:45:09<21:22:26, 7.46s/step, loss=1.461, acc=0.675, acc_len=1.60, thru=16.7, I=18.3, T=27.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:23,768] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:28,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:15:29,626] loop.py:45 INFO TRAIN_STEP step=7706 loss=1.392226 Training: 43%|████▎ | 7706/18018 [16:45:15<20:17:12, 7.08s/step, loss=1.392, acc=0.687, acc_len=1.64, thru=18.3, I=21.8, T=31.8, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:15:29,621] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7705: forward=2848.8ms backward=1127.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:29,744] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:35,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:15:36,093] loop.py:45 INFO TRAIN_STEP step=7707 loss=1.517211 Training: 43%|████▎ | 7707/18018 [16:45:22<19:45:24, 6.90s/step, loss=1.517, acc=0.663, acc_len=1.55, thru=18.9, I=20.4, T=31.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:36,339] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:42,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:15:44,034] loop.py:45 INFO TRAIN_STEP step=7708 loss=1.254647 Training: 43%|████▎ | 7708/18018 [16:45:30<20:38:59, 7.21s/step, loss=1.255, acc=0.710, acc_len=1.72, thru=16.6, I=19.1, T=30.6, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:44,528] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:50,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:15:51,313] loop.py:45 INFO TRAIN_STEP step=7709 loss=1.412561 Training: 43%|████▎ | 7709/18018 [16:45:37<20:42:21, 7.23s/step, loss=1.413, acc=0.681, acc_len=1.62, thru=17.3, I=22.1, T=31.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:51,797] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:57,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:15:59,289] loop.py:380 INFO TIMING step=7710: step=4.165s data=1.900s compute=4.134s [fwd=2.973s bwd=1.146s opt=0.003s] dispatch=3.803s [2026-10-19 06:15:59,289] loop.py:45 INFO TRAIN_STEP step=7710 loss=1.392731 Training: 43%|████▎ | 7710/18018 [16:45:45<21:20:47, 7.46s/step, loss=1.393, acc=0.687, acc_len=1.64, thru=16.5, I=17.8, T=30.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:15:59,667] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:06,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:16:07,422] loop.py:45 INFO TRAIN_STEP step=7711 loss=1.310343 Training: 43%|████▎ | 7711/18018 [16:45:53<21:55:35, 7.66s/step, loss=1.310, acc=0.700, acc_len=1.68, thru=13.8, I=16.7, T=28.1, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:16:07,418] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7710: forward=3168.2ms backward=1329.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:07,742] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:16:15,912] loop.py:45 INFO TRAIN_STEP step=7712 loss=1.618482 Training: 43%|████▎ | 7712/18018 [16:46:02<22:38:26, 7.91s/step, loss=1.618, acc=0.649, acc_len=1.52, thru=13.0, I=17.2, T=28.1, wait=3.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:22,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:16:24,812] loop.py:45 INFO TRAIN_STEP step=7713 loss=1.426680 Training: 43%|████▎ | 7713/18018 [16:46:10<23:29:18, 8.21s/step, loss=1.427, acc=0.680, acc_len=1.62, thru=17.1, I=16.3, T=26.5, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:24,901] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:32,517] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:16:34,049] loop.py:45 INFO TRAIN_STEP step=7714 loss=1.315599 Training: 43%|████▎ | 7714/18018 [16:46:20<24:22:10, 8.51s/step, loss=1.316, acc=0.697, acc_len=1.67, thru=14.0, I=16.8, T=28.8, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:34,531] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:16:40,424] loop.py:380 INFO TIMING step=7715: step=3.774s data=1.712s compute=3.728s [fwd=2.594s bwd=1.119s opt=0.003s] dispatch=2.594s [2026-10-19 06:16:40,424] loop.py:45 INFO TRAIN_STEP step=7715 loss=1.325178 Training: 43%|████▎ | 7715/18018 [16:46:26<22:31:51, 7.87s/step, loss=1.325, acc=0.700, acc_len=1.68, thru=16.9, I=22.1, T=33.9, wait=2.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:47,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:16:49,470] loop.py:45 INFO TRAIN_STEP step=7716 loss=1.286349 Training: 43%|████▎ | 7716/18018 [16:46:35<23:32:11, 8.22s/step, loss=1.286, acc=0.708, acc_len=1.71, thru=16.1, I=17.4, T=27.5, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:16:49,466] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7715: forward=3272.3ms backward=1333.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:49,782] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:55,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:16:56,928] loop.py:45 INFO TRAIN_STEP step=7717 loss=1.320839 Training: 43%|████▎ | 7717/18018 [16:46:43<22:52:31, 7.99s/step, loss=1.321, acc=0.701, acc_len=1.69, thru=14.1, I=20.8, T=26.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:16:56,957] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:17:03,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:17:03,830] loop.py:45 INFO TRAIN_STEP step=7718 loss=1.562698 Training: 43%|████▎ | 7718/18018 [16:46:49<21:56:07, 7.67s/step, loss=1.563, acc=0.656, acc_len=1.54, thru=16.3, I=19.3, T=31.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:17:04,045] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:17:08,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:17:10,568] loop.py:45 INFO TRAIN_STEP step=7719 loss=1.350577 Training: 43%|████▎ | 7719/18018 [16:46:56<21:08:12, 7.39s/step, loss=1.351, acc=0.692, acc_len=1.65, thru=17.2, I=22.2, T=35.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:17:10,926] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:17:18,318] loop.py:380 INFO TIMING step=7720: step=4.384s data=1.985s compute=4.347s [fwd=2.925s bwd=1.406s opt=0.003s] dispatch=3.357s [2026-10-19 06:17:18,319] loop.py:45 INFO TRAIN_STEP step=7720 loss=1.416911 Training: 43%|████▎ | 7720/18018 [16:47:04<21:26:42, 7.50s/step, loss=1.417, acc=0.684, acc_len=1.62, thru=12.8, I=16.8, T=29.2, wait=3.4s, pool=48, epoch=1/2][2026-10-19 06:17:28,055] loop.py:45 INFO TRAIN_STEP step=7721 loss=1.357238 Training: 43%|████▎ | 7721/18018 [16:47:14<23:22:01, 8.17s/step, loss=1.357, acc=0.695, acc_len=1.66, thru=14.4, I=13.9, T=31.7, wait=5.7s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:17:28,051] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7720: forward=2912.8ms backward=1081.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:17:36,919] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:17:37,839] loop.py:45 INFO TRAIN_STEP step=7722 loss=1.351948 Training: 43%|████▎ | 7722/18018 [16:47:23<24:44:51, 8.65s/step, loss=1.352, acc=0.693, acc_len=1.66, thru=13.7, I=15.4, T=32.2, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:17:37,940] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 06:17:45,863] loop.py:45 INFO TRAIN_STEP step=7723 loss=1.505688 Training: 43%|████▎ | 7723/18018 [16:47:31<24:12:21, 8.46s/step, loss=1.506, acc=0.668, acc_len=1.57, thru=13.4, I=13.0, T=33.1, wait=4.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:17:52,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:17:52,808] loop.py:45 INFO TRAIN_STEP step=7724 loss=1.423073 Training: 43%|████▎ | 7724/18018 [16:47:38<22:54:00, 8.01s/step, loss=1.423, acc=0.681, acc_len=1.62, thru=18.6, I=17.8, T=34.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:17:53,251] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:17:59,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:18:00,537] loop.py:380 INFO TIMING step=7725: step=3.915s data=1.694s compute=3.887s [fwd=2.813s bwd=1.058s opt=0.003s] dispatch=3.804s [2026-10-19 06:18:00,537] loop.py:45 INFO TRAIN_STEP step=7725 loss=1.417525 Training: 43%|████▎ | 7725/18018 [16:47:46<22:39:26, 7.92s/step, loss=1.418, acc=0.681, acc_len=1.61, thru=16.5, I=19.6, T=32.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:00,958] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (TrainerActor pid=9310) [2026-10-19 06:18:07,799] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7725: forward=2958.0ms backward=1181.6ms [2026-10-19 06:18:07,807] loop.py:45 INFO TRAIN_STEP step=7726 loss=1.364557 Training: 43%|████▎ | 7726/18018 [16:47:53<22:05:46, 7.73s/step, loss=1.365, acc=0.693, acc_len=1.66, thru=16.1, I=19.9, T=30.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:07,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:08,288] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:18:15,238] loop.py:45 INFO TRAIN_STEP step=7727 loss=1.321065 Training: 43%|████▎ | 7727/18018 [16:48:01<21:50:12, 7.64s/step, loss=1.321, acc=0.695, acc_len=1.66, thru=16.0, I=16.5, T=31.3, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:21,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:18:22,654] loop.py:45 INFO TRAIN_STEP step=7728 loss=1.325724 Training: 43%|████▎ | 7728/18018 [16:48:08<21:38:40, 7.57s/step, loss=1.326, acc=0.697, acc_len=1.68, thru=17.2, I=17.8, T=29.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:22,826] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:29,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:18:30,358] loop.py:45 INFO TRAIN_STEP step=7729 loss=1.368490 Training: 43%|████▎ | 7729/18018 [16:48:16<21:45:11, 7.61s/step, loss=1.368, acc=0.692, acc_len=1.66, thru=17.3, I=20.0, T=31.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:30,844] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:36,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:18:37,858] loop.py:380 INFO TIMING step=7730: step=4.220s data=1.921s compute=4.168s [fwd=2.963s bwd=1.189s opt=0.003s] dispatch=3.272s [2026-10-19 06:18:37,858] loop.py:45 INFO TRAIN_STEP step=7730 loss=1.410108 Training: 43%|████▎ | 7730/18018 [16:48:23<21:39:23, 7.58s/step, loss=1.410, acc=0.682, acc_len=1.62, thru=16.5, I=21.7, T=30.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:38,257] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:18:45,122] loop.py:45 INFO TRAIN_STEP step=7731 loss=1.284709 Training: 43%|████▎ | 7731/18018 [16:48:31<21:23:08, 7.48s/step, loss=1.285, acc=0.703, acc_len=1.69, thru=12.4, I=19.3, T=29.9, wait=3.0s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:18:45,117] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7730: forward=3013.6ms backward=1209.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:53,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:18:55,343] loop.py:45 INFO TRAIN_STEP step=7732 loss=1.554121 Training: 43%|████▎ | 7732/18018 [16:48:41<23:43:48, 8.31s/step, loss=1.554, acc=0.663, acc_len=1.55, thru=14.6, I=15.9, T=27.9, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:18:55,625] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:01,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:19:01,482] loop.py:45 INFO TRAIN_STEP step=7733 loss=1.392942 Training: 43%|████▎ | 7733/18018 [16:48:47<21:52:14, 7.66s/step, loss=1.393, acc=0.686, acc_len=1.62, thru=17.5, I=22.5, T=33.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:01,845] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:06,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:19:08,744] loop.py:45 INFO TRAIN_STEP step=7734 loss=1.326159 Training: 43%|████▎ | 7734/18018 [16:48:54<21:31:56, 7.54s/step, loss=1.326, acc=0.696, acc_len=1.67, thru=18.3, I=22.5, T=27.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:09,104] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:16,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:19:17,483] loop.py:380 INFO TIMING step=7735: step=4.146s data=1.961s compute=4.118s [fwd=2.879s bwd=1.224s opt=0.003s] dispatch=4.583s [2026-10-19 06:19:17,483] loop.py:45 INFO TRAIN_STEP step=7735 loss=1.426109 Training: 43%|████▎ | 7735/18018 [16:49:03<22:33:35, 7.90s/step, loss=1.426, acc=0.680, acc_len=1.61, thru=15.3, I=17.9, T=30.9, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:17,646] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:23,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:19:24,820] loop.py:45 INFO TRAIN_STEP step=7736 loss=1.413790 Training: 43%|████▎ | 7736/18018 [16:49:10<22:04:38, 7.73s/step, loss=1.414, acc=0.684, acc_len=1.63, thru=16.4, I=19.3, T=30.4, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:19:24,817] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7735: forward=2963.0ms backward=1195.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:24,876] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:29,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:19:31,361] loop.py:45 INFO TRAIN_STEP step=7737 loss=1.412356 Training: 43%|████▎ | 7737/18018 [16:49:17<21:03:17, 7.37s/step, loss=1.412, acc=0.686, acc_len=1.64, thru=18.1, I=21.6, T=31.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:31,772] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:37,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:19:38,998] loop.py:45 INFO TRAIN_STEP step=7738 loss=1.386618 Training: 43%|████▎ | 7738/18018 [16:49:25<21:16:51, 7.45s/step, loss=1.387, acc=0.685, acc_len=1.63, thru=14.8, I=20.9, T=28.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:39,397] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:44,919] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:19:46,474] loop.py:45 INFO TRAIN_STEP step=7739 loss=1.498620 Training: 43%|████▎ | 7739/18018 [16:49:32<21:17:47, 7.46s/step, loss=1.499, acc=0.665, acc_len=1.57, thru=15.2, I=22.8, T=31.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:46,933] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:54,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:19:54,771] loop.py:380 INFO TIMING step=7740: step=4.385s data=2.089s compute=4.342s [fwd=3.039s bwd=1.287s opt=0.003s] dispatch=3.904s [2026-10-19 06:19:54,771] loop.py:45 INFO TRAIN_STEP step=7740 loss=1.538750 Training: 43%|████▎ | 7740/18018 [16:49:40<22:00:53, 7.71s/step, loss=1.539, acc=0.661, acc_len=1.56, thru=13.7, I=17.9, T=29.2, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:19:55,012] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:20:01,442] loop.py:45 INFO TRAIN_STEP step=7741 loss=1.358533 Training: 43%|████▎ | 7741/18018 [16:49:47<21:07:18, 7.40s/step, loss=1.359, acc=0.690, acc_len=1.66, thru=16.2, I=19.2, T=32.1, wait=2.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:20:01,437] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7740: forward=2749.8ms backward=1187.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:10,097] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:20:10,459] loop.py:45 INFO TRAIN_STEP step=7742 loss=1.431990 Training: 43%|████▎ | 7742/18018 [16:49:56<22:30:15, 7.88s/step, loss=1.432, acc=0.679, acc_len=1.61, thru=15.3, I=15.5, T=28.0, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:10,607] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:20:19,034] loop.py:45 INFO TRAIN_STEP step=7743 loss=1.336534 Training: 43%|████▎ | 7743/18018 [16:50:05<23:05:38, 8.09s/step, loss=1.337, acc=0.693, acc_len=1.67, thru=11.4, I=13.9, T=30.8, wait=4.4s, pool=32, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 06:20:19,135] eagle_store.py:517 WARNING batch_get_buffer missing keys (k247775-a79d0927_hs, k247775-a79d0927_ids, k247775-a79d0927_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:27,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 06:20:19,231] eagle_store.py:517 WARNING batch_get_buffer missing keys (k247774-bbf441b2_hs, k247774-bbf441b2_ids, k247774-bbf441b2_lhs); sleeping 0.50s. [2026-10-19 06:20:29,333] loop.py:45 INFO TRAIN_STEP step=7744 loss=1.531277 Training: 43%|████▎ | 7744/18018 [16:50:15<24:58:49, 8.75s/step, loss=1.531, acc=0.669, acc_len=1.58, thru=16.0, I=15.9, T=26.6, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:29,834] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:35,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:20:36,019] loop.py:380 INFO TIMING step=7745: step=3.928s data=1.815s compute=3.890s [fwd=2.771s bwd=1.103s opt=0.003s] dispatch=2.750s [2026-10-19 06:20:36,019] loop.py:45 INFO TRAIN_STEP step=7745 loss=1.404674 Training: 43%|████▎ | 7745/18018 [16:50:22<23:12:39, 8.13s/step, loss=1.405, acc=0.684, acc_len=1.63, thru=15.8, I=22.4, T=32.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:36,203] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:44,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:20:44,568] loop.py:45 INFO TRAIN_STEP step=7746 loss=1.612510 Training: 43%|████▎ | 7746/18018 [16:50:30<23:33:44, 8.26s/step, loss=1.613, acc=0.650, acc_len=1.52, thru=15.0, I=14.5, T=30.1, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:20:44,563] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7745: forward=2853.6ms backward=1329.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:44,672] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:50,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:20:51,709] loop.py:45 INFO TRAIN_STEP step=7747 loss=1.514861 Training: 43%|████▎ | 7747/18018 [16:50:37<22:36:22, 7.92s/step, loss=1.515, acc=0.667, acc_len=1.58, thru=16.5, I=18.3, T=34.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:51,881] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:58,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:20:59,920] loop.py:45 INFO TRAIN_STEP step=7748 loss=1.406925 Training: 43%|████▎ | 7748/18018 [16:50:46<22:50:58, 8.01s/step, loss=1.407, acc=0.682, acc_len=1.62, thru=15.2, I=18.4, T=26.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:20:59,973] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 06:21:07,390] loop.py:45 INFO TRAIN_STEP step=7749 loss=1.356427 Training: 43%|████▎ | 7749/18018 [16:50:53<22:23:07, 7.85s/step, loss=1.356, acc=0.690, acc_len=1.65, thru=14.3, I=17.7, T=27.8, wait=2.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:16,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:21:17,552] loop.py:380 INFO TIMING step=7750: step=4.804s data=2.161s compute=4.775s [fwd=3.340s bwd=1.420s opt=0.003s] dispatch=5.348s [2026-10-19 06:21:17,552] loop.py:45 INFO TRAIN_STEP step=7750 loss=1.565951 Training: 43%|████▎ | 7750/18018 [16:51:03<24:21:48, 8.54s/step, loss=1.566, acc=0.658, acc_len=1.54, thru=14.6, I=13.7, T=26.6, wait=5.3s, pool=64, epoch=1/2][2026-10-19 06:21:17,557] loop.py:419 INFO Saving checkpoint at step 7750... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 06:21:21,255] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0007751 (TrainerActor pid=9310) [2026-10-19 06:21:21,264] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0007001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:21,627] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:22,648] inference_manager.py:408 INFO Pool capacity available after 6.1s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:30,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 06:21:21,254] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0007751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 06:21:22,489] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0007001 [2026-10-19 06:21:32,173] loop.py:45 INFO TRAIN_STEP step=7751 loss=1.424628 Training: 43%|████▎ | 7751/18018 [16:51:18<29:33:42, 10.37s/step, loss=1.425, acc=0.682, acc_len=1.62, thru=13.5, I=15.3, T=23.7, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:21:32,168] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7750: forward=4023.9ms backward=1336.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:32,663] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:38,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:21:40,351] loop.py:45 INFO TRAIN_STEP step=7752 loss=1.593525 Training: 43%|████▎ | 7752/18018 [16:51:26<27:41:11, 9.71s/step, loss=1.594, acc=0.653, acc_len=1.52, thru=13.6, I=18.7, T=26.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:40,553] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:47,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:21:47,911] loop.py:45 INFO TRAIN_STEP step=7753 loss=1.387076 Training: 43%|████▎ | 7753/18018 [16:51:33<25:50:45, 9.06s/step, loss=1.387, acc=0.685, acc_len=1.63, thru=14.0, I=16.2, T=29.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:48,062] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:53,736] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:21:55,762] loop.py:45 INFO TRAIN_STEP step=7754 loss=1.482584 Training: 43%|████▎ | 7754/18018 [16:51:41<24:48:18, 8.70s/step, loss=1.483, acc=0.673, acc_len=1.59, thru=16.3, I=20.8, T=27.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:21:56,254] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:22:03,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:22:04,164] loop.py:380 INFO TIMING step=7755: step=4.851s data=2.266s compute=4.813s [fwd=3.165s bwd=1.632s opt=0.003s] dispatch=3.542s [2026-10-19 06:22:04,164] loop.py:45 INFO TRAIN_STEP step=7755 loss=1.272194 Training: 43%|████▎ | 7755/18018 [16:51:50<24:32:53, 8.61s/step, loss=1.272, acc=0.709, acc_len=1.72, thru=16.2, I=16.8, T=26.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:22:04,376] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:22:10,672] loop.py:45 INFO TRAIN_STEP step=7756 loss=1.415014 Training: 43%|████▎ | 7756/18018 [16:51:56<22:44:53, 7.98s/step, loss=1.415, acc=0.682, acc_len=1.62, thru=17.4, I=17.6, T=30.6, wait=2.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:22:10,667] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7755: forward=2928.7ms backward=1204.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:22:15,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:22:17,597] loop.py:45 INFO TRAIN_STEP step=7757 loss=1.446579 Training: 43%|████▎ | 7757/18018 [16:52:03<21:50:39, 7.66s/step, loss=1.447, acc=0.677, acc_len=1.61, thru=17.6, I=20.8, T=29.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:22:17,977] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:22:24,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:22:24,758] loop.py:45 INFO TRAIN_STEP step=7758 loss=1.462720 Training: 43%|████▎ | 7758/18018 [16:52:10<21:24:39, 7.51s/step, loss=1.463, acc=0.678, acc_len=1.61, thru=16.1, I=18.8, T=33.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:22:24,829] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-19 06:22:32,530] loop.py:45 INFO TRAIN_STEP step=7759 loss=1.388696 Training: 43%|████▎ | 7759/18018 [16:52:18<21:37:48, 7.59s/step, loss=1.389, acc=0.685, acc_len=1.63, thru=15.0, I=15.1, T=33.0, wait=3.9s, pool=52, epoch=1/2][2026-10-19 06:22:39,768] loop.py:380 INFO TIMING step=7760: step=4.358s data=2.033s compute=4.329s [fwd=3.063s bwd=1.250s opt=0.003s] dispatch=2.871s [2026-10-19 06:22:39,768] loop.py:45 INFO TRAIN_STEP step=7760 loss=1.447972 Training: 43%|████▎ | 7760/18018 [16:52:25<21:19:46, 7.49s/step, loss=1.448, acc=0.677, acc_len=1.60, thru=16.9, I=16.0, T=29.4, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:22:46,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:22:48,026] loop.py:45 INFO TRAIN_STEP step=7761 loss=1.434463 Training: 43%|████▎ | 7761/18018 [16:52:34<21:59:13, 7.72s/step, loss=1.434, acc=0.680, acc_len=1.62, thru=17.2, I=17.7, T=29.9, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:22:48,022] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7760: forward=2971.5ms backward=1236.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:22:48,067] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 06:22:55,673] loop.py:45 INFO TRAIN_STEP step=7762 loss=1.385183 Training: 43%|████▎ | 7762/18018 [16:52:41<21:55:23, 7.70s/step, loss=1.385, acc=0.689, acc_len=1.65, thru=13.0, I=15.9, T=28.9, wait=3.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:01,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:23:04,629] loop.py:45 INFO TRAIN_STEP step=7763 loss=1.437481 Training: 43%|████▎ | 7763/18018 [16:52:50<22:59:59, 8.07s/step, loss=1.437, acc=0.678, acc_len=1.61, thru=16.1, I=20.0, T=22.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:04,721] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=40/64 [2026-10-19 06:23:11,492] loop.py:45 INFO TRAIN_STEP step=7764 loss=1.381263 Training: 43%|████▎ | 7764/18018 [16:52:57<21:57:51, 7.71s/step, loss=1.381, acc=0.688, acc_len=1.65, thru=17.2, I=18.7, T=28.8, wait=2.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:17,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:23:19,761] loop.py:380 INFO TIMING step=7765: step=4.688s data=2.225s compute=4.660s [fwd=3.240s bwd=1.406s opt=0.003s] dispatch=3.569s [2026-10-19 06:23:19,761] loop.py:45 INFO TRAIN_STEP step=7765 loss=1.294798 Training: 43%|████▎ | 7765/18018 [16:53:05<22:26:13, 7.88s/step, loss=1.295, acc=0.702, acc_len=1.69, thru=16.6, I=18.6, T=27.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:19,927] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:27,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:23:28,183] loop.py:45 INFO TRAIN_STEP step=7766 loss=1.495587 Training: 43%|████▎ | 7766/18018 [16:53:14<22:53:55, 8.04s/step, loss=1.496, acc=0.668, acc_len=1.58, thru=15.5, I=15.9, T=27.5, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:23:28,179] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7765: forward=3191.6ms backward=1408.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:28,617] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:33,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:23:35,759] loop.py:45 INFO TRAIN_STEP step=7767 loss=1.442829 Training: 43%|████▎ | 7767/18018 [16:53:21<22:29:56, 7.90s/step, loss=1.443, acc=0.677, acc_len=1.61, thru=15.9, I=19.1, T=28.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:35,946] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:40,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:23:42,808] loop.py:45 INFO TRAIN_STEP step=7768 loss=1.386363 Training: 43%|████▎ | 7768/18018 [16:53:28<21:46:11, 7.65s/step, loss=1.386, acc=0.682, acc_len=1.62, thru=15.2, I=25.8, T=26.7, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:43,123] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:48,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:23:49,982] loop.py:45 INFO TRAIN_STEP step=7769 loss=1.382406 Training: 43%|████▎ | 7769/18018 [16:53:36<21:21:46, 7.50s/step, loss=1.382, acc=0.684, acc_len=1.63, thru=14.0, I=21.1, T=30.1, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:50,380] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:57,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:23:57,952] loop.py:380 INFO TIMING step=7770: step=4.695s data=2.424s compute=4.665s [fwd=3.373s bwd=1.278s opt=0.003s] dispatch=3.267s [2026-10-19 06:23:57,952] loop.py:45 INFO TRAIN_STEP step=7770 loss=1.484979 Training: 43%|████▎ | 7770/18018 [16:53:44<21:45:35, 7.64s/step, loss=1.485, acc=0.671, acc_len=1.59, thru=15.1, I=17.5, T=27.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:23:58,354] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:24:02,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:24:05,279] loop.py:45 INFO TRAIN_STEP step=7771 loss=1.401372 Training: 43%|████▎ | 7771/18018 [16:53:51<21:29:15, 7.55s/step, loss=1.401, acc=0.684, acc_len=1.64, thru=17.1, I=20.6, T=26.9, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:24:05,274] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7770: forward=3256.1ms backward=1438.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:24:05,315] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 [2026-10-19 06:24:11,655] loop.py:45 INFO TRAIN_STEP step=7772 loss=1.307013 Training: 43%|████▎ | 7772/18018 [16:53:57<20:28:58, 7.20s/step, loss=1.307, acc=0.704, acc_len=1.69, thru=13.9, I=19.2, T=29.9, wait=2.1s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:24:20,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:24:22,933] loop.py:45 INFO TRAIN_STEP step=7773 loss=1.537233 Training: 43%|████▎ | 7773/18018 [16:54:09<23:58:01, 8.42s/step, loss=1.537, acc=0.666, acc_len=1.57, thru=15.6, I=17.4, T=26.4, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:24:23,333] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 06:24:30,586] loop.py:45 INFO TRAIN_STEP step=7774 loss=1.324960 Training: 43%|████▎ | 7774/18018 [16:54:16<23:18:24, 8.19s/step, loss=1.325, acc=0.698, acc_len=1.67, thru=15.4, I=17.6, T=34.4, wait=3.9s, pool=48, epoch=1/2][2026-10-19 06:24:38,087] loop.py:380 INFO TIMING step=7775: step=4.128s data=1.922s compute=4.096s [fwd=2.948s bwd=1.133s opt=0.003s] dispatch=3.364s [2026-10-19 06:24:38,087] loop.py:45 INFO TRAIN_STEP step=7775 loss=1.412032 Training: 43%|████▎ | 7775/18018 [16:54:24<22:43:00, 7.98s/step, loss=1.412, acc=0.682, acc_len=1.62, thru=13.6, I=16.1, T=31.0, wait=3.4s, pool=24, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:24:48,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:24:49,865] loop.py:45 INFO TRAIN_STEP step=7776 loss=1.477464 Training: 43%|████▎ | 7776/18018 [16:54:35<25:57:11, 9.12s/step, loss=1.477, acc=0.672, acc_len=1.59, thru=13.3, I=16.7, T=25.7, wait=6.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:24:49,861] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7775: forward=3512.4ms backward=1423.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:24:50,134] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:24:56,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:24:57,128] loop.py:45 INFO TRAIN_STEP step=7777 loss=1.471232 Training: 43%|████▎ | 7777/18018 [16:54:43<24:21:42, 8.56s/step, loss=1.471, acc=0.675, acc_len=1.60, thru=15.8, I=20.4, T=32.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:24:57,353] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:02,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:25:05,305] loop.py:45 INFO TRAIN_STEP step=7778 loss=1.326610 Training: 43%|████▎ | 7778/18018 [16:54:51<24:01:47, 8.45s/step, loss=1.327, acc=0.700, acc_len=1.69, thru=15.1, I=20.2, T=26.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:05,738] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:11,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:25:12,346] loop.py:45 INFO TRAIN_STEP step=7779 loss=1.385503 Training: 43%|████▎ | 7779/18018 [16:54:58<22:49:41, 8.03s/step, loss=1.386, acc=0.687, acc_len=1.64, thru=14.3, I=23.6, T=30.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:12,752] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:18,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:25:20,166] loop.py:380 INFO TIMING step=7780: step=4.786s data=2.386s compute=4.739s [fwd=3.324s bwd=1.401s opt=0.003s] dispatch=3.024s [2026-10-19 06:25:20,166] loop.py:45 INFO TRAIN_STEP step=7780 loss=1.339410 Training: 43%|████▎ | 7780/18018 [16:55:06<22:38:49, 7.96s/step, loss=1.339, acc=0.694, acc_len=1.67, thru=16.0, I=20.8, T=26.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:20,194] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:26,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:25:26,918] loop.py:45 INFO TRAIN_STEP step=7781 loss=1.523297 Training: 43%|████▎ | 7781/18018 [16:55:13<21:36:46, 7.60s/step, loss=1.523, acc=0.668, acc_len=1.57, thru=16.8, I=20.4, T=29.1, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:25:26,913] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7780: forward=3016.4ms backward=1297.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:27,365] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:25:39,436] loop.py:45 INFO TRAIN_STEP step=7782 loss=1.541189 Training: 43%|████▎ | 7782/18018 [16:55:25<25:48:22, 9.08s/step, loss=1.541, acc=0.661, acc_len=1.55, thru=7.3, I=6.8, T=30.9, wait=8.4s, pool=24, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:49,341] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:25:51,133] loop.py:45 INFO TRAIN_STEP step=7783 loss=1.488313 Training: 43%|████▎ | 7783/18018 [16:55:37<28:02:21, 9.86s/step, loss=1.488, acc=0.669, acc_len=1.58, thru=13.8, I=16.7, T=28.6, wait=7.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:51,352] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:25:59,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:26:00,455] loop.py:45 INFO TRAIN_STEP step=7784 loss=1.442321 Training: 43%|████▎ | 7784/18018 [16:55:46<27:34:27, 9.70s/step, loss=1.442, acc=0.675, acc_len=1.60, thru=14.1, I=14.8, T=26.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:00,915] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:05,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:26:06,488] loop.py:380 INFO TIMING step=7785: step=3.838s data=1.546s compute=3.806s [fwd=2.560s bwd=1.231s opt=0.003s] dispatch=2.186s [2026-10-19 06:26:06,488] loop.py:45 INFO TRAIN_STEP step=7785 loss=1.387692 Training: 43%|████▎ | 7785/18018 [16:55:52<24:26:46, 8.60s/step, loss=1.388, acc=0.681, acc_len=1.62, thru=19.0, I=19.7, T=33.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:06,791] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:16,107] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:26:16,490] loop.py:45 INFO TRAIN_STEP step=7786 loss=1.390207 Training: 43%|████▎ | 7786/18018 [16:56:02<25:38:19, 9.02s/step, loss=1.390, acc=0.685, acc_len=1.63, thru=13.2, I=13.0, T=24.2, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:26:16,486] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7785: forward=3617.8ms backward=1609.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:16,622] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:22,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:26:24,155] loop.py:45 INFO TRAIN_STEP step=7787 loss=1.373639 Training: 43%|████▎ | 7787/18018 [16:56:10<24:28:45, 8.61s/step, loss=1.374, acc=0.692, acc_len=1.66, thru=17.4, I=17.9, T=27.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:24,390] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:29,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:26:30,437] loop.py:45 INFO TRAIN_STEP step=7788 loss=1.423400 Training: 43%|████▎ | 7788/18018 [16:56:16<22:29:25, 7.91s/step, loss=1.423, acc=0.683, acc_len=1.63, thru=18.0, I=22.9, T=32.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:30,905] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:37,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:26:37,305] loop.py:45 INFO TRAIN_STEP step=7789 loss=1.407073 Training: 43%|████▎ | 7789/18018 [16:56:23<21:35:42, 7.60s/step, loss=1.407, acc=0.686, acc_len=1.64, thru=18.1, I=18.7, T=30.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:37,648] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:45,047] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:26:45,531] loop.py:380 INFO TIMING step=7790: step=4.735s data=2.402s compute=4.693s [fwd=3.244s bwd=1.434s opt=0.003s] dispatch=3.481s [2026-10-19 06:26:45,531] loop.py:45 INFO TRAIN_STEP step=7790 loss=1.385479 Training: 43%|████▎ | 7790/18018 [16:56:31<22:07:39, 7.79s/step, loss=1.385, acc=0.688, acc_len=1.64, thru=16.6, I=16.6, T=27.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:45,564] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:53,974] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:26:55,166] loop.py:45 INFO TRAIN_STEP step=7791 loss=1.458369 Training: 43%|████▎ | 7791/18018 [16:56:41<23:41:57, 8.34s/step, loss=1.458, acc=0.675, acc_len=1.59, thru=13.4, I=14.4, T=27.8, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:26:55,161] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7790: forward=3288.6ms backward=1272.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:26:55,483] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:00,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:27:01,744] loop.py:45 INFO TRAIN_STEP step=7792 loss=1.378093 Training: 43%|████▎ | 7792/18018 [16:56:47<22:11:34, 7.81s/step, loss=1.378, acc=0.687, acc_len=1.64, thru=17.7, I=20.2, T=31.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:02,048] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:07,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:27:08,385] loop.py:45 INFO TRAIN_STEP step=7793 loss=1.319435 Training: 43%|████▎ | 7793/18018 [16:56:54<21:11:28, 7.46s/step, loss=1.319, acc=0.695, acc_len=1.67, thru=18.2, I=22.4, T=29.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:08,547] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:13,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:27:15,383] loop.py:45 INFO TRAIN_STEP step=7794 loss=1.423900 Training: 43%|████▎ | 7794/18018 [16:57:01<20:47:48, 7.32s/step, loss=1.424, acc=0.676, acc_len=1.60, thru=16.9, I=20.7, T=28.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:15,418] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:20,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:27:21,613] loop.py:380 INFO TIMING step=7795: step=4.167s data=1.872s compute=4.144s [fwd=2.911s bwd=1.218s opt=0.003s] dispatch=2.052s [2026-10-19 06:27:21,613] loop.py:45 INFO TRAIN_STEP step=7795 loss=1.384330 Training: 43%|████▎ | 7795/18018 [16:57:07<19:51:46, 6.99s/step, loss=1.384, acc=0.688, acc_len=1.64, thru=17.6, I=23.2, T=30.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:22,024] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:28,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:27:29,728] loop.py:45 INFO TRAIN_STEP step=7796 loss=1.449090 (TrainerActor pid=9310) [2026-10-19 06:27:29,722] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7795: forward=3008.4ms backward=1149.5ms Training: 43%|████▎ | 7796/18018 [16:57:15<20:48:56, 7.33s/step, loss=1.449, acc=0.672, acc_len=1.59, thru=14.4, I=18.7, T=30.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:29,883] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:35,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:27:37,487] loop.py:45 INFO TRAIN_STEP step=7797 loss=1.469480 Training: 43%|████▎ | 7797/18018 [16:57:23<21:10:43, 7.46s/step, loss=1.469, acc=0.673, acc_len=1.59, thru=15.6, I=22.8, T=27.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:37,755] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:27:43,589] loop.py:45 INFO TRAIN_STEP step=7798 loss=1.410415 Training: 43%|████▎ | 7798/18018 [16:57:29<20:01:19, 7.05s/step, loss=1.410, acc=0.682, acc_len=1.62, thru=16.4, I=21.5, T=33.6, wait=2.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:53,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:27:55,505] loop.py:45 INFO TRAIN_STEP step=7799 loss=1.380940 Training: 43%|████▎ | 7799/18018 [16:57:41<24:09:35, 8.51s/step, loss=1.381, acc=0.692, acc_len=1.66, thru=13.0, I=14.2, T=24.4, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:27:55,713] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:01,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:28:02,155] loop.py:380 INFO TIMING step=7800: step=4.324s data=2.077s compute=4.298s [fwd=3.045s bwd=1.237s opt=0.003s] dispatch=2.316s [2026-10-19 06:28:02,155] loop.py:45 INFO TRAIN_STEP step=7800 loss=1.327750 Training: 43%|████▎ | 7800/18018 [16:57:48<22:34:16, 7.95s/step, loss=1.328, acc=0.702, acc_len=1.69, thru=15.4, I=23.5, T=29.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:02,557] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:07,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:28:08,924] loop.py:45 INFO TRAIN_STEP step=7801 loss=1.462896 Training: 43%|████▎ | 7801/18018 [16:57:55<21:33:45, 7.60s/step, loss=1.463, acc=0.676, acc_len=1.60, thru=16.5, I=20.1, T=32.2, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:28:08,919] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7800: forward=2781.7ms backward=1132.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:09,167] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:14,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:28:15,693] loop.py:45 INFO TRAIN_STEP step=7802 loss=1.358535 Training: 43%|████▎ | 7802/18018 [16:58:01<20:51:20, 7.35s/step, loss=1.359, acc=0.692, acc_len=1.66, thru=17.7, I=23.0, T=31.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:15,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:24,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:28:24,555] loop.py:45 INFO TRAIN_STEP step=7803 loss=1.393303 Training: 43%|████▎ | 7803/18018 [16:58:10<22:08:25, 7.80s/step, loss=1.393, acc=0.688, acc_len=1.64, thru=13.6, I=16.4, T=28.6, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:24,883] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:28:31,933] loop.py:45 INFO TRAIN_STEP step=7804 loss=1.503502 Training: 43%|████▎ | 7804/18018 [16:58:18<21:46:28, 7.67s/step, loss=1.504, acc=0.669, acc_len=1.58, thru=16.5, I=15.9, T=30.4, wait=3.2s, pool=56, epoch=1/2][2026-10-19 06:28:40,043] loop.py:380 INFO TIMING step=7805: step=4.281s data=1.939s compute=4.244s [fwd=3.073s bwd=1.157s opt=0.003s] dispatch=3.822s [2026-10-19 06:28:40,044] loop.py:45 INFO TRAIN_STEP step=7805 loss=1.288907 Training: 43%|████▎ | 7805/18018 [16:58:26<22:08:45, 7.81s/step, loss=1.289, acc=0.703, acc_len=1.69, thru=14.3, I=14.5, T=29.9, wait=3.8s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:46,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:28:47,649] loop.py:45 INFO TRAIN_STEP step=7806 loss=1.481975 Training: 43%|████▎ | 7806/18018 [16:58:33<21:58:22, 7.75s/step, loss=1.482, acc=0.673, acc_len=1.59, thru=17.5, I=18.2, T=31.9, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:28:47,645] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7805: forward=2858.7ms backward=1109.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:47,790] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:55,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:28:56,167] loop.py:45 INFO TRAIN_STEP step=7807 loss=1.479931 Training: 43%|████▎ | 7807/18018 [16:58:42<22:37:34, 7.98s/step, loss=1.480, acc=0.671, acc_len=1.58, thru=13.3, I=15.9, T=27.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:28:56,300] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:02,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:29:04,500] loop.py:45 INFO TRAIN_STEP step=7808 loss=1.439713 Training: 43%|████▎ | 7808/18018 [16:58:50<22:55:42, 8.08s/step, loss=1.440, acc=0.676, acc_len=1.60, thru=14.8, I=18.5, T=27.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:04,935] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:11,736] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:29:12,619] loop.py:45 INFO TRAIN_STEP step=7809 loss=1.520788 Training: 43%|████▎ | 7809/18018 [16:58:58<22:57:17, 8.09s/step, loss=1.521, acc=0.662, acc_len=1.56, thru=16.7, I=17.9, T=25.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:12,747] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:18,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:29:19,632] loop.py:380 INFO TIMING step=7810: step=4.252s data=1.971s compute=4.213s [fwd=2.951s bwd=1.246s opt=0.003s] dispatch=2.750s [2026-10-19 06:29:19,632] loop.py:45 INFO TRAIN_STEP step=7810 loss=1.473222 Training: 43%|████▎ | 7810/18018 [16:59:05<22:01:52, 7.77s/step, loss=1.473, acc=0.669, acc_len=1.58, thru=16.1, I=20.1, T=30.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:20,107] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:24,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:29:27,113] loop.py:45 INFO TRAIN_STEP step=7811 loss=1.343390 Training: 43%|████▎ | 7811/18018 [16:59:13<21:47:08, 7.68s/step, loss=1.343, acc=0.691, acc_len=1.65, thru=15.1, I=24.5, T=26.8, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:29:27,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7810: forward=3315.5ms backward=1416.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:27,494] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:32,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:29:34,848] loop.py:45 INFO TRAIN_STEP step=7812 loss=1.422512 Training: 43%|████▎ | 7812/18018 [16:59:20<21:49:29, 7.70s/step, loss=1.423, acc=0.683, acc_len=1.62, thru=13.7, I=24.5, T=25.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:35,347] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:41,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:29:42,905] loop.py:45 INFO TRAIN_STEP step=7813 loss=1.388106 Training: 43%|████▎ | 7813/18018 [16:59:28<22:07:41, 7.81s/step, loss=1.388, acc=0.683, acc_len=1.63, thru=16.9, I=20.8, T=26.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:29:43,217] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:29:52,523] loop.py:45 INFO TRAIN_STEP step=7814 loss=1.420739 Training: 43%|████▎ | 7814/18018 [16:59:38<23:40:03, 8.35s/step, loss=1.421, acc=0.679, acc_len=1.62, thru=12.9, I=14.8, T=27.2, wait=4.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:01,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:30:03,525] loop.py:380 INFO TIMING step=7815: step=4.936s data=2.388s compute=4.896s [fwd=3.401s bwd=1.480s opt=0.003s] dispatch=6.056s [2026-10-19 06:30:03,525] loop.py:45 INFO TRAIN_STEP step=7815 loss=1.504412 Training: 43%|████▎ | 7815/18018 [16:59:49<25:55:14, 9.15s/step, loss=1.504, acc=0.668, acc_len=1.57, thru=13.4, I=12.9, T=25.9, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:03,633] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:10,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:30:11,113] loop.py:45 INFO TRAIN_STEP step=7816 loss=1.496689 Training: 43%|████▎ | 7816/18018 [16:59:57<24:35:38, 8.68s/step, loss=1.497, acc=0.670, acc_len=1.57, thru=17.1, I=17.3, T=30.8, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:30:11,109] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7815: forward=2971.5ms backward=1136.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:11,452] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:16,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:30:18,593] loop.py:45 INFO TRAIN_STEP step=7817 loss=1.295080 Training: 43%|████▎ | 7817/18018 [17:00:04<23:34:21, 8.32s/step, loss=1.295, acc=0.703, acc_len=1.70, thru=16.5, I=23.0, T=29.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:18,628] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:23,092] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:30:24,820] loop.py:45 INFO TRAIN_STEP step=7818 loss=1.356087 Training: 43%|████▎ | 7818/18018 [17:00:10<21:47:34, 7.69s/step, loss=1.356, acc=0.693, acc_len=1.66, thru=18.1, I=29.9, T=30.9, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:25,106] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:31,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:30:31,800] loop.py:45 INFO TRAIN_STEP step=7819 loss=1.281450 Training: 43%|████▎ | 7819/18018 [17:00:17<21:11:08, 7.48s/step, loss=1.281, acc=0.707, acc_len=1.71, thru=16.1, I=22.1, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:32,063] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:36,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:30:38,146] loop.py:380 INFO TIMING step=7820: step=3.747s data=1.636s compute=3.715s [fwd=2.659s bwd=1.040s opt=0.003s] dispatch=2.588s [2026-10-19 06:30:38,146] loop.py:45 INFO TRAIN_STEP step=7820 loss=1.534012 Training: 43%|████▎ | 7820/18018 [17:00:24<20:13:14, 7.14s/step, loss=1.534, acc=0.662, acc_len=1.55, thru=18.5, I=23.5, T=34.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:38,504] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:43,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:30:45,515] loop.py:45 INFO TRAIN_STEP step=7821 loss=1.334039 Training: 43%|████▎ | 7821/18018 [17:00:31<20:24:56, 7.21s/step, loss=1.334, acc=0.693, acc_len=1.66, thru=17.5, I=25.8, T=27.0, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:30:45,511] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7820: forward=3387.2ms backward=1301.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:45,549] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:50,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:30:52,199] loop.py:45 INFO TRAIN_STEP step=7822 loss=1.445756 Training: 43%|████▎ | 7822/18018 [17:00:38<19:58:08, 7.05s/step, loss=1.446, acc=0.675, acc_len=1.60, thru=15.7, I=21.9, T=28.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:30:52,354] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-19 06:30:58,860] loop.py:45 INFO TRAIN_STEP step=7823 loss=1.431389 Training: 43%|████▎ | 7823/18018 [17:00:44<19:37:59, 6.93s/step, loss=1.431, acc=0.678, acc_len=1.61, thru=14.7, I=17.1, T=28.9, wait=2.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:08,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:31:08,254] loop.py:45 INFO TRAIN_STEP step=7824 loss=1.515392 Training: 43%|████▎ | 7824/18018 [17:00:54<21:43:20, 7.67s/step, loss=1.515, acc=0.668, acc_len=1.57, thru=15.8, I=16.0, T=31.5, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:08,533] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:31:15,082] loop.py:380 INFO TIMING step=7825: step=4.253s data=2.052s compute=4.219s [fwd=2.960s bwd=1.243s opt=0.003s] dispatch=2.567s [2026-10-19 06:31:15,083] loop.py:45 INFO TRAIN_STEP step=7825 loss=1.419912 Training: 43%|████▎ | 7825/18018 [17:01:01<21:00:18, 7.42s/step, loss=1.420, acc=0.679, acc_len=1.62, thru=16.7, I=20.1, T=30.1, wait=2.6s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:21,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:31:24,545] loop.py:45 INFO TRAIN_STEP step=7826 loss=1.389306 Training: 43%|████▎ | 7826/18018 [17:01:10<22:44:25, 8.03s/step, loss=1.389, acc=0.688, acc_len=1.65, thru=15.3, I=17.0, T=26.3, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:31:24,540] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7825: forward=3323.2ms backward=1449.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:24,793] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:29,535] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:31:31,860] loop.py:45 INFO TRAIN_STEP step=7827 loss=1.343140 Training: 43%|████▎ | 7827/18018 [17:01:17<22:07:42, 7.82s/step, loss=1.343, acc=0.692, acc_len=1.66, thru=18.1, I=27.2, T=25.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:32,049] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:38,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:31:39,765] loop.py:45 INFO TRAIN_STEP step=7828 loss=1.355291 Training: 43%|████▎ | 7828/18018 [17:01:25<22:12:00, 7.84s/step, loss=1.355, acc=0.691, acc_len=1.66, thru=16.6, I=19.8, T=27.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:39,829] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:47,741] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:31:49,355] loop.py:45 INFO TRAIN_STEP step=7829 loss=1.357122 Training: 43%|████▎ | 7829/18018 [17:01:35<23:40:50, 8.37s/step, loss=1.357, acc=0.695, acc_len=1.67, thru=13.5, I=16.3, T=28.6, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:49,754] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:54,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:31:56,001] loop.py:380 INFO TIMING step=7830: step=4.207s data=1.922s compute=4.182s [fwd=2.932s bwd=1.234s opt=0.003s] dispatch=2.431s [2026-10-19 06:31:56,002] loop.py:45 INFO TRAIN_STEP step=7830 loss=1.372876 Training: 43%|████▎ | 7830/18018 [17:01:42<22:13:11, 7.85s/step, loss=1.373, acc=0.690, acc_len=1.65, thru=16.9, I=24.1, T=30.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:31:56,315] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:03,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:32:03,420] loop.py:45 INFO TRAIN_STEP step=7831 loss=1.467748 Training: 43%|████▎ | 7831/18018 [17:01:49<21:50:57, 7.72s/step, loss=1.468, acc=0.675, acc_len=1.59, thru=16.1, I=18.5, T=31.9, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:32:03,415] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7830: forward=2783.0ms backward=1184.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:03,865] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:10,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:32:10,619] loop.py:45 INFO TRAIN_STEP step=7832 loss=1.482588 Training: 43%|████▎ | 7832/18018 [17:01:56<21:24:12, 7.56s/step, loss=1.483, acc=0.671, acc_len=1.59, thru=17.7, I=19.1, T=33.4, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:10,885] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:32:19,530] loop.py:45 INFO TRAIN_STEP step=7833 loss=1.457693 Training: 43%|████▎ | 7833/18018 [17:02:05<22:32:37, 7.97s/step, loss=1.458, acc=0.678, acc_len=1.61, thru=14.1, I=17.2, T=27.7, wait=4.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:26,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:32:26,585] loop.py:45 INFO TRAIN_STEP step=7834 loss=1.226171 Training: 43%|████▎ | 7834/18018 [17:02:12<21:46:03, 7.69s/step, loss=1.226, acc=0.716, acc_len=1.74, thru=15.7, I=16.6, T=30.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:26,621] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:32,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:32:34,135] loop.py:380 INFO TIMING step=7835: step=4.451s data=2.134s compute=4.415s [fwd=3.107s bwd=1.291s opt=0.003s] dispatch=3.088s [2026-10-19 06:32:34,135] loop.py:45 INFO TRAIN_STEP step=7835 loss=1.496187 Training: 43%|████▎ | 7835/18018 [17:02:20<21:38:28, 7.65s/step, loss=1.496, acc=0.668, acc_len=1.58, thru=15.1, I=18.4, T=28.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:34,493] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:38,525] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:32:40,510] loop.py:45 INFO TRAIN_STEP step=7836 loss=1.282151 Training: 43%|████▎ | 7836/18018 [17:02:26<20:33:28, 7.27s/step, loss=1.282, acc=0.707, acc_len=1.71, thru=16.8, I=27.9, T=33.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:40,544] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-19 06:32:40,505] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7835: forward=2681.1ms backward=1114.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:45,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:32:46,658] loop.py:45 INFO TRAIN_STEP step=7837 loss=1.363038 Training: 43%|████▎ | 7837/18018 [17:02:32<19:36:16, 6.93s/step, loss=1.363, acc=0.691, acc_len=1.65, thru=19.5, I=28.2, T=34.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:46,862] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:32:52,392] loop.py:45 INFO TRAIN_STEP step=7838 loss=1.420198 Training: 44%|████▎ | 7838/18018 [17:02:38<18:35:05, 6.57s/step, loss=1.420, acc=0.680, acc_len=1.62, thru=20.6, I=22.9, T=35.9, wait=2.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:32:59,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:33:00,219] loop.py:45 INFO TRAIN_STEP step=7839 loss=1.432654 Training: 44%|████▎ | 7839/18018 [17:02:46<19:38:55, 6.95s/step, loss=1.433, acc=0.685, acc_len=1.63, thru=17.5, I=18.8, T=27.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:00,639] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:08,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:33:09,349] loop.py:380 INFO TIMING step=7840: step=4.092s data=1.901s compute=4.061s [fwd=2.879s bwd=1.167s opt=0.003s] dispatch=5.029s [2026-10-19 06:33:09,349] loop.py:45 INFO TRAIN_STEP step=7840 loss=1.454854 Training: 44%|████▎ | 7840/18018 [17:02:55<21:29:42, 7.60s/step, loss=1.455, acc=0.675, acc_len=1.60, thru=14.7, I=16.3, T=31.3, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:09,377] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:15,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:33:16,844] loop.py:45 INFO TRAIN_STEP step=7841 loss=1.403063 Training: 44%|████▎ | 7841/18018 [17:03:02<21:24:07, 7.57s/step, loss=1.403, acc=0.681, acc_len=1.62, thru=16.4, I=21.1, T=32.2, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:33:16,840] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7840: forward=2741.0ms backward=1171.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:17,294] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:33:24,277] loop.py:45 INFO TRAIN_STEP step=7842 loss=1.452779 Training: 44%|████▎ | 7842/18018 [17:03:10<21:17:05, 7.53s/step, loss=1.453, acc=0.675, acc_len=1.60, thru=14.5, I=19.0, T=34.9, wait=3.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:31,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:33:32,140] loop.py:45 INFO TRAIN_STEP step=7843 loss=1.425523 Training: 44%|████▎ | 7843/18018 [17:03:18<21:33:53, 7.63s/step, loss=1.426, acc=0.682, acc_len=1.62, thru=17.2, I=16.6, T=31.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:32,187] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:38,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:33:38,171] loop.py:45 INFO TRAIN_STEP step=7844 loss=1.510651 Training: 44%|████▎ | 7844/18018 [17:03:24<20:12:33, 7.15s/step, loss=1.511, acc=0.663, acc_len=1.56, thru=19.5, I=20.4, T=32.2, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:38,517] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:44,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:33:45,213] loop.py:380 INFO TIMING step=7845: step=3.799s data=1.539s compute=3.766s [fwd=2.591s bwd=1.161s opt=0.003s] dispatch=3.231s [2026-10-19 06:33:45,213] loop.py:45 INFO TRAIN_STEP step=7845 loss=1.499219 Training: 44%|████▎ | 7845/18018 [17:03:31<20:06:39, 7.12s/step, loss=1.499, acc=0.668, acc_len=1.58, thru=18.0, I=19.5, T=33.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:45,671] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:51,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:33:52,500] loop.py:45 INFO TRAIN_STEP step=7846 loss=1.331342 Training: 44%|████▎ | 7846/18018 [17:03:38<20:15:16, 7.17s/step, loss=1.331, acc=0.700, acc_len=1.68, thru=17.3, I=20.5, T=34.2, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:33:52,496] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7845: forward=2635.1ms backward=1049.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:33:52,943] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:34:00,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:34:01,412] loop.py:45 INFO TRAIN_STEP step=7847 loss=1.380670 Training: 44%|████▎ | 7847/18018 [17:03:47<21:43:51, 7.69s/step, loss=1.381, acc=0.687, acc_len=1.63, thru=13.2, I=15.9, T=28.4, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:34:01,450] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 06:34:07,749] loop.py:45 INFO TRAIN_STEP step=7848 loss=1.376690 Training: 44%|████▎ | 7848/18018 [17:03:53<20:34:41, 7.28s/step, loss=1.377, acc=0.690, acc_len=1.65, thru=17.6, I=18.7, T=32.4, wait=2.4s, pool=56, epoch=1/2][2026-10-19 06:34:15,632] loop.py:45 INFO TRAIN_STEP step=7849 loss=1.505309 Training: 44%|████▎ | 7849/18018 [17:04:01<21:05:03, 7.46s/step, loss=1.505, acc=0.670, acc_len=1.59, thru=14.9, I=14.0, T=31.9, wait=3.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:34:20,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:34:22,880] loop.py:380 INFO TIMING step=7850: step=4.480s data=2.122s compute=4.455s [fwd=3.144s bwd=1.296s opt=0.003s] dispatch=2.761s [2026-10-19 06:34:22,881] loop.py:45 INFO TRAIN_STEP step=7850 loss=1.465079 Training: 44%|████▎ | 7850/18018 [17:04:08<20:54:04, 7.40s/step, loss=1.465, acc=0.673, acc_len=1.59, thru=19.8, I=20.9, T=28.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:34:22,954] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:34:26,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:34:29,221] loop.py:45 INFO TRAIN_STEP step=7851 loss=1.315730 Training: 44%|████▎ | 7851/18018 [17:04:15<19:59:54, 7.08s/step, loss=1.316, acc=0.699, acc_len=1.68, thru=17.4, I=31.8, T=28.5, wait=1.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:34:29,217] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7850: forward=3111.8ms backward=1327.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:34:29,598] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:34:36,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:34:38,092] loop.py:45 INFO TRAIN_STEP step=7852 loss=1.326266 Training: 44%|████▎ | 7852/18018 [17:04:24<21:30:52, 7.62s/step, loss=1.326, acc=0.697, acc_len=1.68, thru=15.1, I=17.4, T=24.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:34:38,336] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:34:45,743] loop.py:45 INFO TRAIN_STEP step=7853 loss=1.412689 Training: 44%|████▎ | 7853/18018 [17:04:31<21:32:18, 7.63s/step, loss=1.413, acc=0.679, acc_len=1.61, thru=13.8, I=14.0, T=29.7, wait=3.3s, pool=48, epoch=1/2][2026-10-19 06:34:53,333] loop.py:45 INFO TRAIN_STEP step=7854 loss=1.417005 Training: 44%|████▎ | 7854/18018 [17:04:39<21:30:18, 7.62s/step, loss=1.417, acc=0.681, acc_len=1.62, thru=16.1, I=15.0, T=33.4, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:34:58,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:35:00,805] loop.py:380 INFO TIMING step=7855: step=4.055s data=1.829s compute=4.024s [fwd=2.889s bwd=1.119s opt=0.003s] dispatch=3.409s [2026-10-19 06:35:00,806] loop.py:45 INFO TRAIN_STEP step=7855 loss=1.499970 Training: 44%|████▎ | 7855/18018 [17:04:46<21:22:50, 7.57s/step, loss=1.500, acc=0.670, acc_len=1.58, thru=18.3, I=21.2, T=31.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:00,881] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:07,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:35:07,785] loop.py:45 INFO TRAIN_STEP step=7856 loss=1.504048 Training: 44%|████▎ | 7856/18018 [17:04:53<20:52:28, 7.40s/step, loss=1.504, acc=0.664, acc_len=1.56, thru=15.0, I=21.1, T=29.4, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:35:07,779] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7855: forward=3002.5ms backward=1299.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:08,157] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:14,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:35:14,981] loop.py:45 INFO TRAIN_STEP step=7857 loss=1.324449 Training: 44%|████▎ | 7857/18018 [17:05:01<20:42:17, 7.34s/step, loss=1.324, acc=0.699, acc_len=1.68, thru=16.0, I=17.4, T=29.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:15,338] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:23,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:35:23,744] loop.py:45 INFO TRAIN_STEP step=7858 loss=1.440516 Training: 44%|████▎ | 7858/18018 [17:05:09<21:54:40, 7.76s/step, loss=1.441, acc=0.679, acc_len=1.61, thru=14.5, I=14.9, T=28.3, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:24,022] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:30,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:35:32,531] loop.py:45 INFO TRAIN_STEP step=7859 loss=1.437284 Training: 44%|████▎ | 7859/18018 [17:05:18<22:46:35, 8.07s/step, loss=1.437, acc=0.677, acc_len=1.61, thru=14.5, I=16.1, T=28.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:32,924] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:35:40,871] loop.py:380 INFO TIMING step=7860: step=4.803s data=2.356s compute=4.773s [fwd=3.283s bwd=1.474s opt=0.003s] dispatch=3.526s [2026-10-19 06:35:40,871] loop.py:45 INFO TRAIN_STEP step=7860 loss=1.420944 Training: 44%|████▎ | 7860/18018 [17:05:26<23:00:10, 8.15s/step, loss=1.421, acc=0.680, acc_len=1.61, thru=9.6, I=18.7, T=26.6, wait=3.5s, pool=28, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:50,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:35:50,843] loop.py:45 INFO TRAIN_STEP step=7861 loss=1.435291 (TrainerActor pid=9310) [2026-10-19 06:35:50,838] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7860: forward=3179.3ms backward=1328.1ms Training: 44%|████▎ | 7861/18018 [17:05:36<24:32:22, 8.70s/step, loss=1.435, acc=0.680, acc_len=1.62, thru=16.5, I=13.3, T=28.0, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:50,878] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:56,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:35:57,759] loop.py:45 INFO TRAIN_STEP step=7862 loss=1.405441 Training: 44%|████▎ | 7862/18018 [17:05:43<23:01:39, 8.16s/step, loss=1.405, acc=0.683, acc_len=1.62, thru=18.3, I=19.6, T=28.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:35:58,240] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:02,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:36:04,722] loop.py:45 INFO TRAIN_STEP step=7863 loss=1.380687 Training: 44%|████▎ | 7863/18018 [17:05:50<22:00:43, 7.80s/step, loss=1.381, acc=0.686, acc_len=1.64, thru=16.0, I=24.4, T=30.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:04,964] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:14,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:36:14,813] loop.py:45 INFO TRAIN_STEP step=7864 loss=1.500102 Training: 44%|████▎ | 7864/18018 [17:06:00<23:56:44, 8.49s/step, loss=1.500, acc=0.669, acc_len=1.57, thru=13.0, I=12.6, T=26.7, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:14,907] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:20,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:36:22,096] loop.py:380 INFO TIMING step=7865: step=4.495s data=2.041s compute=4.472s [fwd=3.170s bwd=1.286s opt=0.003s] dispatch=2.779s [2026-10-19 06:36:22,097] loop.py:45 INFO TRAIN_STEP step=7865 loss=1.476819 Training: 44%|████▎ | 7865/18018 [17:06:08<22:55:19, 8.13s/step, loss=1.477, acc=0.670, acc_len=1.58, thru=17.5, I=19.1, T=28.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:22,248] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:26,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:36:28,722] loop.py:45 INFO TRAIN_STEP step=7866 loss=1.360268 Training: 44%|████▎ | 7866/18018 [17:06:14<21:38:59, 7.68s/step, loss=1.360, acc=0.691, acc_len=1.65, thru=17.5, I=25.0, T=30.3, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:36:28,717] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7865: forward=2819.7ms backward=1346.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:28,759] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 06:36:35,076] loop.py:45 INFO TRAIN_STEP step=7867 loss=1.492213 Training: 44%|████▎ | 7867/18018 [17:06:21<20:31:38, 7.28s/step, loss=1.492, acc=0.669, acc_len=1.58, thru=16.1, I=23.3, T=32.5, wait=2.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:41,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:36:42,550] loop.py:45 INFO TRAIN_STEP step=7868 loss=1.473029 Training: 44%|████▎ | 7868/18018 [17:06:28<20:41:23, 7.34s/step, loss=1.473, acc=0.672, acc_len=1.59, thru=18.1, I=19.0, T=29.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:43,042] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:50,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:36:51,332] loop.py:45 INFO TRAIN_STEP step=7869 loss=1.433044 Training: 44%|████▎ | 7869/18018 [17:06:37<21:54:33, 7.77s/step, loss=1.433, acc=0.679, acc_len=1.61, thru=15.4, I=15.6, T=28.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:36:51,468] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 06:37:00,731] loop.py:380 INFO TIMING step=7870: step=5.112s data=2.610s compute=5.076s [fwd=3.642s bwd=1.420s opt=0.003s] dispatch=4.277s [2026-10-19 06:37:00,732] loop.py:45 INFO TRAIN_STEP step=7870 loss=1.442400 Training: 44%|████▎ | 7870/18018 [17:06:46<23:16:55, 8.26s/step, loss=1.442, acc=0.678, acc_len=1.61, thru=11.7, I=12.2, T=25.0, wait=4.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:09,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:37:11,231] loop.py:45 INFO TRAIN_STEP step=7871 loss=1.463568 Training: 44%|████▎ | 7871/18018 [17:06:57<25:10:31, 8.93s/step, loss=1.464, acc=0.676, acc_len=1.60, thru=14.1, I=16.2, T=24.6, wait=5.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:37:11,226] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7870: forward=3671.5ms backward=1492.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:11,515] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:17,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:37:19,485] loop.py:45 INFO TRAIN_STEP step=7872 loss=1.402264 Training: 44%|████▎ | 7872/18018 [17:07:05<24:36:03, 8.73s/step, loss=1.402, acc=0.683, acc_len=1.63, thru=13.2, I=20.5, T=30.9, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:19,559] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 [2026-10-19 06:37:26,902] loop.py:45 INFO TRAIN_STEP step=7873 loss=1.422949 Training: 44%|████▎ | 7873/18018 [17:07:12<23:29:20, 8.34s/step, loss=1.423, acc=0.683, acc_len=1.63, thru=14.6, I=18.9, T=30.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:26,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:27,409] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:32,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:37:34,109] loop.py:45 INFO TRAIN_STEP step=7874 loss=1.397398 Training: 44%|████▎ | 7874/18018 [17:07:20<22:31:53, 8.00s/step, loss=1.397, acc=0.687, acc_len=1.63, thru=17.2, I=19.6, T=29.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:34,481] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:37:43,182] loop.py:380 INFO TIMING step=7875: step=4.464s data=2.115s compute=4.434s [fwd=3.113s bwd=1.306s opt=0.003s] dispatch=4.601s [2026-10-19 06:37:43,182] loop.py:45 INFO TRAIN_STEP step=7875 loss=1.385840 Training: 44%|████▎ | 7875/18018 [17:07:29<23:26:23, 8.32s/step, loss=1.386, acc=0.685, acc_len=1.63, thru=14.2, I=14.6, T=28.7, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:49,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:37:51,296] loop.py:45 INFO TRAIN_STEP step=7876 loss=1.406051 Training: 44%|████▎ | 7876/18018 [17:07:37<23:15:48, 8.26s/step, loss=1.406, acc=0.688, acc_len=1.64, thru=16.8, I=18.6, T=26.7, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:37:51,291] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7875: forward=3271.8ms backward=1433.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:51,503] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:37:58,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:37:59,989] loop.py:45 INFO TRAIN_STEP step=7877 loss=1.448461 Training: 44%|████▎ | 7877/18018 [17:07:46<23:37:56, 8.39s/step, loss=1.448, acc=0.675, acc_len=1.60, thru=15.1, I=16.7, T=28.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:00,346] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:05,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:38:06,081] loop.py:45 INFO TRAIN_STEP step=7878 loss=1.466842 Training: 44%|████▎ | 7878/18018 [17:07:52<21:41:11, 7.70s/step, loss=1.467, acc=0.673, acc_len=1.59, thru=18.8, I=21.6, T=33.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:06,269] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:13,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:38:14,534] loop.py:45 INFO TRAIN_STEP step=7879 loss=1.521177 Training: 44%|████▎ | 7879/18018 [17:08:00<22:19:11, 7.92s/step, loss=1.521, acc=0.665, acc_len=1.57, thru=15.9, I=17.8, T=25.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:14,955] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:38:23,414] loop.py:380 INFO TIMING step=7880: step=5.270s data=2.312s compute=5.212s [fwd=3.791s bwd=1.406s opt=0.003s] dispatch=3.602s [2026-10-19 06:38:23,414] loop.py:45 INFO TRAIN_STEP step=7880 loss=1.391629 Training: 44%|████▎ | 7880/18018 [17:08:09<23:07:24, 8.21s/step, loss=1.392, acc=0.691, acc_len=1.65, thru=13.8, I=15.1, T=24.3, wait=3.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:30,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:38:31,392] loop.py:45 INFO TRAIN_STEP step=7881 loss=1.418565 Training: 44%|████▎ | 7881/18018 [17:08:17<22:55:32, 8.14s/step, loss=1.419, acc=0.688, acc_len=1.64, thru=17.7, I=17.0, T=26.2, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:38:31,388] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7880: forward=3391.6ms backward=1436.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:31,550] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:38:38,442] loop.py:45 INFO TRAIN_STEP step=7882 loss=1.457956 Training: 44%|████▎ | 7882/18018 [17:08:24<22:00:05, 7.81s/step, loss=1.458, acc=0.675, acc_len=1.60, thru=13.3, I=14.5, T=29.3, wait=2.7s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:47,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:38:48,823] loop.py:45 INFO TRAIN_STEP step=7883 loss=1.464523 Training: 44%|████▍ | 7883/18018 [17:08:34<24:09:55, 8.58s/step, loss=1.465, acc=0.671, acc_len=1.58, thru=15.5, I=17.0, T=22.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:49,325] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:56,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:38:56,583] loop.py:45 INFO TRAIN_STEP step=7884 loss=1.360253 Training: 44%|████▍ | 7884/18018 [17:08:42<23:28:09, 8.34s/step, loss=1.360, acc=0.690, acc_len=1.65, thru=14.5, I=18.6, T=31.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:38:56,991] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:04,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:39:04,316] loop.py:380 INFO TIMING step=7885: step=4.959s data=2.393s compute=4.924s [fwd=3.393s bwd=1.517s opt=0.003s] dispatch=2.764s [2026-10-19 06:39:04,316] loop.py:45 INFO TRAIN_STEP step=7885 loss=1.531016 Training: 44%|████▍ | 7885/18018 [17:08:50<22:57:25, 8.16s/step, loss=1.531, acc=0.660, acc_len=1.55, thru=16.6, I=16.4, T=25.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:04,666] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:10,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:39:10,945] loop.py:45 INFO TRAIN_STEP step=7886 loss=1.434124 Training: 44%|████▍ | 7886/18018 [17:08:57<21:39:53, 7.70s/step, loss=1.434, acc=0.674, acc_len=1.59, thru=18.5, I=18.2, T=30.8, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:39:10,941] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7885: forward=2921.4ms backward=1163.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:11,009] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:18,119] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:39:19,409] loop.py:45 INFO TRAIN_STEP step=7887 loss=1.500342 Training: 44%|████▍ | 7887/18018 [17:09:05<22:18:35, 7.93s/step, loss=1.500, acc=0.663, acc_len=1.56, thru=14.9, I=17.3, T=26.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:19,627] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:26,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:39:26,427] loop.py:45 INFO TRAIN_STEP step=7888 loss=1.365780 Training: 44%|████▍ | 7888/18018 [17:09:12<21:32:21, 7.65s/step, loss=1.366, acc=0.693, acc_len=1.65, thru=16.2, I=19.3, T=30.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:26,612] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:32,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:39:34,205] loop.py:45 INFO TRAIN_STEP step=7889 loss=1.348744 Training: 44%|████▍ | 7889/18018 [17:09:20<21:38:35, 7.69s/step, loss=1.349, acc=0.694, acc_len=1.67, thru=16.8, I=20.3, T=27.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:34,315] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:39,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:39:42,058] loop.py:380 INFO TIMING step=7890: step=5.216s data=2.127s compute=5.175s [fwd=3.834s bwd=1.327s opt=0.003s] dispatch=2.626s [2026-10-19 06:39:42,058] loop.py:45 INFO TRAIN_STEP step=7890 loss=1.383387 Training: 44%|████▍ | 7890/18018 [17:09:28<21:46:28, 7.74s/step, loss=1.383, acc=0.686, acc_len=1.64, thru=13.7, I=19.4, T=24.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:39:42,420] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 06:39:51,360] loop.py:45 INFO TRAIN_STEP step=7891 loss=1.351940 Training: 44%|████▍ | 7891/18018 [17:09:37<23:05:25, 8.21s/step, loss=1.352, acc=0.694, acc_len=1.66, thru=13.4, I=12.1, T=26.9, wait=4.5s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:39:51,355] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7890: forward=3277.4ms backward=1420.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:00,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:40:01,695] loop.py:45 INFO TRAIN_STEP step=7892 loss=1.416980 Training: 44%|████▍ | 7892/18018 [17:09:47<24:53:05, 8.85s/step, loss=1.417, acc=0.688, acc_len=1.64, thru=13.3, I=13.8, T=30.7, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:01,852] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:08,897] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:40:11,194] loop.py:45 INFO TRAIN_STEP step=7893 loss=1.499002 Training: 44%|████▍ | 7893/18018 [17:09:57<25:25:56, 9.04s/step, loss=1.499, acc=0.665, acc_len=1.56, thru=13.7, I=17.9, T=28.5, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:11,415] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:17,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:40:18,516] loop.py:45 INFO TRAIN_STEP step=7894 loss=1.311113 Training: 44%|████▍ | 7894/18018 [17:10:04<23:58:41, 8.53s/step, loss=1.311, acc=0.701, acc_len=1.69, thru=14.6, I=22.4, T=31.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:18,908] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:40:27,008] loop.py:380 INFO TIMING step=7895: step=4.130s data=1.812s compute=4.089s [fwd=2.928s bwd=1.146s opt=0.003s] dispatch=4.352s [2026-10-19 06:40:27,008] loop.py:45 INFO TRAIN_STEP step=7895 loss=1.425536 Training: 44%|████▍ | 7895/18018 [17:10:13<23:56:51, 8.52s/step, loss=1.426, acc=0.680, acc_len=1.62, thru=12.8, I=12.7, T=31.0, wait=4.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:34,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:40:36,225] loop.py:45 INFO TRAIN_STEP step=7896 loss=1.388603 Training: 44%|████▍ | 7896/18018 [17:10:22<24:32:03, 8.73s/step, loss=1.389, acc=0.685, acc_len=1.63, thru=16.6, I=18.6, T=25.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:36,252] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-19 06:40:36,221] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7895: forward=3455.9ms backward=1464.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:40,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:40:41,919] loop.py:45 INFO TRAIN_STEP step=7897 loss=1.233588 Training: 44%|████▍ | 7897/18018 [17:10:28<21:58:31, 7.82s/step, loss=1.234, acc=0.715, acc_len=1.73, thru=17.8, I=25.2, T=34.0, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:42,373] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:50,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:40:50,723] loop.py:45 INFO TRAIN_STEP step=7898 loss=1.476800 Training: 44%|████▍ | 7898/18018 [17:10:36<22:48:24, 8.11s/step, loss=1.477, acc=0.669, acc_len=1.57, thru=13.4, I=16.6, T=32.8, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:40:50,881] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:40:57,423] loop.py:45 INFO TRAIN_STEP step=7899 loss=1.297289 Training: 44%|████▍ | 7899/18018 [17:10:43<21:36:34, 7.69s/step, loss=1.297, acc=0.702, acc_len=1.69, thru=17.7, I=15.4, T=29.4, wait=2.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:04,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:41:05,423] loop.py:380 INFO TIMING step=7900: step=3.945s data=1.776s compute=3.914s [fwd=2.821s bwd=1.079s opt=0.003s] dispatch=4.048s [2026-10-19 06:41:05,423] loop.py:45 INFO TRAIN_STEP step=7900 loss=1.364370 Training: 44%|████▍ | 7900/18018 [17:10:51<21:52:20, 7.78s/step, loss=1.364, acc=0.690, acc_len=1.65, thru=17.4, I=16.9, T=32.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:05,630] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:12,097] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:41:14,196] loop.py:45 INFO TRAIN_STEP step=7901 loss=1.367170 Training: 44%|████▍ | 7901/18018 [17:11:00<22:42:17, 8.08s/step, loss=1.367, acc=0.692, acc_len=1.66, thru=14.4, I=18.5, T=26.8, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:41:14,192] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7900: forward=3375.3ms backward=1354.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:14,615] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:20,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:41:21,214] loop.py:45 INFO TRAIN_STEP step=7902 loss=1.420735 Training: 44%|████▍ | 7902/18018 [17:11:07<21:48:28, 7.76s/step, loss=1.421, acc=0.680, acc_len=1.61, thru=14.9, I=22.2, T=31.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:21,623] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:27,324] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:41:28,605] loop.py:45 INFO TRAIN_STEP step=7903 loss=1.522559 Training: 44%|████▍ | 7903/18018 [17:11:14<21:29:43, 7.65s/step, loss=1.523, acc=0.662, acc_len=1.56, thru=15.7, I=20.1, T=27.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:28,835] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:34,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:41:35,568] loop.py:45 INFO TRAIN_STEP step=7904 loss=1.391639 Training: 44%|████▍ | 7904/18018 [17:11:21<20:54:51, 7.44s/step, loss=1.392, acc=0.689, acc_len=1.65, thru=17.8, I=21.1, T=32.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:35,972] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:41,600] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:41:43,582] loop.py:380 INFO TIMING step=7905: step=3.909s data=1.663s compute=3.877s [fwd=2.718s bwd=1.146s opt=0.003s] dispatch=4.094s [2026-10-19 06:41:43,582] loop.py:45 INFO TRAIN_STEP step=7905 loss=1.441177 Training: 44%|████▍ | 7905/18018 [17:11:29<21:23:27, 7.61s/step, loss=1.441, acc=0.676, acc_len=1.60, thru=15.2, I=19.9, T=32.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:43,610] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 06:41:49,579] loop.py:45 INFO TRAIN_STEP step=7906 loss=1.367060 Training: 44%|████▍ | 7906/18018 [17:11:35<20:01:32, 7.13s/step, loss=1.367, acc=0.690, acc_len=1.65, thru=16.5, I=21.5, T=35.0, wait=2.3s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:41:49,575] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7905: forward=2537.1ms backward=1066.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:55,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:41:56,398] loop.py:45 INFO TRAIN_STEP step=7907 loss=1.458873 Training: 44%|████▍ | 7907/18018 [17:11:42<19:45:45, 7.04s/step, loss=1.459, acc=0.676, acc_len=1.60, thru=19.8, I=19.3, T=35.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:41:56,482] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:42:07,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:42:07,852] loop.py:45 INFO TRAIN_STEP step=7908 loss=1.498398 Training: 44%|████▍ | 7908/18018 [17:11:53<23:28:55, 8.36s/step, loss=1.498, acc=0.668, acc_len=1.58, thru=11.4, I=11.1, T=33.6, wait=7.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:42:08,239] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:42:18,289] loop.py:45 INFO TRAIN_STEP step=7909 loss=1.533001 Training: 44%|████▍ | 7909/18018 [17:12:04<25:13:52, 8.99s/step, loss=1.533, acc=0.661, acc_len=1.55, thru=11.9, I=11.2, T=31.1, wait=6.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:42:24,797] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:42:25,923] loop.py:380 INFO TIMING step=7910: step=3.767s data=1.639s compute=3.727s [fwd=2.636s bwd=1.076s opt=0.003s] dispatch=3.855s [2026-10-19 06:42:25,923] loop.py:45 INFO TRAIN_STEP step=7910 loss=1.403068 Training: 44%|████▍ | 7910/18018 [17:12:12<24:05:20, 8.58s/step, loss=1.403, acc=0.681, acc_len=1.62, thru=17.7, I=18.4, T=34.0, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:42:26,310] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:42:33,451] loop.py:45 INFO TRAIN_STEP step=7911 loss=1.422058 Training: 44%|████▍ | 7911/18018 [17:12:19<23:12:01, 8.26s/step, loss=1.422, acc=0.683, acc_len=1.63, thru=15.0, I=19.1, T=29.9, wait=3.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:42:33,447] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7910: forward=3029.7ms backward=1203.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:42:40,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:42:41,500] loop.py:45 INFO TRAIN_STEP step=7912 loss=1.458426 Training: 44%|████▍ | 7912/18018 [17:12:27<23:01:04, 8.20s/step, loss=1.458, acc=0.678, acc_len=1.61, thru=15.9, I=16.8, T=30.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:42:41,617] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 06:42:47,525] loop.py:45 INFO TRAIN_STEP step=7913 loss=1.390860 Training: 44%|████▍ | 7913/18018 [17:12:33<21:10:57, 7.55s/step, loss=1.391, acc=0.685, acc_len=1.63, thru=18.5, I=21.9, T=34.7, wait=2.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:42:54,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:42:55,223] loop.py:45 INFO TRAIN_STEP step=7914 loss=1.373236 Training: 44%|████▍ | 7914/18018 [17:12:41<21:18:28, 7.59s/step, loss=1.373, acc=0.690, acc_len=1.64, thru=16.1, I=17.1, T=36.5, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:42:55,450] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:03,581] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:43:04,949] loop.py:380 INFO TIMING step=7915: step=3.979s data=1.756s compute=3.944s [fwd=2.824s bwd=1.106s opt=0.003s] dispatch=5.738s [2026-10-19 06:43:04,949] loop.py:45 INFO TRAIN_STEP step=7915 loss=1.478421 Training: 44%|████▍ | 7915/18018 [17:12:51<23:06:11, 8.23s/step, loss=1.478, acc=0.669, acc_len=1.58, thru=13.5, I=15.6, T=32.2, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:05,089] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:11,183] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:43:12,041] loop.py:45 INFO TRAIN_STEP step=7916 loss=1.349874 Training: 44%|████▍ | 7916/18018 [17:12:58<22:08:31, 7.89s/step, loss=1.350, acc=0.690, acc_len=1.65, thru=16.9, I=19.6, T=35.2, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:43:12,037] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7915: forward=2608.3ms backward=981.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:12,195] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-19 06:43:19,228] loop.py:45 INFO TRAIN_STEP step=7917 loss=1.491098 Training: 44%|████▍ | 7917/18018 [17:13:05<21:32:51, 7.68s/step, loss=1.491, acc=0.670, acc_len=1.58, thru=17.0, I=18.6, T=36.1, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:25,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:43:26,334] loop.py:45 INFO TRAIN_STEP step=7918 loss=1.331144 Training: 44%|████▍ | 7918/18018 [17:13:12<21:03:46, 7.51s/step, loss=1.331, acc=0.699, acc_len=1.68, thru=18.9, I=19.6, T=35.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:26,450] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 06:43:32,277] loop.py:45 INFO TRAIN_STEP step=7919 loss=1.409150 Training: 44%|████▍ | 7919/18018 [17:13:18<19:44:31, 7.04s/step, loss=1.409, acc=0.683, acc_len=1.63, thru=20.0, I=20.5, T=36.3, wait=2.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:38,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:43:38,550] loop.py:380 INFO TIMING step=7920: step=3.448s data=1.431s compute=3.399s [fwd=2.433s bwd=0.952s opt=0.003s] dispatch=2.816s [2026-10-19 06:43:38,550] loop.py:45 INFO TRAIN_STEP step=7920 loss=1.459623 Training: 44%|████▍ | 7920/18018 [17:13:24<19:05:54, 6.81s/step, loss=1.460, acc=0.669, acc_len=1.58, thru=20.1, I=22.0, T=37.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:38,867] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:45,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:43:47,045] loop.py:45 INFO TRAIN_STEP step=7921 loss=1.456975 Training: 44%|████▍ | 7921/18018 [17:13:33<20:30:52, 7.31s/step, loss=1.457, acc=0.672, acc_len=1.59, thru=14.8, I=17.2, T=33.5, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:43:47,040] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7920: forward=2626.5ms backward=1122.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:43:47,231] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:43:53,482] loop.py:45 INFO TRAIN_STEP step=7922 loss=1.385208 Training: 44%|████▍ | 7922/18018 [17:13:39<19:46:26, 7.05s/step, loss=1.385, acc=0.690, acc_len=1.65, thru=18.9, I=24.1, T=34.8, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:02,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:44:03,622] loop.py:45 INFO TRAIN_STEP step=7923 loss=1.387357 Training: 44%|████▍ | 7923/18018 [17:13:49<22:22:19, 7.98s/step, loss=1.387, acc=0.687, acc_len=1.64, thru=13.6, I=13.5, T=31.3, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:03,935] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:44:09,049] loop.py:45 INFO TRAIN_STEP step=7924 loss=1.423986 Training: 44%|████▍ | 7924/18018 [17:13:55<20:13:30, 7.21s/step, loss=1.424, acc=0.682, acc_len=1.62, thru=20.4, I=25.7, T=39.2, wait=2.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:15,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:44:16,289] loop.py:380 INFO TIMING step=7925: step=4.081s data=1.659s compute=4.057s [fwd=2.879s bwd=1.163s opt=0.003s] dispatch=3.147s [2026-10-19 06:44:16,289] loop.py:45 INFO TRAIN_STEP step=7925 loss=1.323184 Training: 44%|████▍ | 7925/18018 [17:14:02<20:14:40, 7.22s/step, loss=1.323, acc=0.701, acc_len=1.68, thru=18.9, I=19.7, T=31.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:16,566] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:44:25,252] loop.py:45 INFO TRAIN_STEP step=7926 loss=1.421929 Training: 44%|████▍ | 7926/18018 [17:14:11<21:42:33, 7.74s/step, loss=1.422, acc=0.677, acc_len=1.60, thru=11.7, I=14.7, T=30.4, wait=4.7s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:44:25,248] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7925: forward=2886.9ms backward=1252.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:34,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:44:35,022] loop.py:45 INFO TRAIN_STEP step=7927 loss=1.446795 Training: 44%|████▍ | 7927/18018 [17:14:21<23:24:26, 8.35s/step, loss=1.447, acc=0.677, acc_len=1.61, thru=15.3, I=13.2, T=34.2, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:35,459] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:41,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:44:42,972] loop.py:45 INFO TRAIN_STEP step=7928 loss=1.466942 Training: 44%|████▍ | 7928/18018 [17:14:29<23:04:11, 8.23s/step, loss=1.467, acc=0.674, acc_len=1.59, thru=18.4, I=21.9, T=25.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:43,051] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:48,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:44:49,071] loop.py:45 INFO TRAIN_STEP step=7929 loss=1.435742 Training: 44%|████▍ | 7929/18018 [17:14:35<21:16:25, 7.59s/step, loss=1.436, acc=0.676, acc_len=1.60, thru=17.1, I=20.3, T=34.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:49,302] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:53,693] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:44:54,984] loop.py:380 INFO TIMING step=7930: step=3.657s data=1.588s compute=3.624s [fwd=2.613s bwd=0.996s opt=0.003s] dispatch=2.248s [2026-10-19 06:44:54,984] loop.py:45 INFO TRAIN_STEP step=7930 loss=1.468797 Training: 44%|████▍ | 7930/18018 [17:14:41<19:51:44, 7.09s/step, loss=1.469, acc=0.671, acc_len=1.58, thru=21.1, I=25.3, T=35.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:44:55,204] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:45:01,168] loop.py:45 INFO TRAIN_STEP step=7931 loss=1.518522 Training: 44%|████▍ | 7931/18018 [17:14:47<19:06:11, 6.82s/step, loss=1.519, acc=0.670, acc_len=1.58, thru=18.6, I=22.3, T=35.9, wait=2.6s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:45:01,164] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7930: forward=2556.8ms backward=964.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:06,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:45:08,877] loop.py:45 INFO TRAIN_STEP step=7932 loss=1.339392 Training: 44%|████▍ | 7932/18018 [17:14:54<19:50:46, 7.08s/step, loss=1.339, acc=0.693, acc_len=1.66, thru=17.3, I=21.2, T=29.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:09,104] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:15,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:45:17,059] loop.py:45 INFO TRAIN_STEP step=7933 loss=1.261572 Training: 44%|████▍ | 7933/18018 [17:15:03<20:45:58, 7.41s/step, loss=1.262, acc=0.711, acc_len=1.72, thru=16.1, I=20.5, T=27.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:17,089] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:22,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:45:23,347] loop.py:45 INFO TRAIN_STEP step=7934 loss=1.435956 Training: 44%|████▍ | 7934/18018 [17:15:09<19:49:12, 7.08s/step, loss=1.436, acc=0.682, acc_len=1.62, thru=17.7, I=23.2, T=34.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:23,770] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:45:29,309] loop.py:380 INFO TIMING step=7935: step=3.551s data=1.483s compute=3.523s [fwd=2.492s bwd=1.016s opt=0.003s] dispatch=2.402s [2026-10-19 06:45:29,309] loop.py:45 INFO TRAIN_STEP step=7935 loss=1.289683 Training: 44%|████▍ | 7935/18018 [17:15:15<18:52:53, 6.74s/step, loss=1.290, acc=0.706, acc_len=1.70, thru=18.0, I=21.9, T=36.0, wait=2.4s, pool=40, epoch=1/2][2026-10-19 06:45:38,249] loop.py:45 INFO TRAIN_STEP step=7936 loss=1.370234 Training: 44%|████▍ | 7936/18018 [17:15:24<20:43:36, 7.40s/step, loss=1.370, acc=0.690, acc_len=1.66, thru=13.3, I=14.0, T=33.5, wait=5.1s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:45:38,244] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7935: forward=2709.0ms backward=1067.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:50,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:45:51,182] loop.py:45 INFO TRAIN_STEP step=7937 loss=1.430030 Training: 44%|████▍ | 7937/18018 [17:15:37<25:22:23, 9.06s/step, loss=1.430, acc=0.684, acc_len=1.62, thru=13.2, I=13.6, T=30.1, wait=8.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:51,569] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:57,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:45:58,324] loop.py:45 INFO TRAIN_STEP step=7938 loss=1.329965 Training: 44%|████▍ | 7938/18018 [17:15:44<23:45:30, 8.49s/step, loss=1.330, acc=0.698, acc_len=1.68, thru=17.7, I=19.4, T=33.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:45:58,737] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:46:03,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:46:04,624] loop.py:45 INFO TRAIN_STEP step=7939 loss=1.400718 Training: 44%|████▍ | 7939/18018 [17:15:50<21:55:19, 7.83s/step, loss=1.401, acc=0.684, acc_len=1.63, thru=18.9, I=24.8, T=36.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:46:04,804] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:46:11,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:46:12,372] loop.py:380 INFO TIMING step=7940: step=3.571s data=1.552s compute=3.539s [fwd=2.528s bwd=0.997s opt=0.003s] dispatch=4.166s [2026-10-19 06:46:12,372] loop.py:45 INFO TRAIN_STEP step=7940 loss=1.382221 Training: 44%|████▍ | 7940/18018 [17:15:58<21:50:59, 7.81s/step, loss=1.382, acc=0.688, acc_len=1.64, thru=16.4, I=19.9, T=35.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:46:12,738] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 06:46:19,070] loop.py:45 INFO TRAIN_STEP step=7941 loss=1.356247 Training: 44%|████▍ | 7941/18018 [17:16:05<20:55:08, 7.47s/step, loss=1.356, acc=0.691, acc_len=1.66, thru=18.4, I=22.7, T=30.2, wait=2.4s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:46:19,065] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7940: forward=2927.5ms backward=1266.9ms [2026-10-19 06:46:30,796] loop.py:45 INFO TRAIN_STEP step=7942 loss=1.534226 Training: 44%|████▍ | 7942/18018 [17:16:16<24:29:10, 8.75s/step, loss=1.534, acc=0.663, acc_len=1.56, thru=11.0, I=10.2, T=31.8, wait=7.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:46:38,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:46:40,472] loop.py:45 INFO TRAIN_STEP step=7943 loss=1.431965 Training: 44%|████▍ | 7943/18018 [17:16:26<25:15:45, 9.03s/step, loss=1.432, acc=0.678, acc_len=1.62, thru=14.5, I=16.4, T=29.3, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:46:40,913] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:46:45,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:46:47,181] loop.py:45 INFO TRAIN_STEP step=7944 loss=1.416932 Training: 44%|████▍ | 7944/18018 [17:16:33<23:18:52, 8.33s/step, loss=1.417, acc=0.680, acc_len=1.61, thru=16.9, I=26.5, T=32.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:46:47,565] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:46:53,817] loop.py:380 INFO TIMING step=7945: step=3.739s data=1.650s compute=3.709s [fwd=2.638s bwd=1.057s opt=0.003s] dispatch=2.888s [2026-10-19 06:46:53,818] loop.py:45 INFO TRAIN_STEP step=7945 loss=1.364834 Training: 44%|████▍ | 7945/18018 [17:16:39<21:53:24, 7.82s/step, loss=1.365, acc=0.689, acc_len=1.64, thru=15.2, I=20.9, T=34.2, wait=2.9s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:47:04,897] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:47:05,951] loop.py:45 INFO TRAIN_STEP step=7946 loss=1.427716 Training: 44%|████▍ | 7946/18018 [17:16:52<25:30:15, 9.12s/step, loss=1.428, acc=0.684, acc_len=1.63, thru=13.9, I=12.8, T=27.0, wait=7.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:47:05,946] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7945: forward=3360.7ms backward=1313.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:47:06,405] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:47:13,272] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:47:15,061] loop.py:45 INFO TRAIN_STEP step=7947 loss=1.262493 Training: 44%|████▍ | 7947/18018 [17:17:01<25:29:54, 9.11s/step, loss=1.262, acc=0.708, acc_len=1.71, thru=14.8, I=16.7, T=28.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:47:15,284] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:47:22,082] loop.py:45 INFO TRAIN_STEP step=7948 loss=1.370460 Training: 44%|████▍ | 7948/18018 [17:17:08<23:44:28, 8.49s/step, loss=1.370, acc=0.695, acc_len=1.66, thru=14.2, I=19.4, T=29.8, wait=2.7s, pool=56, epoch=1/2][2026-10-19 06:47:30,010] loop.py:45 INFO TRAIN_STEP step=7949 loss=1.468243 Training: 44%|████▍ | 7949/18018 [17:17:16<23:16:01, 8.32s/step, loss=1.468, acc=0.676, acc_len=1.60, thru=14.5, I=14.7, T=31.4, wait=3.8s, pool=48, epoch=1/2][2026-10-19 06:47:37,115] loop.py:380 INFO TIMING step=7950: step=3.920s data=1.769s compute=3.878s [fwd=2.748s bwd=1.115s opt=0.003s] dispatch=3.175s [2026-10-19 06:47:37,115] loop.py:45 INFO TRAIN_STEP step=7950 loss=1.414692 Training: 44%|████▍ | 7950/18018 [17:17:23<22:14:51, 7.96s/step, loss=1.415, acc=0.680, acc_len=1.61, thru=17.0, I=16.5, T=32.7, wait=3.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:47:43,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:47:43,481] loop.py:45 INFO TRAIN_STEP step=7951 loss=1.442518 Training: 44%|████▍ | 7951/18018 [17:17:29<20:54:39, 7.48s/step, loss=1.443, acc=0.672, acc_len=1.59, thru=18.8, I=19.1, T=35.0, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:47:43,477] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7950: forward=2543.1ms backward=1057.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:47:43,706] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:47:51,015] loop.py:45 INFO TRAIN_STEP step=7952 loss=1.425802 Training: 44%|████▍ | 7952/18018 [17:17:37<20:57:11, 7.49s/step, loss=1.426, acc=0.683, acc_len=1.63, thru=15.5, I=15.7, T=30.1, wait=3.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:47:58,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:47:59,184] loop.py:45 INFO TRAIN_STEP step=7953 loss=1.379349 Training: 44%|████▍ | 7953/18018 [17:17:45<21:31:14, 7.70s/step, loss=1.379, acc=0.688, acc_len=1.64, thru=17.4, I=17.3, T=33.0, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:47:59,283] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:08,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:48:09,038] loop.py:45 INFO TRAIN_STEP step=7954 loss=1.368859 Training: 44%|████▍ | 7954/18018 [17:17:55<23:19:32, 8.34s/step, loss=1.369, acc=0.694, acc_len=1.67, thru=13.2, I=14.1, T=31.2, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:09,085] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:17,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:48:17,651] loop.py:380 INFO TIMING step=7955: step=5.671s data=2.797s compute=5.638s [fwd=3.977s bwd=1.647s opt=0.003s] dispatch=2.933s [2026-10-19 06:48:17,651] loop.py:45 INFO TRAIN_STEP step=7955 loss=1.387632 Training: 44%|████▍ | 7955/18018 [17:18:03<23:33:00, 8.42s/step, loss=1.388, acc=0.689, acc_len=1.65, thru=14.0, I=16.5, T=22.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:17,897] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:25,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:48:27,779] loop.py:45 INFO TRAIN_STEP step=7956 loss=1.611977 Training: 44%|████▍ | 7956/18018 [17:18:13<24:58:33, 8.94s/step, loss=1.612, acc=0.651, acc_len=1.52, thru=13.0, I=14.4, T=25.3, wait=5.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:48:27,774] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7955: forward=3533.8ms backward=1469.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:27,870] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:35,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:48:36,937] loop.py:45 INFO TRAIN_STEP step=7957 loss=1.431117 Training: 44%|████▍ | 7957/18018 [17:18:23<25:09:36, 9.00s/step, loss=1.431, acc=0.679, acc_len=1.61, thru=14.1, I=14.6, T=30.3, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:37,012] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:41,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:48:43,144] loop.py:45 INFO TRAIN_STEP step=7958 loss=1.444642 Training: 44%|████▍ | 7958/18018 [17:18:29<22:48:46, 8.16s/step, loss=1.445, acc=0.681, acc_len=1.61, thru=18.8, I=25.2, T=32.8, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:43,260] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:51,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:48:53,136] loop.py:45 INFO TRAIN_STEP step=7959 loss=1.340135 Training: 44%|████▍ | 7959/18018 [17:18:39<24:20:32, 8.71s/step, loss=1.340, acc=0.698, acc_len=1.67, thru=13.0, I=14.9, T=28.4, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:53,342] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:48:58,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:49:00,064] loop.py:380 INFO TIMING step=7960: step=4.076s data=1.911s compute=4.030s [fwd=2.854s bwd=1.160s opt=0.003s] dispatch=2.843s [2026-10-19 06:49:00,064] loop.py:45 INFO TRAIN_STEP step=7960 loss=1.440309 Training: 44%|████▍ | 7960/18018 [17:18:46<22:50:46, 8.18s/step, loss=1.440, acc=0.678, acc_len=1.61, thru=15.9, I=21.4, T=31.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:00,373] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:49:07,411] loop.py:45 INFO TRAIN_STEP step=7961 loss=1.459678 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:07,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 Training: 44%|████▍ | 7961/18018 [17:18:53<22:08:48, 7.93s/step, loss=1.460, acc=0.674, acc_len=1.60, thru=16.5, I=18.7, T=30.3, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:49:07,407] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7960: forward=3012.6ms backward=1166.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:07,847] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:15,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:49:15,542] loop.py:45 INFO TRAIN_STEP step=7962 loss=1.439243 Training: 44%|████▍ | 7962/18018 [17:19:01<22:18:55, 7.99s/step, loss=1.439, acc=0.681, acc_len=1.61, thru=15.7, I=16.2, T=30.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:15,756] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:22,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:49:25,999] loop.py:45 INFO TRAIN_STEP step=7963 loss=1.409932 Training: 44%|████▍ | 7963/18018 [17:19:11<23:32:36, 8.43s/step, loss=1.410, acc=0.684, acc_len=1.64, thru=13.5, I=19.6, T=25.6, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:25,453] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:30,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:49:31,771] loop.py:45 INFO TRAIN_STEP step=7964 loss=1.431971 Training: 44%|████▍ | 7964/18018 [17:19:17<22:09:12, 7.93s/step, loss=1.432, acc=0.678, acc_len=1.60, thru=14.5, I=23.6, T=34.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:31,958] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:41,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:49:42,238] loop.py:380 INFO TIMING step=7965: step=4.738s data=2.193s compute=4.696s [fwd=3.280s bwd=1.400s opt=0.003s] dispatch=5.719s [2026-10-19 06:49:42,239] loop.py:45 INFO TRAIN_STEP step=7965 loss=1.550574 Training: 44%|████▍ | 7965/18018 [17:19:28<24:16:25, 8.69s/step, loss=1.551, acc=0.663, acc_len=1.56, thru=12.9, I=12.7, T=27.0, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:42,511] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:49,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:49:50,344] loop.py:45 INFO TRAIN_STEP step=7966 loss=1.367585 Training: 44%|████▍ | 7966/18018 [17:19:36<23:46:52, 8.52s/step, loss=1.368, acc=0.690, acc_len=1.65, thru=15.2, I=18.2, T=30.4, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:49:50,339] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7965: forward=2967.5ms backward=1195.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:50,631] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:56,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:49:57,415] loop.py:45 INFO TRAIN_STEP step=7967 loss=1.465723 Training: 44%|████▍ | 7967/18018 [17:19:43<22:33:57, 8.08s/step, loss=1.466, acc=0.671, acc_len=1.58, thru=17.0, I=21.2, T=34.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:49:57,547] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:02,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:50:05,468] loop.py:45 INFO TRAIN_STEP step=7968 loss=1.554636 Training: 44%|████▍ | 7968/18018 [17:19:51<22:32:21, 8.07s/step, loss=1.555, acc=0.658, acc_len=1.54, thru=16.1, I=22.3, T=25.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:05,928] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:12,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:50:13,183] loop.py:45 INFO TRAIN_STEP step=7969 loss=1.409005 Training: 44%|████▍ | 7969/18018 [17:19:59<22:14:14, 7.97s/step, loss=1.409, acc=0.685, acc_len=1.63, thru=17.6, I=18.6, T=33.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:13,658] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:50:19,424] loop.py:380 INFO TIMING step=7970: step=3.326s data=1.322s compute=3.293s [fwd=2.326s bwd=0.952s opt=0.003s] dispatch=2.904s [2026-10-19 06:50:19,424] loop.py:45 INFO TRAIN_STEP step=7970 loss=1.263167 Training: 44%|████▍ | 7970/18018 [17:20:05<20:47:25, 7.45s/step, loss=1.263, acc=0.710, acc_len=1.71, thru=18.7, I=20.8, T=38.5, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:30,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:50:31,367] loop.py:45 INFO TRAIN_STEP step=7971 loss=1.401197 Training: 44%|████▍ | 7971/18018 [17:20:17<24:33:01, 8.80s/step, loss=1.401, acc=0.684, acc_len=1.63, thru=10.2, I=10.6, T=27.7, wait=7.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:50:31,363] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7970: forward=3201.6ms backward=1367.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:31,447] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:38,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:50:39,344] loop.py:45 INFO TRAIN_STEP step=7972 loss=1.389151 Training: 44%|████▍ | 7972/18018 [17:20:25<23:51:45, 8.55s/step, loss=1.389, acc=0.688, acc_len=1.64, thru=15.7, I=16.7, T=30.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:39,445] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 06:50:46,395] loop.py:45 INFO TRAIN_STEP step=7973 loss=1.358085 Training: 44%|████▍ | 7973/18018 [17:20:32<22:36:08, 8.10s/step, loss=1.358, acc=0.691, acc_len=1.65, thru=16.3, I=20.2, T=34.2, wait=3.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:52,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:50:53,955] loop.py:45 INFO TRAIN_STEP step=7974 loss=1.371553 Training: 44%|████▍ | 7974/18018 [17:20:40<22:08:56, 7.94s/step, loss=1.372, acc=0.689, acc_len=1.65, thru=17.1, I=18.7, T=33.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:50:54,396] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:51:01,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:51:03,063] loop.py:380 INFO TIMING step=7975: step=4.014s data=1.857s compute=3.976s [fwd=2.844s bwd=1.116s opt=0.003s] dispatch=5.084s [2026-10-19 06:51:03,063] loop.py:45 INFO TRAIN_STEP step=7975 loss=1.367774 Training: 44%|████▍ | 7975/18018 [17:20:49<23:07:34, 8.29s/step, loss=1.368, acc=0.692, acc_len=1.66, thru=14.8, I=17.9, T=31.9, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:51:03,173] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-19 06:51:10,099] loop.py:45 INFO TRAIN_STEP step=7976 loss=1.471169 Training: 44%|████▍ | 7976/18018 [17:20:56<22:04:26, 7.91s/step, loss=1.471, acc=0.668, acc_len=1.57, thru=16.2, I=19.6, T=29.8, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:51:10,094] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7975: forward=3081.0ms backward=1164.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:51:20,114] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:51:21,486] loop.py:45 INFO TRAIN_STEP step=7977 loss=1.345406 Training: 44%|████▍ | 7977/18018 [17:21:07<24:58:41, 8.96s/step, loss=1.345, acc=0.700, acc_len=1.68, thru=12.7, I=12.8, T=27.3, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:51:21,633] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:51:28,907] loop.py:45 INFO TRAIN_STEP step=7978 loss=1.443937 Training: 44%|████▍ | 7978/18018 [17:21:14<23:41:28, 8.49s/step, loss=1.444, acc=0.679, acc_len=1.61, thru=14.1, I=14.6, T=28.6, wait=2.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:51:35,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:51:36,687] loop.py:45 INFO TRAIN_STEP step=7979 loss=1.287655 Training: 44%|████▍ | 7979/18018 [17:21:22<23:05:28, 8.28s/step, loss=1.288, acc=0.704, acc_len=1.70, thru=17.7, I=17.0, T=30.5, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:51:36,883] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:51:46,266] loop.py:380 INFO TIMING step=7980: step=4.244s data=1.975s compute=4.187s [fwd=2.936s bwd=1.234s opt=0.003s] dispatch=5.327s [2026-10-19 06:51:46,267] loop.py:45 INFO TRAIN_STEP step=7980 loss=1.414051 Training: 44%|████▍ | 7980/18018 [17:21:32<24:10:29, 8.67s/step, loss=1.414, acc=0.684, acc_len=1.63, thru=12.8, I=11.9, T=30.2, wait=5.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:51:56,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:51:56,894] loop.py:45 INFO TRAIN_STEP step=7981 loss=1.454661 Training: 44%|████▍ | 7981/18018 [17:21:42<25:48:41, 9.26s/step, loss=1.455, acc=0.674, acc_len=1.60, thru=12.1, I=12.6, T=33.6, wait=6.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:51:56,889] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7980: forward=2634.7ms backward=1091.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:51:57,230] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:52:04,087] loop.py:45 INFO TRAIN_STEP step=7982 loss=1.435192 Training: 44%|████▍ | 7982/18018 [17:21:50<24:04:51, 8.64s/step, loss=1.435, acc=0.680, acc_len=1.61, thru=16.0, I=17.1, T=34.8, wait=3.5s, pool=44, epoch=1/2][2026-10-19 06:52:11,137] loop.py:45 INFO TRAIN_STEP step=7983 loss=1.389652 Training: 44%|████▍ | 7983/18018 [17:21:57<22:45:01, 8.16s/step, loss=1.390, acc=0.689, acc_len=1.65, thru=18.4, I=17.6, T=36.7, wait=3.6s, pool=60, epoch=1/2][2026-10-19 06:52:19,475] loop.py:45 INFO TRAIN_STEP step=7984 loss=1.444490 Training: 44%|████▍ | 7984/18018 [17:22:05<22:53:41, 8.21s/step, loss=1.444, acc=0.673, acc_len=1.59, thru=12.2, I=14.6, T=35.7, wait=4.7s, pool=24, epoch=1/2][2026-10-19 06:52:30,005] loop.py:380 INFO TIMING step=7985: step=3.855s data=1.671s compute=3.805s [fwd=2.725s bwd=1.067s opt=0.003s] dispatch=6.666s [2026-10-19 06:52:30,005] loop.py:45 INFO TRAIN_STEP step=7985 loss=1.329337 Training: 44%|████▍ | 7985/18018 [17:22:16<24:49:41, 8.91s/step, loss=1.329, acc=0.698, acc_len=1.68, thru=16.4, I=13.9, T=33.2, wait=6.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:52:35,518] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:52:37,060] loop.py:45 INFO TRAIN_STEP step=7986 loss=1.398025 Training: 44%|████▍ | 7986/18018 [17:22:23<23:16:34, 8.35s/step, loss=1.398, acc=0.686, acc_len=1.64, thru=18.5, I=20.7, T=31.7, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:52:37,055] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7985: forward=2758.1ms backward=1234.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:52:37,532] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:52:44,255] loop.py:45 INFO TRAIN_STEP step=7987 loss=1.397073 Training: 44%|████▍ | 7987/18018 [17:22:30<22:18:23, 8.01s/step, loss=1.397, acc=0.684, acc_len=1.63, thru=14.1, I=16.2, T=33.9, wait=3.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:52:49,903] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:52:51,450] loop.py:45 INFO TRAIN_STEP step=7988 loss=1.393092 Training: 44%|████▍ | 7988/18018 [17:22:37<21:37:42, 7.76s/step, loss=1.393, acc=0.686, acc_len=1.64, thru=17.7, I=21.1, T=33.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:52:51,915] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:52:58,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:52:58,788] loop.py:45 INFO TRAIN_STEP step=7989 loss=1.297144 Training: 44%|████▍ | 7989/18018 [17:22:44<21:16:12, 7.64s/step, loss=1.297, acc=0.702, acc_len=1.69, thru=16.7, I=23.6, T=37.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:52:59,186] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:53:03,947] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:53:04,978] loop.py:380 INFO TIMING step=7990: step=3.317s data=1.335s compute=3.291s [fwd=2.359s bwd=0.919s opt=0.003s] dispatch=2.866s [2026-10-19 06:53:04,979] loop.py:45 INFO TRAIN_STEP step=7990 loss=1.378552 Training: 44%|████▍ | 7990/18018 [17:22:51<20:03:35, 7.20s/step, loss=1.379, acc=0.685, acc_len=1.63, thru=20.5, I=25.0, T=38.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:53:05,454] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:53:12,443] loop.py:45 INFO TRAIN_STEP step=7991 loss=1.383587 Training: 44%|████▍ | 7991/18018 [17:22:58<20:16:47, 7.28s/step, loss=1.384, acc=0.685, acc_len=1.64, thru=14.4, I=21.3, T=34.5, wait=3.8s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:53:12,439] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7990: forward=2639.7ms backward=1026.5ms [2026-10-19 06:53:22,101] loop.py:45 INFO TRAIN_STEP step=7992 loss=1.347873 Training: 44%|████▍ | 7992/18018 [17:23:08<22:15:47, 7.99s/step, loss=1.348, acc=0.697, acc_len=1.68, thru=15.0, I=14.3, T=28.9, wait=5.2s, pool=44, epoch=1/2][2026-10-19 06:53:30,795] loop.py:45 INFO TRAIN_STEP step=7993 loss=1.341827 Training: 44%|████▍ | 7993/18018 [17:23:16<22:50:44, 8.20s/step, loss=1.342, acc=0.697, acc_len=1.68, thru=14.4, I=15.6, T=33.0, wait=4.8s, pool=52, epoch=1/2][2026-10-19 06:53:41,773] loop.py:45 INFO TRAIN_STEP step=7994 loss=1.270701 Training: 44%|████▍ | 7994/18018 [17:23:27<25:09:34, 9.04s/step, loss=1.271, acc=0.708, acc_len=1.72, thru=12.3, I=10.5, T=30.4, wait=6.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:53:47,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:53:48,745] loop.py:380 INFO TIMING step=7995: step=3.982s data=1.744s compute=3.945s [fwd=2.780s bwd=1.152s opt=0.003s] dispatch=2.983s [2026-10-19 06:53:48,746] loop.py:45 INFO TRAIN_STEP step=7995 loss=1.395739 Training: 44%|████▍ | 7995/18018 [17:23:34<23:26:06, 8.42s/step, loss=1.396, acc=0.685, acc_len=1.63, thru=18.9, I=18.3, T=32.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:53:49,226] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:53:56,512] loop.py:45 INFO TRAIN_STEP step=7996 loss=1.470926 Training: 44%|████▍ | 7996/18018 [17:23:42<22:53:21, 8.22s/step, loss=1.471, acc=0.670, acc_len=1.59, thru=15.0, I=17.3, T=28.8, wait=3.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:53:56,508] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=7995: forward=3329.2ms backward=1072.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:01,416] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:54:02,611] loop.py:45 INFO TRAIN_STEP step=7997 loss=1.571949 Training: 44%|████▍ | 7997/18018 [17:23:48<21:06:48, 7.58s/step, loss=1.572, acc=0.655, acc_len=1.54, thru=18.4, I=20.0, T=34.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:02,923] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:54:08,583] loop.py:45 INFO TRAIN_STEP step=7998 loss=1.306614 Training: 44%|████▍ | 7998/18018 [17:23:54<19:45:54, 7.10s/step, loss=1.307, acc=0.701, acc_len=1.68, thru=20.4, I=24.7, T=37.4, wait=2.5s, pool=60, epoch=1/2][2026-10-19 06:54:16,006] loop.py:45 INFO TRAIN_STEP step=7999 loss=1.291246 Training: 44%|████▍ | 7999/18018 [17:24:02<20:01:49, 7.20s/step, loss=1.291, acc=0.701, acc_len=1.69, thru=16.4, I=18.5, T=32.9, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:23,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:54:24,111] loop.py:380 INFO TIMING step=8000: step=4.076s data=1.907s compute=4.035s [fwd=2.825s bwd=1.196s opt=0.003s] dispatch=4.022s [2026-10-19 06:54:24,111] loop.py:45 INFO TRAIN_STEP step=8000 loss=1.395246 Training: 44%|████▍ | 8000/18018 [17:24:10<20:47:17, 7.47s/step, loss=1.395, acc=0.687, acc_len=1.63, thru=15.6, I=16.5, T=31.4, wait=4.0s, pool=64, epoch=1/2][2026-10-19 06:54:24,116] loop.py:419 INFO Saving checkpoint at step 8000... (TrainerActor pid=9310) /usr/local/lib/python3.12/dist-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning. (TrainerActor pid=9310) return func(*args, **kwargs) (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:28,193] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 06:54:28,288] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0008001 (TrainerActor pid=9310) [2026-10-19 06:54:28,297] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0007251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:29,701] inference_manager.py:408 INFO Pool capacity available after 6.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:35,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 06:54:28,287] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0008001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 06:54:29,318] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0007251 [2026-10-19 06:54:36,047] loop.py:45 INFO TRAIN_STEP step=8001 loss=1.441654 Training: 44%|████▍ | 8001/18018 [17:24:22<24:30:49, 8.81s/step, loss=1.442, acc=0.677, acc_len=1.61, thru=20.2, I=23.0, T=30.9, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:54:36,042] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8000: forward=2886.3ms backward=1208.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:36,212] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:43,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:54:44,026] loop.py:45 INFO TRAIN_STEP step=8002 loss=1.455104 Training: 44%|████▍ | 8002/18018 [17:24:30<23:48:56, 8.56s/step, loss=1.455, acc=0.677, acc_len=1.61, thru=14.6, I=19.0, T=32.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:44,343] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:52,535] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:54:54,585] loop.py:45 INFO TRAIN_STEP step=8003 loss=1.437519 Training: 44%|████▍ | 8003/18018 [17:24:40<25:28:52, 9.16s/step, loss=1.438, acc=0.678, acc_len=1.61, thru=13.2, I=13.9, T=26.5, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:54:55,041] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 06:55:03,004] loop.py:45 INFO TRAIN_STEP step=8004 loss=1.462419 Training: 44%|████▍ | 8004/18018 [17:24:49<24:51:47, 8.94s/step, loss=1.462, acc=0.675, acc_len=1.60, thru=14.1, I=15.5, T=27.2, wait=3.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:55:09,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:55:11,616] loop.py:380 INFO TIMING step=8005: step=4.394s data=2.021s compute=4.361s [fwd=3.149s bwd=1.198s opt=0.003s] dispatch=4.207s [2026-10-19 06:55:11,616] loop.py:45 INFO TRAIN_STEP step=8005 loss=1.405259 Training: 44%|████▍ | 8005/18018 [17:24:57<24:35:18, 8.84s/step, loss=1.405, acc=0.687, acc_len=1.64, thru=15.5, I=17.6, T=29.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:55:11,969] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:55:18,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:55:18,464] loop.py:45 INFO TRAIN_STEP step=8006 loss=1.330286 Training: 44%|████▍ | 8006/18018 [17:25:04<22:55:25, 8.24s/step, loss=1.330, acc=0.696, acc_len=1.66, thru=17.0, I=21.0, T=27.5, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:55:18,460] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8005: forward=3194.5ms backward=1419.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:55:18,592] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 [2026-10-19 06:55:25,984] loop.py:45 INFO TRAIN_STEP step=8007 loss=1.355625 Training: 44%|████▍ | 8007/18018 [17:25:12<22:19:00, 8.03s/step, loss=1.356, acc=0.694, acc_len=1.66, thru=16.9, I=16.1, T=32.4, wait=3.6s, pool=60, epoch=1/2][2026-10-19 06:55:32,147] loop.py:45 INFO TRAIN_STEP step=8008 loss=1.343250 Training: 44%|████▍ | 8008/18018 [17:25:18<20:45:42, 7.47s/step, loss=1.343, acc=0.697, acc_len=1.67, thru=18.4, I=20.0, T=33.5, wait=2.3s, pool=60, epoch=1/2][2026-10-19 06:55:40,575] loop.py:45 INFO TRAIN_STEP step=8009 loss=1.448868 Training: 44%|████▍ | 8009/18018 [17:25:26<21:33:40, 7.76s/step, loss=1.449, acc=0.678, acc_len=1.61, thru=15.0, I=15.8, T=33.2, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:55:46,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:55:48,026] loop.py:380 INFO TIMING step=8010: step=3.849s data=1.697s compute=3.800s [fwd=2.679s bwd=1.106s opt=0.003s] dispatch=3.593s [2026-10-19 06:55:48,026] loop.py:45 INFO TRAIN_STEP step=8010 loss=1.266370 Training: 44%|████▍ | 8010/18018 [17:25:34<21:18:12, 7.66s/step, loss=1.266, acc=0.703, acc_len=1.70, thru=16.2, I=19.3, T=33.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:55:48,220] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 06:55:55,613] loop.py:45 INFO TRAIN_STEP step=8011 loss=1.366137 Training: 44%|████▍ | 8011/18018 [17:25:41<21:14:25, 7.64s/step, loss=1.366, acc=0.693, acc_len=1.66, thru=11.6, I=20.9, T=34.3, wait=3.8s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:55:55,608] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8010: forward=2629.1ms backward=1048.5ms [2026-10-19 06:56:03,742] loop.py:45 INFO TRAIN_STEP step=8012 loss=1.440281 Training: 44%|████▍ | 8012/18018 [17:25:49<21:38:36, 7.79s/step, loss=1.440, acc=0.677, acc_len=1.61, thru=16.5, I=14.8, T=32.9, wait=4.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:10,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:56:11,270] loop.py:45 INFO TRAIN_STEP step=8013 loss=1.337336 Training: 44%|████▍ | 8013/18018 [17:25:57<21:25:32, 7.71s/step, loss=1.337, acc=0.699, acc_len=1.68, thru=17.1, I=18.1, T=33.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:11,544] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:15,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:56:17,285] loop.py:45 INFO TRAIN_STEP step=8014 loss=1.450875 Training: 44%|████▍ | 8014/18018 [17:26:03<20:00:39, 7.20s/step, loss=1.451, acc=0.678, acc_len=1.61, thru=19.8, I=26.3, T=36.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:17,761] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:23,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:56:24,296] loop.py:380 INFO TIMING step=8015: step=3.733s data=1.649s compute=3.700s [fwd=2.664s bwd=1.022s opt=0.003s] dispatch=3.269s [2026-10-19 06:56:24,296] loop.py:45 INFO TRAIN_STEP step=8015 loss=1.338303 Training: 44%|████▍ | 8015/18018 [17:26:10<19:51:06, 7.14s/step, loss=1.338, acc=0.695, acc_len=1.67, thru=17.8, I=22.1, T=34.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:24,740] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:32,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:56:33,599] loop.py:45 INFO TRAIN_STEP step=8016 loss=1.422640 Training: 44%|████▍ | 8016/18018 [17:26:19<21:38:59, 7.79s/step, loss=1.423, acc=0.683, acc_len=1.62, thru=14.4, I=15.4, T=31.1, wait=5.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:56:33,595] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8015: forward=2914.0ms backward=1152.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:33,867] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:39,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:56:40,982] loop.py:45 INFO TRAIN_STEP step=8017 loss=1.326261 Training: 44%|████▍ | 8017/18018 [17:26:27<21:18:16, 7.67s/step, loss=1.326, acc=0.698, acc_len=1.67, thru=17.1, I=21.9, T=32.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:41,405] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:47,218] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:56:48,658] loop.py:45 INFO TRAIN_STEP step=8018 loss=1.289093 Training: 44%|████▍ | 8018/18018 [17:26:34<21:18:29, 7.67s/step, loss=1.289, acc=0.699, acc_len=1.68, thru=14.6, I=20.9, T=34.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:48,729] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:57,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:56:57,943] loop.py:45 INFO TRAIN_STEP step=8019 loss=1.445835 Training: 45%|████▍ | 8019/18018 [17:26:44<22:39:10, 8.16s/step, loss=1.446, acc=0.675, acc_len=1.60, thru=13.9, I=14.9, T=27.6, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:56:58,156] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:04,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:57:05,223] loop.py:380 INFO TIMING step=8020: step=4.199s data=1.898s compute=4.160s [fwd=2.953s bwd=1.193s opt=0.003s] dispatch=3.070s [2026-10-19 06:57:05,223] loop.py:45 INFO TRAIN_STEP step=8020 loss=1.518207 Training: 45%|████▍ | 8020/18018 [17:26:51<21:55:13, 7.89s/step, loss=1.518, acc=0.662, acc_len=1.55, thru=17.1, I=18.1, T=30.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:05,452] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 06:57:12,889] loop.py:45 INFO TRAIN_STEP step=8021 loss=1.412002 Training: 45%|████▍ | 8021/18018 [17:26:58<21:43:43, 7.82s/step, loss=1.412, acc=0.684, acc_len=1.63, thru=15.1, I=15.8, T=32.7, wait=3.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:57:12,883] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8020: forward=2666.1ms backward=1170.5ms [2026-10-19 06:57:22,438] loop.py:45 INFO TRAIN_STEP step=8022 loss=1.351785 Training: 45%|████▍ | 8022/18018 [17:27:08<23:09:40, 8.34s/step, loss=1.352, acc=0.695, acc_len=1.67, thru=15.8, I=15.7, T=31.2, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:22,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:22,934] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:33,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:57:33,954] loop.py:45 INFO TRAIN_STEP step=8023 loss=1.454531 Training: 45%|████▍ | 8023/18018 [17:27:20<25:48:20, 9.29s/step, loss=1.455, acc=0.675, acc_len=1.59, thru=12.4, I=12.2, T=23.6, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:34,305] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:43,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:57:44,431] loop.py:45 INFO TRAIN_STEP step=8024 loss=1.325783 Training: 45%|████▍ | 8024/18018 [17:27:30<26:47:18, 9.65s/step, loss=1.326, acc=0.700, acc_len=1.68, thru=12.7, I=13.6, T=29.4, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:44,523] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:51,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:57:52,315] loop.py:380 INFO TIMING step=8025: step=3.746s data=1.563s compute=3.714s [fwd=2.591s bwd=1.109s opt=0.003s] dispatch=4.128s [2026-10-19 06:57:52,316] loop.py:45 INFO TRAIN_STEP step=8025 loss=1.462600 Training: 45%|████▍ | 8025/18018 [17:27:38<25:18:47, 9.12s/step, loss=1.463, acc=0.671, acc_len=1.59, thru=15.1, I=16.5, T=34.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:52,500] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:57,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:57:59,018] loop.py:45 INFO TRAIN_STEP step=8026 loss=1.434953 Training: 45%|████▍ | 8026/18018 [17:27:45<23:18:02, 8.39s/step, loss=1.435, acc=0.681, acc_len=1.62, thru=18.3, I=22.4, T=33.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:57:59,013] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8025: forward=2700.8ms backward=1121.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:57:59,089] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-19 06:58:05,723] loop.py:45 INFO TRAIN_STEP step=8027 loss=1.293154 Training: 45%|████▍ | 8027/18018 [17:27:51<21:53:29, 7.89s/step, loss=1.293, acc=0.707, acc_len=1.71, thru=18.1, I=23.3, T=35.6, wait=3.1s, pool=60, epoch=1/2][2026-10-19 06:58:13,367] loop.py:45 INFO TRAIN_STEP step=8028 loss=1.415627 Training: 45%|████▍ | 8028/18018 [17:27:59<21:40:58, 7.81s/step, loss=1.416, acc=0.684, acc_len=1.63, thru=14.9, I=16.9, T=29.9, wait=3.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:58:22,773] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:58:23,215] loop.py:45 INFO TRAIN_STEP step=8029 loss=1.467679 Training: 45%|████▍ | 8029/18018 [17:28:09<23:22:36, 8.42s/step, loss=1.468, acc=0.674, acc_len=1.59, thru=15.0, I=12.9, T=33.3, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:58:23,284] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:58:30,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:58:32,479] loop.py:380 INFO TIMING step=8030: step=4.395s data=2.022s compute=4.357s [fwd=3.060s bwd=1.283s opt=0.003s] dispatch=4.858s [2026-10-19 06:58:32,479] loop.py:45 INFO TRAIN_STEP step=8030 loss=1.467243 Training: 45%|████▍ | 8030/18018 [17:28:18<24:04:15, 8.68s/step, loss=1.467, acc=0.673, acc_len=1.59, thru=13.8, I=18.7, T=29.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:58:32,822] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:58:36,708] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 06:58:39,192] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8030: forward=3088.8ms backward=1202.2ms [2026-10-19 06:58:39,197] loop.py:45 INFO TRAIN_STEP step=8031 loss=1.502308 Training: 45%|████▍ | 8031/18018 [17:28:25<22:26:20, 8.09s/step, loss=1.502, acc=0.664, acc_len=1.56, thru=16.5, I=32.2, T=29.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:58:39,226] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 [2026-10-19 06:58:45,304] loop.py:45 INFO TRAIN_STEP step=8032 loss=1.462701 Training: 45%|████▍ | 8032/18018 [17:28:31<20:47:18, 7.49s/step, loss=1.463, acc=0.678, acc_len=1.61, thru=17.2, I=24.5, T=31.6, wait=2.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:58:51,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:58:52,404] loop.py:45 INFO TRAIN_STEP step=8033 loss=1.361786 Training: 45%|████▍ | 8033/18018 [17:28:38<20:27:29, 7.38s/step, loss=1.362, acc=0.689, acc_len=1.64, thru=16.9, I=18.1, T=32.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:58:52,756] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 06:59:00,049] loop.py:45 INFO TRAIN_STEP step=8034 loss=1.425950 Training: 45%|████▍ | 8034/18018 [17:28:46<20:40:43, 7.46s/step, loss=1.426, acc=0.680, acc_len=1.61, thru=15.1, I=16.7, T=29.9, wait=3.4s, pool=60, epoch=1/2][2026-10-19 06:59:07,565] loop.py:380 INFO TIMING step=8035: step=4.258s data=1.855s compute=4.226s [fwd=3.029s bwd=1.182s opt=0.003s] dispatch=3.251s [2026-10-19 06:59:07,565] loop.py:45 INFO TRAIN_STEP step=8035 loss=1.386800 Training: 45%|████▍ | 8035/18018 [17:28:53<20:43:36, 7.47s/step, loss=1.387, acc=0.686, acc_len=1.64, thru=16.4, I=16.4, T=30.1, wait=3.3s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:14,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:59:16,377] loop.py:45 INFO TRAIN_STEP step=8036 loss=1.400181 Training: 45%|████▍ | 8036/18018 [17:29:02<21:50:22, 7.88s/step, loss=1.400, acc=0.684, acc_len=1.63, thru=16.2, I=16.0, T=28.4, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:59:16,373] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8035: forward=3100.4ms backward=1330.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:16,647] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:23,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:59:24,614] loop.py:45 INFO TRAIN_STEP step=8037 loss=1.404014 Training: 45%|████▍ | 8037/18018 [17:29:10<22:08:07, 7.98s/step, loss=1.404, acc=0.686, acc_len=1.64, thru=13.2, I=16.4, T=32.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:24,998] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:30,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:59:31,954] loop.py:45 INFO TRAIN_STEP step=8038 loss=1.409065 Training: 45%|████▍ | 8038/18018 [17:29:18<21:35:56, 7.79s/step, loss=1.409, acc=0.681, acc_len=1.62, thru=17.3, I=21.9, T=33.5, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:32,165] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:36,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:59:37,535] loop.py:45 INFO TRAIN_STEP step=8039 loss=1.523060 Training: 45%|████▍ | 8039/18018 [17:29:23<19:45:30, 7.13s/step, loss=1.523, acc=0.665, acc_len=1.56, thru=20.8, I=26.7, T=35.2, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:38,011] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:44,341] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:59:44,910] loop.py:380 INFO TIMING step=8040: step=4.269s data=1.971s compute=4.243s [fwd=3.025s bwd=1.203s opt=0.003s] dispatch=3.096s [2026-10-19 06:59:44,910] loop.py:45 INFO TRAIN_STEP step=8040 loss=1.378644 Training: 45%|████▍ | 8040/18018 [17:29:30<19:57:41, 7.20s/step, loss=1.379, acc=0.689, acc_len=1.65, thru=17.4, I=20.7, T=30.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:45,348] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:53,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 06:59:55,588] loop.py:45 INFO TRAIN_STEP step=8041 loss=1.491665 Training: 45%|████▍ | 8041/18018 [17:29:41<22:51:03, 8.25s/step, loss=1.492, acc=0.671, acc_len=1.59, thru=11.8, I=15.2, T=27.6, wait=6.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 06:59:55,583] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8040: forward=3207.4ms backward=1347.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 06:59:55,769] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:00:02,000] loop.py:45 INFO TRAIN_STEP step=8042 loss=1.375610 Training: 45%|████▍ | 8042/18018 [17:29:48<21:19:18, 7.69s/step, loss=1.376, acc=0.686, acc_len=1.63, thru=17.0, I=20.6, T=34.8, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:07,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:00:08,285] loop.py:45 INFO TRAIN_STEP step=8043 loss=1.354920 Training: 45%|████▍ | 8043/18018 [17:29:54<20:08:52, 7.27s/step, loss=1.355, acc=0.691, acc_len=1.65, thru=19.6, I=20.5, T=34.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:08,764] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:15,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:00:15,719] loop.py:45 INFO TRAIN_STEP step=8044 loss=1.489860 Training: 45%|████▍ | 8044/18018 [17:30:01<20:16:58, 7.32s/step, loss=1.490, acc=0.669, acc_len=1.58, thru=17.3, I=18.4, T=33.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:16,047] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:21,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:00:21,558] loop.py:380 INFO TIMING step=8045: step=3.769s data=1.691s compute=3.728s [fwd=2.656s bwd=1.057s opt=0.003s] dispatch=2.060s [2026-10-19 07:00:21,558] loop.py:45 INFO TRAIN_STEP step=8045 loss=1.528482 Training: 45%|████▍ | 8045/18018 [17:30:07<19:02:58, 6.88s/step, loss=1.528, acc=0.664, acc_len=1.56, thru=20.6, I=21.7, T=34.0, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:21,612] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:28,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:00:29,517] loop.py:45 INFO TRAIN_STEP step=8046 loss=1.515818 Training: 45%|████▍ | 8046/18018 [17:30:15<19:56:47, 7.20s/step, loss=1.516, acc=0.665, acc_len=1.56, thru=16.3, I=18.2, T=32.3, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:00:29,513] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8045: forward=2668.6ms backward=1232.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:29,549] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:34,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:00:35,547] loop.py:45 INFO TRAIN_STEP step=8047 loss=1.352832 Training: 45%|████▍ | 8047/18018 [17:30:21<18:58:24, 6.85s/step, loss=1.353, acc=0.695, acc_len=1.66, thru=19.2, I=23.2, T=32.5, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:35,719] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:00:43,170] loop.py:45 INFO TRAIN_STEP step=8048 loss=1.359670 Training: 45%|████▍ | 8048/18018 [17:30:29<19:36:42, 7.08s/step, loss=1.360, acc=0.693, acc_len=1.66, thru=15.0, I=15.6, T=31.9, wait=3.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:50,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:00:51,033] loop.py:45 INFO TRAIN_STEP step=8049 loss=1.319217 Training: 45%|████▍ | 8049/18018 [17:30:37<20:15:33, 7.32s/step, loss=1.319, acc=0.702, acc_len=1.69, thru=17.3, I=17.6, T=30.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:51,219] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:57,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:00:59,039] loop.py:380 INFO TIMING step=8050: step=4.393s data=2.081s compute=4.354s [fwd=3.095s bwd=1.244s opt=0.003s] dispatch=3.604s [2026-10-19 07:00:59,040] loop.py:45 INFO TRAIN_STEP step=8050 loss=1.460751 Training: 45%|████▍ | 8050/18018 [17:30:45<20:49:54, 7.52s/step, loss=1.461, acc=0.674, acc_len=1.60, thru=15.6, I=20.2, T=29.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:00:59,197] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-19 07:01:05,657] loop.py:45 INFO TRAIN_STEP step=8051 loss=1.494234 Training: 45%|████▍ | 8051/18018 [17:30:51<20:04:33, 7.25s/step, loss=1.494, acc=0.666, acc_len=1.56, thru=15.2, I=19.8, T=33.1, wait=2.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:01:05,653] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8050: forward=2706.5ms backward=1122.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:11,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:01:12,498] loop.py:45 INFO TRAIN_STEP step=8052 loss=1.446632 Training: 45%|████▍ | 8052/18018 [17:30:58<19:44:00, 7.13s/step, loss=1.447, acc=0.677, acc_len=1.60, thru=19.2, I=20.0, T=34.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:12,593] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:16,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:01:18,800] loop.py:45 INFO TRAIN_STEP step=8053 loss=1.301715 Training: 45%|████▍ | 8053/18018 [17:31:04<19:02:48, 6.88s/step, loss=1.302, acc=0.702, acc_len=1.69, thru=19.6, I=27.2, T=29.4, wait=1.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:18,982] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:26,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:01:26,261] loop.py:45 INFO TRAIN_STEP step=8054 loss=1.481857 Training: 45%|████▍ | 8054/18018 [17:31:12<19:31:42, 7.06s/step, loss=1.482, acc=0.668, acc_len=1.58, thru=16.1, I=19.2, T=31.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:26,618] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:36,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:01:37,266] loop.py:380 INFO TIMING step=8055: step=5.022s data=2.310s compute=4.994s [fwd=3.515s bwd=1.464s opt=0.003s] dispatch=5.968s [2026-10-19 07:01:37,266] loop.py:45 INFO TRAIN_STEP step=8055 loss=1.395313 Training: 45%|████▍ | 8055/18018 [17:31:23<22:48:10, 8.24s/step, loss=1.395, acc=0.687, acc_len=1.64, thru=12.1, I=12.9, T=25.5, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:37,305] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 07:01:43,755] loop.py:45 INFO TRAIN_STEP step=8056 loss=1.354657 Training: 45%|████▍ | 8056/18018 [17:31:29<21:20:43, 7.71s/step, loss=1.355, acc=0.691, acc_len=1.65, thru=16.2, I=16.6, T=34.0, wait=2.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:01:43,750] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8055: forward=2638.6ms backward=1066.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:52,293] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:01:54,300] loop.py:45 INFO TRAIN_STEP step=8057 loss=1.443552 Training: 45%|████▍ | 8057/18018 [17:31:40<23:41:39, 8.56s/step, loss=1.444, acc=0.676, acc_len=1.61, thru=13.6, I=17.3, T=29.4, wait=6.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:01:54,332] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 07:02:00,549] loop.py:45 INFO TRAIN_STEP step=8058 loss=1.365380 Training: 45%|████▍ | 8058/18018 [17:31:46<21:46:10, 7.87s/step, loss=1.365, acc=0.693, acc_len=1.66, thru=15.6, I=21.5, T=32.9, wait=2.3s, pool=44, epoch=1/2][2026-10-19 07:02:08,751] loop.py:45 INFO TRAIN_STEP step=8059 loss=1.627133 Training: 45%|████▍ | 8059/18018 [17:31:54<22:02:51, 7.97s/step, loss=1.627, acc=0.647, acc_len=1.51, thru=16.5, I=15.5, T=31.2, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:02:15,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:02:15,498] loop.py:380 INFO TIMING step=8060: step=3.903s data=1.713s compute=3.868s [fwd=2.793s bwd=1.059s opt=0.003s] dispatch=2.834s [2026-10-19 07:02:15,498] loop.py:45 INFO TRAIN_STEP step=8060 loss=1.430046 Training: 45%|████▍ | 8060/18018 [17:32:01<21:01:42, 7.60s/step, loss=1.430, acc=0.682, acc_len=1.62, thru=18.7, I=18.4, T=32.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:02:15,544] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-19 07:02:23,732] loop.py:45 INFO TRAIN_STEP step=8061 loss=1.526325 Training: 45%|████▍ | 8061/18018 [17:32:09<21:33:02, 7.79s/step, loss=1.526, acc=0.664, acc_len=1.56, thru=14.1, I=15.9, T=30.7, wait=4.1s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:02:23,727] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8060: forward=2910.1ms backward=1207.7ms [2026-10-19 07:02:35,625] loop.py:45 INFO TRAIN_STEP step=8062 loss=1.450424 Training: 45%|████▍ | 8062/18018 [17:32:21<24:57:09, 9.02s/step, loss=1.450, acc=0.676, acc_len=1.60, thru=12.1, I=10.8, T=31.7, wait=7.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:02:41,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:02:43,289] loop.py:45 INFO TRAIN_STEP step=8063 loss=1.581643 Training: 45%|████▍ | 8063/18018 [17:32:29<23:49:23, 8.62s/step, loss=1.582, acc=0.653, acc_len=1.52, thru=16.1, I=19.5, T=27.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:02:43,380] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:02:48,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:02:50,784] loop.py:45 INFO TRAIN_STEP step=8064 loss=1.362571 Training: 45%|████▍ | 8064/18018 [17:32:36<22:53:24, 8.28s/step, loss=1.363, acc=0.689, acc_len=1.65, thru=14.9, I=23.4, T=28.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:02:51,195] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:02:58,484] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:02:59,003] loop.py:380 INFO TIMING step=8065: step=4.636s data=2.316s compute=4.599s [fwd=3.207s bwd=1.377s opt=0.003s] dispatch=3.574s [2026-10-19 07:02:59,004] loop.py:45 INFO TRAIN_STEP step=8065 loss=1.304888 Training: 45%|████▍ | 8065/18018 [17:32:45<22:50:24, 8.26s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=16.4, I=17.7, T=27.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:02:59,491] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:06,122] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:03:07,438] loop.py:45 INFO TRAIN_STEP step=8066 loss=1.502079 Training: 45%|████▍ | 8066/18018 [17:32:53<22:58:56, 8.31s/step, loss=1.502, acc=0.669, acc_len=1.58, thru=14.4, I=18.1, T=30.8, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:03:07,433] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8065: forward=2937.3ms backward=1171.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:07,632] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:13,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:03:14,670] loop.py:45 INFO TRAIN_STEP step=8067 loss=1.331079 Training: 45%|████▍ | 8067/18018 [17:33:00<22:04:49, 7.99s/step, loss=1.331, acc=0.699, acc_len=1.68, thru=16.4, I=18.9, T=32.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:15,144] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:03:21,094] loop.py:45 INFO TRAIN_STEP step=8068 loss=1.425675 Training: 45%|████▍ | 8068/18018 [17:33:07<20:46:53, 7.52s/step, loss=1.426, acc=0.678, acc_len=1.61, thru=17.7, I=22.7, T=33.8, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:27,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:03:28,547] loop.py:45 INFO TRAIN_STEP step=8069 loss=1.370341 Training: 45%|████▍ | 8069/18018 [17:33:14<20:43:38, 7.50s/step, loss=1.370, acc=0.695, acc_len=1.66, thru=17.2, I=17.7, T=30.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:28,868] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:33,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:03:35,074] loop.py:380 INFO TIMING step=8070: step=4.053s data=1.841s compute=4.029s [fwd=2.829s bwd=1.185s opt=0.003s] dispatch=2.463s [2026-10-19 07:03:35,074] loop.py:45 INFO TRAIN_STEP step=8070 loss=1.423198 Training: 45%|████▍ | 8070/18018 [17:33:21<19:55:04, 7.21s/step, loss=1.423, acc=0.682, acc_len=1.62, thru=18.8, I=21.2, T=31.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:35,197] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:41,168] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:03:41,801] loop.py:45 INFO TRAIN_STEP step=8071 loss=1.471463 Training: 45%|████▍ | 8071/18018 [17:33:27<19:30:59, 7.06s/step, loss=1.471, acc=0.672, acc_len=1.59, thru=17.6, I=21.6, T=33.0, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:03:41,796] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8070: forward=2708.3ms backward=1101.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:42,178] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:47,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:03:49,195] loop.py:45 INFO TRAIN_STEP step=8072 loss=1.386179 Training: 45%|████▍ | 8072/18018 [17:33:35<19:47:15, 7.16s/step, loss=1.386, acc=0.686, acc_len=1.63, thru=16.9, I=20.9, T=27.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:03:49,272] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-19 07:03:56,911] loop.py:45 INFO TRAIN_STEP step=8073 loss=1.417188 Training: 45%|████▍ | 8073/18018 [17:33:43<20:15:01, 7.33s/step, loss=1.417, acc=0.683, acc_len=1.62, thru=12.7, I=17.6, T=30.5, wait=3.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:04,265] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:04:05,885] loop.py:45 INFO TRAIN_STEP step=8074 loss=1.357910 Training: 45%|████▍ | 8074/18018 [17:33:51<21:36:17, 7.82s/step, loss=1.358, acc=0.694, acc_len=1.66, thru=15.1, I=15.1, T=29.2, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:06,278] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:10,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:04:11,936] loop.py:380 INFO TIMING step=8075: step=3.867s data=1.480s compute=3.839s [fwd=2.555s bwd=1.269s opt=0.003s] dispatch=2.174s [2026-10-19 07:04:11,936] loop.py:45 INFO TRAIN_STEP step=8075 loss=1.416449 Training: 45%|████▍ | 8075/18018 [17:33:58<20:08:04, 7.29s/step, loss=1.416, acc=0.682, acc_len=1.62, thru=18.0, I=29.0, T=33.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:12,302] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:04:20,054] loop.py:45 INFO TRAIN_STEP step=8076 loss=1.382499 Training: 45%|████▍ | 8076/18018 [17:34:06<20:49:08, 7.54s/step, loss=1.382, acc=0.688, acc_len=1.64, thru=11.9, I=14.6, T=32.7, wait=4.2s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:04:20,049] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8075: forward=2634.0ms backward=1219.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:28,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:04:28,891] loop.py:45 INFO TRAIN_STEP step=8077 loss=1.552069 Training: 45%|████▍ | 8077/18018 [17:34:14<21:53:38, 7.93s/step, loss=1.552, acc=0.660, acc_len=1.55, thru=15.8, I=15.7, T=27.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:29,037] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:36,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:04:37,838] loop.py:45 INFO TRAIN_STEP step=8078 loss=1.551097 Training: 45%|████▍ | 8078/18018 [17:34:23<22:44:07, 8.23s/step, loss=1.551, acc=0.662, acc_len=1.55, thru=14.6, I=15.3, T=28.8, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:37,933] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:45,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:04:47,014] loop.py:45 INFO TRAIN_STEP step=8079 loss=1.328998 Training: 45%|████▍ | 8079/18018 [17:34:33<23:30:46, 8.52s/step, loss=1.329, acc=0.696, acc_len=1.67, thru=14.1, I=16.8, T=30.8, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:47,471] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:53,341] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:04:53,820] loop.py:380 INFO TIMING step=8080: step=3.682s data=1.567s compute=3.637s [fwd=2.570s bwd=1.052s opt=0.003s] dispatch=3.115s [2026-10-19 07:04:53,820] loop.py:45 INFO TRAIN_STEP step=8080 loss=1.442645 Training: 45%|████▍ | 8080/18018 [17:34:39<22:05:38, 8.00s/step, loss=1.443, acc=0.677, acc_len=1.61, thru=17.3, I=21.4, T=34.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:04:53,854] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:05:00,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:05:00,719] loop.py:45 INFO TRAIN_STEP step=8081 loss=1.414793 Training: 45%|████▍ | 8081/18018 [17:34:46<21:10:38, 7.67s/step, loss=1.415, acc=0.681, acc_len=1.62, thru=18.5, I=20.3, T=36.1, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:05:00,715] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8080: forward=2541.7ms backward=955.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:05:01,132] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:05:06,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:05:07,964] loop.py:45 INFO TRAIN_STEP step=8082 loss=1.413813 Training: 45%|████▍ | 8082/18018 [17:34:54<20:49:11, 7.54s/step, loss=1.414, acc=0.685, acc_len=1.63, thru=18.2, I=21.0, T=35.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:05:08,140] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:05:12,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:05:13,902] loop.py:45 INFO TRAIN_STEP step=8083 loss=1.429375 Training: 45%|████▍ | 8083/18018 [17:34:59<19:29:18, 7.06s/step, loss=1.429, acc=0.679, acc_len=1.61, thru=21.1, I=26.3, T=34.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:05:14,368] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:05:22,068] loop.py:45 INFO TRAIN_STEP step=8084 loss=1.388503 Training: 45%|████▍ | 8084/18018 [17:35:08<20:24:02, 7.39s/step, loss=1.389, acc=0.688, acc_len=1.64, thru=13.8, I=18.1, T=29.1, wait=3.8s, pool=52, epoch=1/2][2026-10-19 07:05:29,785] loop.py:380 INFO TIMING step=8085: step=3.766s data=1.652s compute=3.732s [fwd=2.651s bwd=1.066s opt=0.003s] dispatch=3.943s [2026-10-19 07:05:29,785] loop.py:45 INFO TRAIN_STEP step=8085 loss=1.386676 Training: 45%|████▍ | 8085/18018 [17:35:15<20:40:04, 7.49s/step, loss=1.387, acc=0.687, acc_len=1.64, thru=16.6, I=17.1, T=34.0, wait=3.9s, pool=60, epoch=1/2][2026-10-19 07:05:38,737] loop.py:45 INFO TRAIN_STEP step=8086 loss=1.323013 Training: 45%|████▍ | 8086/18018 [17:35:24<21:52:34, 7.93s/step, loss=1.323, acc=0.698, acc_len=1.68, thru=13.5, I=12.8, T=32.4, wait=5.0s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:05:38,733] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8085: forward=2760.1ms backward=1149.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:05:45,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:05:47,799] loop.py:45 INFO TRAIN_STEP step=8087 loss=1.403373 Training: 45%|████▍ | 8087/18018 [17:35:33<22:48:42, 8.27s/step, loss=1.403, acc=0.687, acc_len=1.64, thru=15.1, I=17.4, T=27.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:05:47,962] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 07:05:55,139] loop.py:45 INFO TRAIN_STEP step=8088 loss=1.353649 Training: 45%|████▍ | 8088/18018 [17:35:41<22:02:16, 7.99s/step, loss=1.354, acc=0.695, acc_len=1.66, thru=13.7, I=17.3, T=34.0, wait=3.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:00,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:06:02,141] loop.py:45 INFO TRAIN_STEP step=8089 loss=1.358569 Training: 45%|████▍ | 8089/18018 [17:35:48<21:13:03, 7.69s/step, loss=1.359, acc=0.695, acc_len=1.66, thru=19.5, I=20.4, T=29.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:02,640] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:06:12,468] loop.py:380 INFO TIMING step=8090: step=4.898s data=2.393s compute=4.862s [fwd=3.423s bwd=1.424s opt=0.003s] dispatch=5.421s [2026-10-19 07:06:12,468] loop.py:45 INFO TRAIN_STEP step=8090 loss=1.342740 Training: 45%|████▍ | 8090/18018 [17:35:58<23:23:40, 8.48s/step, loss=1.343, acc=0.691, acc_len=1.66, thru=11.8, I=11.2, T=26.1, wait=5.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:18,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:06:19,972] loop.py:45 INFO TRAIN_STEP step=8091 loss=1.346643 Training: 45%|████▍ | 8091/18018 [17:36:06<22:35:05, 8.19s/step, loss=1.347, acc=0.697, acc_len=1.67, thru=16.8, I=17.3, T=32.5, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:06:19,967] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8090: forward=2784.5ms backward=1105.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:20,388] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:26,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:06:27,088] loop.py:45 INFO TRAIN_STEP step=8092 loss=1.399350 Training: 45%|████▍ | 8092/18018 [17:36:13<21:41:35, 7.87s/step, loss=1.399, acc=0.683, acc_len=1.62, thru=16.7, I=20.6, T=31.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:27,282] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:34,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:06:35,317] loop.py:45 INFO TRAIN_STEP step=8093 loss=1.385568 Training: 45%|████▍ | 8093/18018 [17:36:21<21:59:26, 7.98s/step, loss=1.386, acc=0.685, acc_len=1.63, thru=16.5, I=16.6, T=29.4, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:35,649] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:40,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:06:41,916] loop.py:45 INFO TRAIN_STEP step=8094 loss=1.346467 Training: 45%|████▍ | 8094/18018 [17:36:28<20:50:57, 7.56s/step, loss=1.346, acc=0.695, acc_len=1.66, thru=18.4, I=21.8, T=30.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:42,118] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:06:48,612] loop.py:380 INFO TIMING step=8095: step=3.891s data=1.658s compute=3.863s [fwd=2.758s bwd=1.090s opt=0.003s] dispatch=2.796s [2026-10-19 07:06:48,612] loop.py:45 INFO TRAIN_STEP step=8095 loss=1.404881 Training: 45%|████▍ | 8095/18018 [17:36:34<20:07:45, 7.30s/step, loss=1.405, acc=0.685, acc_len=1.63, thru=15.6, I=20.7, T=32.9, wait=2.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:56,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:06:57,542] loop.py:45 INFO TRAIN_STEP step=8096 loss=1.345489 Training: 45%|████▍ | 8096/18018 [17:36:43<21:28:21, 7.79s/step, loss=1.345, acc=0.693, acc_len=1.66, thru=16.1, I=16.4, T=32.0, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:06:57,538] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8095: forward=2809.6ms backward=1143.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:06:57,703] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:03,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:07:05,650] loop.py:45 INFO TRAIN_STEP step=8097 loss=1.445871 Training: 45%|████▍ | 8097/18018 [17:36:51<21:44:02, 7.89s/step, loss=1.446, acc=0.678, acc_len=1.61, thru=15.4, I=21.9, T=24.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:05,701] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:12,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:07:13,300] loop.py:45 INFO TRAIN_STEP step=8098 loss=1.432575 Training: 45%|████▍ | 8098/18018 [17:36:59<21:32:09, 7.82s/step, loss=1.433, acc=0.679, acc_len=1.61, thru=16.9, I=18.6, T=29.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:13,525] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:18,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:07:19,648] loop.py:45 INFO TRAIN_STEP step=8099 loss=1.383832 Training: 45%|████▍ | 8099/18018 [17:37:05<20:19:15, 7.38s/step, loss=1.384, acc=0.687, acc_len=1.64, thru=18.4, I=21.0, T=33.4, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:19,721] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-19 07:07:27,842] loop.py:380 INFO TIMING step=8100: step=4.453s data=1.955s compute=4.422s [fwd=3.017s bwd=1.390s opt=0.003s] dispatch=3.731s [2026-10-19 07:07:27,843] loop.py:45 INFO TRAIN_STEP step=8100 loss=1.524686 Training: 45%|████▍ | 8100/18018 [17:37:13<20:59:37, 7.62s/step, loss=1.525, acc=0.665, acc_len=1.56, thru=14.6, I=16.4, T=28.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:27,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:28,261] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:35,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:07:36,951] loop.py:45 INFO TRAIN_STEP step=8101 loss=1.388534 Training: 45%|████▍ | 8101/18018 [17:37:23<22:13:25, 8.07s/step, loss=1.389, acc=0.689, acc_len=1.65, thru=14.8, I=17.8, T=30.8, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:07:36,947] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8100: forward=2818.6ms backward=1283.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:37,243] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:43,880] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:07:44,212] loop.py:45 INFO TRAIN_STEP step=8102 loss=1.352448 Training: 45%|████▍ | 8102/18018 [17:37:30<21:33:12, 7.82s/step, loss=1.352, acc=0.691, acc_len=1.65, thru=15.2, I=19.5, T=32.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:07:44,381] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:07:54,281] loop.py:45 INFO TRAIN_STEP step=8103 loss=1.346003 Training: 45%|████▍ | 8103/18018 [17:37:40<23:24:16, 8.50s/step, loss=1.346, acc=0.695, acc_len=1.67, thru=11.3, I=11.7, T=29.4, wait=5.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:08:03,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:08:04,357] loop.py:45 INFO TRAIN_STEP step=8104 loss=1.522536 Training: 45%|████▍ | 8104/18018 [17:37:50<24:42:23, 8.97s/step, loss=1.523, acc=0.665, acc_len=1.57, thru=14.5, I=14.5, T=31.1, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:08:04,661] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:08:12,796] loop.py:380 INFO TIMING step=8105: step=4.521s data=2.129s compute=4.482s [fwd=3.230s bwd=1.238s opt=0.003s] dispatch=3.910s [2026-10-19 07:08:12,796] loop.py:45 INFO TRAIN_STEP step=8105 loss=1.477926 Training: 45%|████▍ | 8105/18018 [17:37:58<24:15:56, 8.81s/step, loss=1.478, acc=0.669, acc_len=1.58, thru=13.0, I=14.2, T=28.3, wait=3.9s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:08:21,418] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8105: forward=2915.8ms backward=1283.8ms [2026-10-19 07:08:21,422] loop.py:45 INFO TRAIN_STEP step=8106 loss=1.388381 Training: 45%|████▍ | 8106/18018 [17:38:07<24:06:37, 8.76s/step, loss=1.388, acc=0.688, acc_len=1.64, thru=15.7, I=16.0, T=30.0, wait=4.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:08:31,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:08:32,376] loop.py:45 INFO TRAIN_STEP step=8107 loss=1.366884 Training: 45%|████▍ | 8107/18018 [17:38:18<25:55:12, 9.42s/step, loss=1.367, acc=0.691, acc_len=1.65, thru=12.7, I=13.1, T=28.2, wait=6.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:08:32,855] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:08:39,668] loop.py:45 INFO TRAIN_STEP step=8108 loss=1.340420 Training: 45%|████▍ | 8108/18018 [17:38:25<24:09:51, 8.78s/step, loss=1.340, acc=0.695, acc_len=1.67, thru=16.5, I=20.0, T=31.2, wait=3.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:08:47,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:08:47,386] loop.py:45 INFO TRAIN_STEP step=8109 loss=1.406345 Training: 45%|████▌ | 8109/18018 [17:38:33<23:17:25, 8.46s/step, loss=1.406, acc=0.685, acc_len=1.63, thru=16.8, I=15.9, T=31.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:08:47,539] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:08:51,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:08:53,411] loop.py:380 INFO TIMING step=8110: step=3.626s data=1.500s compute=3.595s [fwd=2.490s bwd=1.091s opt=0.003s] dispatch=2.387s [2026-10-19 07:08:53,412] loop.py:45 INFO TRAIN_STEP step=8110 loss=1.377969 Training: 45%|████▌ | 8110/18018 [17:38:39<21:16:23, 7.73s/step, loss=1.378, acc=0.686, acc_len=1.63, thru=20.0, I=22.3, T=35.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:08:53,514] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:00,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:09:01,519] loop.py:45 INFO TRAIN_STEP step=8111 loss=1.467150 Training: 45%|████▌ | 8111/18018 [17:38:47<21:34:55, 7.84s/step, loss=1.467, acc=0.671, acc_len=1.59, thru=14.4, I=18.2, T=30.6, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:09:01,514] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8110: forward=2944.5ms backward=1192.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:01,574] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-19 07:09:09,420] loop.py:45 INFO TRAIN_STEP step=8112 loss=1.383682 Training: 45%|████▌ | 8112/18018 [17:38:55<21:37:50, 7.86s/step, loss=1.384, acc=0.686, acc_len=1.64, thru=13.0, I=17.7, T=31.3, wait=3.8s, pool=44, epoch=1/2][2026-10-19 07:09:20,203] loop.py:45 INFO TRAIN_STEP step=8113 loss=1.381296 Training: 45%|████▌ | 8113/18018 [17:39:06<24:02:20, 8.74s/step, loss=1.381, acc=0.687, acc_len=1.64, thru=13.8, I=11.2, T=27.2, wait=6.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:28,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:09:28,534] loop.py:45 INFO TRAIN_STEP step=8114 loss=1.399344 Training: 45%|████▌ | 8114/18018 [17:39:14<23:42:09, 8.62s/step, loss=1.399, acc=0.686, acc_len=1.64, thru=17.0, I=16.9, T=31.9, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:28,605] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:34,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:09:35,224] loop.py:380 INFO TIMING step=8115: step=4.444s data=2.048s compute=4.407s [fwd=3.114s bwd=1.279s opt=0.003s] dispatch=2.237s [2026-10-19 07:09:35,224] loop.py:45 INFO TRAIN_STEP step=8115 loss=1.439877 Training: 45%|████▌ | 8115/18018 [17:39:21<22:06:39, 8.04s/step, loss=1.440, acc=0.677, acc_len=1.61, thru=19.1, I=20.5, T=28.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:35,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:41,999] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:09:44,130] loop.py:45 INFO TRAIN_STEP step=8116 loss=1.568080 Training: 45%|████▌ | 8116/18018 [17:39:30<22:49:30, 8.30s/step, loss=1.568, acc=0.658, acc_len=1.54, thru=13.7, I=17.1, T=28.8, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:09:44,125] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8115: forward=3088.2ms backward=1280.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:44,514] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:49,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:09:51,345] loop.py:45 INFO TRAIN_STEP step=8117 loss=1.346012 Training: 45%|████▌ | 8117/18018 [17:39:37<21:55:45, 7.97s/step, loss=1.346, acc=0.693, acc_len=1.67, thru=15.2, I=24.4, T=27.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:09:51,699] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:09:58,153] loop.py:45 INFO TRAIN_STEP step=8118 loss=1.425880 Training: 45%|████▌ | 8118/18018 [17:39:44<20:57:47, 7.62s/step, loss=1.426, acc=0.678, acc_len=1.61, thru=14.8, I=19.8, T=29.5, wait=2.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:03,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:10:04,651] loop.py:45 INFO TRAIN_STEP step=8119 loss=1.427004 Training: 45%|████▌ | 8119/18018 [17:39:50<20:02:04, 7.29s/step, loss=1.427, acc=0.678, acc_len=1.60, thru=18.8, I=19.6, T=34.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:04,754] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:11,753] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:10:13,942] loop.py:380 INFO TIMING step=8120: step=4.879s data=2.376s compute=4.850s [fwd=3.503s bwd=1.332s opt=0.003s] dispatch=4.403s [2026-10-19 07:10:13,943] loop.py:45 INFO TRAIN_STEP step=8120 loss=1.366219 Training: 45%|████▌ | 8120/18018 [17:40:00<21:41:06, 7.89s/step, loss=1.366, acc=0.688, acc_len=1.64, thru=13.9, I=17.4, T=26.2, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:14,270] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:21,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:10:22,276] loop.py:45 INFO TRAIN_STEP step=8121 loss=1.441743 Training: 45%|████▌ | 8121/18018 [17:40:08<22:03:12, 8.02s/step, loss=1.442, acc=0.677, acc_len=1.61, thru=16.0, I=16.0, T=26.5, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:10:22,271] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8120: forward=3672.1ms backward=1098.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:22,466] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:28,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:10:29,580] loop.py:45 INFO TRAIN_STEP step=8122 loss=1.383271 Training: 45%|████▌ | 8122/18018 [17:40:15<21:27:33, 7.81s/step, loss=1.383, acc=0.691, acc_len=1.66, thru=17.2, I=17.2, T=29.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:29,773] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:10:37,692] loop.py:45 INFO TRAIN_STEP step=8123 loss=1.431714 Training: 45%|████▌ | 8123/18018 [17:40:23<21:42:23, 7.90s/step, loss=1.432, acc=0.683, acc_len=1.62, thru=13.6, I=13.2, T=28.8, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:43,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:10:45,302] loop.py:45 INFO TRAIN_STEP step=8124 loss=1.431266 Training: 45%|████▌ | 8124/18018 [17:40:31<21:28:10, 7.81s/step, loss=1.431, acc=0.679, acc_len=1.61, thru=17.6, I=20.0, T=29.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:45,790] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:52,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:10:53,040] loop.py:380 INFO TIMING step=8125: step=4.741s data=2.243s compute=4.715s [fwd=3.329s bwd=1.371s opt=0.003s] dispatch=2.988s [2026-10-19 07:10:53,041] loop.py:45 INFO TRAIN_STEP step=8125 loss=1.337071 Training: 45%|████▌ | 8125/18018 [17:40:39<21:24:24, 7.79s/step, loss=1.337, acc=0.699, acc_len=1.67, thru=14.5, I=17.1, T=27.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:10:53,093] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:00,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:11:00,975] loop.py:45 INFO TRAIN_STEP step=8126 loss=1.476741 Training: 45%|████▌ | 8126/18018 [17:40:47<21:31:19, 7.83s/step, loss=1.477, acc=0.671, acc_len=1.59, thru=14.1, I=16.4, T=27.8, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:11:00,970] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8125: forward=3176.3ms backward=1377.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:01,359] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:07,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:11:08,947] loop.py:45 INFO TRAIN_STEP step=8127 loss=1.477527 Training: 45%|████▌ | 8127/18018 [17:40:55<21:38:10, 7.87s/step, loss=1.478, acc=0.672, acc_len=1.59, thru=16.0, I=18.6, T=30.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:09,094] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-19 07:11:15,588] loop.py:45 INFO TRAIN_STEP step=8128 loss=1.279266 Training: 45%|████▌ | 8128/18018 [17:41:01<20:37:05, 7.51s/step, loss=1.279, acc=0.703, acc_len=1.70, thru=14.2, I=19.3, T=35.4, wait=3.0s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:22,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:11:24,129] loop.py:45 INFO TRAIN_STEP step=8129 loss=1.360332 Training: 45%|████▌ | 8129/18018 [17:41:10<21:28:09, 7.82s/step, loss=1.360, acc=0.692, acc_len=1.66, thru=17.9, I=19.1, T=31.1, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:24,638] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:11:31,752] loop.py:380 INFO TIMING step=8130: step=3.720s data=1.648s compute=3.684s [fwd=2.615s bwd=1.054s opt=0.003s] dispatch=3.894s [2026-10-19 07:11:31,753] loop.py:45 INFO TRAIN_STEP step=8130 loss=1.258807 Training: 45%|████▌ | 8130/18018 [17:41:17<21:18:23, 7.76s/step, loss=1.259, acc=0.708, acc_len=1.71, thru=15.4, I=18.4, T=34.4, wait=3.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:42,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:11:44,178] loop.py:45 INFO TRAIN_STEP step=8131 loss=1.275941 Training: 45%|████▌ | 8131/18018 [17:41:30<25:09:06, 9.16s/step, loss=1.276, acc=0.711, acc_len=1.72, thru=11.3, I=13.3, T=27.9, wait=7.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:11:44,173] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8130: forward=3291.3ms backward=1246.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:44,568] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:49,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:11:51,472] loop.py:45 INFO TRAIN_STEP step=8132 loss=1.402351 Training: 45%|████▌ | 8132/18018 [17:41:37<23:36:46, 8.60s/step, loss=1.402, acc=0.685, acc_len=1.63, thru=14.5, I=25.1, T=26.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:11:51,946] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 07:11:59,226] loop.py:45 INFO TRAIN_STEP step=8133 loss=1.491143 Training: 45%|████▌ | 8133/18018 [17:41:45<22:54:56, 8.35s/step, loss=1.491, acc=0.668, acc_len=1.57, thru=13.1, I=16.5, T=26.9, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:08,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:12:08,189] loop.py:45 INFO TRAIN_STEP step=8134 loss=1.571497 Training: 45%|████▌ | 8134/18018 [17:41:54<23:25:16, 8.53s/step, loss=1.571, acc=0.652, acc_len=1.52, thru=15.2, I=14.3, T=29.1, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:08,584] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:15,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:12:17,055] loop.py:380 INFO TIMING step=8135: step=4.114s data=1.748s compute=4.065s [fwd=2.780s bwd=1.270s opt=0.003s] dispatch=4.743s [2026-10-19 07:12:17,055] loop.py:45 INFO TRAIN_STEP step=8135 loss=1.359027 Training: 45%|████▌ | 8135/18018 [17:42:03<23:41:48, 8.63s/step, loss=1.359, acc=0.695, acc_len=1.67, thru=14.9, I=16.3, T=31.1, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:17,240] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:25,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:12:26,340] loop.py:45 INFO TRAIN_STEP step=8136 loss=1.519632 Training: 45%|████▌ | 8136/18018 [17:42:12<24:13:48, 8.83s/step, loss=1.520, acc=0.665, acc_len=1.57, thru=14.1, I=15.1, T=29.3, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:12:26,336] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8135: forward=3024.0ms backward=1284.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:26,683] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:32,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:12:34,071] loop.py:45 INFO TRAIN_STEP step=8137 loss=1.419643 Training: 45%|████▌ | 8137/18018 [17:42:20<23:19:30, 8.50s/step, loss=1.420, acc=0.679, acc_len=1.61, thru=17.4, I=21.8, T=28.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:34,099] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:39,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:12:42,327] loop.py:45 INFO TRAIN_STEP step=8138 loss=1.528047 Training: 45%|████▌ | 8138/18018 [17:42:28<23:07:28, 8.43s/step, loss=1.528, acc=0.666, acc_len=1.57, thru=15.6, I=20.9, T=26.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:42,435] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:47,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:12:48,891] loop.py:45 INFO TRAIN_STEP step=8139 loss=1.301064 Training: 45%|████▌ | 8139/18018 [17:42:34<21:35:25, 7.87s/step, loss=1.301, acc=0.702, acc_len=1.70, thru=15.7, I=22.2, T=32.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:12:49,398] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:12:56,231] loop.py:380 INFO TIMING step=8140: step=4.524s data=2.133s compute=4.479s [fwd=3.064s bwd=1.400s opt=0.003s] dispatch=2.799s [2026-10-19 07:12:56,231] loop.py:45 INFO TRAIN_STEP step=8140 loss=1.235741 Training: 45%|████▌ | 8140/18018 [17:42:42<21:09:17, 7.71s/step, loss=1.236, acc=0.713, acc_len=1.73, thru=16.2, I=19.0, T=28.3, wait=2.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:03,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:13:04,046] loop.py:45 INFO TRAIN_STEP step=8141 loss=1.320013 Training: 45%|████▌ | 8141/18018 [17:42:50<21:14:12, 7.74s/step, loss=1.320, acc=0.698, acc_len=1.67, thru=16.4, I=17.4, T=31.0, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:13:04,042] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8140: forward=2928.9ms backward=1152.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:04,480] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:10,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:13:12,526] loop.py:45 INFO TRAIN_STEP step=8142 loss=1.436834 Training: 45%|████▌ | 8142/18018 [17:42:58<21:50:35, 7.96s/step, loss=1.437, acc=0.680, acc_len=1.61, thru=15.1, I=20.1, T=27.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:12,753] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:20,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:13:21,126] loop.py:45 INFO TRAIN_STEP step=8143 loss=1.472856 Training: 45%|████▌ | 8143/18018 [17:43:07<22:22:06, 8.15s/step, loss=1.473, acc=0.677, acc_len=1.60, thru=15.3, I=16.1, T=25.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:21,633] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:26,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:13:28,665] loop.py:45 INFO TRAIN_STEP step=8144 loss=1.386444 Training: 45%|████▌ | 8144/18018 [17:43:14<21:51:24, 7.97s/step, loss=1.386, acc=0.689, acc_len=1.64, thru=15.6, I=19.4, T=27.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:28,992] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:33,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:13:34,882] loop.py:380 INFO TIMING step=8145: step=3.632s data=1.634s compute=3.603s [fwd=2.595s bwd=0.993s opt=0.003s] dispatch=2.576s [2026-10-19 07:13:34,882] loop.py:45 INFO TRAIN_STEP step=8145 loss=1.379035 Training: 45%|████▌ | 8145/18018 [17:43:20<20:24:47, 7.44s/step, loss=1.379, acc=0.686, acc_len=1.63, thru=17.4, I=26.1, T=35.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:34,991] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:42,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:13:44,499] loop.py:45 INFO TRAIN_STEP step=8146 loss=1.485786 Training: 45%|████▌ | 8146/18018 [17:43:30<22:11:57, 8.10s/step, loss=1.486, acc=0.668, acc_len=1.57, thru=13.5, I=16.5, T=27.9, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:13:44,494] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8145: forward=3257.6ms backward=1280.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:13:44,753] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:13:53,240] loop.py:45 INFO TRAIN_STEP step=8147 loss=1.537110 Training: 45%|████▌ | 8147/18018 [17:43:39<22:43:45, 8.29s/step, loss=1.537, acc=0.661, acc_len=1.54, thru=13.2, I=14.5, T=23.0, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:01,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:14:03,473] loop.py:45 INFO TRAIN_STEP step=8148 loss=1.497378 Training: 45%|████▌ | 8148/18018 [17:43:49<24:19:33, 8.87s/step, loss=1.497, acc=0.668, acc_len=1.58, thru=14.6, I=15.2, T=25.9, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:03,961] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:08,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:14:10,491] loop.py:45 INFO TRAIN_STEP step=8149 loss=1.402522 Training: 45%|████▌ | 8149/18018 [17:43:56<22:47:49, 8.32s/step, loss=1.403, acc=0.684, acc_len=1.62, thru=15.3, I=23.5, T=29.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:10,711] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:15,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:14:17,936] loop.py:380 INFO TIMING step=8150: step=4.208s data=1.933s compute=4.172s [fwd=2.937s bwd=1.219s opt=0.003s] dispatch=3.227s [2026-10-19 07:14:17,936] loop.py:45 INFO TRAIN_STEP step=8150 loss=1.327135 Training: 45%|████▌ | 8150/18018 [17:44:04<22:04:46, 8.06s/step, loss=1.327, acc=0.695, acc_len=1.67, thru=14.7, I=25.2, T=30.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:18,192] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:24,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:14:26,537] loop.py:45 INFO TRAIN_STEP step=8151 loss=1.396572 Training: 45%|████▌ | 8151/18018 [17:44:12<22:31:31, 8.22s/step, loss=1.397, acc=0.688, acc_len=1.65, thru=15.3, I=18.9, T=26.4, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:14:26,532] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8150: forward=3292.7ms backward=1513.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:26,904] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:34,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:14:35,869] loop.py:45 INFO TRAIN_STEP step=8152 loss=1.308753 Training: 45%|████▌ | 8152/18018 [17:44:21<23:26:18, 8.55s/step, loss=1.309, acc=0.704, acc_len=1.69, thru=14.3, I=17.7, T=27.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:36,076] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:42,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:14:44,557] loop.py:45 INFO TRAIN_STEP step=8153 loss=1.429074 Training: 45%|████▌ | 8153/18018 [17:44:30<23:32:50, 8.59s/step, loss=1.429, acc=0.683, acc_len=1.62, thru=15.1, I=18.8, T=27.5, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:14:45,044] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 07:14:52,112] loop.py:45 INFO TRAIN_STEP step=8154 loss=1.460962 Training: 45%|████▌ | 8154/18018 [17:44:38<22:41:29, 8.28s/step, loss=1.461, acc=0.674, acc_len=1.60, thru=13.0, I=18.5, T=31.8, wait=3.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:00,132] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:15:00,542] loop.py:380 INFO TIMING step=8155: step=4.606s data=2.089s compute=4.574s [fwd=3.216s bwd=1.343s opt=0.003s] dispatch=3.814s [2026-10-19 07:15:00,542] loop.py:45 INFO TRAIN_STEP step=8155 loss=1.437483 Training: 45%|████▌ | 8155/18018 [17:44:46<22:48:40, 8.33s/step, loss=1.437, acc=0.679, acc_len=1.60, thru=16.2, I=15.8, T=27.8, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:00,649] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:05,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:15:08,350] loop.py:45 INFO TRAIN_STEP step=8156 loss=1.395412 Training: 45%|████▌ | 8156/18018 [17:44:54<22:22:58, 8.17s/step, loss=1.395, acc=0.683, acc_len=1.63, thru=16.3, I=20.7, T=27.7, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:15:08,345] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8155: forward=3233.9ms backward=1338.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:08,790] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 07:15:16,528] loop.py:45 INFO TRAIN_STEP step=8157 loss=1.372694 Training: 45%|████▌ | 8157/18018 [17:45:02<22:23:17, 8.17s/step, loss=1.373, acc=0.687, acc_len=1.64, thru=16.0, I=17.4, T=32.1, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:22,947] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:15:24,028] loop.py:45 INFO TRAIN_STEP step=8158 loss=1.404095 Training: 45%|████▌ | 8158/18018 [17:45:10<21:49:51, 7.97s/step, loss=1.404, acc=0.682, acc_len=1.62, thru=17.4, I=19.0, T=26.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:24,455] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:30,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:15:30,560] loop.py:45 INFO TRAIN_STEP step=8159 loss=1.401217 Training: 45%|████▌ | 8159/18018 [17:45:16<20:38:46, 7.54s/step, loss=1.401, acc=0.685, acc_len=1.63, thru=17.1, I=20.1, T=31.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:31,054] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:37,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:15:38,146] loop.py:380 INFO TIMING step=8160: step=4.179s data=1.887s compute=4.145s [fwd=2.912s bwd=1.218s opt=0.003s] dispatch=3.399s [2026-10-19 07:15:38,146] loop.py:45 INFO TRAIN_STEP step=8160 loss=1.396629 Training: 45%|████▌ | 8160/18018 [17:45:24<20:41:05, 7.55s/step, loss=1.397, acc=0.689, acc_len=1.64, thru=17.2, I=18.4, T=30.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:38,259] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 07:15:45,044] loop.py:45 INFO TRAIN_STEP step=8161 loss=1.400536 Training: 45%|████▌ | 8161/18018 [17:45:31<20:08:37, 7.36s/step, loss=1.401, acc=0.685, acc_len=1.63, thru=16.7, I=17.9, T=26.6, wait=2.1s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:15:45,040] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8160: forward=3343.4ms backward=1421.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:54,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:15:56,150] loop.py:45 INFO TRAIN_STEP step=8162 loss=1.481162 Training: 45%|████▌ | 8162/18018 [17:45:42<23:13:15, 8.48s/step, loss=1.481, acc=0.673, acc_len=1.59, thru=13.4, I=14.8, T=22.5, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:15:56,206] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:16:00,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:16:03,345] loop.py:45 INFO TRAIN_STEP step=8163 loss=1.444978 Training: 45%|████▌ | 8163/18018 [17:45:49<22:09:39, 8.10s/step, loss=1.445, acc=0.673, acc_len=1.59, thru=15.0, I=27.5, T=27.8, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:16:03,563] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:16:08,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:16:09,993] loop.py:45 INFO TRAIN_STEP step=8164 loss=1.320298 Training: 45%|████▌ | 8164/18018 [17:45:56<20:58:18, 7.66s/step, loss=1.320, acc=0.696, acc_len=1.67, thru=14.5, I=24.5, T=31.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:16:10,205] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:16:18,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:16:18,531] loop.py:380 INFO TIMING step=8165: step=4.695s data=2.210s compute=4.663s [fwd=3.278s bwd=1.370s opt=0.003s] dispatch=3.833s [2026-10-19 07:16:18,532] loop.py:45 INFO TRAIN_STEP step=8165 loss=1.420595 Training: 45%|████▌ | 8165/18018 [17:46:04<21:41:15, 7.92s/step, loss=1.421, acc=0.682, acc_len=1.63, thru=13.4, I=16.0, T=27.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:16:19,012] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:16:26,088] loop.py:45 INFO TRAIN_STEP step=8166 loss=1.420898 Training: 45%|████▌ | 8166/18018 [17:46:12<21:23:08, 7.81s/step, loss=1.421, acc=0.679, acc_len=1.62, thru=14.8, I=18.1, T=29.3, wait=3.2s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:16:26,083] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8165: forward=2981.1ms backward=1346.9ms [2026-10-19 07:16:34,406] loop.py:45 INFO TRAIN_STEP step=8167 loss=1.406617 Training: 45%|████▌ | 8167/18018 [17:46:20<21:47:39, 7.96s/step, loss=1.407, acc=0.679, acc_len=1.62, thru=15.0, I=15.8, T=26.9, wait=3.6s, pool=56, epoch=1/2][2026-10-19 07:16:44,032] loop.py:45 INFO TRAIN_STEP step=8168 loss=1.467332 Training: 45%|████▌ | 8168/18018 [17:46:30<23:09:31, 8.46s/step, loss=1.467, acc=0.678, acc_len=1.61, thru=13.7, I=13.8, T=26.8, wait=4.8s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:16:51,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:16:53,894] loop.py:45 INFO TRAIN_STEP step=8169 loss=1.416125 Training: 45%|████▌ | 8169/18018 [17:46:39<24:18:09, 8.88s/step, loss=1.416, acc=0.679, acc_len=1.61, thru=13.4, I=14.6, T=25.0, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:16:54,161] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:16:59,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:17:00,882] loop.py:380 INFO TIMING step=8170: step=4.408s data=2.035s compute=4.371s [fwd=3.030s bwd=1.325s opt=0.003s] dispatch=2.569s [2026-10-19 07:17:00,882] loop.py:45 INFO TRAIN_STEP step=8170 loss=1.335127 Training: 45%|████▌ | 8170/18018 [17:46:46<22:44:43, 8.31s/step, loss=1.335, acc=0.695, acc_len=1.66, thru=15.2, I=23.9, T=29.0, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:01,351] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:06,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:17:08,813] loop.py:45 INFO TRAIN_STEP step=8171 loss=1.436037 (TrainerActor pid=9310) [2026-10-19 07:17:08,808] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8170: forward=3408.5ms backward=1262.9ms Training: 45%|████▌ | 8171/18018 [17:46:54<22:25:39, 8.20s/step, loss=1.436, acc=0.675, acc_len=1.60, thru=13.8, I=21.7, T=27.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:09,283] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:15,183] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:17:15,634] loop.py:45 INFO TRAIN_STEP step=8172 loss=1.273836 Training: 45%|████▌ | 8172/18018 [17:47:01<21:17:40, 7.79s/step, loss=1.274, acc=0.710, acc_len=1.72, thru=16.2, I=22.6, T=29.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:15,694] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:21,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:17:23,016] loop.py:45 INFO TRAIN_STEP step=8173 loss=1.284052 Training: 45%|████▌ | 8173/18018 [17:47:09<20:57:39, 7.66s/step, loss=1.284, acc=0.705, acc_len=1.71, thru=16.9, I=18.1, T=28.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:23,051] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:28,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:17:30,524] loop.py:45 INFO TRAIN_STEP step=8174 loss=1.402940 Training: 45%|████▌ | 8174/18018 [17:47:16<20:49:51, 7.62s/step, loss=1.403, acc=0.684, acc_len=1.63, thru=15.4, I=21.4, T=25.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:30,686] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:36,027] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:17:36,934] loop.py:380 INFO TIMING step=8175: step=3.927s data=1.731s compute=3.894s [fwd=2.762s bwd=1.116s opt=0.003s] dispatch=2.471s [2026-10-19 07:17:36,934] loop.py:45 INFO TRAIN_STEP step=8175 loss=1.404590 Training: 45%|████▌ | 8175/18018 [17:47:23<19:50:14, 7.26s/step, loss=1.405, acc=0.679, acc_len=1.62, thru=15.6, I=22.6, T=32.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:37,036] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:44,192] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:17:44,384] loop.py:45 INFO TRAIN_STEP step=8176 loss=1.372472 Training: 45%|████▌ | 8176/18018 [17:47:30<19:59:33, 7.31s/step, loss=1.372, acc=0.687, acc_len=1.64, thru=15.6, I=17.0, T=27.3, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:17:44,380] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8175: forward=3288.1ms backward=1355.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:44,696] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:17:51,284] loop.py:45 INFO TRAIN_STEP step=8177 loss=1.381424 Training: 45%|████▌ | 8177/18018 [17:47:37<19:39:19, 7.19s/step, loss=1.381, acc=0.690, acc_len=1.64, thru=17.8, I=16.6, T=28.4, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:57,459] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:17:59,186] loop.py:45 INFO TRAIN_STEP step=8178 loss=1.411093 Training: 45%|████▌ | 8178/18018 [17:47:45<20:14:02, 7.40s/step, loss=1.411, acc=0.682, acc_len=1.62, thru=16.1, I=19.0, T=30.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:17:59,469] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:05,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:18:07,202] loop.py:45 INFO TRAIN_STEP step=8179 loss=1.430129 Training: 45%|████▌ | 8179/18018 [17:47:53<20:44:12, 7.59s/step, loss=1.430, acc=0.681, acc_len=1.62, thru=13.7, I=19.1, T=26.5, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:07,330] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:13,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:18:14,797] loop.py:380 INFO TIMING step=8180: step=3.872s data=1.734s compute=3.818s [fwd=2.635s bwd=1.168s opt=0.003s] dispatch=3.712s [2026-10-19 07:18:14,797] loop.py:45 INFO TRAIN_STEP step=8180 loss=1.500855 Training: 45%|████▌ | 8180/18018 [17:48:00<20:44:22, 7.59s/step, loss=1.501, acc=0.668, acc_len=1.57, thru=15.6, I=19.6, T=33.1, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:15,192] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:23,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:18:23,630] loop.py:45 INFO TRAIN_STEP step=8181 loss=1.445779 Training: 45%|████▌ | 8181/18018 [17:48:09<21:45:33, 7.96s/step, loss=1.446, acc=0.675, acc_len=1.60, thru=15.2, I=16.0, T=32.4, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:18:23,624] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8180: forward=2811.2ms backward=1084.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:23,965] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:18:31,588] loop.py:45 INFO TRAIN_STEP step=8182 loss=1.567290 Training: 45%|████▌ | 8182/18018 [17:48:17<21:45:09, 7.96s/step, loss=1.567, acc=0.658, acc_len=1.54, thru=13.7, I=13.9, T=30.1, wait=3.7s, pool=44, epoch=1/2][2026-10-19 07:18:40,405] loop.py:45 INFO TRAIN_STEP step=8183 loss=1.402694 Training: 45%|████▌ | 8183/18018 [17:48:26<22:27:02, 8.22s/step, loss=1.403, acc=0.680, acc_len=1.62, thru=16.6, I=15.5, T=29.5, wait=4.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:46,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:18:48,412] loop.py:45 INFO TRAIN_STEP step=8184 loss=1.474710 Training: 45%|████▌ | 8184/18018 [17:48:34<22:16:33, 8.15s/step, loss=1.475, acc=0.675, acc_len=1.60, thru=15.6, I=18.7, T=27.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:48,806] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:55,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:18:56,208] loop.py:380 INFO TIMING step=8185: step=4.312s data=2.051s compute=4.285s [fwd=3.009s bwd=1.263s opt=0.003s] dispatch=3.473s [2026-10-19 07:18:56,208] loop.py:45 INFO TRAIN_STEP step=8185 loss=1.510040 Training: 45%|████▌ | 8185/18018 [17:48:42<21:58:45, 8.05s/step, loss=1.510, acc=0.669, acc_len=1.58, thru=13.3, I=19.9, T=29.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:18:56,468] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:19:02,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:19:02,938] loop.py:45 INFO TRAIN_STEP step=8186 loss=1.321001 Training: 45%|████▌ | 8186/18018 [17:48:49<20:53:53, 7.65s/step, loss=1.321, acc=0.698, acc_len=1.68, thru=18.2, I=19.2, T=34.2, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:19:02,934] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8185: forward=2585.9ms backward=1101.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:19:03,131] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:19:11,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:19:12,132] loop.py:45 INFO TRAIN_STEP step=8187 loss=1.343928 Training: 45%|████▌ | 8187/18018 [17:48:58<22:09:30, 8.11s/step, loss=1.344, acc=0.694, acc_len=1.67, thru=14.3, I=15.5, T=30.1, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:19:12,401] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:19:20,272] loop.py:45 INFO TRAIN_STEP step=8188 loss=1.424443 Training: 45%|████▌ | 8188/18018 [17:49:06<22:10:38, 8.12s/step, loss=1.424, acc=0.679, acc_len=1.61, thru=13.4, I=14.8, T=29.1, wait=3.7s, pool=52, epoch=1/2][2026-10-19 07:19:31,204] loop.py:45 INFO TRAIN_STEP step=8189 loss=1.331436 Training: 45%|████▌ | 8189/18018 [17:49:17<24:28:42, 8.97s/step, loss=1.331, acc=0.695, acc_len=1.67, thru=13.3, I=11.5, T=30.0, wait=6.7s, pool=56, epoch=1/2][2026-10-19 07:19:39,496] loop.py:380 INFO TIMING step=8190: step=4.407s data=1.980s compute=4.376s [fwd=3.098s bwd=1.264s opt=0.003s] dispatch=3.872s [2026-10-19 07:19:39,497] loop.py:45 INFO TRAIN_STEP step=8190 loss=1.383220 Training: 45%|████▌ | 8190/18018 [17:49:25<23:55:35, 8.76s/step, loss=1.383, acc=0.686, acc_len=1.63, thru=14.9, I=15.0, T=29.0, wait=3.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:19:47,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:19:49,241] loop.py:45 INFO TRAIN_STEP step=8191 loss=1.500880 Training: 45%|████▌ | 8191/18018 [17:49:35<24:43:23, 9.06s/step, loss=1.501, acc=0.668, acc_len=1.57, thru=14.8, I=16.4, T=29.4, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:19:49,236] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8190: forward=3072.7ms backward=1234.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:19:49,711] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:19:54,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:19:56,185] loop.py:45 INFO TRAIN_STEP step=8192 loss=1.364374 Training: 45%|████▌ | 8192/18018 [17:49:42<22:59:30, 8.42s/step, loss=1.364, acc=0.686, acc_len=1.64, thru=15.7, I=23.9, T=32.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:19:56,473] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:02,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:20:04,150] loop.py:45 INFO TRAIN_STEP step=8193 loss=1.370832 Training: 45%|████▌ | 8193/18018 [17:49:50<22:36:46, 8.29s/step, loss=1.371, acc=0.684, acc_len=1.63, thru=14.1, I=17.1, T=30.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:04,250] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:08,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:20:10,420] loop.py:45 INFO TRAIN_STEP step=8194 loss=1.358326 Training: 45%|████▌ | 8194/18018 [17:49:56<20:57:37, 7.68s/step, loss=1.358, acc=0.692, acc_len=1.65, thru=18.3, I=25.6, T=28.9, wait=1.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:10,565] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:16,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:20:18,581] loop.py:380 INFO TIMING step=8195: step=5.333s data=2.697s compute=5.304s [fwd=3.618s bwd=1.673s opt=0.003s] dispatch=2.819s [2026-10-19 07:20:18,582] loop.py:45 INFO TRAIN_STEP step=8195 loss=1.334314 Training: 45%|████▌ | 8195/18018 [17:50:04<21:21:09, 7.83s/step, loss=1.334, acc=0.691, acc_len=1.65, thru=16.0, I=18.0, T=24.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:18,862] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:20:24,883] loop.py:45 INFO TRAIN_STEP step=8196 loss=1.380498 Training: 45%|████▌ | 8196/18018 [17:50:10<20:06:02, 7.37s/step, loss=1.380, acc=0.688, acc_len=1.64, thru=16.6, I=19.8, T=32.1, wait=2.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:20:24,879] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8195: forward=2788.8ms backward=1157.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:31,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:20:34,126] loop.py:45 INFO TRAIN_STEP step=8197 loss=1.354217 Training: 45%|████▌ | 8197/18018 [17:50:20<21:38:03, 7.93s/step, loss=1.354, acc=0.697, acc_len=1.67, thru=15.2, I=17.6, T=29.7, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:34,301] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:43,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:20:43,414] loop.py:45 INFO TRAIN_STEP step=8198 loss=1.369678 Training: 45%|████▌ | 8198/18018 [17:50:29<22:44:35, 8.34s/step, loss=1.370, acc=0.692, acc_len=1.66, thru=14.1, I=14.9, T=31.2, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:43,874] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:48,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:20:50,828] loop.py:45 INFO TRAIN_STEP step=8199 loss=1.425906 Training: 46%|████▌ | 8199/18018 [17:50:36<21:59:13, 8.06s/step, loss=1.426, acc=0.677, acc_len=1.59, thru=16.9, I=20.5, T=26.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:50,984] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:20:58,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:21:02,014] loop.py:380 INFO TIMING step=8200: step=5.349s data=2.599s compute=5.315s [fwd=3.778s bwd=1.524s opt=0.003s] dispatch=5.826s [2026-10-19 07:21:02,014] loop.py:45 INFO TRAIN_STEP step=8200 loss=1.261487 Training: 46%|████▌ | 8200/18018 [17:50:48<24:32:27, 9.00s/step, loss=1.261, acc=0.706, acc_len=1.70, thru=11.9, I=15.7, T=23.9, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:02,161] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:07,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:21:10,198] loop.py:45 INFO TRAIN_STEP step=8201 loss=1.393903 Training: 46%|████▌ | 8201/18018 [17:50:56<23:52:20, 8.75s/step, loss=1.394, acc=0.687, acc_len=1.64, thru=15.9, I=19.4, T=25.1, wait=3.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:21:10,194] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8200: forward=3418.9ms backward=1618.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:10,469] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:15,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:21:17,243] loop.py:45 INFO TRAIN_STEP step=8202 loss=1.388895 Training: 46%|████▌ | 8202/18018 [17:51:03<22:28:09, 8.24s/step, loss=1.389, acc=0.682, acc_len=1.62, thru=14.1, I=18.9, T=29.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:17,715] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:23,114] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:21:24,460] loop.py:45 INFO TRAIN_STEP step=8203 loss=1.369234 Training: 46%|████▌ | 8203/18018 [17:51:10<21:37:50, 7.93s/step, loss=1.369, acc=0.693, acc_len=1.66, thru=16.0, I=24.4, T=32.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:24,622] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:31,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:21:32,033] loop.py:45 INFO TRAIN_STEP step=8204 loss=1.317059 Training: 46%|████▌ | 8204/18018 [17:51:18<21:20:03, 7.83s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=15.9, I=20.0, T=30.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:32,419] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:38,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:21:38,777] loop.py:380 INFO TIMING step=8205: step=4.245s data=1.906s compute=4.214s [fwd=2.931s bwd=1.271s opt=0.003s] dispatch=2.489s [2026-10-19 07:21:38,777] loop.py:45 INFO TRAIN_STEP step=8205 loss=1.332393 Training: 46%|████▌ | 8205/18018 [17:51:24<20:26:50, 7.50s/step, loss=1.332, acc=0.694, acc_len=1.66, thru=18.5, I=19.9, T=30.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:39,058] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:44,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:21:45,354] loop.py:45 INFO TRAIN_STEP step=8206 loss=1.351092 Training: 46%|████▌ | 8206/18018 [17:51:31<19:41:16, 7.22s/step, loss=1.351, acc=0.691, acc_len=1.65, thru=18.7, I=21.2, T=34.5, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:21:45,350] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8205: forward=2662.4ms backward=999.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:45,591] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:51,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:21:52,917] loop.py:45 INFO TRAIN_STEP step=8207 loss=1.380890 Training: 46%|████▌ | 8207/18018 [17:51:39<19:57:57, 7.33s/step, loss=1.381, acc=0.689, acc_len=1.64, thru=16.0, I=20.2, T=32.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:21:53,171] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:22:00,185] loop.py:45 INFO TRAIN_STEP step=8208 loss=1.367155 Training: 46%|████▌ | 8208/18018 [17:51:46<19:54:52, 7.31s/step, loss=1.367, acc=0.686, acc_len=1.63, thru=16.5, I=20.2, T=32.2, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:22:07,119] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:22:09,046] loop.py:45 INFO TRAIN_STEP step=8209 loss=1.338589 Training: 46%|████▌ | 8209/18018 [17:51:55<21:10:59, 7.77s/step, loss=1.339, acc=0.694, acc_len=1.67, thru=15.2, I=17.9, T=26.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:22:09,140] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 07:22:18,565] loop.py:380 INFO TIMING step=8210: step=4.445s data=1.996s compute=4.411s [fwd=3.078s bwd=1.318s opt=0.003s] dispatch=5.063s [2026-10-19 07:22:18,565] loop.py:45 INFO TRAIN_STEP step=8210 loss=1.373587 Training: 46%|████▌ | 8210/18018 [17:52:04<22:36:27, 8.30s/step, loss=1.374, acc=0.693, acc_len=1.66, thru=9.8, I=10.5, T=28.8, wait=5.1s, pool=28, epoch=1/2] [2026-10-19 07:22:27,475] loop.py:45 INFO TRAIN_STEP step=8211 loss=1.410042 Training: 46%|████▌ | 8211/18018 [17:52:13<23:06:09, 8.48s/step, loss=1.410, acc=0.684, acc_len=1.63, thru=17.4, I=12.7, T=27.6, wait=4.3s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:22:27,470] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8210: forward=3193.8ms backward=1381.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:22:36,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:22:36,641] loop.py:45 INFO TRAIN_STEP step=8212 loss=1.583996 Training: 46%|████▌ | 8212/18018 [17:52:22<23:39:44, 8.69s/step, loss=1.584, acc=0.653, acc_len=1.52, thru=15.4, I=15.2, T=34.0, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:22:36,730] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:22:44,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:22:44,580] loop.py:45 INFO TRAIN_STEP step=8213 loss=1.373275 Training: 46%|████▌ | 8213/18018 [17:52:30<23:02:51, 8.46s/step, loss=1.373, acc=0.692, acc_len=1.65, thru=15.8, I=16.7, T=33.8, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:22:48,232] inference_manager.py:408 INFO Pool capacity available after 4.1s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:22:52,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:22:53,932] loop.py:45 INFO TRAIN_STEP step=8214 loss=1.403226 Training: 46%|████▌ | 8214/18018 [17:52:40<23:46:26, 8.73s/step, loss=1.403, acc=0.684, acc_len=1.63, thru=13.4, I=24.3, T=33.8, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:22:54,224] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:01,416] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:23:02,251] loop.py:380 INFO TIMING step=8215: step=3.754s data=1.581s compute=3.725s [fwd=2.684s bwd=1.026s opt=0.003s] dispatch=4.555s [2026-10-19 07:23:02,252] loop.py:45 INFO TRAIN_STEP step=8215 loss=1.357908 Training: 46%|████▌ | 8215/18018 [17:52:48<23:26:08, 8.61s/step, loss=1.358, acc=0.692, acc_len=1.66, thru=14.5, I=18.5, T=34.1, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:02,421] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:10,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:23:11,210] loop.py:45 INFO TRAIN_STEP step=8216 loss=1.429506 Training: 46%|████▌ | 8216/18018 [17:52:57<23:43:10, 8.71s/step, loss=1.430, acc=0.682, acc_len=1.61, thru=13.7, I=15.4, T=27.0, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:23:11,206] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8215: forward=3240.9ms backward=1441.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:11,680] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:16,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:23:17,711] loop.py:45 INFO TRAIN_STEP step=8217 loss=1.408560 Training: 46%|████▌ | 8217/18018 [17:53:03<21:54:41, 8.05s/step, loss=1.409, acc=0.682, acc_len=1.62, thru=16.4, I=19.4, T=30.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:17,928] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:22,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:23:25,025] loop.py:45 INFO TRAIN_STEP step=8218 loss=1.444913 Training: 46%|████▌ | 8218/18018 [17:53:11<21:18:35, 7.83s/step, loss=1.445, acc=0.680, acc_len=1.61, thru=16.8, I=22.5, T=29.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:25,503] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 07:23:32,820] loop.py:45 INFO TRAIN_STEP step=8219 loss=1.569844 Training: 46%|████▌ | 8219/18018 [17:53:18<21:16:57, 7.82s/step, loss=1.570, acc=0.656, acc_len=1.54, thru=13.4, I=19.6, T=31.7, wait=3.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:40,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:23:41,651] loop.py:380 INFO TIMING step=8220: step=4.720s data=2.288s compute=4.686s [fwd=3.365s bwd=1.306s opt=0.003s] dispatch=4.101s [2026-10-19 07:23:41,651] loop.py:45 INFO TRAIN_STEP step=8220 loss=1.424608 Training: 46%|████▌ | 8220/18018 [17:53:27<22:06:13, 8.12s/step, loss=1.425, acc=0.680, acc_len=1.61, thru=14.6, I=16.2, T=27.1, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:41,893] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:47,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:23:48,241] loop.py:45 INFO TRAIN_STEP step=8221 loss=1.410331 Training: 46%|████▌ | 8221/18018 [17:53:34<20:51:07, 7.66s/step, loss=1.410, acc=0.683, acc_len=1.62, thru=17.7, I=22.3, T=31.4, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:23:48,237] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8220: forward=2852.3ms backward=1175.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:23:48,371] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 [2026-10-19 07:23:55,955] loop.py:45 INFO TRAIN_STEP step=8222 loss=1.382018 Training: 46%|████▌ | 8222/18018 [17:53:42<20:53:29, 7.68s/step, loss=1.382, acc=0.684, acc_len=1.63, thru=14.1, I=15.4, T=29.5, wait=3.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:24:05,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:24:07,922] loop.py:45 INFO TRAIN_STEP step=8223 loss=1.463115 Training: 46%|████▌ | 8223/18018 [17:53:54<24:23:34, 8.97s/step, loss=1.463, acc=0.675, acc_len=1.61, thru=13.8, I=13.7, T=27.5, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:24:08,379] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 07:24:14,533] loop.py:45 INFO TRAIN_STEP step=8224 loss=1.383931 Training: 46%|████▌ | 8224/18018 [17:54:00<22:28:05, 8.26s/step, loss=1.384, acc=0.686, acc_len=1.63, thru=16.1, I=24.9, T=33.1, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:24:19,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:24:22,447] loop.py:380 INFO TIMING step=8225: step=4.847s data=2.010s compute=4.823s [fwd=3.275s bwd=1.534s opt=0.003s] dispatch=3.059s [2026-10-19 07:24:22,448] loop.py:45 INFO TRAIN_STEP step=8225 loss=1.288184 Training: 46%|████▌ | 8225/18018 [17:54:08<22:11:09, 8.16s/step, loss=1.288, acc=0.704, acc_len=1.70, thru=16.6, I=20.3, T=26.4, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:24:22,512] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=40/64 [2026-10-19 07:24:28,965] loop.py:45 INFO TRAIN_STEP step=8226 loss=1.316605 Training: 46%|████▌ | 8226/18018 [17:54:15<20:50:41, 7.66s/step, loss=1.317, acc=0.700, acc_len=1.68, thru=14.2, I=22.3, T=30.8, wait=2.4s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:24:28,960] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8225: forward=2840.7ms backward=1264.1ms [2026-10-19 07:24:39,239] loop.py:45 INFO TRAIN_STEP step=8227 loss=1.544761 Training: 46%|████▌ | 8227/18018 [17:54:25<22:58:19, 8.45s/step, loss=1.545, acc=0.659, acc_len=1.54, thru=14.1, I=13.6, T=28.0, wait=5.7s, pool=60, epoch=1/2][2026-10-19 07:24:49,093] loop.py:45 INFO TRAIN_STEP step=8228 loss=1.362347 Training: 46%|████▌ | 8228/18018 [17:54:35<24:07:18, 8.87s/step, loss=1.362, acc=0.694, acc_len=1.66, thru=13.7, I=13.9, T=26.5, wait=5.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:24:57,207] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:24:57,919] loop.py:45 INFO TRAIN_STEP step=8229 loss=1.364538 Training: 46%|████▌ | 8229/18018 [17:54:44<24:04:51, 8.86s/step, loss=1.365, acc=0.690, acc_len=1.65, thru=14.7, I=14.2, T=28.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:24:58,211] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:04,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:25:06,390] loop.py:380 INFO TIMING step=8230: step=4.194s data=1.622s compute=4.166s [fwd=2.894s bwd=1.257s opt=0.003s] dispatch=4.268s [2026-10-19 07:25:06,391] loop.py:45 INFO TRAIN_STEP step=8230 loss=1.408714 Training: 46%|████▌ | 8230/18018 [17:54:52<23:45:57, 8.74s/step, loss=1.409, acc=0.685, acc_len=1.63, thru=15.0, I=17.7, T=30.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:06,781] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:13,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:25:13,391] loop.py:45 INFO TRAIN_STEP step=8231 loss=1.385384 Training: 46%|████▌ | 8231/18018 [17:54:59<22:20:32, 8.22s/step, loss=1.385, acc=0.684, acc_len=1.63, thru=16.5, I=19.0, T=32.1, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:25:13,386] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8230: forward=2760.7ms backward=1180.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:13,761] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:19,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:25:20,167] loop.py:45 INFO TRAIN_STEP step=8232 loss=1.333424 Training: 46%|████▌ | 8232/18018 [17:55:06<21:09:54, 7.79s/step, loss=1.333, acc=0.698, acc_len=1.68, thru=18.9, I=19.9, T=31.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:20,503] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:25,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:25:27,394] loop.py:45 INFO TRAIN_STEP step=8233 loss=1.428204 Training: 46%|████▌ | 8233/18018 [17:55:13<20:42:26, 7.62s/step, loss=1.428, acc=0.685, acc_len=1.63, thru=16.9, I=22.0, T=29.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:27,544] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:34,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:25:35,362] loop.py:45 INFO TRAIN_STEP step=8234 loss=1.451396 Training: 46%|████▌ | 8234/18018 [17:55:21<20:59:26, 7.72s/step, loss=1.451, acc=0.674, acc_len=1.59, thru=14.1, I=18.9, T=29.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:35,496] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:41,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:25:42,154] loop.py:380 INFO TIMING step=8235: step=3.938s data=1.656s compute=3.896s [fwd=2.651s bwd=1.231s opt=0.003s] dispatch=2.842s [2026-10-19 07:25:42,154] loop.py:45 INFO TRAIN_STEP step=8235 loss=1.387206 Training: 46%|████▌ | 8235/18018 [17:55:28<20:13:37, 7.44s/step, loss=1.387, acc=0.692, acc_len=1.66, thru=18.4, I=19.3, T=32.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:42,299] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:25:49,404] loop.py:45 INFO TRAIN_STEP step=8236 loss=1.392796 Training: 46%|████▌ | 8236/18018 [17:55:35<20:04:08, 7.39s/step, loss=1.393, acc=0.683, acc_len=1.63, thru=16.8, I=20.1, T=28.4, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:25:49,399] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8235: forward=3208.6ms backward=1243.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:55,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:25:58,233] loop.py:45 INFO TRAIN_STEP step=8237 loss=1.409779 Training: 46%|████▌ | 8237/18018 [17:55:44<21:14:36, 7.82s/step, loss=1.410, acc=0.685, acc_len=1.64, thru=15.1, I=19.7, T=27.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:25:58,377] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:05,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:26:05,097] loop.py:45 INFO TRAIN_STEP step=8238 loss=1.480875 Training: 46%|████▌ | 8238/18018 [17:55:51<20:27:38, 7.53s/step, loss=1.481, acc=0.675, acc_len=1.60, thru=14.1, I=20.8, T=32.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:05,574] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:11,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:26:13,761] loop.py:45 INFO TRAIN_STEP step=8239 loss=1.459829 Training: 46%|████▌ | 8239/18018 [17:55:59<21:23:03, 7.87s/step, loss=1.460, acc=0.679, acc_len=1.61, thru=15.2, I=19.4, T=28.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:13,833] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 [2026-10-19 07:26:22,464] loop.py:380 INFO TIMING step=8240: step=4.467s data=2.157s compute=4.437s [fwd=3.132s bwd=1.290s opt=0.003s] dispatch=4.224s [2026-10-19 07:26:22,464] loop.py:45 INFO TRAIN_STEP step=8240 loss=1.461349 Training: 46%|████▌ | 8240/18018 [17:56:08<22:03:21, 8.12s/step, loss=1.461, acc=0.674, acc_len=1.60, thru=13.9, I=13.4, T=28.7, wait=4.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:30,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:26:30,916] loop.py:45 INFO TRAIN_STEP step=8241 loss=1.395237 Training: 46%|████▌ | 8241/18018 [17:56:17<22:19:30, 8.22s/step, loss=1.395, acc=0.683, acc_len=1.62, thru=16.9, I=13.3, T=30.4, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:26:30,911] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8240: forward=3025.4ms backward=1131.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:31,348] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:37,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:26:39,013] loop.py:45 INFO TRAIN_STEP step=8242 loss=1.486030 Training: 46%|████▌ | 8242/18018 [17:56:25<22:13:19, 8.18s/step, loss=1.486, acc=0.670, acc_len=1.58, thru=17.0, I=19.6, T=29.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:39,301] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:43,888] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:26:45,814] loop.py:45 INFO TRAIN_STEP step=8243 loss=1.534189 Training: 46%|████▌ | 8243/18018 [17:56:31<21:05:44, 7.77s/step, loss=1.534, acc=0.662, acc_len=1.56, thru=16.6, I=25.3, T=30.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:26:45,906] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=32/64 [2026-10-19 07:26:51,882] loop.py:45 INFO TRAIN_STEP step=8244 loss=1.248382 Training: 46%|████▌ | 8244/18018 [17:56:37<19:42:20, 7.26s/step, loss=1.248, acc=0.715, acc_len=1.74, thru=18.1, I=24.6, T=36.6, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:03,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:27:03,834] loop.py:380 INFO TIMING step=8245: step=4.689s data=2.213s compute=4.643s [fwd=3.256s bwd=1.372s opt=0.003s] dispatch=7.254s [2026-10-19 07:27:03,834] loop.py:45 INFO TRAIN_STEP step=8245 loss=1.431769 Training: 46%|████▌ | 8245/18018 [17:56:49<23:31:35, 8.67s/step, loss=1.432, acc=0.682, acc_len=1.62, thru=13.1, I=11.9, T=27.3, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:03,863] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:09,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:27:10,837] loop.py:45 INFO TRAIN_STEP step=8246 loss=1.335936 Training: 46%|████▌ | 8246/18018 [17:56:56<22:10:10, 8.17s/step, loss=1.336, acc=0.698, acc_len=1.67, thru=16.0, I=18.5, T=27.0, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:27:10,832] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8245: forward=3292.0ms backward=1383.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:11,164] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:18,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:27:19,107] loop.py:45 INFO TRAIN_STEP step=8247 loss=1.360718 Training: 46%|████▌ | 8247/18018 [17:57:05<22:15:09, 8.20s/step, loss=1.361, acc=0.693, acc_len=1.66, thru=13.9, I=16.5, T=31.5, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:19,142] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 07:27:27,660] loop.py:45 INFO TRAIN_STEP step=8248 loss=1.382136 Training: 46%|████▌ | 8248/18018 [17:57:13<22:32:15, 8.30s/step, loss=1.382, acc=0.688, acc_len=1.64, thru=13.6, I=17.2, T=32.6, wait=4.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:37,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:27:39,369] loop.py:45 INFO TRAIN_STEP step=8249 loss=1.564543 Training: 46%|████▌ | 8249/18018 [17:57:25<25:18:26, 9.33s/step, loss=1.565, acc=0.656, acc_len=1.54, thru=11.7, I=13.3, T=26.6, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:39,590] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:45,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:27:45,788] loop.py:380 INFO TIMING step=8250: step=3.911s data=1.719s compute=3.884s [fwd=2.796s bwd=1.072s opt=0.003s] dispatch=2.499s [2026-10-19 07:27:45,788] loop.py:45 INFO TRAIN_STEP step=8250 loss=1.445137 Training: 46%|████▌ | 8250/18018 [17:57:31<22:56:18, 8.45s/step, loss=1.445, acc=0.682, acc_len=1.62, thru=17.2, I=19.7, T=32.7, wait=2.5s, pool=64, epoch=1/2][2026-10-19 07:27:45,792] loop.py:419 INFO Saving checkpoint at step 8250... (TrainerActor pid=9310) /usr/local/lib/python3.12/dist-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning. (TrainerActor pid=9310) return func(*args, **kwargs) (TrainerActor pid=9310) [2026-10-19 07:27:49,250] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0008251 (TrainerActor pid=9310) [2026-10-19 07:27:49,260] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0007501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:50,711] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:27:50,711] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=0/64 [2026-10-19 07:27:57,500] loop.py:45 INFO TRAIN_STEP step=8251 loss=1.440860 Training: 46%|████▌ | 8251/18018 [17:57:43<25:35:14, 9.43s/step, loss=1.441, acc=0.680, acc_len=1.61, thru=17.1, I=21.1, T=28.6, wait=2.8s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:27:57,495] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8250: forward=3154.0ms backward=1268.7ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 07:27:49,252] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0008251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 07:27:50,203] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0007501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:03,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:28:05,300] loop.py:45 INFO TRAIN_STEP step=8252 loss=1.515331 Training: 46%|████▌ | 8252/18018 [17:57:51<24:15:29, 8.94s/step, loss=1.515, acc=0.661, acc_len=1.55, thru=16.6, I=17.2, T=28.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:05,451] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:12,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:28:14,341] loop.py:45 INFO TRAIN_STEP step=8253 loss=1.451034 Training: 46%|████▌ | 8253/18018 [17:58:00<24:20:02, 8.97s/step, loss=1.451, acc=0.679, acc_len=1.61, thru=14.4, I=15.2, T=27.7, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:14,833] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:28:21,869] loop.py:45 INFO TRAIN_STEP step=8254 loss=1.401865 Training: 46%|████▌ | 8254/18018 [17:58:07<23:09:28, 8.54s/step, loss=1.402, acc=0.685, acc_len=1.63, thru=13.3, I=15.8, T=29.7, wait=3.2s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:27,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:28:29,260] loop.py:380 INFO TIMING step=8255: step=3.960s data=1.695s compute=3.927s [fwd=2.801s bwd=1.111s opt=0.003s] dispatch=3.423s [2026-10-19 07:28:29,261] loop.py:45 INFO TRAIN_STEP step=8255 loss=1.338479 Training: 46%|████▌ | 8255/18018 [17:58:15<22:13:22, 8.19s/step, loss=1.338, acc=0.695, acc_len=1.66, thru=18.2, I=18.7, T=32.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:29,521] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:35,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:28:35,839] loop.py:45 INFO TRAIN_STEP step=8256 loss=1.359436 Training: 46%|████▌ | 8256/18018 [17:58:21<20:54:22, 7.71s/step, loss=1.359, acc=0.688, acc_len=1.64, thru=18.3, I=26.8, T=31.8, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:28:35,834] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8255: forward=2825.4ms backward=1161.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:36,151] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:43,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:28:45,059] loop.py:45 INFO TRAIN_STEP step=8257 loss=1.563301 Training: 46%|████▌ | 8257/18018 [17:58:31<22:07:55, 8.16s/step, loss=1.563, acc=0.656, acc_len=1.53, thru=13.3, I=14.9, T=29.1, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:45,438] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:52,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:28:53,668] loop.py:45 INFO TRAIN_STEP step=8258 loss=1.350216 Training: 46%|████▌ | 8258/18018 [17:58:39<22:29:39, 8.30s/step, loss=1.350, acc=0.693, acc_len=1.66, thru=15.6, I=16.9, T=29.6, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:28:53,962] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:00,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:29:00,994] loop.py:45 INFO TRAIN_STEP step=8259 loss=1.428033 Training: 46%|████▌ | 8259/18018 [17:58:47<21:42:07, 8.01s/step, loss=1.428, acc=0.678, acc_len=1.61, thru=16.9, I=18.8, T=31.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:01,267] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:06,518] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:29:07,115] loop.py:380 INFO TIMING step=8260: step=3.813s data=1.633s compute=3.777s [fwd=2.621s bwd=1.141s opt=0.003s] dispatch=2.298s [2026-10-19 07:29:07,115] loop.py:45 INFO TRAIN_STEP step=8260 loss=1.506265 Training: 46%|████▌ | 8260/18018 [17:58:53<20:09:55, 7.44s/step, loss=1.506, acc=0.668, acc_len=1.57, thru=19.4, I=22.1, T=33.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:07,525] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:12,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:29:13,639] loop.py:45 INFO TRAIN_STEP step=8261 loss=1.315472 Training: 46%|████▌ | 8261/18018 [17:58:59<19:25:15, 7.17s/step, loss=1.315, acc=0.700, acc_len=1.68, thru=18.5, I=22.6, T=34.0, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:29:13,635] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8260: forward=2699.7ms backward=1028.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:14,032] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:29:20,200] loop.py:45 INFO TRAIN_STEP step=8262 loss=1.512204 Training: 46%|████▌ | 8262/18018 [17:59:06<18:55:33, 6.98s/step, loss=1.512, acc=0.671, acc_len=1.58, thru=16.9, I=19.1, T=33.4, wait=2.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:26,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:29:28,390] loop.py:45 INFO TRAIN_STEP step=8263 loss=1.464732 Training: 46%|████▌ | 8263/18018 [17:59:14<19:54:21, 7.35s/step, loss=1.465, acc=0.675, acc_len=1.59, thru=17.3, I=18.3, T=28.7, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:28,641] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:33,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:29:35,689] loop.py:45 INFO TRAIN_STEP step=8264 loss=1.410226 Training: 46%|████▌ | 8264/18018 [17:59:21<19:51:52, 7.33s/step, loss=1.410, acc=0.679, acc_len=1.61, thru=15.3, I=22.0, T=28.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:36,003] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:41,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:29:43,111] loop.py:380 INFO TIMING step=8265: step=4.128s data=1.876s compute=4.096s [fwd=2.928s bwd=1.153s opt=0.003s] dispatch=3.285s [2026-10-19 07:29:43,111] loop.py:45 INFO TRAIN_STEP step=8265 loss=1.456140 Training: 46%|████▌ | 8265/18018 [17:59:29<19:56:16, 7.36s/step, loss=1.456, acc=0.676, acc_len=1.60, thru=14.7, I=20.8, T=31.0, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:43,243] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:49,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:29:50,720] loop.py:45 INFO TRAIN_STEP step=8266 loss=1.361839 Training: 46%|████▌ | 8266/18018 [17:59:36<20:08:13, 7.43s/step, loss=1.362, acc=0.692, acc_len=1.66, thru=16.5, I=20.5, T=23.9, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:29:50,716] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8265: forward=3601.3ms backward=1696.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:51,218] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:29:57,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:29:59,858] loop.py:45 INFO TRAIN_STEP step=8267 loss=1.436861 Training: 46%|████▌ | 8267/18018 [17:59:45<21:31:14, 7.95s/step, loss=1.437, acc=0.677, acc_len=1.61, thru=14.8, I=18.3, T=25.4, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:30:00,269] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 07:30:06,670] loop.py:45 INFO TRAIN_STEP step=8268 loss=1.373177 Training: 46%|████▌ | 8268/18018 [17:59:52<20:35:45, 7.60s/step, loss=1.373, acc=0.688, acc_len=1.65, thru=12.9, I=15.9, T=33.3, wait=3.0s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:30:16,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:30:18,113] loop.py:45 INFO TRAIN_STEP step=8269 loss=1.325141 Training: 46%|████▌ | 8269/18018 [18:00:04<23:42:50, 8.76s/step, loss=1.325, acc=0.694, acc_len=1.67, thru=12.9, I=15.1, T=28.7, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:30:18,478] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:30:24,888] loop.py:380 INFO TIMING step=8270: step=4.228s data=2.014s compute=4.185s [fwd=2.865s bwd=1.305s opt=0.003s] dispatch=2.537s [2026-10-19 07:30:24,888] loop.py:45 INFO TRAIN_STEP step=8270 loss=1.474596 Training: 46%|████▌ | 8270/18018 [18:00:10<22:06:07, 8.16s/step, loss=1.475, acc=0.670, acc_len=1.58, thru=15.7, I=21.4, T=30.3, wait=2.5s, pool=56, epoch=1/2][2026-10-19 07:30:34,463] loop.py:45 INFO TRAIN_STEP step=8271 loss=1.443609 Training: 46%|████▌ | 8271/18018 [18:00:20<23:14:46, 8.59s/step, loss=1.444, acc=0.682, acc_len=1.61, thru=13.8, I=11.7, T=32.1, wait=5.6s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:30:34,458] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8270: forward=2844.3ms backward=1092.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:30:39,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:30:41,589] loop.py:45 INFO TRAIN_STEP step=8272 loss=1.351744 Training: 46%|████▌ | 8272/18018 [18:00:27<22:03:26, 8.15s/step, loss=1.352, acc=0.688, acc_len=1.64, thru=18.4, I=19.8, T=35.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:30:41,931] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:30:48,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:30:48,934] loop.py:45 INFO TRAIN_STEP step=8273 loss=1.342672 Training: 46%|████▌ | 8273/18018 [18:00:35<21:24:14, 7.91s/step, loss=1.343, acc=0.692, acc_len=1.66, thru=16.8, I=20.4, T=27.7, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:30:49,236] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:30:55,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:30:57,874] loop.py:45 INFO TRAIN_STEP step=8274 loss=1.387244 Training: 46%|████▌ | 8274/18018 [18:00:43<22:14:27, 8.22s/step, loss=1.387, acc=0.689, acc_len=1.64, thru=14.9, I=18.1, T=30.3, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:30:57,989] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:31:03,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:31:04,008] loop.py:380 INFO TIMING step=8275: step=3.702s data=1.635s compute=3.672s [fwd=2.621s bwd=1.036s opt=0.003s] dispatch=2.422s [2026-10-19 07:31:04,008] loop.py:45 INFO TRAIN_STEP step=8275 loss=1.404453 Training: 46%|████▌ | 8275/18018 [18:00:50<20:32:50, 7.59s/step, loss=1.404, acc=0.682, acc_len=1.62, thru=18.4, I=25.2, T=34.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:31:04,317] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:31:13,549] loop.py:45 INFO TRAIN_STEP step=8276 loss=1.455025 Training: 46%|████▌ | 8276/18018 [18:00:59<22:07:40, 8.18s/step, loss=1.455, acc=0.677, acc_len=1.61, thru=13.4, I=15.7, T=29.7, wait=5.2s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:31:13,545] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8275: forward=3024.9ms backward=1232.6ms [2026-10-19 07:31:23,625] loop.py:45 INFO TRAIN_STEP step=8277 loss=1.480026 Training: 46%|████▌ | 8277/18018 [18:01:09<23:39:51, 8.75s/step, loss=1.480, acc=0.669, acc_len=1.58, thru=12.7, I=10.8, T=29.8, wait=5.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:31:30,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:31:32,646] loop.py:45 INFO TRAIN_STEP step=8278 loss=1.381106 Training: 46%|████▌ | 8278/18018 [18:01:18<23:53:11, 8.83s/step, loss=1.381, acc=0.686, acc_len=1.64, thru=15.8, I=17.5, T=26.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:31:32,837] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:31:40,449] loop.py:45 INFO TRAIN_STEP step=8279 loss=1.458274 Training: 46%|████▌ | 8279/18018 [18:01:26<23:03:10, 8.52s/step, loss=1.458, acc=0.672, acc_len=1.59, thru=12.9, I=19.5, T=24.9, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:31:48,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:31:48,870] loop.py:380 INFO TIMING step=8280: step=4.172s data=1.833s compute=4.130s [fwd=3.000s bwd=1.115s opt=0.003s] dispatch=4.239s [2026-10-19 07:31:48,870] loop.py:45 INFO TRAIN_STEP step=8280 loss=1.403732 Training: 46%|████▌ | 8280/18018 [18:01:34<22:58:08, 8.49s/step, loss=1.404, acc=0.683, acc_len=1.62, thru=15.9, I=14.1, T=30.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:31:49,157] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:31:54,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:31:55,727] loop.py:45 INFO TRAIN_STEP step=8281 loss=1.440218 Training: 46%|████▌ | 8281/18018 [18:01:41<21:38:26, 8.00s/step, loss=1.440, acc=0.676, acc_len=1.60, thru=18.0, I=20.7, T=29.7, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:31:55,722] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8280: forward=3010.8ms backward=1260.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:31:55,881] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:01,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:32:01,466] loop.py:45 INFO TRAIN_STEP step=8282 loss=1.230155 Training: 46%|████▌ | 8282/18018 [18:01:47<19:48:10, 7.32s/step, loss=1.230, acc=0.715, acc_len=1.73, thru=19.4, I=25.3, T=35.7, wait=2.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:01,697] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:08,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:32:09,852] loop.py:45 INFO TRAIN_STEP step=8283 loss=1.462622 Training: 46%|████▌ | 8283/18018 [18:01:55<20:39:48, 7.64s/step, loss=1.463, acc=0.671, acc_len=1.58, thru=14.4, I=16.0, T=31.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:10,324] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:15,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:32:16,518] loop.py:45 INFO TRAIN_STEP step=8284 loss=1.386996 Training: 46%|████▌ | 8284/18018 [18:02:02<19:52:08, 7.35s/step, loss=1.387, acc=0.691, acc_len=1.65, thru=19.2, I=20.5, T=33.8, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:16,670] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:32:25,556] loop.py:380 INFO TIMING step=8285: step=4.423s data=2.047s compute=4.388s [fwd=3.130s bwd=1.243s opt=0.003s] dispatch=4.607s [2026-10-19 07:32:25,556] loop.py:45 INFO TRAIN_STEP step=8285 loss=1.552324 Training: 46%|████▌ | 8285/18018 [18:02:11<21:14:12, 7.85s/step, loss=1.552, acc=0.659, acc_len=1.54, thru=13.3, I=14.1, T=28.9, wait=4.6s, pool=60, epoch=1/2][2026-10-19 07:32:32,992] loop.py:45 INFO TRAIN_STEP step=8286 loss=1.450409 Training: 46%|████▌ | 8286/18018 [18:02:19<20:53:42, 7.73s/step, loss=1.450, acc=0.677, acc_len=1.60, thru=15.5, I=13.9, T=31.9, wait=3.4s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:32:32,988] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8285: forward=2920.9ms backward=1046.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:39,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:32:40,727] loop.py:45 INFO TRAIN_STEP step=8287 loss=1.357210 Training: 46%|████▌ | 8287/18018 [18:02:26<20:53:57, 7.73s/step, loss=1.357, acc=0.690, acc_len=1.66, thru=18.1, I=17.8, T=33.8, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:41,099] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:46,745] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:32:47,596] loop.py:45 INFO TRAIN_STEP step=8288 loss=1.403384 Training: 46%|████▌ | 8288/18018 [18:02:33<20:11:56, 7.47s/step, loss=1.403, acc=0.683, acc_len=1.62, thru=18.1, I=24.3, T=35.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:47,752] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:54,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:32:54,955] loop.py:45 INFO TRAIN_STEP step=8289 loss=1.452687 Training: 46%|████▌ | 8289/18018 [18:02:41<20:06:09, 7.44s/step, loss=1.453, acc=0.674, acc_len=1.59, thru=17.2, I=20.0, T=28.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:32:55,012] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-19 07:33:03,038] loop.py:380 INFO TIMING step=8290: step=3.891s data=1.647s compute=3.860s [fwd=2.771s bwd=1.074s opt=0.003s] dispatch=4.182s [2026-10-19 07:33:03,039] loop.py:45 INFO TRAIN_STEP step=8290 loss=1.311593 Training: 46%|████▌ | 8290/18018 [18:02:49<20:37:17, 7.63s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=13.6, I=14.3, T=32.9, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:11,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:33:13,310] loop.py:45 INFO TRAIN_STEP step=8291 loss=1.305018 Training: 46%|████▌ | 8291/18018 [18:02:59<22:45:41, 8.42s/step, loss=1.305, acc=0.702, acc_len=1.69, thru=13.2, I=15.4, T=27.4, wait=5.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:33:13,305] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8290: forward=3230.3ms backward=1381.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:13,635] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:19,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:33:20,106] loop.py:45 INFO TRAIN_STEP step=8292 loss=1.489460 Training: 46%|████▌ | 8292/18018 [18:03:06<21:26:20, 7.94s/step, loss=1.489, acc=0.669, acc_len=1.57, thru=15.8, I=21.4, T=32.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:20,297] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:25,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:33:27,222] loop.py:45 INFO TRAIN_STEP step=8293 loss=1.482971 Training: 46%|████▌ | 8293/18018 [18:03:13<20:46:24, 7.69s/step, loss=1.483, acc=0.671, acc_len=1.59, thru=17.7, I=21.1, T=31.2, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:27,525] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:33,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:33:33,633] loop.py:45 INFO TRAIN_STEP step=8294 loss=1.394676 Training: 46%|████▌ | 8294/18018 [18:03:19<19:44:07, 7.31s/step, loss=1.395, acc=0.684, acc_len=1.63, thru=18.4, I=24.1, T=32.0, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:34,141] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:39,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:33:40,004] loop.py:380 INFO TIMING step=8295: step=3.813s data=1.670s compute=3.775s [fwd=2.646s bwd=1.114s opt=0.003s] dispatch=2.547s [2026-10-19 07:33:40,004] loop.py:45 INFO TRAIN_STEP step=8295 loss=1.393773 Training: 46%|████▌ | 8295/18018 [18:03:26<18:58:23, 7.02s/step, loss=1.394, acc=0.685, acc_len=1.63, thru=19.4, I=20.0, T=33.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:40,245] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:33:48,056] loop.py:45 INFO TRAIN_STEP step=8296 loss=1.354129 Training: 46%|████▌ | 8296/18018 [18:03:34<19:48:13, 7.33s/step, loss=1.354, acc=0.698, acc_len=1.67, thru=15.3, I=16.5, T=32.0, wait=4.0s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:33:48,051] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8295: forward=2806.1ms backward=1145.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:54,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:33:56,801] loop.py:45 INFO TRAIN_STEP step=8297 loss=1.377905 Training: 46%|████▌ | 8297/18018 [18:03:42<20:56:41, 7.76s/step, loss=1.378, acc=0.690, acc_len=1.65, thru=15.9, I=18.2, T=27.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:33:57,260] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 07:34:04,034] loop.py:45 INFO TRAIN_STEP step=8298 loss=1.448596 Training: 46%|████▌ | 8298/18018 [18:03:50<20:31:11, 7.60s/step, loss=1.449, acc=0.678, acc_len=1.60, thru=11.2, I=15.9, T=31.3, wait=3.1s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:14,131] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:34:14,931] loop.py:45 INFO TRAIN_STEP step=8299 loss=1.256311 Training: 46%|████▌ | 8299/18018 [18:04:01<23:11:14, 8.59s/step, loss=1.256, acc=0.710, acc_len=1.72, thru=12.6, I=14.0, T=30.1, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:15,136] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:22,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:34:23,844] loop.py:380 INFO TIMING step=8300: step=4.011s data=1.837s compute=3.982s [fwd=2.877s bwd=1.090s opt=0.003s] dispatch=4.893s [2026-10-19 07:34:23,845] loop.py:45 INFO TRAIN_STEP step=8300 loss=1.429168 Training: 46%|████▌ | 8300/18018 [18:04:09<23:26:58, 8.69s/step, loss=1.429, acc=0.683, acc_len=1.63, thru=13.6, I=19.2, T=31.9, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:24,170] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:34:31,259] loop.py:45 INFO TRAIN_STEP step=8301 loss=1.379356 Training: 46%|████▌ | 8301/18018 [18:04:17<22:24:50, 8.30s/step, loss=1.379, acc=0.689, acc_len=1.64, thru=13.5, I=17.5, T=30.7, wait=3.2s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:34:31,254] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8300: forward=2941.5ms backward=1171.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:39,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:34:41,271] loop.py:45 INFO TRAIN_STEP step=8302 loss=1.460669 Training: 46%|████▌ | 8302/18018 [18:04:27<23:47:44, 8.82s/step, loss=1.461, acc=0.678, acc_len=1.61, thru=15.4, I=17.5, T=25.8, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:41,377] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:46,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:34:47,864] loop.py:45 INFO TRAIN_STEP step=8303 loss=1.362873 Training: 46%|████▌ | 8303/18018 [18:04:33<21:59:33, 8.15s/step, loss=1.363, acc=0.689, acc_len=1.65, thru=16.9, I=22.1, T=32.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:48,150] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:52,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:34:54,223] loop.py:45 INFO TRAIN_STEP step=8304 loss=1.417704 Training: 46%|████▌ | 8304/18018 [18:04:40<20:32:34, 7.61s/step, loss=1.418, acc=0.678, acc_len=1.61, thru=18.5, I=24.6, T=35.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:34:54,405] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:02,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:35:03,737] loop.py:380 INFO TIMING step=8305: step=5.195s data=1.859s compute=5.170s [fwd=3.682s bwd=1.473s opt=0.003s] dispatch=4.309s [2026-10-19 07:35:03,737] loop.py:45 INFO TRAIN_STEP step=8305 loss=1.251966 Training: 46%|████▌ | 8305/18018 [18:04:49<22:04:43, 8.18s/step, loss=1.252, acc=0.711, acc_len=1.72, thru=13.7, I=15.6, T=24.6, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:03,988] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:12,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:35:12,873] loop.py:45 INFO TRAIN_STEP step=8306 loss=1.383409 Training: 46%|████▌ | 8306/18018 [18:04:58<22:50:43, 8.47s/step, loss=1.383, acc=0.689, acc_len=1.65, thru=14.4, I=15.4, T=28.5, wait=4.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:35:12,868] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8305: forward=3073.8ms backward=1359.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:13,341] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:19,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:35:20,428] loop.py:45 INFO TRAIN_STEP step=8307 loss=1.353512 Training: 46%|████▌ | 8307/18018 [18:05:06<22:06:17, 8.19s/step, loss=1.354, acc=0.690, acc_len=1.65, thru=17.0, I=19.2, T=31.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:20,719] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:35:26,865] loop.py:45 INFO TRAIN_STEP step=8308 loss=1.405183 Training: 46%|████▌ | 8308/18018 [18:05:12<20:40:48, 7.67s/step, loss=1.405, acc=0.687, acc_len=1.63, thru=18.2, I=19.9, T=31.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:26,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:27,318] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:35:34,776] loop.py:45 INFO TRAIN_STEP step=8309 loss=1.467450 Training: 46%|████▌ | 8309/18018 [18:05:20<20:52:39, 7.74s/step, loss=1.467, acc=0.674, acc_len=1.59, thru=15.7, I=17.8, T=31.4, wait=3.8s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:42,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:35:44,066] loop.py:380 INFO TIMING step=8310: step=4.317s data=1.924s compute=4.292s [fwd=2.990s bwd=1.287s opt=0.003s] dispatch=4.963s [2026-10-19 07:35:44,066] loop.py:45 INFO TRAIN_STEP step=8310 loss=1.297001 Training: 46%|████▌ | 8310/18018 [18:05:30<22:07:34, 8.21s/step, loss=1.297, acc=0.703, acc_len=1.70, thru=14.9, I=15.0, T=29.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:35:44,477] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:35:51,819] loop.py:45 INFO TRAIN_STEP step=8311 loss=1.387364 Training: 46%|████▌ | 8311/18018 [18:05:37<21:45:30, 8.07s/step, loss=1.387, acc=0.688, acc_len=1.64, thru=14.6, I=17.6, T=28.6, wait=3.3s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:35:51,815] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8310: forward=3058.6ms backward=1366.9ms [2026-10-19 07:36:01,863] loop.py:45 INFO TRAIN_STEP step=8312 loss=1.335331 Training: 46%|████▌ | 8312/18018 [18:05:47<23:21:17, 8.66s/step, loss=1.335, acc=0.697, acc_len=1.68, thru=10.6, I=12.0, T=30.2, wait=5.8s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:10,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:36:11,018] loop.py:45 INFO TRAIN_STEP step=8313 loss=1.480249 Training: 46%|████▌ | 8313/18018 [18:05:57<23:44:57, 8.81s/step, loss=1.480, acc=0.672, acc_len=1.58, thru=16.5, I=15.4, T=33.1, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:11,245] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:16,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:36:17,987] loop.py:45 INFO TRAIN_STEP step=8314 loss=1.353661 Training: 46%|████▌ | 8314/18018 [18:06:04<22:15:34, 8.26s/step, loss=1.354, acc=0.692, acc_len=1.65, thru=18.6, I=23.0, T=31.9, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:18,215] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:24,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:36:24,838] loop.py:380 INFO TIMING step=8315: step=4.138s data=1.967s compute=4.107s [fwd=2.884s bwd=1.208s opt=0.003s] dispatch=2.702s [2026-10-19 07:36:24,838] loop.py:45 INFO TRAIN_STEP step=8315 loss=1.269722 Training: 46%|████▌ | 8315/18018 [18:06:10<21:07:03, 7.84s/step, loss=1.270, acc=0.709, acc_len=1.71, thru=16.8, I=21.1, T=30.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:25,284] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:29,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:36:31,376] loop.py:45 INFO TRAIN_STEP step=8316 loss=1.397222 Training: 46%|████▌ | 8316/18018 [18:06:17<20:04:03, 7.45s/step, loss=1.397, acc=0.687, acc_len=1.64, thru=18.8, I=23.6, T=33.3, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:36:31,371] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8315: forward=2649.5ms backward=1147.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:31,511] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:38,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:36:39,631] loop.py:45 INFO TRAIN_STEP step=8317 loss=1.442767 Training: 46%|████▌ | 8317/18018 [18:06:25<20:43:06, 7.69s/step, loss=1.443, acc=0.677, acc_len=1.60, thru=14.6, I=20.4, T=31.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:39,800] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:36:46,265] loop.py:45 INFO TRAIN_STEP step=8318 loss=1.248755 Training: 46%|████▌ | 8318/18018 [18:06:32<19:51:52, 7.37s/step, loss=1.249, acc=0.710, acc_len=1.72, thru=18.0, I=19.9, T=35.0, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:54,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:36:55,746] loop.py:45 INFO TRAIN_STEP step=8319 loss=1.378321 Training: 46%|████▌ | 8319/18018 [18:06:41<21:34:06, 8.01s/step, loss=1.378, acc=0.689, acc_len=1.65, thru=14.5, I=15.1, T=30.2, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:36:56,254] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:37:04,212] loop.py:380 INFO TIMING step=8320: step=3.757s data=1.694s compute=3.712s [fwd=2.633s bwd=1.064s opt=0.003s] dispatch=4.699s [2026-10-19 07:37:04,213] loop.py:45 INFO TRAIN_STEP step=8320 loss=1.449442 Training: 46%|████▌ | 8320/18018 [18:06:50<21:56:13, 8.14s/step, loss=1.449, acc=0.676, acc_len=1.60, thru=12.7, I=14.5, T=34.1, wait=4.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:10,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:37:12,977] loop.py:45 INFO TRAIN_STEP step=8321 loss=1.556248 Training: 46%|████▌ | 8321/18018 [18:06:59<22:26:07, 8.33s/step, loss=1.556, acc=0.659, acc_len=1.55, thru=16.2, I=18.8, T=24.6, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:37:12,973] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8320: forward=3983.1ms backward=1178.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:13,011] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:19,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:37:22,120] loop.py:45 INFO TRAIN_STEP step=8322 loss=1.438933 Training: 46%|████▌ | 8322/18018 [18:07:08<23:05:38, 8.57s/step, loss=1.439, acc=0.678, acc_len=1.61, thru=14.1, I=14.9, T=21.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:22,353] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 07:37:29,059] loop.py:45 INFO TRAIN_STEP step=8323 loss=1.337801 Training: 46%|████▌ | 8323/18018 [18:07:15<21:46:08, 8.08s/step, loss=1.338, acc=0.697, acc_len=1.67, thru=11.6, I=14.8, T=31.4, wait=2.9s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:36,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:37:38,526] loop.py:45 INFO TRAIN_STEP step=8324 loss=1.451312 Training: 46%|████▌ | 8324/18018 [18:07:24<22:53:02, 8.50s/step, loss=1.451, acc=0.675, acc_len=1.59, thru=15.7, I=19.1, T=29.7, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:38,555] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:44,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:37:45,692] loop.py:380 INFO TIMING step=8325: step=4.474s data=2.105s compute=4.437s [fwd=3.092s bwd=1.330s opt=0.003s] dispatch=2.684s [2026-10-19 07:37:45,692] loop.py:45 INFO TRAIN_STEP step=8325 loss=1.227323 Training: 46%|████▌ | 8325/18018 [18:07:31<21:48:23, 8.10s/step, loss=1.227, acc=0.716, acc_len=1.74, thru=15.3, I=21.1, T=28.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:45,839] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:52,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:37:54,046] loop.py:45 INFO TRAIN_STEP step=8326 loss=1.502818 Training: 46%|████▌ | 8326/18018 [18:07:40<22:00:38, 8.18s/step, loss=1.503, acc=0.665, acc_len=1.57, thru=15.1, I=17.8, T=27.4, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:37:54,042] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8325: forward=3244.6ms backward=1379.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:37:54,246] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:00,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:38:00,728] loop.py:45 INFO TRAIN_STEP step=8327 loss=1.368227 Training: 46%|████▌ | 8327/18018 [18:07:46<20:48:10, 7.73s/step, loss=1.368, acc=0.687, acc_len=1.64, thru=17.7, I=20.3, T=31.4, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:01,186] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:07,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:38:08,199] loop.py:45 INFO TRAIN_STEP step=8328 loss=1.451158 Training: 46%|████▌ | 8328/18018 [18:07:54<20:35:34, 7.65s/step, loss=1.451, acc=0.672, acc_len=1.59, thru=16.4, I=20.8, T=30.2, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:08,449] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:14,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:38:15,950] loop.py:45 INFO TRAIN_STEP step=8329 loss=1.501133 Training: 46%|████▌ | 8329/18018 [18:08:02<20:40:17, 7.68s/step, loss=1.501, acc=0.669, acc_len=1.58, thru=15.1, I=18.3, T=31.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:16,237] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:20,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:38:22,701] loop.py:380 INFO TIMING step=8330: step=4.202s data=1.982s compute=4.174s [fwd=2.941s bwd=1.219s opt=0.003s] dispatch=2.540s [2026-10-19 07:38:22,702] loop.py:45 INFO TRAIN_STEP step=8330 loss=1.317486 Training: 46%|████▌ | 8330/18018 [18:08:08<19:55:07, 7.40s/step, loss=1.317, acc=0.697, acc_len=1.68, thru=17.0, I=27.9, T=30.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:22,911] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:28,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:38:29,486] loop.py:45 INFO TRAIN_STEP step=8331 loss=1.364519 Training: 46%|████▌ | 8331/18018 [18:08:15<19:25:08, 7.22s/step, loss=1.365, acc=0.694, acc_len=1.65, thru=16.5, I=23.5, T=33.8, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:38:29,482] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8330: forward=2669.9ms backward=1074.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:29,830] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:34,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:38:35,948] loop.py:45 INFO TRAIN_STEP step=8332 loss=1.467706 Training: 46%|████▌ | 8332/18018 [18:08:22<18:48:27, 6.99s/step, loss=1.468, acc=0.673, acc_len=1.59, thru=18.4, I=24.9, T=30.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:36,379] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:41,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:38:42,140] loop.py:45 INFO TRAIN_STEP step=8333 loss=1.420869 Training: 46%|████▌ | 8333/18018 [18:08:28<18:09:35, 6.75s/step, loss=1.421, acc=0.678, acc_len=1.61, thru=18.0, I=23.9, T=33.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:42,293] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:49,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:38:50,389] loop.py:45 INFO TRAIN_STEP step=8334 loss=1.389535 Training: 46%|████▋ | 8334/18018 [18:08:36<19:22:04, 7.20s/step, loss=1.390, acc=0.687, acc_len=1.64, thru=15.2, I=17.2, T=29.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:50,676] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:56,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:38:57,313] loop.py:380 INFO TIMING step=8335: step=4.193s data=1.948s compute=4.150s [fwd=2.867s bwd=1.268s opt=0.003s] dispatch=2.722s [2026-10-19 07:38:57,313] loop.py:45 INFO TRAIN_STEP step=8335 loss=1.403772 Training: 46%|████▋ | 8335/18018 [18:08:43<19:08:41, 7.12s/step, loss=1.404, acc=0.681, acc_len=1.62, thru=17.4, I=21.2, T=30.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:38:57,626] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:04,736] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:39:05,176] loop.py:45 INFO TRAIN_STEP step=8336 loss=1.314219 Training: 46%|████▋ | 8336/18018 [18:08:51<19:44:31, 7.34s/step, loss=1.314, acc=0.700, acc_len=1.69, thru=14.7, I=18.4, T=36.0, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:39:05,172] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8335: forward=2424.9ms backward=1077.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:05,244] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:10,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:39:11,867] loop.py:45 INFO TRAIN_STEP step=8337 loss=1.389315 Training: 46%|████▋ | 8337/18018 [18:08:57<19:12:56, 7.15s/step, loss=1.389, acc=0.687, acc_len=1.64, thru=19.5, I=20.5, T=32.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:12,182] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:19,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:39:19,859] loop.py:45 INFO TRAIN_STEP step=8338 loss=1.323766 Training: 46%|████▋ | 8338/18018 [18:09:05<19:53:48, 7.40s/step, loss=1.324, acc=0.701, acc_len=1.69, thru=15.7, I=19.6, T=29.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:19,889] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:26,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:39:27,824] loop.py:45 INFO TRAIN_STEP step=8339 loss=1.369962 Training: 46%|████▋ | 8339/18018 [18:09:13<20:21:04, 7.57s/step, loss=1.370, acc=0.687, acc_len=1.64, thru=16.7, I=17.7, T=29.0, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:27,968] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:33,373] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:39:34,949] loop.py:380 INFO TIMING step=8340: step=4.149s data=1.824s compute=4.119s [fwd=2.929s bwd=1.175s opt=0.003s] dispatch=2.968s [2026-10-19 07:39:34,949] loop.py:45 INFO TRAIN_STEP step=8340 loss=1.465757 Training: 46%|████▋ | 8340/18018 [18:09:21<19:59:24, 7.44s/step, loss=1.466, acc=0.669, acc_len=1.58, thru=17.2, I=20.6, T=30.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:35,382] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:41,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:39:42,805] loop.py:45 INFO TRAIN_STEP step=8341 loss=1.428147 Training: 46%|████▋ | 8341/18018 [18:09:28<20:19:36, 7.56s/step, loss=1.428, acc=0.678, acc_len=1.61, thru=14.1, I=19.0, T=30.7, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:39:42,800] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8340: forward=2847.5ms backward=1253.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:43,168] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:49,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:39:50,188] loop.py:45 INFO TRAIN_STEP step=8342 loss=1.465934 Training: 46%|████▋ | 8342/18018 [18:09:36<20:10:50, 7.51s/step, loss=1.466, acc=0.673, acc_len=1.59, thru=16.6, I=20.8, T=32.9, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:50,676] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:56,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:39:57,126] loop.py:45 INFO TRAIN_STEP step=8343 loss=1.368662 Training: 46%|████▋ | 8343/18018 [18:09:43<19:43:13, 7.34s/step, loss=1.369, acc=0.692, acc_len=1.65, thru=17.7, I=21.1, T=30.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:39:57,515] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:03,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:40:04,300] loop.py:45 INFO TRAIN_STEP step=8344 loss=1.402117 Training: 46%|████▋ | 8344/18018 [18:09:50<19:35:07, 7.29s/step, loss=1.402, acc=0.683, acc_len=1.63, thru=17.0, I=22.0, T=30.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:04,748] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:09,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:40:10,435] loop.py:380 INFO TIMING step=8345: step=3.626s data=1.504s compute=3.585s [fwd=2.519s bwd=1.051s opt=0.003s] dispatch=2.500s [2026-10-19 07:40:10,435] loop.py:45 INFO TRAIN_STEP step=8345 loss=1.366256 Training: 46%|████▋ | 8345/18018 [18:09:56<18:39:14, 6.94s/step, loss=1.366, acc=0.690, acc_len=1.65, thru=19.1, I=25.4, T=35.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:10,557] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:16,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:40:16,793] loop.py:45 INFO TRAIN_STEP step=8346 loss=1.343637 Training: 46%|████▋ | 8346/18018 [18:10:02<18:10:53, 6.77s/step, loss=1.344, acc=0.690, acc_len=1.65, thru=19.3, I=24.7, T=30.9, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:40:16,788] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8345: forward=2698.0ms backward=1397.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:17,164] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:23,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:40:24,442] loop.py:45 INFO TRAIN_STEP step=8347 loss=1.381395 Training: 46%|████▋ | 8347/18018 [18:10:10<18:53:23, 7.03s/step, loss=1.381, acc=0.685, acc_len=1.63, thru=15.3, I=18.4, T=32.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:24,476] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:33,213] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:40:33,478] loop.py:45 INFO TRAIN_STEP step=8348 loss=1.240702 Training: 46%|████▋ | 8348/18018 [18:10:19<20:30:12, 7.63s/step, loss=1.241, acc=0.711, acc_len=1.72, thru=13.8, I=14.2, T=31.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:33,726] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:39,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:40:40,432] loop.py:45 INFO TRAIN_STEP step=8349 loss=1.345759 Training: 46%|████▋ | 8349/18018 [18:10:26<19:57:13, 7.43s/step, loss=1.346, acc=0.689, acc_len=1.64, thru=17.7, I=18.4, T=30.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:40,508] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:46,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:40:47,586] loop.py:380 INFO TIMING step=8350: step=4.067s data=1.650s compute=4.033s [fwd=2.777s bwd=1.241s opt=0.003s] dispatch=3.076s [2026-10-19 07:40:47,586] loop.py:45 INFO TRAIN_STEP step=8350 loss=1.517646 Training: 46%|████▋ | 8350/18018 [18:10:33<19:43:43, 7.35s/step, loss=1.518, acc=0.663, acc_len=1.55, thru=17.7, I=20.1, T=31.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:47,986] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:53,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:40:54,787] loop.py:45 INFO TRAIN_STEP step=8351 loss=1.414443 Training: 46%|████▋ | 8351/18018 [18:10:40<19:36:39, 7.30s/step, loss=1.414, acc=0.684, acc_len=1.62, thru=17.1, I=21.1, T=34.0, wait=3.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:40:54,782] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8350: forward=2672.1ms backward=1039.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:54,955] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:40:59,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:41:01,459] loop.py:45 INFO TRAIN_STEP step=8352 loss=1.379798 Training: 46%|████▋ | 8352/18018 [18:10:47<19:05:58, 7.11s/step, loss=1.380, acc=0.685, acc_len=1.64, thru=18.4, I=24.6, T=34.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:41:01,958] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:41:08,590] loop.py:45 INFO TRAIN_STEP step=8353 loss=1.471490 Training: 46%|████▋ | 8353/18018 [18:10:54<19:06:51, 7.12s/step, loss=1.471, acc=0.674, acc_len=1.60, thru=14.7, I=22.3, T=29.5, wait=2.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:41:17,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:41:19,131] loop.py:45 INFO TRAIN_STEP step=8354 loss=1.479723 Training: 46%|████▋ | 8354/18018 [18:11:05<21:51:57, 8.15s/step, loss=1.480, acc=0.675, acc_len=1.59, thru=14.1, I=15.6, T=25.4, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:41:19,376] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:41:25,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:41:26,067] loop.py:380 INFO TIMING step=8355: step=4.112s data=1.895s compute=4.065s [fwd=2.865s bwd=1.184s opt=0.003s] dispatch=2.814s [2026-10-19 07:41:26,067] loop.py:45 INFO TRAIN_STEP step=8355 loss=1.343693 Training: 46%|████▋ | 8355/18018 [18:11:12<20:53:24, 7.78s/step, loss=1.344, acc=0.696, acc_len=1.67, thru=15.5, I=20.1, T=31.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:41:26,339] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:41:31,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:41:32,284] loop.py:45 INFO TRAIN_STEP step=8356 loss=1.452478 Training: 46%|████▋ | 8356/18018 [18:11:18<19:37:36, 7.31s/step, loss=1.452, acc=0.671, acc_len=1.59, thru=18.8, I=21.2, T=32.9, wait=2.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:41:32,279] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8355: forward=2750.1ms backward=1089.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:41:32,560] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:41:40,072] loop.py:45 INFO TRAIN_STEP step=8357 loss=1.487030 Training: 46%|████▋ | 8357/18018 [18:11:26<20:00:31, 7.46s/step, loss=1.487, acc=0.667, acc_len=1.57, thru=14.7, I=16.5, T=27.8, wait=3.2s, pool=60, epoch=1/2][2026-10-19 07:41:49,300] loop.py:45 INFO TRAIN_STEP step=8358 loss=1.588457 Training: 46%|████▋ | 8358/18018 [18:11:35<21:25:46, 7.99s/step, loss=1.588, acc=0.652, acc_len=1.53, thru=12.7, I=12.2, T=28.8, wait=4.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:41:54,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:41:55,844] loop.py:45 INFO TRAIN_STEP step=8359 loss=1.335869 Training: 46%|████▋ | 8359/18018 [18:11:41<20:16:07, 7.55s/step, loss=1.336, acc=0.692, acc_len=1.65, thru=19.0, I=19.8, T=33.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:41:56,025] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:02,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:42:02,852] loop.py:380 INFO TIMING step=8360: step=4.206s data=1.960s compute=4.176s [fwd=2.933s bwd=1.229s opt=0.003s] dispatch=2.792s [2026-10-19 07:42:02,853] loop.py:45 INFO TRAIN_STEP step=8360 loss=1.449550 Training: 46%|████▋ | 8360/18018 [18:11:48<19:49:38, 7.39s/step, loss=1.450, acc=0.676, acc_len=1.60, thru=15.4, I=23.2, T=30.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:02,888] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:12,983] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:42:14,837] loop.py:45 INFO TRAIN_STEP step=8361 loss=1.514261 Training: 46%|████▋ | 8361/18018 [18:12:00<23:31:19, 8.77s/step, loss=1.514, acc=0.665, acc_len=1.57, thru=10.4, I=11.8, T=28.8, wait=7.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:42:14,832] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8360: forward=3199.8ms backward=1193.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:14,994] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:21,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:42:23,486] loop.py:45 INFO TRAIN_STEP step=8362 loss=1.450696 Training: 46%|████▋ | 8362/18018 [18:12:09<23:25:19, 8.73s/step, loss=1.451, acc=0.679, acc_len=1.61, thru=15.1, I=18.6, T=30.1, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:23,518] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:29,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:42:29,757] loop.py:45 INFO TRAIN_STEP step=8363 loss=1.343862 Training: 46%|████▋ | 8363/18018 [18:12:15<21:26:21, 7.99s/step, loss=1.344, acc=0.695, acc_len=1.66, thru=17.6, I=23.2, T=36.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:29,888] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-19 07:42:36,206] loop.py:45 INFO TRAIN_STEP step=8364 loss=1.386981 Training: 46%|████▋ | 8364/18018 [18:12:22<20:11:37, 7.53s/step, loss=1.387, acc=0.686, acc_len=1.64, thru=19.7, I=20.4, T=32.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:36,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:36,667] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:43,136] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:42:43,248] loop.py:380 INFO TIMING step=8365: step=4.106s data=1.917s compute=4.072s [fwd=2.857s bwd=1.201s opt=0.003s] dispatch=2.929s [2026-10-19 07:42:43,248] loop.py:45 INFO TRAIN_STEP step=8365 loss=1.356782 Training: 46%|████▋ | 8365/18018 [18:12:29<19:47:55, 7.38s/step, loss=1.357, acc=0.692, acc_len=1.65, thru=17.9, I=19.5, T=31.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:43,637] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:50,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:42:52,426] loop.py:45 INFO TRAIN_STEP step=8366 loss=1.454558 Training: 46%|████▋ | 8366/18018 [18:12:38<21:14:31, 7.92s/step, loss=1.455, acc=0.679, acc_len=1.61, thru=14.2, I=15.4, T=25.2, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:42:52,421] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8365: forward=3484.9ms backward=1545.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:52,498] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:42:57,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:42:59,992] loop.py:45 INFO TRAIN_STEP step=8367 loss=1.361751 Training: 46%|████▋ | 8367/18018 [18:12:46<20:57:10, 7.82s/step, loss=1.362, acc=0.694, acc_len=1.65, thru=14.4, I=23.5, T=30.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:00,364] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:07,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:43:07,831] loop.py:45 INFO TRAIN_STEP step=8368 loss=1.385084 Training: 46%|████▋ | 8368/18018 [18:12:53<20:58:03, 7.82s/step, loss=1.385, acc=0.683, acc_len=1.63, thru=13.2, I=19.5, T=31.4, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:08,233] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:13,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:43:14,837] loop.py:45 INFO TRAIN_STEP step=8369 loss=1.544453 Training: 46%|████▋ | 8369/18018 [18:13:00<20:18:39, 7.58s/step, loss=1.544, acc=0.657, acc_len=1.54, thru=17.2, I=20.9, T=28.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:15,105] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:19,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:43:21,776] loop.py:380 INFO TIMING step=8370: step=4.374s data=1.890s compute=4.331s [fwd=2.824s bwd=1.492s opt=0.003s] dispatch=2.554s [2026-10-19 07:43:21,777] loop.py:45 INFO TRAIN_STEP step=8370 loss=1.338808 Training: 46%|████▋ | 8370/18018 [18:13:07<19:47:39, 7.39s/step, loss=1.339, acc=0.695, acc_len=1.67, thru=16.6, I=23.5, T=29.3, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:21,917] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:27,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:43:28,127] loop.py:45 INFO TRAIN_STEP step=8371 loss=1.402428 Training: 46%|████▋ | 8371/18018 [18:13:14<18:57:34, 7.08s/step, loss=1.402, acc=0.684, acc_len=1.63, thru=16.5, I=19.7, T=30.2, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:43:28,122] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8370: forward=2982.2ms backward=1207.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:28,335] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:43:35,513] loop.py:45 INFO TRAIN_STEP step=8372 loss=1.507461 Training: 46%|████▋ | 8372/18018 [18:13:21<19:12:32, 7.17s/step, loss=1.507, acc=0.665, acc_len=1.56, thru=17.5, I=18.5, T=30.6, wait=3.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:42,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:43:43,532] loop.py:45 INFO TRAIN_STEP step=8373 loss=1.359779 Training: 46%|████▋ | 8373/18018 [18:13:29<19:53:21, 7.42s/step, loss=1.360, acc=0.690, acc_len=1.65, thru=16.7, I=16.0, T=29.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:44,037] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:50,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:43:51,318] loop.py:45 INFO TRAIN_STEP step=8374 loss=1.360577 Training: 46%|████▋ | 8374/18018 [18:13:37<20:10:38, 7.53s/step, loss=1.361, acc=0.694, acc_len=1.67, thru=15.3, I=18.9, T=28.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:51,547] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:55,974] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:43:57,736] loop.py:380 INFO TIMING step=8375: step=3.959s data=1.796s compute=3.929s [fwd=2.819s bwd=1.095s opt=0.003s] dispatch=2.450s [2026-10-19 07:43:57,737] loop.py:45 INFO TRAIN_STEP step=8375 loss=1.298805 Training: 46%|████▋ | 8375/18018 [18:13:43<19:16:51, 7.20s/step, loss=1.299, acc=0.700, acc_len=1.68, thru=18.3, I=22.9, T=32.3, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:43:57,984] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:44:06,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:44:07,588] loop.py:45 INFO TRAIN_STEP step=8376 loss=1.360148 Training: 46%|████▋ | 8376/18018 [18:13:53<21:24:46, 7.99s/step, loss=1.360, acc=0.694, acc_len=1.67, thru=13.3, I=14.1, T=31.4, wait=5.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:44:07,584] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8375: forward=2803.5ms backward=1221.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:44:07,988] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:44:14,544] loop.py:45 INFO TRAIN_STEP step=8377 loss=1.403200 Training: 46%|████▋ | 8377/18018 [18:14:00<20:34:42, 7.68s/step, loss=1.403, acc=0.683, acc_len=1.62, thru=16.5, I=18.1, T=30.1, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:44:22,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:44:25,956] loop.py:45 INFO TRAIN_STEP step=8378 loss=1.404735 Training: 46%|████▋ | 8378/18018 [18:14:12<23:34:01, 8.80s/step, loss=1.405, acc=0.685, acc_len=1.63, thru=11.4, I=15.3, T=23.9, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:44:26,221] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:44:33,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:44:33,785] loop.py:45 INFO TRAIN_STEP step=8379 loss=1.435118 Training: 47%|████▋ | 8379/18018 [18:14:19<22:47:05, 8.51s/step, loss=1.435, acc=0.678, acc_len=1.61, thru=16.9, I=18.2, T=32.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:44:34,033] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:44:40,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:44:41,539] loop.py:380 INFO TIMING step=8380: step=4.293s data=2.111s compute=4.241s [fwd=2.947s bwd=1.279s opt=0.003s] dispatch=3.451s [2026-10-19 07:44:41,539] loop.py:45 INFO TRAIN_STEP step=8380 loss=1.423315 Training: 47%|████▋ | 8380/18018 [18:14:27<22:10:30, 8.28s/step, loss=1.423, acc=0.679, acc_len=1.61, thru=16.0, I=18.3, T=29.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:44:41,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:44:48,407] loop.py:45 INFO TRAIN_STEP step=8381 loss=1.391970 Training: 47%|████▋ | 8381/18018 [18:14:34<21:02:06, 7.86s/step, loss=1.392, acc=0.686, acc_len=1.63, thru=15.7, I=16.9, T=32.7, wait=2.9s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:44:48,403] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8380: forward=2753.2ms backward=1113.6ms [2026-10-19 07:44:58,671] loop.py:45 INFO TRAIN_STEP step=8382 loss=1.431071 Training: 47%|████▋ | 8382/18018 [18:14:44<22:58:03, 8.58s/step, loss=1.431, acc=0.685, acc_len=1.62, thru=12.6, I=13.3, T=30.9, wait=6.1s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:09,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:45:09,206] loop.py:45 INFO TRAIN_STEP step=8383 loss=1.478618 Training: 47%|████▋ | 8383/18018 [18:14:55<24:31:55, 9.17s/step, loss=1.479, acc=0.673, acc_len=1.59, thru=15.0, I=13.6, T=30.5, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:09,504] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:14,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:45:15,802] loop.py:45 INFO TRAIN_STEP step=8384 loss=1.493545 Training: 47%|████▋ | 8384/18018 [18:15:01<22:28:02, 8.40s/step, loss=1.494, acc=0.669, acc_len=1.58, thru=18.4, I=20.0, T=30.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:16,127] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:21,459] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:45:23,761] loop.py:380 INFO TIMING step=8385: step=4.365s data=1.954s compute=4.323s [fwd=2.994s bwd=1.314s opt=0.003s] dispatch=3.584s [2026-10-19 07:45:23,762] loop.py:45 INFO TRAIN_STEP step=8385 loss=1.333394 Training: 47%|████▋ | 8385/18018 [18:15:09<22:06:55, 8.26s/step, loss=1.333, acc=0.698, acc_len=1.67, thru=14.9, I=20.7, T=29.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:23,973] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:30,209] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:45:31,005] loop.py:45 INFO TRAIN_STEP step=8386 loss=1.440219 Training: 47%|████▋ | 8386/18018 [18:15:17<21:17:37, 7.96s/step, loss=1.440, acc=0.677, acc_len=1.61, thru=14.4, I=23.9, T=26.3, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:45:31,001] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8385: forward=3318.5ms backward=1482.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:31,214] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:38,544] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:45:39,397] loop.py:45 INFO TRAIN_STEP step=8387 loss=1.433167 Training: 47%|████▋ | 8387/18018 [18:15:25<21:38:19, 8.09s/step, loss=1.433, acc=0.679, acc_len=1.61, thru=14.2, I=14.9, T=31.4, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:39,565] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:46,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:45:48,332] loop.py:45 INFO TRAIN_STEP step=8388 loss=1.589715 Training: 47%|████▋ | 8388/18018 [18:15:34<22:18:52, 8.34s/step, loss=1.590, acc=0.652, acc_len=1.51, thru=13.5, I=17.2, T=29.4, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:48,696] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:54,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:45:55,682] loop.py:45 INFO TRAIN_STEP step=8389 loss=1.404412 Training: 47%|████▋ | 8389/18018 [18:15:41<21:30:57, 8.04s/step, loss=1.404, acc=0.685, acc_len=1.63, thru=16.1, I=21.6, T=29.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:45:56,172] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:46:02,231] loop.py:380 INFO TIMING step=8390: step=3.585s data=1.433s compute=3.555s [fwd=2.483s bwd=1.057s opt=0.003s] dispatch=2.957s [2026-10-19 07:46:02,231] loop.py:45 INFO TRAIN_STEP step=8390 loss=1.356368 Training: 47%|████▋ | 8390/18018 [18:15:48<20:18:51, 7.60s/step, loss=1.356, acc=0.693, acc_len=1.66, thru=16.7, I=19.5, T=35.7, wait=3.0s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:13,551] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:46:14,202] loop.py:45 INFO TRAIN_STEP step=8391 loss=1.424664 Training: 47%|████▋ | 8391/18018 [18:16:00<23:49:26, 8.91s/step, loss=1.425, acc=0.681, acc_len=1.62, thru=12.5, I=12.6, T=27.1, wait=7.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:46:14,198] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8390: forward=3330.6ms backward=1323.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:14,556] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:46:23,544] loop.py:45 INFO TRAIN_STEP step=8392 loss=1.412890 Training: 47%|████▋ | 8392/18018 [18:16:09<24:10:11, 9.04s/step, loss=1.413, acc=0.689, acc_len=1.64, thru=13.8, I=12.9, T=25.1, wait=4.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:30,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:46:31,275] loop.py:45 INFO TRAIN_STEP step=8393 loss=1.420937 Training: 47%|████▋ | 8393/18018 [18:16:17<23:06:55, 8.65s/step, loss=1.421, acc=0.681, acc_len=1.62, thru=16.8, I=14.9, T=29.9, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:31,441] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:36,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:46:37,625] loop.py:45 INFO TRAIN_STEP step=8394 loss=1.527371 Training: 47%|████▋ | 8394/18018 [18:16:23<21:16:18, 7.96s/step, loss=1.527, acc=0.667, acc_len=1.57, thru=18.5, I=20.6, T=31.6, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:38,087] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:44,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:46:45,253] loop.py:380 INFO TIMING step=8395: step=4.568s data=2.120s compute=4.539s [fwd=3.155s bwd=1.369s opt=0.003s] dispatch=3.052s [2026-10-19 07:46:45,254] loop.py:45 INFO TRAIN_STEP step=8395 loss=1.411363 Training: 47%|████▋ | 8395/18018 [18:16:31<21:00:22, 7.86s/step, loss=1.411, acc=0.682, acc_len=1.62, thru=15.9, I=19.9, T=28.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:45,436] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:51,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:46:51,939] loop.py:45 INFO TRAIN_STEP step=8396 loss=1.318103 Training: 47%|████▋ | 8396/18018 [18:16:38<20:03:54, 7.51s/step, loss=1.318, acc=0.698, acc_len=1.68, thru=17.8, I=19.1, T=31.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:51,975] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (TrainerActor pid=9310) [2026-10-19 07:46:51,934] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8395: forward=2851.8ms backward=1236.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:58,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:46:59,566] loop.py:45 INFO TRAIN_STEP step=8397 loss=1.365468 Training: 47%|████▋ | 8397/18018 [18:16:45<20:09:26, 7.54s/step, loss=1.365, acc=0.687, acc_len=1.64, thru=16.6, I=19.2, T=31.1, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:46:59,850] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:47:04,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:47:05,339] loop.py:45 INFO TRAIN_STEP step=8398 loss=1.508844 Training: 47%|████▋ | 8398/18018 [18:16:51<18:44:17, 7.01s/step, loss=1.509, acc=0.665, acc_len=1.57, thru=20.1, I=24.3, T=37.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:47:05,679] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 07:47:12,698] loop.py:45 INFO TRAIN_STEP step=8399 loss=1.195996 Training: 47%|████▋ | 8399/18018 [18:16:58<19:00:51, 7.12s/step, loss=1.196, acc=0.717, acc_len=1.74, thru=16.9, I=18.3, T=32.8, wait=3.4s, pool=52, epoch=1/2][2026-10-19 07:47:23,986] loop.py:380 INFO TIMING step=8400: step=4.463s data=1.986s compute=4.434s [fwd=3.119s bwd=1.302s opt=0.003s] dispatch=6.814s [2026-10-19 07:47:23,986] loop.py:45 INFO TRAIN_STEP step=8400 loss=1.341365 Training: 47%|████▋ | 8400/18018 [18:17:10<22:21:14, 8.37s/step, loss=1.341, acc=0.695, acc_len=1.67, thru=9.7, I=8.9, T=28.7, wait=6.8s, pool=40, epoch=1/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:47:31,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:47:32,640] loop.py:45 INFO TRAIN_STEP step=8401 loss=1.499784 Training: 47%|████▋ | 8401/18018 [18:17:18<22:34:55, 8.45s/step, loss=1.500, acc=0.671, acc_len=1.59, thru=17.8, I=17.8, T=29.9, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:47:32,635] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8400: forward=2932.0ms backward=1308.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:47:33,081] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:47:37,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:47:39,388] loop.py:45 INFO TRAIN_STEP step=8402 loss=1.415896 Training: 47%|████▋ | 8402/18018 [18:17:25<21:12:54, 7.94s/step, loss=1.416, acc=0.683, acc_len=1.62, thru=17.0, I=25.9, T=32.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:47:39,797] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:47:46,954] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:47:47,847] loop.py:45 INFO TRAIN_STEP step=8403 loss=1.342910 Training: 47%|████▋ | 8403/18018 [18:17:33<21:37:35, 8.10s/step, loss=1.343, acc=0.693, acc_len=1.66, thru=15.9, I=18.0, T=31.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:47:47,960] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 07:47:54,263] loop.py:45 INFO TRAIN_STEP step=8404 loss=1.345470 Training: 47%|████▋ | 8404/18018 [18:17:40<20:16:35, 7.59s/step, loss=1.345, acc=0.694, acc_len=1.67, thru=18.6, I=19.3, T=33.7, wait=2.6s, pool=60, epoch=1/2][2026-10-19 07:48:05,698] loop.py:380 INFO TIMING step=8405: step=4.948s data=2.258s compute=4.918s [fwd=3.578s bwd=1.325s opt=0.003s] dispatch=6.477s [2026-10-19 07:48:05,699] loop.py:45 INFO TRAIN_STEP step=8405 loss=1.463908 Training: 47%|████▋ | 8405/18018 [18:17:51<23:21:09, 8.75s/step, loss=1.464, acc=0.676, acc_len=1.60, thru=12.0, I=10.8, T=25.9, wait=6.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:48:13,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:48:16,243] loop.py:45 INFO TRAIN_STEP step=8406 loss=1.356631 Training: 47%|████▋ | 8406/18018 [18:18:02<24:47:30, 9.29s/step, loss=1.357, acc=0.691, acc_len=1.65, thru=12.6, I=15.7, T=24.9, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:48:16,238] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8405: forward=3669.5ms backward=1410.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:48:16,739] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:48:22,022] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:48:23,999] loop.py:45 INFO TRAIN_STEP step=8407 loss=1.251204 Training: 47%|████▋ | 8407/18018 [18:18:09<22:45:52, 8.53s/step, loss=1.251, acc=0.706, acc_len=1.70, thru=14.6, I=22.6, T=32.3, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:48:23,036] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 07:48:29,646] loop.py:45 INFO TRAIN_STEP step=8408 loss=1.423366 Training: 47%|████▋ | 8408/18018 [18:18:15<21:15:15, 7.96s/step, loss=1.423, acc=0.681, acc_len=1.61, thru=17.3, I=19.1, T=30.3, wait=2.4s, pool=56, epoch=1/2][2026-10-19 07:48:37,763] loop.py:45 INFO TRAIN_STEP step=8409 loss=1.482976 Training: 47%|████▋ | 8409/18018 [18:18:23<21:22:36, 8.01s/step, loss=1.483, acc=0.671, acc_len=1.58, thru=12.9, I=15.7, T=27.9, wait=3.5s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:48:46,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:48:47,364] loop.py:380 INFO TIMING step=8410: step=4.708s data=2.298s compute=4.683s [fwd=3.293s bwd=1.377s opt=0.003s] dispatch=4.885s [2026-10-19 07:48:47,364] loop.py:45 INFO TRAIN_STEP step=8410 loss=1.429678 Training: 47%|████▋ | 8410/18018 [18:18:33<22:38:57, 8.49s/step, loss=1.430, acc=0.679, acc_len=1.61, thru=17.1, I=15.1, T=27.2, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:48:47,704] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:48:53,788] loop.py:45 INFO TRAIN_STEP step=8411 loss=1.450349 Training: 47%|████▋ | 8411/18018 [18:18:39<20:59:39, 7.87s/step, loss=1.450, acc=0.679, acc_len=1.61, thru=17.6, I=21.9, T=32.8, wait=2.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:48:53,783] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8410: forward=2682.9ms backward=1169.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:00,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:49:01,960] loop.py:45 INFO TRAIN_STEP step=8412 loss=1.387444 Training: 47%|████▋ | 8412/18018 [18:18:48<21:14:16, 7.96s/step, loss=1.387, acc=0.688, acc_len=1.64, thru=15.4, I=16.6, T=33.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:02,202] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:49:10,473] loop.py:45 INFO TRAIN_STEP step=8413 loss=1.449175 Training: 47%|████▋ | 8413/18018 [18:18:56<21:40:47, 8.13s/step, loss=1.449, acc=0.680, acc_len=1.61, thru=13.1, I=15.2, T=30.4, wait=4.3s, pool=60, epoch=1/2][2026-10-19 07:49:17,390] loop.py:45 INFO TRAIN_STEP step=8414 loss=1.342527 Training: 47%|████▋ | 8414/18018 [18:19:03<20:42:31, 7.76s/step, loss=1.343, acc=0.697, acc_len=1.68, thru=16.9, I=15.5, T=32.1, wait=2.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:22,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:49:23,631] loop.py:380 INFO TIMING step=8415: step=3.807s data=1.671s compute=3.778s [fwd=2.730s bwd=1.035s opt=0.003s] dispatch=2.425s [2026-10-19 07:49:23,631] loop.py:45 INFO TRAIN_STEP step=8415 loss=1.446472 Training: 47%|████▋ | 8415/18018 [18:19:09<19:29:28, 7.31s/step, loss=1.446, acc=0.680, acc_len=1.61, thru=19.2, I=22.4, T=33.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:24,048] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:31,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:49:32,248] loop.py:45 INFO TRAIN_STEP step=8416 loss=1.494855 Training: 47%|████▋ | 8416/18018 [18:19:18<20:32:11, 7.70s/step, loss=1.495, acc=0.669, acc_len=1.58, thru=15.6, I=15.8, T=30.0, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:49:32,243] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8415: forward=2985.7ms backward=1230.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:32,332] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:38,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:49:40,039] loop.py:45 INFO TRAIN_STEP step=8417 loss=1.416407 Training: 47%|████▋ | 8417/18018 [18:19:26<20:36:22, 7.73s/step, loss=1.416, acc=0.682, acc_len=1.62, thru=16.2, I=20.6, T=29.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:40,279] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:46,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:49:48,383] loop.py:45 INFO TRAIN_STEP step=8418 loss=1.340122 Training: 47%|████▋ | 8418/18018 [18:19:34<21:05:56, 7.91s/step, loss=1.340, acc=0.695, acc_len=1.67, thru=15.8, I=18.5, T=26.4, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:48,581] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:55,773] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:49:56,964] loop.py:45 INFO TRAIN_STEP step=8419 loss=1.456816 Training: 47%|████▋ | 8419/18018 [18:19:43<21:37:55, 8.11s/step, loss=1.457, acc=0.676, acc_len=1.60, thru=15.3, I=17.7, T=32.8, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:49:57,283] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:50:02,549] loop.py:380 INFO TIMING step=8420: step=3.447s data=1.463s compute=3.418s [fwd=2.419s bwd=0.986s opt=0.003s] dispatch=2.128s [2026-10-19 07:50:02,549] loop.py:45 INFO TRAIN_STEP step=8420 loss=1.413056 Training: 47%|████▋ | 8420/18018 [18:19:48<19:36:31, 7.35s/step, loss=1.413, acc=0.683, acc_len=1.63, thru=19.6, I=24.0, T=37.1, wait=2.1s, pool=52, epoch=1/2][2026-10-19 07:50:09,852] loop.py:45 INFO TRAIN_STEP step=8421 loss=1.517168 Training: 47%|████▋ | 8421/18018 [18:19:55<19:33:46, 7.34s/step, loss=1.517, acc=0.663, acc_len=1.56, thru=17.0, I=17.0, T=33.6, wait=3.5s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:50:09,848] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8420: forward=2744.2ms backward=1024.6ms [2026-10-19 07:50:18,237] loop.py:45 INFO TRAIN_STEP step=8422 loss=1.405069 Training: 47%|████▋ | 8422/18018 [18:20:04<20:24:00, 7.65s/step, loss=1.405, acc=0.685, acc_len=1.63, thru=15.3, I=16.9, T=29.1, wait=4.0s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:50:28,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:50:28,281] loop.py:45 INFO TRAIN_STEP step=8423 loss=1.392508 Training: 47%|████▋ | 8423/18018 [18:20:14<22:18:26, 8.37s/step, loss=1.393, acc=0.688, acc_len=1.65, thru=14.9, I=15.2, T=27.7, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:50:28,580] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:50:36,480] loop.py:45 INFO TRAIN_STEP step=8424 loss=1.253501 Training: 47%|████▋ | 8424/18018 [18:20:22<22:10:05, 8.32s/step, loss=1.254, acc=0.707, acc_len=1.71, thru=14.4, I=15.7, T=31.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:50:36,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:50:36,981] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:50:43,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:50:44,164] loop.py:380 INFO TIMING step=8425: step=4.109s data=1.651s compute=4.067s [fwd=2.747s bwd=1.305s opt=0.003s] dispatch=3.567s [2026-10-19 07:50:44,164] loop.py:45 INFO TRAIN_STEP step=8425 loss=1.380844 Training: 47%|████▋ | 8425/18018 [18:20:30<21:39:40, 8.13s/step, loss=1.381, acc=0.690, acc_len=1.65, thru=15.2, I=16.4, T=31.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:50:44,271] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:50:49,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:50:50,293] loop.py:45 INFO TRAIN_STEP step=8426 loss=1.464129 Training: 47%|████▋ | 8426/18018 [18:20:36<20:03:32, 7.53s/step, loss=1.464, acc=0.676, acc_len=1.61, thru=18.8, I=22.5, T=34.8, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:50:50,288] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8425: forward=2575.8ms backward=1061.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:50:50,758] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:50:59,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:51:00,100] loop.py:45 INFO TRAIN_STEP step=8427 loss=1.515108 Training: 47%|████▋ | 8427/18018 [18:20:46<21:52:44, 8.21s/step, loss=1.515, acc=0.664, acc_len=1.56, thru=13.7, I=15.1, T=32.0, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:51:00,248] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:51:06,666] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:51:08,099] loop.py:45 INFO TRAIN_STEP step=8428 loss=1.408744 Training: 47%|████▋ | 8428/18018 [18:20:54<21:42:23, 8.15s/step, loss=1.409, acc=0.682, acc_len=1.63, thru=16.1, I=19.8, T=29.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:51:08,186] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:51:14,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:51:16,134] loop.py:45 INFO TRAIN_STEP step=8429 loss=1.441115 Training: 47%|████▋ | 8429/18018 [18:21:02<21:36:44, 8.11s/step, loss=1.441, acc=0.680, acc_len=1.61, thru=14.5, I=20.9, T=29.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:51:16,614] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:51:24,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:51:25,675] loop.py:380 INFO TIMING step=8430: step=4.662s data=2.255s compute=4.588s [fwd=3.164s bwd=1.409s opt=0.003s] dispatch=4.870s [2026-10-19 07:51:25,675] loop.py:45 INFO TRAIN_STEP step=8430 loss=1.523261 Training: 47%|████▋ | 8430/18018 [18:21:11<22:45:00, 8.54s/step, loss=1.523, acc=0.667, acc_len=1.57, thru=14.1, I=17.6, T=27.5, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:51:26,160] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:51:33,571] loop.py:45 INFO TRAIN_STEP step=8431 loss=1.230035 Training: 47%|████▋ | 8431/18018 [18:21:19<22:13:58, 8.35s/step, loss=1.230, acc=0.715, acc_len=1.73, thru=13.5, I=14.3, T=30.4, wait=3.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:51:33,567] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8430: forward=2943.5ms backward=1218.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:51:43,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:51:44,197] loop.py:45 INFO TRAIN_STEP step=8432 loss=1.457073 Training: 47%|████▋ | 8432/18018 [18:21:30<24:03:00, 9.03s/step, loss=1.457, acc=0.680, acc_len=1.62, thru=13.8, I=13.1, T=32.6, wait=6.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:51:44,354] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:51:50,761] loop.py:45 INFO TRAIN_STEP step=8433 loss=1.381536 Training: 47%|████▋ | 8433/18018 [18:21:36<22:04:29, 8.29s/step, loss=1.382, acc=0.685, acc_len=1.63, thru=18.5, I=20.1, T=34.7, wait=2.9s, pool=60, epoch=1/2][2026-10-19 07:51:59,453] loop.py:45 INFO TRAIN_STEP step=8434 loss=1.436146 Training: 47%|████▋ | 8434/18018 [18:21:45<22:23:39, 8.41s/step, loss=1.436, acc=0.678, acc_len=1.61, thru=14.6, I=14.1, T=25.7, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:06,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:52:07,416] loop.py:380 INFO TIMING step=8435: step=4.229s data=2.145s compute=4.181s [fwd=3.121s bwd=1.046s opt=0.003s] dispatch=3.723s [2026-10-19 07:52:07,416] loop.py:45 INFO TRAIN_STEP step=8435 loss=1.444600 Training: 47%|████▋ | 8435/18018 [18:21:53<22:01:59, 8.28s/step, loss=1.445, acc=0.676, acc_len=1.60, thru=16.9, I=16.4, T=30.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:07,893] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:13,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:52:15,075] loop.py:45 INFO TRAIN_STEP step=8436 loss=1.513558 Training: 47%|████▋ | 8436/18018 [18:22:01<21:32:05, 8.09s/step, loss=1.514, acc=0.662, acc_len=1.56, thru=15.8, I=22.1, T=26.7, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:52:15,071] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8435: forward=3605.3ms backward=1141.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:15,391] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:21,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:52:22,242] loop.py:45 INFO TRAIN_STEP step=8437 loss=1.472957 Training: 47%|████▋ | 8437/18018 [18:22:08<20:47:49, 7.81s/step, loss=1.473, acc=0.669, acc_len=1.58, thru=16.0, I=22.5, T=30.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:22,733] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:27,773] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:52:29,738] loop.py:45 INFO TRAIN_STEP step=8438 loss=1.476073 Training: 47%|████▋ | 8438/18018 [18:22:15<20:32:29, 7.72s/step, loss=1.476, acc=0.673, acc_len=1.59, thru=15.5, I=20.3, T=29.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:29,797] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:33,983] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:52:36,560] loop.py:45 INFO TRAIN_STEP step=8439 loss=1.322967 Training: 47%|████▋ | 8439/18018 [18:22:22<19:49:20, 7.45s/step, loss=1.323, acc=0.699, acc_len=1.67, thru=16.7, I=29.0, T=27.9, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:37,004] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:42,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:52:44,183] loop.py:380 INFO TIMING step=8440: step=4.693s data=2.325s compute=4.662s [fwd=3.266s bwd=1.382s opt=0.003s] dispatch=2.921s [2026-10-19 07:52:44,184] loop.py:45 INFO TRAIN_STEP step=8440 loss=1.398135 Training: 47%|████▋ | 8440/18018 [18:22:30<19:57:30, 7.50s/step, loss=1.398, acc=0.685, acc_len=1.64, thru=17.8, I=20.8, T=27.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:44,330] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:51,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:52:52,833] loop.py:45 INFO TRAIN_STEP step=8441 loss=1.307149 Training: 47%|████▋ | 8441/18018 [18:22:38<20:52:24, 7.85s/step, loss=1.307, acc=0.697, acc_len=1.67, thru=15.1, I=17.1, T=25.8, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:52:52,828] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8440: forward=3415.1ms backward=1481.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:53,115] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:52:59,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:53:00,810] loop.py:45 INFO TRAIN_STEP step=8442 loss=1.445357 Training: 47%|████▋ | 8442/18018 [18:22:46<20:58:29, 7.89s/step, loss=1.445, acc=0.678, acc_len=1.61, thru=14.0, I=20.0, T=31.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:00,900] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 [2026-10-19 07:53:07,903] loop.py:45 INFO TRAIN_STEP step=8443 loss=1.398802 Training: 47%|████▋ | 8443/18018 [18:22:53<20:20:23, 7.65s/step, loss=1.399, acc=0.680, acc_len=1.62, thru=11.7, I=16.2, T=32.2, wait=3.1s, pool=24, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:17,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:53:18,419] loop.py:45 INFO TRAIN_STEP step=8444 loss=1.482544 Training: 47%|████▋ | 8444/18018 [18:23:04<22:37:43, 8.51s/step, loss=1.483, acc=0.673, acc_len=1.59, thru=16.1, I=16.4, T=31.6, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:18,568] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:24,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:53:24,716] loop.py:380 INFO TIMING step=8445: step=4.247s data=2.031s compute=4.207s [fwd=2.958s bwd=1.234s opt=0.003s] dispatch=2.037s [2026-10-19 07:53:24,717] loop.py:45 INFO TRAIN_STEP step=8445 loss=1.512352 Training: 47%|████▋ | 8445/18018 [18:23:10<20:51:43, 7.85s/step, loss=1.512, acc=0.663, acc_len=1.56, thru=18.8, I=21.8, T=30.1, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:25,058] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:30,614] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:53:31,770] loop.py:45 INFO TRAIN_STEP step=8446 loss=1.408052 Training: 47%|████▋ | 8446/18018 [18:23:17<20:13:42, 7.61s/step, loss=1.408, acc=0.683, acc_len=1.63, thru=17.5, I=20.4, T=33.0, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:53:31,765] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8445: forward=2675.3ms backward=1144.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:32,122] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:39,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:53:41,622] loop.py:45 INFO TRAIN_STEP step=8447 loss=1.397038 Training: 47%|████▋ | 8447/18018 [18:23:27<22:00:50, 8.28s/step, loss=1.397, acc=0.682, acc_len=1.62, thru=13.5, I=16.5, T=28.8, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:41,880] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:49,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:53:49,625] loop.py:45 INFO TRAIN_STEP step=8448 loss=1.287896 Training: 47%|████▋ | 8448/18018 [18:23:35<21:47:24, 8.20s/step, loss=1.288, acc=0.704, acc_len=1.69, thru=16.5, I=17.6, T=33.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:53:49,886] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:53:59,505] loop.py:45 INFO TRAIN_STEP step=8449 loss=1.305878 Training: 47%|████▋ | 8449/18018 [18:23:45<23:07:59, 8.70s/step, loss=1.306, acc=0.705, acc_len=1.70, thru=13.0, I=13.0, T=26.9, wait=5.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:07,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:54:08,229] loop.py:380 INFO TIMING step=8450: step=4.572s data=2.248s compute=4.523s [fwd=3.145s bwd=1.362s opt=0.003s] dispatch=4.141s [2026-10-19 07:54:08,230] loop.py:45 INFO TRAIN_STEP step=8450 loss=1.445863 Training: 47%|████▋ | 8450/18018 [18:23:54<23:08:47, 8.71s/step, loss=1.446, acc=0.677, acc_len=1.61, thru=15.7, I=15.7, T=28.0, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:08,709] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:14,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:54:15,938] loop.py:45 INFO TRAIN_STEP step=8451 loss=1.383304 (TrainerActor pid=9310) [2026-10-19 07:54:15,933] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8450: forward=2788.2ms backward=1151.5ms Training: 47%|████▋ | 8451/18018 [18:24:02<22:20:48, 8.41s/step, loss=1.383, acc=0.690, acc_len=1.65, thru=16.5, I=19.8, T=32.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:15,999] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 07:54:22,698] loop.py:45 INFO TRAIN_STEP step=8452 loss=1.435078 Training: 47%|████▋ | 8452/18018 [18:24:08<21:01:46, 7.91s/step, loss=1.435, acc=0.678, acc_len=1.61, thru=18.2, I=21.9, T=31.8, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:22,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:23,204] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:30,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:54:31,811] loop.py:45 INFO TRAIN_STEP step=8453 loss=1.375959 Training: 47%|████▋ | 8453/18018 [18:24:17<21:58:53, 8.27s/step, loss=1.376, acc=0.688, acc_len=1.64, thru=14.5, I=14.9, T=25.9, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:32,298] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:54:39,864] loop.py:45 INFO TRAIN_STEP step=8454 loss=1.408847 Training: 47%|████▋ | 8454/18018 [18:24:25<21:48:19, 8.21s/step, loss=1.409, acc=0.681, acc_len=1.61, thru=13.3, I=18.8, T=29.6, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:45,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:54:46,715] loop.py:380 INFO TIMING step=8455: step=4.376s data=1.953s compute=4.340s [fwd=2.993s bwd=1.331s opt=0.003s] dispatch=2.463s [2026-10-19 07:54:46,715] loop.py:45 INFO TRAIN_STEP step=8455 loss=1.555786 Training: 47%|████▋ | 8455/18018 [18:24:32<20:43:14, 7.80s/step, loss=1.556, acc=0.657, acc_len=1.54, thru=18.5, I=19.8, T=29.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:47,120] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:54,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:54:54,537] loop.py:45 INFO TRAIN_STEP step=8456 loss=1.357896 Training: 47%|████▋ | 8456/18018 [18:24:40<20:44:13, 7.81s/step, loss=1.358, acc=0.690, acc_len=1.65, thru=14.4, I=18.5, T=30.8, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:54:54,533] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8455: forward=2934.3ms backward=1177.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:54:54,616] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-19 07:55:02,333] loop.py:45 INFO TRAIN_STEP step=8457 loss=1.437123 Training: 47%|████▋ | 8457/18018 [18:24:48<20:43:31, 7.80s/step, loss=1.437, acc=0.675, acc_len=1.59, thru=14.6, I=14.0, T=34.2, wait=4.0s, pool=56, epoch=1/2][2026-10-19 07:55:12,844] loop.py:45 INFO TRAIN_STEP step=8458 loss=1.413061 Training: 47%|████▋ | 8458/18018 [18:24:58<22:52:39, 8.62s/step, loss=1.413, acc=0.678, acc_len=1.61, thru=13.2, I=12.3, T=31.1, wait=6.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:55:19,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:55:21,454] loop.py:45 INFO TRAIN_STEP step=8459 loss=1.461656 Training: 47%|████▋ | 8459/18018 [18:25:07<22:52:26, 8.61s/step, loss=1.462, acc=0.673, acc_len=1.59, thru=16.0, I=16.7, T=26.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:55:21,674] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:55:28,594] loop.py:380 INFO TIMING step=8460: step=3.748s data=1.535s compute=3.712s [fwd=2.547s bwd=1.151s opt=0.003s] dispatch=3.382s [2026-10-19 07:55:28,594] loop.py:45 INFO TRAIN_STEP step=8460 loss=1.370293 Training: 47%|████▋ | 8460/18018 [18:25:14<21:41:44, 8.17s/step, loss=1.370, acc=0.688, acc_len=1.64, thru=15.3, I=18.4, T=34.2, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:55:33,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:55:35,019] loop.py:45 INFO TRAIN_STEP step=8461 loss=1.413344 Training: 47%|████▋ | 8461/18018 [18:25:21<20:18:07, 7.65s/step, loss=1.413, acc=0.681, acc_len=1.62, thru=19.4, I=21.4, T=34.9, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:55:35,015] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8460: forward=2594.3ms backward=1035.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:55:35,510] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:55:40,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:55:41,673] loop.py:45 INFO TRAIN_STEP step=8462 loss=1.461595 Training: 47%|████▋ | 8462/18018 [18:25:27<19:30:38, 7.35s/step, loss=1.462, acc=0.676, acc_len=1.61, thru=19.0, I=25.7, T=34.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:55:41,814] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:55:47,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:55:48,948] loop.py:45 INFO TRAIN_STEP step=8463 loss=1.422966 Training: 47%|████▋ | 8463/18018 [18:25:35<19:26:56, 7.33s/step, loss=1.423, acc=0.681, acc_len=1.62, thru=16.7, I=21.3, T=28.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:55:49,165] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:55:56,983] loop.py:45 INFO TRAIN_STEP step=8464 loss=1.434440 Training: 47%|████▋ | 8464/18018 [18:25:43<20:00:29, 7.54s/step, loss=1.434, acc=0.680, acc_len=1.61, thru=12.8, I=17.2, T=28.5, wait=3.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:06,308] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:56:06,678] loop.py:380 INFO TIMING step=8465: step=4.554s data=1.969s compute=4.504s [fwd=3.100s bwd=1.388s opt=0.003s] dispatch=5.133s [2026-10-19 07:56:06,678] loop.py:45 INFO TRAIN_STEP step=8465 loss=1.425221 Training: 47%|████▋ | 8465/18018 [18:25:52<21:43:24, 8.19s/step, loss=1.425, acc=0.684, acc_len=1.63, thru=15.6, I=14.8, T=28.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:06,814] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:14,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:56:14,392] loop.py:45 INFO TRAIN_STEP step=8466 loss=1.438806 Training: 47%|████▋ | 8466/18018 [18:26:00<21:20:36, 8.04s/step, loss=1.439, acc=0.680, acc_len=1.61, thru=15.9, I=14.7, T=30.8, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:56:14,387] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8465: forward=2903.3ms backward=1198.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:14,835] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:22,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:56:23,775] loop.py:45 INFO TRAIN_STEP step=8467 loss=1.333418 Training: 47%|████▋ | 8467/18018 [18:26:09<22:24:33, 8.45s/step, loss=1.333, acc=0.697, acc_len=1.67, thru=14.2, I=17.9, T=29.1, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:24,142] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:56:31,398] loop.py:45 INFO TRAIN_STEP step=8468 loss=1.496998 Training: 47%|████▋ | 8468/18018 [18:26:17<21:44:56, 8.20s/step, loss=1.497, acc=0.668, acc_len=1.57, thru=14.6, I=17.4, T=33.1, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:37,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:56:38,241] loop.py:45 INFO TRAIN_STEP step=8469 loss=1.485727 Training: 47%|████▋ | 8469/18018 [18:26:24<20:40:13, 7.79s/step, loss=1.486, acc=0.670, acc_len=1.58, thru=20.0, I=19.1, T=33.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:38,513] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:43,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:56:45,190] loop.py:380 INFO TIMING step=8470: step=4.224s data=1.975s compute=4.193s [fwd=2.918s bwd=1.259s opt=0.003s] dispatch=2.715s [2026-10-19 07:56:45,190] loop.py:45 INFO TRAIN_STEP step=8470 loss=1.492258 Training: 47%|████▋ | 8470/18018 [18:26:31<19:59:48, 7.54s/step, loss=1.492, acc=0.667, acc_len=1.57, thru=17.6, I=26.5, T=30.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:45,313] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:51,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:56:51,669] loop.py:45 INFO TRAIN_STEP step=8471 loss=1.382237 Training: 47%|████▋ | 8471/18018 [18:26:37<19:09:02, 7.22s/step, loss=1.382, acc=0.688, acc_len=1.64, thru=18.0, I=22.3, T=34.7, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:56:51,665] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8470: forward=2584.8ms backward=1057.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:51,872] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:56,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:56:58,775] loop.py:45 INFO TRAIN_STEP step=8472 loss=1.441436 Training: 47%|████▋ | 8472/18018 [18:26:44<19:03:21, 7.19s/step, loss=1.441, acc=0.679, acc_len=1.61, thru=17.5, I=20.9, T=32.3, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:56:58,913] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:04,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:57:05,581] loop.py:45 INFO TRAIN_STEP step=8473 loss=1.404573 Training: 47%|████▋ | 8473/18018 [18:26:51<18:45:16, 7.07s/step, loss=1.405, acc=0.683, acc_len=1.62, thru=18.2, I=23.2, T=31.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:05,821] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:14,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:57:15,775] loop.py:45 INFO TRAIN_STEP step=8474 loss=1.535917 Training: 47%|████▋ | 8474/18018 [18:27:01<21:13:51, 8.01s/step, loss=1.536, acc=0.663, acc_len=1.55, thru=12.9, I=13.6, T=24.4, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:16,266] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:22,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:57:24,076] loop.py:380 INFO TIMING step=8475: step=5.015s data=2.352s compute=4.984s [fwd=3.490s bwd=1.479s opt=0.003s] dispatch=3.278s [2026-10-19 07:57:24,076] loop.py:45 INFO TRAIN_STEP step=8475 loss=1.400691 Training: 47%|████▋ | 8475/18018 [18:27:10<21:27:39, 8.10s/step, loss=1.401, acc=0.684, acc_len=1.63, thru=14.5, I=18.3, T=25.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:24,383] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:33,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:57:33,694] loop.py:45 INFO TRAIN_STEP step=8476 loss=1.421040 Training: 47%|████▋ | 8476/18018 [18:27:19<22:40:15, 8.55s/step, loss=1.421, acc=0.677, acc_len=1.61, thru=13.8, I=14.4, T=27.9, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:57:33,690] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8475: forward=3088.7ms backward=1439.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:34,168] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:41,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:57:42,124] loop.py:45 INFO TRAIN_STEP step=8477 loss=1.339361 Training: 47%|████▋ | 8477/18018 [18:27:28<22:34:13, 8.52s/step, loss=1.339, acc=0.698, acc_len=1.67, thru=15.4, I=17.4, T=27.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:42,223] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-19 07:57:48,610] loop.py:45 INFO TRAIN_STEP step=8478 loss=1.358907 Training: 47%|████▋ | 8478/18018 [18:27:34<20:57:12, 7.91s/step, loss=1.359, acc=0.694, acc_len=1.67, thru=17.6, I=17.4, T=32.9, wait=2.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:54,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:57:55,504] loop.py:45 INFO TRAIN_STEP step=8479 loss=1.319482 Training: 47%|████▋ | 8479/18018 [18:27:41<20:08:47, 7.60s/step, loss=1.319, acc=0.695, acc_len=1.67, thru=18.1, I=19.0, T=31.2, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:57:55,884] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:03,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:58:03,960] loop.py:380 INFO TIMING step=8480: step=4.372s data=2.097s compute=4.333s [fwd=2.981s bwd=1.337s opt=0.003s] dispatch=4.073s [2026-10-19 07:58:03,960] loop.py:45 INFO TRAIN_STEP step=8480 loss=1.325652 Training: 47%|████▋ | 8480/18018 [18:27:50<20:49:13, 7.86s/step, loss=1.326, acc=0.697, acc_len=1.67, thru=14.5, I=16.2, T=29.3, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:04,347] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:09,940] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:58:11,664] loop.py:45 INFO TRAIN_STEP step=8481 loss=1.303034 Training: 47%|████▋ | 8481/18018 [18:27:57<20:41:49, 7.81s/step, loss=1.303, acc=0.702, acc_len=1.69, thru=17.4, I=21.0, T=30.9, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:58:11,660] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8480: forward=2913.6ms backward=1185.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:11,955] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 07:58:19,998] loop.py:45 INFO TRAIN_STEP step=8482 loss=1.373342 Training: 47%|████▋ | 8482/18018 [18:28:06<21:06:34, 7.97s/step, loss=1.373, acc=0.689, acc_len=1.64, thru=14.9, I=16.4, T=28.6, wait=3.8s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:29,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:58:29,398] loop.py:45 INFO TRAIN_STEP step=8483 loss=1.439133 Training: 47%|████▋ | 8483/18018 [18:28:15<22:14:39, 8.40s/step, loss=1.439, acc=0.677, acc_len=1.61, thru=14.3, I=14.3, T=28.9, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:29,519] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:36,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:58:37,511] loop.py:45 INFO TRAIN_STEP step=8484 loss=1.526679 Training: 47%|████▋ | 8484/18018 [18:28:23<22:00:53, 8.31s/step, loss=1.527, acc=0.664, acc_len=1.56, thru=15.1, I=16.5, T=29.7, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:37,700] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:43,749] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:58:44,105] loop.py:380 INFO TIMING step=8485: step=4.221s data=1.910s compute=4.188s [fwd=2.946s bwd=1.226s opt=0.003s] dispatch=2.363s [2026-10-19 07:58:44,106] loop.py:45 INFO TRAIN_STEP step=8485 loss=1.366938 Training: 47%|████▋ | 8485/18018 [18:28:30<20:38:47, 7.80s/step, loss=1.367, acc=0.695, acc_len=1.67, thru=17.5, I=21.5, T=30.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:58:44,263] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:58:54,659] loop.py:45 INFO TRAIN_STEP step=8486 loss=1.422733 Training: 47%|████▋ | 8486/18018 [18:28:40<22:50:05, 8.62s/step, loss=1.423, acc=0.679, acc_len=1.61, thru=13.2, I=13.4, T=26.3, wait=5.7s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:58:54,654] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8485: forward=3409.4ms backward=1417.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:03,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:59:04,245] loop.py:45 INFO TRAIN_STEP step=8487 loss=1.473458 Training: 47%|████▋ | 8487/18018 [18:28:50<23:35:47, 8.91s/step, loss=1.473, acc=0.673, acc_len=1.59, thru=13.9, I=12.3, T=26.2, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:04,469] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:10,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:59:12,045] loop.py:45 INFO TRAIN_STEP step=8488 loss=1.342362 Training: 47%|████▋ | 8488/18018 [18:28:58<22:42:34, 8.58s/step, loss=1.342, acc=0.691, acc_len=1.65, thru=16.7, I=21.6, T=29.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:12,342] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:17,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:59:18,701] loop.py:45 INFO TRAIN_STEP step=8489 loss=1.401865 Training: 47%|████▋ | 8489/18018 [18:29:04<21:10:49, 8.00s/step, loss=1.402, acc=0.681, acc_len=1.61, thru=17.4, I=25.3, T=31.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:18,979] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 07:59:25,317] loop.py:380 INFO TIMING step=8490: step=3.858s data=1.689s compute=3.827s [fwd=2.758s bwd=1.055s opt=0.003s] dispatch=2.749s [2026-10-19 07:59:25,317] loop.py:45 INFO TRAIN_STEP step=8490 loss=1.375951 Training: 47%|████▋ | 8490/18018 [18:29:11<20:04:38, 7.59s/step, loss=1.376, acc=0.684, acc_len=1.62, thru=17.5, I=20.4, T=33.2, wait=2.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:31,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:59:32,863] loop.py:45 INFO TRAIN_STEP step=8491 loss=1.464699 Training: 47%|████▋ | 8491/18018 [18:29:18<20:02:42, 7.57s/step, loss=1.465, acc=0.675, acc_len=1.60, thru=17.8, I=18.1, T=29.4, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 07:59:32,858] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8490: forward=3034.9ms backward=1253.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:32,955] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-19 07:59:41,237] loop.py:45 INFO TRAIN_STEP step=8492 loss=1.462663 Training: 47%|████▋ | 8492/18018 [18:29:27<20:40:32, 7.81s/step, loss=1.463, acc=0.674, acc_len=1.59, thru=15.1, I=15.5, T=28.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:41,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:41,686] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 07:59:48,540] loop.py:45 INFO TRAIN_STEP step=8493 loss=1.431201 Training: 47%|████▋ | 8493/18018 [18:29:34<20:16:18, 7.66s/step, loss=1.431, acc=0.679, acc_len=1.61, thru=16.1, I=15.9, T=32.8, wait=3.4s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:54,972] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 07:59:56,286] loop.py:45 INFO TRAIN_STEP step=8494 loss=1.382836 Training: 47%|████▋ | 8494/18018 [18:29:42<20:20:07, 7.69s/step, loss=1.383, acc=0.683, acc_len=1.63, thru=16.6, I=17.8, T=30.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 07:59:56,482] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:01,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:00:02,981] loop.py:380 INFO TIMING step=8495: step=3.820s data=1.660s compute=3.789s [fwd=2.661s bwd=1.113s opt=0.003s] dispatch=2.864s [2026-10-19 08:00:02,981] loop.py:45 INFO TRAIN_STEP step=8495 loss=1.448126 Training: 47%|████▋ | 8495/18018 [18:29:49<19:32:44, 7.39s/step, loss=1.448, acc=0.675, acc_len=1.59, thru=17.4, I=23.2, T=33.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:03,015] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:11,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:00:11,408] loop.py:45 INFO TRAIN_STEP step=8496 loss=1.510249 Training: 47%|████▋ | 8496/18018 [18:29:57<20:22:00, 7.70s/step, loss=1.510, acc=0.663, acc_len=1.56, thru=14.3, I=15.8, T=31.3, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:00:11,404] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8495: forward=2789.8ms backward=1235.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:11,650] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:18,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:00:20,608] loop.py:45 INFO TRAIN_STEP step=8497 loss=1.481930 Training: 47%|████▋ | 8497/18018 [18:30:06<21:33:19, 8.15s/step, loss=1.482, acc=0.665, acc_len=1.57, thru=14.7, I=17.9, T=28.5, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:20,831] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:27,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:00:27,325] loop.py:45 INFO TRAIN_STEP step=8498 loss=1.453802 Training: 47%|████▋ | 8498/18018 [18:30:13<20:24:55, 7.72s/step, loss=1.454, acc=0.675, acc_len=1.60, thru=16.3, I=20.8, T=33.2, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:27,645] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:32,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:00:33,687] loop.py:45 INFO TRAIN_STEP step=8499 loss=1.432191 Training: 47%|████▋ | 8499/18018 [18:30:19<19:20:12, 7.31s/step, loss=1.432, acc=0.679, acc_len=1.61, thru=19.2, I=21.8, T=31.3, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:33,843] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:38,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:00:40,372] loop.py:380 INFO TIMING step=8500: step=4.173s data=1.703s compute=4.144s [fwd=2.947s bwd=1.181s opt=0.003s] dispatch=2.501s [2026-10-19 08:00:40,372] loop.py:45 INFO TRAIN_STEP step=8500 loss=1.285228 Training: 47%|████▋ | 8500/18018 [18:30:26<18:50:10, 7.12s/step, loss=1.285, acc=0.707, acc_len=1.71, thru=17.6, I=24.0, T=30.7, wait=2.5s, pool=64, epoch=1/2][2026-10-19 08:00:40,377] loop.py:419 INFO Saving checkpoint at step 8500... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:43,593] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 08:00:43,843] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0008501 (TrainerActor pid=9310) [2026-10-19 08:00:43,850] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0007751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:00:45,101] inference_manager.py:408 INFO Pool capacity available after 6.5s, pool_size=0/64 [2026-10-19 08:00:52,321] loop.py:45 INFO TRAIN_STEP step=8501 loss=1.339570 Training: 47%|████▋ | 8501/18018 [18:30:38<22:39:34, 8.57s/step, loss=1.340, acc=0.695, acc_len=1.67, thru=15.0, I=17.6, T=30.2, wait=3.2s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:00:52,317] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8500: forward=2986.1ms backward=1199.3ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 08:00:43,844] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0008501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 08:00:44,870] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0007751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:02,538] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:01:02,723] loop.py:45 INFO TRAIN_STEP step=8502 loss=1.609678 Training: 47%|████▋ | 8502/18018 [18:30:48<24:06:31, 9.12s/step, loss=1.610, acc=0.651, acc_len=1.52, thru=14.7, I=14.4, T=27.5, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:03,039] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:08,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:01:09,979] loop.py:45 INFO TRAIN_STEP step=8503 loss=1.373442 Training: 47%|████▋ | 8503/18018 [18:30:56<22:37:38, 8.56s/step, loss=1.373, acc=0.689, acc_len=1.65, thru=16.7, I=19.2, T=30.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:10,111] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:16,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:01:17,116] loop.py:45 INFO TRAIN_STEP step=8504 loss=1.423839 Training: 47%|████▋ | 8504/18018 [18:31:03<21:29:50, 8.13s/step, loss=1.424, acc=0.681, acc_len=1.62, thru=16.9, I=20.4, T=30.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:17,245] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:22,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:01:24,825] loop.py:380 INFO TIMING step=8505: step=4.504s data=2.090s compute=4.470s [fwd=3.150s bwd=1.305s opt=0.003s] dispatch=3.195s [2026-10-19 08:01:24,825] loop.py:45 INFO TRAIN_STEP step=8505 loss=1.474624 Training: 47%|████▋ | 8505/18018 [18:31:10<21:09:22, 8.01s/step, loss=1.475, acc=0.669, acc_len=1.58, thru=16.1, I=21.7, T=28.4, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:25,098] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 08:01:32,208] loop.py:45 INFO TRAIN_STEP step=8506 loss=1.478444 Training: 47%|████▋ | 8506/18018 [18:31:18<20:39:37, 7.82s/step, loss=1.478, acc=0.673, acc_len=1.59, thru=10.4, I=15.2, T=34.2, wait=3.6s, pool=24, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:01:32,203] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8505: forward=2652.6ms backward=1023.8ms (TrainerActor pid=9368) [2026-10-19 08:01:32,260] eagle_store.py:517 WARNING batch_get_buffer missing keys (k272193-8e4c29a4_hs, k272193-8e4c29a4_ids, k272193-8e4c29a4_lhs); sleeping 0.50s. [2026-10-19 08:01:41,624] loop.py:45 INFO TRAIN_STEP step=8507 loss=1.259142 Training: 47%|████▋ | 8507/18018 [18:31:27<21:55:25, 8.30s/step, loss=1.259, acc=0.706, acc_len=1.70, thru=15.8, I=13.7, T=32.8, wait=5.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:48,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9370) [2026-10-19 08:01:32,296] eagle_store.py:517 WARNING batch_get_buffer missing keys (k272195-b4c28f11_hs, k272195-b4c28f11_ids, k272195-b4c28f11_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 08:01:49,339] loop.py:45 INFO TRAIN_STEP step=8508 loss=1.367689 Training: 47%|████▋ | 8508/18018 [18:31:35<21:27:33, 8.12s/step, loss=1.368, acc=0.689, acc_len=1.64, thru=17.9, I=18.7, T=29.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:49,767] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:56,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:01:58,793] loop.py:45 INFO TRAIN_STEP step=8509 loss=1.371847 Training: 47%|████▋ | 8509/18018 [18:31:44<22:30:45, 8.52s/step, loss=1.372, acc=0.685, acc_len=1.64, thru=14.2, I=15.2, T=29.6, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:01:58,974] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:02:07,179] loop.py:380 INFO TIMING step=8510: step=4.191s data=1.994s compute=4.152s [fwd=2.940s bwd=1.197s opt=0.003s] dispatch=4.184s [2026-10-19 08:02:07,179] loop.py:45 INFO TRAIN_STEP step=8510 loss=1.336196 Training: 47%|████▋ | 8510/18018 [18:31:53<22:24:03, 8.48s/step, loss=1.336, acc=0.694, acc_len=1.66, thru=14.6, I=18.3, T=30.5, wait=4.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:02:15,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:02:17,360] loop.py:45 INFO TRAIN_STEP step=8511 loss=1.448420 Training: 47%|████▋ | 8511/18018 [18:32:03<23:44:43, 8.99s/step, loss=1.448, acc=0.681, acc_len=1.62, thru=14.8, I=13.7, T=28.5, wait=5.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:02:17,355] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8510: forward=3181.6ms backward=1239.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:02:17,727] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:02:26,456] loop.py:45 INFO TRAIN_STEP step=8512 loss=1.471151 Training: 47%|████▋ | 8512/18018 [18:32:12<23:49:23, 9.02s/step, loss=1.471, acc=0.675, acc_len=1.59, thru=14.7, I=17.2, T=28.0, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:02:26,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:02:26,916] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:02:33,721] loop.py:45 INFO TRAIN_STEP step=8513 loss=1.489864 Training: 47%|████▋ | 8513/18018 [18:32:19<22:25:53, 8.50s/step, loss=1.490, acc=0.671, acc_len=1.59, thru=12.2, I=16.0, T=31.2, wait=3.2s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:02:40,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:02:42,620] loop.py:45 INFO TRAIN_STEP step=8514 loss=1.461319 Training: 47%|████▋ | 8514/18018 [18:32:28<22:44:54, 8.62s/step, loss=1.461, acc=0.676, acc_len=1.60, thru=16.0, I=16.5, T=27.5, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:02:42,930] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 08:02:48,729] loop.py:380 INFO TIMING step=8515: step=3.779s data=1.657s compute=3.747s [fwd=2.589s bwd=1.144s opt=0.003s] dispatch=2.320s [2026-10-19 08:02:48,729] loop.py:45 INFO TRAIN_STEP step=8515 loss=1.228043 Training: 47%|████▋ | 8515/18018 [18:32:34<20:45:38, 7.86s/step, loss=1.228, acc=0.713, acc_len=1.72, thru=15.8, I=25.8, T=33.9, wait=2.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:02:58,047] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:02:58,255] loop.py:45 INFO TRAIN_STEP step=8516 loss=1.424454 Training: 47%|████▋ | 8516/18018 [18:32:44<22:04:21, 8.36s/step, loss=1.424, acc=0.679, acc_len=1.61, thru=15.6, I=14.8, T=30.8, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:02:58,251] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8515: forward=2971.5ms backward=1125.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:02:58,549] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:03:06,044] loop.py:45 INFO TRAIN_STEP step=8517 loss=1.376837 Training: 47%|████▋ | 8517/18018 [18:32:52<21:36:55, 8.19s/step, loss=1.377, acc=0.688, acc_len=1.64, thru=13.7, I=14.3, T=26.2, wait=2.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:12,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:03:13,418] loop.py:45 INFO TRAIN_STEP step=8518 loss=1.464974 Training: 47%|████▋ | 8518/18018 [18:32:59<20:58:00, 7.95s/step, loss=1.465, acc=0.671, acc_len=1.58, thru=18.7, I=17.0, T=31.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:13,717] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:20,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:03:21,639] loop.py:45 INFO TRAIN_STEP step=8519 loss=1.479957 Training: 47%|████▋ | 8519/18018 [18:33:07<21:10:59, 8.03s/step, loss=1.480, acc=0.674, acc_len=1.59, thru=14.4, I=18.8, T=30.7, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:21,812] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:03:28,187] loop.py:380 INFO TIMING step=8520: step=3.789s data=1.659s compute=3.756s [fwd=2.650s bwd=1.092s opt=0.003s] dispatch=2.750s [2026-10-19 08:03:28,187] loop.py:45 INFO TRAIN_STEP step=8520 loss=1.317739 Training: 47%|████▋ | 8520/18018 [18:33:14<20:00:35, 7.58s/step, loss=1.318, acc=0.699, acc_len=1.68, thru=17.7, I=22.3, T=33.8, wait=2.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:34,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:03:36,935] loop.py:45 INFO TRAIN_STEP step=8521 loss=1.371031 Training: 47%|████▋ | 8521/18018 [18:33:23<20:55:47, 7.93s/step, loss=1.371, acc=0.693, acc_len=1.66, thru=14.9, I=19.6, T=27.8, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:03:36,930] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8520: forward=3248.9ms backward=1316.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:37,011] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:42,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:03:43,556] loop.py:45 INFO TRAIN_STEP step=8522 loss=1.341453 Training: 47%|████▋ | 8522/18018 [18:33:29<19:53:19, 7.54s/step, loss=1.341, acc=0.696, acc_len=1.66, thru=15.5, I=24.6, T=32.5, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:43,613] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-19 08:03:51,270] loop.py:45 INFO TRAIN_STEP step=8523 loss=1.378966 Training: 47%|████▋ | 8523/18018 [18:33:37<20:01:25, 7.59s/step, loss=1.379, acc=0.688, acc_len=1.64, thru=14.8, I=16.5, T=31.2, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:58,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:03:58,824] loop.py:45 INFO TRAIN_STEP step=8524 loss=1.317085 Training: 47%|████▋ | 8524/18018 [18:33:44<19:59:30, 7.58s/step, loss=1.317, acc=0.701, acc_len=1.69, thru=17.9, I=17.1, T=27.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:03:58,879] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:04,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:04:05,372] loop.py:380 INFO TIMING step=8525: step=3.985s data=1.733s compute=3.951s [fwd=2.804s bwd=1.132s opt=0.003s] dispatch=2.552s [2026-10-19 08:04:05,372] loop.py:45 INFO TRAIN_STEP step=8525 loss=1.373408 Training: 47%|████▋ | 8525/18018 [18:33:51<19:10:14, 7.27s/step, loss=1.373, acc=0.688, acc_len=1.64, thru=17.7, I=18.8, T=32.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:05,877] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:04:14,441] loop.py:45 INFO TRAIN_STEP step=8526 loss=1.328359 Training: 47%|████▋ | 8526/18018 [18:34:00<20:35:37, 7.81s/step, loss=1.328, acc=0.698, acc_len=1.67, thru=11.3, I=14.9, T=30.1, wait=4.8s, pool=36, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:04:14,436] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8525: forward=2986.1ms backward=1219.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:24,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:04:25,298] loop.py:45 INFO TRAIN_STEP step=8527 loss=1.329047 Training: 47%|████▋ | 8527/18018 [18:34:11<23:00:04, 8.72s/step, loss=1.329, acc=0.699, acc_len=1.68, thru=15.6, I=12.7, T=26.9, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:25,477] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:32,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:04:34,251] loop.py:45 INFO TRAIN_STEP step=8528 loss=1.490771 Training: 47%|████▋ | 8528/18018 [18:34:20<23:10:39, 8.79s/step, loss=1.491, acc=0.671, acc_len=1.59, thru=14.9, I=16.3, T=30.9, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:34,689] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:04:40,702] loop.py:45 INFO TRAIN_STEP step=8529 loss=1.400218 Training: 47%|████▋ | 8529/18018 [18:34:26<21:19:29, 8.09s/step, loss=1.400, acc=0.683, acc_len=1.63, thru=16.1, I=23.4, T=34.8, wait=2.8s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:47,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:04:48,866] loop.py:380 INFO TIMING step=8530: step=4.500s data=2.191s compute=4.464s [fwd=3.048s bwd=1.402s opt=0.003s] dispatch=3.653s [2026-10-19 08:04:48,866] loop.py:45 INFO TRAIN_STEP step=8530 loss=1.325855 Training: 47%|████▋ | 8530/18018 [18:34:34<21:22:53, 8.11s/step, loss=1.326, acc=0.697, acc_len=1.67, thru=16.9, I=17.7, T=28.4, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:49,136] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:53,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:04:55,913] loop.py:45 INFO TRAIN_STEP step=8531 loss=1.469697 Training: 47%|████▋ | 8531/18018 [18:34:41<20:32:02, 7.79s/step, loss=1.470, acc=0.671, acc_len=1.57, thru=15.9, I=23.1, T=29.4, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:04:55,909] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8530: forward=3004.9ms backward=1305.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:04:56,192] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:00,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:05:02,187] loop.py:45 INFO TRAIN_STEP step=8532 loss=1.254367 Training: 47%|████▋ | 8532/18018 [18:34:48<19:20:00, 7.34s/step, loss=1.254, acc=0.714, acc_len=1.73, thru=17.1, I=26.5, T=33.2, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:02,305] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:08,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:05:09,820] loop.py:45 INFO TRAIN_STEP step=8533 loss=1.450031 Training: 47%|████▋ | 8533/18018 [18:34:55<19:33:56, 7.43s/step, loss=1.450, acc=0.677, acc_len=1.61, thru=15.4, I=20.5, T=27.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:10,204] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:14,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:05:16,695] loop.py:45 INFO TRAIN_STEP step=8534 loss=1.297360 Training: 47%|████▋ | 8534/18018 [18:35:02<19:07:39, 7.26s/step, loss=1.297, acc=0.701, acc_len=1.70, thru=16.9, I=23.4, T=29.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:16,951] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:21,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:05:23,962] loop.py:380 INFO TIMING step=8535: step=4.579s data=2.162s compute=4.542s [fwd=3.138s bwd=1.390s opt=0.003s] dispatch=2.679s [2026-10-19 08:05:23,962] loop.py:45 INFO TRAIN_STEP step=8535 loss=1.338696 Training: 47%|████▋ | 8535/18018 [18:35:10<19:07:52, 7.26s/step, loss=1.339, acc=0.693, acc_len=1.66, thru=15.3, I=24.9, T=28.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:24,047] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:28,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:05:30,770] loop.py:45 INFO TRAIN_STEP step=8536 loss=1.321584 Training: 47%|████▋ | 8536/18018 [18:35:16<18:46:12, 7.13s/step, loss=1.322, acc=0.699, acc_len=1.67, thru=16.1, I=30.2, T=27.7, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:05:30,764] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8535: forward=3116.6ms backward=1449.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:30,985] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:37,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:05:38,057] loop.py:45 INFO TRAIN_STEP step=8537 loss=1.491725 Training: 47%|████▋ | 8537/18018 [18:35:24<18:53:42, 7.17s/step, loss=1.492, acc=0.673, acc_len=1.59, thru=14.5, I=20.5, T=33.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:38,272] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:43,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:05:44,472] loop.py:45 INFO TRAIN_STEP step=8538 loss=1.252357 Training: 47%|████▋ | 8538/18018 [18:35:30<18:17:31, 6.95s/step, loss=1.252, acc=0.710, acc_len=1.72, thru=19.0, I=23.3, T=32.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:44,738] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:05:51,460] loop.py:45 INFO TRAIN_STEP step=8539 loss=1.270703 Training: 47%|████▋ | 8539/18018 [18:35:37<18:19:27, 6.96s/step, loss=1.271, acc=0.704, acc_len=1.70, thru=15.5, I=17.2, T=34.1, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:05:59,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:06:00,181] loop.py:380 INFO TIMING step=8540: step=4.051s data=1.753s compute=4.017s [fwd=2.880s bwd=1.122s opt=0.003s] dispatch=4.660s [2026-10-19 08:06:00,181] loop.py:45 INFO TRAIN_STEP step=8540 loss=1.375405 Training: 47%|████▋ | 8540/18018 [18:35:46<19:42:43, 7.49s/step, loss=1.375, acc=0.688, acc_len=1.64, thru=17.5, I=16.7, T=31.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:06:00,247] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 [2026-10-19 08:06:06,000] loop.py:45 INFO TRAIN_STEP step=8541 loss=1.303257 Training: 47%|████▋ | 8541/18018 [18:35:52<18:23:30, 6.99s/step, loss=1.303, acc=0.697, acc_len=1.67, thru=19.0, I=19.9, T=35.5, wait=2.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:06:05,996] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8540: forward=2483.2ms backward=1056.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:06:10,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:06:11,887] loop.py:45 INFO TRAIN_STEP step=8542 loss=1.218031 Training: 47%|████▋ | 8542/18018 [18:35:57<17:31:28, 6.66s/step, loss=1.218, acc=0.716, acc_len=1.74, thru=21.0, I=23.2, T=33.2, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:06:12,164] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:06:20,633] loop.py:45 INFO TRAIN_STEP step=8543 loss=1.304481 Training: 47%|████▋ | 8543/18018 [18:36:06<19:10:09, 7.28s/step, loss=1.304, acc=0.701, acc_len=1.69, thru=11.2, I=14.6, T=29.8, wait=4.4s, pool=44, epoch=1/2][2026-10-19 08:06:30,673] loop.py:45 INFO TRAIN_STEP step=8544 loss=1.385125 Training: 47%|████▋ | 8544/18018 [18:36:16<21:20:46, 8.11s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=14.5, I=14.2, T=29.5, wait=5.7s, pool=60, epoch=1/2][2026-10-19 08:06:39,445] loop.py:380 INFO TIMING step=8545: step=5.408s data=2.902s compute=5.377s [fwd=4.068s bwd=1.293s opt=0.003s] dispatch=3.354s [2026-10-19 08:06:39,445] loop.py:45 INFO TRAIN_STEP step=8545 loss=1.435822 Training: 47%|████▋ | 8545/18018 [18:36:25<21:51:52, 8.31s/step, loss=1.436, acc=0.676, acc_len=1.60, thru=14.4, I=15.2, T=23.7, wait=3.4s, pool=60, epoch=1/2][2026-10-19 08:06:46,231] loop.py:45 INFO TRAIN_STEP step=8546 loss=1.471606 Training: 47%|████▋ | 8546/18018 [18:36:32<20:39:33, 7.85s/step, loss=1.472, acc=0.674, acc_len=1.59, thru=16.7, I=15.0, T=32.6, wait=2.9s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:06:46,225] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8545: forward=2772.1ms backward=1091.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:06:55,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:06:56,051] loop.py:45 INFO TRAIN_STEP step=8547 loss=1.379012 Training: 47%|████▋ | 8547/18018 [18:36:42<22:12:35, 8.44s/step, loss=1.379, acc=0.689, acc_len=1.64, thru=14.3, I=14.5, T=31.5, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:06:56,081] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:03,011] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:07:03,121] loop.py:45 INFO TRAIN_STEP step=8548 loss=1.533734 Training: 47%|████▋ | 8548/18018 [18:36:49<21:07:25, 8.03s/step, loss=1.534, acc=0.663, acc_len=1.56, thru=17.8, I=18.3, T=32.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:03,515] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:12,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:07:13,022] loop.py:45 INFO TRAIN_STEP step=8549 loss=1.372535 Training: 47%|████▋ | 8549/18018 [18:36:59<22:35:59, 8.59s/step, loss=1.373, acc=0.690, acc_len=1.65, thru=13.5, I=14.5, T=29.1, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:13,215] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:18,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:07:20,795] loop.py:380 INFO TIMING step=8550: step=4.699s data=2.209s compute=4.675s [fwd=3.270s bwd=1.391s opt=0.003s] dispatch=3.063s [2026-10-19 08:07:20,795] loop.py:45 INFO TRAIN_STEP step=8550 loss=1.419722 Training: 47%|████▋ | 8550/18018 [18:37:06<21:57:01, 8.35s/step, loss=1.420, acc=0.681, acc_len=1.61, thru=16.1, I=20.4, T=27.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:21,105] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 08:07:29,593] loop.py:45 INFO TRAIN_STEP step=8551 loss=1.374320 Training: 47%|████▋ | 8551/18018 [18:37:15<22:18:15, 8.48s/step, loss=1.374, acc=0.689, acc_len=1.64, thru=12.3, I=12.3, T=28.6, wait=4.3s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:07:29,588] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8550: forward=3170.0ms backward=1260.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:37,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:07:38,072] loop.py:45 INFO TRAIN_STEP step=8552 loss=1.421021 Training: 47%|████▋ | 8552/18018 [18:37:24<22:18:01, 8.48s/step, loss=1.421, acc=0.681, acc_len=1.62, thru=17.8, I=17.0, T=31.3, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:38,523] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:43,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:07:44,384] loop.py:45 INFO TRAIN_STEP step=8553 loss=1.283348 Training: 47%|████▋ | 8553/18018 [18:37:30<20:35:17, 7.83s/step, loss=1.283, acc=0.707, acc_len=1.71, thru=18.6, I=24.1, T=33.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:44,603] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:50,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:07:50,835] loop.py:45 INFO TRAIN_STEP step=8554 loss=1.393154 Training: 47%|████▋ | 8554/18018 [18:37:36<19:29:46, 7.42s/step, loss=1.393, acc=0.681, acc_len=1.62, thru=19.4, I=22.0, T=33.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:51,222] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:57,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:07:58,176] loop.py:380 INFO TIMING step=8555: step=3.796s data=1.667s compute=3.759s [fwd=2.682s bwd=1.062s opt=0.003s] dispatch=3.536s [2026-10-19 08:07:58,176] loop.py:45 INFO TRAIN_STEP step=8555 loss=1.499851 Training: 47%|████▋ | 8555/18018 [18:37:44<19:26:09, 7.39s/step, loss=1.500, acc=0.668, acc_len=1.57, thru=16.9, I=17.2, T=33.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:07:58,490] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:05,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:08:06,586] loop.py:45 INFO TRAIN_STEP step=8556 loss=1.454158 Training: 47%|████▋ | 8556/18018 [18:37:52<20:14:06, 7.70s/step, loss=1.454, acc=0.676, acc_len=1.60, thru=15.1, I=16.5, T=26.8, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:08:06,580] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8555: forward=3258.6ms backward=1465.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:06,741] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-19 08:08:14,096] loop.py:45 INFO TRAIN_STEP step=8557 loss=1.363584 Training: 47%|████▋ | 8557/18018 [18:38:00<20:04:58, 7.64s/step, loss=1.364, acc=0.690, acc_len=1.65, thru=14.7, I=16.3, T=26.6, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:19,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:08:21,713] loop.py:45 INFO TRAIN_STEP step=8558 loss=1.354024 Training: 47%|████▋ | 8558/18018 [18:38:07<20:03:40, 7.63s/step, loss=1.354, acc=0.690, acc_len=1.64, thru=17.4, I=18.8, T=27.7, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:22,063] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:28,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:08:29,677] loop.py:45 INFO TRAIN_STEP step=8559 loss=1.407899 Training: 48%|████▊ | 8559/18018 [18:38:15<20:19:08, 7.73s/step, loss=1.408, acc=0.677, acc_len=1.61, thru=16.8, I=20.3, T=34.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:30,045] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:08:36,845] loop.py:380 INFO TIMING step=8560: step=3.967s data=1.731s compute=3.938s [fwd=2.775s bwd=1.148s opt=0.003s] dispatch=3.193s [2026-10-19 08:08:36,846] loop.py:45 INFO TRAIN_STEP step=8560 loss=1.298399 Training: 48%|████▊ | 8560/18018 [18:38:22<19:52:16, 7.56s/step, loss=1.298, acc=0.700, acc_len=1.69, thru=17.8, I=19.3, T=32.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:36,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:37,287] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:43,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:08:44,275] loop.py:45 INFO TRAIN_STEP step=8561 loss=1.331519 Training: 48%|████▊ | 8561/18018 [18:38:30<19:45:55, 7.52s/step, loss=1.332, acc=0.694, acc_len=1.67, thru=18.6, I=18.6, T=33.1, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:08:44,271] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8560: forward=2763.2ms backward=1057.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:44,629] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:08:52,290] loop.py:45 INFO TRAIN_STEP step=8562 loss=1.412162 Training: 48%|████▊ | 8562/18018 [18:38:38<20:08:53, 7.67s/step, loss=1.412, acc=0.679, acc_len=1.61, thru=15.4, I=16.0, T=31.4, wait=3.9s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:58,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:08:59,081] loop.py:45 INFO TRAIN_STEP step=8563 loss=1.549961 Training: 48%|████▊ | 8563/18018 [18:38:45<19:27:15, 7.41s/step, loss=1.550, acc=0.658, acc_len=1.54, thru=19.9, I=19.1, T=35.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:08:59,143] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:06,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:09:06,470] loop.py:45 INFO TRAIN_STEP step=8564 loss=1.384871 Training: 48%|████▊ | 8564/18018 [18:38:52<19:26:18, 7.40s/step, loss=1.385, acc=0.683, acc_len=1.63, thru=16.9, I=19.0, T=31.1, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:06,772] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:14,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:09:14,377] loop.py:380 INFO TIMING step=8565: step=4.146s data=1.914s compute=4.116s [fwd=2.920s bwd=1.181s opt=0.003s] dispatch=3.750s [2026-10-19 08:09:14,377] loop.py:45 INFO TRAIN_STEP step=8565 loss=1.411525 Training: 48%|████▊ | 8565/18018 [18:39:00<19:49:55, 7.55s/step, loss=1.412, acc=0.682, acc_len=1.62, thru=16.1, I=15.6, T=30.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:14,804] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:19,328] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:09:21,989] loop.py:45 INFO TRAIN_STEP step=8566 loss=1.412788 Training: 48%|████▊ | 8566/18018 [18:39:08<19:52:38, 7.57s/step, loss=1.413, acc=0.683, acc_len=1.62, thru=18.2, I=21.0, T=25.8, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:09:21,985] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8565: forward=3431.3ms backward=1443.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:22,347] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:27,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:09:28,790] loop.py:45 INFO TRAIN_STEP step=8567 loss=1.454632 Training: 48%|████▊ | 8567/18018 [18:39:14<19:16:09, 7.34s/step, loss=1.455, acc=0.674, acc_len=1.59, thru=14.0, I=22.7, T=28.9, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:29,175] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:35,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:09:36,403] loop.py:45 INFO TRAIN_STEP step=8568 loss=1.427184 Training: 48%|████▊ | 8568/18018 [18:39:22<19:28:59, 7.42s/step, loss=1.427, acc=0.682, acc_len=1.62, thru=14.7, I=16.9, T=25.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:36,704] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:42,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:09:42,711] loop.py:45 INFO TRAIN_STEP step=8569 loss=1.529205 Training: 48%|████▊ | 8569/18018 [18:39:28<18:36:09, 7.09s/step, loss=1.529, acc=0.663, acc_len=1.56, thru=18.0, I=19.2, T=31.9, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:42,881] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:49,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:09:50,813] loop.py:380 INFO TIMING step=8570: step=4.258s data=1.959s compute=4.219s [fwd=2.926s bwd=1.279s opt=0.003s] dispatch=3.835s [2026-10-19 08:09:50,813] loop.py:45 INFO TRAIN_STEP step=8570 loss=1.437393 Training: 48%|████▊ | 8570/18018 [18:39:36<19:24:02, 7.39s/step, loss=1.437, acc=0.680, acc_len=1.62, thru=16.5, I=17.9, T=30.1, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:51,209] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:58,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:09:59,045] loop.py:45 INFO TRAIN_STEP step=8571 loss=1.313453 Training: 48%|████▊ | 8571/18018 [18:39:45<20:03:36, 7.64s/step, loss=1.313, acc=0.700, acc_len=1.69, thru=14.5, I=18.8, T=29.7, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:09:59,040] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8570: forward=3032.2ms backward=1228.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:09:59,245] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:10:07,316] loop.py:45 INFO TRAIN_STEP step=8572 loss=1.391490 Training: 48%|████▊ | 8572/18018 [18:39:53<20:32:59, 7.83s/step, loss=1.391, acc=0.689, acc_len=1.64, thru=13.8, I=16.6, T=25.1, wait=3.2s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:12,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:10:14,288] loop.py:45 INFO TRAIN_STEP step=8573 loss=1.382851 Training: 48%|████▊ | 8573/18018 [18:40:00<19:52:13, 7.57s/step, loss=1.383, acc=0.687, acc_len=1.64, thru=16.4, I=18.0, T=31.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:14,465] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:20,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:10:21,072] loop.py:45 INFO TRAIN_STEP step=8574 loss=1.262456 Training: 48%|████▊ | 8574/18018 [18:40:07<19:14:52, 7.34s/step, loss=1.262, acc=0.713, acc_len=1.72, thru=16.4, I=19.6, T=32.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:21,467] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:25,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:10:28,462] loop.py:380 INFO TIMING step=8575: step=4.917s data=2.434s compute=4.874s [fwd=3.326s bwd=1.533s opt=0.003s] dispatch=2.464s [2026-10-19 08:10:28,463] loop.py:45 INFO TRAIN_STEP step=8575 loss=1.377290 Training: 48%|████▊ | 8575/18018 [18:40:14<19:17:12, 7.35s/step, loss=1.377, acc=0.690, acc_len=1.65, thru=17.7, I=23.1, T=26.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:28,837] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:35,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:10:35,671] loop.py:45 INFO TRAIN_STEP step=8576 loss=1.372117 Training: 48%|████▊ | 8576/18018 [18:40:21<19:10:23, 7.31s/step, loss=1.372, acc=0.690, acc_len=1.65, thru=13.4, I=17.7, T=29.7, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:10:35,667] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8575: forward=2860.1ms backward=1409.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:35,825] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-19 08:10:41,853] loop.py:45 INFO TRAIN_STEP step=8577 loss=1.293454 Training: 48%|████▊ | 8577/18018 [18:40:27<18:16:58, 6.97s/step, loss=1.293, acc=0.705, acc_len=1.69, thru=18.1, I=18.0, T=32.5, wait=2.2s, pool=52, epoch=1/2][2026-10-19 08:10:49,185] loop.py:45 INFO TRAIN_STEP step=8578 loss=1.362543 Training: 48%|████▊ | 8578/18018 [18:40:35<18:33:48, 7.08s/step, loss=1.363, acc=0.693, acc_len=1.66, thru=16.5, I=17.4, T=29.9, wait=3.0s, pool=48, epoch=1/2][2026-10-19 08:10:58,664] loop.py:45 INFO TRAIN_STEP step=8579 loss=1.392594 Training: 48%|████▊ | 8579/18018 [18:40:44<20:26:57, 7.80s/step, loss=1.393, acc=0.687, acc_len=1.64, thru=15.2, I=13.5, T=30.8, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:58,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:10:59,087] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:04,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:11:05,934] loop.py:380 INFO TIMING step=8580: step=4.711s data=2.227s compute=4.677s [fwd=3.296s bwd=1.365s opt=0.003s] dispatch=2.549s [2026-10-19 08:11:05,934] loop.py:45 INFO TRAIN_STEP step=8580 loss=1.413216 Training: 48%|████▊ | 8580/18018 [18:40:52<20:01:49, 7.64s/step, loss=1.413, acc=0.682, acc_len=1.62, thru=17.3, I=21.1, T=27.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:06,056] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:14,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:11:15,162] loop.py:45 INFO TRAIN_STEP step=8581 loss=1.301304 Training: 48%|████▊ | 8581/18018 [18:41:01<21:16:36, 8.12s/step, loss=1.301, acc=0.704, acc_len=1.70, thru=14.9, I=13.1, T=26.2, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:11:15,158] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8580: forward=3434.5ms backward=1391.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:15,323] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:22,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:11:23,580] loop.py:45 INFO TRAIN_STEP step=8582 loss=1.385862 Training: 48%|████▊ | 8582/18018 [18:41:09<21:30:46, 8.21s/step, loss=1.386, acc=0.690, acc_len=1.64, thru=14.6, I=14.2, T=26.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:24,067] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:28,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:11:30,445] loop.py:45 INFO TRAIN_STEP step=8583 loss=1.354646 Training: 48%|████▊ | 8583/18018 [18:41:16<20:27:19, 7.80s/step, loss=1.355, acc=0.693, acc_len=1.66, thru=17.1, I=20.9, T=31.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:30,556] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:35,535] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:11:36,716] loop.py:45 INFO TRAIN_STEP step=8584 loss=1.498152 Training: 48%|████▊ | 8584/18018 [18:41:22<19:14:44, 7.34s/step, loss=1.498, acc=0.668, acc_len=1.58, thru=19.0, I=25.3, T=31.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:37,042] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:42,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:11:44,086] loop.py:380 INFO TIMING step=8585: step=4.675s data=2.183s compute=4.644s [fwd=3.155s bwd=1.472s opt=0.003s] dispatch=2.687s [2026-10-19 08:11:44,086] loop.py:45 INFO TRAIN_STEP step=8585 loss=1.384291 Training: 48%|████▊ | 8585/18018 [18:41:30<19:15:50, 7.35s/step, loss=1.384, acc=0.689, acc_len=1.64, thru=16.4, I=22.4, T=27.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:44,129] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:49,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:11:50,943] loop.py:45 INFO TRAIN_STEP step=8586 loss=1.590221 Training: 48%|████▊ | 8586/18018 [18:41:37<18:52:23, 7.20s/step, loss=1.590, acc=0.651, acc_len=1.52, thru=16.3, I=21.0, T=31.6, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:11:50,939] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8585: forward=2815.3ms backward=1173.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:51,097] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:56,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:11:57,688] loop.py:45 INFO TRAIN_STEP step=8587 loss=1.443443 Training: 48%|████▊ | 8587/18018 [18:41:43<18:30:43, 7.07s/step, loss=1.443, acc=0.677, acc_len=1.60, thru=16.5, I=21.4, T=30.6, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:11:57,771] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:05,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:12:05,953] loop.py:45 INFO TRAIN_STEP step=8588 loss=1.452383 Training: 48%|████▊ | 8588/18018 [18:41:52<19:27:06, 7.43s/step, loss=1.452, acc=0.674, acc_len=1.60, thru=13.9, I=18.2, T=29.3, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:06,226] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:12:13,992] loop.py:45 INFO TRAIN_STEP step=8589 loss=1.407408 Training: 48%|████▊ | 8589/18018 [18:42:00<19:55:57, 7.61s/step, loss=1.407, acc=0.683, acc_len=1.62, thru=12.5, I=13.6, T=31.8, wait=4.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:21,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:12:23,450] loop.py:380 INFO TIMING step=8590: step=4.976s data=2.396s compute=4.929s [fwd=3.380s bwd=1.533s opt=0.003s] dispatch=4.470s [2026-10-19 08:12:23,450] loop.py:45 INFO TRAIN_STEP step=8590 loss=1.271407 Training: 48%|████▊ | 8590/18018 [18:42:09<21:22:52, 8.16s/step, loss=1.271, acc=0.705, acc_len=1.70, thru=16.1, I=18.9, T=25.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:23,519] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:28,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:12:29,955] loop.py:45 INFO TRAIN_STEP step=8591 loss=1.451660 Training: 48%|████▊ | 8591/18018 [18:42:16<20:04:30, 7.67s/step, loss=1.452, acc=0.674, acc_len=1.59, thru=16.2, I=22.7, T=32.5, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:12:29,950] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8590: forward=2691.7ms backward=1176.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:30,424] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:37,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:12:38,239] loop.py:45 INFO TRAIN_STEP step=8592 loss=1.359599 Training: 48%|████▊ | 8592/18018 [18:42:24<20:33:24, 7.85s/step, loss=1.360, acc=0.691, acc_len=1.65, thru=13.8, I=17.2, T=31.7, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:38,321] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 [2026-10-19 08:12:44,560] loop.py:45 INFO TRAIN_STEP step=8593 loss=1.387238 Training: 48%|████▊ | 8593/18018 [18:42:30<19:21:17, 7.39s/step, loss=1.387, acc=0.688, acc_len=1.64, thru=17.2, I=18.9, T=33.7, wait=2.5s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:54,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:12:55,932] loop.py:45 INFO TRAIN_STEP step=8594 loss=1.385477 Training: 48%|████▊ | 8594/18018 [18:42:42<22:28:39, 8.59s/step, loss=1.385, acc=0.685, acc_len=1.64, thru=12.9, I=13.5, T=31.6, wait=7.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:12:55,993] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 [2026-10-19 08:13:02,883] loop.py:380 INFO TIMING step=8595: step=3.957s data=1.796s compute=3.918s [fwd=2.771s bwd=1.132s opt=0.003s] dispatch=2.984s [2026-10-19 08:13:02,883] loop.py:45 INFO TRAIN_STEP step=8595 loss=1.315242 Training: 48%|████▊ | 8595/18018 [18:42:48<21:11:30, 8.10s/step, loss=1.315, acc=0.699, acc_len=1.68, thru=16.6, I=20.6, T=32.4, wait=3.0s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:08,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:13:08,959] loop.py:45 INFO TRAIN_STEP step=8596 loss=1.401319 Training: 48%|████▊ | 8596/18018 [18:42:55<19:36:00, 7.49s/step, loss=1.401, acc=0.680, acc_len=1.62, thru=19.2, I=19.8, T=34.5, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:13:08,955] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8595: forward=2636.8ms backward=1029.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:09,388] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:15,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:13:15,750] loop.py:45 INFO TRAIN_STEP step=8597 loss=1.377582 Training: 48%|████▊ | 8597/18018 [18:43:01<19:03:02, 7.28s/step, loss=1.378, acc=0.688, acc_len=1.64, thru=18.7, I=19.9, T=33.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:15,877] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-19 08:13:22,453] loop.py:45 INFO TRAIN_STEP step=8598 loss=1.348515 Training: 48%|████▊ | 8598/18018 [18:43:08<18:35:47, 7.11s/step, loss=1.349, acc=0.695, acc_len=1.66, thru=17.2, I=21.8, T=33.8, wait=2.9s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:31,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:13:32,791] loop.py:45 INFO TRAIN_STEP step=8599 loss=1.340504 Training: 48%|████▊ | 8599/18018 [18:43:18<21:07:47, 8.08s/step, loss=1.341, acc=0.693, acc_len=1.66, thru=14.1, I=13.1, T=28.6, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:33,253] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:40,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:13:41,575] loop.py:380 INFO TIMING step=8600: step=5.016s data=2.293s compute=4.975s [fwd=3.344s bwd=1.615s opt=0.003s] dispatch=3.760s [2026-10-19 08:13:41,575] loop.py:45 INFO TRAIN_STEP step=8600 loss=1.313182 Training: 48%|████▊ | 8600/18018 [18:43:27<21:41:05, 8.29s/step, loss=1.313, acc=0.702, acc_len=1.69, thru=13.4, I=16.1, T=25.5, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:41,829] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:47,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:13:48,751] loop.py:45 INFO TRAIN_STEP step=8601 loss=1.409579 Training: 48%|████▊ | 8601/18018 [18:43:34<20:48:23, 7.95s/step, loss=1.410, acc=0.680, acc_len=1.62, thru=15.2, I=18.8, T=33.2, wait=3.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:13:48,746] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8600: forward=2713.8ms backward=1084.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:49,202] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:54,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:13:56,074] loop.py:45 INFO TRAIN_STEP step=8602 loss=1.406416 Training: 48%|████▊ | 8602/18018 [18:43:42<20:18:41, 7.77s/step, loss=1.406, acc=0.682, acc_len=1.62, thru=17.4, I=22.5, T=31.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:13:56,372] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:03,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:14:04,615] loop.py:45 INFO TRAIN_STEP step=8603 loss=1.286092 Training: 48%|████▊ | 8603/18018 [18:43:50<20:55:04, 8.00s/step, loss=1.286, acc=0.702, acc_len=1.69, thru=13.7, I=18.7, T=29.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:04,909] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:10,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:14:10,875] loop.py:45 INFO TRAIN_STEP step=8604 loss=1.366298 Training: 48%|████▊ | 8604/18018 [18:43:56<19:33:03, 7.48s/step, loss=1.366, acc=0.689, acc_len=1.64, thru=18.7, I=22.0, T=35.9, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:11,085] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:14:18,302] loop.py:380 INFO TIMING step=8605: step=4.472s data=2.126s compute=4.439s [fwd=3.092s bwd=1.331s opt=0.003s] dispatch=2.947s [2026-10-19 08:14:18,302] loop.py:45 INFO TRAIN_STEP step=8605 loss=1.355387 Training: 48%|████▊ | 8605/18018 [18:44:04<19:30:32, 7.46s/step, loss=1.355, acc=0.695, acc_len=1.67, thru=15.3, I=17.8, T=28.6, wait=2.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:25,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:14:27,715] loop.py:45 INFO TRAIN_STEP step=8606 loss=1.502994 Training: 48%|████▊ | 8606/18018 [18:44:13<21:02:19, 8.05s/step, loss=1.503, acc=0.668, acc_len=1.57, thru=15.6, I=19.1, T=28.9, wait=5.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:14:27,710] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8605: forward=3182.8ms backward=1203.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:27,791] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=40/64 [2026-10-19 08:14:34,643] loop.py:45 INFO TRAIN_STEP step=8607 loss=1.479162 Training: 48%|████▊ | 8607/18018 [18:44:20<20:09:28, 7.71s/step, loss=1.479, acc=0.669, acc_len=1.58, thru=12.8, I=16.7, T=31.0, wait=2.8s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:41,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:14:43,572] loop.py:45 INFO TRAIN_STEP step=8608 loss=1.471192 Training: 48%|████▊ | 8608/18018 [18:44:29<21:06:47, 8.08s/step, loss=1.471, acc=0.672, acc_len=1.59, thru=17.6, I=19.4, T=29.8, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:43,739] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:49,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:14:50,151] loop.py:45 INFO TRAIN_STEP step=8609 loss=1.386104 Training: 48%|████▊ | 8609/18018 [18:44:36<19:56:05, 7.63s/step, loss=1.386, acc=0.684, acc_len=1.62, thru=17.9, I=23.4, T=31.5, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:50,511] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:55,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:14:56,932] loop.py:380 INFO TIMING step=8610: step=4.275s data=1.972s compute=4.230s [fwd=2.953s bwd=1.265s opt=0.003s] dispatch=2.496s [2026-10-19 08:14:56,932] loop.py:45 INFO TRAIN_STEP step=8610 loss=1.388789 Training: 48%|████▊ | 8610/18018 [18:44:43<19:16:10, 7.37s/step, loss=1.389, acc=0.682, acc_len=1.62, thru=18.3, I=22.4, T=29.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:14:57,373] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:15:03,551] loop.py:45 INFO TRAIN_STEP step=8611 loss=1.313668 Training: 48%|████▊ | 8611/18018 [18:44:49<18:40:36, 7.15s/step, loss=1.314, acc=0.699, acc_len=1.68, thru=15.9, I=21.3, T=32.6, wait=2.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:15:03,547] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8610: forward=2785.8ms backward=1093.6ms [2026-10-19 08:15:13,308] loop.py:45 INFO TRAIN_STEP step=8612 loss=1.445863 Training: 48%|████▊ | 8612/18018 [18:44:59<20:43:07, 7.93s/step, loss=1.446, acc=0.674, acc_len=1.60, thru=13.4, I=13.0, T=29.7, wait=5.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:19,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:15:22,459] loop.py:45 INFO TRAIN_STEP step=8613 loss=1.250219 Training: 48%|████▊ | 8613/18018 [18:45:08<21:40:37, 8.30s/step, loss=1.250, acc=0.711, acc_len=1.71, thru=15.8, I=20.4, T=24.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:22,632] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:29,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:15:30,582] loop.py:45 INFO TRAIN_STEP step=8614 loss=1.526055 Training: 48%|████▊ | 8614/18018 [18:45:16<21:32:05, 8.24s/step, loss=1.526, acc=0.664, acc_len=1.56, thru=16.1, I=18.7, T=28.8, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:30,652] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:36,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:15:38,406] loop.py:380 INFO TIMING step=8615: step=3.838s data=1.750s compute=3.794s [fwd=2.734s bwd=1.048s opt=0.003s] dispatch=3.977s [2026-10-19 08:15:38,406] loop.py:45 INFO TRAIN_STEP step=8615 loss=1.369097 Training: 48%|████▊ | 8615/18018 [18:45:24<21:12:18, 8.12s/step, loss=1.369, acc=0.688, acc_len=1.64, thru=14.9, I=19.3, T=33.4, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:38,822] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:45,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:15:47,513] loop.py:45 INFO TRAIN_STEP step=8616 loss=1.295507 Training: 48%|████▊ | 8616/18018 [18:45:33<21:58:38, 8.42s/step, loss=1.296, acc=0.699, acc_len=1.68, thru=14.7, I=16.8, T=30.4, wait=4.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:15:47,508] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8615: forward=2905.0ms backward=1261.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:47,858] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:53,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:15:54,584] loop.py:45 INFO TRAIN_STEP step=8617 loss=1.450619 Training: 48%|████▊ | 8617/18018 [18:45:40<20:55:20, 8.01s/step, loss=1.451, acc=0.677, acc_len=1.61, thru=15.0, I=22.4, T=33.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:15:54,902] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:01,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:16:02,634] loop.py:45 INFO TRAIN_STEP step=8618 loss=1.334371 Training: 48%|████▊ | 8618/18018 [18:45:48<20:56:57, 8.02s/step, loss=1.334, acc=0.694, acc_len=1.67, thru=15.1, I=20.0, T=31.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:03,097] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:11,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:16:11,683] loop.py:45 INFO TRAIN_STEP step=8619 loss=1.282534 Training: 48%|████▊ | 8619/18018 [18:45:57<21:44:57, 8.33s/step, loss=1.283, acc=0.704, acc_len=1.70, thru=14.9, I=15.1, T=33.1, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:12,163] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:17,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:16:18,205] loop.py:380 INFO TIMING step=8620: step=3.943s data=1.788s compute=3.920s [fwd=2.786s bwd=1.119s opt=0.003s] dispatch=2.570s [2026-10-19 08:16:18,205] loop.py:45 INFO TRAIN_STEP step=8620 loss=1.369957 Training: 48%|████▊ | 8620/18018 [18:46:04<20:19:51, 7.79s/step, loss=1.370, acc=0.691, acc_len=1.64, thru=19.1, I=20.4, T=32.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:18,602] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:25,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 08:16:26,603] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8620: forward=2808.6ms backward=1138.3ms [2026-10-19 08:16:26,608] loop.py:45 INFO TRAIN_STEP step=8621 loss=1.384981 Training: 48%|████▊ | 8621/18018 [18:46:12<20:48:40, 7.97s/step, loss=1.385, acc=0.684, acc_len=1.63, thru=14.6, I=16.6, T=32.1, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:26,711] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:32,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:16:33,136] loop.py:45 INFO TRAIN_STEP step=8622 loss=1.360753 Training: 48%|████▊ | 8622/18018 [18:46:19<19:40:41, 7.54s/step, loss=1.361, acc=0.689, acc_len=1.64, thru=17.3, I=20.8, T=31.3, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:33,232] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=32/64 [2026-10-19 08:16:38,813] loop.py:45 INFO TRAIN_STEP step=8623 loss=1.365934 Training: 48%|████▊ | 8623/18018 [18:46:24<18:12:58, 6.98s/step, loss=1.366, acc=0.689, acc_len=1.64, thru=19.0, I=24.1, T=37.3, wait=2.2s, pool=48, epoch=1/2][2026-10-19 08:16:45,543] loop.py:45 INFO TRAIN_STEP step=8624 loss=1.309006 Training: 48%|████▊ | 8624/18018 [18:46:31<18:01:10, 6.91s/step, loss=1.309, acc=0.699, acc_len=1.68, thru=18.8, I=19.4, T=33.7, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:55,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:16:56,820] loop.py:380 INFO TIMING step=8625: step=4.099s data=1.837s compute=4.070s [fwd=2.965s bwd=1.092s opt=0.003s] dispatch=7.168s [2026-10-19 08:16:56,820] loop.py:45 INFO TRAIN_STEP step=8625 loss=1.333571 Training: 48%|████▊ | 8625/18018 [18:46:42<21:26:22, 8.22s/step, loss=1.334, acc=0.694, acc_len=1.67, thru=12.3, I=13.4, T=31.2, wait=7.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:16:57,052] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:17:04,740] loop.py:45 INFO TRAIN_STEP step=8626 loss=1.408018 Training: 48%|████▊ | 8626/18018 [18:46:50<21:12:20, 8.13s/step, loss=1.408, acc=0.679, acc_len=1.61, thru=12.6, I=18.7, T=32.2, wait=3.9s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:17:04,736] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8625: forward=2769.3ms backward=1161.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:15,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:17:15,984] loop.py:45 INFO TRAIN_STEP step=8627 loss=1.427603 Training: 48%|████▊ | 8627/18018 [18:47:02<23:38:22, 9.06s/step, loss=1.428, acc=0.684, acc_len=1.63, thru=13.6, I=12.3, T=27.5, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:16,086] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 08:17:21,832] loop.py:45 INFO TRAIN_STEP step=8628 loss=1.239821 Training: 48%|████▊ | 8628/18018 [18:47:07<21:07:16, 8.10s/step, loss=1.240, acc=0.712, acc_len=1.72, thru=19.0, I=20.3, T=33.4, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:21,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:22,315] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:17:30,110] loop.py:45 INFO TRAIN_STEP step=8629 loss=1.332204 Training: 48%|████▊ | 8629/18018 [18:47:16<21:15:42, 8.15s/step, loss=1.332, acc=0.696, acc_len=1.66, thru=13.2, I=13.0, T=31.7, wait=4.2s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:37,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:17:39,326] loop.py:380 INFO TIMING step=8630: step=4.166s data=1.939s compute=4.124s [fwd=2.894s bwd=1.215s opt=0.003s] dispatch=5.040s [2026-10-19 08:17:39,326] loop.py:45 INFO TRAIN_STEP step=8630 loss=1.517119 Training: 48%|████▊ | 8630/18018 [18:47:25<22:05:32, 8.47s/step, loss=1.517, acc=0.667, acc_len=1.57, thru=16.4, I=18.6, T=30.7, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:39,642] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:45,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:17:45,121] loop.py:45 INFO TRAIN_STEP step=8631 loss=1.338626 Training: 48%|████▊ | 8631/18018 [18:47:31<19:59:41, 7.67s/step, loss=1.339, acc=0.693, acc_len=1.65, thru=19.6, I=23.5, T=37.8, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:17:45,116] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8630: forward=2375.1ms backward=958.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:45,519] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:51,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:17:52,281] loop.py:45 INFO TRAIN_STEP step=8632 loss=1.414957 Training: 48%|████▊ | 8632/18018 [18:47:38<19:35:47, 7.52s/step, loss=1.415, acc=0.678, acc_len=1.62, thru=17.9, I=21.2, T=33.2, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:52,569] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:57,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:17:59,280] loop.py:45 INFO TRAIN_STEP step=8633 loss=1.264303 Training: 48%|████▊ | 8633/18018 [18:47:45<19:11:20, 7.36s/step, loss=1.264, acc=0.709, acc_len=1.71, thru=18.0, I=21.4, T=34.4, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:17:59,504] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:04,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:18:05,725] loop.py:45 INFO TRAIN_STEP step=8634 loss=1.388553 Training: 48%|████▊ | 8634/18018 [18:47:51<18:28:17, 7.09s/step, loss=1.389, acc=0.688, acc_len=1.65, thru=19.2, I=23.9, T=34.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:06,154] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:11,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:18:13,738] loop.py:380 INFO TIMING step=8635: step=4.065s data=1.921s compute=4.028s [fwd=2.871s bwd=1.143s opt=0.003s] dispatch=3.938s [2026-10-19 08:18:13,738] loop.py:45 INFO TRAIN_STEP step=8635 loss=1.371511 Training: 48%|████▊ | 8635/18018 [18:47:59<19:11:41, 7.36s/step, loss=1.372, acc=0.685, acc_len=1.64, thru=15.7, I=22.1, T=31.5, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:13,775] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 08:18:20,293] loop.py:45 INFO TRAIN_STEP step=8636 loss=1.302799 Training: 48%|████▊ | 8636/18018 [18:48:06<18:33:25, 7.12s/step, loss=1.303, acc=0.702, acc_len=1.68, thru=15.7, I=20.6, T=34.8, wait=2.9s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:18:20,288] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8635: forward=2544.1ms backward=1074.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:27,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:18:27,641] loop.py:45 INFO TRAIN_STEP step=8637 loss=1.507520 Training: 48%|████▊ | 8637/18018 [18:48:13<18:44:05, 7.19s/step, loss=1.508, acc=0.670, acc_len=1.58, thru=16.9, I=17.1, T=31.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:27,961] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:33,799] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:18:37,006] loop.py:45 INFO TRAIN_STEP step=8638 loss=1.393932 Training: 48%|████▊ | 8638/18018 [18:48:23<20:26:03, 7.84s/step, loss=1.394, acc=0.688, acc_len=1.64, thru=14.2, I=20.8, T=25.7, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:37,327] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:42,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:18:43,148] loop.py:45 INFO TRAIN_STEP step=8639 loss=1.329049 Training: 48%|████▊ | 8639/18018 [18:48:29<19:06:01, 7.33s/step, loss=1.329, acc=0.697, acc_len=1.67, thru=15.7, I=26.5, T=34.6, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:43,474] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:18:50,174] loop.py:380 INFO TIMING step=8640: step=4.068s data=1.862s compute=4.024s [fwd=2.808s bwd=1.201s opt=0.003s] dispatch=2.951s [2026-10-19 08:18:50,174] loop.py:45 INFO TRAIN_STEP step=8640 loss=1.405716 Training: 48%|████▊ | 8640/18018 [18:48:36<18:51:38, 7.24s/step, loss=1.406, acc=0.682, acc_len=1.62, thru=15.2, I=17.3, T=31.5, wait=3.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:57,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:18:58,614] loop.py:45 INFO TRAIN_STEP step=8641 loss=1.519333 Training: 48%|████▊ | 8641/18018 [18:48:44<19:47:49, 7.60s/step, loss=1.519, acc=0.662, acc_len=1.56, thru=17.4, I=18.2, T=32.5, wait=4.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:18:58,609] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8640: forward=2791.2ms backward=1098.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:18:59,037] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:19:05,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:19:06,119] loop.py:45 INFO TRAIN_STEP step=8642 loss=1.234676 Training: 48%|████▊ | 8642/18018 [18:48:52<19:43:12, 7.57s/step, loss=1.235, acc=0.713, acc_len=1.73, thru=15.0, I=18.5, T=30.7, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:19:06,411] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:19:12,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:19:13,242] loop.py:45 INFO TRAIN_STEP step=8643 loss=1.411024 Training: 48%|████▊ | 8643/18018 [18:48:59<19:22:03, 7.44s/step, loss=1.411, acc=0.682, acc_len=1.63, thru=17.6, I=19.0, T=33.8, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:19:13,295] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 [2026-10-19 08:19:20,918] loop.py:45 INFO TRAIN_STEP step=8644 loss=1.435941 Training: 48%|████▊ | 8644/18018 [18:49:07<19:33:03, 7.51s/step, loss=1.436, acc=0.681, acc_len=1.61, thru=15.1, I=17.8, T=29.1, wait=3.3s, pool=60, epoch=1/2][2026-10-19 08:19:32,612] loop.py:380 INFO TIMING step=8645: step=4.378s data=2.058s compute=4.349s [fwd=3.148s bwd=1.187s opt=0.003s] dispatch=7.307s [2026-10-19 08:19:32,612] loop.py:45 INFO TRAIN_STEP step=8645 loss=1.319579 Training: 48%|████▊ | 8645/18018 [18:49:18<22:49:09, 8.76s/step, loss=1.320, acc=0.701, acc_len=1.69, thru=11.9, I=11.0, T=29.2, wait=7.3s, pool=56, epoch=1/2][2026-10-19 08:19:41,462] loop.py:45 INFO TRAIN_STEP step=8646 loss=1.361766 Training: 48%|████▊ | 8646/18018 [18:49:27<22:53:04, 8.79s/step, loss=1.362, acc=0.693, acc_len=1.66, thru=13.5, I=13.6, T=31.3, wait=4.7s, pool=44, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:19:41,458] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8645: forward=2834.6ms backward=1203.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:19:49,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:19:50,015] loop.py:45 INFO TRAIN_STEP step=8647 loss=1.280556 Training: 48%|████▊ | 8647/18018 [18:49:36<22:41:46, 8.72s/step, loss=1.281, acc=0.704, acc_len=1.70, thru=16.8, I=16.1, T=30.9, wait=4.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:19:50,521] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:19:58,015] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:19:59,467] loop.py:45 INFO TRAIN_STEP step=8648 loss=1.379633 Training: 48%|████▊ | 8648/18018 [18:49:45<23:15:56, 8.94s/step, loss=1.380, acc=0.690, acc_len=1.65, thru=14.3, I=16.9, T=33.5, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:19:59,540] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:05,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:20:07,062] loop.py:45 INFO TRAIN_STEP step=8649 loss=1.391493 Training: 48%|████▊ | 8649/18018 [18:49:53<22:12:47, 8.54s/step, loss=1.391, acc=0.684, acc_len=1.63, thru=15.4, I=21.5, T=36.7, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:07,111] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:11,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:20:13,431] loop.py:380 INFO TIMING step=8650: step=3.940s data=1.697s compute=3.912s [fwd=2.762s bwd=1.134s opt=0.003s] dispatch=2.420s [2026-10-19 08:20:13,432] loop.py:45 INFO TRAIN_STEP step=8650 loss=1.414459 Training: 48%|████▊ | 8650/18018 [18:49:59<20:31:07, 7.89s/step, loss=1.414, acc=0.687, acc_len=1.63, thru=19.8, I=26.2, T=32.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:13,931] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:19,665] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:20:19,904] loop.py:45 INFO TRAIN_STEP step=8651 loss=1.395659 Training: 48%|████▊ | 8651/18018 [18:50:05<19:24:53, 7.46s/step, loss=1.396, acc=0.684, acc_len=1.63, thru=18.5, I=26.2, T=33.1, wait=2.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:20:19,900] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8650: forward=2673.3ms backward=1153.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:20,167] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:20:27,266] loop.py:45 INFO TRAIN_STEP step=8652 loss=1.528934 Training: 48%|████▊ | 8652/18018 [18:50:13<19:20:06, 7.43s/step, loss=1.529, acc=0.666, acc_len=1.56, thru=14.1, I=15.6, T=33.9, wait=3.6s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:34,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:20:35,935] loop.py:45 INFO TRAIN_STEP step=8653 loss=1.515553 Training: 48%|████▊ | 8653/18018 [18:50:22<20:17:59, 7.80s/step, loss=1.516, acc=0.668, acc_len=1.58, thru=16.9, I=17.4, T=31.8, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:36,213] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:41,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:20:42,951] loop.py:45 INFO TRAIN_STEP step=8654 loss=1.235873 Training: 48%|████▊ | 8654/18018 [18:50:29<19:40:59, 7.57s/step, loss=1.236, acc=0.711, acc_len=1.72, thru=16.1, I=25.0, T=29.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:43,097] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:49,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:20:50,134] loop.py:380 INFO TIMING step=8655: step=3.921s data=1.676s compute=3.892s [fwd=2.755s bwd=1.122s opt=0.003s] dispatch=3.252s [2026-10-19 08:20:50,135] loop.py:45 INFO TRAIN_STEP step=8655 loss=1.354047 Training: 48%|████▊ | 8655/18018 [18:50:36<19:22:53, 7.45s/step, loss=1.354, acc=0.696, acc_len=1.67, thru=15.6, I=19.1, T=32.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:50,361] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:54,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:20:56,576] loop.py:45 INFO TRAIN_STEP step=8656 loss=1.379646 Training: 48%|████▊ | 8656/18018 [18:50:42<18:35:22, 7.15s/step, loss=1.380, acc=0.692, acc_len=1.65, thru=18.2, I=23.3, T=30.5, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:20:56,571] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8655: forward=2879.8ms backward=1259.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:20:56,865] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:01,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:21:02,904] loop.py:45 INFO TRAIN_STEP step=8657 loss=1.427578 Training: 48%|████▊ | 8657/18018 [18:50:48<17:56:51, 6.90s/step, loss=1.428, acc=0.682, acc_len=1.62, thru=18.3, I=24.6, T=34.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:03,385] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:08,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:21:10,279] loop.py:45 INFO TRAIN_STEP step=8658 loss=1.434931 Training: 48%|████▊ | 8658/18018 [18:50:56<18:18:52, 7.04s/step, loss=1.435, acc=0.680, acc_len=1.62, thru=17.2, I=23.3, T=29.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:10,327] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:17,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:21:17,402] loop.py:45 INFO TRAIN_STEP step=8659 loss=1.261232 Training: 48%|████▊ | 8659/18018 [18:51:03<18:22:27, 7.07s/step, loss=1.261, acc=0.711, acc_len=1.72, thru=16.1, I=19.5, T=27.7, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:17,569] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:22,033] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:21:24,097] loop.py:380 INFO TIMING step=8660: step=4.006s data=1.709s compute=3.973s [fwd=2.813s bwd=1.145s opt=0.003s] dispatch=2.680s [2026-10-19 08:21:24,097] loop.py:45 INFO TRAIN_STEP step=8660 loss=1.320177 Training: 48%|████▊ | 8660/18018 [18:51:10<18:04:58, 6.96s/step, loss=1.320, acc=0.696, acc_len=1.67, thru=17.4, I=21.0, T=32.0, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:24,550] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:30,314] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:21:30,484] loop.py:45 INFO TRAIN_STEP step=8661 loss=1.315327 Training: 48%|████▊ | 8661/18018 [18:51:16<17:38:08, 6.79s/step, loss=1.315, acc=0.699, acc_len=1.68, thru=18.8, I=28.3, T=35.8, wait=2.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:21:30,480] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8660: forward=2508.5ms backward=1018.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:30,817] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:34,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:21:37,739] loop.py:45 INFO TRAIN_STEP step=8662 loss=1.429976 Training: 48%|████▊ | 8662/18018 [18:51:23<18:00:04, 6.93s/step, loss=1.430, acc=0.678, acc_len=1.61, thru=17.3, I=24.9, T=27.1, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:37,947] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:43,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:21:44,487] loop.py:45 INFO TRAIN_STEP step=8663 loss=1.301728 Training: 48%|████▊ | 8663/18018 [18:51:30<17:51:34, 6.87s/step, loss=1.302, acc=0.701, acc_len=1.69, thru=14.3, I=23.9, T=34.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:44,953] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:49,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:21:51,720] loop.py:45 INFO TRAIN_STEP step=8664 loss=1.434289 Training: 48%|████▊ | 8664/18018 [18:51:37<18:08:16, 6.98s/step, loss=1.434, acc=0.679, acc_len=1.61, thru=17.9, I=26.5, T=28.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:21:52,065] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 08:22:00,973] loop.py:380 INFO TIMING step=8665: step=4.189s data=1.951s compute=4.152s [fwd=3.005s bwd=1.131s opt=0.003s] dispatch=5.056s [2026-10-19 08:22:00,973] loop.py:45 INFO TRAIN_STEP step=8665 loss=1.428030 Training: 48%|████▊ | 8665/18018 [18:51:47<19:54:27, 7.66s/step, loss=1.428, acc=0.681, acc_len=1.62, thru=12.1, I=10.9, T=30.6, wait=5.1s, pool=44, epoch=1/2][2026-10-19 08:22:08,955] loop.py:45 INFO TRAIN_STEP step=8666 loss=1.373628 Training: 48%|████▊ | 8666/18018 [18:51:55<20:09:13, 7.76s/step, loss=1.374, acc=0.691, acc_len=1.65, thru=16.6, I=14.9, T=29.8, wait=3.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:22:08,950] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8665: forward=2996.7ms backward=1248.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:22:19,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:22:20,637] loop.py:45 INFO TRAIN_STEP step=8667 loss=1.280140 Training: 48%|████▊ | 8667/18018 [18:52:06<23:12:34, 8.94s/step, loss=1.280, acc=0.710, acc_len=1.71, thru=12.4, I=13.0, T=29.9, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:22:20,822] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:22:27,274] loop.py:45 INFO TRAIN_STEP step=8668 loss=1.436692 Training: 48%|████▊ | 8668/18018 [18:52:13<21:24:56, 8.25s/step, loss=1.437, acc=0.682, acc_len=1.62, thru=15.7, I=18.6, T=34.3, wait=2.9s, pool=44, epoch=1/2][2026-10-19 08:22:34,732] loop.py:45 INFO TRAIN_STEP step=8669 loss=1.530879 Training: 48%|████▊ | 8669/18018 [18:52:20<20:48:01, 8.01s/step, loss=1.531, acc=0.660, acc_len=1.54, thru=19.0, I=15.9, T=33.3, wait=3.6s, pool=60, epoch=1/2][2026-10-19 08:22:42,314] loop.py:380 INFO TIMING step=8670: step=4.488s data=2.154s compute=4.445s [fwd=3.111s bwd=1.318s opt=0.003s] dispatch=3.085s [2026-10-19 08:22:42,314] loop.py:45 INFO TRAIN_STEP step=8670 loss=1.405021 Training: 48%|████▊ | 8670/18018 [18:52:28<20:27:50, 7.88s/step, loss=1.405, acc=0.682, acc_len=1.62, thru=17.8, I=16.7, T=28.5, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:22:48,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:22:51,000] loop.py:45 INFO TRAIN_STEP step=8671 loss=1.386912 Training: 48%|████▊ | 8671/18018 [18:52:37<21:05:26, 8.12s/step, loss=1.387, acc=0.687, acc_len=1.64, thru=15.8, I=20.3, T=27.2, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:22:50,995] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8670: forward=3397.6ms backward=1251.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:22:51,458] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 08:22:57,127] loop.py:45 INFO TRAIN_STEP step=8672 loss=1.329479 Training: 48%|████▊ | 8672/18018 [18:52:43<19:31:57, 7.52s/step, loss=1.329, acc=0.695, acc_len=1.67, thru=16.8, I=27.6, T=35.7, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:05,237] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:23:06,678] loop.py:45 INFO TRAIN_STEP step=8673 loss=1.323497 Training: 48%|████▊ | 8673/18018 [18:52:52<21:06:35, 8.13s/step, loss=1.323, acc=0.697, acc_len=1.67, thru=14.2, I=14.9, T=28.5, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:06,761] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:12,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:23:13,510] loop.py:45 INFO TRAIN_STEP step=8674 loss=1.483681 Training: 48%|████▊ | 8674/18018 [18:52:59<20:05:38, 7.74s/step, loss=1.484, acc=0.664, acc_len=1.57, thru=16.8, I=20.4, T=31.8, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:13,806] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:20,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:23:20,836] loop.py:380 INFO TIMING step=8675: step=3.904s data=1.748s compute=3.873s [fwd=2.708s bwd=1.150s opt=0.003s] dispatch=3.413s [2026-10-19 08:23:20,836] loop.py:45 INFO TRAIN_STEP step=8675 loss=1.402499 Training: 48%|████▊ | 8675/18018 [18:53:06<19:46:08, 7.62s/step, loss=1.402, acc=0.680, acc_len=1.61, thru=16.8, I=18.8, T=32.8, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:20,894] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:27,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:23:28,349] loop.py:45 INFO TRAIN_STEP step=8676 loss=1.438449 Training: 48%|████▊ | 8676/18018 [18:53:14<19:41:16, 7.59s/step, loss=1.438, acc=0.681, acc_len=1.62, thru=16.7, I=19.2, T=31.6, wait=3.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:23:28,344] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8675: forward=2885.4ms backward=1115.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:28,612] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:33,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:23:36,298] loop.py:45 INFO TRAIN_STEP step=8677 loss=1.280801 Training: 48%|████▊ | 8677/18018 [18:53:22<19:58:00, 7.70s/step, loss=1.281, acc=0.705, acc_len=1.70, thru=15.7, I=21.2, T=27.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:36,507] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:41,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:23:42,828] loop.py:45 INFO TRAIN_STEP step=8678 loss=1.441882 Training: 48%|████▊ | 8678/18018 [18:53:28<19:03:28, 7.35s/step, loss=1.442, acc=0.677, acc_len=1.61, thru=16.2, I=23.9, T=31.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:42,993] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:50,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:23:51,066] loop.py:45 INFO TRAIN_STEP step=8679 loss=1.455383 Training: 48%|████▊ | 8679/18018 [18:53:37<19:45:01, 7.61s/step, loss=1.455, acc=0.672, acc_len=1.59, thru=15.3, I=18.0, T=27.6, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:23:51,294] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:23:56,966] loop.py:380 INFO TIMING step=8680: step=3.515s data=1.369s compute=3.486s [fwd=2.457s bwd=1.014s opt=0.003s] dispatch=2.374s [2026-10-19 08:23:56,966] loop.py:45 INFO TRAIN_STEP step=8680 loss=1.343206 Training: 48%|████▊ | 8680/18018 [18:53:43<18:24:56, 7.10s/step, loss=1.343, acc=0.690, acc_len=1.65, thru=18.4, I=19.8, T=36.4, wait=2.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:03,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:24:04,557] loop.py:45 INFO TRAIN_STEP step=8681 loss=1.312061 Training: 48%|████▊ | 8681/18018 [18:53:50<18:47:39, 7.25s/step, loss=1.312, acc=0.702, acc_len=1.69, thru=16.8, I=17.9, T=32.2, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:24:04,553] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8680: forward=2773.6ms backward=1155.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:04,709] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:09,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:24:10,493] loop.py:45 INFO TRAIN_STEP step=8682 loss=1.324465 Training: 48%|████▊ | 8682/18018 [18:53:56<17:46:20, 6.85s/step, loss=1.324, acc=0.694, acc_len=1.66, thru=19.7, I=22.0, T=34.2, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:10,926] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:24:19,110] loop.py:45 INFO TRAIN_STEP step=8683 loss=1.453876 Training: 48%|████▊ | 8683/18018 [18:54:05<19:08:37, 7.38s/step, loss=1.454, acc=0.674, acc_len=1.60, thru=14.9, I=17.8, T=29.8, wait=4.3s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:26,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:24:27,459] loop.py:45 INFO TRAIN_STEP step=8684 loss=1.448487 Training: 48%|████▊ | 8684/18018 [18:54:13<19:53:38, 7.67s/step, loss=1.448, acc=0.672, acc_len=1.59, thru=15.7, I=15.1, T=26.9, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:27,683] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:34,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:24:34,560] loop.py:380 INFO TIMING step=8685: step=3.641s data=1.501s compute=3.601s [fwd=2.550s bwd=1.036s opt=0.003s] dispatch=3.450s [2026-10-19 08:24:34,561] loop.py:45 INFO TRAIN_STEP step=8685 loss=1.371834 Training: 48%|████▊ | 8685/18018 [18:54:20<19:26:44, 7.50s/step, loss=1.372, acc=0.688, acc_len=1.64, thru=18.0, I=17.1, T=35.2, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:34,915] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:40,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:24:41,671] loop.py:45 INFO TRAIN_STEP step=8686 loss=1.461543 Training: 48%|████▊ | 8686/18018 [18:54:27<19:08:25, 7.38s/step, loss=1.462, acc=0.674, acc_len=1.59, thru=18.8, I=19.2, T=31.0, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:24:41,666] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8685: forward=2884.7ms backward=1193.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:41,831] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:48,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:24:49,466] loop.py:45 INFO TRAIN_STEP step=8687 loss=1.327269 Training: 48%|████▊ | 8687/18018 [18:54:35<19:27:28, 7.51s/step, loss=1.327, acc=0.696, acc_len=1.67, thru=16.1, I=19.8, T=27.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:49,712] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:56,568] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:24:57,735] loop.py:45 INFO TRAIN_STEP step=8688 loss=1.377306 Training: 48%|████▊ | 8688/18018 [18:54:43<20:02:53, 7.74s/step, loss=1.377, acc=0.688, acc_len=1.64, thru=14.0, I=16.3, T=32.6, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:24:58,075] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:06,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:25:07,703] loop.py:45 INFO TRAIN_STEP step=8689 loss=1.457399 Training: 48%|████▊ | 8689/18018 [18:54:53<21:46:54, 8.41s/step, loss=1.457, acc=0.675, acc_len=1.61, thru=13.3, I=15.9, T=28.9, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:08,101] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:25:14,220] loop.py:380 INFO TIMING step=8690: step=3.764s data=1.664s compute=3.723s [fwd=2.628s bwd=1.080s opt=0.003s] dispatch=2.743s [2026-10-19 08:25:14,221] loop.py:45 INFO TRAIN_STEP step=8690 loss=1.370103 Training: 48%|████▊ | 8690/18018 [18:55:00<20:18:38, 7.84s/step, loss=1.370, acc=0.687, acc_len=1.64, thru=12.6, I=17.0, T=34.0, wait=2.7s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:23,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:25:24,565] loop.py:45 INFO TRAIN_STEP step=8691 loss=1.367263 Training: 48%|████▊ | 8691/18018 [18:55:10<22:15:23, 8.59s/step, loss=1.367, acc=0.692, acc_len=1.65, thru=16.0, I=17.2, T=28.5, wait=5.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:25:24,559] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8690: forward=3254.4ms backward=1180.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:25,037] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:29,794] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:25:31,061] loop.py:45 INFO TRAIN_STEP step=8692 loss=1.360026 Training: 48%|████▊ | 8692/18018 [18:55:17<20:37:41, 7.96s/step, loss=1.360, acc=0.690, acc_len=1.65, thru=17.7, I=22.4, T=30.0, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:31,303] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:38,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:25:40,161] loop.py:45 INFO TRAIN_STEP step=8693 loss=1.445407 Training: 48%|████▊ | 8693/18018 [18:55:26<21:30:31, 8.30s/step, loss=1.445, acc=0.680, acc_len=1.61, thru=14.5, I=18.2, T=27.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:40,620] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:48,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:25:49,887] loop.py:45 INFO TRAIN_STEP step=8694 loss=1.499554 Training: 48%|████▊ | 8694/18018 [18:55:35<22:36:47, 8.73s/step, loss=1.500, acc=0.673, acc_len=1.59, thru=13.8, I=14.0, T=28.5, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:25:50,309] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:25:57,088] loop.py:380 INFO TIMING step=8695: step=3.936s data=1.723s compute=3.898s [fwd=2.611s bwd=1.271s opt=0.003s] dispatch=3.255s [2026-10-19 08:25:57,089] loop.py:45 INFO TRAIN_STEP step=8695 loss=1.374972 Training: 48%|████▊ | 8695/18018 [18:55:43<21:25:17, 8.27s/step, loss=1.375, acc=0.691, acc_len=1.65, thru=11.2, I=15.8, T=32.5, wait=3.3s, pool=20, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:07,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:26:08,336] loop.py:45 INFO TRAIN_STEP step=8696 loss=1.362961 Training: 48%|████▊ | 8696/18018 [18:55:54<23:43:48, 9.16s/step, loss=1.363, acc=0.689, acc_len=1.64, thru=17.3, I=13.2, T=31.2, wait=7.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:26:08,331] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8695: forward=2672.4ms backward=1255.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:08,365] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:12,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:26:13,980] loop.py:45 INFO TRAIN_STEP step=8697 loss=1.385368 Training: 48%|████▊ | 8697/18018 [18:56:00<20:59:42, 8.11s/step, loss=1.385, acc=0.688, acc_len=1.64, thru=20.0, I=24.3, T=35.1, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:14,240] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:26:21,636] loop.py:45 INFO TRAIN_STEP step=8698 loss=1.372311 Training: 48%|████▊ | 8698/18018 [18:56:07<20:38:22, 7.97s/step, loss=1.372, acc=0.689, acc_len=1.65, thru=13.7, I=16.5, T=30.4, wait=3.4s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:28,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:26:29,643] loop.py:45 INFO TRAIN_STEP step=8699 loss=1.469012 Training: 48%|████▊ | 8699/18018 [18:56:15<20:39:50, 7.98s/step, loss=1.469, acc=0.677, acc_len=1.60, thru=16.9, I=18.5, T=29.1, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:30,049] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:33,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:26:35,969] loop.py:380 INFO TIMING step=8700: step=4.156s data=1.933s compute=4.119s [fwd=2.949s bwd=1.155s opt=0.003s] dispatch=2.160s [2026-10-19 08:26:35,969] loop.py:45 INFO TRAIN_STEP step=8700 loss=1.432434 Training: 48%|████▊ | 8700/18018 [18:56:22<19:22:33, 7.49s/step, loss=1.432, acc=0.679, acc_len=1.61, thru=18.5, I=26.2, T=30.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:36,440] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 08:26:42,975] loop.py:45 INFO TRAIN_STEP step=8701 loss=1.218923 Training: 48%|████▊ | 8701/18018 [18:56:29<18:59:58, 7.34s/step, loss=1.219, acc=0.717, acc_len=1.75, thru=16.5, I=21.6, T=28.6, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:42,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 08:26:42,970] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8700: forward=3077.2ms backward=1359.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:43,453] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:48,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:26:50,460] loop.py:45 INFO TRAIN_STEP step=8702 loss=1.308716 Training: 48%|████▊ | 8702/18018 [18:56:36<19:06:30, 7.38s/step, loss=1.309, acc=0.698, acc_len=1.69, thru=17.6, I=19.3, T=30.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:50,733] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:56,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:26:57,485] loop.py:45 INFO TRAIN_STEP step=8703 loss=1.469165 Training: 48%|████▊ | 8703/18018 [18:56:43<18:49:41, 7.28s/step, loss=1.469, acc=0.668, acc_len=1.58, thru=16.5, I=22.4, T=30.7, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:26:57,904] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:27:05,781] loop.py:45 INFO TRAIN_STEP step=8704 loss=1.449774 Training: 48%|████▊ | 8704/18018 [18:56:51<19:37:04, 7.58s/step, loss=1.450, acc=0.673, acc_len=1.60, thru=13.7, I=18.3, T=28.1, wait=3.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:14,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:27:16,020] loop.py:380 INFO TIMING step=8705: step=4.413s data=1.858s compute=4.378s [fwd=3.036s bwd=1.327s opt=0.003s] dispatch=5.819s [2026-10-19 08:27:16,020] loop.py:45 INFO TRAIN_STEP step=8705 loss=1.508896 Training: 48%|████▊ | 8705/18018 [18:57:02<21:40:46, 8.38s/step, loss=1.509, acc=0.667, acc_len=1.57, thru=13.7, I=12.6, T=29.0, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:16,210] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:20,352] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:27:22,206] loop.py:45 INFO TRAIN_STEP step=8706 loss=1.388871 Training: 48%|████▊ | 8706/18018 [18:57:08<19:58:27, 7.72s/step, loss=1.389, acc=0.685, acc_len=1.63, thru=19.1, I=23.4, T=32.4, wait=2.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:27:22,202] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8705: forward=2777.3ms backward=1116.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:22,365] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:27:29,714] loop.py:45 INFO TRAIN_STEP step=8707 loss=1.358666 Training: 48%|████▊ | 8707/18018 [18:57:15<19:48:16, 7.66s/step, loss=1.359, acc=0.691, acc_len=1.65, thru=12.9, I=18.5, T=31.8, wait=3.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:38,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:27:39,935] loop.py:45 INFO TRAIN_STEP step=8708 loss=1.445503 Training: 48%|████▊ | 8708/18018 [18:57:26<21:47:33, 8.43s/step, loss=1.446, acc=0.680, acc_len=1.61, thru=15.2, I=14.8, T=30.2, wait=6.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:40,102] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:45,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:27:46,637] loop.py:45 INFO TRAIN_STEP step=8709 loss=1.423510 Training: 48%|████▊ | 8709/18018 [18:57:32<20:27:08, 7.91s/step, loss=1.424, acc=0.678, acc_len=1.61, thru=18.0, I=22.7, T=28.3, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:46,878] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:52,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:27:54,213] loop.py:380 INFO TIMING step=8710: step=4.182s data=1.897s compute=4.141s [fwd=2.936s bwd=1.190s opt=0.003s] dispatch=3.384s [2026-10-19 08:27:54,214] loop.py:45 INFO TRAIN_STEP step=8710 loss=1.369210 Training: 48%|████▊ | 8710/18018 [18:57:40<20:11:26, 7.81s/step, loss=1.369, acc=0.688, acc_len=1.64, thru=15.6, I=22.8, T=30.6, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:54,282] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:27:59,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:28:00,113] loop.py:45 INFO TRAIN_STEP step=8711 loss=1.332023 Training: 48%|████▊ | 8711/18018 [18:57:46<18:42:31, 7.24s/step, loss=1.332, acc=0.698, acc_len=1.68, thru=18.9, I=26.8, T=36.6, wait=2.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:28:00,109] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8710: forward=2463.5ms backward=981.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:28:00,569] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:28:06,473] loop.py:45 INFO TRAIN_STEP step=8712 loss=1.350156 Training: 48%|████▊ | 8712/18018 [18:57:52<18:01:36, 6.97s/step, loss=1.350, acc=0.695, acc_len=1.67, thru=18.6, I=22.0, T=33.7, wait=2.5s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:28:16,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:28:17,043] loop.py:45 INFO TRAIN_STEP step=8713 loss=1.432214 Training: 48%|████▊ | 8713/18018 [18:58:03<20:48:41, 8.05s/step, loss=1.432, acc=0.681, acc_len=1.63, thru=14.9, I=13.1, T=31.2, wait=6.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:28:17,350] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:28:22,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:28:23,693] loop.py:45 INFO TRAIN_STEP step=8714 loss=1.500922 Training: 48%|████▊ | 8714/18018 [18:58:09<19:43:28, 7.63s/step, loss=1.501, acc=0.668, acc_len=1.58, thru=18.1, I=21.6, T=33.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:28:23,859] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:28:30,740] loop.py:380 INFO TIMING step=8715: step=3.919s data=1.737s compute=3.890s [fwd=2.760s bwd=1.114s opt=0.003s] dispatch=3.118s [2026-10-19 08:28:30,740] loop.py:45 INFO TRAIN_STEP step=8715 loss=1.515360 Training: 48%|████▊ | 8715/18018 [18:58:16<19:16:06, 7.46s/step, loss=1.515, acc=0.663, acc_len=1.56, thru=16.0, I=19.6, T=32.7, wait=3.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:28:37,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:28:38,677] loop.py:45 INFO TRAIN_STEP step=8716 loss=1.340625 Training: 48%|████▊ | 8716/18018 [18:58:24<19:38:19, 7.60s/step, loss=1.341, acc=0.694, acc_len=1.67, thru=16.7, I=17.1, T=31.8, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:28:38,672] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8715: forward=2835.7ms backward=1151.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:28:38,859] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:28:44,982] loop.py:45 INFO TRAIN_STEP step=8717 loss=1.400813 Training: 48%|████▊ | 8717/18018 [18:58:31<18:37:55, 7.21s/step, loss=1.401, acc=0.685, acc_len=1.63, thru=18.2, I=20.6, T=34.9, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:28:52,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:28:54,865] loop.py:45 INFO TRAIN_STEP step=8718 loss=1.372283 Training: 48%|████▊ | 8718/18018 [18:58:40<20:42:04, 8.01s/step, loss=1.372, acc=0.689, acc_len=1.65, thru=14.1, I=16.3, T=32.2, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:28:54,899] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 [2026-10-19 08:29:00,524] loop.py:45 INFO TRAIN_STEP step=8719 loss=1.340258 Training: 48%|████▊ | 8719/18018 [18:58:46<18:52:23, 7.31s/step, loss=1.340, acc=0.696, acc_len=1.67, thru=19.4, I=28.3, T=35.9, wait=2.1s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:06,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:29:07,624] loop.py:380 INFO TIMING step=8720: step=3.580s data=1.526s compute=3.546s [fwd=2.525s bwd=1.006s opt=0.003s] dispatch=3.512s [2026-10-19 08:29:07,624] loop.py:45 INFO TRAIN_STEP step=8720 loss=1.372197 Training: 48%|████▊ | 8720/18018 [18:58:53<18:42:43, 7.24s/step, loss=1.372, acc=0.691, acc_len=1.66, thru=17.3, I=20.1, T=35.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:07,792] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:29:13,884] loop.py:45 INFO TRAIN_STEP step=8721 loss=1.397522 Training: 48%|████▊ | 8721/18018 [18:58:59<17:56:56, 6.95s/step, loss=1.398, acc=0.683, acc_len=1.63, thru=18.8, I=25.7, T=32.1, wait=2.3s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:29:13,880] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8720: forward=2863.2ms backward=1081.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:20,450] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:29:22,176] loop.py:45 INFO TRAIN_STEP step=8722 loss=1.406994 Training: 48%|████▊ | 8722/18018 [18:59:08<18:59:04, 7.35s/step, loss=1.407, acc=0.686, acc_len=1.64, thru=16.4, I=18.3, T=27.2, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:22,460] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:27,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:29:29,369] loop.py:45 INFO TRAIN_STEP step=8723 loss=1.301896 Training: 48%|████▊ | 8723/18018 [18:59:15<18:51:36, 7.30s/step, loss=1.302, acc=0.703, acc_len=1.69, thru=14.9, I=22.7, T=35.3, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:29,827] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:35,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:29:35,674] loop.py:45 INFO TRAIN_STEP step=8724 loss=1.398800 Training: 48%|████▊ | 8724/18018 [18:59:21<18:05:03, 7.00s/step, loss=1.399, acc=0.683, acc_len=1.63, thru=19.7, I=27.0, T=36.4, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:35,848] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:41,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:29:43,171] loop.py:380 INFO TIMING step=8725: step=4.224s data=2.043s compute=4.174s [fwd=2.916s bwd=1.242s opt=0.003s] dispatch=3.263s [2026-10-19 08:29:43,171] loop.py:45 INFO TRAIN_STEP step=8725 loss=1.480545 Training: 48%|████▊ | 8725/18018 [18:59:29<18:27:49, 7.15s/step, loss=1.481, acc=0.672, acc_len=1.58, thru=16.6, I=20.3, T=30.3, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:43,239] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=40/64 [2026-10-19 08:29:49,892] loop.py:45 INFO TRAIN_STEP step=8726 loss=1.343186 Training: 48%|████▊ | 8726/18018 [18:59:35<18:07:35, 7.02s/step, loss=1.343, acc=0.694, acc_len=1.66, thru=13.5, I=17.7, T=32.7, wait=2.8s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:29:49,887] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8725: forward=2696.3ms backward=1166.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:58,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:29:58,745] loop.py:45 INFO TRAIN_STEP step=8727 loss=1.328261 Training: 48%|████▊ | 8727/18018 [18:59:44<19:32:30, 7.57s/step, loss=1.328, acc=0.699, acc_len=1.68, thru=18.1, I=15.0, T=32.8, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:29:58,856] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:04,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:30:06,955] loop.py:45 INFO TRAIN_STEP step=8728 loss=1.405154 Training: 48%|████▊ | 8728/18018 [18:59:53<20:02:02, 7.76s/step, loss=1.405, acc=0.685, acc_len=1.63, thru=15.7, I=20.5, T=28.9, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:07,012] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:12,132] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:30:13,543] loop.py:45 INFO TRAIN_STEP step=8729 loss=1.289797 Training: 48%|████▊ | 8729/18018 [18:59:59<19:07:19, 7.41s/step, loss=1.290, acc=0.702, acc_len=1.69, thru=16.7, I=24.0, T=31.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:13,662] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:20,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:30:22,824] loop.py:380 INFO TIMING step=8730: step=4.150s data=1.806s compute=4.111s [fwd=2.930s bwd=1.168s opt=0.003s] dispatch=5.121s [2026-10-19 08:30:22,824] loop.py:45 INFO TRAIN_STEP step=8730 loss=1.365408 Training: 48%|████▊ | 8730/18018 [19:00:08<20:34:01, 7.97s/step, loss=1.365, acc=0.688, acc_len=1.64, thru=14.0, I=17.3, T=30.8, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:22,995] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:27,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:30:28,917] loop.py:45 INFO TRAIN_STEP step=8731 loss=1.231601 Training: 48%|████▊ | 8731/18018 [19:00:15<19:06:35, 7.41s/step, loss=1.232, acc=0.710, acc_len=1.72, thru=18.6, I=27.6, T=32.2, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:30:28,912] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8730: forward=2813.3ms backward=1120.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:29,242] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:34,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:30:36,799] loop.py:45 INFO TRAIN_STEP step=8732 loss=1.348921 Training: 48%|████▊ | 8732/18018 [19:00:22<19:28:30, 7.55s/step, loss=1.349, acc=0.695, acc_len=1.66, thru=15.0, I=22.9, T=29.7, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:37,211] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:41,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:30:43,359] loop.py:45 INFO TRAIN_STEP step=8733 loss=1.477525 Training: 48%|████▊ | 8733/18018 [19:00:29<18:42:27, 7.25s/step, loss=1.478, acc=0.672, acc_len=1.59, thru=16.6, I=26.5, T=32.5, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:43,463] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:50,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:30:51,421] loop.py:45 INFO TRAIN_STEP step=8734 loss=1.346194 Training: 48%|████▊ | 8734/18018 [19:00:37<19:19:51, 7.50s/step, loss=1.346, acc=0.694, acc_len=1.67, thru=14.5, I=18.2, T=31.9, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:30:51,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:30:58,110] loop.py:380 INFO TIMING step=8735: step=4.064s data=1.887s compute=4.033s [fwd=2.882s bwd=1.136s opt=0.003s] dispatch=2.616s [2026-10-19 08:30:58,110] loop.py:45 INFO TRAIN_STEP step=8735 loss=1.349545 Training: 48%|████▊ | 8735/18018 [19:00:44<18:42:17, 7.25s/step, loss=1.350, acc=0.692, acc_len=1.65, thru=18.3, I=17.8, T=31.5, wait=2.6s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:05,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:31:06,955] loop.py:45 INFO TRAIN_STEP step=8736 loss=1.423231 Training: 48%|████▊ | 8736/18018 [19:00:53<19:55:57, 7.73s/step, loss=1.423, acc=0.679, acc_len=1.61, thru=16.0, I=16.5, T=25.4, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:31:06,952] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8735: forward=3477.6ms backward=1512.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:07,435] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:31:13,392] loop.py:45 INFO TRAIN_STEP step=8737 loss=1.474304 Training: 48%|████▊ | 8737/18018 [19:00:59<18:55:54, 7.34s/step, loss=1.474, acc=0.674, acc_len=1.59, thru=14.1, I=21.3, T=31.2, wait=2.3s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:21,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:31:21,637] loop.py:45 INFO TRAIN_STEP step=8738 loss=1.506932 Training: 48%|████▊ | 8738/18018 [19:01:07<19:37:29, 7.61s/step, loss=1.507, acc=0.665, acc_len=1.56, thru=16.2, I=17.0, T=31.4, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:22,003] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:30,746] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:31:31,436] loop.py:45 INFO TRAIN_STEP step=8739 loss=1.332423 Training: 49%|████▊ | 8739/18018 [19:01:17<21:18:51, 8.27s/step, loss=1.332, acc=0.696, acc_len=1.67, thru=13.3, I=13.3, T=30.3, wait=5.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:31,752] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:37,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:31:39,522] loop.py:380 INFO TIMING step=8740: step=4.607s data=1.659s compute=4.566s [fwd=3.436s bwd=1.116s opt=0.003s] dispatch=3.470s [2026-10-19 08:31:39,523] loop.py:45 INFO TRAIN_STEP step=8740 loss=1.290494 Training: 49%|████▊ | 8740/18018 [19:01:25<21:10:15, 8.21s/step, loss=1.290, acc=0.700, acc_len=1.68, thru=15.6, I=21.8, T=27.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:39,649] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:44,906] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 08:31:46,268] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8740: forward=2693.8ms backward=1092.2ms [2026-10-19 08:31:46,273] loop.py:45 INFO TRAIN_STEP step=8741 loss=1.266688 Training: 49%|████▊ | 8741/18018 [19:01:32<20:02:12, 7.78s/step, loss=1.267, acc=0.709, acc_len=1.71, thru=16.2, I=24.8, T=33.4, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:31:46,423] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 [2026-10-19 08:31:52,927] loop.py:45 INFO TRAIN_STEP step=8742 loss=1.321626 Training: 49%|████▊ | 8742/18018 [19:01:39<19:10:03, 7.44s/step, loss=1.322, acc=0.702, acc_len=1.69, thru=17.2, I=24.2, T=30.2, wait=2.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:01,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:32:02,011] loop.py:45 INFO TRAIN_STEP step=8743 loss=1.408683 Training: 49%|████▊ | 8743/18018 [19:01:48<20:26:12, 7.93s/step, loss=1.409, acc=0.680, acc_len=1.62, thru=14.1, I=14.3, T=32.1, wait=5.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:02,070] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=48/64 [2026-10-19 08:32:09,187] loop.py:45 INFO TRAIN_STEP step=8744 loss=1.396147 Training: 49%|████▊ | 8744/18018 [19:01:55<19:51:00, 7.71s/step, loss=1.396, acc=0.685, acc_len=1.64, thru=16.1, I=17.6, T=32.8, wait=3.3s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:16,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:32:18,813] loop.py:380 INFO TIMING step=8745: step=4.374s data=1.957s compute=4.345s [fwd=3.108s bwd=1.222s opt=0.003s] dispatch=5.243s [2026-10-19 08:32:18,813] loop.py:45 INFO TRAIN_STEP step=8745 loss=1.380878 Training: 49%|████▊ | 8745/18018 [19:02:04<21:19:57, 8.28s/step, loss=1.381, acc=0.686, acc_len=1.64, thru=14.0, I=16.0, T=29.3, wait=5.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:18,846] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:23,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:32:24,259] loop.py:45 INFO TRAIN_STEP step=8746 loss=1.408603 Training: 49%|████▊ | 8746/18018 [19:02:10<19:08:15, 7.43s/step, loss=1.409, acc=0.681, acc_len=1.62, thru=19.9, I=26.9, T=38.2, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:32:24,255] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8745: forward=2383.1ms backward=925.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:24,611] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:30,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:32:32,742] loop.py:45 INFO TRAIN_STEP step=8747 loss=1.249240 Training: 49%|████▊ | 8747/18018 [19:02:18<19:57:00, 7.75s/step, loss=1.249, acc=0.713, acc_len=1.73, thru=15.0, I=19.7, T=28.6, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:33,124] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:39,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:32:39,561] loop.py:45 INFO TRAIN_STEP step=8748 loss=1.245120 Training: 49%|████▊ | 8748/18018 [19:02:25<19:13:46, 7.47s/step, loss=1.245, acc=0.711, acc_len=1.72, thru=16.6, I=21.6, T=32.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:39,814] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:45,079] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:32:45,766] loop.py:45 INFO TRAIN_STEP step=8749 loss=1.480607 Training: 49%|████▊ | 8749/18018 [19:02:31<18:15:16, 7.09s/step, loss=1.481, acc=0.669, acc_len=1.58, thru=19.2, I=21.5, T=33.0, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:46,082] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:50,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:32:52,018] loop.py:380 INFO TIMING step=8750: step=4.161s data=1.908s compute=4.130s [fwd=2.900s bwd=1.216s opt=0.003s] dispatch=2.079s [2026-10-19 08:32:52,018] loop.py:45 INFO TRAIN_STEP step=8750 loss=1.469701 Training: 49%|████▊ | 8750/18018 [19:02:38<17:36:15, 6.84s/step, loss=1.470, acc=0.675, acc_len=1.59, thru=19.5, I=24.4, T=30.8, wait=2.1s, pool=64, epoch=1/2][2026-10-19 08:32:52,023] loop.py:419 INFO Saving checkpoint at step 8750... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 08:32:55,471] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0008751 (TrainerActor pid=9310) [2026-10-19 08:32:55,476] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0008001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:55,715] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:32:56,744] inference_manager.py:408 INFO Pool capacity available after 6.1s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:05,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 08:32:55,470] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0008751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 08:32:56,712] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0008001 (TrainerActor pid=9310) [2026-10-19 08:33:06,048] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8750: forward=3063.5ms backward=1252.2ms [2026-10-19 08:33:06,053] loop.py:45 INFO TRAIN_STEP step=8751 loss=1.533883 Training: 49%|████▊ | 8751/18018 [19:02:52<23:09:32, 9.00s/step, loss=1.534, acc=0.659, acc_len=1.54, thru=13.8, I=15.4, T=29.4, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:06,292] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:11,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:33:13,004] loop.py:45 INFO TRAIN_STEP step=8752 loss=1.384573 Training: 49%|████▊ | 8752/18018 [19:02:59<21:34:36, 8.38s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=18.5, I=20.7, T=32.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:13,297] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:33:18,949] loop.py:45 INFO TRAIN_STEP step=8753 loss=1.389573 Training: 49%|████▊ | 8753/18018 [19:03:05<19:41:31, 7.65s/step, loss=1.390, acc=0.688, acc_len=1.64, thru=17.6, I=22.9, T=35.6, wait=2.3s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:26,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:33:27,753] loop.py:45 INFO TRAIN_STEP step=8754 loss=1.362636 Training: 49%|████▊ | 8754/18018 [19:03:13<20:34:48, 8.00s/step, loss=1.363, acc=0.693, acc_len=1.66, thru=16.4, I=17.7, T=27.2, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:27,887] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=8/64 [2026-10-19 08:33:35,607] loop.py:380 INFO TIMING step=8755: step=3.992s data=1.677s compute=3.957s [fwd=2.821s bwd=1.122s opt=0.003s] dispatch=3.853s [2026-10-19 08:33:35,607] loop.py:45 INFO TRAIN_STEP step=8755 loss=1.365525 Training: 49%|████▊ | 8755/18018 [19:03:21<20:28:05, 7.95s/step, loss=1.366, acc=0.693, acc_len=1.66, thru=12.8, I=15.3, T=32.1, wait=3.9s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:44,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:33:45,130] loop.py:45 INFO TRAIN_STEP step=8756 loss=1.386290 Training: 49%|████▊ | 8756/18018 [19:03:31<21:40:34, 8.43s/step, loss=1.386, acc=0.686, acc_len=1.64, thru=15.1, I=14.5, T=32.2, wait=5.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:33:45,125] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8755: forward=2846.4ms backward=1071.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:45,295] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:51,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:33:51,173] loop.py:45 INFO TRAIN_STEP step=8757 loss=1.360530 Training: 49%|████▊ | 8757/18018 [19:03:37<19:49:59, 7.71s/step, loss=1.361, acc=0.693, acc_len=1.65, thru=19.2, I=19.3, T=35.4, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:51,617] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:57,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:33:57,909] loop.py:45 INFO TRAIN_STEP step=8758 loss=1.355650 Training: 49%|████▊ | 8758/18018 [19:03:43<19:04:54, 7.42s/step, loss=1.356, acc=0.695, acc_len=1.66, thru=17.6, I=20.4, T=34.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:33:58,084] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:34:04,657] loop.py:45 INFO TRAIN_STEP step=8759 loss=1.358274 Training: 49%|████▊ | 8759/18018 [19:03:50<18:33:52, 7.22s/step, loss=1.358, acc=0.691, acc_len=1.65, thru=15.6, I=19.9, T=34.4, wait=3.0s, pool=36, epoch=1/2][2026-10-19 08:34:15,771] loop.py:380 INFO TIMING step=8760: step=4.385s data=1.943s compute=4.350s [fwd=3.078s bwd=1.257s opt=0.003s] dispatch=6.717s [2026-10-19 08:34:15,772] loop.py:45 INFO TRAIN_STEP step=8760 loss=1.391876 Training: 49%|████▊ | 8760/18018 [19:04:01<21:33:54, 8.39s/step, loss=1.392, acc=0.683, acc_len=1.63, thru=14.1, I=13.4, T=29.2, wait=6.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:34:26,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:34:29,104] loop.py:45 INFO TRAIN_STEP step=8761 loss=1.442318 Training: 49%|████▊ | 8761/18018 [19:04:15<25:22:48, 9.87s/step, loss=1.442, acc=0.679, acc_len=1.61, thru=9.3, I=10.3, T=22.6, wait=7.7s, pool=64, epoch=1/2] (TrainerActor pid=9310) [2026-10-19 08:34:29,099] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8760: forward=3902.7ms backward=1695.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:34:29,514] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:34:36,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:34:38,964] loop.py:45 INFO TRAIN_STEP step=8762 loss=1.489070 Training: 49%|████▊ | 8762/18018 [19:04:25<25:22:11, 9.87s/step, loss=1.489, acc=0.669, acc_len=1.58, thru=13.6, I=18.7, T=28.1, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:34:39,417] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:34:44,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:34:45,804] loop.py:45 INFO TRAIN_STEP step=8763 loss=1.298813 Training: 49%|████▊ | 8763/18018 [19:04:31<23:01:50, 8.96s/step, loss=1.299, acc=0.700, acc_len=1.68, thru=14.8, I=26.0, T=27.5, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:34:46,279] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:34:52,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:34:54,198] loop.py:45 INFO TRAIN_STEP step=8764 loss=1.356961 Training: 49%|████▊ | 8764/18018 [19:04:40<22:35:43, 8.79s/step, loss=1.357, acc=0.692, acc_len=1.65, thru=13.3, I=17.9, T=27.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:34:54,379] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:34:58,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:35:00,569] loop.py:380 INFO TIMING step=8765: step=4.206s data=1.944s compute=4.165s [fwd=2.918s bwd=1.231s opt=0.003s] dispatch=2.153s [2026-10-19 08:35:00,569] loop.py:45 INFO TRAIN_STEP step=8765 loss=1.452104 Training: 49%|████▊ | 8765/18018 [19:04:46<20:43:32, 8.06s/step, loss=1.452, acc=0.676, acc_len=1.60, thru=17.8, I=22.9, T=30.4, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:00,670] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:07,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:35:08,991] loop.py:45 INFO TRAIN_STEP step=8766 loss=1.365434 Training: 49%|████▊ | 8766/18018 [19:04:55<21:00:02, 8.17s/step, loss=1.365, acc=0.690, acc_len=1.65, thru=15.4, I=17.6, T=26.7, wait=3.6s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:35:08,987] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8765: forward=3365.1ms backward=1380.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:09,382] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:15,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:35:17,075] loop.py:45 INFO TRAIN_STEP step=8767 loss=1.371304 Training: 49%|████▊ | 8767/18018 [19:05:03<20:55:46, 8.14s/step, loss=1.371, acc=0.685, acc_len=1.63, thru=15.1, I=22.3, T=27.2, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:17,177] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:21,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:35:23,388] loop.py:45 INFO TRAIN_STEP step=8768 loss=1.426058 Training: 49%|████▊ | 8768/18018 [19:05:09<19:31:01, 7.60s/step, loss=1.426, acc=0.680, acc_len=1.61, thru=17.2, I=27.3, T=34.0, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:23,634] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:28,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:35:29,465] loop.py:45 INFO TRAIN_STEP step=8769 loss=1.409787 Training: 49%|████▊ | 8769/18018 [19:05:15<18:20:39, 7.14s/step, loss=1.410, acc=0.684, acc_len=1.63, thru=20.6, I=29.7, T=33.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:29,529] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:33,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:35:36,769] loop.py:380 INFO TIMING step=8770: step=5.012s data=2.806s compute=4.977s [fwd=3.735s bwd=1.227s opt=0.003s] dispatch=2.282s [2026-10-19 08:35:36,770] loop.py:45 INFO TRAIN_STEP step=8770 loss=1.235747 Training: 49%|████▊ | 8770/18018 [19:05:22<18:28:02, 7.19s/step, loss=1.236, acc=0.709, acc_len=1.72, thru=16.3, I=27.2, T=25.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:36,978] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 08:35:43,964] loop.py:45 INFO TRAIN_STEP step=8771 loss=1.275484 Training: 49%|████▊ | 8771/18018 [19:05:30<18:28:15, 7.19s/step, loss=1.275, acc=0.705, acc_len=1.70, thru=14.9, I=21.4, T=32.3, wait=3.2s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:35:43,959] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8770: forward=2786.0ms backward=1131.8ms [2026-10-19 08:35:52,895] loop.py:45 INFO TRAIN_STEP step=8772 loss=1.478660 Training: 49%|████▊ | 8772/18018 [19:05:38<19:48:28, 7.71s/step, loss=1.479, acc=0.672, acc_len=1.58, thru=14.4, I=13.2, T=33.9, wait=5.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:59,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:35:59,838] loop.py:45 INFO TRAIN_STEP step=8773 loss=1.386521 Training: 49%|████▊ | 8773/18018 [19:05:45<19:12:54, 7.48s/step, loss=1.387, acc=0.685, acc_len=1.63, thru=18.4, I=17.6, T=32.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:35:59,969] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:07,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:36:07,842] loop.py:45 INFO TRAIN_STEP step=8774 loss=1.415717 Training: 49%|████▊ | 8774/18018 [19:05:53<19:36:45, 7.64s/step, loss=1.416, acc=0.678, acc_len=1.61, thru=15.2, I=15.9, T=31.6, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:08,327] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:14,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:36:16,177] loop.py:380 INFO TIMING step=8775: step=5.762s data=2.882s compute=5.711s [fwd=3.813s bwd=1.883s opt=0.003s] dispatch=2.565s [2026-10-19 08:36:16,177] loop.py:45 INFO TRAIN_STEP step=8775 loss=1.325420 Training: 49%|████▊ | 8775/18018 [19:06:02<20:08:58, 7.85s/step, loss=1.325, acc=0.696, acc_len=1.67, thru=15.4, I=22.3, T=22.2, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:16,388] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:36:24,721] loop.py:45 INFO TRAIN_STEP step=8776 loss=1.495800 Training: 49%|████▊ | 8776/18018 [19:06:10<20:41:00, 8.06s/step, loss=1.496, acc=0.670, acc_len=1.58, thru=14.4, I=13.1, T=29.2, wait=4.2s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:36:24,716] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8775: forward=2946.2ms backward=1384.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:32,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:36:34,463] loop.py:45 INFO TRAIN_STEP step=8777 loss=1.360561 Training: 49%|████▊ | 8777/18018 [19:06:20<21:58:45, 8.56s/step, loss=1.361, acc=0.691, acc_len=1.66, thru=14.1, I=17.4, T=27.1, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:34,722] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:39,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:36:40,658] loop.py:45 INFO TRAIN_STEP step=8778 loss=1.302132 Training: 49%|████▊ | 8778/18018 [19:06:26<20:09:12, 7.85s/step, loss=1.302, acc=0.701, acc_len=1.68, thru=17.2, I=25.4, T=36.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:40,895] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:46,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:36:47,675] loop.py:45 INFO TRAIN_STEP step=8779 loss=1.343776 Training: 49%|████▊ | 8779/18018 [19:06:33<19:30:23, 7.60s/step, loss=1.344, acc=0.690, acc_len=1.65, thru=18.4, I=21.9, T=32.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:47,968] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:52,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:36:54,062] loop.py:380 INFO TIMING step=8780: step=3.894s data=1.707s compute=3.862s [fwd=2.728s bwd=1.120s opt=0.003s] dispatch=2.485s [2026-10-19 08:36:54,062] loop.py:45 INFO TRAIN_STEP step=8780 loss=1.453146 Training: 49%|████▊ | 8780/18018 [19:06:40<18:34:18, 7.24s/step, loss=1.453, acc=0.676, acc_len=1.60, thru=18.9, I=22.5, T=32.9, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:36:54,459] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:02,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:37:03,738] loop.py:45 INFO TRAIN_STEP step=8781 loss=1.547397 Training: 49%|████▊ | 8781/18018 [19:06:49<20:26:45, 7.97s/step, loss=1.547, acc=0.663, acc_len=1.56, thru=13.8, I=15.6, T=30.2, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:37:03,733] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8780: forward=2974.2ms backward=1203.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:04,212] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:09,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:37:10,573] loop.py:45 INFO TRAIN_STEP step=8782 loss=1.385818 Training: 49%|████▊ | 8782/18018 [19:06:56<19:34:22, 7.63s/step, loss=1.386, acc=0.688, acc_len=1.64, thru=17.4, I=22.7, T=32.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:10,782] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:16,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:37:17,171] loop.py:45 INFO TRAIN_STEP step=8783 loss=1.421828 Training: 49%|████▊ | 8783/18018 [19:07:03<18:46:35, 7.32s/step, loss=1.422, acc=0.677, acc_len=1.60, thru=18.4, I=20.9, T=31.9, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:17,669] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:37:24,921] loop.py:45 INFO TRAIN_STEP step=8784 loss=1.374162 Training: 49%|████▉ | 8784/18018 [19:07:11<19:06:18, 7.45s/step, loss=1.374, acc=0.689, acc_len=1.65, thru=11.2, I=13.5, T=33.3, wait=3.9s, pool=24, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 08:37:25,007] eagle_store.py:517 WARNING batch_get_buffer missing keys (k281095-ab10134f_hs, k281095-ab10134f_ids, k281095-ab10134f_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:32,047] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:37:34,142] loop.py:380 INFO TIMING step=8785: step=4.385s data=1.910s compute=4.347s [fwd=2.985s bwd=1.347s opt=0.003s] dispatch=4.827s [2026-10-19 08:37:34,142] loop.py:45 INFO TRAIN_STEP step=8785 loss=1.359635 Training: 49%|████▉ | 8785/18018 [19:07:20<20:28:04, 7.98s/step, loss=1.360, acc=0.690, acc_len=1.64, thru=18.1, I=17.4, T=29.2, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:34,566] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:41,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:37:42,222] loop.py:45 INFO TRAIN_STEP step=8786 loss=1.400558 Training: 49%|████▉ | 8786/18018 [19:07:28<20:32:32, 8.01s/step, loss=1.401, acc=0.683, acc_len=1.62, thru=16.7, I=19.2, T=32.8, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:37:42,218] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8785: forward=2711.6ms backward=1133.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:42,273] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:49,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:37:49,555] loop.py:45 INFO TRAIN_STEP step=8787 loss=1.432475 Training: 49%|████▉ | 8787/18018 [19:07:35<20:01:03, 7.81s/step, loss=1.432, acc=0.679, acc_len=1.61, thru=17.4, I=19.3, T=30.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:49,736] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:56,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:37:57,794] loop.py:45 INFO TRAIN_STEP step=8788 loss=1.382927 Training: 49%|████▉ | 8788/18018 [19:07:43<20:20:57, 7.94s/step, loss=1.383, acc=0.686, acc_len=1.63, thru=16.0, I=17.4, T=32.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:37:57,968] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:06,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:38:06,794] loop.py:45 INFO TRAIN_STEP step=8789 loss=1.476710 Training: 49%|████▉ | 8789/18018 [19:07:52<21:09:50, 8.26s/step, loss=1.477, acc=0.670, acc_len=1.58, thru=13.7, I=15.1, T=29.9, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:06,934] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:12,461] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:38:13,734] loop.py:380 INFO TIMING step=8790: step=3.873s data=1.753s compute=3.841s [fwd=2.719s bwd=1.107s opt=0.003s] dispatch=3.058s [2026-10-19 08:38:13,734] loop.py:45 INFO TRAIN_STEP step=8790 loss=1.394104 Training: 49%|████▉ | 8790/18018 [19:07:59<20:09:02, 7.86s/step, loss=1.394, acc=0.686, acc_len=1.64, thru=18.0, I=20.0, T=33.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:13,970] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:18,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:38:20,071] loop.py:45 INFO TRAIN_STEP step=8791 loss=1.308138 Training: 49%|████▉ | 8791/18018 [19:08:06<18:58:30, 7.40s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=19.4, I=25.6, T=33.2, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:38:20,067] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8790: forward=2719.4ms backward=1086.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:20,251] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:38:25,694] loop.py:45 INFO TRAIN_STEP step=8792 loss=1.415035 Training: 49%|████▉ | 8792/18018 [19:08:11<17:36:22, 6.87s/step, loss=1.415, acc=0.678, acc_len=1.61, thru=18.9, I=24.8, T=34.9, wait=1.9s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:32,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:38:34,872] loop.py:45 INFO TRAIN_STEP step=8793 loss=1.447812 Training: 49%|████▉ | 8793/18018 [19:08:20<19:22:42, 7.56s/step, loss=1.448, acc=0.675, acc_len=1.60, thru=16.1, I=18.3, T=28.3, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:34,983] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:40,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:38:42,097] loop.py:45 INFO TRAIN_STEP step=8794 loss=1.330399 Training: 49%|████▉ | 8794/18018 [19:08:28<19:06:55, 7.46s/step, loss=1.330, acc=0.693, acc_len=1.66, thru=14.5, I=22.8, T=32.6, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:42,458] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:50,665] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:38:51,670] loop.py:380 INFO TIMING step=8795: step=4.534s data=2.132s compute=4.493s [fwd=3.244s bwd=1.234s opt=0.003s] dispatch=5.031s [2026-10-19 08:38:51,670] loop.py:45 INFO TRAIN_STEP step=8795 loss=1.383451 Training: 49%|████▉ | 8795/18018 [19:08:37<20:44:11, 8.09s/step, loss=1.383, acc=0.691, acc_len=1.65, thru=13.9, I=15.3, T=28.2, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:52,171] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:59,207] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:38:59,388] loop.py:45 INFO TRAIN_STEP step=8796 loss=1.321781 Training: 49%|████▉ | 8796/18018 [19:08:45<20:26:44, 7.98s/step, loss=1.322, acc=0.696, acc_len=1.67, thru=17.3, I=20.0, T=34.3, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:38:59,383] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8795: forward=2575.3ms backward=1077.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:38:59,710] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:06,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:39:08,812] loop.py:45 INFO TRAIN_STEP step=8797 loss=1.514436 Training: 49%|████▉ | 8797/18018 [19:08:54<21:33:10, 8.41s/step, loss=1.514, acc=0.664, acc_len=1.56, thru=13.6, I=18.3, T=22.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:09,148] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 08:39:15,514] loop.py:45 INFO TRAIN_STEP step=8798 loss=1.387219 Training: 49%|████▉ | 8798/18018 [19:09:01<20:14:07, 7.90s/step, loss=1.387, acc=0.685, acc_len=1.63, thru=14.1, I=19.7, T=33.9, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:20,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:39:22,667] loop.py:45 INFO TRAIN_STEP step=8799 loss=1.346088 Training: 49%|████▉ | 8799/18018 [19:09:08<19:39:26, 7.68s/step, loss=1.346, acc=0.692, acc_len=1.65, thru=18.3, I=21.3, T=29.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:22,698] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:27,893] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:39:30,136] loop.py:380 INFO TIMING step=8800: step=4.191s data=1.962s compute=4.128s [fwd=2.829s bwd=1.284s opt=0.003s] dispatch=3.268s [2026-10-19 08:39:30,136] loop.py:45 INFO TRAIN_STEP step=8800 loss=1.365428 Training: 49%|████▉ | 8800/18018 [19:09:16<19:29:47, 7.61s/step, loss=1.365, acc=0.690, acc_len=1.65, thru=15.6, I=26.9, T=30.5, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:30,408] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:37,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:39:37,746] loop.py:45 INFO TRAIN_STEP step=8801 loss=1.323560 Training: 49%|████▉ | 8801/18018 [19:09:23<19:29:23, 7.61s/step, loss=1.324, acc=0.695, acc_len=1.66, thru=13.4, I=18.6, T=27.2, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:39:37,742] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8800: forward=3336.6ms backward=1335.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:38,160] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:45,306] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:39:45,610] loop.py:45 INFO TRAIN_STEP step=8802 loss=1.390622 Training: 49%|████▉ | 8802/18018 [19:09:31<19:40:58, 7.69s/step, loss=1.391, acc=0.688, acc_len=1.64, thru=16.1, I=16.7, T=26.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:39:45,809] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:39:52,806] loop.py:45 INFO TRAIN_STEP step=8803 loss=1.320565 Training: 49%|████▉ | 8803/18018 [19:09:38<19:18:06, 7.54s/step, loss=1.321, acc=0.699, acc_len=1.68, thru=14.1, I=15.6, T=31.6, wait=3.1s, pool=40, epoch=1/2][2026-10-19 08:40:00,559] loop.py:45 INFO TRAIN_STEP step=8804 loss=1.364314 Training: 49%|████▉ | 8804/18018 [19:09:46<19:27:42, 7.60s/step, loss=1.364, acc=0.691, acc_len=1.66, thru=14.1, I=13.5, T=33.4, wait=3.9s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:10,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:40:11,533] loop.py:380 INFO TIMING step=8805: step=4.146s data=1.875s compute=4.086s [fwd=2.897s bwd=1.174s opt=0.003s] dispatch=6.821s [2026-10-19 08:40:11,533] loop.py:45 INFO TRAIN_STEP step=8805 loss=1.344731 Training: 49%|████▉ | 8805/18018 [19:09:57<22:02:50, 8.62s/step, loss=1.345, acc=0.698, acc_len=1.68, thru=14.7, I=15.6, T=30.9, wait=6.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:11,917] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:16,885] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 08:40:18,274] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8805: forward=2851.4ms backward=1342.9ms [2026-10-19 08:40:18,279] loop.py:45 INFO TRAIN_STEP step=8806 loss=1.354595 Training: 49%|████▉ | 8806/18018 [19:10:04<20:36:36, 8.05s/step, loss=1.355, acc=0.690, acc_len=1.65, thru=17.3, I=22.9, T=30.2, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:18,396] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:25,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:40:25,742] loop.py:45 INFO TRAIN_STEP step=8807 loss=1.457296 Training: 49%|████▉ | 8807/18018 [19:10:11<20:09:14, 7.88s/step, loss=1.457, acc=0.677, acc_len=1.60, thru=15.7, I=19.1, T=33.0, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:25,950] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:40:32,497] loop.py:45 INFO TRAIN_STEP step=8808 loss=1.315437 Training: 49%|████▉ | 8808/18018 [19:10:18<19:17:30, 7.54s/step, loss=1.315, acc=0.701, acc_len=1.68, thru=17.4, I=17.2, T=28.1, wait=2.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:40,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:40:41,371] loop.py:45 INFO TRAIN_STEP step=8809 loss=1.389816 Training: 49%|████▉ | 8809/18018 [19:10:27<20:18:41, 7.94s/step, loss=1.390, acc=0.687, acc_len=1.63, thru=16.4, I=17.0, T=31.0, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:41,770] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:47,565] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:40:48,441] loop.py:380 INFO TIMING step=8810: step=4.077s data=1.937s compute=4.042s [fwd=2.957s bwd=1.070s opt=0.003s] dispatch=2.984s [2026-10-19 08:40:48,441] loop.py:45 INFO TRAIN_STEP step=8810 loss=1.318019 Training: 49%|████▉ | 8810/18018 [19:10:34<19:38:36, 7.68s/step, loss=1.318, acc=0.695, acc_len=1.67, thru=17.1, I=21.2, T=31.4, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:48,582] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=16/64 [2026-10-19 08:40:55,582] loop.py:45 INFO TRAIN_STEP step=8811 loss=1.403843 Training: 49%|████▉ | 8811/18018 [19:10:41<19:13:29, 7.52s/step, loss=1.404, acc=0.682, acc_len=1.62, thru=17.0, I=17.8, T=28.9, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:40:55,577] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8810: forward=3266.1ms backward=1109.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:55,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:40:56,076] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:03,916] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:41:04,176] loop.py:45 INFO TRAIN_STEP step=8812 loss=1.391372 Training: 49%|████▉ | 8812/18018 [19:10:50<20:02:59, 7.84s/step, loss=1.391, acc=0.685, acc_len=1.63, thru=15.6, I=16.0, T=32.6, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:04,418] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:11,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:41:11,539] loop.py:45 INFO TRAIN_STEP step=8813 loss=1.398569 Training: 49%|████▉ | 8813/18018 [19:10:57<19:40:56, 7.70s/step, loss=1.399, acc=0.685, acc_len=1.62, thru=16.5, I=17.8, T=31.1, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:11,770] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:18,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:41:18,976] loop.py:45 INFO TRAIN_STEP step=8814 loss=1.379572 Training: 49%|████▉ | 8814/18018 [19:11:05<19:28:47, 7.62s/step, loss=1.380, acc=0.691, acc_len=1.65, thru=16.5, I=17.2, T=30.0, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:19,409] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:24,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:41:25,707] loop.py:380 INFO TIMING step=8815: step=4.056s data=1.716s compute=4.023s [fwd=2.771s bwd=1.238s opt=0.003s] dispatch=2.668s [2026-10-19 08:41:25,708] loop.py:45 INFO TRAIN_STEP step=8815 loss=1.379972 Training: 49%|████▉ | 8815/18018 [19:11:11<18:47:48, 7.35s/step, loss=1.380, acc=0.690, acc_len=1.65, thru=18.1, I=21.5, T=31.6, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:26,082] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:41:31,848] loop.py:45 INFO TRAIN_STEP step=8816 loss=1.430704 Training: 49%|████▉ | 8816/18018 [19:11:17<17:51:56, 6.99s/step, loss=1.431, acc=0.679, acc_len=1.61, thru=17.4, I=23.7, T=37.3, wait=2.7s, pool=48, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:41:31,844] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8815: forward=2405.1ms backward=975.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:38,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:41:40,081] loop.py:45 INFO TRAIN_STEP step=8817 loss=1.304116 Training: 49%|████▉ | 8817/18018 [19:11:26<18:49:04, 7.36s/step, loss=1.304, acc=0.702, acc_len=1.69, thru=16.5, I=18.2, T=28.3, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:40,137] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:46,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:41:46,667] loop.py:45 INFO TRAIN_STEP step=8818 loss=1.378764 Training: 49%|████▉ | 8818/18018 [19:11:32<18:13:07, 7.13s/step, loss=1.379, acc=0.685, acc_len=1.63, thru=17.2, I=20.5, T=34.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:46,969] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:54,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:41:55,700] loop.py:45 INFO TRAIN_STEP step=8819 loss=1.195968 Training: 49%|████▉ | 8819/18018 [19:11:41<19:40:36, 7.70s/step, loss=1.196, acc=0.721, acc_len=1.75, thru=14.8, I=17.3, T=32.0, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:41:55,917] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:02,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:42:04,172] loop.py:380 INFO TIMING step=8820: step=4.266s data=1.979s compute=4.235s [fwd=2.916s bwd=1.304s opt=0.003s] dispatch=4.199s [2026-10-19 08:42:04,173] loop.py:45 INFO TRAIN_STEP step=8820 loss=1.241441 Training: 49%|████▉ | 8820/18018 [19:11:50<20:16:04, 7.93s/step, loss=1.241, acc=0.715, acc_len=1.73, thru=13.5, I=19.0, T=30.0, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:04,361] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:10,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:42:11,158] loop.py:45 INFO TRAIN_STEP step=8821 loss=1.326087 Training: 49%|████▉ | 8821/18018 [19:11:57<19:32:18, 7.65s/step, loss=1.326, acc=0.695, acc_len=1.66, thru=16.9, I=22.5, T=32.5, wait=3.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:42:11,154] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8820: forward=2701.1ms backward=1190.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:11,315] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:17,307] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:42:17,678] loop.py:45 INFO TRAIN_STEP step=8822 loss=1.340186 Training: 49%|████▉ | 8822/18018 [19:12:03<18:40:22, 7.31s/step, loss=1.340, acc=0.689, acc_len=1.64, thru=18.0, I=21.1, T=28.3, wait=2.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:17,820] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 [2026-10-19 08:42:25,592] loop.py:45 INFO TRAIN_STEP step=8823 loss=1.285090 Training: 49%|████▉ | 8823/18018 [19:12:11<19:07:55, 7.49s/step, loss=1.285, acc=0.706, acc_len=1.70, thru=15.6, I=15.7, T=32.5, wait=4.0s, pool=60, epoch=1/2][2026-10-19 08:42:33,341] loop.py:45 INFO TRAIN_STEP step=8824 loss=1.393221 Training: 49%|████▉ | 8824/18018 [19:12:19<19:19:54, 7.57s/step, loss=1.393, acc=0.684, acc_len=1.63, thru=16.0, I=16.7, T=31.6, wait=3.7s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:40,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:42:41,726] loop.py:380 INFO TIMING step=8825: step=4.062s data=1.907s compute=4.013s [fwd=2.843s bwd=1.155s opt=0.003s] dispatch=4.311s [2026-10-19 08:42:41,726] loop.py:45 INFO TRAIN_STEP step=8825 loss=1.527954 Training: 49%|████▉ | 8825/18018 [19:12:27<19:57:07, 7.81s/step, loss=1.528, acc=0.663, acc_len=1.55, thru=16.0, I=16.7, T=31.5, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:42,135] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:42:48,792] loop.py:45 INFO TRAIN_STEP step=8826 loss=1.353870 Training: 49%|████▉ | 8826/18018 [19:12:34<19:22:40, 7.59s/step, loss=1.354, acc=0.694, acc_len=1.67, thru=15.0, I=18.0, T=32.4, wait=3.1s, pool=40, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:42:48,789] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8825: forward=2756.4ms backward=1156.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:56,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:42:57,305] loop.py:45 INFO TRAIN_STEP step=8827 loss=1.308593 Training: 49%|████▉ | 8827/18018 [19:12:43<20:04:53, 7.87s/step, loss=1.309, acc=0.701, acc_len=1.68, thru=17.6, I=16.5, T=31.7, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:42:57,703] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:03,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:43:03,770] loop.py:45 INFO TRAIN_STEP step=8828 loss=1.293456 Training: 49%|████▉ | 8828/18018 [19:12:49<19:00:35, 7.45s/step, loss=1.293, acc=0.701, acc_len=1.69, thru=19.0, I=20.9, T=31.7, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:04,157] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:43:10,388] loop.py:45 INFO TRAIN_STEP step=8829 loss=1.315806 Training: 49%|████▉ | 8829/18018 [19:12:56<18:22:11, 7.20s/step, loss=1.316, acc=0.697, acc_len=1.66, thru=15.7, I=16.1, T=35.3, wait=3.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:17,746] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:43:18,178] loop.py:380 INFO TIMING step=8830: step=3.968s data=1.893s compute=3.920s [fwd=2.748s bwd=1.157s opt=0.003s] dispatch=3.814s [2026-10-19 08:43:18,178] loop.py:45 INFO TRAIN_STEP step=8830 loss=1.457126 Training: 49%|████▉ | 8830/18018 [19:13:04<18:49:25, 7.38s/step, loss=1.457, acc=0.673, acc_len=1.59, thru=18.9, I=17.6, T=32.3, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:18,253] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 [2026-10-19 08:43:26,226] loop.py:45 INFO TRAIN_STEP step=8831 loss=1.388488 Training: 49%|████▉ | 8831/18018 [19:13:12<19:20:07, 7.58s/step, loss=1.388, acc=0.683, acc_len=1.63, thru=15.4, I=15.4, T=32.2, wait=4.1s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:43:26,221] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8830: forward=2820.2ms backward=1101.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:32,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:43:33,461] loop.py:45 INFO TRAIN_STEP step=8832 loss=1.422337 Training: 49%|████▉ | 8832/18018 [19:13:19<19:04:25, 7.47s/step, loss=1.422, acc=0.684, acc_len=1.62, thru=18.9, I=18.4, T=31.1, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:33,759] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:43:40,371] loop.py:45 INFO TRAIN_STEP step=8833 loss=1.369781 Training: 49%|████▉ | 8833/18018 [19:13:26<18:38:19, 7.31s/step, loss=1.370, acc=0.686, acc_len=1.64, thru=17.3, I=20.2, T=37.0, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:46,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:43:47,505] loop.py:45 INFO TRAIN_STEP step=8834 loss=1.326358 Training: 49%|████▉ | 8834/18018 [19:13:33<18:30:21, 7.25s/step, loss=1.326, acc=0.699, acc_len=1.68, thru=18.1, I=20.3, T=32.2, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:47,896] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:54,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:43:55,212] loop.py:380 INFO TIMING step=8835: step=3.676s data=1.576s compute=3.638s [fwd=2.565s bwd=1.059s opt=0.003s] dispatch=4.020s [2026-10-19 08:43:55,213] loop.py:45 INFO TRAIN_STEP step=8835 loss=1.422907 Training: 49%|████▉ | 8835/18018 [19:13:41<18:51:01, 7.39s/step, loss=1.423, acc=0.683, acc_len=1.62, thru=16.4, I=20.2, T=34.8, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:43:55,407] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:02,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:44:03,413] loop.py:45 INFO TRAIN_STEP step=8836 loss=1.279822 Training: 49%|████▉ | 8836/18018 [19:13:49<19:28:08, 7.63s/step, loss=1.280, acc=0.701, acc_len=1.69, thru=15.7, I=20.2, T=28.9, wait=3.8s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:44:03,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8835: forward=3069.4ms backward=1320.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:03,530] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:09,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:44:10,537] loop.py:45 INFO TRAIN_STEP step=8837 loss=1.342752 Training: 49%|████▉ | 8837/18018 [19:13:56<19:04:30, 7.48s/step, loss=1.343, acc=0.694, acc_len=1.67, thru=15.7, I=19.5, T=30.5, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:10,846] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:17,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:44:18,670] loop.py:45 INFO TRAIN_STEP step=8838 loss=1.363159 Training: 49%|████▉ | 8838/18018 [19:14:04<19:34:20, 7.68s/step, loss=1.363, acc=0.692, acc_len=1.66, thru=15.2, I=18.2, T=28.9, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:18,746] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:24,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:44:25,065] loop.py:45 INFO TRAIN_STEP step=8839 loss=1.378746 Training: 49%|████▉ | 8839/18018 [19:14:11<18:35:31, 7.29s/step, loss=1.379, acc=0.689, acc_len=1.65, thru=17.9, I=21.2, T=30.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:25,470] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:33,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:44:33,343] loop.py:380 INFO TIMING step=8840: step=4.021s data=1.669s compute=3.990s [fwd=2.830s bwd=1.144s opt=0.003s] dispatch=4.248s [2026-10-19 08:44:33,343] loop.py:45 INFO TRAIN_STEP step=8840 loss=1.366124 Training: 49%|████▉ | 8840/18018 [19:14:19<19:20:37, 7.59s/step, loss=1.366, acc=0.688, acc_len=1.64, thru=16.3, I=18.3, T=31.8, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:33,688] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:44:40,323] loop.py:45 INFO TRAIN_STEP step=8841 loss=1.536211 Training: 49%|████▉ | 8841/18018 [19:14:26<18:52:37, 7.41s/step, loss=1.536, acc=0.659, acc_len=1.55, thru=18.2, I=17.6, T=31.3, wait=2.9s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:44:40,318] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8840: forward=2722.3ms backward=1308.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:47,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:44:47,812] loop.py:45 INFO TRAIN_STEP step=8842 loss=1.338676 Training: 49%|████▉ | 8842/18018 [19:14:33<18:56:22, 7.43s/step, loss=1.339, acc=0.695, acc_len=1.67, thru=17.4, I=17.3, T=28.3, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:44:48,018] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:44:56,111] loop.py:45 INFO TRAIN_STEP step=8843 loss=1.338336 Training: 49%|████▉ | 8843/18018 [19:14:42<19:36:10, 7.69s/step, loss=1.338, acc=0.693, acc_len=1.66, thru=13.6, I=14.9, T=36.1, wait=4.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:45:06,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:45:07,247] loop.py:45 INFO TRAIN_STEP step=8844 loss=1.324803 Training: 49%|████▉ | 8844/18018 [19:14:53<22:13:55, 8.72s/step, loss=1.325, acc=0.695, acc_len=1.66, thru=12.1, I=12.3, T=22.7, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:45:07,700] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:45:15,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:45:16,150] loop.py:380 INFO TIMING step=8845: step=4.303s data=1.987s compute=4.263s [fwd=3.034s bwd=1.214s opt=0.003s] dispatch=4.593s [2026-10-19 08:45:16,150] loop.py:45 INFO TRAIN_STEP step=8845 loss=1.453534 Training: 49%|████▉ | 8845/18018 [19:15:02<22:22:03, 8.78s/step, loss=1.454, acc=0.676, acc_len=1.61, thru=13.8, I=14.9, T=29.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:45:16,502] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:45:24,108] loop.py:45 INFO TRAIN_STEP step=8846 loss=1.398828 Training: 49%|████▉ | 8846/18018 [19:15:10<21:44:13, 8.53s/step, loss=1.399, acc=0.681, acc_len=1.62, thru=14.6, I=16.3, T=28.4, wait=3.4s, pool=32, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:45:24,103] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8845: forward=3451.3ms backward=1012.2ms [2026-10-19 08:45:32,815] loop.py:45 INFO TRAIN_STEP step=8847 loss=1.350181 Training: 49%|████▉ | 8847/18018 [19:15:18<21:52:08, 8.58s/step, loss=1.350, acc=0.688, acc_len=1.64, thru=14.7, I=14.6, T=31.6, wait=4.7s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:45:39,328] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:45:40,876] loop.py:45 INFO TRAIN_STEP step=8848 loss=1.514470 Training: 49%|████▉ | 8848/18018 [19:15:26<21:27:57, 8.43s/step, loss=1.514, acc=0.663, acc_len=1.55, thru=17.9, I=17.3, T=30.0, wait=3.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:45:41,341] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:45:47,842] loop.py:45 INFO TRAIN_STEP step=8849 loss=1.352057 Training: 49%|████▉ | 8849/18018 [19:15:33<20:21:02, 7.99s/step, loss=1.352, acc=0.688, acc_len=1.64, thru=16.1, I=19.9, T=32.8, wait=3.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:45:53,930] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:45:54,806] loop.py:380 INFO TIMING step=8850: step=4.001s data=1.834s compute=3.972s [fwd=2.821s bwd=1.136s opt=0.003s] dispatch=2.952s [2026-10-19 08:45:54,807] loop.py:45 INFO TRAIN_STEP step=8850 loss=1.460321 Training: 49%|████▉ | 8850/18018 [19:15:40<19:33:42, 7.68s/step, loss=1.460, acc=0.677, acc_len=1.60, thru=19.1, I=19.2, T=32.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:45:54,946] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:45:59,712] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:46:01,357] loop.py:45 INFO TRAIN_STEP step=8851 loss=1.366248 Training: 49%|████▉ | 8851/18018 [19:15:47<18:41:44, 7.34s/step, loss=1.366, acc=0.689, acc_len=1.65, thru=17.9, I=23.1, T=33.0, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:46:01,353] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8850: forward=2689.8ms backward=1138.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:01,726] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:06,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:46:07,773] loop.py:45 INFO TRAIN_STEP step=8852 loss=1.352546 Training: 49%|████▉ | 8852/18018 [19:15:53<17:59:10, 7.06s/step, loss=1.353, acc=0.690, acc_len=1.65, thru=18.6, I=25.5, T=35.5, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:08,158] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:14,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:46:14,314] loop.py:45 INFO TRAIN_STEP step=8853 loss=1.456249 Training: 49%|████▉ | 8853/18018 [19:16:00<17:35:03, 6.91s/step, loss=1.456, acc=0.671, acc_len=1.59, thru=19.3, I=22.6, T=31.8, wait=2.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:14,761] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:20,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:46:22,159] loop.py:45 INFO TRAIN_STEP step=8854 loss=1.250028 Training: 49%|████▉ | 8854/18018 [19:16:08<18:18:01, 7.19s/step, loss=1.250, acc=0.711, acc_len=1.73, thru=16.1, I=17.9, T=33.3, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:22,301] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:28,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:46:30,462] loop.py:380 INFO TIMING step=8855: step=4.911s data=2.445s compute=4.875s [fwd=3.388s bwd=1.471s opt=0.003s] dispatch=3.380s [2026-10-19 08:46:30,462] loop.py:45 INFO TRAIN_STEP step=8855 loss=1.291954 Training: 49%|████▉ | 8855/18018 [19:16:16<19:08:52, 7.52s/step, loss=1.292, acc=0.704, acc_len=1.69, thru=13.7, I=19.5, T=26.1, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:30,936] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:38,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:46:39,718] loop.py:45 INFO TRAIN_STEP step=8856 loss=1.417954 Training: 49%|████▉ | 8856/18018 [19:16:25<20:28:06, 8.04s/step, loss=1.418, acc=0.679, acc_len=1.61, thru=14.6, I=16.0, T=26.3, wait=4.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:46:39,713] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8855: forward=3487.1ms backward=1323.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:40,002] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:45,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:46:45,915] loop.py:45 INFO TRAIN_STEP step=8857 loss=1.379096 Training: 49%|████▉ | 8857/18018 [19:16:31<19:03:24, 7.49s/step, loss=1.379, acc=0.687, acc_len=1.64, thru=18.6, I=20.1, T=33.8, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:46,013] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=32/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:52,352] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:46:54,221] loop.py:45 INFO TRAIN_STEP step=8858 loss=1.366467 Training: 49%|████▉ | 8858/18018 [19:16:40<19:40:50, 7.73s/step, loss=1.366, acc=0.690, acc_len=1.64, thru=15.3, I=18.6, T=31.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:46:54,364] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:00,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:47:01,209] loop.py:45 INFO TRAIN_STEP step=8859 loss=1.339411 Training: 49%|████▉ | 8859/18018 [19:16:47<19:06:26, 7.51s/step, loss=1.339, acc=0.693, acc_len=1.66, thru=16.8, I=20.8, T=33.9, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:01,478] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:09,340] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:47:10,105] loop.py:380 INFO TIMING step=8860: step=4.088s data=1.815s compute=4.051s [fwd=2.836s bwd=1.200s opt=0.003s] dispatch=4.799s [2026-10-19 08:47:10,105] loop.py:45 INFO TRAIN_STEP step=8860 loss=1.372140 Training: 49%|████▉ | 8860/18018 [19:16:56<20:09:48, 7.93s/step, loss=1.372, acc=0.690, acc_len=1.65, thru=15.5, I=17.7, T=31.3, wait=4.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:10,347] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:18,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:47:18,530] loop.py:45 INFO TRAIN_STEP step=8861 loss=1.396621 Training: 49%|████▉ | 8861/18018 [19:17:04<20:32:32, 8.08s/step, loss=1.397, acc=0.684, acc_len=1.64, thru=14.8, I=17.1, T=28.6, wait=3.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:47:18,525] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8860: forward=2896.3ms backward=1540.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:18,670] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=16/64 [2026-10-19 08:47:25,612] loop.py:45 INFO TRAIN_STEP step=8862 loss=1.319706 Training: 49%|████▉ | 8862/18018 [19:17:11<19:46:47, 7.78s/step, loss=1.320, acc=0.694, acc_len=1.66, thru=14.9, I=16.7, T=27.2, wait=2.4s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:35,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:47:37,837] loop.py:45 INFO TRAIN_STEP step=8863 loss=1.291107 Training: 49%|████▉ | 8863/18018 [19:17:23<23:10:23, 9.11s/step, loss=1.291, acc=0.704, acc_len=1.69, thru=14.1, I=13.5, T=28.2, wait=7.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:38,256] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:44,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:47:46,588] loop.py:45 INFO TRAIN_STEP step=8864 loss=1.481033 Training: 49%|████▉ | 8864/18018 [19:17:32<22:53:37, 9.00s/step, loss=1.481, acc=0.667, acc_len=1.58, thru=15.4, I=19.2, T=25.5, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:46,938] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 08:47:55,087] loop.py:380 INFO TIMING step=8865: step=3.793s data=1.612s compute=3.756s [fwd=2.692s bwd=1.048s opt=0.003s] dispatch=4.697s [2026-10-19 08:47:55,087] loop.py:45 INFO TRAIN_STEP step=8865 loss=1.402389 Training: 49%|████▉ | 8865/18018 [19:17:41<22:30:20, 8.85s/step, loss=1.402, acc=0.685, acc_len=1.63, thru=15.7, I=16.0, T=33.7, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:55,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:47:55,586] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:48:02,228] loop.py:45 INFO TRAIN_STEP step=8866 loss=1.452690 Training: 49%|████▉ | 8866/18018 [19:17:48<21:12:04, 8.34s/step, loss=1.453, acc=0.676, acc_len=1.60, thru=16.1, I=18.9, T=33.0, wait=3.3s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:48:02,224] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8865: forward=2720.9ms backward=1097.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:11,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:48:11,973] loop.py:45 INFO TRAIN_STEP step=8867 loss=1.249618 Training: 49%|████▉ | 8867/18018 [19:17:58<22:16:09, 8.76s/step, loss=1.250, acc=0.708, acc_len=1.72, thru=14.6, I=14.1, T=28.8, wait=5.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:12,393] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:18,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:48:19,169] loop.py:45 INFO TRAIN_STEP step=8868 loss=1.385076 Training: 49%|████▉ | 8868/18018 [19:18:05<21:04:25, 8.29s/step, loss=1.385, acc=0.688, acc_len=1.63, thru=16.9, I=20.7, T=31.5, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:19,532] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:48:27,127] loop.py:45 INFO TRAIN_STEP step=8869 loss=1.441318 Training: 49%|████▉ | 8869/18018 [19:18:13<20:49:05, 8.19s/step, loss=1.441, acc=0.678, acc_len=1.61, thru=14.9, I=17.9, T=32.9, wait=4.1s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:37,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:48:37,603] loop.py:380 INFO TIMING step=8870: step=4.796s data=2.096s compute=4.758s [fwd=3.276s bwd=1.467s opt=0.003s] dispatch=5.668s [2026-10-19 08:48:37,603] loop.py:45 INFO TRAIN_STEP step=8870 loss=1.521376 Training: 49%|████▉ | 8870/18018 [19:18:23<22:33:18, 8.88s/step, loss=1.521, acc=0.668, acc_len=1.57, thru=12.9, I=13.4, T=26.7, wait=5.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:38,095] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:44,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:48:45,457] loop.py:45 INFO TRAIN_STEP step=8871 loss=1.204732 Training: 49%|████▉ | 8871/18018 [19:18:31<21:46:24, 8.57s/step, loss=1.205, acc=0.719, acc_len=1.75, thru=15.7, I=17.2, T=25.7, wait=2.9s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:48:45,453] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8870: forward=3324.7ms backward=1604.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:45,928] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:52,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:48:53,575] loop.py:45 INFO TRAIN_STEP step=8872 loss=1.492100 Training: 49%|████▉ | 8872/18018 [19:18:39<21:25:39, 8.43s/step, loss=1.492, acc=0.668, acc_len=1.57, thru=14.9, I=17.6, T=32.6, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:48:53,847] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:00,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:49:02,057] loop.py:45 INFO TRAIN_STEP step=8873 loss=1.408594 Training: 49%|████▉ | 8873/18018 [19:18:48<21:27:40, 8.45s/step, loss=1.409, acc=0.679, acc_len=1.61, thru=13.7, I=18.2, T=32.4, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:02,440] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:07,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:49:08,999] loop.py:45 INFO TRAIN_STEP step=8874 loss=1.395199 Training: 49%|████▉ | 8874/18018 [19:18:55<20:18:40, 8.00s/step, loss=1.395, acc=0.683, acc_len=1.62, thru=17.2, I=23.5, T=32.1, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:09,310] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:49:16,728] loop.py:380 INFO TIMING step=8875: step=4.129s data=1.817s compute=4.100s [fwd=2.869s bwd=1.216s opt=0.003s] dispatch=3.591s [2026-10-19 08:49:16,728] loop.py:45 INFO TRAIN_STEP step=8875 loss=1.320883 Training: 49%|████▉ | 8875/18018 [19:19:02<20:06:18, 7.92s/step, loss=1.321, acc=0.700, acc_len=1.68, thru=14.0, I=17.5, T=31.0, wait=3.6s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:25,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:49:26,638] loop.py:45 INFO TRAIN_STEP step=8876 loss=1.416083 Training: 49%|████▉ | 8876/18018 [19:19:12<21:37:19, 8.51s/step, loss=1.416, acc=0.683, acc_len=1.62, thru=13.6, I=15.3, T=24.4, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:49:26,633] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8875: forward=3748.0ms backward=1446.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:26,921] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:49:34,138] loop.py:45 INFO TRAIN_STEP step=8877 loss=1.336669 Training: 49%|████▉ | 8877/18018 [19:19:20<20:50:57, 8.21s/step, loss=1.337, acc=0.697, acc_len=1.67, thru=15.6, I=18.6, T=32.1, wait=3.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:39,929] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:49:42,152] loop.py:45 INFO TRAIN_STEP step=8878 loss=1.392375 Training: 49%|████▉ | 8878/18018 [19:19:28<20:41:44, 8.15s/step, loss=1.392, acc=0.687, acc_len=1.63, thru=16.3, I=18.7, T=24.1, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:42,446] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 08:49:48,667] loop.py:45 INFO TRAIN_STEP step=8879 loss=1.267181 Training: 49%|████▉ | 8879/18018 [19:19:34<19:26:47, 7.66s/step, loss=1.267, acc=0.709, acc_len=1.72, thru=14.2, I=17.5, T=29.5, wait=2.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:55,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:49:56,485] loop.py:380 INFO TIMING step=8880: step=4.302s data=1.944s compute=4.275s [fwd=3.079s bwd=1.181s opt=0.003s] dispatch=3.506s [2026-10-19 08:49:56,485] loop.py:45 INFO TRAIN_STEP step=8880 loss=1.484668 Training: 49%|████▉ | 8880/18018 [19:19:42<19:33:55, 7.71s/step, loss=1.485, acc=0.669, acc_len=1.58, thru=17.6, I=17.9, T=29.8, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:49:56,566] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=40/64 [2026-10-19 08:50:04,504] loop.py:45 INFO TRAIN_STEP step=8881 loss=1.346118 Training: 49%|████▉ | 8881/18018 [19:19:50<19:47:53, 7.80s/step, loss=1.346, acc=0.695, acc_len=1.66, thru=13.2, I=16.7, T=26.5, wait=3.2s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:50:04,499] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8880: forward=3351.8ms backward=1437.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:14,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:50:14,249] loop.py:45 INFO TRAIN_STEP step=8882 loss=1.465603 Training: 49%|████▉ | 8882/18018 [19:20:00<21:16:39, 8.38s/step, loss=1.466, acc=0.676, acc_len=1.60, thru=14.6, I=14.1, T=32.2, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:14,515] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:20,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:50:22,122] loop.py:45 INFO TRAIN_STEP step=8883 loss=1.428121 Training: 49%|████▉ | 8883/18018 [19:20:08<20:53:12, 8.23s/step, loss=1.428, acc=0.679, acc_len=1.61, thru=16.4, I=20.5, T=26.0, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:22,155] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:28,188] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:50:29,446] loop.py:45 INFO TRAIN_STEP step=8884 loss=1.425073 Training: 49%|████▉ | 8884/18018 [19:20:15<20:11:36, 7.96s/step, loss=1.425, acc=0.678, acc_len=1.61, thru=14.4, I=18.0, T=30.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:29,697] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:35,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:50:37,168] loop.py:380 INFO TIMING step=8885: step=4.569s data=2.115s compute=4.538s [fwd=3.251s bwd=1.271s opt=0.003s] dispatch=3.144s [2026-10-19 08:50:37,168] loop.py:45 INFO TRAIN_STEP step=8885 loss=1.474423 Training: 49%|████▉ | 8885/18018 [19:20:23<20:00:39, 7.89s/step, loss=1.474, acc=0.673, acc_len=1.59, thru=15.8, I=23.1, T=28.0, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:37,487] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:43,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:50:43,792] loop.py:45 INFO TRAIN_STEP step=8886 loss=1.414620 Training: 49%|████▉ | 8886/18018 [19:20:29<19:02:47, 7.51s/step, loss=1.415, acc=0.687, acc_len=1.64, thru=16.7, I=21.3, T=32.7, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:50:43,788] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8885: forward=2762.2ms backward=1103.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:50:44,095] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:50:50,177] loop.py:45 INFO TRAIN_STEP step=8887 loss=1.394886 Training: 49%|████▉ | 8887/18018 [19:20:36<18:11:28, 7.17s/step, loss=1.395, acc=0.684, acc_len=1.63, thru=16.5, I=18.1, T=37.2, wait=2.9s, pool=36, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:51:00,402] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:51:00,879] loop.py:45 INFO TRAIN_STEP step=8888 loss=1.423718 Training: 49%|████▉ | 8888/18018 [19:20:46<20:52:25, 8.23s/step, loss=1.424, acc=0.681, acc_len=1.62, thru=14.6, I=15.2, T=28.1, wait=6.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:51:00,916] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:51:07,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:51:08,351] loop.py:45 INFO TRAIN_STEP step=8889 loss=1.349359 Training: 49%|████▉ | 8889/18018 [19:20:54<20:17:41, 8.00s/step, loss=1.349, acc=0.693, acc_len=1.66, thru=16.2, I=17.1, T=33.4, wait=3.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:51:08,563] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 08:51:17,456] loop.py:380 INFO TIMING step=8890: step=3.974s data=1.757s compute=3.946s [fwd=2.792s bwd=1.139s opt=0.003s] dispatch=5.121s [2026-10-19 08:51:17,457] loop.py:45 INFO TRAIN_STEP step=8890 loss=1.425055 Training: 49%|████▉ | 8890/18018 [19:21:03<21:07:42, 8.33s/step, loss=1.425, acc=0.681, acc_len=1.62, thru=11.7, I=12.9, T=32.2, wait=5.1s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:51:25,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:51:26,344] loop.py:45 INFO TRAIN_STEP step=8891 loss=1.338520 Training: 49%|████▉ | 8891/18018 [19:21:12<21:32:59, 8.50s/step, loss=1.339, acc=0.696, acc_len=1.67, thru=16.8, I=16.0, T=30.8, wait=4.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:51:26,339] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8890: forward=3003.5ms backward=1108.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:51:26,758] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:51:34,827] loop.py:45 INFO TRAIN_STEP step=8892 loss=1.299053 Training: 49%|████▉ | 8892/18018 [19:21:20<21:32:00, 8.49s/step, loss=1.299, acc=0.704, acc_len=1.69, thru=12.3, I=13.2, T=26.3, wait=3.6s, pool=48, epoch=1/2][2026-10-19 08:51:44,261] loop.py:45 INFO TRAIN_STEP step=8893 loss=1.305154 Training: 49%|████▉ | 8893/18018 [19:21:30<22:14:48, 8.78s/step, loss=1.305, acc=0.699, acc_len=1.68, thru=13.4, I=14.0, T=30.5, wait=5.2s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:51:51,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:51:53,459] loop.py:45 INFO TRAIN_STEP step=8894 loss=1.389849 Training: 49%|████▉ | 8894/18018 [19:21:39<22:33:52, 8.90s/step, loss=1.390, acc=0.689, acc_len=1.64, thru=16.2, I=16.6, T=27.7, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:51:53,702] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-19 08:52:00,161] loop.py:380 INFO TIMING step=8895: step=3.448s data=1.334s compute=3.420s [fwd=2.442s bwd=0.963s opt=0.003s] dispatch=3.244s [2026-10-19 08:52:00,162] loop.py:45 INFO TRAIN_STEP step=8895 loss=1.292663 Training: 49%|████▉ | 8895/18018 [19:21:46<20:53:13, 8.24s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=15.8, I=20.3, T=37.1, wait=3.2s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:52:06,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:52:09,131] loop.py:45 INFO TRAIN_STEP step=8896 loss=1.510785 Training: 49%|████▉ | 8896/18018 [19:21:55<21:26:23, 8.46s/step, loss=1.511, acc=0.669, acc_len=1.58, thru=15.0, I=19.4, T=26.9, wait=4.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:52:09,126] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8895: forward=3300.8ms backward=1398.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:52:09,611] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 08:52:16,064] loop.py:45 INFO TRAIN_STEP step=8897 loss=1.361676 Training: 49%|████▉ | 8897/18018 [19:22:02<20:16:39, 8.00s/step, loss=1.362, acc=0.694, acc_len=1.66, thru=14.8, I=23.0, T=31.6, wait=2.9s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:52:22,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:52:24,888] loop.py:45 INFO TRAIN_STEP step=8898 loss=1.454581 Training: 49%|████▉ | 8898/18018 [19:22:10<20:53:42, 8.25s/step, loss=1.455, acc=0.675, acc_len=1.60, thru=14.7, I=18.3, T=31.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:52:25,123] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 08:52:31,337] loop.py:45 INFO TRAIN_STEP step=8899 loss=1.369299 Training: 49%|████▉ | 8899/18018 [19:22:17<19:31:46, 7.71s/step, loss=1.369, acc=0.690, acc_len=1.65, thru=16.8, I=21.8, T=33.6, wait=2.6s, pool=52, epoch=1/2][2026-10-19 08:52:42,177] loop.py:380 INFO TIMING step=8900: step=3.951s data=1.670s compute=3.922s [fwd=2.825s bwd=1.082s opt=0.003s] dispatch=6.876s [2026-10-19 08:52:42,177] loop.py:45 INFO TRAIN_STEP step=8900 loss=1.390675 Training: 49%|████▉ | 8900/18018 [19:22:28<21:54:10, 8.65s/step, loss=1.391, acc=0.691, acc_len=1.65, thru=13.2, I=12.9, T=32.4, wait=6.9s, pool=56, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 08:52:42,256] eagle_store.py:517 WARNING batch_get_buffer missing keys (k284807-bef3aab0_hs, k284807-bef3aab0_ids, k284807-bef3aab0_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:52:52,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 08:52:42,249] eagle_store.py:517 WARNING batch_get_buffer missing keys (k284806-a3987c03_hs, k284806-a3987c03_ids, k284806-a3987c03_lhs); sleeping 0.50s. [repeated 3x across cluster] (TrainerActor pid=9310) [2026-10-19 08:52:53,915] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8900: forward=2922.4ms backward=1291.5ms [2026-10-19 08:52:53,920] loop.py:45 INFO TRAIN_STEP step=8901 loss=1.329118 Training: 49%|████▉ | 8901/18018 [19:22:40<24:15:06, 9.58s/step, loss=1.329, acc=0.695, acc_len=1.67, thru=11.8, I=17.1, T=29.8, wait=7.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:52:54,308] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:53:00,654] loop.py:45 INFO TRAIN_STEP step=8902 loss=1.334584 Training: 49%|████▉ | 8902/18018 [19:22:46<22:05:27, 8.72s/step, loss=1.335, acc=0.693, acc_len=1.66, thru=17.6, I=20.5, T=34.4, wait=3.0s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:08,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:53:08,655] loop.py:45 INFO TRAIN_STEP step=8903 loss=1.296212 Training: 49%|████▉ | 8903/18018 [19:22:54<21:32:18, 8.51s/step, loss=1.296, acc=0.701, acc_len=1.69, thru=16.3, I=16.7, T=29.6, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:08,809] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=8/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:15,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:53:15,731] loop.py:45 INFO TRAIN_STEP step=8904 loss=1.371490 Training: 49%|████▉ | 8904/18018 [19:23:01<20:26:57, 8.08s/step, loss=1.371, acc=0.689, acc_len=1.65, thru=17.3, I=19.5, T=31.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:16,123] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:21,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:53:23,741] loop.py:380 INFO TIMING step=8905: step=4.949s data=2.453s compute=4.903s [fwd=3.368s bwd=1.521s opt=0.003s] dispatch=3.052s [2026-10-19 08:53:23,741] loop.py:45 INFO TRAIN_STEP step=8905 loss=1.545223 Training: 49%|████▉ | 8905/18018 [19:23:09<20:23:50, 8.06s/step, loss=1.545, acc=0.657, acc_len=1.54, thru=15.4, I=20.9, T=25.9, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:24,022] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 [2026-10-19 08:53:32,208] loop.py:45 INFO TRAIN_STEP step=8906 loss=1.383543 Training: 49%|████▉ | 8906/18018 [19:23:18<20:42:21, 8.18s/step, loss=1.384, acc=0.690, acc_len=1.65, thru=14.2, I=14.8, T=33.0, wait=4.6s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:53:32,204] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8905: forward=2758.7ms backward=1082.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:38,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:53:40,575] loop.py:45 INFO TRAIN_STEP step=8907 loss=1.275118 Training: 49%|████▉ | 8907/18018 [19:23:26<20:50:42, 8.24s/step, loss=1.275, acc=0.705, acc_len=1.69, thru=15.6, I=18.1, T=30.0, wait=4.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:40,828] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:53:47,565] loop.py:45 INFO TRAIN_STEP step=8908 loss=1.305236 Training: 49%|████▉ | 8908/18018 [19:23:33<19:53:39, 7.86s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=15.8, I=19.3, T=30.1, wait=2.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:58,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:53:58,637] loop.py:45 INFO TRAIN_STEP step=8909 loss=1.434991 Training: 49%|████▉ | 8909/18018 [19:23:44<22:19:49, 8.83s/step, loss=1.435, acc=0.676, acc_len=1.60, thru=12.5, I=12.2, T=31.2, wait=7.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:53:58,883] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:54:04,503] loop.py:380 INFO TIMING step=8910: step=3.833s data=1.706s compute=3.804s [fwd=2.731s bwd=1.059s opt=0.003s] dispatch=2.024s [2026-10-19 08:54:04,504] loop.py:45 INFO TRAIN_STEP step=8910 loss=1.351362 Training: 49%|████▉ | 8910/18018 [19:23:50<20:04:56, 7.94s/step, loss=1.351, acc=0.695, acc_len=1.67, thru=16.0, I=21.5, T=33.4, wait=2.0s, pool=32, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:14,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 08:54:15,336] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8910: forward=3015.0ms backward=1419.2ms [2026-10-19 08:54:15,342] loop.py:45 INFO TRAIN_STEP step=8911 loss=1.393389 Training: 49%|████▉ | 8911/18018 [19:24:01<22:16:48, 8.81s/step, loss=1.393, acc=0.683, acc_len=1.62, thru=16.1, I=13.5, T=28.5, wait=6.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:15,647] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:20,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:54:21,718] loop.py:45 INFO TRAIN_STEP step=8912 loss=1.347321 Training: 49%|████▉ | 8912/18018 [19:24:07<20:25:59, 8.08s/step, loss=1.347, acc=0.690, acc_len=1.65, thru=18.0, I=23.7, T=35.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:21,904] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:30,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:54:32,119] loop.py:45 INFO TRAIN_STEP step=8913 loss=1.489234 Training: 49%|████▉ | 8913/18018 [19:24:18<22:11:44, 8.78s/step, loss=1.489, acc=0.674, acc_len=1.59, thru=12.2, I=13.8, T=27.9, wait=5.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:32,364] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:38,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:54:39,029] loop.py:45 INFO TRAIN_STEP step=8914 loss=1.419743 Training: 49%|████▉ | 8914/18018 [19:24:25<20:46:39, 8.22s/step, loss=1.420, acc=0.679, acc_len=1.61, thru=15.9, I=19.3, T=33.8, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:39,235] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:45,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:54:48,069] loop.py:380 INFO TIMING step=8915: step=5.018s data=2.306s compute=4.992s [fwd=3.329s bwd=1.648s opt=0.003s] dispatch=4.012s [2026-10-19 08:54:48,069] loop.py:45 INFO TRAIN_STEP step=8915 loss=1.393001 Training: 49%|████▉ | 8915/18018 [19:24:34<21:23:55, 8.46s/step, loss=1.393, acc=0.686, acc_len=1.63, thru=14.1, I=19.2, T=25.5, wait=4.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:48,493] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:55,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:54:56,136] loop.py:45 INFO TRAIN_STEP step=8916 loss=1.303936 Training: 49%|████▉ | 8916/18018 [19:24:42<21:05:45, 8.34s/step, loss=1.304, acc=0.696, acc_len=1.67, thru=16.8, I=19.4, T=31.9, wait=4.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:54:56,131] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8915: forward=2852.5ms backward=1104.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:54:56,538] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:02,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:55:03,566] loop.py:45 INFO TRAIN_STEP step=8917 loss=1.228178 Training: 49%|████▉ | 8917/18018 [19:24:49<20:24:02, 8.07s/step, loss=1.228, acc=0.714, acc_len=1.73, thru=16.7, I=20.3, T=29.0, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:03,918] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:11,165] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:55:11,375] loop.py:45 INFO TRAIN_STEP step=8918 loss=1.426917 Training: 49%|████▉ | 8918/18018 [19:24:57<20:12:02, 7.99s/step, loss=1.427, acc=0.680, acc_len=1.62, thru=16.2, I=16.4, T=29.7, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:11,667] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:18,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:55:18,842] loop.py:45 INFO TRAIN_STEP step=8919 loss=1.455993 Training: 50%|████▉ | 8919/18018 [19:25:04<19:48:07, 7.83s/step, loss=1.456, acc=0.677, acc_len=1.61, thru=17.1, I=16.9, T=29.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:18,922] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:24,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:55:25,422] loop.py:380 INFO TIMING step=8920: step=3.848s data=1.637s compute=3.796s [fwd=2.642s bwd=1.137s opt=0.003s] dispatch=2.721s [2026-10-19 08:55:25,422] loop.py:45 INFO TRAIN_STEP step=8920 loss=1.394570 Training: 50%|████▉ | 8920/18018 [19:25:11<18:50:50, 7.46s/step, loss=1.395, acc=0.686, acc_len=1.63, thru=19.0, I=19.1, T=33.3, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:25,568] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:29,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:55:31,530] loop.py:45 INFO TRAIN_STEP step=8921 loss=1.425520 Training: 50%|████▉ | 8921/18018 [19:25:17<17:49:23, 7.05s/step, loss=1.426, acc=0.679, acc_len=1.61, thru=19.6, I=25.9, T=31.6, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:55:31,525] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8920: forward=2783.7ms backward=1226.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:31,833] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:37,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:55:39,244] loop.py:45 INFO TRAIN_STEP step=8922 loss=1.395952 Training: 50%|████▉ | 8922/18018 [19:25:25<18:19:14, 7.25s/step, loss=1.396, acc=0.689, acc_len=1.65, thru=14.9, I=21.8, T=28.9, wait=3.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:39,720] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:55:47,814] loop.py:45 INFO TRAIN_STEP step=8923 loss=1.330984 Training: 50%|████▉ | 8923/18018 [19:25:33<19:19:10, 7.65s/step, loss=1.331, acc=0.697, acc_len=1.67, thru=14.8, I=15.5, T=33.4, wait=4.7s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:53,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:55:55,496] loop.py:45 INFO TRAIN_STEP step=8924 loss=1.325650 Training: 50%|████▉ | 8924/18018 [19:25:41<19:20:36, 7.66s/step, loss=1.326, acc=0.695, acc_len=1.67, thru=16.4, I=20.4, T=30.0, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:55:55,599] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=24/64 [2026-10-19 08:56:02,848] loop.py:380 INFO TIMING step=8925: step=3.638s data=1.478s compute=3.607s [fwd=2.598s bwd=0.994s opt=0.003s] dispatch=3.705s [2026-10-19 08:56:02,848] loop.py:45 INFO TRAIN_STEP step=8925 loss=1.377003 Training: 50%|████▉ | 8925/18018 [19:25:48<19:06:31, 7.57s/step, loss=1.377, acc=0.691, acc_len=1.65, thru=13.7, I=17.3, T=35.2, wait=3.7s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:56:11,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:56:12,959] loop.py:45 INFO TRAIN_STEP step=8926 loss=1.410385 Training: 50%|████▉ | 8926/18018 [19:25:59<21:02:14, 8.33s/step, loss=1.410, acc=0.682, acc_len=1.62, thru=14.8, I=16.5, T=27.1, wait=5.4s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:56:12,954] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8925: forward=3156.3ms backward=1493.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:56:13,176] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:56:22,434] loop.py:45 INFO TRAIN_STEP step=8927 loss=1.346520 Training: 50%|████▉ | 8927/18018 [19:26:08<21:54:09, 8.67s/step, loss=1.347, acc=0.691, acc_len=1.66, thru=11.2, I=12.1, T=28.8, wait=5.0s, pool=40, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:56:32,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:56:32,781] loop.py:45 INFO TRAIN_STEP step=8928 loss=1.435972 Training: 50%|████▉ | 8928/18018 [19:26:18<23:09:57, 9.17s/step, loss=1.436, acc=0.680, acc_len=1.62, thru=14.7, I=15.3, T=28.9, wait=5.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:56:33,007] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:56:38,071] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:56:39,603] loop.py:45 INFO TRAIN_STEP step=8929 loss=1.359270 Training: 50%|████▉ | 8929/18018 [19:26:25<21:22:54, 8.47s/step, loss=1.359, acc=0.692, acc_len=1.66, thru=18.1, I=20.4, T=32.1, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:56:40,081] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 08:56:45,760] loop.py:380 INFO TIMING step=8930: step=3.402s data=1.377s compute=3.370s [fwd=2.389s bwd=0.966s opt=0.003s] dispatch=2.746s [2026-10-19 08:56:45,760] loop.py:45 INFO TRAIN_STEP step=8930 loss=1.343638 Training: 50%|████▉ | 8930/18018 [19:26:31<19:37:55, 7.78s/step, loss=1.344, acc=0.693, acc_len=1.66, thru=18.1, I=21.4, T=37.6, wait=2.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:56:52,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:56:53,466] loop.py:45 INFO TRAIN_STEP step=8931 loss=1.286451 Training: 50%|████▉ | 8931/18018 [19:26:39<19:34:28, 7.75s/step, loss=1.286, acc=0.707, acc_len=1.71, thru=17.9, I=19.3, T=32.2, wait=3.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:56:53,461] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8930: forward=2760.8ms backward=1167.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:56:53,651] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:56:58,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:57:00,484] loop.py:45 INFO TRAIN_STEP step=8932 loss=1.349317 Training: 50%|████▉ | 8932/18018 [19:26:46<19:00:46, 7.53s/step, loss=1.349, acc=0.691, acc_len=1.65, thru=17.1, I=23.7, T=33.3, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:57:00,897] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:57:08,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:57:08,641] loop.py:45 INFO TRAIN_STEP step=8933 loss=1.414419 Training: 50%|████▉ | 8933/18018 [19:26:54<19:28:59, 7.72s/step, loss=1.414, acc=0.681, acc_len=1.62, thru=14.1, I=18.5, T=33.2, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:57:08,856] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:57:15,951] loop.py:45 INFO TRAIN_STEP step=8934 loss=1.441384 Training: 50%|████▉ | 8934/18018 [19:27:02<19:10:13, 7.60s/step, loss=1.441, acc=0.673, acc_len=1.59, thru=12.8, I=15.0, T=32.3, wait=3.3s, pool=24, epoch=1/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 08:57:16,034] eagle_store.py:517 WARNING batch_get_buffer missing keys (k285887-98a389ed_hs, k285887-98a389ed_ids, k285887-98a389ed_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:57:29,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 08:57:16,087] eagle_store.py:517 WARNING batch_get_buffer missing keys (k285885-9644f285_hs, k285885-9644f285_ids, k285885-9644f285_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 08:57:30,574] loop.py:380 INFO TIMING step=8935: step=5.043s data=2.487s compute=4.969s [fwd=3.417s bwd=1.537s opt=0.003s] dispatch=9.572s [2026-10-19 08:57:30,574] loop.py:45 INFO TRAIN_STEP step=8935 loss=1.517333 Training: 50%|████▉ | 8935/18018 [19:27:16<24:29:14, 9.71s/step, loss=1.517, acc=0.668, acc_len=1.56, thru=11.5, I=12.0, T=25.4, wait=9.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:57:30,684] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=24/64 [2026-10-19 08:57:37,315] loop.py:45 INFO TRAIN_STEP step=8936 loss=1.352710 Training: 50%|████▉ | 8936/18018 [19:27:23<22:14:24, 8.82s/step, loss=1.353, acc=0.695, acc_len=1.67, thru=16.0, I=20.7, T=32.1, wait=2.7s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:57:37,310] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8935: forward=2825.1ms backward=1105.0ms [2026-10-19 08:57:45,165] loop.py:45 INFO TRAIN_STEP step=8937 loss=1.350259 Training: 50%|████▉ | 8937/18018 [19:27:31<21:30:34, 8.53s/step, loss=1.350, acc=0.695, acc_len=1.66, thru=14.2, I=14.1, T=33.9, wait=4.1s, pool=52, epoch=1/2][2026-10-19 08:57:53,641] loop.py:45 INFO TRAIN_STEP step=8938 loss=1.401658 Training: 50%|████▉ | 8938/18018 [19:27:39<21:28:01, 8.51s/step, loss=1.402, acc=0.684, acc_len=1.63, thru=15.8, I=15.3, T=31.3, wait=4.4s, pool=56, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:00,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:58:01,786] loop.py:45 INFO TRAIN_STEP step=8939 loss=1.406474 Training: 50%|████▉ | 8939/18018 [19:27:47<21:11:11, 8.40s/step, loss=1.406, acc=0.685, acc_len=1.64, thru=16.6, I=18.0, T=32.2, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:02,253] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 08:58:11,166] loop.py:380 INFO TIMING step=8940: step=4.208s data=1.910s compute=4.161s [fwd=2.898s bwd=1.248s opt=0.003s] dispatch=5.163s [2026-10-19 08:58:11,166] loop.py:45 INFO TRAIN_STEP step=8940 loss=1.372159 Training: 50%|████▉ | 8940/18018 [19:27:57<21:55:28, 8.69s/step, loss=1.372, acc=0.690, acc_len=1.65, thru=12.1, I=12.1, T=30.4, wait=5.2s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:21,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:58:23,752] loop.py:45 INFO TRAIN_STEP step=8941 loss=1.307687 Training: 50%|████▉ | 8941/18018 [19:28:09<24:52:00, 9.86s/step, loss=1.308, acc=0.703, acc_len=1.69, thru=11.2, I=13.1, T=24.1, wait=7.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:58:23,747] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8940: forward=3639.4ms backward=1590.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:23,883] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=16/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:30,856] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:58:32,571] loop.py:45 INFO TRAIN_STEP step=8942 loss=1.290375 Training: 50%|████▉ | 8942/18018 [19:28:18<24:04:29, 9.55s/step, loss=1.290, acc=0.705, acc_len=1.70, thru=14.8, I=18.3, T=32.6, wait=4.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:32,871] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:37,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:58:39,607] loop.py:45 INFO TRAIN_STEP step=8943 loss=1.394558 Training: 50%|████▉ | 8943/18018 [19:28:25<22:10:19, 8.80s/step, loss=1.395, acc=0.683, acc_len=1.63, thru=16.4, I=23.9, T=26.8, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:39,866] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:45,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:58:46,546] loop.py:45 INFO TRAIN_STEP step=8944 loss=1.250716 Training: 50%|████▉ | 8944/18018 [19:28:32<20:45:52, 8.24s/step, loss=1.251, acc=0.709, acc_len=1.71, thru=15.8, I=21.9, T=27.5, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:46,873] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:51,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:58:53,867] loop.py:380 INFO TIMING step=8945: step=4.480s data=2.125s compute=4.449s [fwd=3.088s bwd=1.345s opt=0.003s] dispatch=2.831s [2026-10-19 08:58:53,867] loop.py:45 INFO TRAIN_STEP step=8945 loss=1.355206 Training: 50%|████▉ | 8945/18018 [19:28:39<20:04:11, 7.96s/step, loss=1.355, acc=0.692, acc_len=1.65, thru=16.4, I=21.1, T=28.6, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:58:54,011] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-19 08:59:01,005] loop.py:45 INFO TRAIN_STEP step=8946 loss=1.262656 Training: 50%|████▉ | 8946/18018 [19:28:47<19:26:32, 7.72s/step, loss=1.263, acc=0.706, acc_len=1.71, thru=11.9, I=18.7, T=37.2, wait=3.7s, pool=24, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:59:01,001] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8945: forward=2446.4ms backward=947.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:10,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:59:12,460] loop.py:45 INFO TRAIN_STEP step=8947 loss=1.445879 Training: 50%|████▉ | 8947/18018 [19:28:58<22:16:03, 8.84s/step, loss=1.446, acc=0.677, acc_len=1.60, thru=16.7, I=17.8, T=28.2, wait=6.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:12,761] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:18,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:59:20,279] loop.py:45 INFO TRAIN_STEP step=8948 loss=1.310713 Training: 50%|████▉ | 8948/18018 [19:29:06<21:29:46, 8.53s/step, loss=1.311, acc=0.701, acc_len=1.68, thru=17.0, I=23.9, T=25.2, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:20,628] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 08:59:27,217] loop.py:45 INFO TRAIN_STEP step=8949 loss=1.493464 Training: 50%|████▉ | 8949/18018 [19:29:13<20:17:21, 8.05s/step, loss=1.493, acc=0.667, acc_len=1.57, thru=13.6, I=18.5, T=34.6, wait=3.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:34,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:59:36,378] loop.py:380 INFO TIMING step=8950: step=4.157s data=1.844s compute=4.115s [fwd=2.951s bwd=1.149s opt=0.003s] dispatch=4.993s [2026-10-19 08:59:36,378] loop.py:45 INFO TRAIN_STEP step=8950 loss=1.486876 Training: 50%|████▉ | 8950/18018 [19:29:22<21:07:26, 8.39s/step, loss=1.487, acc=0.672, acc_len=1.58, thru=15.3, I=16.1, T=30.8, wait=5.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:36,419] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:41,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:59:42,594] loop.py:45 INFO TRAIN_STEP step=8951 loss=1.322442 Training: 50%|████▉ | 8951/18018 [19:29:28<19:28:52, 7.73s/step, loss=1.322, acc=0.696, acc_len=1.67, thru=18.6, I=25.2, T=30.5, wait=2.0s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 08:59:42,589] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8950: forward=2969.2ms backward=1188.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:42,628] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:49,555] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 08:59:49,744] loop.py:45 INFO TRAIN_STEP step=8952 loss=1.463549 Training: 50%|████▉ | 8952/18018 [19:29:35<19:02:10, 7.56s/step, loss=1.464, acc=0.674, acc_len=1.59, thru=16.8, I=19.7, T=29.7, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 08:59:50,057] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 [2026-10-19 08:59:57,910] loop.py:45 INFO TRAIN_STEP step=8953 loss=1.414969 Training: 50%|████▉ | 8953/18018 [19:29:43<19:29:37, 7.74s/step, loss=1.415, acc=0.683, acc_len=1.62, thru=16.1, I=16.2, T=33.8, wait=4.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:05,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:00:07,251] loop.py:45 INFO TRAIN_STEP step=8954 loss=1.455289 Training: 50%|████▉ | 8954/18018 [19:29:53<20:42:00, 8.22s/step, loss=1.455, acc=0.672, acc_len=1.59, thru=14.3, I=16.7, T=27.4, wait=4.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:07,310] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:12,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:00:14,286] loop.py:380 INFO TIMING step=8955: step=4.461s data=1.982s compute=4.429s [fwd=3.063s bwd=1.350s opt=0.003s] dispatch=2.564s [2026-10-19 09:00:14,286] loop.py:45 INFO TRAIN_STEP step=8955 loss=1.244482 Training: 50%|████▉ | 8955/18018 [19:30:00<19:48:05, 7.87s/step, loss=1.244, acc=0.709, acc_len=1.72, thru=14.9, I=21.0, T=28.7, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:14,506] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:19,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:00:21,847] loop.py:45 INFO TRAIN_STEP step=8956 loss=1.183622 Training: 50%|████▉ | 8956/18018 [19:30:07<19:34:08, 7.77s/step, loss=1.184, acc=0.723, acc_len=1.76, thru=14.8, I=24.5, T=29.4, wait=3.2s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:00:21,842] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8955: forward=3039.0ms backward=1260.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:21,987] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=8/64 [2026-10-19 09:00:28,350] loop.py:45 INFO TRAIN_STEP step=8957 loss=1.428649 Training: 50%|████▉ | 8957/18018 [19:30:14<18:36:30, 7.39s/step, loss=1.429, acc=0.678, acc_len=1.61, thru=17.1, I=21.8, T=32.2, wait=2.5s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:35,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:00:36,868] loop.py:45 INFO TRAIN_STEP step=8958 loss=1.413985 Training: 50%|████▉ | 8958/18018 [19:30:22<19:27:14, 7.73s/step, loss=1.414, acc=0.679, acc_len=1.61, thru=15.6, I=19.5, T=27.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:37,118] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 09:00:45,685] loop.py:45 INFO TRAIN_STEP step=8959 loss=1.333069 Training: 50%|████▉ | 8959/18018 [19:30:31<20:16:15, 8.06s/step, loss=1.333, acc=0.699, acc_len=1.68, thru=14.5, I=13.3, T=23.7, wait=3.4s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:54,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:00:56,676] loop.py:380 INFO TIMING step=8960: step=4.339s data=1.942s compute=4.309s [fwd=3.071s bwd=1.223s opt=0.003s] dispatch=6.643s [2026-10-19 09:00:56,676] loop.py:45 INFO TRAIN_STEP step=8960 loss=1.453134 Training: 50%|████▉ | 8960/18018 [19:30:42<22:29:10, 8.94s/step, loss=1.453, acc=0.679, acc_len=1.61, thru=12.8, I=14.3, T=29.5, wait=6.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:00:57,005] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:01,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:01:03,920] loop.py:45 INFO TRAIN_STEP step=8961 loss=1.490516 Training: 50%|████▉ | 8961/18018 [19:30:50<21:12:21, 8.43s/step, loss=1.491, acc=0.670, acc_len=1.59, thru=16.8, I=29.8, T=28.1, wait=2.7s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:01:03,915] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8960: forward=3133.0ms backward=1376.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:03,986] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=48/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:10,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:01:11,356] loop.py:45 INFO TRAIN_STEP step=8962 loss=1.304668 Training: 50%|████▉ | 8962/18018 [19:30:57<20:27:14, 8.13s/step, loss=1.305, acc=0.704, acc_len=1.70, thru=13.7, I=19.8, T=27.4, wait=2.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:11,795] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:18,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:01:18,280] loop.py:45 INFO TRAIN_STEP step=8963 loss=1.456080 Training: 50%|████▉ | 8963/18018 [19:31:04<19:32:28, 7.77s/step, loss=1.456, acc=0.673, acc_len=1.59, thru=18.2, I=18.4, T=31.0, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:18,586] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:24,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:01:26,370] loop.py:45 INFO TRAIN_STEP step=8964 loss=1.239780 Training: 50%|████▉ | 8964/18018 [19:31:12<19:46:47, 7.86s/step, loss=1.240, acc=0.711, acc_len=1.72, thru=16.9, I=19.2, T=29.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:26,796] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:32,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:01:33,564] loop.py:380 INFO TIMING step=8965: step=4.327s data=2.109s compute=4.283s [fwd=2.955s bwd=1.312s opt=0.003s] dispatch=2.859s [2026-10-19 09:01:33,564] loop.py:45 INFO TRAIN_STEP step=8965 loss=1.476499 Training: 50%|████▉ | 8965/18018 [19:31:19<19:16:17, 7.66s/step, loss=1.476, acc=0.675, acc_len=1.60, thru=16.7, I=21.2, T=29.6, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:33,950] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 09:01:40,883] loop.py:45 INFO TRAIN_STEP step=8966 loss=1.336883 Training: 50%|████▉ | 8966/18018 [19:31:26<19:00:38, 7.56s/step, loss=1.337, acc=0.697, acc_len=1.67, thru=15.8, I=21.0, T=29.5, wait=3.0s, pool=56, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:01:40,879] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8965: forward=2949.3ms backward=1348.6ms [2026-10-19 09:01:48,178] loop.py:45 INFO TRAIN_STEP step=8967 loss=1.510093 Training: 50%|████▉ | 8967/18018 [19:31:34<18:48:26, 7.48s/step, loss=1.510, acc=0.671, acc_len=1.58, thru=15.7, I=17.0, T=34.5, wait=3.6s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:56,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:01:57,775] loop.py:45 INFO TRAIN_STEP step=8968 loss=1.392193 Training: 50%|████▉ | 8968/18018 [19:31:43<20:24:03, 8.12s/step, loss=1.392, acc=0.687, acc_len=1.64, thru=14.9, I=15.1, T=30.7, wait=5.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:01:58,180] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 09:02:06,043] loop.py:45 INFO TRAIN_STEP step=8969 loss=1.373453 Training: 50%|████▉ | 8969/18018 [19:31:52<20:31:05, 8.16s/step, loss=1.373, acc=0.683, acc_len=1.63, thru=12.2, I=18.3, T=30.6, wait=4.1s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:02:12,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:02:13,851] loop.py:380 INFO TIMING step=8970: step=4.319s data=1.956s compute=4.273s [fwd=3.051s bwd=1.206s opt=0.003s] dispatch=3.475s [2026-10-19 09:02:13,852] loop.py:45 INFO TRAIN_STEP step=8970 loss=1.502755 Training: 50%|████▉ | 8970/18018 [19:31:59<20:14:44, 8.06s/step, loss=1.503, acc=0.665, acc_len=1.56, thru=16.5, I=16.9, T=29.6, wait=3.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:02:14,170] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:02:18,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:02:20,375] loop.py:45 INFO TRAIN_STEP step=8971 loss=1.275648 Training: 50%|████▉ | 8971/18018 [19:32:06<19:05:20, 7.60s/step, loss=1.276, acc=0.708, acc_len=1.71, thru=18.5, I=22.8, T=32.2, wait=2.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:02:20,371] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8970: forward=2836.8ms backward=1094.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:02:20,817] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 09:02:27,854] loop.py:45 INFO TRAIN_STEP step=8972 loss=1.419124 Training: 50%|████▉ | 8972/18018 [19:32:13<18:59:48, 7.56s/step, loss=1.419, acc=0.680, acc_len=1.61, thru=14.2, I=18.2, T=31.9, wait=3.5s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:02:34,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:02:36,513] loop.py:45 INFO TRAIN_STEP step=8973 loss=1.371077 Training: 50%|████▉ | 8973/18018 [19:32:22<19:49:39, 7.89s/step, loss=1.371, acc=0.688, acc_len=1.65, thru=16.9, I=19.6, T=30.9, wait=4.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:02:36,842] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 [2026-10-19 09:02:46,151] loop.py:45 INFO TRAIN_STEP step=8974 loss=1.372547 Training: 50%|████▉ | 8974/18018 [19:32:32<21:08:30, 8.42s/step, loss=1.373, acc=0.691, acc_len=1.65, thru=12.5, I=12.3, T=29.2, wait=5.2s, pool=52, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:02:52,373] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:02:53,171] loop.py:380 INFO TIMING step=8975: step=3.787s data=1.628s compute=3.757s [fwd=2.716s bwd=1.026s opt=0.003s] dispatch=3.219s [2026-10-19 09:02:53,171] loop.py:45 INFO TRAIN_STEP step=8975 loss=1.285127 Training: 50%|████▉ | 8975/18018 [19:32:39<20:05:05, 8.00s/step, loss=1.285, acc=0.705, acc_len=1.70, thru=18.2, I=16.3, T=33.8, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:02:53,380] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 09:02:59,820] loop.py:45 INFO TRAIN_STEP step=8976 loss=1.379311 Training: 50%|████▉ | 8976/18018 [19:32:45<19:03:59, 7.59s/step, loss=1.379, acc=0.689, acc_len=1.64, thru=18.0, I=20.2, T=36.1, wait=3.1s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:02:59,816] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8975: forward=2546.8ms backward=953.3ms [2026-10-19 09:03:09,371] loop.py:45 INFO TRAIN_STEP step=8977 loss=1.468970 Training: 50%|████▉ | 8977/18018 [19:32:55<20:32:39, 8.18s/step, loss=1.469, acc=0.672, acc_len=1.59, thru=12.4, I=13.4, T=33.7, wait=5.7s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:16,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:03:17,687] loop.py:45 INFO TRAIN_STEP step=8978 loss=1.486886 Training: 50%|████▉ | 8978/18018 [19:33:03<20:38:32, 8.22s/step, loss=1.487, acc=0.667, acc_len=1.58, thru=16.0, I=19.6, T=31.1, wait=4.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:18,025] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:24,352] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:03:25,204] loop.py:45 INFO TRAIN_STEP step=8979 loss=1.302352 Training: 50%|████▉ | 8979/18018 [19:33:11<20:06:35, 8.01s/step, loss=1.302, acc=0.700, acc_len=1.69, thru=14.6, I=17.5, T=29.7, wait=3.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:25,365] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:32,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:03:32,990] loop.py:380 INFO TIMING step=8980: step=4.337s data=1.996s compute=4.297s [fwd=3.045s bwd=1.236s opt=0.003s] dispatch=3.439s [2026-10-19 09:03:32,990] loop.py:45 INFO TRAIN_STEP step=8980 loss=1.502123 Training: 50%|████▉ | 8980/18018 [19:33:19<19:56:23, 7.94s/step, loss=1.502, acc=0.666, acc_len=1.57, thru=16.0, I=17.1, T=29.5, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:33,217] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:39,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:03:39,616] loop.py:45 INFO TRAIN_STEP step=8981 loss=1.316820 Training: 50%|████▉ | 8981/18018 [19:33:25<18:56:40, 7.55s/step, loss=1.317, acc=0.694, acc_len=1.66, thru=17.1, I=21.4, T=28.6, wait=2.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:03:39,612] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8980: forward=3045.7ms backward=1376.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:40,023] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:46,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:03:47,522] loop.py:45 INFO TRAIN_STEP step=8982 loss=1.273408 Training: 50%|████▉ | 8982/18018 [19:33:33<19:12:53, 7.66s/step, loss=1.273, acc=0.705, acc_len=1.71, thru=15.4, I=17.5, T=28.3, wait=3.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:03:47,780] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 09:03:55,717] loop.py:45 INFO TRAIN_STEP step=8983 loss=1.429575 Training: 50%|████▉ | 8983/18018 [19:33:41<19:37:04, 7.82s/step, loss=1.430, acc=0.682, acc_len=1.61, thru=12.8, I=18.0, T=30.4, wait=4.0s, pool=44, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:03,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:04:04,440] loop.py:45 INFO TRAIN_STEP step=8984 loss=1.386783 Training: 50%|████▉ | 8984/18018 [19:33:50<20:17:54, 8.09s/step, loss=1.387, acc=0.686, acc_len=1.64, thru=16.2, I=16.2, T=31.5, wait=4.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:04,746] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 09:04:11,514] loop.py:380 INFO TIMING step=8985: step=3.876s data=1.785s compute=3.847s [fwd=2.696s bwd=1.135s opt=0.003s] dispatch=3.188s [2026-10-19 09:04:11,514] loop.py:45 INFO TRAIN_STEP step=8985 loss=1.393103 Training: 50%|████▉ | 8985/18018 [19:33:57<19:31:52, 7.78s/step, loss=1.393, acc=0.688, acc_len=1.64, thru=14.9, I=17.4, T=33.0, wait=3.2s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:22,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:04:22,410] loop.py:45 INFO TRAIN_STEP step=8986 loss=1.412756 Training: 50%|████▉ | 8986/18018 [19:34:08<21:52:18, 8.72s/step, loss=1.413, acc=0.682, acc_len=1.62, thru=14.4, I=14.0, T=28.8, wait=6.5s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:04:22,406] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8985: forward=3081.8ms backward=1298.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:22,527] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:27,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:04:28,358] loop.py:45 INFO TRAIN_STEP step=8987 loss=1.363501 Training: 50%|████▉ | 8987/18018 [19:34:14<19:47:09, 7.89s/step, loss=1.364, acc=0.690, acc_len=1.65, thru=18.8, I=20.1, T=34.1, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:28,604] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 09:04:36,910] loop.py:45 INFO TRAIN_STEP step=8988 loss=1.338377 Training: 50%|████▉ | 8988/18018 [19:34:22<20:16:58, 8.09s/step, loss=1.338, acc=0.698, acc_len=1.68, thru=12.6, I=13.8, T=30.4, wait=4.3s, pool=48, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:44,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:04:44,084] loop.py:45 INFO TRAIN_STEP step=8989 loss=1.462398 Training: 50%|████▉ | 8989/18018 [19:34:30<19:35:37, 7.81s/step, loss=1.462, acc=0.674, acc_len=1.59, thru=17.3, I=16.5, T=30.9, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:44,503] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:49,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:04:50,737] loop.py:380 INFO TIMING step=8990: step=3.677s data=1.464s compute=3.641s [fwd=2.547s bwd=1.080s opt=0.003s] dispatch=2.968s [2026-10-19 09:04:50,738] loop.py:45 INFO TRAIN_STEP step=8990 loss=1.401940 Training: 50%|████▉ | 8990/18018 [19:34:36<18:43:13, 7.46s/step, loss=1.402, acc=0.680, acc_len=1.61, thru=19.2, I=19.3, T=34.8, wait=3.0s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:04:51,184] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=0/64 [2026-10-19 09:04:57,173] loop.py:45 INFO TRAIN_STEP step=8991 loss=1.328231 Training: 50%|████▉ | 8991/18018 [19:34:43<17:56:37, 7.16s/step, loss=1.328, acc=0.696, acc_len=1.66, thru=17.7, I=20.3, T=35.1, wait=2.8s, pool=52, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:04:57,168] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8990: forward=2598.5ms backward=1009.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:03,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:05:04,566] loop.py:45 INFO TRAIN_STEP step=8992 loss=1.479511 Training: 50%|████▉ | 8992/18018 [19:34:50<18:07:10, 7.23s/step, loss=1.480, acc=0.672, acc_len=1.58, thru=18.4, I=19.3, T=29.6, wait=3.1s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:04,627] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=40/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:09,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:05:10,692] loop.py:45 INFO TRAIN_STEP step=8993 loss=1.349224 Training: 50%|████▉ | 8993/18018 [19:34:56<17:17:26, 6.90s/step, loss=1.349, acc=0.694, acc_len=1.66, thru=17.7, I=22.6, T=32.7, wait=2.2s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:10,812] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=24/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:16,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:05:18,451] loop.py:45 INFO TRAIN_STEP step=8994 loss=1.501034 Training: 50%|████▉ | 8994/18018 [19:35:04<17:56:13, 7.16s/step, loss=1.501, acc=0.669, acc_len=1.57, thru=15.4, I=22.0, T=25.1, wait=2.6s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:18,695] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:26,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:05:26,871] loop.py:380 INFO TIMING step=8995: step=4.742s data=2.435s compute=4.701s [fwd=3.337s bwd=1.349s opt=0.003s] dispatch=3.668s [2026-10-19 09:05:26,871] loop.py:45 INFO TRAIN_STEP step=8995 loss=1.405004 Training: 50%|████▉ | 8995/18018 [19:35:12<18:53:11, 7.54s/step, loss=1.405, acc=0.684, acc_len=1.63, thru=15.7, I=17.7, T=27.0, wait=3.7s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:27,079] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:33,551] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:05:35,835] loop.py:45 INFO TRAIN_STEP step=8996 loss=1.480328 Training: 50%|████▉ | 8996/18018 [19:35:21<19:57:30, 7.96s/step, loss=1.480, acc=0.674, acc_len=1.59, thru=13.5, I=18.5, T=27.2, wait=4.3s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:05:35,830] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=8995: forward=3274.7ms backward=1367.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:36,064] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:41,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:05:42,620] loop.py:45 INFO TRAIN_STEP step=8997 loss=1.327782 Training: 50%|████▉ | 8997/18018 [19:35:28<19:04:10, 7.61s/step, loss=1.328, acc=0.699, acc_len=1.68, thru=14.8, I=22.3, T=28.4, wait=2.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:42,655] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=56/64 [2026-10-19 09:05:50,624] loop.py:45 INFO TRAIN_STEP step=8998 loss=1.382361 Training: 50%|████▉ | 8998/18018 [19:35:36<19:21:52, 7.73s/step, loss=1.382, acc=0.689, acc_len=1.64, thru=14.2, I=15.2, T=29.6, wait=3.7s, pool=60, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:58,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:05:59,227] loop.py:45 INFO TRAIN_STEP step=8999 loss=1.414947 Training: 50%|████▉ | 8999/18018 [19:35:45<20:01:05, 7.99s/step, loss=1.415, acc=0.684, acc_len=1.62, thru=15.2, I=15.9, T=30.1, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:05:59,427] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 09:06:07,336] loop.py:380 INFO TIMING step=9000: step=4.071s data=1.827s compute=4.041s [fwd=2.886s bwd=1.140s opt=0.003s] dispatch=4.028s [2026-10-19 09:06:07,336] loop.py:45 INFO TRAIN_STEP step=9000 loss=1.333930 Training: 50%|████▉ | 9000/18018 [19:35:53<20:06:15, 8.03s/step, loss=1.334, acc=0.698, acc_len=1.67, thru=13.5, I=14.4, T=31.4, wait=4.0s, pool=44, epoch=1/2][2026-10-19 09:06:07,339] loop.py:419 INFO Saving checkpoint at step 9000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:08,402] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 09:06:11,460] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009001 (TrainerActor pid=9310) [2026-10-19 09:06:11,466] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0008251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:12,931] inference_manager.py:408 INFO Pool capacity available after 4.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:20,930] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (TrainerActor pid=9310) [2026-10-19 09:06:11,460] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 09:06:12,554] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0008251 [2026-10-19 09:06:21,607] loop.py:45 INFO TRAIN_STEP step=9001 loss=1.298345 Training: 50%|████▉ | 9001/18018 [19:36:07<24:47:41, 9.90s/step, loss=1.298, acc=0.706, acc_len=1.70, thru=14.8, I=15.9, T=26.0, wait=4.1s, pool=64, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:06:21,604] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9000: forward=3486.8ms backward=1386.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:21,935] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:27,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:06:28,085] loop.py:45 INFO TRAIN_STEP step=9002 loss=1.271511 Training: 50%|████▉ | 9002/18018 [19:36:14<22:13:24, 8.87s/step, loss=1.272, acc=0.703, acc_len=1.69, thru=18.0, I=19.2, T=31.5, wait=2.4s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:28,126] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=56/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:35,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:06:36,776] loop.py:45 INFO TRAIN_STEP step=9003 loss=1.437060 Training: 50%|████▉ | 9003/18018 [19:36:22<22:04:55, 8.82s/step, loss=1.437, acc=0.676, acc_len=1.60, thru=14.9, I=17.7, T=26.7, wait=3.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:37,029] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=0/64 [2026-10-19 09:06:45,115] loop.py:45 INFO TRAIN_STEP step=9004 loss=1.342810 Training: 50%|████▉ | 9004/18018 [19:36:31<21:43:11, 8.67s/step, loss=1.343, acc=0.694, acc_len=1.66, thru=15.8, I=16.3, T=31.8, wait=4.3s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:45,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:45,531] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:50,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:06:51,512] loop.py:380 INFO TIMING step=9005: step=3.525s data=1.508s compute=3.496s [fwd=2.511s bwd=0.971s opt=0.003s] dispatch=2.863s [2026-10-19 09:06:51,512] loop.py:45 INFO TRAIN_STEP step=9005 loss=1.382769 Training: 50%|████▉ | 9005/18018 [19:36:37<20:00:29, 7.99s/step, loss=1.383, acc=0.686, acc_len=1.64, thru=19.2, I=20.6, T=36.3, wait=2.9s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:06:51,712] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 09:06:58,146] loop.py:45 INFO TRAIN_STEP step=9006 loss=1.434856 Training: 50%|████▉ | 9006/18018 [19:36:44<18:59:09, 7.58s/step, loss=1.435, acc=0.679, acc_len=1.61, thru=18.8, I=21.7, T=33.2, wait=2.8s, pool=60, epoch=1/2](TrainerActor pid=9310) [2026-10-19 09:06:58,142] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9005: forward=2707.1ms backward=1093.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:07:06,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:07:08,784] loop.py:45 INFO TRAIN_STEP step=9007 loss=1.417369 Training: 50%|████▉ | 9007/18018 [19:36:54<21:16:32, 8.50s/step, loss=1.417, acc=0.683, acc_len=1.62, thru=13.6, I=15.4, T=24.8, wait=5.5s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:07:09,249] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=0/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:07:15,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=64/64 [2026-10-19 09:07:16,073] loop.py:45 INFO TRAIN_STEP step=9008 loss=1.351007 Training: 50%|████▉ | 9008/18018 [19:37:02<20:21:49, 8.14s/step, loss=1.351, acc=0.689, acc_len=1.65, thru=13.7, I=20.5, T=28.9, wait=2.8s, pool=64, epoch=1/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:07:16,287] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=0/64 [2026-10-19 09:07:25,172] loop.py:45 INFO TRAIN_STEP step=9009 loss=1.310032 Training: 50%|█████ | 9009/18018 [19:37:11<21:05:01, 8.43s/step, loss=1.310, acc=0.699, acc_len=1.68, thru=13.0, I=15.7, T=22.8, wait=3.5s, pool=57, epoch=1/2][2026-10-19 09:07:25,174] loop.py:433 INFO Epoch 1 completed (9009 steps). Total steps: 9009/18018 [2026-10-19 09:07:25,174] loop.py:444 INFO Saving checkpoint at end of epoch 1 (step 9009)... (TrainerActor pid=9310) [2026-10-19 09:07:28,962] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009010 (TrainerActor pid=9310) [2026-10-19 09:07:28,969] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0008501 [2026-10-19 09:07:29,906] loop.py:460 INFO Dataset exhausted, reloading (epoch 2)... (AsyncTrainingController pid=8785) [2026-10-19 09:07:30,764] training_controller.py:289 INFO Prepared dataset (1153209 samples, seed 42+1, length group 1) [2026-10-19 09:07:48,668] loop.py:380 INFO TIMING step=9010: step=3.805s data=1.612s compute=3.768s [fwd=2.668s bwd=1.084s opt=0.003s] dispatch=5.137s [2026-10-19 09:07:48,668] loop.py:45 INFO TRAIN_STEP step=9010 loss=1.317335 Training: 50%|█████ | 9010/18018 [19:37:34<32:23:41, 12.95s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=13.3, I=15.3, T=33.6, wait=5.1s, pool=41, epoch=2/2][2026-10-19 09:07:58,933] loop.py:45 INFO TRAIN_STEP step=9011 loss=1.382608 Training: 50%|█████ | 9011/18018 [19:37:45<30:22:47, 12.14s/step, loss=1.383, acc=0.687, acc_len=1.64, thru=12.4, I=12.8, T=31.0, wait=6.1s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:07:58,929] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9010: forward=2900.1ms backward=1157.2ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 09:07:28,963] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009010 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 09:07:29,904] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0008501 [2026-10-19 09:08:08,478] loop.py:45 INFO TRAIN_STEP step=9012 loss=1.310639 Training: 50%|█████ | 9012/18018 [19:37:54<28:25:34, 11.36s/step, loss=1.311, acc=0.703, acc_len=1.69, thru=14.9, I=13.5, T=27.0, wait=4.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:13,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:08:15,848] loop.py:45 INFO TRAIN_STEP step=9013 loss=1.358083 Training: 50%|█████ | 9013/18018 [19:38:01<25:25:40, 10.17s/step, loss=1.358, acc=0.693, acc_len=1.66, thru=18.1, I=18.2, T=31.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:15,989] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:20,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:08:22,219] loop.py:45 INFO TRAIN_STEP step=9014 loss=1.385399 Training: 50%|█████ | 9014/18018 [19:38:08<22:34:40, 9.03s/step, loss=1.385, acc=0.688, acc_len=1.64, thru=18.1, I=27.7, T=32.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:22,369] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:27,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:08:27,879] loop.py:380 INFO TIMING step=9015: step=3.747s data=1.451s compute=3.715s [fwd=2.547s bwd=1.152s opt=0.003s] dispatch=1.903s [2026-10-19 09:08:27,879] loop.py:45 INFO TRAIN_STEP step=9015 loss=1.300078 Training: 50%|█████ | 9015/18018 [19:38:13<20:02:54, 8.02s/step, loss=1.300, acc=0.699, acc_len=1.68, thru=21.0, I=27.4, T=34.2, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:28,276] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:33,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:08:34,985] loop.py:45 INFO TRAIN_STEP step=9016 loss=1.320141 Training: 50%|█████ | 9016/18018 [19:38:21<19:21:51, 7.74s/step, loss=1.320, acc=0.697, acc_len=1.68, thru=17.7, I=20.7, T=33.3, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:08:34,980] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9015: forward=2721.3ms backward=1073.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:35,172] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:42,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:08:43,957] loop.py:45 INFO TRAIN_STEP step=9017 loss=1.453323 Training: 50%|█████ | 9017/18018 [19:38:30<20:16:59, 8.11s/step, loss=1.453, acc=0.676, acc_len=1.60, thru=14.6, I=16.9, T=30.3, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:44,164] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:50,128] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:08:50,724] loop.py:45 INFO TRAIN_STEP step=9018 loss=1.281605 Training: 50%|█████ | 9018/18018 [19:38:36<19:16:14, 7.71s/step, loss=1.282, acc=0.701, acc_len=1.69, thru=17.2, I=23.7, T=31.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:51,131] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:55,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:08:58,348] loop.py:45 INFO TRAIN_STEP step=9019 loss=1.352192 Training: 50%|█████ | 9019/18018 [19:38:44<19:12:22, 7.68s/step, loss=1.352, acc=0.695, acc_len=1.68, thru=16.3, I=21.9, T=27.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:08:58,817] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:03,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:09:05,961] loop.py:380 INFO TIMING step=9020: step=4.451s data=2.210s compute=4.417s [fwd=3.069s bwd=1.333s opt=0.003s] dispatch=3.151s [2026-10-19 09:09:05,961] loop.py:45 INFO TRAIN_STEP step=9020 loss=1.399740 Training: 50%|█████ | 9020/18018 [19:38:52<19:08:57, 7.66s/step, loss=1.400, acc=0.684, acc_len=1.63, thru=17.9, I=26.3, T=28.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:06,384] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:11,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:09:13,896] loop.py:45 INFO TRAIN_STEP step=9021 loss=1.244798 (TrainerActor pid=9310) [2026-10-19 09:09:13,892] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9020: forward=3743.9ms backward=1591.3ms Training: 50%|█████ | 9021/18018 [19:38:59<19:21:10, 7.74s/step, loss=1.245, acc=0.712, acc_len=1.72, thru=17.0, I=20.7, T=23.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:14,201] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:20,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:09:21,519] loop.py:45 INFO TRAIN_STEP step=9022 loss=1.437923 Training: 50%|█████ | 9022/18018 [19:39:07<19:15:40, 7.71s/step, loss=1.438, acc=0.678, acc_len=1.61, thru=13.7, I=19.7, T=30.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:21,770] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:27,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:09:28,944] loop.py:45 INFO TRAIN_STEP step=9023 loss=1.454639 Training: 50%|█████ | 9023/18018 [19:39:15<19:02:47, 7.62s/step, loss=1.455, acc=0.674, acc_len=1.60, thru=17.3, I=21.2, T=26.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:29,067] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:35,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:09:36,084] loop.py:45 INFO TRAIN_STEP step=9024 loss=1.262065 Training: 50%|█████ | 9024/18018 [19:39:22<18:40:58, 7.48s/step, loss=1.262, acc=0.711, acc_len=1.72, thru=15.2, I=19.5, T=31.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:36,355] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 09:09:43,142] loop.py:380 INFO TIMING step=9025: step=3.901s data=1.793s compute=3.867s [fwd=2.768s bwd=1.085s opt=0.003s] dispatch=3.147s [2026-10-19 09:09:43,142] loop.py:45 INFO TRAIN_STEP step=9025 loss=1.330672 Training: 50%|█████ | 9025/18018 [19:39:29<18:21:52, 7.35s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=18.6, I=17.1, T=32.8, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:52,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:09:53,526] loop.py:45 INFO TRAIN_STEP step=9026 loss=1.319148 Training: 50%|█████ | 9026/18018 [19:39:39<20:38:12, 8.26s/step, loss=1.319, acc=0.700, acc_len=1.69, thru=13.4, I=14.6, T=32.9, wait=6.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:09:53,522] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9025: forward=2770.7ms backward=1077.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:09:53,955] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:00,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:10:01,153] loop.py:45 INFO TRAIN_STEP step=9027 loss=1.290405 Training: 50%|█████ | 9027/18018 [19:39:47<20:09:23, 8.07s/step, loss=1.290, acc=0.703, acc_len=1.70, thru=17.3, I=20.0, T=36.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:01,622] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:05,839] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:10:07,796] loop.py:45 INFO TRAIN_STEP step=9028 loss=1.336691 Training: 50%|█████ | 9028/18018 [19:39:53<19:05:09, 7.64s/step, loss=1.337, acc=0.696, acc_len=1.67, thru=19.5, I=23.5, T=29.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:07,860] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 [2026-10-19 09:10:13,724] loop.py:45 INFO TRAIN_STEP step=9029 loss=1.491118 Training: 50%|█████ | 9029/18018 [19:39:59<17:48:03, 7.13s/step, loss=1.491, acc=0.671, acc_len=1.58, thru=18.3, I=24.4, T=33.0, wait=2.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:21,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:10:22,454] loop.py:380 INFO TIMING step=9030: step=4.208s data=1.845s compute=4.178s [fwd=2.996s bwd=1.167s opt=0.003s] dispatch=4.511s [2026-10-19 09:10:22,454] loop.py:45 INFO TRAIN_STEP step=9030 loss=1.371031 Training: 50%|█████ | 9030/18018 [19:40:08<18:59:49, 7.61s/step, loss=1.371, acc=0.691, acc_len=1.66, thru=15.4, I=16.7, T=30.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:22,822] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:30,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:10:30,806] loop.py:45 INFO TRAIN_STEP step=9031 loss=1.527427 (TrainerActor pid=9310) [2026-10-19 09:10:30,802] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9030: forward=3065.8ms backward=1175.4ms Training: 50%|█████ | 9031/18018 [19:40:16<19:33:05, 7.83s/step, loss=1.527, acc=0.663, acc_len=1.55, thru=14.0, I=18.0, T=29.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:30,931] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:38,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:10:40,846] loop.py:45 INFO TRAIN_STEP step=9032 loss=1.436903 Training: 50%|█████ | 9032/18018 [19:40:26<21:12:06, 8.49s/step, loss=1.437, acc=0.679, acc_len=1.61, thru=12.9, I=15.9, T=26.3, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:41,166] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 09:10:48,192] loop.py:45 INFO TRAIN_STEP step=9033 loss=1.378600 Training: 50%|█████ | 9033/18018 [19:40:34<20:20:19, 8.15s/step, loss=1.379, acc=0.690, acc_len=1.65, thru=13.3, I=17.1, T=33.5, wait=3.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:53,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:10:55,281] loop.py:45 INFO TRAIN_STEP step=9034 loss=1.391003 Training: 50%|█████ | 9034/18018 [19:40:41<19:32:39, 7.83s/step, loss=1.391, acc=0.688, acc_len=1.64, thru=19.2, I=19.9, T=33.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:10:55,460] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:01,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:11:02,167] loop.py:380 INFO TIMING step=9035: step=3.869s data=1.726s compute=3.840s [fwd=2.731s bwd=1.095s opt=0.003s] dispatch=3.008s [2026-10-19 09:11:02,167] loop.py:45 INFO TRAIN_STEP step=9035 loss=1.354287 Training: 50%|█████ | 9035/18018 [19:40:48<18:50:00, 7.55s/step, loss=1.354, acc=0.693, acc_len=1.66, thru=17.8, I=23.7, T=33.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:02,526] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:07,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:11:08,179] loop.py:45 INFO TRAIN_STEP step=9036 loss=1.357322 Training: 50%|█████ | 9036/18018 [19:40:54<17:40:55, 7.09s/step, loss=1.357, acc=0.691, acc_len=1.65, thru=19.8, I=25.2, T=33.7, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:11:08,175] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9035: forward=2709.2ms backward=1044.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:08,314] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 09:11:14,918] loop.py:45 INFO TRAIN_STEP step=9037 loss=1.354573 Training: 50%|█████ | 9037/18018 [19:41:01<17:25:06, 6.98s/step, loss=1.355, acc=0.692, acc_len=1.66, thru=16.9, I=19.2, T=31.0, wait=2.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:25,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:11:25,892] loop.py:45 INFO TRAIN_STEP step=9038 loss=1.368324 Training: 50%|█████ | 9038/18018 [19:41:11<20:24:20, 8.18s/step, loss=1.368, acc=0.691, acc_len=1.65, thru=14.4, I=13.0, T=27.4, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:26,202] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:32,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:11:34,375] loop.py:45 INFO TRAIN_STEP step=9039 loss=1.372966 Training: 50%|█████ | 9039/18018 [19:41:20<20:37:47, 8.27s/step, loss=1.373, acc=0.689, acc_len=1.64, thru=15.5, I=16.9, T=30.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:34,808] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:11:41,313] loop.py:380 INFO TIMING step=9040: step=4.547s data=2.100s compute=4.519s [fwd=3.092s bwd=1.413s opt=0.003s] dispatch=2.380s [2026-10-19 09:11:41,313] loop.py:45 INFO TRAIN_STEP step=9040 loss=1.370098 Training: 50%|█████ | 9040/18018 [19:41:27<19:37:49, 7.87s/step, loss=1.370, acc=0.690, acc_len=1.65, thru=15.3, I=22.9, T=28.1, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:48,892] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:11:49,549] loop.py:45 INFO TRAIN_STEP step=9041 loss=1.384029 Training: 50%|█████ | 9041/18018 [19:41:35<19:54:05, 7.98s/step, loss=1.384, acc=0.690, acc_len=1.65, thru=15.2, I=14.9, T=30.1, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:11:49,545] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9040: forward=3043.6ms backward=1157.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:49,898] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:55,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:11:56,034] loop.py:45 INFO TRAIN_STEP step=9042 loss=1.395512 Training: 50%|█████ | 9042/18018 [19:41:42<18:46:44, 7.53s/step, loss=1.396, acc=0.685, acc_len=1.63, thru=18.5, I=19.5, T=32.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:11:56,435] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:12:03,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:12:05,174] loop.py:45 INFO TRAIN_STEP step=9043 loss=1.302760 Training: 50%|█████ | 9043/18018 [19:41:51<19:58:45, 8.01s/step, loss=1.303, acc=0.698, acc_len=1.68, thru=14.9, I=17.8, T=28.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:12:05,671] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:12:12,648] loop.py:45 INFO TRAIN_STEP step=9044 loss=1.424588 Training: 50%|█████ | 9044/18018 [19:41:58<19:34:23, 7.85s/step, loss=1.425, acc=0.680, acc_len=1.62, thru=14.3, I=17.9, T=28.9, wait=3.0s, pool=57, epoch=2/2][2026-10-19 09:12:20,668] loop.py:380 INFO TIMING step=9045: step=4.419s data=2.058s compute=4.365s [fwd=2.970s bwd=1.381s opt=0.003s] dispatch=3.593s [2026-10-19 09:12:20,668] loop.py:45 INFO TRAIN_STEP step=9045 loss=1.441671 Training: 50%|█████ | 9045/18018 [19:42:06<19:41:52, 7.90s/step, loss=1.442, acc=0.679, acc_len=1.61, thru=16.2, I=14.3, T=29.0, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:12:26,418] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 09:12:27,125] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9045: forward=2744.5ms backward=1084.0ms [2026-10-19 09:12:27,130] loop.py:45 INFO TRAIN_STEP step=9046 loss=1.483613 Training: 50%|█████ | 9046/18018 [19:42:13<18:37:08, 7.47s/step, loss=1.484, acc=0.670, acc_len=1.57, thru=19.1, I=19.3, T=33.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:12:27,422] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:12:33,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:12:34,258] loop.py:45 INFO TRAIN_STEP step=9047 loss=1.367482 Training: 50%|█████ | 9047/18018 [19:42:20<18:21:30, 7.37s/step, loss=1.367, acc=0.690, acc_len=1.64, thru=17.9, I=21.3, T=29.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:12:34,710] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:12:41,708] loop.py:45 INFO TRAIN_STEP step=9048 loss=1.345639 Training: 50%|█████ | 9048/18018 [19:42:27<18:25:12, 7.39s/step, loss=1.346, acc=0.692, acc_len=1.66, thru=13.3, I=19.5, T=30.9, wait=3.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:12:52,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:12:53,487] loop.py:45 INFO TRAIN_STEP step=9049 loss=1.413544 Training: 50%|█████ | 9049/18018 [19:42:39<21:41:44, 8.71s/step, loss=1.414, acc=0.683, acc_len=1.62, thru=14.3, I=11.6, T=27.8, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:12:53,893] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:13:00,762] loop.py:380 INFO TIMING step=9050: step=4.174s data=1.901s compute=4.126s [fwd=2.949s bwd=1.161s opt=0.003s] dispatch=3.092s [2026-10-19 09:13:00,762] loop.py:45 INFO TRAIN_STEP step=9050 loss=1.242395 Training: 50%|█████ | 9050/18018 [19:42:46<20:37:23, 8.28s/step, loss=1.242, acc=0.713, acc_len=1.73, thru=15.0, I=17.5, T=30.7, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:07,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:13:08,932] loop.py:45 INFO TRAIN_STEP step=9051 loss=1.507493 Training: 50%|█████ | 9051/18018 [19:42:55<20:32:21, 8.25s/step, loss=1.507, acc=0.665, acc_len=1.56, thru=17.0, I=18.9, T=27.9, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:13:08,928] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9050: forward=3289.2ms backward=1250.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:09,217] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:14,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:13:16,682] loop.py:45 INFO TRAIN_STEP step=9052 loss=1.368407 Training: 50%|█████ | 9052/18018 [19:43:02<20:09:54, 8.10s/step, loss=1.368, acc=0.688, acc_len=1.64, thru=14.0, I=21.7, T=29.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:16,734] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:23,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:13:24,893] loop.py:45 INFO TRAIN_STEP step=9053 loss=1.405387 Training: 50%|█████ | 9053/18018 [19:43:10<20:15:01, 8.13s/step, loss=1.405, acc=0.684, acc_len=1.63, thru=15.7, I=18.7, T=28.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:24,994] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:32,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:13:32,998] loop.py:45 INFO TRAIN_STEP step=9054 loss=1.478396 Training: 50%|█████ | 9054/18018 [19:43:19<20:13:33, 8.12s/step, loss=1.478, acc=0.674, acc_len=1.59, thru=13.7, I=16.0, T=28.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:33,398] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:13:42,143] loop.py:380 INFO TIMING step=9055: step=4.727s data=2.219s compute=4.695s [fwd=3.390s bwd=1.291s opt=0.003s] dispatch=4.411s [2026-10-19 09:13:42,143] loop.py:45 INFO TRAIN_STEP step=9055 loss=1.364374 Training: 50%|█████ | 9055/18018 [19:43:28<20:59:11, 8.43s/step, loss=1.364, acc=0.692, acc_len=1.65, thru=12.3, I=13.6, T=27.1, wait=4.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:51,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:13:51,343] loop.py:45 INFO TRAIN_STEP step=9056 loss=1.220061 Training: 50%|█████ | 9056/18018 [19:43:37<21:33:42, 8.66s/step, loss=1.220, acc=0.718, acc_len=1.74, thru=15.3, I=15.0, T=30.5, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:13:51,339] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9055: forward=2905.0ms backward=1241.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:51,647] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:58,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:13:58,753] loop.py:45 INFO TRAIN_STEP step=9057 loss=1.404414 Training: 50%|█████ | 9057/18018 [19:43:44<20:37:23, 8.29s/step, loss=1.404, acc=0.680, acc_len=1.61, thru=17.3, I=17.3, T=29.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:13:58,814] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:05,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:14:07,210] loop.py:45 INFO TRAIN_STEP step=9058 loss=1.327003 Training: 50%|█████ | 9058/18018 [19:43:53<20:45:00, 8.34s/step, loss=1.327, acc=0.697, acc_len=1.67, thru=14.7, I=18.7, T=25.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:07,476] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:15,021] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:14:16,481] loop.py:45 INFO TRAIN_STEP step=9059 loss=1.344217 Training: 50%|█████ | 9059/18018 [19:44:02<21:26:46, 8.62s/step, loss=1.344, acc=0.694, acc_len=1.67, thru=14.2, I=17.3, T=29.7, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:16,542] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:21,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:14:23,036] loop.py:380 INFO TIMING step=9060: step=4.033s data=1.846s compute=4.002s [fwd=2.837s bwd=1.150s opt=0.003s] dispatch=2.511s [2026-10-19 09:14:23,036] loop.py:45 INFO TRAIN_STEP step=9060 loss=1.330294 Training: 50%|█████ | 9060/18018 [19:44:09<19:54:14, 8.00s/step, loss=1.330, acc=0.698, acc_len=1.68, thru=18.5, I=24.8, T=31.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:23,399] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:30,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:14:30,757] loop.py:45 INFO TRAIN_STEP step=9061 loss=1.364601 Training: 50%|█████ | 9061/18018 [19:44:16<19:41:34, 7.91s/step, loss=1.365, acc=0.688, acc_len=1.64, thru=14.8, I=19.4, T=32.5, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:14:30,753] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9060: forward=2760.2ms backward=1127.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:31,255] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:14:37,669] loop.py:45 INFO TRAIN_STEP step=9062 loss=1.364117 Training: 50%|█████ | 9062/18018 [19:44:23<18:56:29, 7.61s/step, loss=1.364, acc=0.695, acc_len=1.66, thru=18.2, I=19.3, T=33.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:37,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:38,160] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:46,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:14:47,055] loop.py:45 INFO TRAIN_STEP step=9063 loss=1.339812 Training: 50%|█████ | 9063/18018 [19:44:33<20:15:42, 8.15s/step, loss=1.340, acc=0.699, acc_len=1.68, thru=14.0, I=14.7, T=30.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:47,462] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:52,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:14:54,728] loop.py:45 INFO TRAIN_STEP step=9064 loss=1.447901 Training: 50%|█████ | 9064/18018 [19:44:40<19:54:25, 8.00s/step, loss=1.448, acc=0.673, acc_len=1.60, thru=17.2, I=21.4, T=31.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:14:54,947] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:15:03,449] loop.py:380 INFO TIMING step=9065: step=3.822s data=1.493s compute=3.795s [fwd=2.663s bwd=1.117s opt=0.003s] dispatch=4.891s [2026-10-19 09:15:03,449] loop.py:45 INFO TRAIN_STEP step=9065 loss=1.397859 Training: 50%|█████ | 9065/18018 [19:44:49<20:26:26, 8.22s/step, loss=1.398, acc=0.685, acc_len=1.63, thru=13.2, I=14.1, T=33.5, wait=4.9s, pool=49, epoch=2/2][2026-10-19 09:15:11,608] loop.py:45 INFO TRAIN_STEP step=9066 loss=1.483809 Training: 50%|█████ | 9066/18018 [19:44:57<20:23:39, 8.20s/step, loss=1.484, acc=0.671, acc_len=1.58, thru=17.0, I=16.6, T=31.8, wait=4.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:15:11,603] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9065: forward=2769.6ms backward=1179.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:17,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:15:18,717] loop.py:45 INFO TRAIN_STEP step=9067 loss=1.346507 Training: 50%|█████ | 9067/18018 [19:45:04<19:34:39, 7.87s/step, loss=1.347, acc=0.693, acc_len=1.66, thru=17.7, I=18.7, T=29.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:19,068] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:15:26,421] loop.py:45 INFO TRAIN_STEP step=9068 loss=1.380015 Training: 50%|█████ | 9068/18018 [19:45:12<19:26:47, 7.82s/step, loss=1.380, acc=0.689, acc_len=1.65, thru=13.9, I=15.0, T=31.4, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:34,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:15:35,521] loop.py:45 INFO TRAIN_STEP step=9069 loss=1.238659 Training: 50%|█████ | 9069/18018 [19:45:21<20:23:54, 8.21s/step, loss=1.239, acc=0.715, acc_len=1.74, thru=16.2, I=15.5, T=31.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:35,975] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:41,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:15:42,356] loop.py:380 INFO TIMING step=9070: step=4.338s data=2.007s compute=4.303s [fwd=3.010s bwd=1.278s opt=0.003s] dispatch=2.488s [2026-10-19 09:15:42,356] loop.py:45 INFO TRAIN_STEP step=9070 loss=1.453433 Training: 50%|█████ | 9070/18018 [19:45:28<19:22:29, 7.80s/step, loss=1.453, acc=0.674, acc_len=1.59, thru=17.3, I=20.6, T=29.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:42,537] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:47,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:15:48,807] loop.py:45 INFO TRAIN_STEP step=9071 loss=1.351723 Training: 50%|█████ | 9071/18018 [19:45:34<18:22:15, 7.39s/step, loss=1.352, acc=0.689, acc_len=1.64, thru=17.3, I=21.9, T=30.0, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:15:48,803] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9070: forward=3006.4ms backward=1215.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:48,946] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:54,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:15:55,602] loop.py:45 INFO TRAIN_STEP step=9072 loss=1.322561 Training: 50%|█████ | 9072/18018 [19:45:41<17:55:26, 7.21s/step, loss=1.323, acc=0.699, acc_len=1.68, thru=17.0, I=19.6, T=30.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:15:55,927] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:16:01,418] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:16:03,590] loop.py:45 INFO TRAIN_STEP step=9073 loss=1.389331 Training: 50%|█████ | 9073/18018 [19:45:49<18:29:53, 7.44s/step, loss=1.389, acc=0.686, acc_len=1.63, thru=16.2, I=20.5, T=25.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:16:03,930] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 09:16:10,791] loop.py:45 INFO TRAIN_STEP step=9074 loss=1.339783 Training: 50%|█████ | 9074/18018 [19:45:56<18:18:51, 7.37s/step, loss=1.340, acc=0.695, acc_len=1.66, thru=12.2, I=18.1, T=31.8, wait=3.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:16:18,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:16:20,500] loop.py:380 INFO TIMING step=9075: step=4.871s data=2.427s compute=4.824s [fwd=3.333s bwd=1.476s opt=0.003s] dispatch=4.830s [2026-10-19 09:16:20,500] loop.py:45 INFO TRAIN_STEP step=9075 loss=1.490742 Training: 50%|█████ | 9075/18018 [19:46:06<20:03:18, 8.07s/step, loss=1.491, acc=0.669, acc_len=1.58, thru=15.7, I=15.9, T=26.3, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:16:20,576] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 [2026-10-19 09:16:27,646] loop.py:45 INFO TRAIN_STEP step=9076 loss=1.289911 Training: 50%|█████ | 9076/18018 [19:46:13<19:21:40, 7.79s/step, loss=1.290, acc=0.705, acc_len=1.70, thru=14.5, I=17.7, T=30.2, wait=2.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:16:27,642] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9075: forward=2874.6ms backward=1320.1ms [2026-10-19 09:16:34,275] loop.py:45 INFO TRAIN_STEP step=9077 loss=1.373462 Training: 50%|█████ | 9077/18018 [19:46:20<18:29:28, 7.45s/step, loss=1.373, acc=0.685, acc_len=1.63, thru=17.3, I=17.6, T=33.4, wait=2.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:16:41,136] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:16:43,051] loop.py:45 INFO TRAIN_STEP step=9078 loss=1.368486 Training: 50%|█████ | 9078/18018 [19:46:29<19:28:54, 7.84s/step, loss=1.368, acc=0.686, acc_len=1.64, thru=16.1, I=17.6, T=26.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:16:43,160] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 [2026-10-19 09:16:50,230] loop.py:45 INFO TRAIN_STEP step=9079 loss=1.409873 Training: 50%|█████ | 9079/18018 [19:46:36<18:58:52, 7.64s/step, loss=1.410, acc=0.681, acc_len=1.62, thru=13.8, I=19.7, T=27.4, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:16:56,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:16:57,771] loop.py:380 INFO TIMING step=9080: step=4.023s data=1.770s compute=3.989s [fwd=2.867s bwd=1.108s opt=0.003s] dispatch=3.510s [2026-10-19 09:16:57,772] loop.py:45 INFO TRAIN_STEP step=9080 loss=1.443740 Training: 50%|█████ | 9080/18018 [19:46:43<18:54:10, 7.61s/step, loss=1.444, acc=0.678, acc_len=1.61, thru=16.7, I=16.6, T=31.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:16:57,855] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:04,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:17:06,105] loop.py:45 INFO TRAIN_STEP step=9081 loss=1.332067 Training: 50%|█████ | 9081/18018 [19:46:52<19:26:14, 7.83s/step, loss=1.332, acc=0.696, acc_len=1.67, thru=14.2, I=19.0, T=27.9, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:17:06,100] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9080: forward=3164.7ms backward=1369.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:06,201] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:13,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:17:15,060] loop.py:45 INFO TRAIN_STEP step=9082 loss=1.346321 Training: 50%|█████ | 9082/18018 [19:47:01<20:16:21, 8.17s/step, loss=1.346, acc=0.691, acc_len=1.66, thru=14.5, I=16.3, T=28.8, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:15,248] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:20,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:17:21,180] loop.py:45 INFO TRAIN_STEP step=9083 loss=1.334604 Training: 50%|█████ | 9083/18018 [19:47:07<18:44:51, 7.55s/step, loss=1.335, acc=0.695, acc_len=1.67, thru=18.9, I=21.8, T=34.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:21,376] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:17:28,325] loop.py:45 INFO TRAIN_STEP step=9084 loss=1.284361 Training: 50%|█████ | 9084/18018 [19:47:14<18:26:25, 7.43s/step, loss=1.284, acc=0.704, acc_len=1.70, thru=17.5, I=20.2, T=30.3, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:36,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:17:36,703] loop.py:380 INFO TIMING step=9085: step=3.878s data=1.674s compute=3.848s [fwd=2.717s bwd=1.115s opt=0.003s] dispatch=4.491s [2026-10-19 09:17:36,703] loop.py:45 INFO TRAIN_STEP step=9085 loss=1.369541 Training: 50%|█████ | 9085/18018 [19:47:22<19:08:34, 7.71s/step, loss=1.370, acc=0.692, acc_len=1.65, thru=16.5, I=15.7, T=33.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:36,870] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:44,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:17:45,516] loop.py:45 INFO TRAIN_STEP step=9086 loss=1.308126 Training: 50%|█████ | 9086/18018 [19:47:31<19:57:34, 8.04s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=14.9, I=16.4, T=31.4, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:17:45,512] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9085: forward=2911.5ms backward=1110.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:17:45,614] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-19 09:17:51,959] loop.py:45 INFO TRAIN_STEP step=9087 loss=1.260544 Training: 50%|█████ | 9087/18018 [19:47:38<18:45:52, 7.56s/step, loss=1.261, acc=0.710, acc_len=1.71, thru=18.1, I=20.0, T=33.5, wait=2.6s, pool=61, epoch=2/2][2026-10-19 09:18:00,087] loop.py:45 INFO TRAIN_STEP step=9088 loss=1.458974 Training: 50%|█████ | 9088/18018 [19:47:46<19:11:01, 7.73s/step, loss=1.459, acc=0.674, acc_len=1.60, thru=15.7, I=15.0, T=29.7, wait=3.8s, pool=61, epoch=2/2][2026-10-19 09:18:07,549] loop.py:45 INFO TRAIN_STEP step=9089 loss=1.386275 Training: 50%|█████ | 9089/18018 [19:47:53<18:58:51, 7.65s/step, loss=1.386, acc=0.685, acc_len=1.63, thru=15.5, I=15.3, T=31.8, wait=3.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:14,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:18:16,824] loop.py:380 INFO TIMING step=9090: step=4.364s data=1.968s compute=4.331s [fwd=3.060s bwd=1.256s opt=0.003s] dispatch=4.898s [2026-10-19 09:18:16,824] loop.py:45 INFO TRAIN_STEP step=9090 loss=1.308278 Training: 50%|█████ | 9090/18018 [19:48:02<20:11:00, 8.14s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=16.0, I=19.0, T=29.3, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:17,174] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:24,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:18:25,028] loop.py:45 INFO TRAIN_STEP step=9091 loss=1.432028 Training: 50%|█████ | 9091/18018 [19:48:11<20:13:48, 8.16s/step, loss=1.432, acc=0.680, acc_len=1.61, thru=16.3, I=19.5, T=29.8, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:18:25,023] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9090: forward=2982.5ms backward=1266.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:25,175] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:30,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:18:32,744] loop.py:45 INFO TRAIN_STEP step=9092 loss=1.348696 Training: 50%|█████ | 9092/18018 [19:48:18<19:53:56, 8.03s/step, loss=1.349, acc=0.694, acc_len=1.66, thru=16.0, I=19.8, T=28.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:32,958] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 09:18:39,611] loop.py:45 INFO TRAIN_STEP step=9093 loss=1.385487 Training: 50%|█████ | 9093/18018 [19:48:25<19:01:57, 7.68s/step, loss=1.385, acc=0.686, acc_len=1.63, thru=14.9, I=20.4, T=30.6, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:50,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:18:51,752] loop.py:45 INFO TRAIN_STEP step=9094 loss=1.376484 Training: 50%|█████ | 9094/18018 [19:48:37<22:21:04, 9.02s/step, loss=1.376, acc=0.689, acc_len=1.64, thru=10.1, I=11.1, T=23.0, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:51,908] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:57,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:18:58,624] loop.py:380 INFO TIMING step=9095: step=4.260s data=1.934s compute=4.218s [fwd=3.004s bwd=1.199s opt=0.003s] dispatch=2.604s [2026-10-19 09:18:58,624] loop.py:45 INFO TRAIN_STEP step=9095 loss=1.371717 Training: 50%|█████ | 9095/18018 [19:48:44<20:45:17, 8.37s/step, loss=1.372, acc=0.688, acc_len=1.65, thru=18.3, I=18.8, T=30.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:18:58,892] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:05,906] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:19:07,318] loop.py:45 INFO TRAIN_STEP step=9096 loss=1.408023 Training: 50%|█████ | 9096/18018 [19:48:53<20:59:27, 8.47s/step, loss=1.408, acc=0.681, acc_len=1.62, thru=13.3, I=18.2, T=31.0, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:19:07,313] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9095: forward=2923.4ms backward=1154.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:07,424] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:14,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:19:15,125] loop.py:45 INFO TRAIN_STEP step=9097 loss=1.347970 Training: 50%|█████ | 9097/18018 [19:49:01<20:29:45, 8.27s/step, loss=1.348, acc=0.692, acc_len=1.66, thru=14.4, I=16.5, T=28.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:15,612] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:21,308] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:19:22,668] loop.py:45 INFO TRAIN_STEP step=9098 loss=1.296347 Training: 50%|█████ | 9098/18018 [19:49:08<19:57:10, 8.05s/step, loss=1.296, acc=0.702, acc_len=1.69, thru=16.8, I=20.9, T=34.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:22,816] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:28,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:19:30,033] loop.py:45 INFO TRAIN_STEP step=9099 loss=1.242408 Training: 50%|█████ | 9099/18018 [19:49:16<19:26:24, 7.85s/step, loss=1.242, acc=0.713, acc_len=1.72, thru=17.4, I=22.4, T=27.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:30,070] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:37,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:19:38,017] loop.py:380 INFO TIMING step=9100: step=4.981s data=2.378s compute=4.952s [fwd=3.403s bwd=1.534s opt=0.003s] dispatch=2.992s [2026-10-19 09:19:38,017] loop.py:45 INFO TRAIN_STEP step=9100 loss=1.309998 Training: 51%|█████ | 9100/18018 [19:49:24<19:32:20, 7.89s/step, loss=1.310, acc=0.700, acc_len=1.69, thru=14.4, I=17.7, T=25.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:38,261] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:44,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:19:45,408] loop.py:45 INFO TRAIN_STEP step=9101 loss=1.416026 Training: 51%|█████ | 9101/18018 [19:49:31<19:10:03, 7.74s/step, loss=1.416, acc=0.678, acc_len=1.61, thru=16.4, I=17.5, T=27.0, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:19:45,404] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9100: forward=3579.7ms backward=1112.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:19:45,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:19:52,844] loop.py:45 INFO TRAIN_STEP step=9102 loss=1.356203 Training: 51%|█████ | 9102/18018 [19:49:38<18:56:24, 7.65s/step, loss=1.356, acc=0.693, acc_len=1.66, thru=15.3, I=17.3, T=35.6, wait=3.8s, pool=57, epoch=2/2][2026-10-19 09:20:00,100] loop.py:45 INFO TRAIN_STEP step=9103 loss=1.206160 Training: 51%|█████ | 9103/18018 [19:49:46<18:38:46, 7.53s/step, loss=1.206, acc=0.719, acc_len=1.74, thru=17.3, I=17.6, T=30.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:00,092] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:00,594] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:07,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:20:07,989] loop.py:45 INFO TRAIN_STEP step=9104 loss=1.356799 Training: 51%|█████ | 9104/18018 [19:49:54<18:54:47, 7.64s/step, loss=1.357, acc=0.694, acc_len=1.65, thru=16.6, I=18.3, T=31.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:08,078] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:14,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:20:15,873] loop.py:380 INFO TIMING step=9105: step=4.352s data=2.059s compute=4.324s [fwd=3.058s bwd=1.252s opt=0.003s] dispatch=3.522s [2026-10-19 09:20:15,873] loop.py:45 INFO TRAIN_STEP step=9105 loss=1.324549 Training: 51%|█████ | 9105/18018 [19:50:01<19:05:33, 7.71s/step, loss=1.325, acc=0.697, acc_len=1.67, thru=15.3, I=17.7, T=29.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:15,923] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:20,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:20:22,373] loop.py:45 INFO TRAIN_STEP step=9106 loss=1.306834 Training: 51%|█████ | 9106/18018 [19:50:08<18:11:22, 7.35s/step, loss=1.307, acc=0.702, acc_len=1.69, thru=17.9, I=25.2, T=28.9, wait=2.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:20:22,370] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9105: forward=3050.8ms backward=1325.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:22,785] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:28,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:20:29,494] loop.py:45 INFO TRAIN_STEP step=9107 loss=1.393710 Training: 51%|█████ | 9107/18018 [19:50:15<18:01:15, 7.28s/step, loss=1.394, acc=0.686, acc_len=1.63, thru=15.3, I=22.4, T=30.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:29,579] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:37,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:20:37,388] loop.py:45 INFO TRAIN_STEP step=9108 loss=1.325540 Training: 51%|█████ | 9108/18018 [19:50:23<18:28:30, 7.46s/step, loss=1.326, acc=0.697, acc_len=1.67, thru=14.3, I=18.0, T=32.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:37,565] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:45,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:20:45,874] loop.py:45 INFO TRAIN_STEP step=9109 loss=1.321896 Training: 51%|█████ | 9109/18018 [19:50:31<19:13:53, 7.77s/step, loss=1.322, acc=0.697, acc_len=1.67, thru=14.6, I=15.8, T=28.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:46,042] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:52,646] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:20:53,205] loop.py:380 INFO TIMING step=9110: step=4.382s data=1.812s compute=4.352s [fwd=3.067s bwd=1.271s opt=0.003s] dispatch=2.937s [2026-10-19 09:20:53,205] loop.py:45 INFO TRAIN_STEP step=9110 loss=1.313798 Training: 51%|█████ | 9110/18018 [19:50:39<18:54:03, 7.64s/step, loss=1.314, acc=0.706, acc_len=1.71, thru=17.0, I=18.5, T=29.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:20:53,650] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:00,666] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:21:02,310] loop.py:45 INFO TRAIN_STEP step=9111 loss=1.208651 Training: 51%|█████ | 9111/18018 [19:50:48<19:59:17, 8.08s/step, loss=1.209, acc=0.718, acc_len=1.75, thru=13.8, I=17.0, T=30.9, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:21:02,306] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9110: forward=2961.4ms backward=1132.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:02,680] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:08,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:21:09,123] loop.py:45 INFO TRAIN_STEP step=9112 loss=1.277817 Training: 51%|█████ | 9112/18018 [19:50:55<19:02:41, 7.70s/step, loss=1.278, acc=0.703, acc_len=1.70, thru=17.6, I=24.7, T=30.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:09,622] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:15,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:21:16,323] loop.py:45 INFO TRAIN_STEP step=9113 loss=1.291811 Training: 51%|█████ | 9113/18018 [19:51:02<18:40:30, 7.55s/step, loss=1.292, acc=0.701, acc_len=1.68, thru=16.9, I=20.8, T=31.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:16,643] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:21:24,206] loop.py:45 INFO TRAIN_STEP step=9114 loss=1.328529 Training: 51%|█████ | 9114/18018 [19:51:10<18:55:10, 7.65s/step, loss=1.329, acc=0.698, acc_len=1.68, thru=13.7, I=17.1, T=28.3, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:33,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:21:33,881] loop.py:380 INFO TIMING step=9115: step=4.097s data=1.743s compute=4.050s [fwd=2.849s bwd=1.186s opt=0.003s] dispatch=5.567s [2026-10-19 09:21:33,881] loop.py:45 INFO TRAIN_STEP step=9115 loss=1.381850 Training: 51%|█████ | 9115/18018 [19:51:19<20:25:10, 8.26s/step, loss=1.382, acc=0.688, acc_len=1.65, thru=16.9, I=14.9, T=31.2, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:34,255] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:21:40,066] loop.py:45 INFO TRAIN_STEP step=9116 loss=1.331536 Training: 51%|█████ | 9116/18018 [19:51:26<18:52:54, 7.64s/step, loss=1.332, acc=0.693, acc_len=1.65, thru=17.3, I=20.6, T=32.5, wait=2.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:21:40,062] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9115: forward=2726.9ms backward=1136.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:48,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:21:49,695] loop.py:45 INFO TRAIN_STEP step=9117 loss=1.373314 Training: 51%|█████ | 9117/18018 [19:51:35<20:21:23, 8.23s/step, loss=1.373, acc=0.691, acc_len=1.65, thru=15.2, I=15.2, T=28.4, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:49,826] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:55,293] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:21:55,863] loop.py:45 INFO TRAIN_STEP step=9118 loss=1.275026 Training: 51%|█████ | 9118/18018 [19:51:41<18:49:22, 7.61s/step, loss=1.275, acc=0.706, acc_len=1.69, thru=18.8, I=21.2, T=30.6, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:21:56,300] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:03,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:22:05,956] loop.py:45 INFO TRAIN_STEP step=9119 loss=1.335083 Training: 51%|█████ | 9119/18018 [19:51:52<20:39:32, 8.36s/step, loss=1.335, acc=0.694, acc_len=1.66, thru=13.3, I=15.5, T=28.7, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:06,278] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:13,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:22:14,503] loop.py:380 INFO TIMING step=9120: step=4.768s data=2.479s compute=4.734s [fwd=3.455s bwd=1.264s opt=0.003s] dispatch=3.770s [2026-10-19 09:22:14,503] loop.py:45 INFO TRAIN_STEP step=9120 loss=1.373232 Training: 51%|█████ | 9120/18018 [19:52:00<20:47:47, 8.41s/step, loss=1.373, acc=0.687, acc_len=1.64, thru=15.6, I=16.8, T=26.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:14,943] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:20,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:22:22,784] loop.py:45 INFO TRAIN_STEP step=9121 loss=1.541938 Training: 51%|█████ | 9121/18018 [19:52:08<20:41:45, 8.37s/step, loss=1.542, acc=0.660, acc_len=1.54, thru=13.8, I=19.5, T=22.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:22,815] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 09:22:22,779] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9120: forward=4302.6ms backward=1312.6ms [2026-10-19 09:22:30,911] loop.py:45 INFO TRAIN_STEP step=9122 loss=1.255341 Training: 51%|█████ | 9122/18018 [19:52:16<20:30:43, 8.30s/step, loss=1.255, acc=0.706, acc_len=1.70, thru=12.9, I=14.3, T=29.4, wait=3.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:38,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:22:39,724] loop.py:45 INFO TRAIN_STEP step=9123 loss=1.486773 Training: 51%|█████ | 9123/18018 [19:52:25<20:53:17, 8.45s/step, loss=1.487, acc=0.668, acc_len=1.58, thru=17.1, I=16.6, T=26.3, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:40,040] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:45,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:22:47,461] loop.py:45 INFO TRAIN_STEP step=9124 loss=1.386781 Training: 51%|█████ | 9124/18018 [19:52:33<20:21:21, 8.24s/step, loss=1.387, acc=0.686, acc_len=1.63, thru=13.7, I=21.8, T=30.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:47,590] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:52,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:22:53,528] loop.py:380 INFO TIMING step=9125: step=3.752s data=1.695s compute=3.713s [fwd=2.626s bwd=1.072s opt=0.003s] dispatch=2.304s [2026-10-19 09:22:53,528] loop.py:45 INFO TRAIN_STEP step=9125 loss=1.294014 Training: 51%|█████ | 9125/18018 [19:52:39<18:44:36, 7.59s/step, loss=1.294, acc=0.705, acc_len=1.70, thru=17.6, I=26.4, T=34.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:53,659] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:22:59,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 09:23:00,504] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9125: forward=2693.6ms backward=1101.9ms [2026-10-19 09:23:00,509] loop.py:45 INFO TRAIN_STEP step=9126 loss=1.473537 Training: 51%|█████ | 9126/18018 [19:52:46<18:17:27, 7.41s/step, loss=1.474, acc=0.671, acc_len=1.58, thru=18.5, I=22.0, T=33.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:23:00,828] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:23:06,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:23:07,867] loop.py:45 INFO TRAIN_STEP step=9127 loss=1.400403 Training: 51%|█████ | 9127/18018 [19:52:53<18:15:17, 7.39s/step, loss=1.400, acc=0.685, acc_len=1.63, thru=16.8, I=21.8, T=26.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:23:07,937] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 09:23:17,427] loop.py:45 INFO TRAIN_STEP step=9128 loss=1.287557 Training: 51%|█████ | 9128/18018 [19:53:03<19:51:33, 8.04s/step, loss=1.288, acc=0.709, acc_len=1.71, thru=10.1, I=13.3, T=31.4, wait=5.5s, pool=33, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 09:23:17,509] eagle_store.py:517 WARNING batch_get_buffer missing keys (k292094-a1125805_hs, k292094-a1125805_ids, k292094-a1125805_lhs); sleeping 0.50s. [2026-10-19 09:23:26,616] loop.py:45 INFO TRAIN_STEP step=9129 loss=1.300773 Training: 51%|█████ | 9129/18018 [19:53:12<20:42:17, 8.39s/step, loss=1.301, acc=0.699, acc_len=1.68, thru=15.6, I=12.9, T=32.8, wait=5.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:23:32,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 09:23:17,499] eagle_store.py:517 WARNING batch_get_buffer missing keys (k292093-85e111d0_hs, k292093-85e111d0_ids, k292093-85e111d0_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 09:23:33,941] loop.py:380 INFO TIMING step=9130: step=4.126s data=1.864s compute=4.090s [fwd=2.846s bwd=1.229s opt=0.003s] dispatch=3.191s [2026-10-19 09:23:33,941] loop.py:45 INFO TRAIN_STEP step=9130 loss=1.269059 Training: 51%|█████ | 9130/18018 [19:53:20<19:55:07, 8.07s/step, loss=1.269, acc=0.705, acc_len=1.70, thru=17.6, I=18.9, T=31.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:23:34,305] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:23:40,326] loop.py:45 INFO TRAIN_STEP step=9131 loss=1.377385 Training: 51%|█████ | 9131/18018 [19:53:26<18:40:05, 7.56s/step, loss=1.377, acc=0.687, acc_len=1.64, thru=17.5, I=19.9, T=32.3, wait=2.4s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:23:40,322] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9130: forward=2794.0ms backward=1127.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:23:48,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:23:49,930] loop.py:45 INFO TRAIN_STEP step=9132 loss=1.460605 Training: 51%|█████ | 9132/18018 [19:53:36<20:10:49, 8.18s/step, loss=1.461, acc=0.677, acc_len=1.61, thru=15.3, I=16.8, T=27.5, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:23:50,132] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:23:57,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:23:58,306] loop.py:45 INFO TRAIN_STEP step=9133 loss=1.358229 Training: 51%|█████ | 9133/18018 [19:53:44<20:19:28, 8.24s/step, loss=1.358, acc=0.692, acc_len=1.65, thru=14.0, I=16.1, T=29.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:23:58,767] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:04,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:24:04,977] loop.py:45 INFO TRAIN_STEP step=9134 loss=1.252856 Training: 51%|█████ | 9134/18018 [19:53:51<19:09:54, 7.77s/step, loss=1.253, acc=0.706, acc_len=1.71, thru=18.2, I=19.2, T=31.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:05,125] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:10,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:24:10,810] loop.py:380 INFO TIMING step=9135: step=3.874s data=1.655s compute=3.835s [fwd=2.700s bwd=1.120s opt=0.003s] dispatch=1.950s [2026-10-19 09:24:10,810] loop.py:45 INFO TRAIN_STEP step=9135 loss=1.346969 Training: 51%|█████ | 9135/18018 [19:53:56<17:43:54, 7.19s/step, loss=1.347, acc=0.693, acc_len=1.66, thru=20.4, I=22.5, T=33.0, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:11,299] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:18,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:24:19,543] loop.py:45 INFO TRAIN_STEP step=9136 loss=1.432667 Training: 51%|█████ | 9136/18018 [19:54:05<18:52:28, 7.65s/step, loss=1.433, acc=0.680, acc_len=1.61, thru=14.8, I=18.1, T=31.1, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:24:19,538] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9135: forward=2842.9ms backward=1205.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:19,841] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:27,076] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:24:28,728] loop.py:45 INFO TRAIN_STEP step=9137 loss=1.270471 Training: 51%|█████ | 9137/18018 [19:54:14<20:00:34, 8.11s/step, loss=1.270, acc=0.705, acc_len=1.70, thru=14.4, I=17.0, T=30.3, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:29,088] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:34,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:24:36,245] loop.py:45 INFO TRAIN_STEP step=9138 loss=1.295038 Training: 51%|█████ | 9138/18018 [19:54:22<19:33:58, 7.93s/step, loss=1.295, acc=0.703, acc_len=1.70, thru=15.8, I=23.0, T=28.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:36,399] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:24:42,682] loop.py:45 INFO TRAIN_STEP step=9139 loss=1.341656 Training: 51%|█████ | 9139/18018 [19:54:28<18:27:25, 7.48s/step, loss=1.342, acc=0.693, acc_len=1.65, thru=17.3, I=20.6, T=33.1, wait=2.6s, pool=61, epoch=2/2][2026-10-19 09:24:52,185] loop.py:380 INFO TIMING step=9140: step=3.954s data=1.813s compute=3.910s [fwd=2.797s bwd=1.098s opt=0.003s] dispatch=5.542s [2026-10-19 09:24:52,185] loop.py:45 INFO TRAIN_STEP step=9140 loss=1.341926 Training: 51%|█████ | 9140/18018 [19:54:38<19:56:57, 8.09s/step, loss=1.342, acc=0.695, acc_len=1.67, thru=14.2, I=12.6, T=32.4, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:52,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:52,676] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:58,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:24:59,222] loop.py:45 INFO TRAIN_STEP step=9141 loss=1.431470 Training: 51%|█████ | 9141/18018 [19:54:45<19:10:09, 7.77s/step, loss=1.431, acc=0.682, acc_len=1.63, thru=18.6, I=19.3, T=28.9, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:24:59,218] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9140: forward=3086.8ms backward=1295.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:24:59,479] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:06,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:25:08,050] loop.py:45 INFO TRAIN_STEP step=9142 loss=1.308662 Training: 51%|█████ | 9142/18018 [19:54:54<19:56:54, 8.09s/step, loss=1.309, acc=0.700, acc_len=1.68, thru=13.9, I=17.4, T=27.7, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:08,195] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:13,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:25:14,324] loop.py:45 INFO TRAIN_STEP step=9143 loss=1.232930 Training: 51%|█████ | 9143/18018 [19:55:00<18:36:01, 7.54s/step, loss=1.233, acc=0.712, acc_len=1.72, thru=18.2, I=25.2, T=34.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:14,656] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:20,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:25:22,327] loop.py:45 INFO TRAIN_STEP step=9144 loss=1.385183 Training: 51%|█████ | 9144/18018 [19:55:08<18:56:13, 7.68s/step, loss=1.385, acc=0.686, acc_len=1.63, thru=14.7, I=19.1, T=32.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:22,734] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:25:28,864] loop.py:380 INFO TIMING step=9145: step=4.072s data=1.888s compute=4.041s [fwd=2.847s bwd=1.179s opt=0.003s] dispatch=2.455s [2026-10-19 09:25:28,864] loop.py:45 INFO TRAIN_STEP step=9145 loss=1.276608 Training: 51%|█████ | 9145/18018 [19:55:14<18:05:26, 7.34s/step, loss=1.277, acc=0.705, acc_len=1.70, thru=17.9, I=25.0, T=31.4, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:36,415] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:25:38,030] loop.py:45 INFO TRAIN_STEP step=9146 loss=1.410458 Training: 51%|█████ | 9146/18018 [19:55:24<19:26:11, 7.89s/step, loss=1.410, acc=0.684, acc_len=1.63, thru=14.5, I=15.6, T=29.8, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:25:38,026] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9145: forward=2965.2ms backward=1242.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:38,424] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:44,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:25:45,396] loop.py:45 INFO TRAIN_STEP step=9147 loss=1.283580 Training: 51%|█████ | 9147/18018 [19:55:31<19:03:03, 7.73s/step, loss=1.284, acc=0.702, acc_len=1.69, thru=15.5, I=22.6, T=33.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:45,603] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:50,794] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:25:52,717] loop.py:45 INFO TRAIN_STEP step=9148 loss=1.157035 Training: 51%|█████ | 9148/18018 [19:55:38<18:44:41, 7.61s/step, loss=1.157, acc=0.730, acc_len=1.80, thru=16.8, I=23.3, T=30.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:52,833] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:56,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:25:58,206] loop.py:45 INFO TRAIN_STEP step=9149 loss=1.294064 Training: 51%|█████ | 9149/18018 [19:55:44<17:10:37, 6.97s/step, loss=1.294, acc=0.702, acc_len=1.69, thru=20.2, I=28.4, T=38.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:25:58,419] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:03,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:26:03,979] loop.py:380 INFO TIMING step=9150: step=3.807s data=1.676s compute=3.774s [fwd=2.629s bwd=1.130s opt=0.003s] dispatch=1.956s [2026-10-19 09:26:03,979] loop.py:45 INFO TRAIN_STEP step=9150 loss=1.300784 Training: 51%|█████ | 9150/18018 [19:55:50<16:17:19, 6.61s/step, loss=1.301, acc=0.705, acc_len=1.70, thru=22.2, I=26.4, T=33.6, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:04,287] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:09,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:26:10,838] loop.py:45 INFO TRAIN_STEP step=9151 loss=1.176690 Training: 51%|█████ | 9151/18018 [19:55:56<16:28:04, 6.69s/step, loss=1.177, acc=0.726, acc_len=1.77, thru=18.7, I=21.4, T=34.1, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:26:10,833] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9150: forward=2624.4ms backward=1071.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:10,989] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:17,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:26:18,417] loop.py:45 INFO TRAIN_STEP step=9152 loss=1.404274 Training: 51%|█████ | 9152/18018 [19:56:04<17:07:34, 6.95s/step, loss=1.404, acc=0.685, acc_len=1.63, thru=16.1, I=20.8, T=34.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:18,685] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:23,518] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:26:25,075] loop.py:45 INFO TRAIN_STEP step=9153 loss=1.381977 Training: 51%|█████ | 9153/18018 [19:56:11<16:54:22, 6.87s/step, loss=1.382, acc=0.688, acc_len=1.64, thru=19.2, I=23.1, T=31.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:25,530] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:26:31,710] loop.py:45 INFO TRAIN_STEP step=9154 loss=1.256937 Training: 51%|█████ | 9154/18018 [19:56:17<16:43:56, 6.80s/step, loss=1.257, acc=0.702, acc_len=1.68, thru=15.0, I=18.8, T=30.8, wait=2.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:37,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:26:38,697] loop.py:380 INFO TIMING step=9155: step=4.056s data=1.799s compute=4.016s [fwd=2.863s bwd=1.138s opt=0.003s] dispatch=2.923s [2026-10-19 09:26:38,697] loop.py:45 INFO TRAIN_STEP step=9155 loss=1.323386 Training: 51%|█████ | 9155/18018 [19:56:24<16:52:21, 6.85s/step, loss=1.323, acc=0.700, acc_len=1.69, thru=17.5, I=19.8, T=31.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:38,762] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:45,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:26:46,641] loop.py:45 INFO TRAIN_STEP step=9156 loss=1.346760 Training: 51%|█████ | 9156/18018 [19:56:32<17:40:33, 7.18s/step, loss=1.347, acc=0.696, acc_len=1.67, thru=15.6, I=20.9, T=32.2, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:26:46,637] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9155: forward=2858.9ms backward=1075.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:46,687] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:52,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:26:53,485] loop.py:45 INFO TRAIN_STEP step=9157 loss=1.362129 Training: 51%|█████ | 9157/18018 [19:56:39<17:25:36, 7.08s/step, loss=1.362, acc=0.687, acc_len=1.64, thru=17.8, I=20.6, T=37.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:53,598] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:58,306] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:26:59,136] loop.py:45 INFO TRAIN_STEP step=9158 loss=1.416969 Training: 51%|█████ | 9158/18018 [19:56:45<16:22:10, 6.65s/step, loss=1.417, acc=0.680, acc_len=1.61, thru=21.1, I=24.4, T=35.7, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:26:59,310] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:27:05,187] loop.py:45 INFO TRAIN_STEP step=9159 loss=1.302852 Training: 51%|█████ | 9159/18018 [19:56:51<15:55:23, 6.47s/step, loss=1.303, acc=0.700, acc_len=1.69, thru=17.4, I=22.4, T=35.6, wait=2.4s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:15,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:27:15,574] loop.py:380 INFO TIMING step=9160: step=4.193s data=1.900s compute=4.162s [fwd=2.972s bwd=1.175s opt=0.003s] dispatch=6.186s [2026-10-19 09:27:15,575] loop.py:45 INFO TRAIN_STEP step=9160 loss=1.216750 Training: 51%|█████ | 9160/18018 [19:57:01<18:48:48, 7.65s/step, loss=1.217, acc=0.714, acc_len=1.72, thru=15.7, I=13.6, T=30.5, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:15,868] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:23,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 09:27:23,568] loop.py:45 INFO TRAIN_STEP step=9161 loss=1.311130 Training: 51%|█████ | 9161/18018 [19:57:09<19:04:02, 7.75s/step, loss=1.311, acc=0.697, acc_len=1.67, thru=16.6, I=16.4, T=28.5, wait=3.5s, pool=69, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:27:23,564] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9160: forward=3074.9ms backward=1370.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:23,955] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:30,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:27:31,546] loop.py:45 INFO TRAIN_STEP step=9162 loss=1.354876 Training: 51%|█████ | 9162/18018 [19:57:17<19:13:57, 7.82s/step, loss=1.355, acc=0.692, acc_len=1.65, thru=16.9, I=20.2, T=30.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:32,042] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:27:38,620] loop.py:45 INFO TRAIN_STEP step=9163 loss=1.269563 Training: 51%|█████ | 9163/18018 [19:57:24<18:40:55, 7.60s/step, loss=1.270, acc=0.707, acc_len=1.71, thru=15.3, I=19.7, T=37.8, wait=3.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:47,746] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:27:48,297] loop.py:45 INFO TRAIN_STEP step=9164 loss=1.435223 Training: 51%|█████ | 9164/18018 [19:57:34<20:12:58, 8.22s/step, loss=1.435, acc=0.682, acc_len=1.61, thru=15.3, I=16.0, T=33.2, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:48,751] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:53,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:27:54,840] loop.py:380 INFO TIMING step=9165: step=3.695s data=1.647s compute=3.660s [fwd=2.590s bwd=1.056s opt=0.003s] dispatch=2.840s [2026-10-19 09:27:54,841] loop.py:45 INFO TRAIN_STEP step=9165 loss=1.364702 Training: 51%|█████ | 9165/18018 [19:57:40<18:58:40, 7.72s/step, loss=1.365, acc=0.691, acc_len=1.66, thru=19.1, I=22.5, T=34.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:27:54,894] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:28:00,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:28:02,235] loop.py:45 INFO TRAIN_STEP step=9166 loss=1.345829 Training: 51%|█████ | 9166/18018 [19:57:48<18:44:10, 7.62s/step, loss=1.346, acc=0.691, acc_len=1.65, thru=15.6, I=21.4, T=35.8, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:28:02,229] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9165: forward=2529.9ms backward=1003.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:28:02,404] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:28:09,005] loop.py:45 INFO TRAIN_STEP step=9167 loss=1.274060 Training: 51%|█████ | 9167/18018 [19:57:55<18:06:27, 7.36s/step, loss=1.274, acc=0.707, acc_len=1.71, thru=17.6, I=20.2, T=36.6, wait=3.3s, pool=49, epoch=2/2][2026-10-19 09:28:18,362] loop.py:45 INFO TRAIN_STEP step=9168 loss=1.321667 Training: 51%|█████ | 9168/18018 [19:58:04<19:34:33, 7.96s/step, loss=1.322, acc=0.697, acc_len=1.67, thru=13.6, I=13.2, T=33.5, wait=5.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:28:26,528] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:28:28,290] loop.py:45 INFO TRAIN_STEP step=9169 loss=1.358987 Training: 51%|█████ | 9169/18018 [19:58:14<21:01:14, 8.55s/step, loss=1.359, acc=0.691, acc_len=1.65, thru=15.1, I=17.1, T=30.5, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:28:28,534] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:28:36,025] loop.py:380 INFO TIMING step=9170: step=3.688s data=1.459s compute=3.654s [fwd=2.551s bwd=1.087s opt=0.003s] dispatch=4.040s [2026-10-19 09:28:36,025] loop.py:45 INFO TRAIN_STEP step=9170 loss=1.327716 Training: 51%|█████ | 9170/18018 [19:58:22<20:25:03, 8.31s/step, loss=1.328, acc=0.700, acc_len=1.69, thru=11.5, I=16.7, T=34.7, wait=4.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:28:45,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:28:48,575] loop.py:45 INFO TRAIN_STEP step=9171 loss=1.304599 Training: 51%|█████ | 9171/18018 [19:58:34<23:32:31, 9.58s/step, loss=1.305, acc=0.704, acc_len=1.70, thru=13.7, I=14.3, T=28.6, wait=8.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:28:48,569] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9170: forward=3149.3ms backward=1278.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:28:48,985] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:28:53,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:28:55,127] loop.py:45 INFO TRAIN_STEP step=9172 loss=1.306675 Training: 51%|█████ | 9172/18018 [19:58:41<21:18:26, 8.67s/step, loss=1.307, acc=0.695, acc_len=1.66, thru=15.6, I=26.4, T=35.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:28:55,487] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:29:02,494] loop.py:45 INFO TRAIN_STEP step=9173 loss=1.355376 Training: 51%|█████ | 9173/18018 [19:58:48<20:20:44, 8.28s/step, loss=1.355, acc=0.691, acc_len=1.65, thru=15.8, I=20.0, T=32.6, wait=3.4s, pool=53, epoch=2/2][2026-10-19 09:29:10,744] loop.py:45 INFO TRAIN_STEP step=9174 loss=1.241828 Training: 51%|█████ | 9174/18018 [19:58:56<20:19:06, 8.27s/step, loss=1.242, acc=0.713, acc_len=1.72, thru=15.1, I=14.9, T=32.3, wait=4.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:29:17,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:29:18,327] loop.py:380 INFO TIMING step=9175: step=4.114s data=1.852s compute=4.082s [fwd=2.834s bwd=1.232s opt=0.003s] dispatch=3.462s [2026-10-19 09:29:18,327] loop.py:45 INFO TRAIN_STEP step=9175 loss=1.346870 Training: 51%|█████ | 9175/18018 [19:59:04<19:48:35, 8.06s/step, loss=1.347, acc=0.691, acc_len=1.65, thru=15.9, I=17.6, T=31.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:29:18,723] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:29:23,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:29:24,889] loop.py:45 INFO TRAIN_STEP step=9176 loss=1.418383 Training: 51%|█████ | 9176/18018 [19:59:10<18:42:06, 7.61s/step, loss=1.418, acc=0.682, acc_len=1.61, thru=18.5, I=24.6, T=33.0, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:29:24,885] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9175: forward=2735.6ms backward=1098.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:29:25,215] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:29:30,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:29:31,749] loop.py:45 INFO TRAIN_STEP step=9177 loss=1.239526 Training: 51%|█████ | 9177/18018 [19:59:17<18:08:37, 7.39s/step, loss=1.240, acc=0.713, acc_len=1.73, thru=18.0, I=22.5, T=29.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:29:31,854] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:29:39,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:29:40,551] loop.py:45 INFO TRAIN_STEP step=9178 loss=1.329545 Training: 51%|█████ | 9178/18018 [19:59:26<19:10:56, 7.81s/step, loss=1.330, acc=0.701, acc_len=1.69, thru=13.6, I=15.9, T=25.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:29:41,024] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:29:47,518] loop.py:45 INFO TRAIN_STEP step=9179 loss=1.290483 Training: 51%|█████ | 9179/18018 [19:59:33<18:33:33, 7.56s/step, loss=1.290, acc=0.700, acc_len=1.69, thru=14.7, I=18.4, T=29.9, wait=2.7s, pool=61, epoch=2/2][2026-10-19 09:29:55,553] loop.py:380 INFO TIMING step=9180: step=4.241s data=1.840s compute=4.213s [fwd=2.879s bwd=1.319s opt=0.003s] dispatch=3.783s [2026-10-19 09:29:55,553] loop.py:45 INFO TRAIN_STEP step=9180 loss=1.389503 Training: 51%|█████ | 9180/18018 [19:59:41<18:54:22, 7.70s/step, loss=1.390, acc=0.688, acc_len=1.64, thru=16.0, I=16.9, T=30.2, wait=3.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:02,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:30:03,644] loop.py:45 INFO TRAIN_STEP step=9181 loss=1.288472 Training: 51%|█████ | 9181/18018 [19:59:49<19:11:27, 7.82s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=15.5, I=17.0, T=29.4, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:30:03,639] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9180: forward=3105.6ms backward=1194.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:03,675] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:09,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:30:11,014] loop.py:45 INFO TRAIN_STEP step=9182 loss=1.254622 Training: 51%|█████ | 9182/18018 [19:59:57<18:51:37, 7.68s/step, loss=1.255, acc=0.709, acc_len=1.71, thru=14.9, I=21.4, T=27.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:11,266] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:15,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:30:17,241] loop.py:45 INFO TRAIN_STEP step=9183 loss=1.322129 Training: 51%|█████ | 9183/18018 [20:00:03<17:46:59, 7.25s/step, loss=1.322, acc=0.697, acc_len=1.68, thru=17.8, I=26.0, T=33.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:17,560] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:30:23,130] loop.py:45 INFO TRAIN_STEP step=9184 loss=1.396468 Training: 51%|█████ | 9184/18018 [20:00:09<16:46:58, 6.84s/step, loss=1.396, acc=0.685, acc_len=1.63, thru=18.3, I=22.4, T=35.6, wait=2.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:29,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:30:30,450] loop.py:380 INFO TIMING step=9185: step=4.058s data=1.835s compute=4.025s [fwd=2.923s bwd=1.086s opt=0.003s] dispatch=3.253s [2026-10-19 09:30:30,451] loop.py:45 INFO TRAIN_STEP step=9185 loss=1.414428 Training: 51%|█████ | 9185/18018 [20:00:16<17:08:13, 6.98s/step, loss=1.414, acc=0.675, acc_len=1.60, thru=18.2, I=19.4, T=31.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:30,779] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:30:37,211] loop.py:45 INFO TRAIN_STEP step=9186 loss=1.381119 Training: 51%|█████ | 9186/18018 [20:00:23<16:58:08, 6.92s/step, loss=1.381, acc=0.690, acc_len=1.65, thru=12.5, I=21.9, T=33.1, wait=2.9s, pool=13, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:30:37,206] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9185: forward=2751.7ms backward=1062.6ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 09:30:37,344] eagle_store.py:517 WARNING batch_get_buffer missing keys (k293951-a2fe029d_hs, k293951-a2fe029d_ids, k293951-a2fe029d_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:43,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 09:30:37,612] eagle_store.py:517 WARNING batch_get_buffer missing keys (k293952-a2a9255b_hs, k293952-a2a9255b_ids, k293952-a2a9255b_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 09:30:45,546] loop.py:45 INFO TRAIN_STEP step=9187 loss=1.356423 Training: 51%|█████ | 9187/18018 [20:00:31<18:00:39, 7.34s/step, loss=1.356, acc=0.692, acc_len=1.66, thru=21.9, I=19.0, T=31.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:45,667] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:51,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:30:52,342] loop.py:45 INFO TRAIN_STEP step=9188 loss=1.428696 Training: 51%|█████ | 9188/18018 [20:00:38<17:36:30, 7.18s/step, loss=1.429, acc=0.679, acc_len=1.61, thru=17.0, I=25.2, T=33.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:52,439] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:57,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:30:58,976] loop.py:45 INFO TRAIN_STEP step=9189 loss=1.316693 Training: 51%|█████ | 9189/18018 [20:00:45<17:12:17, 7.02s/step, loss=1.317, acc=0.697, acc_len=1.66, thru=18.1, I=21.5, T=34.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:30:59,128] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 09:31:04,987] loop.py:380 INFO TIMING step=9190: step=3.922s data=1.783s compute=3.885s [fwd=2.760s bwd=1.110s opt=0.003s] dispatch=2.078s [2026-10-19 09:31:04,987] loop.py:45 INFO TRAIN_STEP step=9190 loss=1.403423 Training: 51%|█████ | 9190/18018 [20:00:51<16:27:47, 6.71s/step, loss=1.403, acc=0.684, acc_len=1.63, thru=18.3, I=20.4, T=32.6, wait=2.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:12,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:31:13,894] loop.py:45 INFO TRAIN_STEP step=9191 loss=1.378578 Training: 51%|█████ | 9191/18018 [20:00:59<18:04:34, 7.37s/step, loss=1.379, acc=0.687, acc_len=1.63, thru=17.3, I=16.7, T=29.9, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:31:13,891] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9190: forward=2985.2ms backward=1257.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:14,316] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:31:20,028] loop.py:45 INFO TRAIN_STEP step=9192 loss=1.245513 Training: 51%|█████ | 9192/18018 [20:01:06<17:09:51, 7.00s/step, loss=1.246, acc=0.711, acc_len=1.71, thru=18.5, I=23.0, T=35.5, wait=2.5s, pool=61, epoch=2/2][2026-10-19 09:31:28,338] loop.py:45 INFO TRAIN_STEP step=9193 loss=1.384830 Training: 51%|█████ | 9193/18018 [20:01:14<18:07:22, 7.39s/step, loss=1.385, acc=0.691, acc_len=1.64, thru=14.5, I=14.9, T=34.4, wait=4.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:34,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:31:36,750] loop.py:45 INFO TRAIN_STEP step=9194 loss=1.274812 Training: 51%|█████ | 9194/18018 [20:01:22<18:52:16, 7.70s/step, loss=1.275, acc=0.707, acc_len=1.71, thru=17.3, I=18.2, T=30.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:36,900] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:42,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:31:43,350] loop.py:380 INFO TIMING step=9195: step=3.929s data=1.594s compute=3.900s [fwd=2.666s bwd=1.220s opt=0.003s] dispatch=2.660s [2026-10-19 09:31:43,350] loop.py:45 INFO TRAIN_STEP step=9195 loss=1.272793 Training: 51%|█████ | 9195/18018 [20:01:29<18:03:35, 7.37s/step, loss=1.273, acc=0.706, acc_len=1.70, thru=16.9, I=20.6, T=32.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:43,811] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:50,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:31:50,652] loop.py:45 INFO TRAIN_STEP step=9196 loss=1.340524 Training: 51%|█████ | 9196/18018 [20:01:36<18:00:30, 7.35s/step, loss=1.341, acc=0.695, acc_len=1.67, thru=17.2, I=19.1, T=30.1, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:31:50,648] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9195: forward=2975.4ms backward=1205.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:51,045] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:58,492] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:31:59,093] loop.py:45 INFO TRAIN_STEP step=9197 loss=1.408550 Training: 51%|█████ | 9197/18018 [20:01:45<18:48:36, 7.68s/step, loss=1.409, acc=0.682, acc_len=1.62, thru=14.9, I=15.5, T=30.6, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:31:59,499] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:32:05,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:32:06,557] loop.py:45 INFO TRAIN_STEP step=9198 loss=1.205436 Training: 51%|█████ | 9198/18018 [20:01:52<18:39:00, 7.61s/step, loss=1.205, acc=0.717, acc_len=1.75, thru=18.1, I=20.8, T=27.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:32:06,895] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:32:14,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:32:15,240] loop.py:45 INFO TRAIN_STEP step=9199 loss=1.310147 Training: 51%|█████ | 9199/18018 [20:02:01<19:26:12, 7.93s/step, loss=1.310, acc=0.697, acc_len=1.67, thru=13.4, I=15.4, T=27.1, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:32:15,572] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:32:21,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:32:23,482] loop.py:380 INFO TIMING step=9200: step=4.024s data=1.764s compute=3.983s [fwd=2.862s bwd=1.107s opt=0.003s] dispatch=4.207s [2026-10-19 09:32:23,482] loop.py:45 INFO TRAIN_STEP step=9200 loss=1.454014 Training: 51%|█████ | 9200/18018 [20:02:09<19:39:38, 8.03s/step, loss=1.454, acc=0.673, acc_len=1.59, thru=15.0, I=18.6, T=31.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:32:23,714] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:32:31,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:32:32,747] loop.py:45 INFO TRAIN_STEP step=9201 loss=1.336232 Training: 51%|█████ | 9201/18018 [20:02:18<20:33:58, 8.40s/step, loss=1.336, acc=0.696, acc_len=1.67, thru=14.2, I=16.1, T=27.3, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:32:32,742] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9200: forward=3295.5ms backward=1346.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:32:33,034] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:32:41,629] loop.py:45 INFO TRAIN_STEP step=9202 loss=1.413942 Training: 51%|█████ | 9202/18018 [20:02:27<20:55:18, 8.54s/step, loss=1.414, acc=0.685, acc_len=1.63, thru=8.8, I=11.2, T=32.4, wait=4.9s, pool=13, epoch=2/2] (TrainerActor pid=9370) [2026-10-19 09:32:41,857] eagle_store.py:517 WARNING batch_get_buffer missing keys (k294466-bf346267_hs, k294466-bf346267_ids, k294466-bf346267_lhs); sleeping 0.50s. [2026-10-19 09:32:49,621] loop.py:45 INFO TRAIN_STEP step=9203 loss=1.406629 Training: 51%|█████ | 9203/18018 [20:02:35<20:30:54, 8.38s/step, loss=1.407, acc=0.683, acc_len=1.62, thru=17.8, I=12.9, T=36.8, wait=4.5s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:32:59,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 09:32:41,866] eagle_store.py:517 WARNING batch_get_buffer missing keys (k294467-8188a0ed_hs, k294467-8188a0ed_ids, k294467-8188a0ed_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 09:33:01,215] loop.py:45 INFO TRAIN_STEP step=9204 loss=1.384201 Training: 51%|█████ | 9204/18018 [20:02:47<22:52:26, 9.34s/step, loss=1.384, acc=0.686, acc_len=1.64, thru=14.8, I=14.7, T=31.1, wait=7.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:01,558] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:33:07,689] loop.py:380 INFO TIMING step=9205: step=4.125s data=1.873s compute=4.095s [fwd=2.855s bwd=1.227s opt=0.003s] dispatch=2.340s [2026-10-19 09:33:07,689] loop.py:45 INFO TRAIN_STEP step=9205 loss=1.392563 Training: 51%|█████ | 9205/18018 [20:02:53<20:45:44, 8.48s/step, loss=1.393, acc=0.689, acc_len=1.64, thru=18.4, I=21.7, T=31.0, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:07,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:08,187] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 09:33:14,579] loop.py:45 INFO TRAIN_STEP step=9206 loss=1.346812 Training: 51%|█████ | 9206/18018 [20:03:00<19:35:38, 8.00s/step, loss=1.347, acc=0.692, acc_len=1.65, thru=16.6, I=17.9, T=33.2, wait=3.0s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:33:14,575] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9205: forward=2706.2ms backward=1084.0ms [2026-10-19 09:33:21,901] loop.py:45 INFO TRAIN_STEP step=9207 loss=1.277228 Training: 51%|█████ | 9207/18018 [20:03:07<19:05:19, 7.80s/step, loss=1.277, acc=0.703, acc_len=1.69, thru=17.2, I=16.5, T=33.5, wait=3.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:30,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:33:32,175] loop.py:45 INFO TRAIN_STEP step=9208 loss=1.455514 Training: 51%|█████ | 9208/18018 [20:03:18<20:54:17, 8.54s/step, loss=1.456, acc=0.675, acc_len=1.59, thru=13.8, I=15.7, T=32.9, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:32,548] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:36,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:33:38,611] loop.py:45 INFO TRAIN_STEP step=9209 loss=1.269019 Training: 51%|█████ | 9209/18018 [20:03:24<19:21:19, 7.91s/step, loss=1.269, acc=0.708, acc_len=1.70, thru=18.6, I=28.4, T=30.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:38,913] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:46,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:33:48,082] loop.py:380 INFO TIMING step=9210: step=4.986s data=2.439s compute=4.933s [fwd=3.380s bwd=1.539s opt=0.003s] dispatch=4.476s [2026-10-19 09:33:48,083] loop.py:45 INFO TRAIN_STEP step=9210 loss=1.339312 Training: 51%|█████ | 9210/18018 [20:03:34<20:30:01, 8.38s/step, loss=1.339, acc=0.696, acc_len=1.67, thru=14.0, I=15.5, T=25.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:48,533] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:53,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:33:54,650] loop.py:45 INFO TRAIN_STEP step=9211 loss=1.315734 Training: 51%|█████ | 9211/18018 [20:03:40<19:10:05, 7.84s/step, loss=1.316, acc=0.696, acc_len=1.68, thru=16.8, I=22.9, T=31.3, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:33:54,645] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9210: forward=2874.1ms backward=1167.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:54,812] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:33:59,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:34:01,015] loop.py:45 INFO TRAIN_STEP step=9212 loss=1.254258 Training: 51%|█████ | 9212/18018 [20:03:47<18:05:17, 7.39s/step, loss=1.254, acc=0.712, acc_len=1.73, thru=18.5, I=24.5, T=35.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:01,096] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:06,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:34:08,360] loop.py:45 INFO TRAIN_STEP step=9213 loss=1.413805 Training: 51%|█████ | 9213/18018 [20:03:54<18:02:54, 7.38s/step, loss=1.414, acc=0.684, acc_len=1.62, thru=16.5, I=24.9, T=28.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:08,793] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:14,928] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:34:16,067] loop.py:45 INFO TRAIN_STEP step=9214 loss=1.396206 Training: 51%|█████ | 9214/18018 [20:04:02<18:17:10, 7.48s/step, loss=1.396, acc=0.685, acc_len=1.63, thru=14.0, I=21.0, T=27.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:16,437] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:24,492] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:34:25,658] loop.py:380 INFO TIMING step=9215: step=4.051s data=1.786s compute=4.014s [fwd=2.878s bwd=1.121s opt=0.003s] dispatch=5.531s [2026-10-19 09:34:25,659] loop.py:45 INFO TRAIN_STEP step=9215 loss=1.279057 Training: 51%|█████ | 9215/18018 [20:04:11<19:50:11, 8.11s/step, loss=1.279, acc=0.706, acc_len=1.70, thru=13.9, I=15.4, T=31.6, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:26,004] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:31,339] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:34:32,492] loop.py:45 INFO TRAIN_STEP step=9216 loss=1.346793 Training: 51%|█████ | 9216/18018 [20:04:18<18:53:46, 7.73s/step, loss=1.347, acc=0.695, acc_len=1.67, thru=16.5, I=22.1, T=30.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:34:32,486] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9215: forward=2942.1ms backward=1140.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:32,848] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:38,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:34:39,241] loop.py:45 INFO TRAIN_STEP step=9217 loss=1.303973 Training: 51%|█████ | 9217/18018 [20:04:25<18:10:32, 7.43s/step, loss=1.304, acc=0.700, acc_len=1.68, thru=18.4, I=23.0, T=32.7, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:34:39,425] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:34:45,406] loop.py:45 INFO TRAIN_STEP step=9218 loss=1.475394 Training: 51%|█████ | 9218/18018 [20:04:31<17:14:38, 7.05s/step, loss=1.475, acc=0.673, acc_len=1.59, thru=19.0, I=24.0, T=32.8, wait=2.3s, pool=61, epoch=2/2][2026-10-19 09:34:55,216] loop.py:45 INFO TRAIN_STEP step=9219 loss=1.356712 Training: 51%|█████ | 9219/18018 [20:04:41<19:15:35, 7.88s/step, loss=1.357, acc=0.690, acc_len=1.65, thru=13.3, I=12.8, T=29.4, wait=5.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:04,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:35:05,508] loop.py:380 INFO TIMING step=9220: step=4.374s data=2.009s compute=4.335s [fwd=3.037s bwd=1.283s opt=0.003s] dispatch=5.909s [2026-10-19 09:35:05,508] loop.py:45 INFO TRAIN_STEP step=9220 loss=1.349494 Training: 51%|█████ | 9220/18018 [20:04:51<21:01:35, 8.60s/step, loss=1.349, acc=0.695, acc_len=1.66, thru=14.7, I=13.2, T=29.3, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:05,570] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:13,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:35:15,185] loop.py:45 INFO TRAIN_STEP step=9221 loss=1.447688 Training: 51%|█████ | 9221/18018 [20:05:01<21:48:44, 8.93s/step, loss=1.448, acc=0.678, acc_len=1.61, thru=13.3, I=15.0, T=29.5, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:35:15,181] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9220: forward=3015.6ms backward=1276.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:15,391] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:20,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:35:21,809] loop.py:45 INFO TRAIN_STEP step=9222 loss=1.254211 Training: 51%|█████ | 9222/18018 [20:05:07<20:07:13, 8.23s/step, loss=1.254, acc=0.709, acc_len=1.72, thru=17.7, I=21.8, T=35.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:21,976] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:27,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:35:29,117] loop.py:45 INFO TRAIN_STEP step=9223 loss=1.345076 Training: 51%|█████ | 9223/18018 [20:05:15<19:26:25, 7.96s/step, loss=1.345, acc=0.693, acc_len=1.66, thru=17.3, I=23.3, T=32.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:29,364] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:35:35,284] loop.py:45 INFO TRAIN_STEP step=9224 loss=1.360857 Training: 51%|█████ | 9224/18018 [20:05:21<18:07:27, 7.42s/step, loss=1.361, acc=0.693, acc_len=1.65, thru=16.4, I=22.3, T=37.1, wait=2.7s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:41,587] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:35:43,121] loop.py:380 INFO TIMING step=9225: step=4.018s data=1.644s compute=3.970s [fwd=2.782s bwd=1.172s opt=0.003s] dispatch=3.810s [2026-10-19 09:35:43,121] loop.py:45 INFO TRAIN_STEP step=9225 loss=1.361783 Training: 51%|█████ | 9225/18018 [20:05:29<18:25:43, 7.55s/step, loss=1.362, acc=0.689, acc_len=1.64, thru=17.2, I=19.8, T=31.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:43,602] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:35:49,498] loop.py:45 INFO TRAIN_STEP step=9226 loss=1.361690 Training: 51%|█████ | 9226/18018 [20:05:35<17:34:17, 7.19s/step, loss=1.362, acc=0.689, acc_len=1.64, thru=18.6, I=24.0, T=35.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:49,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 09:35:49,492] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9225: forward=2558.3ms backward=959.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:49,949] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:54,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:35:56,509] loop.py:45 INFO TRAIN_STEP step=9227 loss=1.298476 Training: 51%|█████ | 9227/18018 [20:05:42<17:26:05, 7.14s/step, loss=1.298, acc=0.703, acc_len=1.70, thru=17.7, I=21.7, T=29.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:35:56,836] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:36:03,481] loop.py:45 INFO TRAIN_STEP step=9228 loss=1.476610 Training: 51%|█████ | 9228/18018 [20:05:49<17:18:31, 7.09s/step, loss=1.477, acc=0.667, acc_len=1.57, thru=14.6, I=17.9, T=32.5, wait=3.0s, pool=45, epoch=2/2][2026-10-19 09:36:11,669] loop.py:45 INFO TRAIN_STEP step=9229 loss=1.310363 Training: 51%|█████ | 9229/18018 [20:05:57<18:06:46, 7.42s/step, loss=1.310, acc=0.698, acc_len=1.68, thru=15.3, I=13.8, T=32.1, wait=4.2s, pool=45, epoch=2/2][2026-10-19 09:36:24,259] loop.py:380 INFO TIMING step=9230: step=4.270s data=1.859s compute=4.130s [fwd=2.759s bwd=1.356s opt=0.003s] dispatch=8.311s [2026-10-19 09:36:24,259] loop.py:45 INFO TRAIN_STEP step=9230 loss=1.341518 Training: 51%|█████ | 9230/18018 [20:06:10<21:53:53, 8.97s/step, loss=1.342, acc=0.695, acc_len=1.66, thru=12.6, I=12.6, T=30.0, wait=8.3s, pool=53, epoch=2/2][2026-10-19 09:36:32,220] loop.py:45 INFO TRAIN_STEP step=9231 loss=1.327991 (TrainerActor pid=9310) [2026-10-19 09:36:32,216] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9230: forward=2892.6ms backward=1085.6ms Training: 51%|█████ | 9231/18018 [20:06:18<21:09:22, 8.67s/step, loss=1.328, acc=0.695, acc_len=1.67, thru=14.1, I=16.1, T=31.7, wait=3.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:36:40,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:36:40,787] loop.py:45 INFO TRAIN_STEP step=9232 loss=1.451753 Training: 51%|█████ | 9232/18018 [20:06:26<21:04:49, 8.64s/step, loss=1.452, acc=0.675, acc_len=1.60, thru=17.3, I=16.0, T=28.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:36:41,033] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:36:47,484] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:36:48,588] loop.py:45 INFO TRAIN_STEP step=9233 loss=1.433072 Training: 51%|█████ | 9233/18018 [20:06:34<20:27:57, 8.39s/step, loss=1.433, acc=0.679, acc_len=1.61, thru=16.5, I=19.8, T=30.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:36:48,993] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:36:56,510] loop.py:45 INFO TRAIN_STEP step=9234 loss=1.325889 Training: 51%|█████ | 9234/18018 [20:06:42<20:07:15, 8.25s/step, loss=1.326, acc=0.695, acc_len=1.66, thru=15.0, I=17.1, T=29.7, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:01,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:37:02,609] loop.py:380 INFO TIMING step=9235: step=3.983s data=1.732s compute=3.954s [fwd=2.931s bwd=1.007s opt=0.003s] dispatch=2.108s [2026-10-19 09:37:02,609] loop.py:45 INFO TRAIN_STEP step=9235 loss=1.297844 Training: 51%|█████▏ | 9235/18018 [20:06:48<18:32:57, 7.60s/step, loss=1.298, acc=0.699, acc_len=1.68, thru=19.3, I=18.8, T=32.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:02,645] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:10,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:37:11,195] loop.py:45 INFO TRAIN_STEP step=9236 loss=1.328419 Training: 51%|█████▏ | 9236/18018 [20:06:57<19:15:58, 7.90s/step, loss=1.328, acc=0.692, acc_len=1.66, thru=14.0, I=16.1, T=33.7, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:37:11,191] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9235: forward=2686.9ms backward=1068.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:11,647] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:20,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:37:21,221] loop.py:45 INFO TRAIN_STEP step=9237 loss=1.358343 Training: 51%|█████▏ | 9237/18018 [20:07:07<20:49:09, 8.54s/step, loss=1.358, acc=0.691, acc_len=1.66, thru=13.4, I=14.3, T=26.6, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:21,532] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:31,885] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:37:32,023] loop.py:45 INFO TRAIN_STEP step=9238 loss=1.430849 Training: 51%|█████▏ | 9238/18018 [20:07:18<22:28:31, 9.22s/step, loss=1.431, acc=0.681, acc_len=1.62, thru=11.7, I=12.3, T=26.8, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:32,389] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 09:37:39,734] loop.py:45 INFO TRAIN_STEP step=9239 loss=1.396050 Training: 51%|█████▏ | 9239/18018 [20:07:25<21:22:19, 8.76s/step, loss=1.396, acc=0.684, acc_len=1.63, thru=16.5, I=16.3, T=28.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:39,690] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:40,193] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 09:37:46,785] loop.py:380 INFO TIMING step=9240: step=4.423s data=1.802s compute=4.390s [fwd=3.131s bwd=1.243s opt=0.003s] dispatch=2.620s [2026-10-19 09:37:46,785] loop.py:45 INFO TRAIN_STEP step=9240 loss=1.444858 Training: 51%|█████▏ | 9240/18018 [20:07:32<20:07:01, 8.25s/step, loss=1.445, acc=0.679, acc_len=1.61, thru=16.5, I=16.3, T=28.9, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:56,665] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:37:56,959] loop.py:45 INFO TRAIN_STEP step=9241 loss=1.252574 Training: 51%|█████▏ | 9241/18018 [20:07:43<21:31:20, 8.83s/step, loss=1.253, acc=0.713, acc_len=1.73, thru=13.3, I=14.0, T=29.5, wait=5.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:37:56,955] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9240: forward=3088.9ms backward=1200.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:37:57,168] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 09:38:03,760] loop.py:45 INFO TRAIN_STEP step=9242 loss=1.354068 Training: 51%|█████▏ | 9242/18018 [20:07:49<20:02:19, 8.22s/step, loss=1.354, acc=0.691, acc_len=1.65, thru=16.7, I=18.5, T=34.0, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:12,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:38:13,026] loop.py:45 INFO TRAIN_STEP step=9243 loss=1.556133 Training: 51%|█████▏ | 9243/18018 [20:07:59<20:47:57, 8.53s/step, loss=1.556, acc=0.658, acc_len=1.54, thru=15.3, I=14.0, T=31.0, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:13,100] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:19,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:38:20,688] loop.py:45 INFO TRAIN_STEP step=9244 loss=1.342421 Training: 51%|█████▏ | 9244/18018 [20:08:06<20:09:43, 8.27s/step, loss=1.342, acc=0.695, acc_len=1.67, thru=16.8, I=18.6, T=32.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:21,049] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:38:29,957] loop.py:380 INFO TIMING step=9245: step=4.030s data=1.747s compute=3.986s [fwd=2.882s bwd=1.089s opt=0.003s] dispatch=5.228s [2026-10-19 09:38:29,957] loop.py:45 INFO TRAIN_STEP step=9245 loss=1.326660 Training: 51%|█████▏ | 9245/18018 [20:08:16<20:53:10, 8.57s/step, loss=1.327, acc=0.699, acc_len=1.68, thru=13.5, I=12.9, T=31.8, wait=5.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:37,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:38:38,962] loop.py:45 INFO TRAIN_STEP step=9246 loss=1.407042 Training: 51%|█████▏ | 9246/18018 [20:08:25<21:12:09, 8.70s/step, loss=1.407, acc=0.686, acc_len=1.63, thru=15.3, I=15.8, T=29.5, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:38:38,958] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9245: forward=2968.9ms backward=1309.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:39,396] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:45,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:38:45,797] loop.py:45 INFO TRAIN_STEP step=9247 loss=1.456525 Training: 51%|█████▏ | 9247/18018 [20:08:31<19:50:07, 8.14s/step, loss=1.457, acc=0.672, acc_len=1.59, thru=17.7, I=19.9, T=33.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:45,967] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:52,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:38:53,243] loop.py:45 INFO TRAIN_STEP step=9248 loss=1.294234 Training: 51%|█████▏ | 9248/18018 [20:08:39<19:19:29, 7.93s/step, loss=1.294, acc=0.703, acc_len=1.69, thru=17.2, I=19.1, T=28.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:38:53,662] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:38:59,703] loop.py:45 INFO TRAIN_STEP step=9249 loss=1.331753 Training: 51%|█████▏ | 9249/18018 [20:08:45<18:14:52, 7.49s/step, loss=1.332, acc=0.695, acc_len=1.66, thru=16.8, I=23.7, T=36.5, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:06,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:39:08,583] loop.py:380 INFO TIMING step=9250: step=5.295s data=2.914s compute=5.257s [fwd=3.877s bwd=1.364s opt=0.003s] dispatch=3.574s [2026-10-19 09:39:08,583] loop.py:45 INFO TRAIN_STEP step=9250 loss=1.247647 Training: 51%|█████▏ | 9250/18018 [20:08:54<19:15:37, 7.91s/step, loss=1.248, acc=0.707, acc_len=1.71, thru=14.4, I=17.5, T=24.2, wait=3.6s, pool=65, epoch=2/2][2026-10-19 09:39:08,588] loop.py:419 INFO Saving checkpoint at step 9250... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:11,225] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 09:39:12,084] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009251 (TrainerActor pid=9310) [2026-10-19 09:39:12,088] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0008751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:13,263] inference_manager.py:408 INFO Pool capacity available after 7.1s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:17,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 09:39:12,082] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 09:39:13,230] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0008751 [2026-10-19 09:39:19,241] loop.py:45 INFO TRAIN_STEP step=9251 loss=1.346420 Training: 51%|█████▏ | 9251/18018 [20:09:05<21:16:01, 8.73s/step, loss=1.346, acc=0.691, acc_len=1.65, thru=21.5, I=27.8, T=31.2, wait=1.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:39:19,236] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9250: forward=2873.3ms backward=1179.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:19,491] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:39:27,214] loop.py:45 INFO TRAIN_STEP step=9252 loss=1.291512 Training: 51%|█████▏ | 9252/18018 [20:09:13<20:42:39, 8.51s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=15.0, I=24.1, T=29.6, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:34,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:39:35,347] loop.py:45 INFO TRAIN_STEP step=9253 loss=1.448276 Training: 51%|█████▏ | 9253/18018 [20:09:21<20:26:01, 8.39s/step, loss=1.448, acc=0.679, acc_len=1.61, thru=16.9, I=14.2, T=31.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:35,610] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:40,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:39:41,823] loop.py:45 INFO TRAIN_STEP step=9254 loss=1.296066 Training: 51%|█████▏ | 9254/18018 [20:09:27<19:01:59, 7.82s/step, loss=1.296, acc=0.701, acc_len=1.69, thru=18.5, I=23.0, T=33.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:42,235] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:39:47,793] loop.py:380 INFO TIMING step=9255: step=3.401s data=1.423s compute=3.367s [fwd=2.420s bwd=0.931s opt=0.003s] dispatch=2.559s [2026-10-19 09:39:47,793] loop.py:45 INFO TRAIN_STEP step=9255 loss=1.306162 Training: 51%|█████▏ | 9255/18018 [20:09:33<17:40:48, 7.26s/step, loss=1.306, acc=0.703, acc_len=1.69, thru=19.1, I=24.5, T=37.6, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:56,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:39:57,250] loop.py:45 INFO TRAIN_STEP step=9256 loss=1.341609 Training: 51%|█████▏ | 9256/18018 [20:09:43<19:16:47, 7.92s/step, loss=1.342, acc=0.696, acc_len=1.67, thru=14.5, I=15.2, T=33.9, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:39:57,246] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9255: forward=2687.9ms backward=1032.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:39:57,516] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:03,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:40:03,432] loop.py:45 INFO TRAIN_STEP step=9257 loss=1.335395 Training: 51%|█████▏ | 9257/18018 [20:09:49<18:00:27, 7.40s/step, loss=1.335, acc=0.694, acc_len=1.67, thru=19.6, I=20.9, T=36.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:03,765] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:09,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:40:10,231] loop.py:45 INFO TRAIN_STEP step=9258 loss=1.175872 Training: 51%|█████▏ | 9258/18018 [20:09:56<17:34:03, 7.22s/step, loss=1.176, acc=0.722, acc_len=1.76, thru=19.2, I=21.4, T=32.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:10,314] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 09:40:17,469] loop.py:45 INFO TRAIN_STEP step=9259 loss=1.315414 Training: 51%|█████▏ | 9259/18018 [20:10:03<17:34:39, 7.22s/step, loss=1.315, acc=0.699, acc_len=1.69, thru=14.4, I=14.6, T=34.5, wait=3.5s, pool=41, epoch=2/2][2026-10-19 09:40:25,347] loop.py:380 INFO TIMING step=9260: step=3.855s data=1.551s compute=3.824s [fwd=2.740s bwd=1.070s opt=0.003s] dispatch=4.016s [2026-10-19 09:40:25,348] loop.py:45 INFO TRAIN_STEP step=9260 loss=1.354205 Training: 51%|█████▏ | 9260/18018 [20:10:11<18:03:19, 7.42s/step, loss=1.354, acc=0.689, acc_len=1.65, thru=15.6, I=14.1, T=33.2, wait=4.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:35,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:40:36,876] loop.py:45 INFO TRAIN_STEP step=9261 loss=1.451349 Training: 51%|█████▏ | 9261/18018 [20:10:22<21:02:57, 8.65s/step, loss=1.451, acc=0.675, acc_len=1.60, thru=13.6, I=13.9, T=27.7, wait=6.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:40:36,872] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9260: forward=3319.1ms backward=1262.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:37,372] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:43,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:40:44,710] loop.py:45 INFO TRAIN_STEP step=9262 loss=1.289054 Training: 51%|█████▏ | 9262/18018 [20:10:30<20:27:04, 8.41s/step, loss=1.289, acc=0.704, acc_len=1.70, thru=14.6, I=19.9, T=30.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:45,103] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:50,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:40:52,169] loop.py:45 INFO TRAIN_STEP step=9263 loss=1.254884 Training: 51%|█████▏ | 9263/18018 [20:10:38<19:45:10, 8.12s/step, loss=1.255, acc=0.708, acc_len=1.71, thru=15.3, I=21.7, T=29.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:52,318] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:57,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:40:58,792] loop.py:45 INFO TRAIN_STEP step=9264 loss=1.305998 Training: 51%|█████▏ | 9264/18018 [20:10:44<18:39:32, 7.67s/step, loss=1.306, acc=0.700, acc_len=1.68, thru=17.6, I=23.7, T=32.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:40:58,963] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:41:06,777] loop.py:380 INFO TIMING step=9265: step=4.221s data=1.987s compute=4.180s [fwd=2.926s bwd=1.239s opt=0.003s] dispatch=3.754s [2026-10-19 09:41:06,777] loop.py:45 INFO TRAIN_STEP step=9265 loss=1.439254 Training: 51%|█████▏ | 9265/18018 [20:10:52<18:52:57, 7.77s/step, loss=1.439, acc=0.680, acc_len=1.61, thru=15.0, I=17.5, T=30.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:06,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:07,221] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:14,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:41:15,224] loop.py:45 INFO TRAIN_STEP step=9266 loss=1.280511 Training: 51%|█████▏ | 9266/18018 [20:11:01<19:22:37, 7.97s/step, loss=1.281, acc=0.709, acc_len=1.72, thru=15.3, I=16.5, T=32.1, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:41:15,219] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9265: forward=2843.3ms backward=1098.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:15,449] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:21,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:41:21,709] loop.py:45 INFO TRAIN_STEP step=9267 loss=1.362506 Training: 51%|█████▏ | 9267/18018 [20:11:07<18:17:33, 7.53s/step, loss=1.363, acc=0.692, acc_len=1.66, thru=18.6, I=19.2, T=32.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:21,978] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:27,047] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:41:28,361] loop.py:45 INFO TRAIN_STEP step=9268 loss=1.252159 Training: 51%|█████▏ | 9268/18018 [20:11:14<17:39:08, 7.26s/step, loss=1.252, acc=0.711, acc_len=1.72, thru=18.2, I=21.2, T=31.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:28,554] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:33,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:41:34,742] loop.py:45 INFO TRAIN_STEP step=9269 loss=1.310091 Training: 51%|█████▏ | 9269/18018 [20:11:20<17:00:33, 7.00s/step, loss=1.310, acc=0.699, acc_len=1.69, thru=18.4, I=22.2, T=32.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:35,168] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:40,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:41:41,918] loop.py:380 INFO TIMING step=9270: step=4.954s data=2.452s compute=4.928s [fwd=3.498s bwd=1.415s opt=0.003s] dispatch=2.211s [2026-10-19 09:41:41,918] loop.py:45 INFO TRAIN_STEP step=9270 loss=1.379802 Training: 51%|█████▏ | 9270/18018 [20:11:28<17:08:09, 7.05s/step, loss=1.380, acc=0.690, acc_len=1.65, thru=17.8, I=23.0, T=25.8, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:42,064] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:45,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:41:47,658] loop.py:45 INFO TRAIN_STEP step=9271 loss=1.358543 Training: 51%|█████▏ | 9271/18018 [20:11:33<16:10:42, 6.66s/step, loss=1.359, acc=0.692, acc_len=1.66, thru=17.6, I=23.1, T=33.8, wait=1.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:41:47,654] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9270: forward=2559.0ms backward=1117.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:47,975] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:53,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:41:54,139] loop.py:45 INFO TRAIN_STEP step=9272 loss=1.273712 Training: 51%|█████▏ | 9272/18018 [20:11:40<16:02:44, 6.60s/step, loss=1.274, acc=0.705, acc_len=1.70, thru=19.2, I=26.7, T=35.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:41:54,555] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:00,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:42:00,877] loop.py:45 INFO TRAIN_STEP step=9273 loss=1.305860 Training: 51%|█████▏ | 9273/18018 [20:11:46<16:08:27, 6.64s/step, loss=1.306, acc=0.699, acc_len=1.68, thru=18.8, I=24.5, T=34.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:01,377] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:42:07,802] loop.py:45 INFO TRAIN_STEP step=9274 loss=1.381744 Training: 51%|█████▏ | 9274/18018 [20:11:53<16:20:38, 6.73s/step, loss=1.382, acc=0.688, acc_len=1.64, thru=17.7, I=20.6, T=32.5, wait=3.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:17,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:42:18,732] loop.py:380 INFO TIMING step=9275: step=4.018s data=1.708s compute=3.990s [fwd=2.853s bwd=1.121s opt=0.003s] dispatch=6.902s [2026-10-19 09:42:18,732] loop.py:45 INFO TRAIN_STEP step=9275 loss=1.335772 Training: 51%|█████▏ | 9275/18018 [20:12:04<19:24:07, 7.99s/step, loss=1.336, acc=0.694, acc_len=1.65, thru=12.6, I=12.5, T=31.9, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:19,064] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-19 09:42:26,442] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9275: forward=2639.0ms backward=1016.8ms [2026-10-19 09:42:26,447] loop.py:45 INFO TRAIN_STEP step=9276 loss=1.335930 Training: 51%|█████▏ | 9276/18018 [20:12:12<19:12:02, 7.91s/step, loss=1.336, acc=0.694, acc_len=1.66, thru=15.2, I=17.5, T=34.5, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:32,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:42:34,024] loop.py:45 INFO TRAIN_STEP step=9277 loss=1.354263 Training: 51%|█████▏ | 9277/18018 [20:12:20<18:57:34, 7.81s/step, loss=1.354, acc=0.689, acc_len=1.65, thru=17.7, I=18.6, T=33.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:34,445] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:40,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:42:41,556] loop.py:45 INFO TRAIN_STEP step=9278 loss=1.360572 Training: 51%|█████▏ | 9278/18018 [20:12:27<18:45:13, 7.72s/step, loss=1.361, acc=0.687, acc_len=1.64, thru=15.4, I=22.3, T=30.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:42,056] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:47,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:42:48,751] loop.py:45 INFO TRAIN_STEP step=9279 loss=1.311629 Training: 51%|█████▏ | 9279/18018 [20:12:34<18:22:08, 7.57s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=15.6, I=21.6, T=27.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:49,112] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:42:57,415] loop.py:380 INFO TIMING step=9280: step=3.793s data=1.571s compute=3.761s [fwd=2.715s bwd=1.030s opt=0.003s] dispatch=4.859s [2026-10-19 09:42:57,415] loop.py:45 INFO TRAIN_STEP step=9280 loss=1.406366 Training: 52%|█████▏ | 9280/18018 [20:12:43<19:09:45, 7.89s/step, loss=1.406, acc=0.683, acc_len=1.62, thru=13.8, I=15.6, T=33.8, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:57,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:42:57,881] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:02,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:43:03,852] loop.py:45 INFO TRAIN_STEP step=9281 loss=1.336693 Training: 52%|█████▏ | 9281/18018 [20:12:49<18:05:58, 7.46s/step, loss=1.337, acc=0.692, acc_len=1.65, thru=19.2, I=20.7, T=34.6, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:43:03,848] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9280: forward=2619.5ms backward=1035.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:03,948] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 09:43:11,351] loop.py:45 INFO TRAIN_STEP step=9282 loss=1.370250 Training: 52%|█████▏ | 9282/18018 [20:12:57<18:07:47, 7.47s/step, loss=1.370, acc=0.691, acc_len=1.64, thru=15.2, I=18.6, T=32.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:11,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:11,860] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:17,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:43:18,547] loop.py:45 INFO TRAIN_STEP step=9283 loss=1.320599 Training: 52%|█████▏ | 9283/18018 [20:13:04<17:55:35, 7.39s/step, loss=1.321, acc=0.699, acc_len=1.68, thru=17.2, I=18.7, T=33.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:18,816] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:26,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:43:27,313] loop.py:45 INFO TRAIN_STEP step=9284 loss=1.336245 Training: 52%|█████▏ | 9284/18018 [20:13:13<18:55:36, 7.80s/step, loss=1.336, acc=0.694, acc_len=1.66, thru=14.3, I=18.0, T=30.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:27,777] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:43:34,826] loop.py:380 INFO TIMING step=9285: step=4.162s data=1.823s compute=4.132s [fwd=2.945s bwd=1.172s opt=0.003s] dispatch=3.341s [2026-10-19 09:43:34,826] loop.py:45 INFO TRAIN_STEP step=9285 loss=1.278120 Training: 52%|█████▏ | 9285/18018 [20:13:20<18:42:53, 7.71s/step, loss=1.278, acc=0.707, acc_len=1.70, thru=11.7, I=17.5, T=30.8, wait=3.3s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:43,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:43:45,276] loop.py:45 INFO TRAIN_STEP step=9286 loss=1.367887 Training: 52%|█████▏ | 9286/18018 [20:13:31<20:42:04, 8.53s/step, loss=1.368, acc=0.697, acc_len=1.67, thru=16.1, I=19.4, T=31.0, wait=6.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:43:45,272] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9285: forward=2935.6ms backward=1130.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:45,667] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:51,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 09:43:51,924] loop.py:45 INFO TRAIN_STEP step=9287 loss=1.312193 Training: 52%|█████▏ | 9287/18018 [20:13:38<19:19:36, 7.97s/step, loss=1.312, acc=0.697, acc_len=1.66, thru=17.7, I=24.1, T=32.7, wait=2.7s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:43:52,144] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=5/64 [2026-10-19 09:43:59,631] loop.py:45 INFO TRAIN_STEP step=9288 loss=1.247814 Training: 52%|█████▏ | 9288/18018 [20:13:45<19:08:00, 7.89s/step, loss=1.248, acc=0.711, acc_len=1.72, thru=16.4, I=17.4, T=35.3, wait=4.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:06,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:44:08,347] loop.py:45 INFO TRAIN_STEP step=9289 loss=1.406391 Training: 52%|█████▏ | 9289/18018 [20:13:54<19:43:57, 8.14s/step, loss=1.406, acc=0.685, acc_len=1.63, thru=15.7, I=18.2, T=29.4, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:08,615] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:15,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:44:15,849] loop.py:380 INFO TIMING step=9290: step=3.976s data=1.673s compute=3.929s [fwd=2.822s bwd=1.091s opt=0.003s] dispatch=3.518s [2026-10-19 09:44:15,850] loop.py:45 INFO TRAIN_STEP step=9290 loss=1.396348 Training: 52%|█████▏ | 9290/18018 [20:14:01<19:16:04, 7.95s/step, loss=1.396, acc=0.684, acc_len=1.63, thru=15.3, I=16.5, T=32.2, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:16,285] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:21,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:44:22,768] loop.py:45 INFO TRAIN_STEP step=9291 loss=1.341171 Training: 52%|█████▏ | 9291/18018 [20:14:08<18:31:04, 7.64s/step, loss=1.341, acc=0.697, acc_len=1.67, thru=19.0, I=20.0, T=30.2, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:44:22,764] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9290: forward=2949.7ms backward=1234.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:22,956] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:44:29,268] loop.py:45 INFO TRAIN_STEP step=9292 loss=1.352104 Training: 52%|█████▏ | 9292/18018 [20:14:15<17:41:14, 7.30s/step, loss=1.352, acc=0.688, acc_len=1.64, thru=15.5, I=18.3, T=35.8, wait=2.9s, pool=41, epoch=2/2][2026-10-19 09:44:36,688] loop.py:45 INFO TRAIN_STEP step=9293 loss=1.294864 Training: 52%|█████▏ | 9293/18018 [20:14:22<17:46:30, 7.33s/step, loss=1.295, acc=0.699, acc_len=1.68, thru=18.7, I=17.7, T=32.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:36,637] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:37,139] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:43,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:44:44,471] loop.py:45 INFO TRAIN_STEP step=9294 loss=1.295604 Training: 52%|█████▏ | 9294/18018 [20:14:30<18:05:57, 7.47s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=17.0, I=19.0, T=34.1, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:44,802] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:44:50,716] loop.py:380 INFO TIMING step=9295: step=3.460s data=1.398s compute=3.433s [fwd=2.432s bwd=0.986s opt=0.003s] dispatch=2.776s [2026-10-19 09:44:50,717] loop.py:45 INFO TRAIN_STEP step=9295 loss=1.233797 Training: 52%|█████▏ | 9295/18018 [20:14:36<17:12:30, 7.10s/step, loss=1.234, acc=0.713, acc_len=1.72, thru=18.5, I=22.2, T=37.0, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:58,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:44:58,638] loop.py:45 INFO TRAIN_STEP step=9296 loss=1.388893 Training: 52%|█████▏ | 9296/18018 [20:14:44<17:48:02, 7.35s/step, loss=1.389, acc=0.684, acc_len=1.62, thru=17.3, I=17.1, T=32.0, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:44:58,634] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9295: forward=2757.0ms backward=1189.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:44:58,936] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:45:04,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:45:05,445] loop.py:45 INFO TRAIN_STEP step=9297 loss=1.296124 Training: 52%|█████▏ | 9297/18018 [20:14:51<17:24:27, 7.19s/step, loss=1.296, acc=0.701, acc_len=1.69, thru=17.1, I=20.7, T=31.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:45:05,716] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:45:12,274] loop.py:45 INFO TRAIN_STEP step=9298 loss=1.458551 Training: 52%|█████▏ | 9298/18018 [20:14:58<17:08:41, 7.08s/step, loss=1.459, acc=0.670, acc_len=1.58, thru=16.2, I=19.4, T=32.3, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:45:20,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:45:22,063] loop.py:45 INFO TRAIN_STEP step=9299 loss=1.409871 Training: 52%|█████▏ | 9299/18018 [20:15:08<19:06:46, 7.89s/step, loss=1.410, acc=0.681, acc_len=1.62, thru=15.4, I=15.2, T=27.8, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:45:22,413] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:45:27,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:45:29,514] loop.py:380 INFO TIMING step=9300: step=3.993s data=1.699s compute=3.943s [fwd=2.810s bwd=1.117s opt=0.003s] dispatch=3.449s [2026-10-19 09:45:29,514] loop.py:45 INFO TRAIN_STEP step=9300 loss=1.337036 Training: 52%|█████▏ | 9300/18018 [20:15:15<18:47:30, 7.76s/step, loss=1.337, acc=0.693, acc_len=1.66, thru=15.3, I=21.0, T=32.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:45:29,901] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:45:36,091] loop.py:45 INFO TRAIN_STEP step=9301 loss=1.304222 Training: 52%|█████▏ | 9301/18018 [20:15:22<17:55:45, 7.40s/step, loss=1.304, acc=0.702, acc_len=1.69, thru=15.7, I=18.7, T=37.0, wait=3.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:45:36,087] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9300: forward=2477.1ms backward=938.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:45:43,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:45:44,233] loop.py:45 INFO TRAIN_STEP step=9302 loss=1.297803 Training: 52%|█████▏ | 9302/18018 [20:15:30<18:27:46, 7.63s/step, loss=1.298, acc=0.697, acc_len=1.67, thru=17.6, I=18.4, T=33.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:45:44,602] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:45:50,230] loop.py:45 INFO TRAIN_STEP step=9303 loss=1.358418 Training: 52%|█████▏ | 9303/18018 [20:15:36<17:16:37, 7.14s/step, loss=1.358, acc=0.692, acc_len=1.66, thru=18.1, I=21.2, T=36.2, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:45:58,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:46:00,051] loop.py:45 INFO TRAIN_STEP step=9304 loss=1.525426 Training: 52%|█████▏ | 9304/18018 [20:15:46<19:13:30, 7.94s/step, loss=1.525, acc=0.661, acc_len=1.55, thru=15.0, I=15.7, T=31.7, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:00,372] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:04,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:46:06,369] loop.py:380 INFO TIMING step=9305: step=4.251s data=1.817s compute=4.221s [fwd=3.005s bwd=1.200s opt=0.003s] dispatch=2.059s [2026-10-19 09:46:06,369] loop.py:45 INFO TRAIN_STEP step=9305 loss=1.317809 Training: 52%|█████▏ | 9305/18018 [20:15:52<18:02:33, 7.45s/step, loss=1.318, acc=0.700, acc_len=1.68, thru=18.9, I=22.9, T=30.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:06,864] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:12,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:46:13,439] loop.py:45 INFO TRAIN_STEP step=9306 loss=1.419842 Training: 52%|█████▏ | 9306/18018 [20:15:59<17:45:47, 7.34s/step, loss=1.420, acc=0.678, acc_len=1.60, thru=16.3, I=21.5, T=30.4, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:46:13,434] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9305: forward=2922.8ms backward=1252.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:13,582] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 09:46:20,055] loop.py:45 INFO TRAIN_STEP step=9307 loss=1.281585 Training: 52%|█████▏ | 9307/18018 [20:16:06<17:13:59, 7.12s/step, loss=1.282, acc=0.707, acc_len=1.71, thru=16.7, I=19.8, T=31.1, wait=2.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:27,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:46:28,453] loop.py:45 INFO TRAIN_STEP step=9308 loss=1.271267 Training: 52%|█████▏ | 9308/18018 [20:16:14<18:09:34, 7.51s/step, loss=1.271, acc=0.706, acc_len=1.71, thru=16.0, I=16.3, T=32.4, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:28,541] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 09:46:36,696] loop.py:45 INFO TRAIN_STEP step=9309 loss=1.355171 Training: 52%|█████▏ | 9309/18018 [20:16:22<18:41:25, 7.73s/step, loss=1.355, acc=0.693, acc_len=1.65, thru=14.4, I=16.1, T=26.9, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:43,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:46:44,263] loop.py:380 INFO TIMING step=9310: step=4.435s data=1.787s compute=4.408s [fwd=3.078s bwd=1.315s opt=0.003s] dispatch=3.124s [2026-10-19 09:46:44,263] loop.py:45 INFO TRAIN_STEP step=9310 loss=1.293411 Training: 52%|█████▏ | 9310/18018 [20:16:30<18:34:28, 7.68s/step, loss=1.293, acc=0.704, acc_len=1.70, thru=16.8, I=17.4, T=28.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:44,570] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:49,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:46:50,443] loop.py:45 INFO TRAIN_STEP step=9311 loss=1.233930 Training: 52%|█████▏ | 9311/18018 [20:16:36<17:29:06, 7.23s/step, loss=1.234, acc=0.716, acc_len=1.73, thru=18.3, I=21.6, T=32.3, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:46:50,439] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9310: forward=2722.2ms backward=1191.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:50,867] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:57,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:46:57,512] loop.py:45 INFO TRAIN_STEP step=9312 loss=1.390668 Training: 52%|█████▏ | 9312/18018 [20:16:43<17:21:53, 7.18s/step, loss=1.391, acc=0.684, acc_len=1.63, thru=17.0, I=18.9, T=30.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:46:57,824] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:09,373] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:47:09,912] loop.py:45 INFO TRAIN_STEP step=9313 loss=1.411093 Training: 52%|█████▏ | 9313/18018 [20:16:55<21:08:58, 8.75s/step, loss=1.411, acc=0.682, acc_len=1.62, thru=9.1, I=9.0, T=26.5, wait=7.6s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:10,377] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:16,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:47:17,563] loop.py:45 INFO TRAIN_STEP step=9314 loss=1.322778 Training: 52%|█████▏ | 9314/18018 [20:17:03<20:21:09, 8.42s/step, loss=1.323, acc=0.695, acc_len=1.66, thru=17.3, I=17.9, T=32.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:17,778] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:23,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:47:24,006] loop.py:380 INFO TIMING step=9315: step=3.489s data=1.385s compute=3.454s [fwd=2.458s bwd=0.981s opt=0.003s] dispatch=2.946s [2026-10-19 09:47:24,006] loop.py:45 INFO TRAIN_STEP step=9315 loss=1.170973 Training: 52%|█████▏ | 9315/18018 [20:17:10<18:55:07, 7.83s/step, loss=1.171, acc=0.724, acc_len=1.76, thru=18.3, I=19.7, T=36.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:24,367] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:31,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:47:32,931] loop.py:45 INFO TRAIN_STEP step=9316 loss=1.417185 Training: 52%|█████▏ | 9316/18018 [20:17:19<19:42:51, 8.16s/step, loss=1.417, acc=0.680, acc_len=1.61, thru=14.9, I=15.9, T=31.9, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:32,986] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (TrainerActor pid=9310) [2026-10-19 09:47:32,926] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9315: forward=2778.5ms backward=1150.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:37,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:47:39,029] loop.py:45 INFO TRAIN_STEP step=9317 loss=1.333135 Training: 52%|█████▏ | 9317/18018 [20:17:25<18:13:15, 7.54s/step, loss=1.333, acc=0.695, acc_len=1.66, thru=20.2, I=22.4, T=34.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:47:39,330] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:47:47,608] loop.py:45 INFO TRAIN_STEP step=9318 loss=1.375071 Training: 52%|█████▏ | 9318/18018 [20:17:33<18:58:22, 7.85s/step, loss=1.375, acc=0.686, acc_len=1.63, thru=11.8, I=15.5, T=34.3, wait=4.8s, pool=45, epoch=2/2][2026-10-19 09:47:55,314] loop.py:45 INFO TRAIN_STEP step=9319 loss=1.392006 Training: 52%|█████▏ | 9319/18018 [20:17:41<18:51:47, 7.81s/step, loss=1.392, acc=0.682, acc_len=1.62, thru=18.1, I=18.0, T=35.0, wait=4.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:05,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:48:07,090] loop.py:380 INFO TIMING step=9320: step=4.670s data=2.126s compute=4.621s [fwd=3.297s bwd=1.308s opt=0.003s] dispatch=7.099s [2026-10-19 09:48:07,091] loop.py:45 INFO TRAIN_STEP step=9320 loss=1.428869 Training: 52%|█████▏ | 9320/18018 [20:17:53<21:44:23, 9.00s/step, loss=1.429, acc=0.686, acc_len=1.63, thru=12.3, I=12.6, T=27.4, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:07,553] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:48:14,178] loop.py:45 INFO TRAIN_STEP step=9321 loss=1.260371 Training: 52%|█████▏ | 9321/18018 [20:18:00<20:21:06, 8.42s/step, loss=1.260, acc=0.708, acc_len=1.71, thru=15.4, I=20.3, T=28.8, wait=2.6s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:48:14,174] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9320: forward=3038.7ms backward=1355.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:20,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:48:21,309] loop.py:45 INFO TRAIN_STEP step=9322 loss=1.346561 Training: 52%|█████▏ | 9322/18018 [20:18:07<19:24:49, 8.04s/step, loss=1.347, acc=0.694, acc_len=1.66, thru=18.1, I=17.4, T=32.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:21,698] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:26,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:48:27,586] loop.py:45 INFO TRAIN_STEP step=9323 loss=1.343551 Training: 52%|█████▏ | 9323/18018 [20:18:13<18:08:13, 7.51s/step, loss=1.344, acc=0.691, acc_len=1.66, thru=19.0, I=22.0, T=35.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:27,946] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:48:34,138] loop.py:45 INFO TRAIN_STEP step=9324 loss=1.305529 Training: 52%|█████▏ | 9324/18018 [20:18:20<17:26:19, 7.22s/step, loss=1.306, acc=0.700, acc_len=1.69, thru=20.7, I=21.4, T=35.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:34,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:34,621] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 09:48:41,414] loop.py:380 INFO TIMING step=9325: step=4.160s data=1.925s compute=4.111s [fwd=2.894s bwd=1.202s opt=0.003s] dispatch=3.108s [2026-10-19 09:48:41,415] loop.py:45 INFO TRAIN_STEP step=9325 loss=1.307528 Training: 52%|█████▏ | 9325/18018 [20:18:27<17:28:41, 7.24s/step, loss=1.308, acc=0.704, acc_len=1.70, thru=15.8, I=17.2, T=30.8, wait=3.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:49,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:48:49,568] loop.py:45 INFO TRAIN_STEP step=9326 loss=1.398687 Training: 52%|█████▏ | 9326/18018 [20:18:35<18:08:20, 7.51s/step, loss=1.399, acc=0.684, acc_len=1.63, thru=16.6, I=16.9, T=32.6, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:48:49,564] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9325: forward=2724.9ms backward=1135.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:50,032] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:54,674] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:48:56,460] loop.py:45 INFO TRAIN_STEP step=9327 loss=1.400961 Training: 52%|█████▏ | 9327/18018 [20:18:42<17:41:19, 7.33s/step, loss=1.401, acc=0.685, acc_len=1.64, thru=18.5, I=23.3, T=30.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:48:56,688] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:02,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:49:03,698] loop.py:45 INFO TRAIN_STEP step=9328 loss=1.361230 Training: 52%|█████▏ | 9328/18018 [20:18:49<17:37:18, 7.30s/step, loss=1.361, acc=0.691, acc_len=1.66, thru=15.5, I=21.9, T=32.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:03,881] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:09,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:49:10,994] loop.py:45 INFO TRAIN_STEP step=9329 loss=1.456348 Training: 52%|█████▏ | 9329/18018 [20:18:57<17:36:57, 7.30s/step, loss=1.456, acc=0.671, acc_len=1.59, thru=16.2, I=19.5, T=30.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:11,379] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:17,122] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:49:17,721] loop.py:380 INFO TIMING step=9330: step=3.536s data=1.473s compute=3.509s [fwd=2.501s bwd=0.994s opt=0.003s] dispatch=3.181s [2026-10-19 09:49:17,721] loop.py:45 INFO TRAIN_STEP step=9330 loss=1.212244 Training: 52%|█████▏ | 9330/18018 [20:19:03<17:11:57, 7.13s/step, loss=1.212, acc=0.715, acc_len=1.74, thru=18.2, I=21.6, T=36.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:18,127] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:49:24,286] loop.py:45 INFO TRAIN_STEP step=9331 loss=1.324552 Training: 52%|█████▏ | 9331/18018 [20:19:10<16:47:24, 6.96s/step, loss=1.325, acc=0.696, acc_len=1.67, thru=18.4, I=20.3, T=33.1, wait=2.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:49:24,282] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9330: forward=2751.9ms backward=1070.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:32,131] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:49:32,569] loop.py:45 INFO TRAIN_STEP step=9332 loss=1.405785 Training: 52%|█████▏ | 9332/18018 [20:19:18<17:44:56, 7.36s/step, loss=1.406, acc=0.682, acc_len=1.62, thru=17.2, I=16.9, T=34.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:32,640] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:37,746] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:49:39,136] loop.py:45 INFO TRAIN_STEP step=9333 loss=1.403823 Training: 52%|█████▏ | 9333/18018 [20:19:25<17:10:31, 7.12s/step, loss=1.404, acc=0.685, acc_len=1.63, thru=18.9, I=22.3, T=35.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:39,268] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 09:49:46,250] loop.py:45 INFO TRAIN_STEP step=9334 loss=1.369821 Training: 52%|█████▏ | 9334/18018 [20:19:32<17:10:07, 7.12s/step, loss=1.370, acc=0.684, acc_len=1.63, thru=15.2, I=24.0, T=37.1, wait=3.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:53,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:49:54,693] loop.py:380 INFO TIMING step=9335: step=3.947s data=1.747s compute=3.912s [fwd=2.808s bwd=1.088s opt=0.003s] dispatch=4.488s [2026-10-19 09:49:54,694] loop.py:45 INFO TRAIN_STEP step=9335 loss=1.445341 Training: 52%|█████▏ | 9335/18018 [20:19:40<18:07:40, 7.52s/step, loss=1.445, acc=0.678, acc_len=1.60, thru=16.8, I=18.5, T=32.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:49:55,068] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 09:50:01,568] loop.py:45 INFO TRAIN_STEP step=9336 loss=1.398094 Training: 52%|█████▏ | 9336/18018 [20:19:47<17:39:40, 7.32s/step, loss=1.398, acc=0.689, acc_len=1.64, thru=16.6, I=20.4, T=33.3, wait=3.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:50:01,563] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9335: forward=2714.7ms backward=1079.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:07,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:50:09,283] loop.py:45 INFO TRAIN_STEP step=9337 loss=1.366129 Training: 52%|█████▏ | 9337/18018 [20:19:55<17:56:32, 7.44s/step, loss=1.366, acc=0.690, acc_len=1.65, thru=18.9, I=19.5, T=25.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:09,564] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:17,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:50:18,209] loop.py:45 INFO TRAIN_STEP step=9338 loss=1.288149 Training: 52%|█████▏ | 9338/18018 [20:20:04<19:00:54, 7.89s/step, loss=1.288, acc=0.704, acc_len=1.70, thru=14.8, I=15.0, T=29.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:18,455] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:26,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:50:27,303] loop.py:45 INFO TRAIN_STEP step=9339 loss=1.275278 Training: 52%|█████▏ | 9339/18018 [20:20:13<19:53:06, 8.25s/step, loss=1.275, acc=0.703, acc_len=1.70, thru=14.3, I=14.7, T=30.5, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:27,723] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:34,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:50:34,789] loop.py:380 INFO TIMING step=9340: step=3.841s data=1.649s compute=3.809s [fwd=2.714s bwd=1.081s opt=0.003s] dispatch=3.636s [2026-10-19 09:50:34,789] loop.py:45 INFO TRAIN_STEP step=9340 loss=1.306669 Training: 52%|█████▏ | 9340/18018 [20:20:20<19:19:54, 8.02s/step, loss=1.307, acc=0.700, acc_len=1.69, thru=16.5, I=17.4, T=33.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:35,114] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:39,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:50:42,351] loop.py:45 INFO TRAIN_STEP step=9341 loss=1.325219 Training: 52%|█████▏ | 9341/18018 [20:20:28<18:59:53, 7.88s/step, loss=1.325, acc=0.693, acc_len=1.65, thru=16.9, I=21.8, T=26.0, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:50:42,346] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9340: forward=3419.1ms backward=1444.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:42,843] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:48,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:50:50,228] loop.py:45 INFO TRAIN_STEP step=9342 loss=1.302847 Training: 52%|█████▏ | 9342/18018 [20:20:36<18:59:31, 7.88s/step, loss=1.303, acc=0.701, acc_len=1.69, thru=17.3, I=20.3, T=29.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:50,705] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:55,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:50:56,799] loop.py:45 INFO TRAIN_STEP step=9343 loss=1.182354 Training: 52%|█████▏ | 9343/18018 [20:20:42<18:02:35, 7.49s/step, loss=1.182, acc=0.721, acc_len=1.76, thru=17.6, I=26.2, T=34.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:50:57,294] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:51:04,355] loop.py:45 INFO TRAIN_STEP step=9344 loss=1.340769 Training: 52%|█████▏ | 9344/18018 [20:20:50<18:05:39, 7.51s/step, loss=1.341, acc=0.692, acc_len=1.66, thru=15.3, I=21.4, T=32.9, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:51:13,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:51:14,012] loop.py:380 INFO TIMING step=9345: step=4.216s data=1.751s compute=4.187s [fwd=2.938s bwd=1.235s opt=0.003s] dispatch=5.427s [2026-10-19 09:51:14,012] loop.py:45 INFO TRAIN_STEP step=9345 loss=1.459420 Training: 52%|█████▏ | 9345/18018 [20:21:00<19:38:29, 8.15s/step, loss=1.459, acc=0.673, acc_len=1.59, thru=14.3, I=13.3, T=30.4, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:51:14,518] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:51:21,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:51:23,135] loop.py:45 INFO TRAIN_STEP step=9346 loss=1.305505 Training: 52%|█████▏ | 9346/18018 [20:21:09<20:20:27, 8.44s/step, loss=1.306, acc=0.697, acc_len=1.68, thru=14.2, I=16.9, T=30.2, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:51:23,130] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9345: forward=2938.6ms backward=1249.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:51:23,170] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 09:51:29,228] loop.py:45 INFO TRAIN_STEP step=9347 loss=1.335308 Training: 52%|█████▏ | 9347/18018 [20:21:15<18:38:13, 7.74s/step, loss=1.335, acc=0.695, acc_len=1.66, thru=18.5, I=19.9, T=32.2, wait=2.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:51:37,279] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:51:38,214] loop.py:45 INFO TRAIN_STEP step=9348 loss=1.470781 Training: 52%|█████▏ | 9348/18018 [20:21:24<19:32:23, 8.11s/step, loss=1.471, acc=0.676, acc_len=1.61, thru=15.5, I=17.6, T=29.2, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:51:38,299] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:51:45,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:51:46,402] loop.py:45 INFO TRAIN_STEP step=9349 loss=1.351761 Training: 52%|█████▏ | 9349/18018 [20:21:32<19:35:21, 8.13s/step, loss=1.352, acc=0.692, acc_len=1.66, thru=14.2, I=14.8, T=29.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:51:46,483] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 09:51:53,218] loop.py:380 INFO TIMING step=9350: step=3.994s data=1.722s compute=3.955s [fwd=2.797s bwd=1.142s opt=0.003s] dispatch=2.814s [2026-10-19 09:51:53,218] loop.py:45 INFO TRAIN_STEP step=9350 loss=1.296666 Training: 52%|█████▏ | 9350/18018 [20:21:39<18:38:15, 7.74s/step, loss=1.297, acc=0.699, acc_len=1.68, thru=17.5, I=19.1, T=32.1, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:01,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:52:02,551] loop.py:45 INFO TRAIN_STEP step=9351 loss=1.370222 Training: 52%|█████▏ | 9351/18018 [20:21:48<19:46:57, 8.22s/step, loss=1.370, acc=0.690, acc_len=1.65, thru=15.2, I=14.0, T=31.6, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:52:02,546] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9350: forward=2873.9ms backward=1123.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:02,920] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:11,710] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:52:12,596] loop.py:45 INFO TRAIN_STEP step=9352 loss=1.363661 Training: 52%|█████▏ | 9352/18018 [20:21:58<21:06:08, 8.77s/step, loss=1.364, acc=0.691, acc_len=1.65, thru=13.2, I=14.3, T=28.6, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:12,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:18,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:52:19,633] loop.py:45 INFO TRAIN_STEP step=9353 loss=1.357953 Training: 52%|█████▏ | 9353/18018 [20:22:05<19:50:55, 8.25s/step, loss=1.358, acc=0.693, acc_len=1.65, thru=17.4, I=21.3, T=27.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:20,119] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:52:26,762] loop.py:45 INFO TRAIN_STEP step=9354 loss=1.303131 Training: 52%|█████▏ | 9354/18018 [20:22:12<19:02:22, 7.91s/step, loss=1.303, acc=0.704, acc_len=1.70, thru=16.0, I=17.4, T=34.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:26,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:27,204] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:32,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:52:33,964] loop.py:380 INFO TIMING step=9355: step=4.559s data=2.258s compute=4.527s [fwd=3.351s bwd=1.161s opt=0.003s] dispatch=2.636s [2026-10-19 09:52:33,965] loop.py:45 INFO TRAIN_STEP step=9355 loss=1.424604 Training: 52%|█████▏ | 9355/18018 [20:22:20<18:31:41, 7.70s/step, loss=1.425, acc=0.677, acc_len=1.61, thru=17.4, I=22.2, T=28.1, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:34,173] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:42,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:52:43,409] loop.py:45 INFO TRAIN_STEP step=9356 loss=1.351317 Training: 52%|█████▏ | 9356/18018 [20:22:29<19:47:05, 8.22s/step, loss=1.351, acc=0.697, acc_len=1.67, thru=13.9, I=14.3, T=26.8, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:52:43,404] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9355: forward=3366.1ms backward=1365.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:43,541] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:49,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:52:50,736] loop.py:45 INFO TRAIN_STEP step=9357 loss=1.363788 Training: 52%|█████▏ | 9357/18018 [20:22:36<19:08:11, 7.95s/step, loss=1.364, acc=0.690, acc_len=1.65, thru=16.0, I=19.6, T=27.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:52:50,862] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 09:52:57,515] loop.py:45 INFO TRAIN_STEP step=9358 loss=1.301776 Training: 52%|█████▏ | 9358/18018 [20:22:43<18:17:09, 7.60s/step, loss=1.302, acc=0.700, acc_len=1.68, thru=15.2, I=16.7, T=33.6, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:05,551] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:53:06,443] loop.py:45 INFO TRAIN_STEP step=9359 loss=1.339911 Training: 52%|█████▏ | 9359/18018 [20:22:52<19:14:28, 8.00s/step, loss=1.340, acc=0.694, acc_len=1.66, thru=15.9, I=17.4, T=29.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:06,571] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:12,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:53:13,321] loop.py:380 INFO TIMING step=9360: step=4.174s data=1.867s compute=4.131s [fwd=2.897s bwd=1.218s opt=0.003s] dispatch=2.695s [2026-10-19 09:53:13,321] loop.py:45 INFO TRAIN_STEP step=9360 loss=1.412243 Training: 52%|█████▏ | 9360/18018 [20:22:59<18:25:44, 7.66s/step, loss=1.412, acc=0.683, acc_len=1.63, thru=18.4, I=20.2, T=30.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:13,419] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:19,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:53:20,441] loop.py:45 INFO TRAIN_STEP step=9361 loss=1.406850 Training: 52%|█████▏ | 9361/18018 [20:23:06<18:02:10, 7.50s/step, loss=1.407, acc=0.682, acc_len=1.62, thru=16.1, I=19.3, T=32.5, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:53:20,437] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9360: forward=2782.5ms backward=1112.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:20,586] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:26,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:53:27,581] loop.py:45 INFO TRAIN_STEP step=9362 loss=1.348495 Training: 52%|█████▏ | 9362/18018 [20:23:13<17:46:28, 7.39s/step, loss=1.348, acc=0.693, acc_len=1.66, thru=17.0, I=23.1, T=28.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:27,617] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 09:53:36,429] loop.py:45 INFO TRAIN_STEP step=9363 loss=1.351547 Training: 52%|█████▏ | 9363/18018 [20:23:22<18:49:12, 7.83s/step, loss=1.352, acc=0.694, acc_len=1.66, thru=13.7, I=14.1, T=33.9, wait=5.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:44,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:53:44,245] loop.py:45 INFO TRAIN_STEP step=9364 loss=1.461314 Training: 52%|█████▏ | 9364/18018 [20:23:30<18:48:34, 7.82s/step, loss=1.461, acc=0.672, acc_len=1.59, thru=16.9, I=16.2, T=33.1, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:44,575] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:49,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:53:50,880] loop.py:380 INFO TIMING step=9365: step=3.898s data=1.733s compute=3.867s [fwd=2.786s bwd=1.066s opt=0.003s] dispatch=2.730s [2026-10-19 09:53:50,881] loop.py:45 INFO TRAIN_STEP step=9365 loss=1.386113 Training: 52%|█████▏ | 9365/18018 [20:23:36<17:57:03, 7.47s/step, loss=1.386, acc=0.689, acc_len=1.64, thru=18.4, I=22.9, T=32.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:51,260] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:56,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:53:57,336] loop.py:45 INFO TRAIN_STEP step=9366 loss=1.293220 Training: 52%|█████▏ | 9366/18018 [20:23:43<17:13:09, 7.16s/step, loss=1.293, acc=0.704, acc_len=1.70, thru=18.2, I=23.4, T=33.8, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:53:57,332] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9365: forward=2674.3ms backward=1066.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:53:57,413] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:03,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:54:03,947] loop.py:45 INFO TRAIN_STEP step=9367 loss=1.381218 Training: 52%|█████▏ | 9367/18018 [20:23:50<16:48:56, 7.00s/step, loss=1.381, acc=0.689, acc_len=1.65, thru=18.9, I=21.7, T=32.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:04,274] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:54:10,375] loop.py:45 INFO TRAIN_STEP step=9368 loss=1.302162 Training: 52%|█████▏ | 9368/18018 [20:23:56<16:24:11, 6.83s/step, loss=1.302, acc=0.699, acc_len=1.68, thru=16.1, I=18.4, T=33.6, wait=2.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:17,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:54:18,869] loop.py:45 INFO TRAIN_STEP step=9369 loss=1.323450 Training: 52%|█████▏ | 9369/18018 [20:24:04<17:36:13, 7.33s/step, loss=1.323, acc=0.696, acc_len=1.67, thru=18.9, I=17.8, T=33.2, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:19,228] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:25,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:54:26,095] loop.py:380 INFO TIMING step=9370: step=3.785s data=1.602s compute=3.755s [fwd=2.606s bwd=1.134s opt=0.003s] dispatch=3.432s [2026-10-19 09:54:26,095] loop.py:45 INFO TRAIN_STEP step=9370 loss=1.372013 Training: 52%|█████▏ | 9370/18018 [20:24:12<17:31:42, 7.30s/step, loss=1.372, acc=0.688, acc_len=1.64, thru=16.9, I=19.5, T=33.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:26,466] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 09:54:33,536] loop.py:45 INFO TRAIN_STEP step=9371 loss=1.321157 Training: 52%|█████▏ | 9371/18018 [20:24:19<17:38:05, 7.34s/step, loss=1.321, acc=0.699, acc_len=1.68, thru=16.7, I=17.8, T=34.9, wait=3.8s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:54:33,532] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9370: forward=2509.9ms backward=1081.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:41,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:54:42,685] loop.py:45 INFO TRAIN_STEP step=9372 loss=1.351016 Training: 52%|█████▏ | 9372/18018 [20:24:28<18:55:54, 7.88s/step, loss=1.351, acc=0.692, acc_len=1.66, thru=15.6, I=15.5, T=27.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:42,781] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 09:54:49,271] loop.py:45 INFO TRAIN_STEP step=9373 loss=1.475259 Training: 52%|█████▏ | 9373/18018 [20:24:35<17:59:45, 7.49s/step, loss=1.475, acc=0.670, acc_len=1.58, thru=17.3, I=20.2, T=33.8, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:58,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:54:59,065] loop.py:45 INFO TRAIN_STEP step=9374 loss=1.253325 Training: 52%|█████▏ | 9374/18018 [20:24:45<19:39:01, 8.18s/step, loss=1.253, acc=0.708, acc_len=1.70, thru=15.0, I=12.7, T=32.4, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:54:59,290] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 09:55:05,588] loop.py:380 INFO TIMING step=9375: step=4.017s data=1.818s compute=3.985s [fwd=2.792s bwd=1.178s opt=0.003s] dispatch=2.494s [2026-10-19 09:55:05,588] loop.py:45 INFO TRAIN_STEP step=9375 loss=1.401834 Training: 52%|█████▏ | 9375/18018 [20:24:51<18:27:07, 7.69s/step, loss=1.402, acc=0.683, acc_len=1.63, thru=19.0, I=20.9, T=31.9, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:55:11,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:55:13,180] loop.py:45 INFO TRAIN_STEP step=9376 loss=1.403319 Training: 52%|█████▏ | 9376/18018 [20:24:59<18:22:55, 7.66s/step, loss=1.403, acc=0.684, acc_len=1.63, thru=15.8, I=18.4, T=32.0, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:55:13,176] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9375: forward=2790.8ms backward=1148.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:55:13,608] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:55:19,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:55:21,105] loop.py:45 INFO TRAIN_STEP step=9377 loss=1.264857 Training: 52%|█████▏ | 9377/18018 [20:25:07<18:34:17, 7.74s/step, loss=1.265, acc=0.708, acc_len=1.71, thru=14.6, I=19.2, T=30.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:55:21,470] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:55:28,487] loop.py:45 INFO TRAIN_STEP step=9378 loss=1.389455 Training: 52%|█████▏ | 9378/18018 [20:25:14<18:18:50, 7.63s/step, loss=1.389, acc=0.685, acc_len=1.62, thru=15.6, I=17.2, T=32.0, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:55:36,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:55:37,542] loop.py:45 INFO TRAIN_STEP step=9379 loss=1.492158 Training: 52%|█████▏ | 9379/18018 [20:25:23<19:20:12, 8.06s/step, loss=1.492, acc=0.670, acc_len=1.58, thru=15.7, I=16.4, T=28.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:55:37,991] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:55:45,887] loop.py:380 INFO TIMING step=9380: step=4.600s data=1.993s compute=4.548s [fwd=3.116s bwd=1.417s opt=0.003s] dispatch=3.736s [2026-10-19 09:55:45,887] loop.py:45 INFO TRAIN_STEP step=9380 loss=1.246039 Training: 52%|█████▏ | 9380/18018 [20:25:31<19:32:24, 8.14s/step, loss=1.246, acc=0.709, acc_len=1.72, thru=11.9, I=15.2, T=27.8, wait=3.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:55:55,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:55:56,181] loop.py:45 INFO TRAIN_STEP step=9381 loss=1.434181 Training: 52%|█████▏ | 9381/18018 [20:25:42<21:05:16, 8.79s/step, loss=1.434, acc=0.678, acc_len=1.61, thru=15.5, I=13.8, T=30.3, wait=6.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:55:56,177] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9380: forward=2992.3ms backward=1185.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:55:56,482] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:56:02,873] loop.py:45 INFO TRAIN_STEP step=9382 loss=1.394835 Training: 52%|█████▏ | 9382/18018 [20:25:48<19:34:32, 8.16s/step, loss=1.395, acc=0.686, acc_len=1.63, thru=17.4, I=18.3, T=35.7, wait=3.1s, pool=57, epoch=2/2][2026-10-19 09:56:08,209] loop.py:45 INFO TRAIN_STEP step=9383 loss=1.306937 Training: 52%|█████▏ | 9383/18018 [20:25:54<17:32:35, 7.31s/step, loss=1.307, acc=0.698, acc_len=1.68, thru=20.1, I=21.1, T=37.7, wait=1.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:56:16,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:56:17,625] loop.py:45 INFO TRAIN_STEP step=9384 loss=1.428185 Training: 52%|█████▏ | 9384/18018 [20:26:03<19:03:07, 7.94s/step, loss=1.428, acc=0.686, acc_len=1.63, thru=16.3, I=16.3, T=30.6, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:56:17,841] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:56:22,978] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:56:24,327] loop.py:380 INFO TIMING step=9385: step=3.956s data=1.712s compute=3.917s [fwd=2.643s bwd=1.259s opt=0.003s] dispatch=2.735s [2026-10-19 09:56:24,327] loop.py:45 INFO TRAIN_STEP step=9385 loss=1.295748 Training: 52%|█████▏ | 9385/18018 [20:26:10<18:09:12, 7.57s/step, loss=1.296, acc=0.699, acc_len=1.67, thru=18.1, I=23.6, T=32.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:56:24,480] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:56:31,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:56:31,975] loop.py:45 INFO TRAIN_STEP step=9386 loss=1.229871 Training: 52%|█████▏ | 9386/18018 [20:26:18<18:12:28, 7.59s/step, loss=1.230, acc=0.715, acc_len=1.73, thru=16.0, I=19.1, T=30.6, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:56:31,970] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9385: forward=2887.1ms backward=1241.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:56:32,327] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:56:38,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:56:38,989] loop.py:45 INFO TRAIN_STEP step=9387 loss=1.414749 Training: 52%|█████▏ | 9387/18018 [20:26:25<17:47:25, 7.42s/step, loss=1.415, acc=0.681, acc_len=1.62, thru=17.0, I=19.0, T=31.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:56:39,053] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 09:56:45,413] loop.py:45 INFO TRAIN_STEP step=9388 loss=1.385132 Training: 52%|█████▏ | 9388/18018 [20:26:31<17:04:14, 7.12s/step, loss=1.385, acc=0.683, acc_len=1.62, thru=17.4, I=21.9, T=34.1, wait=2.7s, pool=49, epoch=2/2][2026-10-19 09:56:53,538] loop.py:45 INFO TRAIN_STEP step=9389 loss=1.417291 Training: 52%|█████▏ | 9389/18018 [20:26:39<17:47:38, 7.42s/step, loss=1.417, acc=0.685, acc_len=1.62, thru=15.9, I=13.7, T=31.2, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:01,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:57:01,089] loop.py:380 INFO TIMING step=9390: step=3.912s data=1.708s compute=3.875s [fwd=2.767s bwd=1.092s opt=0.003s] dispatch=3.626s [2026-10-19 09:57:01,089] loop.py:45 INFO TRAIN_STEP step=9390 loss=1.386189 Training: 52%|█████▏ | 9390/18018 [20:26:47<17:52:48, 7.46s/step, loss=1.386, acc=0.682, acc_len=1.62, thru=18.4, I=16.1, T=32.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:01,507] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:08,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:57:08,995] loop.py:45 INFO TRAIN_STEP step=9391 loss=1.334080 Training: 52%|█████▏ | 9391/18018 [20:26:55<18:11:59, 7.59s/step, loss=1.334, acc=0.694, acc_len=1.66, thru=18.1, I=18.6, T=28.1, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:57:08,990] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9390: forward=3256.3ms backward=1247.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:09,145] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:15,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:57:17,867] loop.py:45 INFO TRAIN_STEP step=9392 loss=1.292477 Training: 52%|█████▏ | 9392/18018 [20:27:03<19:06:56, 7.98s/step, loss=1.292, acc=0.704, acc_len=1.69, thru=13.5, I=19.6, T=29.8, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:18,257] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:24,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:57:25,962] loop.py:45 INFO TRAIN_STEP step=9393 loss=1.206988 Training: 52%|█████▏ | 9393/18018 [20:27:12<19:11:52, 8.01s/step, loss=1.207, acc=0.716, acc_len=1.74, thru=16.6, I=19.4, T=32.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:26,121] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:33,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:57:34,143] loop.py:45 INFO TRAIN_STEP step=9394 loss=1.279342 Training: 52%|█████▏ | 9394/18018 [20:27:20<19:18:52, 8.06s/step, loss=1.279, acc=0.704, acc_len=1.70, thru=14.6, I=18.2, T=32.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:34,453] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 09:57:41,307] loop.py:380 INFO TIMING step=9395: step=4.014s data=1.782s compute=3.964s [fwd=2.800s bwd=1.149s opt=0.003s] dispatch=3.142s [2026-10-19 09:57:41,307] loop.py:45 INFO TRAIN_STEP step=9395 loss=1.349155 Training: 52%|█████▏ | 9395/18018 [20:27:27<18:40:12, 7.79s/step, loss=1.349, acc=0.698, acc_len=1.67, thru=15.1, I=16.2, T=31.9, wait=3.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:49,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:57:49,616] loop.py:45 INFO TRAIN_STEP step=9396 loss=1.306463 Training: 52%|█████▏ | 9396/18018 [20:27:35<19:02:04, 7.95s/step, loss=1.306, acc=0.700, acc_len=1.68, thru=18.3, I=17.2, T=26.6, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:57:49,612] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9395: forward=3357.3ms backward=1387.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:49,701] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:54,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:57:56,163] loop.py:45 INFO TRAIN_STEP step=9397 loss=1.387543 Training: 52%|█████▏ | 9397/18018 [20:27:42<18:01:39, 7.53s/step, loss=1.388, acc=0.687, acc_len=1.63, thru=18.2, I=20.1, T=31.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:57:56,317] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 09:58:01,860] loop.py:45 INFO TRAIN_STEP step=9398 loss=1.364696 Training: 52%|█████▏ | 9398/18018 [20:27:47<16:42:30, 6.98s/step, loss=1.365, acc=0.693, acc_len=1.66, thru=19.6, I=25.7, T=35.7, wait=2.1s, pool=57, epoch=2/2][2026-10-19 09:58:14,129] loop.py:45 INFO TRAIN_STEP step=9399 loss=1.393150 Training: 52%|█████▏ | 9399/18018 [20:28:00<20:30:32, 8.57s/step, loss=1.393, acc=0.685, acc_len=1.64, thru=10.3, I=8.9, T=29.2, wait=7.9s, pool=45, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:20,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:58:22,078] loop.py:380 INFO TIMING step=9400: step=4.626s data=2.078s compute=4.593s [fwd=3.299s bwd=1.279s opt=0.003s] dispatch=3.312s [2026-10-19 09:58:22,078] loop.py:45 INFO TRAIN_STEP step=9400 loss=1.282650 Training: 52%|█████▏ | 9400/18018 [20:28:08<20:03:38, 8.38s/step, loss=1.283, acc=0.709, acc_len=1.71, thru=17.8, I=18.5, T=27.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:22,578] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:28,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:58:28,486] loop.py:45 INFO TRAIN_STEP step=9401 loss=1.249579 Training: 52%|█████▏ | 9401/18018 [20:28:14<18:38:38, 7.79s/step, loss=1.250, acc=0.708, acc_len=1.72, thru=16.8, I=21.1, T=31.1, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:58:28,482] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9400: forward=2798.0ms backward=1273.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:28,749] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:35,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:58:37,647] loop.py:45 INFO TRAIN_STEP step=9402 loss=1.394995 Training: 52%|█████▏ | 9402/18018 [20:28:23<19:37:36, 8.20s/step, loss=1.395, acc=0.684, acc_len=1.63, thru=14.4, I=15.9, T=26.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:37,806] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:43,339] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:58:43,558] loop.py:45 INFO TRAIN_STEP step=9403 loss=1.217194 Training: 52%|█████▏ | 9403/18018 [20:28:29<17:58:51, 7.51s/step, loss=1.217, acc=0.714, acc_len=1.73, thru=18.6, I=23.1, T=35.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:43,841] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:49,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:58:50,631] loop.py:45 INFO TRAIN_STEP step=9404 loss=1.281351 Training: 52%|█████▏ | 9404/18018 [20:28:36<17:39:41, 7.38s/step, loss=1.281, acc=0.708, acc_len=1.71, thru=18.9, I=20.6, T=33.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:50,682] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:55,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:58:56,615] loop.py:380 INFO TIMING step=9405: step=3.548s data=1.398s compute=3.514s [fwd=2.480s bwd=1.019s opt=0.003s] dispatch=2.426s [2026-10-19 09:58:56,615] loop.py:45 INFO TRAIN_STEP step=9405 loss=1.114909 Training: 52%|█████▏ | 9405/18018 [20:28:42<16:39:22, 6.96s/step, loss=1.115, acc=0.731, acc_len=1.79, thru=20.0, I=24.8, T=36.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:58:56,789] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:05,906] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:59:06,877] loop.py:45 INFO TRAIN_STEP step=9406 loss=1.243005 Training: 52%|█████▏ | 9406/18018 [20:28:52<19:01:21, 7.95s/step, loss=1.243, acc=0.711, acc_len=1.73, thru=12.1, I=14.0, T=32.6, wait=6.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:59:06,872] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9405: forward=2767.5ms backward=1110.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:06,927] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:11,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:59:12,989] loop.py:45 INFO TRAIN_STEP step=9407 loss=1.353587 Training: 52%|█████▏ | 9407/18018 [20:28:59<17:42:04, 7.40s/step, loss=1.354, acc=0.693, acc_len=1.66, thru=19.6, I=20.3, T=34.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:13,332] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 09:59:24,844] loop.py:45 INFO TRAIN_STEP step=9408 loss=1.274319 Training: 52%|█████▏ | 9408/18018 [20:29:10<20:53:40, 8.74s/step, loss=1.274, acc=0.708, acc_len=1.71, thru=7.7, I=7.6, T=31.7, wait=7.8s, pool=45, epoch=2/2] [2026-10-19 09:59:33,023] loop.py:45 INFO TRAIN_STEP step=9409 loss=1.418780 Training: 52%|█████▏ | 9409/18018 [20:29:19<20:29:33, 8.57s/step, loss=1.419, acc=0.676, acc_len=1.60, thru=17.0, I=14.1, T=31.5, wait=4.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:40,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:59:41,129] loop.py:380 INFO TIMING step=9410: step=4.309s data=1.890s compute=4.279s [fwd=2.969s bwd=1.294s opt=0.003s] dispatch=3.787s [2026-10-19 09:59:41,129] loop.py:45 INFO TRAIN_STEP step=9410 loss=1.294437 Training: 52%|█████▏ | 9410/18018 [20:29:27<20:09:30, 8.43s/step, loss=1.294, acc=0.705, acc_len=1.70, thru=17.1, I=15.8, T=29.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:41,164] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:46,581] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:59:47,946] loop.py:45 INFO TRAIN_STEP step=9411 loss=1.377229 Training: 52%|█████▏ | 9411/18018 [20:29:34<18:59:58, 7.95s/step, loss=1.377, acc=0.689, acc_len=1.64, thru=17.3, I=21.5, T=34.7, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 09:59:47,942] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9410: forward=2577.0ms backward=1062.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:48,107] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:52,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 09:59:54,122] loop.py:45 INFO TRAIN_STEP step=9412 loss=1.480039 Training: 52%|█████▏ | 9412/18018 [20:29:40<17:43:35, 7.42s/step, loss=1.480, acc=0.670, acc_len=1.58, thru=19.8, I=25.9, T=34.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 09:59:54,457] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:00:00,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:00:01,131] loop.py:45 INFO TRAIN_STEP step=9413 loss=1.304121 Training: 52%|█████▏ | 9413/18018 [20:29:47<17:25:55, 7.29s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=17.6, I=22.8, T=35.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:00:01,251] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:00:06,947] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:00:07,155] loop.py:45 INFO TRAIN_STEP step=9414 loss=1.367198 Training: 52%|█████▏ | 9414/18018 [20:29:53<16:31:18, 6.91s/step, loss=1.367, acc=0.691, acc_len=1.65, thru=20.2, I=23.3, T=33.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:00:07,449] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:00:14,015] loop.py:380 INFO TIMING step=9415: step=3.646s data=1.558s compute=3.614s [fwd=2.599s bwd=1.000s opt=0.003s] dispatch=3.203s [2026-10-19 10:00:14,015] loop.py:45 INFO TRAIN_STEP step=9415 loss=1.396178 Training: 52%|█████▏ | 9415/18018 [20:30:00<16:28:50, 6.90s/step, loss=1.396, acc=0.687, acc_len=1.64, thru=14.6, I=16.9, T=35.1, wait=3.2s, pool=37, epoch=2/2][2026-10-19 10:00:22,725] loop.py:45 INFO TRAIN_STEP step=9416 loss=1.233019 Training: 52%|█████▏ | 9416/18018 [20:30:08<17:46:40, 7.44s/step, loss=1.233, acc=0.716, acc_len=1.74, thru=16.2, I=15.4, T=32.3, wait=4.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:00:22,720] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9415: forward=2775.4ms backward=1127.8ms [2026-10-19 10:00:31,950] loop.py:45 INFO TRAIN_STEP step=9417 loss=1.316051 Training: 52%|█████▏ | 9417/18018 [20:30:18<19:03:20, 7.98s/step, loss=1.316, acc=0.699, acc_len=1.67, thru=13.9, I=12.9, T=29.8, wait=4.9s, pool=57, epoch=2/2][2026-10-19 10:00:40,278] loop.py:45 INFO TRAIN_STEP step=9418 loss=1.356414 Training: 52%|█████▏ | 9418/18018 [20:30:26<19:18:22, 8.08s/step, loss=1.356, acc=0.692, acc_len=1.66, thru=16.0, I=16.6, T=31.7, wait=4.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:00:49,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:00:50,439] loop.py:45 INFO TRAIN_STEP step=9419 loss=1.403165 Training: 52%|█████▏ | 9419/18018 [20:30:36<20:47:43, 8.71s/step, loss=1.403, acc=0.683, acc_len=1.63, thru=12.8, I=13.7, T=27.1, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:00:50,496] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:00,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:01:01,683] loop.py:380 INFO TIMING step=9420: step=4.472s data=2.039s compute=4.416s [fwd=3.102s bwd=1.300s opt=0.003s] dispatch=6.760s [2026-10-19 10:01:01,683] loop.py:45 INFO TRAIN_STEP step=9420 loss=1.521595 Training: 52%|█████▏ | 9420/18018 [20:30:47<22:36:37, 9.47s/step, loss=1.522, acc=0.663, acc_len=1.56, thru=11.3, I=12.5, T=28.6, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:01,997] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:10,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:01:11,512] loop.py:45 INFO TRAIN_STEP step=9421 loss=1.317886 Training: 52%|█████▏ | 9421/18018 [20:30:57<22:51:57, 9.58s/step, loss=1.318, acc=0.695, acc_len=1.66, thru=13.5, I=13.1, T=30.6, wait=5.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:01:11,507] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9420: forward=2813.6ms backward=1317.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:11,876] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:01:19,442] loop.py:45 INFO TRAIN_STEP step=9422 loss=1.383088 Training: 52%|█████▏ | 9422/18018 [20:31:05<21:41:11, 9.08s/step, loss=1.383, acc=0.686, acc_len=1.63, thru=15.2, I=19.6, T=29.8, wait=3.6s, pool=49, epoch=2/2][2026-10-19 10:01:29,954] loop.py:45 INFO TRAIN_STEP step=9423 loss=1.427745 Training: 52%|█████▏ | 9423/18018 [20:31:16<22:42:27, 9.51s/step, loss=1.428, acc=0.685, acc_len=1.63, thru=13.6, I=12.7, T=31.1, wait=6.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:36,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:01:37,210] loop.py:45 INFO TRAIN_STEP step=9424 loss=1.310265 Training: 52%|█████▏ | 9424/18018 [20:31:23<21:05:20, 8.83s/step, loss=1.310, acc=0.699, acc_len=1.68, thru=16.2, I=18.5, T=33.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:37,555] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:44,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:01:45,049] loop.py:380 INFO TIMING step=9425: step=4.178s data=1.887s compute=4.144s [fwd=2.948s bwd=1.181s opt=0.003s] dispatch=3.652s [2026-10-19 10:01:45,049] loop.py:45 INFO TRAIN_STEP step=9425 loss=1.436374 Training: 52%|█████▏ | 9425/18018 [20:31:31<20:22:32, 8.54s/step, loss=1.436, acc=0.675, acc_len=1.60, thru=15.1, I=19.4, T=30.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:45,178] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 10:01:52,363] loop.py:45 INFO TRAIN_STEP step=9426 loss=1.468309 Training: 52%|█████▏ | 9426/18018 [20:31:38<19:29:52, 8.17s/step, loss=1.468, acc=0.672, acc_len=1.59, thru=15.6, I=16.2, T=27.1, wait=2.6s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:01:52,359] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9425: forward=3586.4ms backward=1083.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:57,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:01:59,676] loop.py:45 INFO TRAIN_STEP step=9427 loss=1.391849 Training: 52%|█████▏ | 9427/18018 [20:31:45<18:52:58, 7.91s/step, loss=1.392, acc=0.687, acc_len=1.64, thru=16.7, I=20.6, T=28.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:01:59,900] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 10:02:05,882] loop.py:45 INFO TRAIN_STEP step=9428 loss=1.250054 Training: 52%|█████▏ | 9428/18018 [20:31:51<17:39:26, 7.40s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=14.0, I=18.7, T=35.3, wait=2.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:13,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:02:16,158] loop.py:45 INFO TRAIN_STEP step=9429 loss=1.336541 Training: 52%|█████▏ | 9429/18018 [20:32:02<19:42:53, 8.26s/step, loss=1.337, acc=0.695, acc_len=1.67, thru=15.7, I=19.2, T=28.2, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:16,283] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:23,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:02:23,535] loop.py:380 INFO TIMING step=9430: step=4.062s data=1.915s compute=4.032s [fwd=2.812s bwd=1.205s opt=0.003s] dispatch=3.305s [2026-10-19 10:02:23,535] loop.py:45 INFO TRAIN_STEP step=9430 loss=1.332545 Training: 52%|█████▏ | 9430/18018 [20:32:09<19:04:32, 8.00s/step, loss=1.333, acc=0.700, acc_len=1.68, thru=14.1, I=20.4, T=31.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:24,036] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:02:30,407] loop.py:45 INFO TRAIN_STEP step=9431 loss=1.394206 Training: 52%|█████▏ | 9431/18018 [20:32:16<18:16:14, 7.66s/step, loss=1.394, acc=0.687, acc_len=1.64, thru=16.1, I=17.4, T=32.2, wait=2.9s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:02:30,402] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9430: forward=2800.2ms backward=1127.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:38,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:02:40,291] loop.py:45 INFO TRAIN_STEP step=9432 loss=1.442788 Training: 52%|█████▏ | 9432/18018 [20:32:26<19:51:39, 8.33s/step, loss=1.443, acc=0.676, acc_len=1.60, thru=14.4, I=16.2, T=29.1, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:40,366] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:44,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:02:46,207] loop.py:45 INFO TRAIN_STEP step=9433 loss=1.169094 Training: 52%|█████▏ | 9433/18018 [20:32:32<18:07:57, 7.60s/step, loss=1.169, acc=0.723, acc_len=1.76, thru=18.9, I=25.8, T=34.6, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:46,675] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:52,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:02:52,653] loop.py:45 INFO TRAIN_STEP step=9434 loss=1.266824 Training: 52%|█████▏ | 9434/18018 [20:32:38<17:18:04, 7.26s/step, loss=1.267, acc=0.703, acc_len=1.70, thru=19.2, I=24.9, T=33.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:53,098] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:59,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:02:59,563] loop.py:380 INFO TIMING step=9435: step=3.704s data=1.535s compute=3.669s [fwd=2.612s bwd=1.042s opt=0.003s] dispatch=3.199s [2026-10-19 10:02:59,564] loop.py:45 INFO TRAIN_STEP step=9435 loss=1.410585 Training: 52%|█████▏ | 9435/18018 [20:32:45<17:03:13, 7.15s/step, loss=1.411, acc=0.684, acc_len=1.63, thru=19.0, I=20.0, T=34.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:02:59,677] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:05,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:03:05,944] loop.py:45 INFO TRAIN_STEP step=9436 loss=1.301147 Training: 52%|█████▏ | 9436/18018 [20:32:52<16:29:52, 6.92s/step, loss=1.301, acc=0.699, acc_len=1.68, thru=19.3, I=21.1, T=36.1, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:03:05,939] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9435: forward=2525.8ms backward=976.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:06,419] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:13,044] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:03:14,095] loop.py:45 INFO TRAIN_STEP step=9437 loss=1.382220 Training: 52%|█████▏ | 9437/18018 [20:33:00<17:22:34, 7.29s/step, loss=1.382, acc=0.689, acc_len=1.64, thru=15.3, I=18.0, T=32.6, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:14,554] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:20,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:03:21,243] loop.py:45 INFO TRAIN_STEP step=9438 loss=1.452734 Training: 52%|█████▏ | 9438/18018 [20:33:07<17:16:21, 7.25s/step, loss=1.453, acc=0.679, acc_len=1.61, thru=16.7, I=21.0, T=32.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:21,667] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:03:27,334] loop.py:45 INFO TRAIN_STEP step=9439 loss=1.380328 Training: 52%|█████▏ | 9439/18018 [20:33:13<16:26:37, 6.90s/step, loss=1.380, acc=0.685, acc_len=1.63, thru=17.7, I=19.9, T=36.4, wait=2.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:32,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:03:34,307] loop.py:380 INFO TIMING step=9440: step=4.033s data=1.750s compute=4.004s [fwd=2.863s bwd=1.126s opt=0.003s] dispatch=2.931s [2026-10-19 10:03:34,308] loop.py:45 INFO TRAIN_STEP step=9440 loss=1.209805 Training: 52%|█████▏ | 9440/18018 [20:33:20<16:29:45, 6.92s/step, loss=1.210, acc=0.718, acc_len=1.75, thru=19.0, I=21.3, T=31.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:34,339] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 10:03:41,649] loop.py:45 INFO TRAIN_STEP step=9441 loss=1.277581 Training: 52%|█████▏ | 9441/18018 [20:33:27<16:47:28, 7.05s/step, loss=1.278, acc=0.705, acc_len=1.70, thru=16.4, I=19.9, T=32.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:41,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 10:03:41,645] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9440: forward=2774.1ms backward=1152.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:42,125] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:47,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:03:48,456] loop.py:45 INFO TRAIN_STEP step=9442 loss=1.287299 Training: 52%|█████▏ | 9442/18018 [20:33:34<16:37:05, 6.98s/step, loss=1.287, acc=0.704, acc_len=1.70, thru=18.4, I=20.0, T=32.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:03:48,554] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-19 10:03:56,010] loop.py:45 INFO TRAIN_STEP step=9443 loss=1.308473 Training: 52%|█████▏ | 9443/18018 [20:33:42<17:01:52, 7.15s/step, loss=1.308, acc=0.702, acc_len=1.69, thru=15.8, I=16.9, T=33.1, wait=3.7s, pool=61, epoch=2/2][2026-10-19 10:04:02,740] loop.py:45 INFO TRAIN_STEP step=9444 loss=1.480097 Training: 52%|█████▏ | 9444/18018 [20:33:48<16:43:33, 7.02s/step, loss=1.480, acc=0.666, acc_len=1.56, thru=14.1, I=14.6, T=38.1, wait=3.4s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:04:11,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:04:11,994] loop.py:380 INFO TIMING step=9445: step=4.376s data=1.942s compute=4.322s [fwd=3.044s bwd=1.262s opt=0.003s] dispatch=4.870s [2026-10-19 10:04:11,994] loop.py:45 INFO TRAIN_STEP step=9445 loss=1.312158 Training: 52%|█████▏ | 9445/18018 [20:33:58<18:19:12, 7.69s/step, loss=1.312, acc=0.702, acc_len=1.70, thru=18.2, I=16.1, T=29.2, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:04:12,052] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 10:04:20,514] loop.py:45 INFO TRAIN_STEP step=9446 loss=1.351771 Training: 52%|█████▏ | 9446/18018 [20:34:06<18:54:25, 7.94s/step, loss=1.352, acc=0.691, acc_len=1.65, thru=14.2, I=14.1, T=31.3, wait=4.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:04:20,509] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9445: forward=2839.7ms backward=1180.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:04:28,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:04:30,129] loop.py:45 INFO TRAIN_STEP step=9447 loss=1.292372 Training: 52%|█████▏ | 9447/18018 [20:34:16<20:06:08, 8.44s/step, loss=1.292, acc=0.704, acc_len=1.70, thru=14.3, I=17.9, T=31.4, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:04:30,482] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:04:38,328] loop.py:45 INFO TRAIN_STEP step=9448 loss=1.457901 Training: 52%|█████▏ | 9448/18018 [20:34:24<19:55:25, 8.37s/step, loss=1.458, acc=0.677, acc_len=1.61, thru=13.0, I=15.9, T=28.7, wait=3.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:04:46,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:04:46,428] loop.py:45 INFO TRAIN_STEP step=9449 loss=1.333359 Training: 52%|█████▏ | 9449/18018 [20:34:32<19:43:45, 8.29s/step, loss=1.333, acc=0.696, acc_len=1.66, thru=16.8, I=15.6, T=28.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:04:46,736] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:04:52,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:04:53,139] loop.py:380 INFO TIMING step=9450: step=3.773s data=1.591s compute=3.732s [fwd=2.668s bwd=1.050s opt=0.003s] dispatch=2.930s [2026-10-19 10:04:53,139] loop.py:45 INFO TRAIN_STEP step=9450 loss=1.323931 Training: 52%|█████▏ | 9450/18018 [20:34:39<18:36:07, 7.82s/step, loss=1.324, acc=0.694, acc_len=1.66, thru=19.6, I=19.3, T=33.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:04:53,510] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:00,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:05:00,638] loop.py:45 INFO TRAIN_STEP step=9451 loss=1.261484 (TrainerActor pid=9310) [2026-10-19 10:05:00,633] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9450: forward=2583.9ms backward=1005.5ms Training: 52%|█████▏ | 9451/18018 [20:34:46<18:22:24, 7.72s/step, loss=1.261, acc=0.707, acc_len=1.71, thru=17.4, I=20.6, T=35.2, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:01,068] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:05:07,506] loop.py:45 INFO TRAIN_STEP step=9452 loss=1.333309 Training: 52%|█████▏ | 9452/18018 [20:34:53<17:45:41, 7.46s/step, loss=1.333, acc=0.695, acc_len=1.67, thru=17.4, I=21.2, T=33.6, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:16,637] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:05:19,237] loop.py:45 INFO TRAIN_STEP step=9453 loss=1.432592 Training: 52%|█████▏ | 9453/18018 [20:35:05<20:48:20, 8.74s/step, loss=1.433, acc=0.680, acc_len=1.61, thru=13.8, I=14.7, T=26.4, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:19,663] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:26,652] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:05:28,492] loop.py:45 INFO TRAIN_STEP step=9454 loss=1.330628 Training: 52%|█████▏ | 9454/18018 [20:35:14<21:09:58, 8.90s/step, loss=1.331, acc=0.697, acc_len=1.68, thru=14.5, I=16.7, T=29.0, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:28,666] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:05:35,793] loop.py:380 INFO TIMING step=9455: step=3.558s data=1.335s compute=3.524s [fwd=2.516s bwd=0.993s opt=0.003s] dispatch=3.734s [2026-10-19 10:05:35,794] loop.py:45 INFO TRAIN_STEP step=9455 loss=1.293046 Training: 52%|█████▏ | 9455/18018 [20:35:21<20:01:26, 8.42s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=14.8, I=16.2, T=36.0, wait=3.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:41,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:05:43,081] loop.py:45 INFO TRAIN_STEP step=9456 loss=1.195882 Training: 52%|█████▏ | 9456/18018 [20:35:29<19:13:00, 8.08s/step, loss=1.196, acc=0.723, acc_len=1.76, thru=17.9, I=20.4, T=29.4, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:05:43,076] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9455: forward=2950.0ms backward=1360.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:43,535] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:49,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:05:50,134] loop.py:45 INFO TRAIN_STEP step=9457 loss=1.352374 Training: 52%|█████▏ | 9457/18018 [20:35:36<18:28:55, 7.77s/step, loss=1.352, acc=0.691, acc_len=1.65, thru=15.6, I=20.7, T=32.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:50,209] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:55,680] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:05:57,183] loop.py:45 INFO TRAIN_STEP step=9458 loss=1.470795 Training: 52%|█████▏ | 9458/18018 [20:35:43<17:57:46, 7.55s/step, loss=1.471, acc=0.674, acc_len=1.59, thru=17.4, I=20.6, T=32.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:05:57,689] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:06:03,553] loop.py:45 INFO TRAIN_STEP step=9459 loss=1.362070 Training: 52%|█████▏ | 9459/18018 [20:35:49<17:06:57, 7.20s/step, loss=1.362, acc=0.689, acc_len=1.64, thru=17.5, I=20.2, T=34.0, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:06:11,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:06:12,392] loop.py:380 INFO TIMING step=9460: step=4.133s data=1.677s compute=4.100s [fwd=2.954s bwd=1.130s opt=0.003s] dispatch=4.697s [2026-10-19 10:06:12,392] loop.py:45 INFO TRAIN_STEP step=9460 loss=1.378862 Training: 53%|█████▎ | 9460/18018 [20:35:58<18:17:02, 7.69s/step, loss=1.379, acc=0.685, acc_len=1.63, thru=16.0, I=17.1, T=31.0, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:06:12,561] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:06:20,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:06:21,199] loop.py:45 INFO TRAIN_STEP step=9461 loss=1.342841 Training: 53%|█████▎ | 9461/18018 [20:36:07<19:04:40, 8.03s/step, loss=1.343, acc=0.696, acc_len=1.67, thru=14.8, I=16.5, T=29.8, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:06:21,194] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9460: forward=2965.5ms backward=1262.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:06:21,424] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:06:26,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:06:27,685] loop.py:45 INFO TRAIN_STEP step=9462 loss=1.422090 Training: 53%|█████▎ | 9462/18018 [20:36:13<17:58:36, 7.56s/step, loss=1.422, acc=0.681, acc_len=1.62, thru=18.4, I=20.2, T=32.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:06:27,951] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:06:34,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:06:35,877] loop.py:45 INFO TRAIN_STEP step=9463 loss=1.374542 Training: 53%|█████▎ | 9463/18018 [20:36:21<18:25:20, 7.75s/step, loss=1.375, acc=0.688, acc_len=1.64, thru=15.7, I=18.0, T=33.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:06:36,383] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:06:44,355] loop.py:45 INFO TRAIN_STEP step=9464 loss=1.358914 Training: 53%|█████▎ | 9464/18018 [20:36:30<18:56:21, 7.97s/step, loss=1.359, acc=0.692, acc_len=1.65, thru=13.0, I=16.8, T=31.3, wait=4.4s, pool=53, epoch=2/2][2026-10-19 10:06:51,507] loop.py:380 INFO TIMING step=9465: step=4.088s data=1.712s compute=4.057s [fwd=2.971s bwd=1.071s opt=0.003s] dispatch=3.053s [2026-10-19 10:06:51,507] loop.py:45 INFO TRAIN_STEP step=9465 loss=1.295114 Training: 53%|█████▎ | 9465/18018 [20:36:37<18:21:10, 7.72s/step, loss=1.295, acc=0.702, acc_len=1.69, thru=13.8, I=17.1, T=31.3, wait=3.1s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:01,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:07:02,234] loop.py:45 INFO TRAIN_STEP step=9466 loss=1.266560 Training: 53%|█████▎ | 9466/18018 [20:36:48<20:29:24, 8.63s/step, loss=1.267, acc=0.706, acc_len=1.71, thru=17.7, I=12.9, T=33.5, wait=6.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:07:02,230] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9465: forward=2759.4ms backward=1007.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:02,374] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:09,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:07:10,324] loop.py:45 INFO TRAIN_STEP step=9467 loss=1.332908 Training: 53%|█████▎ | 9467/18018 [20:36:56<20:06:24, 8.46s/step, loss=1.333, acc=0.696, acc_len=1.66, thru=15.7, I=16.9, T=31.1, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:10,703] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:15,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:07:17,935] loop.py:45 INFO TRAIN_STEP step=9468 loss=1.392040 Training: 53%|█████▎ | 9468/18018 [20:37:04<19:29:43, 8.21s/step, loss=1.392, acc=0.686, acc_len=1.64, thru=15.2, I=20.0, T=27.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:18,380] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:23,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:07:25,368] loop.py:45 INFO TRAIN_STEP step=9469 loss=1.329261 Training: 53%|█████▎ | 9469/18018 [20:37:11<18:56:28, 7.98s/step, loss=1.329, acc=0.696, acc_len=1.67, thru=14.0, I=20.7, T=28.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:25,406] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:30,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:07:31,704] loop.py:380 INFO TIMING step=9470: step=3.770s data=1.549s compute=3.738s [fwd=2.669s bwd=1.054s opt=0.003s] dispatch=2.555s [2026-10-19 10:07:31,704] loop.py:45 INFO TRAIN_STEP step=9470 loss=1.326874 Training: 53%|█████▎ | 9470/18018 [20:37:17<17:46:13, 7.48s/step, loss=1.327, acc=0.696, acc_len=1.68, thru=18.7, I=22.8, T=34.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:32,048] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:36,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:07:37,898] loop.py:45 INFO TRAIN_STEP step=9471 loss=1.393008 Training: 53%|█████▎ | 9471/18018 [20:37:23<16:50:57, 7.10s/step, loss=1.393, acc=0.681, acc_len=1.62, thru=19.4, I=25.1, T=37.2, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:07:37,893] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9470: forward=2435.2ms backward=963.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:38,189] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:07:48,233] loop.py:45 INFO TRAIN_STEP step=9472 loss=1.313486 Training: 53%|█████▎ | 9472/18018 [20:37:34<19:09:03, 8.07s/step, loss=1.313, acc=0.702, acc_len=1.69, thru=11.4, I=12.3, T=28.4, wait=5.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:53,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:07:55,399] loop.py:45 INFO TRAIN_STEP step=9473 loss=1.323966 Training: 53%|█████▎ | 9473/18018 [20:37:41<18:30:33, 7.80s/step, loss=1.324, acc=0.694, acc_len=1.66, thru=17.2, I=18.8, T=33.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:07:55,473] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:00,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:08:00,889] loop.py:45 INFO TRAIN_STEP step=9474 loss=1.325352 Training: 53%|█████▎ | 9474/18018 [20:37:46<16:51:50, 7.11s/step, loss=1.325, acc=0.698, acc_len=1.68, thru=21.7, I=27.3, T=37.2, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:01,033] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:08,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:08:09,923] loop.py:380 INFO TIMING step=9475: step=4.268s data=1.905s compute=4.234s [fwd=3.036s bwd=1.183s opt=0.003s] dispatch=4.754s [2026-10-19 10:08:09,923] loop.py:45 INFO TRAIN_STEP step=9475 loss=1.311731 Training: 53%|█████▎ | 9475/18018 [20:37:56<18:14:00, 7.68s/step, loss=1.312, acc=0.701, acc_len=1.68, thru=13.3, I=17.1, T=30.0, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:10,269] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:16,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:08:17,138] loop.py:45 INFO TRAIN_STEP step=9476 loss=1.391291 Training: 53%|█████▎ | 9476/18018 [20:38:03<17:53:55, 7.54s/step, loss=1.391, acc=0.683, acc_len=1.62, thru=16.1, I=22.0, T=33.7, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:08:17,133] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9475: forward=2684.3ms backward=1065.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:17,292] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:22,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:08:23,596] loop.py:45 INFO TRAIN_STEP step=9477 loss=1.416462 Training: 53%|█████▎ | 9477/18018 [20:38:09<17:07:27, 7.22s/step, loss=1.416, acc=0.680, acc_len=1.61, thru=19.3, I=22.2, T=34.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:23,999] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:29,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:08:30,465] loop.py:45 INFO TRAIN_STEP step=9478 loss=1.344800 Training: 53%|█████▎ | 9478/18018 [20:38:16<16:52:22, 7.11s/step, loss=1.345, acc=0.692, acc_len=1.65, thru=17.9, I=21.6, T=32.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:30,627] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:36,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:08:37,175] loop.py:45 INFO TRAIN_STEP step=9479 loss=1.231551 Training: 53%|█████▎ | 9479/18018 [20:38:23<16:35:02, 6.99s/step, loss=1.232, acc=0.713, acc_len=1.72, thru=17.3, I=22.3, T=32.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:37,295] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:42,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:08:43,377] loop.py:380 INFO TIMING step=9480: step=3.565s data=1.516s compute=3.528s [fwd=2.539s bwd=0.974s opt=0.003s] dispatch=2.629s [2026-10-19 10:08:43,377] loop.py:45 INFO TRAIN_STEP step=9480 loss=1.412942 Training: 53%|█████▎ | 9480/18018 [20:38:29<16:01:14, 6.76s/step, loss=1.413, acc=0.680, acc_len=1.62, thru=19.2, I=21.8, T=35.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:43,444] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:49,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:08:50,785] loop.py:45 INFO TRAIN_STEP step=9481 loss=1.308905 Training: 53%|█████▎ | 9481/18018 [20:38:36<16:29:04, 6.95s/step, loss=1.309, acc=0.701, acc_len=1.69, thru=16.9, I=20.5, T=33.2, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:08:50,780] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9480: forward=2667.3ms backward=1144.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:08:50,876] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 10:08:57,611] loop.py:45 INFO TRAIN_STEP step=9482 loss=1.296395 Training: 53%|█████▎ | 9482/18018 [20:38:43<16:23:28, 6.91s/step, loss=1.296, acc=0.705, acc_len=1.70, thru=16.9, I=18.6, T=32.6, wait=2.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:09:04,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:09:05,879] loop.py:45 INFO TRAIN_STEP step=9483 loss=1.277656 Training: 53%|█████▎ | 9483/18018 [20:38:51<17:21:18, 7.32s/step, loss=1.278, acc=0.705, acc_len=1.70, thru=18.6, I=18.3, T=28.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:09:06,020] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 [2026-10-19 10:09:11,986] loop.py:45 INFO TRAIN_STEP step=9484 loss=1.361031 Training: 53%|█████▎ | 9484/18018 [20:38:58<16:29:19, 6.96s/step, loss=1.361, acc=0.691, acc_len=1.64, thru=16.2, I=23.2, T=35.0, wait=2.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:09:27,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:09:29,730] loop.py:380 INFO TIMING step=9485: step=4.607s data=1.850s compute=4.575s [fwd=3.196s bwd=1.364s opt=0.003s] dispatch=13.128s [2026-10-19 10:09:29,731] loop.py:45 INFO TRAIN_STEP step=9485 loss=1.540340 Training: 53%|█████▎ | 9485/18018 [20:39:15<24:09:36, 10.19s/step, loss=1.540, acc=0.665, acc_len=1.56, thru=14.6, I=8.9, T=27.8, wait=13.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:09:30,000] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:09:36,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:09:36,964] loop.py:45 INFO TRAIN_STEP step=9486 loss=1.319756 Training: 53%|█████▎ | 9486/18018 [20:39:23<22:03:04, 9.30s/step, loss=1.320, acc=0.701, acc_len=1.68, thru=14.5, I=18.3, T=32.9, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:09:36,959] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9485: forward=2774.8ms backward=1068.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:09:37,424] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:09:43,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:09:44,411] loop.py:45 INFO TRAIN_STEP step=9487 loss=1.302204 Training: 53%|█████▎ | 9487/18018 [20:39:30<20:43:48, 8.75s/step, loss=1.302, acc=0.704, acc_len=1.69, thru=18.8, I=18.8, T=30.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:09:44,808] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:09:50,762] loop.py:45 INFO TRAIN_STEP step=9488 loss=1.371324 Training: 53%|█████▎ | 9488/18018 [20:39:36<19:01:18, 8.03s/step, loss=1.371, acc=0.689, acc_len=1.64, thru=18.1, I=19.9, T=35.5, wait=2.7s, pool=61, epoch=2/2][2026-10-19 10:09:59,428] loop.py:45 INFO TRAIN_STEP step=9489 loss=1.264418 Training: 53%|█████▎ | 9489/18018 [20:39:45<19:28:21, 8.22s/step, loss=1.264, acc=0.709, acc_len=1.72, thru=14.2, I=14.8, T=33.4, wait=4.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:07,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:10:08,626] loop.py:380 INFO TIMING step=9490: step=4.735s data=2.273s compute=4.435s [fwd=3.109s bwd=1.311s opt=0.003s] dispatch=4.455s [2026-10-19 10:10:08,626] loop.py:45 INFO TRAIN_STEP step=9490 loss=1.343238 Training: 53%|█████▎ | 9490/18018 [20:39:54<20:10:06, 8.51s/step, loss=1.343, acc=0.696, acc_len=1.67, thru=15.6, I=15.3, T=27.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:08,831] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:16,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:10:18,256] loop.py:45 INFO TRAIN_STEP step=9491 loss=1.440749 Training: 53%|█████▎ | 9491/18018 [20:40:04<20:57:32, 8.85s/step, loss=1.441, acc=0.678, acc_len=1.61, thru=13.6, I=16.8, T=25.1, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:10:18,251] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9490: forward=3710.4ms backward=1320.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:18,340] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 [2026-10-19 10:10:24,131] loop.py:45 INFO TRAIN_STEP step=9492 loss=1.324064 Training: 53%|█████▎ | 9492/18018 [20:40:10<18:50:33, 7.96s/step, loss=1.324, acc=0.698, acc_len=1.68, thru=15.6, I=22.3, T=34.1, wait=2.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:33,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:10:35,113] loop.py:45 INFO TRAIN_STEP step=9493 loss=1.312350 Training: 53%|█████▎ | 9493/18018 [20:40:21<20:59:23, 8.86s/step, loss=1.312, acc=0.703, acc_len=1.69, thru=14.8, I=14.1, T=30.1, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:35,553] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:10:41,132] loop.py:45 INFO TRAIN_STEP step=9494 loss=1.353848 Training: 53%|█████▎ | 9494/18018 [20:40:27<18:58:03, 8.01s/step, loss=1.354, acc=0.694, acc_len=1.66, thru=18.0, I=22.9, T=35.6, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:47,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:10:49,000] loop.py:380 INFO TIMING step=9495: step=4.012s data=1.793s compute=3.974s [fwd=2.819s bwd=1.139s opt=0.003s] dispatch=3.846s [2026-10-19 10:10:49,000] loop.py:45 INFO TRAIN_STEP step=9495 loss=1.294619 Training: 53%|█████▎ | 9495/18018 [20:40:35<18:51:43, 7.97s/step, loss=1.295, acc=0.703, acc_len=1.69, thru=17.9, I=19.7, T=31.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:49,368] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:53,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:10:55,040] loop.py:45 INFO TRAIN_STEP step=9496 loss=1.243937 Training: 53%|█████▎ | 9496/18018 [20:40:41<17:29:34, 7.39s/step, loss=1.244, acc=0.713, acc_len=1.73, thru=19.5, I=27.8, T=32.9, wait=2.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:10:55,035] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9495: forward=2714.0ms backward=1133.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:10:55,503] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:01,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:11:02,185] loop.py:45 INFO TRAIN_STEP step=9497 loss=1.324176 Training: 53%|█████▎ | 9497/18018 [20:40:48<17:19:03, 7.32s/step, loss=1.324, acc=0.695, acc_len=1.66, thru=16.5, I=22.6, T=34.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:02,408] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:11:10,125] loop.py:45 INFO TRAIN_STEP step=9498 loss=1.399419 Training: 53%|█████▎ | 9498/18018 [20:40:56<17:45:29, 7.50s/step, loss=1.399, acc=0.683, acc_len=1.62, thru=15.2, I=18.0, T=28.5, wait=3.4s, pool=57, epoch=2/2][2026-10-19 10:11:18,299] loop.py:45 INFO TRAIN_STEP step=9499 loss=1.427129 Training: 53%|█████▎ | 9499/18018 [20:41:04<18:14:03, 7.71s/step, loss=1.427, acc=0.682, acc_len=1.62, thru=13.8, I=13.2, T=30.6, wait=4.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:27,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:11:28,042] loop.py:380 INFO TIMING step=9500: step=3.947s data=1.706s compute=3.918s [fwd=2.834s bwd=1.069s opt=0.003s] dispatch=5.783s [2026-10-19 10:11:28,042] loop.py:45 INFO TRAIN_STEP step=9500 loss=1.381617 Training: 53%|█████▎ | 9500/18018 [20:41:14<19:40:36, 8.32s/step, loss=1.382, acc=0.687, acc_len=1.64, thru=15.3, I=15.4, T=32.4, wait=5.8s, pool=65, epoch=2/2][2026-10-19 10:11:28,047] loop.py:419 INFO Saving checkpoint at step 9500... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 10:11:31,482] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009501 (TrainerActor pid=9310) [2026-10-19 10:11:31,487] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:32,206] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:33,212] inference_manager.py:408 INFO Pool capacity available after 6.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:38,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 10:11:31,481] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 10:11:32,725] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009001 [2026-10-19 10:11:40,105] loop.py:45 INFO TRAIN_STEP step=9501 loss=1.346829 Training: 53%|█████▎ | 9501/18018 [20:41:26<22:19:51, 9.44s/step, loss=1.347, acc=0.693, acc_len=1.67, thru=18.6, I=22.5, T=28.7, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:11:40,100] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9500: forward=3104.8ms backward=1302.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:40,467] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:45,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:11:46,558] loop.py:45 INFO TRAIN_STEP step=9502 loss=1.352342 Training: 53%|█████▎ | 9502/18018 [20:41:32<20:12:40, 8.54s/step, loss=1.352, acc=0.693, acc_len=1.66, thru=18.0, I=23.3, T=34.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:46,590] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:53,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:11:54,660] loop.py:45 INFO TRAIN_STEP step=9503 loss=1.380587 Training: 53%|█████▎ | 9503/18018 [20:41:40<19:53:45, 8.41s/step, loss=1.381, acc=0.689, acc_len=1.65, thru=14.8, I=17.4, T=30.7, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:11:54,923] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:01,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:12:01,699] loop.py:45 INFO TRAIN_STEP step=9504 loss=1.444172 Training: 53%|█████▎ | 9504/18018 [20:41:47<18:55:11, 8.00s/step, loss=1.444, acc=0.676, acc_len=1.60, thru=16.3, I=18.4, T=33.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:01,975] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:07,892] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:12:09,758] loop.py:380 INFO TIMING step=9505: step=4.185s data=1.886s compute=4.127s [fwd=2.867s bwd=1.244s opt=0.003s] dispatch=3.864s [2026-10-19 10:12:09,758] loop.py:45 INFO TRAIN_STEP step=9505 loss=1.377984 Training: 53%|█████▎ | 9505/18018 [20:41:55<18:57:32, 8.02s/step, loss=1.378, acc=0.687, acc_len=1.63, thru=15.9, I=18.5, T=30.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:09,921] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:16,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:12:17,168] loop.py:45 INFO TRAIN_STEP step=9506 loss=1.435674 Training: 53%|█████▎ | 9506/18018 [20:42:03<18:31:29, 7.83s/step, loss=1.436, acc=0.680, acc_len=1.61, thru=15.1, I=19.6, T=34.0, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:12:17,163] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9505: forward=2670.3ms backward=1046.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:17,569] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:24,705] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:12:24,947] loop.py:45 INFO TRAIN_STEP step=9507 loss=1.256859 Training: 53%|█████▎ | 9507/18018 [20:42:11<18:29:02, 7.82s/step, loss=1.257, acc=0.705, acc_len=1.70, thru=16.5, I=18.6, T=30.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:25,206] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:30,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:12:31,169] loop.py:45 INFO TRAIN_STEP step=9508 loss=1.446239 Training: 53%|█████▎ | 9508/18018 [20:42:17<17:20:58, 7.34s/step, loss=1.446, acc=0.679, acc_len=1.62, thru=19.8, I=20.7, T=33.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:31,551] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:37,377] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:12:37,939] loop.py:45 INFO TRAIN_STEP step=9509 loss=1.395812 Training: 53%|█████▎ | 9509/18018 [20:42:24<16:56:41, 7.17s/step, loss=1.396, acc=0.692, acc_len=1.65, thru=18.5, I=22.6, T=35.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:38,381] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:43,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:12:44,808] loop.py:380 INFO TIMING step=9510: step=3.975s data=1.642s compute=3.945s [fwd=2.775s bwd=1.154s opt=0.003s] dispatch=2.882s [2026-10-19 10:12:44,808] loop.py:45 INFO TRAIN_STEP step=9510 loss=1.279114 Training: 53%|█████▎ | 9510/18018 [20:42:30<16:43:48, 7.08s/step, loss=1.279, acc=0.700, acc_len=1.68, thru=19.2, I=22.6, T=32.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:12:45,245] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:12:52,623] loop.py:45 INFO TRAIN_STEP step=9511 loss=1.435766 Training: 53%|█████▎ | 9511/18018 [20:42:38<17:14:57, 7.30s/step, loss=1.436, acc=0.679, acc_len=1.61, thru=13.1, I=18.2, T=34.0, wait=4.0s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:12:52,618] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9510: forward=2625.0ms backward=1098.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:01,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:13:02,198] loop.py:45 INFO TRAIN_STEP step=9512 loss=1.413756 Training: 53%|█████▎ | 9512/18018 [20:42:48<18:51:32, 7.98s/step, loss=1.414, acc=0.680, acc_len=1.62, thru=14.9, I=15.7, T=32.0, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:02,572] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:13:11,614] loop.py:45 INFO TRAIN_STEP step=9513 loss=1.256418 Training: 53%|█████▎ | 9513/18018 [20:42:57<19:52:23, 8.41s/step, loss=1.256, acc=0.712, acc_len=1.72, thru=13.7, I=13.9, T=32.8, wait=5.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:19,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:13:19,765] loop.py:45 INFO TRAIN_STEP step=9514 loss=1.364743 Training: 53%|█████▎ | 9514/18018 [20:43:05<19:41:15, 8.33s/step, loss=1.365, acc=0.688, acc_len=1.64, thru=17.0, I=16.7, T=33.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:20,090] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:13:27,585] loop.py:380 INFO TIMING step=9515: step=4.034s data=1.814s compute=3.989s [fwd=2.881s bwd=1.092s opt=0.003s] dispatch=3.775s [2026-10-19 10:13:27,586] loop.py:45 INFO TRAIN_STEP step=9515 loss=1.434512 Training: 53%|█████▎ | 9515/18018 [20:43:13<19:19:07, 8.18s/step, loss=1.435, acc=0.678, acc_len=1.61, thru=15.7, I=16.2, T=31.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:27,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:28,040] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:34,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:13:35,411] loop.py:45 INFO TRAIN_STEP step=9516 loss=1.397991 (TrainerActor pid=9310) [2026-10-19 10:13:35,405] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9515: forward=2912.2ms backward=1165.9ms Training: 53%|█████▎ | 9516/18018 [20:43:21<19:04:00, 8.07s/step, loss=1.398, acc=0.681, acc_len=1.62, thru=17.5, I=17.9, T=30.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:35,668] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:40,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:13:41,630] loop.py:45 INFO TRAIN_STEP step=9517 loss=1.265393 Training: 53%|█████▎ | 9517/18018 [20:43:27<17:45:02, 7.52s/step, loss=1.265, acc=0.708, acc_len=1.71, thru=18.6, I=23.6, T=32.0, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:41,757] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:47,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:13:48,119] loop.py:45 INFO TRAIN_STEP step=9518 loss=1.301063 Training: 53%|█████▎ | 9518/18018 [20:43:34<17:01:14, 7.21s/step, loss=1.301, acc=0.704, acc_len=1.70, thru=18.2, I=23.0, T=32.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:48,403] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:54,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:13:56,341] loop.py:45 INFO TRAIN_STEP step=9519 loss=1.349849 Training: 53%|█████▎ | 9519/18018 [20:43:42<17:44:13, 7.51s/step, loss=1.350, acc=0.692, acc_len=1.65, thru=15.2, I=18.3, T=33.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:13:56,501] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:14:02,352] loop.py:380 INFO TIMING step=9520: step=3.279s data=1.242s compute=3.248s [fwd=2.322s bwd=0.911s opt=0.003s] dispatch=2.722s [2026-10-19 10:14:02,353] loop.py:45 INFO TRAIN_STEP step=9520 loss=1.304029 Training: 53%|█████▎ | 9520/18018 [20:43:48<16:40:24, 7.06s/step, loss=1.304, acc=0.701, acc_len=1.69, thru=19.0, I=20.7, T=39.0, wait=2.7s, pool=57, epoch=2/2][2026-10-19 10:14:09,688] loop.py:45 INFO TRAIN_STEP step=9521 loss=1.427939 Training: 53%|█████▎ | 9521/18018 [20:43:55<16:51:44, 7.14s/step, loss=1.428, acc=0.683, acc_len=1.62, thru=16.6, I=17.1, T=36.7, wait=3.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:14:09,684] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9520: forward=2473.2ms backward=975.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:14:17,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:14:19,472] loop.py:45 INFO TRAIN_STEP step=9522 loss=1.441815 Training: 53%|█████▎ | 9522/18018 [20:44:05<18:43:43, 7.94s/step, loss=1.442, acc=0.675, acc_len=1.60, thru=14.9, I=16.6, T=23.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:14:19,503] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:14:24,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:14:26,303] loop.py:45 INFO TRAIN_STEP step=9523 loss=1.321035 Training: 53%|█████▎ | 9523/18018 [20:44:12<17:56:44, 7.61s/step, loss=1.321, acc=0.694, acc_len=1.66, thru=15.4, I=19.0, T=31.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:14:26,505] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:14:33,578] loop.py:45 INFO TRAIN_STEP step=9524 loss=1.275426 Training: 53%|█████▎ | 9524/18018 [20:44:19<17:42:33, 7.51s/step, loss=1.275, acc=0.704, acc_len=1.69, thru=14.0, I=17.1, T=32.0, wait=3.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:14:46,241] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:14:47,515] loop.py:380 INFO TIMING step=9525: step=4.851s data=1.700s compute=4.812s [fwd=3.271s bwd=1.527s opt=0.003s] dispatch=9.076s [2026-10-19 10:14:47,515] loop.py:45 INFO TRAIN_STEP step=9525 loss=1.318377 Training: 53%|█████▎ | 9525/18018 [20:44:33<22:15:29, 9.43s/step, loss=1.318, acc=0.697, acc_len=1.67, thru=10.7, I=11.4, T=26.4, wait=9.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:14:47,751] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:14:52,916] loop.py:45 INFO TRAIN_STEP step=9526 loss=1.323030 Training: 53%|█████▎ | 9526/18018 [20:44:39<19:24:05, 8.22s/step, loss=1.323, acc=0.696, acc_len=1.67, thru=18.1, I=21.1, T=38.3, wait=2.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:14:52,912] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9525: forward=2375.5ms backward=922.0ms [2026-10-19 10:15:01,200] loop.py:45 INFO TRAIN_STEP step=9527 loss=1.458502 Training: 53%|█████▎ | 9527/18018 [20:44:47<19:26:25, 8.24s/step, loss=1.459, acc=0.673, acc_len=1.59, thru=15.8, I=15.3, T=36.1, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:01,168] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:01,671] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:07,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:15:07,377] loop.py:45 INFO TRAIN_STEP step=9528 loss=1.272193 Training: 53%|█████▎ | 9528/18018 [20:44:53<17:58:35, 7.62s/step, loss=1.272, acc=0.704, acc_len=1.70, thru=20.4, I=19.3, T=34.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:07,773] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:13,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:15:14,105] loop.py:45 INFO TRAIN_STEP step=9529 loss=1.335650 Training: 53%|█████▎ | 9529/18018 [20:45:00<17:20:34, 7.35s/step, loss=1.336, acc=0.692, acc_len=1.65, thru=19.0, I=20.5, T=33.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:14,161] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:18,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:15:20,405] loop.py:380 INFO TIMING step=9530: step=4.270s data=2.021s compute=4.233s [fwd=2.967s bwd=1.252s opt=0.003s] dispatch=2.019s [2026-10-19 10:15:20,405] loop.py:45 INFO TRAIN_STEP step=9530 loss=1.273404 Training: 53%|█████▎ | 9530/18018 [20:45:06<16:35:40, 7.04s/step, loss=1.273, acc=0.706, acc_len=1.70, thru=20.0, I=23.8, T=30.0, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:20,827] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:26,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:15:26,885] loop.py:45 INFO TRAIN_STEP step=9531 loss=1.287424 Training: 53%|█████▎ | 9531/18018 [20:45:12<16:11:53, 6.87s/step, loss=1.287, acc=0.702, acc_len=1.69, thru=17.6, I=23.8, T=37.6, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:15:26,879] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9530: forward=2431.3ms backward=917.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:27,100] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:34,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:15:35,231] loop.py:45 INFO TRAIN_STEP step=9532 loss=1.275735 Training: 53%|█████▎ | 9532/18018 [20:45:21<17:14:20, 7.31s/step, loss=1.276, acc=0.705, acc_len=1.70, thru=14.9, I=18.4, T=34.1, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:35,684] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:40,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:15:41,654] loop.py:45 INFO TRAIN_STEP step=9533 loss=1.284735 Training: 53%|█████▎ | 9533/18018 [20:45:27<16:36:20, 7.05s/step, loss=1.285, acc=0.705, acc_len=1.70, thru=18.9, I=24.3, T=37.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:15:42,119] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:15:48,751] loop.py:45 INFO TRAIN_STEP step=9534 loss=1.323439 Training: 53%|█████▎ | 9534/18018 [20:45:34<16:38:33, 7.06s/step, loss=1.323, acc=0.696, acc_len=1.67, thru=17.6, I=20.6, T=34.4, wait=3.4s, pool=61, epoch=2/2][2026-10-19 10:15:56,471] loop.py:380 INFO TIMING step=9535: step=3.699s data=1.492s compute=3.670s [fwd=2.640s bwd=1.014s opt=0.003s] dispatch=4.008s [2026-10-19 10:15:56,471] loop.py:45 INFO TRAIN_STEP step=9535 loss=1.341361 Training: 53%|█████▎ | 9535/18018 [20:45:42<17:06:22, 7.26s/step, loss=1.341, acc=0.692, acc_len=1.66, thru=15.5, I=16.3, T=34.6, wait=4.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:04,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:16:06,308] loop.py:45 INFO TRAIN_STEP step=9536 loss=1.464051 Training: 53%|█████▎ | 9536/18018 [20:45:52<18:55:25, 8.03s/step, loss=1.464, acc=0.670, acc_len=1.58, thru=15.5, I=16.5, T=29.7, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:16:06,303] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9535: forward=2966.3ms backward=1285.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:06,798] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:11,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:16:12,282] loop.py:45 INFO TRAIN_STEP step=9537 loss=1.293090 Training: 53%|█████▎ | 9537/18018 [20:45:58<17:28:08, 7.42s/step, loss=1.293, acc=0.701, acc_len=1.69, thru=18.3, I=25.5, T=34.5, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:12,319] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:18,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:16:20,533] loop.py:45 INFO TRAIN_STEP step=9538 loss=1.385552 Training: 53%|█████▎ | 9538/18018 [20:46:06<18:03:22, 7.67s/step, loss=1.386, acc=0.687, acc_len=1.64, thru=14.7, I=19.3, T=28.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:20,972] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:27,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:16:27,203] loop.py:45 INFO TRAIN_STEP step=9539 loss=1.301684 Training: 53%|█████▎ | 9539/18018 [20:46:13<17:21:07, 7.37s/step, loss=1.302, acc=0.700, acc_len=1.68, thru=18.1, I=24.4, T=31.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:27,597] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:37,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:16:37,422] loop.py:380 INFO TIMING step=9540: step=4.227s data=1.875s compute=4.197s [fwd=3.044s bwd=1.138s opt=0.003s] dispatch=5.982s [2026-10-19 10:16:37,422] loop.py:45 INFO TRAIN_STEP step=9540 loss=1.351958 Training: 53%|█████▎ | 9540/18018 [20:46:23<19:21:53, 8.22s/step, loss=1.352, acc=0.692, acc_len=1.66, thru=13.8, I=12.4, T=30.3, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:37,581] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:42,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:16:43,934] loop.py:45 INFO TRAIN_STEP step=9541 loss=1.341285 Training: 53%|█████▎ | 9541/18018 [20:46:30<18:09:07, 7.71s/step, loss=1.341, acc=0.693, acc_len=1.66, thru=18.7, I=21.1, T=29.9, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:16:43,929] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9540: forward=2956.2ms backward=1268.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:44,048] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:50,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:16:50,569] loop.py:45 INFO TRAIN_STEP step=9542 loss=1.333509 Training: 53%|█████▎ | 9542/18018 [20:46:36<17:23:30, 7.39s/step, loss=1.334, acc=0.692, acc_len=1.65, thru=18.1, I=21.1, T=34.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:51,031] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:58,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:16:59,567] loop.py:45 INFO TRAIN_STEP step=9543 loss=1.357651 Training: 53%|█████▎ | 9543/18018 [20:46:45<18:31:39, 7.87s/step, loss=1.358, acc=0.690, acc_len=1.65, thru=14.4, I=18.4, T=30.1, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:16:59,800] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:17:06,353] loop.py:45 INFO TRAIN_STEP step=9544 loss=1.213339 Training: 53%|█████▎ | 9544/18018 [20:46:52<17:45:44, 7.55s/step, loss=1.213, acc=0.715, acc_len=1.73, thru=14.5, I=19.3, T=34.2, wait=3.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:15,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:17:16,124] loop.py:380 INFO TIMING step=9545: step=4.069s data=1.752s compute=4.041s [fwd=2.944s bwd=1.083s opt=0.003s] dispatch=5.691s [2026-10-19 10:17:16,125] loop.py:45 INFO TRAIN_STEP step=9545 loss=1.314037 Training: 53%|█████▎ | 9545/18018 [20:47:02<19:19:48, 8.21s/step, loss=1.314, acc=0.699, acc_len=1.68, thru=15.7, I=15.4, T=31.5, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:16,468] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:21,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:17:22,025] loop.py:45 INFO TRAIN_STEP step=9546 loss=1.346187 Training: 53%|█████▎ | 9546/18018 [20:47:08<17:41:45, 7.52s/step, loss=1.346, acc=0.694, acc_len=1.66, thru=19.4, I=23.2, T=34.3, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:17:22,020] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9545: forward=2661.3ms backward=1018.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:22,233] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:30,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:17:31,858] loop.py:45 INFO TRAIN_STEP step=9547 loss=1.337143 Training: 53%|█████▎ | 9547/18018 [20:47:17<19:19:31, 8.21s/step, loss=1.337, acc=0.696, acc_len=1.67, thru=13.3, I=14.4, T=32.2, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:31,925] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:38,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:17:39,898] loop.py:45 INFO TRAIN_STEP step=9548 loss=1.238553 Training: 53%|█████▎ | 9548/18018 [20:47:25<19:12:01, 8.16s/step, loss=1.239, acc=0.711, acc_len=1.72, thru=15.3, I=19.1, T=33.7, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:40,302] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:46,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:17:47,397] loop.py:45 INFO TRAIN_STEP step=9549 loss=1.480074 Training: 53%|█████▎ | 9549/18018 [20:47:33<18:43:53, 7.96s/step, loss=1.480, acc=0.670, acc_len=1.58, thru=16.5, I=20.4, T=34.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:47,824] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:55,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:17:56,080] loop.py:380 INFO TIMING step=9550: step=4.520s data=2.081s compute=4.489s [fwd=3.153s bwd=1.320s opt=0.003s] dispatch=4.155s [2026-10-19 10:17:56,080] loop.py:45 INFO TRAIN_STEP step=9550 loss=1.433447 Training: 53%|█████▎ | 9550/18018 [20:47:42<19:14:22, 8.18s/step, loss=1.433, acc=0.675, acc_len=1.59, thru=15.5, I=18.0, T=28.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:17:56,484] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:18:04,070] loop.py:45 INFO TRAIN_STEP step=9551 loss=1.207853 Training: 53%|█████▎ | 9551/18018 [20:47:50<19:06:06, 8.12s/step, loss=1.208, acc=0.715, acc_len=1.73, thru=14.9, I=14.0, T=31.1, wait=3.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:18:04,066] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9550: forward=2819.6ms backward=1238.0ms [2026-10-19 10:18:15,042] loop.py:45 INFO TRAIN_STEP step=9552 loss=1.453591 Training: 53%|█████▎ | 9552/18018 [20:48:01<21:06:40, 8.98s/step, loss=1.454, acc=0.676, acc_len=1.60, thru=11.6, I=12.9, T=30.8, wait=6.8s, pool=61, epoch=2/2][2026-10-19 10:18:23,465] loop.py:45 INFO TRAIN_STEP step=9553 loss=1.267967 Training: 53%|█████▎ | 9553/18018 [20:48:09<20:43:02, 8.81s/step, loss=1.268, acc=0.705, acc_len=1.70, thru=14.8, I=15.3, T=29.0, wait=4.0s, pool=61, epoch=2/2][2026-10-19 10:18:30,642] loop.py:45 INFO TRAIN_STEP step=9554 loss=1.349338 Training: 53%|█████▎ | 9554/18018 [20:48:16<19:33:44, 8.32s/step, loss=1.349, acc=0.692, acc_len=1.66, thru=14.7, I=14.1, T=33.7, wait=3.4s, pool=45, epoch=2/2][2026-10-19 10:18:38,430] loop.py:380 INFO TIMING step=9555: step=3.803s data=1.592s compute=3.776s [fwd=2.731s bwd=1.029s opt=0.003s] dispatch=3.978s [2026-10-19 10:18:38,430] loop.py:45 INFO TRAIN_STEP step=9555 loss=1.402983 Training: 53%|█████▎ | 9555/18018 [20:48:24<19:11:04, 8.16s/step, loss=1.403, acc=0.682, acc_len=1.62, thru=16.8, I=15.0, T=33.7, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:18:47,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:18:47,735] loop.py:45 INFO TRAIN_STEP step=9556 loss=1.341870 (TrainerActor pid=9310) [2026-10-19 10:18:47,731] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9555: forward=2745.8ms backward=998.2ms Training: 53%|█████▎ | 9556/18018 [20:48:33<19:59:21, 8.50s/step, loss=1.342, acc=0.695, acc_len=1.66, thru=15.9, I=14.9, T=33.8, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:18:47,988] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:18:55,043] loop.py:45 INFO TRAIN_STEP step=9557 loss=1.460625 Training: 53%|█████▎ | 9557/18018 [20:48:41<19:08:36, 8.15s/step, loss=1.461, acc=0.674, acc_len=1.59, thru=17.7, I=16.2, T=32.0, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:01,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:19:03,037] loop.py:45 INFO TRAIN_STEP step=9558 loss=1.292425 Training: 53%|█████▎ | 9558/18018 [20:48:49<19:02:09, 8.10s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=17.1, I=17.5, T=33.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:03,374] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:09,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:19:09,498] loop.py:45 INFO TRAIN_STEP step=9559 loss=1.454414 Training: 53%|█████▎ | 9559/18018 [20:48:55<17:52:42, 7.61s/step, loss=1.454, acc=0.672, acc_len=1.58, thru=18.8, I=24.0, T=32.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:09,930] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:17,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:19:17,721] loop.py:380 INFO TIMING step=9560: step=5.141s data=2.296s compute=5.109s [fwd=3.461s bwd=1.632s opt=0.003s] dispatch=3.071s [2026-10-19 10:19:17,721] loop.py:45 INFO TRAIN_STEP step=9560 loss=1.497926 Training: 53%|█████▎ | 9560/18018 [20:49:03<18:18:26, 7.79s/step, loss=1.498, acc=0.667, acc_len=1.57, thru=15.4, I=15.4, T=24.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:17,981] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:24,744] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:19:26,807] loop.py:45 INFO TRAIN_STEP step=9561 loss=1.412595 Training: 53%|█████▎ | 9561/18018 [20:49:12<19:13:07, 8.18s/step, loss=1.413, acc=0.685, acc_len=1.63, thru=14.5, I=17.4, T=26.4, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:19:26,801] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9560: forward=3344.1ms backward=1438.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:27,264] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:33,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:19:33,577] loop.py:45 INFO TRAIN_STEP step=9562 loss=1.343082 Training: 53%|█████▎ | 9562/18018 [20:49:19<18:13:20, 7.76s/step, loss=1.343, acc=0.696, acc_len=1.66, thru=16.5, I=21.6, T=33.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:33,948] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:19:41,643] loop.py:45 INFO TRAIN_STEP step=9563 loss=1.446797 Training: 53%|█████▎ | 9563/18018 [20:49:27<18:26:14, 7.85s/step, loss=1.447, acc=0.674, acc_len=1.59, thru=15.9, I=16.3, T=33.3, wait=4.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:49,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:19:49,966] loop.py:45 INFO TRAIN_STEP step=9564 loss=1.310239 Training: 53%|█████▎ | 9564/18018 [20:49:36<18:45:59, 7.99s/step, loss=1.310, acc=0.698, acc_len=1.68, thru=15.8, I=17.4, T=32.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:50,207] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:55,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:19:56,055] loop.py:380 INFO TIMING step=9565: step=3.873s data=1.683s compute=3.845s [fwd=2.702s bwd=1.128s opt=0.003s] dispatch=2.208s [2026-10-19 10:19:56,055] loop.py:45 INFO TRAIN_STEP step=9565 loss=1.411006 Training: 53%|█████▎ | 9565/18018 [20:49:42<17:25:31, 7.42s/step, loss=1.411, acc=0.684, acc_len=1.62, thru=20.0, I=22.0, T=33.1, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:19:56,128] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:02,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:20:02,762] loop.py:45 INFO TRAIN_STEP step=9566 loss=1.338570 Training: 53%|█████▎ | 9566/18018 [20:49:48<16:55:06, 7.21s/step, loss=1.339, acc=0.695, acc_len=1.66, thru=18.2, I=20.1, T=31.7, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:20:02,758] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9565: forward=2825.4ms backward=1166.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:02,846] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 10:20:11,387] loop.py:45 INFO TRAIN_STEP step=9567 loss=1.312843 Training: 53%|█████▎ | 9567/18018 [20:49:57<17:55:05, 7.63s/step, loss=1.313, acc=0.698, acc_len=1.67, thru=14.7, I=17.9, T=32.7, wait=4.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:19,213] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:20:19,772] loop.py:45 INFO TRAIN_STEP step=9568 loss=1.372569 Training: 53%|█████▎ | 9568/18018 [20:50:05<18:26:41, 7.86s/step, loss=1.373, acc=0.689, acc_len=1.65, thru=15.2, I=14.9, T=31.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:20,219] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:24,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:20:26,354] loop.py:45 INFO TRAIN_STEP step=9569 loss=1.466019 Training: 53%|█████▎ | 9569/18018 [20:50:12<17:32:40, 7.48s/step, loss=1.466, acc=0.673, acc_len=1.59, thru=18.3, I=21.7, T=34.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:26,464] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:32,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:20:34,687] loop.py:380 INFO TIMING step=9570: step=4.311s data=1.693s compute=4.276s [fwd=3.036s bwd=1.225s opt=0.003s] dispatch=4.012s [2026-10-19 10:20:34,687] loop.py:45 INFO TRAIN_STEP step=9570 loss=1.285777 Training: 53%|█████▎ | 9570/18018 [20:50:20<18:08:41, 7.73s/step, loss=1.286, acc=0.704, acc_len=1.69, thru=14.2, I=19.3, T=29.7, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:34,854] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:40,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:20:41,195] loop.py:45 INFO TRAIN_STEP step=9571 loss=1.271654 Training: 53%|█████▎ | 9571/18018 [20:50:27<17:16:55, 7.37s/step, loss=1.272, acc=0.705, acc_len=1.70, thru=17.9, I=22.8, T=33.7, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:20:41,191] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9570: forward=2640.6ms backward=1100.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:41,513] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:47,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:20:48,185] loop.py:45 INFO TRAIN_STEP step=9572 loss=1.391848 Training: 53%|█████▎ | 9572/18018 [20:50:34<17:00:56, 7.25s/step, loss=1.392, acc=0.686, acc_len=1.63, thru=18.5, I=20.3, T=34.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:20:48,220] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 10:20:55,702] loop.py:45 INFO TRAIN_STEP step=9573 loss=1.251615 Training: 53%|█████▎ | 9573/18018 [20:50:41<17:11:59, 7.33s/step, loss=1.252, acc=0.709, acc_len=1.72, thru=14.9, I=16.7, T=35.6, wait=3.9s, pool=49, epoch=2/2][2026-10-19 10:21:05,898] loop.py:45 INFO TRAIN_STEP step=9574 loss=1.295011 Training: 53%|█████▎ | 9574/18018 [20:50:51<19:12:42, 8.19s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=13.1, I=13.1, T=32.0, wait=6.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:21:12,140] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:21:12,782] loop.py:380 INFO TIMING step=9575: step=3.884s data=1.726s compute=3.845s [fwd=2.750s bwd=1.079s opt=0.003s] dispatch=2.992s [2026-10-19 10:21:12,782] loop.py:45 INFO TRAIN_STEP step=9575 loss=1.331473 Training: 53%|█████▎ | 9575/18018 [20:50:58<18:17:24, 7.80s/step, loss=1.331, acc=0.696, acc_len=1.67, thru=18.9, I=18.9, T=33.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:21:13,145] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:21:17,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:21:19,802] loop.py:45 INFO TRAIN_STEP step=9576 loss=1.351438 Training: 53%|█████▎ | 9576/18018 [20:51:05<17:44:31, 7.57s/step, loss=1.351, acc=0.690, acc_len=1.65, thru=17.2, I=21.9, T=30.7, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:21:19,797] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9575: forward=2708.6ms backward=1409.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:21:20,008] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:21:25,615] loop.py:45 INFO TRAIN_STEP step=9577 loss=1.324765 Training: 53%|█████▎ | 9577/18018 [20:51:11<16:30:21, 7.04s/step, loss=1.325, acc=0.697, acc_len=1.67, thru=17.0, I=22.7, T=36.8, wait=2.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:21:32,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:21:34,601] loop.py:45 INFO TRAIN_STEP step=9578 loss=1.424761 Training: 53%|█████▎ | 9578/18018 [20:51:20<17:52:25, 7.62s/step, loss=1.425, acc=0.683, acc_len=1.63, thru=17.2, I=18.1, T=27.8, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:21:34,696] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 [2026-10-19 10:21:42,572] loop.py:45 INFO TRAIN_STEP step=9579 loss=1.431085 Training: 53%|█████▎ | 9579/18018 [20:51:28<18:06:54, 7.73s/step, loss=1.431, acc=0.679, acc_len=1.61, thru=11.3, I=13.8, T=29.2, wait=3.6s, pool=45, epoch=2/2][2026-10-19 10:21:49,531] loop.py:380 INFO TIMING step=9580: step=3.789s data=1.560s compute=3.738s [fwd=2.537s bwd=1.186s opt=0.003s] dispatch=3.160s [2026-10-19 10:21:49,531] loop.py:45 INFO TRAIN_STEP step=9580 loss=1.264099 Training: 53%|█████▎ | 9580/18018 [20:51:35<17:34:13, 7.50s/step, loss=1.264, acc=0.708, acc_len=1.71, thru=18.1, I=17.4, T=33.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:21:49,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:21:50,014] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:21:58,303] loop.py:45 INFO TRAIN_STEP step=9581 loss=1.295583 Training: 53%|█████▎ | 9581/18018 [20:51:44<18:27:59, 7.88s/step, loss=1.296, acc=0.703, acc_len=1.69, thru=13.8, I=13.2, T=35.4, wait=5.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:21:58,299] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9580: forward=2549.7ms backward=1020.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:03,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:22:05,875] loop.py:45 INFO TRAIN_STEP step=9582 loss=1.333537 Training: 53%|█████▎ | 9582/18018 [20:51:51<18:14:53, 7.79s/step, loss=1.334, acc=0.695, acc_len=1.67, thru=18.3, I=22.7, T=28.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:05,905] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:12,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:22:13,267] loop.py:45 INFO TRAIN_STEP step=9583 loss=1.408985 Training: 53%|█████▎ | 9583/18018 [20:51:59<17:58:10, 7.67s/step, loss=1.409, acc=0.681, acc_len=1.62, thru=13.3, I=20.5, T=32.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:13,597] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:20,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:22:21,148] loop.py:45 INFO TRAIN_STEP step=9584 loss=1.295519 Training: 53%|█████▎ | 9584/18018 [20:52:07<18:06:58, 7.73s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=15.4, I=18.7, T=28.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:21,300] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:26,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:22:27,663] loop.py:380 INFO TIMING step=9585: step=4.029s data=1.727s compute=3.988s [fwd=2.825s bwd=1.147s opt=0.003s] dispatch=2.474s [2026-10-19 10:22:27,663] loop.py:45 INFO TRAIN_STEP step=9585 loss=1.391482 Training: 53%|█████▎ | 9585/18018 [20:52:13<17:15:23, 7.37s/step, loss=1.391, acc=0.684, acc_len=1.62, thru=18.0, I=21.0, T=31.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:28,063] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:22:34,150] loop.py:45 INFO TRAIN_STEP step=9586 loss=1.285083 Training: 53%|█████▎ | 9586/18018 [20:52:20<16:38:16, 7.10s/step, loss=1.285, acc=0.702, acc_len=1.69, thru=16.8, I=21.2, T=32.4, wait=2.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:22:34,145] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9585: forward=2779.4ms backward=1125.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:41,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:22:43,309] loop.py:45 INFO TRAIN_STEP step=9587 loss=1.222631 Training: 53%|█████▎ | 9587/18018 [20:52:29<18:04:43, 7.72s/step, loss=1.223, acc=0.712, acc_len=1.72, thru=15.3, I=17.2, T=30.1, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:43,380] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:52,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:22:53,320] loop.py:45 INFO TRAIN_STEP step=9588 loss=1.355388 Training: 53%|█████▎ | 9588/18018 [20:52:39<19:41:14, 8.41s/step, loss=1.355, acc=0.693, acc_len=1.65, thru=12.9, I=23.5, T=33.4, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:22:53,405] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:00,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:23:00,990] loop.py:45 INFO TRAIN_STEP step=9589 loss=1.531863 Training: 53%|█████▎ | 9589/18018 [20:52:47<19:09:58, 8.19s/step, loss=1.532, acc=0.659, acc_len=1.54, thru=17.6, I=18.0, T=33.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:01,363] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:23:10,611] loop.py:380 INFO TIMING step=9590: step=4.478s data=2.033s compute=4.447s [fwd=3.246s bwd=1.185s opt=0.003s] dispatch=5.134s [2026-10-19 10:23:10,612] loop.py:45 INFO TRAIN_STEP step=9590 loss=1.498564 Training: 53%|█████▎ | 9590/18018 [20:52:56<20:10:23, 8.62s/step, loss=1.499, acc=0.668, acc_len=1.58, thru=12.2, I=14.7, T=28.6, wait=5.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:19,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:23:20,087] loop.py:45 INFO TRAIN_STEP step=9591 loss=1.339138 Training: 53%|█████▎ | 9591/18018 [20:53:06<20:46:21, 8.87s/step, loss=1.339, acc=0.692, acc_len=1.66, thru=15.7, I=12.9, T=33.1, wait=5.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:23:20,083] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9590: forward=2665.2ms backward=1125.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:20,397] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:25,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:23:26,767] loop.py:45 INFO TRAIN_STEP step=9592 loss=1.385908 Training: 53%|█████▎ | 9592/18018 [20:53:12<19:13:51, 8.22s/step, loss=1.386, acc=0.684, acc_len=1.63, thru=18.6, I=22.4, T=34.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:26,849] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:31,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:23:33,046] loop.py:45 INFO TRAIN_STEP step=9593 loss=1.310691 Training: 53%|█████▎ | 9593/18018 [20:53:19<17:52:02, 7.63s/step, loss=1.311, acc=0.697, acc_len=1.67, thru=18.9, I=27.3, T=30.8, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:33,528] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:37,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:23:39,583] loop.py:45 INFO TRAIN_STEP step=9594 loss=1.313002 Training: 53%|█████▎ | 9594/18018 [20:53:25<17:05:44, 7.31s/step, loss=1.313, acc=0.699, acc_len=1.68, thru=18.4, I=26.7, T=32.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:39,994] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:44,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:23:45,744] loop.py:380 INFO TIMING step=9595: step=3.766s data=1.636s compute=3.735s [fwd=2.636s bwd=1.084s opt=0.003s] dispatch=2.385s [2026-10-19 10:23:45,744] loop.py:45 INFO TRAIN_STEP step=9595 loss=1.326173 Training: 53%|█████▎ | 9595/18018 [20:53:31<16:17:24, 6.96s/step, loss=1.326, acc=0.692, acc_len=1.66, thru=18.8, I=26.4, T=34.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:23:45,890] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 10:23:52,058] loop.py:45 INFO TRAIN_STEP step=9596 loss=1.326615 Training: 53%|█████▎ | 9596/18018 [20:53:38<15:50:06, 6.77s/step, loss=1.327, acc=0.696, acc_len=1.67, thru=17.7, I=20.2, T=35.1, wait=2.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:23:52,053] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9595: forward=2541.5ms backward=1061.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:02,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:24:03,728] loop.py:45 INFO TRAIN_STEP step=9597 loss=1.372363 Training: 53%|█████▎ | 9597/18018 [20:53:49<19:16:09, 8.24s/step, loss=1.372, acc=0.688, acc_len=1.64, thru=12.9, I=14.7, T=26.9, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:03,810] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:11,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:24:12,599] loop.py:45 INFO TRAIN_STEP step=9598 loss=1.341472 Training: 53%|█████▎ | 9598/18018 [20:53:58<19:42:42, 8.43s/step, loss=1.341, acc=0.692, acc_len=1.65, thru=14.6, I=14.5, T=29.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:13,105] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:18,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:24:20,146] loop.py:45 INFO TRAIN_STEP step=9599 loss=1.342827 Training: 53%|█████▎ | 9599/18018 [20:54:06<19:05:32, 8.16s/step, loss=1.343, acc=0.697, acc_len=1.67, thru=14.8, I=20.3, T=31.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:20,355] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:24:27,538] loop.py:380 INFO TIMING step=9600: step=3.866s data=1.654s compute=3.835s [fwd=2.707s bwd=1.113s opt=0.003s] dispatch=3.514s [2026-10-19 10:24:27,539] loop.py:45 INFO TRAIN_STEP step=9600 loss=1.208565 Training: 53%|█████▎ | 9600/18018 [20:54:13<18:33:02, 7.93s/step, loss=1.209, acc=0.716, acc_len=1.74, thru=15.1, I=21.5, T=33.1, wait=3.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:33,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:24:34,027] loop.py:45 INFO TRAIN_STEP step=9601 loss=1.253782 Training: 53%|█████▎ | 9601/18018 [20:54:20<17:31:51, 7.50s/step, loss=1.254, acc=0.711, acc_len=1.71, thru=19.2, I=19.8, T=33.4, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:24:34,022] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9600: forward=2724.3ms backward=1065.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:34,509] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:24:40,551] loop.py:45 INFO TRAIN_STEP step=9602 loss=1.440345 Training: 53%|█████▎ | 9602/18018 [20:54:26<16:50:51, 7.21s/step, loss=1.440, acc=0.675, acc_len=1.60, thru=17.4, I=22.1, T=35.3, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:51,272] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:24:51,975] loop.py:45 INFO TRAIN_STEP step=9603 loss=1.406080 Training: 53%|█████▎ | 9603/18018 [20:54:38<19:48:12, 8.47s/step, loss=1.406, acc=0.682, acc_len=1.62, thru=11.6, I=13.8, T=29.7, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:52,279] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:58,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:24:59,238] loop.py:45 INFO TRAIN_STEP step=9604 loss=1.502674 Training: 53%|█████▎ | 9604/18018 [20:54:45<18:57:03, 8.11s/step, loss=1.503, acc=0.668, acc_len=1.57, thru=15.8, I=16.5, T=35.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:24:59,358] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:06,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:25:06,646] loop.py:380 INFO TIMING step=9605: step=4.035s data=1.874s compute=4.008s [fwd=2.851s bwd=1.142s opt=0.003s] dispatch=3.365s [2026-10-19 10:25:06,646] loop.py:45 INFO TRAIN_STEP step=9605 loss=1.354285 Training: 53%|█████▎ | 9605/18018 [20:54:52<18:27:29, 7.90s/step, loss=1.354, acc=0.689, acc_len=1.65, thru=18.1, I=20.5, T=31.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:06,991] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:13,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:25:14,811] loop.py:45 INFO TRAIN_STEP step=9606 loss=1.222125 Training: 53%|█████▎ | 9606/18018 [20:55:00<18:38:36, 7.98s/step, loss=1.222, acc=0.716, acc_len=1.74, thru=14.8, I=16.6, T=31.0, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:25:14,806] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9605: forward=2952.5ms backward=1129.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:15,261] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:21,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:25:22,042] loop.py:45 INFO TRAIN_STEP step=9607 loss=1.287159 Training: 53%|█████▎ | 9607/18018 [20:55:08<18:07:05, 7.75s/step, loss=1.287, acc=0.706, acc_len=1.71, thru=16.1, I=20.3, T=32.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:22,075] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 10:25:28,297] loop.py:45 INFO TRAIN_STEP step=9608 loss=1.251181 Training: 53%|█████▎ | 9608/18018 [20:55:14<17:03:52, 7.30s/step, loss=1.251, acc=0.708, acc_len=1.72, thru=18.6, I=21.6, T=36.1, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:35,710] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:25:37,572] loop.py:45 INFO TRAIN_STEP step=9609 loss=1.370264 Training: 53%|█████▎ | 9609/18018 [20:55:23<18:26:37, 7.90s/step, loss=1.370, acc=0.689, acc_len=1.65, thru=14.8, I=17.1, T=26.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:37,720] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 [2026-10-19 10:25:44,344] loop.py:380 INFO TIMING step=9610: step=4.002s data=1.783s compute=3.960s [fwd=2.792s bwd=1.152s opt=0.003s] dispatch=2.760s [2026-10-19 10:25:44,345] loop.py:45 INFO TRAIN_STEP step=9610 loss=1.418996 Training: 53%|█████▎ | 9610/18018 [20:55:30<17:39:15, 7.56s/step, loss=1.419, acc=0.677, acc_len=1.61, thru=14.5, I=24.0, T=32.0, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:51,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 10:25:52,500] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9610: forward=2957.1ms backward=1249.7ms [2026-10-19 10:25:52,506] loop.py:45 INFO TRAIN_STEP step=9611 loss=1.192927 Training: 53%|█████▎ | 9611/18018 [20:55:38<18:04:27, 7.74s/step, loss=1.193, acc=0.723, acc_len=1.76, thru=15.6, I=17.2, T=30.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:52,552] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:58,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:25:59,316] loop.py:45 INFO TRAIN_STEP step=9612 loss=1.316257 Training: 53%|█████▎ | 9612/18018 [20:55:45<17:25:16, 7.46s/step, loss=1.316, acc=0.699, acc_len=1.68, thru=17.2, I=22.3, T=31.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:25:59,604] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:04,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:26:06,735] loop.py:45 INFO TRAIN_STEP step=9613 loss=1.339211 Training: 53%|█████▎ | 9613/18018 [20:55:52<17:23:18, 7.45s/step, loss=1.339, acc=0.697, acc_len=1.67, thru=16.4, I=25.2, T=26.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:06,820] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:12,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:26:12,420] loop.py:45 INFO TRAIN_STEP step=9614 loss=1.326015 Training: 53%|█████▎ | 9614/18018 [20:55:58<16:09:03, 6.92s/step, loss=1.326, acc=0.697, acc_len=1.67, thru=18.4, I=28.4, T=34.4, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:12,725] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:26:21,818] loop.py:380 INFO TIMING step=9615: step=4.362s data=1.937s compute=4.333s [fwd=3.112s bwd=1.206s opt=0.003s] dispatch=5.028s [2026-10-19 10:26:21,818] loop.py:45 INFO TRAIN_STEP step=9615 loss=1.366055 Training: 53%|█████▎ | 9615/18018 [20:56:07<17:53:15, 7.66s/step, loss=1.366, acc=0.691, acc_len=1.65, thru=13.2, I=13.9, T=29.3, wait=5.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:26,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:26:28,779] loop.py:45 INFO TRAIN_STEP step=9616 loss=1.254763 Training: 53%|█████▎ | 9616/18018 [20:56:14<17:23:32, 7.45s/step, loss=1.255, acc=0.710, acc_len=1.73, thru=17.2, I=20.7, T=30.7, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:26:28,774] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9615: forward=2947.5ms backward=1179.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:28,952] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:34,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:26:34,844] loop.py:45 INFO TRAIN_STEP step=9617 loss=1.248730 Training: 53%|█████▎ | 9617/18018 [20:56:20<16:25:08, 7.04s/step, loss=1.249, acc=0.710, acc_len=1.72, thru=19.0, I=24.9, T=36.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:35,267] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:40,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:26:42,038] loop.py:45 INFO TRAIN_STEP step=9618 loss=1.301492 Training: 53%|█████▎ | 9618/18018 [20:56:28<16:31:46, 7.08s/step, loss=1.301, acc=0.698, acc_len=1.67, thru=17.3, I=22.1, T=30.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:42,340] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:46,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:26:47,482] loop.py:45 INFO TRAIN_STEP step=9619 loss=1.300441 Training: 53%|█████▎ | 9619/18018 [20:56:33<15:22:45, 6.59s/step, loss=1.300, acc=0.700, acc_len=1.68, thru=21.3, I=26.1, T=37.8, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:47,545] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:53,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:26:53,821] loop.py:380 INFO TIMING step=9620: step=3.733s data=1.595s compute=3.697s [fwd=2.646s bwd=1.036s opt=0.003s] dispatch=2.596s [2026-10-19 10:26:53,821] loop.py:45 INFO TRAIN_STEP step=9620 loss=1.380800 Training: 53%|█████▎ | 9620/18018 [20:56:39<15:12:02, 6.52s/step, loss=1.381, acc=0.686, acc_len=1.63, thru=20.9, I=24.8, T=34.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:53,898] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:26:59,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:27:01,085] loop.py:45 INFO TRAIN_STEP step=9621 loss=1.264292 Training: 53%|█████▎ | 9621/18018 [20:56:47<15:43:13, 6.74s/step, loss=1.264, acc=0.707, acc_len=1.71, thru=17.8, I=20.9, T=35.1, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:27:01,080] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9620: forward=2621.0ms backward=980.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:01,378] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:08,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:27:09,685] loop.py:45 INFO TRAIN_STEP step=9622 loss=1.302819 Training: 53%|█████▎ | 9622/18018 [20:56:55<17:01:18, 7.30s/step, loss=1.303, acc=0.700, acc_len=1.69, thru=13.4, I=16.7, T=32.5, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:09,922] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:16,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:27:17,224] loop.py:45 INFO TRAIN_STEP step=9623 loss=1.346489 Training: 53%|█████▎ | 9623/18018 [20:57:03<17:11:11, 7.37s/step, loss=1.346, acc=0.693, acc_len=1.67, thru=15.2, I=20.5, T=34.6, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:17,305] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:24,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:27:25,108] loop.py:45 INFO TRAIN_STEP step=9624 loss=1.353906 Training: 53%|█████▎ | 9624/18018 [20:57:11<17:32:44, 7.52s/step, loss=1.354, acc=0.691, acc_len=1.65, thru=16.8, I=20.2, T=29.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:25,198] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:31,660] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:27:33,014] loop.py:380 INFO TIMING step=9625: step=4.147s data=1.892s compute=3.911s [fwd=2.737s bwd=1.159s opt=0.003s] dispatch=3.748s [2026-10-19 10:27:33,015] loop.py:45 INFO TRAIN_STEP step=9625 loss=1.180452 Training: 53%|█████▎ | 9625/18018 [20:57:19<17:48:35, 7.64s/step, loss=1.180, acc=0.722, acc_len=1.76, thru=14.8, I=17.6, T=30.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:33,176] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:38,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:27:39,314] loop.py:45 INFO TRAIN_STEP step=9626 loss=1.339864 Training: 53%|█████▎ | 9626/18018 [20:57:25<16:52:11, 7.24s/step, loss=1.340, acc=0.694, acc_len=1.66, thru=18.7, I=22.6, T=30.9, wait=2.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:27:39,309] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9625: forward=2935.8ms backward=1158.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:39,365] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:44,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:27:45,799] loop.py:45 INFO TRAIN_STEP step=9627 loss=1.332174 Training: 53%|█████▎ | 9627/18018 [20:57:31<16:20:37, 7.01s/step, loss=1.332, acc=0.696, acc_len=1.67, thru=17.9, I=21.2, T=28.8, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:46,184] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:52,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:27:52,364] loop.py:45 INFO TRAIN_STEP step=9628 loss=1.397673 Training: 53%|█████▎ | 9628/18018 [20:57:38<16:01:39, 6.88s/step, loss=1.398, acc=0.688, acc_len=1.64, thru=17.6, I=21.3, T=34.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:52,790] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:58,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:27:59,549] loop.py:45 INFO TRAIN_STEP step=9629 loss=1.316645 Training: 53%|█████▎ | 9629/18018 [20:57:45<16:14:30, 6.97s/step, loss=1.317, acc=0.698, acc_len=1.68, thru=18.2, I=20.7, T=32.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:27:59,812] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:06,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:28:06,496] loop.py:380 INFO TIMING step=9630: step=3.655s data=1.526s compute=3.624s [fwd=2.575s bwd=1.034s opt=0.003s] dispatch=3.281s [2026-10-19 10:28:06,496] loop.py:45 INFO TRAIN_STEP step=9630 loss=1.263622 Training: 53%|█████▎ | 9630/18018 [20:57:52<16:13:26, 6.96s/step, loss=1.264, acc=0.706, acc_len=1.70, thru=18.1, I=21.5, T=35.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:06,573] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 10:28:12,619] loop.py:45 INFO TRAIN_STEP step=9631 loss=1.291963 Training: 53%|█████▎ | 9631/18018 [20:57:58<15:38:01, 6.71s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=19.0, I=21.5, T=33.3, wait=2.3s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:28:12,615] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9630: forward=2673.4ms backward=1117.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:22,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:28:22,677] loop.py:45 INFO TRAIN_STEP step=9632 loss=1.395115 Training: 53%|█████▎ | 9632/18018 [20:58:08<17:58:20, 7.72s/step, loss=1.395, acc=0.688, acc_len=1.64, thru=14.4, I=14.1, T=31.1, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:22,713] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:28,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:28:29,453] loop.py:45 INFO TRAIN_STEP step=9633 loss=1.374795 Training: 53%|█████▎ | 9633/18018 [20:58:15<17:18:50, 7.43s/step, loss=1.375, acc=0.688, acc_len=1.64, thru=17.9, I=20.6, T=29.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:29,597] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:35,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:28:36,673] loop.py:45 INFO TRAIN_STEP step=9634 loss=1.396653 Training: 53%|█████▎ | 9634/18018 [20:58:22<17:09:46, 7.37s/step, loss=1.397, acc=0.684, acc_len=1.62, thru=15.9, I=20.4, T=32.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:37,068] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:28:43,413] loop.py:380 INFO TIMING step=9635: step=3.792s data=1.630s compute=3.761s [fwd=2.721s bwd=1.025s opt=0.003s] dispatch=2.937s [2026-10-19 10:28:43,413] loop.py:45 INFO TRAIN_STEP step=9635 loss=1.443421 Training: 53%|█████▎ | 9635/18018 [20:58:29<16:43:18, 7.18s/step, loss=1.443, acc=0.677, acc_len=1.61, thru=16.8, I=21.6, T=33.8, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:50,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:28:52,963] loop.py:45 INFO TRAIN_STEP step=9636 loss=1.327536 Training: 53%|█████▎ | 9636/18018 [20:58:39<18:22:26, 7.89s/step, loss=1.328, acc=0.695, acc_len=1.66, thru=14.8, I=17.0, T=26.0, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:28:52,959] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9635: forward=3487.2ms backward=1383.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:53,054] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:58,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:28:59,759] loop.py:45 INFO TRAIN_STEP step=9637 loss=1.287557 Training: 53%|█████▎ | 9637/18018 [20:58:45<17:36:24, 7.56s/step, loss=1.288, acc=0.700, acc_len=1.68, thru=15.7, I=24.0, T=36.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:28:59,880] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:29:06,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:29:07,319] loop.py:45 INFO TRAIN_STEP step=9638 loss=1.154585 Training: 53%|█████▎ | 9638/18018 [20:58:53<17:36:07, 7.56s/step, loss=1.155, acc=0.729, acc_len=1.79, thru=15.7, I=19.7, T=33.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:29:07,706] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:29:13,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:29:14,213] loop.py:45 INFO TRAIN_STEP step=9639 loss=1.420884 Training: 53%|█████▎ | 9639/18018 [20:59:00<17:07:58, 7.36s/step, loss=1.421, acc=0.679, acc_len=1.61, thru=19.1, I=20.0, T=31.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:29:14,328] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 10:29:20,630] loop.py:380 INFO TIMING step=9640: step=3.667s data=1.524s compute=3.633s [fwd=2.585s bwd=1.033s opt=0.003s] dispatch=2.741s [2026-10-19 10:29:20,631] loop.py:45 INFO TRAIN_STEP step=9640 loss=1.364516 Training: 54%|█████▎ | 9640/18018 [20:59:06<16:28:19, 7.08s/step, loss=1.365, acc=0.688, acc_len=1.64, thru=17.3, I=19.9, T=34.9, wait=2.7s, pool=57, epoch=2/2][2026-10-19 10:29:29,096] loop.py:45 INFO TRAIN_STEP step=9641 loss=1.417682 Training: 54%|█████▎ | 9641/18018 [20:59:15<17:26:17, 7.49s/step, loss=1.418, acc=0.681, acc_len=1.62, thru=13.7, I=14.1, T=30.8, wait=4.3s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:29:29,091] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9640: forward=2925.4ms backward=1176.8ms [2026-10-19 10:29:39,074] loop.py:45 INFO TRAIN_STEP step=9642 loss=1.294415 Training: 54%|█████▎ | 9642/18018 [20:59:25<19:10:18, 8.24s/step, loss=1.294, acc=0.706, acc_len=1.70, thru=14.1, I=14.9, T=28.3, wait=5.4s, pool=57, epoch=2/2][2026-10-19 10:29:46,229] loop.py:45 INFO TRAIN_STEP step=9643 loss=1.410062 Training: 54%|█████▎ | 9643/18018 [20:59:32<18:24:37, 7.91s/step, loss=1.410, acc=0.681, acc_len=1.62, thru=16.1, I=16.0, T=33.2, wait=3.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:29:55,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:29:56,849] loop.py:45 INFO TRAIN_STEP step=9644 loss=1.370998 Training: 54%|█████▎ | 9644/18018 [20:59:42<20:17:49, 8.73s/step, loss=1.371, acc=0.688, acc_len=1.64, thru=13.9, I=15.9, T=29.1, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:29:57,193] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:30:03,128] loop.py:380 INFO TIMING step=9645: step=3.689s data=1.424s compute=3.661s [fwd=2.525s bwd=1.120s opt=0.003s] dispatch=2.582s [2026-10-19 10:30:03,128] loop.py:45 INFO TRAIN_STEP step=9645 loss=1.354751 Training: 54%|█████▎ | 9645/18018 [20:59:49<18:35:17, 7.99s/step, loss=1.355, acc=0.690, acc_len=1.65, thru=18.1, I=24.0, T=34.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:03,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:03,552] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:08,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:30:09,878] loop.py:45 INFO TRAIN_STEP step=9646 loss=1.307040 Training: 54%|█████▎ | 9646/18018 [20:59:55<17:43:09, 7.62s/step, loss=1.307, acc=0.699, acc_len=1.68, thru=17.9, I=20.3, T=34.1, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:30:09,873] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9645: forward=2626.0ms backward=1053.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:10,232] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:16,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:30:16,708] loop.py:45 INFO TRAIN_STEP step=9647 loss=1.273683 Training: 54%|█████▎ | 9647/18018 [21:00:02<17:09:59, 7.38s/step, loss=1.274, acc=0.706, acc_len=1.71, thru=18.1, I=21.6, T=34.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:16,921] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:21,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:30:22,716] loop.py:45 INFO TRAIN_STEP step=9648 loss=1.337528 Training: 54%|█████▎ | 9648/18018 [21:00:08<16:12:22, 6.97s/step, loss=1.338, acc=0.694, acc_len=1.66, thru=20.2, I=21.7, T=35.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:22,951] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:30,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:30:31,425] loop.py:45 INFO TRAIN_STEP step=9649 loss=1.334851 Training: 54%|█████▎ | 9649/18018 [21:00:17<17:25:01, 7.49s/step, loss=1.335, acc=0.696, acc_len=1.67, thru=14.5, I=17.5, T=30.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:31,799] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:37,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:30:38,520] loop.py:380 INFO TIMING step=9650: step=4.037s data=1.839s compute=4.005s [fwd=2.796s bwd=1.193s opt=0.003s] dispatch=3.049s [2026-10-19 10:30:38,521] loop.py:45 INFO TRAIN_STEP step=9650 loss=1.444954 Training: 54%|█████▎ | 9650/18018 [21:00:24<17:08:12, 7.37s/step, loss=1.445, acc=0.673, acc_len=1.59, thru=17.0, I=22.8, T=31.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:38,608] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-19 10:30:46,333] loop.py:45 INFO TRAIN_STEP step=9651 loss=1.267684 Training: 54%|█████▎ | 9651/18018 [21:00:32<17:26:34, 7.50s/step, loss=1.268, acc=0.706, acc_len=1.71, thru=13.2, I=15.8, T=34.0, wait=4.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:30:46,329] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9650: forward=2715.3ms backward=997.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:54,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:30:56,372] loop.py:45 INFO TRAIN_STEP step=9652 loss=1.437042 Training: 54%|█████▎ | 9652/18018 [21:00:42<19:12:23, 8.26s/step, loss=1.437, acc=0.677, acc_len=1.60, thru=14.3, I=16.0, T=29.0, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:30:56,474] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:00,880] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:31:02,402] loop.py:45 INFO TRAIN_STEP step=9653 loss=1.277507 Training: 54%|█████▎ | 9653/18018 [21:00:48<17:38:49, 7.59s/step, loss=1.278, acc=0.704, acc_len=1.70, thru=18.6, I=25.5, T=34.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:02,896] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:09,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:31:10,559] loop.py:45 INFO TRAIN_STEP step=9654 loss=1.406249 Training: 54%|█████▎ | 9654/18018 [21:00:56<18:02:10, 7.76s/step, loss=1.406, acc=0.684, acc_len=1.63, thru=13.6, I=20.7, T=33.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:11,039] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:31:19,995] loop.py:380 INFO TIMING step=9655: step=4.441s data=2.005s compute=4.390s [fwd=3.089s bwd=1.284s opt=0.003s] dispatch=4.985s [2026-10-19 10:31:19,995] loop.py:45 INFO TRAIN_STEP step=9655 loss=1.311125 Training: 54%|█████▎ | 9655/18018 [21:01:06<19:11:59, 8.26s/step, loss=1.311, acc=0.701, acc_len=1.68, thru=10.3, I=17.6, T=28.8, wait=5.0s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:31,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:31:33,126] loop.py:45 INFO TRAIN_STEP step=9656 loss=1.299456 (TrainerActor pid=9310) [2026-10-19 10:31:33,121] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9655: forward=2947.8ms backward=1206.5ms Training: 54%|█████▎ | 9656/18018 [21:01:19<22:35:22, 9.73s/step, loss=1.299, acc=0.704, acc_len=1.70, thru=12.8, I=14.6, T=30.4, wait=8.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:33,469] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:40,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:31:42,257] loop.py:45 INFO TRAIN_STEP step=9657 loss=1.268945 Training: 54%|█████▎ | 9657/18018 [21:01:28<22:10:22, 9.55s/step, loss=1.269, acc=0.706, acc_len=1.70, thru=14.6, I=15.8, T=29.2, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:42,450] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:49,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:31:50,672] loop.py:45 INFO TRAIN_STEP step=9658 loss=1.341372 Training: 54%|█████▎ | 9658/18018 [21:01:36<21:22:54, 9.21s/step, loss=1.341, acc=0.692, acc_len=1.65, thru=13.6, I=17.1, T=28.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:50,755] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:57,326] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:31:58,168] loop.py:45 INFO TRAIN_STEP step=9659 loss=1.326111 Training: 54%|█████▎ | 9659/18018 [21:01:44<20:11:14, 8.69s/step, loss=1.326, acc=0.698, acc_len=1.68, thru=16.0, I=18.1, T=31.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:31:58,339] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 10:32:05,259] loop.py:380 INFO TIMING step=9660: step=3.782s data=1.568s compute=3.747s [fwd=2.682s bwd=1.050s opt=0.003s] dispatch=3.298s [2026-10-19 10:32:05,259] loop.py:45 INFO TRAIN_STEP step=9660 loss=1.403559 Training: 54%|█████▎ | 9660/18018 [21:01:51<19:03:54, 8.21s/step, loss=1.404, acc=0.679, acc_len=1.62, thru=15.2, I=15.6, T=33.8, wait=3.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:15,693] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:32:15,757] loop.py:45 INFO TRAIN_STEP step=9661 loss=1.492629 Training: 54%|█████▎ | 9661/18018 [21:02:01<20:39:19, 8.90s/step, loss=1.493, acc=0.669, acc_len=1.58, thru=13.8, I=13.3, T=29.7, wait=6.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:32:15,753] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9660: forward=3097.9ms backward=1156.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:16,194] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:22,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:32:22,919] loop.py:45 INFO TRAIN_STEP step=9662 loss=1.371369 Training: 54%|█████▎ | 9662/18018 [21:02:09<19:26:42, 8.38s/step, loss=1.371, acc=0.690, acc_len=1.65, thru=19.3, I=18.5, T=31.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:23,139] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:32,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:32:32,335] loop.py:45 INFO TRAIN_STEP step=9663 loss=1.419475 Training: 54%|█████▎ | 9663/18018 [21:02:18<20:10:00, 8.69s/step, loss=1.419, acc=0.683, acc_len=1.63, thru=13.9, I=13.3, T=28.9, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:32,546] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:38,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:32:39,079] loop.py:45 INFO TRAIN_STEP step=9664 loss=1.158711 Training: 54%|█████▎ | 9664/18018 [21:02:25<18:48:31, 8.11s/step, loss=1.159, acc=0.726, acc_len=1.77, thru=17.3, I=18.7, T=33.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:39,483] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:32:46,808] loop.py:380 INFO TIMING step=9665: step=3.972s data=1.743s compute=3.926s [fwd=2.809s bwd=1.100s opt=0.003s] dispatch=3.749s [2026-10-19 10:32:46,809] loop.py:45 INFO TRAIN_STEP step=9665 loss=1.289599 Training: 54%|█████▎ | 9665/18018 [21:02:32<18:32:40, 7.99s/step, loss=1.290, acc=0.700, acc_len=1.68, thru=14.4, I=16.0, T=32.2, wait=3.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:55,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:32:56,220] loop.py:45 INFO TRAIN_STEP step=9666 loss=1.471866 Training: 54%|█████▎ | 9666/18018 [21:02:42<19:31:55, 8.42s/step, loss=1.472, acc=0.673, acc_len=1.59, thru=15.9, I=15.0, T=33.1, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:32:56,216] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9665: forward=2757.0ms backward=1057.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:32:56,289] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 10:33:03,247] loop.py:45 INFO TRAIN_STEP step=9667 loss=1.419793 Training: 54%|█████▎ | 9667/18018 [21:02:49<18:33:34, 8.00s/step, loss=1.420, acc=0.679, acc_len=1.62, thru=16.1, I=18.2, T=36.1, wait=3.5s, pool=49, epoch=2/2][2026-10-19 10:33:13,486] loop.py:45 INFO TRAIN_STEP step=9668 loss=1.323638 Training: 54%|█████▎ | 9668/18018 [21:02:59<20:06:57, 8.67s/step, loss=1.324, acc=0.699, acc_len=1.68, thru=11.8, I=13.5, T=32.1, wait=6.2s, pool=41, epoch=2/2][2026-10-19 10:33:22,160] loop.py:45 INFO TRAIN_STEP step=9669 loss=1.548832 Training: 54%|█████▎ | 9669/18018 [21:03:08<20:06:46, 8.67s/step, loss=1.549, acc=0.662, acc_len=1.55, thru=14.0, I=13.5, T=33.8, wait=4.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:33:30,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:33:30,997] loop.py:380 INFO TIMING step=9670: step=4.209s data=1.802s compute=4.185s [fwd=3.006s bwd=1.163s opt=0.003s] dispatch=4.620s [2026-10-19 10:33:30,997] loop.py:45 INFO TRAIN_STEP step=9670 loss=1.417330 Training: 54%|█████▎ | 9670/18018 [21:03:17<20:13:28, 8.72s/step, loss=1.417, acc=0.683, acc_len=1.63, thru=17.4, I=15.4, T=30.4, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:33:31,029] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:33:36,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:33:38,667] loop.py:45 INFO TRAIN_STEP step=9671 loss=1.203932 Training: 54%|█████▎ | 9671/18018 [21:03:24<19:29:28, 8.41s/step, loss=1.204, acc=0.721, acc_len=1.76, thru=18.3, I=19.9, T=28.7, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:33:38,662] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9670: forward=3094.3ms backward=1312.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:33:38,831] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:33:44,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:33:46,059] loop.py:45 INFO TRAIN_STEP step=9672 loss=1.353683 Training: 54%|█████▎ | 9672/18018 [21:03:32<18:47:08, 8.10s/step, loss=1.354, acc=0.692, acc_len=1.66, thru=15.4, I=22.2, T=31.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:33:46,442] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:33:53,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:33:54,901] loop.py:45 INFO TRAIN_STEP step=9673 loss=1.315855 Training: 54%|█████▎ | 9673/18018 [21:03:40<19:17:47, 8.32s/step, loss=1.316, acc=0.696, acc_len=1.67, thru=13.2, I=16.6, T=34.8, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:33:54,996] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:01,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:34:01,946] loop.py:45 INFO TRAIN_STEP step=9674 loss=1.281117 Training: 54%|█████▎ | 9674/18018 [21:03:48<18:24:10, 7.94s/step, loss=1.281, acc=0.704, acc_len=1.69, thru=17.2, I=19.9, T=36.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:01,973] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 10:34:09,290] loop.py:380 INFO TIMING step=9675: step=3.937s data=1.723s compute=3.909s [fwd=2.770s bwd=1.124s opt=0.003s] dispatch=3.399s [2026-10-19 10:34:09,291] loop.py:45 INFO TRAIN_STEP step=9675 loss=1.357961 Training: 54%|█████▎ | 9675/18018 [21:03:55<17:59:11, 7.76s/step, loss=1.358, acc=0.695, acc_len=1.67, thru=17.4, I=16.7, T=32.5, wait=3.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:15,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:34:17,019] loop.py:45 INFO TRAIN_STEP step=9676 loss=1.295898 Training: 54%|█████▎ | 9676/18018 [21:04:03<17:57:47, 7.75s/step, loss=1.296, acc=0.701, acc_len=1.68, thru=18.1, I=18.4, T=31.9, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:34:17,014] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9675: forward=2803.8ms backward=1163.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:17,480] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:24,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:34:25,223] loop.py:45 INFO TRAIN_STEP step=9677 loss=1.363908 Training: 54%|█████▎ | 9677/18018 [21:04:11<18:16:28, 7.89s/step, loss=1.364, acc=0.690, acc_len=1.64, thru=14.8, I=17.6, T=29.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:25,462] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:31,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:34:32,204] loop.py:45 INFO TRAIN_STEP step=9678 loss=1.363092 Training: 54%|█████▎ | 9678/18018 [21:04:18<17:38:36, 7.62s/step, loss=1.363, acc=0.693, acc_len=1.66, thru=16.6, I=19.3, T=33.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:32,377] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:34:38,313] loop.py:45 INFO TRAIN_STEP step=9679 loss=1.280423 Training: 54%|█████▎ | 9679/18018 [21:04:24<16:35:33, 7.16s/step, loss=1.280, acc=0.703, acc_len=1.69, thru=19.7, I=19.5, T=34.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:38,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:38,783] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:44,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:34:44,771] loop.py:380 INFO TIMING step=9680: step=3.523s data=1.407s compute=3.501s [fwd=2.545s bwd=0.941s opt=0.003s] dispatch=2.926s [2026-10-19 10:34:44,771] loop.py:45 INFO TRAIN_STEP step=9680 loss=1.378487 Training: 54%|█████▎ | 9680/18018 [21:04:30<16:06:06, 6.95s/step, loss=1.378, acc=0.685, acc_len=1.63, thru=19.3, I=20.0, T=36.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:44,946] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:49,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:34:50,790] loop.py:45 INFO TRAIN_STEP step=9681 loss=1.323241 Training: 54%|█████▎ | 9681/18018 [21:04:36<15:27:05, 6.67s/step, loss=1.323, acc=0.696, acc_len=1.67, thru=20.5, I=22.9, T=35.4, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:34:50,785] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9680: forward=2581.0ms backward=987.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:51,005] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:55,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:34:56,599] loop.py:45 INFO TRAIN_STEP step=9682 loss=1.213407 Training: 54%|█████▎ | 9682/18018 [21:04:42<14:51:00, 6.41s/step, loss=1.213, acc=0.715, acc_len=1.73, thru=21.1, I=26.1, T=36.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:34:56,853] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:35:04,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:35:04,944] loop.py:45 INFO TRAIN_STEP step=9683 loss=1.321993 Training: 54%|█████▎ | 9683/18018 [21:04:51<16:11:22, 6.99s/step, loss=1.322, acc=0.698, acc_len=1.68, thru=15.7, I=17.6, T=32.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:35:05,366] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:35:12,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:35:14,057] loop.py:45 INFO TRAIN_STEP step=9684 loss=1.251954 Training: 54%|█████▎ | 9684/18018 [21:05:00<17:39:41, 7.63s/step, loss=1.252, acc=0.710, acc_len=1.72, thru=13.6, I=17.9, T=31.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:35:14,273] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:35:21,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:35:21,996] loop.py:380 INFO TIMING step=9685: step=3.733s data=1.569s compute=3.702s [fwd=2.633s bwd=1.056s opt=0.003s] dispatch=4.195s [2026-10-19 10:35:21,997] loop.py:45 INFO TRAIN_STEP step=9685 loss=1.365219 Training: 54%|█████▍ | 9685/18018 [21:05:08<17:52:22, 7.72s/step, loss=1.365, acc=0.689, acc_len=1.64, thru=14.2, I=16.2, T=34.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:35:22,379] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:35:29,262] loop.py:45 INFO TRAIN_STEP step=9686 loss=1.304094 Training: 54%|█████▍ | 9686/18018 [21:05:15<17:33:18, 7.59s/step, loss=1.304, acc=0.697, acc_len=1.66, thru=15.0, I=15.7, T=33.8, wait=3.5s, pool=25, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:35:29,258] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9685: forward=2640.9ms backward=1091.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:35:38,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:35:39,430] loop.py:45 INFO TRAIN_STEP step=9687 loss=1.356644 Training: 54%|█████▍ | 9687/18018 [21:05:25<19:20:44, 8.36s/step, loss=1.357, acc=0.693, acc_len=1.66, thru=16.8, I=15.6, T=28.4, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:35:39,561] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 10:35:45,945] loop.py:45 INFO TRAIN_STEP step=9688 loss=1.282877 Training: 54%|█████▍ | 9688/18018 [21:05:32<18:03:55, 7.81s/step, loss=1.283, acc=0.701, acc_len=1.69, thru=14.3, I=19.0, T=31.4, wait=2.4s, pool=45, epoch=2/2][2026-10-19 10:35:55,011] loop.py:45 INFO TRAIN_STEP step=9689 loss=1.374556 Training: 54%|█████▍ | 9689/18018 [21:05:41<18:56:08, 8.18s/step, loss=1.375, acc=0.692, acc_len=1.65, thru=15.9, I=14.5, T=31.5, wait=5.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:36:02,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:36:03,921] loop.py:380 INFO TIMING step=9690: step=4.016s data=1.727s compute=3.992s [fwd=2.848s bwd=1.129s opt=0.003s] dispatch=4.884s [2026-10-19 10:36:03,921] loop.py:45 INFO TRAIN_STEP step=9690 loss=1.483885 Training: 54%|█████▍ | 9690/18018 [21:05:50<19:26:06, 8.40s/step, loss=1.484, acc=0.671, acc_len=1.58, thru=15.0, I=16.4, T=31.9, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:36:04,361] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-19 10:36:10,771] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9690: forward=2560.9ms backward=1017.3ms [2026-10-19 10:36:10,776] loop.py:45 INFO TRAIN_STEP step=9691 loss=1.435938 Training: 54%|█████▍ | 9691/18018 [21:05:56<18:21:42, 7.94s/step, loss=1.436, acc=0.679, acc_len=1.61, thru=14.3, I=20.6, T=35.3, wait=3.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:36:17,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:36:18,935] loop.py:45 INFO TRAIN_STEP step=9692 loss=1.403684 Training: 54%|█████▍ | 9692/18018 [21:06:05<18:30:48, 8.00s/step, loss=1.404, acc=0.687, acc_len=1.64, thru=17.4, I=18.4, T=28.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:36:19,100] inference_manager.py:408 INFO Pool capacity available after 2.1s, pool_size=41/64 [2026-10-19 10:36:27,159] loop.py:45 INFO TRAIN_STEP step=9693 loss=1.268102 Training: 54%|█████▍ | 9693/18018 [21:06:13<18:39:38, 8.07s/step, loss=1.268, acc=0.707, acc_len=1.70, thru=15.4, I=14.7, T=35.3, wait=4.6s, pool=61, epoch=2/2][2026-10-19 10:36:33,722] loop.py:45 INFO TRAIN_STEP step=9694 loss=1.431395 Training: 54%|█████▍ | 9694/18018 [21:06:19<17:36:50, 7.62s/step, loss=1.431, acc=0.679, acc_len=1.61, thru=16.4, I=15.9, T=33.5, wait=2.7s, pool=33, epoch=2/2][2026-10-19 10:36:42,336] loop.py:380 INFO TIMING step=9695: step=3.674s data=1.477s compute=3.634s [fwd=2.601s bwd=1.019s opt=0.003s] dispatch=4.931s [2026-10-19 10:36:42,336] loop.py:45 INFO TRAIN_STEP step=9695 loss=1.243075 Training: 54%|█████▍ | 9695/18018 [21:06:28<18:18:14, 7.92s/step, loss=1.243, acc=0.710, acc_len=1.72, thru=17.0, I=15.2, T=34.8, wait=4.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:36:49,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:36:49,296] loop.py:45 INFO TRAIN_STEP step=9696 loss=1.376554 Training: 54%|█████▍ | 9696/18018 [21:06:35<17:38:10, 7.63s/step, loss=1.377, acc=0.691, acc_len=1.65, thru=19.2, I=17.6, T=34.1, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:36:49,291] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9695: forward=2703.6ms backward=1000.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:36:49,615] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:36:56,324] loop.py:45 INFO TRAIN_STEP step=9697 loss=1.483715 Training: 54%|█████▍ | 9697/18018 [21:06:42<17:13:01, 7.45s/step, loss=1.484, acc=0.671, acc_len=1.59, thru=14.8, I=18.1, T=34.2, wait=3.3s, pool=33, epoch=2/2][2026-10-19 10:37:06,352] loop.py:45 INFO TRAIN_STEP step=9698 loss=1.284379 Training: 54%|█████▍ | 9698/18018 [21:06:52<19:00:19, 8.22s/step, loss=1.284, acc=0.705, acc_len=1.70, thru=14.0, I=13.0, T=27.8, wait=5.4s, pool=33, epoch=2/2][2026-10-19 10:37:15,676] loop.py:45 INFO TRAIN_STEP step=9699 loss=1.319813 Training: 54%|█████▍ | 9699/18018 [21:07:01<19:45:52, 8.55s/step, loss=1.320, acc=0.699, acc_len=1.68, thru=13.2, I=13.0, T=31.7, wait=5.3s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:37:24,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:37:28,602] loop.py:380 INFO TIMING step=9700: step=5.798s data=2.658s compute=5.767s [fwd=4.583s bwd=1.170s opt=0.003s] dispatch=7.119s [2026-10-19 10:37:28,602] loop.py:45 INFO TRAIN_STEP step=9700 loss=1.406156 Training: 54%|█████▍ | 9700/18018 [21:07:14<22:47:40, 9.87s/step, loss=1.406, acc=0.686, acc_len=1.64, thru=14.6, I=15.6, T=22.1, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:37:28,985] inference_manager.py:408 INFO Pool capacity available after 4.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:37:33,741] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:37:35,541] loop.py:45 INFO TRAIN_STEP step=9701 loss=1.302638 Training: 54%|█████▍ | 9701/18018 [21:07:21<20:45:48, 8.99s/step, loss=1.303, acc=0.700, acc_len=1.68, thru=19.5, I=26.6, T=28.5, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:37:35,536] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9700: forward=3414.8ms backward=1030.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:37:35,757] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:37:41,718] loop.py:45 INFO TRAIN_STEP step=9702 loss=1.311702 Training: 54%|█████▍ | 9702/18018 [21:07:27<18:48:49, 8.14s/step, loss=1.312, acc=0.700, acc_len=1.68, thru=16.9, I=23.8, T=35.5, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:37:49,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:37:50,805] loop.py:45 INFO TRAIN_STEP step=9703 loss=1.380708 Training: 54%|█████▍ | 9703/18018 [21:07:36<19:27:49, 8.43s/step, loss=1.381, acc=0.688, acc_len=1.64, thru=14.7, I=15.0, T=31.4, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:37:51,262] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:37:56,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:37:57,268] loop.py:45 INFO TRAIN_STEP step=9704 loss=1.449695 Training: 54%|█████▍ | 9704/18018 [21:07:43<18:05:58, 7.84s/step, loss=1.450, acc=0.673, acc_len=1.59, thru=19.6, I=21.1, T=31.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:37:57,453] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:38:07,135] loop.py:380 INFO TIMING step=9705: step=4.128s data=1.782s compute=4.092s [fwd=2.930s bwd=1.147s opt=0.003s] dispatch=5.730s [2026-10-19 10:38:07,135] loop.py:45 INFO TRAIN_STEP step=9705 loss=1.461327 Training: 54%|█████▍ | 9705/18018 [21:07:53<19:30:20, 8.45s/step, loss=1.461, acc=0.680, acc_len=1.61, thru=11.2, I=10.7, T=31.0, wait=5.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:38:17,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:38:18,573] loop.py:45 INFO TRAIN_STEP step=9706 loss=1.367250 Training: 54%|█████▍ | 9706/18018 [21:08:04<21:34:22, 9.34s/step, loss=1.367, acc=0.690, acc_len=1.65, thru=12.6, I=13.4, T=31.3, wait=7.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:38:18,568] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9705: forward=2914.3ms backward=1117.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:38:18,759] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:38:23,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:38:24,955] loop.py:45 INFO TRAIN_STEP step=9707 loss=1.358623 Training: 54%|█████▍ | 9707/18018 [21:08:11<19:31:16, 8.46s/step, loss=1.359, acc=0.687, acc_len=1.64, thru=19.6, I=22.1, T=34.1, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:38:25,463] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:38:32,471] loop.py:45 INFO TRAIN_STEP step=9708 loss=1.317942 Training: 54%|█████▍ | 9708/18018 [21:08:18<18:52:01, 8.17s/step, loss=1.318, acc=0.696, acc_len=1.67, thru=14.9, I=19.5, T=37.3, wait=4.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:38:40,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:38:42,232] loop.py:45 INFO TRAIN_STEP step=9709 loss=1.356453 Training: 54%|█████▍ | 9709/18018 [21:08:28<19:57:51, 8.65s/step, loss=1.356, acc=0.688, acc_len=1.64, thru=15.0, I=15.7, T=30.5, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:38:42,351] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:38:48,755] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:38:50,376] loop.py:380 INFO TIMING step=9710: step=3.927s data=1.835s compute=3.756s [fwd=2.564s bwd=1.177s opt=0.003s] dispatch=4.207s [2026-10-19 10:38:50,376] loop.py:45 INFO TRAIN_STEP step=9710 loss=1.222592 Training: 54%|█████▍ | 9710/18018 [21:08:36<19:36:39, 8.50s/step, loss=1.223, acc=0.716, acc_len=1.74, thru=14.0, I=18.4, T=32.6, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:38:50,769] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:38:57,972] loop.py:45 INFO TRAIN_STEP step=9711 loss=1.277897 Training: 54%|█████▍ | 9711/18018 [21:08:44<18:58:59, 8.23s/step, loss=1.278, acc=0.705, acc_len=1.70, thru=12.3, I=13.0, T=32.0, wait=3.6s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:38:57,968] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9710: forward=2836.7ms backward=1113.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:04,241] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:39:05,041] loop.py:45 INFO TRAIN_STEP step=9712 loss=1.455429 Training: 54%|█████▍ | 9712/18018 [21:08:51<18:10:53, 7.88s/step, loss=1.455, acc=0.675, acc_len=1.60, thru=20.9, I=18.1, T=35.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:05,245] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:12,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:39:13,621] loop.py:45 INFO TRAIN_STEP step=9713 loss=1.336535 Training: 54%|█████▍ | 9713/18018 [21:08:59<18:39:45, 8.09s/step, loss=1.337, acc=0.693, acc_len=1.66, thru=14.4, I=17.9, T=31.8, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:13,667] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:19,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:39:20,212] loop.py:45 INFO TRAIN_STEP step=9714 loss=1.386324 Training: 54%|█████▍ | 9714/18018 [21:09:06<17:37:28, 7.64s/step, loss=1.386, acc=0.685, acc_len=1.62, thru=18.5, I=23.1, T=36.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:20,381] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:26,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:39:26,627] loop.py:380 INFO TIMING step=9715: step=3.692s data=1.484s compute=3.656s [fwd=2.632s bwd=1.009s opt=0.003s] dispatch=2.712s [2026-10-19 10:39:26,627] loop.py:45 INFO TRAIN_STEP step=9715 loss=1.405020 Training: 54%|█████▍ | 9715/18018 [21:09:12<16:46:22, 7.27s/step, loss=1.405, acc=0.685, acc_len=1.63, thru=19.6, I=22.4, T=34.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:27,047] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:34,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:39:34,705] loop.py:45 INFO TRAIN_STEP step=9716 loss=1.232156 Training: 54%|█████▍ | 9716/18018 [21:09:20<17:19:41, 7.51s/step, loss=1.232, acc=0.713, acc_len=1.73, thru=15.7, I=18.0, T=33.2, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:39:34,701] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9715: forward=2707.0ms backward=1091.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:34,796] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:41,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:39:42,691] loop.py:45 INFO TRAIN_STEP step=9717 loss=1.262129 Training: 54%|█████▍ | 9717/18018 [21:09:28<17:39:09, 7.66s/step, loss=1.262, acc=0.708, acc_len=1.71, thru=17.0, I=17.5, T=35.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:43,003] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:39:49,798] loop.py:45 INFO TRAIN_STEP step=9718 loss=1.314730 Training: 54%|█████▍ | 9718/18018 [21:09:35<17:16:24, 7.49s/step, loss=1.315, acc=0.696, acc_len=1.67, thru=18.2, I=19.6, T=32.3, wait=3.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:55,733] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:39:57,224] loop.py:45 INFO TRAIN_STEP step=9719 loss=1.326985 Training: 54%|█████▍ | 9719/18018 [21:09:43<17:13:29, 7.47s/step, loss=1.327, acc=0.700, acc_len=1.68, thru=17.9, I=20.1, T=29.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:39:57,261] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:02,885] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:40:03,613] loop.py:380 INFO TIMING step=9720: step=3.556s data=1.497s compute=3.528s [fwd=2.531s bwd=0.983s opt=0.003s] dispatch=2.821s [2026-10-19 10:40:03,614] loop.py:45 INFO TRAIN_STEP step=9720 loss=1.139875 Training: 54%|█████▍ | 9720/18018 [21:09:49<16:28:26, 7.15s/step, loss=1.140, acc=0.729, acc_len=1.78, thru=17.7, I=23.9, T=36.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:03,892] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:11,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:40:12,351] loop.py:45 INFO TRAIN_STEP step=9721 loss=1.319521 Training: 54%|█████▍ | 9721/18018 [21:09:58<17:34:18, 7.62s/step, loss=1.320, acc=0.700, acc_len=1.68, thru=14.0, I=16.8, T=31.3, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:40:12,345] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9720: forward=2861.3ms backward=1180.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:12,436] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 10:40:18,117] loop.py:45 INFO TRAIN_STEP step=9722 loss=1.206476 Training: 54%|█████▍ | 9722/18018 [21:10:04<16:16:58, 7.07s/step, loss=1.206, acc=0.714, acc_len=1.73, thru=19.4, I=22.2, T=34.6, wait=2.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:24,076] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:40:25,628] loop.py:45 INFO TRAIN_STEP step=9723 loss=1.315887 Training: 54%|█████▍ | 9723/18018 [21:10:11<16:35:17, 7.20s/step, loss=1.316, acc=0.700, acc_len=1.68, thru=17.6, I=20.5, T=31.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:26,080] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:30,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:40:32,271] loop.py:45 INFO TRAIN_STEP step=9724 loss=1.295716 Training: 54%|█████▍ | 9724/18018 [21:10:18<16:12:09, 7.03s/step, loss=1.296, acc=0.699, acc_len=1.68, thru=17.8, I=27.0, T=33.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:32,692] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:39,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:40:40,707] loop.py:380 INFO TIMING step=9725: step=3.880s data=1.733s compute=3.854s [fwd=2.774s bwd=1.065s opt=0.003s] dispatch=4.548s [2026-10-19 10:40:40,707] loop.py:45 INFO TRAIN_STEP step=9725 loss=1.326908 Training: 54%|█████▍ | 9725/18018 [21:10:26<17:10:19, 7.45s/step, loss=1.327, acc=0.698, acc_len=1.68, thru=16.0, I=18.3, T=33.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:41,099] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:46,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:40:47,627] loop.py:45 INFO TRAIN_STEP step=9726 loss=1.288077 Training: 54%|█████▍ | 9726/18018 [21:10:33<16:48:03, 7.29s/step, loss=1.288, acc=0.701, acc_len=1.68, thru=17.3, I=23.0, T=32.5, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:40:47,622] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9725: forward=2737.9ms backward=1135.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:47,831] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:54,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:40:54,525] loop.py:45 INFO TRAIN_STEP step=9727 loss=1.357742 Training: 54%|█████▍ | 9727/18018 [21:10:40<16:31:28, 7.18s/step, loss=1.358, acc=0.687, acc_len=1.64, thru=18.1, I=20.4, T=33.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:40:54,660] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:00,308] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:41:01,685] loop.py:45 INFO TRAIN_STEP step=9728 loss=1.414735 Training: 54%|█████▍ | 9728/18018 [21:10:47<16:30:41, 7.17s/step, loss=1.415, acc=0.685, acc_len=1.63, thru=18.2, I=19.3, T=35.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:01,817] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:07,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:41:08,739] loop.py:45 INFO TRAIN_STEP step=9729 loss=1.366577 Training: 54%|█████▍ | 9729/18018 [21:10:54<16:25:44, 7.14s/step, loss=1.367, acc=0.687, acc_len=1.64, thru=18.1, I=22.9, T=31.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:08,869] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 10:41:16,946] loop.py:380 INFO TIMING step=9730: step=4.535s data=2.035s compute=4.491s [fwd=3.162s bwd=1.313s opt=0.003s] dispatch=3.664s [2026-10-19 10:41:16,946] loop.py:45 INFO TRAIN_STEP step=9730 loss=1.367001 Training: 54%|█████▍ | 9730/18018 [21:11:03<17:10:01, 7.46s/step, loss=1.367, acc=0.691, acc_len=1.65, thru=12.4, I=16.7, T=28.2, wait=3.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:25,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:41:27,445] loop.py:45 INFO TRAIN_STEP step=9731 loss=1.391906 Training: 54%|█████▍ | 9731/18018 [21:11:13<19:16:02, 8.37s/step, loss=1.392, acc=0.684, acc_len=1.63, thru=15.2, I=13.3, T=31.9, wait=6.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:41:27,440] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9730: forward=2839.1ms backward=1113.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:27,623] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 [2026-10-19 10:41:35,279] loop.py:45 INFO TRAIN_STEP step=9732 loss=1.354725 Training: 54%|█████▍ | 9732/18018 [21:11:21<18:53:35, 8.21s/step, loss=1.355, acc=0.694, acc_len=1.66, thru=14.1, I=15.5, T=34.9, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:35,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:35,755] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:41:42,278] loop.py:45 INFO TRAIN_STEP step=9733 loss=1.391150 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:42,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 54%|█████▍ | 9733/18018 [21:11:28<18:03:21, 7.85s/step, loss=1.391, acc=0.684, acc_len=1.63, thru=19.2, I=18.2, T=33.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:42,717] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:48,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 10:41:50,322] loop.py:45 INFO TRAIN_STEP step=9734 loss=1.353227 Training: 54%|█████▍ | 9734/18018 [21:11:36<18:11:23, 7.90s/step, loss=1.353, acc=0.690, acc_len=1.65, thru=17.3, I=19.5, T=32.3, wait=4.1s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:41:50,790] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=5/64 [2026-10-19 10:41:58,512] loop.py:380 INFO TIMING step=9735: step=3.818s data=1.677s compute=3.784s [fwd=2.666s bwd=1.102s opt=0.003s] dispatch=4.365s [2026-10-19 10:41:58,512] loop.py:45 INFO TRAIN_STEP step=9735 loss=1.370358 Training: 54%|█████▍ | 9735/18018 [21:11:44<18:23:11, 7.99s/step, loss=1.370, acc=0.689, acc_len=1.64, thru=13.7, I=17.2, T=33.5, wait=4.4s, pool=61, epoch=2/2][2026-10-19 10:42:05,294] loop.py:45 INFO TRAIN_STEP step=9736 loss=1.334011 Training: 54%|█████▍ | 9736/18018 [21:11:51<17:32:51, 7.63s/step, loss=1.334, acc=0.692, acc_len=1.65, thru=18.3, I=15.0, T=33.3, wait=2.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:42:05,289] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9735: forward=2697.9ms backward=1097.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:12,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:42:12,783] loop.py:45 INFO TRAIN_STEP step=9737 loss=1.373984 Training: 54%|█████▍ | 9737/18018 [21:11:58<17:27:02, 7.59s/step, loss=1.374, acc=0.688, acc_len=1.64, thru=17.6, I=16.7, T=31.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:13,213] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:22,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:42:22,668] loop.py:45 INFO TRAIN_STEP step=9738 loss=1.383337 Training: 54%|█████▍ | 9738/18018 [21:12:08<19:02:05, 8.28s/step, loss=1.383, acc=0.689, acc_len=1.64, thru=13.6, I=14.5, T=28.3, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:22,696] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:28,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:42:29,927] loop.py:45 INFO TRAIN_STEP step=9739 loss=1.441823 Training: 54%|█████▍ | 9739/18018 [21:12:16<18:19:56, 7.97s/step, loss=1.442, acc=0.672, acc_len=1.59, thru=16.1, I=17.4, T=30.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:30,014] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 10:42:38,005] loop.py:380 INFO TIMING step=9740: step=4.161s data=1.822s compute=4.131s [fwd=2.887s bwd=1.228s opt=0.003s] dispatch=3.906s [2026-10-19 10:42:38,005] loop.py:45 INFO TRAIN_STEP step=9740 loss=1.452238 Training: 54%|█████▍ | 9740/18018 [21:12:24<18:24:07, 8.00s/step, loss=1.452, acc=0.679, acc_len=1.61, thru=12.2, I=16.8, T=30.8, wait=3.9s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:46,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:42:48,039] loop.py:45 INFO TRAIN_STEP step=9741 loss=1.404646 Training: 54%|█████▍ | 9741/18018 [21:12:34<19:48:08, 8.61s/step, loss=1.405, acc=0.681, acc_len=1.62, thru=15.7, I=15.3, T=29.8, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:42:48,034] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9740: forward=2955.2ms backward=1287.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:48,230] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:55,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:42:55,863] loop.py:45 INFO TRAIN_STEP step=9742 loss=1.258373 Training: 54%|█████▍ | 9742/18018 [21:12:41<19:15:16, 8.38s/step, loss=1.258, acc=0.712, acc_len=1.73, thru=14.7, I=16.5, T=31.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:42:56,013] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:03,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:43:04,512] loop.py:45 INFO TRAIN_STEP step=9743 loss=1.366470 Training: 54%|█████▍ | 9743/18018 [21:12:50<19:26:31, 8.46s/step, loss=1.366, acc=0.696, acc_len=1.68, thru=14.9, I=16.7, T=31.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:04,568] inference_manager.py:408 INFO Pool capacity available after 1.1s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:11,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:43:12,504] loop.py:45 INFO TRAIN_STEP step=9744 loss=1.381526 Training: 54%|█████▍ | 9744/18018 [21:12:58<19:07:07, 8.32s/step, loss=1.382, acc=0.691, acc_len=1.65, thru=15.7, I=17.9, T=29.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:12,937] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:19,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:43:19,766] loop.py:380 INFO TIMING step=9745: step=4.273s data=1.700s compute=4.236s [fwd=2.856s bwd=1.364s opt=0.003s] dispatch=2.978s [2026-10-19 10:43:19,767] loop.py:45 INFO TRAIN_STEP step=9745 loss=1.364213 Training: 54%|█████▍ | 9745/18018 [21:13:05<18:23:12, 8.00s/step, loss=1.364, acc=0.689, acc_len=1.64, thru=16.7, I=18.2, T=30.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:20,046] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:26,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:43:26,402] loop.py:45 INFO TRAIN_STEP step=9746 loss=1.234527 Training: 54%|█████▍ | 9746/18018 [21:13:12<17:26:39, 7.59s/step, loss=1.235, acc=0.710, acc_len=1.72, thru=17.5, I=18.5, T=35.5, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:43:26,397] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9745: forward=2538.4ms backward=1019.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:26,617] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:32,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:43:34,132] loop.py:45 INFO TRAIN_STEP step=9747 loss=1.307805 Training: 54%|█████▍ | 9747/18018 [21:13:20<17:32:14, 7.63s/step, loss=1.308, acc=0.701, acc_len=1.69, thru=17.3, I=20.4, T=33.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:34,354] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:38,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:43:40,105] loop.py:45 INFO TRAIN_STEP step=9748 loss=1.223516 Training: 54%|█████▍ | 9748/18018 [21:13:26<16:23:29, 7.14s/step, loss=1.224, acc=0.714, acc_len=1.73, thru=19.8, I=26.1, T=33.7, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:40,435] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:46,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:43:47,294] loop.py:45 INFO TRAIN_STEP step=9749 loss=1.390036 Training: 54%|█████▍ | 9749/18018 [21:13:33<16:25:32, 7.15s/step, loss=1.390, acc=0.684, acc_len=1.63, thru=16.4, I=20.6, T=32.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:47,395] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 10:43:57,547] loop.py:380 INFO TIMING step=9750: step=4.198s data=1.952s compute=4.165s [fwd=3.042s bwd=1.109s opt=0.003s] dispatch=6.046s [2026-10-19 10:43:57,548] loop.py:45 INFO TRAIN_STEP step=9750 loss=1.546067 Training: 54%|█████▍ | 9750/18018 [21:13:43<18:33:37, 8.08s/step, loss=1.546, acc=0.658, acc_len=1.54, thru=11.2, I=11.2, T=30.5, wait=6.0s, pool=53, epoch=2/2][2026-10-19 10:43:57,551] loop.py:419 INFO Saving checkpoint at step 9750... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:43:57,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 10:44:00,964] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009751 (TrainerActor pid=9310) [2026-10-19 10:44:00,970] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009010 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:02,371] inference_manager.py:408 INFO Pool capacity available after 4.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:09,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 10:44:00,965] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0009751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 10:44:02,045] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009010 [2026-10-19 10:44:10,944] loop.py:45 INFO TRAIN_STEP step=9751 loss=1.311235 Training: 54%|█████▍ | 9751/18018 [21:13:57<22:13:12, 9.68s/step, loss=1.311, acc=0.698, acc_len=1.67, thru=14.9, I=17.5, T=26.6, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:44:10,940] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9750: forward=3359.4ms backward=1404.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:11,229] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:44:18,301] loop.py:45 INFO TRAIN_STEP step=9752 loss=1.286597 Training: 54%|█████▍ | 9752/18018 [21:14:04<20:37:07, 8.98s/step, loss=1.287, acc=0.703, acc_len=1.69, thru=15.2, I=16.2, T=34.0, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:29,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:44:30,731] loop.py:45 INFO TRAIN_STEP step=9753 loss=1.522259 Training: 54%|█████▍ | 9753/18018 [21:14:16<22:59:44, 10.02s/step, loss=1.522, acc=0.660, acc_len=1.55, thru=11.6, I=12.0, T=28.9, wait=8.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:31,054] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:37,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:44:38,247] loop.py:45 INFO TRAIN_STEP step=9754 loss=1.399474 Training: 54%|█████▍ | 9754/18018 [21:14:24<21:16:06, 9.27s/step, loss=1.399, acc=0.683, acc_len=1.62, thru=15.7, I=18.4, T=33.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:38,323] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:45,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:44:46,073] loop.py:380 INFO TIMING step=9755: step=4.591s data=2.205s compute=4.556s [fwd=3.154s bwd=1.387s opt=0.003s] dispatch=3.227s [2026-10-19 10:44:46,073] loop.py:45 INFO TRAIN_STEP step=9755 loss=1.266702 Training: 54%|█████▍ | 9755/18018 [21:14:32<20:16:37, 8.83s/step, loss=1.267, acc=0.704, acc_len=1.70, thru=18.3, I=20.2, T=27.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:46,478] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:52,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:44:53,091] loop.py:45 INFO TRAIN_STEP step=9756 loss=1.327563 Training: 54%|█████▍ | 9756/18018 [21:14:39<19:01:23, 8.29s/step, loss=1.328, acc=0.700, acc_len=1.68, thru=17.7, I=19.1, T=32.4, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:44:53,087] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9755: forward=2795.2ms backward=1109.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:44:53,421] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:02,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:45:02,864] loop.py:45 INFO TRAIN_STEP step=9757 loss=1.277708 Training: 54%|█████▍ | 9757/18018 [21:14:48<20:02:29, 8.73s/step, loss=1.278, acc=0.704, acc_len=1.69, thru=13.3, I=13.4, T=33.6, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:03,251] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:08,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:45:09,258] loop.py:45 INFO TRAIN_STEP step=9758 loss=1.451297 Training: 54%|█████▍ | 9758/18018 [21:14:55<18:25:42, 8.03s/step, loss=1.451, acc=0.673, acc_len=1.59, thru=19.2, I=23.0, T=34.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:09,713] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:45:16,969] loop.py:45 INFO TRAIN_STEP step=9759 loss=1.226834 Training: 54%|█████▍ | 9759/18018 [21:15:03<18:12:20, 7.94s/step, loss=1.227, acc=0.712, acc_len=1.72, thru=15.7, I=18.5, T=34.3, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:22,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:45:23,933] loop.py:380 INFO TIMING step=9760: step=3.950s data=1.710s compute=3.909s [fwd=2.775s bwd=1.118s opt=0.003s] dispatch=3.005s [2026-10-19 10:45:23,933] loop.py:45 INFO TRAIN_STEP step=9760 loss=1.348985 Training: 54%|█████▍ | 9760/18018 [21:15:10<17:32:06, 7.64s/step, loss=1.349, acc=0.691, acc_len=1.65, thru=17.6, I=20.4, T=32.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:24,254] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:45:30,710] loop.py:45 INFO TRAIN_STEP step=9761 loss=1.318849 Training: 54%|█████▍ | 9761/18018 [21:15:16<16:56:05, 7.38s/step, loss=1.319, acc=0.697, acc_len=1.68, thru=16.7, I=21.0, T=33.3, wait=2.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:45:30,706] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9760: forward=2687.5ms backward=1100.2ms [2026-10-19 10:45:37,858] loop.py:45 INFO TRAIN_STEP step=9762 loss=1.256988 Training: 54%|█████▍ | 9762/18018 [21:15:23<16:46:20, 7.31s/step, loss=1.257, acc=0.709, acc_len=1.72, thru=17.4, I=18.3, T=31.6, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:45,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:45:46,150] loop.py:45 INFO TRAIN_STEP step=9763 loss=1.370701 Training: 54%|█████▍ | 9763/18018 [21:15:32<17:26:38, 7.61s/step, loss=1.371, acc=0.691, acc_len=1.65, thru=15.3, I=17.1, T=30.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:46,222] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:52,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:45:53,205] loop.py:45 INFO TRAIN_STEP step=9764 loss=1.256273 Training: 54%|█████▍ | 9764/18018 [21:15:39<17:03:39, 7.44s/step, loss=1.256, acc=0.708, acc_len=1.71, thru=16.1, I=18.5, T=30.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:45:53,685] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:46:00,139] loop.py:380 INFO TIMING step=9765: step=3.673s data=1.502s compute=3.634s [fwd=2.538s bwd=1.080s opt=0.003s] dispatch=3.252s [2026-10-19 10:46:00,139] loop.py:45 INFO TRAIN_STEP step=9765 loss=1.486548 Training: 54%|█████▍ | 9765/18018 [21:15:46<16:42:39, 7.29s/step, loss=1.487, acc=0.667, acc_len=1.57, thru=16.4, I=19.7, T=34.8, wait=3.3s, pool=61, epoch=2/2][2026-10-19 10:46:08,423] loop.py:45 INFO TRAIN_STEP step=9766 loss=1.325750 Training: 54%|█████▍ | 9766/18018 [21:15:54<17:23:29, 7.59s/step, loss=1.326, acc=0.697, acc_len=1.68, thru=14.5, I=13.9, T=31.9, wait=4.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:46:08,419] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9765: forward=2878.1ms backward=1086.8ms [2026-10-19 10:46:17,382] loop.py:45 INFO TRAIN_STEP step=9767 loss=1.267456 Training: 54%|█████▍ | 9767/18018 [21:16:03<18:20:03, 8.00s/step, loss=1.267, acc=0.704, acc_len=1.70, thru=15.2, I=12.7, T=30.4, wait=4.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:46:22,883] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:46:24,571] loop.py:45 INFO TRAIN_STEP step=9768 loss=1.311638 Training: 54%|█████▍ | 9768/18018 [21:16:10<17:46:30, 7.76s/step, loss=1.312, acc=0.700, acc_len=1.68, thru=18.5, I=18.5, T=31.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:46:24,898] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:46:31,984] loop.py:45 INFO TRAIN_STEP step=9769 loss=1.300766 Training: 54%|█████▍ | 9769/18018 [21:16:18<17:32:14, 7.65s/step, loss=1.301, acc=0.701, acc_len=1.69, thru=12.7, I=17.1, T=35.9, wait=3.8s, pool=37, epoch=2/2][2026-10-19 10:46:42,987] loop.py:380 INFO TIMING step=9770: step=3.828s data=1.602s compute=3.785s [fwd=2.714s bwd=1.055s opt=0.003s] dispatch=7.164s [2026-10-19 10:46:42,988] loop.py:45 INFO TRAIN_STEP step=9770 loss=1.305500 Training: 54%|█████▍ | 9770/18018 [21:16:29<19:50:10, 8.66s/step, loss=1.306, acc=0.703, acc_len=1.69, thru=11.6, I=12.9, T=33.4, wait=7.2s, pool=33, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 10:46:43,056] eagle_store.py:517 WARNING batch_get_buffer missing keys (k312642-9ec1ad68_hs, k312642-9ec1ad68_ids, k312642-9ec1ad68_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:46:52,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 10:46:43,059] eagle_store.py:517 WARNING batch_get_buffer missing keys (k312641-babbe124_hs, k312641-babbe124_ids, k312641-babbe124_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 10:46:53,302] loop.py:45 INFO TRAIN_STEP step=9771 loss=1.365350 Training: 54%|█████▍ | 9771/18018 [21:16:39<20:58:25, 9.16s/step, loss=1.365, acc=0.688, acc_len=1.64, thru=17.7, I=14.2, T=32.1, wait=6.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:46:53,298] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9770: forward=2777.5ms backward=1152.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:46:53,433] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:46:57,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:46:58,995] loop.py:45 INFO TRAIN_STEP step=9772 loss=1.374310 Training: 54%|█████▍ | 9772/18018 [21:16:45<18:35:28, 8.12s/step, loss=1.374, acc=0.689, acc_len=1.64, thru=20.6, I=25.6, T=36.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:46:59,336] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:47:07,428] loop.py:45 INFO TRAIN_STEP step=9773 loss=1.240671 Training: 54%|█████▍ | 9773/18018 [21:16:53<18:48:17, 8.21s/step, loss=1.241, acc=0.713, acc_len=1.73, thru=11.9, I=14.3, T=31.7, wait=4.4s, pool=33, epoch=2/2][2026-10-19 10:47:16,794] loop.py:45 INFO TRAIN_STEP step=9774 loss=1.215908 Training: 54%|█████▍ | 9774/18018 [21:17:02<19:35:47, 8.56s/step, loss=1.216, acc=0.716, acc_len=1.74, thru=15.2, I=15.3, T=32.5, wait=5.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:47:24,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:47:25,785] loop.py:380 INFO TIMING step=9775: step=4.058s data=1.743s compute=4.024s [fwd=2.869s bwd=1.139s opt=0.003s] dispatch=4.926s [2026-10-19 10:47:25,785] loop.py:45 INFO TRAIN_STEP step=9775 loss=1.315850 Training: 54%|█████▍ | 9775/18018 [21:17:11<19:53:35, 8.69s/step, loss=1.316, acc=0.697, acc_len=1.67, thru=16.9, I=16.5, T=31.5, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:47:26,003] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:47:32,362] loop.py:45 INFO TRAIN_STEP step=9776 loss=1.335745 Training: 54%|█████▍ | 9776/18018 [21:17:18<18:26:27, 8.05s/step, loss=1.336, acc=0.694, acc_len=1.66, thru=15.7, I=23.4, T=34.1, wait=2.8s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:47:32,357] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9775: forward=2633.7ms backward=1057.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:47:42,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:47:43,146] loop.py:45 INFO TRAIN_STEP step=9777 loss=1.398268 Training: 54%|█████▍ | 9777/18018 [21:17:29<20:18:42, 8.87s/step, loss=1.398, acc=0.686, acc_len=1.63, thru=14.8, I=13.8, T=27.2, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:47:43,177] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:47:47,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:47:48,853] loop.py:45 INFO TRAIN_STEP step=9778 loss=1.376659 Training: 54%|█████▍ | 9778/18018 [21:17:34<18:08:06, 7.92s/step, loss=1.377, acc=0.689, acc_len=1.65, thru=18.6, I=21.2, T=33.1, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:47:49,162] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:47:54,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:47:55,399] loop.py:45 INFO TRAIN_STEP step=9779 loss=1.297224 Training: 54%|█████▍ | 9779/18018 [21:17:41<17:11:18, 7.51s/step, loss=1.297, acc=0.705, acc_len=1.70, thru=18.9, I=22.9, T=37.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:47:55,792] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:03,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:48:03,904] loop.py:380 INFO TIMING step=9780: step=3.989s data=1.681s compute=3.958s [fwd=2.783s bwd=1.159s opt=0.003s] dispatch=4.506s [2026-10-19 10:48:03,904] loop.py:45 INFO TRAIN_STEP step=9780 loss=1.356277 Training: 54%|█████▍ | 9780/18018 [21:17:49<17:52:09, 7.81s/step, loss=1.356, acc=0.693, acc_len=1.66, thru=16.4, I=17.8, T=32.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:04,130] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:09,492] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:48:10,517] loop.py:45 INFO TRAIN_STEP step=9781 loss=1.312650 Training: 54%|█████▍ | 9781/18018 [21:17:56<17:02:42, 7.45s/step, loss=1.313, acc=0.699, acc_len=1.68, thru=18.4, I=20.6, T=34.3, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:48:10,512] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9780: forward=2656.1ms backward=1027.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:11,000] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:48:18,086] loop.py:45 INFO TRAIN_STEP step=9782 loss=1.299979 Training: 54%|█████▍ | 9782/18018 [21:18:04<17:07:36, 7.49s/step, loss=1.300, acc=0.698, acc_len=1.68, thru=14.9, I=20.7, T=33.1, wait=3.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:24,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:48:25,888] loop.py:45 INFO TRAIN_STEP step=9783 loss=1.320621 Training: 54%|█████▍ | 9783/18018 [21:18:11<17:20:28, 7.58s/step, loss=1.321, acc=0.699, acc_len=1.67, thru=16.6, I=18.6, T=28.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:26,210] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:32,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:48:32,479] loop.py:45 INFO TRAIN_STEP step=9784 loss=1.218726 Training: 54%|█████▍ | 9784/18018 [21:18:18<16:39:35, 7.28s/step, loss=1.219, acc=0.714, acc_len=1.73, thru=16.0, I=21.1, T=31.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:32,545] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:37,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:48:38,439] loop.py:380 INFO TIMING step=9785: step=3.587s data=1.546s compute=3.557s [fwd=2.519s bwd=1.023s opt=0.003s] dispatch=2.362s [2026-10-19 10:48:38,439] loop.py:45 INFO TRAIN_STEP step=9785 loss=1.362662 Training: 54%|█████▍ | 9785/18018 [21:18:24<15:44:57, 6.89s/step, loss=1.363, acc=0.689, acc_len=1.64, thru=19.9, I=22.9, T=35.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:38,565] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 10:48:47,896] loop.py:45 INFO TRAIN_STEP step=9786 loss=1.417064 Training: 54%|█████▍ | 9786/18018 [21:18:33<17:30:45, 7.66s/step, loss=1.417, acc=0.681, acc_len=1.61, thru=13.3, I=13.6, T=31.4, wait=5.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:48:47,891] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9785: forward=2786.9ms backward=1153.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:56,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:48:56,508] loop.py:45 INFO TRAIN_STEP step=9787 loss=1.316375 Training: 54%|█████▍ | 9787/18018 [21:18:42<18:09:41, 7.94s/step, loss=1.316, acc=0.697, acc_len=1.67, thru=15.5, I=14.6, T=30.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:48:56,971] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:04,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:49:04,490] loop.py:45 INFO TRAIN_STEP step=9788 loss=1.268318 Training: 54%|█████▍ | 9788/18018 [21:18:50<18:11:11, 7.96s/step, loss=1.268, acc=0.710, acc_len=1.72, thru=16.9, I=16.8, T=32.6, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:04,665] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:12,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:49:13,192] loop.py:45 INFO TRAIN_STEP step=9789 loss=1.339685 Training: 54%|█████▍ | 9789/18018 [21:18:59<18:41:48, 8.18s/step, loss=1.340, acc=0.697, acc_len=1.67, thru=14.5, I=16.5, T=29.7, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:13,270] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 10:49:20,062] loop.py:380 INFO TIMING step=9790: step=4.272s data=1.787s compute=4.243s [fwd=2.916s bwd=1.311s opt=0.003s] dispatch=2.588s [2026-10-19 10:49:20,062] loop.py:45 INFO TRAIN_STEP step=9790 loss=1.323153 Training: 54%|█████▍ | 9790/18018 [21:19:06<17:47:40, 7.79s/step, loss=1.323, acc=0.692, acc_len=1.65, thru=16.6, I=16.7, T=30.0, wait=2.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:26,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:49:28,492] loop.py:45 INFO TRAIN_STEP step=9791 loss=1.466720 (TrainerActor pid=9310) [2026-10-19 10:49:28,487] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9790: forward=2843.4ms backward=1112.3ms Training: 54%|█████▍ | 9791/18018 [21:19:14<18:14:12, 7.98s/step, loss=1.467, acc=0.672, acc_len=1.59, thru=17.6, I=17.2, T=31.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:28,592] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=33/64 [2026-10-19 10:49:37,404] loop.py:45 INFO TRAIN_STEP step=9792 loss=1.231385 Training: 54%|█████▍ | 9792/18018 [21:19:23<18:52:27, 8.26s/step, loss=1.231, acc=0.714, acc_len=1.73, thru=14.1, I=16.5, T=32.1, wait=4.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:46,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:49:47,547] loop.py:45 INFO TRAIN_STEP step=9793 loss=1.290261 Training: 54%|█████▍ | 9793/18018 [21:19:33<20:09:40, 8.82s/step, loss=1.290, acc=0.705, acc_len=1.70, thru=13.5, I=13.3, T=30.5, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:47,789] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:53,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:49:54,731] loop.py:45 INFO TRAIN_STEP step=9794 loss=1.406827 Training: 54%|█████▍ | 9794/18018 [21:19:40<19:02:00, 8.33s/step, loss=1.407, acc=0.682, acc_len=1.62, thru=15.9, I=20.7, T=31.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:49:54,807] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 10:50:00,487] loop.py:380 INFO TIMING step=9795: step=3.421s data=1.431s compute=3.388s [fwd=2.439s bwd=0.934s opt=0.003s] dispatch=2.328s [2026-10-19 10:50:00,488] loop.py:45 INFO TRAIN_STEP step=9795 loss=1.406327 Training: 54%|█████▍ | 9795/18018 [21:19:46<17:15:59, 7.56s/step, loss=1.406, acc=0.683, acc_len=1.63, thru=17.9, I=20.1, T=37.4, wait=2.3s, pool=45, epoch=2/2][2026-10-19 10:50:14,322] loop.py:45 INFO TRAIN_STEP step=9796 loss=1.338371 Training: 54%|█████▍ | 9796/18018 [21:20:00<21:33:49, 9.44s/step, loss=1.338, acc=0.695, acc_len=1.67, thru=9.3, I=9.6, T=28.6, wait=9.3s, pool=49, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 10:50:14,317] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9795: forward=3115.0ms backward=1304.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:50:23,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 10:50:24,291] loop.py:45 INFO TRAIN_STEP step=9797 loss=1.306478 Training: 54%|█████▍ | 9797/18018 [21:20:10<21:55:19, 9.60s/step, loss=1.306, acc=0.704, acc_len=1.70, thru=15.1, I=13.4, T=27.6, wait=5.3s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:50:24,761] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:50:29,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:50:30,536] loop.py:45 INFO TRAIN_STEP step=9798 loss=1.369592 Training: 54%|█████▍ | 9798/18018 [21:20:16<19:37:19, 8.59s/step, loss=1.370, acc=0.690, acc_len=1.64, thru=19.7, I=21.5, T=33.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:50:30,823] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:50:35,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:50:37,105] loop.py:45 INFO TRAIN_STEP step=9799 loss=1.288876 Training: 54%|█████▍ | 9799/18018 [21:20:23<18:13:55, 7.99s/step, loss=1.289, acc=0.701, acc_len=1.68, thru=18.9, I=25.2, T=30.3, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:50:37,581] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:50:45,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:50:45,607] loop.py:380 INFO TIMING step=9800: step=4.519s data=2.018s compute=4.475s [fwd=3.115s bwd=1.346s opt=0.003s] dispatch=3.975s [2026-10-19 10:50:45,607] loop.py:45 INFO TRAIN_STEP step=9800 loss=1.177019 Training: 54%|█████▍ | 9800/18018 [21:20:31<18:35:01, 8.14s/step, loss=1.177, acc=0.727, acc_len=1.78, thru=16.0, I=16.9, T=28.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:50:45,698] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 10:50:53,281] loop.py:45 INFO TRAIN_STEP step=9801 loss=1.394587 Training: 54%|█████▍ | 9801/18018 [21:20:39<18:15:42, 8.00s/step, loss=1.395, acc=0.683, acc_len=1.62, thru=15.5, I=15.6, T=30.1, wait=3.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:50:53,277] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9800: forward=2897.8ms backward=1304.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:01,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:51:02,170] loop.py:45 INFO TRAIN_STEP step=9802 loss=1.404653 Training: 54%|█████▍ | 9802/18018 [21:20:48<18:52:03, 8.27s/step, loss=1.405, acc=0.680, acc_len=1.62, thru=15.1, I=17.0, T=29.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:02,638] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:12,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:51:13,139] loop.py:45 INFO TRAIN_STEP step=9803 loss=1.362236 Training: 54%|█████▍ | 9803/18018 [21:20:59<20:42:54, 9.08s/step, loss=1.362, acc=0.691, acc_len=1.65, thru=11.9, I=12.2, T=29.5, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:13,241] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:19,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:51:20,749] loop.py:45 INFO TRAIN_STEP step=9804 loss=1.307604 Training: 54%|█████▍ | 9804/18018 [21:21:06<19:42:33, 8.64s/step, loss=1.308, acc=0.700, acc_len=1.69, thru=17.1, I=19.1, T=31.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:21,035] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:51:27,322] loop.py:380 INFO TIMING step=9805: step=3.731s data=1.667s compute=3.681s [fwd=2.560s bwd=1.107s opt=0.003s] dispatch=2.832s [2026-10-19 10:51:27,322] loop.py:45 INFO TRAIN_STEP step=9805 loss=1.211797 Training: 54%|█████▍ | 9805/18018 [21:21:13<18:17:31, 8.02s/step, loss=1.212, acc=0.714, acc_len=1.74, thru=17.2, I=18.9, T=34.3, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:35,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:51:35,903] loop.py:45 INFO TRAIN_STEP step=9806 loss=1.278340 Training: 54%|█████▍ | 9806/18018 [21:21:21<18:40:35, 8.19s/step, loss=1.278, acc=0.707, acc_len=1.71, thru=16.4, I=17.3, T=34.4, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:51:35,898] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9805: forward=2503.0ms backward=1139.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:36,167] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:51:42,485] loop.py:45 INFO TRAIN_STEP step=9807 loss=1.266327 Training: 54%|█████▍ | 9807/18018 [21:21:28<17:34:29, 7.71s/step, loss=1.266, acc=0.704, acc_len=1.70, thru=18.9, I=18.4, T=32.4, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:51,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:51:51,710] loop.py:45 INFO TRAIN_STEP step=9808 loss=1.394708 Training: 54%|█████▍ | 9808/18018 [21:21:37<18:36:41, 8.16s/step, loss=1.395, acc=0.687, acc_len=1.64, thru=14.7, I=14.8, T=28.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:51:52,203] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:51:59,449] loop.py:45 INFO TRAIN_STEP step=9809 loss=1.353828 Training: 54%|█████▍ | 9809/18018 [21:21:45<18:19:28, 8.04s/step, loss=1.354, acc=0.695, acc_len=1.67, thru=13.5, I=17.6, T=28.1, wait=3.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:11,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:52:13,312] loop.py:380 INFO TIMING step=9810: step=5.266s data=2.797s compute=5.197s [fwd=3.766s bwd=1.415s opt=0.003s] dispatch=8.583s [2026-10-19 10:52:13,312] loop.py:45 INFO TRAIN_STEP step=9810 loss=1.357046 Training: 54%|█████▍ | 9810/18018 [21:21:59<22:18:16, 9.78s/step, loss=1.357, acc=0.691, acc_len=1.65, thru=11.5, I=13.5, T=24.3, wait=8.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:13,788] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:19,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:52:20,033] loop.py:45 INFO TRAIN_STEP step=9811 loss=1.297386 Training: 54%|█████▍ | 9811/18018 [21:22:06<20:12:27, 8.86s/step, loss=1.297, acc=0.701, acc_len=1.68, thru=15.0, I=17.1, T=33.5, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:52:20,028] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9810: forward=2709.9ms backward=1064.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:20,478] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:26,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:52:28,454] loop.py:45 INFO TRAIN_STEP step=9812 loss=1.190702 Training: 54%|█████▍ | 9812/18018 [21:22:14<19:54:08, 8.73s/step, loss=1.191, acc=0.717, acc_len=1.74, thru=15.8, I=18.7, T=30.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:28,506] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:34,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:52:36,347] loop.py:45 INFO TRAIN_STEP step=9813 loss=1.402516 Training: 54%|█████▍ | 9813/18018 [21:22:22<19:19:34, 8.48s/step, loss=1.403, acc=0.682, acc_len=1.62, thru=14.9, I=20.2, T=29.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:36,819] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:41,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:52:43,124] loop.py:45 INFO TRAIN_STEP step=9814 loss=1.409488 Training: 54%|█████▍ | 9814/18018 [21:22:29<18:09:41, 7.97s/step, loss=1.409, acc=0.684, acc_len=1.62, thru=16.9, I=22.1, T=30.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:43,374] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:52:51,471] loop.py:380 INFO TIMING step=9815: step=3.668s data=1.497s compute=3.633s [fwd=2.598s bwd=1.020s opt=0.003s] dispatch=4.669s [2026-10-19 10:52:51,471] loop.py:45 INFO TRAIN_STEP step=9815 loss=1.571978 Training: 54%|█████▍ | 9815/18018 [21:22:37<18:25:00, 8.08s/step, loss=1.572, acc=0.656, acc_len=1.53, thru=13.7, I=17.5, T=34.9, wait=4.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:58,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:52:59,206] loop.py:45 INFO TRAIN_STEP step=9816 loss=1.339083 Training: 54%|█████▍ | 9816/18018 [21:22:45<18:10:40, 7.98s/step, loss=1.339, acc=0.695, acc_len=1.67, thru=16.7, I=18.2, T=31.1, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:52:59,200] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9815: forward=2895.2ms backward=1154.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:52:59,287] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:05,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:53:06,352] loop.py:45 INFO TRAIN_STEP step=9817 loss=1.446755 Training: 54%|█████▍ | 9817/18018 [21:22:52<17:36:19, 7.73s/step, loss=1.447, acc=0.674, acc_len=1.59, thru=17.9, I=21.1, T=28.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:06,730] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:12,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:53:13,616] loop.py:45 INFO TRAIN_STEP step=9818 loss=1.314667 Training: 54%|█████▍ | 9818/18018 [21:22:59<17:17:12, 7.59s/step, loss=1.315, acc=0.700, acc_len=1.68, thru=16.9, I=21.0, T=32.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:14,033] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:19,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:53:20,891] loop.py:45 INFO TRAIN_STEP step=9819 loss=1.244273 Training: 54%|█████▍ | 9819/18018 [21:23:06<17:04:12, 7.50s/step, loss=1.244, acc=0.712, acc_len=1.73, thru=16.9, I=20.4, T=33.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:21,052] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:27,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:53:28,952] loop.py:380 INFO TIMING step=9820: step=4.272s data=1.923s compute=4.221s [fwd=2.928s bwd=1.278s opt=0.003s] dispatch=3.778s [2026-10-19 10:53:28,952] loop.py:45 INFO TRAIN_STEP step=9820 loss=1.394234 Training: 55%|█████▍ | 9820/18018 [21:23:15<17:27:12, 7.66s/step, loss=1.394, acc=0.689, acc_len=1.64, thru=14.8, I=19.5, T=30.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:29,325] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:53:35,901] loop.py:45 INFO TRAIN_STEP step=9821 loss=1.343250 Training: 55%|█████▍ | 9821/18018 [21:23:21<16:57:51, 7.45s/step, loss=1.343, acc=0.692, acc_len=1.66, thru=15.8, I=17.9, T=36.0, wait=3.4s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:53:35,896] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9820: forward=2503.6ms backward=1000.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:41,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:53:42,400] loop.py:45 INFO TRAIN_STEP step=9822 loss=1.308601 Training: 55%|█████▍ | 9822/18018 [21:23:28<16:18:38, 7.16s/step, loss=1.309, acc=0.698, acc_len=1.67, thru=19.5, I=20.7, T=31.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:42,498] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:46,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:53:48,997] loop.py:45 INFO TRAIN_STEP step=9823 loss=1.259762 Training: 55%|█████▍ | 9823/18018 [21:23:35<15:55:15, 6.99s/step, loss=1.260, acc=0.714, acc_len=1.73, thru=17.7, I=22.3, T=30.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:49,487] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:54,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:53:56,516] loop.py:45 INFO TRAIN_STEP step=9824 loss=1.240216 Training: 55%|█████▍ | 9824/18018 [21:23:42<16:16:42, 7.15s/step, loss=1.240, acc=0.711, acc_len=1.72, thru=14.3, I=25.0, T=28.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:53:56,644] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:04,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:54:04,700] loop.py:380 INFO TIMING step=9825: step=3.885s data=1.713s compute=3.842s [fwd=2.736s bwd=1.091s opt=0.003s] dispatch=4.291s [2026-10-19 10:54:04,700] loop.py:45 INFO TRAIN_STEP step=9825 loss=1.249555 Training: 55%|█████▍ | 9825/18018 [21:23:50<16:58:54, 7.46s/step, loss=1.250, acc=0.712, acc_len=1.72, thru=15.9, I=17.3, T=33.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:04,735] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:09,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:54:11,134] loop.py:45 INFO TRAIN_STEP step=9826 loss=1.313745 Training: 55%|█████▍ | 9826/18018 [21:23:57<16:16:43, 7.15s/step, loss=1.314, acc=0.698, acc_len=1.68, thru=19.3, I=20.1, T=35.4, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:54:11,130] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9825: forward=2535.9ms backward=1036.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:11,467] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:17,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:54:17,742] loop.py:45 INFO TRAIN_STEP step=9827 loss=1.303790 Training: 55%|█████▍ | 9827/18018 [21:24:03<15:54:12, 6.99s/step, loss=1.304, acc=0.702, acc_len=1.69, thru=18.2, I=21.2, T=33.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:18,048] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:24,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:54:25,239] loop.py:45 INFO TRAIN_STEP step=9828 loss=1.389731 Training: 55%|█████▍ | 9828/18018 [21:24:11<16:14:53, 7.14s/step, loss=1.390, acc=0.687, acc_len=1.64, thru=17.4, I=18.5, T=34.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:25,354] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:33,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:54:34,477] loop.py:45 INFO TRAIN_STEP step=9829 loss=1.371681 Training: 55%|█████▍ | 9829/18018 [21:24:20<17:40:37, 7.77s/step, loss=1.372, acc=0.692, acc_len=1.65, thru=14.0, I=16.7, T=28.8, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:34,884] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:54:41,238] loop.py:380 INFO TIMING step=9830: step=4.294s data=1.938s compute=4.266s [fwd=2.925s bwd=1.326s opt=0.003s] dispatch=2.456s [2026-10-19 10:54:41,238] loop.py:45 INFO TRAIN_STEP step=9830 loss=1.349636 Training: 55%|█████▍ | 9830/18018 [21:24:27<16:59:05, 7.47s/step, loss=1.350, acc=0.692, acc_len=1.65, thru=16.0, I=18.7, T=29.8, wait=2.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:49,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:54:51,388] loop.py:45 INFO TRAIN_STEP step=9831 loss=1.349695 Training: 55%|█████▍ | 9831/18018 [21:24:37<18:48:47, 8.27s/step, loss=1.350, acc=0.691, acc_len=1.65, thru=15.2, I=16.0, T=26.4, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:54:51,384] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9830: forward=3414.7ms backward=1380.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:51,623] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:55,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:54:57,425] loop.py:45 INFO TRAIN_STEP step=9832 loss=1.339037 Training: 55%|█████▍ | 9832/18018 [21:24:43<17:17:04, 7.60s/step, loss=1.339, acc=0.693, acc_len=1.65, thru=17.7, I=25.2, T=34.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:54:57,493] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:03,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:55:04,608] loop.py:45 INFO TRAIN_STEP step=9833 loss=1.434951 Training: 55%|█████▍ | 9833/18018 [21:24:50<16:59:55, 7.48s/step, loss=1.435, acc=0.674, acc_len=1.59, thru=17.0, I=21.3, T=33.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:04,645] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 10:55:11,539] loop.py:45 INFO TRAIN_STEP step=9834 loss=1.421194 Training: 55%|█████▍ | 9834/18018 [21:24:57<16:37:19, 7.31s/step, loss=1.421, acc=0.681, acc_len=1.61, thru=19.0, I=18.5, T=35.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:11,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:12,035] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:55:19,775] loop.py:380 INFO TIMING step=9835: step=3.980s data=1.723s compute=3.952s [fwd=2.752s bwd=1.185s opt=0.003s] dispatch=4.247s [2026-10-19 10:55:19,775] loop.py:45 INFO TRAIN_STEP step=9835 loss=1.347125 Training: 55%|█████▍ | 9835/18018 [21:25:05<17:15:10, 7.59s/step, loss=1.347, acc=0.694, acc_len=1.66, thru=14.6, I=15.8, T=32.2, wait=4.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:27,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:55:28,296] loop.py:45 INFO TRAIN_STEP step=9836 loss=1.272349 Training: 55%|█████▍ | 9836/18018 [21:25:14<17:53:05, 7.87s/step, loss=1.272, acc=0.705, acc_len=1.71, thru=17.2, I=18.0, T=33.0, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:55:28,292] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9835: forward=2750.5ms backward=1076.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:28,591] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:34,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:55:35,430] loop.py:45 INFO TRAIN_STEP step=9837 loss=1.325644 Training: 55%|█████▍ | 9837/18018 [21:25:21<17:22:49, 7.65s/step, loss=1.326, acc=0.692, acc_len=1.66, thru=16.9, I=23.1, T=31.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:35,624] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:42,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:55:44,022] loop.py:45 INFO TRAIN_STEP step=9838 loss=1.300297 Training: 55%|█████▍ | 9838/18018 [21:25:30<18:01:23, 7.93s/step, loss=1.300, acc=0.699, acc_len=1.69, thru=13.3, I=16.8, T=28.4, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:44,189] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 10:55:50,773] loop.py:45 INFO TRAIN_STEP step=9839 loss=1.259568 Training: 55%|█████▍ | 9839/18018 [21:25:36<17:12:50, 7.58s/step, loss=1.260, acc=0.707, acc_len=1.71, thru=17.0, I=19.5, T=28.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:50,757] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:51,259] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:56,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:55:57,756] loop.py:380 INFO TIMING step=9840: step=3.854s data=1.667s compute=3.821s [fwd=2.686s bwd=1.120s opt=0.003s] dispatch=3.120s [2026-10-19 10:55:57,757] loop.py:45 INFO TRAIN_STEP step=9840 loss=1.244350 Training: 55%|█████▍ | 9840/18018 [21:25:43<16:48:34, 7.40s/step, loss=1.244, acc=0.708, acc_len=1.71, thru=18.5, I=18.5, T=33.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:55:57,810] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:04,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:56:05,053] loop.py:45 INFO TRAIN_STEP step=9841 loss=1.331469 Training: 55%|█████▍ | 9841/18018 [21:25:51<16:44:08, 7.37s/step, loss=1.331, acc=0.697, acc_len=1.67, thru=17.2, I=18.9, T=33.2, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:56:05,049] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9840: forward=2625.6ms backward=1162.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:05,333] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:09,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:56:11,206] loop.py:45 INFO TRAIN_STEP step=9842 loss=1.406657 Training: 55%|█████▍ | 9842/18018 [21:25:57<15:54:19, 7.00s/step, loss=1.407, acc=0.684, acc_len=1.63, thru=19.4, I=22.4, T=33.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:11,315] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:17,066] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:56:18,221] loop.py:45 INFO TRAIN_STEP step=9843 loss=1.371765 Training: 55%|█████▍ | 9843/18018 [21:26:04<15:54:39, 7.01s/step, loss=1.372, acc=0.692, acc_len=1.66, thru=16.2, I=20.4, T=30.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:18,575] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:23,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:56:24,967] loop.py:45 INFO TRAIN_STEP step=9844 loss=1.380037 Training: 55%|█████▍ | 9844/18018 [21:26:11<15:44:01, 6.93s/step, loss=1.380, acc=0.688, acc_len=1.64, thru=17.3, I=24.2, T=33.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:25,044] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 10:56:32,853] loop.py:380 INFO TIMING step=9845: step=4.013s data=1.776s compute=3.983s [fwd=2.858s bwd=1.109s opt=0.003s] dispatch=3.862s [2026-10-19 10:56:32,853] loop.py:45 INFO TRAIN_STEP step=9845 loss=1.370122 Training: 55%|█████▍ | 9845/18018 [21:26:18<16:22:59, 7.22s/step, loss=1.370, acc=0.689, acc_len=1.64, thru=11.2, I=15.8, T=31.9, wait=3.9s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:44,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:56:44,972] loop.py:45 INFO TRAIN_STEP step=9846 loss=1.364952 Training: 55%|█████▍ | 9846/18018 [21:26:31<19:43:11, 8.69s/step, loss=1.365, acc=0.693, acc_len=1.66, thru=13.4, I=13.1, T=30.1, wait=7.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:56:44,968] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9845: forward=3060.3ms backward=1144.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:45,249] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:50,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:56:51,029] loop.py:45 INFO TRAIN_STEP step=9847 loss=1.342199 Training: 55%|█████▍ | 9847/18018 [21:26:37<17:55:34, 7.90s/step, loss=1.342, acc=0.697, acc_len=1.67, thru=18.9, I=20.5, T=36.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:51,216] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:57,551] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:56:57,726] loop.py:45 INFO TRAIN_STEP step=9848 loss=1.440419 Training: 55%|█████▍ | 9848/18018 [21:26:43<17:06:19, 7.54s/step, loss=1.440, acc=0.678, acc_len=1.60, thru=19.1, I=21.1, T=33.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:56:58,053] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:04,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:57:04,969] loop.py:45 INFO TRAIN_STEP step=9849 loss=1.406741 Training: 55%|█████▍ | 9849/18018 [21:26:51<16:54:16, 7.45s/step, loss=1.407, acc=0.685, acc_len=1.63, thru=16.9, I=18.8, T=33.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:05,268] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:57:11,159] loop.py:380 INFO TIMING step=9850: step=3.439s data=1.412s compute=3.404s [fwd=2.435s bwd=0.955s opt=0.003s] dispatch=2.741s [2026-10-19 10:57:11,159] loop.py:45 INFO TRAIN_STEP step=9850 loss=1.242602 Training: 55%|█████▍ | 9850/18018 [21:26:57<16:02:34, 7.07s/step, loss=1.243, acc=0.709, acc_len=1.71, thru=18.6, I=20.2, T=37.2, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:16,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:57:18,219] loop.py:45 INFO TRAIN_STEP step=9851 loss=1.303783 Training: 55%|█████▍ | 9851/18018 [21:27:04<16:02:06, 7.07s/step, loss=1.304, acc=0.696, acc_len=1.66, thru=19.0, I=21.3, T=34.3, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:57:18,215] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9850: forward=2601.0ms backward=1086.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:18,419] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:57:23,894] loop.py:45 INFO TRAIN_STEP step=9852 loss=1.363206 Training: 55%|█████▍ | 9852/18018 [21:27:09<15:05:15, 6.65s/step, loss=1.363, acc=0.687, acc_len=1.63, thru=18.0, I=22.8, T=37.2, wait=2.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:33,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:57:34,387] loop.py:45 INFO TRAIN_STEP step=9853 loss=1.365944 Training: 55%|█████▍ | 9853/18018 [21:27:20<17:41:48, 7.80s/step, loss=1.366, acc=0.689, acc_len=1.64, thru=14.8, I=15.0, T=29.1, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:34,487] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:39,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:57:40,685] loop.py:45 INFO TRAIN_STEP step=9854 loss=1.352878 Training: 55%|█████▍ | 9854/18018 [21:27:26<16:40:18, 7.35s/step, loss=1.353, acc=0.692, acc_len=1.66, thru=18.9, I=18.3, T=35.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:40,719] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 10:57:48,188] loop.py:380 INFO TIMING step=9855: step=3.905s data=1.747s compute=3.879s [fwd=2.760s bwd=1.102s opt=0.003s] dispatch=3.587s [2026-10-19 10:57:48,188] loop.py:45 INFO TRAIN_STEP step=9855 loss=1.344959 Training: 55%|█████▍ | 9855/18018 [21:27:34<16:46:21, 7.40s/step, loss=1.345, acc=0.694, acc_len=1.66, thru=13.8, I=16.5, T=32.8, wait=3.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:55,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:57:56,232] loop.py:45 INFO TRAIN_STEP step=9856 loss=1.332963 Training: 55%|█████▍ | 9856/18018 [21:27:42<17:12:36, 7.59s/step, loss=1.333, acc=0.696, acc_len=1.66, thru=18.9, I=17.2, T=29.5, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:57:56,227] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9855: forward=3057.4ms backward=1227.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:57:56,564] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 10:58:02,373] loop.py:45 INFO TRAIN_STEP step=9857 loss=1.253360 Training: 55%|█████▍ | 9857/18018 [21:27:48<16:13:16, 7.16s/step, loss=1.253, acc=0.711, acc_len=1.72, thru=14.7, I=17.5, T=34.7, wait=2.4s, pool=29, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 10:58:02,444] eagle_store.py:517 WARNING batch_get_buffer missing keys (k315434-afd34d3e_hs, k315434-afd34d3e_ids, k315434-afd34d3e_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:58:09,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 10:58:02,442] eagle_store.py:517 WARNING batch_get_buffer missing keys (k315432-bd02166b_hs, k315432-bd02166b_ids, k315432-bd02166b_lhs); sleeping 0.50s. [2026-10-19 10:58:11,131] loop.py:45 INFO TRAIN_STEP step=9858 loss=1.332122 Training: 55%|█████▍ | 9858/18018 [21:27:57<17:18:29, 7.64s/step, loss=1.332, acc=0.699, acc_len=1.68, thru=18.2, I=18.9, T=32.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:58:11,610] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:58:19,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:58:20,878] loop.py:45 INFO TRAIN_STEP step=9859 loss=1.308896 Training: 55%|█████▍ | 9859/18018 [21:28:06<18:44:35, 8.27s/step, loss=1.309, acc=0.697, acc_len=1.67, thru=13.8, I=16.4, T=29.3, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:58:21,221] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:58:27,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:58:28,009] loop.py:380 INFO TIMING step=9860: step=4.073s data=1.818s compute=4.020s [fwd=2.808s bwd=1.197s opt=0.003s] dispatch=3.048s [2026-10-19 10:58:28,009] loop.py:45 INFO TRAIN_STEP step=9860 loss=1.374023 Training: 55%|█████▍ | 9860/18018 [21:28:14<17:57:55, 7.93s/step, loss=1.374, acc=0.686, acc_len=1.64, thru=15.9, I=20.6, T=31.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:58:28,375] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:58:34,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:58:34,990] loop.py:45 INFO TRAIN_STEP step=9861 loss=1.384640 Training: 55%|█████▍ | 9861/18018 [21:28:21<17:19:13, 7.64s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=16.9, I=21.0, T=29.8, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:58:34,985] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9860: forward=2998.6ms backward=1245.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:58:35,090] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 10:58:41,755] loop.py:45 INFO TRAIN_STEP step=9862 loss=1.355382 Training: 55%|█████▍ | 9862/18018 [21:28:27<16:43:14, 7.38s/step, loss=1.355, acc=0.689, acc_len=1.65, thru=16.8, I=17.4, T=32.4, wait=2.8s, pool=57, epoch=2/2][2026-10-19 10:58:50,487] loop.py:45 INFO TRAIN_STEP step=9863 loss=1.366830 Training: 55%|█████▍ | 9863/18018 [21:28:36<17:38:12, 7.79s/step, loss=1.367, acc=0.692, acc_len=1.66, thru=15.7, I=16.6, T=31.9, wait=4.7s, pool=61, epoch=2/2][2026-10-19 10:58:57,124] loop.py:45 INFO TRAIN_STEP step=9864 loss=1.347322 Training: 55%|█████▍ | 9864/18018 [21:28:43<16:51:13, 7.44s/step, loss=1.347, acc=0.691, acc_len=1.64, thru=17.2, I=19.0, T=34.5, wait=2.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:02,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:59:04,547] loop.py:380 INFO TIMING step=9865: step=3.663s data=1.566s compute=3.621s [fwd=2.553s bwd=1.053s opt=0.003s] dispatch=3.750s [2026-10-19 10:59:04,547] loop.py:45 INFO TRAIN_STEP step=9865 loss=1.442599 Training: 55%|█████▍ | 9865/18018 [21:28:50<16:50:22, 7.44s/step, loss=1.443, acc=0.676, acc_len=1.60, thru=17.4, I=20.0, T=34.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:04,867] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 10:59:11,630] loop.py:45 INFO TRAIN_STEP step=9866 loss=1.216325 Training: 55%|█████▍ | 9866/18018 [21:28:57<16:35:55, 7.33s/step, loss=1.216, acc=0.716, acc_len=1.74, thru=15.7, I=22.9, T=31.4, wait=3.0s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:59:11,625] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9865: forward=2902.2ms backward=1131.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:18,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:59:19,499] loop.py:45 INFO TRAIN_STEP step=9867 loss=1.412863 Training: 55%|█████▍ | 9867/18018 [21:29:05<16:57:41, 7.49s/step, loss=1.413, acc=0.685, acc_len=1.63, thru=16.1, I=17.0, T=33.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:19,924] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 10:59:26,716] loop.py:45 INFO TRAIN_STEP step=9868 loss=1.288035 Training: 55%|█████▍ | 9868/18018 [21:29:12<16:46:21, 7.41s/step, loss=1.288, acc=0.704, acc_len=1.69, thru=16.6, I=17.8, T=34.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:26,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:27,153] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 10:59:32,918] loop.py:45 INFO TRAIN_STEP step=9869 loss=1.490316 Training: 55%|█████▍ | 9869/18018 [21:29:19<15:57:11, 7.05s/step, loss=1.490, acc=0.669, acc_len=1.57, thru=18.8, I=22.2, T=34.5, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:41,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:59:41,892] loop.py:380 INFO TIMING step=9870: step=3.922s data=1.797s compute=3.861s [fwd=2.710s bwd=1.136s opt=0.003s] dispatch=5.043s [2026-10-19 10:59:41,893] loop.py:45 INFO TRAIN_STEP step=9870 loss=1.290925 Training: 55%|█████▍ | 9870/18018 [21:29:27<17:15:28, 7.62s/step, loss=1.291, acc=0.703, acc_len=1.69, thru=14.5, I=14.6, T=32.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:42,145] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:49,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:59:50,592] loop.py:45 INFO TRAIN_STEP step=9871 loss=1.349361 Training: 55%|█████▍ | 9871/18018 [21:29:36<17:59:03, 7.95s/step, loss=1.349, acc=0.694, acc_len=1.66, thru=14.0, I=15.4, T=26.9, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 10:59:50,587] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9870: forward=3286.8ms backward=1410.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:51,005] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:57,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 10:59:57,604] loop.py:45 INFO TRAIN_STEP step=9872 loss=1.423501 Training: 55%|█████▍ | 9872/18018 [21:29:43<17:20:54, 7.67s/step, loss=1.424, acc=0.680, acc_len=1.60, thru=15.9, I=17.6, T=34.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 10:59:58,003] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:05,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:00:06,133] loop.py:45 INFO TRAIN_STEP step=9873 loss=1.393970 Training: 55%|█████▍ | 9873/18018 [21:29:52<17:55:53, 7.93s/step, loss=1.394, acc=0.684, acc_len=1.63, thru=15.7, I=16.5, T=31.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:06,490] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:11,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:00:12,525] loop.py:45 INFO TRAIN_STEP step=9874 loss=1.371006 Training: 55%|█████▍ | 9874/18018 [21:29:58<16:53:20, 7.47s/step, loss=1.371, acc=0.692, acc_len=1.65, thru=19.4, I=18.8, T=32.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:12,888] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:20,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:00:20,884] loop.py:380 INFO TIMING step=9875: step=4.109s data=1.681s compute=4.072s [fwd=2.814s bwd=1.243s opt=0.003s] dispatch=4.240s [2026-10-19 11:00:20,884] loop.py:45 INFO TRAIN_STEP step=9875 loss=1.379856 Training: 55%|█████▍ | 9875/18018 [21:30:06<17:29:33, 7.73s/step, loss=1.380, acc=0.684, acc_len=1.63, thru=16.0, I=19.0, T=31.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:21,017] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:25,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:00:27,354] loop.py:45 INFO TRAIN_STEP step=9876 loss=1.288832 Training: 55%|█████▍ | 9876/18018 [21:30:13<16:38:04, 7.35s/step, loss=1.289, acc=0.699, acc_len=1.68, thru=18.8, I=24.9, T=29.5, wait=2.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:00:27,350] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9875: forward=3052.8ms backward=1242.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:27,806] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:32,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:00:34,277] loop.py:45 INFO TRAIN_STEP step=9877 loss=1.410812 Training: 55%|█████▍ | 9877/18018 [21:30:20<16:20:21, 7.23s/step, loss=1.411, acc=0.680, acc_len=1.61, thru=16.6, I=26.2, T=32.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:34,784] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:39,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:00:41,117] loop.py:45 INFO TRAIN_STEP step=9878 loss=1.374230 Training: 55%|█████▍ | 9878/18018 [21:30:27<16:04:34, 7.11s/step, loss=1.374, acc=0.686, acc_len=1.64, thru=17.9, I=26.9, T=30.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:41,397] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:48,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:00:48,944] loop.py:45 INFO TRAIN_STEP step=9879 loss=1.246831 Training: 55%|█████▍ | 9879/18018 [21:30:35<16:33:32, 7.32s/step, loss=1.247, acc=0.710, acc_len=1.72, thru=14.9, I=19.6, T=34.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:49,085] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:55,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:00:56,299] loop.py:380 INFO TIMING step=9880: step=3.600s data=1.488s compute=3.557s [fwd=2.536s bwd=1.006s opt=0.003s] dispatch=3.747s [2026-10-19 11:00:56,300] loop.py:45 INFO TRAIN_STEP step=9880 loss=1.394542 Training: 55%|█████▍ | 9880/18018 [21:30:42<16:34:42, 7.33s/step, loss=1.395, acc=0.686, acc_len=1.64, thru=18.0, I=18.7, T=35.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:00:56,329] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:01:03,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:01:04,647] loop.py:45 INFO TRAIN_STEP step=9881 loss=1.393688 Training: 55%|█████▍ | 9881/18018 [21:30:50<17:15:51, 7.64s/step, loss=1.394, acc=0.686, acc_len=1.63, thru=15.3, I=17.8, T=31.1, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:01:04,642] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9880: forward=2859.5ms backward=1210.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:01:04,949] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:01:11,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:01:11,783] loop.py:45 INFO TRAIN_STEP step=9882 loss=1.365365 Training: 55%|█████▍ | 9882/18018 [21:30:57<16:55:19, 7.49s/step, loss=1.365, acc=0.688, acc_len=1.65, thru=17.1, I=19.7, T=35.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:01:12,044] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:01:22,689] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:01:23,091] loop.py:45 INFO TRAIN_STEP step=9883 loss=1.346427 Training: 55%|█████▍ | 9883/18018 [21:31:09<19:30:31, 8.63s/step, loss=1.346, acc=0.695, acc_len=1.66, thru=11.5, I=11.2, T=27.5, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:01:23,191] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 11:01:31,162] loop.py:45 INFO TRAIN_STEP step=9884 loss=1.361110 Training: 55%|█████▍ | 9884/18018 [21:31:17<19:07:29, 8.46s/step, loss=1.361, acc=0.688, acc_len=1.65, thru=14.7, I=14.4, T=31.3, wait=4.0s, pool=61, epoch=2/2][2026-10-19 11:01:39,081] loop.py:380 INFO TIMING step=9885: step=3.920s data=1.704s compute=3.880s [fwd=2.711s bwd=1.154s opt=0.003s] dispatch=3.991s [2026-10-19 11:01:39,081] loop.py:45 INFO TRAIN_STEP step=9885 loss=1.352341 Training: 55%|█████▍ | 9885/18018 [21:31:25<18:45:16, 8.30s/step, loss=1.352, acc=0.694, acc_len=1.66, thru=15.5, I=15.6, T=32.7, wait=4.0s, pool=49, epoch=2/2][2026-10-19 11:01:45,902] loop.py:45 INFO TRAIN_STEP step=9886 loss=1.297632 Training: 55%|█████▍ | 9886/18018 [21:31:31<17:44:49, 7.86s/step, loss=1.298, acc=0.699, acc_len=1.69, thru=15.5, I=16.2, T=32.2, wait=2.8s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:01:45,897] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9885: forward=2795.5ms backward=1126.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:01:54,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:01:55,218] loop.py:45 INFO TRAIN_STEP step=9887 loss=1.410853 Training: 55%|█████▍ | 9887/18018 [21:31:41<18:44:08, 8.30s/step, loss=1.411, acc=0.681, acc_len=1.62, thru=16.7, I=15.8, T=29.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:01:55,519] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:02:03,059] loop.py:45 INFO TRAIN_STEP step=9888 loss=1.497418 Training: 55%|█████▍ | 9888/18018 [21:31:49<18:25:28, 8.16s/step, loss=1.497, acc=0.665, acc_len=1.56, thru=13.4, I=14.9, T=32.9, wait=3.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:10,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:02:10,970] loop.py:45 INFO TRAIN_STEP step=9889 loss=1.363347 Training: 55%|█████▍ | 9889/18018 [21:31:57<18:15:18, 8.08s/step, loss=1.363, acc=0.687, acc_len=1.64, thru=16.8, I=16.1, T=29.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:11,157] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:17,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:02:18,208] loop.py:380 INFO TIMING step=9890: step=4.248s data=1.936s compute=4.211s [fwd=2.966s bwd=1.229s opt=0.003s] dispatch=2.980s [2026-10-19 11:02:18,208] loop.py:45 INFO TRAIN_STEP step=9890 loss=1.405522 Training: 55%|█████▍ | 9890/18018 [21:32:04<17:40:45, 7.83s/step, loss=1.406, acc=0.682, acc_len=1.62, thru=16.7, I=18.3, T=30.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:18,373] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:28,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:02:28,278] loop.py:45 INFO TRAIN_STEP step=9891 loss=1.416063 Training: 55%|█████▍ | 9891/18018 [21:32:14<19:11:40, 8.50s/step, loss=1.416, acc=0.681, acc_len=1.62, thru=12.9, I=12.4, T=31.8, wait=6.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:02:28,273] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9890: forward=2799.2ms backward=1159.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:28,584] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:34,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:02:35,410] loop.py:45 INFO TRAIN_STEP step=9892 loss=1.338063 Training: 55%|█████▍ | 9892/18018 [21:32:21<18:15:47, 8.09s/step, loss=1.338, acc=0.694, acc_len=1.66, thru=17.4, I=19.0, T=30.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:35,557] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:41,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:02:42,860] loop.py:45 INFO TRAIN_STEP step=9893 loss=1.359324 Training: 55%|█████▍ | 9893/18018 [21:32:28<17:49:34, 7.90s/step, loss=1.359, acc=0.693, acc_len=1.65, thru=16.0, I=21.9, T=29.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:43,310] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:02:51,348] loop.py:45 INFO TRAIN_STEP step=9894 loss=1.388637 Training: 55%|█████▍ | 9894/18018 [21:32:37<18:13:29, 8.08s/step, loss=1.389, acc=0.682, acc_len=1.62, thru=11.4, I=13.9, T=35.8, wait=4.9s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:02:58,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:03:00,135] loop.py:380 INFO TIMING step=9895: step=3.862s data=1.567s compute=3.831s [fwd=2.767s bwd=1.049s opt=0.003s] dispatch=4.916s [2026-10-19 11:03:00,136] loop.py:45 INFO TRAIN_STEP step=9895 loss=1.307348 Training: 55%|█████▍ | 9895/18018 [21:32:46<18:42:16, 8.29s/step, loss=1.307, acc=0.697, acc_len=1.67, thru=18.9, I=17.7, T=33.1, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:00,331] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:05,825] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:03:06,500] loop.py:45 INFO TRAIN_STEP step=9896 loss=1.327440 Training: 55%|█████▍ | 9896/18018 [21:32:52<17:23:52, 7.71s/step, loss=1.327, acc=0.696, acc_len=1.67, thru=18.9, I=24.0, T=37.0, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:03:06,496] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9895: forward=2417.5ms backward=994.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:06,831] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:12,551] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:03:12,733] loop.py:45 INFO TRAIN_STEP step=9897 loss=1.207356 Training: 55%|█████▍ | 9897/18018 [21:32:58<16:23:41, 7.27s/step, loss=1.207, acc=0.719, acc_len=1.75, thru=20.2, I=22.6, T=32.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:13,053] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:20,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:03:21,284] loop.py:45 INFO TRAIN_STEP step=9898 loss=1.283685 Training: 55%|█████▍ | 9898/18018 [21:33:07<17:15:46, 7.65s/step, loss=1.284, acc=0.700, acc_len=1.68, thru=15.1, I=15.8, T=32.6, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:21,347] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 11:03:29,527] loop.py:45 INFO TRAIN_STEP step=9899 loss=1.438226 Training: 55%|█████▍ | 9899/18018 [21:33:15<17:39:30, 7.83s/step, loss=1.438, acc=0.676, acc_len=1.60, thru=14.1, I=17.0, T=31.6, wait=4.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:36,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:03:38,252] loop.py:380 INFO TIMING step=9900: step=4.269s data=1.911s compute=4.243s [fwd=3.054s bwd=1.175s opt=0.003s] dispatch=4.446s [2026-10-19 11:03:38,252] loop.py:45 INFO TRAIN_STEP step=9900 loss=1.403534 Training: 55%|█████▍ | 9900/18018 [21:33:24<18:15:45, 8.10s/step, loss=1.404, acc=0.685, acc_len=1.62, thru=16.2, I=17.0, T=30.0, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:38,345] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-19 11:03:48,395] loop.py:45 INFO TRAIN_STEP step=9901 loss=1.352175 Training: 55%|█████▍ | 9901/18018 [21:33:34<19:38:27, 8.71s/step, loss=1.352, acc=0.689, acc_len=1.65, thru=11.8, I=11.8, T=30.6, wait=5.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:03:48,390] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9900: forward=2956.0ms backward=1181.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:58,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:03:59,629] loop.py:45 INFO TRAIN_STEP step=9902 loss=1.304883 Training: 55%|█████▍ | 9902/18018 [21:33:45<21:20:51, 9.47s/step, loss=1.305, acc=0.702, acc_len=1.70, thru=14.3, I=12.5, T=33.5, wait=7.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:03:59,893] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:04:03,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:04:05,728] loop.py:45 INFO TRAIN_STEP step=9903 loss=1.325347 Training: 55%|█████▍ | 9903/18018 [21:33:51<19:03:51, 8.46s/step, loss=1.325, acc=0.696, acc_len=1.67, thru=19.7, I=26.1, T=32.4, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:04:05,923] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:04:11,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:04:12,637] loop.py:45 INFO TRAIN_STEP step=9904 loss=1.249221 Training: 55%|█████▍ | 9904/18018 [21:33:58<18:01:00, 7.99s/step, loss=1.249, acc=0.709, acc_len=1.72, thru=16.4, I=24.5, T=34.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:04:12,928] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:04:19,696] loop.py:380 INFO TIMING step=9905: step=4.229s data=2.016s compute=4.192s [fwd=2.926s bwd=1.251s opt=0.003s] dispatch=2.820s [2026-10-19 11:04:19,696] loop.py:45 INFO TRAIN_STEP step=9905 loss=1.363294 Training: 55%|█████▍ | 9905/18018 [21:34:05<17:22:52, 7.71s/step, loss=1.363, acc=0.693, acc_len=1.66, thru=17.5, I=20.9, T=30.3, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:04:28,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:04:28,530] loop.py:45 INFO TRAIN_STEP step=9906 loss=1.580806 Training: 55%|█████▍ | 9906/18018 [21:34:14<18:08:17, 8.05s/step, loss=1.581, acc=0.654, acc_len=1.53, thru=15.3, I=13.4, T=29.5, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:04:28,525] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9905: forward=2995.6ms backward=1293.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:04:28,699] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:04:35,387] loop.py:45 INFO TRAIN_STEP step=9907 loss=1.247639 Training: 55%|█████▍ | 9907/18018 [21:34:21<17:19:39, 7.69s/step, loss=1.248, acc=0.713, acc_len=1.73, thru=14.4, I=14.9, T=34.9, wait=3.2s, pool=29, epoch=2/2][2026-10-19 11:04:45,089] loop.py:45 INFO TRAIN_STEP step=9908 loss=1.327867 Training: 55%|█████▍ | 9908/18018 [21:34:31<18:41:11, 8.29s/step, loss=1.328, acc=0.700, acc_len=1.68, thru=16.4, I=13.4, T=32.7, wait=5.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:04:52,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:04:54,133] loop.py:45 INFO TRAIN_STEP step=9909 loss=1.350484 Training: 55%|█████▍ | 9909/18018 [21:34:40<19:11:26, 8.52s/step, loss=1.350, acc=0.693, acc_len=1.65, thru=15.3, I=17.9, T=32.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:04:54,403] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:05:00,450] loop.py:380 INFO TIMING step=9910: step=3.691s data=1.667s compute=3.659s [fwd=2.613s bwd=1.034s opt=0.003s] dispatch=2.616s [2026-10-19 11:05:00,451] loop.py:45 INFO TRAIN_STEP step=9910 loss=1.313357 Training: 55%|█████▌ | 9910/18018 [21:34:46<17:41:54, 7.86s/step, loss=1.313, acc=0.694, acc_len=1.66, thru=18.4, I=22.4, T=34.7, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:06,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:05:08,389] loop.py:45 INFO TRAIN_STEP step=9911 loss=1.291545 Training: 55%|█████▌ | 9911/18018 [21:34:54<17:45:05, 7.88s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=18.0, I=19.8, T=27.9, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:05:08,384] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9910: forward=3141.0ms backward=1388.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:08,580] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:13,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:05:14,642] loop.py:45 INFO TRAIN_STEP step=9912 loss=1.278526 Training: 55%|█████▌ | 9912/18018 [21:35:00<16:38:57, 7.39s/step, loss=1.279, acc=0.706, acc_len=1.71, thru=19.0, I=24.4, T=28.1, wait=1.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:14,756] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 11:05:21,566] loop.py:45 INFO TRAIN_STEP step=9913 loss=1.412933 Training: 55%|█████▌ | 9913/18018 [21:35:07<16:19:39, 7.25s/step, loss=1.413, acc=0.687, acc_len=1.64, thru=15.6, I=18.0, T=31.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:21,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:22,054] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:26,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:05:27,765] loop.py:45 INFO TRAIN_STEP step=9914 loss=1.323734 Training: 55%|█████▌ | 9914/18018 [21:35:13<15:36:54, 6.94s/step, loss=1.324, acc=0.694, acc_len=1.66, thru=19.0, I=21.4, T=32.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:27,795] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:34,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:05:35,739] loop.py:380 INFO TIMING step=9915: step=3.747s data=1.592s compute=3.718s [fwd=2.709s bwd=0.994s opt=0.003s] dispatch=4.219s [2026-10-19 11:05:35,739] loop.py:45 INFO TRAIN_STEP step=9915 loss=1.348976 Training: 55%|█████▌ | 9915/18018 [21:35:21<16:18:54, 7.25s/step, loss=1.349, acc=0.693, acc_len=1.66, thru=16.0, I=19.2, T=34.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:35,920] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:41,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:05:42,583] loop.py:45 INFO TRAIN_STEP step=9916 loss=1.378511 Training: 55%|█████▌ | 9916/18018 [21:35:28<16:02:22, 7.13s/step, loss=1.379, acc=0.684, acc_len=1.63, thru=19.0, I=22.4, T=35.5, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:05:42,578] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9915: forward=2591.0ms backward=971.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:05:42,930] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:05:51,392] loop.py:45 INFO TRAIN_STEP step=9917 loss=1.285182 Training: 55%|█████▌ | 9917/18018 [21:35:37<17:10:23, 7.63s/step, loss=1.285, acc=0.704, acc_len=1.70, thru=13.1, I=13.6, T=31.4, wait=4.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:00,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:06:01,324] loop.py:45 INFO TRAIN_STEP step=9918 loss=1.445016 Training: 55%|█████▌ | 9918/18018 [21:35:47<18:43:22, 8.32s/step, loss=1.445, acc=0.674, acc_len=1.59, thru=15.0, I=15.9, T=34.1, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:01,665] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:06,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:06:08,008] loop.py:45 INFO TRAIN_STEP step=9919 loss=1.328653 Training: 55%|█████▌ | 9919/18018 [21:35:54<17:36:57, 7.83s/step, loss=1.329, acc=0.696, acc_len=1.67, thru=19.5, I=23.8, T=34.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:08,332] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:13,183] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:06:13,792] loop.py:380 INFO TIMING step=9920: step=3.368s data=1.348s compute=3.334s [fwd=2.393s bwd=0.926s opt=0.003s] dispatch=2.408s [2026-10-19 11:06:13,792] loop.py:45 INFO TRAIN_STEP step=9920 loss=1.264034 Training: 55%|█████▌ | 9920/18018 [21:35:59<16:13:53, 7.22s/step, loss=1.264, acc=0.709, acc_len=1.72, thru=20.9, I=26.4, T=38.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:14,186] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:19,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:06:21,336] loop.py:45 INFO TRAIN_STEP step=9921 loss=1.220202 Training: 55%|█████▌ | 9921/18018 [21:36:07<16:27:07, 7.31s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=17.8, I=25.5, T=29.7, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:06:21,331] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9920: forward=3141.4ms backward=1116.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:21,568] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 11:06:28,770] loop.py:45 INFO TRAIN_STEP step=9922 loss=1.286691 Training: 55%|█████▌ | 9922/18018 [21:36:14<16:31:50, 7.35s/step, loss=1.287, acc=0.703, acc_len=1.70, thru=13.4, I=18.5, T=33.3, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:34,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:06:35,720] loop.py:45 INFO TRAIN_STEP step=9923 loss=1.294835 Training: 55%|█████▌ | 9923/18018 [21:36:21<16:15:26, 7.23s/step, loss=1.295, acc=0.701, acc_len=1.69, thru=17.8, I=21.1, T=34.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:36,223] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:41,722] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:06:41,750] loop.py:45 INFO TRAIN_STEP step=9924 loss=1.302065 Training: 55%|█████▌ | 9924/18018 [21:36:27<15:26:46, 6.87s/step, loss=1.302, acc=0.700, acc_len=1.68, thru=20.0, I=25.0, T=38.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:42,224] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:47,091] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:06:47,605] loop.py:380 INFO TIMING step=9925: step=3.478s data=1.442s compute=3.442s [fwd=2.452s bwd=0.976s opt=0.003s] dispatch=2.369s [2026-10-19 11:06:47,605] loop.py:45 INFO TRAIN_STEP step=9925 loss=1.264811 Training: 55%|█████▌ | 9925/18018 [21:36:33<14:45:35, 6.57s/step, loss=1.265, acc=0.709, acc_len=1.71, thru=21.3, I=23.2, T=36.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:48,098] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:54,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:06:55,326] loop.py:45 INFO TRAIN_STEP step=9926 loss=1.277177 Training: 55%|█████▌ | 9926/18018 [21:36:41<15:32:16, 6.91s/step, loss=1.277, acc=0.705, acc_len=1.70, thru=16.2, I=19.4, T=27.4, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:06:55,321] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9925: forward=3263.8ms backward=1357.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:06:55,757] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:01,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:07:03,163] loop.py:45 INFO TRAIN_STEP step=9927 loss=1.310803 Training: 55%|█████▌ | 9927/18018 [21:36:49<16:09:34, 7.19s/step, loss=1.311, acc=0.700, acc_len=1.68, thru=15.2, I=17.6, T=33.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:03,498] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:10,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:07:11,787] loop.py:45 INFO TRAIN_STEP step=9928 loss=1.375929 Training: 55%|█████▌ | 9928/18018 [21:36:57<17:07:29, 7.62s/step, loss=1.376, acc=0.688, acc_len=1.64, thru=13.7, I=18.1, T=32.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:11,956] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:07:18,084] loop.py:45 INFO TRAIN_STEP step=9929 loss=1.364994 Training: 55%|█████▌ | 9929/18018 [21:37:04<16:13:53, 7.22s/step, loss=1.365, acc=0.689, acc_len=1.64, thru=18.6, I=22.6, T=32.8, wait=2.4s, pool=61, epoch=2/2][2026-10-19 11:07:25,575] loop.py:380 INFO TIMING step=9930: step=4.391s data=2.126s compute=4.341s [fwd=2.983s bwd=1.343s opt=0.003s] dispatch=3.088s [2026-10-19 11:07:25,575] loop.py:45 INFO TRAIN_STEP step=9930 loss=1.284351 Training: 55%|█████▌ | 9930/18018 [21:37:11<16:24:30, 7.30s/step, loss=1.284, acc=0.706, acc_len=1.71, thru=15.6, I=16.6, T=29.2, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:32,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:07:32,590] loop.py:45 INFO TRAIN_STEP step=9931 loss=1.265047 Training: 55%|█████▌ | 9931/18018 [21:37:18<16:12:38, 7.22s/step, loss=1.265, acc=0.704, acc_len=1.70, thru=17.6, I=17.2, T=32.3, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:07:32,585] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9930: forward=2873.4ms backward=1049.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:32,809] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:07:38,560] loop.py:45 INFO TRAIN_STEP step=9932 loss=1.323447 Training: 55%|█████▌ | 9932/18018 [21:37:24<15:22:06, 6.84s/step, loss=1.323, acc=0.701, acc_len=1.68, thru=19.7, I=19.7, T=34.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:38,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:39,055] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:07:44,565] loop.py:45 INFO TRAIN_STEP step=9933 loss=1.271309 Training: 55%|█████▌ | 9933/18018 [21:37:30<14:48:08, 6.59s/step, loss=1.271, acc=0.704, acc_len=1.69, thru=19.8, I=21.4, T=37.4, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:49,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:07:51,647] loop.py:45 INFO TRAIN_STEP step=9934 loss=1.464247 Training: 55%|█████▌ | 9934/18018 [21:37:37<15:07:59, 6.74s/step, loss=1.464, acc=0.673, acc_len=1.59, thru=17.6, I=21.6, T=32.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:51,966] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:57,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:07:58,079] loop.py:380 INFO TIMING step=9935: step=3.671s data=1.609s compute=3.643s [fwd=2.611s bwd=1.016s opt=0.003s] dispatch=2.750s [2026-10-19 11:07:58,079] loop.py:45 INFO TRAIN_STEP step=9935 loss=1.380073 Training: 55%|█████▌ | 9935/18018 [21:37:44<14:55:21, 6.65s/step, loss=1.380, acc=0.681, acc_len=1.61, thru=18.0, I=23.7, T=34.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:07:58,454] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:08:05,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:08:07,830] loop.py:45 INFO TRAIN_STEP step=9936 loss=1.352740 Training: 55%|█████▌ | 9936/18018 [21:37:53<17:00:49, 7.58s/step, loss=1.353, acc=0.693, acc_len=1.66, thru=13.3, I=16.7, T=27.8, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:08:07,825] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9935: forward=3266.9ms backward=1267.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:08:07,957] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:08:12,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:08:15,105] loop.py:45 INFO TRAIN_STEP step=9937 loss=1.301150 Training: 55%|█████▌ | 9937/18018 [21:38:01<16:48:26, 7.49s/step, loss=1.301, acc=0.701, acc_len=1.68, thru=13.7, I=23.3, T=26.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:08:15,162] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=49/64 [2026-10-19 11:08:21,657] loop.py:45 INFO TRAIN_STEP step=9938 loss=1.284730 Training: 55%|█████▌ | 9938/18018 [21:38:07<16:10:26, 7.21s/step, loss=1.285, acc=0.703, acc_len=1.69, thru=13.4, I=20.9, T=32.6, wait=2.6s, pool=49, epoch=2/2][2026-10-19 11:08:31,757] loop.py:45 INFO TRAIN_STEP step=9939 loss=1.316326 Training: 55%|█████▌ | 9939/18018 [21:38:17<18:07:16, 8.07s/step, loss=1.316, acc=0.699, acc_len=1.68, thru=15.7, I=12.7, T=30.1, wait=5.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:08:40,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:08:40,410] loop.py:380 INFO TIMING step=9940: step=4.599s data=2.306s compute=4.567s [fwd=3.417s bwd=1.135s opt=0.003s] dispatch=4.043s [2026-10-19 11:08:40,410] loop.py:45 INFO TRAIN_STEP step=9940 loss=1.216756 Training: 55%|█████▌ | 9940/18018 [21:38:26<18:30:29, 8.25s/step, loss=1.217, acc=0.714, acc_len=1.73, thru=16.3, I=16.3, T=27.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:08:40,783] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:08:48,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:08:48,436] loop.py:45 INFO TRAIN_STEP step=9941 loss=1.268143 Training: 55%|█████▌ | 9941/18018 [21:38:34<18:21:22, 8.18s/step, loss=1.268, acc=0.710, acc_len=1.72, thru=15.6, I=16.3, T=29.9, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:08:48,431] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9940: forward=2937.7ms backward=1270.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:08:48,762] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:08:57,970] loop.py:45 INFO TRAIN_STEP step=9942 loss=1.367475 Training: 55%|█████▌ | 9942/18018 [21:38:44<19:15:56, 8.59s/step, loss=1.367, acc=0.687, acc_len=1.64, thru=11.5, I=12.6, T=31.3, wait=5.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:03,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:09:05,460] loop.py:45 INFO TRAIN_STEP step=9943 loss=1.226699 Training: 55%|█████▌ | 9943/18018 [21:38:51<18:31:23, 8.26s/step, loss=1.227, acc=0.711, acc_len=1.72, thru=18.4, I=18.6, T=34.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:05,622] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:10,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:09:12,283] loop.py:45 INFO TRAIN_STEP step=9944 loss=1.239869 Training: 55%|█████▌ | 9944/18018 [21:38:58<17:33:19, 7.83s/step, loss=1.240, acc=0.713, acc_len=1.73, thru=16.8, I=28.4, T=29.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:12,614] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:20,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:09:21,113] loop.py:380 INFO TIMING step=9945: step=3.815s data=1.744s compute=3.779s [fwd=2.711s bwd=1.054s opt=0.003s] dispatch=5.004s [2026-10-19 11:09:21,114] loop.py:45 INFO TRAIN_STEP step=9945 loss=1.244239 Training: 55%|█████▌ | 9945/18018 [21:39:07<18:13:34, 8.13s/step, loss=1.244, acc=0.710, acc_len=1.72, thru=15.1, I=17.1, T=33.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:21,600] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:09:28,439] loop.py:45 INFO TRAIN_STEP step=9946 loss=1.399832 Training: 55%|█████▌ | 9946/18018 [21:39:14<17:41:10, 7.89s/step, loss=1.400, acc=0.685, acc_len=1.63, thru=16.9, I=19.4, T=33.4, wait=3.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:09:28,434] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9945: forward=2733.9ms backward=1053.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:35,258] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:09:36,221] loop.py:45 INFO TRAIN_STEP step=9947 loss=1.409992 Training: 55%|█████▌ | 9947/18018 [21:39:22<17:36:44, 7.86s/step, loss=1.410, acc=0.679, acc_len=1.62, thru=16.9, I=18.0, T=33.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:36,266] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:43,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:09:44,470] loop.py:45 INFO TRAIN_STEP step=9948 loss=1.268280 Training: 55%|█████▌ | 9948/18018 [21:39:30<17:52:29, 7.97s/step, loss=1.268, acc=0.710, acc_len=1.72, thru=14.6, I=16.4, T=32.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:09:44,906] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:09:53,137] loop.py:45 INFO TRAIN_STEP step=9949 loss=1.244836 Training: 55%|█████▌ | 9949/18018 [21:39:39<18:20:20, 8.18s/step, loss=1.245, acc=0.708, acc_len=1.71, thru=13.6, I=15.6, T=34.0, wait=4.9s, pool=53, epoch=2/2][2026-10-19 11:10:03,966] loop.py:380 INFO TIMING step=9950: step=4.134s data=1.768s compute=4.105s [fwd=2.931s bwd=1.159s opt=0.003s] dispatch=6.685s [2026-10-19 11:10:03,966] loop.py:45 INFO TRAIN_STEP step=9950 loss=1.382346 Training: 55%|█████▌ | 9950/18018 [21:39:50<20:06:55, 8.98s/step, loss=1.382, acc=0.687, acc_len=1.64, thru=13.6, I=12.5, T=31.0, wait=6.7s, pool=57, epoch=2/2][2026-10-19 11:10:12,747] loop.py:45 INFO TRAIN_STEP step=9951 loss=1.380139 Training: 55%|█████▌ | 9951/18018 [21:39:58<19:58:52, 8.92s/step, loss=1.380, acc=0.688, acc_len=1.64, thru=15.1, I=13.7, T=33.5, wait=5.0s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:10:12,742] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9950: forward=2655.5ms backward=1100.0ms [2026-10-19 11:10:20,391] loop.py:45 INFO TRAIN_STEP step=9952 loss=1.269162 Training: 55%|█████▌ | 9952/18018 [21:40:06<19:07:31, 8.54s/step, loss=1.269, acc=0.711, acc_len=1.72, thru=16.1, I=17.6, T=29.5, wait=3.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:26,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:10:28,420] loop.py:45 INFO TRAIN_STEP step=9953 loss=1.363150 Training: 55%|█████▌ | 9953/18018 [21:40:14<18:46:55, 8.38s/step, loss=1.363, acc=0.689, acc_len=1.64, thru=17.2, I=18.3, T=26.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:28,559] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:34,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:10:36,187] loop.py:45 INFO TRAIN_STEP step=9954 loss=1.387456 Training: 55%|█████▌ | 9954/18018 [21:40:22<18:21:49, 8.20s/step, loss=1.387, acc=0.687, acc_len=1.64, thru=16.8, I=21.9, T=31.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:36,678] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 11:10:43,658] loop.py:380 INFO TIMING step=9955: step=3.581s data=1.532s compute=3.554s [fwd=2.559s bwd=0.981s opt=0.003s] dispatch=3.882s [2026-10-19 11:10:43,658] loop.py:45 INFO TRAIN_STEP step=9955 loss=1.291108 Training: 55%|█████▌ | 9955/18018 [21:40:29<17:52:25, 7.98s/step, loss=1.291, acc=0.702, acc_len=1.69, thru=14.6, I=19.3, T=35.7, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:43,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:44,164] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:51,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:10:52,446] loop.py:45 INFO TRAIN_STEP step=9956 loss=1.333223 (TrainerActor pid=9310) [2026-10-19 11:10:52,442] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9955: forward=3146.5ms backward=1293.5ms Training: 55%|█████▌ | 9956/18018 [21:40:38<18:24:52, 8.22s/step, loss=1.333, acc=0.699, acc_len=1.68, thru=14.9, I=18.3, T=28.5, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:52,548] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:56,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:10:58,281] loop.py:45 INFO TRAIN_STEP step=9957 loss=1.302248 Training: 55%|█████▌ | 9957/18018 [21:40:44<16:48:30, 7.51s/step, loss=1.302, acc=0.699, acc_len=1.68, thru=18.6, I=26.3, T=36.3, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:10:58,339] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:02,928] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:11:04,100] loop.py:45 INFO TRAIN_STEP step=9958 loss=1.259707 Training: 55%|█████▌ | 9958/18018 [21:40:50<15:40:22, 7.00s/step, loss=1.260, acc=0.709, acc_len=1.72, thru=21.9, I=29.1, T=32.5, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:04,437] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:09,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:11:10,464] loop.py:45 INFO TRAIN_STEP step=9959 loss=1.416471 Training: 55%|█████▌ | 9959/18018 [21:40:56<15:14:37, 6.81s/step, loss=1.416, acc=0.681, acc_len=1.62, thru=18.5, I=23.5, T=32.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:10,528] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:16,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:11:16,660] loop.py:380 INFO TIMING step=9960: step=3.809s data=1.555s compute=3.782s [fwd=2.720s bwd=1.047s opt=0.003s] dispatch=2.376s [2026-10-19 11:11:16,660] loop.py:45 INFO TRAIN_STEP step=9960 loss=1.290318 Training: 55%|█████▌ | 9960/18018 [21:41:02<14:49:44, 6.63s/step, loss=1.290, acc=0.705, acc_len=1.70, thru=18.9, I=22.1, T=33.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:17,156] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:22,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:11:23,855] loop.py:45 INFO TRAIN_STEP step=9961 loss=1.280051 Training: 55%|█████▌ | 9961/18018 [21:41:09<15:12:36, 6.80s/step, loss=1.280, acc=0.705, acc_len=1.70, thru=17.9, I=20.9, T=33.6, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:11:23,851] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9960: forward=2647.4ms backward=1114.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:24,321] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:31,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:11:32,390] loop.py:45 INFO TRAIN_STEP step=9962 loss=1.206319 Training: 55%|█████▌ | 9962/18018 [21:41:18<16:22:27, 7.32s/step, loss=1.206, acc=0.718, acc_len=1.73, thru=14.1, I=18.0, T=33.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:32,774] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:11:39,931] loop.py:45 INFO TRAIN_STEP step=9963 loss=1.308764 Training: 55%|█████▌ | 9963/18018 [21:41:26<16:31:27, 7.39s/step, loss=1.309, acc=0.701, acc_len=1.69, thru=16.0, I=22.3, T=31.6, wait=3.5s, pool=61, epoch=2/2][2026-10-19 11:11:48,486] loop.py:45 INFO TRAIN_STEP step=9964 loss=1.269344 Training: 55%|█████▌ | 9964/18018 [21:41:34<17:18:20, 7.74s/step, loss=1.269, acc=0.707, acc_len=1.70, thru=15.4, I=12.6, T=30.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:48,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:11:48,974] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:11:56,133] loop.py:380 INFO TIMING step=9965: step=3.723s data=1.473s compute=3.675s [fwd=2.617s bwd=1.042s opt=0.003s] dispatch=3.915s [2026-10-19 11:11:56,134] loop.py:45 INFO TRAIN_STEP step=9965 loss=1.356244 Training: 55%|█████▌ | 9965/18018 [21:41:42<17:14:40, 7.71s/step, loss=1.356, acc=0.688, acc_len=1.64, thru=16.4, I=15.8, T=34.4, wait=3.9s, pool=53, epoch=2/2][2026-10-19 11:12:03,882] loop.py:45 INFO TRAIN_STEP step=9966 loss=1.532338 Training: 55%|█████▌ | 9966/18018 [21:41:49<17:16:11, 7.72s/step, loss=1.532, acc=0.661, acc_len=1.55, thru=17.6, I=17.0, T=32.4, wait=3.8s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:12:03,878] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9965: forward=2839.0ms backward=1073.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:12:14,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:12:14,891] loop.py:45 INFO TRAIN_STEP step=9967 loss=1.278060 Training: 55%|█████▌ | 9967/18018 [21:42:00<19:28:26, 8.71s/step, loss=1.278, acc=0.706, acc_len=1.70, thru=12.4, I=12.1, T=34.0, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:12:14,999] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:12:19,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:12:21,417] loop.py:45 INFO TRAIN_STEP step=9968 loss=1.365617 Training: 55%|█████▌ | 9968/18018 [21:42:07<18:00:27, 8.05s/step, loss=1.366, acc=0.688, acc_len=1.64, thru=18.9, I=23.0, T=31.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:12:21,742] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:12:26,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:12:27,236] loop.py:45 INFO TRAIN_STEP step=9969 loss=1.256631 Training: 55%|█████▌ | 9969/18018 [21:42:13<16:30:19, 7.38s/step, loss=1.257, acc=0.706, acc_len=1.70, thru=20.5, I=27.7, T=34.0, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:12:27,668] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:12:34,456] loop.py:380 INFO TIMING step=9970: step=3.910s data=1.634s compute=3.854s [fwd=2.649s bwd=1.190s opt=0.003s] dispatch=3.303s [2026-10-19 11:12:34,457] loop.py:45 INFO TRAIN_STEP step=9970 loss=1.252126 Training: 55%|█████▌ | 9970/18018 [21:42:20<16:23:43, 7.33s/step, loss=1.252, acc=0.708, acc_len=1.71, thru=15.9, I=19.6, T=32.7, wait=3.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:12:42,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:12:43,330] loop.py:45 INFO TRAIN_STEP step=9971 loss=1.337693 Training: 55%|█████▌ | 9971/18018 [21:42:29<17:25:30, 7.80s/step, loss=1.338, acc=0.696, acc_len=1.67, thru=15.6, I=14.9, T=32.7, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:12:43,325] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9970: forward=2764.8ms backward=1100.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:12:43,754] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:12:50,792] loop.py:45 INFO TRAIN_STEP step=9972 loss=1.366792 Training: 55%|█████▌ | 9972/18018 [21:42:36<17:11:58, 7.70s/step, loss=1.367, acc=0.687, acc_len=1.63, thru=15.3, I=17.2, T=32.5, wait=3.5s, pool=49, epoch=2/2][2026-10-19 11:12:57,866] loop.py:45 INFO TRAIN_STEP step=9973 loss=1.189514 Training: 55%|█████▌ | 9973/18018 [21:42:43<16:46:52, 7.51s/step, loss=1.190, acc=0.723, acc_len=1.76, thru=15.6, I=15.9, T=32.3, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:06,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:13:06,896] loop.py:45 INFO TRAIN_STEP step=9974 loss=1.462896 Training: 55%|█████▌ | 9974/18018 [21:42:52<17:47:50, 7.97s/step, loss=1.463, acc=0.670, acc_len=1.58, thru=16.9, I=14.7, T=32.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:07,277] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:11,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:13:13,033] loop.py:380 INFO TIMING step=9975: step=3.680s data=1.564s compute=3.647s [fwd=2.649s bwd=0.983s opt=0.003s] dispatch=2.449s [2026-10-19 11:13:13,033] loop.py:45 INFO TRAIN_STEP step=9975 loss=1.434381 Training: 55%|█████▌ | 9975/18018 [21:42:59<16:34:21, 7.42s/step, loss=1.434, acc=0.683, acc_len=1.63, thru=20.2, I=21.6, T=34.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:13,495] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:20,750] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:13:21,934] loop.py:45 INFO TRAIN_STEP step=9976 loss=1.200247 Training: 55%|█████▌ | 9976/18018 [21:43:08<17:33:46, 7.86s/step, loss=1.200, acc=0.720, acc_len=1.75, thru=13.3, I=18.2, T=31.1, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:13:21,929] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9975: forward=2934.4ms backward=1135.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:22,257] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:27,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:13:28,437] loop.py:45 INFO TRAIN_STEP step=9977 loss=1.315280 Training: 55%|█████▌ | 9977/18018 [21:43:14<16:38:58, 7.45s/step, loss=1.315, acc=0.700, acc_len=1.68, thru=18.2, I=23.2, T=34.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:28,932] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:13:35,278] loop.py:45 INFO TRAIN_STEP step=9978 loss=1.389233 Training: 55%|█████▌ | 9978/18018 [21:43:21<16:14:19, 7.27s/step, loss=1.389, acc=0.685, acc_len=1.63, thru=16.1, I=20.1, T=32.9, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:41,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:13:42,184] loop.py:45 INFO TRAIN_STEP step=9979 loss=1.300225 Training: 55%|█████▌ | 9979/18018 [21:43:28<15:59:28, 7.16s/step, loss=1.300, acc=0.699, acc_len=1.68, thru=18.7, I=18.2, T=32.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:42,358] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:49,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:13:50,453] loop.py:380 INFO TIMING step=9980: step=4.473s data=2.069s compute=4.437s [fwd=3.124s bwd=1.298s opt=0.003s] dispatch=3.786s [2026-10-19 11:13:50,454] loop.py:45 INFO TRAIN_STEP step=9980 loss=1.344131 Training: 55%|█████▌ | 9980/18018 [21:43:36<16:43:56, 7.49s/step, loss=1.344, acc=0.694, acc_len=1.66, thru=16.1, I=17.9, T=28.6, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:50,894] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:55,978] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:13:56,700] loop.py:45 INFO TRAIN_STEP step=9981 loss=1.416392 Training: 55%|█████▌ | 9981/18018 [21:43:42<15:53:41, 7.12s/step, loss=1.416, acc=0.681, acc_len=1.62, thru=18.8, I=20.9, T=36.3, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:13:56,695] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9980: forward=2506.7ms backward=975.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:13:56,986] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:01,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:14:03,368] loop.py:45 INFO TRAIN_STEP step=9982 loss=1.385377 Training: 55%|█████▌ | 9982/18018 [21:43:49<15:35:26, 6.98s/step, loss=1.385, acc=0.689, acc_len=1.64, thru=18.8, I=23.8, T=34.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:03,468] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:08,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:14:09,377] loop.py:45 INFO TRAIN_STEP step=9983 loss=1.305460 Training: 55%|█████▌ | 9983/18018 [21:43:55<14:56:02, 6.69s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=20.1, I=27.3, T=32.8, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:09,826] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:17,218] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:14:17,690] loop.py:45 INFO TRAIN_STEP step=9984 loss=1.329640 Training: 55%|█████▌ | 9984/18018 [21:44:03<16:01:04, 7.18s/step, loss=1.330, acc=0.695, acc_len=1.66, thru=15.2, I=19.1, T=30.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:17,720] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:24,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:14:25,741] loop.py:380 INFO TIMING step=9985: step=4.052s data=1.784s compute=4.014s [fwd=2.845s bwd=1.152s opt=0.003s] dispatch=3.991s [2026-10-19 11:14:25,742] loop.py:45 INFO TRAIN_STEP step=9985 loss=1.312428 Training: 55%|█████▌ | 9985/18018 [21:44:11<16:36:04, 7.44s/step, loss=1.312, acc=0.699, acc_len=1.67, thru=15.5, I=19.3, T=31.6, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:26,165] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:14:32,691] loop.py:45 INFO TRAIN_STEP step=9986 loss=1.286386 Training: 55%|█████▌ | 9986/18018 [21:44:18<16:16:18, 7.29s/step, loss=1.286, acc=0.705, acc_len=1.70, thru=16.6, I=20.6, T=36.7, wait=3.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:14:32,687] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9985: forward=2482.0ms backward=961.3ms [2026-10-19 11:14:39,643] loop.py:45 INFO TRAIN_STEP step=9987 loss=1.394990 Training: 55%|█████▌ | 9987/18018 [21:44:25<16:02:29, 7.19s/step, loss=1.395, acc=0.686, acc_len=1.64, thru=18.0, I=17.8, T=31.2, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:45,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:14:46,343] loop.py:45 INFO TRAIN_STEP step=9988 loss=1.221832 Training: 55%|█████▌ | 9988/18018 [21:44:32<15:42:35, 7.04s/step, loss=1.222, acc=0.716, acc_len=1.73, thru=18.7, I=20.4, T=30.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:46,824] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:51,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:14:53,056] loop.py:45 INFO TRAIN_STEP step=9989 loss=1.181115 Training: 55%|█████▌ | 9989/18018 [21:44:39<15:29:17, 6.94s/step, loss=1.181, acc=0.720, acc_len=1.75, thru=17.7, I=22.0, T=32.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:14:53,265] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:01,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:15:01,860] loop.py:380 INFO TIMING step=9990: step=3.843s data=1.717s compute=3.816s [fwd=2.717s bwd=1.083s opt=0.003s] dispatch=4.952s [2026-10-19 11:15:01,860] loop.py:45 INFO TRAIN_STEP step=9990 loss=1.275951 Training: 55%|█████▌ | 9990/18018 [21:44:47<16:43:45, 7.50s/step, loss=1.276, acc=0.705, acc_len=1.70, thru=14.9, I=16.4, T=33.3, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:02,226] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:15:09,416] loop.py:45 INFO TRAIN_STEP step=9991 loss=1.287026 Training: 55%|█████▌ | 9991/18018 [21:44:55<16:45:49, 7.52s/step, loss=1.287, acc=0.704, acc_len=1.69, thru=15.9, I=16.9, T=34.2, wait=3.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:15:09,412] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9990: forward=2641.3ms backward=1051.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:16,340] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:15:18,292] loop.py:45 INFO TRAIN_STEP step=9992 loss=1.259389 Training: 55%|█████▌ | 9992/18018 [21:45:04<17:40:15, 7.93s/step, loss=1.259, acc=0.708, acc_len=1.71, thru=16.6, I=19.8, T=31.0, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:18,359] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:24,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:15:25,840] loop.py:45 INFO TRAIN_STEP step=9993 loss=1.393981 Training: 55%|█████▌ | 9993/18018 [21:45:11<17:24:52, 7.81s/step, loss=1.394, acc=0.683, acc_len=1.62, thru=15.0, I=22.3, T=32.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:26,262] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:15:31,869] loop.py:45 INFO TRAIN_STEP step=9994 loss=1.372078 Training: 55%|█████▌ | 9994/18018 [21:45:17<16:13:09, 7.28s/step, loss=1.372, acc=0.690, acc_len=1.64, thru=18.8, I=22.4, T=35.8, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:38,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:15:39,243] loop.py:380 INFO TIMING step=9995: step=3.967s data=1.708s compute=3.935s [fwd=2.808s bwd=1.111s opt=0.003s] dispatch=3.399s [2026-10-19 11:15:39,243] loop.py:45 INFO TRAIN_STEP step=9995 loss=1.312551 Training: 55%|█████▌ | 9995/18018 [21:45:25<16:16:59, 7.31s/step, loss=1.313, acc=0.696, acc_len=1.67, thru=18.4, I=19.8, T=32.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:39,565] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:45,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:15:45,669] loop.py:45 INFO TRAIN_STEP step=9996 loss=1.259382 Training: 55%|█████▌ | 9996/18018 [21:45:31<15:41:36, 7.04s/step, loss=1.259, acc=0.708, acc_len=1.70, thru=18.8, I=22.4, T=35.3, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:15:45,665] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=9995: forward=2551.6ms backward=1026.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:15:45,974] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:15:54,085] loop.py:45 INFO TRAIN_STEP step=9997 loss=1.424880 Training: 55%|█████▌ | 9997/18018 [21:45:40<16:36:26, 7.45s/step, loss=1.425, acc=0.682, acc_len=1.62, thru=14.8, I=15.2, T=31.8, wait=4.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:01,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:16:02,182] loop.py:45 INFO TRAIN_STEP step=9998 loss=1.252477 Training: 55%|█████▌ | 9998/18018 [21:45:48<17:02:07, 7.65s/step, loss=1.252, acc=0.713, acc_len=1.73, thru=15.7, I=17.0, T=33.4, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:02,624] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:16:10,161] loop.py:45 INFO TRAIN_STEP step=9999 loss=1.219515 Training: 55%|█████▌ | 9999/18018 [21:45:56<17:15:18, 7.75s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=14.1, I=17.0, T=37.4, wait=4.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:16,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:16:16,910] loop.py:380 INFO TIMING step=10000: step=3.795s data=1.639s compute=3.754s [fwd=2.694s bwd=1.044s opt=0.003s] dispatch=2.944s [2026-10-19 11:16:16,910] loop.py:45 INFO TRAIN_STEP step=10000 loss=1.423480 Training: 56%|█████▌ | 10000/18018 [21:46:02<16:35:20, 7.45s/step, loss=1.423, acc=0.678, acc_len=1.61, thru=19.7, I=18.6, T=33.7, wait=2.9s, pool=65, epoch=2/2][2026-10-19 11:16:16,916] loop.py:419 INFO Saving checkpoint at step 10000... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 11:16:20,277] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0010001 (TrainerActor pid=9310) [2026-10-19 11:16:20,283] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:21,447] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:21,447] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:27,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 11:16:20,276] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0010001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 11:16:21,298] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009251 [2026-10-19 11:16:28,166] loop.py:45 INFO TRAIN_STEP step=10001 loss=1.416033 Training: 56%|█████▌ | 10001/18018 [21:46:14<19:07:49, 8.59s/step, loss=1.416, acc=0.679, acc_len=1.61, thru=19.1, I=19.7, T=32.1, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:16:28,160] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10000: forward=2725.6ms backward=1125.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:28,342] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:34,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:16:34,905] loop.py:45 INFO TRAIN_STEP step=10002 loss=1.295021 Training: 56%|█████▌ | 10002/18018 [21:46:20<17:53:22, 8.03s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=17.7, I=19.9, T=34.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:35,063] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:39,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:16:40,817] loop.py:45 INFO TRAIN_STEP step=10003 loss=1.359997 Training: 56%|█████▌ | 10003/18018 [21:46:26<16:28:13, 7.40s/step, loss=1.360, acc=0.691, acc_len=1.65, thru=20.5, I=26.5, T=35.0, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:41,147] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:47,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:16:47,915] loop.py:45 INFO TRAIN_STEP step=10004 loss=1.324894 Training: 56%|█████▌ | 10004/18018 [21:46:34<16:16:08, 7.31s/step, loss=1.325, acc=0.696, acc_len=1.67, thru=17.6, I=21.2, T=33.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:48,068] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:54,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:16:55,754] loop.py:380 INFO TIMING step=10005: step=3.542s data=1.441s compute=3.515s [fwd=2.496s bwd=1.004s opt=0.003s] dispatch=4.287s [2026-10-19 11:16:55,755] loop.py:45 INFO TRAIN_STEP step=10005 loss=1.291371 Training: 56%|█████▌ | 10005/18018 [21:46:41<16:37:18, 7.47s/step, loss=1.291, acc=0.705, acc_len=1.70, thru=16.5, I=18.9, T=36.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:16:56,086] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:00,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:17:03,313] loop.py:45 INFO TRAIN_STEP step=10006 loss=1.260911 Training: 56%|█████▌ | 10006/18018 [21:46:49<16:40:43, 7.49s/step, loss=1.261, acc=0.708, acc_len=1.72, thru=16.3, I=26.6, T=27.3, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:17:03,309] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10005: forward=3504.1ms backward=1128.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:03,550] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:08,485] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:17:09,851] loop.py:45 INFO TRAIN_STEP step=10007 loss=1.332179 Training: 56%|█████▌ | 10007/18018 [21:46:55<16:02:25, 7.21s/step, loss=1.332, acc=0.699, acc_len=1.67, thru=16.0, I=26.6, T=35.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:09,996] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:17:15,742] loop.py:45 INFO TRAIN_STEP step=10008 loss=1.288470 Training: 56%|█████▌ | 10008/18018 [21:47:01<15:09:30, 6.81s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=20.0, I=24.7, T=33.0, wait=2.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:25,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:17:27,594] loop.py:45 INFO TRAIN_STEP step=10009 loss=1.378083 Training: 56%|█████▌ | 10009/18018 [21:47:13<18:31:13, 8.32s/step, loss=1.378, acc=0.688, acc_len=1.64, thru=11.5, I=13.4, T=29.0, wait=7.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:27,835] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:33,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:17:34,669] loop.py:380 INFO TIMING step=10010: step=4.027s data=1.826s compute=3.994s [fwd=2.811s bwd=1.168s opt=0.003s] dispatch=3.037s [2026-10-19 11:17:34,669] loop.py:45 INFO TRAIN_STEP step=10010 loss=1.314245 Training: 56%|█████▌ | 10010/18018 [21:47:20<17:41:03, 7.95s/step, loss=1.314, acc=0.697, acc_len=1.67, thru=16.4, I=21.9, T=31.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:34,916] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:40,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:17:41,077] loop.py:45 INFO TRAIN_STEP step=10011 loss=1.266011 Training: 56%|█████▌ | 10011/18018 [21:47:27<16:39:05, 7.49s/step, loss=1.266, acc=0.709, acc_len=1.71, thru=18.4, I=20.1, T=35.1, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:17:41,073] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10010: forward=2471.0ms backward=1097.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:41,356] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:47,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:17:47,562] loop.py:45 INFO TRAIN_STEP step=10012 loss=1.277060 Training: 56%|█████▌ | 10012/18018 [21:47:33<15:58:55, 7.19s/step, loss=1.277, acc=0.704, acc_len=1.69, thru=19.6, I=20.9, T=32.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:47,781] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:54,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:17:54,393] loop.py:45 INFO TRAIN_STEP step=10013 loss=1.407135 Training: 56%|█████▌ | 10013/18018 [21:47:40<15:44:35, 7.08s/step, loss=1.407, acc=0.678, acc_len=1.61, thru=17.7, I=19.3, T=32.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:17:54,618] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:18:03,393] loop.py:45 INFO TRAIN_STEP step=10014 loss=1.348035 Training: 56%|█████▌ | 10014/18018 [21:47:49<17:01:22, 7.66s/step, loss=1.348, acc=0.696, acc_len=1.67, thru=13.8, I=13.7, T=30.6, wait=4.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:11,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:18:12,383] loop.py:380 INFO TIMING step=10015: step=3.805s data=1.614s compute=3.772s [fwd=2.676s bwd=1.082s opt=0.003s] dispatch=5.173s [2026-10-19 11:18:12,384] loop.py:45 INFO TRAIN_STEP step=10015 loss=1.397680 Training: 56%|█████▌ | 10015/18018 [21:47:58<17:54:34, 8.06s/step, loss=1.398, acc=0.684, acc_len=1.63, thru=14.5, I=14.5, T=33.6, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:12,423] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 11:18:18,952] loop.py:45 INFO TRAIN_STEP step=10016 loss=1.372702 Training: 56%|█████▌ | 10016/18018 [21:48:05<16:54:53, 7.61s/step, loss=1.373, acc=0.690, acc_len=1.65, thru=18.0, I=17.3, T=33.5, wait=2.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:18:18,947] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10015: forward=2669.4ms backward=1100.6ms [2026-10-19 11:18:28,825] loop.py:45 INFO TRAIN_STEP step=10017 loss=1.465073 Training: 56%|█████▌ | 10017/18018 [21:48:14<18:25:27, 8.29s/step, loss=1.465, acc=0.676, acc_len=1.60, thru=14.3, I=13.3, T=25.2, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:28,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:29,334] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:34,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:18:36,598] loop.py:45 INFO TRAIN_STEP step=10018 loss=1.342218 Training: 56%|█████▌ | 10018/18018 [21:48:22<18:04:25, 8.13s/step, loss=1.342, acc=0.696, acc_len=1.67, thru=16.8, I=18.2, T=31.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:36,828] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:42,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:18:42,696] loop.py:45 INFO TRAIN_STEP step=10019 loss=1.383935 Training: 56%|█████▌ | 10019/18018 [21:48:28<16:42:53, 7.52s/step, loss=1.384, acc=0.688, acc_len=1.64, thru=19.1, I=23.9, T=33.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:42,726] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:49,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:18:50,981] loop.py:380 INFO TIMING step=10020: step=4.251s data=1.873s compute=4.197s [fwd=2.912s bwd=1.270s opt=0.003s] dispatch=4.024s [2026-10-19 11:18:50,981] loop.py:45 INFO TRAIN_STEP step=10020 loss=1.351448 Training: 56%|█████▌ | 10020/18018 [21:48:37<17:13:15, 7.75s/step, loss=1.351, acc=0.691, acc_len=1.64, thru=16.5, I=19.1, T=30.1, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:51,175] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:55,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:18:57,024] loop.py:45 INFO TRAIN_STEP step=10021 loss=1.229413 (TrainerActor pid=9310) [2026-10-19 11:18:57,018] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10020: forward=2940.6ms backward=1209.7ms Training: 56%|█████▌ | 10021/18018 [21:48:43<16:04:56, 7.24s/step, loss=1.229, acc=0.717, acc_len=1.74, thru=19.4, I=24.8, T=30.5, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:18:57,316] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:04,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:19:05,228] loop.py:45 INFO TRAIN_STEP step=10022 loss=1.211647 Training: 56%|█████▌ | 10022/18018 [21:48:51<16:43:14, 7.53s/step, loss=1.212, acc=0.716, acc_len=1.74, thru=13.3, I=17.6, T=29.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:05,311] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:10,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:19:11,881] loop.py:45 INFO TRAIN_STEP step=10023 loss=1.484853 Training: 56%|█████▌ | 10023/18018 [21:48:57<16:08:14, 7.27s/step, loss=1.485, acc=0.668, acc_len=1.57, thru=17.5, I=19.5, T=34.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:11,943] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 11:19:17,841] loop.py:45 INFO TRAIN_STEP step=10024 loss=1.371956 Training: 56%|█████▌ | 10024/18018 [21:49:03<15:15:56, 6.87s/step, loss=1.372, acc=0.690, acc_len=1.65, thru=20.2, I=21.7, T=37.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:17,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:18,281] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:24,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:19:24,924] loop.py:380 INFO TIMING step=10025: step=3.803s data=1.619s compute=3.771s [fwd=2.657s bwd=1.099s opt=0.003s] dispatch=3.268s [2026-10-19 11:19:24,924] loop.py:45 INFO TRAIN_STEP step=10025 loss=1.282257 Training: 56%|█████▌ | 10025/18018 [21:49:11<15:23:59, 6.94s/step, loss=1.282, acc=0.706, acc_len=1.71, thru=18.2, I=19.8, T=33.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:25,282] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:19:31,921] loop.py:45 INFO TRAIN_STEP step=10026 loss=1.319953 Training: 56%|█████▌ | 10026/18018 [21:49:18<15:26:24, 6.96s/step, loss=1.320, acc=0.694, acc_len=1.66, thru=16.1, I=18.0, T=35.7, wait=3.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:19:31,917] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10025: forward=2556.4ms backward=985.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:38,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:19:39,104] loop.py:45 INFO TRAIN_STEP step=10027 loss=1.255006 Training: 56%|█████▌ | 10027/18018 [21:49:25<15:35:17, 7.02s/step, loss=1.255, acc=0.710, acc_len=1.72, thru=18.5, I=18.5, T=33.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:39,589] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:45,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:19:46,634] loop.py:45 INFO TRAIN_STEP step=10028 loss=1.292246 Training: 56%|█████▌ | 10028/18018 [21:49:32<15:55:29, 7.18s/step, loss=1.292, acc=0.698, acc_len=1.68, thru=16.9, I=20.0, T=31.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:46,666] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:51,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:19:52,682] loop.py:45 INFO TRAIN_STEP step=10029 loss=1.388176 Training: 56%|█████▌ | 10029/18018 [21:49:38<15:10:26, 6.84s/step, loss=1.388, acc=0.688, acc_len=1.64, thru=19.2, I=24.8, T=35.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:19:52,844] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:01,492] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:20:02,129] loop.py:380 INFO TIMING step=10030: step=4.354s data=2.026s compute=4.178s [fwd=2.827s bwd=1.336s opt=0.003s] dispatch=5.082s [2026-10-19 11:20:02,129] loop.py:45 INFO TRAIN_STEP step=10030 loss=1.331052 Training: 56%|█████▌ | 10030/18018 [21:49:48<16:54:30, 7.62s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=13.8, I=14.1, T=29.4, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:02,497] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:09,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:20:10,939] loop.py:45 INFO TRAIN_STEP step=10031 loss=1.365808 Training: 56%|█████▌ | 10031/18018 [21:49:57<17:41:55, 7.98s/step, loss=1.366, acc=0.688, acc_len=1.65, thru=14.1, I=16.4, T=29.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:11,017] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (TrainerActor pid=9310) [2026-10-19 11:20:10,935] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10030: forward=3114.0ms backward=1222.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:16,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:20:16,908] loop.py:45 INFO TRAIN_STEP step=10032 loss=1.224594 Training: 56%|█████▌ | 10032/18018 [21:50:02<16:21:29, 7.37s/step, loss=1.225, acc=0.711, acc_len=1.72, thru=18.6, I=22.3, T=34.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:17,041] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 11:20:25,518] loop.py:45 INFO TRAIN_STEP step=10033 loss=1.321462 Training: 56%|█████▌ | 10033/18018 [21:50:11<17:10:51, 7.75s/step, loss=1.321, acc=0.700, acc_len=1.68, thru=13.1, I=14.8, T=35.3, wait=5.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:32,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:20:33,886] loop.py:45 INFO TRAIN_STEP step=10034 loss=1.316122 Training: 56%|█████▌ | 10034/18018 [21:50:19<17:35:29, 7.93s/step, loss=1.316, acc=0.699, acc_len=1.67, thru=16.2, I=18.6, T=31.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:34,241] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:40,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:20:40,933] loop.py:380 INFO TIMING step=10035: step=4.747s data=2.240s compute=4.713s [fwd=3.261s bwd=1.435s opt=0.003s] dispatch=2.290s [2026-10-19 11:20:40,934] loop.py:45 INFO TRAIN_STEP step=10035 loss=1.149737 Training: 56%|█████▌ | 10035/18018 [21:50:27<17:00:03, 7.67s/step, loss=1.150, acc=0.728, acc_len=1.78, thru=16.4, I=20.9, T=27.0, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:41,076] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:47,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:20:48,677] loop.py:45 INFO TRAIN_STEP step=10036 loss=1.294924 Training: 56%|█████▌ | 10036/18018 [21:50:34<17:03:08, 7.69s/step, loss=1.295, acc=0.697, acc_len=1.67, thru=16.0, I=17.2, T=29.7, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:20:48,670] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10035: forward=3044.7ms backward=1214.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:48,753] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:54,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:20:55,284] loop.py:45 INFO TRAIN_STEP step=10037 loss=1.314635 Training: 56%|█████▌ | 10037/18018 [21:50:41<16:19:36, 7.36s/step, loss=1.315, acc=0.699, acc_len=1.67, thru=17.6, I=23.5, T=32.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:20:55,628] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:01,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:21:01,802] loop.py:45 INFO TRAIN_STEP step=10038 loss=1.320231 Training: 56%|█████▌ | 10038/18018 [21:50:47<15:45:42, 7.11s/step, loss=1.320, acc=0.701, acc_len=1.69, thru=18.1, I=22.5, T=35.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:01,837] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:07,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:21:08,285] loop.py:45 INFO TRAIN_STEP step=10039 loss=1.322104 Training: 56%|█████▌ | 10039/18018 [21:50:54<15:20:36, 6.92s/step, loss=1.322, acc=0.699, acc_len=1.68, thru=19.0, I=21.5, T=35.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:08,470] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:13,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:21:14,381] loop.py:380 INFO TIMING step=10040: step=3.755s data=1.719s compute=3.716s [fwd=2.741s bwd=0.960s opt=0.003s] dispatch=2.329s [2026-10-19 11:21:14,382] loop.py:45 INFO TRAIN_STEP step=10040 loss=1.415969 Training: 56%|█████▌ | 10040/18018 [21:51:00<14:47:29, 6.67s/step, loss=1.416, acc=0.680, acc_len=1.61, thru=20.0, I=24.0, T=34.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:14,760] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:21,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:21:21,856] loop.py:45 INFO TRAIN_STEP step=10041 loss=1.430091 Training: 56%|█████▌ | 10041/18018 [21:51:07<15:19:13, 6.91s/step, loss=1.430, acc=0.673, acc_len=1.59, thru=16.2, I=18.1, T=35.1, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:21:21,852] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10040: forward=2599.8ms backward=990.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:22,205] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:28,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:21:29,167] loop.py:45 INFO TRAIN_STEP step=10042 loss=1.327839 Training: 56%|█████▌ | 10042/18018 [21:51:15<15:34:56, 7.03s/step, loss=1.328, acc=0.698, acc_len=1.68, thru=18.5, I=20.9, T=33.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:29,650] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:34,601] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:21:36,788] loop.py:45 INFO TRAIN_STEP step=10043 loss=1.407887 Training: 56%|█████▌ | 10043/18018 [21:51:22<15:58:20, 7.21s/step, loss=1.408, acc=0.682, acc_len=1.62, thru=16.1, I=23.3, T=31.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:37,119] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:43,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:21:44,797] loop.py:45 INFO TRAIN_STEP step=10044 loss=1.283939 Training: 56%|█████▌ | 10044/18018 [21:51:30<16:30:04, 7.45s/step, loss=1.284, acc=0.702, acc_len=1.69, thru=16.7, I=18.6, T=31.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:44,866] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:50,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:21:52,254] loop.py:380 INFO TIMING step=10045: step=4.086s data=1.835s compute=4.059s [fwd=2.872s bwd=1.172s opt=0.003s] dispatch=3.360s [2026-10-19 11:21:52,254] loop.py:45 INFO TRAIN_STEP step=10045 loss=1.325460 Training: 56%|█████▌ | 10045/18018 [21:51:38<16:30:07, 7.45s/step, loss=1.325, acc=0.694, acc_len=1.66, thru=16.9, I=22.3, T=31.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:52,717] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:58,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:21:59,406] loop.py:45 INFO TRAIN_STEP step=10046 loss=1.260157 Training: 56%|█████▌ | 10046/18018 [21:51:45<16:18:07, 7.36s/step, loss=1.260, acc=0.707, acc_len=1.70, thru=15.4, I=23.1, T=31.8, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:21:59,402] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10045: forward=2864.3ms backward=1107.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:21:59,637] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:04,860] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:22:06,674] loop.py:45 INFO TRAIN_STEP step=10047 loss=1.475238 Training: 56%|█████▌ | 10047/18018 [21:51:52<16:14:19, 7.33s/step, loss=1.475, acc=0.671, acc_len=1.58, thru=16.9, I=23.0, T=29.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:06,871] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:22:13,584] loop.py:45 INFO TRAIN_STEP step=10048 loss=1.278291 Training: 56%|█████▌ | 10048/18018 [21:51:59<15:57:13, 7.21s/step, loss=1.278, acc=0.709, acc_len=1.71, thru=15.1, I=19.7, T=28.8, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:22,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:22:23,466] loop.py:45 INFO TRAIN_STEP step=10049 loss=1.344412 Training: 56%|█████▌ | 10049/18018 [21:52:09<17:43:43, 8.01s/step, loss=1.344, acc=0.696, acc_len=1.66, thru=14.0, I=14.0, T=31.0, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:23,722] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:29,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:22:29,686] loop.py:380 INFO TIMING step=10050: step=3.798s data=1.557s compute=3.769s [fwd=2.679s bwd=1.075s opt=0.003s] dispatch=2.413s [2026-10-19 11:22:29,686] loop.py:45 INFO TRAIN_STEP step=10050 loss=1.338300 Training: 56%|█████▌ | 10050/18018 [21:52:15<16:32:22, 7.47s/step, loss=1.338, acc=0.698, acc_len=1.67, thru=19.3, I=22.5, T=33.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:29,956] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:22:36,314] loop.py:45 INFO TRAIN_STEP step=10051 loss=1.185147 Training: 56%|█████▌ | 10051/18018 [21:52:22<15:58:35, 7.22s/step, loss=1.185, acc=0.724, acc_len=1.77, thru=15.6, I=18.0, T=34.1, wait=2.9s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:22:36,310] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10050: forward=2650.7ms backward=1057.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:45,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:22:46,535] loop.py:45 INFO TRAIN_STEP step=10052 loss=1.138849 Training: 56%|█████▌ | 10052/18018 [21:52:32<17:58:02, 8.12s/step, loss=1.139, acc=0.731, acc_len=1.79, thru=15.7, I=15.4, T=31.4, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:46,712] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:53,498] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:22:54,366] loop.py:45 INFO TRAIN_STEP step=10053 loss=1.333483 Training: 56%|█████▌ | 10053/18018 [21:52:40<17:46:25, 8.03s/step, loss=1.333, acc=0.696, acc_len=1.67, thru=15.8, I=18.4, T=30.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:22:54,504] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:00,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:23:00,592] loop.py:45 INFO TRAIN_STEP step=10054 loss=1.284451 Training: 56%|█████▌ | 10054/18018 [21:52:46<16:34:15, 7.49s/step, loss=1.284, acc=0.705, acc_len=1.70, thru=18.4, I=21.6, T=35.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:00,870] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:06,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:23:07,459] loop.py:380 INFO TIMING step=10055: step=3.555s data=1.498s compute=3.520s [fwd=2.526s bwd=0.978s opt=0.003s] dispatch=3.302s [2026-10-19 11:23:07,459] loop.py:45 INFO TRAIN_STEP step=10055 loss=1.288454 Training: 56%|█████▌ | 10055/18018 [21:52:53<16:09:18, 7.30s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=18.6, I=20.6, T=36.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:07,654] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:13,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:23:13,328] loop.py:45 INFO TRAIN_STEP step=10056 loss=1.256585 Training: 56%|█████▌ | 10056/18018 [21:52:59<15:12:01, 6.87s/step, loss=1.257, acc=0.710, acc_len=1.72, thru=21.0, I=22.7, T=36.3, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:23:13,323] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10055: forward=2542.7ms backward=938.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:13,802] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:23:20,446] loop.py:45 INFO TRAIN_STEP step=10057 loss=1.349406 Training: 56%|█████▌ | 10057/18018 [21:53:06<15:21:42, 6.95s/step, loss=1.349, acc=0.691, acc_len=1.65, thru=16.6, I=17.5, T=31.5, wait=3.0s, pool=53, epoch=2/2][2026-10-19 11:23:27,077] loop.py:45 INFO TRAIN_STEP step=10058 loss=1.361065 Training: 56%|█████▌ | 10058/18018 [21:53:13<15:09:01, 6.85s/step, loss=1.361, acc=0.690, acc_len=1.65, thru=16.7, I=16.5, T=36.0, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:32,854] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:23:34,124] loop.py:45 INFO TRAIN_STEP step=10059 loss=1.299966 Training: 56%|█████▌ | 10059/18018 [21:53:20<15:16:43, 6.91s/step, loss=1.300, acc=0.703, acc_len=1.69, thru=18.9, I=20.4, T=34.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:34,363] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:41,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:23:43,334] loop.py:380 INFO TIMING step=10060: step=4.468s data=2.296s compute=4.435s [fwd=3.392s bwd=1.028s opt=0.003s] dispatch=4.731s [2026-10-19 11:23:43,334] loop.py:45 INFO TRAIN_STEP step=10060 loss=1.304627 Training: 56%|█████▌ | 10060/18018 [21:53:29<16:48:06, 7.60s/step, loss=1.305, acc=0.699, acc_len=1.68, thru=14.3, I=17.9, T=28.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:43,463] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:48,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:23:50,613] loop.py:45 INFO TRAIN_STEP step=10061 loss=1.366135 Training: 56%|█████▌ | 10061/18018 [21:53:36<16:35:11, 7.50s/step, loss=1.366, acc=0.689, acc_len=1.65, thru=15.8, I=24.7, T=30.5, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:23:50,609] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10060: forward=3113.5ms backward=1031.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:51,005] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:56,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:23:57,510] loop.py:45 INFO TRAIN_STEP step=10062 loss=1.421072 Training: 56%|█████▌ | 10062/18018 [21:53:43<16:10:49, 7.32s/step, loss=1.421, acc=0.678, acc_len=1.61, thru=17.4, I=23.6, T=33.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:23:57,961] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:24:04,457] loop.py:45 INFO TRAIN_STEP step=10063 loss=1.264453 Training: 56%|█████▌ | 10063/18018 [21:53:50<15:55:47, 7.21s/step, loss=1.264, acc=0.711, acc_len=1.72, thru=16.9, I=19.2, T=36.0, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:24:11,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:24:12,487] loop.py:45 INFO TRAIN_STEP step=10064 loss=1.257435 Training: 56%|█████▌ | 10064/18018 [21:53:58<16:28:25, 7.46s/step, loss=1.257, acc=0.709, acc_len=1.71, thru=16.2, I=17.1, T=33.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:24:12,791] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:24:19,188] loop.py:380 INFO TIMING step=10065: step=3.669s data=1.549s compute=3.636s [fwd=2.585s bwd=1.036s opt=0.003s] dispatch=3.022s [2026-10-19 11:24:19,189] loop.py:45 INFO TRAIN_STEP step=10065 loss=1.338308 Training: 56%|█████▌ | 10065/18018 [21:54:05<15:58:11, 7.23s/step, loss=1.338, acc=0.695, acc_len=1.66, thru=16.0, I=16.8, T=34.9, wait=3.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:24:25,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:24:25,978] loop.py:45 INFO TRAIN_STEP step=10066 loss=1.302061 Training: 56%|█████▌ | 10066/18018 [21:54:12<15:40:39, 7.10s/step, loss=1.302, acc=0.700, acc_len=1.68, thru=19.8, I=17.8, T=35.8, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:24:25,973] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10065: forward=2538.1ms backward=979.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:24:26,328] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:24:30,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:24:31,826] loop.py:45 INFO TRAIN_STEP step=10067 loss=1.376918 Training: 56%|█████▌ | 10067/18018 [21:54:17<14:50:53, 6.72s/step, loss=1.377, acc=0.687, acc_len=1.63, thru=21.1, I=25.4, T=34.6, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:24:32,267] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:24:38,909] loop.py:45 INFO TRAIN_STEP step=10068 loss=1.319295 Training: 56%|█████▌ | 10068/18018 [21:54:24<15:05:03, 6.83s/step, loss=1.319, acc=0.698, acc_len=1.68, thru=14.8, I=20.3, T=37.3, wait=3.6s, pool=41, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 11:24:38,985] eagle_store.py:517 WARNING batch_get_buffer missing keys (k322182-8d628c9b_hs, k322182-8d628c9b_ids, k322182-8d628c9b_lhs); sleeping 0.50s. [2026-10-19 11:24:46,923] loop.py:45 INFO TRAIN_STEP step=10069 loss=1.388176 Training: 56%|█████▌ | 10069/18018 [21:54:33<15:52:01, 7.19s/step, loss=1.388, acc=0.685, acc_len=1.63, thru=16.2, I=17.3, T=28.7, wait=3.5s, pool=53, epoch=2/2][2026-10-19 11:24:57,143] loop.py:380 INFO TIMING step=10070: step=4.247s data=1.868s compute=4.207s [fwd=2.973s bwd=1.218s opt=0.003s] dispatch=5.963s [2026-10-19 11:24:57,144] loop.py:45 INFO TRAIN_STEP step=10070 loss=1.343853 Training: 56%|█████▌ | 10070/18018 [21:54:43<17:52:25, 8.10s/step, loss=1.344, acc=0.696, acc_len=1.67, thru=10.3, I=10.7, T=30.1, wait=6.0s, pool=25, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 11:24:57,214] eagle_store.py:517 WARNING batch_get_buffer missing keys (k322238-a69389f3_hs, k322238-a69389f3_ids, k322238-a69389f3_lhs); sleeping 0.50s. [repeated 3x across cluster] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:04,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 11:24:57,241] eagle_store.py:517 WARNING batch_get_buffer missing keys (k322236-92564cab_hs, k322236-92564cab_ids, k322236-92564cab_lhs); sleeping 0.50s. [2026-10-19 11:25:06,247] loop.py:45 INFO TRAIN_STEP step=10071 loss=1.468834 Training: 56%|█████▌ | 10071/18018 [21:54:52<18:32:19, 8.40s/step, loss=1.469, acc=0.669, acc_len=1.58, thru=18.0, I=16.6, T=31.4, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:25:06,243] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10070: forward=2797.0ms backward=1220.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:06,730] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:25:12,360] loop.py:45 INFO TRAIN_STEP step=10072 loss=1.318729 Training: 56%|█████▌ | 10072/18018 [21:54:58<17:01:30, 7.71s/step, loss=1.319, acc=0.695, acc_len=1.67, thru=15.4, I=20.3, T=36.0, wait=2.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:19,652] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:25:19,893] loop.py:45 INFO TRAIN_STEP step=10073 loss=1.367629 Training: 56%|█████▌ | 10073/18018 [21:55:05<16:54:10, 7.66s/step, loss=1.368, acc=0.689, acc_len=1.65, thru=18.5, I=17.8, T=32.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:20,154] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:27,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:25:27,954] loop.py:45 INFO TRAIN_STEP step=10074 loss=1.407290 Training: 56%|█████▌ | 10074/18018 [21:55:14<17:10:02, 7.78s/step, loss=1.407, acc=0.688, acc_len=1.63, thru=16.0, I=16.1, T=29.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:28,192] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:34,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:25:34,722] loop.py:380 INFO TIMING step=10075: step=4.073s data=1.866s compute=4.039s [fwd=2.869s bwd=1.154s opt=0.003s] dispatch=2.684s [2026-10-19 11:25:34,722] loop.py:45 INFO TRAIN_STEP step=10075 loss=1.459616 Training: 56%|█████▌ | 10075/18018 [21:55:20<16:29:43, 7.48s/step, loss=1.460, acc=0.674, acc_len=1.60, thru=17.0, I=19.0, T=31.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:34,938] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:25:41,979] loop.py:45 INFO TRAIN_STEP step=10076 loss=1.202716 Training: 56%|█████▌ | 10076/18018 [21:55:28<16:20:50, 7.41s/step, loss=1.203, acc=0.715, acc_len=1.74, thru=15.3, I=17.3, T=29.2, wait=2.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:25:41,973] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10075: forward=3203.8ms backward=1129.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:51,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:25:52,167] loop.py:45 INFO TRAIN_STEP step=10077 loss=1.344488 Training: 56%|█████▌ | 10077/18018 [21:55:38<18:10:56, 8.24s/step, loss=1.344, acc=0.694, acc_len=1.66, thru=14.6, I=14.7, T=31.6, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:52,598] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:58,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:25:58,860] loop.py:45 INFO TRAIN_STEP step=10078 loss=1.351171 Training: 56%|█████▌ | 10078/18018 [21:55:44<17:09:23, 7.78s/step, loss=1.351, acc=0.692, acc_len=1.65, thru=18.2, I=21.1, T=36.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:25:59,030] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:26:05,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:26:05,374] loop.py:45 INFO TRAIN_STEP step=10079 loss=1.314515 Training: 56%|█████▌ | 10079/18018 [21:55:51<16:19:03, 7.40s/step, loss=1.315, acc=0.698, acc_len=1.68, thru=19.5, I=21.1, T=36.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:26:05,618] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:26:12,340] loop.py:380 INFO TIMING step=10080: step=3.907s data=1.701s compute=3.864s [fwd=2.740s bwd=1.109s opt=0.003s] dispatch=3.049s [2026-10-19 11:26:12,340] loop.py:45 INFO TRAIN_STEP step=10080 loss=1.422240 Training: 56%|█████▌ | 10080/18018 [21:55:58<16:01:42, 7.27s/step, loss=1.422, acc=0.677, acc_len=1.60, thru=17.7, I=18.7, T=32.8, wait=3.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:26:17,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:26:18,412] loop.py:45 INFO TRAIN_STEP step=10081 loss=1.382470 Training: 56%|█████▌ | 10081/18018 [21:56:04<15:14:07, 6.91s/step, loss=1.382, acc=0.684, acc_len=1.63, thru=19.3, I=21.3, T=35.6, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:26:18,407] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10080: forward=2535.2ms backward=1006.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:26:18,466] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 11:26:26,486] loop.py:45 INFO TRAIN_STEP step=10082 loss=1.440545 Training: 56%|█████▌ | 10082/18018 [21:56:12<16:00:06, 7.26s/step, loss=1.441, acc=0.676, acc_len=1.60, thru=13.7, I=18.2, T=32.4, wait=4.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:26:38,567] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:26:38,820] loop.py:45 INFO TRAIN_STEP step=10083 loss=1.452244 Training: 56%|█████▌ | 10083/18018 [21:56:24<19:21:23, 8.78s/step, loss=1.452, acc=0.675, acc_len=1.60, thru=15.3, I=12.0, T=28.9, wait=7.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:26:39,069] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:26:48,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:26:50,038] loop.py:45 INFO TRAIN_STEP step=10084 loss=1.386899 Training: 56%|█████▌ | 10084/18018 [21:56:36<20:57:55, 9.51s/step, loss=1.387, acc=0.689, acc_len=1.65, thru=12.3, I=12.9, T=28.3, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:26:50,145] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 11:26:57,118] loop.py:380 INFO TIMING step=10085: step=3.743s data=1.517s compute=3.704s [fwd=2.597s bwd=1.091s opt=0.003s] dispatch=3.326s [2026-10-19 11:26:57,119] loop.py:45 INFO TRAIN_STEP step=10085 loss=1.392390 Training: 56%|█████▌ | 10085/18018 [21:56:43<19:21:24, 8.78s/step, loss=1.392, acc=0.685, acc_len=1.64, thru=14.9, I=19.1, T=34.2, wait=3.3s, pool=45, epoch=2/2][2026-10-19 11:27:05,525] loop.py:45 INFO TRAIN_STEP step=10086 loss=1.292660 Training: 56%|█████▌ | 10086/18018 [21:56:51<19:06:05, 8.67s/step, loss=1.293, acc=0.705, acc_len=1.70, thru=14.6, I=15.1, T=31.2, wait=4.3s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:27:05,520] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10085: forward=2688.2ms backward=1352.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:15,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:27:15,651] loop.py:45 INFO TRAIN_STEP step=10087 loss=1.328257 Training: 56%|█████▌ | 10087/18018 [21:57:01<20:03:45, 9.11s/step, loss=1.328, acc=0.694, acc_len=1.66, thru=15.6, I=14.3, T=26.7, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:16,095] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:23,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:27:24,260] loop.py:45 INFO TRAIN_STEP step=10088 loss=1.250244 Training: 56%|█████▌ | 10088/18018 [21:57:10<19:43:46, 8.96s/step, loss=1.250, acc=0.711, acc_len=1.72, thru=14.5, I=17.5, T=28.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:24,745] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:30,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:27:30,891] loop.py:45 INFO TRAIN_STEP step=10089 loss=1.336174 Training: 56%|█████▌ | 10089/18018 [21:57:16<18:11:31, 8.26s/step, loss=1.336, acc=0.694, acc_len=1.66, thru=18.3, I=18.9, T=34.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:30,964] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:35,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:27:37,708] loop.py:380 INFO TIMING step=10090: step=3.950s data=1.760s compute=3.919s [fwd=2.785s bwd=1.119s opt=0.003s] dispatch=2.857s [2026-10-19 11:27:37,708] loop.py:45 INFO TRAIN_STEP step=10090 loss=1.297276 Training: 56%|█████▌ | 10090/18018 [21:57:23<17:14:04, 7.83s/step, loss=1.297, acc=0.702, acc_len=1.69, thru=19.1, I=23.1, T=32.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:38,195] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:43,079] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:27:44,911] loop.py:45 INFO TRAIN_STEP step=10091 loss=1.351578 Training: 56%|█████▌ | 10091/18018 [21:57:30<16:49:19, 7.64s/step, loss=1.352, acc=0.689, acc_len=1.64, thru=15.7, I=26.3, T=29.8, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:27:44,907] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10090: forward=2913.5ms backward=1322.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:45,093] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:27:51,382] loop.py:45 INFO TRAIN_STEP step=10092 loss=1.356977 Training: 56%|█████▌ | 10092/18018 [21:57:37<16:02:55, 7.29s/step, loss=1.357, acc=0.690, acc_len=1.65, thru=15.5, I=20.3, T=32.0, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:27:59,498] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:28:01,397] loop.py:45 INFO TRAIN_STEP step=10093 loss=1.273384 Training: 56%|█████▌ | 10093/18018 [21:57:47<17:50:48, 8.11s/step, loss=1.273, acc=0.704, acc_len=1.69, thru=14.2, I=15.9, T=25.9, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:28:01,517] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:28:06,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:28:08,686] loop.py:45 INFO TRAIN_STEP step=10094 loss=1.257007 Training: 56%|█████▌ | 10094/18018 [21:57:54<17:18:14, 7.86s/step, loss=1.257, acc=0.708, acc_len=1.71, thru=14.3, I=20.5, T=31.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:28:08,960] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:28:16,018] loop.py:380 INFO TIMING step=10095: step=4.475s data=2.131s compute=4.441s [fwd=3.117s bwd=1.307s opt=0.003s] dispatch=2.848s [2026-10-19 11:28:16,018] loop.py:45 INFO TRAIN_STEP step=10095 loss=1.233034 Training: 56%|█████▌ | 10095/18018 [21:58:02<16:57:04, 7.70s/step, loss=1.233, acc=0.712, acc_len=1.73, thru=14.6, I=20.7, T=28.6, wait=2.8s, pool=49, epoch=2/2](TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 11:28:16,075] eagle_store.py:517 WARNING batch_get_buffer missing keys (k323043-a4debec3_hs, k323043-a4debec3_ids, k323043-a4debec3_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:28:21,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:28:23,075] loop.py:45 INFO TRAIN_STEP step=10096 loss=1.261039 Training: 56%|█████▌ | 10096/18018 [21:58:09<16:31:26, 7.51s/step, loss=1.261, acc=0.712, acc_len=1.72, thru=18.6, I=19.1, T=33.1, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:28:23,070] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10095: forward=2708.5ms backward=1096.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:28:23,330] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:28:30,130] loop.py:45 INFO TRAIN_STEP step=10097 loss=1.325876 Training: 56%|█████▌ | 10097/18018 [21:58:16<16:13:23, 7.37s/step, loss=1.326, acc=0.699, acc_len=1.67, thru=13.2, I=17.2, T=30.8, wait=2.9s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:28:38,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:28:39,786] loop.py:45 INFO TRAIN_STEP step=10098 loss=1.368087 Training: 56%|█████▌ | 10098/18018 [21:58:25<17:43:38, 8.06s/step, loss=1.368, acc=0.689, acc_len=1.65, thru=17.1, I=17.1, T=29.3, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:28:39,820] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 11:28:47,444] loop.py:45 INFO TRAIN_STEP step=10099 loss=1.306976 Training: 56%|█████▌ | 10099/18018 [21:58:33<17:27:34, 7.94s/step, loss=1.307, acc=0.701, acc_len=1.69, thru=14.5, I=18.9, T=29.3, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:28:53,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:28:55,099] loop.py:380 INFO TIMING step=10100: step=4.195s data=1.652s compute=4.159s [fwd=2.783s bwd=1.360s opt=0.003s] dispatch=3.452s [2026-10-19 11:28:55,100] loop.py:45 INFO TRAIN_STEP step=10100 loss=1.316623 Training: 56%|█████▌ | 10100/18018 [21:58:41<17:16:18, 7.85s/step, loss=1.317, acc=0.697, acc_len=1.67, thru=14.8, I=17.3, T=30.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:28:55,587] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:02,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:29:03,883] loop.py:45 INFO TRAIN_STEP step=10101 loss=1.376122 Training: 56%|█████▌ | 10101/18018 [21:58:49<17:53:04, 8.13s/step, loss=1.376, acc=0.689, acc_len=1.65, thru=15.4, I=17.4, T=29.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:03,939] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (TrainerActor pid=9310) [2026-10-19 11:29:03,879] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10100: forward=3059.3ms backward=1280.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:09,207] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:29:10,705] loop.py:45 INFO TRAIN_STEP step=10102 loss=1.251401 Training: 56%|█████▌ | 10102/18018 [21:58:56<17:01:02, 7.74s/step, loss=1.251, acc=0.710, acc_len=1.72, thru=17.4, I=21.3, T=30.1, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:10,735] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:18,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:29:19,518] loop.py:45 INFO TRAIN_STEP step=10103 loss=1.419269 Training: 56%|█████▌ | 10103/18018 [21:59:05<17:43:23, 8.06s/step, loss=1.419, acc=0.680, acc_len=1.62, thru=14.6, I=16.0, T=29.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:19,737] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:26,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:29:26,404] loop.py:45 INFO TRAIN_STEP step=10104 loss=1.408448 Training: 56%|█████▌ | 10104/18018 [21:59:12<16:56:50, 7.71s/step, loss=1.408, acc=0.683, acc_len=1.62, thru=17.0, I=19.4, T=32.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:26,664] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:32,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:29:33,860] loop.py:380 INFO TIMING step=10105: step=4.103s data=1.708s compute=4.062s [fwd=2.835s bwd=1.210s opt=0.003s] dispatch=3.342s [2026-10-19 11:29:33,860] loop.py:45 INFO TRAIN_STEP step=10105 loss=1.390108 Training: 56%|█████▌ | 10105/18018 [21:59:19<16:46:41, 7.63s/step, loss=1.390, acc=0.684, acc_len=1.63, thru=17.4, I=18.2, T=31.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:34,114] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:39,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:29:39,985] loop.py:45 INFO TRAIN_STEP step=10106 loss=1.261195 Training: 56%|█████▌ | 10106/18018 [21:59:26<15:46:53, 7.18s/step, loss=1.261, acc=0.706, acc_len=1.71, thru=19.2, I=23.4, T=33.1, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:29:39,980] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10105: forward=2721.2ms backward=1100.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:29:40,237] inference_manager.py:408 INFO Pool capacity available after 1.1s, pool_size=1/64 [2026-10-19 11:29:46,047] loop.py:45 INFO TRAIN_STEP step=10107 loss=1.364526 Training: 56%|█████▌ | 10107/18018 [21:59:32<15:02:28, 6.84s/step, loss=1.365, acc=0.689, acc_len=1.64, thru=19.6, I=22.5, T=36.0, wait=2.5s, pool=61, epoch=2/2][2026-10-19 11:29:57,677] loop.py:45 INFO TRAIN_STEP step=10108 loss=1.386314 Training: 56%|█████▌ | 10108/18018 [21:59:43<18:11:34, 8.28s/step, loss=1.386, acc=0.688, acc_len=1.64, thru=9.7, I=10.4, T=31.2, wait=7.5s, pool=45, epoch=2/2] [2026-10-19 11:30:10,778] loop.py:45 INFO TRAIN_STEP step=10109 loss=1.277152 Training: 56%|█████▌ | 10109/18018 [21:59:56<21:22:04, 9.73s/step, loss=1.277, acc=0.703, acc_len=1.69, thru=10.7, I=9.8, T=26.4, wait=8.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:10,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:11,261] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:17,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:30:18,987] loop.py:380 INFO TIMING step=10110: step=4.220s data=1.871s compute=4.180s [fwd=3.011s bwd=1.154s opt=0.003s] dispatch=3.981s [2026-10-19 11:30:18,987] loop.py:45 INFO TRAIN_STEP step=10110 loss=1.295024 Training: 56%|█████▌ | 10110/18018 [22:00:05<20:21:58, 9.27s/step, loss=1.295, acc=0.701, acc_len=1.69, thru=16.1, I=16.9, T=30.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:19,260] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:24,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:30:25,597] loop.py:45 INFO TRAIN_STEP step=10111 loss=1.366319 Training: 56%|█████▌ | 10111/18018 [22:00:11<18:36:36, 8.47s/step, loss=1.366, acc=0.693, acc_len=1.66, thru=16.8, I=23.7, T=32.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:30:25,593] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10110: forward=2771.8ms backward=1066.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:25,877] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:31,019] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:30:32,489] loop.py:45 INFO TRAIN_STEP step=10112 loss=1.237767 Training: 56%|█████▌ | 10112/18018 [22:00:18<17:33:56, 8.00s/step, loss=1.238, acc=0.711, acc_len=1.72, thru=17.4, I=22.9, T=31.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:32,539] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:40,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:30:40,633] loop.py:45 INFO TRAIN_STEP step=10113 loss=1.366483 Training: 56%|█████▌ | 10113/18018 [22:00:26<17:39:32, 8.04s/step, loss=1.366, acc=0.688, acc_len=1.64, thru=14.1, I=16.3, T=30.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:41,033] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:48,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:30:50,223] loop.py:45 INFO TRAIN_STEP step=10114 loss=1.287082 Training: 56%|█████▌ | 10114/18018 [22:00:36<18:40:35, 8.51s/step, loss=1.287, acc=0.702, acc_len=1.69, thru=13.8, I=16.8, T=28.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:50,384] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:54,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:30:56,138] loop.py:380 INFO TIMING step=10115: step=3.587s data=1.530s compute=3.556s [fwd=2.535s bwd=1.006s opt=0.003s] dispatch=2.320s [2026-10-19 11:30:56,138] loop.py:45 INFO TRAIN_STEP step=10115 loss=1.421982 Training: 56%|█████▌ | 10115/18018 [22:00:42<16:58:07, 7.73s/step, loss=1.422, acc=0.680, acc_len=1.61, thru=18.8, I=25.4, T=35.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:30:56,489] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:02,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:31:02,812] loop.py:45 INFO TRAIN_STEP step=10116 loss=1.306591 Training: 56%|█████▌ | 10116/18018 [22:00:48<16:16:17, 7.41s/step, loss=1.307, acc=0.699, acc_len=1.68, thru=19.1, I=23.4, T=35.7, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:31:02,807] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10115: forward=2569.0ms backward=967.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:03,272] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:31:09,816] loop.py:45 INFO TRAIN_STEP step=10117 loss=1.284276 Training: 56%|█████▌ | 10117/18018 [22:00:55<15:59:56, 7.29s/step, loss=1.284, acc=0.704, acc_len=1.70, thru=16.9, I=18.1, T=35.3, wait=3.4s, pool=57, epoch=2/2][2026-10-19 11:31:18,112] loop.py:45 INFO TRAIN_STEP step=10118 loss=1.290116 Training: 56%|█████▌ | 10118/18018 [22:01:04<16:39:33, 7.59s/step, loss=1.290, acc=0.705, acc_len=1.70, thru=15.9, I=15.2, T=32.1, wait=4.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:25,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:31:25,953] loop.py:45 INFO TRAIN_STEP step=10119 loss=1.384707 Training: 56%|█████▌ | 10119/18018 [22:01:12<16:49:21, 7.67s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=16.8, I=16.8, T=31.1, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:26,405] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:33,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:31:33,450] loop.py:380 INFO TIMING step=10120: step=4.155s data=1.783s compute=4.123s [fwd=2.937s bwd=1.170s opt=0.003s] dispatch=3.331s [2026-10-19 11:31:33,451] loop.py:45 INFO TRAIN_STEP step=10120 loss=1.434237 Training: 56%|█████▌ | 10120/18018 [22:01:19<16:42:32, 7.62s/step, loss=1.434, acc=0.677, acc_len=1.60, thru=16.9, I=19.3, T=30.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:33,605] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:40,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 11:31:41,251] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10120: forward=3214.9ms backward=1384.7ms [2026-10-19 11:31:41,255] loop.py:45 INFO TRAIN_STEP step=10121 loss=1.366375 Training: 56%|█████▌ | 10121/18018 [22:01:27<16:49:50, 7.67s/step, loss=1.366, acc=0.691, acc_len=1.65, thru=15.0, I=15.4, T=27.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:41,471] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:46,648] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:31:47,249] loop.py:45 INFO TRAIN_STEP step=10122 loss=1.306134 Training: 56%|█████▌ | 10122/18018 [22:01:33<15:43:23, 7.17s/step, loss=1.306, acc=0.699, acc_len=1.68, thru=19.9, I=19.5, T=35.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:47,652] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:53,119] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:31:54,027] loop.py:45 INFO TRAIN_STEP step=10123 loss=1.401939 Training: 56%|█████▌ | 10123/18018 [22:01:40<15:27:54, 7.05s/step, loss=1.402, acc=0.681, acc_len=1.62, thru=18.9, I=22.3, T=30.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:31:54,141] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:03,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:32:03,894] loop.py:45 INFO TRAIN_STEP step=10124 loss=1.308323 Training: 56%|█████▌ | 10124/18018 [22:01:49<17:18:50, 7.90s/step, loss=1.308, acc=0.701, acc_len=1.69, thru=13.2, I=14.1, T=27.8, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:04,061] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:09,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:32:10,707] loop.py:380 INFO TIMING step=10125: step=4.289s data=1.847s compute=4.250s [fwd=3.103s bwd=1.131s opt=0.003s] dispatch=2.515s [2026-10-19 11:32:10,707] loop.py:45 INFO TRAIN_STEP step=10125 loss=1.228404 Training: 56%|█████▌ | 10125/18018 [22:01:56<16:35:58, 7.57s/step, loss=1.228, acc=0.713, acc_len=1.73, thru=17.6, I=22.1, T=29.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:11,116] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:16,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:32:19,278] loop.py:45 INFO TRAIN_STEP step=10126 loss=1.260735 Training: 56%|█████▌ | 10126/18018 [22:02:05<17:15:16, 7.87s/step, loss=1.261, acc=0.709, acc_len=1.71, thru=15.7, I=22.1, T=28.4, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:32:19,273] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10125: forward=3161.6ms backward=1298.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:19,704] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:26,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:32:26,890] loop.py:45 INFO TRAIN_STEP step=10127 loss=1.333664 Training: 56%|█████▌ | 10127/18018 [22:02:12<17:05:02, 7.79s/step, loss=1.334, acc=0.694, acc_len=1.66, thru=17.8, I=19.9, T=33.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:27,034] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:33,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:32:33,859] loop.py:45 INFO TRAIN_STEP step=10128 loss=1.245266 Training: 56%|█████▌ | 10128/18018 [22:02:19<16:32:15, 7.55s/step, loss=1.245, acc=0.711, acc_len=1.72, thru=18.4, I=20.8, T=32.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:34,307] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:39,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:32:40,438] loop.py:45 INFO TRAIN_STEP step=10129 loss=1.313961 Training: 56%|█████▌ | 10129/18018 [22:02:26<15:54:06, 7.26s/step, loss=1.314, acc=0.696, acc_len=1.67, thru=18.4, I=21.2, T=32.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:32:40,630] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:32:47,259] loop.py:380 INFO TIMING step=10130: step=3.793s data=1.655s compute=3.766s [fwd=2.640s bwd=1.112s opt=0.003s] dispatch=3.017s [2026-10-19 11:32:47,259] loop.py:45 INFO TRAIN_STEP step=10130 loss=1.380868 Training: 56%|█████▌ | 10130/18018 [22:02:33<15:36:45, 7.13s/step, loss=1.381, acc=0.689, acc_len=1.64, thru=16.6, I=19.9, T=33.7, wait=3.0s, pool=53, epoch=2/2][2026-10-19 11:32:57,180] loop.py:45 INFO TRAIN_STEP step=10131 loss=1.353660 Training: 56%|█████▌ | 10131/18018 [22:02:43<17:26:54, 7.96s/step, loss=1.354, acc=0.690, acc_len=1.65, thru=13.0, I=12.6, T=32.5, wait=6.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:32:57,176] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10130: forward=2709.4ms backward=1165.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:33:04,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:33:05,873] loop.py:45 INFO TRAIN_STEP step=10132 loss=1.358892 Training: 56%|█████▌ | 10132/18018 [22:02:51<17:55:26, 8.18s/step, loss=1.359, acc=0.692, acc_len=1.65, thru=16.6, I=18.9, T=29.8, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:33:06,028] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:33:12,913] loop.py:45 INFO TRAIN_STEP step=10133 loss=1.283568 Training: 56%|█████▌ | 10133/18018 [22:02:58<17:10:13, 7.84s/step, loss=1.284, acc=0.703, acc_len=1.70, thru=15.3, I=19.5, T=30.8, wait=2.9s, pool=61, epoch=2/2][2026-10-19 11:33:20,685] loop.py:45 INFO TRAIN_STEP step=10134 loss=1.346248 Training: 56%|█████▌ | 10134/18018 [22:03:06<17:07:31, 7.82s/step, loss=1.346, acc=0.698, acc_len=1.67, thru=16.4, I=15.7, T=30.4, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:33:27,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:33:28,488] loop.py:380 INFO TIMING step=10135: step=4.712s data=2.254s compute=4.680s [fwd=3.291s bwd=1.375s opt=0.003s] dispatch=3.081s [2026-10-19 11:33:28,488] loop.py:45 INFO TRAIN_STEP step=10135 loss=1.348369 Training: 56%|█████▌ | 10135/18018 [22:03:14<17:06:46, 7.82s/step, loss=1.348, acc=0.691, acc_len=1.65, thru=17.0, I=18.1, T=27.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:33:28,900] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:33:34,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:33:35,065] loop.py:45 INFO TRAIN_STEP step=10136 loss=1.230585 Training: 56%|█████▋ | 10136/18018 [22:03:21<16:17:47, 7.44s/step, loss=1.231, acc=0.710, acc_len=1.72, thru=15.8, I=19.9, T=33.8, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:33:35,060] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10135: forward=2620.8ms backward=1123.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:33:35,117] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:33:42,425] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:33:43,070] loop.py:45 INFO TRAIN_STEP step=10137 loss=1.265928 Training: 56%|█████▋ | 10137/18018 [22:03:29<16:39:51, 7.61s/step, loss=1.266, acc=0.709, acc_len=1.72, thru=16.3, I=18.1, T=28.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:33:43,431] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:33:51,960] loop.py:45 INFO TRAIN_STEP step=10138 loss=1.318295 Training: 56%|█████▋ | 10138/18018 [22:03:38<17:30:03, 8.00s/step, loss=1.318, acc=0.695, acc_len=1.67, thru=13.3, I=14.2, T=33.2, wait=5.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:33:59,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:34:00,353] loop.py:45 INFO TRAIN_STEP step=10139 loss=1.442203 Training: 56%|█████▋ | 10139/18018 [22:03:46<17:45:28, 8.11s/step, loss=1.442, acc=0.680, acc_len=1.62, thru=16.9, I=17.3, T=33.6, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:34:00,552] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:34:08,450] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:34:09,114] loop.py:380 INFO TIMING step=10140: step=4.124s data=1.830s compute=4.096s [fwd=2.925s bwd=1.156s opt=0.003s] dispatch=4.630s [2026-10-19 11:34:09,114] loop.py:45 INFO TRAIN_STEP step=10140 loss=1.388643 Training: 56%|█████▋ | 10140/18018 [22:03:55<18:10:57, 8.31s/step, loss=1.389, acc=0.682, acc_len=1.62, thru=14.9, I=15.4, T=31.0, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:34:09,458] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:34:16,549] loop.py:45 INFO TRAIN_STEP step=10141 loss=1.285833 Training: 56%|█████▋ | 10141/18018 [22:04:02<17:36:18, 8.05s/step, loss=1.286, acc=0.704, acc_len=1.71, thru=14.4, I=16.0, T=32.7, wait=3.5s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:34:16,545] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10140: forward=2797.0ms backward=1079.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:34:23,581] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:34:24,843] loop.py:45 INFO TRAIN_STEP step=10142 loss=1.488677 Training: 56%|█████▋ | 10142/18018 [22:04:10<17:46:01, 8.12s/step, loss=1.489, acc=0.666, acc_len=1.56, thru=17.5, I=17.6, T=32.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:34:25,095] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:34:31,900] loop.py:45 INFO TRAIN_STEP step=10143 loss=1.246161 Training: 56%|█████▋ | 10143/18018 [22:04:17<17:04:02, 7.80s/step, loss=1.246, acc=0.709, acc_len=1.72, thru=15.2, I=17.6, T=30.9, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:34:40,688] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:34:40,945] loop.py:45 INFO TRAIN_STEP step=10144 loss=1.461835 Training: 56%|█████▋ | 10144/18018 [22:04:27<17:52:43, 8.17s/step, loss=1.462, acc=0.674, acc_len=1.59, thru=16.8, I=15.3, T=34.5, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:34:41,190] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:34:47,976] loop.py:380 INFO TIMING step=10145: step=4.308s data=2.030s compute=4.271s [fwd=3.038s bwd=1.218s opt=0.003s] dispatch=2.715s [2026-10-19 11:34:47,977] loop.py:45 INFO TRAIN_STEP step=10145 loss=1.326300 Training: 56%|█████▋ | 10145/18018 [22:04:34<17:07:42, 7.83s/step, loss=1.326, acc=0.699, acc_len=1.68, thru=17.9, I=19.3, T=29.7, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:34:59,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:35:00,947] loop.py:45 INFO TRAIN_STEP step=10146 loss=1.368637 Training: 56%|█████▋ | 10146/18018 [22:04:47<20:29:43, 9.37s/step, loss=1.369, acc=0.690, acc_len=1.65, thru=11.3, I=11.6, T=31.6, wait=8.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:35:00,943] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10145: forward=2851.1ms backward=1128.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:35:01,108] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:35:06,306] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:35:07,519] loop.py:45 INFO TRAIN_STEP step=10147 loss=1.292660 Training: 56%|█████▋ | 10147/18018 [22:04:53<18:39:25, 8.53s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=18.5, I=23.6, T=29.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:35:07,817] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:35:15,278] loop.py:45 INFO TRAIN_STEP step=10148 loss=1.289589 Training: 56%|█████▋ | 10148/18018 [22:05:01<18:08:45, 8.30s/step, loss=1.290, acc=0.704, acc_len=1.69, thru=15.8, I=18.7, T=30.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:35:15,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:35:15,784] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:35:23,291] loop.py:45 INFO TRAIN_STEP step=10149 loss=1.233234 Training: 56%|█████▋ | 10149/18018 [22:05:09<17:57:17, 8.21s/step, loss=1.233, acc=0.715, acc_len=1.73, thru=14.1, I=14.2, T=30.9, wait=3.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:35:29,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:35:30,716] loop.py:380 INFO TIMING step=10150: step=4.294s data=1.946s compute=4.240s [fwd=3.005s bwd=1.220s opt=0.003s] dispatch=3.122s [2026-10-19 11:35:30,716] loop.py:45 INFO TRAIN_STEP step=10150 loss=1.381715 Training: 56%|█████▋ | 10150/18018 [22:05:16<17:26:05, 7.98s/step, loss=1.382, acc=0.686, acc_len=1.63, thru=18.3, I=18.1, T=29.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:35:30,794] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 11:35:36,702] loop.py:45 INFO TRAIN_STEP step=10151 loss=1.192942 Training: 56%|█████▋ | 10151/18018 [22:05:22<16:07:36, 7.38s/step, loss=1.193, acc=0.721, acc_len=1.76, thru=18.0, I=20.8, T=34.7, wait=2.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:35:36,697] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10150: forward=2626.6ms backward=1005.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:35:43,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:35:44,369] loop.py:45 INFO TRAIN_STEP step=10152 loss=1.452166 Training: 56%|█████▋ | 10152/18018 [22:05:30<16:18:44, 7.47s/step, loss=1.452, acc=0.676, acc_len=1.60, thru=17.1, I=18.5, T=30.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:35:44,869] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:35:51,027] loop.py:45 INFO TRAIN_STEP step=10153 loss=1.342060 Training: 56%|█████▋ | 10153/18018 [22:05:37<15:47:01, 7.22s/step, loss=1.342, acc=0.698, acc_len=1.67, thru=14.8, I=19.8, T=32.9, wait=2.8s, pool=45, epoch=2/2][2026-10-19 11:35:58,366] loop.py:45 INFO TRAIN_STEP step=10154 loss=1.293389 Training: 56%|█████▋ | 10154/18018 [22:05:44<15:51:19, 7.26s/step, loss=1.293, acc=0.699, acc_len=1.68, thru=14.8, I=15.9, T=31.6, wait=3.3s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:05,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:36:07,640] loop.py:380 INFO TIMING step=10155: step=4.072s data=1.788s compute=4.034s [fwd=2.931s bwd=1.089s opt=0.003s] dispatch=5.191s [2026-10-19 11:36:07,640] loop.py:45 INFO TRAIN_STEP step=10155 loss=1.456071 Training: 56%|█████▋ | 10155/18018 [22:05:53<17:10:29, 7.86s/step, loss=1.456, acc=0.676, acc_len=1.59, thru=17.7, I=17.6, T=31.4, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:07,950] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:13,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:36:14,789] loop.py:45 INFO TRAIN_STEP step=10156 loss=1.237083 Training: 56%|█████▋ | 10156/18018 [22:06:00<16:42:17, 7.65s/step, loss=1.237, acc=0.711, acc_len=1.72, thru=15.3, I=21.3, T=29.2, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:36:14,784] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10155: forward=3180.3ms backward=1149.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:14,967] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:20,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:36:20,781] loop.py:45 INFO TRAIN_STEP step=10157 loss=1.364421 Training: 56%|█████▋ | 10157/18018 [22:06:06<15:36:54, 7.15s/step, loss=1.364, acc=0.688, acc_len=1.64, thru=18.4, I=21.7, T=34.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:21,271] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:36:28,378] loop.py:45 INFO TRAIN_STEP step=10158 loss=1.359966 Training: 56%|█████▋ | 10158/18018 [22:06:14<15:54:25, 7.29s/step, loss=1.360, acc=0.692, acc_len=1.66, thru=14.9, I=17.0, T=33.6, wait=3.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:37,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:36:37,475] loop.py:45 INFO TRAIN_STEP step=10159 loss=1.348947 Training: 56%|█████▋ | 10159/18018 [22:06:23<17:05:27, 7.83s/step, loss=1.349, acc=0.688, acc_len=1.64, thru=15.3, I=15.1, T=30.5, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:37,794] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:44,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:36:44,627] loop.py:380 INFO TIMING step=10160: step=3.699s data=1.543s compute=3.651s [fwd=2.557s bwd=1.080s opt=0.003s] dispatch=3.444s [2026-10-19 11:36:44,628] loop.py:45 INFO TRAIN_STEP step=10160 loss=1.276740 Training: 56%|█████▋ | 10160/18018 [22:06:30<16:38:44, 7.63s/step, loss=1.277, acc=0.705, acc_len=1.70, thru=17.6, I=17.5, T=34.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:36:44,787] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 11:36:50,833] loop.py:45 INFO TRAIN_STEP step=10161 loss=1.277679 Training: 56%|█████▋ | 10161/18018 [22:06:36<15:42:49, 7.20s/step, loss=1.278, acc=0.704, acc_len=1.70, thru=18.0, I=18.8, T=36.7, wait=2.7s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:36:50,828] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10160: forward=2518.2ms backward=920.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:01,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:37:01,954] loop.py:45 INFO TRAIN_STEP step=10162 loss=1.503716 Training: 56%|█████▋ | 10162/18018 [22:06:48<18:16:44, 8.38s/step, loss=1.504, acc=0.666, acc_len=1.57, thru=13.7, I=13.3, T=32.7, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:02,025] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:09,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:37:10,941] loop.py:45 INFO TRAIN_STEP step=10163 loss=1.425990 Training: 56%|█████▋ | 10163/18018 [22:06:57<18:40:34, 8.56s/step, loss=1.426, acc=0.681, acc_len=1.62, thru=14.3, I=15.6, T=28.5, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:10,977] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 11:37:17,817] loop.py:45 INFO TRAIN_STEP step=10164 loss=1.293283 Training: 56%|█████▋ | 10164/18018 [22:07:03<17:34:15, 8.05s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=14.5, I=18.5, T=32.6, wait=2.9s, pool=25, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 11:37:17,923] eagle_store.py:517 WARNING batch_get_buffer missing keys (k325243-a36e5b04_hs, k325243-a36e5b04_ids, k325243-a36e5b04_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:26,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 11:37:17,893] eagle_store.py:517 WARNING batch_get_buffer missing keys (k325242-87da1fae_hs, k325242-87da1fae_ids, k325242-87da1fae_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 11:37:27,538] loop.py:380 INFO TIMING step=10165: step=4.163s data=1.618s compute=4.118s [fwd=2.948s bwd=1.153s opt=0.003s] dispatch=5.549s [2026-10-19 11:37:27,538] loop.py:45 INFO TRAIN_STEP step=10165 loss=1.295151 Training: 56%|█████▋ | 10165/18018 [22:07:13<18:39:39, 8.55s/step, loss=1.295, acc=0.705, acc_len=1.70, thru=17.5, I=14.5, T=30.7, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:27,703] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:32,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 11:37:33,532] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10165: forward=2634.9ms backward=1042.7ms [2026-10-19 11:37:33,536] loop.py:45 INFO TRAIN_STEP step=10166 loss=1.228353 Training: 56%|█████▋ | 10166/18018 [22:07:19<16:59:09, 7.79s/step, loss=1.228, acc=0.712, acc_len=1.72, thru=19.1, I=24.1, T=34.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:33,663] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 11:37:43,288] loop.py:45 INFO TRAIN_STEP step=10167 loss=1.389732 Training: 56%|█████▋ | 10167/18018 [22:07:29<18:15:59, 8.38s/step, loss=1.390, acc=0.686, acc_len=1.63, thru=13.3, I=12.2, T=31.6, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:43,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:37:43,791] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:37:50,746] loop.py:45 INFO TRAIN_STEP step=10168 loss=1.305137 Training: 56%|█████▋ | 10168/18018 [22:07:36<17:39:50, 8.10s/step, loss=1.305, acc=0.702, acc_len=1.69, thru=17.3, I=16.4, T=29.6, wait=3.1s, pool=61, epoch=2/2][2026-10-19 11:37:58,729] loop.py:45 INFO TRAIN_STEP step=10169 loss=1.390980 Training: 56%|█████▋ | 10169/18018 [22:07:44<17:35:10, 8.07s/step, loss=1.391, acc=0.681, acc_len=1.62, thru=15.3, I=14.0, T=29.8, wait=3.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:38:09,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:38:12,137] loop.py:380 INFO TIMING step=10170: step=5.478s data=2.600s compute=5.450s [fwd=3.661s bwd=1.773s opt=0.003s] dispatch=7.919s [2026-10-19 11:38:12,138] loop.py:45 INFO TRAIN_STEP step=10170 loss=1.354715 Training: 56%|█████▋ | 10170/18018 [22:07:58<21:04:36, 9.67s/step, loss=1.355, acc=0.695, acc_len=1.66, thru=10.8, I=13.2, T=23.4, wait=7.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:38:12,235] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=33/64 (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 11:38:17,834] eagle_store.py:517 WARNING batch_get_buffer missing keys (k325470-86d066a4_hs, k325470-86d066a4_ids, k325470-86d066a4_lhs); sleeping 0.50s. [2026-10-19 11:38:18,987] loop.py:45 INFO TRAIN_STEP step=10171 loss=1.367404 Training: 56%|█████▋ | 10171/18018 [22:08:05<19:13:55, 8.82s/step, loss=1.367, acc=0.689, acc_len=1.64, thru=12.5, I=19.9, T=34.9, wait=3.2s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:38:18,982] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10170: forward=2570.5ms backward=1059.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:38:26,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 11:38:17,834] eagle_store.py:517 WARNING batch_get_buffer missing keys (k325469-b7fb2a00_hs, k325469-b7fb2a00_ids, k325469-b7fb2a00_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 11:38:28,697] loop.py:45 INFO TRAIN_STEP step=10172 loss=1.339953 Training: 56%|█████▋ | 10172/18018 [22:08:14<19:48:27, 9.09s/step, loss=1.340, acc=0.696, acc_len=1.67, thru=16.3, I=17.6, T=24.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:38:28,769] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:38:33,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:38:35,574] loop.py:45 INFO TRAIN_STEP step=10173 loss=1.300348 Training: 56%|█████▋ | 10173/18018 [22:08:21<18:21:41, 8.43s/step, loss=1.300, acc=0.698, acc_len=1.68, thru=14.1, I=23.4, T=28.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:38:35,949] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:38:42,760] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:38:43,860] loop.py:45 INFO TRAIN_STEP step=10174 loss=1.381568 Training: 56%|█████▋ | 10174/18018 [22:08:29<18:15:58, 8.38s/step, loss=1.382, acc=0.686, acc_len=1.63, thru=13.3, I=19.4, T=29.7, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:38:44,267] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:38:52,469] loop.py:380 INFO TIMING step=10175: step=4.516s data=2.128s compute=4.485s [fwd=3.168s bwd=1.302s opt=0.003s] dispatch=4.085s [2026-10-19 11:38:52,470] loop.py:45 INFO TRAIN_STEP step=10175 loss=1.290323 Training: 56%|█████▋ | 10175/18018 [22:08:38<18:24:46, 8.45s/step, loss=1.290, acc=0.701, acc_len=1.67, thru=11.8, I=15.0, T=28.3, wait=4.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:39:01,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:39:04,483] loop.py:45 INFO TRAIN_STEP step=10176 loss=1.296890 Training: 56%|█████▋ | 10176/18018 [22:08:50<20:44:13, 9.52s/step, loss=1.297, acc=0.702, acc_len=1.69, thru=13.6, I=16.0, T=23.8, wait=6.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:39:04,479] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10175: forward=3831.5ms backward=1496.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:39:04,788] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 [2026-10-19 11:39:11,927] loop.py:45 INFO TRAIN_STEP step=10177 loss=1.252378 Training: 56%|█████▋ | 10177/18018 [22:08:58<19:22:42, 8.90s/step, loss=1.252, acc=0.708, acc_len=1.71, thru=16.8, I=18.4, T=28.3, wait=2.9s, pool=57, epoch=2/2][2026-10-19 11:39:19,783] loop.py:45 INFO TRAIN_STEP step=10178 loss=1.354133 Training: 56%|█████▋ | 10178/18018 [22:09:05<18:41:47, 8.59s/step, loss=1.354, acc=0.691, acc_len=1.65, thru=14.0, I=15.8, T=26.0, wait=2.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:39:27,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:39:28,676] loop.py:45 INFO TRAIN_STEP step=10179 loss=1.280669 Training: 56%|█████▋ | 10179/18018 [22:09:14<18:53:43, 8.68s/step, loss=1.281, acc=0.707, acc_len=1.70, thru=17.4, I=14.7, T=27.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:39:29,023] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:39:36,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:39:37,076] loop.py:380 INFO TIMING step=10180: step=4.808s data=2.438s compute=4.774s [fwd=3.515s bwd=1.244s opt=0.003s] dispatch=3.582s [2026-10-19 11:39:37,076] loop.py:45 INFO TRAIN_STEP step=10180 loss=1.329679 Training: 56%|█████▋ | 10180/18018 [22:09:23<18:42:41, 8.59s/step, loss=1.330, acc=0.696, acc_len=1.67, thru=13.8, I=15.6, T=26.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:39:37,155] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 11:39:47,182] loop.py:45 INFO TRAIN_STEP step=10181 loss=1.451699 Training: 57%|█████▋ | 10181/18018 [22:09:33<19:41:40, 9.05s/step, loss=1.452, acc=0.673, acc_len=1.59, thru=10.8, I=10.9, T=29.9, wait=5.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:39:47,178] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10180: forward=3089.0ms backward=1144.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:39:56,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:39:56,619] loop.py:45 INFO TRAIN_STEP step=10182 loss=1.323052 Training: 57%|█████▋ | 10182/18018 [22:09:42<19:56:51, 9.16s/step, loss=1.323, acc=0.700, acc_len=1.68, thru=15.7, I=15.6, T=28.8, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:39:56,731] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 11:40:02,840] loop.py:45 INFO TRAIN_STEP step=10183 loss=1.409029 Training: 57%|█████▋ | 10183/18018 [22:09:48<18:01:30, 8.28s/step, loss=1.409, acc=0.681, acc_len=1.61, thru=16.9, I=18.3, T=33.2, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:11,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:40:12,062] loop.py:45 INFO TRAIN_STEP step=10184 loss=1.181092 Training: 57%|█████▋ | 10184/18018 [22:09:58<18:38:00, 8.56s/step, loss=1.181, acc=0.725, acc_len=1.77, thru=14.7, I=13.8, T=28.7, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:12,233] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:19,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:40:20,088] loop.py:380 INFO TIMING step=10185: step=4.897s data=2.329s compute=4.857s [fwd=3.377s bwd=1.465s opt=0.003s] dispatch=3.123s [2026-10-19 11:40:20,088] loop.py:45 INFO TRAIN_STEP step=10185 loss=1.475461 Training: 57%|█████▋ | 10185/18018 [22:10:06<18:16:58, 8.40s/step, loss=1.475, acc=0.673, acc_len=1.58, thru=16.5, I=17.6, T=26.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:20,144] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 11:40:27,666] loop.py:45 INFO TRAIN_STEP step=10186 loss=1.247487 Training: 57%|█████▋ | 10186/18018 [22:10:13<17:44:25, 8.15s/step, loss=1.247, acc=0.710, acc_len=1.72, thru=16.9, I=16.9, T=28.9, wait=3.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:40:27,662] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10185: forward=3090.0ms backward=1281.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:35,955] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:40:36,317] loop.py:45 INFO TRAIN_STEP step=10187 loss=1.193359 Training: 57%|█████▋ | 10187/18018 [22:10:22<18:03:51, 8.30s/step, loss=1.193, acc=0.720, acc_len=1.76, thru=15.0, I=15.6, T=29.5, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:36,466] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:42,198] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:40:44,236] loop.py:45 INFO TRAIN_STEP step=10188 loss=1.318692 Training: 57%|█████▋ | 10188/18018 [22:10:30<17:48:37, 8.19s/step, loss=1.319, acc=0.699, acc_len=1.68, thru=14.7, I=18.2, T=27.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:44,716] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:49,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:40:50,509] loop.py:45 INFO TRAIN_STEP step=10189 loss=1.375165 Training: 57%|█████▋ | 10189/18018 [22:10:36<16:33:29, 7.61s/step, loss=1.375, acc=0.690, acc_len=1.64, thru=17.4, I=26.4, T=34.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:50,838] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:56,773] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:40:56,863] loop.py:380 INFO TIMING step=10190: step=3.827s data=1.637s compute=3.798s [fwd=2.701s bwd=1.083s opt=0.003s] dispatch=2.517s [2026-10-19 11:40:56,864] loop.py:45 INFO TRAIN_STEP step=10190 loss=1.304163 Training: 57%|█████▋ | 10190/18018 [22:10:42<15:43:57, 7.24s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=19.4, I=23.0, T=33.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:40:57,274] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:41:04,022] loop.py:45 INFO TRAIN_STEP step=10191 loss=1.383192 Training: 57%|█████▋ | 10191/18018 [22:10:50<15:40:52, 7.21s/step, loss=1.383, acc=0.685, acc_len=1.62, thru=17.2, I=18.0, T=35.9, wait=3.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:41:04,017] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10190: forward=2554.2ms backward=962.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:09,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:41:11,143] loop.py:45 INFO TRAIN_STEP step=10192 loss=1.410753 Training: 57%|█████▋ | 10192/18018 [22:10:57<15:37:14, 7.19s/step, loss=1.411, acc=0.682, acc_len=1.62, thru=18.7, I=20.0, T=33.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:11,441] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:16,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:41:18,632] loop.py:45 INFO TRAIN_STEP step=10193 loss=1.301649 Training: 57%|█████▋ | 10193/18018 [22:11:04<15:48:52, 7.28s/step, loss=1.302, acc=0.699, acc_len=1.69, thru=16.4, I=24.5, T=28.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:18,660] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:23,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:41:25,254] loop.py:45 INFO TRAIN_STEP step=10194 loss=1.323999 Training: 57%|█████▋ | 10194/18018 [22:11:11<15:23:17, 7.08s/step, loss=1.324, acc=0.695, acc_len=1.67, thru=16.1, I=26.0, T=32.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:25,388] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:30,377] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:41:32,224] loop.py:380 INFO TIMING step=10195: step=4.381s data=1.999s compute=4.340s [fwd=2.950s bwd=1.374s opt=0.003s] dispatch=2.578s [2026-10-19 11:41:32,224] loop.py:45 INFO TRAIN_STEP step=10195 loss=1.446961 Training: 57%|█████▋ | 10195/18018 [22:11:18<15:18:47, 7.05s/step, loss=1.447, acc=0.678, acc_len=1.60, thru=18.4, I=26.5, T=29.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:32,386] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:37,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:41:38,143] loop.py:45 INFO TRAIN_STEP step=10196 loss=1.268487 Training: 57%|█████▋ | 10196/18018 [22:11:24<14:34:38, 6.71s/step, loss=1.268, acc=0.704, acc_len=1.69, thru=18.4, I=23.7, T=35.7, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:41:38,139] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10195: forward=2557.4ms backward=980.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:38,219] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:45,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:41:45,811] loop.py:45 INFO TRAIN_STEP step=10197 loss=1.377538 Training: 57%|█████▋ | 10197/18018 [22:11:31<15:11:54, 7.00s/step, loss=1.378, acc=0.688, acc_len=1.64, thru=16.7, I=17.4, T=30.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:45,929] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:54,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:41:55,125] loop.py:45 INFO TRAIN_STEP step=10198 loss=1.439245 Training: 57%|█████▋ | 10198/18018 [22:11:41<16:42:33, 7.69s/step, loss=1.439, acc=0.680, acc_len=1.62, thru=14.1, I=15.8, T=31.5, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:41:55,442] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:00,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:42:01,640] loop.py:45 INFO TRAIN_STEP step=10199 loss=1.306708 Training: 57%|█████▋ | 10199/18018 [22:11:47<15:56:25, 7.34s/step, loss=1.307, acc=0.698, acc_len=1.67, thru=18.1, I=22.4, T=35.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:01,789] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:06,915] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:42:08,329] loop.py:380 INFO TIMING step=10200: step=4.097s data=1.712s compute=4.062s [fwd=2.777s bwd=1.269s opt=0.003s] dispatch=2.580s [2026-10-19 11:42:08,329] loop.py:45 INFO TRAIN_STEP step=10200 loss=1.299669 Training: 57%|█████▋ | 10200/18018 [22:11:54<15:30:50, 7.14s/step, loss=1.300, acc=0.701, acc_len=1.69, thru=19.0, I=23.0, T=31.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:08,430] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:16,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:42:17,238] loop.py:45 INFO TRAIN_STEP step=10201 loss=1.317718 Training: 57%|█████▋ | 10201/18018 [22:12:03<16:39:43, 7.67s/step, loss=1.318, acc=0.699, acc_len=1.69, thru=14.6, I=16.1, T=30.4, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:42:17,234] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10200: forward=2909.6ms backward=1239.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:17,400] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:25,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:42:27,300] loop.py:45 INFO TRAIN_STEP step=10202 loss=1.526420 Training: 57%|█████▋ | 10202/18018 [22:12:13<18:12:50, 8.39s/step, loss=1.526, acc=0.666, acc_len=1.57, thru=12.9, I=14.8, T=27.4, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:27,780] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:34,498] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:42:35,594] loop.py:45 INFO TRAIN_STEP step=10203 loss=1.399404 Training: 57%|█████▋ | 10203/18018 [22:12:21<18:09:05, 8.36s/step, loss=1.399, acc=0.683, acc_len=1.62, thru=13.7, I=19.2, T=30.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:36,011] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:40,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:42:41,906] loop.py:45 INFO TRAIN_STEP step=10204 loss=1.253982 Training: 57%|█████▋ | 10204/18018 [22:12:27<16:48:48, 7.75s/step, loss=1.254, acc=0.709, acc_len=1.71, thru=18.0, I=22.4, T=32.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:42,244] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:46,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:42:48,567] loop.py:380 INFO TIMING step=10205: step=4.120s data=1.958s compute=4.074s [fwd=2.876s bwd=1.184s opt=0.003s] dispatch=2.532s [2026-10-19 11:42:48,567] loop.py:45 INFO TRAIN_STEP step=10205 loss=1.291870 Training: 57%|█████▋ | 10205/18018 [22:12:34<16:06:17, 7.42s/step, loss=1.292, acc=0.701, acc_len=1.69, thru=17.8, I=22.8, T=31.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:48,771] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:55,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:42:55,655] loop.py:45 INFO TRAIN_STEP step=10206 loss=1.316373 Training: 57%|█████▋ | 10206/18018 [22:12:41<15:53:12, 7.32s/step, loss=1.316, acc=0.698, acc_len=1.68, thru=15.8, I=19.2, T=32.3, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:42:55,650] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10205: forward=2766.1ms backward=1143.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:42:55,734] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:02,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:43:03,379] loop.py:45 INFO TRAIN_STEP step=10207 loss=1.328418 Training: 57%|█████▋ | 10207/18018 [22:12:49<16:08:47, 7.44s/step, loss=1.328, acc=0.694, acc_len=1.66, thru=17.8, I=19.1, T=32.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:03,847] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:43:10,900] loop.py:45 INFO TRAIN_STEP step=10208 loss=1.409811 Training: 57%|█████▋ | 10208/18018 [22:12:56<16:11:49, 7.47s/step, loss=1.410, acc=0.679, acc_len=1.61, thru=13.7, I=18.4, T=34.1, wait=3.8s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:20,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:43:22,761] loop.py:45 INFO TRAIN_STEP step=10209 loss=1.305618 Training: 57%|█████▋ | 10209/18018 [22:13:08<19:03:16, 8.78s/step, loss=1.306, acc=0.696, acc_len=1.67, thru=15.4, I=16.3, T=26.6, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:23,255] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 [2026-10-19 11:43:29,389] loop.py:380 INFO TIMING step=10210: step=3.739s data=1.609s compute=3.713s [fwd=2.677s bwd=1.020s opt=0.003s] dispatch=2.879s [2026-10-19 11:43:29,389] loop.py:45 INFO TRAIN_STEP step=10210 loss=1.377888 Training: 57%|█████▋ | 10210/18018 [22:13:15<17:38:51, 8.14s/step, loss=1.378, acc=0.685, acc_len=1.63, thru=12.8, I=17.1, T=34.2, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:36,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:43:37,416] loop.py:45 INFO TRAIN_STEP step=10211 loss=1.309889 Training: 57%|█████▋ | 10211/18018 [22:13:23<17:34:28, 8.10s/step, loss=1.310, acc=0.698, acc_len=1.67, thru=18.0, I=18.5, T=29.5, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:43:37,412] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10210: forward=2959.5ms backward=1334.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:37,680] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:43,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:43:44,781] loop.py:45 INFO TRAIN_STEP step=10212 loss=1.412670 Training: 57%|█████▋ | 10212/18018 [22:13:30<17:05:28, 7.88s/step, loss=1.413, acc=0.680, acc_len=1.61, thru=15.8, I=20.4, T=31.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:45,008] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:51,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:43:52,125] loop.py:45 INFO TRAIN_STEP step=10213 loss=1.355771 Training: 57%|█████▋ | 10213/18018 [22:13:38<16:44:20, 7.72s/step, loss=1.356, acc=0.693, acc_len=1.66, thru=15.8, I=18.3, T=29.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:43:52,550] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:00,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:44:02,228] loop.py:45 INFO TRAIN_STEP step=10214 loss=1.384750 Training: 57%|█████▋ | 10214/18018 [22:13:48<18:17:11, 8.44s/step, loss=1.385, acc=0.686, acc_len=1.63, thru=13.2, I=14.5, T=29.3, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:02,481] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:07,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:44:08,841] loop.py:380 INFO TIMING step=10215: step=3.928s data=1.673s compute=3.884s [fwd=2.703s bwd=1.165s opt=0.003s] dispatch=2.676s [2026-10-19 11:44:08,841] loop.py:45 INFO TRAIN_STEP step=10215 loss=1.311345 Training: 57%|█████▋ | 10215/18018 [22:13:54<17:05:56, 7.89s/step, loss=1.311, acc=0.697, acc_len=1.67, thru=17.4, I=21.4, T=32.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:08,955] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 11:44:15,737] loop.py:45 INFO TRAIN_STEP step=10216 loss=1.362438 Training: 57%|█████▋ | 10216/18018 [22:14:01<16:27:03, 7.59s/step, loss=1.362, acc=0.690, acc_len=1.64, thru=14.3, I=19.4, T=32.5, wait=2.9s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:44:15,732] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10215: forward=2814.3ms backward=1087.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:24,915] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:44:25,873] loop.py:45 INFO TRAIN_STEP step=10217 loss=1.426215 Training: 57%|█████▋ | 10217/18018 [22:14:11<18:06:19, 8.36s/step, loss=1.426, acc=0.682, acc_len=1.62, thru=17.3, I=14.4, T=27.9, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:25,927] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 11:44:32,939] loop.py:45 INFO TRAIN_STEP step=10218 loss=1.214804 Training: 57%|█████▋ | 10218/18018 [22:14:19<17:15:48, 7.97s/step, loss=1.215, acc=0.716, acc_len=1.73, thru=16.2, I=19.4, T=28.0, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:40,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:44:40,487] loop.py:45 INFO TRAIN_STEP step=10219 loss=1.225515 Training: 57%|█████▋ | 10219/18018 [22:14:26<16:59:20, 7.84s/step, loss=1.226, acc=0.712, acc_len=1.72, thru=16.8, I=17.6, T=30.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:40,735] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:47,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:44:48,962] loop.py:380 INFO TIMING step=10220: step=4.422s data=2.014s compute=4.389s [fwd=3.121s bwd=1.252s opt=0.003s] dispatch=4.043s [2026-10-19 11:44:48,963] loop.py:45 INFO TRAIN_STEP step=10220 loss=1.393109 Training: 57%|█████▋ | 10220/18018 [22:14:35<17:23:56, 8.03s/step, loss=1.393, acc=0.686, acc_len=1.63, thru=14.8, I=17.0, T=28.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:49,105] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:55,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:44:56,491] loop.py:45 INFO TRAIN_STEP step=10221 loss=1.352360 Training: 57%|█████▋ | 10221/18018 [22:14:42<17:04:09, 7.88s/step, loss=1.352, acc=0.689, acc_len=1.65, thru=15.3, I=19.6, T=27.6, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:44:56,486] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10220: forward=3207.5ms backward=1390.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:44:56,523] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:01,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:45:03,509] loop.py:45 INFO TRAIN_STEP step=10222 loss=1.378242 Training: 57%|█████▋ | 10222/18018 [22:14:49<16:30:17, 7.62s/step, loss=1.378, acc=0.691, acc_len=1.65, thru=16.7, I=21.2, T=33.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:03,810] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:45:10,998] loop.py:45 INFO TRAIN_STEP step=10223 loss=1.292950 Training: 57%|█████▋ | 10223/18018 [22:14:57<16:25:04, 7.58s/step, loss=1.293, acc=0.701, acc_len=1.69, thru=15.0, I=21.3, T=29.8, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:17,323] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:45:18,568] loop.py:45 INFO TRAIN_STEP step=10224 loss=1.400978 Training: 57%|█████▋ | 10224/18018 [22:15:04<16:24:25, 7.58s/step, loss=1.401, acc=0.681, acc_len=1.62, thru=16.3, I=18.1, T=32.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:18,830] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:45:24,227] loop.py:380 INFO TIMING step=10225: step=3.407s data=1.360s compute=3.379s [fwd=2.400s bwd=0.965s opt=0.003s] dispatch=2.244s [2026-10-19 11:45:24,228] loop.py:45 INFO TRAIN_STEP step=10225 loss=1.289324 Training: 57%|█████▋ | 10225/18018 [22:15:10<15:09:35, 7.00s/step, loss=1.289, acc=0.701, acc_len=1.68, thru=19.2, I=23.4, T=37.6, wait=2.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:32,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:45:33,615] loop.py:45 INFO TRAIN_STEP step=10226 loss=1.262530 (TrainerActor pid=9310) [2026-10-19 11:45:33,610] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10225: forward=2928.6ms backward=1306.1ms Training: 57%|█████▋ | 10226/18018 [22:15:19<16:42:20, 7.72s/step, loss=1.263, acc=0.710, acc_len=1.71, thru=15.0, I=17.3, T=29.8, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:33,672] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:37,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:45:39,974] loop.py:45 INFO TRAIN_STEP step=10227 loss=1.239517 Training: 57%|█████▋ | 10227/18018 [22:15:26<15:49:11, 7.31s/step, loss=1.240, acc=0.711, acc_len=1.72, thru=18.2, I=27.6, T=29.3, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:40,394] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:48,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:45:49,759] loop.py:45 INFO TRAIN_STEP step=10228 loss=1.356618 Training: 57%|█████▋ | 10228/18018 [22:15:35<17:25:29, 8.05s/step, loss=1.357, acc=0.695, acc_len=1.66, thru=13.7, I=15.7, T=31.0, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:49,824] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:45:57,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:46:00,405] loop.py:45 INFO TRAIN_STEP step=10229 loss=1.359439 Training: 57%|█████▋ | 10229/18018 [22:15:46<19:06:27, 8.83s/step, loss=1.359, acc=0.692, acc_len=1.66, thru=11.2, I=15.5, T=24.9, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:00,468] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:06,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:46:08,383] loop.py:380 INFO TIMING step=10230: step=4.726s data=2.323s compute=4.680s [fwd=3.252s bwd=1.413s opt=0.003s] dispatch=3.239s [2026-10-19 11:46:08,383] loop.py:45 INFO TRAIN_STEP step=10230 loss=1.324433 Training: 57%|█████▋ | 10230/18018 [22:15:54<18:32:59, 8.57s/step, loss=1.324, acc=0.697, acc_len=1.67, thru=16.2, I=19.9, T=27.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:08,776] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:15,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:46:16,610] loop.py:45 INFO TRAIN_STEP step=10231 loss=1.394631 Training: 57%|█████▋ | 10231/18018 [22:16:02<18:19:18, 8.47s/step, loss=1.395, acc=0.681, acc_len=1.62, thru=14.2, I=19.7, T=28.9, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:46:16,605] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10230: forward=3017.8ms backward=1335.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:16,724] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:23,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:46:23,951] loop.py:45 INFO TRAIN_STEP step=10232 loss=1.306206 Training: 57%|█████▋ | 10232/18018 [22:16:10<17:35:12, 8.13s/step, loss=1.306, acc=0.701, acc_len=1.69, thru=15.7, I=20.1, T=30.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:24,017] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:31,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:46:31,486] loop.py:45 INFO TRAIN_STEP step=10233 loss=1.215378 Training: 57%|█████▋ | 10233/18018 [22:16:17<17:11:49, 7.95s/step, loss=1.215, acc=0.717, acc_len=1.74, thru=15.1, I=15.9, T=32.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:31,758] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:37,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:46:39,104] loop.py:45 INFO TRAIN_STEP step=10234 loss=1.314550 Training: 57%|█████▋ | 10234/18018 [22:16:25<16:58:44, 7.85s/step, loss=1.315, acc=0.699, acc_len=1.68, thru=18.0, I=21.7, T=26.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:39,525] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:45,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:46:46,611] loop.py:380 INFO TIMING step=10235: step=4.334s data=2.015s compute=4.306s [fwd=3.084s bwd=1.207s opt=0.003s] dispatch=3.162s [2026-10-19 11:46:46,611] loop.py:45 INFO TRAIN_STEP step=10235 loss=1.270455 Training: 57%|█████▋ | 10235/18018 [22:16:32<16:45:08, 7.75s/step, loss=1.270, acc=0.708, acc_len=1.72, thru=14.1, I=21.8, T=29.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:46,889] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:51,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:46:53,656] loop.py:45 INFO TRAIN_STEP step=10236 loss=1.269199 Training: 57%|█████▋ | 10236/18018 [22:16:39<16:17:35, 7.54s/step, loss=1.269, acc=0.705, acc_len=1.70, thru=17.0, I=23.2, T=31.6, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:46:53,650] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10235: forward=2823.6ms backward=1186.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:53,807] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:46:59,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:47:01,541] loop.py:45 INFO TRAIN_STEP step=10237 loss=1.338758 Training: 57%|█████▋ | 10237/18018 [22:16:47<16:31:03, 7.64s/step, loss=1.339, acc=0.694, acc_len=1.66, thru=13.8, I=22.2, T=29.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:47:01,666] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 [2026-10-19 11:47:08,415] loop.py:45 INFO TRAIN_STEP step=10238 loss=1.367656 Training: 57%|█████▋ | 10238/18018 [22:16:54<16:01:00, 7.41s/step, loss=1.368, acc=0.688, acc_len=1.64, thru=15.0, I=19.7, T=32.3, wait=2.9s, pool=57, epoch=2/2][2026-10-19 11:47:16,411] loop.py:45 INFO TRAIN_STEP step=10239 loss=1.339996 Training: 57%|█████▋ | 10239/18018 [22:17:02<16:23:38, 7.59s/step, loss=1.340, acc=0.693, acc_len=1.66, thru=16.4, I=14.8, T=28.0, wait=3.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:47:22,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:47:22,721] loop.py:380 INFO TIMING step=10240: step=3.928s data=1.732s compute=3.879s [fwd=2.725s bwd=1.138s opt=0.003s] dispatch=2.374s [2026-10-19 11:47:22,721] loop.py:45 INFO TRAIN_STEP step=10240 loss=1.502164 Training: 57%|█████▋ | 10240/18018 [22:17:08<15:33:50, 7.20s/step, loss=1.502, acc=0.671, acc_len=1.58, thru=19.1, I=18.8, T=32.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:47:23,061] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:47:28,615] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:47:30,413] loop.py:45 INFO TRAIN_STEP step=10241 loss=1.273671 Training: 57%|█████▋ | 10241/18018 [22:17:16<15:52:46, 7.35s/step, loss=1.274, acc=0.705, acc_len=1.70, thru=16.0, I=19.0, T=26.6, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:47:30,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10240: forward=3317.7ms backward=1433.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:47:30,625] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:47:37,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:47:37,820] loop.py:45 INFO TRAIN_STEP step=10242 loss=1.329402 Training: 57%|█████▋ | 10242/18018 [22:17:23<15:54:43, 7.37s/step, loss=1.329, acc=0.695, acc_len=1.67, thru=14.8, I=18.7, T=29.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:47:38,320] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:47:44,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:47:46,046] loop.py:45 INFO TRAIN_STEP step=10243 loss=1.422504 Training: 57%|█████▋ | 10243/18018 [22:17:32<16:28:01, 7.62s/step, loss=1.423, acc=0.683, acc_len=1.62, thru=15.1, I=18.4, T=30.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:47:46,462] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:47:54,302] loop.py:45 INFO TRAIN_STEP step=10244 loss=1.288132 Training: 57%|█████▋ | 10244/18018 [22:17:40<16:52:31, 7.81s/step, loss=1.288, acc=0.704, acc_len=1.69, thru=10.9, I=16.1, T=36.0, wait=4.7s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:05,140] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:48:05,825] loop.py:380 INFO TIMING step=10245: step=4.615s data=2.154s compute=4.579s [fwd=3.182s bwd=1.382s opt=0.003s] dispatch=6.898s [2026-10-19 11:48:05,825] loop.py:45 INFO TRAIN_STEP step=10245 loss=1.374025 Training: 57%|█████▋ | 10245/18018 [22:17:51<19:16:31, 8.93s/step, loss=1.374, acc=0.693, acc_len=1.65, thru=15.1, I=14.6, T=27.7, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:06,145] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:13,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:48:14,654] loop.py:45 INFO TRAIN_STEP step=10246 loss=1.313239 Training: 57%|█████▋ | 10246/18018 [22:18:00<19:12:33, 8.90s/step, loss=1.313, acc=0.699, acc_len=1.68, thru=13.6, I=16.1, T=24.7, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:48:14,650] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10245: forward=3465.9ms backward=1647.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:14,972] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:20,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:48:22,428] loop.py:45 INFO TRAIN_STEP step=10247 loss=1.280598 Training: 57%|█████▋ | 10247/18018 [22:18:08<18:28:43, 8.56s/step, loss=1.281, acc=0.705, acc_len=1.70, thru=14.7, I=19.5, T=32.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:22,461] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:26,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:48:27,935] loop.py:45 INFO TRAIN_STEP step=10248 loss=1.310964 Training: 57%|█████▋ | 10248/18018 [22:18:14<16:29:58, 7.64s/step, loss=1.311, acc=0.701, acc_len=1.68, thru=20.2, I=30.9, T=36.1, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:28,048] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:36,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:48:36,876] loop.py:45 INFO TRAIN_STEP step=10249 loss=1.363142 Training: 57%|█████▋ | 10249/18018 [22:18:22<17:20:10, 8.03s/step, loss=1.363, acc=0.688, acc_len=1.64, thru=14.5, I=15.8, T=30.3, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:37,209] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:43,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:48:45,560] loop.py:380 INFO TIMING step=10250: step=4.670s data=2.283s compute=4.635s [fwd=3.241s bwd=1.378s opt=0.003s] dispatch=4.003s [2026-10-19 11:48:45,560] loop.py:45 INFO TRAIN_STEP step=10250 loss=1.235830 Training: 57%|█████▋ | 10250/18018 [22:18:31<17:45:21, 8.23s/step, loss=1.236, acc=0.715, acc_len=1.74, thru=14.4, I=19.2, T=27.4, wait=4.0s, pool=65, epoch=2/2][2026-10-19 11:48:45,565] loop.py:419 INFO Saving checkpoint at step 10250... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:48,349] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 11:48:48,996] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0010251 (TrainerActor pid=9310) [2026-10-19 11:48:49,001] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:50,366] inference_manager.py:408 INFO Pool capacity available after 7.1s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:55,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 11:48:48,997] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0010251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 11:48:50,139] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009501 [2026-10-19 11:48:57,430] loop.py:45 INFO TRAIN_STEP step=10251 loss=1.222241 Training: 57%|█████▋ | 10251/18018 [22:18:43<20:06:38, 9.32s/step, loss=1.222, acc=0.712, acc_len=1.72, thru=18.1, I=22.2, T=27.7, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:48:57,425] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10250: forward=3222.7ms backward=1356.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:48:57,508] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 11:49:03,914] loop.py:45 INFO TRAIN_STEP step=10252 loss=1.290874 Training: 57%|█████▋ | 10252/18018 [22:18:49<18:16:14, 8.47s/step, loss=1.291, acc=0.701, acc_len=1.69, thru=15.6, I=20.1, T=33.6, wait=2.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:11,756] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:49:11,839] loop.py:45 INFO TRAIN_STEP step=10253 loss=1.374858 Training: 57%|█████▋ | 10253/18018 [22:18:57<17:54:54, 8.31s/step, loss=1.375, acc=0.687, acc_len=1.64, thru=16.5, I=16.5, T=31.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:12,258] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:18,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:49:19,348] loop.py:45 INFO TRAIN_STEP step=10254 loss=1.422228 Training: 57%|█████▋ | 10254/18018 [22:19:05<17:23:56, 8.07s/step, loss=1.422, acc=0.678, acc_len=1.61, thru=16.4, I=16.8, T=33.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:19,451] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:24,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:49:25,986] loop.py:380 INFO TIMING step=10255: step=4.202s data=1.899s compute=4.168s [fwd=2.953s bwd=1.201s opt=0.003s] dispatch=2.425s [2026-10-19 11:49:25,986] loop.py:45 INFO TRAIN_STEP step=10255 loss=1.352168 Training: 57%|█████▋ | 10255/18018 [22:19:12<16:28:18, 7.64s/step, loss=1.352, acc=0.691, acc_len=1.65, thru=18.8, I=21.8, T=30.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:26,482] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:34,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:49:36,641] loop.py:45 INFO TRAIN_STEP step=10256 loss=1.325571 Training: 57%|█████▋ | 10256/18018 [22:19:22<18:25:09, 8.54s/step, loss=1.326, acc=0.700, acc_len=1.68, thru=12.1, I=15.5, T=28.7, wait=6.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:49:36,636] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10255: forward=3097.3ms backward=1288.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:36,755] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:42,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:49:43,811] loop.py:45 INFO TRAIN_STEP step=10257 loss=1.295378 Training: 57%|█████▋ | 10257/18018 [22:19:29<17:31:51, 8.13s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=14.7, I=20.2, T=29.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:49:44,093] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:49:50,503] loop.py:45 INFO TRAIN_STEP step=10258 loss=1.278365 Training: 57%|█████▋ | 10258/18018 [22:19:36<16:35:45, 7.70s/step, loss=1.278, acc=0.705, acc_len=1.69, thru=17.3, I=18.8, T=34.2, wait=2.9s, pool=61, epoch=2/2][2026-10-19 11:49:57,275] loop.py:45 INFO TRAIN_STEP step=10259 loss=1.331460 Training: 57%|█████▋ | 10259/18018 [22:19:43<15:59:41, 7.42s/step, loss=1.331, acc=0.695, acc_len=1.66, thru=17.5, I=19.1, T=32.0, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:06,192] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:50:08,719] loop.py:380 INFO TIMING step=10260: step=4.557s data=2.018s compute=4.522s [fwd=3.140s bwd=1.368s opt=0.003s] dispatch=6.878s [2026-10-19 11:50:08,720] loop.py:45 INFO TRAIN_STEP step=10260 loss=1.307096 Training: 57%|█████▋ | 10260/18018 [22:19:54<18:35:34, 8.63s/step, loss=1.307, acc=0.702, acc_len=1.69, thru=12.5, I=14.7, T=28.1, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:09,218] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 [2026-10-19 11:50:15,949] loop.py:45 INFO TRAIN_STEP step=10261 loss=1.351076 Training: 57%|█████▋ | 10261/18018 [22:20:02<17:41:16, 8.21s/step, loss=1.351, acc=0.695, acc_len=1.67, thru=12.5, I=18.5, T=34.8, wait=3.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:50:15,944] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10260: forward=2563.3ms backward=1062.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:22,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:50:23,137] loop.py:45 INFO TRAIN_STEP step=10262 loss=1.315559 Training: 57%|█████▋ | 10262/18018 [22:20:09<17:01:35, 7.90s/step, loss=1.316, acc=0.699, acc_len=1.68, thru=18.6, I=18.7, T=31.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:23,175] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:30,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:50:32,930] loop.py:45 INFO TRAIN_STEP step=10263 loss=1.330553 Training: 57%|█████▋ | 10263/18018 [22:20:19<18:14:39, 8.47s/step, loss=1.331, acc=0.698, acc_len=1.68, thru=13.2, I=14.9, T=27.7, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:33,386] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 11:50:39,489] loop.py:45 INFO TRAIN_STEP step=10264 loss=1.297478 Training: 57%|█████▋ | 10264/18018 [22:20:25<17:00:23, 7.90s/step, loss=1.297, acc=0.704, acc_len=1.70, thru=16.6, I=22.1, T=34.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:39,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:39,964] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:46,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:50:47,672] loop.py:380 INFO TIMING step=10265: step=4.243s data=2.068s compute=4.214s [fwd=3.163s bwd=1.037s opt=0.003s] dispatch=3.934s [2026-10-19 11:50:47,673] loop.py:45 INFO TRAIN_STEP step=10265 loss=1.256796 Training: 57%|█████▋ | 10265/18018 [22:20:33<17:11:31, 7.98s/step, loss=1.257, acc=0.708, acc_len=1.71, thru=16.4, I=16.0, T=30.2, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:50:47,706] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 11:50:54,507] loop.py:45 INFO TRAIN_STEP step=10266 loss=1.385231 Training: 57%|█████▋ | 10266/18018 [22:20:40<16:26:52, 7.64s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=16.9, I=21.0, T=29.6, wait=2.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:50:54,502] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10265: forward=3029.4ms backward=1247.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:01,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:51:02,354] loop.py:45 INFO TRAIN_STEP step=10267 loss=1.277867 Training: 57%|█████▋ | 10267/18018 [22:20:48<16:34:46, 7.70s/step, loss=1.278, acc=0.706, acc_len=1.70, thru=16.3, I=17.2, T=30.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:02,418] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:07,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:51:08,527] loop.py:45 INFO TRAIN_STEP step=10268 loss=1.405988 Training: 57%|█████▋ | 10268/18018 [22:20:54<15:35:28, 7.24s/step, loss=1.406, acc=0.683, acc_len=1.62, thru=19.4, I=22.7, T=30.9, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:08,556] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:15,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:51:15,409] loop.py:45 INFO TRAIN_STEP step=10269 loss=1.365811 Training: 57%|█████▋ | 10269/18018 [22:21:01<15:21:26, 7.13s/step, loss=1.366, acc=0.692, acc_len=1.65, thru=17.3, I=19.8, T=35.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:15,726] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:20,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:51:22,020] loop.py:380 INFO TIMING step=10270: step=4.453s data=2.083s compute=4.420s [fwd=3.110s bwd=1.296s opt=0.003s] dispatch=2.147s [2026-10-19 11:51:22,020] loop.py:45 INFO TRAIN_STEP step=10270 loss=1.319779 Training: 57%|█████▋ | 10270/18018 [22:21:08<15:00:58, 6.98s/step, loss=1.320, acc=0.701, acc_len=1.68, thru=19.7, I=24.6, T=28.7, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:22,095] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:30,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:51:30,876] loop.py:45 INFO TRAIN_STEP step=10271 loss=1.267351 Training: 57%|█████▋ | 10271/18018 [22:21:16<16:13:37, 7.54s/step, loss=1.267, acc=0.706, acc_len=1.71, thru=14.6, I=13.1, T=32.7, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:51:30,871] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10270: forward=2819.9ms backward=1043.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:31,152] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:39,373] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:51:39,533] loop.py:45 INFO TRAIN_STEP step=10272 loss=1.354908 Training: 57%|█████▋ | 10272/18018 [22:21:25<16:56:48, 7.88s/step, loss=1.355, acc=0.690, acc_len=1.65, thru=15.6, I=17.2, T=27.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:39,875] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:50,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:51:51,225] loop.py:45 INFO TRAIN_STEP step=10273 loss=1.339272 Training: 57%|█████▋ | 10273/18018 [22:21:37<19:24:26, 9.02s/step, loss=1.339, acc=0.696, acc_len=1.67, thru=11.3, I=11.1, T=26.4, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:51,622] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:56,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:51:57,473] loop.py:45 INFO TRAIN_STEP step=10274 loss=1.287890 Training: 57%|█████▋ | 10274/18018 [22:21:43<17:36:52, 8.19s/step, loss=1.288, acc=0.703, acc_len=1.70, thru=17.3, I=20.9, T=32.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:51:57,615] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:04,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:52:05,014] loop.py:380 INFO TIMING step=10275: step=3.730s data=1.546s compute=3.697s [fwd=2.609s bwd=1.074s opt=0.003s] dispatch=3.802s [2026-10-19 11:52:05,014] loop.py:45 INFO TRAIN_STEP step=10275 loss=1.202034 Training: 57%|█████▋ | 10275/18018 [22:21:51<17:11:42, 7.99s/step, loss=1.202, acc=0.719, acc_len=1.75, thru=17.3, I=18.9, T=34.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:05,330] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:12,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:52:15,208] loop.py:45 INFO TRAIN_STEP step=10276 loss=1.244071 Training: 57%|█████▋ | 10276/18018 [22:22:01<18:36:44, 8.65s/step, loss=1.244, acc=0.712, acc_len=1.73, thru=13.0, I=16.5, T=25.6, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:52:15,204] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10275: forward=3403.6ms backward=1518.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:15,612] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:22,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:52:22,715] loop.py:45 INFO TRAIN_STEP step=10277 loss=1.290872 Training: 57%|█████▋ | 10277/18018 [22:22:08<17:52:11, 8.31s/step, loss=1.291, acc=0.701, acc_len=1.69, thru=18.0, I=20.4, T=28.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:22,788] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:29,008] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:52:30,671] loop.py:45 INFO TRAIN_STEP step=10278 loss=1.433371 Training: 57%|█████▋ | 10278/18018 [22:22:16<17:38:16, 8.20s/step, loss=1.433, acc=0.680, acc_len=1.62, thru=14.2, I=16.5, T=25.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:31,022] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:38,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:52:39,303] loop.py:45 INFO TRAIN_STEP step=10279 loss=1.374368 Training: 57%|█████▋ | 10279/18018 [22:22:25<17:54:41, 8.33s/step, loss=1.374, acc=0.689, acc_len=1.65, thru=15.5, I=16.2, T=28.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:39,622] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:52:47,160] loop.py:380 INFO TIMING step=10280: step=4.514s data=2.144s compute=4.086s [fwd=2.925s bwd=1.146s opt=0.003s] dispatch=3.226s [2026-10-19 11:52:47,161] loop.py:45 INFO TRAIN_STEP step=10280 loss=1.420310 Training: 57%|█████▋ | 10280/18018 [22:22:35<18:52:48, 8.78s/step, loss=1.420, acc=0.681, acc_len=1.61, thru=9.7, I=14.7, T=28.4, wait=3.2s, pool=29, epoch=2/2] [2026-10-19 11:52:57,307] loop.py:45 INFO TRAIN_STEP step=10281 loss=1.383610 Training: 57%|█████▋ | 10281/18018 [22:22:43<18:28:44, 8.60s/step, loss=1.384, acc=0.685, acc_len=1.63, thru=20.6, I=18.8, T=30.0, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:52:57,302] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10280: forward=3010.8ms backward=1221.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:57,307] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:52:57,810] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:06,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:53:06,563] loop.py:45 INFO TRAIN_STEP step=10282 loss=1.339751 Training: 57%|█████▋ | 10282/18018 [22:22:52<18:54:04, 8.80s/step, loss=1.340, acc=0.695, acc_len=1.66, thru=14.3, I=15.5, T=28.2, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:06,708] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:13,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:53:14,092] loop.py:45 INFO TRAIN_STEP step=10283 loss=1.265403 Training: 57%|█████▋ | 10283/18018 [22:23:00<18:04:59, 8.42s/step, loss=1.265, acc=0.703, acc_len=1.69, thru=16.4, I=16.1, T=29.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:14,253] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:20,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:53:20,810] loop.py:45 INFO TRAIN_STEP step=10284 loss=1.328155 Training: 57%|█████▋ | 10284/18018 [22:23:06<16:59:11, 7.91s/step, loss=1.328, acc=0.697, acc_len=1.67, thru=18.7, I=18.0, T=35.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:21,085] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:53:28,146] loop.py:380 INFO TIMING step=10285: step=3.843s data=1.736s compute=3.802s [fwd=2.716s bwd=1.071s opt=0.003s] dispatch=3.482s [2026-10-19 11:53:28,146] loop.py:45 INFO TRAIN_STEP step=10285 loss=1.211872 Training: 57%|█████▋ | 10285/18018 [22:23:14<16:37:01, 7.74s/step, loss=1.212, acc=0.718, acc_len=1.74, thru=14.4, I=15.5, T=33.3, wait=3.5s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:35,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:53:37,501] loop.py:45 INFO TRAIN_STEP step=10286 loss=1.336712 Training: 57%|█████▋ | 10286/18018 [22:23:23<17:39:26, 8.22s/step, loss=1.337, acc=0.698, acc_len=1.67, thru=17.5, I=18.7, T=31.0, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:53:37,496] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10285: forward=2971.2ms backward=1108.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:37,721] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:42,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:53:43,824] loop.py:45 INFO TRAIN_STEP step=10287 loss=1.369017 Training: 57%|█████▋ | 10287/18018 [22:23:29<16:25:51, 7.65s/step, loss=1.369, acc=0.688, acc_len=1.64, thru=17.8, I=25.3, T=33.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:44,281] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:49,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:53:50,174] loop.py:45 INFO TRAIN_STEP step=10288 loss=1.324375 Training: 57%|█████▋ | 10288/18018 [22:23:36<15:35:32, 7.26s/step, loss=1.324, acc=0.699, acc_len=1.68, thru=18.9, I=25.4, T=36.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:50,273] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:56,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:53:57,019] loop.py:45 INFO TRAIN_STEP step=10289 loss=1.210916 Training: 57%|█████▋ | 10289/18018 [22:23:43<15:19:16, 7.14s/step, loss=1.211, acc=0.714, acc_len=1.73, thru=18.1, I=21.1, T=34.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:53:57,512] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:02,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:54:03,708] loop.py:380 INFO TIMING step=10290: step=3.863s data=1.751s compute=3.831s [fwd=2.733s bwd=1.084s opt=0.003s] dispatch=2.816s [2026-10-19 11:54:03,709] loop.py:45 INFO TRAIN_STEP step=10290 loss=1.388985 Training: 57%|█████▋ | 10290/18018 [22:23:49<15:01:49, 7.00s/step, loss=1.389, acc=0.686, acc_len=1.64, thru=18.5, I=22.7, T=33.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:04,106] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:11,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:54:11,793] loop.py:45 INFO TRAIN_STEP step=10291 loss=1.357836 Training: 57%|█████▋ | 10291/18018 [22:23:57<15:43:33, 7.33s/step, loss=1.358, acc=0.688, acc_len=1.64, thru=16.1, I=17.9, T=28.6, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:54:11,789] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10290: forward=3118.7ms backward=1306.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:12,179] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:19,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:54:22,139] loop.py:45 INFO TRAIN_STEP step=10292 loss=1.292892 Training: 57%|█████▋ | 10292/18018 [22:24:08<17:40:06, 8.23s/step, loss=1.293, acc=0.705, acc_len=1.70, thru=13.5, I=16.8, T=25.1, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:22,440] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:28,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:54:29,354] loop.py:45 INFO TRAIN_STEP step=10293 loss=1.290843 Training: 57%|█████▋ | 10293/18018 [22:24:15<17:00:33, 7.93s/step, loss=1.291, acc=0.704, acc_len=1.70, thru=13.3, I=21.8, T=29.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:29,695] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:37,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:54:37,985] loop.py:45 INFO TRAIN_STEP step=10294 loss=1.394439 Training: 57%|█████▋ | 10294/18018 [22:24:24<17:27:39, 8.14s/step, loss=1.394, acc=0.685, acc_len=1.64, thru=13.9, I=15.2, T=33.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:38,374] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:54:45,342] loop.py:380 INFO TIMING step=10295: step=3.907s data=1.720s compute=3.878s [fwd=2.744s bwd=1.119s opt=0.003s] dispatch=3.442s [2026-10-19 11:54:45,342] loop.py:45 INFO TRAIN_STEP step=10295 loss=1.264530 Training: 57%|█████▋ | 10295/18018 [22:24:31<16:57:22, 7.90s/step, loss=1.265, acc=0.703, acc_len=1.69, thru=16.4, I=16.8, T=32.8, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:50,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:54:52,887] loop.py:45 INFO TRAIN_STEP step=10296 loss=1.311447 Training: 57%|█████▋ | 10296/18018 [22:24:38<16:43:23, 7.80s/step, loss=1.311, acc=0.699, acc_len=1.68, thru=17.5, I=20.2, T=28.6, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:54:52,882] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10295: forward=3058.5ms backward=1360.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:53,340] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:54:59,639] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:55:00,058] loop.py:45 INFO TRAIN_STEP step=10297 loss=1.226830 Training: 57%|█████▋ | 10297/18018 [22:24:46<16:19:11, 7.61s/step, loss=1.227, acc=0.712, acc_len=1.73, thru=15.3, I=22.3, T=30.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:00,142] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:05,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:55:06,921] loop.py:45 INFO TRAIN_STEP step=10298 loss=1.434948 Training: 57%|█████▋ | 10298/18018 [22:24:53<15:50:14, 7.39s/step, loss=1.435, acc=0.678, acc_len=1.61, thru=17.1, I=19.6, T=29.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:07,311] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:11,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:55:13,295] loop.py:45 INFO TRAIN_STEP step=10299 loss=1.245451 Training: 57%|█████▋ | 10299/18018 [22:24:59<15:11:05, 7.08s/step, loss=1.245, acc=0.710, acc_len=1.72, thru=17.9, I=23.9, T=30.7, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:13,776] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:20,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:55:21,404] loop.py:380 INFO TIMING step=10300: step=4.017s data=1.578s compute=3.983s [fwd=2.756s bwd=1.212s opt=0.003s] dispatch=4.083s [2026-10-19 11:55:21,405] loop.py:45 INFO TRAIN_STEP step=10300 loss=1.340317 Training: 57%|█████▋ | 10300/18018 [22:25:07<15:50:32, 7.39s/step, loss=1.340, acc=0.695, acc_len=1.67, thru=14.0, I=18.9, T=31.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:21,906] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:27,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:55:28,302] loop.py:45 INFO TRAIN_STEP step=10301 loss=1.297205 Training: 57%|█████▋ | 10301/18018 [22:25:14<15:31:28, 7.24s/step, loss=1.297, acc=0.700, acc_len=1.68, thru=17.0, I=18.9, T=32.7, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:55:28,297] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10300: forward=2812.6ms backward=1047.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:28,490] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:35,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:55:36,922] loop.py:45 INFO TRAIN_STEP step=10302 loss=1.334933 Training: 57%|█████▋ | 10302/18018 [22:25:23<16:24:30, 7.66s/step, loss=1.335, acc=0.696, acc_len=1.67, thru=15.3, I=19.1, T=27.1, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:37,184] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 11:55:44,762] loop.py:45 INFO TRAIN_STEP step=10303 loss=1.294629 Training: 57%|█████▋ | 10303/18018 [22:25:30<16:31:37, 7.71s/step, loss=1.295, acc=0.704, acc_len=1.69, thru=12.8, I=17.5, T=29.6, wait=3.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:52,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:55:54,295] loop.py:45 INFO TRAIN_STEP step=10304 loss=1.339713 Training: 57%|█████▋ | 10304/18018 [22:25:40<17:41:40, 8.26s/step, loss=1.340, acc=0.693, acc_len=1.66, thru=15.3, I=14.8, T=30.6, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:55:54,622] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:00,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:56:01,197] loop.py:380 INFO TIMING step=10305: step=3.858s data=1.644s compute=3.820s [fwd=2.710s bwd=1.094s opt=0.003s] dispatch=3.034s [2026-10-19 11:56:01,197] loop.py:45 INFO TRAIN_STEP step=10305 loss=1.397490 Training: 57%|█████▋ | 10305/18018 [22:25:47<16:49:11, 7.85s/step, loss=1.397, acc=0.687, acc_len=1.64, thru=17.5, I=20.9, T=33.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:01,230] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 11:56:08,112] loop.py:45 INFO TRAIN_STEP step=10306 loss=1.305051 Training: 57%|█████▋ | 10306/18018 [22:25:54<16:13:02, 7.57s/step, loss=1.305, acc=0.699, acc_len=1.68, thru=15.6, I=18.6, T=32.8, wait=3.0s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:56:08,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10305: forward=2625.9ms backward=1226.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:15,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:56:17,488] loop.py:45 INFO TRAIN_STEP step=10307 loss=1.419443 Training: 57%|█████▋ | 10307/18018 [22:26:03<17:22:27, 8.11s/step, loss=1.419, acc=0.677, acc_len=1.60, thru=15.6, I=17.9, T=30.8, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:17,681] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:27,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:56:29,095] loop.py:45 INFO TRAIN_STEP step=10308 loss=1.316474 Training: 57%|█████▋ | 10308/18018 [22:26:15<19:37:04, 9.16s/step, loss=1.316, acc=0.700, acc_len=1.68, thru=11.2, I=11.4, T=24.5, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:29,366] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:33,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:56:35,412] loop.py:45 INFO TRAIN_STEP step=10309 loss=1.338008 Training: 57%|█████▋ | 10309/18018 [22:26:21<17:47:17, 8.31s/step, loss=1.338, acc=0.697, acc_len=1.67, thru=17.6, I=23.7, T=33.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:35,814] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:41,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:56:42,493] loop.py:380 INFO TIMING step=10310: step=4.212s data=1.956s compute=4.184s [fwd=2.984s bwd=1.184s opt=0.003s] dispatch=2.862s [2026-10-19 11:56:42,493] loop.py:45 INFO TRAIN_STEP step=10310 loss=1.358749 Training: 57%|█████▋ | 10310/18018 [22:26:28<17:00:00, 7.94s/step, loss=1.359, acc=0.692, acc_len=1.65, thru=18.4, I=25.2, T=30.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:42,766] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:48,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:56:49,439] loop.py:45 INFO TRAIN_STEP step=10311 loss=1.287633 Training: 57%|█████▋ | 10311/18018 [22:26:35<16:21:34, 7.64s/step, loss=1.288, acc=0.705, acc_len=1.71, thru=18.2, I=22.5, T=30.9, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:56:49,434] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10310: forward=2910.0ms backward=1183.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:49,593] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:53,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:56:55,307] loop.py:45 INFO TRAIN_STEP step=10312 loss=1.220028 Training: 57%|█████▋ | 10312/18018 [22:26:41<15:13:03, 7.11s/step, loss=1.220, acc=0.715, acc_len=1.74, thru=19.6, I=24.8, T=31.4, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:56:55,397] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:02,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:57:03,243] loop.py:45 INFO TRAIN_STEP step=10313 loss=1.297833 Training: 57%|█████▋ | 10313/18018 [22:26:49<15:44:44, 7.36s/step, loss=1.298, acc=0.703, acc_len=1.69, thru=15.0, I=17.7, T=29.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:03,743] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:08,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:57:10,886] loop.py:45 INFO TRAIN_STEP step=10314 loss=1.243870 Training: 57%|█████▋ | 10314/18018 [22:26:56<15:55:42, 7.44s/step, loss=1.244, acc=0.713, acc_len=1.73, thru=17.2, I=21.3, T=30.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:11,130] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:15,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:57:16,940] loop.py:380 INFO TIMING step=10315: step=3.646s data=1.452s compute=3.620s [fwd=2.554s bwd=1.050s opt=0.003s] dispatch=2.398s [2026-10-19 11:57:16,940] loop.py:45 INFO TRAIN_STEP step=10315 loss=1.382005 Training: 57%|█████▋ | 10315/18018 [22:27:03<15:02:06, 7.03s/step, loss=1.382, acc=0.686, acc_len=1.63, thru=18.5, I=28.3, T=35.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:17,301] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:23,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:57:24,115] loop.py:45 INFO TRAIN_STEP step=10316 loss=1.376574 Training: 57%|█████▋ | 10316/18018 [22:27:10<15:07:38, 7.07s/step, loss=1.377, acc=0.686, acc_len=1.64, thru=16.4, I=21.1, T=31.6, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:57:24,110] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10315: forward=2903.2ms backward=1097.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:24,598] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:57:30,310] loop.py:45 INFO TRAIN_STEP step=10317 loss=1.154759 Training: 57%|█████▋ | 10317/18018 [22:27:16<14:33:49, 6.81s/step, loss=1.155, acc=0.727, acc_len=1.78, thru=17.2, I=18.8, T=31.8, wait=2.2s, pool=49, epoch=2/2][2026-10-19 11:57:42,909] loop.py:45 INFO TRAIN_STEP step=10318 loss=1.343157 Training: 57%|█████▋ | 10318/18018 [22:27:28<18:16:39, 8.55s/step, loss=1.343, acc=0.700, acc_len=1.68, thru=9.7, I=9.1, T=26.3, wait=7.7s, pool=49, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:50,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:57:50,523] loop.py:45 INFO TRAIN_STEP step=10319 loss=1.278455 Training: 57%|█████▋ | 10319/18018 [22:27:36<17:40:39, 8.27s/step, loss=1.278, acc=0.705, acc_len=1.70, thru=16.7, I=16.0, T=31.3, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:50,960] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:56,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:57:57,891] loop.py:380 INFO TIMING step=10320: step=4.150s data=1.798s compute=4.122s [fwd=2.902s bwd=1.205s opt=0.003s] dispatch=3.209s [2026-10-19 11:57:57,891] loop.py:45 INFO TRAIN_STEP step=10320 loss=1.378364 Training: 57%|█████▋ | 10320/18018 [22:27:43<17:05:57, 8.00s/step, loss=1.378, acc=0.687, acc_len=1.63, thru=17.6, I=19.6, T=30.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:57:58,181] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:05,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:58:05,406] loop.py:45 INFO TRAIN_STEP step=10321 loss=1.327624 Training: 57%|█████▋ | 10321/18018 [22:27:51<16:47:18, 7.85s/step, loss=1.328, acc=0.696, acc_len=1.66, thru=16.4, I=19.5, T=32.8, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:58:05,402] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10320: forward=2726.6ms backward=1110.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:05,833] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 11:58:11,923] loop.py:45 INFO TRAIN_STEP step=10322 loss=1.416879 Training: 57%|█████▋ | 10322/18018 [22:27:58<15:55:48, 7.45s/step, loss=1.417, acc=0.679, acc_len=1.61, thru=17.7, I=20.5, T=34.6, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:20,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:58:21,011] loop.py:45 INFO TRAIN_STEP step=10323 loss=1.420854 Training: 57%|█████▋ | 10323/18018 [22:28:07<16:58:40, 7.94s/step, loss=1.421, acc=0.681, acc_len=1.62, thru=14.9, I=15.1, T=32.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:21,429] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:25,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:58:27,295] loop.py:45 INFO TRAIN_STEP step=10324 loss=1.246830 Training: 57%|█████▋ | 10324/18018 [22:28:13<15:54:37, 7.44s/step, loss=1.247, acc=0.707, acc_len=1.70, thru=19.3, I=25.1, T=35.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:27,792] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:32,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:58:34,661] loop.py:380 INFO TIMING step=10325: step=4.456s data=2.089s compute=4.420s [fwd=3.031s bwd=1.374s opt=0.003s] dispatch=2.902s [2026-10-19 11:58:34,661] loop.py:45 INFO TRAIN_STEP step=10325 loss=1.296460 Training: 57%|█████▋ | 10325/18018 [22:28:20<15:51:30, 7.42s/step, loss=1.296, acc=0.705, acc_len=1.70, thru=16.3, I=25.8, T=28.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:34,759] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 [2026-10-19 11:58:42,360] loop.py:45 INFO TRAIN_STEP step=10326 loss=1.339406 Training: 57%|█████▋ | 10326/18018 [22:28:28<16:02:12, 7.51s/step, loss=1.339, acc=0.690, acc_len=1.65, thru=13.7, I=18.7, T=32.6, wait=3.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:58:42,356] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10325: forward=2797.6ms backward=1075.4ms [2026-10-19 11:58:51,157] loop.py:45 INFO TRAIN_STEP step=10327 loss=1.533570 Training: 57%|█████▋ | 10327/18018 [22:28:37<16:51:40, 7.89s/step, loss=1.534, acc=0.662, acc_len=1.55, thru=12.4, I=12.8, T=29.5, wait=4.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:57,601] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:58:59,072] loop.py:45 INFO TRAIN_STEP step=10328 loss=1.419238 Training: 57%|█████▋ | 10328/18018 [22:28:45<16:52:26, 7.90s/step, loss=1.419, acc=0.682, acc_len=1.62, thru=16.9, I=17.4, T=31.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:58:59,130] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:04,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:59:05,601] loop.py:45 INFO TRAIN_STEP step=10329 loss=1.397257 Training: 57%|█████▋ | 10329/18018 [22:28:51<15:59:36, 7.49s/step, loss=1.397, acc=0.685, acc_len=1.63, thru=18.4, I=25.0, T=35.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:05,902] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 11:59:11,307] loop.py:380 INFO TIMING step=10330: step=3.481s data=1.468s compute=3.450s [fwd=2.439s bwd=0.997s opt=0.003s] dispatch=2.216s [2026-10-19 11:59:11,308] loop.py:45 INFO TRAIN_STEP step=10330 loss=1.233079 Training: 57%|█████▋ | 10330/18018 [22:28:57<14:50:52, 6.95s/step, loss=1.233, acc=0.713, acc_len=1.72, thru=21.0, I=24.7, T=36.8, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:11,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:11,802] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:17,733] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:59:18,727] loop.py:45 INFO TRAIN_STEP step=10331 loss=1.295005 Training: 57%|█████▋ | 10331/18018 [22:29:04<15:08:49, 7.09s/step, loss=1.295, acc=0.701, acc_len=1.69, thru=18.2, I=19.5, T=35.1, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 11:59:18,722] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10330: forward=2555.5ms backward=1039.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:18,760] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 11:59:24,453] loop.py:45 INFO TRAIN_STEP step=10332 loss=1.354163 Training: 57%|█████▋ | 10332/18018 [22:29:10<14:16:04, 6.68s/step, loss=1.354, acc=0.688, acc_len=1.64, thru=20.0, I=21.3, T=33.5, wait=1.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:32,484] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:59:33,133] loop.py:45 INFO TRAIN_STEP step=10333 loss=1.410199 Training: 57%|█████▋ | 10333/18018 [22:29:19<15:32:47, 7.28s/step, loss=1.410, acc=0.683, acc_len=1.63, thru=15.3, I=16.0, T=31.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:33,489] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 11:59:40,420] loop.py:45 INFO TRAIN_STEP step=10334 loss=1.447687 Training: 57%|█████▋ | 10334/18018 [22:29:26<15:32:48, 7.28s/step, loss=1.448, acc=0.671, acc_len=1.58, thru=16.4, I=19.2, T=27.8, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:48,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:59:49,642] loop.py:380 INFO TIMING step=10335: step=4.191s data=1.914s compute=4.159s [fwd=2.958s bwd=1.185s opt=0.003s] dispatch=5.021s [2026-10-19 11:59:49,642] loop.py:45 INFO TRAIN_STEP step=10335 loss=1.322906 Training: 57%|█████▋ | 10335/18018 [22:29:35<16:47:02, 7.86s/step, loss=1.323, acc=0.700, acc_len=1.68, thru=14.4, I=13.8, T=30.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:50,142] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:55,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 11:59:55,785] loop.py:45 INFO TRAIN_STEP step=10336 loss=1.306767 (TrainerActor pid=9310) [2026-10-19 11:59:55,780] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10335: forward=2488.2ms backward=1034.9ms Training: 57%|█████▋ | 10336/18018 [22:29:41<15:40:55, 7.35s/step, loss=1.307, acc=0.700, acc_len=1.67, thru=18.8, I=22.6, T=35.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 11:59:56,048] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:00:03,297] loop.py:45 INFO TRAIN_STEP step=10337 loss=1.301423 Training: 57%|█████▋ | 10337/18018 [22:29:49<15:46:56, 7.40s/step, loss=1.301, acc=0.697, acc_len=1.67, thru=15.6, I=15.9, T=34.6, wait=3.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:00:09,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:00:12,432] loop.py:45 INFO TRAIN_STEP step=10338 loss=1.334899 Training: 57%|█████▋ | 10338/18018 [22:29:58<16:53:35, 7.92s/step, loss=1.335, acc=0.697, acc_len=1.67, thru=16.1, I=20.7, T=26.0, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:00:12,510] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=33/64 [2026-10-19 12:00:19,939] loop.py:45 INFO TRAIN_STEP step=10339 loss=1.247639 Training: 57%|█████▋ | 10339/18018 [22:30:06<16:37:42, 7.80s/step, loss=1.248, acc=0.704, acc_len=1.70, thru=16.2, I=17.2, T=29.6, wait=3.2s, pool=57, epoch=2/2][2026-10-19 12:00:29,322] loop.py:380 INFO TIMING step=10340: step=4.839s data=2.406s compute=4.801s [fwd=3.315s bwd=1.471s opt=0.003s] dispatch=4.534s [2026-10-19 12:00:29,323] loop.py:45 INFO TRAIN_STEP step=10340 loss=1.421032 Training: 57%|█████▋ | 10340/18018 [22:30:15<17:38:27, 8.27s/step, loss=1.421, acc=0.687, acc_len=1.64, thru=13.8, I=13.1, T=26.5, wait=4.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:00:33,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:00:36,596] loop.py:45 INFO TRAIN_STEP step=10341 loss=1.277594 Training: 57%|█████▋ | 10341/18018 [22:30:22<17:00:06, 7.97s/step, loss=1.278, acc=0.707, acc_len=1.70, thru=16.6, I=20.2, T=27.7, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:00:36,591] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10340: forward=3184.8ms backward=1388.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:00:36,752] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 [2026-10-19 12:00:42,319] loop.py:45 INFO TRAIN_STEP step=10342 loss=1.236797 Training: 57%|█████▋ | 10342/18018 [22:30:28<15:33:36, 7.30s/step, loss=1.237, acc=0.711, acc_len=1.71, thru=17.4, I=26.1, T=35.2, wait=2.1s, pool=53, epoch=2/2][2026-10-19 12:00:50,149] loop.py:45 INFO TRAIN_STEP step=10343 loss=1.371562 Training: 57%|█████▋ | 10343/18018 [22:30:36<15:53:50, 7.46s/step, loss=1.372, acc=0.690, acc_len=1.64, thru=15.6, I=14.7, T=33.3, wait=4.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:00:57,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:00:58,988] loop.py:45 INFO TRAIN_STEP step=10344 loss=1.437940 Training: 57%|█████▋ | 10344/18018 [22:30:45<16:46:47, 7.87s/step, loss=1.438, acc=0.681, acc_len=1.62, thru=17.3, I=17.2, T=33.8, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:00:59,458] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:08,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:01:09,078] loop.py:380 INFO TIMING step=10345: step=4.049s data=1.776s compute=4.008s [fwd=2.781s bwd=1.211s opt=0.003s] dispatch=6.031s [2026-10-19 12:01:09,078] loop.py:45 INFO TRAIN_STEP step=10345 loss=1.278231 Training: 57%|█████▋ | 10345/18018 [22:30:55<18:11:49, 8.54s/step, loss=1.278, acc=0.705, acc_len=1.71, thru=13.3, I=13.5, T=31.6, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:09,215] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:14,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:01:16,312] loop.py:45 INFO TRAIN_STEP step=10346 loss=1.398239 Training: 57%|█████▋ | 10346/18018 [22:31:02<17:21:35, 8.15s/step, loss=1.398, acc=0.684, acc_len=1.62, thru=18.9, I=21.7, T=31.1, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:01:16,308] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10345: forward=2747.0ms backward=1313.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:16,404] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:23,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:01:24,626] loop.py:45 INFO TRAIN_STEP step=10347 loss=1.299855 Training: 57%|█████▋ | 10347/18018 [22:31:10<17:27:55, 8.20s/step, loss=1.300, acc=0.701, acc_len=1.68, thru=15.6, I=18.4, T=30.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:25,046] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:31,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:01:31,590] loop.py:45 INFO TRAIN_STEP step=10348 loss=1.237610 Training: 57%|█████▋ | 10348/18018 [22:31:17<16:40:28, 7.83s/step, loss=1.238, acc=0.710, acc_len=1.72, thru=17.2, I=20.8, T=33.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:32,021] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:38,822] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:01:39,071] loop.py:45 INFO TRAIN_STEP step=10349 loss=1.258648 Training: 57%|█████▋ | 10349/18018 [22:31:25<16:27:10, 7.72s/step, loss=1.259, acc=0.709, acc_len=1.71, thru=16.5, I=19.9, T=27.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:39,324] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:47,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:01:48,945] loop.py:380 INFO TIMING step=10350: step=5.148s data=2.556s compute=5.084s [fwd=3.532s bwd=1.537s opt=0.003s] dispatch=4.716s [2026-10-19 12:01:48,946] loop.py:45 INFO TRAIN_STEP step=10350 loss=1.363635 Training: 57%|█████▋ | 10350/18018 [22:31:35<17:49:29, 8.37s/step, loss=1.364, acc=0.690, acc_len=1.64, thru=13.3, I=14.6, T=24.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:49,095] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:55,794] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:01:57,074] loop.py:45 INFO TRAIN_STEP step=10351 loss=1.220063 Training: 57%|█████▋ | 10351/18018 [22:31:43<17:40:12, 8.30s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=13.9, I=17.4, T=31.0, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:01:57,071] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10350: forward=2904.7ms backward=1168.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:01:57,301] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:01,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:02:02,763] loop.py:45 INFO TRAIN_STEP step=10352 loss=1.291785 Training: 57%|█████▋ | 10352/18018 [22:31:48<16:00:01, 7.51s/step, loss=1.292, acc=0.704, acc_len=1.70, thru=20.2, I=25.2, T=37.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:02,829] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:07,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:02:09,436] loop.py:45 INFO TRAIN_STEP step=10353 loss=1.306531 Training: 57%|█████▋ | 10353/18018 [22:31:55<15:27:46, 7.26s/step, loss=1.307, acc=0.705, acc_len=1.70, thru=19.2, I=24.0, T=29.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:09,496] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:18,097] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:02:18,185] loop.py:45 INFO TRAIN_STEP step=10354 loss=1.388929 Training: 57%|█████▋ | 10354/18018 [22:32:04<16:24:30, 7.71s/step, loss=1.389, acc=0.684, acc_len=1.63, thru=14.8, I=14.4, T=32.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:18,599] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:23,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:02:24,607] loop.py:380 INFO TIMING step=10355: step=3.883s data=1.714s compute=3.840s [fwd=2.716s bwd=1.110s opt=0.003s] dispatch=2.531s [2026-10-19 12:02:24,607] loop.py:45 INFO TRAIN_STEP step=10355 loss=1.339760 Training: 57%|█████▋ | 10355/18018 [22:32:10<15:35:08, 7.32s/step, loss=1.340, acc=0.693, acc_len=1.65, thru=19.3, I=21.7, T=33.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:24,882] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:30,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:02:30,890] loop.py:45 INFO TRAIN_STEP step=10356 loss=1.155596 Training: 57%|█████▋ | 10356/18018 [22:32:16<14:55:19, 7.01s/step, loss=1.156, acc=0.726, acc_len=1.77, thru=18.9, I=24.3, T=33.1, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:02:30,886] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10355: forward=2669.7ms backward=1149.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:31,007] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 12:02:36,504] loop.py:45 INFO TRAIN_STEP step=10357 loss=1.470690 Training: 57%|█████▋ | 10357/18018 [22:32:22<14:01:44, 6.59s/step, loss=1.471, acc=0.670, acc_len=1.58, thru=20.6, I=26.9, T=36.2, wait=2.1s, pool=61, epoch=2/2][2026-10-19 12:02:44,116] loop.py:45 INFO TRAIN_STEP step=10358 loss=1.267355 Training: 57%|█████▋ | 10358/18018 [22:32:30<14:40:34, 6.90s/step, loss=1.267, acc=0.708, acc_len=1.71, thru=13.9, I=14.8, T=31.6, wait=3.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:51,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:02:51,919] loop.py:45 INFO TRAIN_STEP step=10359 loss=1.388567 Training: 57%|█████▋ | 10359/18018 [22:32:38<15:15:05, 7.17s/step, loss=1.389, acc=0.686, acc_len=1.63, thru=17.2, I=16.6, T=34.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:52,177] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:56,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:02:57,619] loop.py:380 INFO TIMING step=10360: step=3.748s data=1.617s compute=3.705s [fwd=2.576s bwd=1.114s opt=0.003s] dispatch=1.944s [2026-10-19 12:02:57,619] loop.py:45 INFO TRAIN_STEP step=10360 loss=1.196901 Training: 57%|█████▋ | 10360/18018 [22:32:43<14:18:48, 6.73s/step, loss=1.197, acc=0.716, acc_len=1.73, thru=20.8, I=25.2, T=34.2, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:02:58,002] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:03:02,897] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:03:04,186] loop.py:45 INFO TRAIN_STEP step=10361 loss=1.264849 Training: 58%|█████▊ | 10361/18018 [22:32:50<14:12:25, 6.68s/step, loss=1.265, acc=0.708, acc_len=1.71, thru=18.5, I=25.4, T=32.0, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:03:04,181] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10360: forward=2801.0ms backward=1152.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:03:04,406] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:03:10,612] loop.py:45 INFO TRAIN_STEP step=10362 loss=1.335104 Training: 58%|█████▊ | 10362/18018 [22:32:56<14:02:34, 6.60s/step, loss=1.335, acc=0.695, acc_len=1.67, thru=17.4, I=20.3, T=30.9, wait=2.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:03:19,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:03:20,253] loop.py:45 INFO TRAIN_STEP step=10363 loss=1.269328 Training: 58%|█████▊ | 10363/18018 [22:33:06<15:58:48, 7.52s/step, loss=1.269, acc=0.703, acc_len=1.69, thru=14.5, I=16.0, T=31.7, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:03:20,648] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:03:27,259] loop.py:45 INFO TRAIN_STEP step=10364 loss=1.272336 Training: 58%|█████▊ | 10364/18018 [22:33:13<15:39:05, 7.36s/step, loss=1.272, acc=0.707, acc_len=1.70, thru=16.0, I=18.4, T=30.9, wait=2.9s, pool=45, epoch=2/2][2026-10-19 12:03:35,832] loop.py:380 INFO TIMING step=10365: step=4.073s data=1.654s compute=4.046s [fwd=2.914s bwd=1.116s opt=0.003s] dispatch=4.494s [2026-10-19 12:03:35,832] loop.py:45 INFO TRAIN_STEP step=10365 loss=1.368800 Training: 58%|█████▊ | 10365/18018 [22:33:21<16:25:27, 7.73s/step, loss=1.369, acc=0.689, acc_len=1.64, thru=14.4, I=13.4, T=31.4, wait=4.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:03:43,494] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:03:44,990] loop.py:45 INFO TRAIN_STEP step=10366 loss=1.292655 Training: 58%|█████▊ | 10366/18018 [22:33:31<17:20:08, 8.16s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=16.7, I=16.0, T=31.5, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:03:44,985] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10365: forward=2879.6ms backward=1141.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:03:45,029] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:03:50,128] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:03:52,310] loop.py:45 INFO TRAIN_STEP step=10367 loss=1.339381 Training: 58%|█████▊ | 10367/18018 [22:33:38<16:47:57, 7.90s/step, loss=1.339, acc=0.688, acc_len=1.65, thru=16.6, I=23.4, T=28.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:03:52,647] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:03:59,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:03:59,796] loop.py:45 INFO TRAIN_STEP step=10368 loss=1.309214 Training: 58%|█████▊ | 10368/18018 [22:33:45<16:31:47, 7.78s/step, loss=1.309, acc=0.701, acc_len=1.68, thru=14.1, I=19.4, T=33.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:00,021] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:06,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:04:07,485] loop.py:45 INFO TRAIN_STEP step=10369 loss=1.309619 Training: 58%|█████▊ | 10369/18018 [22:33:53<16:28:13, 7.75s/step, loss=1.310, acc=0.699, acc_len=1.68, thru=16.4, I=18.5, T=35.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:07,799] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:13,131] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:04:13,734] loop.py:380 INFO TIMING step=10370: step=4.076s data=1.909s compute=4.041s [fwd=2.848s bwd=1.178s opt=0.003s] dispatch=2.164s [2026-10-19 12:04:13,734] loop.py:45 INFO TRAIN_STEP step=10370 loss=1.296223 Training: 58%|█████▊ | 10370/18018 [22:33:59<15:30:41, 7.30s/step, loss=1.296, acc=0.699, acc_len=1.68, thru=19.9, I=22.9, T=31.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:14,137] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:18,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:04:19,887] loop.py:45 INFO TRAIN_STEP step=10371 loss=1.170958 Training: 58%|█████▊ | 10371/18018 [22:34:05<14:46:39, 6.96s/step, loss=1.171, acc=0.727, acc_len=1.77, thru=19.4, I=23.9, T=34.7, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:04:19,882] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10370: forward=2590.2ms backward=1057.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:20,373] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:25,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:04:27,272] loop.py:45 INFO TRAIN_STEP step=10372 loss=1.316381 Training: 58%|█████▊ | 10372/18018 [22:34:13<15:02:51, 7.08s/step, loss=1.316, acc=0.696, acc_len=1.67, thru=16.7, I=24.7, T=32.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:27,606] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:33,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:04:34,446] loop.py:45 INFO TRAIN_STEP step=10373 loss=1.343648 Training: 58%|█████▊ | 10373/18018 [22:34:20<15:06:12, 7.11s/step, loss=1.344, acc=0.691, acc_len=1.65, thru=15.8, I=20.9, T=33.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:34,724] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:41,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:04:42,546] loop.py:45 INFO TRAIN_STEP step=10374 loss=1.396351 Training: 58%|█████▊ | 10374/18018 [22:34:28<15:43:47, 7.41s/step, loss=1.396, acc=0.684, acc_len=1.63, thru=16.4, I=17.7, T=30.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:42,921] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:04:49,509] loop.py:380 INFO TIMING step=10375: step=3.816s data=1.766s compute=3.782s [fwd=2.706s bwd=1.061s opt=0.003s] dispatch=3.138s [2026-10-19 12:04:49,510] loop.py:45 INFO TRAIN_STEP step=10375 loss=1.366527 Training: 58%|█████▊ | 10375/18018 [22:34:35<15:26:40, 7.27s/step, loss=1.367, acc=0.687, acc_len=1.64, thru=15.7, I=19.8, T=33.5, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:55,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:04:57,307] loop.py:45 INFO TRAIN_STEP step=10376 loss=1.282851 Training: 58%|█████▊ | 10376/18018 [22:34:43<15:46:31, 7.43s/step, loss=1.283, acc=0.700, acc_len=1.68, thru=16.4, I=18.8, T=32.0, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:04:57,303] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10375: forward=2836.8ms backward=1118.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:04:57,443] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 12:05:02,905] loop.py:45 INFO TRAIN_STEP step=10377 loss=1.376889 Training: 58%|█████▊ | 10377/18018 [22:34:48<14:36:26, 6.88s/step, loss=1.377, acc=0.685, acc_len=1.63, thru=20.3, I=24.8, T=36.5, wait=2.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:05:09,341] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:05:10,270] loop.py:45 INFO TRAIN_STEP step=10378 loss=1.376274 Training: 58%|█████▊ | 10378/18018 [22:34:56<14:54:46, 7.03s/step, loss=1.376, acc=0.693, acc_len=1.65, thru=18.2, I=19.9, T=33.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:05:10,359] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:05:14,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:05:16,101] loop.py:45 INFO TRAIN_STEP step=10379 loss=1.313291 Training: 58%|█████▊ | 10379/18018 [22:35:02<14:08:54, 6.67s/step, loss=1.313, acc=0.701, acc_len=1.69, thru=20.4, I=26.9, T=32.4, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:05:16,201] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:05:23,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:05:25,146] loop.py:380 INFO TIMING step=10380: step=4.124s data=1.913s compute=4.083s [fwd=2.850s bwd=1.218s opt=0.003s] dispatch=4.912s [2026-10-19 12:05:25,146] loop.py:45 INFO TRAIN_STEP step=10380 loss=1.312078 Training: 58%|█████▊ | 10380/18018 [22:35:11<15:39:38, 7.38s/step, loss=1.312, acc=0.698, acc_len=1.68, thru=14.3, I=16.1, T=31.0, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:05:25,363] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:05:31,394] loop.py:45 INFO TRAIN_STEP step=10381 loss=1.382618 Training: 58%|█████▊ | 10381/18018 [22:35:17<14:56:13, 7.04s/step, loss=1.383, acc=0.686, acc_len=1.64, thru=17.0, I=18.9, T=32.6, wait=2.3s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:05:31,390] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10380: forward=2750.1ms backward=1134.6ms [2026-10-19 12:05:42,320] loop.py:45 INFO TRAIN_STEP step=10382 loss=1.295593 Training: 58%|█████▊ | 10382/18018 [22:35:28<17:24:27, 8.21s/step, loss=1.296, acc=0.703, acc_len=1.69, thru=11.5, I=14.2, T=28.3, wait=6.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:05:48,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:05:50,001] loop.py:45 INFO TRAIN_STEP step=10383 loss=1.421862 Training: 58%|█████▊ | 10383/18018 [22:35:36<17:04:09, 8.05s/step, loss=1.422, acc=0.682, acc_len=1.62, thru=15.2, I=17.0, T=31.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:05:50,285] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:05:57,133] loop.py:45 INFO TRAIN_STEP step=10384 loss=1.343754 Training: 58%|█████▊ | 10384/18018 [22:35:43<16:29:02, 7.77s/step, loss=1.344, acc=0.696, acc_len=1.67, thru=14.4, I=17.9, T=32.7, wait=3.2s, pool=49, epoch=2/2][2026-10-19 12:06:05,568] loop.py:380 INFO TIMING step=10385: step=4.138s data=1.812s compute=4.105s [fwd=2.943s bwd=1.147s opt=0.003s] dispatch=4.288s [2026-10-19 12:06:05,568] loop.py:45 INFO TRAIN_STEP step=10385 loss=1.277868 Training: 58%|█████▊ | 10385/18018 [22:35:51<16:54:20, 7.97s/step, loss=1.278, acc=0.703, acc_len=1.69, thru=15.3, I=16.9, T=30.9, wait=4.3s, pool=53, epoch=2/2][2026-10-19 12:06:14,413] loop.py:45 INFO TRAIN_STEP step=10386 loss=1.290194 Training: 58%|█████▊ | 10386/18018 [22:36:00<17:27:21, 8.23s/step, loss=1.290, acc=0.703, acc_len=1.69, thru=11.3, I=14.0, T=32.3, wait=4.9s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:06:14,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10385: forward=2853.8ms backward=1062.8ms [2026-10-19 12:06:25,328] loop.py:45 INFO TRAIN_STEP step=10387 loss=1.341086 Training: 58%|█████▊ | 10387/18018 [22:36:11<19:09:30, 9.04s/step, loss=1.341, acc=0.692, acc_len=1.65, thru=16.2, I=12.9, T=31.6, wait=6.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:06:34,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:06:35,844] loop.py:45 INFO TRAIN_STEP step=10388 loss=1.417347 Training: 58%|█████▊ | 10388/18018 [22:36:21<20:05:43, 9.48s/step, loss=1.417, acc=0.682, acc_len=1.62, thru=13.4, I=15.3, T=29.1, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:06:36,046] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:06:43,238] loop.py:45 INFO TRAIN_STEP step=10389 loss=1.308003 Training: 58%|█████▊ | 10389/18018 [22:36:29<18:45:54, 8.85s/step, loss=1.308, acc=0.698, acc_len=1.68, thru=13.3, I=16.0, T=33.9, wait=3.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:06:53,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:06:53,817] loop.py:380 INFO TIMING step=10390: step=4.003s data=1.724s compute=3.956s [fwd=2.793s bwd=1.146s opt=0.003s] dispatch=6.566s [2026-10-19 12:06:53,818] loop.py:45 INFO TRAIN_STEP step=10390 loss=1.192470 Training: 58%|█████▊ | 10390/18018 [22:36:39<19:51:32, 9.37s/step, loss=1.192, acc=0.716, acc_len=1.74, thru=14.2, I=14.1, T=32.0, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:06:54,198] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:07:02,486] loop.py:45 INFO TRAIN_STEP step=10391 loss=1.309163 Training: 58%|█████▊ | 10391/18018 [22:36:48<19:24:35, 9.16s/step, loss=1.309, acc=0.699, acc_len=1.68, thru=14.3, I=14.5, T=34.4, wait=4.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:07:02,482] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10390: forward=2587.0ms backward=1076.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:09,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:07:10,279] loop.py:45 INFO TRAIN_STEP step=10392 loss=1.276818 Training: 58%|█████▊ | 10392/18018 [22:36:56<18:32:15, 8.75s/step, loss=1.277, acc=0.708, acc_len=1.71, thru=16.9, I=17.3, T=27.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:10,435] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:16,940] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:07:18,657] loop.py:45 INFO TRAIN_STEP step=10393 loss=1.317127 Training: 58%|█████▊ | 10393/18018 [22:37:04<18:17:52, 8.64s/step, loss=1.317, acc=0.698, acc_len=1.68, thru=15.3, I=19.4, T=28.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:18,956] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:25,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:07:26,826] loop.py:45 INFO TRAIN_STEP step=10394 loss=1.431642 Training: 58%|█████▊ | 10394/18018 [22:37:12<17:59:49, 8.50s/step, loss=1.432, acc=0.679, acc_len=1.60, thru=13.3, I=19.9, T=24.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:27,130] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:34,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:07:34,652] loop.py:380 INFO TIMING step=10395: step=3.935s data=1.583s compute=3.899s [fwd=2.756s bwd=1.126s opt=0.003s] dispatch=3.881s [2026-10-19 12:07:34,652] loop.py:45 INFO TRAIN_STEP step=10395 loss=1.235132 Training: 58%|█████▊ | 10395/18018 [22:37:20<17:34:03, 8.30s/step, loss=1.235, acc=0.716, acc_len=1.74, thru=13.7, I=17.0, T=32.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:34,825] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:43,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:07:44,719] loop.py:45 INFO TRAIN_STEP step=10396 loss=1.337644 Training: 58%|█████▊ | 10396/18018 [22:37:30<18:41:18, 8.83s/step, loss=1.338, acc=0.694, acc_len=1.67, thru=12.9, I=14.3, T=30.2, wait=5.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:07:44,714] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10395: forward=3020.6ms backward=1173.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:45,084] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:52,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:07:53,573] loop.py:45 INFO TRAIN_STEP step=10397 loss=1.335354 Training: 58%|█████▊ | 10397/18018 [22:37:39<18:42:17, 8.84s/step, loss=1.335, acc=0.693, acc_len=1.65, thru=15.1, I=16.6, T=33.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:07:53,945] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:07:59,746] loop.py:45 INFO TRAIN_STEP step=10398 loss=1.404465 Training: 58%|█████▊ | 10398/18018 [22:37:45<17:00:40, 8.04s/step, loss=1.404, acc=0.681, acc_len=1.62, thru=18.2, I=19.1, T=38.3, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:06,631] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:08:08,044] loop.py:45 INFO TRAIN_STEP step=10399 loss=1.289903 Training: 58%|█████▊ | 10399/18018 [22:37:54<17:10:27, 8.11s/step, loss=1.290, acc=0.701, acc_len=1.69, thru=17.5, I=20.0, T=33.0, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:08,144] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-19 12:08:14,051] loop.py:380 INFO TIMING step=10400: step=3.867s data=1.702s compute=3.836s [fwd=2.697s bwd=1.125s opt=0.003s] dispatch=2.131s [2026-10-19 12:08:14,052] loop.py:45 INFO TRAIN_STEP step=10400 loss=1.265603 Training: 58%|█████▊ | 10400/18018 [22:38:00<15:50:00, 7.48s/step, loss=1.266, acc=0.706, acc_len=1.71, thru=20.1, I=24.4, T=33.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:14,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:14,552] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:24,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:08:25,167] loop.py:45 INFO TRAIN_STEP step=10401 loss=1.406050 Training: 58%|█████▊ | 10401/18018 [22:38:11<18:08:15, 8.57s/step, loss=1.406, acc=0.683, acc_len=1.62, thru=13.5, I=11.5, T=27.9, wait=6.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:08:25,162] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10400: forward=3216.3ms backward=1299.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:25,422] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:08:33,741] loop.py:45 INFO TRAIN_STEP step=10402 loss=1.365773 Training: 58%|█████▊ | 10402/18018 [22:38:19<18:08:10, 8.57s/step, loss=1.366, acc=0.687, acc_len=1.64, thru=15.0, I=14.8, T=32.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:33,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:34,221] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:08:43,505] loop.py:45 INFO TRAIN_STEP step=10403 loss=1.498732 Training: 58%|█████▊ | 10403/18018 [22:38:29<18:53:23, 8.93s/step, loss=1.499, acc=0.668, acc_len=1.57, thru=13.5, I=13.1, T=32.1, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:43,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:43,999] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:51,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:08:53,493] loop.py:45 INFO TRAIN_STEP step=10404 loss=1.333081 Training: 58%|█████▊ | 10404/18018 [22:38:39<19:33:35, 9.25s/step, loss=1.333, acc=0.696, acc_len=1.67, thru=13.5, I=16.7, T=30.0, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:08:53,891] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 12:09:00,324] loop.py:380 INFO TIMING step=10405: step=3.895s data=1.678s compute=3.862s [fwd=2.680s bwd=1.167s opt=0.003s] dispatch=2.926s [2026-10-19 12:09:00,324] loop.py:45 INFO TRAIN_STEP step=10405 loss=1.202183 Training: 58%|█████▊ | 10405/18018 [22:38:46<18:01:22, 8.52s/step, loss=1.202, acc=0.716, acc_len=1.73, thru=13.4, I=20.0, T=32.9, wait=2.9s, pool=41, epoch=2/2][2026-10-19 12:09:07,820] loop.py:45 INFO TRAIN_STEP step=10406 loss=1.339504 Training: 58%|█████▊ | 10406/18018 [22:38:53<17:22:08, 8.21s/step, loss=1.340, acc=0.690, acc_len=1.64, thru=18.4, I=18.4, T=29.4, wait=3.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:09:07,815] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10405: forward=2966.8ms backward=1324.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:09:15,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:09:16,764] loop.py:45 INFO TRAIN_STEP step=10407 loss=1.328402 Training: 58%|█████▊ | 10407/18018 [22:39:02<17:49:50, 8.43s/step, loss=1.328, acc=0.696, acc_len=1.67, thru=14.7, I=15.2, T=32.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:09:16,960] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:09:23,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:09:24,226] loop.py:45 INFO TRAIN_STEP step=10408 loss=1.261665 Training: 58%|█████▊ | 10408/18018 [22:39:10<17:12:35, 8.14s/step, loss=1.262, acc=0.707, acc_len=1.71, thru=17.3, I=21.5, T=25.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:09:24,257] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:09:29,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:09:31,570] loop.py:45 INFO TRAIN_STEP step=10409 loss=1.356789 Training: 58%|█████▊ | 10409/18018 [22:39:17<16:42:10, 7.90s/step, loss=1.357, acc=0.690, acc_len=1.65, thru=16.2, I=19.8, T=27.1, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:09:31,729] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 12:09:38,239] loop.py:380 INFO TIMING step=10410: step=3.542s data=1.470s compute=3.516s [fwd=2.546s bwd=0.956s opt=0.003s] dispatch=3.118s [2026-10-19 12:09:38,239] loop.py:45 INFO TRAIN_STEP step=10410 loss=1.355254 Training: 58%|█████▊ | 10410/18018 [22:39:24<15:55:05, 7.53s/step, loss=1.355, acc=0.692, acc_len=1.66, thru=13.4, I=22.5, T=36.1, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:09:48,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:09:49,296] loop.py:45 INFO TRAIN_STEP step=10411 loss=1.299314 Training: 58%|█████▊ | 10411/18018 [22:39:35<18:09:05, 8.59s/step, loss=1.299, acc=0.702, acc_len=1.70, thru=13.9, I=14.8, T=31.4, wait=7.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:09:49,291] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10410: forward=2852.4ms backward=1168.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:09:49,557] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:09:55,817] loop.py:45 INFO TRAIN_STEP step=10412 loss=1.420630 Training: 58%|█████▊ | 10412/18018 [22:39:41<16:50:18, 7.97s/step, loss=1.421, acc=0.678, acc_len=1.60, thru=16.4, I=21.7, T=34.6, wait=2.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:04,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:10:07,159] loop.py:45 INFO TRAIN_STEP step=10413 loss=1.278449 Training: 58%|█████▊ | 10413/18018 [22:39:53<18:58:21, 8.98s/step, loss=1.278, acc=0.703, acc_len=1.69, thru=13.3, I=14.2, T=28.5, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:07,383] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:14,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:10:15,104] loop.py:45 INFO TRAIN_STEP step=10414 loss=1.358753 Training: 58%|█████▊ | 10414/18018 [22:40:01<18:18:46, 8.67s/step, loss=1.359, acc=0.693, acc_len=1.65, thru=16.6, I=17.5, T=31.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:15,527] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:10:22,672] loop.py:380 INFO TIMING step=10415: step=4.368s data=2.013s compute=4.330s [fwd=3.095s bwd=1.220s opt=0.003s] dispatch=3.192s [2026-10-19 12:10:22,672] loop.py:45 INFO TRAIN_STEP step=10415 loss=1.296181 Training: 58%|█████▊ | 10415/18018 [22:40:08<17:36:41, 8.34s/step, loss=1.296, acc=0.699, acc_len=1.68, thru=14.6, I=15.4, T=29.3, wait=3.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:33,972] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:10:35,920] loop.py:45 INFO TRAIN_STEP step=10416 loss=1.384460 Training: 58%|█████▊ | 10416/18018 [22:40:22<20:43:16, 9.81s/step, loss=1.384, acc=0.686, acc_len=1.63, thru=14.9, I=11.8, T=28.9, wait=8.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:10:35,915] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10415: forward=3197.5ms backward=1197.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:36,005] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 [2026-10-19 12:10:41,502] loop.py:45 INFO TRAIN_STEP step=10417 loss=1.308439 Training: 58%|█████▊ | 10417/18018 [22:40:27<18:02:17, 8.54s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=18.8, I=26.4, T=37.4, wait=2.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:48,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:10:51,159] loop.py:45 INFO TRAIN_STEP step=10418 loss=1.345810 Training: 58%|█████▊ | 10418/18018 [22:40:37<18:44:27, 8.88s/step, loss=1.346, acc=0.693, acc_len=1.66, thru=13.6, I=16.6, T=26.9, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:51,324] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:56,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:10:57,418] loop.py:45 INFO TRAIN_STEP step=10419 loss=1.177845 Training: 58%|█████▊ | 10419/18018 [22:40:43<17:04:52, 8.09s/step, loss=1.178, acc=0.722, acc_len=1.75, thru=16.4, I=25.6, T=31.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:10:57,588] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:04,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:11:05,416] loop.py:380 INFO TIMING step=10420: step=4.608s data=2.226s compute=4.573s [fwd=3.183s bwd=1.375s opt=0.003s] dispatch=3.379s [2026-10-19 12:11:05,417] loop.py:45 INFO TRAIN_STEP step=10420 loss=1.439184 Training: 58%|█████▊ | 10420/18018 [22:40:51<17:01:09, 8.06s/step, loss=1.439, acc=0.672, acc_len=1.59, thru=14.6, I=18.7, T=27.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:05,686] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:11:13,893] loop.py:45 INFO TRAIN_STEP step=10421 loss=1.286889 Training: 58%|█████▊ | 10421/18018 [22:40:59<17:16:38, 8.19s/step, loss=1.287, acc=0.700, acc_len=1.68, thru=12.4, I=17.7, T=31.8, wait=4.4s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:11:13,889] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10420: forward=2882.5ms backward=1096.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:24,555] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:11:26,355] loop.py:45 INFO TRAIN_STEP step=10422 loss=1.376532 Training: 58%|█████▊ | 10422/18018 [22:41:12<19:58:50, 9.47s/step, loss=1.377, acc=0.688, acc_len=1.64, thru=12.5, I=12.9, T=26.4, wait=7.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:26,567] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:31,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:11:33,900] loop.py:45 INFO TRAIN_STEP step=10423 loss=1.257109 Training: 58%|█████▊ | 10423/18018 [22:41:19<18:45:40, 8.89s/step, loss=1.257, acc=0.710, acc_len=1.71, thru=15.1, I=21.3, T=27.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:34,202] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:39,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:11:39,894] loop.py:45 INFO TRAIN_STEP step=10424 loss=1.374184 Training: 58%|█████▊ | 10424/18018 [22:41:25<16:55:30, 8.02s/step, loss=1.374, acc=0.689, acc_len=1.64, thru=17.2, I=26.6, T=35.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:40,245] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:45,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:11:47,160] loop.py:380 INFO TIMING step=10425: step=4.194s data=1.956s compute=4.165s [fwd=2.995s bwd=1.155s opt=0.003s] dispatch=3.060s [2026-10-19 12:11:47,160] loop.py:45 INFO TRAIN_STEP step=10425 loss=1.340236 Training: 58%|█████▊ | 10425/18018 [22:41:33<16:26:25, 7.79s/step, loss=1.340, acc=0.692, acc_len=1.65, thru=16.8, I=23.1, T=30.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:11:47,657] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 12:11:55,581] loop.py:45 INFO TRAIN_STEP step=10426 loss=1.291308 Training: 58%|█████▊ | 10426/18018 [22:41:41<16:50:06, 7.98s/step, loss=1.291, acc=0.701, acc_len=1.68, thru=11.6, I=13.7, T=33.6, wait=4.6s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:11:55,577] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10425: forward=2674.2ms backward=1091.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:02,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:12:03,584] loop.py:45 INFO TRAIN_STEP step=10427 loss=1.261614 Training: 58%|█████▊ | 10427/18018 [22:41:49<16:50:47, 7.99s/step, loss=1.262, acc=0.709, acc_len=1.71, thru=19.5, I=16.1, T=30.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:04,005] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:12,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:12:13,820] loop.py:45 INFO TRAIN_STEP step=10428 loss=1.347137 Training: 58%|█████▊ | 10428/18018 [22:41:59<18:15:59, 8.66s/step, loss=1.347, acc=0.692, acc_len=1.66, thru=13.0, I=14.3, T=28.4, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:14,084] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:18,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:12:20,744] loop.py:45 INFO TRAIN_STEP step=10429 loss=1.327802 Training: 58%|█████▊ | 10429/18018 [22:42:06<17:09:44, 8.14s/step, loss=1.328, acc=0.698, acc_len=1.68, thru=17.2, I=23.6, T=27.0, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:20,844] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:26,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:12:26,996] loop.py:380 INFO TIMING step=10430: step=3.695s data=1.639s compute=3.665s [fwd=2.617s bwd=1.034s opt=0.003s] dispatch=2.548s [2026-10-19 12:12:26,996] loop.py:45 INFO TRAIN_STEP step=10430 loss=1.272314 Training: 58%|█████▊ | 10430/18018 [22:42:13<15:57:58, 7.57s/step, loss=1.272, acc=0.705, acc_len=1.70, thru=16.4, I=22.7, T=34.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:27,353] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:33,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:12:34,099] loop.py:45 INFO TRAIN_STEP step=10431 loss=1.376112 Training: 58%|█████▊ | 10431/18018 [22:42:20<15:39:51, 7.43s/step, loss=1.376, acc=0.688, acc_len=1.64, thru=17.6, I=20.1, T=33.3, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:12:34,094] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10430: forward=2687.5ms backward=1093.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:34,570] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:12:41,551] loop.py:45 INFO TRAIN_STEP step=10432 loss=1.255671 Training: 58%|█████▊ | 10432/18018 [22:42:27<15:40:31, 7.44s/step, loss=1.256, acc=0.712, acc_len=1.72, thru=14.4, I=19.3, T=30.3, wait=3.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:48,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:12:49,189] loop.py:45 INFO TRAIN_STEP step=10433 loss=1.413797 Training: 58%|█████▊ | 10433/18018 [22:42:35<15:47:59, 7.50s/step, loss=1.414, acc=0.679, acc_len=1.61, thru=17.6, I=17.7, T=33.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:49,569] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:55,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:12:56,401] loop.py:45 INFO TRAIN_STEP step=10434 loss=1.273666 Training: 58%|█████▊ | 10434/18018 [22:42:42<15:36:58, 7.41s/step, loss=1.274, acc=0.707, acc_len=1.70, thru=17.1, I=20.9, T=34.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:12:56,745] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:13:04,256] loop.py:380 INFO TIMING step=10435: step=3.682s data=1.619s compute=3.647s [fwd=2.568s bwd=1.064s opt=0.003s] dispatch=4.163s [2026-10-19 12:13:04,256] loop.py:45 INFO TRAIN_STEP step=10435 loss=1.505762 Training: 58%|█████▊ | 10435/18018 [22:42:50<15:53:31, 7.54s/step, loss=1.506, acc=0.664, acc_len=1.56, thru=13.3, I=15.2, T=34.8, wait=4.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:13:11,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:13:13,391] loop.py:45 INFO TRAIN_STEP step=10436 loss=1.311452 Training: 58%|█████▊ | 10436/18018 [22:42:59<16:53:44, 8.02s/step, loss=1.311, acc=0.698, acc_len=1.67, thru=16.5, I=18.9, T=29.1, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:13:13,386] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10435: forward=3070.6ms backward=1276.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:13:13,674] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:13:20,352] loop.py:45 INFO TRAIN_STEP step=10437 loss=1.299259 Training: 58%|█████▊ | 10437/18018 [22:43:06<16:13:18, 7.70s/step, loss=1.299, acc=0.703, acc_len=1.69, thru=14.3, I=17.2, T=31.2, wait=2.9s, pool=41, epoch=2/2][2026-10-19 12:13:33,420] loop.py:45 INFO TRAIN_STEP step=10438 loss=1.241106 Training: 58%|█████▊ | 10438/18018 [22:43:19<19:36:40, 9.31s/step, loss=1.241, acc=0.715, acc_len=1.73, thru=8.9, I=7.9, T=30.0, wait=8.8s, pool=37, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:13:44,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:13:44,933] loop.py:45 INFO TRAIN_STEP step=10439 loss=1.347585 Training: 58%|█████▊ | 10439/18018 [22:43:31<20:59:48, 9.97s/step, loss=1.348, acc=0.691, acc_len=1.65, thru=14.8, I=14.8, T=26.5, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:13:45,166] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:13:49,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:13:51,260] loop.py:380 INFO TIMING step=10440: step=3.861s data=1.610s compute=3.835s [fwd=2.725s bwd=1.096s opt=0.003s] dispatch=2.454s [2026-10-19 12:13:51,261] loop.py:45 INFO TRAIN_STEP step=10440 loss=1.308509 Training: 58%|█████▊ | 10440/18018 [22:43:37<18:41:25, 8.88s/step, loss=1.309, acc=0.700, acc_len=1.67, thru=17.4, I=22.5, T=33.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:13:51,667] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:13:56,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:13:58,638] loop.py:45 INFO TRAIN_STEP step=10441 loss=1.297933 Training: 58%|█████▊ | 10441/18018 [22:43:44<17:44:26, 8.43s/step, loss=1.298, acc=0.704, acc_len=1.69, thru=16.5, I=25.9, T=32.3, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:13:58,633] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10440: forward=2681.9ms backward=1231.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:13:58,755] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:05,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:14:06,052] loop.py:45 INFO TRAIN_STEP step=10442 loss=1.266358 Training: 58%|█████▊ | 10442/18018 [22:43:52<17:05:49, 8.12s/step, loss=1.266, acc=0.704, acc_len=1.70, thru=15.6, I=20.7, T=33.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:06,553] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:11,307] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:14:12,046] loop.py:45 INFO TRAIN_STEP step=10443 loss=1.349468 Training: 58%|█████▊ | 10443/18018 [22:43:58<15:44:58, 7.48s/step, loss=1.349, acc=0.693, acc_len=1.66, thru=20.0, I=22.8, T=36.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:12,311] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:18,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:14:19,155] loop.py:45 INFO TRAIN_STEP step=10444 loss=1.301562 Training: 58%|█████▊ | 10444/18018 [22:44:05<15:30:38, 7.37s/step, loss=1.302, acc=0.703, acc_len=1.70, thru=18.4, I=21.9, T=34.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:19,611] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:14:25,890] loop.py:380 INFO TIMING step=10445: step=3.580s data=1.514s compute=3.541s [fwd=2.551s bwd=0.975s opt=0.003s] dispatch=3.147s [2026-10-19 12:14:25,890] loop.py:45 INFO TRAIN_STEP step=10445 loss=1.317090 Training: 58%|█████▊ | 10445/18018 [22:44:11<15:06:22, 7.18s/step, loss=1.317, acc=0.696, acc_len=1.67, thru=17.3, I=20.3, T=35.8, wait=3.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:32,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:14:33,318] loop.py:45 INFO TRAIN_STEP step=10446 loss=1.244895 Training: 58%|█████▊ | 10446/18018 [22:44:19<15:15:38, 7.26s/step, loss=1.245, acc=0.711, acc_len=1.72, thru=17.8, I=18.8, T=29.6, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:14:33,314] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10445: forward=2999.6ms backward=1266.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:33,497] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 12:14:40,211] loop.py:45 INFO TRAIN_STEP step=10447 loss=1.258462 Training: 58%|█████▊ | 10447/18018 [22:44:26<15:01:46, 7.15s/step, loss=1.258, acc=0.707, acc_len=1.70, thru=17.2, I=20.0, T=31.7, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:48,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:14:49,300] loop.py:45 INFO TRAIN_STEP step=10448 loss=1.294458 Training: 58%|█████▊ | 10448/18018 [22:44:35<16:15:07, 7.73s/step, loss=1.294, acc=0.700, acc_len=1.68, thru=15.2, I=14.4, T=28.8, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:49,724] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:54,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:14:55,516] loop.py:45 INFO TRAIN_STEP step=10449 loss=1.222131 Training: 58%|█████▊ | 10449/18018 [22:44:41<15:17:46, 7.28s/step, loss=1.222, acc=0.715, acc_len=1.73, thru=20.0, I=20.4, T=34.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:14:55,832] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:15:01,718] loop.py:380 INFO TIMING step=10450: step=3.369s data=1.336s compute=3.341s [fwd=2.398s bwd=0.928s opt=0.003s] dispatch=2.823s [2026-10-19 12:15:01,718] loop.py:45 INFO TRAIN_STEP step=10450 loss=1.200514 Training: 58%|█████▊ | 10450/18018 [22:44:47<14:37:09, 6.95s/step, loss=1.201, acc=0.722, acc_len=1.76, thru=19.6, I=24.7, T=38.0, wait=2.8s, pool=61, epoch=2/2][2026-10-19 12:15:09,425] loop.py:45 INFO TRAIN_STEP step=10451 loss=1.294264 Training: 58%|█████▊ | 10451/18018 [22:44:55<15:05:28, 7.18s/step, loss=1.294, acc=0.702, acc_len=1.69, thru=16.8, I=17.4, T=33.8, wait=3.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:15:09,419] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10450: forward=2579.1ms backward=1166.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:15:14,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:15:16,105] loop.py:45 INFO TRAIN_STEP step=10452 loss=1.382099 Training: 58%|█████▊ | 10452/18018 [22:45:02<14:46:22, 7.03s/step, loss=1.382, acc=0.688, acc_len=1.64, thru=18.7, I=20.8, T=34.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:15:16,275] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:15:23,222] loop.py:45 INFO TRAIN_STEP step=10453 loss=1.186174 Training: 58%|█████▊ | 10453/18018 [22:45:09<14:49:31, 7.06s/step, loss=1.186, acc=0.722, acc_len=1.76, thru=17.2, I=19.4, T=34.3, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:15:31,308] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:15:32,810] loop.py:45 INFO TRAIN_STEP step=10454 loss=1.276103 Training: 58%|█████▊ | 10454/18018 [22:45:18<16:25:19, 7.82s/step, loss=1.276, acc=0.705, acc_len=1.70, thru=14.7, I=16.6, T=30.2, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:15:33,317] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:15:38,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:15:39,943] loop.py:380 INFO TIMING step=10455: step=4.248s data=1.976s compute=4.217s [fwd=2.930s bwd=1.271s opt=0.003s] dispatch=2.874s [2026-10-19 12:15:39,943] loop.py:45 INFO TRAIN_STEP step=10455 loss=1.390520 Training: 58%|█████▊ | 10455/18018 [22:45:26<15:59:24, 7.61s/step, loss=1.391, acc=0.685, acc_len=1.63, thru=16.7, I=24.3, T=30.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:15:40,130] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:15:45,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:15:47,136] loop.py:45 INFO TRAIN_STEP step=10456 loss=1.435904 Training: 58%|█████▊ | 10456/18018 [22:45:33<15:43:25, 7.49s/step, loss=1.436, acc=0.677, acc_len=1.60, thru=17.3, I=22.9, T=31.4, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:15:47,132] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10455: forward=2893.9ms backward=1127.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:15:47,468] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:15:55,812] loop.py:45 INFO TRAIN_STEP step=10457 loss=1.289593 Training: 58%|█████▊ | 10457/18018 [22:45:41<16:28:17, 7.84s/step, loss=1.290, acc=0.703, acc_len=1.69, thru=13.4, I=13.9, T=34.1, wait=4.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:16:06,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:16:06,609] loop.py:45 INFO TRAIN_STEP step=10458 loss=1.289693 Training: 58%|█████▊ | 10458/18018 [22:45:52<18:19:48, 8.73s/step, loss=1.290, acc=0.704, acc_len=1.70, thru=13.4, I=13.5, T=27.4, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:16:06,722] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 12:16:12,717] loop.py:45 INFO TRAIN_STEP step=10459 loss=1.294812 Training: 58%|█████▊ | 10459/18018 [22:45:58<16:40:35, 7.94s/step, loss=1.295, acc=0.704, acc_len=1.69, thru=17.2, I=18.1, T=34.2, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:16:19,680] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:16:20,310] loop.py:380 INFO TIMING step=10460: step=3.942s data=1.620s compute=3.921s [fwd=2.720s bwd=1.185s opt=0.003s] dispatch=3.642s [2026-10-19 12:16:20,311] loop.py:45 INFO TRAIN_STEP step=10460 loss=1.376334 Training: 58%|█████▊ | 10460/18018 [22:46:06<16:27:21, 7.84s/step, loss=1.376, acc=0.690, acc_len=1.65, thru=18.0, I=18.5, T=32.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:16:20,685] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:16:26,524] loop.py:45 INFO TRAIN_STEP step=10461 loss=1.262219 Training: 58%|█████▊ | 10461/18018 [22:46:12<15:25:51, 7.35s/step, loss=1.262, acc=0.706, acc_len=1.70, thru=18.2, I=20.3, T=35.5, wait=2.6s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:16:26,520] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10460: forward=2588.2ms backward=970.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:16:34,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:16:35,410] loop.py:45 INFO TRAIN_STEP step=10462 loss=1.276051 Training: 58%|█████▊ | 10462/18018 [22:46:21<16:23:42, 7.81s/step, loss=1.276, acc=0.702, acc_len=1.68, thru=15.8, I=16.3, T=29.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:16:35,849] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:16:42,708] loop.py:45 INFO TRAIN_STEP step=10463 loss=1.358774 Training: 58%|█████▊ | 10463/18018 [22:46:28<16:04:09, 7.66s/step, loss=1.359, acc=0.689, acc_len=1.64, thru=15.0, I=19.7, T=33.6, wait=3.5s, pool=57, epoch=2/2][2026-10-19 12:16:52,581] loop.py:45 INFO TRAIN_STEP step=10464 loss=1.377861 Training: 58%|█████▊ | 10464/18018 [22:46:38<17:27:45, 8.32s/step, loss=1.378, acc=0.692, acc_len=1.65, thru=14.2, I=11.9, T=30.8, wait=5.7s, pool=57, epoch=2/2][2026-10-19 12:16:59,362] loop.py:380 INFO TIMING step=10465: step=3.657s data=1.489s compute=3.623s [fwd=2.572s bwd=1.036s opt=0.003s] dispatch=3.114s [2026-10-19 12:16:59,363] loop.py:45 INFO TRAIN_STEP step=10465 loss=1.272243 Training: 58%|█████▊ | 10465/18018 [22:46:45<16:29:20, 7.86s/step, loss=1.272, acc=0.706, acc_len=1.70, thru=18.0, I=16.7, T=35.0, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:04,860] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:17:05,793] loop.py:45 INFO TRAIN_STEP step=10466 loss=1.232634 Training: 58%|█████▊ | 10466/18018 [22:46:51<15:35:20, 7.43s/step, loss=1.233, acc=0.713, acc_len=1.73, thru=18.8, I=21.4, T=34.0, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:17:05,788] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10465: forward=2612.6ms backward=1102.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:05,877] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:11,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:17:12,487] loop.py:45 INFO TRAIN_STEP step=10467 loss=1.397685 Training: 58%|█████▊ | 10467/18018 [22:46:58<15:07:23, 7.21s/step, loss=1.398, acc=0.683, acc_len=1.63, thru=18.6, I=23.1, T=35.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:12,855] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:19,866] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:17:20,567] loop.py:45 INFO TRAIN_STEP step=10468 loss=1.271977 Training: 58%|█████▊ | 10468/18018 [22:47:06<15:40:07, 7.47s/step, loss=1.272, acc=0.708, acc_len=1.70, thru=15.8, I=19.1, T=28.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:20,875] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:27,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:17:28,795] loop.py:45 INFO TRAIN_STEP step=10469 loss=1.387146 Training: 58%|█████▊ | 10469/18018 [22:47:14<16:08:31, 7.70s/step, loss=1.387, acc=0.685, acc_len=1.63, thru=15.8, I=18.7, T=28.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:29,200] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:37,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:17:38,245] loop.py:380 INFO TIMING step=10470: step=4.195s data=1.827s compute=4.159s [fwd=2.893s bwd=1.251s opt=0.003s] dispatch=5.244s [2026-10-19 12:17:38,245] loop.py:45 INFO TRAIN_STEP step=10470 loss=1.468662 Training: 58%|█████▊ | 10470/18018 [22:47:24<17:14:31, 8.22s/step, loss=1.469, acc=0.666, acc_len=1.57, thru=14.2, I=14.9, T=30.5, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:38,564] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:44,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:17:45,256] loop.py:45 INFO TRAIN_STEP step=10471 loss=1.353591 Training: 58%|█████▊ | 10471/18018 [22:47:31<16:28:36, 7.86s/step, loss=1.354, acc=0.692, acc_len=1.66, thru=17.4, I=21.4, T=31.0, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:17:45,251] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10470: forward=2877.3ms backward=1199.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:45,543] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:51,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:17:52,584] loop.py:45 INFO TRAIN_STEP step=10472 loss=1.353468 Training: 58%|█████▊ | 10472/18018 [22:47:38<16:08:23, 7.70s/step, loss=1.353, acc=0.691, acc_len=1.65, thru=16.7, I=20.2, T=30.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:17:52,727] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:18:00,377] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:18:00,511] loop.py:45 INFO TRAIN_STEP step=10473 loss=1.327447 Training: 58%|█████▊ | 10473/18018 [22:47:46<16:16:50, 7.77s/step, loss=1.327, acc=0.702, acc_len=1.68, thru=16.2, I=17.4, T=36.6, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:18:00,880] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:18:06,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 12:18:09,610] loop.py:45 INFO TRAIN_STEP step=10474 loss=1.483944 Training: 58%|█████▊ | 10474/18018 [22:47:55<17:06:57, 8.17s/step, loss=1.484, acc=0.664, acc_len=1.56, thru=15.6, I=20.6, T=28.3, wait=4.6s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:18:09,833] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=5/64 [2026-10-19 12:18:15,594] loop.py:380 INFO TIMING step=10475: step=3.482s data=1.482s compute=3.456s [fwd=2.468s bwd=0.974s opt=0.003s] dispatch=2.491s [2026-10-19 12:18:15,594] loop.py:45 INFO TRAIN_STEP step=10475 loss=1.302850 Training: 58%|█████▊ | 10475/18018 [22:48:01<15:44:27, 7.51s/step, loss=1.303, acc=0.703, acc_len=1.69, thru=15.2, I=24.4, T=36.8, wait=2.5s, pool=49, epoch=2/2][2026-10-19 12:18:23,324] loop.py:45 INFO TRAIN_STEP step=10476 loss=1.231309 Training: 58%|█████▊ | 10476/18018 [22:48:09<15:52:28, 7.58s/step, loss=1.231, acc=0.713, acc_len=1.74, thru=18.2, I=16.8, T=33.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:18:23,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 12:18:23,319] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10475: forward=2755.0ms backward=1072.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:18:23,758] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:18:30,438] loop.py:45 INFO TRAIN_STEP step=10477 loss=1.388770 Training: 58%|█████▊ | 10477/18018 [22:48:16<15:34:52, 7.44s/step, loss=1.389, acc=0.685, acc_len=1.63, thru=16.2, I=16.3, T=34.3, wait=3.4s, pool=49, epoch=2/2][2026-10-19 12:18:40,855] loop.py:45 INFO TRAIN_STEP step=10478 loss=1.275011 Training: 58%|█████▊ | 10478/18018 [22:48:26<17:27:02, 8.33s/step, loss=1.275, acc=0.706, acc_len=1.71, thru=13.7, I=15.8, T=26.3, wait=5.5s, pool=57, epoch=2/2][2026-10-19 12:18:48,728] loop.py:45 INFO TRAIN_STEP step=10479 loss=1.260316 Training: 58%|█████▊ | 10479/18018 [22:48:34<17:09:41, 8.19s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=15.1, I=16.0, T=33.1, wait=4.0s, pool=57, epoch=2/2][2026-10-19 12:18:56,411] loop.py:380 INFO TIMING step=10480: step=3.656s data=1.549s compute=3.612s [fwd=2.518s bwd=1.079s opt=0.003s] dispatch=4.015s [2026-10-19 12:18:56,411] loop.py:45 INFO TRAIN_STEP step=10480 loss=1.225992 Training: 58%|█████▊ | 10480/18018 [22:48:42<16:50:11, 8.04s/step, loss=1.226, acc=0.716, acc_len=1.74, thru=15.1, I=15.6, T=35.0, wait=4.0s, pool=53, epoch=2/2][2026-10-19 12:19:03,798] loop.py:45 INFO TRAIN_STEP step=10481 loss=1.313566 Training: 58%|█████▊ | 10481/18018 [22:48:49<16:25:24, 7.84s/step, loss=1.314, acc=0.702, acc_len=1.69, thru=16.8, I=17.5, T=34.6, wait=3.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:19:03,793] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10480: forward=2576.9ms backward=1067.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:19:11,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:19:12,744] loop.py:45 INFO TRAIN_STEP step=10482 loss=1.288453 Training: 58%|█████▊ | 10482/18018 [22:48:58<17:06:51, 8.18s/step, loss=1.288, acc=0.701, acc_len=1.68, thru=15.4, I=16.5, T=31.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:19:13,204] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:19:20,759] loop.py:45 INFO TRAIN_STEP step=10483 loss=1.322922 Training: 58%|█████▊ | 10483/18018 [22:49:06<17:00:38, 8.13s/step, loss=1.323, acc=0.699, acc_len=1.67, thru=13.3, I=18.1, T=31.1, wait=3.9s, pool=57, epoch=2/2][2026-10-19 12:19:28,651] loop.py:45 INFO TRAIN_STEP step=10484 loss=1.437333 Training: 58%|█████▊ | 10484/18018 [22:49:14<16:51:36, 8.06s/step, loss=1.437, acc=0.676, acc_len=1.61, thru=15.7, I=17.7, T=28.4, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:19:35,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:19:35,958] loop.py:380 INFO TIMING step=10485: step=4.149s data=1.780s compute=4.118s [fwd=2.866s bwd=1.237s opt=0.003s] dispatch=3.149s [2026-10-19 12:19:35,958] loop.py:45 INFO TRAIN_STEP step=10485 loss=1.304416 Training: 58%|█████▊ | 10485/18018 [22:49:22<16:23:17, 7.83s/step, loss=1.304, acc=0.700, acc_len=1.68, thru=16.3, I=16.9, T=30.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:19:36,140] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:19:41,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:19:43,755] loop.py:45 INFO TRAIN_STEP step=10486 loss=1.307597 Training: 58%|█████▊ | 10486/18018 [22:49:29<16:21:48, 7.82s/step, loss=1.308, acc=0.697, acc_len=1.68, thru=15.4, I=22.0, T=25.7, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:19:43,750] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10485: forward=3433.6ms backward=1492.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:19:43,891] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=9/64 [2026-10-19 12:19:51,073] loop.py:45 INFO TRAIN_STEP step=10487 loss=1.378590 Training: 58%|█████▊ | 10487/18018 [22:49:37<16:02:46, 7.67s/step, loss=1.379, acc=0.690, acc_len=1.64, thru=11.9, I=16.7, T=29.5, wait=3.0s, pool=49, epoch=2/2][2026-10-19 12:19:58,304] loop.py:45 INFO TRAIN_STEP step=10488 loss=1.473578 Training: 58%|█████▊ | 10488/18018 [22:49:44<15:46:08, 7.54s/step, loss=1.474, acc=0.669, acc_len=1.58, thru=17.8, I=18.4, T=34.0, wait=3.5s, pool=57, epoch=2/2][2026-10-19 12:20:06,296] loop.py:45 INFO TRAIN_STEP step=10489 loss=1.403532 Training: 58%|█████▊ | 10489/18018 [22:49:52<16:02:59, 7.67s/step, loss=1.404, acc=0.681, acc_len=1.62, thru=12.9, I=13.6, T=32.2, wait=4.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:14,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:20:15,171] loop.py:380 INFO TIMING step=10490: step=3.609s data=1.505s compute=3.579s [fwd=2.555s bwd=1.008s opt=0.003s] dispatch=5.257s [2026-10-19 12:20:15,171] loop.py:45 INFO TRAIN_STEP step=10490 loss=1.307272 Training: 58%|█████▊ | 10490/18018 [22:50:01<16:48:07, 8.03s/step, loss=1.307, acc=0.697, acc_len=1.67, thru=16.8, I=15.9, T=35.5, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:15,426] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:20:24,577] loop.py:45 INFO TRAIN_STEP step=10491 loss=1.416048 Training: 58%|█████▊ | 10491/18018 [22:50:10<17:39:38, 8.45s/step, loss=1.416, acc=0.679, acc_len=1.61, thru=11.2, I=11.5, T=33.2, wait=5.5s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:20:24,572] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10490: forward=2701.9ms backward=1094.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:33,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:20:34,745] loop.py:45 INFO TRAIN_STEP step=10492 loss=1.321897 Training: 58%|█████▊ | 10492/18018 [22:50:20<18:44:12, 8.96s/step, loss=1.322, acc=0.701, acc_len=1.68, thru=14.8, I=17.0, T=25.3, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:35,062] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:39,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:20:41,593] loop.py:45 INFO TRAIN_STEP step=10493 loss=1.311787 Training: 58%|█████▊ | 10493/18018 [22:50:27<17:24:24, 8.33s/step, loss=1.312, acc=0.696, acc_len=1.67, thru=17.0, I=23.5, T=29.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:42,034] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:49,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:20:50,064] loop.py:45 INFO TRAIN_STEP step=10494 loss=1.267246 Training: 58%|█████▊ | 10494/18018 [22:50:36<17:29:41, 8.37s/step, loss=1.267, acc=0.707, acc_len=1.70, thru=15.9, I=16.5, T=28.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:50,201] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:55,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:20:56,510] loop.py:380 INFO TIMING step=10495: step=3.818s data=1.631s compute=3.783s [fwd=2.737s bwd=1.031s opt=0.003s] dispatch=2.620s [2026-10-19 12:20:56,510] loop.py:45 INFO TRAIN_STEP step=10495 loss=1.357584 Training: 58%|█████▊ | 10495/18018 [22:50:42<16:17:06, 7.79s/step, loss=1.358, acc=0.687, acc_len=1.64, thru=18.3, I=20.0, T=33.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:20:56,970] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:02,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:21:04,201] loop.py:45 INFO TRAIN_STEP step=10496 loss=1.352789 Training: 58%|█████▊ | 10496/18018 [22:50:50<16:13:15, 7.76s/step, loss=1.353, acc=0.693, acc_len=1.66, thru=16.0, I=22.0, T=28.7, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:21:04,196] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10495: forward=3084.8ms backward=1332.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:04,506] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:10,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:21:10,995] loop.py:45 INFO TRAIN_STEP step=10497 loss=1.385682 Training: 58%|█████▊ | 10497/18018 [22:50:57<15:36:40, 7.47s/step, loss=1.386, acc=0.683, acc_len=1.63, thru=16.1, I=21.7, T=35.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:11,030] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 12:21:17,668] loop.py:45 INFO TRAIN_STEP step=10498 loss=1.344663 Training: 58%|█████▊ | 10498/18018 [22:51:03<15:06:35, 7.23s/step, loss=1.345, acc=0.695, acc_len=1.66, thru=17.5, I=19.4, T=32.4, wait=2.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:24,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:21:26,897] loop.py:45 INFO TRAIN_STEP step=10499 loss=1.375607 Training: 58%|█████▊ | 10499/18018 [22:51:12<16:21:24, 7.83s/step, loss=1.376, acc=0.690, acc_len=1.64, thru=15.5, I=19.8, T=26.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:26,960] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=49/64 [2026-10-19 12:21:33,929] loop.py:380 INFO TIMING step=10500: step=3.542s data=1.471s compute=3.506s [fwd=2.510s bwd=0.982s opt=0.003s] dispatch=3.479s [2026-10-19 12:21:33,929] loop.py:45 INFO TRAIN_STEP step=10500 loss=1.423282 Training: 58%|█████▊ | 10500/18018 [22:51:20<15:51:14, 7.59s/step, loss=1.423, acc=0.679, acc_len=1.60, thru=10.7, I=17.6, T=36.1, wait=3.5s, pool=33, epoch=2/2][2026-10-19 12:21:33,934] loop.py:419 INFO Saving checkpoint at step 10500... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:35,415] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 12:21:36,825] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0010501 (TrainerActor pid=9310) [2026-10-19 12:21:36,829] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:38,431] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:44,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 12:21:36,823] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0010501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 12:21:38,094] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0009751 [2026-10-19 12:21:44,725] loop.py:45 INFO TRAIN_STEP step=10501 loss=1.183335 Training: 58%|█████▊ | 10501/18018 [22:51:30<17:51:31, 8.55s/step, loss=1.183, acc=0.723, acc_len=1.76, thru=14.9, I=19.6, T=32.2, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:21:44,720] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10500: forward=2802.0ms backward=1118.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:21:44,868] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 12:21:52,486] loop.py:45 INFO TRAIN_STEP step=10502 loss=1.364284 Training: 58%|█████▊ | 10502/18018 [22:51:38<17:21:31, 8.31s/step, loss=1.364, acc=0.688, acc_len=1.65, thru=15.3, I=15.5, T=29.3, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:01,893] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:22:03,933] loop.py:45 INFO TRAIN_STEP step=10503 loss=1.385790 Training: 58%|█████▊ | 10503/18018 [22:51:50<19:19:06, 9.25s/step, loss=1.386, acc=0.691, acc_len=1.65, thru=13.3, I=13.4, T=25.3, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:04,408] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:10,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:22:11,307] loop.py:45 INFO TRAIN_STEP step=10504 loss=1.410849 Training: 58%|█████▊ | 10504/18018 [22:51:57<18:08:19, 8.69s/step, loss=1.411, acc=0.683, acc_len=1.63, thru=14.3, I=20.4, T=32.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:11,740] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:19,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:22:20,530] loop.py:380 INFO TIMING step=10505: step=4.735s data=2.193s compute=4.707s [fwd=3.304s bwd=1.390s opt=0.003s] dispatch=4.479s [2026-10-19 12:22:20,530] loop.py:45 INFO TRAIN_STEP step=10505 loss=1.351639 Training: 58%|█████▊ | 10505/18018 [22:52:06<18:28:15, 8.85s/step, loss=1.352, acc=0.694, acc_len=1.66, thru=14.6, I=16.8, T=27.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:20,884] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:26,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:22:27,699] loop.py:45 INFO TRAIN_STEP step=10506 loss=1.404968 Training: 58%|█████▊ | 10506/18018 [22:52:13<17:24:56, 8.35s/step, loss=1.405, acc=0.680, acc_len=1.61, thru=16.2, I=19.9, T=28.5, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:22:27,694] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10505: forward=3160.0ms backward=1283.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:27,972] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:22:34,953] loop.py:45 INFO TRAIN_STEP step=10507 loss=1.307568 Training: 58%|█████▊ | 10507/18018 [22:52:21<16:43:42, 8.02s/step, loss=1.308, acc=0.697, acc_len=1.68, thru=14.1, I=18.2, T=32.4, wait=3.3s, pool=41, epoch=2/2][2026-10-19 12:22:42,936] loop.py:45 INFO TRAIN_STEP step=10508 loss=1.298296 Training: 58%|█████▊ | 10508/18018 [22:52:29<16:42:18, 8.01s/step, loss=1.298, acc=0.702, acc_len=1.68, thru=14.1, I=15.2, T=34.3, wait=4.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:49,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:22:51,171] loop.py:45 INFO TRAIN_STEP step=10509 loss=1.405329 Training: 58%|█████▊ | 10509/18018 [22:52:37<16:50:39, 8.08s/step, loss=1.405, acc=0.683, acc_len=1.62, thru=17.6, I=18.5, T=32.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:51,203] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:57,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:22:58,662] loop.py:380 INFO TIMING step=10510: step=3.828s data=1.717s compute=3.797s [fwd=2.726s bwd=1.057s opt=0.003s] dispatch=3.655s [2026-10-19 12:22:58,662] loop.py:45 INFO TRAIN_STEP step=10510 loss=1.290361 Training: 58%|█████▊ | 10510/18018 [22:52:44<16:28:31, 7.90s/step, loss=1.290, acc=0.705, acc_len=1.70, thru=16.7, I=20.4, T=33.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:22:58,988] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:05,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:23:05,520] loop.py:45 INFO TRAIN_STEP step=10511 loss=1.312663 Training: 58%|█████▊ | 10511/18018 [22:52:51<15:49:26, 7.59s/step, loss=1.313, acc=0.698, acc_len=1.68, thru=18.9, I=20.3, T=33.9, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:23:05,516] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10510: forward=2676.6ms backward=1041.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:05,577] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:12,192] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:23:13,093] loop.py:45 INFO TRAIN_STEP step=10512 loss=1.261419 Training: 58%|█████▊ | 10512/18018 [22:52:59<15:48:42, 7.58s/step, loss=1.261, acc=0.707, acc_len=1.71, thru=16.8, I=19.0, T=31.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:13,204] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 12:23:21,125] loop.py:45 INFO TRAIN_STEP step=10513 loss=1.310150 Training: 58%|█████▊ | 10513/18018 [22:53:07<16:05:17, 7.72s/step, loss=1.310, acc=0.700, acc_len=1.68, thru=14.0, I=16.9, T=32.9, wait=4.1s, pool=45, epoch=2/2][2026-10-19 12:23:29,302] loop.py:45 INFO TRAIN_STEP step=10514 loss=1.318481 Training: 58%|█████▊ | 10514/18018 [22:53:15<16:22:30, 7.86s/step, loss=1.318, acc=0.696, acc_len=1.66, thru=16.8, I=17.2, T=29.2, wait=3.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:36,885] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:23:38,580] loop.py:380 INFO TIMING step=10515: step=4.786s data=2.345s compute=4.751s [fwd=3.430s bwd=1.306s opt=0.003s] dispatch=4.483s [2026-10-19 12:23:38,580] loop.py:45 INFO TRAIN_STEP step=10515 loss=1.491012 Training: 58%|█████▊ | 10515/18018 [22:53:24<17:15:45, 8.28s/step, loss=1.491, acc=0.671, acc_len=1.58, thru=14.4, I=15.2, T=26.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:38,897] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:45,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:23:47,697] loop.py:45 INFO TRAIN_STEP step=10516 loss=1.330493 Training: 58%|█████▊ | 10516/18018 [22:53:33<17:46:53, 8.53s/step, loss=1.330, acc=0.696, acc_len=1.66, thru=14.6, I=18.7, T=31.7, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:23:47,693] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10515: forward=2794.3ms backward=1190.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:48,201] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:55,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:23:55,840] loop.py:45 INFO TRAIN_STEP step=10517 loss=1.400325 Training: 58%|█████▊ | 10517/18018 [22:53:41<17:32:09, 8.42s/step, loss=1.400, acc=0.681, acc_len=1.61, thru=16.8, I=18.7, T=34.0, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:23:55,872] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:01,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:24:03,536] loop.py:45 INFO TRAIN_STEP step=10518 loss=1.397981 Training: 58%|█████▊ | 10518/18018 [22:53:49<17:04:59, 8.20s/step, loss=1.398, acc=0.685, acc_len=1.63, thru=16.7, I=21.6, T=26.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:03,941] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:11,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:24:11,688] loop.py:45 INFO TRAIN_STEP step=10519 loss=1.291289 Training: 58%|█████▊ | 10519/18018 [22:53:57<17:03:01, 8.19s/step, loss=1.291, acc=0.704, acc_len=1.70, thru=16.5, I=17.4, T=30.3, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:11,780] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 12:24:18,158] loop.py:380 INFO TIMING step=10520: step=3.694s data=1.629s compute=3.658s [fwd=2.652s bwd=0.991s opt=0.003s] dispatch=2.767s [2026-10-19 12:24:18,158] loop.py:45 INFO TRAIN_STEP step=10520 loss=1.417412 Training: 58%|█████▊ | 10520/18018 [22:54:04<15:58:30, 7.67s/step, loss=1.417, acc=0.678, acc_len=1.61, thru=18.1, I=17.4, T=34.7, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:25,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:24:25,743] loop.py:45 INFO TRAIN_STEP step=10521 loss=1.300128 Training: 58%|█████▊ | 10521/18018 [22:54:11<15:55:12, 7.64s/step, loss=1.300, acc=0.703, acc_len=1.69, thru=18.1, I=17.8, T=32.8, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:24:25,738] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10520: forward=2793.5ms backward=1061.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:25,968] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:30,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:24:31,796] loop.py:45 INFO TRAIN_STEP step=10522 loss=1.219008 Training: 58%|█████▊ | 10522/18018 [22:54:17<14:55:28, 7.17s/step, loss=1.219, acc=0.712, acc_len=1.72, thru=19.6, I=22.0, T=34.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:32,224] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:24:38,885] loop.py:45 INFO TRAIN_STEP step=10523 loss=1.268755 Training: 58%|█████▊ | 10523/18018 [22:54:24<14:52:26, 7.14s/step, loss=1.269, acc=0.706, acc_len=1.70, thru=16.8, I=21.2, T=34.6, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:45,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:24:46,638] loop.py:45 INFO TRAIN_STEP step=10524 loss=1.185276 Training: 58%|█████▊ | 10524/18018 [22:54:32<15:15:09, 7.33s/step, loss=1.185, acc=0.722, acc_len=1.76, thru=16.4, I=17.5, T=29.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:46,802] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:51,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:24:53,238] loop.py:380 INFO TIMING step=10525: step=3.905s data=1.772s compute=3.876s [fwd=2.754s bwd=1.107s opt=0.003s] dispatch=2.684s [2026-10-19 12:24:53,238] loop.py:45 INFO TRAIN_STEP step=10525 loss=1.232240 Training: 58%|█████▊ | 10525/18018 [22:54:39<14:47:47, 7.11s/step, loss=1.232, acc=0.713, acc_len=1.73, thru=17.3, I=20.9, T=32.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:53,402] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:59,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:24:59,196] loop.py:45 INFO TRAIN_STEP step=10526 loss=1.334418 Training: 58%|█████▊ | 10526/18018 [22:54:45<14:04:30, 6.76s/step, loss=1.334, acc=0.693, acc_len=1.66, thru=20.1, I=23.2, T=36.3, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:24:59,192] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10525: forward=2497.2ms backward=981.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:24:59,507] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:25:06,758] loop.py:45 INFO TRAIN_STEP step=10527 loss=1.295202 Training: 58%|█████▊ | 10527/18018 [22:54:52<14:34:18, 7.00s/step, loss=1.295, acc=0.700, acc_len=1.69, thru=14.8, I=16.7, T=32.6, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:16,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:25:16,880] loop.py:45 INFO TRAIN_STEP step=10528 loss=1.263855 Training: 58%|█████▊ | 10528/18018 [22:55:02<16:31:03, 7.94s/step, loss=1.264, acc=0.706, acc_len=1.71, thru=14.2, I=14.6, T=32.5, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:17,184] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:23,047] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:25:23,969] loop.py:45 INFO TRAIN_STEP step=10529 loss=1.361307 Training: 58%|█████▊ | 10529/18018 [22:55:10<15:59:07, 7.68s/step, loss=1.361, acc=0.689, acc_len=1.64, thru=17.5, I=21.0, T=30.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:24,067] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:29,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:25:30,212] loop.py:380 INFO TIMING step=10530: step=3.767s data=1.619s compute=3.737s [fwd=2.662s bwd=1.059s opt=0.003s] dispatch=2.465s [2026-10-19 12:25:30,212] loop.py:45 INFO TRAIN_STEP step=10530 loss=1.359614 Training: 58%|█████▊ | 10530/18018 [22:55:16<15:04:56, 7.25s/step, loss=1.360, acc=0.689, acc_len=1.65, thru=18.2, I=22.5, T=34.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:30,241] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 12:25:37,172] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10530: forward=2537.2ms backward=996.5ms [2026-10-19 12:25:37,177] loop.py:45 INFO TRAIN_STEP step=10531 loss=1.395213 Training: 58%|█████▊ | 10531/18018 [22:55:23<14:54:02, 7.16s/step, loss=1.395, acc=0.683, acc_len=1.62, thru=16.9, I=18.2, T=35.6, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:44,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:25:46,749] loop.py:45 INFO TRAIN_STEP step=10532 loss=1.354174 Training: 58%|█████▊ | 10532/18018 [22:55:32<16:24:06, 7.89s/step, loss=1.354, acc=0.694, acc_len=1.66, thru=15.7, I=17.9, T=29.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:46,954] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:53,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:25:54,314] loop.py:45 INFO TRAIN_STEP step=10533 loss=1.281195 Training: 58%|█████▊ | 10533/18018 [22:55:40<16:11:56, 7.79s/step, loss=1.281, acc=0.701, acc_len=1.68, thru=13.7, I=18.4, T=29.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:25:54,441] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:00,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:26:01,530] loop.py:45 INFO TRAIN_STEP step=10534 loss=1.309969 Training: 58%|█████▊ | 10534/18018 [22:55:47<15:50:13, 7.62s/step, loss=1.310, acc=0.700, acc_len=1.68, thru=17.4, I=17.8, T=32.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:01,905] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:26:09,053] loop.py:380 INFO TIMING step=10535: step=4.408s data=2.064s compute=4.378s [fwd=3.062s bwd=1.301s opt=0.003s] dispatch=3.108s [2026-10-19 12:26:09,053] loop.py:45 INFO TRAIN_STEP step=10535 loss=1.377480 Training: 58%|█████▊ | 10535/18018 [22:55:55<15:46:32, 7.59s/step, loss=1.377, acc=0.685, acc_len=1.63, thru=15.9, I=16.4, T=29.0, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:14,140] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:26:16,070] loop.py:45 INFO TRAIN_STEP step=10536 loss=1.298841 Training: 58%|█████▊ | 10536/18018 [22:56:02<15:24:56, 7.42s/step, loss=1.299, acc=0.702, acc_len=1.68, thru=20.1, I=19.5, T=27.9, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:26:16,065] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10535: forward=3256.1ms backward=1273.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:16,147] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:22,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:26:24,582] loop.py:45 INFO TRAIN_STEP step=10537 loss=1.308624 Training: 58%|█████▊ | 10537/18018 [22:56:10<16:05:50, 7.75s/step, loss=1.309, acc=0.699, acc_len=1.68, thru=15.2, I=21.1, T=25.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:25,050] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:32,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:26:33,511] loop.py:45 INFO TRAIN_STEP step=10538 loss=1.256170 Training: 58%|█████▊ | 10538/18018 [22:56:19<16:50:01, 8.10s/step, loss=1.256, acc=0.706, acc_len=1.70, thru=15.1, I=17.5, T=29.8, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:33,764] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:41,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:26:42,495] loop.py:45 INFO TRAIN_STEP step=10539 loss=1.358274 Training: 58%|█████▊ | 10539/18018 [22:56:28<17:22:49, 8.37s/step, loss=1.358, acc=0.692, acc_len=1.66, thru=14.7, I=18.2, T=27.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:42,809] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:48,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:26:49,367] loop.py:380 INFO TIMING step=10540: step=3.873s data=1.674s compute=3.844s [fwd=2.744s bwd=1.085s opt=0.003s] dispatch=2.988s [2026-10-19 12:26:49,367] loop.py:45 INFO TRAIN_STEP step=10540 loss=1.212016 Training: 58%|█████▊ | 10540/18018 [22:56:35<16:26:46, 7.92s/step, loss=1.212, acc=0.719, acc_len=1.75, thru=16.5, I=18.6, T=33.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:49,836] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:54,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:26:55,952] loop.py:45 INFO TRAIN_STEP step=10541 loss=1.373542 Training: 59%|█████▊ | 10541/18018 [22:56:42<15:36:49, 7.52s/step, loss=1.374, acc=0.691, acc_len=1.65, thru=18.8, I=23.1, T=34.7, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:26:55,946] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10540: forward=2623.3ms backward=1018.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:26:56,173] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:27:02,033] loop.py:45 INFO TRAIN_STEP step=10542 loss=1.283589 Training: 59%|█████▊ | 10542/18018 [22:56:48<14:43:10, 7.09s/step, loss=1.284, acc=0.700, acc_len=1.68, thru=19.8, I=25.1, T=34.8, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:10,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:27:11,001] loop.py:45 INFO TRAIN_STEP step=10543 loss=1.289906 Training: 59%|█████▊ | 10543/18018 [22:56:57<15:53:13, 7.65s/step, loss=1.290, acc=0.703, acc_len=1.69, thru=14.8, I=15.3, T=33.0, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:11,054] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:20,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:27:20,664] loop.py:45 INFO TRAIN_STEP step=10544 loss=1.269304 Training: 59%|█████▊ | 10544/18018 [22:57:06<17:08:14, 8.25s/step, loss=1.269, acc=0.709, acc_len=1.71, thru=13.8, I=12.6, T=31.6, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:20,868] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:26,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:27:27,733] loop.py:380 INFO TIMING step=10545: step=3.746s data=1.644s compute=3.710s [fwd=2.687s bwd=1.009s opt=0.003s] dispatch=3.314s [2026-10-19 12:27:27,733] loop.py:45 INFO TRAIN_STEP step=10545 loss=1.434120 Training: 59%|█████▊ | 10545/18018 [22:57:13<16:23:48, 7.90s/step, loss=1.434, acc=0.677, acc_len=1.60, thru=18.8, I=17.7, T=34.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:27,923] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:27:34,467] loop.py:45 INFO TRAIN_STEP step=10546 loss=1.131047 Training: 59%|█████▊ | 10546/18018 [22:57:20<15:40:04, 7.55s/step, loss=1.131, acc=0.729, acc_len=1.78, thru=17.3, I=22.0, T=30.5, wait=2.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:27:34,462] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10545: forward=2697.1ms backward=1449.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:41,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:27:43,160] loop.py:45 INFO TRAIN_STEP step=10547 loss=1.311189 Training: 59%|█████▊ | 10547/18018 [22:57:29<16:22:46, 7.89s/step, loss=1.311, acc=0.701, acc_len=1.69, thru=16.1, I=16.0, T=30.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:43,655] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:48,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:27:49,598] loop.py:45 INFO TRAIN_STEP step=10548 loss=1.245492 Training: 59%|█████▊ | 10548/18018 [22:57:35<15:28:18, 7.46s/step, loss=1.245, acc=0.711, acc_len=1.71, thru=18.5, I=23.2, T=33.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:27:50,065] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:27:56,390] loop.py:45 INFO TRAIN_STEP step=10549 loss=1.345604 Training: 59%|█████▊ | 10549/18018 [22:57:42<15:03:22, 7.26s/step, loss=1.346, acc=0.693, acc_len=1.66, thru=15.3, I=19.3, T=33.7, wait=3.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:28:05,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:28:07,473] loop.py:380 INFO TIMING step=10550: step=4.813s data=2.202s compute=4.787s [fwd=3.372s bwd=1.400s opt=0.003s] dispatch=6.260s [2026-10-19 12:28:07,473] loop.py:45 INFO TRAIN_STEP step=10550 loss=1.257126 Training: 59%|█████▊ | 10550/18018 [22:57:53<17:26:03, 8.40s/step, loss=1.257, acc=0.707, acc_len=1.71, thru=13.5, I=15.4, T=26.6, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:28:07,627] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:28:13,802] loop.py:45 INFO TRAIN_STEP step=10551 loss=1.190012 Training: 59%|█████▊ | 10551/18018 [22:57:59<16:08:29, 7.78s/step, loss=1.190, acc=0.718, acc_len=1.74, thru=15.7, I=18.7, T=35.3, wait=2.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:28:13,796] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10550: forward=2534.4ms backward=1044.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:28:25,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:28:27,332] loop.py:45 INFO TRAIN_STEP step=10552 loss=1.330203 Training: 59%|█████▊ | 10552/18018 [22:58:13<19:42:54, 9.51s/step, loss=1.330, acc=0.697, acc_len=1.67, thru=10.4, I=11.7, T=24.2, wait=8.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:28:27,695] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:28:32,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:28:33,592] loop.py:45 INFO TRAIN_STEP step=10553 loss=1.297570 Training: 59%|█████▊ | 10553/18018 [22:58:19<17:41:38, 8.53s/step, loss=1.298, acc=0.699, acc_len=1.67, thru=17.2, I=21.3, T=36.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:28:33,879] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:28:41,080] loop.py:45 INFO TRAIN_STEP step=10554 loss=1.461470 Training: 59%|█████▊ | 10554/18018 [22:58:27<17:02:26, 8.22s/step, loss=1.461, acc=0.665, acc_len=1.57, thru=15.3, I=18.9, T=32.0, wait=3.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:28:48,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:28:49,478] loop.py:380 INFO TIMING step=10555: step=3.899s data=1.597s compute=3.869s [fwd=2.782s bwd=1.072s opt=0.003s] dispatch=4.489s [2026-10-19 12:28:49,478] loop.py:45 INFO TRAIN_STEP step=10555 loss=1.276747 Training: 59%|█████▊ | 10555/18018 [22:58:35<17:08:59, 8.27s/step, loss=1.277, acc=0.704, acc_len=1.69, thru=16.4, I=17.2, T=32.8, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:28:49,859] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:28:56,871] loop.py:45 INFO TRAIN_STEP step=10556 loss=1.262070 Training: 59%|█████▊ | 10556/18018 [22:58:42<16:36:02, 8.01s/step, loss=1.262, acc=0.712, acc_len=1.72, thru=17.3, I=20.1, T=34.7, wait=3.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:28:56,866] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10555: forward=2539.4ms backward=1100.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:04,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:29:05,066] loop.py:45 INFO TRAIN_STEP step=10557 loss=1.387607 Training: 59%|█████▊ | 10557/18018 [22:58:51<16:42:49, 8.06s/step, loss=1.388, acc=0.687, acc_len=1.63, thru=16.7, I=17.1, T=32.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:05,095] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:13,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:29:15,000] loop.py:45 INFO TRAIN_STEP step=10558 loss=1.409060 Training: 59%|█████▊ | 10558/18018 [22:59:01<17:52:22, 8.62s/step, loss=1.409, acc=0.685, acc_len=1.63, thru=12.9, I=14.4, T=29.7, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:15,400] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:20,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:29:21,193] loop.py:45 INFO TRAIN_STEP step=10559 loss=1.303952 Training: 59%|█████▊ | 10559/18018 [22:59:07<16:21:34, 7.90s/step, loss=1.304, acc=0.701, acc_len=1.69, thru=18.8, I=24.5, T=35.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:21,633] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:31,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:29:31,519] loop.py:380 INFO TIMING step=10560: step=4.320s data=1.953s compute=4.280s [fwd=3.020s bwd=1.245s opt=0.003s] dispatch=5.997s [2026-10-19 12:29:31,519] loop.py:45 INFO TRAIN_STEP step=10560 loss=1.338623 Training: 59%|█████▊ | 10560/18018 [22:59:17<17:52:02, 8.62s/step, loss=1.339, acc=0.694, acc_len=1.66, thru=12.9, I=13.5, T=29.6, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:31,784] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:29:37,726] loop.py:45 INFO TRAIN_STEP step=10561 loss=1.214131 Training: 59%|█████▊ | 10561/18018 [22:59:23<16:21:45, 7.90s/step, loss=1.214, acc=0.715, acc_len=1.73, thru=17.9, I=17.9, T=36.1, wait=2.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:29:37,721] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10560: forward=2502.6ms backward=988.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:44,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:29:44,545] loop.py:45 INFO TRAIN_STEP step=10562 loss=1.316925 Training: 59%|█████▊ | 10562/18018 [22:59:30<15:41:20, 7.58s/step, loss=1.317, acc=0.696, acc_len=1.67, thru=18.7, I=19.6, T=31.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:44,610] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:51,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:29:52,747] loop.py:45 INFO TRAIN_STEP step=10563 loss=1.460070 Training: 59%|█████▊ | 10563/18018 [22:59:38<16:04:39, 7.76s/step, loss=1.460, acc=0.675, acc_len=1.59, thru=16.1, I=17.6, T=33.5, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:29:52,996] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:30:00,178] loop.py:45 INFO TRAIN_STEP step=10564 loss=1.357182 Training: 59%|█████▊ | 10564/18018 [22:59:46<15:52:06, 7.66s/step, loss=1.357, acc=0.690, acc_len=1.64, thru=15.1, I=17.5, T=33.8, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:08,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:30:10,039] loop.py:380 INFO TIMING step=10565: step=4.582s data=1.975s compute=4.548s [fwd=3.151s bwd=1.381s opt=0.003s] dispatch=5.268s [2026-10-19 12:30:10,039] loop.py:45 INFO TRAIN_STEP step=10565 loss=1.271600 Training: 59%|█████▊ | 10565/18018 [22:59:56<17:13:53, 8.32s/step, loss=1.272, acc=0.707, acc_len=1.71, thru=14.7, I=16.1, T=27.9, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:10,179] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:16,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:30:16,585] loop.py:45 INFO TRAIN_STEP step=10566 loss=1.353599 Training: 59%|█████▊ | 10566/18018 [23:00:02<16:07:29, 7.79s/step, loss=1.354, acc=0.691, acc_len=1.65, thru=17.2, I=21.4, T=31.5, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:30:16,580] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10565: forward=2873.3ms backward=1143.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:16,733] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:23,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:30:24,806] loop.py:45 INFO TRAIN_STEP step=10567 loss=1.445768 Training: 59%|█████▊ | 10567/18018 [23:00:10<16:23:26, 7.92s/step, loss=1.446, acc=0.680, acc_len=1.61, thru=16.0, I=18.8, T=27.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:24,961] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:29,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:30:31,148] loop.py:45 INFO TRAIN_STEP step=10568 loss=1.340879 Training: 59%|█████▊ | 10568/18018 [23:00:17<15:24:29, 7.45s/step, loss=1.341, acc=0.694, acc_len=1.65, thru=18.4, I=25.6, T=33.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:31,340] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:30:38,020] loop.py:45 INFO TRAIN_STEP step=10569 loss=1.476179 Training: 59%|█████▊ | 10569/18018 [23:00:24<15:03:07, 7.27s/step, loss=1.476, acc=0.674, acc_len=1.59, thru=17.1, I=21.6, T=30.9, wait=2.7s, pool=61, epoch=2/2][2026-10-19 12:30:46,206] loop.py:380 INFO TIMING step=10570: step=3.837s data=1.591s compute=3.793s [fwd=2.596s bwd=1.182s opt=0.003s] dispatch=4.337s [2026-10-19 12:30:46,206] loop.py:45 INFO TRAIN_STEP step=10570 loss=1.251302 Training: 59%|█████▊ | 10570/18018 [23:00:32<15:36:56, 7.55s/step, loss=1.251, acc=0.710, acc_len=1.72, thru=16.9, I=15.7, T=33.4, wait=4.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:52,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:30:53,394] loop.py:45 INFO TRAIN_STEP step=10571 loss=1.424012 Training: 59%|█████▊ | 10571/18018 [23:00:39<15:23:22, 7.44s/step, loss=1.424, acc=0.677, acc_len=1.61, thru=18.1, I=18.1, T=36.0, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:30:53,389] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10570: forward=2514.6ms backward=991.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:30:53,447] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 12:30:59,333] loop.py:45 INFO TRAIN_STEP step=10572 loss=1.301249 Training: 59%|█████▊ | 10572/18018 [23:00:45<14:27:24, 6.99s/step, loss=1.301, acc=0.697, acc_len=1.67, thru=19.4, I=21.2, T=36.6, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:04,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:31:06,030] loop.py:45 INFO TRAIN_STEP step=10573 loss=1.314105 Training: 59%|█████▊ | 10573/18018 [23:00:52<14:16:16, 6.90s/step, loss=1.314, acc=0.698, acc_len=1.67, thru=18.9, I=21.8, T=34.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:06,485] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:13,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:31:13,918] loop.py:45 INFO TRAIN_STEP step=10574 loss=1.351540 Training: 59%|█████▊ | 10574/18018 [23:01:00<14:52:53, 7.20s/step, loss=1.352, acc=0.691, acc_len=1.65, thru=15.1, I=19.1, T=34.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:14,254] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:21,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:31:23,565] loop.py:380 INFO TIMING step=10575: step=4.838s data=2.288s compute=4.800s [fwd=3.380s bwd=1.404s opt=0.003s] dispatch=4.803s [2026-10-19 12:31:23,565] loop.py:45 INFO TRAIN_STEP step=10575 loss=1.363983 Training: 59%|█████▊ | 10575/18018 [23:01:09<16:24:03, 7.93s/step, loss=1.364, acc=0.690, acc_len=1.64, thru=13.8, I=18.3, T=26.5, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:24,029] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:33,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:31:33,346] loop.py:45 INFO TRAIN_STEP step=10576 loss=1.340080 Training: 59%|█████▊ | 10576/18018 [23:01:19<17:32:39, 8.49s/step, loss=1.340, acc=0.696, acc_len=1.67, thru=13.7, I=14.1, T=28.0, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:31:33,342] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10575: forward=3071.0ms backward=1364.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:33,582] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:42,753] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:31:43,595] loop.py:45 INFO TRAIN_STEP step=10577 loss=1.360082 Training: 59%|█████▊ | 10577/18018 [23:01:29<18:38:07, 9.02s/step, loss=1.360, acc=0.690, acc_len=1.65, thru=13.1, I=13.1, T=30.1, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:43,765] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:50,888] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:31:52,010] loop.py:45 INFO TRAIN_STEP step=10578 loss=1.306841 Training: 59%|█████▊ | 10578/18018 [23:01:38<18:15:36, 8.84s/step, loss=1.307, acc=0.698, acc_len=1.68, thru=15.2, I=17.6, T=29.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:31:52,397] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:01,401] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:32:02,253] loop.py:45 INFO TRAIN_STEP step=10579 loss=1.268518 Training: 59%|█████▊ | 10579/18018 [23:01:48<19:07:47, 9.26s/step, loss=1.269, acc=0.706, acc_len=1.70, thru=13.0, I=13.9, T=30.7, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:02,420] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:09,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:32:09,539] loop.py:380 INFO TIMING step=10580: step=4.373s data=2.000s compute=4.325s [fwd=3.087s bwd=1.223s opt=0.003s] dispatch=2.903s [2026-10-19 12:32:09,540] loop.py:45 INFO TRAIN_STEP step=10580 loss=1.277391 Training: 59%|█████▊ | 10580/18018 [23:01:55<17:54:18, 8.67s/step, loss=1.277, acc=0.705, acc_len=1.69, thru=16.1, I=18.7, T=29.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:09,849] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:16,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:32:16,372] loop.py:45 INFO TRAIN_STEP step=10581 loss=1.318665 Training: 59%|█████▊ | 10581/18018 [23:02:02<16:46:02, 8.12s/step, loss=1.319, acc=0.701, acc_len=1.69, thru=17.9, I=19.6, T=29.9, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:32:16,368] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10580: forward=3024.6ms backward=1197.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:16,589] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:32:23,034] loop.py:45 INFO TRAIN_STEP step=10582 loss=1.422741 Training: 59%|█████▊ | 10582/18018 [23:02:09<15:51:46, 7.68s/step, loss=1.423, acc=0.681, acc_len=1.62, thru=16.6, I=18.1, T=30.8, wait=2.5s, pool=57, epoch=2/2][2026-10-19 12:32:31,171] loop.py:45 INFO TRAIN_STEP step=10583 loss=1.345104 Training: 59%|█████▊ | 10583/18018 [23:02:17<16:08:36, 7.82s/step, loss=1.345, acc=0.692, acc_len=1.65, thru=15.4, I=14.4, T=31.0, wait=4.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:38,324] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:32:39,167] loop.py:45 INFO TRAIN_STEP step=10584 loss=1.292611 Training: 59%|█████▊ | 10584/18018 [23:02:25<16:15:12, 7.87s/step, loss=1.293, acc=0.701, acc_len=1.69, thru=17.5, I=18.6, T=28.3, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:39,350] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:43,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:32:45,260] loop.py:380 INFO TIMING step=10585: step=3.688s data=1.645s compute=3.647s [fwd=2.596s bwd=1.036s opt=0.003s] dispatch=2.397s [2026-10-19 12:32:45,261] loop.py:45 INFO TRAIN_STEP step=10585 loss=1.330285 Training: 59%|█████▊ | 10585/18018 [23:02:31<15:08:58, 7.34s/step, loss=1.330, acc=0.695, acc_len=1.67, thru=19.3, I=24.4, T=34.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:32:45,689] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:32:51,534] loop.py:45 INFO TRAIN_STEP step=10586 loss=1.310774 Training: 59%|█████▉ | 10586/18018 [23:02:37<14:29:20, 7.02s/step, loss=1.311, acc=0.698, acc_len=1.68, thru=18.0, I=23.7, T=35.3, wait=2.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:32:51,530] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10585: forward=2550.5ms backward=1028.5ms [2026-10-19 12:33:01,315] loop.py:45 INFO TRAIN_STEP step=10587 loss=1.311159 Training: 59%|█████▉ | 10587/18018 [23:02:47<16:11:50, 7.85s/step, loss=1.311, acc=0.696, acc_len=1.67, thru=14.0, I=14.0, T=30.2, wait=5.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:10,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:33:11,383] loop.py:45 INFO TRAIN_STEP step=10588 loss=1.502387 Training: 59%|█████▉ | 10588/18018 [23:02:57<17:34:12, 8.51s/step, loss=1.502, acc=0.669, acc_len=1.58, thru=14.4, I=13.7, T=29.5, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:11,879] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:19,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:33:20,425] loop.py:45 INFO TRAIN_STEP step=10589 loss=1.463520 Training: 59%|█████▉ | 10589/18018 [23:03:06<17:53:42, 8.67s/step, loss=1.464, acc=0.675, acc_len=1.60, thru=14.5, I=16.7, T=31.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:20,823] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:26,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:33:27,724] loop.py:380 INFO TIMING step=10590: step=3.851s data=1.692s compute=3.813s [fwd=2.767s bwd=1.030s opt=0.003s] dispatch=3.440s [2026-10-19 12:33:27,724] loop.py:45 INFO TRAIN_STEP step=10590 loss=1.218679 Training: 59%|█████▉ | 10590/18018 [23:03:13<17:02:39, 8.26s/step, loss=1.219, acc=0.714, acc_len=1.73, thru=16.1, I=19.5, T=33.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:27,901] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:32,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:33:34,497] loop.py:45 INFO TRAIN_STEP step=10591 loss=1.336164 Training: 59%|█████▉ | 10591/18018 [23:03:20<16:07:14, 7.81s/step, loss=1.336, acc=0.695, acc_len=1.67, thru=18.4, I=25.1, T=30.7, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:33:34,493] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10590: forward=2979.1ms backward=1133.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:34,566] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:39,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:33:41,007] loop.py:45 INFO TRAIN_STEP step=10592 loss=1.317528 Training: 59%|█████▉ | 10592/18018 [23:03:27<15:18:45, 7.42s/step, loss=1.318, acc=0.696, acc_len=1.66, thru=17.5, I=25.6, T=31.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:41,125] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 12:33:47,583] loop.py:45 INFO TRAIN_STEP step=10593 loss=1.201430 Training: 59%|█████▉ | 10593/18018 [23:03:33<14:47:02, 7.17s/step, loss=1.201, acc=0.718, acc_len=1.74, thru=16.0, I=19.3, T=31.8, wait=2.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:55,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:33:55,680] loop.py:45 INFO TRAIN_STEP step=10594 loss=1.283900 Training: 59%|█████▉ | 10594/18018 [23:03:41<15:21:25, 7.45s/step, loss=1.284, acc=0.703, acc_len=1.70, thru=17.6, I=15.7, T=36.5, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:33:56,113] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:34:00,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:34:02,240] loop.py:380 INFO TIMING step=10595: step=3.556s data=1.474s compute=3.522s [fwd=2.546s bwd=0.961s opt=0.003s] dispatch=2.996s [2026-10-19 12:34:02,240] loop.py:45 INFO TRAIN_STEP step=10595 loss=1.408570 Training: 59%|█████▉ | 10595/18018 [23:03:48<14:48:29, 7.18s/step, loss=1.409, acc=0.683, acc_len=1.62, thru=20.0, I=23.0, T=36.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:34:02,434] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:34:07,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:34:07,988] loop.py:45 INFO TRAIN_STEP step=10596 loss=1.135083 Training: 59%|█████▉ | 10596/18018 [23:03:54<13:55:05, 6.75s/step, loss=1.135, acc=0.728, acc_len=1.78, thru=22.1, I=25.3, T=37.1, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:34:07,983] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10595: forward=2391.3ms backward=1019.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:34:08,322] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:34:14,594] loop.py:45 INFO TRAIN_STEP step=10597 loss=1.256682 Training: 59%|█████▉ | 10597/18018 [23:04:00<13:49:38, 6.71s/step, loss=1.257, acc=0.711, acc_len=1.72, thru=18.2, I=21.6, T=32.8, wait=2.7s, pool=53, epoch=2/2][2026-10-19 12:34:27,115] loop.py:45 INFO TRAIN_STEP step=10598 loss=1.273535 Training: 59%|█████▉ | 10598/18018 [23:04:13<17:25:10, 8.45s/step, loss=1.274, acc=0.705, acc_len=1.69, thru=11.5, I=9.0, T=30.3, wait=8.3s, pool=45, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:34:36,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:34:38,460] loop.py:45 INFO TRAIN_STEP step=10599 loss=1.212316 Training: 59%|█████▉ | 10599/18018 [23:04:24<19:12:18, 9.32s/step, loss=1.212, acc=0.715, acc_len=1.73, thru=13.0, I=15.8, T=29.6, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:34:38,943] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 12:34:46,082] loop.py:380 INFO TIMING step=10600: step=4.110s data=1.943s compute=4.075s [fwd=2.904s bwd=1.156s opt=0.003s] dispatch=3.506s [2026-10-19 12:34:46,082] loop.py:45 INFO TRAIN_STEP step=10600 loss=1.380129 Training: 59%|█████▉ | 10600/18018 [23:04:32<18:09:20, 8.81s/step, loss=1.380, acc=0.687, acc_len=1.64, thru=10.8, I=22.0, T=31.1, wait=3.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:34:55,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:34:56,286] loop.py:45 INFO TRAIN_STEP step=10601 loss=1.343574 Training: 59%|█████▉ | 10601/18018 [23:04:42<19:00:53, 9.23s/step, loss=1.344, acc=0.692, acc_len=1.65, thru=17.5, I=12.8, T=35.0, wait=6.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:34:56,280] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10600: forward=2621.2ms backward=990.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:34:56,329] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:03,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:35:03,777] loop.py:45 INFO TRAIN_STEP step=10602 loss=1.368445 Training: 59%|█████▉ | 10602/18018 [23:04:49<17:56:10, 8.71s/step, loss=1.368, acc=0.688, acc_len=1.65, thru=16.6, I=18.1, T=32.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:04,082] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:10,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:35:11,861] loop.py:45 INFO TRAIN_STEP step=10603 loss=1.450894 Training: 59%|█████▉ | 10603/18018 [23:04:57<17:32:54, 8.52s/step, loss=1.451, acc=0.673, acc_len=1.58, thru=15.2, I=18.7, T=29.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:12,259] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:19,370] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:35:20,190] loop.py:45 INFO TRAIN_STEP step=10604 loss=1.265843 Training: 59%|█████▉ | 10604/18018 [23:05:06<17:25:45, 8.46s/step, loss=1.266, acc=0.703, acc_len=1.69, thru=13.3, I=17.6, T=30.0, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:20,378] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:26,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 12:35:28,127] loop.py:380 INFO TIMING step=10605: step=3.875s data=1.713s compute=3.839s [fwd=2.722s bwd=1.104s opt=0.003s] dispatch=4.051s [2026-10-19 12:35:28,127] loop.py:45 INFO TRAIN_STEP step=10605 loss=1.217518 Training: 59%|█████▉ | 10605/18018 [23:05:14<17:06:13, 8.31s/step, loss=1.218, acc=0.721, acc_len=1.76, thru=16.3, I=19.2, T=33.0, wait=4.1s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:28,455] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=5/64 [2026-10-19 12:35:34,606] loop.py:45 INFO TRAIN_STEP step=10606 loss=1.347142 Training: 59%|█████▉ | 10606/18018 [23:05:20<15:58:11, 7.76s/step, loss=1.347, acc=0.693, acc_len=1.65, thru=18.4, I=21.7, T=30.5, wait=2.3s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:35:34,602] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10605: forward=2917.6ms backward=1225.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:40,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:35:42,473] loop.py:45 INFO TRAIN_STEP step=10607 loss=1.315647 Training: 59%|█████▉ | 10607/18018 [23:05:28<16:02:14, 7.79s/step, loss=1.316, acc=0.696, acc_len=1.67, thru=16.0, I=19.5, T=29.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:42,773] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:49,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:35:49,127] loop.py:45 INFO TRAIN_STEP step=10608 loss=1.228297 Training: 59%|█████▉ | 10608/18018 [23:05:35<15:20:01, 7.45s/step, loss=1.228, acc=0.713, acc_len=1.73, thru=15.6, I=20.6, T=32.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:49,522] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:55,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:35:55,781] loop.py:45 INFO TRAIN_STEP step=10609 loss=1.403697 Training: 59%|█████▉ | 10609/18018 [23:05:41<14:50:21, 7.21s/step, loss=1.404, acc=0.683, acc_len=1.62, thru=18.4, I=20.2, T=34.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:35:56,084] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:36:01,624] loop.py:380 INFO TIMING step=10610: step=3.573s data=1.494s compute=3.549s [fwd=2.561s bwd=0.976s opt=0.003s] dispatch=2.263s [2026-10-19 12:36:01,625] loop.py:45 INFO TRAIN_STEP step=10610 loss=1.262294 Training: 59%|█████▉ | 10610/18018 [23:05:47<13:59:36, 6.80s/step, loss=1.262, acc=0.710, acc_len=1.72, thru=18.8, I=22.4, T=35.8, wait=2.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:08,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:36:09,273] loop.py:45 INFO TRAIN_STEP step=10611 loss=1.366644 Training: 59%|█████▉ | 10611/18018 [23:05:55<14:30:58, 7.06s/step, loss=1.367, acc=0.688, acc_len=1.64, thru=17.2, I=18.0, T=26.0, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:36:09,268] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10610: forward=3365.4ms backward=1493.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:09,455] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:36:16,027] loop.py:45 INFO TRAIN_STEP step=10612 loss=1.324504 Training: 59%|█████▉ | 10612/18018 [23:06:02<14:19:38, 6.96s/step, loss=1.325, acc=0.697, acc_len=1.67, thru=16.9, I=18.6, T=30.2, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:23,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:36:24,070] loop.py:45 INFO TRAIN_STEP step=10613 loss=1.358251 Training: 59%|█████▉ | 10613/18018 [23:06:10<14:59:29, 7.29s/step, loss=1.358, acc=0.694, acc_len=1.66, thru=15.9, I=15.9, T=32.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:24,569] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:36:30,756] loop.py:45 INFO TRAIN_STEP step=10614 loss=1.372069 Training: 59%|█████▉ | 10614/18018 [23:06:16<14:36:58, 7.11s/step, loss=1.372, acc=0.690, acc_len=1.64, thru=14.4, I=15.5, T=34.7, wait=3.0s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:41,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:36:42,549] loop.py:380 INFO TIMING step=10615: step=5.231s data=2.043s compute=5.195s [fwd=3.812s bwd=1.369s opt=0.003s] dispatch=6.555s [2026-10-19 12:36:42,549] loop.py:45 INFO TRAIN_STEP step=10615 loss=1.330213 Training: 59%|█████▉ | 10615/18018 [23:06:28<17:30:22, 8.51s/step, loss=1.330, acc=0.696, acc_len=1.67, thru=13.4, I=14.5, T=24.5, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:42,772] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:47,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:36:49,492] loop.py:45 INFO TRAIN_STEP step=10616 loss=1.309670 Training: 59%|█████▉ | 10616/18018 [23:06:35<16:32:11, 8.04s/step, loss=1.310, acc=0.697, acc_len=1.67, thru=16.2, I=20.7, T=27.9, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:36:49,487] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10615: forward=3406.3ms backward=1137.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:49,691] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:55,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:36:56,052] loop.py:45 INFO TRAIN_STEP step=10617 loss=1.412945 Training: 59%|█████▉ | 10617/18018 [23:06:42<15:37:12, 7.60s/step, loss=1.413, acc=0.678, acc_len=1.61, thru=17.6, I=22.4, T=31.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:36:56,317] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:02,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:37:03,404] loop.py:45 INFO TRAIN_STEP step=10618 loss=1.458611 Training: 59%|█████▉ | 10618/18018 [23:06:49<15:27:55, 7.52s/step, loss=1.459, acc=0.675, acc_len=1.60, thru=17.4, I=19.9, T=35.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:03,831] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:08,947] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:37:09,705] loop.py:45 INFO TRAIN_STEP step=10619 loss=1.272678 Training: 59%|█████▉ | 10619/18018 [23:06:55<14:42:36, 7.16s/step, loss=1.273, acc=0.707, acc_len=1.70, thru=20.1, I=23.1, T=32.0, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:09,951] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:16,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:37:16,804] loop.py:380 INFO TIMING step=10620: step=3.697s data=1.584s compute=3.668s [fwd=2.651s bwd=1.003s opt=0.003s] dispatch=3.391s [2026-10-19 12:37:16,804] loop.py:45 INFO TRAIN_STEP step=10620 loss=1.339332 Training: 59%|█████▉ | 10620/18018 [23:07:02<14:40:15, 7.14s/step, loss=1.339, acc=0.690, acc_len=1.65, thru=16.4, I=18.9, T=34.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:17,051] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:22,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:37:24,187] loop.py:45 INFO TRAIN_STEP step=10621 loss=1.310752 Training: 59%|█████▉ | 10621/18018 [23:07:10<14:49:09, 7.21s/step, loss=1.311, acc=0.698, acc_len=1.68, thru=16.7, I=21.3, T=26.9, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:37:24,182] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10620: forward=3084.3ms backward=1621.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:24,390] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:37:30,277] loop.py:45 INFO TRAIN_STEP step=10622 loss=1.240307 Training: 59%|█████▉ | 10622/18018 [23:07:16<14:07:41, 6.88s/step, loss=1.240, acc=0.716, acc_len=1.74, thru=16.1, I=19.6, T=34.2, wait=2.3s, pool=53, epoch=2/2][2026-10-19 12:37:39,442] loop.py:45 INFO TRAIN_STEP step=10623 loss=1.375984 Training: 59%|█████▉ | 10623/18018 [23:07:25<15:32:02, 7.56s/step, loss=1.376, acc=0.688, acc_len=1.64, thru=15.5, I=16.4, T=32.8, wait=5.2s, pool=61, epoch=2/2][2026-10-19 12:37:48,128] loop.py:45 INFO TRAIN_STEP step=10624 loss=1.300853 Training: 59%|█████▉ | 10624/18018 [23:07:34<16:13:24, 7.90s/step, loss=1.301, acc=0.698, acc_len=1.67, thru=14.0, I=14.4, T=32.9, wait=4.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:54,979] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:37:55,465] loop.py:380 INFO TIMING step=10625: step=3.615s data=1.592s compute=3.585s [fwd=2.616s bwd=0.954s opt=0.003s] dispatch=3.715s [2026-10-19 12:37:55,466] loop.py:45 INFO TRAIN_STEP step=10625 loss=1.408728 Training: 59%|█████▉ | 10625/18018 [23:07:41<15:52:37, 7.73s/step, loss=1.409, acc=0.684, acc_len=1.62, thru=18.3, I=17.4, T=35.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:37:55,497] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:02,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:38:03,040] loop.py:45 INFO TRAIN_STEP step=10626 loss=1.315634 Training: 59%|█████▉ | 10626/18018 [23:07:49<15:46:41, 7.68s/step, loss=1.316, acc=0.702, acc_len=1.69, thru=17.4, I=18.7, T=35.7, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:38:03,036] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10625: forward=2499.7ms backward=1036.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:03,302] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:09,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:38:10,327] loop.py:45 INFO TRAIN_STEP step=10627 loss=1.332120 Training: 59%|█████▉ | 10627/18018 [23:07:56<15:31:54, 7.57s/step, loss=1.332, acc=0.696, acc_len=1.67, thru=17.7, I=21.5, T=35.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:10,528] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:38:20,270] loop.py:45 INFO TRAIN_STEP step=10628 loss=1.333899 Training: 59%|█████▉ | 10628/18018 [23:08:06<16:59:31, 8.28s/step, loss=1.334, acc=0.694, acc_len=1.66, thru=12.7, I=13.1, T=25.5, wait=4.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:25,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:38:27,326] loop.py:45 INFO TRAIN_STEP step=10629 loss=1.292732 Training: 59%|█████▉ | 10629/18018 [23:08:13<16:14:22, 7.91s/step, loss=1.293, acc=0.703, acc_len=1.69, thru=16.9, I=17.6, T=30.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:27,397] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 [2026-10-19 12:38:35,034] loop.py:380 INFO TIMING step=10630: step=4.570s data=2.354s compute=4.538s [fwd=3.429s bwd=1.094s opt=0.003s] dispatch=3.128s [2026-10-19 12:38:35,034] loop.py:45 INFO TRAIN_STEP step=10630 loss=1.319961 Training: 59%|█████▉ | 10630/18018 [23:08:21<16:06:34, 7.85s/step, loss=1.320, acc=0.696, acc_len=1.67, thru=14.0, I=17.7, T=28.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:34,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:35,460] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:41,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:38:41,859] loop.py:45 INFO TRAIN_STEP step=10631 loss=1.380862 Training: 59%|█████▉ | 10631/18018 [23:08:27<15:28:37, 7.54s/step, loss=1.381, acc=0.685, acc_len=1.62, thru=17.3, I=19.3, T=33.8, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:38:41,856] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10630: forward=2802.2ms backward=939.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:42,311] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:48,631] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:38:50,798] loop.py:45 INFO TRAIN_STEP step=10632 loss=1.381322 Training: 59%|█████▉ | 10632/18018 [23:08:36<16:20:04, 7.96s/step, loss=1.381, acc=0.687, acc_len=1.64, thru=13.9, I=18.6, T=29.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:51,143] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:56,647] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:38:57,292] loop.py:45 INFO TRAIN_STEP step=10633 loss=1.323292 Training: 59%|█████▉ | 10633/18018 [23:08:43<15:25:49, 7.52s/step, loss=1.323, acc=0.696, acc_len=1.67, thru=16.2, I=23.7, T=35.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:38:57,653] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:03,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:39:04,761] loop.py:45 INFO TRAIN_STEP step=10634 loss=1.262456 Training: 59%|█████▉ | 10634/18018 [23:08:50<15:23:41, 7.51s/step, loss=1.262, acc=0.710, acc_len=1.72, thru=17.9, I=21.5, T=29.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:04,788] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:09,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:39:10,799] loop.py:380 INFO TIMING step=10635: step=3.987s data=1.829s compute=3.958s [fwd=2.795s bwd=1.148s opt=0.003s] dispatch=2.042s [2026-10-19 12:39:10,799] loop.py:45 INFO TRAIN_STEP step=10635 loss=1.301741 Training: 59%|█████▉ | 10635/18018 [23:08:56<14:29:23, 7.07s/step, loss=1.302, acc=0.698, acc_len=1.68, thru=18.4, I=23.8, T=32.1, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:11,120] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:15,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:39:17,031] loop.py:45 INFO TRAIN_STEP step=10636 loss=1.188769 Training: 59%|█████▉ | 10636/18018 [23:09:03<13:58:33, 6.82s/step, loss=1.189, acc=0.723, acc_len=1.76, thru=18.9, I=24.8, T=32.9, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:39:17,026] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10635: forward=2743.4ms backward=1102.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:17,174] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:21,733] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:39:23,507] loop.py:45 INFO TRAIN_STEP step=10637 loss=1.244039 Training: 59%|█████▉ | 10637/18018 [23:09:09<13:45:52, 6.71s/step, loss=1.244, acc=0.710, acc_len=1.72, thru=18.5, I=25.0, T=35.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:23,747] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:39:31,102] loop.py:45 INFO TRAIN_STEP step=10638 loss=1.309753 Training: 59%|█████▉ | 10638/18018 [23:09:17<14:18:15, 6.98s/step, loss=1.310, acc=0.696, acc_len=1.67, thru=14.4, I=18.5, T=35.4, wait=4.0s, pool=53, epoch=2/2][2026-10-19 12:39:38,292] loop.py:45 INFO TRAIN_STEP step=10639 loss=1.383313 Training: 59%|█████▉ | 10639/18018 [23:09:24<14:26:00, 7.04s/step, loss=1.383, acc=0.686, acc_len=1.64, thru=17.2, I=16.5, T=32.2, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:44,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:39:45,660] loop.py:380 INFO TIMING step=10640: step=3.759s data=1.608s compute=3.728s [fwd=2.631s bwd=1.083s opt=0.003s] dispatch=3.601s [2026-10-19 12:39:45,660] loop.py:45 INFO TRAIN_STEP step=10640 loss=1.319603 Training: 59%|█████▉ | 10640/18018 [23:09:31<14:37:53, 7.14s/step, loss=1.320, acc=0.697, acc_len=1.67, thru=17.9, I=18.6, T=34.1, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:45,687] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:53,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:39:54,023] loop.py:45 INFO TRAIN_STEP step=10641 loss=1.431258 Training: 59%|█████▉ | 10641/18018 [23:09:40<15:22:53, 7.51s/step, loss=1.431, acc=0.679, acc_len=1.61, thru=14.2, I=15.3, T=34.5, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:39:54,019] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10640: forward=2635.1ms backward=1006.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:39:54,245] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:39:59,462] loop.py:45 INFO TRAIN_STEP step=10642 loss=1.291585 Training: 59%|█████▉ | 10642/18018 [23:09:45<14:06:35, 6.89s/step, loss=1.292, acc=0.700, acc_len=1.68, thru=20.4, I=20.4, T=35.5, wait=1.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:07,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:40:07,375] loop.py:45 INFO TRAIN_STEP step=10643 loss=1.264733 Training: 59%|█████▉ | 10643/18018 [23:09:53<14:44:16, 7.19s/step, loss=1.265, acc=0.709, acc_len=1.71, thru=16.5, I=16.4, T=33.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:07,660] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:14,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:40:15,282] loop.py:45 INFO TRAIN_STEP step=10644 loss=1.243905 Training: 59%|█████▉ | 10644/18018 [23:10:01<15:10:31, 7.41s/step, loss=1.244, acc=0.709, acc_len=1.71, thru=15.7, I=18.3, T=30.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:15,560] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:21,450] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:40:21,798] loop.py:380 INFO TIMING step=10645: step=3.751s data=1.621s compute=3.708s [fwd=2.599s bwd=1.094s opt=0.003s] dispatch=2.754s [2026-10-19 12:40:21,798] loop.py:45 INFO TRAIN_STEP step=10645 loss=1.329373 Training: 59%|█████▉ | 10645/18018 [23:10:07<14:37:29, 7.14s/step, loss=1.329, acc=0.698, acc_len=1.68, thru=17.8, I=21.4, T=34.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:21,950] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:27,576] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:40:29,259] loop.py:45 INFO TRAIN_STEP step=10646 loss=1.207168 Training: 59%|█████▉ | 10646/18018 [23:10:15<14:49:10, 7.24s/step, loss=1.207, acc=0.719, acc_len=1.75, thru=16.9, I=20.9, T=28.6, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:40:29,254] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10645: forward=3252.3ms backward=1183.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:29,591] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:40:35,063] loop.py:45 INFO TRAIN_STEP step=10647 loss=1.414171 Training: 59%|█████▉ | 10647/18018 [23:10:21<13:56:09, 6.81s/step, loss=1.414, acc=0.678, acc_len=1.61, thru=18.3, I=23.1, T=34.7, wait=2.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:40,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:40:41,693] loop.py:45 INFO TRAIN_STEP step=10648 loss=1.295183 Training: 59%|█████▉ | 10648/18018 [23:10:27<13:49:37, 6.75s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=19.1, I=20.5, T=34.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:41,941] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:48,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:40:49,547] loop.py:45 INFO TRAIN_STEP step=10649 loss=1.284887 Training: 59%|█████▉ | 10649/18018 [23:10:35<14:30:02, 7.08s/step, loss=1.285, acc=0.701, acc_len=1.68, thru=16.5, I=21.8, T=35.0, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:49,754] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:54,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:40:55,274] loop.py:380 INFO TIMING step=10650: step=3.591s data=1.520s compute=3.563s [fwd=2.538s bwd=1.011s opt=0.003s] dispatch=2.126s [2026-10-19 12:40:55,275] loop.py:45 INFO TRAIN_STEP step=10650 loss=1.256154 Training: 59%|█████▉ | 10650/18018 [23:10:41<13:39:52, 6.68s/step, loss=1.256, acc=0.708, acc_len=1.70, thru=20.4, I=21.4, T=35.6, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:40:55,677] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:01,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:41:01,688] loop.py:45 INFO TRAIN_STEP step=10651 loss=1.315281 Training: 59%|█████▉ | 10651/18018 [23:10:47<13:30:02, 6.60s/step, loss=1.315, acc=0.699, acc_len=1.68, thru=19.5, I=22.3, T=36.5, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:41:01,683] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10650: forward=2484.6ms backward=969.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:02,191] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:06,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:41:08,540] loop.py:45 INFO TRAIN_STEP step=10652 loss=1.266599 Training: 59%|█████▉ | 10652/18018 [23:10:54<13:39:21, 6.67s/step, loss=1.267, acc=0.707, acc_len=1.70, thru=19.0, I=25.8, T=34.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:08,984] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 12:41:15,043] loop.py:45 INFO TRAIN_STEP step=10653 loss=1.189172 Training: 59%|█████▉ | 10653/18018 [23:11:01<13:32:56, 6.62s/step, loss=1.189, acc=0.719, acc_len=1.75, thru=16.9, I=29.2, T=35.8, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:25,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:41:26,035] loop.py:45 INFO TRAIN_STEP step=10654 loss=1.401177 Training: 59%|█████▉ | 10654/18018 [23:11:12<16:13:44, 7.93s/step, loss=1.401, acc=0.686, acc_len=1.64, thru=12.5, I=13.9, T=29.9, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:26,236] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:41:32,780] loop.py:380 INFO TIMING step=10655: step=3.756s data=1.584s compute=3.709s [fwd=2.638s bwd=1.055s opt=0.003s] dispatch=2.980s [2026-10-19 12:41:32,780] loop.py:45 INFO TRAIN_STEP step=10655 loss=1.312974 Training: 59%|█████▉ | 10655/18018 [23:11:18<15:29:51, 7.58s/step, loss=1.313, acc=0.696, acc_len=1.67, thru=16.4, I=16.6, T=34.1, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:39,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:41:41,914] loop.py:45 INFO TRAIN_STEP step=10656 loss=1.151176 Training: 59%|█████▉ | 10656/18018 [23:11:28<16:27:03, 8.04s/step, loss=1.151, acc=0.728, acc_len=1.78, thru=16.2, I=17.9, T=30.0, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:41,967] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=49/64 (TrainerActor pid=9310) [2026-10-19 12:41:41,909] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10655: forward=2977.4ms backward=1239.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:48,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:41:49,110] loop.py:45 INFO TRAIN_STEP step=10657 loss=1.469700 Training: 59%|█████▉ | 10657/18018 [23:11:35<15:55:40, 7.79s/step, loss=1.470, acc=0.671, acc_len=1.58, thru=15.8, I=22.6, T=33.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:49,584] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:55,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:41:55,803] loop.py:45 INFO TRAIN_STEP step=10658 loss=1.235650 Training: 59%|█████▉ | 10658/18018 [23:11:41<15:15:10, 7.46s/step, loss=1.236, acc=0.711, acc_len=1.72, thru=18.4, I=21.7, T=35.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:41:55,990] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:42:01,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:42:03,200] loop.py:45 INFO TRAIN_STEP step=10659 loss=1.255284 Training: 59%|█████▉ | 10659/18018 [23:11:49<15:12:39, 7.44s/step, loss=1.255, acc=0.709, acc_len=1.71, thru=17.3, I=20.0, T=33.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:42:03,458] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:42:10,170] loop.py:380 INFO TIMING step=10660: step=3.685s data=1.585s compute=3.656s [fwd=2.562s bwd=1.079s opt=0.003s] dispatch=3.277s [2026-10-19 12:42:10,170] loop.py:45 INFO TRAIN_STEP step=10660 loss=1.232190 Training: 59%|█████▉ | 10660/18018 [23:11:56<14:55:11, 7.30s/step, loss=1.232, acc=0.712, acc_len=1.72, thru=16.2, I=20.2, T=34.7, wait=3.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:42:18,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:42:18,763] loop.py:45 INFO TRAIN_STEP step=10661 loss=1.335907 Training: 59%|█████▉ | 10661/18018 [23:12:04<15:42:39, 7.69s/step, loss=1.336, acc=0.695, acc_len=1.66, thru=16.0, I=16.2, T=29.1, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:42:18,759] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10660: forward=3023.7ms backward=1328.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:42:19,250] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:42:25,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:42:26,770] loop.py:45 INFO TRAIN_STEP step=10662 loss=1.244768 Training: 59%|█████▉ | 10662/18018 [23:12:12<15:54:19, 7.78s/step, loss=1.245, acc=0.714, acc_len=1.74, thru=14.9, I=17.1, T=29.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:42:26,934] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:42:31,666] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:42:32,820] loop.py:45 INFO TRAIN_STEP step=10663 loss=1.222343 Training: 59%|█████▉ | 10663/18018 [23:12:18<14:50:25, 7.26s/step, loss=1.222, acc=0.714, acc_len=1.73, thru=20.3, I=23.8, T=32.6, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:42:33,176] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:42:39,808] loop.py:45 INFO TRAIN_STEP step=10664 loss=1.178426 Training: 59%|█████▉ | 10664/18018 [23:12:25<14:40:08, 7.18s/step, loss=1.178, acc=0.722, acc_len=1.76, thru=16.5, I=21.7, T=34.7, wait=3.3s, pool=53, epoch=2/2][2026-10-19 12:42:48,270] loop.py:380 INFO TIMING step=10665: step=3.951s data=1.729s compute=3.920s [fwd=2.832s bwd=1.073s opt=0.003s] dispatch=4.501s [2026-10-19 12:42:48,270] loop.py:45 INFO TRAIN_STEP step=10665 loss=1.374192 Training: 59%|█████▉ | 10665/18018 [23:12:34<15:27:09, 7.57s/step, loss=1.374, acc=0.686, acc_len=1.63, thru=15.1, I=15.6, T=32.4, wait=4.5s, pool=57, epoch=2/2][2026-10-19 12:42:55,109] loop.py:45 INFO TRAIN_STEP step=10666 loss=1.299964 Training: 59%|█████▉ | 10666/18018 [23:12:41<15:00:15, 7.35s/step, loss=1.300, acc=0.701, acc_len=1.68, thru=15.8, I=18.1, T=34.7, wait=3.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:42:55,105] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10665: forward=2514.0ms backward=1105.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:43:04,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:43:04,585] loop.py:45 INFO TRAIN_STEP step=10667 loss=1.197821 Training: 59%|█████▉ | 10667/18018 [23:12:50<16:18:19, 7.99s/step, loss=1.198, acc=0.719, acc_len=1.75, thru=15.3, I=14.9, T=32.4, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:43:04,668] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 12:43:11,456] loop.py:45 INFO TRAIN_STEP step=10668 loss=1.389760 Training: 59%|█████▉ | 10668/18018 [23:12:57<15:37:15, 7.65s/step, loss=1.390, acc=0.685, acc_len=1.63, thru=18.7, I=18.8, T=31.4, wait=2.8s, pool=61, epoch=2/2][2026-10-19 12:43:19,193] loop.py:45 INFO TRAIN_STEP step=10669 loss=1.323753 Training: 59%|█████▉ | 10669/18018 [23:13:05<15:40:19, 7.68s/step, loss=1.324, acc=0.699, acc_len=1.68, thru=14.2, I=14.1, T=33.9, wait=4.0s, pool=41, epoch=2/2][2026-10-19 12:43:27,282] loop.py:380 INFO TIMING step=10670: step=3.874s data=1.643s compute=3.839s [fwd=2.790s bwd=1.034s opt=0.003s] dispatch=4.206s [2026-10-19 12:43:27,282] loop.py:45 INFO TRAIN_STEP step=10670 loss=1.355239 Training: 59%|█████▉ | 10670/18018 [23:13:13<15:55:23, 7.80s/step, loss=1.355, acc=0.695, acc_len=1.66, thru=15.8, I=17.6, T=33.0, wait=4.2s, pool=49, epoch=2/2][2026-10-19 12:43:35,559] loop.py:45 INFO TRAIN_STEP step=10671 loss=1.184684 Training: 59%|█████▉ | 10671/18018 [23:13:21<16:12:43, 7.94s/step, loss=1.185, acc=0.723, acc_len=1.76, thru=14.7, I=13.9, T=32.7, wait=4.3s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:43:35,555] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10670: forward=2663.9ms backward=1195.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:43:43,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:43:44,187] loop.py:45 INFO TRAIN_STEP step=10672 loss=1.370396 Training: 59%|█████▉ | 10672/18018 [23:13:30<16:37:41, 8.15s/step, loss=1.370, acc=0.688, acc_len=1.64, thru=16.9, I=16.6, T=31.8, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:43:44,510] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:43:51,176] loop.py:45 INFO TRAIN_STEP step=10673 loss=1.322121 Training: 59%|█████▉ | 10673/18018 [23:13:37<15:54:54, 7.80s/step, loss=1.322, acc=0.698, acc_len=1.67, thru=16.2, I=18.8, T=33.3, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:43:57,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:43:58,209] loop.py:45 INFO TRAIN_STEP step=10674 loss=1.290593 Training: 59%|█████▉ | 10674/18018 [23:13:44<15:26:36, 7.57s/step, loss=1.291, acc=0.705, acc_len=1.70, thru=17.6, I=18.5, T=36.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:43:58,422] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:44:03,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:44:04,849] loop.py:380 INFO TIMING step=10675: step=3.701s data=1.581s compute=3.663s [fwd=2.660s bwd=0.989s opt=0.003s] dispatch=2.932s [2026-10-19 12:44:04,850] loop.py:45 INFO TRAIN_STEP step=10675 loss=1.378569 Training: 59%|█████▉ | 10675/18018 [23:13:50<14:52:21, 7.29s/step, loss=1.379, acc=0.684, acc_len=1.63, thru=19.1, I=21.4, T=34.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:44:04,968] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:44:13,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:44:13,645] loop.py:45 INFO TRAIN_STEP step=10676 loss=1.369153 Training: 59%|█████▉ | 10676/18018 [23:13:59<15:47:26, 7.74s/step, loss=1.369, acc=0.691, acc_len=1.65, thru=14.8, I=14.6, T=32.0, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:44:13,640] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10675: forward=2757.5ms backward=1204.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:44:13,888] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:44:20,670] loop.py:45 INFO TRAIN_STEP step=10677 loss=1.328999 Training: 59%|█████▉ | 10677/18018 [23:14:06<15:21:01, 7.53s/step, loss=1.329, acc=0.700, acc_len=1.68, thru=14.0, I=16.9, T=31.9, wait=3.0s, pool=37, epoch=2/2][2026-10-19 12:44:30,179] loop.py:45 INFO TRAIN_STEP step=10678 loss=1.339221 Training: 59%|█████▉ | 10678/18018 [23:14:16<16:33:36, 8.12s/step, loss=1.339, acc=0.695, acc_len=1.66, thru=14.1, I=13.7, T=30.9, wait=5.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:44:39,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:44:40,642] loop.py:45 INFO TRAIN_STEP step=10679 loss=1.429258 Training: 59%|█████▉ | 10679/18018 [23:14:26<17:59:18, 8.82s/step, loss=1.429, acc=0.680, acc_len=1.62, thru=14.9, I=14.5, T=34.6, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:44:41,142] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:44:46,422] loop.py:380 INFO TIMING step=10680: step=3.333s data=1.374s compute=3.302s [fwd=2.361s bwd=0.926s opt=0.003s] dispatch=2.438s [2026-10-19 12:44:46,422] loop.py:45 INFO TRAIN_STEP step=10680 loss=1.257036 Training: 59%|█████▉ | 10680/18018 [23:14:32<16:07:30, 7.91s/step, loss=1.257, acc=0.706, acc_len=1.70, thru=19.3, I=23.6, T=38.4, wait=2.4s, pool=53, epoch=2/2][2026-10-19 12:44:55,353] loop.py:45 INFO TRAIN_STEP step=10681 loss=1.251467 Training: 59%|█████▉ | 10681/18018 [23:14:41<16:44:47, 8.22s/step, loss=1.251, acc=0.708, acc_len=1.71, thru=14.6, I=13.6, T=34.0, wait=5.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:44:55,348] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10680: forward=2565.8ms backward=1133.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:03,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:45:03,696] loop.py:45 INFO TRAIN_STEP step=10682 loss=1.312593 Training: 59%|█████▉ | 10682/18018 [23:14:49<16:49:16, 8.25s/step, loss=1.313, acc=0.698, acc_len=1.67, thru=17.7, I=17.2, T=32.5, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:04,053] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:11,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:45:11,774] loop.py:45 INFO TRAIN_STEP step=10683 loss=1.312394 Training: 59%|█████▉ | 10683/18018 [23:14:57<16:42:42, 8.20s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=15.3, I=16.4, T=31.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:11,934] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 12:45:19,239] loop.py:45 INFO TRAIN_STEP step=10684 loss=1.282447 Training: 59%|█████▉ | 10684/18018 [23:15:05<16:15:29, 7.98s/step, loss=1.282, acc=0.706, acc_len=1.70, thru=16.2, I=17.1, T=33.5, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:27,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:45:28,206] loop.py:380 INFO TIMING step=10685: step=3.794s data=1.548s compute=3.760s [fwd=2.710s bwd=1.035s opt=0.003s] dispatch=5.165s [2026-10-19 12:45:28,207] loop.py:45 INFO TRAIN_STEP step=10685 loss=1.317317 Training: 59%|█████▉ | 10685/18018 [23:15:14<16:51:32, 8.28s/step, loss=1.317, acc=0.698, acc_len=1.67, thru=15.7, I=15.5, T=33.7, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:28,249] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 12:45:35,325] loop.py:45 INFO TRAIN_STEP step=10686 loss=1.316905 Training: 59%|█████▉ | 10686/18018 [23:15:21<16:08:56, 7.93s/step, loss=1.317, acc=0.697, acc_len=1.67, thru=17.5, I=20.9, T=29.7, wait=2.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:45:35,321] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10685: forward=3025.5ms backward=1240.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:41,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:45:43,354] loop.py:45 INFO TRAIN_STEP step=10687 loss=1.457595 Training: 59%|█████▉ | 10687/18018 [23:15:29<16:12:34, 7.96s/step, loss=1.458, acc=0.675, acc_len=1.59, thru=16.1, I=16.9, T=33.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:43,388] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:48,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:45:49,411] loop.py:45 INFO TRAIN_STEP step=10688 loss=1.405376 Training: 59%|█████▉ | 10688/18018 [23:15:35<15:02:39, 7.39s/step, loss=1.405, acc=0.680, acc_len=1.60, thru=20.1, I=24.3, T=34.5, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:45:49,823] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:45:55,889] loop.py:45 INFO TRAIN_STEP step=10689 loss=1.304593 Training: 59%|█████▉ | 10689/18018 [23:15:41<14:29:06, 7.12s/step, loss=1.305, acc=0.701, acc_len=1.68, thru=15.0, I=19.1, T=35.7, wait=2.9s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:06,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:46:06,697] loop.py:380 INFO TIMING step=10690: step=3.976s data=1.780s compute=3.943s [fwd=2.832s bwd=1.096s opt=0.003s] dispatch=6.823s [2026-10-19 12:46:06,697] loop.py:45 INFO TRAIN_STEP step=10690 loss=1.320970 Training: 59%|█████▉ | 10690/18018 [23:15:52<16:44:18, 8.22s/step, loss=1.321, acc=0.694, acc_len=1.66, thru=15.8, I=14.9, T=32.2, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:06,858] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:15,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:46:16,045] loop.py:45 INFO TRAIN_STEP step=10691 loss=1.426881 Training: 59%|█████▉ | 10691/18018 [23:16:02<17:25:21, 8.56s/step, loss=1.427, acc=0.676, acc_len=1.60, thru=13.6, I=13.4, T=34.6, wait=5.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:46:16,041] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10690: forward=2644.2ms backward=1009.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:16,388] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:23,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:46:23,586] loop.py:45 INFO TRAIN_STEP step=10692 loss=1.368760 Training: 59%|█████▉ | 10692/18018 [23:16:09<16:47:53, 8.25s/step, loss=1.369, acc=0.689, acc_len=1.65, thru=16.9, I=16.9, T=34.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:23,998] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:30,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:46:30,410] loop.py:45 INFO TRAIN_STEP step=10693 loss=1.336337 Training: 59%|█████▉ | 10693/18018 [23:16:16<15:55:23, 7.83s/step, loss=1.336, acc=0.690, acc_len=1.65, thru=18.6, I=18.9, T=32.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:30,611] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:46:36,583] loop.py:45 INFO TRAIN_STEP step=10694 loss=1.207144 Training: 59%|█████▉ | 10694/18018 [23:16:22<14:54:41, 7.33s/step, loss=1.207, acc=0.719, acc_len=1.75, thru=17.9, I=20.0, T=35.1, wait=2.5s, pool=49, epoch=2/2][2026-10-19 12:46:43,450] loop.py:380 INFO TIMING step=10695: step=3.492s data=1.490s compute=3.460s [fwd=2.469s bwd=0.977s opt=0.003s] dispatch=3.364s [2026-10-19 12:46:43,451] loop.py:45 INFO TRAIN_STEP step=10695 loss=1.265834 Training: 59%|█████▉ | 10695/18018 [23:16:29<14:37:46, 7.19s/step, loss=1.266, acc=0.704, acc_len=1.71, thru=17.7, I=18.5, T=36.7, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:48,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:46:50,358] loop.py:45 INFO TRAIN_STEP step=10696 loss=1.328876 Training: 59%|█████▉ | 10696/18018 [23:16:36<14:27:06, 7.11s/step, loss=1.329, acc=0.695, acc_len=1.66, thru=18.6, I=21.2, T=32.7, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:46:50,354] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10695: forward=2737.4ms backward=1134.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:50,853] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:57,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:46:57,922] loop.py:45 INFO TRAIN_STEP step=10697 loss=1.381063 Training: 59%|█████▉ | 10697/18018 [23:16:44<14:43:51, 7.24s/step, loss=1.381, acc=0.688, acc_len=1.64, thru=14.7, I=16.9, T=36.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:46:58,199] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:47:04,735] loop.py:45 INFO TRAIN_STEP step=10698 loss=1.321924 Training: 59%|█████▉ | 10698/18018 [23:16:50<14:27:56, 7.11s/step, loss=1.322, acc=0.695, acc_len=1.67, thru=18.5, I=17.9, T=34.5, wait=3.1s, pool=61, epoch=2/2][2026-10-19 12:47:16,122] loop.py:45 INFO TRAIN_STEP step=10699 loss=1.504430 Training: 59%|█████▉ | 10699/18018 [23:17:02<17:04:11, 8.40s/step, loss=1.504, acc=0.665, acc_len=1.56, thru=11.7, I=11.6, T=28.5, wait=6.9s, pool=45, epoch=2/2][2026-10-19 12:47:25,047] loop.py:380 INFO TIMING step=10700: step=3.883s data=1.697s compute=3.847s [fwd=2.766s bwd=1.065s opt=0.003s] dispatch=5.033s [2026-10-19 12:47:25,048] loop.py:45 INFO TRAIN_STEP step=10700 loss=1.221014 Training: 59%|█████▉ | 10700/18018 [23:17:11<17:23:24, 8.55s/step, loss=1.221, acc=0.716, acc_len=1.74, thru=16.5, I=13.2, T=33.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:47:25,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:47:25,513] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:47:30,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:47:31,556] loop.py:45 INFO TRAIN_STEP step=10701 loss=1.328614 Training: 59%|█████▉ | 10701/18018 [23:17:17<16:08:24, 7.94s/step, loss=1.329, acc=0.693, acc_len=1.66, thru=18.5, I=21.5, T=35.4, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:47:31,551] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10700: forward=2613.1ms backward=957.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:47:31,740] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 12:47:38,062] loop.py:45 INFO TRAIN_STEP step=10702 loss=1.295787 Training: 59%|█████▉ | 10702/18018 [23:17:24<15:15:46, 7.51s/step, loss=1.296, acc=0.699, acc_len=1.68, thru=18.2, I=23.4, T=33.7, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:47:47,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:47:48,563] loop.py:45 INFO TRAIN_STEP step=10703 loss=1.239939 Training: 59%|█████▉ | 10703/18018 [23:17:34<17:05:02, 8.41s/step, loss=1.240, acc=0.713, acc_len=1.72, thru=14.0, I=13.6, T=31.0, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:47:48,855] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:47:54,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:47:55,816] loop.py:45 INFO TRAIN_STEP step=10704 loss=1.301686 Training: 59%|█████▉ | 10704/18018 [23:17:41<16:22:37, 8.06s/step, loss=1.302, acc=0.702, acc_len=1.69, thru=16.7, I=19.5, T=35.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:47:55,984] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:01,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:48:01,318] loop.py:380 INFO TIMING step=10705: step=3.266s data=1.307s compute=3.235s [fwd=2.341s bwd=0.880s opt=0.003s] dispatch=2.228s [2026-10-19 12:48:01,318] loop.py:45 INFO TRAIN_STEP step=10705 loss=1.323510 Training: 59%|█████▉ | 10705/18018 [23:17:47<14:48:54, 7.29s/step, loss=1.324, acc=0.697, acc_len=1.67, thru=21.7, I=23.4, T=39.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:01,592] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:08,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 12:48:09,337] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10705: forward=2638.3ms backward=1137.0ms [2026-10-19 12:48:09,342] loop.py:45 INFO TRAIN_STEP step=10706 loss=1.349726 Training: 59%|█████▉ | 10706/18018 [23:17:55<15:15:36, 7.51s/step, loss=1.350, acc=0.692, acc_len=1.66, thru=17.1, I=17.0, T=33.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:09,419] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 12:48:14,962] loop.py:45 INFO TRAIN_STEP step=10707 loss=1.112778 Training: 59%|█████▉ | 10707/18018 [23:18:01<14:06:23, 6.95s/step, loss=1.113, acc=0.734, acc_len=1.80, thru=19.2, I=20.1, T=37.1, wait=2.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:23,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:48:23,338] loop.py:45 INFO TRAIN_STEP step=10708 loss=1.374938 Training: 59%|█████▉ | 10708/18018 [23:18:09<14:58:23, 7.37s/step, loss=1.375, acc=0.688, acc_len=1.64, thru=16.1, I=17.2, T=28.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:23,538] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:29,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:48:29,388] loop.py:45 INFO TRAIN_STEP step=10709 loss=1.338204 Training: 59%|█████▉ | 10709/18018 [23:18:15<14:09:53, 6.98s/step, loss=1.338, acc=0.694, acc_len=1.66, thru=18.5, I=19.4, T=34.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:29,552] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:48:37,769] loop.py:380 INFO TIMING step=10710: step=3.874s data=1.731s compute=3.832s [fwd=2.728s bwd=1.091s opt=0.003s] dispatch=4.498s [2026-10-19 12:48:37,770] loop.py:45 INFO TRAIN_STEP step=10710 loss=1.409139 Training: 59%|█████▉ | 10710/18018 [23:18:23<15:01:03, 7.40s/step, loss=1.409, acc=0.680, acc_len=1.62, thru=13.7, I=15.1, T=33.0, wait=4.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:44,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:48:44,984] loop.py:45 INFO TRAIN_STEP step=10711 loss=1.389335 Training: 59%|█████▉ | 10711/18018 [23:18:31<14:54:17, 7.34s/step, loss=1.389, acc=0.686, acc_len=1.64, thru=18.3, I=18.2, T=33.3, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:48:44,980] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10710: forward=2660.2ms backward=1115.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:45,165] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:48:52,968] loop.py:45 INFO TRAIN_STEP step=10712 loss=1.419128 Training: 59%|█████▉ | 10712/18018 [23:18:39<15:17:31, 7.54s/step, loss=1.419, acc=0.685, acc_len=1.63, thru=15.6, I=15.7, T=36.4, wait=4.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:48:59,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:49:00,862] loop.py:45 INFO TRAIN_STEP step=10713 loss=1.346318 Training: 59%|█████▉ | 10713/18018 [23:18:46<15:30:33, 7.64s/step, loss=1.346, acc=0.688, acc_len=1.65, thru=19.1, I=18.6, T=35.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:49:01,108] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:49:07,916] loop.py:45 INFO TRAIN_STEP step=10714 loss=1.348052 Training: 59%|█████▉ | 10714/18018 [23:18:54<15:08:56, 7.47s/step, loss=1.348, acc=0.692, acc_len=1.65, thru=17.8, I=23.4, T=31.9, wait=3.0s, pool=61, epoch=2/2][2026-10-19 12:49:17,382] loop.py:380 INFO TIMING step=10715: step=4.004s data=1.718s compute=3.972s [fwd=2.847s bwd=1.112s opt=0.003s] dispatch=5.451s [2026-10-19 12:49:17,382] loop.py:45 INFO TRAIN_STEP step=10715 loss=1.348996 Training: 59%|█████▉ | 10715/18018 [23:19:03<16:21:42, 8.07s/step, loss=1.349, acc=0.690, acc_len=1.65, thru=10.4, I=11.8, T=32.0, wait=5.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:49:26,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:49:27,860] loop.py:45 INFO TRAIN_STEP step=10716 loss=1.256475 Training: 59%|█████▉ | 10716/18018 [23:19:13<17:49:41, 8.79s/step, loss=1.256, acc=0.712, acc_len=1.72, thru=15.7, I=16.1, T=28.4, wait=6.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:49:27,855] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10715: forward=3166.7ms backward=1306.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:49:28,286] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:49:33,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:49:34,640] loop.py:45 INFO TRAIN_STEP step=10717 loss=1.461913 Training: 59%|█████▉ | 10717/18018 [23:19:20<16:36:13, 8.19s/step, loss=1.462, acc=0.671, acc_len=1.58, thru=18.1, I=20.5, T=33.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:49:34,892] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:49:41,828] loop.py:45 INFO TRAIN_STEP step=10718 loss=1.409054 Training: 59%|█████▉ | 10718/18018 [23:19:27<15:59:31, 7.89s/step, loss=1.409, acc=0.685, acc_len=1.63, thru=15.7, I=18.3, T=33.2, wait=3.3s, pool=53, epoch=2/2][2026-10-19 12:49:49,732] loop.py:45 INFO TRAIN_STEP step=10719 loss=1.328504 Training: 59%|█████▉ | 10719/18018 [23:19:35<16:00:01, 7.89s/step, loss=1.329, acc=0.697, acc_len=1.67, thru=14.6, I=15.6, T=33.8, wait=4.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:49:55,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:49:57,287] loop.py:380 INFO TIMING step=10720: step=3.581s data=1.488s compute=3.553s [fwd=2.518s bwd=1.020s opt=0.003s] dispatch=3.968s [2026-10-19 12:49:57,287] loop.py:45 INFO TRAIN_STEP step=10720 loss=1.260207 Training: 59%|█████▉ | 10720/18018 [23:19:43<15:47:42, 7.79s/step, loss=1.260, acc=0.705, acc_len=1.69, thru=17.4, I=19.2, T=35.7, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:49:57,640] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:50:05,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:50:05,460] loop.py:45 INFO TRAIN_STEP step=10721 loss=1.302031 Training: 60%|█████▉ | 10721/18018 [23:19:51<16:01:26, 7.91s/step, loss=1.302, acc=0.702, acc_len=1.69, thru=13.8, I=16.7, T=34.7, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:50:05,456] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10720: forward=2583.9ms backward=1062.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:50:05,600] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 12:50:14,331] loop.py:45 INFO TRAIN_STEP step=10722 loss=1.391744 Training: 60%|█████▉ | 10722/18018 [23:20:00<16:36:36, 8.20s/step, loss=1.392, acc=0.684, acc_len=1.63, thru=12.2, I=15.8, T=34.6, wait=5.2s, pool=33, epoch=2/2][2026-10-19 12:50:21,737] loop.py:45 INFO TRAIN_STEP step=10723 loss=1.390466 Training: 60%|█████▉ | 10723/18018 [23:20:07<16:07:38, 7.96s/step, loss=1.390, acc=0.687, acc_len=1.64, thru=17.7, I=15.7, T=35.7, wait=3.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:50:26,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:50:28,376] loop.py:45 INFO TRAIN_STEP step=10724 loss=1.272175 Training: 60%|█████▉ | 10724/18018 [23:20:14<15:19:25, 7.56s/step, loss=1.272, acc=0.703, acc_len=1.69, thru=20.4, I=23.0, T=35.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:50:28,681] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:50:35,019] loop.py:380 INFO TIMING step=10725: step=3.281s data=1.332s compute=3.246s [fwd=2.307s bwd=0.924s opt=0.003s] dispatch=3.352s [2026-10-19 12:50:35,020] loop.py:45 INFO TRAIN_STEP step=10725 loss=1.345260 Training: 60%|█████▉ | 10725/18018 [23:20:21<14:45:37, 7.29s/step, loss=1.345, acc=0.691, acc_len=1.65, thru=14.9, I=21.7, T=39.0, wait=3.4s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:50:44,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:50:46,377] loop.py:45 INFO TRAIN_STEP step=10726 loss=1.330045 Training: 60%|█████▉ | 10726/18018 [23:20:32<17:13:58, 8.51s/step, loss=1.330, acc=0.700, acc_len=1.67, thru=14.7, I=16.2, T=31.4, wait=7.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:50:46,373] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10725: forward=2847.4ms backward=1184.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:50:46,501] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:50:52,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:50:54,375] loop.py:45 INFO TRAIN_STEP step=10727 loss=1.237271 Training: 60%|█████▉ | 10727/18018 [23:20:40<16:55:19, 8.36s/step, loss=1.237, acc=0.714, acc_len=1.73, thru=15.7, I=22.2, T=32.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:50:54,687] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:51:02,164] loop.py:45 INFO TRAIN_STEP step=10728 loss=1.334502 Training: 60%|█████▉ | 10728/18018 [23:20:48<16:34:33, 8.19s/step, loss=1.335, acc=0.695, acc_len=1.66, thru=12.9, I=18.8, T=35.9, wait=4.2s, pool=49, epoch=2/2][2026-10-19 12:51:10,616] loop.py:45 INFO TRAIN_STEP step=10729 loss=1.286063 Training: 60%|█████▉ | 10729/18018 [23:20:56<16:44:07, 8.27s/step, loss=1.286, acc=0.703, acc_len=1.69, thru=14.9, I=14.5, T=31.1, wait=4.3s, pool=57, epoch=2/2][2026-10-19 12:51:18,332] loop.py:380 INFO TIMING step=10730: step=3.763s data=1.666s compute=3.715s [fwd=2.628s bwd=1.073s opt=0.003s] dispatch=3.941s [2026-10-19 12:51:18,332] loop.py:45 INFO TRAIN_STEP step=10730 loss=1.424491 Training: 60%|█████▉ | 10730/18018 [23:21:04<16:23:57, 8.10s/step, loss=1.424, acc=0.680, acc_len=1.61, thru=14.1, I=15.1, T=34.0, wait=3.9s, pool=41, epoch=2/2][2026-10-19 12:51:26,959] loop.py:45 INFO TRAIN_STEP step=10731 loss=1.426047 Training: 60%|█████▉ | 10731/18018 [23:21:13<16:42:59, 8.26s/step, loss=1.426, acc=0.680, acc_len=1.61, thru=16.1, I=14.8, T=27.9, wait=4.0s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:51:26,954] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10730: forward=3505.8ms backward=1029.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:51:33,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:51:33,751] loop.py:45 INFO TRAIN_STEP step=10732 loss=1.287189 Training: 60%|█████▉ | 10732/18018 [23:21:19<15:49:25, 7.82s/step, loss=1.287, acc=0.703, acc_len=1.69, thru=18.2, I=18.9, T=36.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:51:34,134] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:51:40,887] loop.py:45 INFO TRAIN_STEP step=10733 loss=1.219298 Training: 60%|█████▉ | 10733/18018 [23:21:26<15:24:23, 7.61s/step, loss=1.219, acc=0.716, acc_len=1.74, thru=14.1, I=16.6, T=36.1, wait=3.6s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:51:48,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:51:49,792] loop.py:45 INFO TRAIN_STEP step=10734 loss=1.325188 Training: 60%|█████▉ | 10734/18018 [23:21:35<16:11:17, 8.00s/step, loss=1.325, acc=0.697, acc_len=1.67, thru=17.5, I=17.7, T=33.6, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:51:50,181] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:51:57,332] loop.py:380 INFO TIMING step=10735: step=3.447s data=1.377s compute=3.415s [fwd=2.418s bwd=0.983s opt=0.003s] dispatch=4.086s [2026-10-19 12:51:57,332] loop.py:45 INFO TRAIN_STEP step=10735 loss=1.295635 Training: 60%|█████▉ | 10735/18018 [23:21:43<15:54:32, 7.86s/step, loss=1.296, acc=0.704, acc_len=1.70, thru=12.9, I=14.7, T=37.1, wait=4.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:05,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:52:06,112] loop.py:45 INFO TRAIN_STEP step=10736 loss=1.241727 Training: 60%|█████▉ | 10736/18018 [23:21:52<16:27:39, 8.14s/step, loss=1.242, acc=0.712, acc_len=1.72, thru=16.1, I=16.5, T=33.4, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:52:06,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10735: forward=2703.8ms backward=1067.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:06,197] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:11,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:52:12,491] loop.py:45 INFO TRAIN_STEP step=10737 loss=1.252893 Training: 60%|█████▉ | 10737/18018 [23:21:58<15:23:29, 7.61s/step, loss=1.253, acc=0.709, acc_len=1.71, thru=19.5, I=20.2, T=36.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:12,997] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:19,928] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:52:20,734] loop.py:45 INFO TRAIN_STEP step=10738 loss=1.332523 Training: 60%|█████▉ | 10738/18018 [23:22:06<15:46:26, 7.80s/step, loss=1.333, acc=0.696, acc_len=1.66, thru=14.6, I=16.3, T=35.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:20,935] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:26,741] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:52:27,318] loop.py:45 INFO TRAIN_STEP step=10739 loss=1.250925 Training: 60%|█████▉ | 10739/18018 [23:22:13<15:01:55, 7.43s/step, loss=1.251, acc=0.709, acc_len=1.71, thru=18.9, I=21.8, T=37.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:27,745] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:32,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:52:33,291] loop.py:380 INFO TIMING step=10740: step=3.366s data=1.374s compute=3.339s [fwd=2.404s bwd=0.921s opt=0.003s] dispatch=2.600s [2026-10-19 12:52:33,292] loop.py:45 INFO TRAIN_STEP step=10740 loss=1.211589 Training: 60%|█████▉ | 10740/18018 [23:22:19<14:08:42, 7.00s/step, loss=1.212, acc=0.715, acc_len=1.73, thru=20.5, I=23.1, T=38.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:33,468] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:40,587] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:52:40,947] loop.py:45 INFO TRAIN_STEP step=10741 loss=1.309856 Training: 60%|█████▉ | 10741/18018 [23:22:27<14:32:30, 7.19s/step, loss=1.310, acc=0.699, acc_len=1.68, thru=17.4, I=18.2, T=35.6, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:52:40,942] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10740: forward=2515.8ms backward=1034.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:41,093] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:45,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:52:47,901] loop.py:45 INFO TRAIN_STEP step=10742 loss=1.399166 Training: 60%|█████▉ | 10742/18018 [23:22:33<14:23:39, 7.12s/step, loss=1.399, acc=0.683, acc_len=1.62, thru=18.1, I=22.4, T=28.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:52:47,985] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=33/64 [2026-10-19 12:52:53,860] loop.py:45 INFO TRAIN_STEP step=10743 loss=1.334684 Training: 60%|█████▉ | 10743/18018 [23:22:39<13:41:19, 6.77s/step, loss=1.335, acc=0.694, acc_len=1.66, thru=14.8, I=25.6, T=37.1, wait=2.5s, pool=29, epoch=2/2][2026-10-19 12:53:04,505] loop.py:45 INFO TRAIN_STEP step=10744 loss=1.376533 Training: 60%|█████▉ | 10744/18018 [23:22:50<16:01:52, 7.93s/step, loss=1.377, acc=0.689, acc_len=1.65, thru=14.2, I=13.8, T=32.4, wait=6.7s, pool=53, epoch=2/2][2026-10-19 12:53:10,509] loop.py:380 INFO TIMING step=10745: step=3.453s data=1.480s compute=3.414s [fwd=2.429s bwd=0.971s opt=0.003s] dispatch=2.544s [2026-10-19 12:53:10,509] loop.py:45 INFO TRAIN_STEP step=10745 loss=1.331075 Training: 60%|█████▉ | 10745/18018 [23:22:56<14:51:36, 7.36s/step, loss=1.331, acc=0.695, acc_len=1.66, thru=18.0, I=16.4, T=37.1, wait=2.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:20,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:53:21,040] loop.py:45 INFO TRAIN_STEP step=10746 loss=1.256491 Training: 60%|█████▉ | 10746/18018 [23:23:07<16:46:59, 8.31s/step, loss=1.256, acc=0.713, acc_len=1.73, thru=14.1, I=13.6, T=33.0, wait=6.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:53:21,036] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10745: forward=2763.6ms backward=1067.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:21,173] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:28,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:53:28,906] loop.py:45 INFO TRAIN_STEP step=10747 loss=1.280312 Training: 60%|█████▉ | 10747/18018 [23:23:14<16:30:46, 8.18s/step, loss=1.280, acc=0.704, acc_len=1.70, thru=15.7, I=17.4, T=35.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:29,155] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:33,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:53:34,071] loop.py:45 INFO TRAIN_STEP step=10748 loss=1.261953 Training: 60%|█████▉ | 10748/18018 [23:23:20<14:41:06, 7.27s/step, loss=1.262, acc=0.706, acc_len=1.70, thru=22.5, I=27.9, T=38.5, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:34,515] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:41,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:53:42,000] loop.py:45 INFO TRAIN_STEP step=10749 loss=1.305434 Training: 60%|█████▉ | 10749/18018 [23:23:28<15:04:56, 7.47s/step, loss=1.305, acc=0.696, acc_len=1.67, thru=16.5, I=19.4, T=37.3, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:42,288] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:47,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:53:48,944] loop.py:380 INFO TIMING step=10750: step=3.503s data=1.441s compute=3.472s [fwd=2.501s bwd=0.956s opt=0.003s] dispatch=3.431s [2026-10-19 12:53:48,944] loop.py:45 INFO TRAIN_STEP step=10750 loss=1.368419 Training: 60%|█████▉ | 10750/18018 [23:23:35<14:45:38, 7.31s/step, loss=1.368, acc=0.689, acc_len=1.64, thru=18.0, I=21.2, T=36.5, wait=3.4s, pool=65, epoch=2/2][2026-10-19 12:53:48,946] loop.py:419 INFO Saving checkpoint at step 10750... (TrainerActor pid=9310) [2026-10-19 12:53:52,371] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0010751 (TrainerActor pid=9310) [2026-10-19 12:53:52,377] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0010001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:52,879] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:53,887] inference_manager.py:408 INFO Pool capacity available after 6.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:53:59,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 12:53:52,372] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0010751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 12:53:53,503] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0010001 [2026-10-19 12:53:59,884] loop.py:45 INFO TRAIN_STEP step=10751 loss=1.276662 Training: 60%|█████▉ | 10751/18018 [23:23:45<16:57:23, 8.40s/step, loss=1.277, acc=0.704, acc_len=1.70, thru=21.4, I=22.4, T=38.6, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:53:59,879] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10750: forward=2318.5ms backward=939.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:00,027] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 12:54:07,834] loop.py:45 INFO TRAIN_STEP step=10752 loss=1.202668 Training: 60%|█████▉ | 10752/18018 [23:23:53<16:41:01, 8.27s/step, loss=1.203, acc=0.716, acc_len=1.73, thru=12.3, I=16.4, T=35.2, wait=4.3s, pool=21, epoch=2/2](TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 12:54:08,043] eagle_store.py:517 WARNING batch_get_buffer missing keys (k344061-ab32d76b_hs, k344061-ab32d76b_ids, k344061-ab32d76b_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:17,395] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 12:54:08,050] eagle_store.py:517 WARNING batch_get_buffer missing keys (k344062-9503a4f6_hs, k344062-9503a4f6_ids, k344062-9503a4f6_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 12:54:19,908] loop.py:45 INFO TRAIN_STEP step=10753 loss=1.474092 Training: 60%|█████▉ | 10753/18018 [23:24:05<18:59:09, 9.41s/step, loss=1.474, acc=0.672, acc_len=1.58, thru=13.0, I=16.7, T=26.9, wait=7.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:20,414] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:26,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:54:26,425] loop.py:45 INFO TRAIN_STEP step=10754 loss=1.138927 Training: 60%|█████▉ | 10754/18018 [23:24:12<17:13:53, 8.54s/step, loss=1.139, acc=0.728, acc_len=1.78, thru=16.9, I=22.6, T=36.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:26,869] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:33,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:54:33,972] loop.py:380 INFO TIMING step=10755: step=3.474s data=1.439s compute=3.442s [fwd=2.454s bwd=0.973s opt=0.003s] dispatch=4.065s [2026-10-19 12:54:33,972] loop.py:45 INFO TRAIN_STEP step=10755 loss=1.473750 Training: 60%|█████▉ | 10755/18018 [23:24:20<16:37:40, 8.24s/step, loss=1.474, acc=0.671, acc_len=1.58, thru=18.7, I=19.2, T=36.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:34,094] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:38,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:54:39,390] loop.py:45 INFO TRAIN_STEP step=10756 loss=1.357836 Training: 60%|█████▉ | 10756/18018 [23:24:25<14:55:07, 7.40s/step, loss=1.358, acc=0.687, acc_len=1.65, thru=22.1, I=27.0, T=36.1, wait=1.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:54:39,385] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10755: forward=2538.3ms backward=966.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:39,731] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:45,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:54:45,685] loop.py:45 INFO TRAIN_STEP step=10757 loss=1.238561 Training: 60%|█████▉ | 10757/18018 [23:24:31<14:15:02, 7.07s/step, loss=1.239, acc=0.711, acc_len=1.73, thru=20.2, I=24.0, T=38.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:45,795] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:50,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:54:51,362] loop.py:45 INFO TRAIN_STEP step=10758 loss=1.247515 Training: 60%|█████▉ | 10758/18018 [23:24:37<13:24:27, 6.65s/step, loss=1.248, acc=0.708, acc_len=1.71, thru=21.2, I=24.4, T=37.0, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:54:51,596] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:54:57,315] loop.py:45 INFO TRAIN_STEP step=10759 loss=1.174660 Training: 60%|█████▉ | 10759/18018 [23:24:43<12:59:09, 6.44s/step, loss=1.175, acc=0.720, acc_len=1.74, thru=20.4, I=23.7, T=38.9, wait=2.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:55:04,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:55:05,620] loop.py:380 INFO TIMING step=10760: step=3.806s data=1.646s compute=3.777s [fwd=2.741s bwd=1.021s opt=0.003s] dispatch=4.490s [2026-10-19 12:55:05,620] loop.py:45 INFO TRAIN_STEP step=10760 loss=1.381657 Training: 60%|█████▉ | 10760/18018 [23:24:51<14:06:44, 7.00s/step, loss=1.382, acc=0.686, acc_len=1.63, thru=16.4, I=18.8, T=33.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:55:05,764] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:55:13,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:55:13,355] loop.py:45 INFO TRAIN_STEP step=10761 loss=1.229651 Training: 60%|█████▉ | 10761/18018 [23:24:59<14:33:10, 7.22s/step, loss=1.230, acc=0.711, acc_len=1.72, thru=15.1, I=16.6, T=36.9, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:55:13,350] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10760: forward=2462.9ms backward=959.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:55:13,806] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:55:19,619] loop.py:45 INFO TRAIN_STEP step=10762 loss=1.372905 Training: 60%|█████▉ | 10762/18018 [23:25:05<13:58:30, 6.93s/step, loss=1.373, acc=0.687, acc_len=1.63, thru=17.6, I=18.5, T=37.6, wait=2.8s, pool=41, epoch=2/2][2026-10-19 12:55:25,688] loop.py:45 INFO TRAIN_STEP step=10763 loss=1.385679 Training: 60%|█████▉ | 10763/18018 [23:25:11<13:26:59, 6.67s/step, loss=1.386, acc=0.691, acc_len=1.64, thru=20.4, I=21.2, T=36.2, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:55:32,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:55:33,604] loop.py:45 INFO TRAIN_STEP step=10764 loss=1.393161 Training: 60%|█████▉ | 10764/18018 [23:25:19<14:11:57, 7.05s/step, loss=1.393, acc=0.682, acc_len=1.62, thru=16.7, I=17.5, T=34.4, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:55:33,916] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:55:41,654] loop.py:380 INFO TIMING step=10765: step=3.524s data=1.403s compute=3.481s [fwd=2.461s bwd=1.005s opt=0.003s] dispatch=4.516s [2026-10-19 12:55:41,655] loop.py:45 INFO TRAIN_STEP step=10765 loss=1.275899 Training: 60%|█████▉ | 10765/18018 [23:25:27<14:48:15, 7.35s/step, loss=1.276, acc=0.706, acc_len=1.70, thru=12.0, I=15.4, T=36.3, wait=4.5s, pool=25, epoch=2/2][2026-10-19 12:55:54,432] loop.py:45 INFO TRAIN_STEP step=10766 loss=1.352798 Training: 60%|█████▉ | 10766/18018 [23:25:40<18:04:57, 8.98s/step, loss=1.353, acc=0.691, acc_len=1.65, thru=13.2, I=13.8, T=32.0, wait=8.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:55:54,428] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10765: forward=2895.9ms backward=1059.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:56:01,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:56:02,745] loop.py:45 INFO TRAIN_STEP step=10767 loss=1.246326 Training: 60%|█████▉ | 10767/18018 [23:25:48<17:40:43, 8.78s/step, loss=1.246, acc=0.707, acc_len=1.71, thru=15.5, I=16.1, T=33.3, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:56:03,069] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:56:10,748] loop.py:45 INFO TRAIN_STEP step=10768 loss=1.285287 Training: 60%|█████▉ | 10768/18018 [23:25:56<17:12:33, 8.55s/step, loss=1.285, acc=0.706, acc_len=1.70, thru=12.8, I=17.6, T=35.2, wait=4.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:56:21,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:56:22,520] loop.py:45 INFO TRAIN_STEP step=10769 loss=1.364812 Training: 60%|█████▉ | 10769/18018 [23:26:08<19:09:22, 9.51s/step, loss=1.365, acc=0.690, acc_len=1.65, thru=13.9, I=13.7, T=35.6, wait=8.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:56:22,700] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:56:28,624] loop.py:380 INFO TIMING step=10770: step=3.291s data=1.243s compute=3.254s [fwd=2.345s bwd=0.894s opt=0.003s] dispatch=2.804s [2026-10-19 12:56:28,624] loop.py:45 INFO TRAIN_STEP step=10770 loss=1.405020 Training: 60%|█████▉ | 10770/18018 [23:26:14<17:05:40, 8.49s/step, loss=1.405, acc=0.676, acc_len=1.61, thru=19.7, I=20.2, T=38.9, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:56:33,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:56:35,265] loop.py:45 INFO TRAIN_STEP step=10771 loss=1.282768 Training: 60%|█████▉ | 10771/18018 [23:26:21<15:58:32, 7.94s/step, loss=1.283, acc=0.698, acc_len=1.68, thru=20.0, I=24.0, T=33.3, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:56:35,261] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10770: forward=2747.6ms backward=1045.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:56:35,499] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 12:56:40,869] loop.py:45 INFO TRAIN_STEP step=10772 loss=1.312497 Training: 60%|█████▉ | 10772/18018 [23:26:26<14:33:50, 7.24s/step, loss=1.312, acc=0.698, acc_len=1.68, thru=20.0, I=25.6, T=38.2, wait=2.2s, pool=53, epoch=2/2][2026-10-19 12:56:53,117] loop.py:45 INFO TRAIN_STEP step=10773 loss=1.339626 Training: 60%|█████▉ | 10773/18018 [23:26:39<17:35:25, 8.74s/step, loss=1.340, acc=0.694, acc_len=1.66, thru=9.8, I=9.3, T=34.6, wait=8.5s, pool=45, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:00,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:57:01,657] loop.py:45 INFO TRAIN_STEP step=10774 loss=1.340106 Training: 60%|█████▉ | 10774/18018 [23:26:47<17:27:56, 8.68s/step, loss=1.340, acc=0.695, acc_len=1.67, thru=17.6, I=17.0, T=34.8, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:01,877] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:06,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:57:07,633] loop.py:380 INFO TIMING step=10775: step=3.376s data=1.380s compute=3.349s [fwd=2.427s bwd=0.907s opt=0.003s] dispatch=2.590s [2026-10-19 12:57:07,634] loop.py:45 INFO TRAIN_STEP step=10775 loss=1.458255 Training: 60%|█████▉ | 10775/18018 [23:26:53<15:49:53, 7.87s/step, loss=1.458, acc=0.672, acc_len=1.59, thru=20.8, I=24.4, T=37.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:07,835] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:14,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:57:14,712] loop.py:45 INFO TRAIN_STEP step=10776 loss=1.383887 Training: 60%|█████▉ | 10776/18018 [23:27:00<15:21:09, 7.63s/step, loss=1.384, acc=0.689, acc_len=1.64, thru=18.5, I=20.2, T=36.1, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:57:14,708] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10775: forward=2519.4ms backward=978.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:15,072] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:57:21,098] loop.py:45 INFO TRAIN_STEP step=10777 loss=1.393824 Training: 60%|█████▉ | 10777/18018 [23:27:07<14:35:52, 7.26s/step, loss=1.394, acc=0.684, acc_len=1.62, thru=18.5, I=20.1, T=35.5, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:28,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:57:28,882] loop.py:45 INFO TRAIN_STEP step=10778 loss=1.315159 Training: 60%|█████▉ | 10778/18018 [23:27:14<14:54:43, 7.41s/step, loss=1.315, acc=0.699, acc_len=1.67, thru=18.6, I=17.3, T=34.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:29,366] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:34,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:57:35,182] loop.py:45 INFO TRAIN_STEP step=10779 loss=1.325957 Training: 60%|█████▉ | 10779/18018 [23:27:21<14:14:22, 7.08s/step, loss=1.326, acc=0.693, acc_len=1.66, thru=20.0, I=21.7, T=36.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:35,544] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:40,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:57:41,610] loop.py:380 INFO TIMING step=10780: step=3.593s data=1.560s compute=3.553s [fwd=2.497s bwd=1.041s opt=0.003s] dispatch=2.825s [2026-10-19 12:57:41,610] loop.py:45 INFO TRAIN_STEP step=10780 loss=1.367488 Training: 60%|█████▉ | 10780/18018 [23:27:27<13:50:34, 6.89s/step, loss=1.367, acc=0.690, acc_len=1.64, thru=20.1, I=26.2, T=35.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:42,116] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:47,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:57:47,601] loop.py:45 INFO TRAIN_STEP step=10781 loss=1.303771 Training: 60%|█████▉ | 10781/18018 [23:27:33<13:18:03, 6.62s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=20.9, I=26.4, T=40.4, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:57:47,597] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10780: forward=2260.2ms backward=862.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:48,076] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:53,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:57:53,672] loop.py:45 INFO TRAIN_STEP step=10782 loss=1.314454 Training: 60%|█████▉ | 10782/18018 [23:27:39<12:58:10, 6.45s/step, loss=1.314, acc=0.700, acc_len=1.68, thru=21.1, I=22.2, T=38.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:54,095] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:57:59,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:58:00,891] loop.py:45 INFO TRAIN_STEP step=10783 loss=1.218643 Training: 60%|█████▉ | 10783/18018 [23:27:46<13:25:50, 6.68s/step, loss=1.219, acc=0.715, acc_len=1.74, thru=18.5, I=21.3, T=36.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:58:01,228] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:58:07,331] loop.py:45 INFO TRAIN_STEP step=10784 loss=1.349163 Training: 60%|█████▉ | 10784/18018 [23:27:53<13:17:06, 6.61s/step, loss=1.349, acc=0.687, acc_len=1.63, thru=17.6, I=23.5, T=38.2, wait=3.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:58:14,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:58:15,304] loop.py:380 INFO TIMING step=10785: step=3.461s data=1.412s compute=3.433s [fwd=2.450s bwd=0.968s opt=0.003s] dispatch=4.500s [2026-10-19 12:58:15,304] loop.py:45 INFO TRAIN_STEP step=10785 loss=1.353429 Training: 60%|█████▉ | 10785/18018 [23:28:01<14:06:09, 7.02s/step, loss=1.353, acc=0.687, acc_len=1.64, thru=16.9, I=17.4, T=37.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:58:15,401] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:58:22,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:58:24,867] loop.py:45 INFO TRAIN_STEP step=10786 loss=1.394033 Training: 60%|█████▉ | 10786/18018 [23:28:10<15:37:59, 7.78s/step, loss=1.394, acc=0.685, acc_len=1.64, thru=13.2, I=16.9, T=30.8, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:58:24,863] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10785: forward=3032.0ms backward=1074.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:58:25,209] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:58:30,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:58:30,359] loop.py:45 INFO TRAIN_STEP step=10787 loss=1.339962 Training: 60%|█████▉ | 10787/18018 [23:28:16<14:15:04, 7.10s/step, loss=1.340, acc=0.693, acc_len=1.66, thru=20.2, I=28.0, T=38.4, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:58:30,639] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 12:58:37,911] loop.py:45 INFO TRAIN_STEP step=10788 loss=1.340148 Training: 60%|█████▉ | 10788/18018 [23:28:23<14:31:29, 7.23s/step, loss=1.340, acc=0.694, acc_len=1.66, thru=16.7, I=17.5, T=35.9, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:58:44,091] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:58:45,101] loop.py:45 INFO TRAIN_STEP step=10789 loss=1.349184 Training: 60%|█████▉ | 10789/18018 [23:28:31<14:29:49, 7.22s/step, loss=1.349, acc=0.695, acc_len=1.67, thru=18.7, I=19.2, T=36.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:58:45,600] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 12:58:52,314] loop.py:380 INFO TIMING step=10790: step=3.549s data=1.500s compute=3.516s [fwd=2.478s bwd=1.024s opt=0.003s] dispatch=3.654s [2026-10-19 12:58:52,314] loop.py:45 INFO TRAIN_STEP step=10790 loss=1.359789 Training: 60%|█████▉ | 10790/18018 [23:28:38<14:29:23, 7.22s/step, loss=1.360, acc=0.691, acc_len=1.65, thru=16.0, I=18.8, T=36.1, wait=3.7s, pool=45, epoch=2/2][2026-10-19 12:59:00,785] loop.py:45 INFO TRAIN_STEP step=10791 loss=1.274788 Training: 60%|█████▉ | 10791/18018 [23:28:46<15:14:41, 7.59s/step, loss=1.275, acc=0.706, acc_len=1.70, thru=15.4, I=14.9, T=35.5, wait=4.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:59:00,780] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10790: forward=2471.2ms backward=1078.8ms [2026-10-19 12:59:07,739] loop.py:45 INFO TRAIN_STEP step=10792 loss=1.341341 Training: 60%|█████▉ | 10792/18018 [23:28:53<14:51:20, 7.40s/step, loss=1.341, acc=0.693, acc_len=1.65, thru=17.4, I=18.2, T=37.2, wait=3.5s, pool=49, epoch=2/2][2026-10-19 12:59:16,501] loop.py:45 INFO TRAIN_STEP step=10793 loss=1.412707 Training: 60%|█████▉ | 10793/18018 [23:29:02<15:40:29, 7.81s/step, loss=1.413, acc=0.682, acc_len=1.62, thru=14.3, I=15.1, T=34.8, wait=5.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:59:25,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:59:26,692] loop.py:45 INFO TRAIN_STEP step=10794 loss=1.320808 Training: 60%|█████▉ | 10794/18018 [23:29:12<17:06:16, 8.52s/step, loss=1.321, acc=0.696, acc_len=1.67, thru=15.6, I=13.9, T=34.9, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:59:27,019] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:59:32,352] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:59:33,819] loop.py:380 INFO TIMING step=10795: step=3.486s data=1.397s compute=3.448s [fwd=2.444s bwd=0.988s opt=0.003s] dispatch=3.632s [2026-10-19 12:59:33,820] loop.py:45 INFO TRAIN_STEP step=10795 loss=1.207276 Training: 60%|█████▉ | 10795/18018 [23:29:19<16:15:46, 8.11s/step, loss=1.207, acc=0.718, acc_len=1.74, thru=17.1, I=22.5, T=36.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:59:33,874] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:59:38,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:59:39,444] loop.py:45 INFO TRAIN_STEP step=10796 loss=1.317810 Training: 60%|█████▉ | 10796/18018 [23:29:25<14:46:01, 7.36s/step, loss=1.318, acc=0.699, acc_len=1.67, thru=21.9, I=27.7, T=38.2, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 12:59:39,440] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10795: forward=2348.8ms backward=962.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:59:39,656] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:59:45,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 12:59:46,200] loop.py:45 INFO TRAIN_STEP step=10797 loss=1.345605 Training: 60%|█████▉ | 10797/18018 [23:29:32<14:24:01, 7.18s/step, loss=1.346, acc=0.691, acc_len=1.65, thru=19.5, I=22.3, T=35.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 12:59:46,593] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 12:59:53,712] loop.py:45 INFO TRAIN_STEP step=10798 loss=1.298479 Training: 60%|█████▉ | 10798/18018 [23:29:39<14:35:56, 7.28s/step, loss=1.298, acc=0.702, acc_len=1.69, thru=15.5, I=17.0, T=36.4, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:02,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:00:02,731] loop.py:45 INFO TRAIN_STEP step=10799 loss=1.246833 Training: 60%|█████▉ | 10799/18018 [23:29:48<15:38:38, 7.80s/step, loss=1.247, acc=0.708, acc_len=1.71, thru=15.3, I=15.0, T=35.7, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:02,990] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:07,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:00:08,620] loop.py:380 INFO TIMING step=10800: step=3.355s data=1.368s compute=3.310s [fwd=2.333s bwd=0.963s opt=0.003s] dispatch=2.523s [2026-10-19 13:00:08,620] loop.py:45 INFO TRAIN_STEP step=10800 loss=1.419592 Training: 60%|█████▉ | 10800/18018 [23:29:54<14:29:27, 7.23s/step, loss=1.420, acc=0.683, acc_len=1.62, thru=20.5, I=21.3, T=38.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:08,912] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:00:16,821] loop.py:45 INFO TRAIN_STEP step=10801 loss=1.341187 Training: 60%|█████▉ | 10801/18018 [23:30:02<15:04:27, 7.52s/step, loss=1.341, acc=0.693, acc_len=1.66, thru=14.8, I=16.2, T=36.8, wait=4.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:00:16,816] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10800: forward=2442.2ms backward=975.9ms [2026-10-19 13:00:25,340] loop.py:45 INFO TRAIN_STEP step=10802 loss=1.314528 Training: 60%|█████▉ | 10802/18018 [23:30:11<15:40:19, 7.82s/step, loss=1.315, acc=0.699, acc_len=1.68, thru=15.6, I=14.9, T=35.8, wait=4.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:31,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:00:32,587] loop.py:45 INFO TRAIN_STEP step=10803 loss=1.201703 Training: 60%|█████▉ | 10803/18018 [23:30:18<15:19:36, 7.65s/step, loss=1.202, acc=0.718, acc_len=1.75, thru=19.1, I=20.8, T=36.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:33,023] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:00:38,692] loop.py:45 INFO TRAIN_STEP step=10804 loss=1.328550 Training: 60%|█████▉ | 10804/18018 [23:30:24<14:23:54, 7.19s/step, loss=1.329, acc=0.696, acc_len=1.67, thru=18.6, I=24.2, T=37.1, wait=2.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:48,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:00:49,461] loop.py:380 INFO TIMING step=10805: step=3.806s data=1.635s compute=3.760s [fwd=2.589s bwd=1.155s opt=0.003s] dispatch=6.954s [2026-10-19 13:00:49,461] loop.py:45 INFO TRAIN_STEP step=10805 loss=1.277236 Training: 60%|█████▉ | 10805/18018 [23:30:35<16:32:56, 8.26s/step, loss=1.277, acc=0.709, acc_len=1.71, thru=13.6, I=13.6, T=33.6, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:49,699] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:55,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:00:55,774] loop.py:45 INFO TRAIN_STEP step=10806 loss=1.354862 Training: 60%|█████▉ | 10806/18018 [23:30:41<15:22:43, 7.68s/step, loss=1.355, acc=0.693, acc_len=1.65, thru=19.4, I=19.8, T=36.0, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:00:55,769] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10805: forward=2491.1ms backward=1013.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:00:55,885] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:01:00,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:01:01,377] loop.py:45 INFO TRAIN_STEP step=10807 loss=1.244430 Training: 60%|█████▉ | 10807/18018 [23:30:47<14:07:49, 7.05s/step, loss=1.244, acc=0.713, acc_len=1.73, thru=22.5, I=26.5, T=37.5, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:01:01,740] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:01:08,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:01:09,014] loop.py:45 INFO TRAIN_STEP step=10808 loss=1.346702 Training: 60%|█████▉ | 10808/18018 [23:30:55<14:28:38, 7.23s/step, loss=1.347, acc=0.692, acc_len=1.65, thru=16.4, I=19.2, T=34.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:01:09,512] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:01:17,443] loop.py:45 INFO TRAIN_STEP step=10809 loss=1.371846 Training: 60%|█████▉ | 10809/18018 [23:31:03<15:11:50, 7.59s/step, loss=1.372, acc=0.685, acc_len=1.63, thru=14.4, I=15.7, T=35.4, wait=4.8s, pool=61, epoch=2/2][2026-10-19 13:01:23,417] loop.py:380 INFO TIMING step=10810: step=3.670s data=1.595s compute=3.639s [fwd=2.590s bwd=1.034s opt=0.003s] dispatch=2.295s [2026-10-19 13:01:23,417] loop.py:45 INFO TRAIN_STEP step=10810 loss=1.256934 Training: 60%|█████▉ | 10810/18018 [23:31:09<14:13:24, 7.10s/step, loss=1.257, acc=0.706, acc_len=1.70, thru=19.9, I=18.6, T=34.9, wait=2.3s, pool=61, epoch=2/2][2026-10-19 13:01:29,173] loop.py:45 INFO TRAIN_STEP step=10811 loss=1.244120 Training: 60%|██████ | 10811/18018 [23:31:15<13:24:45, 6.70s/step, loss=1.244, acc=0.707, acc_len=1.71, thru=19.6, I=20.9, T=38.8, wait=2.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:01:29,169] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10810: forward=2353.8ms backward=895.8ms [2026-10-19 13:01:36,495] loop.py:45 INFO TRAIN_STEP step=10812 loss=1.360200 Training: 60%|██████ | 10812/18018 [23:31:22<13:47:08, 6.89s/step, loss=1.360, acc=0.691, acc_len=1.65, thru=15.7, I=17.1, T=37.1, wait=3.9s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:01:44,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:01:46,523] loop.py:45 INFO TRAIN_STEP step=10813 loss=1.360757 Training: 60%|██████ | 10813/18018 [23:31:32<15:40:09, 7.83s/step, loss=1.361, acc=0.690, acc_len=1.65, thru=15.6, I=16.8, T=32.6, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:01:46,696] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:01:52,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:01:52,484] loop.py:45 INFO TRAIN_STEP step=10814 loss=1.235186 Training: 60%|██████ | 10814/18018 [23:31:38<14:32:43, 7.27s/step, loss=1.235, acc=0.712, acc_len=1.73, thru=20.0, I=24.5, T=36.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:01:52,674] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:01:58,693] loop.py:380 INFO TIMING step=10815: step=3.410s data=1.396s compute=3.365s [fwd=2.396s bwd=0.954s opt=0.003s] dispatch=2.789s [2026-10-19 13:01:58,693] loop.py:45 INFO TRAIN_STEP step=10815 loss=1.354815 Training: 60%|██████ | 10815/18018 [23:31:44<13:54:25, 6.95s/step, loss=1.355, acc=0.689, acc_len=1.65, thru=16.4, I=19.3, T=37.5, wait=2.8s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:06,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:02:07,789] loop.py:45 INFO TRAIN_STEP step=10816 loss=1.370393 Training: 60%|██████ | 10816/18018 [23:31:53<15:11:30, 7.59s/step, loss=1.370, acc=0.690, acc_len=1.65, thru=17.9, I=18.4, T=36.4, wait=5.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:02:07,784] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10815: forward=2510.2ms backward=957.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:08,286] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:13,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:02:13,863] loop.py:45 INFO TRAIN_STEP step=10817 loss=1.342625 Training: 60%|██████ | 10817/18018 [23:31:59<14:16:45, 7.14s/step, loss=1.343, acc=0.691, acc_len=1.65, thru=20.4, I=25.0, T=37.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:14,088] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:18,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:02:19,778] loop.py:45 INFO TRAIN_STEP step=10818 loss=1.344203 Training: 60%|██████ | 10818/18018 [23:32:05<13:32:31, 6.77s/step, loss=1.344, acc=0.691, acc_len=1.65, thru=21.6, I=25.5, T=37.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:19,953] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:02:26,272] loop.py:45 INFO TRAIN_STEP step=10819 loss=1.441112 Training: 60%|██████ | 10819/18018 [23:32:12<13:22:25, 6.69s/step, loss=1.441, acc=0.675, acc_len=1.59, thru=17.8, I=23.3, T=36.1, wait=2.9s, pool=45, epoch=2/2][2026-10-19 13:02:33,045] loop.py:380 INFO TIMING step=10820: step=3.585s data=1.433s compute=3.555s [fwd=2.511s bwd=1.029s opt=0.003s] dispatch=3.179s [2026-10-19 13:02:33,045] loop.py:45 INFO TRAIN_STEP step=10820 loss=1.329403 Training: 60%|██████ | 10820/18018 [23:32:19<13:25:19, 6.71s/step, loss=1.329, acc=0.700, acc_len=1.68, thru=18.9, I=18.6, T=35.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:33,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:33,532] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:39,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:02:40,558] loop.py:45 INFO TRAIN_STEP step=10821 loss=1.277427 Training: 60%|██████ | 10821/18018 [23:32:26<13:54:05, 6.95s/step, loss=1.277, acc=0.708, acc_len=1.71, thru=17.6, I=19.0, T=36.5, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:02:40,554] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10820: forward=2489.5ms backward=975.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:40,955] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:46,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:02:46,223] loop.py:45 INFO TRAIN_STEP step=10822 loss=1.240194 Training: 60%|██████ | 10822/18018 [23:32:32<13:07:36, 6.57s/step, loss=1.240, acc=0.712, acc_len=1.73, thru=20.6, I=21.0, T=36.3, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:46,545] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:54,209] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:02:55,501] loop.py:45 INFO TRAIN_STEP step=10823 loss=1.502880 Training: 60%|██████ | 10823/18018 [23:32:41<14:44:59, 7.38s/step, loss=1.503, acc=0.667, acc_len=1.57, thru=13.9, I=15.2, T=32.0, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:02:55,721] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:02,276] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:03:03,146] loop.py:45 INFO TRAIN_STEP step=10824 loss=1.380923 Training: 60%|██████ | 10824/18018 [23:32:49<14:54:28, 7.46s/step, loss=1.381, acc=0.687, acc_len=1.64, thru=15.7, I=19.2, T=35.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:03,290] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:09,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:03:09,628] loop.py:380 INFO TIMING step=10825: step=3.308s data=1.335s compute=3.278s [fwd=2.363s bwd=0.901s opt=0.003s] dispatch=3.163s [2026-10-19 13:03:09,628] loop.py:45 INFO TRAIN_STEP step=10825 loss=1.213360 Training: 60%|██████ | 10825/18018 [23:32:55<14:19:07, 7.17s/step, loss=1.213, acc=0.713, acc_len=1.73, thru=19.5, I=22.0, T=38.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:09,678] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:14,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:03:15,264] loop.py:45 INFO TRAIN_STEP step=10826 loss=1.381283 Training: 60%|██████ | 10826/18018 [23:33:01<13:24:00, 6.71s/step, loss=1.381, acc=0.689, acc_len=1.64, thru=22.4, I=24.1, T=37.5, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:15,302] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 13:03:15,259] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10825: forward=2448.1ms backward=921.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:22,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:03:23,188] loop.py:45 INFO TRAIN_STEP step=10827 loss=1.371467 Training: 60%|██████ | 10827/18018 [23:33:09<14:07:38, 7.07s/step, loss=1.371, acc=0.690, acc_len=1.64, thru=17.1, I=18.3, T=34.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:23,274] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:29,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:03:29,738] loop.py:45 INFO TRAIN_STEP step=10828 loss=1.260654 Training: 60%|██████ | 10828/18018 [23:33:15<13:48:44, 6.92s/step, loss=1.261, acc=0.707, acc_len=1.71, thru=18.2, I=20.0, T=35.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:30,083] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:34,968] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:03:35,430] loop.py:45 INFO TRAIN_STEP step=10829 loss=1.235606 Training: 60%|██████ | 10829/18018 [23:33:21<13:04:37, 6.55s/step, loss=1.236, acc=0.716, acc_len=1.73, thru=21.5, I=23.1, T=37.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:35,476] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:40,525] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:03:41,929] loop.py:380 INFO TIMING step=10830: step=3.588s data=1.519s compute=3.559s [fwd=2.543s bwd=1.001s opt=0.003s] dispatch=2.901s [2026-10-19 13:03:41,929] loop.py:45 INFO TRAIN_STEP step=10830 loss=1.304702 Training: 60%|██████ | 10830/18018 [23:33:28<13:02:43, 6.53s/step, loss=1.305, acc=0.701, acc_len=1.68, thru=19.8, I=23.0, T=35.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:42,047] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:50,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:03:51,248] loop.py:45 INFO TRAIN_STEP step=10831 loss=1.329971 Training: 60%|██████ | 10831/18018 [23:33:37<14:42:39, 7.37s/step, loss=1.330, acc=0.693, acc_len=1.67, thru=13.9, I=15.0, T=33.1, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:03:51,244] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10830: forward=2675.6ms backward=1139.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:03:51,685] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:03:57,765] loop.py:45 INFO TRAIN_STEP step=10832 loss=1.338982 Training: 60%|██████ | 10832/18018 [23:33:43<14:11:53, 7.11s/step, loss=1.339, acc=0.691, acc_len=1.65, thru=18.0, I=18.9, T=38.6, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:04:04,183] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:04:06,205] loop.py:45 INFO TRAIN_STEP step=10833 loss=1.325944 Training: 60%|██████ | 10833/18018 [23:33:52<14:59:31, 7.51s/step, loss=1.326, acc=0.697, acc_len=1.67, thru=17.0, I=20.5, T=28.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:04:06,700] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:04:12,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:04:13,363] loop.py:45 INFO TRAIN_STEP step=10834 loss=1.370403 Training: 60%|██████ | 10834/18018 [23:33:59<14:46:43, 7.41s/step, loss=1.370, acc=0.691, acc_len=1.65, thru=15.7, I=22.1, T=38.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:04:13,451] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 13:04:18,685] loop.py:380 INFO TIMING step=10835: step=3.398s data=1.396s compute=3.358s [fwd=2.374s bwd=0.970s opt=0.003s] dispatch=1.913s [2026-10-19 13:04:18,685] loop.py:45 INFO TRAIN_STEP step=10835 loss=1.343970 Training: 60%|██████ | 10835/18018 [23:34:04<13:31:43, 6.78s/step, loss=1.344, acc=0.689, acc_len=1.64, thru=21.0, I=22.3, T=37.7, wait=1.9s, pool=57, epoch=2/2][2026-10-19 13:04:25,270] loop.py:45 INFO TRAIN_STEP step=10836 loss=1.322696 Training: 60%|██████ | 10836/18018 [23:34:11<13:24:38, 6.72s/step, loss=1.323, acc=0.694, acc_len=1.66, thru=17.9, I=18.7, T=38.6, wait=3.3s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:04:25,266] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10835: forward=2356.0ms backward=912.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:04:32,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:04:32,778] loop.py:45 INFO TRAIN_STEP step=10837 loss=1.226004 Training: 60%|██████ | 10837/18018 [23:34:18<13:52:39, 6.96s/step, loss=1.226, acc=0.713, acc_len=1.73, thru=18.4, I=18.3, T=31.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:04:32,960] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:04:42,435] loop.py:45 INFO TRAIN_STEP step=10838 loss=1.371360 Training: 60%|██████ | 10838/18018 [23:34:28<15:29:28, 7.77s/step, loss=1.371, acc=0.689, acc_len=1.64, thru=12.8, I=12.5, T=30.6, wait=5.5s, pool=61, epoch=2/2][2026-10-19 13:04:49,170] loop.py:45 INFO TRAIN_STEP step=10839 loss=1.351103 Training: 60%|██████ | 10839/18018 [23:34:35<14:52:17, 7.46s/step, loss=1.351, acc=0.691, acc_len=1.64, thru=16.9, I=15.6, T=32.3, wait=2.8s, pool=49, epoch=2/2][2026-10-19 13:04:59,775] loop.py:380 INFO TIMING step=10840: step=3.937s data=1.667s compute=3.907s [fwd=2.849s bwd=1.044s opt=0.003s] dispatch=6.660s [2026-10-19 13:04:59,776] loop.py:45 INFO TRAIN_STEP step=10840 loss=1.275247 Training: 60%|██████ | 10840/18018 [23:34:45<16:45:06, 8.40s/step, loss=1.275, acc=0.707, acc_len=1.71, thru=12.8, I=11.3, T=32.5, wait=6.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:05,498] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:05:06,054] loop.py:45 INFO TRAIN_STEP step=10841 loss=1.287821 Training: 60%|██████ | 10841/18018 [23:34:52<15:28:47, 7.76s/step, loss=1.288, acc=0.704, acc_len=1.70, thru=19.1, I=20.1, T=37.0, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:05:06,050] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10840: forward=2429.9ms backward=977.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:06,503] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:12,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:05:13,954] loop.py:45 INFO TRAIN_STEP step=10842 loss=1.328778 Training: 60%|██████ | 10842/18018 [23:35:00<15:33:32, 7.81s/step, loss=1.329, acc=0.695, acc_len=1.66, thru=16.7, I=20.8, T=34.7, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:14,147] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:05:20,942] loop.py:45 INFO TRAIN_STEP step=10843 loss=1.317370 Training: 60%|██████ | 10843/18018 [23:35:07<15:04:05, 7.56s/step, loss=1.317, acc=0.696, acc_len=1.67, thru=15.4, I=18.2, T=36.9, wait=3.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:28,822] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:05:29,939] loop.py:45 INFO TRAIN_STEP step=10844 loss=1.295899 Training: 60%|██████ | 10844/18018 [23:35:16<15:55:31, 7.99s/step, loss=1.296, acc=0.703, acc_len=1.70, thru=18.1, I=17.0, T=34.6, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:30,331] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:34,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:05:35,519] loop.py:380 INFO TIMING step=10845: step=3.262s data=1.291s compute=3.235s [fwd=2.316s bwd=0.906s opt=0.003s] dispatch=2.309s [2026-10-19 13:05:35,519] loop.py:45 INFO TRAIN_STEP step=10845 loss=1.279879 Training: 60%|██████ | 10845/18018 [23:35:21<14:28:52, 7.27s/step, loss=1.280, acc=0.703, acc_len=1.69, thru=21.6, I=27.1, T=39.2, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:35,682] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:40,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:05:41,506] loop.py:45 INFO TRAIN_STEP step=10846 loss=1.265792 (TrainerActor pid=9310) [2026-10-19 13:05:41,501] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10845: forward=2387.1ms backward=950.4ms Training: 60%|██████ | 10846/18018 [23:35:27<13:42:52, 6.88s/step, loss=1.266, acc=0.710, acc_len=1.71, thru=22.8, I=26.2, T=37.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:05:41,843] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:05:49,756] loop.py:45 INFO TRAIN_STEP step=10847 loss=1.434386 Training: 60%|██████ | 10847/18018 [23:35:35<14:31:39, 7.29s/step, loss=1.434, acc=0.676, acc_len=1.60, thru=11.7, I=18.3, T=34.5, wait=4.5s, pool=37, epoch=2/2][2026-10-19 13:05:56,451] loop.py:45 INFO TRAIN_STEP step=10848 loss=1.363938 Training: 60%|██████ | 10848/18018 [23:35:42<14:10:06, 7.11s/step, loss=1.364, acc=0.689, acc_len=1.64, thru=18.5, I=18.1, T=37.2, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:06:04,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:06:05,453] loop.py:45 INFO TRAIN_STEP step=10849 loss=1.257078 Training: 60%|██████ | 10849/18018 [23:35:51<15:17:43, 7.68s/step, loss=1.257, acc=0.710, acc_len=1.73, thru=15.1, I=15.8, T=34.0, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:06:05,528] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:06:11,021] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:06:12,347] loop.py:380 INFO TIMING step=10850: step=3.385s data=1.296s compute=3.356s [fwd=2.422s bwd=0.920s opt=0.003s] dispatch=3.499s [2026-10-19 13:06:12,347] loop.py:45 INFO TRAIN_STEP step=10850 loss=1.302835 Training: 60%|██████ | 10850/18018 [23:35:58<14:49:21, 7.44s/step, loss=1.303, acc=0.697, acc_len=1.67, thru=18.1, I=23.2, T=37.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:06:12,528] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:06:19,510] loop.py:45 INFO TRAIN_STEP step=10851 loss=1.306040 Training: 60%|██████ | 10851/18018 [23:36:05<14:39:07, 7.36s/step, loss=1.306, acc=0.701, acc_len=1.69, thru=16.6, I=20.5, T=37.1, wait=3.7s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:06:19,507] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10850: forward=2409.8ms backward=973.3ms [2026-10-19 13:06:29,433] loop.py:45 INFO TRAIN_STEP step=10852 loss=1.314573 Training: 60%|██████ | 10852/18018 [23:36:15<16:10:59, 8.13s/step, loss=1.315, acc=0.700, acc_len=1.68, thru=13.3, I=13.1, T=34.8, wait=6.2s, pool=53, epoch=2/2][2026-10-19 13:06:38,689] loop.py:45 INFO TRAIN_STEP step=10853 loss=1.348050 Training: 60%|██████ | 10853/18018 [23:36:24<16:51:04, 8.47s/step, loss=1.348, acc=0.688, acc_len=1.64, thru=13.4, I=13.4, T=34.4, wait=5.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:06:45,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:06:46,306] loop.py:45 INFO TRAIN_STEP step=10854 loss=1.429494 Training: 60%|██████ | 10854/18018 [23:36:32<16:20:27, 8.21s/step, loss=1.429, acc=0.676, acc_len=1.60, thru=19.0, I=17.7, T=36.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:06:46,755] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:06:51,992] loop.py:380 INFO TIMING step=10855: step=3.289s data=1.235s compute=3.259s [fwd=2.364s bwd=0.881s opt=0.003s] dispatch=2.389s [2026-10-19 13:06:51,992] loop.py:45 INFO TRAIN_STEP step=10855 loss=1.295552 Training: 60%|██████ | 10855/18018 [23:36:38<14:49:57, 7.45s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=19.2, I=23.0, T=38.9, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:06:58,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:06:59,800] loop.py:45 INFO TRAIN_STEP step=10856 loss=1.242291 Training: 60%|██████ | 10856/18018 [23:36:45<15:02:26, 7.56s/step, loss=1.242, acc=0.714, acc_len=1.73, thru=17.2, I=19.3, T=36.4, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:06:59,795] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10855: forward=2474.3ms backward=992.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:06:59,962] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:07:08,238] loop.py:45 INFO TRAIN_STEP step=10857 loss=1.145195 Training: 60%|██████ | 10857/18018 [23:36:54<15:33:46, 7.82s/step, loss=1.145, acc=0.727, acc_len=1.77, thru=12.2, I=15.2, T=38.9, wait=5.1s, pool=49, epoch=2/2][2026-10-19 13:07:15,975] loop.py:45 INFO TRAIN_STEP step=10858 loss=1.354048 Training: 60%|██████ | 10858/18018 [23:37:02<15:30:28, 7.80s/step, loss=1.354, acc=0.695, acc_len=1.67, thru=17.3, I=17.0, T=35.5, wait=4.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:07:22,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:07:22,977] loop.py:45 INFO TRAIN_STEP step=10859 loss=1.405596 Training: 60%|██████ | 10859/18018 [23:37:09<15:01:52, 7.56s/step, loss=1.406, acc=0.682, acc_len=1.62, thru=18.6, I=19.8, T=37.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:07:23,121] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:07:28,607] loop.py:380 INFO TIMING step=10860: step=3.339s data=1.318s compute=3.309s [fwd=2.390s bwd=0.904s opt=0.003s] dispatch=2.284s [2026-10-19 13:07:28,608] loop.py:45 INFO TRAIN_STEP step=10860 loss=1.284838 Training: 60%|██████ | 10860/18018 [23:37:14<13:52:44, 6.98s/step, loss=1.285, acc=0.701, acc_len=1.68, thru=20.6, I=22.8, T=38.3, wait=2.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:07:36,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:07:37,580] loop.py:45 INFO TRAIN_STEP step=10861 loss=1.350045 Training: 60%|██████ | 10861/18018 [23:37:23<15:03:59, 7.58s/step, loss=1.350, acc=0.692, acc_len=1.66, thru=15.3, I=16.5, T=34.3, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:07:37,576] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10860: forward=2647.7ms backward=1039.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:07:37,870] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:07:43,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:07:43,300] loop.py:45 INFO TRAIN_STEP step=10862 loss=1.302995 Training: 60%|██████ | 10862/18018 [23:37:29<13:57:17, 7.02s/step, loss=1.303, acc=0.700, acc_len=1.68, thru=20.4, I=20.8, T=39.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:07:43,728] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:07:49,100] loop.py:45 INFO TRAIN_STEP step=10863 loss=1.382222 Training: 60%|██████ | 10863/18018 [23:37:35<13:13:29, 6.65s/step, loss=1.382, acc=0.687, acc_len=1.63, thru=21.2, I=22.9, T=37.7, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:07:54,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:07:55,217] loop.py:45 INFO TRAIN_STEP step=10864 loss=1.217003 Training: 60%|██████ | 10864/18018 [23:37:41<12:54:15, 6.49s/step, loss=1.217, acc=0.715, acc_len=1.73, thru=21.2, I=21.5, T=37.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:07:55,434] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:00,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:08:01,435] loop.py:380 INFO TIMING step=10865: step=3.335s data=1.290s compute=3.303s [fwd=2.363s bwd=0.925s opt=0.003s] dispatch=2.873s [2026-10-19 13:08:01,435] loop.py:45 INFO TRAIN_STEP step=10865 loss=1.301741 Training: 60%|██████ | 10865/18018 [23:37:47<12:44:18, 6.41s/step, loss=1.302, acc=0.702, acc_len=1.68, thru=21.2, I=24.5, T=38.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:01,623] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:06,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:08:08,940] loop.py:45 INFO TRAIN_STEP step=10866 loss=1.414192 Training: 60%|██████ | 10866/18018 [23:37:55<13:23:18, 6.74s/step, loss=1.414, acc=0.685, acc_len=1.63, thru=17.6, I=24.0, T=30.4, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:08:08,935] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10865: forward=3051.5ms backward=1118.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:09,026] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:15,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:08:15,415] loop.py:45 INFO TRAIN_STEP step=10867 loss=1.291465 Training: 60%|██████ | 10867/18018 [23:38:01<13:13:45, 6.66s/step, loss=1.291, acc=0.702, acc_len=1.69, thru=16.3, I=21.3, T=36.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:15,710] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:21,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:08:22,768] loop.py:45 INFO TRAIN_STEP step=10868 loss=1.228923 Training: 60%|██████ | 10868/18018 [23:38:08<13:38:21, 6.87s/step, loss=1.229, acc=0.713, acc_len=1.73, thru=18.6, I=20.7, T=28.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:22,889] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 13:08:28,940] loop.py:45 INFO TRAIN_STEP step=10869 loss=1.298379 Training: 60%|██████ | 10869/18018 [23:38:15<13:13:21, 6.66s/step, loss=1.298, acc=0.699, acc_len=1.67, thru=17.2, I=19.6, T=36.2, wait=2.6s, pool=61, epoch=2/2][2026-10-19 13:08:36,916] loop.py:380 INFO TIMING step=10870: step=3.436s data=1.385s compute=3.409s [fwd=2.463s bwd=0.932s opt=0.003s] dispatch=4.533s [2026-10-19 13:08:36,917] loop.py:45 INFO TRAIN_STEP step=10870 loss=1.366145 Training: 60%|██████ | 10870/18018 [23:38:23<14:00:25, 7.05s/step, loss=1.366, acc=0.689, acc_len=1.64, thru=15.8, I=15.3, T=37.3, wait=4.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:44,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:08:44,807] loop.py:45 INFO TRAIN_STEP step=10871 loss=1.208839 Training: 60%|██████ | 10871/18018 [23:38:30<14:30:08, 7.30s/step, loss=1.209, acc=0.718, acc_len=1.74, thru=16.7, I=17.1, T=36.9, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:08:44,802] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10870: forward=2381.7ms backward=1032.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:45,023] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:49,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:08:50,566] loop.py:45 INFO TRAIN_STEP step=10872 loss=1.270196 Training: 60%|██████ | 10872/18018 [23:38:36<13:34:46, 6.84s/step, loss=1.270, acc=0.705, acc_len=1.70, thru=20.9, I=23.7, T=36.1, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:08:50,850] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:08:58,576] loop.py:45 INFO TRAIN_STEP step=10873 loss=1.230270 Training: 60%|██████ | 10873/18018 [23:38:44<14:16:27, 7.19s/step, loss=1.230, acc=0.712, acc_len=1.72, thru=14.7, I=16.0, T=37.6, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:05,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:09:06,466] loop.py:45 INFO TRAIN_STEP step=10874 loss=1.349751 Training: 60%|██████ | 10874/18018 [23:38:52<14:41:14, 7.40s/step, loss=1.350, acc=0.692, acc_len=1.66, thru=18.0, I=18.0, T=35.5, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:06,728] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:09:12,702] loop.py:380 INFO TIMING step=10875: step=3.542s data=1.457s compute=3.514s [fwd=2.493s bwd=1.007s opt=0.003s] dispatch=2.685s [2026-10-19 13:09:12,703] loop.py:45 INFO TRAIN_STEP step=10875 loss=1.476379 Training: 60%|██████ | 10875/18018 [23:38:58<13:59:28, 7.05s/step, loss=1.476, acc=0.669, acc_len=1.57, thru=19.4, I=21.6, T=36.1, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:20,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:09:21,031] loop.py:45 INFO TRAIN_STEP step=10876 loss=1.292583 Training: 60%|██████ | 10876/18018 [23:39:07<14:45:03, 7.44s/step, loss=1.293, acc=0.704, acc_len=1.70, thru=16.5, I=17.8, T=33.6, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:09:21,027] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10875: forward=2640.5ms backward=1109.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:21,071] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 13:09:26,728] loop.py:45 INFO TRAIN_STEP step=10877 loss=1.399025 Training: 60%|██████ | 10877/18018 [23:39:12<13:42:51, 6.91s/step, loss=1.399, acc=0.679, acc_len=1.60, thru=19.5, I=20.7, T=38.8, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:34,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:09:36,075] loop.py:45 INFO TRAIN_STEP step=10878 loss=1.441342 Training: 60%|██████ | 10878/18018 [23:39:22<15:09:35, 7.64s/step, loss=1.441, acc=0.680, acc_len=1.61, thru=15.3, I=17.8, T=33.1, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:36,129] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:41,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:09:43,577] loop.py:45 INFO TRAIN_STEP step=10879 loss=1.219667 Training: 60%|██████ | 10879/18018 [23:39:29<15:04:26, 7.60s/step, loss=1.220, acc=0.718, acc_len=1.74, thru=15.6, I=21.0, T=36.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:43,737] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:09:51,700] loop.py:380 INFO TIMING step=10880: step=3.857s data=1.675s compute=3.831s [fwd=2.701s bwd=1.114s opt=0.003s] dispatch=4.255s [2026-10-19 13:09:51,700] loop.py:45 INFO TRAIN_STEP step=10880 loss=1.278411 Training: 60%|██████ | 10880/18018 [23:39:37<15:22:51, 7.76s/step, loss=1.278, acc=0.704, acc_len=1.70, thru=12.4, I=15.4, T=33.2, wait=4.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:58,272] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:09:59,586] loop.py:45 INFO TRAIN_STEP step=10881 loss=1.367818 Training: 60%|██████ | 10881/18018 [23:39:45<15:27:24, 7.80s/step, loss=1.368, acc=0.688, acc_len=1.64, thru=16.6, I=17.9, T=36.1, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:09:59,581] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10880: forward=2493.3ms backward=1009.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:09:59,789] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 13:10:05,389] loop.py:45 INFO TRAIN_STEP step=10882 loss=1.309861 Training: 60%|██████ | 10882/18018 [23:39:51<14:16:04, 7.20s/step, loss=1.310, acc=0.697, acc_len=1.67, thru=20.1, I=23.8, T=38.8, wait=2.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:11,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:10:12,054] loop.py:45 INFO TRAIN_STEP step=10883 loss=1.202000 Training: 60%|██████ | 10883/18018 [23:39:58<13:56:55, 7.04s/step, loss=1.202, acc=0.715, acc_len=1.73, thru=19.7, I=20.6, T=37.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:12,390] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:19,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:10:19,754] loop.py:45 INFO TRAIN_STEP step=10884 loss=1.394885 Training: 60%|██████ | 10884/18018 [23:40:05<14:20:27, 7.24s/step, loss=1.395, acc=0.681, acc_len=1.63, thru=16.8, I=18.5, T=38.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:19,956] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:24,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:10:25,666] loop.py:380 INFO TIMING step=10885: step=3.446s data=1.353s compute=3.410s [fwd=2.437s bwd=0.957s opt=0.003s] dispatch=2.457s [2026-10-19 13:10:25,666] loop.py:45 INFO TRAIN_STEP step=10885 loss=1.429730 Training: 60%|██████ | 10885/18018 [23:40:11<13:33:00, 6.84s/step, loss=1.430, acc=0.682, acc_len=1.62, thru=21.0, I=22.4, T=37.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:25,876] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:31,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:10:31,858] loop.py:45 INFO TRAIN_STEP step=10886 loss=1.104158 Training: 60%|██████ | 10886/18018 [23:40:17<13:09:54, 6.65s/step, loss=1.104, acc=0.735, acc_len=1.81, thru=20.4, I=23.1, T=36.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:10:31,853] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10885: forward=2461.3ms backward=957.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:32,073] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:37,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:10:38,659] loop.py:45 INFO TRAIN_STEP step=10887 loss=1.285553 Training: 60%|██████ | 10887/18018 [23:40:24<13:15:18, 6.69s/step, loss=1.286, acc=0.701, acc_len=1.69, thru=19.4, I=22.5, T=35.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:38,793] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 13:10:44,250] loop.py:45 INFO TRAIN_STEP step=10888 loss=1.295532 Training: 60%|██████ | 10888/18018 [23:40:30<12:35:56, 6.36s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=20.0, I=23.1, T=39.5, wait=2.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:50,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:10:51,189] loop.py:45 INFO TRAIN_STEP step=10889 loss=1.367228 Training: 60%|██████ | 10889/18018 [23:40:37<12:56:26, 6.53s/step, loss=1.367, acc=0.687, acc_len=1.64, thru=18.8, I=19.3, T=36.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:10:51,436] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:10:58,534] loop.py:380 INFO TIMING step=10890: step=3.567s data=1.494s compute=3.531s [fwd=2.524s bwd=0.992s opt=0.003s] dispatch=3.770s [2026-10-19 13:10:58,535] loop.py:45 INFO TRAIN_STEP step=10890 loss=1.241043 Training: 60%|██████ | 10890/18018 [23:40:44<13:25:11, 6.78s/step, loss=1.241, acc=0.710, acc_len=1.72, thru=16.9, I=18.6, T=35.9, wait=3.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:11:04,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:11:05,495] loop.py:45 INFO TRAIN_STEP step=10891 loss=1.379476 Training: 60%|██████ | 10891/18018 [23:40:51<13:31:38, 6.83s/step, loss=1.379, acc=0.688, acc_len=1.64, thru=17.6, I=20.5, T=37.9, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:11:05,491] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10890: forward=2389.9ms backward=931.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:11:05,584] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:11:12,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:11:13,263] loop.py:45 INFO TRAIN_STEP step=10892 loss=1.274068 Training: 60%|██████ | 10892/18018 [23:40:59<14:04:50, 7.11s/step, loss=1.274, acc=0.706, acc_len=1.70, thru=15.3, I=20.2, T=35.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:11:13,576] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:11:21,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:11:21,580] loop.py:45 INFO TRAIN_STEP step=10893 loss=1.255275 Training: 60%|██████ | 10893/18018 [23:41:07<14:47:34, 7.47s/step, loss=1.255, acc=0.709, acc_len=1.71, thru=15.8, I=18.0, T=36.3, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:11:21,631] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 13:11:31,078] loop.py:45 INFO TRAIN_STEP step=10894 loss=1.379881 Training: 60%|██████ | 10894/18018 [23:41:17<15:59:36, 8.08s/step, loss=1.380, acc=0.689, acc_len=1.64, thru=12.4, I=13.7, T=34.6, wait=5.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:11:41,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:11:41,966] loop.py:380 INFO TIMING step=10895: step=3.656s data=1.417s compute=3.618s [fwd=2.587s bwd=1.016s opt=0.003s] dispatch=7.222s [2026-10-19 13:11:41,966] loop.py:45 INFO TRAIN_STEP step=10895 loss=1.320081 Training: 60%|██████ | 10895/18018 [23:41:28<17:39:21, 8.92s/step, loss=1.320, acc=0.705, acc_len=1.69, thru=15.6, I=12.7, T=35.0, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:11:42,065] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 13:11:51,698] loop.py:45 INFO TRAIN_STEP step=10896 loss=1.353642 Training: 60%|██████ | 10896/18018 [23:41:37<18:07:57, 9.17s/step, loss=1.354, acc=0.689, acc_len=1.64, thru=12.5, I=12.7, T=34.2, wait=6.0s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:11:51,694] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10895: forward=2549.1ms backward=1143.1ms [2026-10-19 13:12:00,194] loop.py:45 INFO TRAIN_STEP step=10897 loss=1.395682 Training: 60%|██████ | 10897/18018 [23:41:46<17:44:04, 8.97s/step, loss=1.396, acc=0.687, acc_len=1.63, thru=15.2, I=17.4, T=28.6, wait=4.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:12:07,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:12:08,606] loop.py:45 INFO TRAIN_STEP step=10898 loss=1.242435 Training: 60%|██████ | 10898/18018 [23:41:54<17:24:13, 8.80s/step, loss=1.242, acc=0.710, acc_len=1.71, thru=14.4, I=15.6, T=33.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:12:08,644] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:12:13,943] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:12:14,603] loop.py:45 INFO TRAIN_STEP step=10899 loss=1.379310 Training: 60%|██████ | 10899/18018 [23:42:00<15:44:18, 7.96s/step, loss=1.379, acc=0.688, acc_len=1.64, thru=20.0, I=22.6, T=38.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:12:14,947] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:12:21,917] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:12:22,423] loop.py:380 INFO TIMING step=10900: step=3.526s data=1.507s compute=3.493s [fwd=2.495s bwd=0.983s opt=0.003s] dispatch=4.284s [2026-10-19 13:12:22,423] loop.py:45 INFO TRAIN_STEP step=10900 loss=1.383696 Training: 60%|██████ | 10900/18018 [23:42:08<15:39:09, 7.92s/step, loss=1.384, acc=0.687, acc_len=1.63, thru=17.2, I=19.6, T=36.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:12:22,923] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:12:28,853] loop.py:45 INFO TRAIN_STEP step=10901 loss=1.189329 Training: 61%|██████ | 10901/18018 [23:42:14<14:46:09, 7.47s/step, loss=1.189, acc=0.720, acc_len=1.75, thru=16.4, I=19.1, T=38.3, wait=3.1s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:12:28,849] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10900: forward=2377.7ms backward=922.2ms [2026-10-19 13:12:38,839] loop.py:45 INFO TRAIN_STEP step=10902 loss=1.391871 Training: 61%|██████ | 10902/18018 [23:42:24<16:15:28, 8.22s/step, loss=1.392, acc=0.687, acc_len=1.64, thru=15.3, I=13.9, T=33.9, wait=6.2s, pool=57, epoch=2/2][2026-10-19 13:12:48,157] loop.py:45 INFO TRAIN_STEP step=10903 loss=1.379685 Training: 61%|██████ | 10903/18018 [23:42:34<16:54:24, 8.55s/step, loss=1.380, acc=0.684, acc_len=1.63, thru=14.1, I=15.7, T=37.7, wait=5.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:12:54,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:12:56,182] loop.py:45 INFO TRAIN_STEP step=10904 loss=1.203650 Training: 61%|██████ | 10904/18018 [23:42:42<16:35:20, 8.39s/step, loss=1.204, acc=0.719, acc_len=1.75, thru=16.5, I=18.6, T=35.7, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:12:56,234] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 13:13:02,458] loop.py:380 INFO TIMING step=10905: step=3.370s data=1.381s compute=3.342s [fwd=2.380s bwd=0.948s opt=0.003s] dispatch=2.896s [2026-10-19 13:13:02,458] loop.py:45 INFO TRAIN_STEP step=10905 loss=1.278302 Training: 61%|██████ | 10905/18018 [23:42:48<15:19:44, 7.76s/step, loss=1.278, acc=0.705, acc_len=1.70, thru=17.6, I=22.3, T=38.0, wait=2.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:13:10,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:13:10,849] loop.py:45 INFO TRAIN_STEP step=10906 loss=1.309420 Training: 61%|██████ | 10906/18018 [23:42:56<15:42:09, 7.95s/step, loss=1.309, acc=0.702, acc_len=1.69, thru=18.2, I=17.4, T=35.1, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:13:10,846] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10905: forward=2579.7ms backward=1021.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:13:11,068] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:13:17,794] loop.py:45 INFO TRAIN_STEP step=10907 loss=1.384366 Training: 61%|██████ | 10907/18018 [23:43:03<15:06:19, 7.65s/step, loss=1.384, acc=0.687, acc_len=1.64, thru=17.6, I=19.5, T=34.3, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:13:23,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:13:24,926] loop.py:45 INFO TRAIN_STEP step=10908 loss=1.291213 Training: 61%|██████ | 10908/18018 [23:43:11<14:47:56, 7.49s/step, loss=1.291, acc=0.700, acc_len=1.68, thru=18.0, I=19.8, T=34.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:13:24,976] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:13:29,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:13:30,852] loop.py:45 INFO TRAIN_STEP step=10909 loss=1.278928 Training: 61%|██████ | 10909/18018 [23:43:16<13:52:13, 7.02s/step, loss=1.279, acc=0.705, acc_len=1.69, thru=20.9, I=26.5, T=37.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:13:30,957] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 13:13:36,863] loop.py:380 INFO TIMING step=10910: step=3.325s data=1.366s compute=3.295s [fwd=2.375s bwd=0.906s opt=0.003s] dispatch=2.672s [2026-10-19 13:13:36,863] loop.py:45 INFO TRAIN_STEP step=10910 loss=1.253618 Training: 61%|██████ | 10910/18018 [23:43:22<13:15:57, 6.72s/step, loss=1.254, acc=0.709, acc_len=1.71, thru=19.1, I=23.6, T=38.5, wait=2.7s, pool=49, epoch=2/2][2026-10-19 13:13:44,548] loop.py:45 INFO TRAIN_STEP step=10911 loss=1.398696 Training: 61%|██████ | 10911/18018 [23:43:30<13:50:08, 7.01s/step, loss=1.399, acc=0.684, acc_len=1.63, thru=15.6, I=15.1, T=35.4, wait=4.1s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:13:44,543] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10910: forward=2563.1ms backward=998.6ms [2026-10-19 13:13:51,636] loop.py:45 INFO TRAIN_STEP step=10912 loss=1.125191 Training: 61%|██████ | 10912/18018 [23:43:37<13:52:52, 7.03s/step, loss=1.125, acc=0.731, acc_len=1.79, thru=19.1, I=17.9, T=36.3, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:13:58,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:13:59,100] loop.py:45 INFO TRAIN_STEP step=10913 loss=1.443125 Training: 61%|██████ | 10913/18018 [23:43:45<14:08:06, 7.16s/step, loss=1.443, acc=0.673, acc_len=1.59, thru=17.9, I=17.9, T=36.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:13:59,322] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:14:05,719] loop.py:45 INFO TRAIN_STEP step=10914 loss=1.125195 Training: 61%|██████ | 10914/18018 [23:43:51<13:48:44, 7.00s/step, loss=1.125, acc=0.732, acc_len=1.79, thru=18.3, I=19.2, T=37.5, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:12,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:14:12,764] loop.py:380 INFO TIMING step=10915: step=3.590s data=1.543s compute=3.554s [fwd=2.519s bwd=1.020s opt=0.003s] dispatch=3.445s [2026-10-19 13:14:12,764] loop.py:45 INFO TRAIN_STEP step=10915 loss=1.315370 Training: 61%|██████ | 10915/18018 [23:43:58<13:50:11, 7.01s/step, loss=1.315, acc=0.697, acc_len=1.67, thru=18.8, I=19.8, T=35.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:13,029] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:20,818] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:14:21,427] loop.py:45 INFO TRAIN_STEP step=10916 loss=1.264159 Training: 61%|██████ | 10916/18018 [23:44:07<14:48:38, 7.51s/step, loss=1.264, acc=0.708, acc_len=1.71, thru=14.4, I=15.4, T=35.0, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:14:21,422] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10915: forward=2483.3ms backward=1112.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:21,823] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:30,999] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:14:31,815] loop.py:45 INFO TRAIN_STEP step=10917 loss=1.408010 Training: 61%|██████ | 10917/18018 [23:44:17<16:30:53, 8.37s/step, loss=1.408, acc=0.680, acc_len=1.61, thru=12.8, I=13.6, T=34.2, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:32,006] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:14:37,882] loop.py:45 INFO TRAIN_STEP step=10918 loss=1.293091 Training: 61%|██████ | 10918/18018 [23:44:23<15:08:50, 7.68s/step, loss=1.293, acc=0.700, acc_len=1.68, thru=17.9, I=20.4, T=36.3, wait=2.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:44,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:14:45,084] loop.py:45 INFO TRAIN_STEP step=10919 loss=1.229854 Training: 61%|██████ | 10919/18018 [23:44:31<14:51:46, 7.54s/step, loss=1.230, acc=0.714, acc_len=1.73, thru=19.6, I=18.2, T=36.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:45,206] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:51,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:14:51,687] loop.py:380 INFO TIMING step=10920: step=3.539s data=1.519s compute=3.501s [fwd=2.478s bwd=1.008s opt=0.003s] dispatch=3.055s [2026-10-19 13:14:51,687] loop.py:45 INFO TRAIN_STEP step=10920 loss=1.310703 Training: 61%|██████ | 10920/18018 [23:44:37<14:18:26, 7.26s/step, loss=1.311, acc=0.699, acc_len=1.68, thru=19.5, I=20.5, T=36.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:51,738] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:56,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:14:57,299] loop.py:45 INFO TRAIN_STEP step=10921 loss=1.320851 Training: 61%|██████ | 10921/18018 [23:44:43<13:20:00, 6.76s/step, loss=1.321, acc=0.697, acc_len=1.67, thru=21.1, I=22.4, T=38.0, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:14:57,294] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10920: forward=2383.2ms backward=940.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:14:57,466] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 13:15:04,625] loop.py:45 INFO TRAIN_STEP step=10922 loss=1.178775 Training: 61%|██████ | 10922/18018 [23:44:50<13:39:47, 6.93s/step, loss=1.179, acc=0.722, acc_len=1.76, thru=17.4, I=17.2, T=37.3, wait=3.9s, pool=61, epoch=2/2][2026-10-19 13:15:12,553] loop.py:45 INFO TRAIN_STEP step=10923 loss=1.234144 Training: 61%|██████ | 10923/18018 [23:44:58<14:15:01, 7.23s/step, loss=1.234, acc=0.716, acc_len=1.74, thru=15.0, I=13.9, T=36.2, wait=4.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:19,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:15:19,662] loop.py:45 INFO TRAIN_STEP step=10924 loss=1.354880 Training: 61%|██████ | 10924/18018 [23:45:05<14:10:34, 7.19s/step, loss=1.355, acc=0.691, acc_len=1.65, thru=19.9, I=18.1, T=36.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:20,056] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:26,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:15:27,153] loop.py:380 INFO TIMING step=10925: step=3.847s data=1.660s compute=3.811s [fwd=2.707s bwd=1.089s opt=0.003s] dispatch=3.635s [2026-10-19 13:15:27,153] loop.py:45 INFO TRAIN_STEP step=10925 loss=1.181421 Training: 61%|██████ | 10925/18018 [23:45:13<14:21:02, 7.28s/step, loss=1.181, acc=0.721, acc_len=1.76, thru=17.7, I=19.2, T=33.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:27,616] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:34,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:15:35,418] loop.py:45 INFO TRAIN_STEP step=10926 loss=1.399881 Training: 61%|██████ | 10926/18018 [23:45:21<14:55:41, 7.58s/step, loss=1.400, acc=0.685, acc_len=1.63, thru=15.0, I=16.3, T=35.4, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:35,466] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (TrainerActor pid=9310) [2026-10-19 13:15:35,413] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10925: forward=2521.3ms backward=1034.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:40,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:15:40,802] loop.py:45 INFO TRAIN_STEP step=10927 loss=1.361474 Training: 61%|██████ | 10927/18018 [23:45:26<13:37:50, 6.92s/step, loss=1.361, acc=0.693, acc_len=1.65, thru=21.6, I=20.4, T=36.8, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:41,127] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:46,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:15:47,182] loop.py:45 INFO TRAIN_STEP step=10928 loss=1.423498 Training: 61%|██████ | 10928/18018 [23:45:33<13:18:31, 6.76s/step, loss=1.423, acc=0.681, acc_len=1.62, thru=20.1, I=22.1, T=37.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:47,362] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:54,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:15:55,141] loop.py:45 INFO TRAIN_STEP step=10929 loss=1.334082 Training: 61%|██████ | 10929/18018 [23:45:41<14:00:57, 7.12s/step, loss=1.334, acc=0.695, acc_len=1.66, thru=16.1, I=19.5, T=32.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:15:55,638] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:16:03,248] loop.py:380 INFO TIMING step=10930: step=3.628s data=1.516s compute=3.587s [fwd=2.538s bwd=1.034s opt=0.003s] dispatch=4.471s [2026-10-19 13:16:03,248] loop.py:45 INFO TRAIN_STEP step=10930 loss=1.322209 Training: 61%|██████ | 10930/18018 [23:45:49<14:35:59, 7.42s/step, loss=1.322, acc=0.692, acc_len=1.66, thru=13.7, I=15.7, T=35.3, wait=4.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:11,845] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:16:13,119] loop.py:45 INFO TRAIN_STEP step=10931 loss=1.265178 Training: 61%|██████ | 10931/18018 [23:45:59<16:02:53, 8.15s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=14.3, I=15.0, T=30.3, wait=5.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:16:13,114] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10930: forward=3064.7ms backward=1115.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:13,354] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:16:19,889] loop.py:45 INFO TRAIN_STEP step=10932 loss=1.376451 Training: 61%|██████ | 10932/18018 [23:46:05<15:13:50, 7.74s/step, loss=1.376, acc=0.686, acc_len=1.63, thru=17.3, I=19.3, T=36.0, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:27,670] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:16:28,548] loop.py:45 INFO TRAIN_STEP step=10933 loss=1.309359 Training: 61%|██████ | 10933/18018 [23:46:14<15:46:18, 8.01s/step, loss=1.309, acc=0.699, acc_len=1.68, thru=15.8, I=16.2, T=33.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:28,692] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:33,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:16:34,097] loop.py:45 INFO TRAIN_STEP step=10934 loss=1.446839 Training: 61%|██████ | 10934/18018 [23:46:20<14:18:53, 7.27s/step, loss=1.447, acc=0.679, acc_len=1.60, thru=20.4, I=21.6, T=38.3, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:34,493] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:40,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:16:41,547] loop.py:380 INFO TIMING step=10935: step=3.661s data=1.540s compute=3.624s [fwd=2.622s bwd=0.988s opt=0.003s] dispatch=3.778s [2026-10-19 13:16:41,548] loop.py:45 INFO TRAIN_STEP step=10935 loss=1.336158 Training: 61%|██████ | 10935/18018 [23:46:27<14:24:56, 7.33s/step, loss=1.336, acc=0.696, acc_len=1.67, thru=17.9, I=19.4, T=35.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:41,891] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:46,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:16:47,655] loop.py:45 INFO TRAIN_STEP step=10936 loss=1.323793 Training: 61%|██████ | 10936/18018 [23:46:33<13:41:32, 6.96s/step, loss=1.324, acc=0.695, acc_len=1.67, thru=20.2, I=24.7, T=37.3, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:16:47,651] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10935: forward=2435.9ms backward=950.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:48,122] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:52,971] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:16:53,718] loop.py:45 INFO TRAIN_STEP step=10937 loss=1.232775 Training: 61%|██████ | 10937/18018 [23:46:39<13:09:47, 6.69s/step, loss=1.233, acc=0.709, acc_len=1.71, thru=20.9, I=26.4, T=39.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:53,979] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:59,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:16:59,451] loop.py:45 INFO TRAIN_STEP step=10938 loss=1.312724 Training: 61%|██████ | 10938/18018 [23:46:45<12:35:40, 6.40s/step, loss=1.313, acc=0.699, acc_len=1.68, thru=22.8, I=24.7, T=37.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:16:59,648] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:17:05,001] loop.py:45 INFO TRAIN_STEP step=10939 loss=1.262283 Training: 61%|██████ | 10939/18018 [23:46:51<12:05:22, 6.15s/step, loss=1.262, acc=0.710, acc_len=1.71, thru=20.6, I=26.8, T=37.2, wait=2.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:12,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:17:13,291] loop.py:380 INFO TIMING step=10940: step=3.499s data=1.377s compute=3.474s [fwd=2.492s bwd=0.969s opt=0.003s] dispatch=4.781s [2026-10-19 13:17:13,291] loop.py:45 INFO TRAIN_STEP step=10940 loss=1.279997 Training: 61%|██████ | 10940/18018 [23:46:59<13:21:04, 6.79s/step, loss=1.280, acc=0.703, acc_len=1.69, thru=17.4, I=17.5, T=36.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:13,635] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:18,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:17:19,150] loop.py:45 INFO TRAIN_STEP step=10941 loss=1.135605 Training: 61%|██████ | 10941/18018 [23:47:05<12:47:55, 6.51s/step, loss=1.136, acc=0.729, acc_len=1.78, thru=21.4, I=26.8, T=38.6, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:17:19,146] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10940: forward=2353.6ms backward=920.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:19,633] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:17:25,420] loop.py:45 INFO TRAIN_STEP step=10942 loss=1.193432 Training: 61%|██████ | 10942/18018 [23:47:11<12:39:17, 6.44s/step, loss=1.193, acc=0.722, acc_len=1.76, thru=19.7, I=21.6, T=38.4, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:37,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:17:39,000] loop.py:45 INFO TRAIN_STEP step=10943 loss=1.450442 Training: 61%|██████ | 10943/18018 [23:47:25<16:51:58, 8.58s/step, loss=1.450, acc=0.675, acc_len=1.60, thru=15.7, I=9.8, T=32.1, wait=9.6s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:39,224] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:44,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:17:45,093] loop.py:45 INFO TRAIN_STEP step=10944 loss=1.407089 Training: 61%|██████ | 10944/18018 [23:47:31<15:23:42, 7.83s/step, loss=1.407, acc=0.680, acc_len=1.61, thru=19.0, I=23.9, T=39.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:45,390] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:49,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:17:50,755] loop.py:380 INFO TIMING step=10945: step=3.332s data=1.380s compute=3.298s [fwd=2.380s bwd=0.904s opt=0.003s] dispatch=2.320s [2026-10-19 13:17:50,755] loop.py:45 INFO TRAIN_STEP step=10945 loss=1.196521 Training: 61%|██████ | 10945/18018 [23:47:36<14:06:44, 7.18s/step, loss=1.197, acc=0.719, acc_len=1.74, thru=22.5, I=25.5, T=38.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:17:50,869] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:18:01,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:18:01,458] loop.py:45 INFO TRAIN_STEP step=10946 loss=1.274743 Training: 61%|██████ | 10946/18018 [23:47:47<16:11:02, 8.24s/step, loss=1.275, acc=0.705, acc_len=1.70, thru=12.1, I=12.4, T=27.9, wait=6.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:18:01,453] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10945: forward=3492.0ms backward=1054.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:18:01,549] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 13:18:09,733] loop.py:45 INFO TRAIN_STEP step=10947 loss=1.330584 Training: 61%|██████ | 10947/18018 [23:47:55<16:12:11, 8.25s/step, loss=1.331, acc=0.695, acc_len=1.66, thru=15.7, I=15.5, T=33.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:18:09,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:18:10,181] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:18:16,557] loop.py:45 INFO TRAIN_STEP step=10948 loss=1.307624 Training: 61%|██████ | 10948/18018 [23:48:02<15:21:42, 7.82s/step, loss=1.308, acc=0.698, acc_len=1.67, thru=18.1, I=18.6, T=36.4, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:18:21,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:18:23,457] loop.py:45 INFO TRAIN_STEP step=10949 loss=1.349467 Training: 61%|██████ | 10949/18018 [23:48:09<14:48:57, 7.55s/step, loss=1.349, acc=0.692, acc_len=1.65, thru=18.1, I=22.3, T=32.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:18:23,630] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:18:28,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:18:29,517] loop.py:380 INFO TIMING step=10950: step=3.643s data=1.519s compute=3.612s [fwd=2.526s bwd=1.072s opt=0.003s] dispatch=2.407s [2026-10-19 13:18:29,517] loop.py:45 INFO TRAIN_STEP step=10950 loss=1.166282 Training: 61%|██████ | 10950/18018 [23:48:15<13:56:23, 7.10s/step, loss=1.166, acc=0.722, acc_len=1.75, thru=19.3, I=28.2, T=35.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:18:29,550] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 13:18:38,136] loop.py:45 INFO TRAIN_STEP step=10951 loss=1.404148 Training: 61%|██████ | 10951/18018 [23:48:24<14:49:53, 7.56s/step, loss=1.404, acc=0.684, acc_len=1.63, thru=12.6, I=12.5, T=30.6, wait=4.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:18:38,133] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10950: forward=2849.7ms backward=1236.8ms [2026-10-19 13:18:46,641] loop.py:45 INFO TRAIN_STEP step=10952 loss=1.336720 Training: 61%|██████ | 10952/18018 [23:48:32<15:23:16, 7.84s/step, loss=1.337, acc=0.697, acc_len=1.67, thru=16.6, I=13.7, T=37.1, wait=5.0s, pool=53, epoch=2/2][2026-10-19 13:18:54,069] loop.py:45 INFO TRAIN_STEP step=10953 loss=1.347969 Training: 61%|██████ | 10953/18018 [23:48:40<15:08:41, 7.72s/step, loss=1.348, acc=0.694, acc_len=1.66, thru=18.5, I=18.9, T=34.2, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:01,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:19:01,279] loop.py:45 INFO TRAIN_STEP step=10954 loss=1.233642 Training: 61%|██████ | 10954/18018 [23:48:47<14:50:37, 7.56s/step, loss=1.234, acc=0.710, acc_len=1.72, thru=17.3, I=18.2, T=36.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:01,564] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:07,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:19:08,029] loop.py:380 INFO TIMING step=10955: step=3.468s data=1.467s compute=3.442s [fwd=2.460s bwd=0.967s opt=0.003s] dispatch=3.273s [2026-10-19 13:19:08,029] loop.py:45 INFO TRAIN_STEP step=10955 loss=1.390096 Training: 61%|██████ | 10955/18018 [23:48:54<14:21:41, 7.32s/step, loss=1.390, acc=0.685, acc_len=1.63, thru=18.9, I=20.0, T=36.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:08,335] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:15,373] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:19:16,128] loop.py:45 INFO TRAIN_STEP step=10956 loss=1.246502 (TrainerActor pid=9310) [2026-10-19 13:19:16,123] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10955: forward=2554.2ms backward=963.5ms Training: 61%|██████ | 10956/18018 [23:49:02<14:49:07, 7.55s/step, loss=1.247, acc=0.712, acc_len=1.73, thru=17.7, I=18.2, T=35.9, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:16,378] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:19:22,656] loop.py:45 INFO TRAIN_STEP step=10957 loss=1.240404 Training: 61%|██████ | 10957/18018 [23:49:08<14:12:43, 7.25s/step, loss=1.240, acc=0.707, acc_len=1.71, thru=20.1, I=22.8, T=38.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:22,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:23,115] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:19:30,740] loop.py:45 INFO TRAIN_STEP step=10958 loss=1.343771 Training: 61%|██████ | 10958/18018 [23:49:16<14:42:10, 7.50s/step, loss=1.344, acc=0.692, acc_len=1.65, thru=14.4, I=17.0, T=34.6, wait=4.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:38,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:19:38,493] loop.py:45 INFO TRAIN_STEP step=10959 loss=1.322320 Training: 61%|██████ | 10959/18018 [23:49:24<14:51:01, 7.57s/step, loss=1.322, acc=0.698, acc_len=1.67, thru=15.7, I=16.2, T=34.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:38,960] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:44,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:19:45,247] loop.py:380 INFO TIMING step=10960: step=3.583s data=1.523s compute=3.558s [fwd=2.539s bwd=1.005s opt=0.003s] dispatch=3.164s [2026-10-19 13:19:45,248] loop.py:45 INFO TRAIN_STEP step=10960 loss=1.259113 Training: 61%|██████ | 10960/18018 [23:49:31<14:22:03, 7.33s/step, loss=1.259, acc=0.710, acc_len=1.72, thru=18.3, I=20.1, T=35.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:45,700] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:52,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:19:52,837] loop.py:45 INFO TRAIN_STEP step=10961 loss=1.307559 Training: 61%|██████ | 10961/18018 [23:49:38<14:31:09, 7.41s/step, loss=1.308, acc=0.697, acc_len=1.66, thru=16.9, I=19.4, T=38.0, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:19:52,832] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10960: forward=2350.2ms backward=971.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:52,970] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:58,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:19:58,819] loop.py:45 INFO TRAIN_STEP step=10962 loss=1.344499 Training: 61%|██████ | 10962/18018 [23:49:44<13:40:46, 6.98s/step, loss=1.344, acc=0.693, acc_len=1.66, thru=20.6, I=21.5, T=37.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:19:59,026] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:03,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:20:05,155] loop.py:45 INFO TRAIN_STEP step=10963 loss=1.396455 Training: 61%|██████ | 10963/18018 [23:49:51<13:18:00, 6.79s/step, loss=1.396, acc=0.690, acc_len=1.65, thru=20.1, I=23.9, T=36.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:05,502] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:20:12,153] loop.py:45 INFO TRAIN_STEP step=10964 loss=1.338209 Training: 61%|██████ | 10964/18018 [23:49:58<13:25:19, 6.85s/step, loss=1.338, acc=0.690, acc_len=1.65, thru=15.7, I=19.5, T=36.7, wait=3.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:18,755] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:20:19,279] loop.py:380 INFO TIMING step=10965: step=3.577s data=1.401s compute=3.546s [fwd=2.538s bwd=0.992s opt=0.003s] dispatch=3.540s [2026-10-19 13:20:19,279] loop.py:45 INFO TRAIN_STEP step=10965 loss=1.366113 Training: 61%|██████ | 10965/18018 [23:50:05<13:34:54, 6.93s/step, loss=1.366, acc=0.691, acc_len=1.65, thru=18.8, I=18.9, T=35.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:19,761] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:25,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:20:25,969] loop.py:45 INFO TRAIN_STEP step=10966 loss=1.240756 Training: 61%|██████ | 10966/18018 [23:50:12<13:26:18, 6.86s/step, loss=1.241, acc=0.711, acc_len=1.72, thru=18.6, I=22.5, T=34.9, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:20:25,965] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10965: forward=2623.8ms backward=993.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:26,183] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:20:32,195] loop.py:45 INFO TRAIN_STEP step=10967 loss=1.304665 Training: 61%|██████ | 10967/18018 [23:50:18<13:03:45, 6.67s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=15.7, I=24.5, T=36.1, wait=2.7s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:41,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:20:42,877] loop.py:45 INFO TRAIN_STEP step=10968 loss=1.348634 Training: 61%|██████ | 10968/18018 [23:50:28<15:25:06, 7.87s/step, loss=1.349, acc=0.698, acc_len=1.68, thru=16.5, I=15.1, T=25.9, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:42,909] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 13:20:49,869] loop.py:45 INFO TRAIN_STEP step=10969 loss=1.332651 Training: 61%|██████ | 10969/18018 [23:50:35<14:53:55, 7.61s/step, loss=1.333, acc=0.694, acc_len=1.67, thru=14.1, I=16.4, T=36.5, wait=3.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:57,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:20:58,231] loop.py:380 INFO TIMING step=10970: step=3.907s data=1.757s compute=3.870s [fwd=2.843s bwd=1.012s opt=0.003s] dispatch=4.446s [2026-10-19 13:20:58,231] loop.py:45 INFO TRAIN_STEP step=10970 loss=1.371163 Training: 61%|██████ | 10970/18018 [23:50:44<15:20:17, 7.83s/step, loss=1.371, acc=0.689, acc_len=1.65, thru=17.9, I=18.8, T=32.8, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:20:58,729] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:21:04,487] loop.py:45 INFO TRAIN_STEP step=10971 loss=1.349659 Training: 61%|██████ | 10971/18018 [23:50:50<14:24:33, 7.36s/step, loss=1.350, acc=0.690, acc_len=1.64, thru=18.0, I=18.5, T=38.0, wait=2.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:21:04,482] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10970: forward=2390.1ms backward=935.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:21:12,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:21:12,891] loop.py:45 INFO TRAIN_STEP step=10972 loss=1.335182 Training: 61%|██████ | 10972/18018 [23:50:58<15:01:12, 7.67s/step, loss=1.335, acc=0.693, acc_len=1.66, thru=17.9, I=17.5, T=35.3, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:21:13,295] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:21:22,065] loop.py:45 INFO TRAIN_STEP step=10973 loss=1.437359 Training: 61%|██████ | 10973/18018 [23:51:08<15:54:01, 8.13s/step, loss=1.437, acc=0.675, acc_len=1.59, thru=13.1, I=16.2, T=34.0, wait=5.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:21:29,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:21:29,331] loop.py:45 INFO TRAIN_STEP step=10974 loss=1.245968 Training: 61%|██████ | 10974/18018 [23:51:15<15:23:31, 7.87s/step, loss=1.246, acc=0.711, acc_len=1.71, thru=16.4, I=16.6, T=36.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:21:29,654] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:21:35,944] loop.py:380 INFO TIMING step=10975: step=3.592s data=1.492s compute=3.559s [fwd=2.492s bwd=1.053s opt=0.003s] dispatch=3.012s [2026-10-19 13:21:35,944] loop.py:45 INFO TRAIN_STEP step=10975 loss=1.238349 Training: 61%|██████ | 10975/18018 [23:51:22<14:39:17, 7.49s/step, loss=1.238, acc=0.709, acc_len=1.71, thru=19.0, I=18.2, T=35.6, wait=3.0s, pool=49, epoch=2/2][2026-10-19 13:21:42,464] loop.py:45 INFO TRAIN_STEP step=10976 loss=1.378565 Training: 61%|██████ | 10976/18018 [23:51:28<14:05:07, 7.20s/step, loss=1.379, acc=0.687, acc_len=1.63, thru=19.1, I=19.7, T=34.9, wait=2.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:21:42,459] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10975: forward=2586.0ms backward=1038.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:21:48,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:21:48,801] loop.py:45 INFO TRAIN_STEP step=10977 loss=1.275435 Training: 61%|██████ | 10977/18018 [23:51:34<13:34:28, 6.94s/step, loss=1.275, acc=0.702, acc_len=1.69, thru=19.2, I=21.1, T=37.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:21:49,152] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:21:54,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:21:55,741] loop.py:45 INFO TRAIN_STEP step=10978 loss=1.268121 Training: 61%|██████ | 10978/18018 [23:51:41<13:34:19, 6.94s/step, loss=1.268, acc=0.707, acc_len=1.71, thru=18.8, I=22.6, T=32.5, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:21:55,811] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:01,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:22:02,917] loop.py:45 INFO TRAIN_STEP step=10979 loss=1.283053 Training: 61%|██████ | 10979/18018 [23:51:49<13:42:29, 7.01s/step, loss=1.283, acc=0.702, acc_len=1.69, thru=16.9, I=19.7, T=35.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:03,417] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:10,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:22:11,168] loop.py:380 INFO TIMING step=10980: step=3.779s data=1.758s compute=3.739s [fwd=2.733s bwd=0.992s opt=0.003s] dispatch=4.463s [2026-10-19 13:22:11,169] loop.py:45 INFO TRAIN_STEP step=10980 loss=1.235931 Training: 61%|██████ | 10980/18018 [23:51:57<14:26:01, 7.38s/step, loss=1.236, acc=0.707, acc_len=1.70, thru=14.9, I=18.1, T=33.9, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:11,608] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:17,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:22:18,076] loop.py:45 INFO TRAIN_STEP step=10981 loss=1.315142 Training: 61%|██████ | 10981/18018 [23:52:04<14:09:09, 7.24s/step, loss=1.315, acc=0.702, acc_len=1.69, thru=18.8, I=21.5, T=30.7, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:22:18,071] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10980: forward=3111.0ms backward=1001.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:18,272] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:23,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:22:24,834] loop.py:45 INFO TRAIN_STEP step=10982 loss=1.345198 Training: 61%|██████ | 10982/18018 [23:52:10<13:52:03, 7.10s/step, loss=1.345, acc=0.690, acc_len=1.64, thru=18.2, I=22.0, T=30.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:25,275] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:32,525] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:22:32,875] loop.py:45 INFO TRAIN_STEP step=10983 loss=1.394181 Training: 61%|██████ | 10983/18018 [23:52:18<14:25:17, 7.38s/step, loss=1.394, acc=0.680, acc_len=1.61, thru=14.8, I=16.3, T=30.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:33,038] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:38,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:22:39,159] loop.py:45 INFO TRAIN_STEP step=10984 loss=1.364473 Training: 61%|██████ | 10984/18018 [23:52:25<13:46:34, 7.05s/step, loss=1.364, acc=0.689, acc_len=1.65, thru=19.6, I=21.4, T=36.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:39,382] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:51,669] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:22:52,769] loop.py:380 INFO TIMING step=10985: step=3.662s data=1.549s compute=3.625s [fwd=2.561s bwd=1.050s opt=0.003s] dispatch=9.937s [2026-10-19 13:22:52,769] loop.py:45 INFO TRAIN_STEP step=10985 loss=1.319314 Training: 61%|██████ | 10985/18018 [23:52:38<17:37:04, 9.02s/step, loss=1.319, acc=0.699, acc_len=1.68, thru=8.5, I=13.6, T=35.0, wait=9.9s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:22:53,179] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:22:58,953] loop.py:45 INFO TRAIN_STEP step=10986 loss=1.511585 Training: 61%|██████ | 10986/18018 [23:52:45<15:57:20, 8.17s/step, loss=1.512, acc=0.663, acc_len=1.56, thru=19.0, I=23.2, T=34.8, wait=2.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:22:58,948] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10985: forward=2584.2ms backward=1044.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:23:05,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:23:05,849] loop.py:45 INFO TRAIN_STEP step=10987 loss=1.244718 Training: 61%|██████ | 10987/18018 [23:52:51<15:12:26, 7.79s/step, loss=1.245, acc=0.711, acc_len=1.72, thru=18.2, I=19.8, T=34.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:23:06,151] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:23:14,125] loop.py:45 INFO TRAIN_STEP step=10988 loss=1.482886 Training: 61%|██████ | 10988/18018 [23:53:00<15:29:31, 7.93s/step, loss=1.483, acc=0.671, acc_len=1.58, thru=14.2, I=16.5, T=33.3, wait=4.4s, pool=61, epoch=2/2][2026-10-19 13:23:21,300] loop.py:45 INFO TRAIN_STEP step=10989 loss=1.345918 Training: 61%|██████ | 10989/18018 [23:53:07<15:02:44, 7.71s/step, loss=1.346, acc=0.691, acc_len=1.65, thru=15.3, I=16.1, T=34.9, wait=3.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:23:28,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:23:30,930] loop.py:380 INFO TIMING step=10990: step=3.950s data=1.764s compute=3.914s [fwd=2.846s bwd=1.054s opt=0.003s] dispatch=5.670s [2026-10-19 13:23:30,930] loop.py:45 INFO TRAIN_STEP step=10990 loss=1.361276 Training: 61%|██████ | 10990/18018 [23:53:17<16:10:11, 8.28s/step, loss=1.361, acc=0.692, acc_len=1.65, thru=15.7, I=18.6, T=32.4, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:23:31,347] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:23:36,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:23:37,131] loop.py:45 INFO TRAIN_STEP step=10991 loss=1.278447 Training: 61%|██████ | 10991/18018 [23:53:23<14:56:54, 7.66s/step, loss=1.278, acc=0.706, acc_len=1.70, thru=18.8, I=28.0, T=37.2, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:23:37,127] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10990: forward=2440.7ms backward=949.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:23:37,332] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:23:43,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:23:44,020] loop.py:45 INFO TRAIN_STEP step=10992 loss=1.376622 Training: 61%|██████ | 10992/18018 [23:53:30<14:29:42, 7.43s/step, loss=1.377, acc=0.684, acc_len=1.62, thru=18.6, I=21.6, T=34.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:23:44,434] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:23:50,780] loop.py:45 INFO TRAIN_STEP step=10993 loss=1.381275 Training: 61%|██████ | 10993/18018 [23:53:36<14:06:12, 7.23s/step, loss=1.381, acc=0.687, acc_len=1.64, thru=17.7, I=17.9, T=35.9, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:23:59,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:24:00,339] loop.py:45 INFO TRAIN_STEP step=10994 loss=1.256954 Training: 61%|██████ | 10994/18018 [23:53:46<15:27:56, 7.93s/step, loss=1.257, acc=0.708, acc_len=1.71, thru=14.7, I=16.7, T=34.0, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:24:00,765] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:24:05,799] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:24:07,092] loop.py:380 INFO TIMING step=10995: step=3.823s data=1.684s compute=3.798s [fwd=2.717s bwd=1.066s opt=0.003s] dispatch=2.921s [2026-10-19 13:24:07,092] loop.py:45 INFO TRAIN_STEP step=10995 loss=1.257856 Training: 61%|██████ | 10995/18018 [23:53:53<14:46:37, 7.57s/step, loss=1.258, acc=0.708, acc_len=1.71, thru=18.7, I=23.6, T=33.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:24:07,310] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:24:14,009] loop.py:45 INFO TRAIN_STEP step=10996 loss=1.140364 Training: 61%|██████ | 10996/18018 [23:54:00<14:23:23, 7.38s/step, loss=1.140, acc=0.733, acc_len=1.80, thru=15.3, I=18.1, T=36.8, wait=3.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:24:14,005] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=10995: forward=2491.1ms backward=936.6ms [2026-10-19 13:24:24,147] loop.py:45 INFO TRAIN_STEP step=10997 loss=1.311090 Training: 61%|██████ | 10997/18018 [23:54:10<16:00:22, 8.21s/step, loss=1.311, acc=0.699, acc_len=1.68, thru=11.6, I=13.0, T=35.2, wait=6.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:24:31,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:24:33,911] loop.py:45 INFO TRAIN_STEP step=10998 loss=1.315462 Training: 61%|██████ | 10998/18018 [23:54:19<16:54:42, 8.67s/step, loss=1.315, acc=0.704, acc_len=1.70, thru=17.7, I=17.7, T=30.7, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:24:34,309] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 13:24:40,714] loop.py:45 INFO TRAIN_STEP step=10999 loss=1.263676 Training: 61%|██████ | 10999/18018 [23:54:26<15:48:53, 8.11s/step, loss=1.264, acc=0.706, acc_len=1.71, thru=14.8, I=21.5, T=34.6, wait=3.1s, pool=49, epoch=2/2][2026-10-19 13:24:49,409] loop.py:380 INFO TIMING step=11000: step=3.529s data=1.418s compute=3.502s [fwd=2.503s bwd=0.984s opt=0.003s] dispatch=5.158s [2026-10-19 13:24:49,409] loop.py:45 INFO TRAIN_STEP step=11000 loss=1.337130 Training: 61%|██████ | 11000/18018 [23:54:35<16:09:16, 8.29s/step, loss=1.337, acc=0.692, acc_len=1.65, thru=14.0, I=13.7, T=36.3, wait=5.2s, pool=37, epoch=2/2][2026-10-19 13:24:49,412] loop.py:419 INFO Saving checkpoint at step 11000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:24:50,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 13:24:52,844] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0011001 (TrainerActor pid=9310) [2026-10-19 13:24:52,853] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0010251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:24:54,071] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:01,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 13:24:52,845] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0011001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 13:24:53,865] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0010251 [2026-10-19 13:25:01,912] loop.py:45 INFO TRAIN_STEP step=11001 loss=1.193704 Training: 61%|██████ | 11001/18018 [23:54:47<18:37:07, 9.55s/step, loss=1.194, acc=0.717, acc_len=1.75, thru=16.3, I=17.2, T=33.0, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:25:01,908] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11000: forward=2773.4ms backward=1057.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:01,968] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:08,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:25:09,249] loop.py:45 INFO TRAIN_STEP step=11002 loss=1.221181 Training: 61%|██████ | 11002/18018 [23:54:55<17:19:14, 8.89s/step, loss=1.221, acc=0.715, acc_len=1.74, thru=17.6, I=19.4, T=33.3, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:09,566] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:16,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:25:16,952] loop.py:45 INFO TRAIN_STEP step=11003 loss=1.404011 Training: 61%|██████ | 11003/18018 [23:55:03<16:37:32, 8.53s/step, loss=1.404, acc=0.682, acc_len=1.62, thru=16.9, I=20.7, T=35.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:16,985] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:24,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:25:25,252] loop.py:45 INFO TRAIN_STEP step=11004 loss=1.400000 Training: 61%|██████ | 11004/18018 [23:55:11<16:29:13, 8.46s/step, loss=1.400, acc=0.685, acc_len=1.63, thru=15.8, I=17.4, T=35.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:25,313] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:32,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:25:34,261] loop.py:380 INFO TIMING step=11005: step=4.059s data=1.738s compute=4.029s [fwd=2.895s bwd=1.118s opt=0.003s] dispatch=4.942s [2026-10-19 13:25:34,262] loop.py:45 INFO TRAIN_STEP step=11005 loss=1.337946 Training: 61%|██████ | 11005/18018 [23:55:20<16:48:13, 8.63s/step, loss=1.338, acc=0.695, acc_len=1.67, thru=13.6, I=16.7, T=31.5, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:34,585] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:25:40,828] loop.py:45 INFO TRAIN_STEP step=11006 loss=1.342039 Training: 61%|██████ | 11006/18018 [23:55:26<15:36:00, 8.01s/step, loss=1.342, acc=0.696, acc_len=1.66, thru=16.0, I=19.3, T=35.0, wait=2.9s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:25:40,824] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11005: forward=2563.3ms backward=1043.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:51,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:25:51,266] loop.py:45 INFO TRAIN_STEP step=11007 loss=1.385691 Training: 61%|██████ | 11007/18018 [23:55:37<17:00:56, 8.74s/step, loss=1.386, acc=0.686, acc_len=1.63, thru=14.4, I=14.9, T=30.1, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:51,559] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:57,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:25:58,071] loop.py:45 INFO TRAIN_STEP step=11008 loss=1.185613 Training: 61%|██████ | 11008/18018 [23:55:44<15:53:04, 8.16s/step, loss=1.186, acc=0.719, acc_len=1.75, thru=16.9, I=17.3, T=35.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:25:58,321] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:26:04,918] loop.py:45 INFO TRAIN_STEP step=11009 loss=1.292535 Training: 61%|██████ | 11009/18018 [23:55:51<15:07:05, 7.77s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=17.6, I=18.6, T=36.4, wait=3.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:26:12,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:26:13,674] loop.py:380 INFO TIMING step=11010: step=3.736s data=1.612s compute=3.701s [fwd=2.667s bwd=1.020s opt=0.003s] dispatch=5.009s [2026-10-19 13:26:13,675] loop.py:45 INFO TRAIN_STEP step=11010 loss=1.457230 Training: 61%|██████ | 11010/18018 [23:55:59<15:41:36, 8.06s/step, loss=1.457, acc=0.672, acc_len=1.59, thru=16.1, I=17.2, T=34.3, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:26:13,886] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:26:18,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:26:19,838] loop.py:45 INFO TRAIN_STEP step=11011 loss=1.203378 Training: 61%|██████ | 11011/18018 [23:56:05<14:34:58, 7.49s/step, loss=1.203, acc=0.716, acc_len=1.74, thru=19.7, I=21.8, T=34.9, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:26:19,833] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11010: forward=2505.7ms backward=1118.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:26:19,866] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 13:26:26,700] loop.py:45 INFO TRAIN_STEP step=11012 loss=1.415216 Training: 61%|██████ | 11012/18018 [23:56:12<14:12:49, 7.30s/step, loss=1.415, acc=0.681, acc_len=1.62, thru=17.8, I=18.5, T=35.6, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:26:34,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:26:34,982] loop.py:45 INFO TRAIN_STEP step=11013 loss=1.305147 Training: 61%|██████ | 11013/18018 [23:56:21<14:46:56, 7.60s/step, loss=1.305, acc=0.702, acc_len=1.69, thru=17.2, I=17.5, T=33.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:26:35,347] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:26:43,167] loop.py:45 INFO TRAIN_STEP step=11014 loss=1.412117 Training: 61%|██████ | 11014/18018 [23:56:29<15:07:29, 7.77s/step, loss=1.412, acc=0.682, acc_len=1.62, thru=12.7, I=15.4, T=33.2, wait=4.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:26:52,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:26:53,423] loop.py:380 INFO TIMING step=11015: step=3.936s data=1.799s compute=3.880s [fwd=2.685s bwd=1.180s opt=0.003s] dispatch=6.307s [2026-10-19 13:26:53,423] loop.py:45 INFO TRAIN_STEP step=11015 loss=1.368038 Training: 61%|██████ | 11015/18018 [23:56:39<16:34:10, 8.52s/step, loss=1.368, acc=0.689, acc_len=1.64, thru=15.7, I=15.0, T=32.5, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:26:53,883] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:01,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:27:01,674] loop.py:45 INFO TRAIN_STEP step=11016 loss=1.471062 Training: 61%|██████ | 11016/18018 [23:56:47<16:24:36, 8.44s/step, loss=1.471, acc=0.671, acc_len=1.58, thru=14.1, I=14.9, T=35.1, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:27:01,669] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11015: forward=2597.6ms backward=986.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:01,913] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:27:09,581] loop.py:45 INFO TRAIN_STEP step=11017 loss=1.392609 Training: 61%|██████ | 11017/18018 [23:56:55<16:06:09, 8.28s/step, loss=1.393, acc=0.685, acc_len=1.64, thru=15.7, I=15.8, T=37.4, wait=4.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:15,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:27:17,063] loop.py:45 INFO TRAIN_STEP step=11018 loss=1.315433 Training: 61%|██████ | 11018/18018 [23:57:03<15:37:57, 8.04s/step, loss=1.315, acc=0.701, acc_len=1.68, thru=18.0, I=20.3, T=33.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:17,261] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:21,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:27:22,588] loop.py:45 INFO TRAIN_STEP step=11019 loss=1.190888 Training: 61%|██████ | 11019/18018 [23:57:08<14:09:48, 7.29s/step, loss=1.191, acc=0.720, acc_len=1.75, thru=22.2, I=30.9, T=39.3, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:23,008] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:27:29,938] loop.py:380 INFO TIMING step=11020: step=3.644s data=1.515s compute=3.618s [fwd=2.615s bwd=0.988s opt=0.003s] dispatch=3.697s [2026-10-19 13:27:29,938] loop.py:45 INFO TRAIN_STEP step=11020 loss=1.341674 Training: 61%|██████ | 11020/18018 [23:57:16<14:11:57, 7.30s/step, loss=1.342, acc=0.696, acc_len=1.67, thru=15.4, I=20.9, T=35.1, wait=3.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:40,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 13:27:40,793] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11020: forward=2820.3ms backward=1155.8ms [2026-10-19 13:27:40,797] loop.py:45 INFO TRAIN_STEP step=11021 loss=1.346967 Training: 61%|██████ | 11021/18018 [23:57:26<16:16:08, 8.37s/step, loss=1.347, acc=0.697, acc_len=1.67, thru=16.7, I=12.5, T=31.8, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:40,827] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 13:27:47,763] loop.py:45 INFO TRAIN_STEP step=11022 loss=1.334428 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:47,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 61%|██████ | 11022/18018 [23:57:33<15:26:49, 7.95s/step, loss=1.334, acc=0.697, acc_len=1.67, thru=18.6, I=18.0, T=35.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:48,206] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:54,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:27:54,614] loop.py:45 INFO TRAIN_STEP step=11023 loss=1.345378 Training: 61%|██████ | 11023/18018 [23:57:40<14:48:21, 7.62s/step, loss=1.345, acc=0.692, acc_len=1.65, thru=18.6, I=18.6, T=36.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:27:55,073] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:01,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:28:02,040] loop.py:45 INFO TRAIN_STEP step=11024 loss=1.407026 Training: 61%|██████ | 11024/18018 [23:57:48<14:41:26, 7.56s/step, loss=1.407, acc=0.686, acc_len=1.63, thru=17.6, I=19.2, T=34.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:02,371] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:08,033] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:28:08,941] loop.py:380 INFO TIMING step=11025: step=3.693s data=1.636s compute=3.662s [fwd=2.578s bwd=1.070s opt=0.003s] dispatch=3.198s [2026-10-19 13:28:08,941] loop.py:45 INFO TRAIN_STEP step=11025 loss=1.222921 Training: 61%|██████ | 11025/18018 [23:57:55<14:18:11, 7.36s/step, loss=1.223, acc=0.713, acc_len=1.73, thru=18.4, I=21.9, T=34.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:09,052] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:13,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:28:15,284] loop.py:45 INFO TRAIN_STEP step=11026 loss=1.280371 Training: 61%|██████ | 11026/18018 [23:58:01<13:42:28, 7.06s/step, loss=1.280, acc=0.705, acc_len=1.70, thru=19.4, I=24.4, T=33.8, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:28:15,280] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11025: forward=2723.7ms backward=1021.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:15,547] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:20,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:28:21,209] loop.py:45 INFO TRAIN_STEP step=11027 loss=1.338354 Training: 61%|██████ | 11027/18018 [23:58:07<13:02:45, 6.72s/step, loss=1.338, acc=0.690, acc_len=1.65, thru=20.3, I=26.0, T=37.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:21,327] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 13:28:29,161] loop.py:45 INFO TRAIN_STEP step=11028 loss=1.379288 Training: 61%|██████ | 11028/18018 [23:58:15<13:45:41, 7.09s/step, loss=1.379, acc=0.689, acc_len=1.65, thru=15.6, I=15.5, T=30.9, wait=3.8s, pool=49, epoch=2/2][2026-10-19 13:28:38,511] loop.py:45 INFO TRAIN_STEP step=11029 loss=1.269793 Training: 61%|██████ | 11029/18018 [23:58:24<15:04:36, 7.77s/step, loss=1.270, acc=0.706, acc_len=1.70, thru=15.9, I=15.7, T=34.2, wait=5.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:43,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:28:44,837] loop.py:380 INFO TIMING step=11030: step=3.398s data=1.393s compute=3.372s [fwd=2.442s bwd=0.916s opt=0.003s] dispatch=2.920s [2026-10-19 13:28:44,837] loop.py:45 INFO TRAIN_STEP step=11030 loss=1.386692 Training: 61%|██████ | 11030/18018 [23:58:30<14:14:11, 7.33s/step, loss=1.387, acc=0.689, acc_len=1.65, thru=20.0, I=19.6, T=37.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:45,210] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:51,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:28:52,885] loop.py:45 INFO TRAIN_STEP step=11031 loss=1.333818 Training: 61%|██████ | 11031/18018 [23:58:38<14:39:06, 7.55s/step, loss=1.334, acc=0.689, acc_len=1.65, thru=15.6, I=22.0, T=33.7, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:28:52,880] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11030: forward=2704.6ms backward=1056.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:53,160] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:28:59,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:29:00,604] loop.py:45 INFO TRAIN_STEP step=11032 loss=1.418215 Training: 61%|██████ | 11032/18018 [23:58:46<14:44:48, 7.60s/step, loss=1.418, acc=0.682, acc_len=1.62, thru=15.0, I=19.7, T=37.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:29:00,894] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:29:08,334] loop.py:45 INFO TRAIN_STEP step=11033 loss=1.224607 Training: 61%|██████ | 11033/18018 [23:58:54<14:49:19, 7.64s/step, loss=1.225, acc=0.711, acc_len=1.73, thru=15.4, I=18.9, T=32.3, wait=3.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:29:18,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:29:19,658] loop.py:45 INFO TRAIN_STEP step=11034 loss=1.392724 Training: 61%|██████ | 11034/18018 [23:59:05<16:57:47, 8.74s/step, loss=1.393, acc=0.687, acc_len=1.64, thru=12.8, I=13.8, T=32.1, wait=7.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:29:20,112] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:29:28,009] loop.py:380 INFO TIMING step=11035: step=3.764s data=1.576s compute=3.740s [fwd=2.666s bwd=1.059s opt=0.003s] dispatch=4.580s [2026-10-19 13:29:28,009] loop.py:45 INFO TRAIN_STEP step=11035 loss=1.137038 Training: 61%|██████ | 11035/18018 [23:59:14<16:43:55, 8.63s/step, loss=1.137, acc=0.728, acc_len=1.78, thru=12.9, I=14.1, T=34.0, wait=4.6s, pool=53, epoch=2/2][2026-10-19 13:29:34,449] loop.py:45 INFO TRAIN_STEP step=11036 loss=1.428337 Training: 61%|██████ | 11036/18018 [23:59:20<15:27:28, 7.97s/step, loss=1.428, acc=0.674, acc_len=1.59, thru=18.4, I=18.9, T=36.0, wait=2.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:29:34,445] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11035: forward=2525.2ms backward=982.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:29:42,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:29:43,041] loop.py:45 INFO TRAIN_STEP step=11037 loss=1.384720 Training: 61%|██████▏ | 11037/18018 [23:59:29<15:49:03, 8.16s/step, loss=1.385, acc=0.684, acc_len=1.63, thru=15.6, I=16.9, T=33.3, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:29:43,355] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:29:50,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:29:50,854] loop.py:45 INFO TRAIN_STEP step=11038 loss=1.360046 Training: 61%|██████▏ | 11038/18018 [23:59:36<15:36:59, 8.05s/step, loss=1.360, acc=0.693, acc_len=1.66, thru=15.6, I=16.6, T=34.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:29:51,021] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:29:55,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:29:56,479] loop.py:45 INFO TRAIN_STEP step=11039 loss=1.218026 Training: 61%|██████▏ | 11039/18018 [23:59:42<14:12:05, 7.33s/step, loss=1.218, acc=0.717, acc_len=1.74, thru=21.0, I=22.7, T=35.5, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:29:56,662] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:30:05,166] loop.py:380 INFO TIMING step=11040: step=4.105s data=1.763s compute=4.067s [fwd=2.911s bwd=1.140s opt=0.003s] dispatch=4.572s [2026-10-19 13:30:05,167] loop.py:45 INFO TRAIN_STEP step=11040 loss=1.315287 Training: 61%|██████▏ | 11040/18018 [23:59:51<14:59:26, 7.73s/step, loss=1.315, acc=0.698, acc_len=1.67, thru=14.1, I=15.2, T=31.2, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:30:10,551] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:30:11,844] loop.py:45 INFO TRAIN_STEP step=11041 loss=1.308146 Training: 61%|██████▏ | 11041/18018 [23:59:57<14:22:29, 7.42s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=17.6, I=19.9, T=34.1, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:30:11,840] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11040: forward=2632.6ms backward=1084.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:30:12,058] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:30:17,731] loop.py:45 INFO TRAIN_STEP step=11042 loss=1.333008 Training: 61%|██████▏ | 11042/18018 [24:00:03<13:29:06, 6.96s/step, loss=1.333, acc=0.692, acc_len=1.65, thru=18.9, I=25.5, T=39.0, wait=2.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:30:24,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:30:25,889] loop.py:45 INFO TRAIN_STEP step=11043 loss=1.259925 Training: 61%|██████▏ | 11043/18018 [24:00:11<14:10:38, 7.32s/step, loss=1.260, acc=0.705, acc_len=1.70, thru=16.9, I=18.5, T=31.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:30:26,267] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:30:31,630] loop.py:45 INFO TRAIN_STEP step=11044 loss=1.255641 Training: 61%|██████▏ | 11044/18018 [24:00:17<13:15:32, 6.84s/step, loss=1.256, acc=0.708, acc_len=1.71, thru=18.8, I=24.0, T=38.2, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:30:36,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:30:38,580] loop.py:380 INFO TIMING step=11045: step=3.608s data=1.533s compute=3.578s [fwd=2.532s bwd=1.032s opt=0.003s] dispatch=3.334s [2026-10-19 13:30:38,581] loop.py:45 INFO TRAIN_STEP step=11045 loss=1.223045 Training: 61%|██████▏ | 11045/18018 [24:00:24<13:19:07, 6.88s/step, loss=1.223, acc=0.713, acc_len=1.72, thru=19.3, I=22.1, T=35.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:30:38,828] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:30:44,900] loop.py:45 INFO TRAIN_STEP step=11046 loss=1.348274 Training: 61%|██████▏ | 11046/18018 [24:00:30<12:59:37, 6.71s/step, loss=1.348, acc=0.691, acc_len=1.65, thru=18.1, I=22.7, T=38.5, wait=3.0s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:30:44,895] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11045: forward=2350.1ms backward=920.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:30:53,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:30:53,804] loop.py:45 INFO TRAIN_STEP step=11047 loss=1.275422 Training: 61%|██████▏ | 11047/18018 [24:00:39<14:16:00, 7.37s/step, loss=1.275, acc=0.705, acc_len=1.70, thru=16.4, I=16.8, T=35.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:30:54,006] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:31:00,571] loop.py:45 INFO TRAIN_STEP step=11048 loss=1.196259 Training: 61%|██████▏ | 11048/18018 [24:00:46<13:54:58, 7.19s/step, loss=1.196, acc=0.720, acc_len=1.75, thru=18.5, I=18.7, T=36.9, wait=3.3s, pool=57, epoch=2/2][2026-10-19 13:31:07,536] loop.py:45 INFO TRAIN_STEP step=11049 loss=1.259758 Training: 61%|██████▏ | 11049/18018 [24:00:53<13:47:02, 7.12s/step, loss=1.260, acc=0.712, acc_len=1.72, thru=17.7, I=17.7, T=34.5, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:13,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:31:14,489] loop.py:380 INFO TIMING step=11050: step=3.604s data=1.459s compute=3.568s [fwd=2.567s bwd=0.986s opt=0.003s] dispatch=3.342s [2026-10-19 13:31:14,489] loop.py:45 INFO TRAIN_STEP step=11050 loss=1.255668 Training: 61%|██████▏ | 11050/18018 [24:01:00<13:41:09, 7.07s/step, loss=1.256, acc=0.707, acc_len=1.70, thru=18.5, I=20.6, T=35.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:14,755] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:22,340] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:31:22,435] loop.py:45 INFO TRAIN_STEP step=11051 loss=1.293700 Training: 61%|██████▏ | 11051/18018 [24:01:08<14:11:28, 7.33s/step, loss=1.294, acc=0.702, acc_len=1.69, thru=15.0, I=16.1, T=35.2, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:31:22,431] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11050: forward=2613.0ms backward=976.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:22,841] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:30,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:31:31,078] loop.py:45 INFO TRAIN_STEP step=11052 loss=1.481137 Training: 61%|██████▏ | 11052/18018 [24:01:17<14:56:58, 7.73s/step, loss=1.481, acc=0.672, acc_len=1.59, thru=14.6, I=17.2, T=31.5, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:31,161] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 13:31:37,894] loop.py:45 INFO TRAIN_STEP step=11053 loss=1.251106 Training: 61%|██████▏ | 11053/18018 [24:01:23<14:25:07, 7.45s/step, loss=1.251, acc=0.710, acc_len=1.72, thru=17.2, I=19.0, T=37.2, wait=3.4s, pool=53, epoch=2/2][2026-10-19 13:31:45,061] loop.py:45 INFO TRAIN_STEP step=11054 loss=1.352369 Training: 61%|██████▏ | 11054/18018 [24:01:31<14:15:08, 7.37s/step, loss=1.352, acc=0.689, acc_len=1.65, thru=17.3, I=17.4, T=35.9, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:52,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 13:31:53,665] loop.py:380 INFO TIMING step=11055: step=3.694s data=1.650s compute=3.655s [fwd=2.614s bwd=1.026s opt=0.003s] dispatch=4.901s [2026-10-19 13:31:53,665] loop.py:45 INFO TRAIN_STEP step=11055 loss=1.295129 Training: 61%|██████▏ | 11055/18018 [24:01:39<14:57:59, 7.74s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=16.1, I=17.4, T=34.7, wait=4.9s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:54,127] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:58,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:31:59,520] loop.py:45 INFO TRAIN_STEP step=11056 loss=1.213862 Training: 61%|██████▏ | 11056/18018 [24:01:45<13:52:23, 7.17s/step, loss=1.214, acc=0.712, acc_len=1.73, thru=20.1, I=23.3, T=38.4, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:31:59,516] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11055: forward=2372.9ms backward=912.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:31:59,939] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:32:08,852] loop.py:45 INFO TRAIN_STEP step=11057 loss=1.214711 Training: 61%|██████▏ | 11057/18018 [24:01:54<15:07:22, 7.82s/step, loss=1.215, acc=0.715, acc_len=1.73, thru=11.7, I=13.2, T=33.2, wait=5.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:32:16,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:32:18,249] loop.py:45 INFO TRAIN_STEP step=11058 loss=1.317554 Training: 61%|██████▏ | 11058/18018 [24:02:04<16:02:02, 8.29s/step, loss=1.318, acc=0.697, acc_len=1.68, thru=15.3, I=15.3, T=33.0, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:32:18,643] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:32:23,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:32:24,721] loop.py:45 INFO TRAIN_STEP step=11059 loss=1.414277 Training: 61%|██████▏ | 11059/18018 [24:02:10<14:58:37, 7.75s/step, loss=1.414, acc=0.677, acc_len=1.61, thru=18.7, I=25.6, T=36.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:32:24,865] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 13:32:30,549] loop.py:380 INFO TIMING step=11060: step=3.343s data=1.395s compute=3.316s [fwd=2.373s bwd=0.928s opt=0.003s] dispatch=2.473s [2026-10-19 13:32:30,550] loop.py:45 INFO TRAIN_STEP step=11060 loss=1.268831 Training: 61%|██████▏ | 11060/18018 [24:02:16<13:51:45, 7.17s/step, loss=1.269, acc=0.708, acc_len=1.71, thru=20.3, I=26.2, T=38.3, wait=2.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:32:36,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:32:37,507] loop.py:45 INFO TRAIN_STEP step=11061 loss=1.379560 Training: 61%|██████▏ | 11061/18018 [24:02:23<13:44:01, 7.11s/step, loss=1.380, acc=0.684, acc_len=1.62, thru=19.1, I=20.8, T=36.7, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:32:37,500] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11060: forward=2334.6ms backward=1009.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:32:37,920] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:32:42,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:32:43,445] loop.py:45 INFO TRAIN_STEP step=11062 loss=1.281198 Training: 61%|██████▏ | 11062/18018 [24:02:29<13:03:17, 6.76s/step, loss=1.281, acc=0.702, acc_len=1.69, thru=21.6, I=26.2, T=36.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:32:43,639] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:32:49,381] loop.py:45 INFO TRAIN_STEP step=11063 loss=1.242361 Training: 61%|██████▏ | 11063/18018 [24:02:35<12:34:41, 6.51s/step, loss=1.242, acc=0.710, acc_len=1.72, thru=19.3, I=25.7, T=37.8, wait=2.5s, pool=49, epoch=2/2][2026-10-19 13:32:58,314] loop.py:45 INFO TRAIN_STEP step=11064 loss=1.397079 Training: 61%|██████▏ | 11064/18018 [24:02:44<13:58:46, 7.24s/step, loss=1.397, acc=0.685, acc_len=1.62, thru=13.7, I=13.2, T=33.8, wait=5.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:33:06,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:33:07,658] loop.py:380 INFO TIMING step=11065: step=3.766s data=1.489s compute=3.704s [fwd=2.568s bwd=1.120s opt=0.003s] dispatch=5.570s [2026-10-19 13:33:07,659] loop.py:45 INFO TRAIN_STEP step=11065 loss=1.330368 Training: 61%|██████▏ | 11065/18018 [24:02:53<15:11:56, 7.87s/step, loss=1.330, acc=0.694, acc_len=1.66, thru=15.8, I=16.5, T=34.0, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:33:07,982] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:33:13,218] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:33:14,218] loop.py:45 INFO TRAIN_STEP step=11066 loss=1.260804 Training: 61%|██████▏ | 11066/18018 [24:03:00<14:26:17, 7.48s/step, loss=1.261, acc=0.706, acc_len=1.70, thru=18.6, I=23.0, T=36.5, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:33:14,214] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11065: forward=2436.8ms backward=1033.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:33:14,726] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:33:20,246] loop.py:45 INFO TRAIN_STEP step=11067 loss=1.320543 Training: 61%|██████▏ | 11067/18018 [24:03:06<13:35:54, 7.04s/step, loss=1.321, acc=0.699, acc_len=1.68, thru=20.2, I=22.9, T=37.5, wait=2.6s, pool=61, epoch=2/2][2026-10-19 13:33:27,132] loop.py:45 INFO TRAIN_STEP step=11068 loss=1.353969 Training: 61%|██████▏ | 11068/18018 [24:03:13<13:30:09, 6.99s/step, loss=1.354, acc=0.690, acc_len=1.64, thru=18.4, I=18.0, T=36.5, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:33:34,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:33:35,909] loop.py:45 INFO TRAIN_STEP step=11069 loss=1.292861 Training: 61%|██████▏ | 11069/18018 [24:03:21<14:32:01, 7.53s/step, loss=1.293, acc=0.702, acc_len=1.68, thru=16.4, I=17.5, T=33.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:33:35,941] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 13:33:42,701] loop.py:380 INFO TIMING step=11070: step=3.514s data=1.491s compute=3.486s [fwd=2.493s bwd=0.978s opt=0.003s] dispatch=3.268s [2026-10-19 13:33:42,701] loop.py:45 INFO TRAIN_STEP step=11070 loss=1.382652 Training: 61%|██████▏ | 11070/18018 [24:03:28<14:06:15, 7.31s/step, loss=1.383, acc=0.687, acc_len=1.64, thru=17.1, I=19.4, T=36.4, wait=3.3s, pool=53, epoch=2/2][2026-10-19 13:33:50,628] loop.py:45 INFO TRAIN_STEP step=11071 loss=1.265965 Training: 61%|██████▏ | 11071/18018 [24:03:36<14:27:44, 7.49s/step, loss=1.266, acc=0.707, acc_len=1.71, thru=16.6, I=18.0, T=34.9, wait=4.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:33:50,624] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11070: forward=2576.5ms backward=1026.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:33:56,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:33:57,297] loop.py:45 INFO TRAIN_STEP step=11072 loss=1.282925 Training: 61%|██████▏ | 11072/18018 [24:03:43<13:58:53, 7.25s/step, loss=1.283, acc=0.702, acc_len=1.69, thru=18.7, I=19.6, T=35.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:33:57,350] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:04,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:34:04,860] loop.py:45 INFO TRAIN_STEP step=11073 loss=1.320091 Training: 61%|██████▏ | 11073/18018 [24:03:50<14:09:47, 7.34s/step, loss=1.320, acc=0.697, acc_len=1.68, thru=16.8, I=18.9, T=35.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:05,342] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:11,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:34:12,136] loop.py:45 INFO TRAIN_STEP step=11074 loss=1.246352 Training: 61%|██████▏ | 11074/18018 [24:03:58<14:07:23, 7.32s/step, loss=1.246, acc=0.710, acc_len=1.71, thru=18.5, I=20.7, T=34.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:12,359] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:19,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:34:19,895] loop.py:380 INFO TIMING step=11075: step=3.790s data=1.491s compute=3.742s [fwd=2.482s bwd=1.245s opt=0.003s] dispatch=3.959s [2026-10-19 13:34:19,895] loop.py:45 INFO TRAIN_STEP step=11075 loss=1.253460 Training: 61%|██████▏ | 11075/18018 [24:04:05<14:22:26, 7.45s/step, loss=1.253, acc=0.706, acc_len=1.70, thru=17.3, I=19.5, T=33.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:20,162] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:26,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:34:26,695] loop.py:45 INFO TRAIN_STEP step=11076 loss=1.302551 Training: 61%|██████▏ | 11076/18018 [24:04:12<13:59:35, 7.26s/step, loss=1.303, acc=0.702, acc_len=1.69, thru=19.5, I=19.7, T=28.8, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:34:26,691] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11075: forward=3316.1ms backward=1073.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:26,991] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:31,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:34:32,543] loop.py:45 INFO TRAIN_STEP step=11077 loss=1.206860 Training: 61%|██████▏ | 11077/18018 [24:04:18<13:10:36, 6.83s/step, loss=1.207, acc=0.717, acc_len=1.75, thru=19.8, I=20.7, T=36.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:32,926] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:34:40,040] loop.py:45 INFO TRAIN_STEP step=11078 loss=1.210512 Training: 61%|██████▏ | 11078/18018 [24:04:26<13:33:32, 7.03s/step, loss=1.211, acc=0.717, acc_len=1.73, thru=15.7, I=17.0, T=36.7, wait=4.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:44,674] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:34:46,159] loop.py:45 INFO TRAIN_STEP step=11079 loss=1.321487 Training: 61%|██████▏ | 11079/18018 [24:04:32<13:01:42, 6.76s/step, loss=1.321, acc=0.696, acc_len=1.66, thru=20.3, I=21.5, T=36.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:46,194] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:50,943] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:34:51,531] loop.py:380 INFO TIMING step=11080: step=3.374s data=1.399s compute=3.339s [fwd=2.369s bwd=0.957s opt=0.003s] dispatch=1.987s [2026-10-19 13:34:51,531] loop.py:45 INFO TRAIN_STEP step=11080 loss=1.279066 Training: 61%|██████▏ | 11080/18018 [24:04:37<12:13:23, 6.34s/step, loss=1.279, acc=0.701, acc_len=1.68, thru=22.2, I=28.1, T=37.9, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:34:51,948] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:34:57,858] loop.py:45 INFO TRAIN_STEP step=11081 loss=1.306249 Training: 61%|██████▏ | 11081/18018 [24:04:43<12:12:55, 6.34s/step, loss=1.306, acc=0.699, acc_len=1.68, thru=19.4, I=22.8, T=35.9, wait=2.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:34:57,853] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11080: forward=2533.9ms backward=978.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:04,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:35:05,168] loop.py:45 INFO TRAIN_STEP step=11082 loss=1.307869 Training: 62%|██████▏ | 11082/18018 [24:04:51<12:46:21, 6.63s/step, loss=1.308, acc=0.701, acc_len=1.68, thru=18.0, I=19.7, T=30.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:05,373] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:14,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:35:15,446] loop.py:45 INFO TRAIN_STEP step=11083 loss=1.313537 Training: 62%|██████▏ | 11083/18018 [24:05:01<14:52:41, 7.72s/step, loss=1.314, acc=0.696, acc_len=1.67, thru=13.2, I=12.6, T=31.6, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:15,900] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:35:21,353] loop.py:45 INFO TRAIN_STEP step=11084 loss=1.413441 Training: 62%|██████▏ | 11084/18018 [24:05:07<13:49:39, 7.18s/step, loss=1.413, acc=0.679, acc_len=1.61, thru=19.1, I=22.1, T=36.4, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:29,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:35:30,359] loop.py:380 INFO TIMING step=11085: step=3.839s data=1.672s compute=3.812s [fwd=2.756s bwd=1.041s opt=0.003s] dispatch=5.156s [2026-10-19 13:35:30,359] loop.py:45 INFO TRAIN_STEP step=11085 loss=1.268954 Training: 62%|██████▏ | 11085/18018 [24:05:16<14:52:48, 7.73s/step, loss=1.269, acc=0.710, acc_len=1.72, thru=15.7, I=15.3, T=33.3, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:30,840] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:35:39,495] loop.py:45 INFO TRAIN_STEP step=11086 loss=1.422163 Training: 62%|██████▏ | 11086/18018 [24:05:25<15:41:36, 8.15s/step, loss=1.422, acc=0.680, acc_len=1.61, thru=11.6, I=12.6, T=31.3, wait=5.0s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:35:39,490] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11085: forward=2875.5ms backward=1151.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:44,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:35:46,427] loop.py:45 INFO TRAIN_STEP step=11087 loss=1.332837 Training: 62%|██████▏ | 11087/18018 [24:05:32<14:59:14, 7.78s/step, loss=1.333, acc=0.698, acc_len=1.67, thru=18.6, I=20.1, T=34.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:46,512] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:52,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:35:52,745] loop.py:45 INFO TRAIN_STEP step=11088 loss=1.265114 Training: 62%|██████▏ | 11088/18018 [24:05:38<14:08:13, 7.34s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=19.6, I=24.0, T=35.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:53,077] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:35:59,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:36:00,375] loop.py:45 INFO TRAIN_STEP step=11089 loss=1.367507 Training: 62%|██████▏ | 11089/18018 [24:05:46<14:18:05, 7.43s/step, loss=1.368, acc=0.688, acc_len=1.63, thru=19.0, I=19.5, T=35.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:00,622] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:05,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:36:06,630] loop.py:380 INFO TIMING step=11090: step=3.413s data=1.389s compute=3.379s [fwd=2.415s bwd=0.950s opt=0.003s] dispatch=2.832s [2026-10-19 13:36:06,630] loop.py:45 INFO TRAIN_STEP step=11090 loss=1.389090 Training: 62%|██████▏ | 11090/18018 [24:05:52<13:37:15, 7.08s/step, loss=1.389, acc=0.685, acc_len=1.63, thru=19.7, I=26.3, T=37.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:06,865] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:11,182] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:36:13,590] loop.py:45 INFO TRAIN_STEP step=11091 loss=1.293321 Training: 62%|██████▏ | 11091/18018 [24:05:59<13:32:59, 7.04s/step, loss=1.293, acc=0.702, acc_len=1.68, thru=18.1, I=28.2, T=28.0, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:36:13,585] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11090: forward=3490.0ms backward=1038.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:13,709] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:19,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:36:22,175] loop.py:45 INFO TRAIN_STEP step=11092 loss=1.329794 Training: 62%|██████▏ | 11092/18018 [24:06:08<14:26:25, 7.51s/step, loss=1.330, acc=0.692, acc_len=1.66, thru=15.1, I=18.9, T=22.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:22,222] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=57/64 [2026-10-19 13:36:29,769] loop.py:45 INFO TRAIN_STEP step=11093 loss=1.293412 Training: 62%|██████▏ | 11093/18018 [24:06:15<14:29:16, 7.53s/step, loss=1.293, acc=0.700, acc_len=1.68, thru=16.5, I=15.5, T=34.5, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:37,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:36:37,315] loop.py:45 INFO TRAIN_STEP step=11094 loss=1.341490 Training: 62%|██████▏ | 11094/18018 [24:06:23<14:29:35, 7.54s/step, loss=1.341, acc=0.693, acc_len=1.66, thru=17.8, I=17.1, T=34.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:37,570] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:43,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:36:43,890] loop.py:380 INFO TIMING step=11095: step=3.431s data=1.371s compute=3.386s [fwd=2.372s bwd=0.999s opt=0.003s] dispatch=3.136s [2026-10-19 13:36:43,890] loop.py:45 INFO TRAIN_STEP step=11095 loss=1.320486 Training: 62%|██████▏ | 11095/18018 [24:06:29<13:56:13, 7.25s/step, loss=1.320, acc=0.699, acc_len=1.67, thru=19.4, I=18.1, T=37.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:44,233] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:49,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:36:50,067] loop.py:45 INFO TRAIN_STEP step=11096 loss=1.142483 Training: 62%|██████▏ | 11096/18018 [24:06:36<13:19:05, 6.93s/step, loss=1.142, acc=0.729, acc_len=1.78, thru=20.7, I=21.8, T=36.0, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:36:50,063] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11095: forward=2513.4ms backward=1003.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:36:50,244] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:00,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:37:01,378] loop.py:45 INFO TRAIN_STEP step=11097 loss=1.343728 Training: 62%|██████▏ | 11097/18018 [24:06:47<15:50:43, 8.24s/step, loss=1.344, acc=0.692, acc_len=1.65, thru=11.9, I=12.5, T=31.3, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:01,524] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:07,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:37:07,934] loop.py:45 INFO TRAIN_STEP step=11098 loss=1.316548 Training: 62%|██████▏ | 11098/18018 [24:06:54<14:52:15, 7.74s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=18.9, I=20.1, T=31.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:08,082] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:13,165] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:37:13,891] loop.py:45 INFO TRAIN_STEP step=11099 loss=1.329589 Training: 62%|██████▏ | 11099/18018 [24:06:59<13:50:35, 7.20s/step, loss=1.330, acc=0.693, acc_len=1.65, thru=19.2, I=21.8, T=35.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:14,172] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:37:20,738] loop.py:380 INFO TIMING step=11100: step=3.669s data=1.596s compute=3.634s [fwd=2.572s bwd=1.047s opt=0.003s] dispatch=3.167s [2026-10-19 13:37:20,739] loop.py:45 INFO TRAIN_STEP step=11100 loss=1.214259 Training: 62%|██████▏ | 11100/18018 [24:07:06<13:38:06, 7.10s/step, loss=1.214, acc=0.715, acc_len=1.73, thru=18.0, I=18.5, T=34.9, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:28,879] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:37:29,565] loop.py:45 INFO TRAIN_STEP step=11101 loss=1.359267 Training: 62%|██████▏ | 11101/18018 [24:07:15<14:37:54, 7.62s/step, loss=1.359, acc=0.692, acc_len=1.66, thru=16.2, I=16.2, T=32.1, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:37:29,561] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11100: forward=2798.1ms backward=1148.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:29,883] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:37:37,404] loop.py:45 INFO TRAIN_STEP step=11102 loss=1.265788 Training: 62%|██████▏ | 11102/18018 [24:07:23<14:45:31, 7.68s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=13.5, I=15.4, T=34.4, wait=4.1s, pool=41, epoch=2/2][2026-10-19 13:37:45,838] loop.py:45 INFO TRAIN_STEP step=11103 loss=1.526343 Training: 62%|██████▏ | 11103/18018 [24:07:31<15:11:20, 7.91s/step, loss=1.526, acc=0.664, acc_len=1.56, thru=16.5, I=13.8, T=35.5, wait=4.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:53,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:37:53,523] loop.py:45 INFO TRAIN_STEP step=11104 loss=1.411892 Training: 62%|██████▏ | 11104/18018 [24:07:39<15:03:35, 7.84s/step, loss=1.412, acc=0.681, acc_len=1.61, thru=18.1, I=16.3, T=34.2, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:37:53,907] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:00,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:38:00,959] loop.py:380 INFO TIMING step=11105: step=3.608s data=1.536s compute=3.571s [fwd=2.556s bwd=1.000s opt=0.003s] dispatch=3.818s [2026-10-19 13:38:00,960] loop.py:45 INFO TRAIN_STEP step=11105 loss=1.408919 Training: 62%|██████▏ | 11105/18018 [24:07:47<14:49:22, 7.72s/step, loss=1.409, acc=0.679, acc_len=1.61, thru=17.6, I=18.4, T=35.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:01,244] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:05,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:38:07,112] loop.py:45 INFO TRAIN_STEP step=11106 loss=1.210886 Training: 62%|██████▏ | 11106/18018 [24:07:53<13:55:06, 7.25s/step, loss=1.211, acc=0.716, acc_len=1.74, thru=20.3, I=25.2, T=35.2, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:38:07,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11105: forward=2566.7ms backward=1016.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:07,539] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 13:38:14,250] loop.py:45 INFO TRAIN_STEP step=11107 loss=1.211251 Training: 62%|██████▏ | 11107/18018 [24:08:00<13:51:13, 7.22s/step, loss=1.211, acc=0.715, acc_len=1.73, thru=14.8, I=20.5, T=32.4, wait=3.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:20,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:38:21,612] loop.py:45 INFO TRAIN_STEP step=11108 loss=1.314560 Training: 62%|██████▏ | 11108/18018 [24:08:07<13:56:06, 7.26s/step, loss=1.315, acc=0.699, acc_len=1.67, thru=18.0, I=18.4, T=33.2, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:21,810] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:27,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:38:28,390] loop.py:45 INFO TRAIN_STEP step=11109 loss=1.364055 Training: 62%|██████▏ | 11109/18018 [24:08:14<13:39:17, 7.11s/step, loss=1.364, acc=0.686, acc_len=1.63, thru=18.7, I=23.8, T=36.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:28,756] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:38:35,390] loop.py:380 INFO TIMING step=11110: step=3.457s data=1.393s compute=3.433s [fwd=2.466s bwd=0.951s opt=0.003s] dispatch=3.535s [2026-10-19 13:38:35,390] loop.py:45 INFO TRAIN_STEP step=11110 loss=1.339533 Training: 62%|██████▏ | 11110/18018 [24:08:21<13:35:10, 7.08s/step, loss=1.340, acc=0.693, acc_len=1.66, thru=16.8, I=19.3, T=37.0, wait=3.5s, pool=57, epoch=2/2][2026-10-19 13:38:43,075] loop.py:45 INFO TRAIN_STEP step=11111 loss=1.279648 Training: 62%|██████▏ | 11111/18018 [24:08:29<13:55:54, 7.26s/step, loss=1.280, acc=0.701, acc_len=1.69, thru=16.4, I=15.7, T=33.8, wait=3.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:38:43,070] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11110: forward=2680.6ms backward=1072.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:49,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:38:51,250] loop.py:45 INFO TRAIN_STEP step=11112 loss=1.392598 Training: 62%|██████▏ | 11112/18018 [24:08:37<14:27:26, 7.54s/step, loss=1.393, acc=0.684, acc_len=1.63, thru=17.5, I=19.6, T=35.5, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:38:51,507] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:38:57,097] loop.py:45 INFO TRAIN_STEP step=11113 loss=1.204951 Training: 62%|██████▏ | 11113/18018 [24:08:43<13:28:53, 7.03s/step, loss=1.205, acc=0.715, acc_len=1.74, thru=20.0, I=23.0, T=37.4, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:04,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:39:04,946] loop.py:45 INFO TRAIN_STEP step=11114 loss=1.259815 Training: 62%|██████▏ | 11114/18018 [24:08:51<13:57:05, 7.27s/step, loss=1.260, acc=0.710, acc_len=1.71, thru=18.4, I=18.1, T=33.7, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:05,387] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:39:13,621] loop.py:380 INFO TIMING step=11115: step=3.956s data=1.731s compute=3.924s [fwd=2.829s bwd=1.079s opt=0.003s] dispatch=4.712s [2026-10-19 13:39:13,621] loop.py:45 INFO TRAIN_STEP step=11115 loss=1.401822 Training: 62%|██████▏ | 11115/18018 [24:08:59<14:45:20, 7.70s/step, loss=1.402, acc=0.681, acc_len=1.61, thru=14.4, I=13.8, T=32.4, wait=4.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:20,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:39:21,577] loop.py:45 INFO TRAIN_STEP step=11116 loss=1.310359 (TrainerActor pid=9310) [2026-10-19 13:39:21,572] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11115: forward=2394.2ms backward=1128.3ms Training: 62%|██████▏ | 11116/18018 [24:09:07<14:54:15, 7.77s/step, loss=1.310, acc=0.695, acc_len=1.66, thru=17.4, I=18.7, T=35.7, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:22,063] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:39:28,047] loop.py:45 INFO TRAIN_STEP step=11117 loss=1.234829 Training: 62%|██████▏ | 11117/18018 [24:09:14<14:09:03, 7.38s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=18.1, I=21.1, T=37.6, wait=3.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:35,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:39:36,595] loop.py:45 INFO TRAIN_STEP step=11118 loss=1.388803 Training: 62%|██████▏ | 11118/18018 [24:09:22<14:49:13, 7.73s/step, loss=1.389, acc=0.685, acc_len=1.63, thru=16.8, I=18.0, T=32.6, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:36,791] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:42,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:39:43,391] loop.py:45 INFO TRAIN_STEP step=11119 loss=1.437541 Training: 62%|██████▏ | 11119/18018 [24:09:29<14:16:49, 7.45s/step, loss=1.438, acc=0.673, acc_len=1.58, thru=17.7, I=21.8, T=32.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:43,510] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:50,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:39:50,386] loop.py:380 INFO TIMING step=11120: step=3.536s data=1.480s compute=3.510s [fwd=2.552s bwd=0.945s opt=0.003s] dispatch=3.448s [2026-10-19 13:39:50,387] loop.py:45 INFO TRAIN_STEP step=11120 loss=1.267866 Training: 62%|██████▏ | 11120/18018 [24:09:36<14:00:56, 7.31s/step, loss=1.268, acc=0.703, acc_len=1.69, thru=16.3, I=18.8, T=36.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:50,635] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:39:58,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:39:59,980] loop.py:45 INFO TRAIN_STEP step=11121 loss=1.290604 Training: 62%|██████▏ | 11121/18018 [24:09:46<15:19:20, 8.00s/step, loss=1.291, acc=0.702, acc_len=1.69, thru=13.4, I=14.4, T=34.9, wait=5.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:39:59,976] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11120: forward=2553.9ms backward=1070.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:00,402] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:05,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:40:06,836] loop.py:45 INFO TRAIN_STEP step=11122 loss=1.322936 Training: 62%|██████▏ | 11122/18018 [24:09:52<14:39:52, 7.66s/step, loss=1.323, acc=0.700, acc_len=1.68, thru=18.2, I=24.6, T=35.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:06,917] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:16,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:40:16,510] loop.py:45 INFO TRAIN_STEP step=11123 loss=1.382518 Training: 62%|██████▏ | 11123/18018 [24:10:02<15:49:21, 8.26s/step, loss=1.383, acc=0.686, acc_len=1.64, thru=13.4, I=13.4, T=31.9, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:16,807] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:22,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:40:23,099] loop.py:45 INFO TRAIN_STEP step=11124 loss=1.342290 Training: 62%|██████▏ | 11124/18018 [24:10:09<14:51:31, 7.76s/step, loss=1.342, acc=0.694, acc_len=1.66, thru=18.6, I=19.4, T=33.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:23,529] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:28,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:40:28,926] loop.py:380 INFO TIMING step=11125: step=3.434s data=1.426s compute=3.403s [fwd=2.465s bwd=0.923s opt=0.003s] dispatch=2.385s [2026-10-19 13:40:28,926] loop.py:45 INFO TRAIN_STEP step=11125 loss=1.331082 Training: 62%|██████▏ | 11125/18018 [24:10:15<13:44:51, 7.18s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=20.8, I=24.3, T=37.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:29,200] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:38,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:40:40,169] loop.py:45 INFO TRAIN_STEP step=11126 loss=1.393938 Training: 62%|██████▏ | 11126/18018 [24:10:26<16:04:44, 8.40s/step, loss=1.394, acc=0.686, acc_len=1.63, thru=11.4, I=13.4, T=33.9, wait=7.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:40:40,165] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11125: forward=2592.0ms backward=1114.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:40:40,548] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:40:46,196] loop.py:45 INFO TRAIN_STEP step=11127 loss=1.205588 Training: 62%|██████▏ | 11127/18018 [24:10:32<14:42:51, 7.69s/step, loss=1.206, acc=0.715, acc_len=1.73, thru=19.1, I=21.1, T=37.0, wait=2.6s, pool=57, epoch=2/2][2026-10-19 13:40:53,359] loop.py:45 INFO TRAIN_STEP step=11128 loss=1.414178 Training: 62%|██████▏ | 11128/18018 [24:10:39<14:24:38, 7.53s/step, loss=1.414, acc=0.678, acc_len=1.61, thru=16.0, I=19.1, T=32.7, wait=3.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:02,459] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:41:03,423] loop.py:45 INFO TRAIN_STEP step=11129 loss=1.377128 Training: 62%|██████▏ | 11129/18018 [24:10:49<15:51:49, 8.29s/step, loss=1.377, acc=0.687, acc_len=1.63, thru=15.3, I=14.3, T=32.7, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:03,467] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 13:41:11,328] loop.py:380 INFO TIMING step=11130: step=3.585s data=1.506s compute=3.547s [fwd=2.520s bwd=1.013s opt=0.003s] dispatch=4.311s [2026-10-19 13:41:11,328] loop.py:45 INFO TRAIN_STEP step=11130 loss=1.186374 Training: 62%|██████▏ | 11130/18018 [24:10:57<15:38:27, 8.17s/step, loss=1.186, acc=0.721, acc_len=1.75, thru=15.2, I=17.0, T=35.7, wait=4.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:16,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:41:18,599] loop.py:45 INFO TRAIN_STEP step=11131 loss=1.257247 Training: 62%|██████▏ | 11131/18018 [24:11:04<15:07:07, 7.90s/step, loss=1.257, acc=0.708, acc_len=1.71, thru=18.3, I=20.8, T=30.9, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:41:18,595] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11130: forward=2937.9ms backward=1149.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:18,769] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:41:25,850] loop.py:45 INFO TRAIN_STEP step=11132 loss=1.310585 Training: 62%|██████▏ | 11132/18018 [24:11:11<14:44:39, 7.71s/step, loss=1.311, acc=0.700, acc_len=1.68, thru=13.2, I=16.5, T=35.1, wait=3.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:33,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:41:34,551] loop.py:45 INFO TRAIN_STEP step=11133 loss=1.394113 Training: 62%|██████▏ | 11133/18018 [24:11:20<15:18:36, 8.01s/step, loss=1.394, acc=0.683, acc_len=1.62, thru=17.2, I=18.4, T=33.7, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:35,035] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:41,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:41:42,630] loop.py:45 INFO TRAIN_STEP step=11134 loss=1.231387 Training: 62%|██████▏ | 11134/18018 [24:11:28<15:21:01, 8.03s/step, loss=1.231, acc=0.713, acc_len=1.72, thru=14.1, I=19.6, T=32.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:42,659] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 13:41:48,749] loop.py:380 INFO TIMING step=11135: step=3.438s data=1.462s compute=3.412s [fwd=2.442s bwd=0.955s opt=0.003s] dispatch=2.673s [2026-10-19 13:41:48,749] loop.py:45 INFO TRAIN_STEP step=11135 loss=1.362881 Training: 62%|██████▏ | 11135/18018 [24:11:34<14:15:10, 7.45s/step, loss=1.363, acc=0.686, acc_len=1.63, thru=17.4, I=19.1, T=37.2, wait=2.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:55,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:41:56,670] loop.py:45 INFO TRAIN_STEP step=11136 loss=1.267470 Training: 62%|██████▏ | 11136/18018 [24:11:42<14:31:12, 7.60s/step, loss=1.267, acc=0.706, acc_len=1.71, thru=18.0, I=18.3, T=32.7, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:41:56,666] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11135: forward=2697.0ms backward=1158.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:41:56,787] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:04,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:42:05,392] loop.py:45 INFO TRAIN_STEP step=11137 loss=1.298538 Training: 62%|██████▏ | 11137/18018 [24:11:51<15:09:47, 7.93s/step, loss=1.299, acc=0.703, acc_len=1.69, thru=14.9, I=16.2, T=35.2, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:05,732] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:42:11,778] loop.py:45 INFO TRAIN_STEP step=11138 loss=1.333906 Training: 62%|██████▏ | 11138/18018 [24:11:57<14:16:23, 7.47s/step, loss=1.334, acc=0.693, acc_len=1.65, thru=17.8, I=19.1, T=38.3, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:20,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:42:20,433] loop.py:45 INFO TRAIN_STEP step=11139 loss=1.397078 Training: 62%|██████▏ | 11139/18018 [24:12:06<14:57:07, 7.82s/step, loss=1.397, acc=0.687, acc_len=1.64, thru=17.3, I=15.5, T=35.3, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:20,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:42:26,989] loop.py:380 INFO TIMING step=11140: step=3.722s data=1.594s compute=3.685s [fwd=2.628s bwd=1.042s opt=0.003s] dispatch=2.825s [2026-10-19 13:42:26,990] loop.py:45 INFO TRAIN_STEP step=11140 loss=1.292690 Training: 62%|██████▏ | 11140/18018 [24:12:13<14:13:26, 7.44s/step, loss=1.293, acc=0.704, acc_len=1.69, thru=17.7, I=20.0, T=34.4, wait=2.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:35,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:42:35,835] loop.py:45 INFO TRAIN_STEP step=11141 loss=1.197206 Training: 62%|██████▏ | 11141/18018 [24:12:21<15:01:26, 7.86s/step, loss=1.197, acc=0.717, acc_len=1.74, thru=16.8, I=14.8, T=33.3, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:42:35,830] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11140: forward=2573.7ms backward=1206.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:36,095] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:40,352] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:42:41,259] loop.py:45 INFO TRAIN_STEP step=11142 loss=1.303014 Training: 62%|██████▏ | 11142/18018 [24:12:27<13:37:24, 7.13s/step, loss=1.303, acc=0.701, acc_len=1.68, thru=20.8, I=25.2, T=37.6, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:41,373] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:47,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:42:47,730] loop.py:45 INFO TRAIN_STEP step=11143 loss=1.163268 Training: 62%|██████▏ | 11143/18018 [24:12:33<13:14:29, 6.93s/step, loss=1.163, acc=0.727, acc_len=1.78, thru=20.9, I=22.9, T=34.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:47,890] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:55,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:42:56,018] loop.py:45 INFO TRAIN_STEP step=11144 loss=1.257565 Training: 62%|██████▏ | 11144/18018 [24:12:42<14:00:57, 7.34s/step, loss=1.258, acc=0.708, acc_len=1.71, thru=16.6, I=16.3, T=35.7, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:42:56,239] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:43:02,369] loop.py:380 INFO TIMING step=11145: step=3.381s data=1.433s compute=3.336s [fwd=2.358s bwd=0.964s opt=0.003s] dispatch=2.960s [2026-10-19 13:43:02,369] loop.py:45 INFO TRAIN_STEP step=11145 loss=1.176994 Training: 62%|██████▏ | 11145/18018 [24:12:48<13:26:54, 7.04s/step, loss=1.177, acc=0.719, acc_len=1.75, thru=16.1, I=17.2, T=37.9, wait=3.0s, pool=29, epoch=2/2][2026-10-19 13:43:14,404] loop.py:45 INFO TRAIN_STEP step=11146 loss=1.255196 Training: 62%|██████▏ | 11146/18018 [24:13:00<16:18:08, 8.54s/step, loss=1.255, acc=0.710, acc_len=1.72, thru=12.5, I=12.9, T=30.8, wait=7.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:43:14,400] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11145: forward=2903.8ms backward=1205.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:43:21,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:43:22,368] loop.py:45 INFO TRAIN_STEP step=11147 loss=1.336590 Training: 62%|██████▏ | 11147/18018 [24:13:08<15:58:16, 8.37s/step, loss=1.337, acc=0.696, acc_len=1.67, thru=16.1, I=17.8, T=33.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:43:22,536] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 13:43:28,815] loop.py:45 INFO TRAIN_STEP step=11148 loss=1.231174 Training: 62%|██████▏ | 11148/18018 [24:13:14<14:52:08, 7.79s/step, loss=1.231, acc=0.714, acc_len=1.73, thru=17.6, I=22.5, T=37.3, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:43:35,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:43:36,846] loop.py:45 INFO TRAIN_STEP step=11149 loss=1.352508 Training: 62%|██████▏ | 11149/18018 [24:13:22<15:00:10, 7.86s/step, loss=1.353, acc=0.696, acc_len=1.67, thru=17.3, I=19.3, T=30.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:43:37,289] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:43:41,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:43:43,624] loop.py:380 INFO TIMING step=11150: step=3.905s data=1.728s compute=3.873s [fwd=2.716s bwd=1.142s opt=0.003s] dispatch=2.865s [2026-10-19 13:43:43,625] loop.py:45 INFO TRAIN_STEP step=11150 loss=1.343836 Training: 62%|██████▏ | 11150/18018 [24:13:29<14:22:53, 7.54s/step, loss=1.344, acc=0.696, acc_len=1.67, thru=18.1, I=27.9, T=32.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:43:43,669] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=57/64 [2026-10-19 13:43:52,994] loop.py:45 INFO TRAIN_STEP step=11151 loss=1.232668 Training: 62%|██████▏ | 11151/18018 [24:13:39<15:25:33, 8.09s/step, loss=1.233, acc=0.710, acc_len=1.71, thru=12.9, I=14.8, T=33.1, wait=5.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:43:52,990] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11150: forward=2744.5ms backward=1076.2ms [2026-10-19 13:44:01,879] loop.py:45 INFO TRAIN_STEP step=11152 loss=1.371573 Training: 62%|██████▏ | 11152/18018 [24:13:47<15:52:47, 8.33s/step, loss=1.372, acc=0.688, acc_len=1.64, thru=13.6, I=13.5, T=32.7, wait=5.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:08,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:44:09,096] loop.py:45 INFO TRAIN_STEP step=11153 loss=1.288078 Training: 62%|██████▏ | 11153/18018 [24:13:55<15:14:38, 7.99s/step, loss=1.288, acc=0.702, acc_len=1.69, thru=18.6, I=18.6, T=31.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:09,448] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:16,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:44:16,620] loop.py:45 INFO TRAIN_STEP step=11154 loss=1.294055 Training: 62%|██████▏ | 11154/18018 [24:14:02<14:58:33, 7.85s/step, loss=1.294, acc=0.703, acc_len=1.69, thru=16.7, I=18.7, T=34.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:16,884] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 13:44:22,507] loop.py:380 INFO TIMING step=11155: step=3.373s data=1.398s compute=3.336s [fwd=2.376s bwd=0.945s opt=0.003s] dispatch=2.499s [2026-10-19 13:44:22,508] loop.py:45 INFO TRAIN_STEP step=11155 loss=1.370675 Training: 62%|██████▏ | 11155/18018 [24:14:08<13:50:48, 7.26s/step, loss=1.371, acc=0.689, acc_len=1.64, thru=20.1, I=21.9, T=37.9, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:28,340] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:44:29,263] loop.py:45 INFO TRAIN_STEP step=11156 loss=1.370607 Training: 62%|██████▏ | 11156/18018 [24:14:15<13:33:10, 7.11s/step, loss=1.371, acc=0.687, acc_len=1.64, thru=18.8, I=21.4, T=37.3, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:44:29,258] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11155: forward=2410.8ms backward=969.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:29,353] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 13:44:35,502] loop.py:45 INFO TRAIN_STEP step=11157 loss=1.390597 Training: 62%|██████▏ | 11157/18018 [24:14:21<13:03:12, 6.85s/step, loss=1.391, acc=0.680, acc_len=1.62, thru=19.4, I=22.2, T=33.1, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:40,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:44:42,213] loop.py:45 INFO TRAIN_STEP step=11158 loss=1.336255 Training: 62%|██████▏ | 11158/18018 [24:14:28<12:58:24, 6.81s/step, loss=1.336, acc=0.692, acc_len=1.65, thru=18.7, I=22.4, T=31.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:42,390] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:47,893] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:44:48,902] loop.py:45 INFO TRAIN_STEP step=11159 loss=1.385211 Training: 62%|██████▏ | 11159/18018 [24:14:34<12:54:03, 6.77s/step, loss=1.385, acc=0.687, acc_len=1.63, thru=17.6, I=24.3, T=32.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:49,404] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:55,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:44:56,643] loop.py:380 INFO TIMING step=11160: step=3.646s data=1.623s compute=3.612s [fwd=2.591s bwd=1.007s opt=0.003s] dispatch=4.088s [2026-10-19 13:44:56,643] loop.py:45 INFO TRAIN_STEP step=11160 loss=1.251860 Training: 62%|██████▏ | 11160/18018 [24:14:42<13:27:15, 7.06s/step, loss=1.252, acc=0.708, acc_len=1.71, thru=15.2, I=19.2, T=35.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:44:56,882] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:45:04,356] loop.py:45 INFO TRAIN_STEP step=11161 loss=1.268547 Training: 62%|██████▏ | 11161/18018 [24:14:50<13:49:24, 7.26s/step, loss=1.269, acc=0.705, acc_len=1.70, thru=16.0, I=16.4, T=33.0, wait=3.8s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:45:04,352] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11160: forward=2738.5ms backward=1087.1ms [2026-10-19 13:45:12,146] loop.py:45 INFO TRAIN_STEP step=11162 loss=1.271377 Training: 62%|██████▏ | 11162/18018 [24:14:58<14:07:34, 7.42s/step, loss=1.271, acc=0.705, acc_len=1.70, thru=16.9, I=18.9, T=34.0, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:17,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:45:19,835] loop.py:45 INFO TRAIN_STEP step=11163 loss=1.370966 Training: 62%|██████▏ | 11163/18018 [24:15:05<14:16:48, 7.50s/step, loss=1.371, acc=0.687, acc_len=1.63, thru=16.5, I=20.0, T=32.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:19,885] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:24,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:45:25,973] loop.py:45 INFO TRAIN_STEP step=11164 loss=1.211659 Training: 62%|██████▏ | 11164/18018 [24:15:12<13:29:55, 7.09s/step, loss=1.212, acc=0.714, acc_len=1.73, thru=19.2, I=26.7, T=33.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:26,461] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:45:32,312] loop.py:380 INFO TIMING step=11165: step=3.443s data=1.459s compute=3.409s [fwd=2.471s bwd=0.923s opt=0.003s] dispatch=2.889s [2026-10-19 13:45:32,312] loop.py:45 INFO TRAIN_STEP step=11165 loss=1.281101 Training: 62%|██████▏ | 11165/18018 [24:15:18<13:04:04, 6.86s/step, loss=1.281, acc=0.708, acc_len=1.71, thru=18.9, I=22.0, T=37.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:32,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:32,744] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:38,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:45:39,179] loop.py:45 INFO TRAIN_STEP step=11166 loss=1.295081 Training: 62%|██████▏ | 11166/18018 [24:15:25<13:04:05, 6.87s/step, loss=1.295, acc=0.702, acc_len=1.69, thru=19.5, I=21.7, T=36.2, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:45:39,175] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11165: forward=2508.4ms backward=972.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:39,661] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:45:48,645] loop.py:45 INFO TRAIN_STEP step=11167 loss=1.267921 Training: 62%|██████▏ | 11167/18018 [24:15:34<14:33:00, 7.65s/step, loss=1.268, acc=0.707, acc_len=1.71, thru=13.4, I=13.8, T=35.5, wait=5.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:57,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:45:57,677] loop.py:45 INFO TRAIN_STEP step=11168 loss=1.303536 Training: 62%|██████▏ | 11168/18018 [24:15:43<15:20:25, 8.06s/step, loss=1.304, acc=0.702, acc_len=1.69, thru=15.4, I=15.3, T=30.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:45:58,127] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:46:04,116] loop.py:45 INFO TRAIN_STEP step=11169 loss=1.297743 Training: 62%|██████▏ | 11169/18018 [24:15:50<14:24:38, 7.57s/step, loss=1.298, acc=0.697, acc_len=1.67, thru=16.7, I=20.1, T=36.8, wait=2.9s, pool=49, epoch=2/2][2026-10-19 13:46:14,545] loop.py:380 INFO TIMING step=11170: step=3.822s data=1.621s compute=3.794s [fwd=2.710s bwd=1.069s opt=0.003s] dispatch=6.598s [2026-10-19 13:46:14,545] loop.py:45 INFO TRAIN_STEP step=11170 loss=1.334821 Training: 62%|██████▏ | 11170/18018 [24:16:00<16:02:18, 8.43s/step, loss=1.335, acc=0.693, acc_len=1.66, thru=14.5, I=12.5, T=33.5, wait=6.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:46:20,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:46:21,280] loop.py:45 INFO TRAIN_STEP step=11171 loss=1.362387 Training: 62%|██████▏ | 11171/18018 [24:16:07<15:04:05, 7.92s/step, loss=1.362, acc=0.693, acc_len=1.66, thru=18.5, I=19.9, T=32.5, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:46:21,276] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11170: forward=2815.9ms backward=1078.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:46:21,779] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:46:28,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:46:28,490] loop.py:45 INFO TRAIN_STEP step=11172 loss=1.304773 Training: 62%|██████▏ | 11172/18018 [24:16:14<14:39:31, 7.71s/step, loss=1.305, acc=0.699, acc_len=1.67, thru=17.3, I=20.1, T=33.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:46:28,600] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 13:46:36,400] loop.py:45 INFO TRAIN_STEP step=11173 loss=1.185187 Training: 62%|██████▏ | 11173/18018 [24:16:22<14:46:21, 7.77s/step, loss=1.185, acc=0.721, acc_len=1.75, thru=14.5, I=19.0, T=31.8, wait=3.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:46:44,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:46:46,946] loop.py:45 INFO TRAIN_STEP step=11174 loss=1.372409 Training: 62%|██████▏ | 11174/18018 [24:16:33<16:21:14, 8.60s/step, loss=1.372, acc=0.688, acc_len=1.64, thru=17.1, I=15.1, T=31.4, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:46:47,248] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 13:46:53,346] loop.py:380 INFO TIMING step=11175: step=3.446s data=1.435s compute=3.417s [fwd=2.480s bwd=0.922s opt=0.003s] dispatch=2.944s [2026-10-19 13:46:53,346] loop.py:45 INFO TRAIN_STEP step=11175 loss=1.350324 Training: 62%|██████▏ | 11175/18018 [24:16:39<15:05:46, 7.94s/step, loss=1.350, acc=0.692, acc_len=1.66, thru=15.7, I=24.4, T=37.1, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:47:00,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:47:01,140] loop.py:45 INFO TRAIN_STEP step=11176 loss=1.417035 Training: 62%|██████▏ | 11176/18018 [24:16:47<15:00:33, 7.90s/step, loss=1.417, acc=0.676, acc_len=1.59, thru=17.4, I=18.1, T=35.6, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:47:01,135] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11175: forward=2580.5ms backward=971.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:47:01,452] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:47:07,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:47:08,489] loop.py:45 INFO TRAIN_STEP step=11177 loss=1.393797 Training: 62%|██████▏ | 11177/18018 [24:16:54<14:41:37, 7.73s/step, loss=1.394, acc=0.687, acc_len=1.63, thru=17.2, I=20.5, T=31.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:47:08,518] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:47:13,741] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:47:16,454] loop.py:45 INFO TRAIN_STEP step=11178 loss=1.354680 Training: 62%|██████▏ | 11178/18018 [24:17:02<14:49:23, 7.80s/step, loss=1.355, acc=0.691, acc_len=1.65, thru=16.1, I=24.1, T=25.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:47:16,747] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 [2026-10-19 13:47:23,456] loop.py:45 INFO TRAIN_STEP step=11179 loss=1.200917 Training: 62%|██████▏ | 11179/18018 [24:17:09<14:21:57, 7.56s/step, loss=1.201, acc=0.716, acc_len=1.73, thru=13.9, I=19.3, T=32.0, wait=3.0s, pool=61, epoch=2/2][2026-10-19 13:47:30,723] loop.py:380 INFO TIMING step=11180: step=3.788s data=1.594s compute=3.744s [fwd=2.673s bwd=1.055s opt=0.003s] dispatch=3.470s [2026-10-19 13:47:30,723] loop.py:45 INFO TRAIN_STEP step=11180 loss=1.237991 Training: 62%|██████▏ | 11180/18018 [24:17:16<14:11:42, 7.47s/step, loss=1.238, acc=0.712, acc_len=1.72, thru=17.7, I=16.8, T=33.8, wait=3.5s, pool=61, epoch=2/2][2026-10-19 13:47:38,735] loop.py:45 INFO TRAIN_STEP step=11181 loss=1.241366 Training: 62%|██████▏ | 11181/18018 [24:17:24<14:30:08, 7.64s/step, loss=1.241, acc=0.709, acc_len=1.71, thru=17.0, I=18.1, T=32.9, wait=4.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:47:38,731] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11180: forward=2782.2ms backward=1066.1ms [2026-10-19 13:47:48,218] loop.py:45 INFO TRAIN_STEP step=11182 loss=1.309545 Training: 62%|██████▏ | 11182/18018 [24:17:34<15:33:02, 8.19s/step, loss=1.310, acc=0.700, acc_len=1.68, thru=12.9, I=11.3, T=29.4, wait=5.1s, pool=53, epoch=2/2][2026-10-19 13:47:56,820] loop.py:45 INFO TRAIN_STEP step=11183 loss=1.218583 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:47:56,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 62%|██████▏ | 11183/18018 [24:17:42<15:46:59, 8.31s/step, loss=1.219, acc=0.715, acc_len=1.73, thru=17.4, I=15.0, T=32.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:47:57,271] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:03,660] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:48:04,280] loop.py:45 INFO TRAIN_STEP step=11184 loss=1.254512 Training: 62%|██████▏ | 11184/18018 [24:17:50<15:17:42, 8.06s/step, loss=1.255, acc=0.711, acc_len=1.72, thru=16.9, I=18.1, T=33.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:04,666] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:48:10,415] loop.py:380 INFO TIMING step=11185: step=3.691s data=1.645s compute=3.663s [fwd=2.619s bwd=1.030s opt=0.003s] dispatch=2.436s [2026-10-19 13:48:10,415] loop.py:45 INFO TRAIN_STEP step=11185 loss=1.384918 Training: 62%|██████▏ | 11185/18018 [24:17:56<14:11:53, 7.48s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=18.0, I=23.1, T=34.7, wait=2.4s, pool=49, epoch=2/2](TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 13:48:10,477] eagle_store.py:517 WARNING batch_get_buffer missing keys (k357920-9926d6ae_hs, k357920-9926d6ae_ids, k357920-9926d6ae_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:19,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9371) [2026-10-19 13:48:10,471] eagle_store.py:517 WARNING batch_get_buffer missing keys (k357918-8049c62f_hs, k357918-8049c62f_ids, k357918-8049c62f_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 13:48:20,671] loop.py:45 INFO TRAIN_STEP step=11186 loss=1.315125 Training: 62%|██████▏ | 11186/18018 [24:18:06<15:46:38, 8.31s/step, loss=1.315, acc=0.699, acc_len=1.68, thru=14.8, I=13.9, T=30.2, wait=6.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:48:20,667] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11185: forward=3041.3ms backward=1159.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:21,055] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:29,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:48:29,505] loop.py:45 INFO TRAIN_STEP step=11187 loss=1.243031 Training: 62%|██████▏ | 11187/18018 [24:18:15<16:04:15, 8.47s/step, loss=1.243, acc=0.712, acc_len=1.73, thru=13.2, I=15.4, T=31.7, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:29,641] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 13:48:36,012] loop.py:45 INFO TRAIN_STEP step=11188 loss=1.281531 Training: 62%|██████▏ | 11188/18018 [24:18:22<14:57:06, 7.88s/step, loss=1.282, acc=0.707, acc_len=1.70, thru=18.6, I=17.8, T=29.5, wait=2.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:43,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:48:43,524] loop.py:45 INFO TRAIN_STEP step=11189 loss=1.340948 Training: 62%|██████▏ | 11189/18018 [24:18:29<14:44:25, 7.77s/step, loss=1.341, acc=0.693, acc_len=1.66, thru=17.4, I=16.5, T=31.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:43,632] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:51,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:48:51,773] loop.py:380 INFO TIMING step=11190: step=5.065s data=2.018s compute=5.038s [fwd=3.773s bwd=1.251s opt=0.003s] dispatch=3.173s [2026-10-19 13:48:51,773] loop.py:45 INFO TRAIN_STEP step=11190 loss=1.324391 Training: 62%|██████▏ | 11190/18018 [24:18:37<15:00:36, 7.91s/step, loss=1.324, acc=0.697, acc_len=1.67, thru=16.2, I=17.8, T=25.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:52,165] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:57,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:48:58,067] loop.py:45 INFO TRAIN_STEP step=11191 loss=1.282381 Training: 62%|██████▏ | 11191/18018 [24:18:44<14:05:10, 7.43s/step, loss=1.282, acc=0.705, acc_len=1.70, thru=18.8, I=18.5, T=34.9, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:48:58,062] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11190: forward=2606.6ms backward=1014.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:48:58,275] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:03,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:49:04,511] loop.py:45 INFO TRAIN_STEP step=11192 loss=1.261123 Training: 62%|██████▏ | 11192/18018 [24:18:50<13:31:27, 7.13s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=19.2, I=22.0, T=35.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:04,841] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:10,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:49:11,171] loop.py:45 INFO TRAIN_STEP step=11193 loss=1.311303 Training: 62%|██████▏ | 11193/18018 [24:18:57<13:15:11, 6.99s/step, loss=1.311, acc=0.697, acc_len=1.67, thru=18.7, I=22.3, T=33.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:11,561] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:17,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:49:17,524] loop.py:45 INFO TRAIN_STEP step=11194 loss=1.283397 Training: 62%|██████▏ | 11194/18018 [24:19:03<12:53:18, 6.80s/step, loss=1.283, acc=0.700, acc_len=1.68, thru=18.7, I=21.8, T=34.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:17,585] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:23,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:49:24,906] loop.py:380 INFO TIMING step=11195: step=4.026s data=1.773s compute=3.983s [fwd=2.836s bwd=1.132s opt=0.003s] dispatch=3.348s [2026-10-19 13:49:24,907] loop.py:45 INFO TRAIN_STEP step=11195 loss=1.338889 Training: 62%|██████▏ | 11195/18018 [24:19:10<13:13:04, 6.97s/step, loss=1.339, acc=0.692, acc_len=1.66, thru=17.8, I=20.9, T=31.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:25,161] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:49:31,459] loop.py:45 INFO TRAIN_STEP step=11196 loss=1.245780 Training: 62%|██████▏ | 11196/18018 [24:19:17<12:58:35, 6.85s/step, loss=1.246, acc=0.709, acc_len=1.71, thru=15.1, I=22.4, T=33.6, wait=2.7s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:49:31,455] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11195: forward=2605.5ms backward=1149.2ms [2026-10-19 13:49:40,780] loop.py:45 INFO TRAIN_STEP step=11197 loss=1.230244 Training: 62%|██████▏ | 11197/18018 [24:19:26<14:22:53, 7.59s/step, loss=1.230, acc=0.715, acc_len=1.73, thru=16.1, I=14.3, T=33.7, wait=5.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:48,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:49:49,385] loop.py:45 INFO TRAIN_STEP step=11198 loss=1.278490 Training: 62%|██████▏ | 11198/18018 [24:19:35<14:57:21, 7.89s/step, loss=1.278, acc=0.706, acc_len=1.70, thru=15.3, I=15.6, T=31.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:49,420] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:55,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:49:56,083] loop.py:45 INFO TRAIN_STEP step=11199 loss=1.422495 Training: 62%|██████▏ | 11199/18018 [24:19:42<14:16:20, 7.53s/step, loss=1.422, acc=0.682, acc_len=1.63, thru=18.7, I=19.7, T=34.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:49:56,554] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:50:03,824] loop.py:380 INFO TIMING step=11200: step=3.581s data=1.566s compute=3.550s [fwd=2.522s bwd=1.014s opt=0.003s] dispatch=4.152s [2026-10-19 13:50:03,825] loop.py:45 INFO TRAIN_STEP step=11200 loss=1.365219 Training: 62%|██████▏ | 11200/18018 [24:19:49<14:23:15, 7.60s/step, loss=1.365, acc=0.687, acc_len=1.64, thru=14.9, I=16.3, T=35.7, wait=4.2s, pool=53, epoch=2/2][2026-10-19 13:50:13,148] loop.py:45 INFO TRAIN_STEP step=11201 loss=1.389726 Training: 62%|██████▏ | 11201/18018 [24:19:59<15:22:04, 8.12s/step, loss=1.390, acc=0.687, acc_len=1.64, thru=14.5, I=15.6, T=33.3, wait=5.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:50:13,143] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11200: forward=2723.7ms backward=1072.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:50:18,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:50:19,944] loop.py:45 INFO TRAIN_STEP step=11202 loss=1.409274 Training: 62%|██████▏ | 11202/18018 [24:20:06<14:36:56, 7.72s/step, loss=1.409, acc=0.682, acc_len=1.62, thru=17.8, I=20.0, T=36.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:50:19,976] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:50:27,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:50:29,452] loop.py:45 INFO TRAIN_STEP step=11203 loss=1.341597 Training: 62%|██████▏ | 11203/18018 [24:20:15<15:37:44, 8.26s/step, loss=1.342, acc=0.691, acc_len=1.66, thru=13.5, I=16.9, T=24.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:50:29,838] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:50:35,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:50:36,704] loop.py:45 INFO TRAIN_STEP step=11204 loss=1.202846 Training: 62%|██████▏ | 11204/18018 [24:20:22<15:03:24, 7.95s/step, loss=1.203, acc=0.718, acc_len=1.75, thru=14.1, I=20.4, T=28.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:50:36,795] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:50:44,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:50:44,744] loop.py:380 INFO TIMING step=11205: step=4.038s data=1.980s compute=3.980s [fwd=2.917s bwd=1.049s opt=0.003s] dispatch=3.992s [2026-10-19 13:50:44,744] loop.py:45 INFO TRAIN_STEP step=11205 loss=1.399775 Training: 62%|██████▏ | 11205/18018 [24:20:30<15:06:09, 7.98s/step, loss=1.400, acc=0.679, acc_len=1.61, thru=14.0, I=14.4, T=31.7, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:50:44,826] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 13:50:50,513] loop.py:45 INFO TRAIN_STEP step=11206 loss=1.233124 Training: 62%|██████▏ | 11206/18018 [24:20:36<13:50:41, 7.32s/step, loss=1.233, acc=0.712, acc_len=1.72, thru=18.3, I=19.9, T=34.8, wait=2.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:50:50,509] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11205: forward=2594.2ms backward=1039.4ms [2026-10-19 13:50:58,555] loop.py:45 INFO TRAIN_STEP step=11207 loss=1.293177 Training: 62%|██████▏ | 11207/18018 [24:20:44<14:15:17, 7.53s/step, loss=1.293, acc=0.703, acc_len=1.70, thru=16.4, I=17.5, T=33.1, wait=4.2s, pool=61, epoch=2/2][2026-10-19 13:51:05,734] loop.py:45 INFO TRAIN_STEP step=11208 loss=1.373084 Training: 62%|██████▏ | 11208/18018 [24:20:51<14:03:00, 7.43s/step, loss=1.373, acc=0.688, acc_len=1.64, thru=16.1, I=16.1, T=32.2, wait=3.2s, pool=61, epoch=2/2][2026-10-19 13:51:13,021] loop.py:45 INFO TRAIN_STEP step=11209 loss=1.306798 Training: 62%|██████▏ | 11209/18018 [24:20:59<13:58:11, 7.39s/step, loss=1.307, acc=0.698, acc_len=1.68, thru=16.1, I=17.8, T=31.9, wait=3.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:19,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:51:20,451] loop.py:380 INFO TIMING step=11210: step=3.767s data=1.549s compute=3.733s [fwd=2.616s bwd=1.102s opt=0.003s] dispatch=3.652s [2026-10-19 13:51:20,452] loop.py:45 INFO TRAIN_STEP step=11210 loss=1.303394 Training: 62%|██████▏ | 11210/18018 [24:21:06<13:59:30, 7.40s/step, loss=1.303, acc=0.697, acc_len=1.68, thru=17.2, I=18.1, T=34.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:20,699] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:25,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:51:26,154] loop.py:45 INFO TRAIN_STEP step=11211 loss=1.317076 Training: 62%|██████▏ | 11211/18018 [24:21:12<13:01:39, 6.89s/step, loss=1.317, acc=0.696, acc_len=1.67, thru=20.0, I=24.8, T=36.4, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:51:26,149] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11210: forward=2491.4ms backward=973.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:26,553] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:32,693] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:51:33,058] loop.py:45 INFO TRAIN_STEP step=11212 loss=1.211404 Training: 62%|██████▏ | 11212/18018 [24:21:19<13:02:01, 6.89s/step, loss=1.211, acc=0.718, acc_len=1.75, thru=19.2, I=22.8, T=36.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:33,210] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:38,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:51:39,546] loop.py:45 INFO TRAIN_STEP step=11213 loss=1.287838 Training: 62%|██████▏ | 11213/18018 [24:21:25<12:48:05, 6.77s/step, loss=1.288, acc=0.702, acc_len=1.69, thru=19.4, I=21.2, T=36.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:39,798] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:44,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:51:45,605] loop.py:45 INFO TRAIN_STEP step=11214 loss=1.307358 Training: 62%|██████▏ | 11214/18018 [24:21:31<12:23:46, 6.56s/step, loss=1.307, acc=0.698, acc_len=1.67, thru=21.2, I=26.3, T=34.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:45,799] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:50,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:51:51,573] loop.py:380 INFO TIMING step=11215: step=3.388s data=1.414s compute=3.358s [fwd=2.407s bwd=0.936s opt=0.003s] dispatch=2.569s [2026-10-19 13:51:51,573] loop.py:45 INFO TRAIN_STEP step=11215 loss=1.182626 Training: 62%|██████▏ | 11215/18018 [24:21:37<12:03:32, 6.38s/step, loss=1.183, acc=0.721, acc_len=1.76, thru=20.4, I=27.1, T=37.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:51,793] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:55,915] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:51:57,213] loop.py:45 INFO TRAIN_STEP step=11216 loss=1.280706 Training: 62%|██████▏ | 11216/18018 [24:21:43<11:38:12, 6.16s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=22.3, I=30.9, T=38.6, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:51:57,209] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11215: forward=2372.0ms backward=897.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:51:57,426] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:52:03,048] loop.py:45 INFO TRAIN_STEP step=11217 loss=1.254713 Training: 62%|██████▏ | 11217/18018 [24:21:49<11:27:03, 6.06s/step, loss=1.255, acc=0.709, acc_len=1.72, thru=18.5, I=27.1, T=38.9, wait=2.5s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:13,868] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:52:13,948] loop.py:45 INFO TRAIN_STEP step=11218 loss=1.306275 Training: 62%|██████▏ | 11218/18018 [24:22:00<14:11:26, 7.51s/step, loss=1.306, acc=0.695, acc_len=1.66, thru=15.3, I=14.3, T=33.9, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:14,370] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:20,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:52:22,016] loop.py:45 INFO TRAIN_STEP step=11219 loss=1.393065 Training: 62%|██████▏ | 11219/18018 [24:22:08<14:30:17, 7.68s/step, loss=1.393, acc=0.682, acc_len=1.63, thru=15.5, I=18.4, T=32.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:22,138] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:26,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:52:28,237] loop.py:380 INFO TIMING step=11220: step=3.624s data=1.511s compute=3.592s [fwd=2.532s bwd=1.045s opt=0.003s] dispatch=2.586s [2026-10-19 13:52:28,237] loop.py:45 INFO TRAIN_STEP step=11220 loss=1.309052 Training: 62%|██████▏ | 11220/18018 [24:22:14<13:40:33, 7.24s/step, loss=1.309, acc=0.698, acc_len=1.67, thru=18.6, I=27.7, T=35.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:28,288] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:33,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:52:34,616] loop.py:45 INFO TRAIN_STEP step=11221 loss=1.274623 Training: 62%|██████▏ | 11221/18018 [24:22:20<13:11:05, 6.98s/step, loss=1.275, acc=0.702, acc_len=1.68, thru=18.8, I=26.2, T=34.1, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:52:34,611] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11220: forward=2665.3ms backward=1039.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:34,739] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:40,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:52:42,869] loop.py:45 INFO TRAIN_STEP step=11222 loss=1.316418 Training: 62%|██████▏ | 11222/18018 [24:22:28<13:54:02, 7.36s/step, loss=1.316, acc=0.699, acc_len=1.67, thru=13.7, I=19.6, T=30.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:43,239] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:49,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:52:51,123] loop.py:45 INFO TRAIN_STEP step=11223 loss=1.200489 Training: 62%|██████▏ | 11223/18018 [24:22:37<14:24:10, 7.63s/step, loss=1.200, acc=0.714, acc_len=1.73, thru=16.2, I=19.0, T=30.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:52:51,589] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:52:56,995] loop.py:45 INFO TRAIN_STEP step=11224 loss=1.273622 Training: 62%|██████▏ | 11224/18018 [24:22:43<13:24:19, 7.10s/step, loss=1.274, acc=0.708, acc_len=1.71, thru=19.0, I=22.8, T=36.8, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:02,408] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:53:03,722] loop.py:380 INFO TIMING step=11225: step=3.702s data=1.672s compute=3.669s [fwd=2.591s bwd=1.064s opt=0.003s] dispatch=3.017s [2026-10-19 13:53:03,722] loop.py:45 INFO TRAIN_STEP step=11225 loss=1.338336 Training: 62%|██████▏ | 11225/18018 [24:22:49<13:11:29, 6.99s/step, loss=1.338, acc=0.693, acc_len=1.65, thru=18.5, I=21.0, T=34.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:03,917] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:53:09,970] loop.py:45 INFO TRAIN_STEP step=11226 loss=1.299603 Training: 62%|██████▏ | 11226/18018 [24:22:56<12:46:02, 6.77s/step, loss=1.300, acc=0.700, acc_len=1.68, thru=18.0, I=21.2, T=36.3, wait=2.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:53:09,965] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11225: forward=2507.2ms backward=973.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:19,418] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:53:19,594] loop.py:45 INFO TRAIN_STEP step=11227 loss=1.387276 Training: 62%|██████▏ | 11227/18018 [24:23:05<14:22:59, 7.62s/step, loss=1.387, acc=0.688, acc_len=1.64, thru=15.3, I=14.7, T=34.4, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:19,920] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:25,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:53:26,300] loop.py:45 INFO TRAIN_STEP step=11228 loss=1.358825 Training: 62%|██████▏ | 11228/18018 [24:23:12<13:51:38, 7.35s/step, loss=1.359, acc=0.690, acc_len=1.64, thru=18.9, I=18.4, T=35.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:26,681] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:53:32,118] loop.py:45 INFO TRAIN_STEP step=11229 loss=1.254879 Training: 62%|██████▏ | 11229/18018 [24:23:18<12:59:35, 6.89s/step, loss=1.255, acc=0.709, acc_len=1.71, thru=20.8, I=24.4, T=38.3, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:37,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:53:38,547] loop.py:380 INFO TIMING step=11230: step=3.579s data=1.482s compute=3.556s [fwd=2.501s bwd=1.041s opt=0.003s] dispatch=2.841s [2026-10-19 13:53:38,547] loop.py:45 INFO TRAIN_STEP step=11230 loss=1.255882 Training: 62%|██████▏ | 11230/18018 [24:23:24<12:43:52, 6.75s/step, loss=1.256, acc=0.708, acc_len=1.71, thru=20.5, I=22.5, T=35.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:38,719] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:53:46,818] loop.py:45 INFO TRAIN_STEP step=11231 loss=1.277991 Training: 62%|██████▏ | 11231/18018 [24:23:32<13:35:17, 7.21s/step, loss=1.278, acc=0.703, acc_len=1.69, thru=13.2, I=15.2, T=30.7, wait=4.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:53:46,813] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11230: forward=2922.2ms backward=1206.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:55,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:53:56,362] loop.py:45 INFO TRAIN_STEP step=11232 loss=1.314676 Training: 62%|██████▏ | 11232/18018 [24:23:42<14:54:28, 7.91s/step, loss=1.315, acc=0.698, acc_len=1.67, thru=15.4, I=16.1, T=35.2, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:53:56,454] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:02,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:54:03,073] loop.py:45 INFO TRAIN_STEP step=11233 loss=1.353908 Training: 62%|██████▏ | 11233/18018 [24:23:49<14:13:40, 7.55s/step, loss=1.354, acc=0.691, acc_len=1.65, thru=19.4, I=19.4, T=37.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:03,548] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:08,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:54:09,811] loop.py:45 INFO TRAIN_STEP step=11234 loss=1.476287 Training: 62%|██████▏ | 11234/18018 [24:23:55<13:46:00, 7.31s/step, loss=1.476, acc=0.672, acc_len=1.58, thru=19.1, I=22.3, T=32.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:10,228] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:54:16,356] loop.py:380 INFO TIMING step=11235: step=3.842s data=1.758s compute=3.815s [fwd=2.718s bwd=1.082s opt=0.003s] dispatch=2.693s [2026-10-19 13:54:16,356] loop.py:45 INFO TRAIN_STEP step=11235 loss=1.377313 Training: 62%|██████▏ | 11235/18018 [24:24:02<13:20:07, 7.08s/step, loss=1.377, acc=0.684, acc_len=1.63, thru=18.2, I=22.1, T=33.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:16,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:16,837] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:22,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:54:23,871] loop.py:45 INFO TRAIN_STEP step=11236 loss=1.200717 Training: 62%|██████▏ | 11236/18018 [24:24:09<13:34:48, 7.21s/step, loss=1.201, acc=0.716, acc_len=1.74, thru=16.9, I=18.8, T=34.0, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:54:23,867] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11235: forward=2688.2ms backward=1033.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:24,267] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:31,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:54:32,129] loop.py:45 INFO TRAIN_STEP step=11237 loss=1.383762 Training: 62%|██████▏ | 11237/18018 [24:24:18<14:10:13, 7.52s/step, loss=1.384, acc=0.686, acc_len=1.63, thru=14.2, I=15.8, T=29.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:32,405] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:37,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:54:38,315] loop.py:45 INFO TRAIN_STEP step=11238 loss=1.331570 Training: 62%|██████▏ | 11238/18018 [24:24:24<13:24:49, 7.12s/step, loss=1.332, acc=0.695, acc_len=1.67, thru=18.5, I=20.2, T=35.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:38,593] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:44,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:54:45,015] loop.py:45 INFO TRAIN_STEP step=11239 loss=1.416759 Training: 62%|██████▏ | 11239/18018 [24:24:31<13:10:22, 7.00s/step, loss=1.417, acc=0.678, acc_len=1.61, thru=19.0, I=20.4, T=35.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:45,416] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:53,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:54:55,240] loop.py:380 INFO TIMING step=11240: step=3.922s data=1.777s compute=3.893s [fwd=2.760s bwd=1.118s opt=0.003s] dispatch=6.294s [2026-10-19 13:54:55,240] loop.py:45 INFO TRAIN_STEP step=11240 loss=1.403016 Training: 62%|██████▏ | 11240/18018 [24:24:41<14:59:46, 7.96s/step, loss=1.403, acc=0.682, acc_len=1.63, thru=13.0, I=14.6, T=32.6, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:54:55,386] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 13:55:00,692] loop.py:45 INFO TRAIN_STEP step=11241 loss=1.169211 Training: 62%|██████▏ | 11241/18018 [24:24:46<13:34:47, 7.21s/step, loss=1.169, acc=0.722, acc_len=1.75, thru=19.1, I=21.0, T=37.2, wait=2.0s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:55:00,686] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11240: forward=2438.1ms backward=960.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:07,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:55:09,239] loop.py:45 INFO TRAIN_STEP step=11242 loss=1.240156 Training: 62%|██████▏ | 11242/18018 [24:24:55<14:19:27, 7.61s/step, loss=1.240, acc=0.712, acc_len=1.72, thru=17.4, I=19.5, T=32.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:09,681] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:14,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:55:15,923] loop.py:45 INFO TRAIN_STEP step=11243 loss=1.191437 Training: 62%|██████▏ | 11243/18018 [24:25:02<13:48:01, 7.33s/step, loss=1.191, acc=0.719, acc_len=1.75, thru=18.1, I=25.1, T=34.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:16,281] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 13:55:23,125] loop.py:45 INFO TRAIN_STEP step=11244 loss=1.380473 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:23,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 62%|██████▏ | 11244/18018 [24:25:09<13:43:20, 7.29s/step, loss=1.380, acc=0.684, acc_len=1.62, thru=16.9, I=20.4, T=33.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:23,625] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:30,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:55:31,562] loop.py:380 INFO TIMING step=11245: step=3.741s data=1.542s compute=3.705s [fwd=2.669s bwd=1.021s opt=0.003s] dispatch=4.690s [2026-10-19 13:55:31,562] loop.py:45 INFO TRAIN_STEP step=11245 loss=1.307480 Training: 62%|██████▏ | 11245/18018 [24:25:17<14:22:02, 7.64s/step, loss=1.307, acc=0.701, acc_len=1.69, thru=16.9, I=18.6, T=34.2, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:32,005] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:55:37,911] loop.py:45 INFO TRAIN_STEP step=11246 loss=1.318007 Training: 62%|██████▏ | 11246/18018 [24:25:23<13:38:17, 7.25s/step, loss=1.318, acc=0.693, acc_len=1.65, thru=18.2, I=20.1, T=34.6, wait=2.6s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:55:37,906] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11245: forward=2549.6ms backward=1098.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:44,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:55:45,122] loop.py:45 INFO TRAIN_STEP step=11247 loss=1.296502 Training: 62%|██████▏ | 11247/18018 [24:25:31<13:36:55, 7.24s/step, loss=1.297, acc=0.701, acc_len=1.69, thru=19.3, I=18.7, T=33.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:45,255] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 13:55:50,370] loop.py:45 INFO TRAIN_STEP step=11248 loss=1.317547 Training: 62%|██████▏ | 11248/18018 [24:25:36<12:29:19, 6.64s/step, loss=1.318, acc=0.695, acc_len=1.66, thru=22.1, I=22.7, T=37.6, wait=1.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:56,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:55:57,775] loop.py:45 INFO TRAIN_STEP step=11249 loss=1.297210 Training: 62%|██████▏ | 11249/18018 [24:25:43<12:55:06, 6.87s/step, loss=1.297, acc=0.704, acc_len=1.69, thru=18.5, I=20.6, T=35.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:55:57,961] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:56:03,348] loop.py:380 INFO TIMING step=11250: step=3.242s data=1.322s compute=3.213s [fwd=2.287s bwd=0.912s opt=0.003s] dispatch=2.322s [2026-10-19 13:56:03,349] loop.py:45 INFO TRAIN_STEP step=11250 loss=1.348515 Training: 62%|██████▏ | 11250/18018 [24:25:49<12:11:06, 6.48s/step, loss=1.349, acc=0.691, acc_len=1.65, thru=20.8, I=23.8, T=39.5, wait=2.3s, pool=53, epoch=2/2][2026-10-19 13:56:03,352] loop.py:419 INFO Saving checkpoint at step 11250... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:56:03,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 13:56:06,698] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0011251 (TrainerActor pid=9310) [2026-10-19 13:56:06,705] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0010501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:56:07,926] inference_manager.py:408 INFO Pool capacity available after 4.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:56:11,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 13:56:06,699] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0011251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 13:56:07,875] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0010501 [2026-10-19 13:56:13,882] loop.py:45 INFO TRAIN_STEP step=11251 loss=1.324183 Training: 62%|██████▏ | 11251/18018 [24:25:59<14:28:05, 7.70s/step, loss=1.324, acc=0.697, acc_len=1.67, thru=21.5, I=28.8, T=32.9, wait=2.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:56:13,878] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11250: forward=2743.5ms backward=1098.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:56:14,313] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:56:21,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:56:22,081] loop.py:45 INFO TRAIN_STEP step=11252 loss=1.315791 Training: 62%|██████▏ | 11252/18018 [24:26:08<14:44:57, 7.85s/step, loss=1.316, acc=0.696, acc_len=1.67, thru=16.5, I=17.2, T=34.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:56:22,566] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:56:28,427] loop.py:45 INFO TRAIN_STEP step=11253 loss=1.106837 Training: 62%|██████▏ | 11253/18018 [24:26:14<13:53:57, 7.40s/step, loss=1.107, acc=0.733, acc_len=1.80, thru=19.0, I=18.9, T=37.2, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:56:34,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:56:35,439] loop.py:45 INFO TRAIN_STEP step=11254 loss=1.375526 Training: 62%|██████▏ | 11254/18018 [24:26:21<13:40:56, 7.28s/step, loss=1.376, acc=0.689, acc_len=1.64, thru=19.7, I=21.2, T=33.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:56:35,631] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:56:41,940] loop.py:380 INFO TIMING step=11255: step=3.607s data=1.572s compute=3.579s [fwd=2.581s bwd=0.984s opt=0.003s] dispatch=2.883s [2026-10-19 13:56:41,940] loop.py:45 INFO TRAIN_STEP step=11255 loss=1.278499 Training: 62%|██████▏ | 11255/18018 [24:26:28<13:14:21, 7.05s/step, loss=1.278, acc=0.703, acc_len=1.69, thru=17.7, I=22.5, T=35.5, wait=2.9s, pool=57, epoch=2/2][2026-10-19 13:56:49,705] loop.py:45 INFO TRAIN_STEP step=11256 loss=1.322134 Training: 62%|██████▏ | 11256/18018 [24:26:35<13:38:29, 7.26s/step, loss=1.322, acc=0.694, acc_len=1.66, thru=13.5, I=14.1, T=34.8, wait=4.1s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:56:49,701] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11255: forward=2494.1ms backward=1124.4ms [2026-10-19 13:56:56,951] loop.py:45 INFO TRAIN_STEP step=11257 loss=1.258537 Training: 62%|██████▏ | 11257/18018 [24:26:43<13:37:54, 7.26s/step, loss=1.259, acc=0.710, acc_len=1.71, thru=16.9, I=17.6, T=34.6, wait=3.5s, pool=45, epoch=2/2][2026-10-19 13:57:05,071] loop.py:45 INFO TRAIN_STEP step=11258 loss=1.243620 Training: 62%|██████▏ | 11258/18018 [24:26:51<14:06:49, 7.52s/step, loss=1.244, acc=0.708, acc_len=1.71, thru=16.5, I=15.9, T=36.2, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:57:16,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:57:16,876] loop.py:45 INFO TRAIN_STEP step=11259 loss=1.348221 Training: 62%|██████▏ | 11259/18018 [24:27:02<16:31:37, 8.80s/step, loss=1.348, acc=0.694, acc_len=1.66, thru=12.3, I=12.8, T=31.6, wait=7.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:57:16,907] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:57:23,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:57:24,147] loop.py:380 INFO TIMING step=11260: step=3.609s data=1.486s compute=3.573s [fwd=2.496s bwd=1.063s opt=0.003s] dispatch=3.653s [2026-10-19 13:57:24,147] loop.py:45 INFO TRAIN_STEP step=11260 loss=1.479325 Training: 62%|██████▏ | 11260/18018 [24:27:10<15:39:47, 8.34s/step, loss=1.479, acc=0.669, acc_len=1.56, thru=18.1, I=17.3, T=35.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:57:24,384] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 13:57:29,935] loop.py:45 INFO TRAIN_STEP step=11261 loss=1.141141 Training: 62%|██████▏ | 11261/18018 [24:27:16<14:13:12, 7.58s/step, loss=1.141, acc=0.729, acc_len=1.78, thru=19.8, I=22.7, T=35.5, wait=2.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:57:29,931] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11260: forward=2535.0ms backward=1023.4ms [2026-10-19 13:57:38,146] loop.py:45 INFO TRAIN_STEP step=11262 loss=1.254424 Training: 63%|██████▎ | 11262/18018 [24:27:24<14:34:30, 7.77s/step, loss=1.254, acc=0.707, acc_len=1.71, thru=16.0, I=15.9, T=33.6, wait=4.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:57:44,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:57:44,984] loop.py:45 INFO TRAIN_STEP step=11263 loss=1.428395 Training: 63%|██████▎ | 11263/18018 [24:27:31<14:03:07, 7.49s/step, loss=1.428, acc=0.679, acc_len=1.60, thru=19.1, I=19.2, T=34.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:57:45,384] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:57:53,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:57:53,656] loop.py:45 INFO TRAIN_STEP step=11264 loss=1.271350 Training: 63%|██████▎ | 11264/18018 [24:27:39<14:42:50, 7.84s/step, loss=1.271, acc=0.707, acc_len=1.69, thru=14.8, I=16.3, T=31.5, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:57:54,140] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:01,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:58:01,958] loop.py:380 INFO TIMING step=11265: step=3.656s data=1.560s compute=3.621s [fwd=2.639s bwd=0.967s opt=0.003s] dispatch=4.639s [2026-10-19 13:58:01,958] loop.py:45 INFO TRAIN_STEP step=11265 loss=1.337294 Training: 63%|██████▎ | 11265/18018 [24:27:48<14:58:12, 7.98s/step, loss=1.337, acc=0.693, acc_len=1.66, thru=15.3, I=17.3, T=35.0, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:02,128] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:07,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:58:08,053] loop.py:45 INFO TRAIN_STEP step=11266 loss=1.320907 Training: 63%|██████▎ | 11266/18018 [24:27:54<13:54:29, 7.42s/step, loss=1.321, acc=0.694, acc_len=1.66, thru=20.0, I=23.6, T=34.4, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:58:08,049] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11265: forward=2576.8ms backward=1095.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:08,081] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:13,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:58:14,399] loop.py:45 INFO TRAIN_STEP step=11267 loss=1.262176 Training: 63%|██████▎ | 11267/18018 [24:28:00<13:18:17, 7.09s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=19.2, I=24.1, T=34.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:14,570] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 13:58:22,016] loop.py:45 INFO TRAIN_STEP step=11268 loss=1.295094 Training: 63%|██████▎ | 11268/18018 [24:28:08<13:35:46, 7.25s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=13.7, I=16.5, T=29.3, wait=3.2s, pool=41, epoch=2/2][2026-10-19 13:58:31,078] loop.py:45 INFO TRAIN_STEP step=11269 loss=1.264321 Training: 63%|██████▎ | 11269/18018 [24:28:17<14:36:45, 7.79s/step, loss=1.264, acc=0.708, acc_len=1.71, thru=16.1, I=15.4, T=31.9, wait=5.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:36,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:58:37,867] loop.py:380 INFO TIMING step=11270: step=4.154s data=1.870s compute=4.122s [fwd=2.860s bwd=1.247s opt=0.003s] dispatch=2.625s [2026-10-19 13:58:37,867] loop.py:45 INFO TRAIN_STEP step=11270 loss=1.431571 Training: 63%|██████▎ | 11270/18018 [24:28:23<14:02:41, 7.49s/step, loss=1.432, acc=0.678, acc_len=1.60, thru=17.9, I=19.8, T=30.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:38,194] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:45,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:58:46,101] loop.py:45 INFO TRAIN_STEP step=11271 loss=1.317269 Training: 63%|██████▎ | 11271/18018 [24:28:32<14:27:36, 7.72s/step, loss=1.317, acc=0.703, acc_len=1.69, thru=14.1, I=16.4, T=32.8, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:58:46,097] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11270: forward=2780.1ms backward=1084.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:46,178] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:52,328] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:58:53,716] loop.py:45 INFO TRAIN_STEP step=11272 loss=1.276477 Training: 63%|██████▎ | 11272/18018 [24:28:39<14:24:05, 7.69s/step, loss=1.276, acc=0.707, acc_len=1.70, thru=17.0, I=19.1, T=32.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:58:53,838] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 13:58:59,971] loop.py:45 INFO TRAIN_STEP step=11273 loss=1.293984 Training: 63%|██████▎ | 11273/18018 [24:28:46<13:35:39, 7.26s/step, loss=1.294, acc=0.704, acc_len=1.70, thru=17.8, I=22.3, T=37.3, wait=2.8s, pool=53, epoch=2/2][2026-10-19 13:59:09,061] loop.py:45 INFO TRAIN_STEP step=11274 loss=1.270009 Training: 63%|██████▎ | 11274/18018 [24:28:55<14:37:29, 7.81s/step, loss=1.270, acc=0.708, acc_len=1.71, thru=14.3, I=15.1, T=32.8, wait=5.2s, pool=53, epoch=2/2][2026-10-19 13:59:19,877] loop.py:380 INFO TIMING step=11275: step=4.162s data=1.776s compute=4.132s [fwd=2.893s bwd=1.224s opt=0.003s] dispatch=6.642s [2026-10-19 13:59:19,877] loop.py:45 INFO TRAIN_STEP step=11275 loss=1.321994 Training: 63%|██████▎ | 11275/18018 [24:29:05<16:18:47, 8.71s/step, loss=1.322, acc=0.697, acc_len=1.67, thru=12.0, I=12.7, T=30.8, wait=6.6s, pool=45, epoch=2/2][2026-10-19 13:59:28,984] loop.py:45 INFO TRAIN_STEP step=11276 loss=1.344035 Training: 63%|██████▎ | 11276/18018 [24:29:15<16:31:58, 8.83s/step, loss=1.344, acc=0.689, acc_len=1.64, thru=14.6, I=12.4, T=33.5, wait=5.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 13:59:28,980] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11275: forward=2686.3ms backward=1082.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:59:36,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:59:38,088] loop.py:45 INFO TRAIN_STEP step=11277 loss=1.374132 Training: 63%|██████▎ | 11277/18018 [24:29:24<16:41:08, 8.91s/step, loss=1.374, acc=0.686, acc_len=1.63, thru=16.1, I=17.2, T=31.3, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:59:38,264] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:59:45,659] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:59:47,494] loop.py:45 INFO TRAIN_STEP step=11278 loss=1.272101 Training: 63%|██████▎ | 11278/18018 [24:29:33<16:57:43, 9.06s/step, loss=1.272, acc=0.700, acc_len=1.68, thru=13.9, I=16.4, T=27.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:59:47,672] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:59:53,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 13:59:54,684] loop.py:45 INFO TRAIN_STEP step=11279 loss=1.271811 Training: 63%|██████▎ | 11279/18018 [24:29:40<15:54:34, 8.50s/step, loss=1.272, acc=0.706, acc_len=1.70, thru=15.2, I=21.3, T=34.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 13:59:55,052] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:01,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:00:01,310] loop.py:380 INFO TIMING step=11280: step=3.357s data=1.346s compute=3.319s [fwd=2.367s bwd=0.937s opt=0.003s] dispatch=3.259s [2026-10-19 14:00:01,310] loop.py:45 INFO TRAIN_STEP step=11280 loss=1.179686 Training: 63%|██████▎ | 11280/18018 [24:29:47<14:51:13, 7.94s/step, loss=1.180, acc=0.721, acc_len=1.75, thru=18.8, I=22.2, T=38.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:01,717] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:00:08,141] loop.py:45 INFO TRAIN_STEP step=11281 loss=1.255980 Training: 63%|██████▎ | 11281/18018 [24:29:54<14:13:52, 7.60s/step, loss=1.256, acc=0.709, acc_len=1.71, thru=16.9, I=17.6, T=35.1, wait=3.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:00:08,137] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11280: forward=2578.7ms backward=1027.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:16,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:00:17,631] loop.py:45 INFO TRAIN_STEP step=11282 loss=1.385926 Training: 63%|██████▎ | 11282/18018 [24:30:03<15:17:17, 8.17s/step, loss=1.386, acc=0.685, acc_len=1.63, thru=15.6, I=16.7, T=30.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:17,877] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:00:24,258] loop.py:45 INFO TRAIN_STEP step=11283 loss=1.276515 Training: 63%|██████▎ | 11283/18018 [24:30:10<14:25:08, 7.71s/step, loss=1.277, acc=0.702, acc_len=1.69, thru=17.3, I=18.1, T=35.2, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:32,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:00:32,784] loop.py:45 INFO TRAIN_STEP step=11284 loss=1.332868 Training: 63%|██████▎ | 11284/18018 [24:30:18<14:52:35, 7.95s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=15.2, I=16.7, T=35.0, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:33,235] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:39,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:00:41,207] loop.py:380 INFO TIMING step=11285: step=3.907s data=1.676s compute=3.880s [fwd=2.718s bwd=1.148s opt=0.003s] dispatch=4.508s [2026-10-19 14:00:41,207] loop.py:45 INFO TRAIN_STEP step=11285 loss=1.455175 Training: 63%|██████▎ | 11285/18018 [24:30:27<15:08:21, 8.09s/step, loss=1.455, acc=0.675, acc_len=1.59, thru=14.6, I=19.7, T=32.8, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:41,405] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:00:47,242] loop.py:45 INFO TRAIN_STEP step=11286 loss=1.131195 Training: 63%|██████▎ | 11286/18018 [24:30:33<13:58:46, 7.48s/step, loss=1.131, acc=0.732, acc_len=1.79, thru=18.9, I=21.8, T=36.2, wait=2.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:00:47,237] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11285: forward=2479.6ms backward=1000.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:54,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:00:56,035] loop.py:45 INFO TRAIN_STEP step=11287 loss=1.345544 Training: 63%|██████▎ | 11287/18018 [24:30:42<14:43:06, 7.87s/step, loss=1.346, acc=0.692, acc_len=1.65, thru=16.5, I=18.8, T=31.2, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:00:56,171] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:01:01,922] loop.py:45 INFO TRAIN_STEP step=11288 loss=1.281980 Training: 63%|██████▎ | 11288/18018 [24:30:48<13:36:09, 7.28s/step, loss=1.282, acc=0.703, acc_len=1.69, thru=17.9, I=25.1, T=40.2, wait=2.7s, pool=49, epoch=2/2][2026-10-19 14:01:11,578] loop.py:45 INFO TRAIN_STEP step=11289 loss=1.378212 Training: 63%|██████▎ | 11289/18018 [24:30:57<14:56:06, 7.99s/step, loss=1.378, acc=0.687, acc_len=1.64, thru=13.4, I=13.2, T=30.1, wait=5.4s, pool=37, epoch=2/2][2026-10-19 14:01:23,642] loop.py:380 INFO TIMING step=11290: step=3.868s data=1.732s compute=3.812s [fwd=2.625s bwd=1.173s opt=0.003s] dispatch=8.185s [2026-10-19 14:01:23,643] loop.py:45 INFO TRAIN_STEP step=11290 loss=1.400745 Training: 63%|██████▎ | 11290/18018 [24:31:09<17:13:07, 9.21s/step, loss=1.401, acc=0.685, acc_len=1.63, thru=13.5, I=12.4, T=33.1, wait=8.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:23,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:24,114] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:28,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:01:29,668] loop.py:45 INFO TRAIN_STEP step=11291 loss=1.306309 Training: 63%|██████▎ | 11291/18018 [24:31:15<15:25:38, 8.26s/step, loss=1.306, acc=0.698, acc_len=1.68, thru=19.7, I=21.3, T=35.8, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:01:29,663] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11290: forward=2466.4ms backward=1049.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:29,866] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:34,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:01:35,486] loop.py:45 INFO TRAIN_STEP step=11292 loss=1.294192 Training: 63%|██████▎ | 11292/18018 [24:31:21<14:03:25, 7.52s/step, loss=1.294, acc=0.698, acc_len=1.68, thru=21.0, I=24.6, T=37.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:35,976] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:42,279] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:01:43,195] loop.py:45 INFO TRAIN_STEP step=11293 loss=1.344785 Training: 63%|██████▎ | 11293/18018 [24:31:29<14:09:37, 7.58s/step, loss=1.345, acc=0.690, acc_len=1.65, thru=16.9, I=20.2, T=32.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:43,287] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:50,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:01:51,536] loop.py:45 INFO TRAIN_STEP step=11294 loss=1.346000 Training: 63%|██████▎ | 11294/18018 [24:31:37<14:35:03, 7.81s/step, loss=1.346, acc=0.692, acc_len=1.65, thru=14.2, I=16.8, T=32.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:01:51,980] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:01:57,971] loop.py:380 INFO TIMING step=11295: step=3.375s data=1.377s compute=3.339s [fwd=2.391s bwd=0.934s opt=0.003s] dispatch=3.051s [2026-10-19 14:01:57,971] loop.py:45 INFO TRAIN_STEP step=11295 loss=1.271211 Training: 63%|██████▎ | 11295/18018 [24:31:44<13:48:47, 7.40s/step, loss=1.271, acc=0.703, acc_len=1.69, thru=18.4, I=19.4, T=37.9, wait=3.1s, pool=61, epoch=2/2][2026-10-19 14:02:05,082] loop.py:45 INFO TRAIN_STEP step=11296 loss=1.280063 Training: 63%|██████▎ | 11296/18018 [24:31:51<13:39:01, 7.31s/step, loss=1.280, acc=0.701, acc_len=1.69, thru=17.3, I=16.9, T=38.3, wait=3.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:02:05,078] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11295: forward=2386.9ms backward=905.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:02:13,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:02:14,399] loop.py:45 INFO TRAIN_STEP step=11297 loss=1.273267 Training: 63%|██████▎ | 11297/18018 [24:32:00<14:46:16, 7.91s/step, loss=1.273, acc=0.706, acc_len=1.70, thru=15.7, I=16.7, T=31.7, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:02:14,843] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:02:20,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:02:21,255] loop.py:45 INFO TRAIN_STEP step=11298 loss=1.475663 Training: 63%|██████▎ | 11298/18018 [24:32:07<14:10:43, 7.60s/step, loss=1.476, acc=0.669, acc_len=1.58, thru=16.3, I=19.4, T=33.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:02:21,398] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 14:02:28,051] loop.py:45 INFO TRAIN_STEP step=11299 loss=1.225172 Training: 63%|██████▎ | 11299/18018 [24:32:14<13:43:41, 7.36s/step, loss=1.225, acc=0.715, acc_len=1.73, thru=18.6, I=18.7, T=33.2, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:02:35,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:02:36,617] loop.py:380 INFO TIMING step=11300: step=3.654s data=1.551s compute=3.616s [fwd=2.565s bwd=1.036s opt=0.003s] dispatch=4.903s [2026-10-19 14:02:36,617] loop.py:45 INFO TRAIN_STEP step=11300 loss=1.225962 Training: 63%|██████▎ | 11300/18018 [24:32:22<14:24:12, 7.72s/step, loss=1.226, acc=0.710, acc_len=1.72, thru=16.7, I=16.7, T=35.0, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:02:36,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:02:43,529] loop.py:45 INFO TRAIN_STEP step=11301 loss=1.353758 Training: 63%|██████▎ | 11301/18018 [24:32:29<13:57:00, 7.48s/step, loss=1.354, acc=0.692, acc_len=1.65, thru=15.1, I=16.8, T=36.2, wait=3.4s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:02:43,524] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11300: forward=2492.7ms backward=998.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:02:55,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:02:56,353] loop.py:45 INFO TRAIN_STEP step=11302 loss=1.426577 Training: 63%|██████▎ | 11302/18018 [24:32:42<16:56:30, 9.08s/step, loss=1.427, acc=0.679, acc_len=1.61, thru=11.9, I=12.8, T=29.5, wait=8.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:02:56,519] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:03:03,538] loop.py:45 INFO TRAIN_STEP step=11303 loss=1.327256 Training: 63%|██████▎ | 11303/18018 [24:32:49<15:52:35, 8.51s/step, loss=1.327, acc=0.696, acc_len=1.67, thru=15.7, I=17.0, T=35.3, wait=3.5s, pool=49, epoch=2/2][2026-10-19 14:03:10,924] loop.py:45 INFO TRAIN_STEP step=11304 loss=1.368390 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:10,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 63%|██████▎ | 11304/18018 [24:32:57<15:14:40, 8.17s/step, loss=1.368, acc=0.691, acc_len=1.64, thru=18.1, I=17.9, T=32.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:11,400] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:16,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:03:16,801] loop.py:380 INFO TIMING step=11305: step=3.485s data=1.397s compute=3.456s [fwd=2.467s bwd=0.974s opt=0.003s] dispatch=2.384s [2026-10-19 14:03:16,802] loop.py:45 INFO TRAIN_STEP step=11305 loss=1.207288 Training: 63%|██████▎ | 11305/18018 [24:33:02<13:57:28, 7.49s/step, loss=1.207, acc=0.716, acc_len=1.74, thru=20.4, I=21.7, T=36.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:17,007] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:21,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:03:22,972] loop.py:45 INFO TRAIN_STEP step=11306 loss=1.258996 Training: 63%|██████▎ | 11306/18018 [24:33:09<13:13:16, 7.09s/step, loss=1.259, acc=0.704, acc_len=1.69, thru=20.2, I=24.6, T=33.9, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:03:22,967] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11305: forward=2623.9ms backward=1100.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:23,299] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:29,693] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:03:30,355] loop.py:45 INFO TRAIN_STEP step=11307 loss=1.326672 Training: 63%|██████▎ | 11307/18018 [24:33:16<13:22:57, 7.18s/step, loss=1.327, acc=0.695, acc_len=1.67, thru=15.9, I=20.2, T=35.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:30,697] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:03:37,513] loop.py:45 INFO TRAIN_STEP step=11308 loss=1.366875 Training: 63%|██████▎ | 11308/18018 [24:33:23<13:22:06, 7.17s/step, loss=1.367, acc=0.690, acc_len=1.64, thru=17.7, I=20.2, T=32.8, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:43,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:03:44,686] loop.py:45 INFO TRAIN_STEP step=11309 loss=1.180007 Training: 63%|██████▎ | 11309/18018 [24:33:30<13:21:57, 7.17s/step, loss=1.180, acc=0.721, acc_len=1.75, thru=18.5, I=19.6, T=33.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:45,132] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:49,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:03:50,550] loop.py:380 INFO TIMING step=11310: step=3.396s data=1.467s compute=3.364s [fwd=2.406s bwd=0.944s opt=0.003s] dispatch=2.460s [2026-10-19 14:03:50,550] loop.py:45 INFO TRAIN_STEP step=11310 loss=1.217794 Training: 63%|██████▎ | 11310/18018 [24:33:36<12:38:02, 6.78s/step, loss=1.218, acc=0.713, acc_len=1.73, thru=20.3, I=26.6, T=37.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:03:50,585] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:00,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:04:01,765] loop.py:45 INFO TRAIN_STEP step=11311 loss=1.304518 Training: 63%|██████▎ | 11311/18018 [24:33:47<15:06:39, 8.11s/step, loss=1.305, acc=0.698, acc_len=1.67, thru=11.9, I=13.0, T=30.3, wait=7.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:04:01,760] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11310: forward=3033.5ms backward=1116.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:02,272] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:07,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:04:08,604] loop.py:45 INFO TRAIN_STEP step=11312 loss=1.204253 Training: 63%|██████▎ | 11312/18018 [24:33:54<14:23:47, 7.73s/step, loss=1.204, acc=0.715, acc_len=1.74, thru=18.1, I=21.6, T=29.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:09,020] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:14,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:04:15,973] loop.py:45 INFO TRAIN_STEP step=11313 loss=1.192885 Training: 63%|██████▎ | 11313/18018 [24:34:02<14:11:34, 7.62s/step, loss=1.193, acc=0.719, acc_len=1.75, thru=15.3, I=22.4, T=29.5, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:16,129] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:04:21,890] loop.py:45 INFO TRAIN_STEP step=11314 loss=1.404464 Training: 63%|██████▎ | 11314/18018 [24:34:07<13:14:22, 7.11s/step, loss=1.404, acc=0.679, acc_len=1.61, thru=18.1, I=20.4, T=35.2, wait=2.3s, pool=53, epoch=2/2][2026-10-19 14:04:28,065] loop.py:380 INFO TIMING step=11315: step=3.850s data=1.690s compute=3.820s [fwd=2.716s bwd=1.089s opt=0.003s] dispatch=2.317s [2026-10-19 14:04:28,065] loop.py:45 INFO TRAIN_STEP step=11315 loss=1.321870 Training: 63%|██████▎ | 11315/18018 [24:34:14<12:42:57, 6.83s/step, loss=1.322, acc=0.698, acc_len=1.68, thru=19.3, I=22.8, T=33.2, wait=2.3s, pool=53, epoch=2/2][2026-10-19 14:04:35,661] loop.py:45 INFO TRAIN_STEP step=11316 loss=1.265980 Training: 63%|██████▎ | 11316/18018 [24:34:21<13:08:32, 7.06s/step, loss=1.266, acc=0.710, acc_len=1.71, thru=15.4, I=16.7, T=35.1, wait=3.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:04:35,656] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11315: forward=2600.8ms backward=1008.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:42,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:04:43,447] loop.py:45 INFO TRAIN_STEP step=11317 loss=1.268828 Training: 63%|██████▎ | 11317/18018 [24:34:29<13:32:45, 7.28s/step, loss=1.269, acc=0.706, acc_len=1.71, thru=16.9, I=17.8, T=30.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:43,609] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:49,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:04:50,133] loop.py:45 INFO TRAIN_STEP step=11318 loss=1.261461 Training: 63%|██████▎ | 11318/18018 [24:34:36<13:12:49, 7.10s/step, loss=1.261, acc=0.706, acc_len=1.70, thru=17.9, I=19.3, T=35.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:04:50,603] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:04:56,344] loop.py:45 INFO TRAIN_STEP step=11319 loss=1.220675 Training: 63%|██████▎ | 11319/18018 [24:34:42<12:42:55, 6.83s/step, loss=1.221, acc=0.716, acc_len=1.74, thru=19.0, I=21.2, T=38.9, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:06,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:05:06,546] loop.py:380 INFO TIMING step=11320: step=3.822s data=1.585s compute=3.783s [fwd=2.675s bwd=1.093s opt=0.003s] dispatch=6.371s [2026-10-19 14:05:06,546] loop.py:45 INFO TRAIN_STEP step=11320 loss=1.314342 Training: 63%|██████▎ | 11320/18018 [24:34:52<14:35:42, 7.84s/step, loss=1.314, acc=0.699, acc_len=1.68, thru=13.8, I=14.1, T=33.5, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:06,677] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:12,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:05:12,580] loop.py:45 INFO TRAIN_STEP step=11321 loss=1.360870 Training: 63%|██████▎ | 11321/18018 [24:34:58<13:34:50, 7.30s/step, loss=1.361, acc=0.690, acc_len=1.64, thru=20.2, I=22.2, T=33.2, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:05:12,576] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11320: forward=2684.0ms backward=1113.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:12,975] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:05:20,439] loop.py:45 INFO TRAIN_STEP step=11322 loss=1.325023 Training: 63%|██████▎ | 11322/18018 [24:35:06<13:53:31, 7.47s/step, loss=1.325, acc=0.699, acc_len=1.67, thru=16.0, I=16.4, T=35.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:20,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:20,883] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:26,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:05:27,577] loop.py:45 INFO TRAIN_STEP step=11323 loss=1.276420 Training: 63%|██████▎ | 11323/18018 [24:35:13<13:42:20, 7.37s/step, loss=1.276, acc=0.706, acc_len=1.70, thru=18.8, I=19.6, T=33.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:27,792] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:05:34,291] loop.py:45 INFO TRAIN_STEP step=11324 loss=1.332894 Training: 63%|██████▎ | 11324/18018 [24:35:20<13:20:15, 7.17s/step, loss=1.333, acc=0.693, acc_len=1.66, thru=17.3, I=19.2, T=36.6, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:40,647] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:05:40,884] loop.py:380 INFO TIMING step=11325: step=3.546s data=1.456s compute=3.510s [fwd=2.472s bwd=1.023s opt=0.003s] dispatch=3.037s [2026-10-19 14:05:40,884] loop.py:45 INFO TRAIN_STEP step=11325 loss=1.300168 Training: 63%|██████▎ | 11325/18018 [24:35:26<13:00:40, 7.00s/step, loss=1.300, acc=0.698, acc_len=1.68, thru=19.5, I=19.9, T=36.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:41,151] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:46,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:05:47,031] loop.py:45 INFO TRAIN_STEP step=11326 loss=1.304892 Training: 63%|██████▎ | 11326/18018 [24:35:33<12:32:04, 6.74s/step, loss=1.305, acc=0.699, acc_len=1.68, thru=20.5, I=21.9, T=36.5, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:05:47,026] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11325: forward=2481.4ms backward=970.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:47,446] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:53,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:05:54,336] loop.py:45 INFO TRAIN_STEP step=11327 loss=1.316690 Training: 63%|██████▎ | 11327/18018 [24:35:40<12:50:48, 6.91s/step, loss=1.317, acc=0.702, acc_len=1.69, thru=17.3, I=21.8, T=35.2, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:54,786] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:05:59,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:06:00,067] loop.py:45 INFO TRAIN_STEP step=11328 loss=1.314606 Training: 63%|██████▎ | 11328/18018 [24:35:46<12:11:06, 6.56s/step, loss=1.315, acc=0.698, acc_len=1.68, thru=21.1, I=25.0, T=40.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:06:00,310] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:06:06,719] loop.py:45 INFO TRAIN_STEP step=11329 loss=1.274722 Training: 63%|██████▎ | 11329/18018 [24:35:52<12:14:10, 6.59s/step, loss=1.275, acc=0.706, acc_len=1.69, thru=19.1, I=19.4, T=35.6, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:06:13,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:06:14,002] loop.py:380 INFO TIMING step=11330: step=3.565s data=1.472s compute=3.535s [fwd=2.495s bwd=1.026s opt=0.003s] dispatch=3.710s [2026-10-19 14:06:14,002] loop.py:45 INFO TRAIN_STEP step=11330 loss=1.303226 Training: 63%|██████▎ | 11330/18018 [24:36:00<12:37:29, 6.80s/step, loss=1.303, acc=0.702, acc_len=1.68, thru=18.4, I=19.6, T=35.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:06:14,184] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 14:06:21,592] loop.py:45 INFO TRAIN_STEP step=11331 loss=1.334889 Training: 63%|██████▎ | 11331/18018 [24:36:07<13:03:55, 7.03s/step, loss=1.335, acc=0.692, acc_len=1.65, thru=16.1, I=17.4, T=36.2, wait=4.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:06:21,588] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11330: forward=2516.0ms backward=974.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:06:31,218] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:06:31,960] loop.py:45 INFO TRAIN_STEP step=11332 loss=1.298557 Training: 63%|██████▎ | 11332/18018 [24:36:18<14:55:11, 8.03s/step, loss=1.299, acc=0.700, acc_len=1.68, thru=13.8, I=13.9, T=33.1, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:06:32,223] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:06:37,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:06:38,594] loop.py:45 INFO TRAIN_STEP step=11333 loss=1.286591 Training: 63%|██████▎ | 11333/18018 [24:36:24<14:08:20, 7.61s/step, loss=1.287, acc=0.702, acc_len=1.69, thru=18.1, I=22.6, T=36.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:06:38,765] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:06:45,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:06:45,623] loop.py:45 INFO TRAIN_STEP step=11334 loss=1.242226 Training: 63%|██████▎ | 11334/18018 [24:36:31<13:48:40, 7.44s/step, loss=1.242, acc=0.714, acc_len=1.73, thru=18.0, I=20.9, T=36.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:06:46,132] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:06:52,971] loop.py:380 INFO TIMING step=11335: step=3.621s data=1.482s compute=3.593s [fwd=2.547s bwd=1.031s opt=0.003s] dispatch=3.717s [2026-10-19 14:06:52,971] loop.py:45 INFO TRAIN_STEP step=11335 loss=1.294270 Training: 63%|██████▎ | 11335/18018 [24:36:39<13:45:39, 7.41s/step, loss=1.294, acc=0.705, acc_len=1.70, thru=14.6, I=20.4, T=35.4, wait=3.7s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:04,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:07:04,198] loop.py:45 INFO TRAIN_STEP step=11336 loss=1.190496 Training: 63%|██████▎ | 11336/18018 [24:36:50<15:52:45, 8.56s/step, loss=1.190, acc=0.719, acc_len=1.75, thru=16.3, I=13.3, T=33.7, wait=7.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:07:04,193] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11335: forward=2710.3ms backward=1036.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:04,516] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:14,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:07:15,375] loop.py:45 INFO TRAIN_STEP step=11337 loss=1.300373 Training: 63%|██████▎ | 11337/18018 [24:37:01<17:20:16, 9.34s/step, loss=1.300, acc=0.703, acc_len=1.69, thru=13.6, I=11.3, T=33.5, wait=7.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:15,665] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:07:22,403] loop.py:45 INFO TRAIN_STEP step=11338 loss=1.224137 Training: 63%|██████▎ | 11338/18018 [24:37:08<16:02:43, 8.65s/step, loss=1.224, acc=0.709, acc_len=1.71, thru=17.7, I=18.4, T=37.0, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:30,398] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:07:32,219] loop.py:45 INFO TRAIN_STEP step=11339 loss=1.436689 Training: 63%|██████▎ | 11339/18018 [24:37:18<16:41:35, 9.00s/step, loss=1.437, acc=0.674, acc_len=1.58, thru=14.1, I=17.1, T=30.7, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:32,407] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:38,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:07:38,756] loop.py:380 INFO TIMING step=11340: step=3.568s data=1.524s compute=3.530s [fwd=2.513s bwd=1.003s opt=0.003s] dispatch=2.962s [2026-10-19 14:07:38,756] loop.py:45 INFO TRAIN_STEP step=11340 loss=1.286450 Training: 63%|██████▎ | 11340/18018 [24:37:24<15:19:21, 8.26s/step, loss=1.286, acc=0.703, acc_len=1.70, thru=18.2, I=22.5, T=35.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:39,163] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:43,930] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:07:45,796] loop.py:45 INFO TRAIN_STEP step=11341 loss=1.243115 Training: 63%|██████▎ | 11341/18018 [24:37:31<14:38:26, 7.89s/step, loss=1.243, acc=0.712, acc_len=1.72, thru=18.4, I=23.3, T=33.1, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:07:45,792] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11340: forward=2727.5ms backward=1101.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:07:45,964] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 [2026-10-19 14:07:52,003] loop.py:45 INFO TRAIN_STEP step=11342 loss=1.234397 Training: 63%|██████▎ | 11342/18018 [24:37:38<13:41:59, 7.39s/step, loss=1.234, acc=0.714, acc_len=1.72, thru=17.2, I=19.9, T=39.1, wait=2.9s, pool=45, epoch=2/2][2026-10-19 14:08:00,333] loop.py:45 INFO TRAIN_STEP step=11343 loss=1.281680 Training: 63%|██████▎ | 11343/18018 [24:37:46<14:13:22, 7.67s/step, loss=1.282, acc=0.701, acc_len=1.68, thru=17.6, I=16.4, T=38.3, wait=5.0s, pool=53, epoch=2/2][2026-10-19 14:08:11,768] loop.py:45 INFO TRAIN_STEP step=11344 loss=1.358501 Training: 63%|██████▎ | 11344/18018 [24:37:57<16:18:50, 8.80s/step, loss=1.359, acc=0.688, acc_len=1.64, thru=9.7, I=10.1, T=28.7, wait=7.0s, pool=49, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:17,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:08:19,101] loop.py:380 INFO TIMING step=11345: step=4.220s data=2.023s compute=4.194s [fwd=3.152s bwd=1.027s opt=0.003s] dispatch=3.104s [2026-10-19 14:08:19,101] loop.py:45 INFO TRAIN_STEP step=11345 loss=1.345160 Training: 63%|██████▎ | 11345/18018 [24:38:05<15:29:44, 8.36s/step, loss=1.345, acc=0.690, acc_len=1.65, thru=18.0, I=19.1, T=30.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:19,231] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 [2026-10-19 14:08:25,058] loop.py:45 INFO TRAIN_STEP step=11346 loss=1.324666 Training: 63%|██████▎ | 11346/18018 [24:38:11<14:09:24, 7.64s/step, loss=1.325, acc=0.697, acc_len=1.67, thru=18.8, I=22.4, T=34.2, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:08:25,052] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11345: forward=2529.4ms backward=1156.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:24,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:25,490] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:30,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:08:31,308] loop.py:45 INFO TRAIN_STEP step=11347 loss=1.293093 Training: 63%|██████▎ | 11347/18018 [24:38:17<13:23:03, 7.22s/step, loss=1.293, acc=0.699, acc_len=1.68, thru=20.0, I=21.5, T=36.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:31,513] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:38,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:08:38,480] loop.py:45 INFO TRAIN_STEP step=11348 loss=1.267300 Training: 63%|██████▎ | 11348/18018 [24:38:24<13:21:15, 7.21s/step, loss=1.267, acc=0.711, acc_len=1.72, thru=17.7, I=20.2, T=31.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:38,523] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 14:08:44,334] loop.py:45 INFO TRAIN_STEP step=11349 loss=1.206530 Training: 63%|██████▎ | 11349/18018 [24:38:30<12:35:56, 6.80s/step, loss=1.207, acc=0.714, acc_len=1.73, thru=18.2, I=18.5, T=37.2, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:51,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:08:52,994] loop.py:380 INFO TIMING step=11350: step=3.706s data=1.554s compute=3.669s [fwd=2.637s bwd=1.017s opt=0.003s] dispatch=4.946s [2026-10-19 14:08:52,995] loop.py:45 INFO TRAIN_STEP step=11350 loss=1.202499 Training: 63%|██████▎ | 11350/18018 [24:38:39<13:37:46, 7.36s/step, loss=1.202, acc=0.718, acc_len=1.75, thru=17.1, I=17.5, T=34.5, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:53,331] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:58,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:08:59,227] loop.py:45 INFO TRAIN_STEP step=11351 loss=1.271524 Training: 63%|██████▎ | 11351/18018 [24:38:45<13:00:09, 7.02s/step, loss=1.272, acc=0.707, acc_len=1.70, thru=19.7, I=22.9, T=36.7, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:08:59,222] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11350: forward=2498.5ms backward=948.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:08:59,464] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:09:07,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:09:07,308] loop.py:45 INFO TRAIN_STEP step=11352 loss=1.349071 Training: 63%|██████▎ | 11352/18018 [24:38:53<13:35:20, 7.34s/step, loss=1.349, acc=0.687, acc_len=1.63, thru=16.4, I=16.0, T=35.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:09:07,770] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:09:16,899] loop.py:45 INFO TRAIN_STEP step=11353 loss=1.441360 Training: 63%|██████▎ | 11353/18018 [24:39:02<14:50:16, 8.01s/step, loss=1.441, acc=0.675, acc_len=1.60, thru=12.8, I=12.7, T=34.5, wait=5.9s, pool=53, epoch=2/2][2026-10-19 14:09:24,014] loop.py:45 INFO TRAIN_STEP step=11354 loss=1.228820 Training: 63%|██████▎ | 11354/18018 [24:39:10<14:20:10, 7.74s/step, loss=1.229, acc=0.714, acc_len=1.74, thru=16.7, I=15.9, T=34.7, wait=3.4s, pool=41, epoch=2/2][2026-10-19 14:09:32,664] loop.py:380 INFO TIMING step=11355: step=3.748s data=1.590s compute=3.709s [fwd=2.633s bwd=1.061s opt=0.003s] dispatch=4.895s [2026-10-19 14:09:32,664] loop.py:45 INFO TRAIN_STEP step=11355 loss=1.357704 Training: 63%|██████▎ | 11355/18018 [24:39:18<14:50:13, 8.02s/step, loss=1.358, acc=0.689, acc_len=1.65, thru=14.4, I=13.0, T=34.1, wait=4.9s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:09:42,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:09:43,500] loop.py:45 INFO TRAIN_STEP step=11356 loss=1.322479 Training: 63%|██████▎ | 11356/18018 [24:39:29<16:24:03, 8.86s/step, loss=1.322, acc=0.700, acc_len=1.67, thru=15.2, I=16.0, T=30.5, wait=6.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:09:43,496] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11355: forward=3010.2ms backward=1143.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:09:43,534] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 14:09:49,615] loop.py:45 INFO TRAIN_STEP step=11357 loss=1.424719 Training: 63%|██████▎ | 11357/18018 [24:39:35<14:52:25, 8.04s/step, loss=1.425, acc=0.673, acc_len=1.59, thru=17.5, I=19.2, T=36.1, wait=2.6s, pool=53, epoch=2/2][2026-10-19 14:09:57,979] loop.py:45 INFO TRAIN_STEP step=11358 loss=1.236280 Training: 63%|██████▎ | 11358/18018 [24:39:44<15:03:09, 8.14s/step, loss=1.236, acc=0.712, acc_len=1.72, thru=14.8, I=16.7, T=35.5, wait=4.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:05,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:10:06,283] loop.py:45 INFO TRAIN_STEP step=11359 loss=1.452723 Training: 63%|██████▎ | 11359/18018 [24:39:52<15:08:29, 8.19s/step, loss=1.453, acc=0.677, acc_len=1.60, thru=15.7, I=17.5, T=35.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:06,597] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:11,637] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:10:13,083] loop.py:380 INFO TIMING step=11360: step=3.815s data=1.684s compute=3.785s [fwd=2.630s bwd=1.141s opt=0.003s] dispatch=2.977s [2026-10-19 14:10:13,084] loop.py:45 INFO TRAIN_STEP step=11360 loss=1.242905 Training: 63%|██████▎ | 11360/18018 [24:39:59<14:22:16, 7.77s/step, loss=1.243, acc=0.708, acc_len=1.70, thru=17.7, I=21.4, T=33.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:13,152] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:18,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 14:10:20,226] loop.py:45 INFO TRAIN_STEP step=11361 loss=1.253935 Training: 63%|██████▎ | 11361/18018 [24:40:06<14:01:15, 7.58s/step, loss=1.254, acc=0.706, acc_len=1.70, thru=16.9, I=22.9, T=32.7, wait=3.2s, pool=69, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:10:20,221] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11360: forward=2698.2ms backward=1160.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:20,304] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=45/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:28,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:10:29,089] loop.py:45 INFO TRAIN_STEP step=11362 loss=1.317947 Training: 63%|██████▎ | 11362/18018 [24:40:15<14:43:44, 7.97s/step, loss=1.318, acc=0.700, acc_len=1.69, thru=14.2, I=14.3, T=29.8, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:29,146] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:36,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:10:37,325] loop.py:45 INFO TRAIN_STEP step=11363 loss=1.282448 Training: 63%|██████▎ | 11363/18018 [24:40:23<14:52:32, 8.05s/step, loss=1.282, acc=0.704, acc_len=1.69, thru=15.1, I=14.9, T=34.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:37,429] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:43,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:10:45,641] loop.py:45 INFO TRAIN_STEP step=11364 loss=1.215206 Training: 63%|██████▎ | 11364/18018 [24:40:31<15:01:25, 8.13s/step, loss=1.215, acc=0.717, acc_len=1.74, thru=16.9, I=22.3, T=26.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:45,823] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 14:10:52,231] loop.py:380 INFO TIMING step=11365: step=3.292s data=1.346s compute=3.269s [fwd=2.351s bwd=0.904s opt=0.003s] dispatch=3.286s [2026-10-19 14:10:52,231] loop.py:45 INFO TRAIN_STEP step=11365 loss=1.365734 Training: 63%|██████▎ | 11365/18018 [24:40:38<14:10:02, 7.67s/step, loss=1.366, acc=0.689, acc_len=1.64, thru=14.1, I=21.2, T=38.9, wait=3.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:10:59,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:11:00,614] loop.py:45 INFO TRAIN_STEP step=11366 loss=1.327223 Training: 63%|██████▎ | 11366/18018 [24:40:46<14:33:45, 7.88s/step, loss=1.327, acc=0.696, acc_len=1.66, thru=17.1, I=19.0, T=34.7, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:11:00,609] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11365: forward=2575.1ms backward=1060.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:11:01,042] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:11:07,183] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:11:07,306] loop.py:45 INFO TRAIN_STEP step=11367 loss=1.428043 Training: 63%|██████▎ | 11367/18018 [24:40:53<13:54:03, 7.52s/step, loss=1.428, acc=0.678, acc_len=1.61, thru=18.7, I=21.8, T=37.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:11:07,686] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:11:13,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:11:14,553] loop.py:45 INFO TRAIN_STEP step=11368 loss=1.262865 Training: 63%|██████▎ | 11368/18018 [24:41:00<13:44:48, 7.44s/step, loss=1.263, acc=0.708, acc_len=1.70, thru=17.6, I=20.2, T=35.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:11:14,586] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 14:11:21,856] loop.py:45 INFO TRAIN_STEP step=11369 loss=1.174063 Training: 63%|██████▎ | 11369/18018 [24:41:07<13:40:08, 7.40s/step, loss=1.174, acc=0.725, acc_len=1.77, thru=15.2, I=20.2, T=35.0, wait=3.6s, pool=45, epoch=2/2][2026-10-19 14:11:31,882] loop.py:380 INFO TIMING step=11370: step=4.714s data=2.173s compute=4.679s [fwd=3.258s bwd=1.406s opt=0.003s] dispatch=5.300s [2026-10-19 14:11:31,883] loop.py:45 INFO TRAIN_STEP step=11370 loss=1.311763 Training: 63%|██████▎ | 11370/18018 [24:41:17<15:07:08, 8.19s/step, loss=1.312, acc=0.701, acc_len=1.68, thru=13.8, I=12.8, T=27.2, wait=5.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:11:38,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:11:39,807] loop.py:45 INFO TRAIN_STEP step=11371 loss=1.420294 Training: 63%|██████▎ | 11371/18018 [24:41:25<14:58:14, 8.11s/step, loss=1.420, acc=0.677, acc_len=1.60, thru=16.5, I=18.5, T=34.6, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:11:39,802] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11370: forward=2631.5ms backward=1004.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:11:40,229] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:11:46,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:11:47,364] loop.py:45 INFO TRAIN_STEP step=11372 loss=1.253477 Training: 63%|██████▎ | 11372/18018 [24:41:33<14:39:54, 7.94s/step, loss=1.253, acc=0.705, acc_len=1.69, thru=15.5, I=19.4, T=34.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:11:47,746] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:11:53,896] loop.py:45 INFO TRAIN_STEP step=11373 loss=1.299474 Training: 63%|██████▎ | 11373/18018 [24:41:39<13:52:45, 7.52s/step, loss=1.299, acc=0.701, acc_len=1.68, thru=17.3, I=19.8, T=38.1, wait=3.2s, pool=49, epoch=2/2][2026-10-19 14:12:00,672] loop.py:45 INFO TRAIN_STEP step=11374 loss=1.361477 Training: 63%|██████▎ | 11374/18018 [24:41:46<13:28:03, 7.30s/step, loss=1.361, acc=0.688, acc_len=1.64, thru=18.3, I=19.3, T=35.3, wait=3.1s, pool=49, epoch=2/2][2026-10-19 14:12:07,955] loop.py:380 INFO TIMING step=11375: step=3.687s data=1.513s compute=3.649s [fwd=2.538s bwd=1.095s opt=0.003s] dispatch=3.585s [2026-10-19 14:12:07,956] loop.py:45 INFO TRAIN_STEP step=11375 loss=1.317554 Training: 63%|██████▎ | 11375/18018 [24:41:54<13:27:26, 7.29s/step, loss=1.318, acc=0.698, acc_len=1.68, thru=16.4, I=17.4, T=34.7, wait=3.6s, pool=53, epoch=2/2][2026-10-19 14:12:16,544] loop.py:45 INFO TRAIN_STEP step=11376 loss=1.357585 Training: 63%|██████▎ | 11376/18018 [24:42:02<14:10:17, 7.68s/step, loss=1.358, acc=0.691, acc_len=1.65, thru=14.0, I=15.1, T=30.2, wait=4.3s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:12:16,539] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11375: forward=2989.7ms backward=1180.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:12:25,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:12:26,310] loop.py:45 INFO TRAIN_STEP step=11377 loss=1.307915 Training: 63%|██████▎ | 11377/18018 [24:42:12<15:19:25, 8.31s/step, loss=1.308, acc=0.698, acc_len=1.67, thru=15.2, I=15.1, T=32.1, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:12:26,599] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:12:32,642] loop.py:45 INFO TRAIN_STEP step=11378 loss=1.269344 Training: 63%|██████▎ | 11378/18018 [24:42:18<14:13:46, 7.71s/step, loss=1.269, acc=0.705, acc_len=1.69, thru=16.5, I=19.8, T=37.3, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:12:41,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:12:43,149] loop.py:45 INFO TRAIN_STEP step=11379 loss=1.373644 Training: 63%|██████▎ | 11379/18018 [24:42:29<15:46:20, 8.55s/step, loss=1.374, acc=0.690, acc_len=1.65, thru=15.1, I=15.3, T=29.8, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:12:43,297] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:12:47,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:12:49,270] loop.py:380 INFO TIMING step=11380: step=3.807s data=1.774s compute=3.758s [fwd=2.683s bwd=1.060s opt=0.003s] dispatch=2.302s [2026-10-19 14:12:49,271] loop.py:45 INFO TRAIN_STEP step=11380 loss=1.352054 Training: 63%|██████▎ | 11380/18018 [24:42:35<14:25:31, 7.82s/step, loss=1.352, acc=0.691, acc_len=1.65, thru=17.7, I=25.3, T=33.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:12:49,732] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:12:56,543] loop.py:45 INFO TRAIN_STEP step=11381 loss=1.330856 Training: 63%|██████▎ | 11381/18018 [24:42:42<14:07:03, 7.66s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=14.5, I=18.8, T=35.9, wait=3.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:12:56,539] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11380: forward=2525.0ms backward=986.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:13:05,705] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:13:06,459] loop.py:45 INFO TRAIN_STEP step=11382 loss=1.345056 Training: 63%|██████▎ | 11382/18018 [24:42:52<15:21:53, 8.34s/step, loss=1.345, acc=0.694, acc_len=1.65, thru=14.8, I=14.9, T=33.9, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:13:06,712] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:13:12,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:13:12,553] loop.py:45 INFO TRAIN_STEP step=11383 loss=1.319032 Training: 63%|██████▎ | 11383/18018 [24:42:58<14:07:21, 7.66s/step, loss=1.319, acc=0.695, acc_len=1.66, thru=20.0, I=20.1, T=34.2, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:13:12,713] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:13:17,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:13:18,607] loop.py:45 INFO TRAIN_STEP step=11384 loss=1.241001 Training: 63%|██████▎ | 11384/18018 [24:43:04<13:13:55, 7.18s/step, loss=1.241, acc=0.710, acc_len=1.72, thru=20.1, I=22.2, T=37.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:13:18,898] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:13:28,201] loop.py:380 INFO TIMING step=11385: step=4.024s data=1.567s compute=3.999s [fwd=2.947s bwd=1.036s opt=0.003s] dispatch=5.559s [2026-10-19 14:13:28,202] loop.py:45 INFO TRAIN_STEP step=11385 loss=1.223340 Training: 63%|██████▎ | 11385/18018 [24:43:14<14:33:49, 7.90s/step, loss=1.223, acc=0.713, acc_len=1.72, thru=12.0, I=13.5, T=31.8, wait=5.6s, pool=49, epoch=2/2][2026-10-19 14:13:34,945] loop.py:45 INFO TRAIN_STEP step=11386 loss=1.323802 Training: 63%|██████▎ | 11386/18018 [24:43:21<13:55:14, 7.56s/step, loss=1.324, acc=0.696, acc_len=1.67, thru=17.3, I=19.9, T=35.4, wait=3.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:13:34,941] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11385: forward=2562.1ms backward=1010.7ms [2026-10-19 14:13:42,170] loop.py:45 INFO TRAIN_STEP step=11387 loss=1.343958 Training: 63%|██████▎ | 11387/18018 [24:43:28<13:44:01, 7.46s/step, loss=1.344, acc=0.691, acc_len=1.65, thru=14.8, I=17.3, T=30.8, wait=3.1s, pool=37, epoch=2/2][2026-10-19 14:13:51,739] loop.py:45 INFO TRAIN_STEP step=11388 loss=1.308533 Training: 63%|██████▎ | 11388/18018 [24:43:37<14:53:56, 8.09s/step, loss=1.309, acc=0.695, acc_len=1.66, thru=16.2, I=14.5, T=34.8, wait=5.9s, pool=61, epoch=2/2][2026-10-19 14:14:00,126] loop.py:45 INFO TRAIN_STEP step=11389 loss=1.259925 Training: 63%|██████▎ | 11389/18018 [24:43:46<15:03:40, 8.18s/step, loss=1.260, acc=0.711, acc_len=1.71, thru=12.9, I=14.8, T=35.1, wait=4.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:09,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:14:09,638] loop.py:380 INFO TIMING step=11390: step=4.030s data=1.798s compute=3.999s [fwd=2.824s bwd=1.161s opt=0.003s] dispatch=5.474s [2026-10-19 14:14:09,638] loop.py:45 INFO TRAIN_STEP step=11390 loss=1.344072 Training: 63%|██████▎ | 11390/18018 [24:43:55<15:47:46, 8.58s/step, loss=1.344, acc=0.691, acc_len=1.65, thru=16.0, I=15.5, T=31.8, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:09,929] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:15,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:14:16,274] loop.py:45 INFO TRAIN_STEP step=11391 loss=1.433099 Training: 63%|██████▎ | 11391/18018 [24:44:02<14:43:09, 8.00s/step, loss=1.433, acc=0.679, acc_len=1.61, thru=18.1, I=20.2, T=34.5, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:14:16,270] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11390: forward=2667.0ms backward=1007.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:16,709] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:14:22,341] loop.py:45 INFO TRAIN_STEP step=11392 loss=1.372406 Training: 63%|██████▎ | 11392/18018 [24:44:08<13:39:06, 7.42s/step, loss=1.372, acc=0.689, acc_len=1.64, thru=18.7, I=22.1, T=36.7, wait=2.6s, pool=49, epoch=2/2][2026-10-19 14:14:30,318] loop.py:45 INFO TRAIN_STEP step=11393 loss=1.204894 Training: 63%|██████▎ | 11393/18018 [24:44:16<13:57:36, 7.59s/step, loss=1.205, acc=0.714, acc_len=1.73, thru=13.5, I=16.2, T=29.6, wait=3.6s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:39,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:14:40,918] loop.py:45 INFO TRAIN_STEP step=11394 loss=1.305485 Training: 63%|██████▎ | 11394/18018 [24:44:27<15:37:13, 8.49s/step, loss=1.305, acc=0.697, acc_len=1.68, thru=16.6, I=16.0, T=32.4, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:41,305] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:47,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:14:47,863] loop.py:380 INFO TIMING step=11395: step=3.612s data=1.612s compute=3.587s [fwd=2.606s bwd=0.968s opt=0.003s] dispatch=3.324s [2026-10-19 14:14:47,863] loop.py:45 INFO TRAIN_STEP step=11395 loss=1.222194 Training: 63%|██████▎ | 11395/18018 [24:44:33<14:45:56, 8.03s/step, loss=1.222, acc=0.712, acc_len=1.72, thru=17.6, I=19.9, T=35.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:48,099] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:52,517] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:14:53,584] loop.py:45 INFO TRAIN_STEP step=11396 loss=1.285123 Training: 63%|██████▎ | 11396/18018 [24:44:39<13:29:28, 7.33s/step, loss=1.285, acc=0.706, acc_len=1.70, thru=21.2, I=24.0, T=36.1, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:14:53,579] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11395: forward=2544.2ms backward=959.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:54,024] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:58,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:14:59,788] loop.py:45 INFO TRAIN_STEP step=11397 loss=1.332175 Training: 63%|██████▎ | 11397/18018 [24:44:45<12:51:59, 7.00s/step, loss=1.332, acc=0.692, acc_len=1.65, thru=20.8, I=26.7, T=37.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:14:59,891] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:08,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:15:10,183] loop.py:45 INFO TRAIN_STEP step=11398 loss=1.373101 Training: 63%|██████▎ | 11398/18018 [24:44:56<14:44:22, 8.02s/step, loss=1.373, acc=0.690, acc_len=1.64, thru=12.2, I=16.1, T=26.9, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:10,625] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:15,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:15:16,367] loop.py:45 INFO TRAIN_STEP step=11399 loss=1.191130 Training: 63%|██████▎ | 11399/18018 [24:45:02<13:43:39, 7.47s/step, loss=1.191, acc=0.718, acc_len=1.74, thru=17.8, I=24.9, T=34.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:16,561] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:25,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:15:25,861] loop.py:380 INFO TIMING step=11400: step=3.940s data=1.809s compute=3.896s [fwd=2.789s bwd=1.093s opt=0.003s] dispatch=5.544s [2026-10-19 14:15:25,861] loop.py:45 INFO TRAIN_STEP step=11400 loss=1.370679 Training: 63%|██████▎ | 11400/18018 [24:45:11<14:50:36, 8.07s/step, loss=1.371, acc=0.686, acc_len=1.63, thru=13.8, I=16.0, T=32.5, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:26,194] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:31,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:15:33,582] loop.py:45 INFO TRAIN_STEP step=11401 loss=1.185269 Training: 63%|██████▎ | 11401/18018 [24:45:19<14:38:49, 7.97s/step, loss=1.185, acc=0.721, acc_len=1.75, thru=15.3, I=18.1, T=29.1, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:15:33,578] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11400: forward=3355.3ms backward=993.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:33,790] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:40,237] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:15:40,920] loop.py:45 INFO TRAIN_STEP step=11402 loss=1.420033 Training: 63%|██████▎ | 11402/18018 [24:45:27<14:17:47, 7.78s/step, loss=1.420, acc=0.681, acc_len=1.62, thru=14.9, I=19.5, T=36.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:41,245] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:48,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:15:49,199] loop.py:45 INFO TRAIN_STEP step=11403 loss=1.413965 Training: 63%|██████▎ | 11403/18018 [24:45:35<14:34:08, 7.93s/step, loss=1.414, acc=0.682, acc_len=1.62, thru=15.8, I=19.7, T=32.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:49,654] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:54,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:15:55,684] loop.py:45 INFO TRAIN_STEP step=11404 loss=1.169852 Training: 63%|██████▎ | 11404/18018 [24:45:41<13:46:19, 7.50s/step, loss=1.170, acc=0.722, acc_len=1.76, thru=18.9, I=24.5, T=35.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:15:56,002] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:16:02,773] loop.py:380 INFO TIMING step=11405: step=3.481s data=1.475s compute=3.443s [fwd=2.476s bwd=0.953s opt=0.003s] dispatch=3.598s [2026-10-19 14:16:02,773] loop.py:45 INFO TRAIN_STEP step=11405 loss=1.347984 Training: 63%|██████▎ | 11405/18018 [24:45:48<13:32:42, 7.37s/step, loss=1.348, acc=0.692, acc_len=1.66, thru=15.9, I=19.7, T=36.8, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:13,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:16:13,334] loop.py:45 INFO TRAIN_STEP step=11406 loss=1.200445 Training: 63%|██████▎ | 11406/18018 [24:45:59<15:17:57, 8.33s/step, loss=1.200, acc=0.721, acc_len=1.75, thru=14.2, I=13.7, T=31.8, wait=6.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:16:13,329] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11405: forward=2812.4ms backward=1144.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:13,625] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:18,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:16:20,222] loop.py:45 INFO TRAIN_STEP step=11407 loss=1.191587 Training: 63%|██████▎ | 11407/18018 [24:46:06<14:30:04, 7.90s/step, loss=1.192, acc=0.721, acc_len=1.76, thru=19.1, I=18.2, T=33.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:20,492] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:25,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:16:25,756] loop.py:45 INFO TRAIN_STEP step=11408 loss=1.229373 Training: 63%|██████▎ | 11408/18018 [24:46:11<13:11:53, 7.19s/step, loss=1.229, acc=0.714, acc_len=1.73, thru=21.4, I=25.1, T=35.6, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:26,204] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:33,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:16:33,253] loop.py:45 INFO TRAIN_STEP step=11409 loss=1.256368 Training: 63%|██████▎ | 11409/18018 [24:46:19<13:21:59, 7.28s/step, loss=1.256, acc=0.707, acc_len=1.71, thru=17.3, I=17.9, T=33.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:33,590] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:40,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:16:41,072] loop.py:380 INFO TIMING step=11410: step=3.754s data=1.574s compute=3.719s [fwd=2.672s bwd=1.032s opt=0.003s] dispatch=4.056s [2026-10-19 14:16:41,072] loop.py:45 INFO TRAIN_STEP step=11410 loss=1.272466 Training: 63%|██████▎ | 11410/18018 [24:46:27<13:39:42, 7.44s/step, loss=1.272, acc=0.705, acc_len=1.69, thru=16.8, I=17.8, T=34.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:41,308] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:50,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:16:51,096] loop.py:45 INFO TRAIN_STEP step=11411 loss=1.291620 Training: 63%|██████▎ | 11411/18018 [24:46:37<15:04:49, 8.22s/step, loss=1.292, acc=0.705, acc_len=1.70, thru=13.1, I=13.5, T=32.2, wait=6.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:16:51,091] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11410: forward=2855.4ms backward=1060.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:51,402] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:57,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:16:58,191] loop.py:45 INFO TRAIN_STEP step=11412 loss=1.241587 Training: 63%|██████▎ | 11412/18018 [24:46:44<14:27:36, 7.88s/step, loss=1.242, acc=0.711, acc_len=1.72, thru=17.8, I=18.4, T=38.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:16:58,404] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:05,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:17:05,872] loop.py:45 INFO TRAIN_STEP step=11413 loss=1.228942 Training: 63%|██████▎ | 11413/18018 [24:46:51<14:20:56, 7.82s/step, loss=1.229, acc=0.715, acc_len=1.73, thru=16.7, I=17.7, T=36.0, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:06,127] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:11,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:17:12,895] loop.py:45 INFO TRAIN_STEP step=11414 loss=1.319548 Training: 63%|██████▎ | 11414/18018 [24:46:58<13:54:25, 7.58s/step, loss=1.320, acc=0.696, acc_len=1.67, thru=18.8, I=20.4, T=35.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:13,399] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:18,926] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:17:19,705] loop.py:380 INFO TIMING step=11415: step=3.468s data=1.482s compute=3.443s [fwd=2.472s bwd=0.957s opt=0.003s] dispatch=3.334s [2026-10-19 14:17:19,706] loop.py:45 INFO TRAIN_STEP step=11415 loss=1.172228 Training: 63%|██████▎ | 11415/18018 [24:47:05<13:28:54, 7.35s/step, loss=1.172, acc=0.725, acc_len=1.77, thru=19.1, I=21.6, T=36.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:19,932] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:26,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:17:27,659] loop.py:45 INFO TRAIN_STEP step=11416 loss=1.415887 Training: 63%|██████▎ | 11416/18018 [24:47:13<13:48:42, 7.53s/step, loss=1.416, acc=0.680, acc_len=1.61, thru=16.4, I=19.6, T=34.7, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:17:27,654] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11415: forward=2615.9ms backward=1035.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:27,960] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:17:35,616] loop.py:45 INFO TRAIN_STEP step=11417 loss=1.197044 Training: 63%|██████▎ | 11417/18018 [24:47:21<14:02:36, 7.66s/step, loss=1.197, acc=0.721, acc_len=1.75, thru=13.9, I=15.8, T=33.3, wait=4.1s, pool=53, epoch=2/2][2026-10-19 14:17:42,048] loop.py:45 INFO TRAIN_STEP step=11418 loss=1.204253 Training: 63%|██████▎ | 11418/18018 [24:47:28<13:21:58, 7.29s/step, loss=1.204, acc=0.716, acc_len=1.74, thru=17.9, I=20.1, T=36.3, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:50,929] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:17:51,048] loop.py:45 INFO TRAIN_STEP step=11419 loss=1.259611 Training: 63%|██████▎ | 11419/18018 [24:47:37<14:18:16, 7.80s/step, loss=1.260, acc=0.708, acc_len=1.71, thru=14.9, I=13.9, T=35.8, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:17:51,433] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:17:59,431] loop.py:380 INFO TIMING step=11420: step=3.726s data=1.506s compute=3.703s [fwd=2.662s bwd=1.027s opt=0.003s] dispatch=4.648s [2026-10-19 14:17:59,431] loop.py:45 INFO TRAIN_STEP step=11420 loss=1.238783 Training: 63%|██████▎ | 11420/18018 [24:47:45<14:37:19, 7.98s/step, loss=1.239, acc=0.712, acc_len=1.73, thru=16.0, I=17.1, T=34.4, wait=4.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:06,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:18:07,952] loop.py:45 INFO TRAIN_STEP step=11421 loss=1.449324 Training: 63%|██████▎ | 11421/18018 [24:47:54<14:54:56, 8.14s/step, loss=1.449, acc=0.671, acc_len=1.58, thru=15.1, I=16.7, T=33.8, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:18:07,946] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11420: forward=2602.5ms backward=1134.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:08,453] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:15,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:18:15,949] loop.py:45 INFO TRAIN_STEP step=11422 loss=1.359822 Training: 63%|██████▎ | 11422/18018 [24:48:02<14:50:12, 8.10s/step, loss=1.360, acc=0.691, acc_len=1.65, thru=15.3, I=17.8, T=33.6, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:16,255] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:18:24,774] loop.py:45 INFO TRAIN_STEP step=11423 loss=1.303107 Training: 63%|██████▎ | 11423/18018 [24:48:10<15:13:59, 8.32s/step, loss=1.303, acc=0.699, acc_len=1.68, thru=13.6, I=14.9, T=31.9, wait=4.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:31,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:18:32,171] loop.py:45 INFO TRAIN_STEP step=11424 loss=1.339732 Training: 63%|██████▎ | 11424/18018 [24:48:18<14:43:39, 8.04s/step, loss=1.340, acc=0.692, acc_len=1.65, thru=17.3, I=17.8, T=32.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:32,559] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:18:38,667] loop.py:380 INFO TIMING step=11425: step=4.164s data=2.136s compute=4.139s [fwd=3.170s bwd=0.955s opt=0.003s] dispatch=2.322s [2026-10-19 14:18:38,667] loop.py:45 INFO TRAIN_STEP step=11425 loss=1.322459 Training: 63%|██████▎ | 11425/18018 [24:48:24<13:52:32, 7.58s/step, loss=1.322, acc=0.695, acc_len=1.66, thru=15.8, I=19.0, T=30.7, wait=2.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:45,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:18:46,772] loop.py:45 INFO TRAIN_STEP step=11426 loss=1.362417 Training: 63%|██████▎ | 11426/18018 [24:48:32<14:09:52, 7.74s/step, loss=1.362, acc=0.689, acc_len=1.65, thru=18.0, I=17.8, T=36.3, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:18:46,768] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11425: forward=2538.9ms backward=948.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:46,967] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:18:52,708] loop.py:45 INFO TRAIN_STEP step=11427 loss=1.290887 Training: 63%|██████▎ | 11427/18018 [24:48:38<13:10:19, 7.19s/step, loss=1.291, acc=0.696, acc_len=1.67, thru=19.6, I=22.4, T=38.6, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:59,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:18:59,249] loop.py:45 INFO TRAIN_STEP step=11428 loss=1.281227 Training: 63%|██████▎ | 11428/18018 [24:48:45<12:48:42, 7.00s/step, loss=1.281, acc=0.705, acc_len=1.70, thru=20.2, I=20.3, T=37.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:18:59,508] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:05,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:19:06,456] loop.py:45 INFO TRAIN_STEP step=11429 loss=1.266989 Training: 63%|██████▎ | 11429/18018 [24:48:52<12:55:31, 7.06s/step, loss=1.267, acc=0.707, acc_len=1.70, thru=17.9, I=20.2, T=31.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:06,669] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:12,182] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:19:12,941] loop.py:380 INFO TIMING step=11430: step=3.637s data=1.601s compute=3.604s [fwd=2.579s bwd=1.011s opt=0.003s] dispatch=2.838s [2026-10-19 14:19:12,941] loop.py:45 INFO TRAIN_STEP step=11430 loss=1.260868 Training: 63%|██████▎ | 11430/18018 [24:48:59<12:36:20, 6.89s/step, loss=1.261, acc=0.705, acc_len=1.71, thru=18.7, I=20.8, T=35.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:13,187] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:19:19,666] loop.py:45 INFO TRAIN_STEP step=11431 loss=1.334878 Training: 63%|██████▎ | 11431/18018 [24:49:05<12:30:48, 6.84s/step, loss=1.335, acc=0.695, acc_len=1.65, thru=15.6, I=18.9, T=32.6, wait=2.8s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:19:19,661] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11430: forward=2730.4ms backward=1153.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:32,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:19:33,078] loop.py:45 INFO TRAIN_STEP step=11432 loss=1.390179 Training: 63%|██████▎ | 11432/18018 [24:49:19<16:07:11, 8.81s/step, loss=1.390, acc=0.688, acc_len=1.64, thru=13.3, I=11.7, T=30.6, wait=9.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:33,147] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:39,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:19:40,031] loop.py:45 INFO TRAIN_STEP step=11433 loss=1.206143 Training: 63%|██████▎ | 11433/18018 [24:49:26<15:05:55, 8.25s/step, loss=1.206, acc=0.718, acc_len=1.75, thru=17.4, I=19.5, T=33.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:40,361] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:45,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:19:45,866] loop.py:45 INFO TRAIN_STEP step=11434 loss=1.236316 Training: 63%|██████▎ | 11434/18018 [24:49:31<13:46:02, 7.53s/step, loss=1.236, acc=0.713, acc_len=1.72, thru=21.0, I=23.6, T=36.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:45,930] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 14:19:51,549] loop.py:380 INFO TIMING step=11435: step=3.449s data=1.441s compute=3.416s [fwd=2.431s bwd=0.971s opt=0.003s] dispatch=2.225s [2026-10-19 14:19:51,549] loop.py:45 INFO TRAIN_STEP step=11435 loss=1.312538 Training: 63%|██████▎ | 11435/18018 [24:49:37<12:45:15, 6.97s/step, loss=1.313, acc=0.696, acc_len=1.67, thru=21.1, I=23.1, T=37.1, wait=2.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:57,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:19:58,305] loop.py:45 INFO TRAIN_STEP step=11436 loss=1.234586 Training: 63%|██████▎ | 11436/18018 [24:49:44<12:37:56, 6.91s/step, loss=1.235, acc=0.710, acc_len=1.72, thru=19.0, I=19.9, T=35.1, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:19:58,301] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11435: forward=2559.1ms backward=999.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:19:58,490] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:20:05,889] loop.py:45 INFO TRAIN_STEP step=11437 loss=1.454898 Training: 63%|██████▎ | 11437/18018 [24:49:51<13:00:04, 7.11s/step, loss=1.455, acc=0.676, acc_len=1.60, thru=16.8, I=17.9, T=35.1, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:11,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:20:12,531] loop.py:45 INFO TRAIN_STEP step=11438 loss=1.331660 Training: 63%|██████▎ | 11438/18018 [24:49:58<12:44:25, 6.97s/step, loss=1.332, acc=0.694, acc_len=1.66, thru=19.6, I=20.9, T=35.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:12,634] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:16,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:20:18,407] loop.py:45 INFO TRAIN_STEP step=11439 loss=1.415774 Training: 63%|██████▎ | 11439/18018 [24:50:04<12:08:19, 6.64s/step, loss=1.416, acc=0.679, acc_len=1.61, thru=20.9, I=26.3, T=35.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:18,748] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:20:25,790] loop.py:380 INFO TIMING step=11440: step=3.683s data=1.581s compute=3.654s [fwd=2.579s bwd=1.060s opt=0.003s] dispatch=3.692s [2026-10-19 14:20:25,791] loop.py:45 INFO TRAIN_STEP step=11440 loss=1.238218 Training: 63%|██████▎ | 11440/18018 [24:50:11<12:32:32, 6.86s/step, loss=1.238, acc=0.711, acc_len=1.72, thru=13.9, I=17.0, T=34.8, wait=3.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:31,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:20:32,410] loop.py:45 INFO TRAIN_STEP step=11441 loss=1.259730 Training: 63%|██████▎ | 11441/18018 [24:50:18<12:24:23, 6.79s/step, loss=1.260, acc=0.708, acc_len=1.71, thru=19.3, I=20.2, T=32.8, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:20:32,405] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11440: forward=2845.6ms backward=1003.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:32,750] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:39,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:20:40,041] loop.py:45 INFO TRAIN_STEP step=11442 loss=1.325746 Training: 64%|██████▎ | 11442/18018 [24:50:26<12:51:54, 7.04s/step, loss=1.326, acc=0.694, acc_len=1.66, thru=17.1, I=19.6, T=31.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:40,418] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:20:45,748] loop.py:45 INFO TRAIN_STEP step=11443 loss=1.334098 Training: 64%|██████▎ | 11443/18018 [24:50:31<12:07:54, 6.64s/step, loss=1.334, acc=0.694, acc_len=1.66, thru=19.6, I=21.2, T=35.9, wait=2.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:52,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:20:52,433] loop.py:45 INFO TRAIN_STEP step=11444 loss=1.288883 Training: 64%|██████▎ | 11444/18018 [24:50:38<12:09:08, 6.65s/step, loss=1.289, acc=0.706, acc_len=1.70, thru=19.6, I=19.7, T=37.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:20:52,712] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:00,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:21:01,720] loop.py:380 INFO TIMING step=11445: step=4.093s data=1.730s compute=4.057s [fwd=3.008s bwd=1.034s opt=0.003s] dispatch=5.185s [2026-10-19 14:21:01,720] loop.py:45 INFO TRAIN_STEP step=11445 loss=1.266325 Training: 64%|██████▎ | 11445/18018 [24:50:47<13:35:33, 7.44s/step, loss=1.266, acc=0.708, acc_len=1.71, thru=14.0, I=15.8, T=31.3, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:02,166] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:08,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:21:08,858] loop.py:45 INFO TRAIN_STEP step=11446 loss=1.255984 Training: 64%|██████▎ | 11446/18018 [24:50:54<13:25:23, 7.35s/step, loss=1.256, acc=0.705, acc_len=1.70, thru=17.4, I=20.6, T=35.3, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:21:08,854] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11445: forward=2557.6ms backward=1028.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:09,335] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:21:15,786] loop.py:45 INFO TRAIN_STEP step=11447 loss=1.301654 Training: 64%|██████▎ | 11447/18018 [24:51:01<13:11:14, 7.22s/step, loss=1.302, acc=0.698, acc_len=1.67, thru=17.2, I=19.6, T=35.0, wait=3.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:21,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:21:22,894] loop.py:45 INFO TRAIN_STEP step=11448 loss=1.385436 Training: 64%|██████▎ | 11448/18018 [24:51:08<13:07:21, 7.19s/step, loss=1.385, acc=0.683, acc_len=1.63, thru=20.0, I=20.8, T=33.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:23,100] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:29,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:21:29,391] loop.py:45 INFO TRAIN_STEP step=11449 loss=1.157623 Training: 64%|██████▎ | 11449/18018 [24:51:15<12:44:25, 6.98s/step, loss=1.158, acc=0.727, acc_len=1.78, thru=18.9, I=24.2, T=32.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:29,576] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:34,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:21:35,747] loop.py:380 INFO TIMING step=11450: step=3.893s data=1.783s compute=3.856s [fwd=2.741s bwd=1.101s opt=0.003s] dispatch=2.454s [2026-10-19 14:21:35,748] loop.py:45 INFO TRAIN_STEP step=11450 loss=1.265626 Training: 64%|██████▎ | 11450/18018 [24:51:21<12:23:48, 6.79s/step, loss=1.266, acc=0.705, acc_len=1.70, thru=19.5, I=21.5, T=32.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:36,006] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:40,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:21:41,857] loop.py:45 INFO TRAIN_STEP step=11451 loss=1.283300 Training: 64%|██████▎ | 11451/18018 [24:51:27<12:01:10, 6.59s/step, loss=1.283, acc=0.706, acc_len=1.70, thru=19.4, I=26.0, T=35.0, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:21:41,853] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11450: forward=2576.9ms backward=1031.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:42,096] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:21:48,270] loop.py:45 INFO TRAIN_STEP step=11452 loss=1.259347 Training: 64%|██████▎ | 11452/18018 [24:51:34<11:55:18, 6.54s/step, loss=1.259, acc=0.711, acc_len=1.73, thru=16.5, I=23.9, T=34.4, wait=2.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:56,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:21:56,966] loop.py:45 INFO TRAIN_STEP step=11453 loss=1.262419 Training: 64%|██████▎ | 11453/18018 [24:51:43<13:06:02, 7.18s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=16.4, I=15.8, T=34.9, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:21:57,290] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:02,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:22:03,214] loop.py:45 INFO TRAIN_STEP step=11454 loss=1.245570 Training: 64%|██████▎ | 11454/18018 [24:51:49<12:35:11, 6.90s/step, loss=1.246, acc=0.711, acc_len=1.72, thru=19.7, I=21.3, T=35.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:03,393] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:22:10,560] loop.py:380 INFO TIMING step=11455: step=3.876s data=1.748s compute=3.837s [fwd=2.770s bwd=1.052s opt=0.003s] dispatch=3.462s [2026-10-19 14:22:10,560] loop.py:45 INFO TRAIN_STEP step=11455 loss=1.229193 Training: 64%|██████▎ | 11455/18018 [24:51:56<12:49:37, 7.04s/step, loss=1.229, acc=0.714, acc_len=1.73, thru=16.7, I=18.2, T=33.0, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:17,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:22:18,844] loop.py:45 INFO TRAIN_STEP step=11456 loss=1.383660 Training: 64%|██████▎ | 11456/18018 [24:52:04<13:30:30, 7.41s/step, loss=1.384, acc=0.688, acc_len=1.64, thru=15.7, I=17.3, T=32.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:18,878] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 14:22:18,840] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11455: forward=2736.3ms backward=1166.8ms [2026-10-19 14:22:26,113] loop.py:45 INFO TRAIN_STEP step=11457 loss=1.239585 Training: 64%|██████▎ | 11457/18018 [24:52:12<13:25:43, 7.37s/step, loss=1.240, acc=0.715, acc_len=1.74, thru=15.2, I=17.6, T=35.1, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:31,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:22:34,188] loop.py:45 INFO TRAIN_STEP step=11458 loss=1.349851 Training: 64%|██████▎ | 11458/18018 [24:52:20<13:48:45, 7.58s/step, loss=1.350, acc=0.694, acc_len=1.66, thru=16.5, I=20.7, T=28.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:34,366] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:39,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:22:40,336] loop.py:45 INFO TRAIN_STEP step=11459 loss=1.320668 Training: 64%|██████▎ | 11459/18018 [24:52:26<13:01:36, 7.15s/step, loss=1.321, acc=0.696, acc_len=1.66, thru=18.1, I=26.3, T=34.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:40,469] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:46,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:22:46,462] loop.py:380 INFO TIMING step=11460: step=3.385s data=1.381s compute=3.352s [fwd=2.376s bwd=0.962s opt=0.003s] dispatch=2.733s [2026-10-19 14:22:46,462] loop.py:45 INFO TRAIN_STEP step=11460 loss=1.335262 Training: 64%|██████▎ | 11460/18018 [24:52:32<12:27:54, 6.84s/step, loss=1.335, acc=0.697, acc_len=1.67, thru=19.9, I=21.8, T=37.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:46,680] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:52,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:22:53,563] loop.py:45 INFO TRAIN_STEP step=11461 loss=1.305011 Training: 64%|██████▎ | 11461/18018 [24:52:39<12:36:15, 6.92s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=18.6, I=21.1, T=37.4, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:22:53,559] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11460: forward=2414.0ms backward=970.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:22:53,986] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:22:59,954] loop.py:45 INFO TRAIN_STEP step=11462 loss=1.273440 Training: 64%|██████▎ | 11462/18018 [24:52:46<12:18:48, 6.76s/step, loss=1.273, acc=0.706, acc_len=1.70, thru=18.3, I=17.8, T=36.9, wait=2.9s, pool=49, epoch=2/2][2026-10-19 14:23:07,387] loop.py:45 INFO TRAIN_STEP step=11463 loss=1.384873 Training: 64%|██████▎ | 11463/18018 [24:52:53<12:40:46, 6.96s/step, loss=1.385, acc=0.688, acc_len=1.64, thru=19.6, I=16.6, T=33.8, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:14,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:23:15,578] loop.py:45 INFO TRAIN_STEP step=11464 loss=1.258415 Training: 64%|██████▎ | 11464/18018 [24:53:01<13:20:49, 7.33s/step, loss=1.258, acc=0.705, acc_len=1.70, thru=17.6, I=17.3, T=29.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:15,976] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:21,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:23:22,880] loop.py:380 INFO TIMING step=11465: step=4.000s data=1.831s compute=3.977s [fwd=2.806s bwd=1.157s opt=0.003s] dispatch=3.294s [2026-10-19 14:23:22,880] loop.py:45 INFO TRAIN_STEP step=11465 loss=1.387901 Training: 64%|██████▎ | 11465/18018 [24:53:08<13:19:46, 7.32s/step, loss=1.388, acc=0.687, acc_len=1.64, thru=15.4, I=20.0, T=32.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:23,292] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:28,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:23:28,877] loop.py:45 INFO TRAIN_STEP step=11466 loss=1.272562 Training: 64%|██████▎ | 11466/18018 [24:53:14<12:36:12, 6.92s/step, loss=1.273, acc=0.704, acc_len=1.70, thru=19.6, I=24.1, T=37.6, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:23:28,873] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11465: forward=2426.2ms backward=936.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:29,052] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:34,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:23:36,290] loop.py:45 INFO TRAIN_STEP step=11467 loss=1.332708 Training: 64%|██████▎ | 11467/18018 [24:53:22<12:52:06, 7.07s/step, loss=1.333, acc=0.693, acc_len=1.66, thru=17.3, I=20.3, T=32.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:36,693] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:41,272] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:23:42,623] loop.py:45 INFO TRAIN_STEP step=11468 loss=1.325759 Training: 64%|██████▎ | 11468/18018 [24:53:28<12:27:49, 6.85s/step, loss=1.326, acc=0.693, acc_len=1.66, thru=19.2, I=25.5, T=35.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:23:42,782] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:23:49,152] loop.py:45 INFO TRAIN_STEP step=11469 loss=1.287934 Training: 64%|██████▎ | 11469/18018 [24:53:35<12:17:03, 6.75s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=17.6, I=20.7, T=36.0, wait=3.0s, pool=53, epoch=2/2][2026-10-19 14:23:56,711] loop.py:380 INFO TIMING step=11470: step=3.331s data=1.347s compute=3.284s [fwd=2.352s bwd=0.917s opt=0.003s] dispatch=4.222s [2026-10-19 14:23:56,712] loop.py:45 INFO TRAIN_STEP step=11470 loss=1.392675 Training: 64%|██████▎ | 11470/18018 [24:53:42<12:43:22, 6.99s/step, loss=1.393, acc=0.686, acc_len=1.63, thru=16.6, I=15.5, T=38.4, wait=4.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:02,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:24:04,195] loop.py:45 INFO TRAIN_STEP step=11471 loss=1.233695 Training: 64%|██████▎ | 11471/18018 [24:53:50<12:59:17, 7.14s/step, loss=1.234, acc=0.712, acc_len=1.73, thru=19.9, I=21.4, T=35.1, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:24:04,191] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11470: forward=2614.6ms backward=984.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:04,569] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:13,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:24:13,699] loop.py:45 INFO TRAIN_STEP step=11472 loss=1.278797 Training: 64%|██████▎ | 11472/18018 [24:53:59<14:16:31, 7.85s/step, loss=1.279, acc=0.706, acc_len=1.70, thru=14.0, I=13.5, T=32.4, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:13,998] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:18,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:24:19,410] loop.py:45 INFO TRAIN_STEP step=11473 loss=1.343691 Training: 64%|██████▎ | 11473/18018 [24:54:05<13:06:22, 7.21s/step, loss=1.344, acc=0.690, acc_len=1.65, thru=20.2, I=22.2, T=35.0, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:19,649] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:23,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:24:25,151] loop.py:45 INFO TRAIN_STEP step=11474 loss=1.324068 Training: 64%|██████▎ | 11474/18018 [24:54:11<12:18:14, 6.77s/step, loss=1.324, acc=0.695, acc_len=1.67, thru=21.6, I=25.7, T=36.6, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:25,468] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:30,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:24:30,784] loop.py:380 INFO TIMING step=11475: step=3.421s data=1.459s compute=3.394s [fwd=2.462s bwd=0.917s opt=0.003s] dispatch=2.201s [2026-10-19 14:24:30,784] loop.py:45 INFO TRAIN_STEP step=11475 loss=1.315049 Training: 64%|██████▎ | 11475/18018 [24:54:16<11:40:54, 6.43s/step, loss=1.315, acc=0.696, acc_len=1.66, thru=22.5, I=26.4, T=37.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:31,172] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:24:36,940] loop.py:45 INFO TRAIN_STEP step=11476 loss=1.262325 Training: 64%|██████▎ | 11476/18018 [24:54:23<11:31:55, 6.35s/step, loss=1.262, acc=0.706, acc_len=1.70, thru=17.7, I=21.1, T=35.9, wait=2.6s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:24:36,936] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11475: forward=2521.2ms backward=998.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:44,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:24:45,919] loop.py:45 INFO TRAIN_STEP step=11477 loss=1.346593 Training: 64%|██████▎ | 11477/18018 [24:54:32<12:57:59, 7.14s/step, loss=1.347, acc=0.693, acc_len=1.67, thru=16.8, I=17.5, T=35.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:24:46,099] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:24:54,928] loop.py:45 INFO TRAIN_STEP step=11478 loss=1.443486 Training: 64%|██████▎ | 11478/18018 [24:54:41<13:59:06, 7.70s/step, loss=1.443, acc=0.676, acc_len=1.61, thru=11.3, I=12.6, T=34.4, wait=5.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:02,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:25:03,032] loop.py:45 INFO TRAIN_STEP step=11479 loss=1.233747 Training: 64%|██████▎ | 11479/18018 [24:54:49<14:12:14, 7.82s/step, loss=1.234, acc=0.709, acc_len=1.71, thru=17.3, I=17.8, T=36.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:03,217] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:09,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:25:10,171] loop.py:380 INFO TIMING step=11480: step=3.835s data=1.708s compute=3.801s [fwd=2.640s bwd=1.147s opt=0.003s] dispatch=3.294s [2026-10-19 14:25:10,172] loop.py:45 INFO TRAIN_STEP step=11480 loss=1.376638 Training: 64%|██████▎ | 11480/18018 [24:54:56<13:49:51, 7.62s/step, loss=1.377, acc=0.687, acc_len=1.63, thru=18.2, I=20.0, T=33.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:10,555] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:25:19,269] loop.py:45 INFO TRAIN_STEP step=11481 loss=1.359983 Training: 64%|██████▎ | 11481/18018 [24:55:05<14:38:09, 8.06s/step, loss=1.360, acc=0.690, acc_len=1.65, thru=12.5, I=13.9, T=31.6, wait=5.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:25:19,265] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11480: forward=2867.0ms backward=1140.7ms [2026-10-19 14:25:28,206] loop.py:45 INFO TRAIN_STEP step=11482 loss=1.372514 Training: 64%|██████▎ | 11482/18018 [24:55:14<15:06:37, 8.32s/step, loss=1.373, acc=0.688, acc_len=1.64, thru=14.5, I=13.7, T=32.8, wait=5.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:35,799] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:25:36,712] loop.py:45 INFO TRAIN_STEP step=11483 loss=1.276263 Training: 64%|██████▎ | 11483/18018 [24:55:22<15:12:28, 8.38s/step, loss=1.276, acc=0.703, acc_len=1.69, thru=16.9, I=17.6, T=32.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:36,816] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:42,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:25:43,681] loop.py:45 INFO TRAIN_STEP step=11484 loss=1.250722 Training: 64%|██████▎ | 11484/18018 [24:55:29<14:26:18, 7.96s/step, loss=1.251, acc=0.708, acc_len=1.71, thru=18.5, I=21.5, T=32.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:43,895] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:25:50,043] loop.py:380 INFO TIMING step=11485: step=3.701s data=1.685s compute=3.653s [fwd=2.601s bwd=1.037s opt=0.003s] dispatch=2.653s [2026-10-19 14:25:50,044] loop.py:45 INFO TRAIN_STEP step=11485 loss=1.234364 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:49,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 64%|██████▎ | 11485/18018 [24:55:36<13:34:07, 7.48s/step, loss=1.234, acc=0.713, acc_len=1.72, thru=19.0, I=19.1, T=34.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:50,496] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:54,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:25:55,657] loop.py:45 INFO TRAIN_STEP step=11486 loss=1.169156 Training: 64%|██████▎ | 11486/18018 [24:55:41<12:33:14, 6.92s/step, loss=1.169, acc=0.725, acc_len=1.76, thru=20.8, I=23.6, T=36.2, wait=2.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:25:55,652] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11485: forward=2476.3ms backward=1006.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:25:55,727] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 14:26:02,981] loop.py:45 INFO TRAIN_STEP step=11487 loss=1.288253 Training: 64%|██████▍ | 11487/18018 [24:55:49<12:46:15, 7.04s/step, loss=1.288, acc=0.705, acc_len=1.70, thru=16.4, I=18.8, T=37.0, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:11,674] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:26:12,230] loop.py:45 INFO TRAIN_STEP step=11488 loss=1.457970 Training: 64%|██████▍ | 11488/18018 [24:55:58<13:58:21, 7.70s/step, loss=1.458, acc=0.674, acc_len=1.59, thru=14.2, I=14.7, T=28.7, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:12,682] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:17,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:26:18,544] loop.py:45 INFO TRAIN_STEP step=11489 loss=1.323792 Training: 64%|██████▍ | 11489/18018 [24:56:04<13:12:48, 7.29s/step, loss=1.324, acc=0.692, acc_len=1.66, thru=19.4, I=20.9, T=36.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:18,716] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:24,207] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:26:24,693] loop.py:380 INFO TIMING step=11490: step=3.287s data=1.319s compute=3.259s [fwd=2.351s bwd=0.894s opt=0.003s] dispatch=2.854s [2026-10-19 14:26:24,693] loop.py:45 INFO TRAIN_STEP step=11490 loss=1.152621 Training: 64%|██████▍ | 11490/18018 [24:56:10<12:35:37, 6.95s/step, loss=1.153, acc=0.726, acc_len=1.77, thru=20.3, I=24.6, T=38.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:24,744] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:29,485] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:26:30,123] loop.py:45 INFO TRAIN_STEP step=11491 loss=1.148760 Training: 64%|██████▍ | 11491/18018 [24:56:16<11:46:04, 6.49s/step, loss=1.149, acc=0.727, acc_len=1.77, thru=23.1, I=25.0, T=36.7, wait=1.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:26:30,118] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11490: forward=2421.5ms backward=1019.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:30,491] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:35,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:26:36,211] loop.py:45 INFO TRAIN_STEP step=11492 loss=1.378900 Training: 64%|██████▍ | 11492/18018 [24:56:22<11:32:49, 6.37s/step, loss=1.379, acc=0.688, acc_len=1.64, thru=20.8, I=24.2, T=33.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:36,404] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:42,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:26:43,099] loop.py:45 INFO TRAIN_STEP step=11493 loss=1.364876 Training: 64%|██████▍ | 11493/18018 [24:56:29<11:49:37, 6.53s/step, loss=1.365, acc=0.688, acc_len=1.64, thru=18.3, I=21.0, T=33.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:43,341] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:47,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:26:48,751] loop.py:45 INFO TRAIN_STEP step=11494 loss=1.186404 Training: 64%|██████▍ | 11494/18018 [24:56:34<11:21:00, 6.26s/step, loss=1.186, acc=0.723, acc_len=1.76, thru=21.4, I=26.4, T=38.0, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:26:49,035] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:26:54,828] loop.py:380 INFO TIMING step=11495: step=3.432s data=1.452s compute=3.402s [fwd=2.433s bwd=0.954s opt=0.003s] dispatch=2.636s [2026-10-19 14:26:54,829] loop.py:45 INFO TRAIN_STEP step=11495 loss=1.205182 Training: 64%|██████▍ | 11495/18018 [24:56:40<11:14:49, 6.21s/step, loss=1.205, acc=0.717, acc_len=1.75, thru=21.2, I=24.8, T=37.3, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:00,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:27:01,588] loop.py:45 INFO TRAIN_STEP step=11496 loss=1.269355 Training: 64%|██████▍ | 11496/18018 [24:56:47<11:32:47, 6.37s/step, loss=1.269, acc=0.705, acc_len=1.70, thru=19.3, I=20.4, T=35.1, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:27:01,583] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11495: forward=2527.8ms backward=1061.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:01,764] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:27:08,203] loop.py:45 INFO TRAIN_STEP step=11497 loss=1.304788 Training: 64%|██████▍ | 11497/18018 [24:56:54<11:40:28, 6.45s/step, loss=1.305, acc=0.702, acc_len=1.69, thru=18.9, I=20.4, T=36.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:08,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:08,661] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:14,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:27:15,221] loop.py:45 INFO TRAIN_STEP step=11498 loss=1.286628 Training: 64%|██████▍ | 11498/18018 [24:57:01<11:59:07, 6.62s/step, loss=1.287, acc=0.704, acc_len=1.70, thru=18.1, I=18.4, T=31.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:15,493] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:20,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:27:21,985] loop.py:45 INFO TRAIN_STEP step=11499 loss=1.306849 Training: 64%|██████▍ | 11499/18018 [24:57:08<12:03:44, 6.66s/step, loss=1.307, acc=0.698, acc_len=1.67, thru=18.5, I=21.2, T=32.7, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:22,033] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 14:27:28,184] loop.py:380 INFO TIMING step=11500: step=4.005s data=1.814s compute=3.976s [fwd=2.823s bwd=1.138s opt=0.003s] dispatch=2.186s [2026-10-19 14:27:28,184] loop.py:45 INFO TRAIN_STEP step=11500 loss=1.386669 Training: 64%|██████▍ | 11500/18018 [24:57:14<11:48:35, 6.52s/step, loss=1.387, acc=0.685, acc_len=1.63, thru=18.2, I=21.2, T=32.0, wait=2.2s, pool=57, epoch=2/2][2026-10-19 14:27:28,188] loop.py:419 INFO Saving checkpoint at step 11500... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:28,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 14:27:31,621] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0011501 (TrainerActor pid=9310) [2026-10-19 14:27:31,630] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0010751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:33,360] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:33,360] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=1/64 [2026-10-19 14:27:39,507] loop.py:45 INFO TRAIN_STEP step=11501 loss=1.305432 Training: 64%|██████▍ | 11501/18018 [24:57:25<14:24:50, 7.96s/step, loss=1.305, acc=0.698, acc_len=1.67, thru=18.2, I=15.7, T=32.6, wait=2.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:27:39,502] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11500: forward=2795.4ms backward=1085.4ms (TrainerActor pid=9310) [2026-10-19 14:27:31,620] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0011501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 14:27:32,861] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0010751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:46,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:27:47,639] loop.py:45 INFO TRAIN_STEP step=11502 loss=1.314555 Training: 64%|██████▍ | 11502/18018 [24:57:33<14:30:17, 8.01s/step, loss=1.315, acc=0.700, acc_len=1.68, thru=18.2, I=18.0, T=31.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:48,115] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:52,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:27:53,369] loop.py:45 INFO TRAIN_STEP step=11503 loss=1.350785 Training: 64%|██████▍ | 11503/18018 [24:57:39<13:15:46, 7.33s/step, loss=1.351, acc=0.689, acc_len=1.65, thru=20.2, I=27.7, T=38.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:53,765] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:58,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:27:59,522] loop.py:45 INFO TRAIN_STEP step=11504 loss=1.358310 Training: 64%|██████▍ | 11504/18018 [24:57:45<12:37:23, 6.98s/step, loss=1.358, acc=0.691, acc_len=1.65, thru=21.6, I=26.8, T=39.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:27:59,573] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:28:07,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:28:07,667] loop.py:380 INFO TIMING step=11505: step=3.757s data=1.734s compute=3.710s [fwd=2.564s bwd=1.132s opt=0.003s] dispatch=4.378s [2026-10-19 14:28:07,667] loop.py:45 INFO TRAIN_STEP step=11505 loss=1.329498 Training: 64%|██████▍ | 11505/18018 [24:57:53<13:15:16, 7.33s/step, loss=1.329, acc=0.696, acc_len=1.66, thru=15.6, I=18.1, T=34.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:28:07,735] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:28:12,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:28:14,024] loop.py:45 INFO TRAIN_STEP step=11506 loss=1.531492 Training: 64%|██████▍ | 11506/18018 [24:58:00<12:43:35, 7.04s/step, loss=1.531, acc=0.660, acc_len=1.55, thru=19.7, I=21.0, T=35.9, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:28:14,020] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11505: forward=2532.7ms backward=990.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:28:14,100] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=33/64 [2026-10-19 14:28:21,023] loop.py:45 INFO TRAIN_STEP step=11507 loss=1.323366 Training: 64%|██████▍ | 11507/18018 [24:58:07<12:42:19, 7.02s/step, loss=1.323, acc=0.697, acc_len=1.67, thru=16.2, I=20.4, T=36.6, wait=3.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:28:27,027] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:28:28,523] loop.py:45 INFO TRAIN_STEP step=11508 loss=1.348816 Training: 64%|██████▍ | 11508/18018 [24:58:14<12:57:41, 7.17s/step, loss=1.349, acc=0.695, acc_len=1.66, thru=18.4, I=20.0, T=30.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:28:28,557] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 14:28:34,028] loop.py:45 INFO TRAIN_STEP step=11509 loss=1.179637 Training: 64%|██████▍ | 11509/18018 [24:58:20<12:03:22, 6.67s/step, loss=1.180, acc=0.721, acc_len=1.75, thru=18.9, I=21.5, T=37.1, wait=2.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:28:45,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:28:46,530] loop.py:380 INFO TIMING step=11510: step=4.018s data=1.872s compute=3.940s [fwd=2.748s bwd=1.177s opt=0.003s] dispatch=8.476s [2026-10-19 14:28:46,530] loop.py:45 INFO TRAIN_STEP step=11510 loss=1.369636 Training: 64%|██████▍ | 11510/18018 [24:58:32<15:13:11, 8.42s/step, loss=1.370, acc=0.691, acc_len=1.65, thru=10.6, I=11.3, T=31.9, wait=8.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:28:46,668] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 14:28:53,448] loop.py:45 INFO TRAIN_STEP step=11511 loss=1.196032 Training: 64%|██████▍ | 11511/18018 [24:58:39<14:24:10, 7.97s/step, loss=1.196, acc=0.715, acc_len=1.73, thru=14.1, I=17.8, T=35.9, wait=3.3s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:28:53,442] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11510: forward=2553.0ms backward=971.2ms [2026-10-19 14:29:07,111] loop.py:45 INFO TRAIN_STEP step=11512 loss=1.470559 Training: 64%|██████▍ | 11512/18018 [24:58:53<17:29:13, 9.68s/step, loss=1.471, acc=0.675, acc_len=1.60, thru=12.9, I=11.5, T=32.8, wait=9.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:29:07,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:29:07,608] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:29:14,211] loop.py:45 INFO TRAIN_STEP step=11513 loss=1.301659 Training: 64%|██████▍ | 11513/18018 [24:59:00<16:05:18, 8.90s/step, loss=1.302, acc=0.698, acc_len=1.67, thru=16.9, I=16.4, T=31.1, wait=3.0s, pool=49, epoch=2/2][2026-10-19 14:29:23,443] loop.py:45 INFO TRAIN_STEP step=11514 loss=1.484469 Training: 64%|██████▍ | 11514/18018 [24:59:09<16:15:58, 9.00s/step, loss=1.484, acc=0.670, acc_len=1.58, thru=13.8, I=12.8, T=33.8, wait=5.4s, pool=45, epoch=2/2][2026-10-19 14:29:31,410] loop.py:380 INFO TIMING step=11515: step=3.644s data=1.499s compute=3.593s [fwd=2.528s bwd=1.051s opt=0.003s] dispatch=4.310s [2026-10-19 14:29:31,410] loop.py:45 INFO TRAIN_STEP step=11515 loss=1.183493 Training: 64%|██████▍ | 11515/18018 [24:59:17<15:42:02, 8.69s/step, loss=1.183, acc=0.724, acc_len=1.76, thru=18.0, I=17.5, T=35.1, wait=4.3s, pool=61, epoch=2/2][2026-10-19 14:29:38,638] loop.py:45 INFO TRAIN_STEP step=11516 loss=1.252261 Training: 64%|██████▍ | 11516/18018 [24:59:24<14:54:15, 8.25s/step, loss=1.252, acc=0.708, acc_len=1.71, thru=16.7, I=17.9, T=34.9, wait=3.6s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:29:38,633] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11515: forward=2604.3ms backward=1008.3ms [2026-10-19 14:29:47,181] loop.py:45 INFO TRAIN_STEP step=11517 loss=1.318148 Training: 64%|██████▍ | 11517/18018 [24:59:33<15:03:33, 8.34s/step, loss=1.318, acc=0.690, acc_len=1.65, thru=13.6, I=14.1, T=31.8, wait=4.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:29:53,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:29:54,745] loop.py:45 INFO TRAIN_STEP step=11518 loss=1.413497 Training: 64%|██████▍ | 11518/18018 [24:59:40<14:38:15, 8.11s/step, loss=1.413, acc=0.683, acc_len=1.62, thru=17.9, I=18.5, T=36.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:29:55,024] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:30:00,966] loop.py:45 INFO TRAIN_STEP step=11519 loss=1.277295 Training: 64%|██████▍ | 11519/18018 [24:59:47<13:36:50, 7.54s/step, loss=1.277, acc=0.703, acc_len=1.68, thru=19.4, I=26.8, T=39.7, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:30:09,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:30:11,331] loop.py:380 INFO TIMING step=11520: step=4.270s data=2.039s compute=4.231s [fwd=3.036s bwd=1.180s opt=0.003s] dispatch=6.085s [2026-10-19 14:30:11,332] loop.py:45 INFO TRAIN_STEP step=11520 loss=1.297777 Training: 64%|██████▍ | 11520/18018 [24:59:57<15:08:30, 8.39s/step, loss=1.298, acc=0.702, acc_len=1.68, thru=13.6, I=13.9, T=30.0, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:30:11,389] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:30:17,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:30:17,554] loop.py:45 INFO TRAIN_STEP step=11521 loss=1.256657 (TrainerActor pid=9310) [2026-10-19 14:30:17,549] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11520: forward=2658.5ms backward=1073.2ms Training: 64%|██████▍ | 11521/18018 [25:00:03<13:57:58, 7.74s/step, loss=1.257, acc=0.707, acc_len=1.70, thru=19.3, I=21.6, T=33.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:30:17,590] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 14:30:25,183] loop.py:45 INFO TRAIN_STEP step=11522 loss=1.303677 Training: 64%|██████▍ | 11522/18018 [25:00:11<13:54:18, 7.71s/step, loss=1.304, acc=0.702, acc_len=1.69, thru=15.8, I=16.2, T=34.4, wait=3.9s, pool=61, epoch=2/2][2026-10-19 14:30:33,079] loop.py:45 INFO TRAIN_STEP step=11523 loss=1.295566 Training: 64%|██████▍ | 11523/18018 [25:00:19<14:00:17, 7.76s/step, loss=1.296, acc=0.703, acc_len=1.69, thru=16.0, I=14.8, T=33.2, wait=4.0s, pool=61, epoch=2/2][2026-10-19 14:30:40,131] loop.py:45 INFO TRAIN_STEP step=11524 loss=1.447011 Training: 64%|██████▍ | 11524/18018 [25:00:26<13:37:04, 7.55s/step, loss=1.447, acc=0.677, acc_len=1.60, thru=17.2, I=17.1, T=34.6, wait=3.3s, pool=49, epoch=2/2][2026-10-19 14:30:49,469] loop.py:380 INFO TIMING step=11525: step=4.457s data=2.073s compute=4.412s [fwd=3.181s bwd=1.217s opt=0.003s] dispatch=4.873s [2026-10-19 14:30:49,469] loop.py:45 INFO TRAIN_STEP step=11525 loss=1.243065 Training: 64%|██████▍ | 11525/18018 [25:00:35<14:34:59, 8.09s/step, loss=1.243, acc=0.710, acc_len=1.71, thru=15.2, I=16.2, T=28.7, wait=4.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:30:56,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:30:56,990] loop.py:45 INFO TRAIN_STEP step=11526 loss=1.331346 Training: 64%|██████▍ | 11526/18018 [25:00:43<14:16:32, 7.92s/step, loss=1.331, acc=0.696, acc_len=1.67, thru=17.3, I=16.3, T=34.4, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:30:56,985] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11525: forward=2636.5ms backward=1030.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:30:57,360] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:31:03,959] loop.py:45 INFO TRAIN_STEP step=11527 loss=1.400572 Training: 64%|██████▍ | 11527/18018 [25:00:50<13:45:45, 7.63s/step, loss=1.401, acc=0.684, acc_len=1.62, thru=16.6, I=16.6, T=36.7, wait=3.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:31:11,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:31:12,459] loop.py:45 INFO TRAIN_STEP step=11528 loss=1.365914 Training: 64%|██████▍ | 11528/18018 [25:00:58<14:13:43, 7.89s/step, loss=1.366, acc=0.690, acc_len=1.65, thru=17.8, I=16.6, T=31.5, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:31:12,836] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:31:20,054] loop.py:45 INFO TRAIN_STEP step=11529 loss=1.469621 Training: 64%|██████▍ | 11529/18018 [25:01:06<14:03:51, 7.80s/step, loss=1.470, acc=0.669, acc_len=1.58, thru=15.5, I=16.8, T=35.2, wait=4.0s, pool=49, epoch=2/2][2026-10-19 14:31:29,368] loop.py:380 INFO TIMING step=11530: step=4.129s data=1.752s compute=4.099s [fwd=2.913s bwd=1.171s opt=0.003s] dispatch=5.178s [2026-10-19 14:31:29,368] loop.py:45 INFO TRAIN_STEP step=11530 loss=1.285941 Training: 64%|██████▍ | 11530/18018 [25:01:15<14:52:44, 8.26s/step, loss=1.286, acc=0.705, acc_len=1.70, thru=14.1, I=13.7, T=31.0, wait=5.2s, pool=53, epoch=2/2][2026-10-19 14:31:35,778] loop.py:45 INFO TRAIN_STEP step=11531 loss=1.217285 Training: 64%|██████▍ | 11531/18018 [25:01:21<13:52:51, 7.70s/step, loss=1.217, acc=0.716, acc_len=1.74, thru=19.4, I=19.7, T=30.5, wait=2.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:31:35,773] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11530: forward=2961.0ms backward=1187.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:31:41,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:31:42,360] loop.py:45 INFO TRAIN_STEP step=11532 loss=1.268315 Training: 64%|██████▍ | 11532/18018 [25:01:28<13:16:20, 7.37s/step, loss=1.268, acc=0.708, acc_len=1.70, thru=17.1, I=18.8, T=35.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:31:42,708] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:31:49,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:31:50,118] loop.py:45 INFO TRAIN_STEP step=11533 loss=1.348115 Training: 64%|██████▍ | 11533/18018 [25:01:36<13:28:50, 7.48s/step, loss=1.348, acc=0.691, acc_len=1.65, thru=15.7, I=18.9, T=35.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:31:50,511] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:31:57,407] loop.py:45 INFO TRAIN_STEP step=11534 loss=1.330782 Training: 64%|██████▍ | 11534/18018 [25:01:43<13:22:23, 7.42s/step, loss=1.331, acc=0.695, acc_len=1.67, thru=15.3, I=19.8, T=35.9, wait=3.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:07,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:32:07,904] loop.py:380 INFO TIMING step=11535: step=3.806s data=1.675s compute=3.762s [fwd=2.631s bwd=1.116s opt=0.003s] dispatch=6.684s [2026-10-19 14:32:07,904] loop.py:45 INFO TRAIN_STEP step=11535 loss=1.333670 Training: 64%|██████▍ | 11535/18018 [25:01:53<15:01:55, 8.35s/step, loss=1.334, acc=0.699, acc_len=1.68, thru=15.3, I=13.7, T=33.6, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:08,266] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:12,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:32:13,886] loop.py:45 INFO TRAIN_STEP step=11536 loss=1.252465 Training: 64%|██████▍ | 11536/18018 [25:01:59<13:45:04, 7.64s/step, loss=1.252, acc=0.709, acc_len=1.71, thru=20.3, I=26.7, T=33.6, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:32:13,881] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11535: forward=2638.4ms backward=1120.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:14,151] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:19,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:32:20,101] loop.py:45 INFO TRAIN_STEP step=11537 loss=1.266109 Training: 64%|██████▍ | 11537/18018 [25:02:06<12:58:55, 7.21s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=19.2, I=27.2, T=38.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:20,607] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:24,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 14:32:26,753] loop.py:45 INFO TRAIN_STEP step=11538 loss=1.308368 Training: 64%|██████▍ | 11538/18018 [25:02:12<12:40:41, 7.04s/step, loss=1.308, acc=0.701, acc_len=1.68, thru=19.6, I=30.6, T=31.1, wait=2.5s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:26,967] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:31,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:32:32,060] loop.py:45 INFO TRAIN_STEP step=11539 loss=1.329754 Training: 64%|██████▍ | 11539/18018 [25:02:18<11:44:19, 6.52s/step, loss=1.330, acc=0.700, acc_len=1.68, thru=21.2, I=30.2, T=37.0, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:32:32,425] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:32:38,449] loop.py:380 INFO TIMING step=11540: step=3.579s data=1.512s compute=3.547s [fwd=2.550s bwd=0.983s opt=0.003s] dispatch=2.800s [2026-10-19 14:32:38,449] loop.py:45 INFO TRAIN_STEP step=11540 loss=1.232085 Training: 64%|██████▍ | 11540/18018 [25:02:24<11:39:58, 6.48s/step, loss=1.232, acc=0.714, acc_len=1.73, thru=17.4, I=21.5, T=35.8, wait=2.8s, pool=37, epoch=2/2][2026-10-19 14:32:47,079] loop.py:45 INFO TRAIN_STEP step=11541 loss=1.245420 Training: 64%|██████▍ | 11541/18018 [25:02:33<12:49:18, 7.13s/step, loss=1.245, acc=0.711, acc_len=1.72, thru=16.3, I=14.3, T=35.1, wait=5.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:32:47,074] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11540: forward=2617.6ms backward=978.0ms [2026-10-19 14:32:55,710] loop.py:45 INFO TRAIN_STEP step=11542 loss=1.228288 Training: 64%|██████▍ | 11542/18018 [25:02:41<13:37:48, 7.58s/step, loss=1.228, acc=0.716, acc_len=1.74, thru=16.9, I=14.8, T=26.5, wait=3.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:00,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:33:02,952] loop.py:45 INFO TRAIN_STEP step=11543 loss=1.244025 Training: 64%|██████▍ | 11543/18018 [25:02:49<13:26:56, 7.48s/step, loss=1.244, acc=0.714, acc_len=1.73, thru=18.1, I=20.9, T=28.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:03,097] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:07,856] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:33:09,485] loop.py:45 INFO TRAIN_STEP step=11544 loss=1.273044 Training: 64%|██████▍ | 11544/18018 [25:02:55<12:56:13, 7.19s/step, loss=1.273, acc=0.708, acc_len=1.71, thru=17.0, I=28.2, T=35.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:09,870] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:33:15,584] loop.py:380 INFO TIMING step=11545: step=3.511s data=1.462s compute=3.484s [fwd=2.502s bwd=0.967s opt=0.003s] dispatch=2.577s [2026-10-19 14:33:15,584] loop.py:45 INFO TRAIN_STEP step=11545 loss=1.260404 Training: 64%|██████▍ | 11545/18018 [25:03:01<12:20:39, 6.87s/step, loss=1.260, acc=0.710, acc_len=1.71, thru=17.0, I=26.1, T=36.5, wait=2.6s, pool=29, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 14:33:15,652] eagle_store.py:517 WARNING batch_get_buffer missing keys (k369442-8d33cf0e_hs, k369442-8d33cf0e_ids, k369442-8d33cf0e_lhs); sleeping 0.50s. [2026-10-19 14:33:28,003] loop.py:45 INFO TRAIN_STEP step=11546 loss=1.269431 Training: 64%|██████▍ | 11546/18018 [25:03:14<15:20:20, 8.53s/step, loss=1.269, acc=0.708, acc_len=1.71, thru=13.9, I=13.5, T=31.3, wait=8.3s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:33:27,998] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11545: forward=2844.2ms backward=1181.8ms (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 14:33:15,682] eagle_store.py:517 WARNING batch_get_buffer missing keys (k369441-98f3102f_hs, k369441-98f3102f_ids, k369441-98f3102f_lhs); sleeping 0.50s. [repeated 2x across cluster] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:36,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:33:37,783] loop.py:45 INFO TRAIN_STEP step=11547 loss=1.270146 Training: 64%|██████▍ | 11547/18018 [25:03:23<16:00:25, 8.91s/step, loss=1.270, acc=0.707, acc_len=1.71, thru=13.8, I=15.0, T=30.7, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:38,214] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:43,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:33:44,938] loop.py:45 INFO TRAIN_STEP step=11548 loss=1.348147 Training: 64%|██████▍ | 11548/18018 [25:03:31<15:03:41, 8.38s/step, loss=1.348, acc=0.695, acc_len=1.66, thru=16.2, I=22.8, T=37.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:45,443] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:50,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:33:51,788] loop.py:45 INFO TRAIN_STEP step=11549 loss=1.295867 Training: 64%|██████▍ | 11549/18018 [25:03:37<14:14:01, 7.92s/step, loss=1.296, acc=0.702, acc_len=1.69, thru=18.3, I=25.9, T=34.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:33:52,259] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:33:57,922] loop.py:380 INFO TIMING step=11550: step=3.497s data=1.494s compute=3.460s [fwd=2.523s bwd=0.923s opt=0.003s] dispatch=2.629s [2026-10-19 14:33:57,922] loop.py:45 INFO TRAIN_STEP step=11550 loss=1.393725 Training: 64%|██████▍ | 11550/18018 [25:03:44<13:16:05, 7.38s/step, loss=1.394, acc=0.686, acc_len=1.63, thru=19.4, I=21.6, T=36.6, wait=2.6s, pool=61, epoch=2/2][2026-10-19 14:34:06,493] loop.py:45 INFO TRAIN_STEP step=11551 loss=1.200088 Training: 64%|██████▍ | 11551/18018 [25:03:52<13:54:23, 7.74s/step, loss=1.200, acc=0.715, acc_len=1.73, thru=16.2, I=16.5, T=33.2, wait=4.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:34:06,489] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11550: forward=2648.3ms backward=1153.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:14,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:34:14,587] loop.py:45 INFO TRAIN_STEP step=11552 loss=1.243650 Training: 64%|██████▍ | 11552/18018 [25:04:00<14:05:34, 7.85s/step, loss=1.244, acc=0.706, acc_len=1.70, thru=15.7, I=15.1, T=33.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:14,934] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:19,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:34:20,829] loop.py:45 INFO TRAIN_STEP step=11553 loss=1.440267 Training: 64%|██████▍ | 11553/18018 [25:04:06<13:13:40, 7.37s/step, loss=1.440, acc=0.680, acc_len=1.61, thru=19.8, I=20.7, T=34.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:21,247] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:29,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:34:30,863] loop.py:45 INFO TRAIN_STEP step=11554 loss=1.295897 Training: 64%|██████▍ | 11554/18018 [25:04:16<14:39:43, 8.17s/step, loss=1.296, acc=0.701, acc_len=1.68, thru=13.3, I=15.5, T=31.2, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:30,931] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:35,825] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:34:37,117] loop.py:380 INFO TIMING step=11555: step=3.807s data=1.681s compute=3.762s [fwd=2.598s bwd=1.149s opt=0.003s] dispatch=2.438s [2026-10-19 14:34:37,117] loop.py:45 INFO TRAIN_STEP step=11555 loss=1.214574 Training: 64%|██████▍ | 11555/18018 [25:04:23<13:37:49, 7.59s/step, loss=1.215, acc=0.711, acc_len=1.72, thru=18.4, I=23.2, T=33.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:37,334] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:43,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:34:43,452] loop.py:45 INFO TRAIN_STEP step=11556 loss=1.223718 Training: 64%|██████▍ | 11556/18018 [25:04:29<12:57:02, 7.21s/step, loss=1.224, acc=0.714, acc_len=1.73, thru=19.3, I=22.9, T=37.9, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:34:43,447] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11555: forward=2413.0ms backward=923.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:34:43,774] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:34:50,537] loop.py:45 INFO TRAIN_STEP step=11557 loss=1.327728 Training: 64%|██████▍ | 11557/18018 [25:04:36<12:52:45, 7.18s/step, loss=1.328, acc=0.695, acc_len=1.66, thru=18.0, I=20.5, T=33.4, wait=3.2s, pool=57, epoch=2/2][2026-10-19 14:34:58,758] loop.py:45 INFO TRAIN_STEP step=11558 loss=1.233199 Training: 64%|██████▍ | 11558/18018 [25:04:44<13:26:20, 7.49s/step, loss=1.233, acc=0.715, acc_len=1.73, thru=12.6, I=13.6, T=34.9, wait=4.5s, pool=41, epoch=2/2][2026-10-19 14:35:06,904] loop.py:45 INFO TRAIN_STEP step=11559 loss=1.347896 Training: 64%|██████▍ | 11559/18018 [25:04:52<13:47:29, 7.69s/step, loss=1.348, acc=0.692, acc_len=1.65, thru=17.2, I=17.4, T=35.7, wait=4.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:13,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:35:15,016] loop.py:380 INFO TIMING step=11560: step=4.024s data=1.845s compute=3.986s [fwd=2.829s bwd=1.143s opt=0.003s] dispatch=4.078s [2026-10-19 14:35:15,016] loop.py:45 INFO TRAIN_STEP step=11560 loss=1.412912 Training: 64%|██████▍ | 11560/18018 [25:05:01<14:01:03, 7.81s/step, loss=1.413, acc=0.685, acc_len=1.63, thru=15.9, I=17.9, T=31.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:15,246] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:35:22,785] loop.py:45 INFO TRAIN_STEP step=11561 loss=1.206767 Training: 64%|██████▍ | 11561/18018 [25:05:08<13:59:34, 7.80s/step, loss=1.207, acc=0.719, acc_len=1.75, thru=14.1, I=19.1, T=34.9, wait=4.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:35:22,781] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11560: forward=2581.3ms backward=1038.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:29,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:35:29,924] loop.py:45 INFO TRAIN_STEP step=11562 loss=1.270731 Training: 64%|██████▍ | 11562/18018 [25:05:16<13:38:00, 7.60s/step, loss=1.271, acc=0.705, acc_len=1.70, thru=17.2, I=17.8, T=32.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:30,231] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:36,749] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:35:37,476] loop.py:45 INFO TRAIN_STEP step=11563 loss=1.450421 Training: 64%|██████▍ | 11563/18018 [25:05:23<13:36:16, 7.59s/step, loss=1.450, acc=0.676, acc_len=1.60, thru=17.2, I=18.5, T=32.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:37,752] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:45,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:35:46,300] loop.py:45 INFO TRAIN_STEP step=11564 loss=1.297020 Training: 64%|██████▍ | 11564/18018 [25:05:32<14:15:58, 7.96s/step, loss=1.297, acc=0.705, acc_len=1.70, thru=13.7, I=15.5, T=34.2, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:46,589] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:51,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:35:52,742] loop.py:380 INFO TIMING step=11565: step=3.821s data=1.706s compute=3.792s [fwd=2.708s bwd=1.069s opt=0.003s] dispatch=2.613s [2026-10-19 14:35:52,742] loop.py:45 INFO TRAIN_STEP step=11565 loss=1.261165 Training: 64%|██████▍ | 11565/18018 [25:05:38<13:26:57, 7.50s/step, loss=1.261, acc=0.707, acc_len=1.71, thru=19.2, I=21.9, T=33.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:53,217] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:58,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:35:59,441] loop.py:45 INFO TRAIN_STEP step=11566 loss=1.310022 Training: 64%|██████▍ | 11566/18018 [25:05:45<13:00:56, 7.26s/step, loss=1.310, acc=0.701, acc_len=1.68, thru=17.7, I=22.5, T=33.0, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:35:59,437] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11565: forward=2739.6ms backward=1085.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:35:59,601] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:36:05,917] loop.py:45 INFO TRAIN_STEP step=11567 loss=1.347010 Training: 64%|██████▍ | 11567/18018 [25:05:52<12:35:25, 7.03s/step, loss=1.347, acc=0.688, acc_len=1.64, thru=16.6, I=18.8, T=37.3, wait=3.0s, pool=41, epoch=2/2][2026-10-19 14:36:15,487] loop.py:45 INFO TRAIN_STEP step=11568 loss=1.182991 Training: 64%|██████▍ | 11568/18018 [25:06:01<13:57:22, 7.79s/step, loss=1.183, acc=0.723, acc_len=1.76, thru=13.7, I=15.3, T=32.8, wait=5.7s, pool=41, epoch=2/2][2026-10-19 14:36:24,093] loop.py:45 INFO TRAIN_STEP step=11569 loss=1.375787 Training: 64%|██████▍ | 11569/18018 [25:06:10<14:23:34, 8.03s/step, loss=1.376, acc=0.689, acc_len=1.64, thru=16.8, I=15.1, T=33.2, wait=4.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:29,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:36:30,851] loop.py:380 INFO TIMING step=11570: step=3.756s data=1.584s compute=3.719s [fwd=2.571s bwd=1.132s opt=0.003s] dispatch=2.992s [2026-10-19 14:36:30,851] loop.py:45 INFO TRAIN_STEP step=11570 loss=1.350365 Training: 64%|██████▍ | 11570/18018 [25:06:16<13:42:14, 7.65s/step, loss=1.350, acc=0.695, acc_len=1.66, thru=20.2, I=21.2, T=34.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:31,123] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:35,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:36:36,942] loop.py:45 INFO TRAIN_STEP step=11571 loss=1.407348 Training: 64%|██████▍ | 11571/18018 [25:06:23<12:51:48, 7.18s/step, loss=1.407, acc=0.684, acc_len=1.62, thru=19.8, I=27.2, T=33.5, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:36:36,937] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11570: forward=2709.6ms backward=1065.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:37,179] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:43,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:36:44,836] loop.py:45 INFO TRAIN_STEP step=11572 loss=1.219380 Training: 64%|██████▍ | 11572/18018 [25:06:30<13:14:43, 7.40s/step, loss=1.219, acc=0.715, acc_len=1.74, thru=15.6, I=20.3, T=30.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:44,920] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:50,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:36:50,839] loop.py:45 INFO TRAIN_STEP step=11573 loss=1.278532 Training: 64%|██████▍ | 11573/18018 [25:06:36<12:29:36, 6.98s/step, loss=1.279, acc=0.708, acc_len=1.71, thru=20.4, I=21.6, T=32.5, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:50,923] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:57,666] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:36:58,543] loop.py:45 INFO TRAIN_STEP step=11574 loss=1.490324 Training: 64%|██████▍ | 11574/18018 [25:06:44<12:52:52, 7.20s/step, loss=1.490, acc=0.666, acc_len=1.58, thru=17.3, I=19.1, T=36.7, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:36:58,674] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:37:05,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:37:07,326] loop.py:380 INFO TIMING step=11575: step=4.091s data=2.012s compute=4.057s [fwd=3.043s bwd=0.999s opt=0.003s] dispatch=4.682s [2026-10-19 14:37:07,326] loop.py:45 INFO TRAIN_STEP step=11575 loss=1.342533 Training: 64%|██████▍ | 11575/18018 [25:06:53<13:43:52, 7.67s/step, loss=1.343, acc=0.694, acc_len=1.66, thru=13.7, I=17.6, T=31.3, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:37:07,484] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:37:12,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:37:13,666] loop.py:45 INFO TRAIN_STEP step=11576 loss=1.239512 Training: 64%|██████▍ | 11576/18018 [25:06:59<13:00:49, 7.27s/step, loss=1.240, acc=0.713, acc_len=1.72, thru=19.4, I=22.5, T=31.8, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:37:13,661] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11575: forward=2754.9ms backward=1218.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:37:13,771] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 14:37:20,158] loop.py:45 INFO TRAIN_STEP step=11577 loss=1.281842 Training: 64%|██████▍ | 11577/18018 [25:07:06<12:35:28, 7.04s/step, loss=1.282, acc=0.704, acc_len=1.70, thru=16.3, I=22.0, T=34.9, wait=2.8s, pool=49, epoch=2/2][2026-10-19 14:37:26,786] loop.py:45 INFO TRAIN_STEP step=11578 loss=1.324305 Training: 64%|██████▍ | 11578/18018 [25:07:12<12:22:17, 6.92s/step, loss=1.324, acc=0.698, acc_len=1.68, thru=17.8, I=17.4, T=34.1, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:37:32,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:37:33,485] loop.py:45 INFO TRAIN_STEP step=11579 loss=1.388791 Training: 64%|██████▍ | 11579/18018 [25:07:19<12:15:10, 6.85s/step, loss=1.389, acc=0.685, acc_len=1.63, thru=20.9, I=20.1, T=32.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:37:33,842] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:37:40,134] loop.py:380 INFO TIMING step=11580: step=3.412s data=1.402s compute=3.374s [fwd=2.446s bwd=0.912s opt=0.003s] dispatch=3.226s [2026-10-19 14:37:40,134] loop.py:45 INFO TRAIN_STEP step=11580 loss=1.451993 Training: 64%|██████▍ | 11580/18018 [25:07:26<12:08:33, 6.79s/step, loss=1.452, acc=0.669, acc_len=1.58, thru=17.7, I=20.9, T=37.5, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:37:47,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:37:48,047] loop.py:45 INFO TRAIN_STEP step=11581 loss=1.281121 Training: 64%|██████▍ | 11581/18018 [25:07:34<12:44:33, 7.13s/step, loss=1.281, acc=0.700, acc_len=1.68, thru=16.9, I=17.4, T=32.7, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:37:48,043] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11580: forward=2730.7ms backward=1129.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:37:48,101] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 14:37:54,470] loop.py:45 INFO TRAIN_STEP step=11582 loss=1.361216 Training: 64%|██████▍ | 11582/18018 [25:07:40<12:21:50, 6.92s/step, loss=1.361, acc=0.688, acc_len=1.64, thru=17.9, I=17.9, T=37.8, wait=3.0s, pool=49, epoch=2/2][2026-10-19 14:38:02,855] loop.py:45 INFO TRAIN_STEP step=11583 loss=1.284842 Training: 64%|██████▍ | 11583/18018 [25:07:48<13:08:59, 7.36s/step, loss=1.285, acc=0.707, acc_len=1.70, thru=13.3, I=14.3, T=34.1, wait=4.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:38:11,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:38:12,099] loop.py:45 INFO TRAIN_STEP step=11584 loss=1.272661 Training: 64%|██████▍ | 11584/18018 [25:07:58<14:09:35, 7.92s/step, loss=1.273, acc=0.705, acc_len=1.70, thru=16.8, I=15.8, T=32.9, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:38:12,539] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:38:19,454] loop.py:380 INFO TIMING step=11585: step=3.748s data=1.602s compute=3.713s [fwd=2.642s bwd=1.055s opt=0.003s] dispatch=3.598s [2026-10-19 14:38:19,454] loop.py:45 INFO TRAIN_STEP step=11585 loss=1.344519 Training: 64%|██████▍ | 11585/18018 [25:08:05<13:51:09, 7.75s/step, loss=1.345, acc=0.694, acc_len=1.66, thru=14.2, I=17.6, T=34.2, wait=3.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:38:29,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:38:30,847] loop.py:45 INFO TRAIN_STEP step=11586 loss=1.420844 Training: 64%|██████▍ | 11586/18018 [25:08:16<15:48:05, 8.84s/step, loss=1.421, acc=0.682, acc_len=1.62, thru=14.9, I=14.1, T=30.0, wait=7.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:38:30,844] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11585: forward=3074.2ms backward=1140.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:38:31,338] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:38:37,467] loop.py:45 INFO TRAIN_STEP step=11587 loss=1.347408 Training: 64%|██████▍ | 11587/18018 [25:08:23<14:36:23, 8.18s/step, loss=1.347, acc=0.693, acc_len=1.66, thru=15.0, I=16.9, T=37.5, wait=3.2s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:38:45,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:38:46,805] loop.py:45 INFO TRAIN_STEP step=11588 loss=1.342707 Training: 64%|██████▍ | 11588/18018 [25:08:32<15:13:38, 8.53s/step, loss=1.343, acc=0.695, acc_len=1.66, thru=17.4, I=16.3, T=33.6, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:38:46,958] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:38:54,348] loop.py:45 INFO TRAIN_STEP step=11589 loss=1.354478 Training: 64%|██████▍ | 11589/18018 [25:08:40<14:41:56, 8.23s/step, loss=1.354, acc=0.690, acc_len=1.65, thru=16.1, I=17.9, T=36.2, wait=4.0s, pool=61, epoch=2/2][2026-10-19 14:39:03,499] loop.py:380 INFO TIMING step=11590: step=3.808s data=1.627s compute=3.774s [fwd=2.672s bwd=1.088s opt=0.003s] dispatch=5.333s [2026-10-19 14:39:03,499] loop.py:45 INFO TRAIN_STEP step=11590 loss=1.300274 Training: 64%|██████▍ | 11590/18018 [25:08:49<15:11:30, 8.51s/step, loss=1.300, acc=0.702, acc_len=1.69, thru=13.1, I=12.9, T=33.6, wait=5.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:10,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:39:10,738] loop.py:45 INFO TRAIN_STEP step=11591 loss=1.235252 Training: 64%|██████▍ | 11591/18018 [25:08:56<14:30:23, 8.13s/step, loss=1.235, acc=0.709, acc_len=1.71, thru=19.0, I=17.9, T=37.0, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:39:10,733] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11590: forward=2467.3ms backward=943.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:11,222] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:18,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:39:18,361] loop.py:45 INFO TRAIN_STEP step=11592 loss=1.442248 Training: 64%|██████▍ | 11592/18018 [25:09:04<14:14:08, 7.98s/step, loss=1.442, acc=0.679, acc_len=1.60, thru=18.2, I=20.3, T=32.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:18,518] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:25,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:39:25,943] loop.py:45 INFO TRAIN_STEP step=11593 loss=1.223889 Training: 64%|██████▍ | 11593/18018 [25:09:12<14:01:22, 7.86s/step, loss=1.224, acc=0.714, acc_len=1.73, thru=16.2, I=17.7, T=36.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:26,194] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:30,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:39:31,481] loop.py:45 INFO TRAIN_STEP step=11594 loss=1.152091 Training: 64%|██████▍ | 11594/18018 [25:09:17<12:46:48, 7.16s/step, loss=1.152, acc=0.724, acc_len=1.77, thru=22.4, I=27.0, T=38.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:31,614] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:37,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:39:37,374] loop.py:380 INFO TIMING step=11595: step=3.645s data=1.637s compute=3.613s [fwd=2.558s bwd=1.040s opt=0.003s] dispatch=2.239s [2026-10-19 14:39:37,374] loop.py:45 INFO TRAIN_STEP step=11595 loss=1.383231 Training: 64%|██████▍ | 11595/18018 [25:09:23<12:05:52, 6.78s/step, loss=1.383, acc=0.684, acc_len=1.63, thru=22.1, I=24.2, T=35.1, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:37,663] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:44,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:39:44,954] loop.py:45 INFO TRAIN_STEP step=11596 loss=1.301285 Training: 64%|██████▍ | 11596/18018 [25:09:31<12:31:24, 7.02s/step, loss=1.301, acc=0.703, acc_len=1.69, thru=18.1, I=18.6, T=30.5, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:39:44,949] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11595: forward=2939.1ms backward=1212.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:39:45,305] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:39:53,001] loop.py:45 INFO TRAIN_STEP step=11597 loss=1.304869 Training: 64%|██████▍ | 11597/18018 [25:09:39<13:04:14, 7.33s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=14.6, I=15.6, T=34.6, wait=4.3s, pool=57, epoch=2/2][2026-10-19 14:40:00,064] loop.py:45 INFO TRAIN_STEP step=11598 loss=1.338301 Training: 64%|██████▍ | 11598/18018 [25:09:46<12:55:36, 7.25s/step, loss=1.338, acc=0.690, acc_len=1.65, thru=16.4, I=17.1, T=37.8, wait=3.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:40:05,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:40:06,747] loop.py:45 INFO TRAIN_STEP step=11599 loss=1.357399 Training: 64%|██████▍ | 11599/18018 [25:09:52<12:37:24, 7.08s/step, loss=1.357, acc=0.692, acc_len=1.65, thru=19.2, I=20.7, T=37.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:40:06,953] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:40:12,182] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:40:13,804] loop.py:380 INFO TIMING step=11600: step=3.953s data=1.727s compute=3.922s [fwd=2.757s bwd=1.150s opt=0.003s] dispatch=3.094s [2026-10-19 14:40:13,805] loop.py:45 INFO TRAIN_STEP step=11600 loss=1.323396 Training: 64%|██████▍ | 11600/18018 [25:09:59<12:36:31, 7.07s/step, loss=1.323, acc=0.697, acc_len=1.67, thru=17.7, I=21.7, T=32.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:40:14,191] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:40:19,796] loop.py:45 INFO TRAIN_STEP step=11601 loss=1.243804 Training: 64%|██████▍ | 11601/18018 [25:10:05<12:01:42, 6.75s/step, loss=1.244, acc=0.709, acc_len=1.70, thru=17.2, I=20.6, T=35.9, wait=2.4s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:40:19,792] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11600: forward=2530.3ms backward=991.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:40:26,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:40:26,820] loop.py:45 INFO TRAIN_STEP step=11602 loss=1.339552 Training: 64%|██████▍ | 11602/18018 [25:10:12<12:10:30, 6.83s/step, loss=1.340, acc=0.693, acc_len=1.66, thru=19.5, I=19.3, T=34.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:40:26,852] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 14:40:32,255] loop.py:45 INFO TRAIN_STEP step=11603 loss=1.265672 Training: 64%|██████▍ | 11603/18018 [25:10:18<11:25:32, 6.41s/step, loss=1.266, acc=0.704, acc_len=1.69, thru=20.6, I=23.5, T=36.7, wait=1.9s, pool=57, epoch=2/2][2026-10-19 14:40:39,747] loop.py:45 INFO TRAIN_STEP step=11604 loss=1.188192 Training: 64%|██████▍ | 11604/18018 [25:10:25<12:00:13, 6.74s/step, loss=1.188, acc=0.719, acc_len=1.75, thru=15.2, I=18.2, T=35.2, wait=3.8s, pool=45, epoch=2/2][2026-10-19 14:40:48,090] loop.py:380 INFO TIMING step=11605: step=3.546s data=1.485s compute=3.500s [fwd=2.461s bwd=1.025s opt=0.003s] dispatch=4.784s [2026-10-19 14:40:48,090] loop.py:45 INFO TRAIN_STEP step=11605 loss=1.420481 Training: 64%|██████▍ | 11605/18018 [25:10:34<12:51:29, 7.22s/step, loss=1.420, acc=0.681, acc_len=1.62, thru=16.4, I=16.0, T=36.1, wait=4.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:40:54,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:40:55,395] loop.py:45 INFO TRAIN_STEP step=11606 loss=1.430476 Training: 64%|██████▍ | 11606/18018 [25:10:41<12:54:07, 7.24s/step, loss=1.430, acc=0.680, acc_len=1.61, thru=18.0, I=19.8, T=36.3, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:40:55,391] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11605: forward=2516.8ms backward=963.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:40:55,761] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:00,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:41:02,000] loop.py:45 INFO TRAIN_STEP step=11607 loss=1.288276 Training: 64%|██████▍ | 11607/18018 [25:10:48<12:33:35, 7.05s/step, loss=1.288, acc=0.699, acc_len=1.67, thru=18.8, I=23.5, T=36.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:02,333] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:07,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:41:08,084] loop.py:45 INFO TRAIN_STEP step=11608 loss=1.171651 Training: 64%|██████▍ | 11608/18018 [25:10:54<12:02:25, 6.76s/step, loss=1.172, acc=0.722, acc_len=1.76, thru=20.4, I=25.3, T=36.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:08,160] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 14:41:14,738] loop.py:45 INFO TRAIN_STEP step=11609 loss=1.236853 Training: 64%|██████▍ | 11609/18018 [25:11:00<11:58:53, 6.73s/step, loss=1.237, acc=0.713, acc_len=1.72, thru=18.4, I=20.9, T=31.6, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:21,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:41:23,064] loop.py:380 INFO TIMING step=11610: step=4.112s data=1.995s compute=3.711s [fwd=2.569s bwd=1.126s opt=0.003s] dispatch=4.200s [2026-10-19 14:41:23,065] loop.py:45 INFO TRAIN_STEP step=11610 loss=1.246834 Training: 64%|██████▍ | 11610/18018 [25:11:09<12:49:53, 7.21s/step, loss=1.247, acc=0.707, acc_len=1.71, thru=16.7, I=18.5, T=31.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:23,316] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:28,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:41:28,365] loop.py:45 INFO TRAIN_STEP step=11611 loss=1.301675 Training: 64%|██████▍ | 11611/18018 [25:11:14<11:48:37, 6.64s/step, loss=1.302, acc=0.698, acc_len=1.67, thru=21.2, I=26.8, T=39.5, wait=2.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:41:28,361] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11610: forward=2312.2ms backward=879.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:28,536] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:33,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:41:34,249] loop.py:45 INFO TRAIN_STEP step=11612 loss=1.339577 Training: 64%|██████▍ | 11612/18018 [25:11:20<11:24:25, 6.41s/step, loss=1.340, acc=0.693, acc_len=1.66, thru=21.8, I=25.9, T=37.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:34,585] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:41:41,197] loop.py:45 INFO TRAIN_STEP step=11613 loss=1.299524 Training: 64%|██████▍ | 11613/18018 [25:11:27<11:41:35, 6.57s/step, loss=1.300, acc=0.698, acc_len=1.68, thru=16.5, I=21.2, T=35.6, wait=3.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:49,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:41:49,909] loop.py:45 INFO TRAIN_STEP step=11614 loss=1.350165 Training: 64%|██████▍ | 11614/18018 [25:11:35<12:49:55, 7.21s/step, loss=1.350, acc=0.691, acc_len=1.65, thru=15.5, I=15.7, T=36.0, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:50,068] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:55,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:41:56,253] loop.py:380 INFO TIMING step=11615: step=3.398s data=1.392s compute=3.377s [fwd=2.451s bwd=0.914s opt=0.003s] dispatch=2.935s [2026-10-19 14:41:56,253] loop.py:45 INFO TRAIN_STEP step=11615 loss=1.416370 Training: 64%|██████▍ | 11615/18018 [25:11:42<12:21:54, 6.95s/step, loss=1.416, acc=0.685, acc_len=1.63, thru=19.7, I=23.4, T=37.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:41:56,516] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:01,760] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:42:02,150] loop.py:45 INFO TRAIN_STEP step=11616 loss=1.314285 Training: 64%|██████▍ | 11616/18018 [25:11:48<11:48:05, 6.64s/step, loss=1.314, acc=0.699, acc_len=1.68, thru=20.6, I=24.1, T=39.1, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:42:02,146] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11615: forward=2337.4ms backward=898.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:02,262] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:06,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:42:07,823] loop.py:45 INFO TRAIN_STEP step=11617 loss=1.300101 Training: 64%|██████▍ | 11617/18018 [25:11:53<11:17:05, 6.35s/step, loss=1.300, acc=0.699, acc_len=1.68, thru=22.5, I=25.4, T=38.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:08,007] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 14:42:14,712] loop.py:45 INFO TRAIN_STEP step=11618 loss=1.255619 Training: 64%|██████▍ | 11618/18018 [25:12:00<11:34:18, 6.51s/step, loss=1.256, acc=0.708, acc_len=1.71, thru=18.7, I=19.9, T=34.3, wait=3.1s, pool=61, epoch=2/2][2026-10-19 14:42:22,029] loop.py:45 INFO TRAIN_STEP step=11619 loss=1.436657 Training: 64%|██████▍ | 11619/18018 [25:12:08<12:00:06, 6.75s/step, loss=1.437, acc=0.679, acc_len=1.61, thru=17.7, I=17.6, T=33.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:22,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:22,520] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:42:30,950] loop.py:380 INFO TIMING step=11620: step=4.795s data=1.969s compute=4.753s [fwd=3.552s bwd=1.188s opt=0.003s] dispatch=4.116s [2026-10-19 14:42:30,950] loop.py:45 INFO TRAIN_STEP step=11620 loss=1.308011 Training: 64%|██████▍ | 11620/18018 [25:12:17<13:09:21, 7.40s/step, loss=1.308, acc=0.698, acc_len=1.68, thru=13.6, I=15.5, T=26.7, wait=4.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:38,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:42:38,730] loop.py:45 INFO TRAIN_STEP step=11621 loss=1.419149 Training: 64%|██████▍ | 11621/18018 [25:12:24<13:21:16, 7.52s/step, loss=1.419, acc=0.681, acc_len=1.62, thru=17.2, I=14.8, T=34.4, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:42:38,725] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11620: forward=2635.3ms backward=1043.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:39,082] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:44,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:42:46,155] loop.py:45 INFO TRAIN_STEP step=11622 loss=1.255966 Training: 65%|██████▍ | 11622/18018 [25:12:32<13:18:18, 7.49s/step, loss=1.256, acc=0.707, acc_len=1.70, thru=18.1, I=19.3, T=35.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:46,405] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:50,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:42:51,891] loop.py:45 INFO TRAIN_STEP step=11623 loss=1.251127 Training: 65%|██████▍ | 11623/18018 [25:12:37<12:22:03, 6.96s/step, loss=1.251, acc=0.709, acc_len=1.71, thru=21.4, I=27.2, T=37.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:52,254] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:57,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:42:58,282] loop.py:45 INFO TRAIN_STEP step=11624 loss=1.254023 Training: 65%|██████▍ | 11624/18018 [25:12:44<12:03:43, 6.79s/step, loss=1.254, acc=0.709, acc_len=1.72, thru=20.2, I=26.2, T=35.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:42:58,525] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:43:03,998] loop.py:380 INFO TIMING step=11625: step=3.225s data=1.272s compute=3.192s [fwd=2.277s bwd=0.901s opt=0.003s] dispatch=2.483s [2026-10-19 14:43:03,999] loop.py:45 INFO TRAIN_STEP step=11625 loss=1.240104 Training: 65%|██████▍ | 11625/18018 [25:12:50<11:29:13, 6.47s/step, loss=1.240, acc=0.717, acc_len=1.73, thru=21.4, I=23.7, T=39.7, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:10,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:43:10,769] loop.py:45 INFO TRAIN_STEP step=11626 loss=1.306671 Training: 65%|██████▍ | 11626/18018 [25:12:56<11:38:45, 6.56s/step, loss=1.307, acc=0.699, acc_len=1.67, thru=19.9, I=19.0, T=36.5, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:43:10,765] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11625: forward=2519.5ms backward=945.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:11,157] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:16,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:43:16,568] loop.py:45 INFO TRAIN_STEP step=11627 loss=1.229398 Training: 65%|██████▍ | 11627/18018 [25:13:02<11:14:27, 6.33s/step, loss=1.229, acc=0.711, acc_len=1.72, thru=21.1, I=22.6, T=35.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:16,660] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:21,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:43:22,752] loop.py:45 INFO TRAIN_STEP step=11628 loss=1.184942 Training: 65%|██████▍ | 11628/18018 [25:13:08<11:09:38, 6.29s/step, loss=1.185, acc=0.718, acc_len=1.74, thru=20.3, I=22.1, T=36.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:22,875] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:28,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:43:29,401] loop.py:45 INFO TRAIN_STEP step=11629 loss=1.288011 Training: 65%|██████▍ | 11629/18018 [25:13:15<11:21:01, 6.40s/step, loss=1.288, acc=0.704, acc_len=1.69, thru=19.3, I=21.4, T=36.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:29,508] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:35,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:43:36,128] loop.py:380 INFO TIMING step=11630: step=3.445s data=1.407s compute=3.405s [fwd=2.425s bwd=0.966s opt=0.003s] dispatch=3.274s [2026-10-19 14:43:36,128] loop.py:45 INFO TRAIN_STEP step=11630 loss=1.375686 Training: 65%|██████▍ | 11630/18018 [25:13:22<11:31:33, 6.50s/step, loss=1.376, acc=0.687, acc_len=1.64, thru=19.2, I=20.8, T=37.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:36,480] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:43:45,050] loop.py:45 INFO TRAIN_STEP step=11631 loss=1.385644 Training: 65%|██████▍ | 11631/18018 [25:13:31<12:48:50, 7.22s/step, loss=1.386, acc=0.689, acc_len=1.64, thru=12.9, I=14.0, T=30.8, wait=4.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:43:45,045] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11630: forward=2962.2ms backward=1137.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:50,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:43:51,927] loop.py:45 INFO TRAIN_STEP step=11632 loss=1.427263 Training: 65%|██████▍ | 11632/18018 [25:13:38<12:37:45, 7.12s/step, loss=1.427, acc=0.677, acc_len=1.60, thru=18.9, I=19.3, T=32.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:43:52,243] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:43:59,294] loop.py:45 INFO TRAIN_STEP step=11633 loss=1.228373 Training: 65%|██████▍ | 11633/18018 [25:13:45<12:45:30, 7.19s/step, loss=1.228, acc=0.713, acc_len=1.72, thru=15.6, I=18.8, T=37.5, wait=3.9s, pool=45, epoch=2/2][2026-10-19 14:44:07,555] loop.py:45 INFO TRAIN_STEP step=11634 loss=1.234591 Training: 65%|██████▍ | 11634/18018 [25:13:53<13:19:25, 7.51s/step, loss=1.235, acc=0.712, acc_len=1.72, thru=15.8, I=15.2, T=31.9, wait=4.2s, pool=57, epoch=2/2][2026-10-19 14:44:14,068] loop.py:380 INFO TIMING step=11635: step=3.631s data=1.522s compute=3.603s [fwd=2.593s bwd=0.996s opt=0.003s] dispatch=2.875s [2026-10-19 14:44:14,068] loop.py:45 INFO TRAIN_STEP step=11635 loss=1.312946 Training: 65%|██████▍ | 11635/18018 [25:14:00<12:47:22, 7.21s/step, loss=1.313, acc=0.697, acc_len=1.67, thru=17.5, I=16.7, T=35.3, wait=2.9s, pool=53, epoch=2/2][2026-10-19 14:44:20,667] loop.py:45 INFO TRAIN_STEP step=11636 loss=1.281889 Training: 65%|██████▍ | 11636/18018 [25:14:06<12:27:41, 7.03s/step, loss=1.282, acc=0.703, acc_len=1.68, thru=19.4, I=18.0, T=36.2, wait=3.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:44:20,662] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11635: forward=2475.3ms backward=1005.8ms [2026-10-19 14:44:29,254] loop.py:45 INFO TRAIN_STEP step=11637 loss=1.315275 Training: 65%|██████▍ | 11637/18018 [25:14:15<13:17:12, 7.50s/step, loss=1.315, acc=0.699, acc_len=1.68, thru=14.2, I=14.9, T=35.8, wait=5.0s, pool=41, epoch=2/2][2026-10-19 14:44:37,918] loop.py:45 INFO TRAIN_STEP step=11638 loss=1.262972 Training: 65%|██████▍ | 11638/18018 [25:14:24<13:54:26, 7.85s/step, loss=1.263, acc=0.711, acc_len=1.72, thru=15.2, I=14.4, T=36.0, wait=5.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:44:46,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:44:46,952] loop.py:45 INFO TRAIN_STEP step=11639 loss=1.293393 Training: 65%|██████▍ | 11639/18018 [25:14:33<14:32:10, 8.20s/step, loss=1.293, acc=0.705, acc_len=1.70, thru=15.4, I=16.1, T=31.8, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:44:47,046] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:44:51,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:44:53,036] loop.py:380 INFO TIMING step=11640: step=3.586s data=1.534s compute=3.554s [fwd=2.510s bwd=1.030s opt=0.003s] dispatch=2.487s [2026-10-19 14:44:53,037] loop.py:45 INFO TRAIN_STEP step=11640 loss=1.241018 Training: 65%|██████▍ | 11640/18018 [25:14:39<13:24:28, 7.57s/step, loss=1.241, acc=0.712, acc_len=1.72, thru=20.1, I=21.2, T=35.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:44:53,499] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:44:58,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:44:58,933] loop.py:45 INFO TRAIN_STEP step=11641 loss=1.311180 Training: 65%|██████▍ | 11641/18018 [25:14:45<12:30:56, 7.07s/step, loss=1.311, acc=0.696, acc_len=1.67, thru=21.6, I=24.4, T=38.0, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:44:58,928] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11640: forward=2378.7ms backward=937.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:44:59,222] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:05,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:45:06,031] loop.py:45 INFO TRAIN_STEP step=11642 loss=1.230027 Training: 65%|██████▍ | 11642/18018 [25:14:52<12:31:53, 7.08s/step, loss=1.230, acc=0.717, acc_len=1.74, thru=18.6, I=20.3, T=36.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:06,126] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:11,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:45:12,074] loop.py:45 INFO TRAIN_STEP step=11643 loss=1.311424 Training: 65%|██████▍ | 11643/18018 [25:14:58<11:58:53, 6.77s/step, loss=1.311, acc=0.695, acc_len=1.67, thru=20.7, I=25.1, T=34.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:12,128] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 14:45:19,385] loop.py:45 INFO TRAIN_STEP step=11644 loss=1.169615 Training: 65%|██████▍ | 11644/18018 [25:15:05<12:16:06, 6.93s/step, loss=1.170, acc=0.723, acc_len=1.76, thru=15.3, I=16.9, T=35.0, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:27,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:45:27,758] loop.py:380 INFO TIMING step=11645: step=3.866s data=1.645s compute=3.833s [fwd=2.757s bwd=1.062s opt=0.003s] dispatch=4.498s [2026-10-19 14:45:27,758] loop.py:45 INFO TRAIN_STEP step=11645 loss=1.361699 Training: 65%|██████▍ | 11645/18018 [25:15:13<13:02:01, 7.36s/step, loss=1.362, acc=0.692, acc_len=1.65, thru=17.4, I=18.6, T=33.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:28,155] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:33,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:45:34,788] loop.py:45 INFO TRAIN_STEP step=11646 loss=1.412302 Training: 65%|██████▍ | 11646/18018 [25:15:20<12:51:21, 7.26s/step, loss=1.412, acc=0.682, acc_len=1.62, thru=17.2, I=20.8, T=37.5, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:45:34,784] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11645: forward=2363.6ms backward=1009.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:35,017] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:41,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:45:42,444] loop.py:45 INFO TRAIN_STEP step=11647 loss=1.322215 Training: 65%|██████▍ | 11647/18018 [25:15:28<13:03:42, 7.38s/step, loss=1.322, acc=0.696, acc_len=1.66, thru=17.0, I=22.0, T=34.3, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:42,532] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:47,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:45:48,441] loop.py:45 INFO TRAIN_STEP step=11648 loss=1.315299 Training: 65%|██████▍ | 11648/18018 [25:15:34<12:19:29, 6.97s/step, loss=1.315, acc=0.702, acc_len=1.68, thru=20.5, I=26.0, T=36.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:48,771] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:53,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:45:54,359] loop.py:45 INFO TRAIN_STEP step=11649 loss=1.191888 Training: 65%|██████▍ | 11649/18018 [25:15:40<11:46:01, 6.65s/step, loss=1.192, acc=0.718, acc_len=1.74, thru=21.6, I=26.8, T=36.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:45:54,668] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:46:00,861] loop.py:380 INFO TIMING step=11650: step=3.942s data=1.972s compute=3.912s [fwd=2.974s bwd=0.923s opt=0.003s] dispatch=2.551s [2026-10-19 14:46:00,861] loop.py:45 INFO TRAIN_STEP step=11650 loss=1.351796 Training: 65%|██████▍ | 11650/18018 [25:15:46<11:41:09, 6.61s/step, loss=1.352, acc=0.693, acc_len=1.65, thru=18.9, I=20.8, T=32.5, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:07,272] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:46:08,325] loop.py:45 INFO TRAIN_STEP step=11651 loss=1.249001 Training: 65%|██████▍ | 11651/18018 [25:15:54<12:08:18, 6.86s/step, loss=1.249, acc=0.707, acc_len=1.70, thru=16.9, I=18.3, T=33.8, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:46:08,320] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11650: forward=2617.3ms backward=1109.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:08,783] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:14,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:46:16,279] loop.py:45 INFO TRAIN_STEP step=11652 loss=1.309271 Training: 65%|██████▍ | 11652/18018 [25:16:02<12:42:59, 7.19s/step, loss=1.309, acc=0.698, acc_len=1.67, thru=15.9, I=21.3, T=33.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:16,395] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:21,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:46:22,581] loop.py:45 INFO TRAIN_STEP step=11653 loss=1.239586 Training: 65%|██████▍ | 11653/18018 [25:16:08<12:14:30, 6.92s/step, loss=1.240, acc=0.710, acc_len=1.72, thru=19.0, I=21.8, T=35.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:22,968] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:27,688] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:46:28,901] loop.py:45 INFO TRAIN_STEP step=11654 loss=1.297270 Training: 65%|██████▍ | 11654/18018 [25:16:14<11:55:11, 6.74s/step, loss=1.297, acc=0.701, acc_len=1.68, thru=19.7, I=24.8, T=31.5, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:29,198] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:34,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:46:35,594] loop.py:380 INFO TIMING step=11655: step=4.289s data=2.113s compute=4.251s [fwd=3.150s bwd=1.086s opt=0.003s] dispatch=2.395s [2026-10-19 14:46:35,594] loop.py:45 INFO TRAIN_STEP step=11655 loss=1.415183 Training: 65%|██████▍ | 11655/18018 [25:16:21<11:53:32, 6.73s/step, loss=1.415, acc=0.679, acc_len=1.61, thru=17.6, I=22.9, T=29.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:35,732] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:41,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:46:41,911] loop.py:45 INFO TRAIN_STEP step=11656 loss=1.204307 Training: 65%|██████▍ | 11656/18018 [25:16:27<11:40:21, 6.61s/step, loss=1.204, acc=0.719, acc_len=1.75, thru=18.1, I=21.9, T=34.8, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:46:41,907] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11655: forward=2631.1ms backward=1006.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:42,139] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:48,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:46:49,040] loop.py:45 INFO TRAIN_STEP step=11657 loss=1.228394 Training: 65%|██████▍ | 11657/18018 [25:16:35<11:56:54, 6.76s/step, loss=1.228, acc=0.716, acc_len=1.74, thru=17.6, I=19.4, T=35.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:49,345] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:56,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:46:56,806] loop.py:45 INFO TRAIN_STEP step=11658 loss=1.293999 Training: 65%|██████▍ | 11658/18018 [25:16:42<12:28:37, 7.06s/step, loss=1.294, acc=0.703, acc_len=1.68, thru=16.4, I=17.2, T=35.9, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:46:57,244] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:03,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:47:03,708] loop.py:45 INFO TRAIN_STEP step=11659 loss=1.202046 Training: 65%|██████▍ | 11659/18018 [25:16:49<12:23:27, 7.01s/step, loss=1.202, acc=0.721, acc_len=1.76, thru=19.4, I=17.3, T=36.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:03,975] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:47:10,825] loop.py:380 INFO TIMING step=11660: step=3.781s data=1.564s compute=3.747s [fwd=2.612s bwd=1.121s opt=0.003s] dispatch=3.326s [2026-10-19 14:47:10,825] loop.py:45 INFO TRAIN_STEP step=11660 loss=1.254360 Training: 65%|██████▍ | 11660/18018 [25:16:56<12:26:31, 7.04s/step, loss=1.254, acc=0.709, acc_len=1.72, thru=17.7, I=17.9, T=33.9, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:18,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:47:19,144] loop.py:45 INFO TRAIN_STEP step=11661 loss=1.244268 Training: 65%|██████▍ | 11661/18018 [25:17:05<13:06:55, 7.43s/step, loss=1.244, acc=0.706, acc_len=1.70, thru=16.9, I=17.8, T=35.9, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:47:19,139] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11660: forward=2555.4ms backward=967.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:19,514] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:24,927] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:47:25,855] loop.py:45 INFO TRAIN_STEP step=11662 loss=1.227944 Training: 65%|██████▍ | 11662/18018 [25:17:11<12:44:06, 7.21s/step, loss=1.228, acc=0.712, acc_len=1.72, thru=18.6, I=21.3, T=35.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:25,934] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:32,708] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:47:33,323] loop.py:45 INFO TRAIN_STEP step=11663 loss=1.261720 Training: 65%|██████▍ | 11663/18018 [25:17:19<12:51:59, 7.29s/step, loss=1.262, acc=0.706, acc_len=1.70, thru=16.5, I=18.9, T=33.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:33,714] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:40,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:47:40,223] loop.py:45 INFO TRAIN_STEP step=11664 loss=1.393271 Training: 65%|██████▍ | 11664/18018 [25:17:26<12:39:32, 7.17s/step, loss=1.393, acc=0.689, acc_len=1.64, thru=18.1, I=20.1, T=36.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:40,554] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:46,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:47:46,703] loop.py:380 INFO TIMING step=11665: step=3.595s data=1.499s compute=3.564s [fwd=2.580s bwd=0.969s opt=0.003s] dispatch=2.878s [2026-10-19 14:47:46,703] loop.py:45 INFO TRAIN_STEP step=11665 loss=1.225480 Training: 65%|██████▍ | 11665/18018 [25:17:32<12:17:26, 6.96s/step, loss=1.225, acc=0.715, acc_len=1.73, thru=19.1, I=20.7, T=35.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:47:47,166] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:47:54,040] loop.py:45 INFO TRAIN_STEP step=11666 loss=1.294841 Training: 65%|██████▍ | 11666/18018 [25:17:40<12:29:07, 7.08s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=14.6, I=16.4, T=36.3, wait=3.8s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:47:54,034] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11665: forward=2491.0ms backward=986.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:48:01,892] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:48:03,336] loop.py:45 INFO TRAIN_STEP step=11667 loss=1.237117 Training: 65%|██████▍ | 11667/18018 [25:17:49<13:39:32, 7.74s/step, loss=1.237, acc=0.713, acc_len=1.72, thru=17.7, I=18.3, T=33.7, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:48:03,410] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 14:48:10,005] loop.py:45 INFO TRAIN_STEP step=11668 loss=1.225528 Training: 65%|██████▍ | 11668/18018 [25:17:56<13:05:23, 7.42s/step, loss=1.226, acc=0.710, acc_len=1.72, thru=15.6, I=20.7, T=35.2, wait=3.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:48:18,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:48:18,873] loop.py:45 INFO TRAIN_STEP step=11669 loss=1.296200 Training: 65%|██████▍ | 11669/18018 [25:18:04<13:51:10, 7.85s/step, loss=1.296, acc=0.702, acc_len=1.69, thru=17.0, I=15.5, T=34.2, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:48:19,339] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:48:27,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:48:27,513] loop.py:380 INFO TIMING step=11670: step=4.010s data=1.934s compute=3.969s [fwd=2.862s bwd=1.091s opt=0.003s] dispatch=4.620s [2026-10-19 14:48:27,513] loop.py:45 INFO TRAIN_STEP step=11670 loss=1.323250 Training: 65%|██████▍ | 11670/18018 [25:18:13<14:15:56, 8.09s/step, loss=1.323, acc=0.695, acc_len=1.66, thru=14.3, I=15.3, T=31.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:48:27,914] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:48:33,555] loop.py:45 INFO TRAIN_STEP step=11671 loss=1.315231 Training: 65%|██████▍ | 11671/18018 [25:18:19<13:10:47, 7.48s/step, loss=1.315, acc=0.698, acc_len=1.67, thru=18.6, I=21.1, T=34.0, wait=2.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:48:33,550] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11670: forward=2663.5ms backward=1044.2ms [2026-10-19 14:48:41,351] loop.py:45 INFO TRAIN_STEP step=11672 loss=1.373460 Training: 65%|██████▍ | 11672/18018 [25:18:27<13:20:52, 7.57s/step, loss=1.373, acc=0.685, acc_len=1.63, thru=14.4, I=14.3, T=37.0, wait=4.3s, pool=53, epoch=2/2][2026-10-19 14:48:50,033] loop.py:45 INFO TRAIN_STEP step=11673 loss=1.294347 Training: 65%|██████▍ | 11673/18018 [25:18:36<13:55:53, 7.90s/step, loss=1.294, acc=0.702, acc_len=1.69, thru=15.8, I=15.5, T=33.2, wait=4.8s, pool=57, epoch=2/2][2026-10-19 14:48:57,766] loop.py:45 INFO TRAIN_STEP step=11674 loss=1.370231 Training: 65%|██████▍ | 11674/18018 [25:18:43<13:50:22, 7.85s/step, loss=1.370, acc=0.691, acc_len=1.65, thru=15.6, I=16.3, T=34.2, wait=4.0s, pool=53, epoch=2/2][2026-10-19 14:49:06,719] loop.py:380 INFO TIMING step=11675: step=3.815s data=1.436s compute=3.789s [fwd=2.794s bwd=0.981s opt=0.003s] dispatch=5.128s [2026-10-19 14:49:06,719] loop.py:45 INFO TRAIN_STEP step=11675 loss=1.285729 Training: 65%|██████▍ | 11675/18018 [25:18:52<14:25:07, 8.18s/step, loss=1.286, acc=0.702, acc_len=1.69, thru=14.0, I=13.2, T=33.5, wait=5.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:49:18,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:49:19,246] loop.py:45 INFO TRAIN_STEP step=11676 loss=1.421967 Training: 65%|██████▍ | 11676/18018 [25:19:05<16:42:43, 9.49s/step, loss=1.422, acc=0.682, acc_len=1.62, thru=14.8, I=11.3, T=30.4, wait=8.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:49:19,241] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11675: forward=2919.8ms backward=1238.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:49:19,488] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:49:26,514] loop.py:45 INFO TRAIN_STEP step=11677 loss=1.392906 Training: 65%|██████▍ | 11677/18018 [25:19:12<15:32:08, 8.82s/step, loss=1.393, acc=0.683, acc_len=1.63, thru=15.3, I=17.5, T=33.6, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:49:37,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:49:38,742] loop.py:45 INFO TRAIN_STEP step=11678 loss=1.445233 Training: 65%|██████▍ | 11678/18018 [25:19:24<17:20:07, 9.84s/step, loss=1.445, acc=0.677, acc_len=1.60, thru=11.8, I=13.2, T=28.7, wait=7.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:49:38,873] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:49:46,745] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:49:47,291] loop.py:45 INFO TRAIN_STEP step=11679 loss=1.332825 Training: 65%|██████▍ | 11679/18018 [25:19:33<16:38:51, 9.45s/step, loss=1.333, acc=0.690, acc_len=1.64, thru=13.3, I=16.8, T=34.3, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:49:47,752] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:49:52,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:49:53,290] loop.py:380 INFO TIMING step=11680: step=3.637s data=1.654s compute=3.601s [fwd=2.578s bwd=1.010s opt=0.003s] dispatch=2.355s [2026-10-19 14:49:53,290] loop.py:45 INFO TRAIN_STEP step=11680 loss=1.235154 Training: 65%|██████▍ | 11680/18018 [25:19:39<14:49:10, 8.42s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=20.1, I=24.0, T=35.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:49:53,757] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:50:02,441] loop.py:45 INFO TRAIN_STEP step=11681 loss=1.269592 Training: 65%|██████▍ | 11681/18018 [25:19:48<15:12:17, 8.64s/step, loss=1.270, acc=0.704, acc_len=1.69, thru=12.4, I=11.4, T=35.7, wait=5.6s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:50:02,437] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11680: forward=2550.5ms backward=990.5ms [2026-10-19 14:50:11,956] loop.py:45 INFO TRAIN_STEP step=11682 loss=1.311680 Training: 65%|██████▍ | 11682/18018 [25:19:58<15:39:59, 8.90s/step, loss=1.312, acc=0.697, acc_len=1.67, thru=13.5, I=14.0, T=35.8, wait=5.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:50:17,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:50:19,314] loop.py:45 INFO TRAIN_STEP step=11683 loss=1.283681 Training: 65%|██████▍ | 11683/18018 [25:20:05<14:50:54, 8.44s/step, loss=1.284, acc=0.701, acc_len=1.68, thru=17.7, I=20.0, T=33.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:50:19,790] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:50:24,587] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:50:25,155] loop.py:45 INFO TRAIN_STEP step=11684 loss=1.213002 Training: 65%|██████▍ | 11684/18018 [25:20:11<13:28:31, 7.66s/step, loss=1.213, acc=0.712, acc_len=1.73, thru=20.9, I=26.2, T=35.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:50:25,594] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:50:36,180] loop.py:380 INFO TIMING step=11685: step=4.030s data=1.754s compute=3.986s [fwd=2.676s bwd=1.295s opt=0.003s] dispatch=6.986s [2026-10-19 14:50:36,180] loop.py:45 INFO TRAIN_STEP step=11685 loss=1.298328 Training: 65%|██████▍ | 11685/18018 [25:20:22<15:14:57, 8.67s/step, loss=1.298, acc=0.701, acc_len=1.68, thru=11.0, I=10.4, T=31.8, wait=7.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:50:43,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:50:43,485] loop.py:45 INFO TRAIN_STEP step=11686 loss=1.322379 Training: 65%|██████▍ | 11686/18018 [25:20:29<14:31:43, 8.26s/step, loss=1.322, acc=0.694, acc_len=1.66, thru=18.5, I=17.2, T=33.3, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:50:43,481] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11685: forward=2709.7ms backward=1081.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:50:43,834] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:50:48,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:50:49,399] loop.py:45 INFO TRAIN_STEP step=11687 loss=1.219824 Training: 65%|██████▍ | 11687/18018 [25:20:35<13:17:16, 7.56s/step, loss=1.220, acc=0.712, acc_len=1.72, thru=20.2, I=22.1, T=36.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:50:49,427] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 14:50:54,924] loop.py:45 INFO TRAIN_STEP step=11688 loss=1.165066 Training: 65%|██████▍ | 11688/18018 [25:20:41<12:12:57, 6.95s/step, loss=1.165, acc=0.723, acc_len=1.77, thru=21.7, I=23.9, T=37.6, wait=2.1s, pool=61, epoch=2/2][2026-10-19 14:51:02,230] loop.py:45 INFO TRAIN_STEP step=11689 loss=1.324324 Training: 65%|██████▍ | 11689/18018 [25:20:48<12:24:09, 7.05s/step, loss=1.324, acc=0.697, acc_len=1.67, thru=17.5, I=18.6, T=36.0, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:08,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:51:08,510] loop.py:380 INFO TIMING step=11690: step=3.707s data=1.622s compute=3.668s [fwd=2.611s bwd=1.043s opt=0.003s] dispatch=2.563s [2026-10-19 14:51:08,510] loop.py:45 INFO TRAIN_STEP step=11690 loss=1.219900 Training: 65%|██████▍ | 11690/18018 [25:20:54<11:59:29, 6.82s/step, loss=1.220, acc=0.718, acc_len=1.74, thru=19.9, I=20.6, T=34.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:08,752] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:13,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:51:14,347] loop.py:45 INFO TRAIN_STEP step=11691 loss=1.242404 Training: 65%|██████▍ | 11691/18018 [25:21:00<11:28:11, 6.53s/step, loss=1.242, acc=0.709, acc_len=1.71, thru=20.9, I=23.7, T=33.7, wait=2.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:51:14,342] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11690: forward=2705.3ms backward=1043.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:14,430] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:19,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:51:21,152] loop.py:45 INFO TRAIN_STEP step=11692 loss=1.203835 Training: 65%|██████▍ | 11692/18018 [25:21:07<11:36:56, 6.61s/step, loss=1.204, acc=0.716, acc_len=1.74, thru=18.1, I=22.3, T=27.3, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:21,373] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:51:28,073] loop.py:45 INFO TRAIN_STEP step=11693 loss=1.278611 Training: 65%|██████▍ | 11693/18018 [25:21:14<11:46:35, 6.70s/step, loss=1.279, acc=0.702, acc_len=1.69, thru=15.5, I=17.9, T=29.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:28,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:28,549] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:33,393] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:51:34,705] loop.py:45 INFO TRAIN_STEP step=11694 loss=1.427675 Training: 65%|██████▍ | 11694/18018 [25:21:20<11:44:16, 6.68s/step, loss=1.428, acc=0.677, acc_len=1.60, thru=18.1, I=19.8, T=31.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:34,899] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:51:42,008] loop.py:380 INFO TIMING step=11695: step=3.972s data=1.831s compute=3.945s [fwd=2.779s bwd=1.151s opt=0.003s] dispatch=3.322s [2026-10-19 14:51:42,008] loop.py:45 INFO TRAIN_STEP step=11695 loss=1.328959 Training: 65%|██████▍ | 11695/18018 [25:21:28<12:03:50, 6.87s/step, loss=1.329, acc=0.693, acc_len=1.66, thru=15.3, I=17.6, T=32.2, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:48,935] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:51:50,075] loop.py:45 INFO TRAIN_STEP step=11696 loss=1.318871 Training: 65%|██████▍ | 11696/18018 [25:21:36<12:41:31, 7.23s/step, loss=1.319, acc=0.696, acc_len=1.66, thru=17.4, I=17.8, T=33.8, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:51:50,070] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11695: forward=2680.9ms backward=1067.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:50,442] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:55,171] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:51:55,677] loop.py:45 INFO TRAIN_STEP step=11697 loss=1.203484 Training: 65%|██████▍ | 11697/18018 [25:21:41<11:50:02, 6.74s/step, loss=1.203, acc=0.716, acc_len=1.74, thru=20.8, I=23.7, T=38.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:51:56,177] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:52:02,640] loop.py:45 INFO TRAIN_STEP step=11698 loss=1.224292 Training: 65%|██████▍ | 11698/18018 [25:21:48<11:57:06, 6.81s/step, loss=1.224, acc=0.715, acc_len=1.73, thru=15.8, I=19.2, T=33.7, wait=3.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:52:12,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:52:12,758] loop.py:45 INFO TRAIN_STEP step=11699 loss=1.257011 Training: 65%|██████▍ | 11699/18018 [25:21:58<13:41:28, 7.80s/step, loss=1.257, acc=0.706, acc_len=1.71, thru=16.1, I=14.8, T=33.0, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:52:12,853] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:52:20,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:52:21,643] loop.py:380 INFO TIMING step=11700: step=4.229s data=1.799s compute=4.200s [fwd=2.951s bwd=1.234s opt=0.003s] dispatch=4.647s [2026-10-19 14:52:21,643] loop.py:45 INFO TRAIN_STEP step=11700 loss=1.335442 Training: 65%|██████▍ | 11700/18018 [25:22:07<14:15:40, 8.13s/step, loss=1.335, acc=0.696, acc_len=1.67, thru=14.3, I=17.0, T=30.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:52:21,908] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:52:26,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:52:27,566] loop.py:45 INFO TRAIN_STEP step=11701 loss=1.288581 Training: 65%|██████▍ | 11701/18018 [25:22:13<13:05:55, 7.46s/step, loss=1.289, acc=0.702, acc_len=1.68, thru=19.4, I=27.6, T=33.9, wait=2.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:52:27,562] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11700: forward=2705.6ms backward=1027.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:52:27,934] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:52:34,878] loop.py:45 INFO TRAIN_STEP step=11702 loss=1.250043 Training: 65%|██████▍ | 11702/18018 [25:22:20<13:01:00, 7.42s/step, loss=1.250, acc=0.711, acc_len=1.72, thru=13.6, I=18.6, T=34.1, wait=3.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:52:42,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:52:44,368] loop.py:45 INFO TRAIN_STEP step=11703 loss=1.265804 Training: 65%|██████▍ | 11703/18018 [25:22:30<14:06:12, 8.04s/step, loss=1.266, acc=0.710, acc_len=1.71, thru=18.7, I=15.5, T=33.2, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:52:49,013] inference_manager.py:398 INFO Still waiting for pool capacity (waited 6.2s): pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:52:49,013] inference_manager.py:408 INFO Pool capacity available after 6.2s, pool_size=1/64 [2026-10-19 14:52:54,457] loop.py:45 INFO TRAIN_STEP step=11704 loss=1.279627 Training: 65%|██████▍ | 11704/18018 [25:22:40<15:10:48, 8.66s/step, loss=1.280, acc=0.699, acc_len=1.68, thru=19.1, I=22.6, T=36.9, wait=6.6s, pool=41, epoch=2/2](TrainerActor pid=9370) [2026-10-19 14:52:54,524] eagle_store.py:517 WARNING batch_get_buffer missing keys (k374527-a070a471_hs, k374527-a070a471_ids, k374527-a070a471_lhs); sleeping 0.50s. [2026-10-19 14:53:03,440] loop.py:380 INFO TIMING step=11705: step=3.843s data=1.676s compute=3.798s [fwd=2.718s bwd=1.066s opt=0.003s] dispatch=5.130s [2026-10-19 14:53:03,440] loop.py:45 INFO TRAIN_STEP step=11705 loss=1.331835 Training: 65%|██████▍ | 11705/18018 [25:22:49<15:20:57, 8.75s/step, loss=1.332, acc=0.694, acc_len=1.66, thru=16.8, I=15.4, T=33.3, wait=5.1s, pool=57, epoch=2/2][2026-10-19 14:53:12,335] loop.py:45 INFO TRAIN_STEP step=11706 loss=1.321590 Training: 65%|██████▍ | 11706/18018 [25:22:58<15:25:21, 8.80s/step, loss=1.322, acc=0.696, acc_len=1.66, thru=13.7, I=12.8, T=35.5, wait=5.3s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:53:12,331] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11705: forward=2548.9ms backward=1011.1ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 14:52:54,527] eagle_store.py:517 WARNING batch_get_buffer missing keys (k374528-869e70fa_hs, k374528-869e70fa_ids, k374528-869e70fa_lhs); sleeping 0.50s. [repeated 2x across cluster] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:20,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:53:21,314] loop.py:45 INFO TRAIN_STEP step=11707 loss=1.252654 Training: 65%|██████▍ | 11707/18018 [25:23:07<15:30:55, 8.85s/step, loss=1.253, acc=0.712, acc_len=1.72, thru=16.6, I=16.1, T=33.8, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:21,611] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 14:53:27,966] loop.py:45 INFO TRAIN_STEP step=11708 loss=1.347448 Training: 65%|██████▍ | 11708/18018 [25:23:14<14:21:24, 8.19s/step, loss=1.347, acc=0.691, acc_len=1.65, thru=18.0, I=21.4, T=34.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:27,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:28,445] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:33,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:53:35,454] loop.py:45 INFO TRAIN_STEP step=11709 loss=1.287662 Training: 65%|██████▍ | 11709/18018 [25:23:21<13:59:10, 7.98s/step, loss=1.288, acc=0.701, acc_len=1.69, thru=17.4, I=20.4, T=31.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:35,796] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:53:41,196] loop.py:380 INFO TIMING step=11710: step=3.326s data=1.311s compute=3.302s [fwd=2.369s bwd=0.919s opt=0.003s] dispatch=2.406s [2026-10-19 14:53:41,196] loop.py:45 INFO TRAIN_STEP step=11710 loss=1.245335 Training: 65%|██████▍ | 11710/18018 [25:23:27<12:48:25, 7.31s/step, loss=1.245, acc=0.708, acc_len=1.71, thru=19.2, I=25.5, T=38.5, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:50,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:53:51,288] loop.py:45 INFO TRAIN_STEP step=11711 loss=1.324145 Training: 65%|██████▍ | 11711/18018 [25:23:37<14:16:00, 8.14s/step, loss=1.324, acc=0.699, acc_len=1.68, thru=13.7, I=14.8, T=32.7, wait=6.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:53:51,284] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11710: forward=2750.4ms backward=1121.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:51,576] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:56,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:53:57,326] loop.py:45 INFO TRAIN_STEP step=11712 loss=1.276814 Training: 65%|██████▌ | 11712/18018 [25:23:43<13:09:27, 7.51s/step, loss=1.277, acc=0.705, acc_len=1.70, thru=19.6, I=21.2, T=38.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:53:57,371] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:54:05,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:54:05,387] loop.py:45 INFO TRAIN_STEP step=11713 loss=1.281438 Training: 65%|██████▌ | 11713/18018 [25:23:51<13:26:41, 7.68s/step, loss=1.281, acc=0.707, acc_len=1.70, thru=17.3, I=18.9, T=33.4, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:54:05,659] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:54:11,673] loop.py:45 INFO TRAIN_STEP step=11714 loss=1.347623 Training: 65%|██████▌ | 11714/18018 [25:23:57<12:42:43, 7.26s/step, loss=1.348, acc=0.691, acc_len=1.65, thru=19.3, I=21.4, T=33.2, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:54:17,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:54:18,916] loop.py:380 INFO TIMING step=11715: step=3.679s data=1.570s compute=3.647s [fwd=2.607s bwd=1.026s opt=0.003s] dispatch=3.556s [2026-10-19 14:54:18,916] loop.py:45 INFO TRAIN_STEP step=11715 loss=1.325140 Training: 65%|██████▌ | 11715/18018 [25:24:05<12:42:08, 7.26s/step, loss=1.325, acc=0.698, acc_len=1.68, thru=16.5, I=19.1, T=34.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:54:19,020] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:54:24,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:54:26,196] loop.py:45 INFO TRAIN_STEP step=11716 loss=1.348326 Training: 65%|██████▌ | 11716/18018 [25:24:12<12:42:47, 7.26s/step, loss=1.348, acc=0.695, acc_len=1.66, thru=16.8, I=22.4, T=34.8, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:54:26,191] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11715: forward=2613.7ms backward=1027.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:54:26,549] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:54:33,065] loop.py:45 INFO TRAIN_STEP step=11717 loss=1.219688 Training: 65%|██████▌ | 11717/18018 [25:24:19<12:30:18, 7.14s/step, loss=1.220, acc=0.712, acc_len=1.73, thru=16.5, I=21.0, T=33.4, wait=3.0s, pool=49, epoch=2/2][2026-10-19 14:54:41,016] loop.py:45 INFO TRAIN_STEP step=11718 loss=1.340301 Training: 65%|██████▌ | 11718/18018 [25:24:27<12:55:37, 7.39s/step, loss=1.340, acc=0.690, acc_len=1.64, thru=16.9, I=16.2, T=33.2, wait=4.1s, pool=53, epoch=2/2][2026-10-19 14:54:50,086] loop.py:45 INFO TRAIN_STEP step=11719 loss=1.362860 Training: 65%|██████▌ | 11719/18018 [25:24:36<13:48:25, 7.89s/step, loss=1.363, acc=0.691, acc_len=1.65, thru=15.1, I=14.3, T=34.7, wait=5.4s, pool=61, epoch=2/2][2026-10-19 14:54:57,863] loop.py:380 INFO TIMING step=11720: step=3.609s data=1.553s compute=3.575s [fwd=2.556s bwd=1.004s opt=0.003s] dispatch=4.159s [2026-10-19 14:54:57,863] loop.py:45 INFO TRAIN_STEP step=11720 loss=1.370437 Training: 65%|██████▌ | 11720/18018 [25:24:43<13:44:39, 7.86s/step, loss=1.370, acc=0.687, acc_len=1.64, thru=13.5, I=14.1, T=35.5, wait=4.2s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:03,706] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:55:05,081] loop.py:45 INFO TRAIN_STEP step=11721 loss=1.234578 Training: 65%|██████▌ | 11721/18018 [25:24:51<13:24:28, 7.67s/step, loss=1.235, acc=0.712, acc_len=1.72, thru=19.2, I=18.8, T=36.6, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:55:05,074] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11720: forward=2453.0ms backward=1004.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:05,221] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 14:55:12,455] loop.py:45 INFO TRAIN_STEP step=11722 loss=1.292856 Training: 65%|██████▌ | 11722/18018 [25:24:58<13:15:09, 7.58s/step, loss=1.293, acc=0.700, acc_len=1.68, thru=15.2, I=17.7, T=36.1, wait=3.8s, pool=53, epoch=2/2][2026-10-19 14:55:18,664] loop.py:45 INFO TRAIN_STEP step=11723 loss=1.283591 Training: 65%|██████▌ | 11723/18018 [25:25:04<12:31:55, 7.17s/step, loss=1.284, acc=0.707, acc_len=1.70, thru=19.6, I=19.1, T=35.1, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:27,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:55:28,121] loop.py:45 INFO TRAIN_STEP step=11724 loss=1.305968 Training: 65%|██████▌ | 11724/18018 [25:25:14<13:43:51, 7.85s/step, loss=1.306, acc=0.703, acc_len=1.69, thru=14.7, I=15.0, T=31.6, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:28,569] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:55:35,614] loop.py:380 INFO TIMING step=11725: step=4.088s data=1.821s compute=4.060s [fwd=2.835s bwd=1.211s opt=0.003s] dispatch=3.398s [2026-10-19 14:55:35,615] loop.py:45 INFO TRAIN_STEP step=11725 loss=1.431597 Training: 65%|██████▌ | 11725/18018 [25:25:21<13:32:30, 7.75s/step, loss=1.432, acc=0.675, acc_len=1.59, thru=16.6, I=20.1, T=31.3, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:43,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:55:44,053] loop.py:45 INFO TRAIN_STEP step=11726 loss=1.348268 Training: 65%|██████▌ | 11726/18018 [25:25:30<13:54:06, 7.95s/step, loss=1.348, acc=0.693, acc_len=1.65, thru=15.1, I=15.3, T=31.0, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:55:44,048] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11725: forward=2918.3ms backward=1156.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:44,262] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:49,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:55:49,970] loop.py:45 INFO TRAIN_STEP step=11727 loss=1.154773 Training: 65%|██████▌ | 11727/18018 [25:25:36<12:49:50, 7.34s/step, loss=1.155, acc=0.724, acc_len=1.77, thru=19.3, I=20.9, T=36.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:50,221] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:54,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:55:56,421] loop.py:45 INFO TRAIN_STEP step=11728 loss=1.259346 Training: 65%|██████▌ | 11728/18018 [25:25:42<12:21:42, 7.08s/step, loss=1.259, acc=0.708, acc_len=1.71, thru=19.6, I=26.5, T=32.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:55:56,699] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:01,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:56:02,657] loop.py:45 INFO TRAIN_STEP step=11729 loss=1.349569 Training: 65%|██████▌ | 11729/18018 [25:25:48<11:55:11, 6.82s/step, loss=1.350, acc=0.694, acc_len=1.66, thru=18.8, I=25.5, T=38.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:02,813] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:56:10,656] loop.py:380 INFO TIMING step=11730: step=3.686s data=1.609s compute=3.646s [fwd=2.621s bwd=1.011s opt=0.003s] dispatch=4.304s [2026-10-19 14:56:10,656] loop.py:45 INFO TRAIN_STEP step=11730 loss=1.317860 Training: 65%|██████▌ | 11730/18018 [25:25:56<12:32:03, 7.18s/step, loss=1.318, acc=0.699, acc_len=1.68, thru=14.3, I=15.6, T=34.7, wait=4.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:16,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:56:18,208] loop.py:45 INFO TRAIN_STEP step=11731 loss=1.478711 Training: 65%|██████▌ | 11731/18018 [25:26:04<12:43:44, 7.29s/step, loss=1.479, acc=0.668, acc_len=1.57, thru=17.6, I=19.4, T=34.4, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:56:18,203] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11730: forward=2616.0ms backward=1048.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:18,482] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:24,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:56:26,104] loop.py:45 INFO TRAIN_STEP step=11732 loss=1.333025 Training: 65%|██████▌ | 11732/18018 [25:26:12<13:02:46, 7.47s/step, loss=1.333, acc=0.695, acc_len=1.66, thru=14.8, I=20.5, T=31.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:26,366] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 14:56:32,851] loop.py:45 INFO TRAIN_STEP step=11733 loss=1.239521 Training: 65%|██████▌ | 11733/18018 [25:26:18<12:39:52, 7.25s/step, loss=1.240, acc=0.711, acc_len=1.72, thru=16.2, I=22.1, T=32.2, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:40,797] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:56:41,690] loop.py:45 INFO TRAIN_STEP step=11734 loss=1.307263 Training: 65%|██████▌ | 11734/18018 [25:26:27<13:29:31, 7.73s/step, loss=1.307, acc=0.699, acc_len=1.67, thru=15.1, I=15.5, T=31.2, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:41,818] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:48,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:56:48,246] loop.py:380 INFO TIMING step=11735: step=3.536s data=1.453s compute=3.502s [fwd=2.499s bwd=0.987s opt=0.003s] dispatch=3.010s [2026-10-19 14:56:48,246] loop.py:45 INFO TRAIN_STEP step=11735 loss=1.353213 Training: 65%|██████▌ | 11735/18018 [25:26:34<12:52:31, 7.38s/step, loss=1.353, acc=0.689, acc_len=1.64, thru=19.3, I=18.8, T=36.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:56:48,539] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:56:55,481] loop.py:45 INFO TRAIN_STEP step=11736 loss=1.286685 Training: 65%|██████▌ | 11736/18018 [25:26:41<12:47:52, 7.33s/step, loss=1.287, acc=0.706, acc_len=1.70, thru=15.8, I=17.1, T=31.9, wait=3.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:56:55,476] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11735: forward=2756.1ms backward=1216.3ms [2026-10-19 14:57:02,444] loop.py:45 INFO TRAIN_STEP step=11737 loss=1.287886 Training: 65%|██████▌ | 11737/18018 [25:26:48<12:36:04, 7.22s/step, loss=1.288, acc=0.705, acc_len=1.70, thru=16.1, I=14.8, T=34.1, wait=3.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:11,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:57:12,149] loop.py:45 INFO TRAIN_STEP step=11738 loss=1.337674 Training: 65%|██████▌ | 11738/18018 [25:26:58<13:54:02, 7.97s/step, loss=1.338, acc=0.692, acc_len=1.65, thru=16.8, I=15.4, T=30.5, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:12,304] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:17,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:57:18,713] loop.py:45 INFO TRAIN_STEP step=11739 loss=1.369192 Training: 65%|██████▌ | 11739/18018 [25:27:04<13:09:43, 7.55s/step, loss=1.369, acc=0.686, acc_len=1.63, thru=17.6, I=20.5, T=34.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:19,041] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:24,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:57:24,989] loop.py:380 INFO TIMING step=11740: step=3.589s data=1.612s compute=3.561s [fwd=2.591s bwd=0.956s opt=0.003s] dispatch=2.679s [2026-10-19 14:57:24,990] loop.py:45 INFO TRAIN_STEP step=11740 loss=1.263823 Training: 65%|██████▌ | 11740/18018 [25:27:11<12:29:44, 7.17s/step, loss=1.264, acc=0.708, acc_len=1.71, thru=19.2, I=23.1, T=35.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:25,282] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:30,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:57:31,773] loop.py:45 INFO TRAIN_STEP step=11741 loss=1.193971 Training: 65%|██████▌ | 11741/18018 [25:27:17<12:17:37, 7.05s/step, loss=1.194, acc=0.719, acc_len=1.76, thru=19.0, I=20.6, T=35.3, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:57:31,768] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11740: forward=2525.6ms backward=1044.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:31,939] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:57:38,664] loop.py:45 INFO TRAIN_STEP step=11742 loss=1.294462 Training: 65%|██████▌ | 11742/18018 [25:27:24<12:12:27, 7.00s/step, loss=1.294, acc=0.703, acc_len=1.69, thru=18.4, I=18.6, T=30.6, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:44,377] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:57:44,854] loop.py:45 INFO TRAIN_STEP step=11743 loss=1.281690 Training: 65%|██████▌ | 11743/18018 [25:27:30<11:46:56, 6.76s/step, loss=1.282, acc=0.706, acc_len=1.70, thru=19.8, I=19.8, T=34.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:44,895] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:49,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:57:51,003] loop.py:45 INFO TRAIN_STEP step=11744 loss=1.124745 Training: 65%|██████▌ | 11744/18018 [25:27:37<11:27:37, 6.58s/step, loss=1.125, acc=0.734, acc_len=1.80, thru=20.5, I=24.4, T=34.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:51,315] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:57:59,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:58:00,404] loop.py:380 INFO TIMING step=11745: step=3.661s data=1.496s compute=3.629s [fwd=2.603s bwd=1.012s opt=0.003s] dispatch=5.732s [2026-10-19 14:58:00,404] loop.py:45 INFO TRAIN_STEP step=11745 loss=1.373820 Training: 65%|██████▌ | 11745/18018 [25:27:46<12:56:07, 7.42s/step, loss=1.374, acc=0.690, acc_len=1.64, thru=14.1, I=14.8, T=35.0, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:00,624] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 14:58:10,526] loop.py:45 INFO TRAIN_STEP step=11746 loss=1.401488 Training: 65%|██████▌ | 11746/18018 [25:27:56<14:20:38, 8.23s/step, loss=1.401, acc=0.685, acc_len=1.63, thru=9.3, I=8.6, T=32.0, wait=6.1s, pool=29, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 14:58:10,521] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11745: forward=2833.7ms backward=1113.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:16,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:58:18,159] loop.py:45 INFO TRAIN_STEP step=11747 loss=1.355557 Training: 65%|██████▌ | 11747/18018 [25:28:04<14:01:43, 8.05s/step, loss=1.356, acc=0.695, acc_len=1.67, thru=19.3, I=19.1, T=33.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:18,322] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:24,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:58:25,685] loop.py:45 INFO TRAIN_STEP step=11748 loss=1.329683 Training: 65%|██████▌ | 11748/18018 [25:28:11<13:45:03, 7.90s/step, loss=1.330, acc=0.695, acc_len=1.67, thru=16.6, I=20.7, T=36.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:25,999] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:32,652] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:58:32,730] loop.py:45 INFO TRAIN_STEP step=11749 loss=1.273981 Training: 65%|██████▌ | 11749/18018 [25:28:18<13:18:09, 7.64s/step, loss=1.274, acc=0.703, acc_len=1.69, thru=18.4, I=18.7, T=33.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:33,154] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:38,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:58:40,512] loop.py:380 INFO TIMING step=11750: step=4.116s data=1.881s compute=4.073s [fwd=2.891s bwd=1.168s opt=0.003s] dispatch=3.659s [2026-10-19 14:58:40,512] loop.py:45 INFO TRAIN_STEP step=11750 loss=1.269016 Training: 65%|██████▌ | 11750/18018 [25:28:26<13:22:33, 7.68s/step, loss=1.269, acc=0.707, acc_len=1.70, thru=18.1, I=19.7, T=31.1, wait=3.7s, pool=65, epoch=2/2][2026-10-19 14:58:40,515] loop.py:419 INFO Saving checkpoint at step 11750... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 14:58:43,957] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0011751 (TrainerActor pid=9310) [2026-10-19 14:58:43,966] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0011001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:43,998] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:58:45,508] inference_manager.py:408 INFO Pool capacity available after 6.5s, pool_size=1/64 [2026-10-19 14:58:51,271] loop.py:45 INFO TRAIN_STEP step=11751 loss=1.294572 Training: 65%|██████▌ | 11751/18018 [25:28:37<14:58:50, 8.61s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=20.1, I=24.1, T=38.2, wait=2.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:58:51,266] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11750: forward=2357.9ms backward=946.9ms (TrainerActor pid=9310) [2026-10-19 14:58:43,956] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0011751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 14:58:45,077] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0011001 [2026-10-19 14:58:59,841] loop.py:45 INFO TRAIN_STEP step=11752 loss=1.378015 Training: 65%|██████▌ | 11752/18018 [25:28:45<14:57:40, 8.60s/step, loss=1.378, acc=0.684, acc_len=1.63, thru=14.4, I=13.7, T=36.5, wait=5.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:06,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:59:06,403] loop.py:45 INFO TRAIN_STEP step=11753 loss=1.296331 Training: 65%|██████▌ | 11753/18018 [25:28:52<13:53:44, 7.98s/step, loss=1.296, acc=0.700, acc_len=1.69, thru=19.6, I=20.2, T=33.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:06,687] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:13,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:59:14,074] loop.py:45 INFO TRAIN_STEP step=11754 loss=1.225231 Training: 65%|██████▌ | 11754/18018 [25:29:00<13:43:47, 7.89s/step, loss=1.225, acc=0.713, acc_len=1.73, thru=17.4, I=19.0, T=34.2, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:14,544] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:21,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:59:23,561] loop.py:380 INFO TIMING step=11755: step=4.350s data=1.977s compute=4.317s [fwd=3.073s bwd=1.230s opt=0.003s] dispatch=5.128s [2026-10-19 14:59:23,561] loop.py:45 INFO TRAIN_STEP step=11755 loss=1.247971 Training: 65%|██████▌ | 11755/18018 [25:29:09<14:33:38, 8.37s/step, loss=1.248, acc=0.710, acc_len=1.72, thru=14.2, I=15.8, T=29.4, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:23,592] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:29,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:59:30,005] loop.py:45 INFO TRAIN_STEP step=11756 loss=1.278866 Training: 65%|██████▌ | 11756/18018 [25:29:16<13:33:14, 7.79s/step, loss=1.279, acc=0.703, acc_len=1.70, thru=17.4, I=22.2, T=31.9, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 14:59:30,000] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11755: forward=2822.3ms backward=1143.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:30,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:35,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:59:36,646] loop.py:45 INFO TRAIN_STEP step=11757 loss=1.268557 Training: 65%|██████▌ | 11757/18018 [25:29:22<12:57:05, 7.45s/step, loss=1.269, acc=0.707, acc_len=1.71, thru=19.6, I=21.4, T=34.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:36,709] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 14:59:44,046] loop.py:45 INFO TRAIN_STEP step=11758 loss=1.343240 Training: 65%|██████▌ | 11758/18018 [25:29:30<12:55:30, 7.43s/step, loss=1.343, acc=0.693, acc_len=1.66, thru=18.0, I=19.7, T=35.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:44,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:44,548] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:52,615] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 14:59:52,750] loop.py:45 INFO TRAIN_STEP step=11759 loss=1.331670 Training: 65%|██████▌ | 11759/18018 [25:29:38<13:35:05, 7.81s/step, loss=1.332, acc=0.696, acc_len=1.67, thru=15.4, I=15.7, T=34.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 14:59:53,117] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 14:59:59,620] loop.py:380 INFO TIMING step=11760: step=3.508s data=1.456s compute=3.480s [fwd=2.510s bwd=0.956s opt=0.003s] dispatch=3.353s [2026-10-19 14:59:59,620] loop.py:45 INFO TRAIN_STEP step=11760 loss=1.292287 Training: 65%|██████▌ | 11760/18018 [25:29:45<13:05:29, 7.53s/step, loss=1.292, acc=0.702, acc_len=1.69, thru=17.7, I=19.0, T=36.5, wait=3.4s, pool=53, epoch=2/2][2026-10-19 15:00:09,042] loop.py:45 INFO TRAIN_STEP step=11761 loss=1.308019 Training: 65%|██████▌ | 11761/18018 [25:29:55<14:04:30, 8.10s/step, loss=1.308, acc=0.701, acc_len=1.68, thru=13.4, I=14.3, T=34.5, wait=5.7s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:00:09,038] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11760: forward=2644.3ms backward=1025.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:14,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:00:15,694] loop.py:45 INFO TRAIN_STEP step=11762 loss=1.325165 Training: 65%|██████▌ | 11762/18018 [25:30:01<13:19:08, 7.66s/step, loss=1.325, acc=0.697, acc_len=1.67, thru=19.0, I=18.8, T=35.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:15,869] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:21,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:00:21,693] loop.py:45 INFO TRAIN_STEP step=11763 loss=1.124908 Training: 65%|██████▌ | 11763/18018 [25:30:07<12:26:52, 7.16s/step, loss=1.125, acc=0.730, acc_len=1.78, thru=21.1, I=23.3, T=35.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:22,088] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:28,917] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:00:30,790] loop.py:45 INFO TRAIN_STEP step=11764 loss=1.349227 Training: 65%|██████▌ | 11764/18018 [25:30:16<13:27:09, 7.74s/step, loss=1.349, acc=0.691, acc_len=1.65, thru=14.7, I=17.5, T=29.5, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:30,935] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:36,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:00:37,166] loop.py:380 INFO TIMING step=11765: step=3.997s data=1.887s compute=3.955s [fwd=2.778s bwd=1.162s opt=0.003s] dispatch=2.371s [2026-10-19 15:00:37,166] loop.py:45 INFO TRAIN_STEP step=11765 loss=1.298458 Training: 65%|██████▌ | 11765/18018 [25:30:23<12:44:21, 7.33s/step, loss=1.298, acc=0.701, acc_len=1.68, thru=18.9, I=24.4, T=32.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:37,298] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:41,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:00:42,674] loop.py:45 INFO TRAIN_STEP step=11766 loss=1.297323 Training: 65%|██████▌ | 11766/18018 [25:30:28<11:47:11, 6.79s/step, loss=1.297, acc=0.700, acc_len=1.68, thru=20.6, I=23.6, T=34.0, wait=1.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:42,710] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 15:00:42,670] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11765: forward=2604.4ms backward=1095.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:47,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:00:48,669] loop.py:45 INFO TRAIN_STEP step=11767 loss=1.439739 Training: 65%|██████▌ | 11767/18018 [25:30:34<11:22:18, 6.55s/step, loss=1.440, acc=0.675, acc_len=1.60, thru=20.1, I=22.9, T=34.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:48,900] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:54,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:00:55,070] loop.py:45 INFO TRAIN_STEP step=11768 loss=1.191898 Training: 65%|██████▌ | 11768/18018 [25:30:41<11:17:33, 6.50s/step, loss=1.192, acc=0.718, acc_len=1.74, thru=19.9, I=22.5, T=36.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:00:55,421] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:01:03,845] loop.py:45 INFO TRAIN_STEP step=11769 loss=1.311091 Training: 65%|██████▌ | 11769/18018 [25:30:49<12:28:18, 7.18s/step, loss=1.311, acc=0.702, acc_len=1.69, thru=13.3, I=16.7, T=35.5, wait=5.2s, pool=53, epoch=2/2][2026-10-19 15:01:11,626] loop.py:380 INFO TIMING step=11770: step=3.701s data=1.584s compute=3.673s [fwd=2.603s bwd=1.056s opt=0.003s] dispatch=4.071s [2026-10-19 15:01:11,626] loop.py:45 INFO TRAIN_STEP step=11770 loss=1.270007 Training: 65%|██████▌ | 11770/18018 [25:30:57<12:46:50, 7.36s/step, loss=1.270, acc=0.706, acc_len=1.70, thru=15.3, I=17.2, T=34.6, wait=4.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:01:19,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:01:19,831] loop.py:45 INFO TRAIN_STEP step=11771 loss=1.274797 Training: 65%|██████▌ | 11771/18018 [25:31:05<13:12:58, 7.62s/step, loss=1.275, acc=0.707, acc_len=1.71, thru=15.7, I=16.6, T=33.5, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:01:19,826] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11770: forward=2734.2ms backward=1028.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:01:20,066] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:01:25,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:01:26,289] loop.py:45 INFO TRAIN_STEP step=11772 loss=1.303300 Training: 65%|██████▌ | 11772/18018 [25:31:12<12:36:43, 7.27s/step, loss=1.303, acc=0.703, acc_len=1.69, thru=19.9, I=18.7, T=33.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:01:26,455] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:01:33,077] loop.py:45 INFO TRAIN_STEP step=11773 loss=1.358458 Training: 65%|██████▌ | 11773/18018 [25:31:19<12:21:34, 7.12s/step, loss=1.358, acc=0.690, acc_len=1.64, thru=17.4, I=17.8, T=35.1, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:01:39,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:01:41,541] loop.py:45 INFO TRAIN_STEP step=11774 loss=1.340863 Training: 65%|██████▌ | 11774/18018 [25:31:27<13:03:16, 7.53s/step, loss=1.341, acc=0.694, acc_len=1.66, thru=16.5, I=18.5, T=31.8, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:01:41,853] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:01:47,595] loop.py:380 INFO TIMING step=11775: step=3.520s data=1.532s compute=3.483s [fwd=2.490s bwd=0.979s opt=0.003s] dispatch=2.524s [2026-10-19 15:01:47,596] loop.py:45 INFO TRAIN_STEP step=11775 loss=1.294966 Training: 65%|██████▌ | 11775/18018 [25:31:33<12:17:08, 7.08s/step, loss=1.295, acc=0.701, acc_len=1.69, thru=18.7, I=24.9, T=36.4, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:01:53,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:01:54,915] loop.py:45 INFO TRAIN_STEP step=11776 loss=1.248312 Training: 65%|██████▌ | 11776/18018 [25:31:41<12:24:24, 7.16s/step, loss=1.248, acc=0.711, acc_len=1.72, thru=17.2, I=19.9, T=34.2, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:01:54,910] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11775: forward=2624.5ms backward=1060.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:01:55,197] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:02:01,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:02:02,622] loop.py:45 INFO TRAIN_STEP step=11777 loss=1.336470 Training: 65%|██████▌ | 11777/18018 [25:31:48<12:41:29, 7.32s/step, loss=1.336, acc=0.695, acc_len=1.66, thru=16.5, I=22.5, T=35.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:02:02,674] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 15:02:08,076] loop.py:45 INFO TRAIN_STEP step=11778 loss=1.309451 Training: 65%|██████▌ | 11778/18018 [25:31:54<11:43:11, 6.76s/step, loss=1.309, acc=0.699, acc_len=1.67, thru=19.6, I=21.6, T=39.5, wait=2.2s, pool=45, epoch=2/2][2026-10-19 15:02:15,132] loop.py:45 INFO TRAIN_STEP step=11779 loss=1.283525 Training: 65%|██████▌ | 11779/18018 [25:32:01<11:52:10, 6.85s/step, loss=1.284, acc=0.706, acc_len=1.70, thru=17.5, I=18.6, T=35.2, wait=3.4s, pool=57, epoch=2/2][2026-10-19 15:02:23,087] loop.py:380 INFO TIMING step=11780: step=3.876s data=1.720s compute=3.847s [fwd=2.810s bwd=1.022s opt=0.003s] dispatch=4.070s [2026-10-19 15:02:23,087] loop.py:45 INFO TRAIN_STEP step=11780 loss=1.388526 Training: 65%|██████▌ | 11780/18018 [25:32:09<12:26:40, 7.18s/step, loss=1.389, acc=0.686, acc_len=1.64, thru=16.8, I=17.0, T=33.0, wait=4.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:02:29,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:02:30,085] loop.py:45 INFO TRAIN_STEP step=11781 loss=1.347827 Training: 65%|██████▌ | 11781/18018 [25:32:16<12:20:45, 7.13s/step, loss=1.348, acc=0.688, acc_len=1.64, thru=17.4, I=19.9, T=32.9, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:02:30,081] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11780: forward=2781.0ms backward=1063.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:02:30,233] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 15:02:36,826] loop.py:45 INFO TRAIN_STEP step=11782 loss=1.205241 Training: 65%|██████▌ | 11782/18018 [25:32:22<12:08:33, 7.01s/step, loss=1.205, acc=0.717, acc_len=1.73, thru=17.7, I=18.8, T=33.3, wait=2.9s, pool=57, epoch=2/2][2026-10-19 15:02:42,927] loop.py:45 INFO TRAIN_STEP step=11783 loss=1.270775 Training: 65%|██████▌ | 11783/18018 [25:32:29<11:40:09, 6.74s/step, loss=1.271, acc=0.703, acc_len=1.68, thru=17.2, I=19.8, T=37.3, wait=2.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:02:54,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:02:54,545] loop.py:45 INFO TRAIN_STEP step=11784 loss=1.421732 Training: 65%|██████▌ | 11784/18018 [25:32:40<14:12:05, 8.20s/step, loss=1.422, acc=0.682, acc_len=1.62, thru=13.9, I=13.2, T=31.5, wait=7.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:02:54,849] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:03:03,518] loop.py:380 INFO TIMING step=11785: step=3.835s data=1.596s compute=3.796s [fwd=2.675s bwd=1.106s opt=0.003s] dispatch=5.130s [2026-10-19 15:03:03,518] loop.py:45 INFO TRAIN_STEP step=11785 loss=1.201158 Training: 65%|██████▌ | 11785/18018 [25:32:49<14:36:02, 8.43s/step, loss=1.201, acc=0.721, acc_len=1.76, thru=13.3, I=14.3, T=33.4, wait=5.1s, pool=49, epoch=2/2][2026-10-19 15:03:15,257] loop.py:45 INFO TRAIN_STEP step=11786 loss=1.226960 Training: 65%|██████▌ | 11786/18018 [25:33:01<16:18:57, 9.43s/step, loss=1.227, acc=0.713, acc_len=1.73, thru=12.9, I=10.6, T=32.5, wait=7.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:03:15,252] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11785: forward=2760.5ms backward=1134.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:03:23,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:03:24,217] loop.py:45 INFO TRAIN_STEP step=11787 loss=1.413624 Training: 65%|██████▌ | 11787/18018 [25:33:10<16:04:17, 9.29s/step, loss=1.414, acc=0.678, acc_len=1.60, thru=15.4, I=16.1, T=33.1, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:03:24,459] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:03:30,944] loop.py:45 INFO TRAIN_STEP step=11788 loss=1.337668 Training: 65%|██████▌ | 11788/18018 [25:33:17<14:44:25, 8.52s/step, loss=1.338, acc=0.690, acc_len=1.65, thru=16.1, I=18.3, T=34.4, wait=3.0s, pool=49, epoch=2/2][2026-10-19 15:03:38,711] loop.py:45 INFO TRAIN_STEP step=11789 loss=1.292647 Training: 65%|██████▌ | 11789/18018 [25:33:24<14:20:57, 8.29s/step, loss=1.293, acc=0.698, acc_len=1.67, thru=17.3, I=17.9, T=29.7, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:03:46,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:03:47,871] loop.py:380 INFO TIMING step=11790: step=4.167s data=1.828s compute=4.138s [fwd=3.018s bwd=1.105s opt=0.003s] dispatch=4.983s [2026-10-19 15:03:47,871] loop.py:45 INFO TRAIN_STEP step=11790 loss=1.319129 Training: 65%|██████▌ | 11790/18018 [25:33:33<14:47:45, 8.55s/step, loss=1.319, acc=0.702, acc_len=1.69, thru=14.7, I=15.5, T=30.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:03:48,017] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:03:52,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:03:53,594] loop.py:45 INFO TRAIN_STEP step=11791 loss=1.205491 Training: 65%|██████▌ | 11791/18018 [25:33:39<13:19:34, 7.70s/step, loss=1.205, acc=0.718, acc_len=1.74, thru=20.1, I=24.4, T=34.9, wait=2.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:03:53,590] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11790: forward=2609.1ms backward=1020.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:03:53,793] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:03:59,766] loop.py:45 INFO TRAIN_STEP step=11792 loss=1.233598 Training: 65%|██████▌ | 11792/18018 [25:33:45<12:31:42, 7.24s/step, loss=1.234, acc=0.713, acc_len=1.73, thru=19.8, I=22.5, T=38.1, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:05,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:04:06,549] loop.py:45 INFO TRAIN_STEP step=11793 loss=1.344050 Training: 65%|██████▌ | 11793/18018 [25:33:52<12:17:10, 7.11s/step, loss=1.344, acc=0.695, acc_len=1.66, thru=19.2, I=20.0, T=36.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:06,937] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:04:13,167] loop.py:45 INFO TRAIN_STEP step=11794 loss=1.338685 Training: 65%|██████▌ | 11794/18018 [25:33:59<12:01:58, 6.96s/step, loss=1.339, acc=0.694, acc_len=1.65, thru=18.2, I=20.9, T=37.1, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:20,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:04:20,998] loop.py:380 INFO TIMING step=11795: step=3.681s data=1.586s compute=3.646s [fwd=2.524s bwd=1.107s opt=0.003s] dispatch=4.140s [2026-10-19 15:04:20,998] loop.py:45 INFO TRAIN_STEP step=11795 loss=1.226960 Training: 65%|██████▌ | 11795/18018 [25:34:07<12:28:58, 7.22s/step, loss=1.227, acc=0.717, acc_len=1.73, thru=16.9, I=17.2, T=34.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:21,200] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:26,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:04:27,037] loop.py:45 INFO TRAIN_STEP step=11796 loss=1.312472 Training: 65%|██████▌ | 11796/18018 [25:34:13<11:52:00, 6.87s/step, loss=1.312, acc=0.698, acc_len=1.67, thru=20.1, I=21.9, T=33.5, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:04:27,032] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11795: forward=2715.7ms backward=1058.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:27,163] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 15:04:32,895] loop.py:45 INFO TRAIN_STEP step=11797 loss=1.233921 Training: 65%|██████▌ | 11797/18018 [25:34:18<11:20:35, 6.56s/step, loss=1.234, acc=0.708, acc_len=1.71, thru=19.3, I=21.6, T=38.5, wait=2.5s, pool=53, epoch=2/2][2026-10-19 15:04:40,904] loop.py:45 INFO TRAIN_STEP step=11798 loss=1.380214 Training: 65%|██████▌ | 11798/18018 [25:34:26<12:05:29, 7.00s/step, loss=1.380, acc=0.687, acc_len=1.64, thru=15.7, I=17.6, T=33.9, wait=4.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:46,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:04:48,123] loop.py:45 INFO TRAIN_STEP step=11799 loss=1.327873 Training: 65%|██████▌ | 11799/18018 [25:34:34<12:12:06, 7.06s/step, loss=1.328, acc=0.697, acc_len=1.67, thru=17.0, I=19.3, T=30.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:48,289] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:53,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:04:54,364] loop.py:380 INFO TIMING step=11800: step=3.438s data=1.437s compute=3.389s [fwd=2.352s bwd=1.023s opt=0.003s] dispatch=2.794s [2026-10-19 15:04:54,364] loop.py:45 INFO TRAIN_STEP step=11800 loss=1.292736 Training: 65%|██████▌ | 11800/18018 [25:34:40<11:46:28, 6.82s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=18.1, I=22.8, T=37.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:04:54,838] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:05:00,243] loop.py:45 INFO TRAIN_STEP step=11801 loss=1.290234 Training: 65%|██████▌ | 11801/18018 [25:34:46<11:17:07, 6.53s/step, loss=1.290, acc=0.703, acc_len=1.68, thru=19.9, I=22.2, T=40.0, wait=2.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:05:00,238] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11800: forward=2270.3ms backward=881.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:08,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:05:09,554] loop.py:45 INFO TRAIN_STEP step=11802 loss=1.379797 Training: 66%|██████▌ | 11802/18018 [25:34:55<12:43:24, 7.37s/step, loss=1.380, acc=0.688, acc_len=1.64, thru=15.7, I=16.0, T=33.5, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:09,791] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:05:16,697] loop.py:45 INFO TRAIN_STEP step=11803 loss=1.169692 Training: 66%|██████▌ | 11803/18018 [25:35:02<12:36:17, 7.30s/step, loss=1.170, acc=0.722, acc_len=1.76, thru=17.1, I=18.6, T=30.4, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:23,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:05:25,559] loop.py:45 INFO TRAIN_STEP step=11804 loss=1.294582 Training: 66%|██████▌ | 11804/18018 [25:35:11<13:24:37, 7.77s/step, loss=1.295, acc=0.698, acc_len=1.67, thru=15.9, I=18.9, T=28.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:25,893] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:32,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:05:32,244] loop.py:380 INFO TIMING step=11805: step=3.703s data=1.589s compute=3.675s [fwd=2.629s bwd=1.032s opt=0.003s] dispatch=2.972s [2026-10-19 15:05:32,244] loop.py:45 INFO TRAIN_STEP step=11805 loss=1.190766 Training: 66%|██████▌ | 11805/18018 [25:35:18<12:50:47, 7.44s/step, loss=1.191, acc=0.719, acc_len=1.75, thru=16.5, I=20.4, T=34.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:32,567] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:39,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:05:39,553] loop.py:45 INFO TRAIN_STEP step=11806 loss=1.284374 Training: 66%|██████▌ | 11806/18018 [25:35:25<12:46:27, 7.40s/step, loss=1.284, acc=0.703, acc_len=1.69, thru=17.4, I=16.9, T=35.5, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:05:39,548] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11805: forward=2498.5ms backward=1054.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:39,950] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:43,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:05:45,474] loop.py:45 INFO TRAIN_STEP step=11807 loss=1.244203 Training: 66%|██████▌ | 11807/18018 [25:35:31<12:00:22, 6.96s/step, loss=1.244, acc=0.707, acc_len=1.70, thru=20.9, I=25.1, T=36.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:05:45,955] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:05:51,455] loop.py:45 INFO TRAIN_STEP step=11808 loss=1.346353 Training: 66%|██████▌ | 11808/18018 [25:35:37<11:29:51, 6.67s/step, loss=1.346, acc=0.694, acc_len=1.66, thru=19.4, I=27.5, T=37.6, wait=2.6s, pool=49, epoch=2/2][2026-10-19 15:05:59,091] loop.py:45 INFO TRAIN_STEP step=11809 loss=1.187509 Training: 66%|██████▌ | 11809/18018 [25:35:45<11:59:51, 6.96s/step, loss=1.188, acc=0.719, acc_len=1.75, thru=15.4, I=16.8, T=34.2, wait=3.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:06:07,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:06:07,670] loop.py:380 INFO TIMING step=11810: step=3.758s data=1.496s compute=3.728s [fwd=2.590s bwd=1.124s opt=0.003s] dispatch=4.813s [2026-10-19 15:06:07,671] loop.py:45 INFO TRAIN_STEP step=11810 loss=1.334904 Training: 66%|██████▌ | 11810/18018 [25:35:53<12:50:09, 7.44s/step, loss=1.335, acc=0.695, acc_len=1.66, thru=17.2, I=16.6, T=34.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:06:07,726] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:06:14,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:06:14,705] loop.py:45 INFO TRAIN_STEP step=11811 loss=1.243974 Training: 66%|██████▌ | 11811/18018 [25:36:00<12:37:22, 7.32s/step, loss=1.244, acc=0.711, acc_len=1.72, thru=18.0, I=18.6, T=35.5, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:06:14,701] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11810: forward=2560.0ms backward=988.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:06:14,962] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:06:22,781] loop.py:45 INFO TRAIN_STEP step=11812 loss=1.237241 Training: 66%|██████▌ | 11812/18018 [25:36:08<13:00:35, 7.55s/step, loss=1.237, acc=0.711, acc_len=1.72, thru=15.4, I=15.3, T=32.7, wait=4.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:06:27,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:06:28,958] loop.py:45 INFO TRAIN_STEP step=11813 loss=1.326019 Training: 66%|██████▌ | 11813/18018 [25:36:15<12:18:02, 7.14s/step, loss=1.326, acc=0.694, acc_len=1.65, thru=19.8, I=20.6, T=37.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:06:29,018] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 15:06:35,062] loop.py:45 INFO TRAIN_STEP step=11814 loss=1.308388 Training: 66%|██████▌ | 11814/18018 [25:36:21<11:45:54, 6.83s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=18.0, I=22.0, T=36.8, wait=2.6s, pool=41, epoch=2/2][2026-10-19 15:06:42,422] loop.py:380 INFO TIMING step=11815: step=3.563s data=1.524s compute=3.530s [fwd=2.518s bwd=0.998s opt=0.003s] dispatch=3.786s [2026-10-19 15:06:42,422] loop.py:45 INFO TRAIN_STEP step=11815 loss=1.345217 Training: 66%|██████▌ | 11815/18018 [25:36:28<12:02:16, 6.99s/step, loss=1.345, acc=0.694, acc_len=1.67, thru=18.9, I=19.0, T=35.9, wait=3.8s, pool=61, epoch=2/2][2026-10-19 15:06:50,019] loop.py:45 INFO TRAIN_STEP step=11816 loss=1.303241 Training: 66%|██████▌ | 11816/18018 [25:36:36<12:21:03, 7.17s/step, loss=1.303, acc=0.700, acc_len=1.68, thru=17.0, I=16.9, T=35.4, wait=4.0s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:06:50,015] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11815: forward=2503.1ms backward=1047.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:06:56,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:06:57,639] loop.py:45 INFO TRAIN_STEP step=11817 loss=1.367878 Training: 66%|██████▌ | 11817/18018 [25:36:43<12:34:54, 7.30s/step, loss=1.368, acc=0.687, acc_len=1.64, thru=16.9, I=18.4, T=37.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:06:57,925] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:02,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:07:03,443] loop.py:45 INFO TRAIN_STEP step=11818 loss=1.240010 Training: 66%|██████▌ | 11818/18018 [25:36:49<11:48:20, 6.85s/step, loss=1.240, acc=0.710, acc_len=1.71, thru=20.7, I=23.3, T=39.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:03,938] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:08,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:07:09,605] loop.py:45 INFO TRAIN_STEP step=11819 loss=1.202029 Training: 66%|██████▌ | 11819/18018 [25:36:55<11:26:45, 6.65s/step, loss=1.202, acc=0.715, acc_len=1.73, thru=21.5, I=27.3, T=38.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:09,658] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:16,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:07:17,462] loop.py:380 INFO TIMING step=11820: step=3.558s data=1.522s compute=3.525s [fwd=2.511s bwd=1.000s opt=0.003s] dispatch=4.289s [2026-10-19 15:07:17,462] loop.py:45 INFO TRAIN_STEP step=11820 loss=1.233002 Training: 66%|██████▌ | 11820/18018 [25:37:03<12:04:05, 7.01s/step, loss=1.233, acc=0.711, acc_len=1.72, thru=15.5, I=18.0, T=36.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:17,496] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:23,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:07:24,272] loop.py:45 INFO TRAIN_STEP step=11821 loss=1.344341 Training: 66%|██████▌ | 11821/18018 [25:37:10<11:57:48, 6.95s/step, loss=1.344, acc=0.690, acc_len=1.64, thru=18.4, I=21.4, T=28.7, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:07:24,267] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11820: forward=3473.3ms backward=939.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:24,560] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:29,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:07:31,197] loop.py:45 INFO TRAIN_STEP step=11822 loss=1.293113 Training: 66%|██████▌ | 11822/18018 [25:37:17<11:56:55, 6.94s/step, loss=1.293, acc=0.700, acc_len=1.68, thru=17.3, I=22.6, T=31.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:31,456] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:07:38,153] loop.py:45 INFO TRAIN_STEP step=11823 loss=1.212800 Training: 66%|██████▌ | 11823/18018 [25:37:24<11:57:14, 6.95s/step, loss=1.213, acc=0.717, acc_len=1.74, thru=14.0, I=19.7, T=30.3, wait=2.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:44,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:07:46,021] loop.py:45 INFO TRAIN_STEP step=11824 loss=1.315080 Training: 66%|██████▌ | 11824/18018 [25:37:32<12:25:35, 7.22s/step, loss=1.315, acc=0.702, acc_len=1.69, thru=17.1, I=17.4, T=33.1, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:46,503] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:53,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:07:54,300] loop.py:380 INFO TIMING step=11825: step=3.569s data=1.496s compute=3.528s [fwd=2.493s bwd=1.022s opt=0.003s] dispatch=4.702s [2026-10-19 15:07:54,301] loop.py:45 INFO TRAIN_STEP step=11825 loss=1.361562 Training: 66%|██████▌ | 11825/18018 [25:37:40<12:58:13, 7.54s/step, loss=1.362, acc=0.684, acc_len=1.63, thru=13.6, I=19.1, T=35.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:07:54,383] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:03,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:08:03,900] loop.py:45 INFO TRAIN_STEP step=11826 loss=1.301119 Training: 66%|██████▌ | 11826/18018 [25:37:49<14:01:54, 8.16s/step, loss=1.301, acc=0.700, acc_len=1.68, thru=13.5, I=14.0, T=34.4, wait=5.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:08:03,896] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11825: forward=2588.4ms backward=1077.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:04,393] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:10,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:08:10,931] loop.py:45 INFO TRAIN_STEP step=11827 loss=1.222853 Training: 66%|██████▌ | 11827/18018 [25:37:57<13:26:53, 7.82s/step, loss=1.223, acc=0.712, acc_len=1.72, thru=17.7, I=18.6, T=35.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:11,032] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 15:08:19,374] loop.py:45 INFO TRAIN_STEP step=11828 loss=1.235364 Training: 66%|██████▌ | 11828/18018 [25:38:05<13:45:58, 8.01s/step, loss=1.235, acc=0.710, acc_len=1.71, thru=16.3, I=15.6, T=32.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:19,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:19,849] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:26,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:08:27,052] loop.py:45 INFO TRAIN_STEP step=11829 loss=1.267935 Training: 66%|██████▌ | 11829/18018 [25:38:13<13:35:47, 7.91s/step, loss=1.268, acc=0.705, acc_len=1.69, thru=17.0, I=17.7, T=28.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:27,210] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:08:34,613] loop.py:380 INFO TIMING step=11830: step=3.465s data=1.355s compute=3.442s [fwd=2.474s bwd=0.954s opt=0.003s] dispatch=4.085s [2026-10-19 15:08:34,613] loop.py:45 INFO TRAIN_STEP step=11830 loss=1.332685 Training: 66%|██████▌ | 11830/18018 [25:38:20<13:24:54, 7.80s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=14.8, I=15.4, T=36.9, wait=4.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:42,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:08:42,637] loop.py:45 INFO TRAIN_STEP step=11831 loss=1.274939 Training: 66%|██████▌ | 11831/18018 [25:38:28<13:31:28, 7.87s/step, loss=1.275, acc=0.705, acc_len=1.70, thru=18.0, I=18.1, T=32.6, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:08:42,633] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11830: forward=2765.5ms backward=1100.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:42,881] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:08:48,892] loop.py:45 INFO TRAIN_STEP step=11832 loss=1.321647 Training: 66%|██████▌ | 11832/18018 [25:38:34<12:41:26, 7.39s/step, loss=1.322, acc=0.700, acc_len=1.68, thru=18.1, I=20.3, T=36.0, wait=2.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:56,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:08:56,891] loop.py:45 INFO TRAIN_STEP step=11833 loss=1.316720 Training: 66%|██████▌ | 11833/18018 [25:38:42<13:00:15, 7.57s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=16.9, I=18.1, T=31.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:08:57,140] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:09:03,367] loop.py:45 INFO TRAIN_STEP step=11834 loss=1.166443 Training: 66%|██████▌ | 11834/18018 [25:38:49<12:26:23, 7.24s/step, loss=1.166, acc=0.727, acc_len=1.77, thru=15.8, I=17.0, T=37.9, wait=3.1s, pool=37, epoch=2/2][2026-10-19 15:09:11,844] loop.py:380 INFO TIMING step=11835: step=3.677s data=1.508s compute=3.646s [fwd=2.636s bwd=0.995s opt=0.003s] dispatch=4.789s [2026-10-19 15:09:11,844] loop.py:45 INFO TRAIN_STEP step=11835 loss=1.234226 Training: 66%|██████▌ | 11835/18018 [25:38:57<13:04:22, 7.61s/step, loss=1.234, acc=0.716, acc_len=1.74, thru=15.4, I=15.2, T=34.8, wait=4.8s, pool=41, epoch=2/2][2026-10-19 15:09:21,059] loop.py:45 INFO TRAIN_STEP step=11836 loss=1.311692 Training: 66%|██████▌ | 11836/18018 [25:39:07<13:53:50, 8.09s/step, loss=1.312, acc=0.696, acc_len=1.67, thru=14.9, I=16.1, T=33.8, wait=5.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:09:21,055] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11835: forward=2687.1ms backward=1062.7ms [2026-10-19 15:09:30,107] loop.py:45 INFO TRAIN_STEP step=11837 loss=1.219316 Training: 66%|██████▌ | 11837/18018 [25:39:16<14:23:12, 8.38s/step, loss=1.219, acc=0.717, acc_len=1.74, thru=13.7, I=12.3, T=34.4, wait=5.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:09:37,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:09:38,899] loop.py:45 INFO TRAIN_STEP step=11838 loss=1.368433 Training: 66%|██████▌ | 11838/18018 [25:39:24<14:35:53, 8.50s/step, loss=1.368, acc=0.687, acc_len=1.64, thru=17.2, I=16.8, T=35.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:09:39,269] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:09:45,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:09:45,555] loop.py:45 INFO TRAIN_STEP step=11839 loss=1.229715 Training: 66%|██████▌ | 11839/18018 [25:39:31<13:38:35, 7.95s/step, loss=1.230, acc=0.715, acc_len=1.74, thru=19.2, I=23.3, T=36.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:09:45,899] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:09:52,294] loop.py:380 INFO TIMING step=11840: step=3.518s data=1.514s compute=3.480s [fwd=2.486s bwd=0.980s opt=0.003s] dispatch=3.212s [2026-10-19 15:09:52,294] loop.py:45 INFO TRAIN_STEP step=11840 loss=1.206160 Training: 66%|██████▌ | 11840/18018 [25:39:38<13:01:08, 7.59s/step, loss=1.206, acc=0.715, acc_len=1.74, thru=17.5, I=19.0, T=36.4, wait=3.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:09:57,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:09:58,686] loop.py:45 INFO TRAIN_STEP step=11841 loss=1.248212 Training: 66%|██████▌ | 11841/18018 [25:39:44<12:24:07, 7.23s/step, loss=1.248, acc=0.716, acc_len=1.74, thru=19.7, I=22.1, T=36.2, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:09:58,681] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11840: forward=2476.5ms backward=986.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:09:58,775] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:10:04,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:10:04,395] loop.py:45 INFO TRAIN_STEP step=11842 loss=1.337545 Training: 66%|██████▌ | 11842/18018 [25:39:50<11:37:05, 6.77s/step, loss=1.338, acc=0.691, acc_len=1.64, thru=21.7, I=28.1, T=36.6, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:10:04,681] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:10:14,306] loop.py:45 INFO TRAIN_STEP step=11843 loss=1.451565 Training: 66%|██████▌ | 11843/18018 [25:40:00<13:13:51, 7.71s/step, loss=1.452, acc=0.676, acc_len=1.60, thru=11.8, I=10.8, T=32.1, wait=5.9s, pool=45, epoch=2/2][2026-10-19 15:10:23,943] loop.py:45 INFO TRAIN_STEP step=11844 loss=1.295879 Training: 66%|██████▌ | 11844/18018 [25:40:10<14:13:05, 8.29s/step, loss=1.296, acc=0.702, acc_len=1.69, thru=14.0, I=14.9, T=32.4, wait=5.7s, pool=49, epoch=2/2][2026-10-19 15:10:34,062] loop.py:380 INFO TIMING step=11845: step=3.836s data=1.631s compute=3.795s [fwd=2.668s bwd=1.112s opt=0.003s] dispatch=6.275s [2026-10-19 15:10:34,062] loop.py:45 INFO TRAIN_STEP step=11845 loss=1.327544 Training: 66%|██████▌ | 11845/18018 [25:40:20<15:09:26, 8.84s/step, loss=1.328, acc=0.693, acc_len=1.65, thru=12.8, I=12.4, T=33.4, wait=6.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:10:43,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:10:44,481] loop.py:45 INFO TRAIN_STEP step=11846 loss=1.396154 Training: 66%|██████▌ | 11846/18018 [25:40:30<15:58:02, 9.31s/step, loss=1.396, acc=0.682, acc_len=1.62, thru=15.0, I=14.3, T=32.1, wait=6.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:10:44,476] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11845: forward=2761.6ms backward=1168.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:10:44,799] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:10:51,101] loop.py:45 INFO TRAIN_STEP step=11847 loss=1.310070 Training: 66%|██████▌ | 11847/18018 [25:40:37<14:34:46, 8.51s/step, loss=1.310, acc=0.700, acc_len=1.68, thru=15.3, I=19.3, T=36.6, wait=3.1s, pool=41, epoch=2/2][2026-10-19 15:10:59,211] loop.py:45 INFO TRAIN_STEP step=11848 loss=1.418189 Training: 66%|██████▌ | 11848/18018 [25:40:45<14:22:26, 8.39s/step, loss=1.418, acc=0.679, acc_len=1.61, thru=17.4, I=17.6, T=34.3, wait=4.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:05,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:11:05,243] loop.py:45 INFO TRAIN_STEP step=11849 loss=1.256272 Training: 66%|██████▌ | 11849/18018 [25:40:51<13:09:39, 7.68s/step, loss=1.256, acc=0.709, acc_len=1.71, thru=20.0, I=20.0, T=36.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:05,547] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:13,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:11:14,451] loop.py:380 INFO TIMING step=11850: step=4.225s data=1.865s compute=4.192s [fwd=2.898s bwd=1.280s opt=0.003s] dispatch=4.973s [2026-10-19 15:11:14,451] loop.py:45 INFO TRAIN_STEP step=11850 loss=1.329165 Training: 66%|██████▌ | 11850/18018 [25:41:00<13:56:39, 8.14s/step, loss=1.329, acc=0.695, acc_len=1.66, thru=14.0, I=15.1, T=30.3, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:14,791] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:21,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:11:21,512] loop.py:45 INFO TRAIN_STEP step=11851 loss=1.344992 Training: 66%|██████▌ | 11851/18018 [25:41:07<13:23:17, 7.82s/step, loss=1.345, acc=0.693, acc_len=1.66, thru=16.4, I=19.0, T=31.4, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:11:21,508] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11850: forward=2840.9ms backward=1175.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:21,692] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:11:28,149] loop.py:45 INFO TRAIN_STEP step=11852 loss=1.349242 Training: 66%|██████▌ | 11852/18018 [25:41:14<12:46:50, 7.46s/step, loss=1.349, acc=0.690, acc_len=1.65, thru=16.1, I=19.9, T=33.2, wait=2.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:35,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:11:36,071] loop.py:45 INFO TRAIN_STEP step=11853 loss=1.338019 Training: 66%|██████▌ | 11853/18018 [25:41:22<13:00:50, 7.60s/step, loss=1.338, acc=0.694, acc_len=1.66, thru=16.2, I=16.6, T=35.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:36,482] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:42,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:11:43,609] loop.py:45 INFO TRAIN_STEP step=11854 loss=1.260955 Training: 66%|██████▌ | 11854/18018 [25:41:29<12:58:49, 7.58s/step, loss=1.261, acc=0.706, acc_len=1.70, thru=17.2, I=19.6, T=35.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:43,927] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:49,885] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:11:51,078] loop.py:380 INFO TIMING step=11855: step=3.555s data=1.520s compute=3.518s [fwd=2.548s bwd=0.955s opt=0.003s] dispatch=3.905s [2026-10-19 15:11:51,078] loop.py:45 INFO TRAIN_STEP step=11855 loss=1.327592 Training: 66%|██████▌ | 11855/18018 [25:41:37<12:55:13, 7.55s/step, loss=1.328, acc=0.697, acc_len=1.67, thru=16.5, I=23.4, T=36.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:51,394] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:56,173] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:11:57,138] loop.py:45 INFO TRAIN_STEP step=11856 loss=1.254256 Training: 66%|██████▌ | 11856/18018 [25:41:43<12:09:18, 7.10s/step, loss=1.254, acc=0.710, acc_len=1.72, thru=20.9, I=26.2, T=38.7, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:11:57,133] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11855: forward=2341.5ms backward=922.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:11:57,185] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:02,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:12:03,623] loop.py:45 INFO TRAIN_STEP step=11857 loss=1.199181 Training: 66%|██████▌ | 11857/18018 [25:41:49<11:50:11, 6.92s/step, loss=1.199, acc=0.717, acc_len=1.74, thru=20.3, I=24.9, T=37.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:03,750] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:11,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:12:12,637] loop.py:45 INFO TRAIN_STEP step=11858 loss=1.310248 Training: 66%|██████▌ | 11858/18018 [25:41:58<12:54:42, 7.55s/step, loss=1.310, acc=0.699, acc_len=1.68, thru=14.4, I=14.7, T=35.5, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:12,894] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:12:20,674] loop.py:45 INFO TRAIN_STEP step=11859 loss=1.319854 Training: 66%|██████▌ | 11859/18018 [25:42:06<13:09:44, 7.69s/step, loss=1.320, acc=0.695, acc_len=1.66, thru=12.8, I=15.7, T=34.3, wait=4.3s, pool=33, epoch=2/2][2026-10-19 15:12:28,713] loop.py:380 INFO TIMING step=11860: step=3.998s data=1.820s compute=3.958s [fwd=2.856s bwd=1.086s opt=0.003s] dispatch=4.029s [2026-10-19 15:12:28,713] loop.py:45 INFO TRAIN_STEP step=11860 loss=1.323162 Training: 66%|██████▌ | 11860/18018 [25:42:14<13:20:11, 7.80s/step, loss=1.323, acc=0.693, acc_len=1.65, thru=16.0, I=17.6, T=32.0, wait=4.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:36,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 15:12:37,194] loop.py:45 INFO TRAIN_STEP step=11861 loss=1.433194 Training: 66%|██████▌ | 11861/18018 [25:42:23<13:41:05, 8.00s/step, loss=1.433, acc=0.675, acc_len=1.60, thru=16.1, I=17.1, T=34.1, wait=4.7s, pool=69, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:12:37,190] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11860: forward=2619.8ms backward=1070.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:37,590] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:43,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:12:44,615] loop.py:45 INFO TRAIN_STEP step=11862 loss=1.264609 Training: 66%|██████▌ | 11862/18018 [25:42:30<13:23:08, 7.83s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=16.5, I=21.8, T=34.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:44,668] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 15:12:52,076] loop.py:45 INFO TRAIN_STEP step=11863 loss=1.192752 Training: 66%|██████▌ | 11863/18018 [25:42:38<13:11:44, 7.72s/step, loss=1.193, acc=0.721, acc_len=1.75, thru=15.6, I=19.3, T=35.6, wait=3.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:58,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:12:59,631] loop.py:45 INFO TRAIN_STEP step=11864 loss=1.382900 Training: 66%|██████▌ | 11864/18018 [25:42:45<13:06:34, 7.67s/step, loss=1.383, acc=0.685, acc_len=1.63, thru=18.0, I=18.6, T=34.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:12:59,799] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:13:06,292] loop.py:380 INFO TIMING step=11865: step=3.391s data=1.425s compute=3.356s [fwd=2.423s bwd=0.918s opt=0.003s] dispatch=3.261s [2026-10-19 15:13:06,292] loop.py:45 INFO TRAIN_STEP step=11865 loss=1.353407 Training: 66%|██████▌ | 11865/18018 [25:42:52<12:35:30, 7.37s/step, loss=1.353, acc=0.692, acc_len=1.65, thru=18.5, I=22.7, T=37.8, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:13,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:13:14,618] loop.py:45 INFO TRAIN_STEP step=11866 loss=1.289992 Training: 66%|██████▌ | 11866/18018 [25:43:00<13:04:51, 7.65s/step, loss=1.290, acc=0.700, acc_len=1.68, thru=15.2, I=16.5, T=32.0, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:13:14,613] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11865: forward=2858.9ms backward=1093.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:14,896] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:21,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:13:22,692] loop.py:45 INFO TRAIN_STEP step=11867 loss=1.352604 Training: 66%|██████▌ | 11867/18018 [25:43:08<13:17:38, 7.78s/step, loss=1.353, acc=0.691, acc_len=1.65, thru=15.4, I=20.2, T=30.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:23,058] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:13:29,428] loop.py:45 INFO TRAIN_STEP step=11868 loss=1.288785 Training: 66%|██████▌ | 11868/18018 [25:43:15<12:45:16, 7.47s/step, loss=1.289, acc=0.699, acc_len=1.68, thru=17.0, I=20.8, T=34.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:29,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:29,926] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:13:37,026] loop.py:45 INFO TRAIN_STEP step=11869 loss=1.259655 Training: 66%|██████▌ | 11869/18018 [25:43:23<12:49:16, 7.51s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=15.8, I=16.7, T=36.0, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:45,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:13:47,560] loop.py:380 INFO TIMING step=11870: step=4.067s data=1.923s compute=4.020s [fwd=2.860s bwd=1.145s opt=0.003s] dispatch=6.457s [2026-10-19 15:13:47,560] loop.py:45 INFO TRAIN_STEP step=11870 loss=1.359395 Training: 66%|██████▌ | 11870/18018 [25:43:33<14:22:11, 8.41s/step, loss=1.359, acc=0.694, acc_len=1.66, thru=14.9, I=15.9, T=31.5, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:47,651] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:53,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:13:54,890] loop.py:45 INFO TRAIN_STEP step=11871 loss=1.367047 Training: 66%|██████▌ | 11871/18018 [25:43:40<13:48:46, 8.09s/step, loss=1.367, acc=0.687, acc_len=1.63, thru=13.8, I=22.0, T=35.7, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:13:54,886] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11870: forward=2527.0ms backward=1009.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:13:55,323] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:14:01,186] loop.py:45 INFO TRAIN_STEP step=11872 loss=1.281733 Training: 66%|██████▌ | 11872/18018 [25:43:47<12:53:27, 7.55s/step, loss=1.282, acc=0.704, acc_len=1.70, thru=16.5, I=22.1, T=37.8, wait=2.9s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:12,587] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:14:13,373] loop.py:45 INFO TRAIN_STEP step=11873 loss=1.367676 Training: 66%|██████▌ | 11873/18018 [25:43:59<15:15:50, 8.94s/step, loss=1.368, acc=0.685, acc_len=1.62, thru=14.1, I=12.6, T=31.7, wait=8.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:13,594] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:21,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:14:21,658] loop.py:45 INFO TRAIN_STEP step=11874 loss=1.295276 Training: 66%|██████▌ | 11874/18018 [25:44:07<14:55:29, 8.75s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=14.9, I=16.8, T=34.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:21,837] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:14:28,476] loop.py:380 INFO TIMING step=11875: step=3.662s data=1.618s compute=3.625s [fwd=2.630s bwd=0.980s opt=0.003s] dispatch=3.145s [2026-10-19 15:14:28,476] loop.py:45 INFO TRAIN_STEP step=11875 loss=1.206321 Training: 66%|██████▌ | 11875/18018 [25:44:14<13:56:07, 8.17s/step, loss=1.206, acc=0.717, acc_len=1.73, thru=17.6, I=18.0, T=34.9, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:35,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:14:35,520] loop.py:45 INFO TRAIN_STEP step=11876 loss=1.275287 Training: 66%|██████▌ | 11876/18018 [25:44:21<13:21:33, 7.83s/step, loss=1.275, acc=0.704, acc_len=1.69, thru=18.2, I=18.8, T=32.5, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:14:35,515] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11875: forward=2797.9ms backward=1086.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:35,639] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:41,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:14:42,319] loop.py:45 INFO TRAIN_STEP step=11877 loss=1.326225 Training: 66%|██████▌ | 11877/18018 [25:44:28<12:49:41, 7.52s/step, loss=1.326, acc=0.695, acc_len=1.67, thru=18.1, I=19.6, T=35.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:42,463] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:48,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:14:49,551] loop.py:45 INFO TRAIN_STEP step=11878 loss=1.227921 Training: 66%|██████▌ | 11878/18018 [25:44:35<12:40:47, 7.43s/step, loss=1.228, acc=0.716, acc_len=1.74, thru=17.8, I=19.5, T=32.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:49,587] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:55,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:14:56,106] loop.py:45 INFO TRAIN_STEP step=11879 loss=1.385354 Training: 66%|██████▌ | 11879/18018 [25:44:42<12:13:33, 7.17s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=18.1, I=21.4, T=35.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:14:56,164] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:01,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:15:02,198] loop.py:380 INFO TIMING step=11880: step=3.401s data=1.410s compute=3.366s [fwd=2.410s bwd=0.942s opt=0.003s] dispatch=2.684s [2026-10-19 15:15:02,198] loop.py:45 INFO TRAIN_STEP step=11880 loss=1.304461 Training: 66%|██████▌ | 11880/18018 [25:44:48<11:40:24, 6.85s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=20.0, I=21.9, T=37.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:02,662] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:08,032] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:15:08,164] loop.py:45 INFO TRAIN_STEP step=11881 loss=1.261715 Training: 66%|██████▌ | 11881/18018 [25:44:54<11:13:18, 6.58s/step, loss=1.262, acc=0.707, acc_len=1.71, thru=20.5, I=21.9, T=37.1, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:15:08,159] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11880: forward=2447.8ms backward=954.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:08,532] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:15:14,316] loop.py:45 INFO TRAIN_STEP step=11882 loss=1.176890 Training: 66%|██████▌ | 11882/18018 [25:45:00<10:59:55, 6.45s/step, loss=1.177, acc=0.720, acc_len=1.75, thru=19.2, I=20.8, T=37.4, wait=2.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:22,002] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:15:22,749] loop.py:45 INFO TRAIN_STEP step=11883 loss=1.301545 Training: 66%|██████▌ | 11883/18018 [25:45:08<12:00:37, 7.05s/step, loss=1.302, acc=0.700, acc_len=1.68, thru=17.8, I=18.4, T=33.1, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:23,009] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:27,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:15:28,674] loop.py:45 INFO TRAIN_STEP step=11884 loss=1.215672 Training: 66%|██████▌ | 11884/18018 [25:45:14<11:26:01, 6.71s/step, loss=1.216, acc=0.717, acc_len=1.74, thru=19.9, I=21.6, T=37.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:28,992] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:15:35,737] loop.py:380 INFO TIMING step=11885: step=3.884s data=1.732s compute=3.854s [fwd=2.720s bwd=1.118s opt=0.003s] dispatch=3.170s [2026-10-19 15:15:35,737] loop.py:45 INFO TRAIN_STEP step=11885 loss=1.261755 Training: 66%|██████▌ | 11885/18018 [25:45:21<11:36:43, 6.82s/step, loss=1.262, acc=0.703, acc_len=1.70, thru=17.1, I=18.5, T=33.0, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:43,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:15:43,739] loop.py:45 INFO TRAIN_STEP step=11886 loss=1.303346 Training: 66%|██████▌ | 11886/18018 [25:45:29<12:12:58, 7.17s/step, loss=1.303, acc=0.702, acc_len=1.68, thru=17.9, I=18.0, T=34.6, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:15:43,734] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11885: forward=2525.5ms backward=1120.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:44,183] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:49,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:15:49,963] loop.py:45 INFO TRAIN_STEP step=11887 loss=1.255066 Training: 66%|██████▌ | 11887/18018 [25:45:36<11:43:47, 6.89s/step, loss=1.255, acc=0.709, acc_len=1.71, thru=19.7, I=22.5, T=39.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:15:50,188] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:15:56,274] loop.py:45 INFO TRAIN_STEP step=11888 loss=1.238604 Training: 66%|██████▌ | 11888/18018 [25:45:42<11:26:01, 6.71s/step, loss=1.239, acc=0.711, acc_len=1.72, thru=20.1, I=23.1, T=34.9, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:02,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:16:02,818] loop.py:45 INFO TRAIN_STEP step=11889 loss=1.326111 Training: 66%|██████▌ | 11889/18018 [25:45:48<11:20:42, 6.66s/step, loss=1.326, acc=0.694, acc_len=1.66, thru=19.6, I=19.4, T=37.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:03,230] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:07,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:16:08,822] loop.py:380 INFO TIMING step=11890: step=3.429s data=1.419s compute=3.407s [fwd=2.433s bwd=0.959s opt=0.003s] dispatch=2.566s [2026-10-19 15:16:08,823] loop.py:45 INFO TRAIN_STEP step=11890 loss=1.235878 Training: 66%|██████▌ | 11890/18018 [25:45:54<11:00:24, 6.47s/step, loss=1.236, acc=0.712, acc_len=1.72, thru=21.0, I=25.3, T=37.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:09,024] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:14,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:16:14,845] loop.py:45 INFO TRAIN_STEP step=11891 loss=1.282800 Training: 66%|██████▌ | 11891/18018 [25:46:00<10:46:41, 6.33s/step, loss=1.283, acc=0.701, acc_len=1.68, thru=20.9, I=24.5, T=37.9, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:16:14,840] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11890: forward=2351.8ms backward=968.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:15,096] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:21,008] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:16:21,390] loop.py:45 INFO TRAIN_STEP step=11892 loss=1.200461 Training: 66%|██████▌ | 11892/18018 [25:46:07<10:53:06, 6.40s/step, loss=1.200, acc=0.714, acc_len=1.73, thru=19.7, I=22.1, T=36.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:21,515] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:26,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:16:27,855] loop.py:45 INFO TRAIN_STEP step=11893 loss=1.246430 Training: 66%|██████▌ | 11893/18018 [25:46:13<10:55:05, 6.42s/step, loss=1.246, acc=0.707, acc_len=1.71, thru=19.6, I=21.7, T=36.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:27,996] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 15:16:33,966] loop.py:45 INFO TRAIN_STEP step=11894 loss=1.335390 Training: 66%|██████▌ | 11894/18018 [25:46:20<10:45:34, 6.33s/step, loss=1.335, acc=0.695, acc_len=1.66, thru=19.0, I=21.5, T=34.6, wait=2.4s, pool=53, epoch=2/2][2026-10-19 15:16:42,830] loop.py:380 INFO TIMING step=11895: step=3.581s data=1.461s compute=3.541s [fwd=2.509s bwd=1.018s opt=0.003s] dispatch=5.274s [2026-10-19 15:16:42,830] loop.py:45 INFO TRAIN_STEP step=11895 loss=1.331532 Training: 66%|██████▌ | 11895/18018 [25:46:28<12:03:08, 7.09s/step, loss=1.332, acc=0.696, acc_len=1.68, thru=15.4, I=16.4, T=35.7, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:42,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:16:43,277] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:16:49,174] loop.py:45 INFO TRAIN_STEP step=11896 loss=1.150893 Training: 66%|██████▌ | 11896/18018 [25:46:35<11:40:23, 6.86s/step, loss=1.151, acc=0.726, acc_len=1.77, thru=18.1, I=19.1, T=35.8, wait=2.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:16:49,170] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11895: forward=2585.9ms backward=949.3ms [2026-10-19 15:16:59,158] loop.py:45 INFO TRAIN_STEP step=11897 loss=1.343398 Training: 66%|██████▌ | 11897/18018 [25:46:45<13:15:40, 7.80s/step, loss=1.343, acc=0.691, acc_len=1.65, thru=12.0, I=14.8, T=34.6, wait=6.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:10,307] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:17:10,586] loop.py:45 INFO TRAIN_STEP step=11898 loss=1.355680 Training: 66%|██████▌ | 11898/18018 [25:46:56<15:06:38, 8.89s/step, loss=1.356, acc=0.694, acc_len=1.66, thru=16.4, I=12.8, T=29.9, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:10,808] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:18,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:17:18,727] loop.py:45 INFO TRAIN_STEP step=11899 loss=1.275077 Training: 66%|██████▌ | 11899/18018 [25:47:04<14:43:38, 8.66s/step, loss=1.275, acc=0.706, acc_len=1.70, thru=15.7, I=15.9, T=33.8, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:19,028] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:23,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:17:24,597] loop.py:380 INFO TIMING step=11900: step=3.562s data=1.513s compute=3.527s [fwd=2.511s bwd=1.001s opt=0.003s] dispatch=2.297s [2026-10-19 15:17:24,597] loop.py:45 INFO TRAIN_STEP step=11900 loss=1.437726 Training: 66%|██████▌ | 11900/18018 [25:47:10<13:17:57, 7.83s/step, loss=1.438, acc=0.674, acc_len=1.59, thru=20.2, I=20.4, T=35.9, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:24,708] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:32,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:17:32,879] loop.py:45 INFO TRAIN_STEP step=11901 loss=1.223737 Training: 66%|██████▌ | 11901/18018 [25:47:18<13:31:46, 7.96s/step, loss=1.224, acc=0.714, acc_len=1.73, thru=15.3, I=17.0, T=32.6, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:17:32,875] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11900: forward=2739.4ms backward=1139.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:32,907] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:40,401] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:17:41,240] loop.py:45 INFO TRAIN_STEP step=11902 loss=1.284868 Training: 66%|██████▌ | 11902/18018 [25:47:27<13:43:52, 8.08s/step, loss=1.285, acc=0.706, acc_len=1.70, thru=16.0, I=16.4, T=34.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:41,420] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:45,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:17:47,018] loop.py:45 INFO TRAIN_STEP step=11903 loss=1.243898 Training: 66%|██████▌ | 11903/18018 [25:47:33<12:33:14, 7.39s/step, loss=1.244, acc=0.709, acc_len=1.70, thru=20.6, I=24.2, T=35.7, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:17:47,306] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:17:54,471] loop.py:45 INFO TRAIN_STEP step=11904 loss=1.215931 Training: 66%|██████▌ | 11904/18018 [25:47:40<12:34:58, 7.41s/step, loss=1.216, acc=0.713, acc_len=1.73, thru=16.2, I=17.9, T=37.2, wait=4.0s, pool=57, epoch=2/2][2026-10-19 15:18:01,326] loop.py:380 INFO TIMING step=11905: step=3.365s data=1.393s compute=3.323s [fwd=2.352s bwd=0.957s opt=0.003s] dispatch=3.483s [2026-10-19 15:18:01,326] loop.py:45 INFO TRAIN_STEP step=11905 loss=1.138576 Training: 66%|██████▌ | 11905/18018 [25:47:47<12:17:58, 7.24s/step, loss=1.139, acc=0.731, acc_len=1.79, thru=18.8, I=19.4, T=38.0, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:08,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:18:10,149] loop.py:45 INFO TRAIN_STEP step=11906 loss=1.181005 Training: 66%|██████▌ | 11906/18018 [25:47:56<13:06:09, 7.72s/step, loss=1.181, acc=0.721, acc_len=1.76, thru=14.7, I=15.8, T=35.5, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:18:10,144] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11905: forward=2410.8ms backward=1140.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:10,472] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:18:17,338] loop.py:45 INFO TRAIN_STEP step=11907 loss=1.295667 Training: 66%|██████▌ | 11907/18018 [25:48:03<12:49:56, 7.56s/step, loss=1.296, acc=0.699, acc_len=1.68, thru=17.0, I=19.7, T=34.6, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:22,794] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:18:23,934] loop.py:45 INFO TRAIN_STEP step=11908 loss=1.346706 Training: 66%|██████▌ | 11908/18018 [25:48:10<12:20:15, 7.27s/step, loss=1.347, acc=0.691, acc_len=1.65, thru=18.4, I=21.1, T=30.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:24,303] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:18:29,935] loop.py:45 INFO TRAIN_STEP step=11909 loss=1.233227 Training: 66%|██████▌ | 11909/18018 [25:48:16<11:41:24, 6.89s/step, loss=1.233, acc=0.717, acc_len=1.74, thru=17.9, I=21.4, T=37.7, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:37,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:18:38,381] loop.py:380 INFO TIMING step=11910: step=4.249s data=1.965s compute=4.207s [fwd=2.994s bwd=1.199s opt=0.003s] dispatch=4.190s [2026-10-19 15:18:38,381] loop.py:45 INFO TRAIN_STEP step=11910 loss=1.357863 Training: 66%|██████▌ | 11910/18018 [25:48:24<12:28:54, 7.36s/step, loss=1.358, acc=0.693, acc_len=1.66, thru=16.4, I=17.5, T=30.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:38,487] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:44,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:18:45,774] loop.py:45 INFO TRAIN_STEP step=11911 loss=1.201910 Training: 66%|██████▌ | 11911/18018 [25:48:31<12:29:49, 7.37s/step, loss=1.202, acc=0.721, acc_len=1.76, thru=17.0, I=19.4, T=32.8, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:18:45,770] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11910: forward=2696.2ms backward=1156.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:45,909] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:51,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:18:53,113] loop.py:45 INFO TRAIN_STEP step=11912 loss=1.415514 Training: 66%|██████▌ | 11912/18018 [25:48:39<12:28:53, 7.36s/step, loss=1.416, acc=0.682, acc_len=1.62, thru=15.8, I=20.4, T=34.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:53,371] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:58,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:18:59,358] loop.py:45 INFO TRAIN_STEP step=11913 loss=1.404690 Training: 66%|██████▌ | 11913/18018 [25:48:45<11:54:49, 7.03s/step, loss=1.405, acc=0.677, acc_len=1.60, thru=19.7, I=24.2, T=36.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:18:59,507] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:04,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:19:06,035] loop.py:45 INFO TRAIN_STEP step=11914 loss=1.271212 Training: 66%|██████▌ | 11914/18018 [25:48:52<11:44:00, 6.92s/step, loss=1.271, acc=0.702, acc_len=1.70, thru=18.9, I=21.8, T=36.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:06,270] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:12,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:19:12,866] loop.py:380 INFO TIMING step=11915: step=3.495s data=1.444s compute=3.465s [fwd=2.459s bwd=0.991s opt=0.003s] dispatch=3.327s [2026-10-19 15:19:12,866] loop.py:45 INFO TRAIN_STEP step=11915 loss=1.292009 Training: 66%|██████▌ | 11915/18018 [25:48:58<11:41:08, 6.89s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=18.5, I=21.1, T=36.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:13,317] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:19:20,603] loop.py:45 INFO TRAIN_STEP step=11916 loss=1.382623 Training: 66%|██████▌ | 11916/18018 [25:49:06<12:06:51, 7.15s/step, loss=1.383, acc=0.684, acc_len=1.62, thru=15.8, I=16.9, T=34.3, wait=4.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:19:20,599] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11915: forward=2685.5ms backward=999.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:26,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:19:28,611] loop.py:45 INFO TRAIN_STEP step=11917 loss=1.351748 Training: 66%|██████▌ | 11917/18018 [25:49:14<12:32:57, 7.40s/step, loss=1.352, acc=0.692, acc_len=1.66, thru=17.2, I=19.8, T=26.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:28,714] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 [2026-10-19 15:19:34,767] loop.py:45 INFO TRAIN_STEP step=11918 loss=1.177534 Training: 66%|██████▌ | 11918/18018 [25:49:20<11:54:42, 7.03s/step, loss=1.178, acc=0.722, acc_len=1.75, thru=13.8, I=20.1, T=37.1, wait=2.7s, pool=25, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 15:19:34,891] eagle_store.py:517 WARNING batch_get_buffer missing keys (k381374-a14cfe9a_hs, k381374-a14cfe9a_ids, k381374-a14cfe9a_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:42,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 15:19:34,914] eagle_store.py:517 WARNING batch_get_buffer missing keys (k381373-853116e4_hs, k381373-853116e4_ids, k381373-853116e4_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 15:19:43,248] loop.py:45 INFO TRAIN_STEP step=11919 loss=1.393446 Training: 66%|██████▌ | 11919/18018 [25:49:29<12:38:54, 7.47s/step, loss=1.393, acc=0.682, acc_len=1.62, thru=18.1, I=16.8, T=35.4, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:43,285] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:49,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:19:49,942] loop.py:380 INFO TIMING step=11920: step=3.514s data=1.457s compute=3.484s [fwd=2.478s bwd=0.992s opt=0.003s] dispatch=3.170s [2026-10-19 15:19:49,942] loop.py:45 INFO TRAIN_STEP step=11920 loss=1.264507 Training: 66%|██████▌ | 11920/18018 [25:49:36<12:15:10, 7.23s/step, loss=1.265, acc=0.708, acc_len=1.70, thru=19.7, I=20.4, T=36.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:19:50,124] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:19:56,066] loop.py:45 INFO TRAIN_STEP step=11921 loss=1.232594 Training: 66%|██████▌ | 11921/18018 [25:49:42<11:41:13, 6.90s/step, loss=1.233, acc=0.710, acc_len=1.71, thru=19.4, I=19.7, T=38.1, wait=2.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:19:56,061] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11920: forward=2407.1ms backward=905.8ms [2026-10-19 15:20:03,512] loop.py:45 INFO TRAIN_STEP step=11922 loss=1.275907 Training: 66%|██████▌ | 11922/18018 [25:49:49<11:57:51, 7.07s/step, loss=1.276, acc=0.708, acc_len=1.71, thru=15.8, I=16.7, T=32.4, wait=3.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:11,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:20:12,150] loop.py:45 INFO TRAIN_STEP step=11923 loss=1.395325 Training: 66%|██████▌ | 11923/18018 [25:49:58<12:45:30, 7.54s/step, loss=1.395, acc=0.682, acc_len=1.62, thru=16.9, I=15.8, T=33.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:12,630] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:20,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:20:21,149] loop.py:45 INFO TRAIN_STEP step=11924 loss=1.316453 Training: 66%|██████▌ | 11924/18018 [25:50:07<13:29:59, 7.98s/step, loss=1.316, acc=0.697, acc_len=1.67, thru=15.0, I=16.1, T=35.0, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:21,632] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:26,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:20:26,944] loop.py:380 INFO TIMING step=11925: step=3.276s data=1.274s compute=3.246s [fwd=2.341s bwd=0.890s opt=0.003s] dispatch=2.511s [2026-10-19 15:20:26,944] loop.py:45 INFO TRAIN_STEP step=11925 loss=1.221701 Training: 66%|██████▌ | 11925/18018 [25:50:13<12:23:27, 7.32s/step, loss=1.222, acc=0.714, acc_len=1.74, thru=20.6, I=22.5, T=39.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:27,429] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:32,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:20:32,962] loop.py:45 INFO TRAIN_STEP step=11926 loss=1.190967 Training: 66%|██████▌ | 11926/18018 [25:50:19<11:43:39, 6.93s/step, loss=1.191, acc=0.719, acc_len=1.75, thru=21.8, I=25.8, T=35.7, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:20:32,957] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11925: forward=2473.8ms backward=1069.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:33,362] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:39,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:20:40,207] loop.py:45 INFO TRAIN_STEP step=11927 loss=1.273535 Training: 66%|██████▌ | 11927/18018 [25:50:26<11:53:07, 7.02s/step, loss=1.274, acc=0.706, acc_len=1.70, thru=17.2, I=19.8, T=32.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:20:40,313] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 15:20:46,737] loop.py:45 INFO TRAIN_STEP step=11928 loss=1.288470 Training: 66%|██████▌ | 11928/18018 [25:50:32<11:37:54, 6.88s/step, loss=1.288, acc=0.699, acc_len=1.67, thru=16.6, I=17.0, T=35.1, wait=2.9s, pool=41, epoch=2/2][2026-10-19 15:20:57,321] loop.py:45 INFO TRAIN_STEP step=11929 loss=1.384296 Training: 66%|██████▌ | 11929/18018 [25:50:43<13:30:51, 7.99s/step, loss=1.384, acc=0.685, acc_len=1.63, thru=9.7, I=12.3, T=32.2, wait=6.6s, pool=29, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:07,621] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:21:08,649] loop.py:380 INFO TIMING step=11930: step=4.459s data=2.078s compute=4.387s [fwd=3.067s bwd=1.305s opt=0.003s] dispatch=6.855s [2026-10-19 15:21:08,649] loop.py:45 INFO TRAIN_STEP step=11930 loss=1.368489 Training: 66%|██████▌ | 11930/18018 [25:50:54<15:12:09, 8.99s/step, loss=1.368, acc=0.690, acc_len=1.65, thru=14.4, I=14.7, T=28.7, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:09,129] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:15,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:21:16,378] loop.py:45 INFO TRAIN_STEP step=11931 loss=1.301612 Training: 66%|██████▌ | 11931/18018 [25:51:02<14:33:43, 8.61s/step, loss=1.302, acc=0.701, acc_len=1.69, thru=15.2, I=18.8, T=31.9, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:21:16,375] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11930: forward=2837.4ms backward=1122.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:16,511] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:24,459] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:21:24,753] loop.py:45 INFO TRAIN_STEP step=11932 loss=1.321092 Training: 66%|██████▌ | 11932/18018 [25:51:10<14:26:20, 8.54s/step, loss=1.321, acc=0.696, acc_len=1.66, thru=15.1, I=17.4, T=29.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:24,960] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:31,461] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:21:31,905] loop.py:45 INFO TRAIN_STEP step=11933 loss=1.263080 Training: 66%|██████▌ | 11933/18018 [25:51:17<13:43:54, 8.12s/step, loss=1.263, acc=0.711, acc_len=1.71, thru=16.4, I=16.5, T=31.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:31,974] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 15:21:37,876] loop.py:45 INFO TRAIN_STEP step=11934 loss=1.175056 Training: 66%|██████▌ | 11934/18018 [25:51:23<12:38:17, 7.48s/step, loss=1.175, acc=0.723, acc_len=1.76, thru=16.8, I=17.4, T=37.0, wait=2.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:48,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:21:49,445] loop.py:380 INFO TIMING step=11935: step=5.611s data=2.864s compute=5.555s [fwd=4.207s bwd=1.332s opt=0.003s] dispatch=5.949s [2026-10-19 15:21:49,445] loop.py:45 INFO TRAIN_STEP step=11935 loss=1.394754 Training: 66%|██████▌ | 11935/18018 [25:51:35<14:42:32, 8.71s/step, loss=1.395, acc=0.686, acc_len=1.63, thru=12.0, I=13.3, T=22.8, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:49,931] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:56,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 15:21:57,328] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11935: forward=3046.5ms backward=1385.6ms [2026-10-19 15:21:57,333] loop.py:45 INFO TRAIN_STEP step=11936 loss=1.278259 Training: 66%|██████▌ | 11936/18018 [25:51:43<14:17:33, 8.46s/step, loss=1.278, acc=0.706, acc_len=1.70, thru=14.0, I=14.6, T=28.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:21:57,496] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:04,646] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:22:06,240] loop.py:45 INFO TRAIN_STEP step=11937 loss=1.236011 Training: 66%|██████▋ | 11937/18018 [25:51:52<14:31:03, 8.59s/step, loss=1.236, acc=0.708, acc_len=1.71, thru=13.9, I=16.0, T=27.9, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:06,655] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:11,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:22:13,040] loop.py:45 INFO TRAIN_STEP step=11938 loss=1.271183 Training: 66%|██████▋ | 11938/18018 [25:51:59<13:36:21, 8.06s/step, loss=1.271, acc=0.705, acc_len=1.69, thru=16.9, I=22.9, T=31.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:13,494] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:19,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:22:20,149] loop.py:45 INFO TRAIN_STEP step=11939 loss=1.354508 Training: 66%|██████▋ | 11939/18018 [25:52:06<13:07:24, 7.77s/step, loss=1.355, acc=0.689, acc_len=1.65, thru=16.5, I=19.8, T=28.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:20,604] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:25,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:22:27,929] loop.py:380 INFO TIMING step=11940: step=4.638s data=2.127s compute=4.611s [fwd=3.211s bwd=1.385s opt=0.003s] dispatch=3.134s [2026-10-19 15:22:27,929] loop.py:45 INFO TRAIN_STEP step=11940 loss=1.278071 Training: 66%|██████▋ | 11940/18018 [25:52:14<13:07:30, 7.77s/step, loss=1.278, acc=0.703, acc_len=1.69, thru=14.9, I=18.3, T=27.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:28,425] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 15:22:34,133] loop.py:45 INFO TRAIN_STEP step=11941 loss=1.307219 Training: 66%|██████▋ | 11941/18018 [25:52:20<12:19:43, 7.30s/step, loss=1.307, acc=0.697, acc_len=1.68, thru=16.0, I=26.3, T=34.9, wait=2.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:22:34,128] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11940: forward=2585.9ms backward=1032.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:41,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:22:41,548] loop.py:45 INFO TRAIN_STEP step=11942 loss=1.324301 Training: 66%|██████▋ | 11942/18018 [25:52:27<12:22:56, 7.34s/step, loss=1.324, acc=0.699, acc_len=1.68, thru=17.8, I=17.7, T=31.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:41,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:22:47,862] loop.py:45 INFO TRAIN_STEP step=11943 loss=1.360251 Training: 66%|██████▋ | 11943/18018 [25:52:33<11:51:46, 7.03s/step, loss=1.360, acc=0.690, acc_len=1.65, thru=19.4, I=17.7, T=31.0, wait=2.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:53,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:22:55,351] loop.py:45 INFO TRAIN_STEP step=11944 loss=1.374411 Training: 66%|██████▋ | 11944/18018 [25:52:41<12:05:38, 7.17s/step, loss=1.374, acc=0.685, acc_len=1.63, thru=18.5, I=18.7, T=27.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:55,620] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:22:59,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:23:00,857] loop.py:380 INFO TIMING step=11945: step=3.464s data=1.339s compute=3.436s [fwd=2.461s bwd=0.960s opt=0.003s] dispatch=2.033s [2026-10-19 15:23:00,858] loop.py:45 INFO TRAIN_STEP step=11945 loss=1.330777 Training: 66%|██████▋ | 11945/18018 [25:52:46<11:15:04, 6.67s/step, loss=1.331, acc=0.692, acc_len=1.64, thru=19.0, I=23.3, T=37.0, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:01,230] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:06,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:23:07,260] loop.py:45 INFO TRAIN_STEP step=11946 loss=1.205680 Training: 66%|██████▋ | 11946/18018 [25:52:53<11:06:50, 6.59s/step, loss=1.206, acc=0.718, acc_len=1.74, thru=20.4, I=26.1, T=32.0, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:23:07,255] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11945: forward=2791.0ms backward=1154.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:07,578] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:11,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:23:13,439] loop.py:45 INFO TRAIN_STEP step=11947 loss=1.234037 Training: 66%|██████▋ | 11947/18018 [25:52:59<10:54:18, 6.47s/step, loss=1.234, acc=0.709, acc_len=1.72, thru=19.4, I=29.4, T=32.1, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:13,675] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 15:23:19,164] loop.py:45 INFO TRAIN_STEP step=11948 loss=1.136782 Training: 66%|██████▋ | 11948/18018 [25:53:05<10:31:41, 6.24s/step, loss=1.137, acc=0.726, acc_len=1.77, thru=18.5, I=28.3, T=36.4, wait=2.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:26,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:23:27,794] loop.py:45 INFO TRAIN_STEP step=11949 loss=1.330079 Training: 66%|██████▋ | 11949/18018 [25:53:13<11:43:51, 6.96s/step, loss=1.330, acc=0.695, acc_len=1.66, thru=16.8, I=16.9, T=32.8, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:28,295] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:33,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:23:34,624] loop.py:380 INFO TIMING step=11950: step=3.693s data=1.652s compute=3.649s [fwd=2.566s bwd=1.068s opt=0.003s] dispatch=3.130s [2026-10-19 15:23:34,624] loop.py:45 INFO TRAIN_STEP step=11950 loss=1.259140 Training: 66%|██████▋ | 11950/18018 [25:53:20<11:39:54, 6.92s/step, loss=1.259, acc=0.707, acc_len=1.70, thru=17.8, I=21.9, T=34.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:34,968] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:39,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 15:23:41,567] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11950: forward=3218.7ms backward=1440.3ms [2026-10-19 15:23:41,572] loop.py:45 INFO TRAIN_STEP step=11951 loss=1.252884 Training: 66%|██████▋ | 11951/18018 [25:53:27<11:40:36, 6.93s/step, loss=1.253, acc=0.709, acc_len=1.71, thru=18.1, I=25.6, T=27.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:42,070] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:47,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:23:50,585] loop.py:45 INFO TRAIN_STEP step=11952 loss=1.280576 Training: 66%|██████▋ | 11952/18018 [25:53:36<12:43:42, 7.55s/step, loss=1.281, acc=0.705, acc_len=1.69, thru=15.0, I=21.4, T=24.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:50,697] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:57,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:23:58,536] loop.py:45 INFO TRAIN_STEP step=11953 loss=1.263762 Training: 66%|██████▋ | 11953/18018 [25:53:44<12:55:37, 7.67s/step, loss=1.264, acc=0.710, acc_len=1.71, thru=16.4, I=18.6, T=28.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:23:58,998] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:04,880] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:24:05,834] loop.py:45 INFO TRAIN_STEP step=11954 loss=1.291972 Training: 66%|██████▋ | 11954/18018 [25:53:51<12:44:11, 7.56s/step, loss=1.292, acc=0.700, acc_len=1.68, thru=17.4, I=21.6, T=34.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:05,885] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:12,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:24:13,529] loop.py:380 INFO TIMING step=11955: step=4.052s data=1.749s compute=4.015s [fwd=2.818s bwd=1.182s opt=0.003s] dispatch=3.634s [2026-10-19 15:24:13,530] loop.py:45 INFO TRAIN_STEP step=11955 loss=1.309126 Training: 66%|██████▋ | 11955/18018 [25:53:59<12:48:07, 7.60s/step, loss=1.309, acc=0.698, acc_len=1.67, thru=16.8, I=19.3, T=31.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:13,819] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:19,047] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:24:20,867] loop.py:45 INFO TRAIN_STEP step=11956 loss=1.219931 Training: 66%|██████▋ | 11956/18018 [25:54:06<12:39:57, 7.52s/step, loss=1.220, acc=0.716, acc_len=1.74, thru=16.6, I=21.6, T=30.5, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:24:20,863] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11955: forward=2913.0ms backward=1231.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:21,057] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:25,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:24:27,399] loop.py:45 INFO TRAIN_STEP step=11957 loss=1.316260 Training: 66%|██████▋ | 11957/18018 [25:54:13<12:09:52, 7.23s/step, loss=1.316, acc=0.693, acc_len=1.66, thru=17.4, I=26.3, T=31.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:27,571] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:34,584] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:24:35,731] loop.py:45 INFO TRAIN_STEP step=11958 loss=1.308616 Training: 66%|██████▋ | 11958/18018 [25:54:21<12:43:13, 7.56s/step, loss=1.309, acc=0.703, acc_len=1.69, thru=15.7, I=18.0, T=30.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:36,096] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:24:42,943] loop.py:45 INFO TRAIN_STEP step=11959 loss=1.222118 Training: 66%|██████▋ | 11959/18018 [25:54:29<12:32:43, 7.45s/step, loss=1.222, acc=0.712, acc_len=1.73, thru=14.1, I=17.2, T=31.9, wait=3.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:53,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:24:54,674] loop.py:380 INFO TIMING step=11960: step=4.624s data=2.119s compute=4.587s [fwd=3.247s bwd=1.325s opt=0.003s] dispatch=7.096s [2026-10-19 15:24:54,674] loop.py:45 INFO TRAIN_STEP step=11960 loss=1.270180 Training: 66%|██████▋ | 11960/18018 [25:54:40<14:42:08, 8.74s/step, loss=1.270, acc=0.707, acc_len=1.70, thru=14.2, I=14.8, T=27.7, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:24:54,946] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:25:01,276] loop.py:45 INFO TRAIN_STEP step=11961 loss=1.259138 Training: 66%|██████▋ | 11961/18018 [25:54:47<13:37:20, 8.10s/step, loss=1.259, acc=0.709, acc_len=1.71, thru=15.3, I=18.4, T=34.0, wait=2.8s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:25:01,272] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11960: forward=2636.6ms backward=1064.6ms [2026-10-19 15:25:09,942] loop.py:45 INFO TRAIN_STEP step=11962 loss=1.362201 Training: 66%|██████▋ | 11962/18018 [25:54:56<13:54:21, 8.27s/step, loss=1.362, acc=0.691, acc_len=1.65, thru=16.4, I=14.8, T=32.1, wait=4.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:14,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:25:16,820] loop.py:45 INFO TRAIN_STEP step=11963 loss=1.193922 Training: 66%|██████▋ | 11963/18018 [25:55:02<13:12:18, 7.85s/step, loss=1.194, acc=0.717, acc_len=1.74, thru=18.8, I=21.6, T=29.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:16,860] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:21,968] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:25:23,121] loop.py:45 INFO TRAIN_STEP step=11964 loss=1.310079 Training: 66%|██████▋ | 11964/18018 [25:55:09<12:25:10, 7.39s/step, loss=1.310, acc=0.698, acc_len=1.67, thru=17.8, I=26.2, T=31.9, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:23,474] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:28,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:25:29,318] loop.py:380 INFO TIMING step=11965: step=3.653s data=1.509s compute=3.621s [fwd=2.568s bwd=1.037s opt=0.003s] dispatch=2.536s [2026-10-19 15:25:29,319] loop.py:45 INFO TRAIN_STEP step=11965 loss=1.178062 Training: 66%|██████▋ | 11965/18018 [25:55:15<11:49:09, 7.03s/step, loss=1.178, acc=0.724, acc_len=1.77, thru=18.8, I=23.2, T=35.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:29,773] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:34,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:25:36,118] loop.py:45 INFO TRAIN_STEP step=11966 loss=1.273589 (TrainerActor pid=9310) [2026-10-19 15:25:36,114] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11965: forward=2485.1ms backward=1011.2ms Training: 66%|██████▋ | 11966/18018 [25:55:22<11:42:05, 6.96s/step, loss=1.274, acc=0.705, acc_len=1.70, thru=18.3, I=23.3, T=36.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:36,472] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:25:42,143] loop.py:45 INFO TRAIN_STEP step=11967 loss=1.247206 Training: 66%|██████▋ | 11967/18018 [25:55:28<11:13:39, 6.68s/step, loss=1.247, acc=0.709, acc_len=1.72, thru=19.0, I=24.0, T=36.5, wait=2.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:49,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:25:50,237] loop.py:45 INFO TRAIN_STEP step=11968 loss=1.283232 Training: 66%|██████▋ | 11968/18018 [25:55:36<11:56:17, 7.10s/step, loss=1.283, acc=0.705, acc_len=1.70, thru=17.7, I=18.2, T=33.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:25:50,324] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 15:25:58,605] loop.py:45 INFO TRAIN_STEP step=11969 loss=1.319026 Training: 66%|██████▋ | 11969/18018 [25:55:44<12:34:27, 7.48s/step, loss=1.319, acc=0.698, acc_len=1.68, thru=11.2, I=12.6, T=33.5, wait=4.5s, pool=25, epoch=2/2][2026-10-19 15:26:08,524] loop.py:380 INFO TIMING step=11970: step=3.810s data=1.498s compute=3.781s [fwd=2.704s bwd=1.063s opt=0.003s] dispatch=6.099s [2026-10-19 15:26:08,524] loop.py:45 INFO TRAIN_STEP step=11970 loss=1.283071 Training: 66%|██████▋ | 11970/18018 [25:55:54<13:47:59, 8.21s/step, loss=1.283, acc=0.703, acc_len=1.69, thru=16.8, I=15.6, T=33.6, wait=6.1s, pool=61, epoch=2/2][2026-10-19 15:26:16,291] loop.py:45 INFO TRAIN_STEP step=11971 loss=1.206798 Training: 66%|██████▋ | 11971/18018 [25:56:02<13:34:15, 8.08s/step, loss=1.207, acc=0.715, acc_len=1.73, thru=15.6, I=16.0, T=34.7, wait=4.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:26:16,286] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11970: forward=2516.8ms backward=1127.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:26:22,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:26:23,367] loop.py:45 INFO TRAIN_STEP step=11972 loss=1.155109 Training: 66%|██████▋ | 11972/18018 [25:56:09<13:03:50, 7.78s/step, loss=1.155, acc=0.728, acc_len=1.78, thru=17.5, I=19.4, T=35.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:26:23,670] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:26:29,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:26:29,906] loop.py:45 INFO TRAIN_STEP step=11973 loss=1.204953 Training: 66%|██████▋ | 11973/18018 [25:56:15<12:26:14, 7.41s/step, loss=1.205, acc=0.720, acc_len=1.75, thru=19.3, I=20.6, T=32.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:26:30,094] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:26:37,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:26:38,674] loop.py:45 INFO TRAIN_STEP step=11974 loss=1.324994 Training: 66%|██████▋ | 11974/18018 [25:56:24<13:07:14, 7.82s/step, loss=1.325, acc=0.694, acc_len=1.67, thru=15.2, I=17.0, T=30.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:26:39,138] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:26:45,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:26:46,025] loop.py:380 INFO TIMING step=11975: step=4.026s data=1.809s compute=3.975s [fwd=2.797s bwd=1.163s opt=0.003s] dispatch=3.315s [2026-10-19 15:26:46,026] loop.py:45 INFO TRAIN_STEP step=11975 loss=1.344493 Training: 66%|██████▋ | 11975/18018 [25:56:32<12:53:06, 7.68s/step, loss=1.344, acc=0.691, acc_len=1.65, thru=16.0, I=19.7, T=31.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:26:46,459] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:26:54,306] loop.py:45 INFO TRAIN_STEP step=11976 loss=1.208336 Training: 66%|██████▋ | 11976/18018 [25:56:40<13:11:16, 7.86s/step, loss=1.208, acc=0.715, acc_len=1.73, thru=14.3, I=14.3, T=30.6, wait=4.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:26:54,302] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11975: forward=2948.4ms backward=1185.2ms [2026-10-19 15:27:01,687] loop.py:45 INFO TRAIN_STEP step=11977 loss=1.367604 Training: 66%|██████▋ | 11977/18018 [25:56:47<12:56:50, 7.72s/step, loss=1.368, acc=0.686, acc_len=1.63, thru=15.4, I=15.4, T=33.7, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:09,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:27:09,464] loop.py:45 INFO TRAIN_STEP step=11978 loss=1.307961 Training: 66%|██████▋ | 11978/18018 [25:56:55<12:58:27, 7.73s/step, loss=1.308, acc=0.699, acc_len=1.68, thru=17.4, I=17.5, T=33.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:09,598] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:15,601] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:27:16,126] loop.py:45 INFO TRAIN_STEP step=11979 loss=1.400499 Training: 66%|██████▋ | 11979/18018 [25:57:02<12:25:56, 7.41s/step, loss=1.400, acc=0.682, acc_len=1.61, thru=18.1, I=19.5, T=31.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:16,608] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:22,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:27:23,372] loop.py:380 INFO TIMING step=11980: step=4.040s data=1.841s compute=4.007s [fwd=2.778s bwd=1.214s opt=0.003s] dispatch=3.197s [2026-10-19 15:27:23,372] loop.py:45 INFO TRAIN_STEP step=11980 loss=1.203310 Training: 66%|██████▋ | 11980/18018 [25:57:09<12:20:49, 7.36s/step, loss=1.203, acc=0.715, acc_len=1.73, thru=17.4, I=19.1, T=31.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:23,833] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:29,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:27:30,131] loop.py:45 INFO TRAIN_STEP step=11981 loss=1.318354 Training: 66%|██████▋ | 11981/18018 [25:57:16<12:02:32, 7.18s/step, loss=1.318, acc=0.699, acc_len=1.68, thru=18.1, I=21.8, T=34.3, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:27:30,125] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11980: forward=2647.5ms backward=1035.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:30,512] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:36,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:27:36,850] loop.py:45 INFO TRAIN_STEP step=11982 loss=1.300264 Training: 67%|██████▋ | 11982/18018 [25:57:22<11:48:26, 7.04s/step, loss=1.300, acc=0.697, acc_len=1.67, thru=18.1, I=21.3, T=32.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:37,131] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:27:43,743] loop.py:45 INFO TRAIN_STEP step=11983 loss=1.230786 Training: 67%|██████▋ | 11983/18018 [25:57:29<11:43:53, 7.00s/step, loss=1.231, acc=0.712, acc_len=1.73, thru=15.1, I=17.8, T=31.3, wait=2.8s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:50,952] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:27:52,399] loop.py:45 INFO TRAIN_STEP step=11984 loss=1.382301 Training: 67%|██████▋ | 11984/18018 [25:57:38<12:33:46, 7.50s/step, loss=1.382, acc=0.683, acc_len=1.63, thru=17.7, I=17.5, T=31.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:52,461] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:58,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:27:59,672] loop.py:380 INFO TIMING step=11985: step=3.921s data=1.770s compute=3.876s [fwd=2.745s bwd=1.116s opt=0.003s] dispatch=3.342s [2026-10-19 15:27:59,673] loop.py:45 INFO TRAIN_STEP step=11985 loss=1.429488 Training: 67%|██████▋ | 11985/18018 [25:57:45<12:26:59, 7.43s/step, loss=1.429, acc=0.674, acc_len=1.60, thru=15.7, I=19.6, T=32.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:27:59,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:05,558] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:28:06,977] loop.py:45 INFO TRAIN_STEP step=11986 loss=1.249978 Training: 67%|██████▋ | 11986/18018 [25:57:53<12:23:03, 7.39s/step, loss=1.250, acc=0.713, acc_len=1.72, thru=17.8, I=22.0, T=33.7, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:28:06,972] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11985: forward=2653.7ms backward=1100.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:07,069] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:13,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:28:13,678] loop.py:45 INFO TRAIN_STEP step=11987 loss=1.337526 Training: 67%|██████▋ | 11987/18018 [25:57:59<12:02:08, 7.18s/step, loss=1.338, acc=0.695, acc_len=1.66, thru=18.8, I=23.5, T=32.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:14,108] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:28:20,238] loop.py:45 INFO TRAIN_STEP step=11988 loss=1.225644 Training: 67%|██████▋ | 11988/18018 [25:58:06<11:43:06, 7.00s/step, loss=1.226, acc=0.715, acc_len=1.74, thru=19.3, I=22.6, T=32.1, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:28,966] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:28:30,436] loop.py:45 INFO TRAIN_STEP step=11989 loss=1.271811 Training: 67%|██████▋ | 11989/18018 [25:58:16<13:19:36, 7.96s/step, loss=1.272, acc=0.705, acc_len=1.70, thru=13.7, I=13.1, T=32.8, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:30,474] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:36,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:28:37,010] loop.py:380 INFO TIMING step=11990: step=4.083s data=1.875s compute=4.052s [fwd=2.842s bwd=1.194s opt=0.003s] dispatch=2.481s [2026-10-19 15:28:37,010] loop.py:45 INFO TRAIN_STEP step=11990 loss=1.277490 Training: 67%|██████▋ | 11990/18018 [25:58:23<12:37:46, 7.54s/step, loss=1.277, acc=0.701, acc_len=1.69, thru=18.6, I=22.5, T=31.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:37,247] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:42,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:28:42,555] loop.py:45 INFO TRAIN_STEP step=11991 loss=1.328680 Training: 67%|██████▋ | 11991/18018 [25:58:28<11:37:27, 6.94s/step, loss=1.329, acc=0.694, acc_len=1.66, thru=20.0, I=22.7, T=35.4, wait=1.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:28:42,551] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11990: forward=2523.0ms backward=1053.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:42,623] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 15:28:48,425] loop.py:45 INFO TRAIN_STEP step=11992 loss=1.278869 Training: 67%|██████▋ | 11992/18018 [25:58:34<11:04:57, 6.62s/step, loss=1.279, acc=0.703, acc_len=1.69, thru=20.1, I=22.7, T=34.2, wait=2.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:55,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:28:56,122] loop.py:45 INFO TRAIN_STEP step=11993 loss=1.273394 Training: 67%|██████▋ | 11993/18018 [25:58:42<11:37:15, 6.94s/step, loss=1.273, acc=0.703, acc_len=1.70, thru=18.3, I=17.0, T=33.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:28:56,381] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:01,797] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:29:03,751] loop.py:45 INFO TRAIN_STEP step=11994 loss=1.333609 Training: 67%|██████▋ | 11994/18018 [25:58:49<11:57:50, 7.15s/step, loss=1.334, acc=0.691, acc_len=1.65, thru=16.8, I=20.3, T=33.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:03,811] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:09,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:29:10,038] loop.py:380 INFO TIMING step=11995: step=3.827s data=1.703s compute=3.798s [fwd=2.721s bwd=1.061s opt=0.003s] dispatch=2.450s [2026-10-19 15:29:10,038] loop.py:45 INFO TRAIN_STEP step=11995 loss=1.195024 Training: 67%|██████▋ | 11995/18018 [25:58:56<11:31:45, 6.89s/step, loss=1.195, acc=0.720, acc_len=1.74, thru=19.1, I=25.1, T=33.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:10,097] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:16,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:29:16,770] loop.py:45 INFO TRAIN_STEP step=11996 loss=1.289832 Training: 67%|██████▋ | 11996/18018 [25:59:02<11:26:45, 6.84s/step, loss=1.290, acc=0.701, acc_len=1.68, thru=18.0, I=19.9, T=35.5, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:29:16,765] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=11995: forward=2572.3ms backward=990.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:17,231] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:22,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:29:23,963] loop.py:45 INFO TRAIN_STEP step=11997 loss=1.168582 Training: 67%|██████▋ | 11997/18018 [25:59:10<11:37:12, 6.95s/step, loss=1.169, acc=0.718, acc_len=1.75, thru=17.3, I=20.6, T=29.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:24,385] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:28,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:29:30,397] loop.py:45 INFO TRAIN_STEP step=11998 loss=1.311787 Training: 67%|██████▋ | 11998/18018 [25:59:16<11:21:40, 6.79s/step, loss=1.312, acc=0.698, acc_len=1.67, thru=18.6, I=23.1, T=30.5, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:30,627] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:34,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:29:36,316] loop.py:45 INFO TRAIN_STEP step=11999 loss=1.259498 Training: 67%|██████▋ | 11999/18018 [25:59:22<10:55:10, 6.53s/step, loss=1.259, acc=0.706, acc_len=1.70, thru=19.0, I=29.4, T=32.7, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:36,467] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:29:43,299] loop.py:380 INFO TIMING step=12000: step=4.580s data=2.251s compute=4.551s [fwd=3.142s bwd=1.393s opt=0.003s] dispatch=2.394s [2026-10-19 15:29:43,299] loop.py:45 INFO TRAIN_STEP step=12000 loss=1.293596 Training: 67%|██████▋ | 12000/18018 [25:59:29<11:08:43, 6.67s/step, loss=1.294, acc=0.699, acc_len=1.68, thru=14.1, I=21.5, T=27.9, wait=2.4s, pool=41, epoch=2/2][2026-10-19 15:29:43,304] loop.py:419 INFO Saving checkpoint at step 12000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:44,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 15:29:46,674] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0012001 (TrainerActor pid=9310) [2026-10-19 15:29:46,685] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0011251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:29:47,900] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=1/64 [2026-10-19 15:29:55,786] loop.py:45 INFO TRAIN_STEP step=12001 loss=1.348134 Training: 67%|██████▋ | 12001/18018 [25:59:41<14:03:40, 8.41s/step, loss=1.348, acc=0.690, acc_len=1.65, thru=14.7, I=14.8, T=30.1, wait=3.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:29:55,782] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12000: forward=3028.8ms backward=1168.1ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 15:29:46,675] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0012001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 15:29:47,707] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0011251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:01,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:30:02,349] loop.py:45 INFO TRAIN_STEP step=12002 loss=1.282700 Training: 67%|██████▋ | 12002/18018 [25:59:48<13:07:51, 7.86s/step, loss=1.283, acc=0.701, acc_len=1.68, thru=19.2, I=19.2, T=35.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:02,824] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:30:09,700] loop.py:45 INFO TRAIN_STEP step=12003 loss=1.275097 Training: 67%|██████▋ | 12003/18018 [25:59:55<12:52:31, 7.71s/step, loss=1.275, acc=0.705, acc_len=1.69, thru=15.7, I=18.8, T=37.7, wait=3.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:18,091] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:30:19,457] loop.py:45 INFO TRAIN_STEP step=12004 loss=1.313241 Training: 67%|██████▋ | 12004/18018 [26:00:05<13:54:02, 8.32s/step, loss=1.313, acc=0.702, acc_len=1.69, thru=14.5, I=15.9, T=31.6, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:19,617] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:24,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:30:25,159] loop.py:380 INFO TIMING step=12005: step=3.632s data=1.533s compute=3.599s [fwd=2.560s bwd=1.024s opt=0.003s] dispatch=2.061s [2026-10-19 15:30:25,159] loop.py:45 INFO TRAIN_STEP step=12005 loss=1.307713 Training: 67%|██████▋ | 12005/18018 [26:00:11<12:35:12, 7.54s/step, loss=1.308, acc=0.699, acc_len=1.68, thru=20.4, I=25.5, T=35.2, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:25,647] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:30:31,583] loop.py:45 INFO TRAIN_STEP step=12006 loss=1.366628 (TrainerActor pid=9310) [2026-10-19 15:30:31,578] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12005: forward=2519.8ms backward=982.5ms Training: 67%|██████▋ | 12006/18018 [26:00:17<12:01:44, 7.20s/step, loss=1.367, acc=0.687, acc_len=1.63, thru=18.1, I=23.4, T=36.1, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:37,773] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:30:38,726] loop.py:45 INFO TRAIN_STEP step=12007 loss=1.331439 Training: 67%|██████▋ | 12007/18018 [26:00:24<11:59:44, 7.18s/step, loss=1.331, acc=0.698, acc_len=1.67, thru=18.0, I=20.2, T=31.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:38,782] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:46,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:30:46,702] loop.py:45 INFO TRAIN_STEP step=12008 loss=1.301323 Training: 67%|██████▋ | 12008/18018 [26:00:32<12:23:24, 7.42s/step, loss=1.301, acc=0.701, acc_len=1.68, thru=15.7, I=17.6, T=30.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:47,065] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:53,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:30:54,711] loop.py:45 INFO TRAIN_STEP step=12009 loss=1.216350 Training: 67%|██████▋ | 12009/18018 [26:00:40<12:40:52, 7.60s/step, loss=1.216, acc=0.718, acc_len=1.75, thru=16.0, I=20.3, T=33.0, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:55,040] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:30:59,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:00,313] loop.py:380 INFO TIMING step=12010: step=3.713s data=1.563s compute=3.682s [fwd=2.566s bwd=1.100s opt=0.003s] dispatch=1.880s [2026-10-19 15:31:00,313] loop.py:45 INFO TRAIN_STEP step=12010 loss=1.323731 Training: 67%|██████▋ | 12010/18018 [26:00:46<11:40:52, 7.00s/step, loss=1.324, acc=0.697, acc_len=1.67, thru=21.1, I=26.7, T=34.5, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:00,408] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:06,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:07,282] loop.py:45 INFO TRAIN_STEP step=12011 loss=1.350966 Training: 67%|██████▋ | 12011/18018 [26:00:53<11:39:46, 6.99s/step, loss=1.351, acc=0.689, acc_len=1.64, thru=18.9, I=20.7, T=33.8, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:31:07,278] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12010: forward=2669.0ms backward=1061.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:07,755] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:13,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:14,626] loop.py:45 INFO TRAIN_STEP step=12012 loss=1.264433 Training: 67%|██████▋ | 12012/18018 [26:01:00<11:50:17, 7.10s/step, loss=1.264, acc=0.711, acc_len=1.72, thru=17.1, I=20.0, T=37.1, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:15,105] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:19,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:21,369] loop.py:45 INFO TRAIN_STEP step=12013 loss=1.297591 Training: 67%|██████▋ | 12013/18018 [26:01:07<11:39:35, 6.99s/step, loss=1.298, acc=0.700, acc_len=1.68, thru=18.9, I=24.8, T=32.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:21,875] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:28,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:28,233] loop.py:45 INFO TRAIN_STEP step=12014 loss=1.226644 Training: 67%|██████▋ | 12014/18018 [26:01:14<11:35:43, 6.95s/step, loss=1.227, acc=0.714, acc_len=1.73, thru=17.3, I=20.9, T=32.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:28,559] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:33,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:34,660] loop.py:380 INFO TIMING step=12015: step=3.571s data=1.485s compute=3.543s [fwd=2.538s bwd=0.990s opt=0.003s] dispatch=2.846s [2026-10-19 15:31:34,660] loop.py:45 INFO TRAIN_STEP step=12015 loss=1.253342 Training: 67%|██████▋ | 12015/18018 [26:01:20<11:19:51, 6.80s/step, loss=1.253, acc=0.712, acc_len=1.73, thru=19.5, I=21.4, T=35.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:35,029] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:39,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:40,152] loop.py:45 INFO TRAIN_STEP step=12016 loss=1.164581 Training: 67%|██████▋ | 12016/18018 [26:01:26<10:40:34, 6.40s/step, loss=1.165, acc=0.722, acc_len=1.76, thru=22.1, I=28.3, T=38.9, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:31:40,147] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12015: forward=2319.9ms backward=922.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:40,482] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:44,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:46,432] loop.py:45 INFO TRAIN_STEP step=12017 loss=1.293155 Training: 67%|██████▋ | 12017/18018 [26:01:32<10:36:49, 6.37s/step, loss=1.293, acc=0.699, acc_len=1.68, thru=20.8, I=26.4, T=32.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:46,741] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:52,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:52,755] loop.py:45 INFO TRAIN_STEP step=12018 loss=1.301055 Training: 67%|██████▋ | 12018/18018 [26:01:38<10:35:16, 6.35s/step, loss=1.301, acc=0.701, acc_len=1.68, thru=18.9, I=24.6, T=34.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:53,028] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:57,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:31:58,183] loop.py:45 INFO TRAIN_STEP step=12019 loss=1.126544 Training: 67%|██████▋ | 12019/18018 [26:01:44<10:07:29, 6.08s/step, loss=1.127, acc=0.736, acc_len=1.80, thru=21.4, I=23.7, T=35.2, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:31:58,364] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:04,122] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:32:04,769] loop.py:380 INFO TIMING step=12020: step=3.927s data=1.708s compute=3.898s [fwd=2.734s bwd=1.150s opt=0.003s] dispatch=2.650s [2026-10-19 15:32:04,769] loop.py:45 INFO TRAIN_STEP step=12020 loss=1.215813 Training: 67%|██████▋ | 12020/18018 [26:01:50<10:22:44, 6.23s/step, loss=1.216, acc=0.717, acc_len=1.74, thru=19.4, I=21.0, T=32.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:05,126] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:09,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:32:10,755] loop.py:45 INFO TRAIN_STEP step=12021 loss=1.283863 Training: 67%|██████▋ | 12021/18018 [26:01:56<10:15:16, 6.16s/step, loss=1.284, acc=0.703, acc_len=1.69, thru=19.7, I=25.4, T=32.5, wait=2.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:32:10,750] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12020: forward=2804.0ms backward=1076.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:10,858] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 15:32:17,823] loop.py:45 INFO TRAIN_STEP step=12022 loss=1.252403 Training: 67%|██████▋ | 12022/18018 [26:02:03<10:42:33, 6.43s/step, loss=1.252, acc=0.709, acc_len=1.71, thru=17.1, I=20.6, T=32.0, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:23,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:32:26,097] loop.py:45 INFO TRAIN_STEP step=12023 loss=1.280692 Training: 67%|██████▋ | 12023/18018 [26:02:12<11:37:45, 6.98s/step, loss=1.281, acc=0.708, acc_len=1.71, thru=16.4, I=20.8, T=23.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:26,446] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 [2026-10-19 15:32:32,256] loop.py:45 INFO TRAIN_STEP step=12024 loss=1.195241 Training: 67%|██████▋ | 12024/18018 [26:02:18<11:12:57, 6.74s/step, loss=1.195, acc=0.719, acc_len=1.75, thru=16.4, I=27.1, T=34.3, wait=2.4s, pool=61, epoch=2/2][2026-10-19 15:32:39,700] loop.py:380 INFO TIMING step=12025: step=3.582s data=1.424s compute=3.545s [fwd=2.567s bwd=0.964s opt=0.003s] dispatch=3.850s [2026-10-19 15:32:39,700] loop.py:45 INFO TRAIN_STEP step=12025 loss=1.309543 Training: 67%|██████▋ | 12025/18018 [26:02:25<11:33:57, 6.95s/step, loss=1.310, acc=0.695, acc_len=1.66, thru=16.8, I=16.4, T=35.7, wait=3.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:46,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:32:48,283] loop.py:45 INFO TRAIN_STEP step=12026 loss=1.343832 Training: 67%|██████▋ | 12026/18018 [26:02:34<12:22:51, 7.44s/step, loss=1.344, acc=0.697, acc_len=1.67, thru=17.2, I=17.0, T=26.3, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:32:48,279] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12025: forward=3402.8ms backward=1416.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:48,506] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:55,141] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:32:55,495] loop.py:45 INFO TRAIN_STEP step=12027 loss=1.260397 Training: 67%|██████▋ | 12027/18018 [26:02:41<12:15:58, 7.37s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=15.3, I=16.9, T=33.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:32:55,658] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:33:02,563] loop.py:45 INFO TRAIN_STEP step=12028 loss=1.271345 Training: 67%|██████▋ | 12028/18018 [26:02:48<12:06:45, 7.28s/step, loss=1.271, acc=0.703, acc_len=1.69, thru=15.7, I=16.1, T=31.4, wait=3.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:33:10,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:33:11,783] loop.py:45 INFO TRAIN_STEP step=12029 loss=1.305000 Training: 67%|██████▋ | 12029/18018 [26:02:57<13:04:46, 7.86s/step, loss=1.305, acc=0.701, acc_len=1.69, thru=16.2, I=16.1, T=31.8, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:33:11,983] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:33:19,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:33:19,894] loop.py:380 INFO TIMING step=12030: step=4.036s data=1.777s compute=4.007s [fwd=2.883s bwd=1.110s opt=0.003s] dispatch=4.065s [2026-10-19 15:33:19,894] loop.py:45 INFO TRAIN_STEP step=12030 loss=1.382527 Training: 67%|██████▋ | 12030/18018 [26:03:05<13:12:04, 7.94s/step, loss=1.383, acc=0.688, acc_len=1.64, thru=14.6, I=16.1, T=31.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:33:20,230] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:33:26,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:33:26,977] loop.py:45 INFO TRAIN_STEP step=12031 loss=1.406838 Training: 67%|██████▋ | 12031/18018 [26:03:13<12:46:24, 7.68s/step, loss=1.407, acc=0.680, acc_len=1.61, thru=17.7, I=19.7, T=35.1, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:33:26,972] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12030: forward=2526.1ms backward=1069.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:33:27,168] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:33:33,627] loop.py:45 INFO TRAIN_STEP step=12032 loss=1.202497 Training: 67%|██████▋ | 12032/18018 [26:03:19<12:15:24, 7.37s/step, loss=1.202, acc=0.720, acc_len=1.75, thru=17.1, I=18.5, T=34.9, wait=3.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:33:42,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:33:43,981] loop.py:45 INFO TRAIN_STEP step=12033 loss=1.277071 Training: 67%|██████▋ | 12033/18018 [26:03:30<13:44:35, 8.27s/step, loss=1.277, acc=0.707, acc_len=1.70, thru=14.6, I=16.2, T=32.7, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:33:44,269] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:33:52,239] loop.py:45 INFO TRAIN_STEP step=12034 loss=1.177174 Training: 67%|██████▋ | 12034/18018 [26:03:38<13:44:10, 8.26s/step, loss=1.177, acc=0.720, acc_len=1.75, thru=8.4, I=15.3, T=35.9, wait=4.7s, pool=17, epoch=2/2] [2026-10-19 15:34:02,020] loop.py:380 INFO TIMING step=12035: step=3.492s data=1.338s compute=3.462s [fwd=2.470s bwd=0.978s opt=0.003s] dispatch=6.279s [2026-10-19 15:34:02,020] loop.py:45 INFO TRAIN_STEP step=12035 loss=1.235547 Training: 67%|██████▋ | 12035/18018 [26:03:48<14:29:19, 8.72s/step, loss=1.236, acc=0.714, acc_len=1.73, thru=18.2, I=13.6, T=36.7, wait=6.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:08,722] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:34:09,856] loop.py:45 INFO TRAIN_STEP step=12036 loss=1.222230 Training: 67%|██████▋ | 12036/18018 [26:03:55<14:02:50, 8.45s/step, loss=1.222, acc=0.714, acc_len=1.73, thru=17.8, I=18.2, T=30.9, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:34:09,852] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12035: forward=2953.1ms backward=1133.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:10,228] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:34:16,291] loop.py:45 INFO TRAIN_STEP step=12037 loss=1.322882 Training: 67%|██████▋ | 12037/18018 [26:04:02<13:02:24, 7.85s/step, loss=1.323, acc=0.700, acc_len=1.68, thru=17.3, I=24.1, T=32.9, wait=2.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:26,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:34:27,735] loop.py:45 INFO TRAIN_STEP step=12038 loss=1.281411 Training: 67%|██████▋ | 12038/18018 [26:04:13<14:49:44, 8.93s/step, loss=1.281, acc=0.701, acc_len=1.69, thru=14.1, I=15.0, T=28.5, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:28,196] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:34:35,082] loop.py:45 INFO TRAIN_STEP step=12039 loss=1.278833 Training: 67%|██████▋ | 12039/18018 [26:04:21<14:02:20, 8.45s/step, loss=1.279, acc=0.704, acc_len=1.70, thru=13.7, I=14.7, T=32.5, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:41,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:34:43,073] loop.py:380 INFO TIMING step=12040: step=4.209s data=1.958s compute=4.170s [fwd=2.951s bwd=1.205s opt=0.003s] dispatch=3.773s [2026-10-19 15:34:43,074] loop.py:45 INFO TRAIN_STEP step=12040 loss=1.307267 Training: 67%|██████▋ | 12040/18018 [26:04:29<13:48:25, 8.31s/step, loss=1.307, acc=0.700, acc_len=1.67, thru=16.8, I=19.4, T=30.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:43,313] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:48,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:34:49,403] loop.py:45 INFO TRAIN_STEP step=12041 loss=1.228412 Training: 67%|██████▋ | 12041/18018 [26:04:35<12:48:54, 7.72s/step, loss=1.228, acc=0.712, acc_len=1.72, thru=17.6, I=26.0, T=34.5, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:34:49,398] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12040: forward=2607.3ms backward=1050.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:49,469] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:57,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:34:59,255] loop.py:45 INFO TRAIN_STEP step=12042 loss=1.258512 Training: 67%|██████▋ | 12042/18018 [26:04:45<13:52:31, 8.36s/step, loss=1.259, acc=0.711, acc_len=1.72, thru=13.7, I=14.0, T=33.3, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:34:59,737] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:03,943] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:35:05,968] loop.py:45 INFO TRAIN_STEP step=12043 loss=1.220887 Training: 67%|██████▋ | 12043/18018 [26:04:52<13:03:17, 7.87s/step, loss=1.221, acc=0.716, acc_len=1.73, thru=18.4, I=25.9, T=34.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:06,461] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:11,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:35:12,168] loop.py:45 INFO TRAIN_STEP step=12044 loss=1.379224 Training: 67%|██████▋ | 12044/18018 [26:04:58<12:13:23, 7.37s/step, loss=1.379, acc=0.685, acc_len=1.63, thru=19.8, I=29.1, T=36.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:12,401] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:17,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:35:18,718] loop.py:380 INFO TIMING step=12045: step=3.941s data=1.778s compute=3.911s [fwd=2.785s bwd=1.111s opt=0.003s] dispatch=2.599s [2026-10-19 15:35:18,719] loop.py:45 INFO TRAIN_STEP step=12045 loss=1.291933 Training: 67%|██████▋ | 12045/18018 [26:05:04<11:48:54, 7.12s/step, loss=1.292, acc=0.700, acc_len=1.68, thru=19.2, I=23.5, T=32.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:18,823] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:24,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:35:25,154] loop.py:45 INFO TRAIN_STEP step=12046 loss=1.348950 Training: 67%|██████▋ | 12046/18018 [26:05:11<11:28:19, 6.92s/step, loss=1.349, acc=0.691, acc_len=1.65, thru=18.9, I=22.5, T=33.6, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:35:25,151] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12045: forward=2733.9ms backward=1022.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:25,420] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:31,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:35:32,216] loop.py:45 INFO TRAIN_STEP step=12047 loss=1.216214 Training: 67%|██████▋ | 12047/18018 [26:05:18<11:32:34, 6.96s/step, loss=1.216, acc=0.713, acc_len=1.73, thru=17.3, I=19.9, T=34.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:32,598] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:39,274] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:35:40,590] loop.py:45 INFO TRAIN_STEP step=12048 loss=1.346813 Training: 67%|██████▋ | 12048/18018 [26:05:26<12:14:41, 7.38s/step, loss=1.347, acc=0.691, acc_len=1.64, thru=15.1, I=18.3, T=30.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:40,784] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:35:47,803] loop.py:45 INFO TRAIN_STEP step=12049 loss=1.291513 Training: 67%|██████▋ | 12049/18018 [26:05:33<12:09:25, 7.33s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=14.8, I=18.2, T=31.9, wait=3.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:56,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:35:56,452] loop.py:380 INFO TIMING step=12050: step=3.991s data=1.518s compute=3.953s [fwd=2.935s bwd=1.002s opt=0.003s] dispatch=4.649s [2026-10-19 15:35:56,452] loop.py:45 INFO TRAIN_STEP step=12050 loss=1.284092 Training: 67%|██████▋ | 12050/18018 [26:05:42<12:48:35, 7.73s/step, loss=1.284, acc=0.704, acc_len=1.70, thru=17.3, I=16.8, T=32.1, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:35:56,539] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:02,128] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:36:03,246] loop.py:45 INFO TRAIN_STEP step=12051 loss=1.335160 Training: 67%|██████▋ | 12051/18018 [26:05:49<12:20:37, 7.45s/step, loss=1.335, acc=0.693, acc_len=1.66, thru=18.5, I=20.4, T=29.6, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:36:03,241] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12050: forward=3225.8ms backward=1042.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:03,634] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:08,079] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:36:08,942] loop.py:45 INFO TRAIN_STEP step=12052 loss=1.253031 Training: 67%|██████▋ | 12052/18018 [26:05:55<11:28:18, 6.92s/step, loss=1.253, acc=0.709, acc_len=1.72, thru=19.6, I=24.2, T=37.1, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:09,099] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:15,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:36:16,508] loop.py:45 INFO TRAIN_STEP step=12053 loss=1.326655 Training: 67%|██████▋ | 12053/18018 [26:06:02<11:47:22, 7.12s/step, loss=1.327, acc=0.693, acc_len=1.67, thru=17.0, I=21.3, T=36.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:16,855] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:36:21,965] loop.py:45 INFO TRAIN_STEP step=12054 loss=1.200224 Training: 67%|██████▋ | 12054/18018 [26:06:08<10:57:46, 6.62s/step, loss=1.200, acc=0.715, acc_len=1.74, thru=20.2, I=23.9, T=38.6, wait=2.1s, pool=53, epoch=2/2][2026-10-19 15:36:30,226] loop.py:380 INFO TIMING step=12055: step=4.330s data=2.088s compute=4.294s [fwd=3.118s bwd=1.161s opt=0.003s] dispatch=3.921s [2026-10-19 15:36:30,226] loop.py:45 INFO TRAIN_STEP step=12055 loss=1.386146 Training: 67%|██████▋ | 12055/18018 [26:06:16<11:46:43, 7.11s/step, loss=1.386, acc=0.683, acc_len=1.62, thru=15.2, I=15.3, T=29.6, wait=3.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:36,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:36:37,566] loop.py:45 INFO TRAIN_STEP step=12056 loss=1.349028 Training: 67%|██████▋ | 12056/18018 [26:06:23<11:53:21, 7.18s/step, loss=1.349, acc=0.694, acc_len=1.65, thru=19.4, I=17.1, T=29.1, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:36:37,561] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12055: forward=3305.1ms backward=1051.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:37,824] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:36:44,897] loop.py:45 INFO TRAIN_STEP step=12057 loss=1.284219 Training: 67%|██████▋ | 12057/18018 [26:06:30<11:57:46, 7.22s/step, loss=1.284, acc=0.702, acc_len=1.70, thru=15.1, I=16.6, T=32.3, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:50,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:36:52,206] loop.py:45 INFO TRAIN_STEP step=12058 loss=1.404262 Training: 67%|██████▋ | 12058/18018 [26:06:38<12:00:13, 7.25s/step, loss=1.404, acc=0.688, acc_len=1.64, thru=18.2, I=19.8, T=33.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:52,383] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:57,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:36:57,819] loop.py:45 INFO TRAIN_STEP step=12059 loss=1.249207 Training: 67%|██████▋ | 12059/18018 [26:06:43<11:11:19, 6.76s/step, loss=1.249, acc=0.708, acc_len=1.71, thru=20.6, I=25.6, T=37.0, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:36:57,853] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 15:37:05,881] loop.py:380 INFO TIMING step=12060: step=3.559s data=1.377s compute=3.531s [fwd=2.501s bwd=1.016s opt=0.003s] dispatch=4.493s [2026-10-19 15:37:05,881] loop.py:45 INFO TRAIN_STEP step=12060 loss=1.408482 Training: 67%|██████▋ | 12060/18018 [26:06:51<11:49:54, 7.15s/step, loss=1.408, acc=0.679, acc_len=1.61, thru=16.1, I=16.4, T=36.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:05,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:06,381] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:11,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:37:12,164] loop.py:45 INFO TRAIN_STEP step=12061 loss=1.245746 Training: 67%|██████▋ | 12061/18018 [26:06:58<11:24:01, 6.89s/step, loss=1.246, acc=0.712, acc_len=1.72, thru=19.8, I=21.4, T=34.4, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:37:12,159] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12060: forward=2680.1ms backward=995.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:12,315] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:17,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:37:18,027] loop.py:45 INFO TRAIN_STEP step=12062 loss=1.244232 Training: 67%|██████▋ | 12062/18018 [26:07:04<10:53:20, 6.58s/step, loss=1.244, acc=0.708, acc_len=1.71, thru=20.9, I=25.2, T=38.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:18,435] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:37:27,843] loop.py:45 INFO TRAIN_STEP step=12063 loss=1.420568 Training: 67%|██████▋ | 12063/18018 [26:07:13<12:29:32, 7.55s/step, loss=1.421, acc=0.681, acc_len=1.62, thru=11.7, I=12.3, T=35.1, wait=6.2s, pool=49, epoch=2/2][2026-10-19 15:37:34,911] loop.py:45 INFO TRAIN_STEP step=12064 loss=1.261446 Training: 67%|██████▋ | 12064/18018 [26:07:20<12:14:58, 7.41s/step, loss=1.261, acc=0.708, acc_len=1.70, thru=16.1, I=13.3, T=35.3, wait=3.4s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:41,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:37:42,618] loop.py:380 INFO TIMING step=12065: step=3.694s data=1.508s compute=3.662s [fwd=2.632s bwd=1.016s opt=0.003s] dispatch=4.005s [2026-10-19 15:37:42,618] loop.py:45 INFO TRAIN_STEP step=12065 loss=1.336249 Training: 67%|██████▋ | 12065/18018 [26:07:28<12:23:50, 7.50s/step, loss=1.336, acc=0.698, acc_len=1.68, thru=17.8, I=19.0, T=34.6, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:42,814] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:37:50,841] loop.py:45 INFO TRAIN_STEP step=12066 loss=1.206838 Training: 67%|██████▋ | 12066/18018 [26:07:36<12:45:22, 7.72s/step, loss=1.207, acc=0.717, acc_len=1.74, thru=10.4, I=14.9, T=35.2, wait=4.6s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:37:50,836] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12065: forward=2595.6ms backward=993.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:58,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:37:58,831] loop.py:45 INFO TRAIN_STEP step=12067 loss=1.259753 Training: 67%|██████▋ | 12067/18018 [26:07:44<12:53:23, 7.80s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=18.6, I=16.9, T=35.4, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:37:59,109] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:06,666] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:38:07,083] loop.py:45 INFO TRAIN_STEP step=12068 loss=1.261010 Training: 67%|██████▋ | 12068/18018 [26:07:53<13:06:43, 7.93s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=15.1, I=16.2, T=35.0, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:07,168] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 15:38:12,881] loop.py:45 INFO TRAIN_STEP step=12069 loss=1.356646 Training: 67%|██████▋ | 12069/18018 [26:07:58<12:03:10, 7.29s/step, loss=1.357, acc=0.692, acc_len=1.66, thru=18.3, I=21.8, T=35.3, wait=2.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:24,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:38:25,093] loop.py:380 INFO TIMING step=12070: step=4.692s data=2.465s compute=4.654s [fwd=3.571s bwd=1.067s opt=0.003s] dispatch=7.509s [2026-10-19 15:38:25,093] loop.py:45 INFO TRAIN_STEP step=12070 loss=1.298458 Training: 67%|██████▋ | 12070/18018 [26:08:11<14:29:17, 8.77s/step, loss=1.298, acc=0.697, acc_len=1.67, thru=11.5, I=11.5, T=27.3, wait=7.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:25,330] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:38:31,609] loop.py:45 INFO TRAIN_STEP step=12071 loss=1.227707 Training: 67%|██████▋ | 12071/18018 [26:08:17<13:22:07, 8.09s/step, loss=1.228, acc=0.711, acc_len=1.72, thru=16.8, I=17.6, T=31.7, wait=2.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:38:31,604] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12070: forward=2839.8ms backward=1142.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:37,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:38:38,460] loop.py:45 INFO TRAIN_STEP step=12072 loss=1.158134 Training: 67%|██████▋ | 12072/18018 [26:08:24<12:45:05, 7.72s/step, loss=1.158, acc=0.726, acc_len=1.77, thru=18.5, I=18.9, T=30.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:38,758] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:46,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:38:46,545] loop.py:45 INFO TRAIN_STEP step=12073 loss=1.261399 Training: 67%|██████▋ | 12073/18018 [26:08:32<12:55:48, 7.83s/step, loss=1.261, acc=0.704, acc_len=1.69, thru=15.3, I=16.9, T=35.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:46,579] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:51,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:38:52,507] loop.py:45 INFO TRAIN_STEP step=12074 loss=1.256712 Training: 67%|██████▋ | 12074/18018 [26:08:38<12:00:09, 7.27s/step, loss=1.257, acc=0.708, acc_len=1.71, thru=20.3, I=22.0, T=36.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:52,537] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:58,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:38:59,141] loop.py:380 INFO TIMING step=12075: step=3.471s data=1.443s compute=3.427s [fwd=2.430s bwd=0.982s opt=0.003s] dispatch=3.153s [2026-10-19 15:38:59,141] loop.py:45 INFO TRAIN_STEP step=12075 loss=1.250374 Training: 67%|██████▋ | 12075/18018 [26:08:45<11:41:04, 7.08s/step, loss=1.250, acc=0.706, acc_len=1.70, thru=19.0, I=23.7, T=36.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:38:59,596] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:04,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:39:06,346] loop.py:45 INFO TRAIN_STEP step=12076 loss=1.407614 Training: 67%|██████▋ | 12076/18018 [26:08:52<11:44:47, 7.12s/step, loss=1.408, acc=0.681, acc_len=1.61, thru=18.0, I=24.2, T=35.2, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:39:06,341] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12075: forward=2567.0ms backward=1021.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:06,826] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:39:12,986] loop.py:45 INFO TRAIN_STEP step=12077 loss=1.292707 Training: 67%|██████▋ | 12077/18018 [26:08:59<11:30:30, 6.97s/step, loss=1.293, acc=0.701, acc_len=1.69, thru=17.5, I=23.4, T=34.5, wait=2.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:19,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:39:20,940] loop.py:45 INFO TRAIN_STEP step=12078 loss=1.225111 Training: 67%|██████▋ | 12078/18018 [26:09:07<11:59:31, 7.27s/step, loss=1.225, acc=0.715, acc_len=1.74, thru=17.2, I=19.3, T=32.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:20,979] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:26,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:39:26,747] loop.py:45 INFO TRAIN_STEP step=12079 loss=1.232484 Training: 67%|██████▋ | 12079/18018 [26:09:12<11:16:01, 6.83s/step, loss=1.232, acc=0.712, acc_len=1.72, thru=19.9, I=24.5, T=37.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:26,890] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 15:39:33,068] loop.py:380 INFO TIMING step=12080: step=3.706s data=1.626s compute=3.673s [fwd=2.616s bwd=1.041s opt=0.003s] dispatch=2.605s [2026-10-19 15:39:33,068] loop.py:45 INFO TRAIN_STEP step=12080 loss=1.374263 Training: 67%|██████▋ | 12080/18018 [26:09:19<11:00:44, 6.68s/step, loss=1.374, acc=0.692, acc_len=1.65, thru=18.4, I=21.3, T=34.5, wait=2.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:39,485] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:39:41,128] loop.py:45 INFO TRAIN_STEP step=12081 loss=1.249925 Training: 67%|██████▋ | 12081/18018 [26:09:27<11:41:42, 7.09s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=16.5, I=18.3, T=31.7, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:39:41,123] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12080: forward=2815.5ms backward=1179.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:41,500] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:47,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:39:48,196] loop.py:45 INFO TRAIN_STEP step=12082 loss=1.377806 Training: 67%|██████▋ | 12082/18018 [26:09:34<11:40:57, 7.09s/step, loss=1.378, acc=0.685, acc_len=1.64, thru=17.3, I=23.0, T=37.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:48,455] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:54,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:39:55,458] loop.py:45 INFO TRAIN_STEP step=12083 loss=1.344436 Training: 67%|██████▋ | 12083/18018 [26:09:41<11:46:03, 7.14s/step, loss=1.344, acc=0.692, acc_len=1.66, thru=17.4, I=20.5, T=36.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:39:55,492] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:03,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:40:03,358] loop.py:45 INFO TRAIN_STEP step=12084 loss=1.375013 Training: 67%|██████▋ | 12084/18018 [26:09:49<12:08:30, 7.37s/step, loss=1.375, acc=0.689, acc_len=1.64, thru=17.2, I=17.6, T=34.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:03,565] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:09,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:40:09,722] loop.py:380 INFO TIMING step=12085: step=3.423s data=1.433s compute=3.390s [fwd=2.444s bwd=0.932s opt=0.003s] dispatch=2.932s [2026-10-19 15:40:09,722] loop.py:45 INFO TRAIN_STEP step=12085 loss=1.392395 Training: 67%|██████▋ | 12085/18018 [26:09:55<11:38:39, 7.07s/step, loss=1.392, acc=0.685, acc_len=1.63, thru=18.9, I=19.3, T=37.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:10,129] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:17,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:40:18,132] loop.py:45 INFO TRAIN_STEP step=12086 loss=1.466131 Training: 67%|██████▋ | 12086/18018 [26:10:04<12:18:28, 7.47s/step, loss=1.466, acc=0.672, acc_len=1.59, thru=16.9, I=17.9, T=34.1, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:40:18,127] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12085: forward=2595.8ms backward=1082.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:18,205] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:24,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:40:25,184] loop.py:45 INFO TRAIN_STEP step=12087 loss=1.329664 Training: 67%|██████▋ | 12087/18018 [26:10:11<12:05:56, 7.34s/step, loss=1.330, acc=0.697, acc_len=1.67, thru=18.2, I=20.8, T=33.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:25,550] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:30,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:40:31,216] loop.py:45 INFO TRAIN_STEP step=12088 loss=1.175246 Training: 67%|██████▋ | 12088/18018 [26:10:17<11:26:57, 6.95s/step, loss=1.175, acc=0.722, acc_len=1.76, thru=20.3, I=24.7, T=37.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:31,250] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:37,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:40:38,001] loop.py:45 INFO TRAIN_STEP step=12089 loss=1.244399 Training: 67%|██████▋ | 12089/18018 [26:10:24<11:21:52, 6.90s/step, loss=1.244, acc=0.709, acc_len=1.70, thru=19.2, I=20.4, T=32.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:38,220] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:44,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:40:44,643] loop.py:380 INFO TIMING step=12090: step=3.535s data=1.432s compute=3.504s [fwd=2.523s bwd=0.967s opt=0.003s] dispatch=3.098s [2026-10-19 15:40:44,643] loop.py:45 INFO TRAIN_STEP step=12090 loss=1.293527 Training: 67%|██████▋ | 12090/18018 [26:10:30<11:14:05, 6.82s/step, loss=1.294, acc=0.700, acc_len=1.69, thru=18.6, I=19.3, T=36.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:40:45,030] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:40:52,154] loop.py:45 INFO TRAIN_STEP step=12091 loss=1.335405 Training: 67%|██████▋ | 12091/18018 [26:10:38<11:34:23, 7.03s/step, loss=1.335, acc=0.694, acc_len=1.66, thru=14.2, I=16.8, T=35.9, wait=3.9s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:40:52,149] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12090: forward=2499.2ms backward=1009.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:00,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:41:00,684] loop.py:45 INFO TRAIN_STEP step=12092 loss=1.259899 Training: 67%|██████▋ | 12092/18018 [26:10:46<12:18:43, 7.48s/step, loss=1.260, acc=0.709, acc_len=1.71, thru=16.7, I=17.0, T=33.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:01,171] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:06,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:41:08,425] loop.py:45 INFO TRAIN_STEP step=12093 loss=1.369669 Training: 67%|██████▋ | 12093/18018 [26:10:54<12:26:21, 7.56s/step, loss=1.370, acc=0.686, acc_len=1.64, thru=16.8, I=21.9, T=32.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:08,813] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:13,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:41:14,694] loop.py:45 INFO TRAIN_STEP step=12094 loss=1.224603 Training: 67%|██████▋ | 12094/18018 [26:11:00<11:48:06, 7.17s/step, loss=1.225, acc=0.712, acc_len=1.73, thru=18.5, I=26.9, T=34.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:14,907] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:20,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:41:21,637] loop.py:380 INFO TIMING step=12095: step=3.678s data=1.535s compute=3.640s [fwd=2.612s bwd=1.013s opt=0.003s] dispatch=3.255s [2026-10-19 15:41:21,638] loop.py:45 INFO TRAIN_STEP step=12095 loss=1.234411 Training: 67%|██████▋ | 12095/18018 [26:11:07<11:41:11, 7.10s/step, loss=1.234, acc=0.709, acc_len=1.71, thru=17.4, I=21.5, T=34.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:21,893] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:27,821] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:41:28,138] loop.py:45 INFO TRAIN_STEP step=12096 loss=1.358799 Training: 67%|██████▋ | 12096/18018 [26:11:14<11:23:12, 6.92s/step, loss=1.359, acc=0.688, acc_len=1.64, thru=19.1, I=23.3, T=38.4, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:41:28,134] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12095: forward=2388.5ms backward=902.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:28,339] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:41:34,431] loop.py:45 INFO TRAIN_STEP step=12097 loss=1.357654 Training: 67%|██████▋ | 12097/18018 [26:11:20<11:04:31, 6.73s/step, loss=1.358, acc=0.690, acc_len=1.64, thru=18.0, I=19.6, T=35.1, wait=2.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:42,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:41:43,294] loop.py:45 INFO TRAIN_STEP step=12098 loss=1.425802 Training: 67%|██████▋ | 12098/18018 [26:11:29<12:07:24, 7.37s/step, loss=1.426, acc=0.680, acc_len=1.62, thru=17.2, I=17.6, T=34.7, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:43,499] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:49,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:41:49,928] loop.py:45 INFO TRAIN_STEP step=12099 loss=1.313991 Training: 67%|██████▋ | 12099/18018 [26:11:36<11:45:23, 7.15s/step, loss=1.314, acc=0.696, acc_len=1.67, thru=18.8, I=21.9, T=34.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:50,217] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:55,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:41:57,141] loop.py:380 INFO TIMING step=12100: step=3.824s data=1.713s compute=3.793s [fwd=2.714s bwd=1.065s opt=0.003s] dispatch=3.380s [2026-10-19 15:41:57,141] loop.py:45 INFO TRAIN_STEP step=12100 loss=1.209536 Training: 67%|██████▋ | 12100/18018 [26:11:43<11:47:04, 7.17s/step, loss=1.210, acc=0.718, acc_len=1.74, thru=17.6, I=21.0, T=33.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:41:57,166] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 15:42:04,486] loop.py:45 INFO TRAIN_STEP step=12101 loss=1.277605 Training: 67%|██████▋ | 12101/18018 [26:11:50<11:52:13, 7.22s/step, loss=1.278, acc=0.704, acc_len=1.69, thru=16.1, I=18.5, T=37.3, wait=3.9s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:42:04,482] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12100: forward=2410.8ms backward=967.5ms [2026-10-19 15:42:11,411] loop.py:45 INFO TRAIN_STEP step=12102 loss=1.333934 Training: 67%|██████▋ | 12102/18018 [26:11:57<11:43:23, 7.13s/step, loss=1.334, acc=0.697, acc_len=1.67, thru=18.8, I=18.2, T=36.9, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:16,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:42:17,760] loop.py:45 INFO TRAIN_STEP step=12103 loss=1.333174 Training: 67%|██████▋ | 12103/18018 [26:12:03<11:20:01, 6.90s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=19.9, I=20.9, T=37.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:17,873] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:24,258] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:42:26,286] loop.py:45 INFO TRAIN_STEP step=12104 loss=1.387221 Training: 67%|██████▋ | 12104/18018 [26:12:12<12:08:00, 7.39s/step, loss=1.387, acc=0.684, acc_len=1.63, thru=15.1, I=21.4, T=32.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:26,775] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 15:42:32,895] loop.py:380 INFO TIMING step=12105: step=3.648s data=1.600s compute=3.624s [fwd=2.611s bwd=0.998s opt=0.003s] dispatch=2.952s [2026-10-19 15:42:32,895] loop.py:45 INFO TRAIN_STEP step=12105 loss=1.226424 Training: 67%|██████▋ | 12105/18018 [26:12:18<11:44:54, 7.15s/step, loss=1.226, acc=0.716, acc_len=1.74, thru=17.7, I=23.8, T=35.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:32,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:33,363] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:38,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:42:39,762] loop.py:45 INFO TRAIN_STEP step=12106 loss=1.183962 Training: 67%|██████▋ | 12106/18018 [26:12:25<11:36:20, 7.07s/step, loss=1.184, acc=0.720, acc_len=1.76, thru=18.8, I=21.2, T=33.1, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:42:39,758] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12105: forward=2741.7ms backward=1077.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:40,224] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:46,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:42:46,729] loop.py:45 INFO TRAIN_STEP step=12107 loss=1.181103 Training: 67%|██████▋ | 12107/18018 [26:12:32<11:33:15, 7.04s/step, loss=1.181, acc=0.725, acc_len=1.77, thru=17.2, I=21.0, T=34.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:42:46,849] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 15:42:54,540] loop.py:45 INFO TRAIN_STEP step=12108 loss=1.267226 Training: 67%|██████▋ | 12108/18018 [26:12:40<11:56:03, 7.27s/step, loss=1.267, acc=0.708, acc_len=1.72, thru=13.8, I=15.3, T=33.0, wait=3.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:01,241] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:43:01,466] loop.py:45 INFO TRAIN_STEP step=12109 loss=1.217800 Training: 67%|██████▋ | 12109/18018 [26:12:47<11:45:43, 7.17s/step, loss=1.218, acc=0.719, acc_len=1.75, thru=17.9, I=17.3, T=35.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:01,743] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:07,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:43:09,077] loop.py:380 INFO TIMING step=12110: step=3.614s data=1.540s compute=3.568s [fwd=2.515s bwd=1.038s opt=0.003s] dispatch=3.989s [2026-10-19 15:43:09,077] loop.py:45 INFO TRAIN_STEP step=12110 loss=1.402720 Training: 67%|██████▋ | 12110/18018 [26:12:55<11:58:45, 7.30s/step, loss=1.403, acc=0.681, acc_len=1.61, thru=16.9, I=19.3, T=35.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:09,359] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:14,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:43:15,201] loop.py:45 INFO TRAIN_STEP step=12111 loss=1.155154 Training: 67%|██████▋ | 12111/18018 [26:13:01<11:23:59, 6.95s/step, loss=1.155, acc=0.724, acc_len=1.77, thru=20.1, I=22.4, T=35.6, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:43:15,197] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12110: forward=2590.0ms backward=968.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:15,259] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:19,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:43:20,611] loop.py:45 INFO TRAIN_STEP step=12112 loss=1.411833 Training: 67%|██████▋ | 12112/18018 [26:13:06<10:38:27, 6.49s/step, loss=1.412, acc=0.679, acc_len=1.61, thru=22.2, I=26.6, T=39.0, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:20,701] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:25,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:43:26,220] loop.py:45 INFO TRAIN_STEP step=12113 loss=1.133996 Training: 67%|██████▋ | 12113/18018 [26:13:12<10:12:27, 6.22s/step, loss=1.134, acc=0.728, acc_len=1.78, thru=22.3, I=26.4, T=37.8, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:26,643] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:34,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:43:35,273] loop.py:45 INFO TRAIN_STEP step=12114 loss=1.302463 Training: 67%|██████▋ | 12114/18018 [26:13:21<11:35:53, 7.07s/step, loss=1.302, acc=0.699, acc_len=1.69, thru=14.8, I=15.6, T=33.8, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:35,461] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:40,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:43:42,019] loop.py:380 INFO TIMING step=12115: step=3.441s data=1.390s compute=3.416s [fwd=2.446s bwd=0.955s opt=0.003s] dispatch=3.294s [2026-10-19 15:43:42,019] loop.py:45 INFO TRAIN_STEP step=12115 loss=1.298628 Training: 67%|██████▋ | 12115/18018 [26:13:28<11:26:06, 6.97s/step, loss=1.299, acc=0.702, acc_len=1.68, thru=17.8, I=21.3, T=37.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:42,397] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:46,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:43:47,453] loop.py:45 INFO TRAIN_STEP step=12116 loss=1.370785 Training: 67%|██████▋ | 12116/18018 [26:13:33<10:40:34, 6.51s/step, loss=1.371, acc=0.690, acc_len=1.64, thru=22.2, I=28.2, T=36.7, wait=1.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:43:47,449] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12115: forward=2452.5ms backward=988.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:47,569] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:53,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:43:54,791] loop.py:45 INFO TRAIN_STEP step=12117 loss=1.395601 Training: 67%|██████▋ | 12117/18018 [26:13:40<11:04:51, 6.76s/step, loss=1.396, acc=0.684, acc_len=1.62, thru=17.1, I=20.7, T=36.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:43:54,843] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 15:44:01,719] loop.py:45 INFO TRAIN_STEP step=12118 loss=1.272094 Training: 67%|██████▋ | 12118/18018 [26:13:47<11:09:40, 6.81s/step, loss=1.272, acc=0.706, acc_len=1.70, thru=14.9, I=17.5, T=40.2, wait=3.7s, pool=29, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 15:44:01,797] eagle_store.py:517 WARNING batch_get_buffer missing keys (k387778-91e9d705_hs, k387778-91e9d705_ids, k387778-91e9d705_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:10,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 15:44:01,789] eagle_store.py:517 WARNING batch_get_buffer missing keys (k387777-967689ac_hs, k387777-967689ac_ids, k387777-967689ac_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 15:44:11,278] loop.py:45 INFO TRAIN_STEP step=12119 loss=1.366192 Training: 67%|██████▋ | 12119/18018 [26:13:57<12:30:38, 7.63s/step, loss=1.366, acc=0.687, acc_len=1.63, thru=17.7, I=16.6, T=32.4, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:11,671] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:44:19,688] loop.py:380 INFO TIMING step=12120: step=3.634s data=1.447s compute=3.598s [fwd=2.492s bwd=1.091s opt=0.003s] dispatch=4.766s [2026-10-19 15:44:19,689] loop.py:45 INFO TRAIN_STEP step=12120 loss=1.210277 Training: 67%|██████▋ | 12120/18018 [26:14:05<12:53:22, 7.87s/step, loss=1.210, acc=0.720, acc_len=1.75, thru=13.8, I=16.1, T=35.2, wait=4.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:24,840] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:44:26,240] loop.py:45 INFO TRAIN_STEP step=12121 loss=1.221125 Training: 67%|██████▋ | 12121/18018 [26:14:12<12:14:28, 7.47s/step, loss=1.221, acc=0.715, acc_len=1.74, thru=19.0, I=21.6, T=35.9, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:44:26,236] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12120: forward=2527.9ms backward=991.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:26,359] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:33,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:44:34,895] loop.py:45 INFO TRAIN_STEP step=12122 loss=1.327881 Training: 67%|██████▋ | 12122/18018 [26:14:20<12:49:08, 7.83s/step, loss=1.328, acc=0.695, acc_len=1.66, thru=15.0, I=18.0, T=30.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:35,074] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:40,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:44:41,136] loop.py:45 INFO TRAIN_STEP step=12123 loss=1.217027 Training: 67%|██████▋ | 12123/18018 [26:14:27<12:02:14, 7.35s/step, loss=1.217, acc=0.715, acc_len=1.73, thru=19.4, I=20.0, T=35.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:41,487] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:48,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:44:49,225] loop.py:45 INFO TRAIN_STEP step=12124 loss=1.354421 Training: 67%|██████▋ | 12124/18018 [26:14:35<12:23:53, 7.57s/step, loss=1.354, acc=0.693, acc_len=1.65, thru=16.4, I=17.2, T=35.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:49,356] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:55,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:44:56,076] loop.py:380 INFO TIMING step=12125: step=3.546s data=1.484s compute=3.510s [fwd=2.483s bwd=1.014s opt=0.003s] dispatch=3.296s [2026-10-19 15:44:56,077] loop.py:45 INFO TRAIN_STEP step=12125 loss=1.345340 Training: 67%|██████▋ | 12125/18018 [26:14:42<12:02:33, 7.36s/step, loss=1.345, acc=0.689, acc_len=1.65, thru=18.5, I=20.8, T=36.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:44:56,391] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:45:04,741] loop.py:45 INFO TRAIN_STEP step=12126 loss=1.244384 Training: 67%|██████▋ | 12126/18018 [26:14:50<12:40:53, 7.75s/step, loss=1.244, acc=0.710, acc_len=1.72, thru=13.8, I=14.0, T=33.2, wait=4.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:45:04,736] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12125: forward=2738.1ms backward=1065.4ms [2026-10-19 15:45:11,158] loop.py:45 INFO TRAIN_STEP step=12127 loss=1.298784 Training: 67%|██████▋ | 12127/18018 [26:14:57<12:01:31, 7.35s/step, loss=1.299, acc=0.702, acc_len=1.70, thru=18.8, I=18.4, T=36.9, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:45:16,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:45:17,131] loop.py:45 INFO TRAIN_STEP step=12128 loss=1.263168 Training: 67%|██████▋ | 12128/18018 [26:15:03<11:20:57, 6.94s/step, loss=1.263, acc=0.708, acc_len=1.71, thru=20.9, I=21.4, T=34.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:45:17,377] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:45:24,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:45:24,561] loop.py:45 INFO TRAIN_STEP step=12129 loss=1.308489 Training: 67%|██████▋ | 12129/18018 [26:15:10<11:35:19, 7.08s/step, loss=1.308, acc=0.697, acc_len=1.67, thru=16.3, I=17.7, T=31.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:45:24,795] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:45:30,371] loop.py:380 INFO TIMING step=12130: step=3.614s data=1.554s compute=3.584s [fwd=2.586s bwd=0.985s opt=0.003s] dispatch=2.188s [2026-10-19 15:45:30,371] loop.py:45 INFO TRAIN_STEP step=12130 loss=1.225542 Training: 67%|██████▋ | 12130/18018 [26:15:16<10:57:42, 6.70s/step, loss=1.226, acc=0.715, acc_len=1.73, thru=16.1, I=23.4, T=35.4, wait=2.2s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:45:40,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:45:41,296] loop.py:45 INFO TRAIN_STEP step=12131 loss=1.193882 Training: 67%|██████▋ | 12131/18018 [26:15:27<13:01:53, 7.97s/step, loss=1.194, acc=0.723, acc_len=1.76, thru=18.4, I=13.7, T=33.7, wait=7.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:45:41,291] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12130: forward=2710.5ms backward=1035.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:45:41,652] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:45:50,548] loop.py:45 INFO TRAIN_STEP step=12132 loss=1.355166 Training: 67%|██████▋ | 12132/18018 [26:15:36<13:39:28, 8.35s/step, loss=1.355, acc=0.690, acc_len=1.65, thru=10.8, I=10.5, T=36.8, wait=5.8s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:45:57,999] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:45:58,522] loop.py:45 INFO TRAIN_STEP step=12133 loss=1.233434 Training: 67%|██████▋ | 12133/18018 [26:15:44<13:28:09, 8.24s/step, loss=1.233, acc=0.712, acc_len=1.72, thru=19.2, I=14.7, T=37.5, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:45:59,004] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:03,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:46:04,702] loop.py:45 INFO TRAIN_STEP step=12134 loss=1.278746 Training: 67%|██████▋ | 12134/18018 [26:15:50<12:27:31, 7.62s/step, loss=1.279, acc=0.704, acc_len=1.69, thru=20.6, I=23.7, T=36.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:04,988] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:10,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:46:10,650] loop.py:380 INFO TIMING step=12135: step=3.532s data=1.544s compute=3.506s [fwd=2.515s bwd=0.977s opt=0.003s] dispatch=2.405s [2026-10-19 15:46:10,650] loop.py:45 INFO TRAIN_STEP step=12135 loss=1.324878 Training: 67%|██████▋ | 12135/18018 [26:15:56<11:38:07, 7.12s/step, loss=1.325, acc=0.700, acc_len=1.68, thru=21.3, I=24.1, T=36.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:10,797] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:16,927] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:46:19,015] loop.py:45 INFO TRAIN_STEP step=12136 loss=1.215210 Training: 67%|██████▋ | 12136/18018 [26:16:05<12:14:34, 7.49s/step, loss=1.215, acc=0.713, acc_len=1.73, thru=15.7, I=18.2, T=30.2, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:46:19,011] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12135: forward=3169.2ms backward=1011.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:19,440] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:25,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:46:26,250] loop.py:45 INFO TRAIN_STEP step=12137 loss=1.286965 Training: 67%|██████▋ | 12137/18018 [26:16:12<12:06:55, 7.42s/step, loss=1.287, acc=0.701, acc_len=1.68, thru=15.0, I=19.5, T=35.1, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:26,347] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:33,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:46:34,069] loop.py:45 INFO TRAIN_STEP step=12138 loss=1.176050 Training: 67%|██████▋ | 12138/18018 [26:16:20<12:18:36, 7.54s/step, loss=1.176, acc=0.718, acc_len=1.74, thru=15.9, I=18.2, T=36.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:34,250] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:41,851] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:46:43,031] loop.py:45 INFO TRAIN_STEP step=12139 loss=1.259603 Training: 67%|██████▋ | 12139/18018 [26:16:29<13:00:22, 7.96s/step, loss=1.260, acc=0.714, acc_len=1.73, thru=14.1, I=16.1, T=31.8, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:46:43,364] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:46:49,218] loop.py:380 INFO TIMING step=12140: step=3.512s data=1.505s compute=3.485s [fwd=2.492s bwd=0.980s opt=0.003s] dispatch=2.666s [2026-10-19 15:46:49,218] loop.py:45 INFO TRAIN_STEP step=12140 loss=1.227227 Training: 67%|██████▋ | 12140/18018 [26:16:35<12:07:59, 7.43s/step, loss=1.227, acc=0.714, acc_len=1.73, thru=18.6, I=22.5, T=36.4, wait=2.7s, pool=57, epoch=2/2][2026-10-19 15:46:56,161] loop.py:45 INFO TRAIN_STEP step=12141 loss=1.339833 (TrainerActor pid=9310) [2026-10-19 15:46:56,156] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12140: forward=2556.4ms backward=992.9ms Training: 67%|██████▋ | 12141/18018 [26:16:42<11:53:31, 7.28s/step, loss=1.340, acc=0.693, acc_len=1.66, thru=16.9, I=18.4, T=35.5, wait=3.3s, pool=49, epoch=2/2][2026-10-19 15:47:03,102] loop.py:45 INFO TRAIN_STEP step=12142 loss=1.269622 Training: 67%|██████▋ | 12142/18018 [26:16:49<11:43:17, 7.18s/step, loss=1.270, acc=0.701, acc_len=1.68, thru=18.8, I=18.4, T=36.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:03,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:03,602] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:08,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:47:08,623] loop.py:45 INFO TRAIN_STEP step=12143 loss=1.419477 Training: 67%|██████▋ | 12143/18018 [26:16:54<10:54:21, 6.68s/step, loss=1.419, acc=0.678, acc_len=1.61, thru=22.6, I=23.7, T=37.3, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:08,706] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:14,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:47:15,041] loop.py:45 INFO TRAIN_STEP step=12144 loss=1.157149 Training: 67%|██████▋ | 12144/18018 [26:17:01<10:46:35, 6.60s/step, loss=1.157, acc=0.724, acc_len=1.76, thru=20.1, I=23.3, T=35.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:15,317] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:47:22,968] loop.py:380 INFO TIMING step=12145: step=3.679s data=1.529s compute=3.632s [fwd=2.566s bwd=1.050s opt=0.003s] dispatch=4.239s [2026-10-19 15:47:22,969] loop.py:45 INFO TRAIN_STEP step=12145 loss=1.279807 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:22,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 67%|██████▋ | 12145/18018 [26:17:09<11:25:14, 7.00s/step, loss=1.280, acc=0.708, acc_len=1.70, thru=15.8, I=17.4, T=34.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:23,415] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:47:31,555] loop.py:45 INFO TRAIN_STEP step=12146 loss=1.197955 Training: 67%|██████▋ | 12146/18018 [26:17:17<12:11:48, 7.48s/step, loss=1.198, acc=0.718, acc_len=1.74, thru=15.2, I=15.9, T=32.1, wait=4.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:47:31,551] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12145: forward=2841.2ms backward=1092.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:36,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:47:38,642] loop.py:45 INFO TRAIN_STEP step=12147 loss=1.370849 Training: 67%|██████▋ | 12147/18018 [26:17:24<12:00:08, 7.36s/step, loss=1.371, acc=0.687, acc_len=1.63, thru=17.6, I=19.6, T=33.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:38,989] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:47:45,337] loop.py:45 INFO TRAIN_STEP step=12148 loss=1.341115 Training: 67%|██████▋ | 12148/18018 [26:17:31<11:40:28, 7.16s/step, loss=1.341, acc=0.691, acc_len=1.64, thru=16.1, I=20.3, T=35.4, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:54,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:47:54,460] loop.py:45 INFO TRAIN_STEP step=12149 loss=1.438093 Training: 67%|██████▋ | 12149/18018 [26:17:40<12:37:57, 7.75s/step, loss=1.438, acc=0.675, acc_len=1.59, thru=16.5, I=15.2, T=32.5, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:54,804] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:47:58,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:48:00,105] loop.py:380 INFO TIMING step=12150: step=3.250s data=1.304s compute=3.226s [fwd=2.308s bwd=0.903s opt=0.003s] dispatch=2.387s [2026-10-19 15:48:00,105] loop.py:45 INFO TRAIN_STEP step=12150 loss=1.312594 Training: 67%|██████▋ | 12150/18018 [26:17:46<11:36:07, 7.12s/step, loss=1.313, acc=0.700, acc_len=1.68, thru=20.3, I=22.0, T=39.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:00,466] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:04,927] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:48:06,670] loop.py:45 INFO TRAIN_STEP step=12151 loss=1.235379 Training: 67%|██████▋ | 12151/18018 [26:17:52<11:19:50, 6.95s/step, loss=1.235, acc=0.713, acc_len=1.72, thru=20.4, I=28.5, T=37.5, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:48:06,667] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12150: forward=2373.8ms backward=983.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:06,941] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:48:13,024] loop.py:45 INFO TRAIN_STEP step=12152 loss=1.196539 Training: 67%|██████▋ | 12152/18018 [26:17:59<11:02:11, 6.77s/step, loss=1.197, acc=0.717, acc_len=1.74, thru=20.6, I=27.2, T=35.6, wait=2.7s, pool=61, epoch=2/2][2026-10-19 15:48:20,758] loop.py:45 INFO TRAIN_STEP step=12153 loss=1.293093 Training: 67%|██████▋ | 12153/18018 [26:18:06<11:30:14, 7.06s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=14.6, I=16.0, T=35.7, wait=4.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:27,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:48:28,296] loop.py:45 INFO TRAIN_STEP step=12154 loss=1.328078 Training: 67%|██████▋ | 12154/18018 [26:18:14<11:44:01, 7.20s/step, loss=1.328, acc=0.695, acc_len=1.66, thru=17.4, I=18.5, T=34.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:28,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:34,544] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:48:35,202] loop.py:380 INFO TIMING step=12155: step=3.640s data=1.556s compute=3.613s [fwd=2.579s bwd=1.019s opt=0.003s] dispatch=3.256s [2026-10-19 15:48:35,202] loop.py:45 INFO TRAIN_STEP step=12155 loss=1.162741 Training: 67%|██████▋ | 12155/18018 [26:18:21<11:35:16, 7.12s/step, loss=1.163, acc=0.724, acc_len=1.76, thru=19.3, I=20.6, T=35.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:35,550] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:48:41,638] loop.py:45 INFO TRAIN_STEP step=12156 loss=1.314338 Training: 67%|██████▋ | 12156/18018 [26:18:27<11:15:08, 6.91s/step, loss=1.314, acc=0.697, acc_len=1.67, thru=18.4, I=19.2, T=34.4, wait=2.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:48:41,634] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12155: forward=2596.7ms backward=1085.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:48,377] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:48:48,553] loop.py:45 INFO TRAIN_STEP step=12157 loss=1.267590 Training: 67%|██████▋ | 12157/18018 [26:18:34<11:15:12, 6.91s/step, loss=1.268, acc=0.704, acc_len=1.69, thru=18.9, I=18.9, T=36.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:48,879] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:55,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:48:56,817] loop.py:45 INFO TRAIN_STEP step=12158 loss=1.311554 Training: 67%|██████▋ | 12158/18018 [26:18:42<11:54:43, 7.32s/step, loss=1.312, acc=0.701, acc_len=1.68, thru=15.6, I=18.8, T=33.9, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:48:56,905] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-19 15:49:03,671] loop.py:45 INFO TRAIN_STEP step=12159 loss=1.225588 Training: 67%|██████▋ | 12159/18018 [26:18:49<11:40:59, 7.18s/step, loss=1.226, acc=0.711, acc_len=1.72, thru=17.7, I=22.7, T=36.4, wait=3.3s, pool=57, epoch=2/2][2026-10-19 15:49:11,665] loop.py:380 INFO TIMING step=12160: step=3.513s data=1.385s compute=3.480s [fwd=2.484s bwd=0.983s opt=0.003s] dispatch=4.472s [2026-10-19 15:49:11,666] loop.py:45 INFO TRAIN_STEP step=12160 loss=1.369277 Training: 67%|██████▋ | 12160/18018 [26:18:57<12:04:44, 7.42s/step, loss=1.369, acc=0.691, acc_len=1.65, thru=13.7, I=14.3, T=36.4, wait=4.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:49:19,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:49:20,896] loop.py:45 INFO TRAIN_STEP step=12161 loss=1.300464 Training: 67%|██████▋ | 12161/18018 [26:19:06<12:57:34, 7.97s/step, loss=1.300, acc=0.702, acc_len=1.69, thru=16.2, I=16.7, T=33.1, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:49:20,891] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12160: forward=2765.2ms backward=1051.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:49:21,229] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:49:25,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:49:26,229] loop.py:45 INFO TRAIN_STEP step=12162 loss=1.279155 Training: 67%|██████▋ | 12162/18018 [26:19:12<11:40:22, 7.18s/step, loss=1.279, acc=0.703, acc_len=1.69, thru=20.8, I=24.5, T=37.1, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:49:26,729] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:49:34,342] loop.py:45 INFO TRAIN_STEP step=12163 loss=1.225806 Training: 68%|██████▊ | 12163/18018 [26:19:20<12:07:38, 7.46s/step, loss=1.226, acc=0.712, acc_len=1.73, thru=14.7, I=18.3, T=33.9, wait=4.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:49:42,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:49:43,489] loop.py:45 INFO TRAIN_STEP step=12164 loss=1.314729 Training: 68%|██████▊ | 12164/18018 [26:19:29<12:56:59, 7.96s/step, loss=1.315, acc=0.698, acc_len=1.68, thru=15.2, I=15.0, T=27.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:49:43,784] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:49:52,987] loop.py:380 INFO TIMING step=12165: step=3.782s data=1.559s compute=3.749s [fwd=2.711s bwd=1.024s opt=0.003s] dispatch=5.708s [2026-10-19 15:49:52,987] loop.py:45 INFO TRAIN_STEP step=12165 loss=1.372232 Training: 68%|██████▊ | 12165/18018 [26:19:39<13:41:43, 8.42s/step, loss=1.372, acc=0.694, acc_len=1.66, thru=12.2, I=11.2, T=33.8, wait=5.7s, pool=49, epoch=2/2][2026-10-19 15:50:01,204] loop.py:45 INFO TRAIN_STEP step=12166 loss=1.355161 Training: 68%|██████▊ | 12166/18018 [26:19:47<13:35:34, 8.36s/step, loss=1.355, acc=0.691, acc_len=1.65, thru=16.1, I=15.4, T=35.1, wait=4.6s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:50:01,200] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12165: forward=2587.0ms backward=1013.1ms [2026-10-19 15:50:06,863] loop.py:45 INFO TRAIN_STEP step=12167 loss=1.304882 Training: 68%|██████▊ | 12167/18018 [26:19:52<12:16:25, 7.55s/step, loss=1.305, acc=0.696, acc_len=1.67, thru=17.9, I=19.7, T=38.0, wait=2.3s, pool=41, epoch=2/2][2026-10-19 15:50:18,167] loop.py:45 INFO TRAIN_STEP step=12168 loss=1.314475 Training: 68%|██████▊ | 12168/18018 [26:20:04<14:06:05, 8.68s/step, loss=1.314, acc=0.693, acc_len=1.66, thru=11.3, I=10.2, T=33.4, wait=7.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:50:27,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:50:28,920] loop.py:45 INFO TRAIN_STEP step=12169 loss=1.323828 Training: 68%|██████▊ | 12169/18018 [26:20:15<15:06:31, 9.30s/step, loss=1.324, acc=0.697, acc_len=1.67, thru=14.5, I=17.0, T=30.2, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:50:29,268] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:50:35,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:50:35,846] loop.py:380 INFO TIMING step=12170: step=4.232s data=2.089s compute=4.190s [fwd=3.154s bwd=1.021s opt=0.003s] dispatch=2.686s [2026-10-19 15:50:35,846] loop.py:45 INFO TRAIN_STEP step=12170 loss=1.287665 Training: 68%|██████▊ | 12170/18018 [26:20:21<13:57:02, 8.59s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=16.8, I=21.5, T=30.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:50:36,226] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:50:43,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:50:44,934] loop.py:45 INFO TRAIN_STEP step=12171 loss=1.316465 Training: 68%|██████▊ | 12171/18018 [26:20:31<14:11:28, 8.74s/step, loss=1.316, acc=0.698, acc_len=1.67, thru=13.6, I=16.9, T=33.7, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:50:44,929] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12170: forward=2704.1ms backward=1053.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:50:45,155] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:50:51,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:50:52,467] loop.py:45 INFO TRAIN_STEP step=12172 loss=1.315176 Training: 68%|██████▊ | 12172/18018 [26:20:38<13:36:06, 8.38s/step, loss=1.315, acc=0.699, acc_len=1.67, thru=16.0, I=18.8, T=36.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:50:52,829] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:51:02,054] loop.py:45 INFO TRAIN_STEP step=12173 loss=1.212013 Training: 68%|██████▊ | 12173/18018 [26:20:48<14:11:22, 8.74s/step, loss=1.212, acc=0.718, acc_len=1.74, thru=10.8, I=10.9, T=35.9, wait=6.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:51:09,526] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:51:10,349] loop.py:45 INFO TRAIN_STEP step=12174 loss=1.364359 Training: 68%|██████▊ | 12174/18018 [26:20:56<13:58:16, 8.61s/step, loss=1.364, acc=0.689, acc_len=1.64, thru=18.1, I=19.4, T=34.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:51:10,529] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:51:16,967] loop.py:380 INFO TIMING step=12175: step=3.675s data=1.558s compute=3.638s [fwd=2.525s bwd=1.098s opt=0.003s] dispatch=2.933s [2026-10-19 15:51:16,967] loop.py:45 INFO TRAIN_STEP step=12175 loss=1.201237 Training: 68%|██████▊ | 12175/18018 [26:21:03<13:00:02, 8.01s/step, loss=1.201, acc=0.718, acc_len=1.75, thru=17.7, I=20.7, T=34.8, wait=2.9s, pool=57, epoch=2/2][2026-10-19 15:51:26,093] loop.py:45 INFO TRAIN_STEP step=12176 loss=1.279283 Training: 68%|██████▊ | 12176/18018 [26:21:12<13:32:30, 8.34s/step, loss=1.279, acc=0.703, acc_len=1.68, thru=13.7, I=14.6, T=35.6, wait=5.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:51:26,089] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12175: forward=2580.9ms backward=966.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:51:35,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:51:36,122] loop.py:45 INFO TRAIN_STEP step=12177 loss=1.271625 Training: 68%|██████▊ | 12177/18018 [26:21:22<14:21:32, 8.85s/step, loss=1.272, acc=0.706, acc_len=1.70, thru=13.7, I=13.6, T=32.2, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:51:36,445] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:51:41,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:51:41,957] loop.py:45 INFO TRAIN_STEP step=12178 loss=1.271330 Training: 68%|██████▊ | 12178/18018 [26:21:28<12:53:18, 7.94s/step, loss=1.271, acc=0.703, acc_len=1.69, thru=19.9, I=22.4, T=36.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:51:42,216] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:51:48,525] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:51:48,630] loop.py:45 INFO TRAIN_STEP step=12179 loss=1.388515 Training: 68%|██████▊ | 12179/18018 [26:21:34<12:16:04, 7.56s/step, loss=1.389, acc=0.684, acc_len=1.63, thru=18.5, I=19.9, T=32.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:51:49,027] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:51:55,287] loop.py:380 INFO TIMING step=12180: step=3.573s data=1.481s compute=3.545s [fwd=2.549s bwd=0.982s opt=0.003s] dispatch=3.074s [2026-10-19 15:51:55,288] loop.py:45 INFO TRAIN_STEP step=12180 loss=1.387930 Training: 68%|██████▊ | 12180/18018 [26:21:41<11:49:36, 7.29s/step, loss=1.388, acc=0.684, acc_len=1.63, thru=18.0, I=18.3, T=35.8, wait=3.1s, pool=61, epoch=2/2][2026-10-19 15:52:01,409] loop.py:45 INFO TRAIN_STEP step=12181 loss=1.228616 Training: 68%|██████▊ | 12181/18018 [26:21:47<11:15:09, 6.94s/step, loss=1.229, acc=0.713, acc_len=1.72, thru=19.3, I=18.8, T=38.7, wait=2.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:52:01,405] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12180: forward=2340.9ms backward=918.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:10,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:52:11,294] loop.py:45 INFO TRAIN_STEP step=12182 loss=1.392462 Training: 68%|██████▊ | 12182/18018 [26:21:57<12:41:02, 7.82s/step, loss=1.392, acc=0.685, acc_len=1.63, thru=14.7, I=16.9, T=28.6, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:11,615] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:15,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:52:16,501] loop.py:45 INFO TRAIN_STEP step=12183 loss=1.174380 Training: 68%|██████▊ | 12183/18018 [26:22:02<11:24:31, 7.04s/step, loss=1.174, acc=0.724, acc_len=1.77, thru=20.1, I=25.6, T=39.0, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:16,927] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:21,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:52:22,339] loop.py:45 INFO TRAIN_STEP step=12184 loss=1.152556 Training: 68%|██████▊ | 12184/18018 [26:22:08<10:49:19, 6.68s/step, loss=1.153, acc=0.725, acc_len=1.76, thru=23.1, I=26.7, T=34.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:22,800] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:27,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:52:28,470] loop.py:380 INFO TIMING step=12185: step=3.493s data=1.435s compute=3.452s [fwd=2.484s bwd=0.953s opt=0.003s] dispatch=2.630s [2026-10-19 15:52:28,470] loop.py:45 INFO TRAIN_STEP step=12185 loss=1.214644 Training: 68%|██████▊ | 12185/18018 [26:22:14<10:33:17, 6.51s/step, loss=1.215, acc=0.715, acc_len=1.73, thru=20.6, I=24.8, T=36.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:28,924] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:33,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:52:35,009] loop.py:45 INFO TRAIN_STEP step=12186 loss=1.249752 Training: 68%|██████▊ | 12186/18018 [26:22:21<10:33:55, 6.52s/step, loss=1.250, acc=0.706, acc_len=1.70, thru=19.3, I=26.2, T=36.5, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:52:35,004] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12185: forward=2513.9ms backward=953.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:35,262] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:40,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:52:41,330] loop.py:45 INFO TRAIN_STEP step=12187 loss=1.192036 Training: 68%|██████▊ | 12187/18018 [26:22:27<10:27:54, 6.46s/step, loss=1.192, acc=0.720, acc_len=1.75, thru=20.1, I=25.0, T=37.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:41,701] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:46,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:52:47,645] loop.py:45 INFO TRAIN_STEP step=12188 loss=1.316173 Training: 68%|██████▊ | 12188/18018 [26:22:33<10:23:33, 6.42s/step, loss=1.316, acc=0.700, acc_len=1.68, thru=20.0, I=23.3, T=35.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:52:48,049] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:52:54,586] loop.py:45 INFO TRAIN_STEP step=12189 loss=1.220661 Training: 68%|██████▊ | 12189/18018 [26:22:40<10:38:44, 6.57s/step, loss=1.221, acc=0.714, acc_len=1.73, thru=17.2, I=19.8, T=35.4, wait=3.3s, pool=57, epoch=2/2][2026-10-19 15:53:00,872] loop.py:380 INFO TIMING step=12190: step=3.707s data=1.633s compute=3.662s [fwd=2.587s bwd=1.061s opt=0.003s] dispatch=2.569s [2026-10-19 15:53:00,872] loop.py:45 INFO TRAIN_STEP step=12190 loss=1.447651 Training: 68%|██████▊ | 12190/18018 [26:22:46<10:30:10, 6.49s/step, loss=1.448, acc=0.676, acc_len=1.59, thru=19.3, I=18.9, T=34.5, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:08,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 15:53:08,584] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12190: forward=2694.0ms backward=1047.4ms [2026-10-19 15:53:08,588] loop.py:45 INFO TRAIN_STEP step=12191 loss=1.381133 Training: 68%|██████▊ | 12191/18018 [26:22:54<11:05:53, 6.86s/step, loss=1.381, acc=0.686, acc_len=1.63, thru=18.0, I=17.0, T=33.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:08,727] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:15,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:53:17,047] loop.py:45 INFO TRAIN_STEP step=12192 loss=1.322360 Training: 68%|██████▊ | 12192/18018 [26:23:03<11:52:24, 7.34s/step, loss=1.322, acc=0.694, acc_len=1.66, thru=14.7, I=17.5, T=30.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:17,411] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:23,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:53:24,608] loop.py:45 INFO TRAIN_STEP step=12193 loss=1.333505 Training: 68%|██████▊ | 12193/18018 [26:23:10<11:58:49, 7.40s/step, loss=1.334, acc=0.696, acc_len=1.66, thru=15.4, I=17.7, T=35.2, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:24,845] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:31,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 15:53:32,766] loop.py:45 INFO TRAIN_STEP step=12194 loss=1.224529 Training: 68%|██████▊ | 12194/18018 [26:23:18<12:20:40, 7.63s/step, loss=1.225, acc=0.711, acc_len=1.72, thru=15.7, I=17.5, T=31.0, wait=4.0s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:32,993] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=5/64 [2026-10-19 15:53:39,378] loop.py:380 INFO TIMING step=12195: step=3.450s data=1.443s compute=3.420s [fwd=2.453s bwd=0.951s opt=0.003s] dispatch=3.154s [2026-10-19 15:53:39,379] loop.py:45 INFO TRAIN_STEP step=12195 loss=1.341892 Training: 68%|██████▊ | 12195/18018 [26:23:25<11:50:50, 7.32s/step, loss=1.342, acc=0.695, acc_len=1.66, thru=17.6, I=19.0, T=37.1, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:45,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:53:46,667] loop.py:45 INFO TRAIN_STEP step=12196 loss=1.235569 Training: 68%|██████▊ | 12196/18018 [26:23:32<11:49:43, 7.31s/step, loss=1.236, acc=0.714, acc_len=1.73, thru=18.8, I=19.1, T=34.2, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:53:46,663] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12195: forward=2629.5ms backward=1063.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:46,967] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:53,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:53:53,951] loop.py:45 INFO TRAIN_STEP step=12197 loss=1.330894 Training: 68%|██████▊ | 12197/18018 [26:23:40<11:48:40, 7.30s/step, loss=1.331, acc=0.697, acc_len=1.68, thru=16.9, I=19.7, T=36.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:54,436] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:53:59,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:54:00,485] loop.py:45 INFO TRAIN_STEP step=12198 loss=1.362852 Training: 68%|██████▊ | 12198/18018 [26:23:46<11:26:12, 7.07s/step, loss=1.363, acc=0.688, acc_len=1.64, thru=19.4, I=22.3, T=39.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:00,765] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:05,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:54:06,702] loop.py:45 INFO TRAIN_STEP step=12199 loss=1.223469 Training: 68%|██████▊ | 12199/18018 [26:23:52<11:01:08, 6.82s/step, loss=1.223, acc=0.713, acc_len=1.72, thru=20.6, I=23.9, T=32.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:06,897] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:11,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:54:13,158] loop.py:380 INFO TIMING step=12200: step=3.734s data=1.670s compute=3.682s [fwd=2.529s bwd=1.138s opt=0.003s] dispatch=2.712s [2026-10-19 15:54:13,158] loop.py:45 INFO TRAIN_STEP step=12200 loss=1.346917 Training: 68%|██████▊ | 12200/18018 [26:23:59<10:50:27, 6.71s/step, loss=1.347, acc=0.691, acc_len=1.65, thru=18.4, I=24.1, T=34.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:13,610] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:20,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:54:20,671] loop.py:45 INFO TRAIN_STEP step=12201 loss=1.377873 Training: 68%|██████▊ | 12201/18018 [26:24:06<11:13:48, 6.95s/step, loss=1.378, acc=0.687, acc_len=1.64, thru=15.3, I=20.1, T=34.6, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:54:20,667] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12200: forward=2647.3ms backward=1008.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:21,073] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:29,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:54:30,098] loop.py:45 INFO TRAIN_STEP step=12202 loss=1.184600 Training: 68%|██████▊ | 12202/18018 [26:24:16<12:25:44, 7.69s/step, loss=1.185, acc=0.721, acc_len=1.75, thru=14.2, I=15.2, T=29.5, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:30,155] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:36,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:54:36,595] loop.py:45 INFO TRAIN_STEP step=12203 loss=1.386086 Training: 68%|██████▊ | 12203/18018 [26:24:22<11:50:49, 7.33s/step, loss=1.386, acc=0.685, acc_len=1.63, thru=19.0, I=19.1, T=35.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:37,087] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:42,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:54:43,457] loop.py:45 INFO TRAIN_STEP step=12204 loss=1.276260 Training: 68%|██████▊ | 12204/18018 [26:24:29<11:36:57, 7.19s/step, loss=1.276, acc=0.705, acc_len=1.69, thru=18.9, I=20.7, T=34.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:54:43,938] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 15:54:50,753] loop.py:380 INFO TIMING step=12205: step=3.403s data=1.409s compute=3.368s [fwd=2.415s bwd=0.938s opt=0.003s] dispatch=3.883s [2026-10-19 15:54:50,754] loop.py:45 INFO TRAIN_STEP step=12205 loss=1.338415 Training: 68%|██████▊ | 12205/18018 [26:24:36<11:39:50, 7.22s/step, loss=1.338, acc=0.693, acc_len=1.66, thru=16.0, I=19.6, T=37.6, wait=3.9s, pool=61, epoch=2/2][2026-10-19 15:54:58,646] loop.py:45 INFO TRAIN_STEP step=12206 loss=1.287195 Training: 68%|██████▊ | 12206/18018 [26:24:44<11:59:11, 7.42s/step, loss=1.287, acc=0.704, acc_len=1.69, thru=16.3, I=16.9, T=34.6, wait=4.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:54:58,641] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12205: forward=2614.4ms backward=1021.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:55:04,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:55:04,763] loop.py:45 INFO TRAIN_STEP step=12207 loss=1.314874 Training: 68%|██████▊ | 12207/18018 [26:24:50<11:21:01, 7.03s/step, loss=1.315, acc=0.698, acc_len=1.67, thru=19.9, I=21.1, T=36.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:55:05,266] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:55:10,953] loop.py:45 INFO TRAIN_STEP step=12208 loss=1.301793 Training: 68%|██████▊ | 12208/18018 [26:24:57<10:56:27, 6.78s/step, loss=1.302, acc=0.697, acc_len=1.67, thru=18.1, I=22.0, T=37.0, wait=2.7s, pool=45, epoch=2/2][2026-10-19 15:55:19,964] loop.py:45 INFO TRAIN_STEP step=12209 loss=1.345703 Training: 68%|██████▊ | 12209/18018 [26:25:06<12:01:14, 7.45s/step, loss=1.346, acc=0.695, acc_len=1.66, thru=14.6, I=13.0, T=34.6, wait=5.3s, pool=45, epoch=2/2][2026-10-19 15:55:30,375] loop.py:380 INFO TIMING step=12210: step=3.898s data=1.552s compute=3.855s [fwd=2.710s bwd=1.129s opt=0.003s] dispatch=6.503s [2026-10-19 15:55:30,375] loop.py:45 INFO TRAIN_STEP step=12210 loss=1.311531 Training: 68%|██████▊ | 12210/18018 [26:25:16<13:27:00, 8.34s/step, loss=1.312, acc=0.697, acc_len=1.67, thru=15.0, I=13.6, T=32.8, wait=6.5s, pool=61, epoch=2/2][2026-10-19 15:55:35,734] loop.py:45 INFO TRAIN_STEP step=12211 loss=1.245003 Training: 68%|██████▊ | 12211/18018 [26:25:21<12:00:30, 7.44s/step, loss=1.245, acc=0.706, acc_len=1.71, thru=19.9, I=18.9, T=38.2, wait=2.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:55:35,728] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12210: forward=2384.2ms backward=920.0ms [2026-10-19 15:55:43,577] loop.py:45 INFO TRAIN_STEP step=12212 loss=1.232273 Training: 68%|██████▊ | 12212/18018 [26:25:29<12:11:55, 7.56s/step, loss=1.232, acc=0.711, acc_len=1.72, thru=13.2, I=14.5, T=35.4, wait=4.2s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:55:50,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:55:51,473] loop.py:45 INFO TRAIN_STEP step=12213 loss=1.328076 Training: 68%|██████▊ | 12213/18018 [26:25:37<12:21:28, 7.66s/step, loss=1.328, acc=0.694, acc_len=1.66, thru=18.0, I=17.7, T=36.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:55:51,606] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:55:58,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:55:58,237] loop.py:45 INFO TRAIN_STEP step=12214 loss=1.359524 Training: 68%|██████▊ | 12214/18018 [26:25:44<11:55:10, 7.39s/step, loss=1.360, acc=0.690, acc_len=1.64, thru=18.8, I=21.3, T=37.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:55:58,635] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:56:02,772] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:56:04,031] loop.py:380 INFO TIMING step=12215: step=3.479s data=1.398s compute=3.447s [fwd=2.502s bwd=0.931s opt=0.003s] dispatch=2.307s [2026-10-19 15:56:04,031] loop.py:45 INFO TRAIN_STEP step=12215 loss=1.238451 Training: 68%|██████▊ | 12215/18018 [26:25:50<11:08:38, 6.91s/step, loss=1.238, acc=0.709, acc_len=1.71, thru=21.6, I=25.4, T=36.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:56:04,279] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:56:09,666] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:56:10,651] loop.py:45 INFO TRAIN_STEP step=12216 loss=1.285704 (TrainerActor pid=9310) [2026-10-19 15:56:10,647] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12215: forward=2456.9ms backward=968.9ms Training: 68%|██████▊ | 12216/18018 [26:25:56<11:00:04, 6.83s/step, loss=1.286, acc=0.706, acc_len=1.70, thru=19.1, I=23.9, T=36.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:56:10,685] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 15:56:16,496] loop.py:45 INFO TRAIN_STEP step=12217 loss=1.289383 Training: 68%|██████▊ | 12217/18018 [26:26:02<10:31:27, 6.53s/step, loss=1.289, acc=0.703, acc_len=1.69, thru=19.9, I=23.1, T=36.1, wait=2.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:56:29,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:56:30,235] loop.py:45 INFO TRAIN_STEP step=12218 loss=1.403137 Training: 68%|██████▊ | 12218/18018 [26:26:16<14:00:24, 8.69s/step, loss=1.403, acc=0.683, acc_len=1.62, thru=13.5, I=10.3, T=31.8, wait=9.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:56:30,386] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:56:36,524] loop.py:45 INFO TRAIN_STEP step=12219 loss=1.209343 Training: 68%|██████▊ | 12219/18018 [26:26:22<12:50:29, 7.97s/step, loss=1.209, acc=0.713, acc_len=1.72, thru=19.0, I=18.6, T=35.6, wait=2.7s, pool=61, epoch=2/2][2026-10-19 15:56:42,870] loop.py:380 INFO TIMING step=12220: step=3.456s data=1.378s compute=3.429s [fwd=2.465s bwd=0.949s opt=0.003s] dispatch=2.880s [2026-10-19 15:56:42,870] loop.py:45 INFO TRAIN_STEP step=12220 loss=1.297262 Training: 68%|██████▊ | 12220/18018 [26:26:28<12:03:17, 7.48s/step, loss=1.297, acc=0.703, acc_len=1.70, thru=19.2, I=20.4, T=37.0, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:56:49,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:56:51,623] loop.py:45 INFO TRAIN_STEP step=12221 loss=1.272110 Training: 68%|██████▊ | 12221/18018 [26:26:37<12:39:50, 7.86s/step, loss=1.272, acc=0.703, acc_len=1.69, thru=15.2, I=17.4, T=30.8, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:56:51,618] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12220: forward=2921.8ms backward=1183.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:56:51,887] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 15:56:57,437] loop.py:45 INFO TRAIN_STEP step=12222 loss=1.350310 Training: 68%|██████▊ | 12222/18018 [26:26:43<11:40:22, 7.25s/step, loss=1.350, acc=0.690, acc_len=1.64, thru=19.0, I=24.0, T=37.7, wait=2.4s, pool=61, epoch=2/2][2026-10-19 15:57:05,540] loop.py:45 INFO TRAIN_STEP step=12223 loss=1.263967 Training: 68%|██████▊ | 12223/18018 [26:26:51<12:04:56, 7.51s/step, loss=1.264, acc=0.706, acc_len=1.70, thru=14.7, I=14.8, T=35.8, wait=4.5s, pool=45, epoch=2/2][2026-10-19 15:57:13,864] loop.py:45 INFO TRAIN_STEP step=12224 loss=1.262653 Training: 68%|██████▊ | 12224/18018 [26:26:59<12:28:28, 7.75s/step, loss=1.263, acc=0.707, acc_len=1.71, thru=16.1, I=15.8, T=35.2, wait=4.7s, pool=45, epoch=2/2][2026-10-19 15:57:20,914] loop.py:380 INFO TIMING step=12225: step=3.834s data=1.616s compute=3.802s [fwd=2.723s bwd=1.065s opt=0.003s] dispatch=3.208s [2026-10-19 15:57:20,914] loop.py:45 INFO TRAIN_STEP step=12225 loss=1.303818 Training: 68%|██████▊ | 12225/18018 [26:27:06<12:07:59, 7.54s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=17.3, I=19.3, T=33.4, wait=3.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:57:27,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:57:28,712] loop.py:45 INFO TRAIN_STEP step=12226 loss=1.335168 (TrainerActor pid=9310) [2026-10-19 15:57:28,708] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12225: forward=2567.1ms backward=1013.3ms Training: 68%|██████▊ | 12226/18018 [26:27:14<12:15:26, 7.62s/step, loss=1.335, acc=0.695, acc_len=1.67, thru=16.9, I=18.7, T=35.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:57:28,861] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:57:34,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:57:35,176] loop.py:45 INFO TRAIN_STEP step=12227 loss=1.357557 Training: 68%|██████▊ | 12227/18018 [26:27:21<11:41:53, 7.27s/step, loss=1.358, acc=0.686, acc_len=1.63, thru=19.5, I=24.6, T=39.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:57:35,354] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:57:42,363] loop.py:45 INFO TRAIN_STEP step=12228 loss=1.333786 Training: 68%|██████▊ | 12228/18018 [26:27:28<11:39:18, 7.25s/step, loss=1.334, acc=0.694, acc_len=1.65, thru=16.2, I=16.4, T=33.3, wait=3.3s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:57:42,421] eagle_store.py:517 WARNING batch_get_buffer missing keys (k391303-8fa0fd34_hs, k391303-8fa0fd34_ids, k391303-8fa0fd34_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:57:49,756] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 15:57:42,416] eagle_store.py:517 WARNING batch_get_buffer missing keys (k391301-99c45e27_hs, k391301-99c45e27_ids, k391301-99c45e27_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 15:57:50,064] loop.py:45 INFO TRAIN_STEP step=12229 loss=1.232559 Training: 68%|██████▊ | 12229/18018 [26:27:36<11:52:14, 7.38s/step, loss=1.233, acc=0.710, acc_len=1.72, thru=19.3, I=18.1, T=34.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:57:50,259] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:57:56,347] loop.py:380 INFO TIMING step=12230: step=3.454s data=1.450s compute=3.429s [fwd=2.449s bwd=0.965s opt=0.003s] dispatch=2.822s [2026-10-19 15:57:56,347] loop.py:45 INFO TRAIN_STEP step=12230 loss=1.401418 Training: 68%|██████▊ | 12230/18018 [26:27:42<11:20:19, 7.05s/step, loss=1.401, acc=0.679, acc_len=1.60, thru=18.0, I=19.7, T=37.1, wait=2.8s, pool=49, epoch=2/2][2026-10-19 15:58:05,561] loop.py:45 INFO TRAIN_STEP step=12231 loss=1.295750 Training: 68%|██████▊ | 12231/18018 [26:27:51<12:22:47, 7.70s/step, loss=1.296, acc=0.699, acc_len=1.68, thru=14.8, I=13.4, T=31.2, wait=5.1s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:58:05,556] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12230: forward=2904.5ms backward=1153.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:11,535] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:58:13,156] loop.py:45 INFO TRAIN_STEP step=12232 loss=1.434379 Training: 68%|██████▊ | 12232/18018 [26:27:59<12:19:34, 7.67s/step, loss=1.434, acc=0.679, acc_len=1.61, thru=18.9, I=19.4, T=35.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:13,548] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:19,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:58:19,501] loop.py:45 INFO TRAIN_STEP step=12233 loss=1.234700 Training: 68%|██████▊ | 12233/18018 [26:28:05<11:41:04, 7.27s/step, loss=1.235, acc=0.712, acc_len=1.72, thru=19.5, I=24.7, T=37.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:19,701] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:58:27,719] loop.py:45 INFO TRAIN_STEP step=12234 loss=1.399894 Training: 68%|██████▊ | 12234/18018 [26:28:13<12:08:26, 7.56s/step, loss=1.400, acc=0.680, acc_len=1.62, thru=14.3, I=15.1, T=30.9, wait=4.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:33,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:58:34,638] loop.py:380 INFO TIMING step=12235: step=3.566s data=1.444s compute=3.536s [fwd=2.594s bwd=0.928s opt=0.003s] dispatch=3.342s [2026-10-19 15:58:34,638] loop.py:45 INFO TRAIN_STEP step=12235 loss=1.241361 Training: 68%|██████▊ | 12235/18018 [26:28:20<11:49:52, 7.37s/step, loss=1.241, acc=0.710, acc_len=1.72, thru=18.2, I=18.3, T=35.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:34,998] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 15:58:40,644] loop.py:45 INFO TRAIN_STEP step=12236 loss=1.227995 Training: 68%|██████▊ | 12236/18018 [26:28:26<11:10:25, 6.96s/step, loss=1.228, acc=0.715, acc_len=1.73, thru=19.3, I=24.5, T=38.6, wait=2.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:58:40,640] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12235: forward=2376.3ms backward=896.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:49,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:58:49,451] loop.py:45 INFO TRAIN_STEP step=12237 loss=1.292751 Training: 68%|██████▊ | 12237/18018 [26:28:35<12:03:47, 7.51s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=15.9, I=15.5, T=34.2, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:49,622] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:55,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:58:55,812] loop.py:45 INFO TRAIN_STEP step=12238 loss=1.212452 Training: 68%|██████▊ | 12238/18018 [26:28:41<11:30:25, 7.17s/step, loss=1.212, acc=0.711, acc_len=1.72, thru=19.6, I=19.7, T=39.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:58:56,157] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:59:03,001] loop.py:45 INFO TRAIN_STEP step=12239 loss=1.277920 Training: 68%|██████▊ | 12239/18018 [26:28:49<11:30:51, 7.17s/step, loss=1.278, acc=0.704, acc_len=1.69, thru=18.1, I=19.1, T=36.2, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:11,556] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:59:12,604] loop.py:380 INFO TIMING step=12240: step=3.777s data=1.703s compute=3.725s [fwd=2.561s bwd=1.150s opt=0.003s] dispatch=5.819s [2026-10-19 15:59:12,604] loop.py:45 INFO TRAIN_STEP step=12240 loss=1.253331 Training: 68%|██████▊ | 12240/18018 [26:28:58<12:40:58, 7.90s/step, loss=1.253, acc=0.709, acc_len=1.71, thru=14.5, I=15.3, T=33.9, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:13,066] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:18,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:59:18,630] loop.py:45 INFO TRAIN_STEP step=12241 loss=1.202771 Training: 68%|██████▊ | 12241/18018 [26:29:04<11:46:40, 7.34s/step, loss=1.203, acc=0.717, acc_len=1.74, thru=20.2, I=22.8, T=36.7, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:59:18,625] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12240: forward=2488.3ms backward=952.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:18,837] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:25,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:59:25,996] loop.py:45 INFO TRAIN_STEP step=12242 loss=1.357485 Training: 68%|██████▊ | 12242/18018 [26:29:12<11:47:17, 7.35s/step, loss=1.357, acc=0.692, acc_len=1.65, thru=17.7, I=18.2, T=33.2, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:26,219] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:31,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:59:32,326] loop.py:45 INFO TRAIN_STEP step=12243 loss=1.251741 Training: 68%|██████▊ | 12243/18018 [26:29:18<11:17:51, 7.04s/step, loss=1.252, acc=0.708, acc_len=1.71, thru=19.7, I=19.9, T=37.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:32,471] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:36,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:59:37,812] loop.py:45 INFO TRAIN_STEP step=12244 loss=1.234124 Training: 68%|██████▊ | 12244/18018 [26:29:23<10:32:44, 6.58s/step, loss=1.234, acc=0.714, acc_len=1.73, thru=22.5, I=25.3, T=35.5, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:37,880] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:43,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:59:44,117] loop.py:380 INFO TIMING step=12245: step=3.989s data=1.900s compute=3.965s [fwd=2.830s bwd=1.120s opt=0.003s] dispatch=2.307s [2026-10-19 15:59:44,117] loop.py:45 INFO TRAIN_STEP step=12245 loss=1.260935 Training: 68%|██████▊ | 12245/18018 [26:29:30<10:24:48, 6.49s/step, loss=1.261, acc=0.709, acc_len=1.71, thru=20.8, I=25.0, T=32.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:44,535] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:51,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 15:59:51,883] loop.py:45 INFO TRAIN_STEP step=12246 loss=1.175301 Training: 68%|██████▊ | 12246/18018 [26:29:37<11:01:25, 6.88s/step, loss=1.175, acc=0.721, acc_len=1.75, thru=15.5, I=18.0, T=31.9, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 15:59:51,878] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12245: forward=2806.0ms backward=1162.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 15:59:52,292] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 15:59:59,172] loop.py:45 INFO TRAIN_STEP step=12247 loss=1.414094 Training: 68%|██████▊ | 12247/18018 [26:29:45<11:13:14, 7.00s/step, loss=1.414, acc=0.681, acc_len=1.62, thru=15.7, I=15.4, T=36.6, wait=3.8s, pool=45, epoch=2/2][2026-10-19 16:00:08,274] loop.py:45 INFO TRAIN_STEP step=12248 loss=1.243971 Training: 68%|██████▊ | 12248/18018 [26:29:54<12:13:47, 7.63s/step, loss=1.244, acc=0.710, acc_len=1.71, thru=12.9, I=14.2, T=34.7, wait=5.4s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:17,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:00:19,149] loop.py:45 INFO TRAIN_STEP step=12249 loss=1.384519 Training: 68%|██████▊ | 12249/18018 [26:30:05<13:47:18, 8.60s/step, loss=1.385, acc=0.689, acc_len=1.64, thru=14.5, I=16.5, T=28.6, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:19,421] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:26,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:00:26,530] loop.py:380 INFO TIMING step=12250: step=4.018s data=1.813s compute=3.989s [fwd=2.825s bwd=1.150s opt=0.003s] dispatch=3.352s [2026-10-19 16:00:26,530] loop.py:45 INFO TRAIN_STEP step=12250 loss=1.269507 Training: 68%|██████▊ | 12250/18018 [26:30:12<13:11:49, 8.24s/step, loss=1.270, acc=0.704, acc_len=1.70, thru=14.9, I=17.7, T=31.9, wait=3.4s, pool=65, epoch=2/2][2026-10-19 16:00:26,533] loop.py:419 INFO Saving checkpoint at step 12250... (TrainerActor pid=9310) [2026-10-19 16:00:29,629] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0012251 (TrainerActor pid=9310) [2026-10-19 16:00:29,637] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0011501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:31,097] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:31,097] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=1/64 [2026-10-19 16:00:37,112] loop.py:45 INFO TRAIN_STEP step=12251 loss=1.248618 Training: 68%|██████▊ | 12251/18018 [26:30:23<14:19:23, 8.94s/step, loss=1.249, acc=0.711, acc_len=1.72, thru=16.6, I=14.3, T=35.7, wait=2.7s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:00:37,109] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12250: forward=2548.2ms backward=994.7ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 16:00:29,631] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0012251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 16:00:30,786] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0011501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:43,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:00:44,428] loop.py:45 INFO TRAIN_STEP step=12252 loss=1.376243 Training: 68%|██████▊ | 12252/18018 [26:30:30<13:32:22, 8.45s/step, loss=1.376, acc=0.686, acc_len=1.64, thru=19.2, I=19.7, T=34.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:44,464] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:49,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:00:49,981] loop.py:45 INFO TRAIN_STEP step=12253 loss=1.388400 Training: 68%|██████▊ | 12253/18018 [26:30:36<12:08:34, 7.58s/step, loss=1.388, acc=0.685, acc_len=1.63, thru=22.4, I=26.0, T=35.0, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:50,174] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:56,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:00:57,637] loop.py:45 INFO TRAIN_STEP step=12254 loss=1.234515 Training: 68%|██████▊ | 12254/18018 [26:30:43<12:10:37, 7.61s/step, loss=1.235, acc=0.712, acc_len=1.72, thru=16.1, I=18.6, T=36.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:00:57,735] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:01:02,457] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:01:03,515] loop.py:380 INFO TIMING step=12255: step=3.338s data=1.369s compute=3.309s [fwd=2.395s bwd=0.899s opt=0.003s] dispatch=2.531s [2026-10-19 16:01:03,516] loop.py:45 INFO TRAIN_STEP step=12255 loss=1.304857 Training: 68%|██████▊ | 12255/18018 [26:30:49<11:20:42, 7.09s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=20.9, I=25.9, T=38.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:01:03,965] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:01:08,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:01:10,130] loop.py:45 INFO TRAIN_STEP step=12256 loss=1.309795 Training: 68%|██████▊ | 12256/18018 [26:30:56<11:06:55, 6.94s/step, loss=1.310, acc=0.701, acc_len=1.69, thru=19.6, I=25.5, T=37.1, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:01:10,125] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12255: forward=2452.2ms backward=942.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:01:10,505] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:01:17,851] loop.py:45 INFO TRAIN_STEP step=12257 loss=1.313696 Training: 68%|██████▊ | 12257/18018 [26:31:03<11:29:11, 7.18s/step, loss=1.314, acc=0.695, acc_len=1.67, thru=12.2, I=18.6, T=36.5, wait=4.2s, pool=21, epoch=2/2][2026-10-19 16:01:29,588] loop.py:45 INFO TRAIN_STEP step=12258 loss=1.182785 Training: 68%|██████▊ | 12258/18018 [26:31:15<13:40:30, 8.55s/step, loss=1.183, acc=0.723, acc_len=1.76, thru=14.1, I=13.7, T=32.6, wait=7.8s, pool=45, epoch=2/2][2026-10-19 16:01:39,538] loop.py:45 INFO TRAIN_STEP step=12259 loss=1.212374 Training: 68%|██████▊ | 12259/18018 [26:31:25<14:20:37, 8.97s/step, loss=1.212, acc=0.712, acc_len=1.73, thru=10.1, I=11.0, T=35.0, wait=6.3s, pool=17, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:01:46,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:01:48,458] loop.py:380 INFO TIMING step=12260: step=3.671s data=1.521s compute=3.616s [fwd=2.526s bwd=1.075s opt=0.003s] dispatch=5.241s [2026-10-19 16:01:48,459] loop.py:45 INFO TRAIN_STEP step=12260 loss=1.284644 Training: 68%|██████▊ | 12260/18018 [26:31:34<14:19:09, 8.95s/step, loss=1.285, acc=0.707, acc_len=1.71, thru=19.3, I=17.6, T=34.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:01:48,697] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:01:56,036] loop.py:45 INFO TRAIN_STEP step=12261 loss=1.395481 Training: 68%|██████▊ | 12261/18018 [26:31:42<13:39:28, 8.54s/step, loss=1.395, acc=0.681, acc_len=1.61, thru=12.9, I=21.2, T=37.2, wait=4.1s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:01:56,032] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12260: forward=2426.7ms backward=977.1ms [2026-10-19 16:02:06,815] loop.py:45 INFO TRAIN_STEP step=12262 loss=1.260348 Training: 68%|██████▊ | 12262/18018 [26:31:52<14:43:42, 9.21s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=15.1, I=11.5, T=37.1, wait=7.3s, pool=53, epoch=2/2][2026-10-19 16:02:13,980] loop.py:45 INFO TRAIN_STEP step=12263 loss=1.396487 Training: 68%|██████▊ | 12263/18018 [26:32:00<13:44:38, 8.60s/step, loss=1.396, acc=0.678, acc_len=1.61, thru=16.9, I=17.7, T=36.6, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:20,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:02:21,167] loop.py:45 INFO TRAIN_STEP step=12264 loss=1.269783 Training: 68%|██████▊ | 12264/18018 [26:32:07<13:03:58, 8.18s/step, loss=1.270, acc=0.711, acc_len=1.71, thru=18.3, I=18.2, T=36.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:21,340] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:27,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:02:28,343] loop.py:380 INFO TIMING step=12265: step=3.524s data=1.421s compute=3.484s [fwd=2.469s bwd=1.000s opt=0.003s] dispatch=3.642s [2026-10-19 16:02:28,343] loop.py:45 INFO TRAIN_STEP step=12265 loss=1.301228 Training: 68%|██████▊ | 12265/18018 [26:32:14<12:35:03, 7.87s/step, loss=1.301, acc=0.699, acc_len=1.67, thru=18.9, I=19.3, T=36.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:28,805] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:35,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:02:36,364] loop.py:45 INFO TRAIN_STEP step=12266 loss=1.248174 Training: 68%|██████▊ | 12266/18018 [26:32:22<12:39:10, 7.92s/step, loss=1.248, acc=0.711, acc_len=1.72, thru=15.7, I=17.8, T=35.1, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:02:36,359] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12265: forward=2554.3ms backward=1038.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:36,458] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:43,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:02:44,091] loop.py:45 INFO TRAIN_STEP step=12267 loss=1.163227 Training: 68%|██████▊ | 12267/18018 [26:32:30<12:33:28, 7.86s/step, loss=1.163, acc=0.723, acc_len=1.76, thru=18.1, I=19.0, T=35.4, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:44,183] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:49,665] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:02:49,840] loop.py:45 INFO TRAIN_STEP step=12268 loss=1.337755 Training: 68%|██████▊ | 12268/18018 [26:32:35<11:32:36, 7.23s/step, loss=1.338, acc=0.691, acc_len=1.65, thru=20.6, I=23.4, T=38.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:50,167] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:54,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:02:55,271] loop.py:45 INFO TRAIN_STEP step=12269 loss=1.236219 Training: 68%|██████▊ | 12269/18018 [26:32:41<10:40:54, 6.69s/step, loss=1.236, acc=0.711, acc_len=1.71, thru=23.1, I=24.5, T=38.7, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:02:55,629] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:00,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:03:02,025] loop.py:380 INFO TIMING step=12270: step=3.477s data=1.459s compute=3.436s [fwd=2.484s bwd=0.938s opt=0.003s] dispatch=3.266s [2026-10-19 16:03:02,026] loop.py:45 INFO TRAIN_STEP step=12270 loss=1.266891 Training: 68%|██████▊ | 12270/18018 [26:32:48<10:42:43, 6.71s/step, loss=1.267, acc=0.704, acc_len=1.70, thru=19.3, I=23.2, T=36.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:02,067] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 16:03:07,862] loop.py:45 INFO TRAIN_STEP step=12271 loss=1.234300 Training: 68%|██████▊ | 12271/18018 [26:32:53<10:17:30, 6.45s/step, loss=1.234, acc=0.710, acc_len=1.72, thru=20.6, I=25.3, T=38.8, wait=2.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:03:07,858] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12270: forward=2347.0ms backward=907.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:16,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:03:17,582] loop.py:45 INFO TRAIN_STEP step=12272 loss=1.374741 Training: 68%|██████▊ | 12272/18018 [26:33:03<11:51:26, 7.43s/step, loss=1.375, acc=0.683, acc_len=1.62, thru=14.4, I=15.4, T=34.0, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:17,883] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:24,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:03:24,668] loop.py:45 INFO TRAIN_STEP step=12273 loss=1.122476 Training: 68%|██████▊ | 12273/18018 [26:33:10<11:41:27, 7.33s/step, loss=1.122, acc=0.734, acc_len=1.80, thru=17.4, I=21.0, T=35.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:24,756] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 16:03:30,283] loop.py:45 INFO TRAIN_STEP step=12274 loss=1.296376 Training: 68%|██████▊ | 12274/18018 [26:33:16<10:52:12, 6.81s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=21.0, I=20.1, T=33.7, wait=1.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:36,701] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:03:37,318] loop.py:380 INFO TIMING step=12275: step=3.685s data=1.576s compute=3.642s [fwd=2.626s bwd=1.002s opt=0.003s] dispatch=3.340s [2026-10-19 16:03:37,319] loop.py:45 INFO TRAIN_STEP step=12275 loss=1.295285 Training: 68%|██████▊ | 12275/18018 [26:33:23<10:58:29, 6.88s/step, loss=1.295, acc=0.703, acc_len=1.69, thru=18.1, I=18.7, T=34.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:37,706] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:43,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:03:44,842] loop.py:45 INFO TRAIN_STEP step=12276 loss=1.331131 Training: 68%|██████▊ | 12276/18018 [26:33:30<11:16:49, 7.07s/step, loss=1.331, acc=0.697, acc_len=1.68, thru=17.2, I=21.2, T=31.8, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:03:44,838] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12275: forward=2882.1ms backward=1086.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:45,148] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:51,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:03:51,492] loop.py:45 INFO TRAIN_STEP step=12277 loss=1.344717 Training: 68%|██████▊ | 12277/18018 [26:33:37<11:04:36, 6.95s/step, loss=1.345, acc=0.692, acc_len=1.65, thru=18.0, I=22.5, T=32.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:51,651] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:56,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:03:57,742] loop.py:45 INFO TRAIN_STEP step=12278 loss=1.243493 Training: 68%|██████▊ | 12278/18018 [26:33:43<10:44:31, 6.74s/step, loss=1.243, acc=0.709, acc_len=1.71, thru=18.8, I=22.3, T=33.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:03:57,819] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:02,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:04:03,556] loop.py:45 INFO TRAIN_STEP step=12279 loss=1.310028 Training: 68%|██████▊ | 12279/18018 [26:33:49<10:17:57, 6.46s/step, loss=1.310, acc=0.698, acc_len=1.67, thru=20.0, I=24.6, T=37.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:03,814] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:04:10,231] loop.py:380 INFO TIMING step=12280: step=3.558s data=1.579s compute=3.512s [fwd=2.488s bwd=1.009s opt=0.003s] dispatch=3.106s [2026-10-19 16:04:10,231] loop.py:45 INFO TRAIN_STEP step=12280 loss=1.327920 Training: 68%|██████▊ | 12280/18018 [26:33:56<10:23:56, 6.52s/step, loss=1.328, acc=0.693, acc_len=1.65, thru=18.8, I=20.0, T=36.0, wait=3.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:15,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:04:16,402] loop.py:45 INFO TRAIN_STEP step=12281 loss=1.281007 Training: 68%|██████▊ | 12281/18018 [26:34:02<10:13:41, 6.42s/step, loss=1.281, acc=0.707, acc_len=1.70, thru=20.1, I=21.7, T=32.5, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:04:16,397] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12280: forward=2827.1ms backward=1066.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:16,736] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:20,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:04:22,361] loop.py:45 INFO TRAIN_STEP step=12282 loss=1.282855 Training: 68%|██████▊ | 12282/18018 [26:34:08<10:00:25, 6.28s/step, loss=1.283, acc=0.708, acc_len=1.70, thru=20.0, I=25.2, T=35.5, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:22,731] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:27,652] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:04:28,164] loop.py:45 INFO TRAIN_STEP step=12283 loss=1.210453 Training: 68%|██████▊ | 12283/18018 [26:34:14<9:46:38, 6.14s/step, loss=1.210, acc=0.719, acc_len=1.75, thru=20.8, I=31.6, T=38.8, wait=2.5s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:28,657] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:04:35,295] loop.py:45 INFO TRAIN_STEP step=12284 loss=1.268806 Training: 68%|██████▊ | 12284/18018 [26:34:21<10:14:56, 6.43s/step, loss=1.269, acc=0.707, acc_len=1.70, thru=17.2, I=19.5, T=34.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:35,279] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:35,780] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:42,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:04:45,184] loop.py:380 INFO TIMING step=12285: step=4.953s data=2.269s compute=4.882s [fwd=3.728s bwd=1.138s opt=0.003s] dispatch=4.928s [2026-10-19 16:04:45,184] loop.py:45 INFO TRAIN_STEP step=12285 loss=1.382422 Training: 68%|██████▊ | 12285/18018 [26:34:31<11:53:56, 7.47s/step, loss=1.382, acc=0.686, acc_len=1.64, thru=13.3, I=16.0, T=25.8, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:45,457] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:51,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:04:51,494] loop.py:45 INFO TRAIN_STEP step=12286 loss=1.224296 Training: 68%|██████▊ | 12286/18018 [26:34:37<11:20:26, 7.12s/step, loss=1.224, acc=0.714, acc_len=1.73, thru=14.9, I=20.3, T=34.7, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:04:51,490] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12285: forward=2560.7ms backward=1082.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:51,898] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:56,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:04:57,389] loop.py:45 INFO TRAIN_STEP step=12287 loss=1.147371 Training: 68%|██████▊ | 12287/18018 [26:34:43<10:45:10, 6.75s/step, loss=1.147, acc=0.725, acc_len=1.78, thru=20.4, I=21.5, T=36.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:04:57,481] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:03,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:05:04,552] loop.py:45 INFO TRAIN_STEP step=12288 loss=1.331683 Training: 68%|██████▊ | 12288/18018 [26:34:50<10:56:48, 6.88s/step, loss=1.332, acc=0.694, acc_len=1.65, thru=18.4, I=20.7, T=37.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:04,904] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:09,326] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:05:10,397] loop.py:45 INFO TRAIN_STEP step=12289 loss=1.250055 Training: 68%|██████▊ | 12289/18018 [26:34:56<10:27:06, 6.57s/step, loss=1.250, acc=0.712, acc_len=1.71, thru=22.0, I=28.9, T=38.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:10,837] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:15,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:05:16,354] loop.py:380 INFO TIMING step=12290: step=3.598s data=1.680s compute=3.568s [fwd=2.647s bwd=0.907s opt=0.003s] dispatch=2.349s [2026-10-19 16:05:16,354] loop.py:45 INFO TRAIN_STEP step=12290 loss=1.365322 Training: 68%|██████▊ | 12290/18018 [26:35:02<10:09:31, 6.38s/step, loss=1.365, acc=0.691, acc_len=1.65, thru=22.1, I=27.5, T=35.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:16,519] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 16:05:22,386] loop.py:45 INFO TRAIN_STEP step=12291 loss=1.207162 Training: 68%|██████▊ | 12291/18018 [26:35:08<9:59:14, 6.28s/step, loss=1.207, acc=0.719, acc_len=1.74, thru=18.1, I=23.3, T=39.3, wait=2.8s, pool=41, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 16:05:22,381] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12290: forward=2295.6ms backward=908.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:30,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:05:31,113] loop.py:45 INFO TRAIN_STEP step=12292 loss=1.468764 Training: 68%|██████▊ | 12292/18018 [26:35:17<11:09:16, 7.01s/step, loss=1.469, acc=0.670, acc_len=1.58, thru=16.8, I=17.4, T=34.3, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:31,240] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 16:05:37,638] loop.py:45 INFO TRAIN_STEP step=12293 loss=1.289044 Training: 68%|██████▊ | 12293/18018 [26:35:23<10:55:11, 6.87s/step, loss=1.289, acc=0.701, acc_len=1.68, thru=16.6, I=20.5, T=34.2, wait=2.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:43,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:05:44,121] loop.py:45 INFO TRAIN_STEP step=12294 loss=1.249342 Training: 68%|██████▊ | 12294/18018 [26:35:30<10:44:05, 6.75s/step, loss=1.249, acc=0.707, acc_len=1.70, thru=18.9, I=21.1, T=35.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:44,184] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 16:05:51,204] loop.py:380 INFO TIMING step=12295: step=3.510s data=1.511s compute=3.486s [fwd=2.513s bwd=0.958s opt=0.003s] dispatch=3.565s [2026-10-19 16:05:51,205] loop.py:45 INFO TRAIN_STEP step=12295 loss=1.308288 Training: 68%|██████▊ | 12295/18018 [26:35:37<10:53:26, 6.85s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=17.2, I=19.2, T=36.5, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:55,600] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:05:57,050] loop.py:45 INFO TRAIN_STEP step=12296 loss=1.266595 Training: 68%|██████▊ | 12296/18018 [26:35:43<10:24:38, 6.55s/step, loss=1.267, acc=0.704, acc_len=1.70, thru=20.1, I=21.7, T=34.7, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:05:57,046] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12295: forward=2634.0ms backward=1007.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:05:57,111] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:06:02,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:06:04,193] loop.py:45 INFO TRAIN_STEP step=12297 loss=1.186507 Training: 68%|██████▊ | 12297/18018 [26:35:50<10:41:30, 6.73s/step, loss=1.187, acc=0.722, acc_len=1.77, thru=17.3, I=22.9, T=32.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:06:04,561] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:06:10,626] loop.py:45 INFO TRAIN_STEP step=12298 loss=1.282244 Training: 68%|██████▊ | 12298/18018 [26:35:56<10:32:54, 6.64s/step, loss=1.282, acc=0.701, acc_len=1.68, thru=18.4, I=23.4, T=37.6, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:06:17,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:06:18,603] loop.py:45 INFO TRAIN_STEP step=12299 loss=1.386968 Training: 68%|██████▊ | 12299/18018 [26:36:04<11:11:03, 7.04s/step, loss=1.387, acc=0.687, acc_len=1.64, thru=16.7, I=18.7, T=33.4, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:06:18,737] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:06:25,486] loop.py:380 INFO TIMING step=12300: step=3.566s data=1.517s compute=3.527s [fwd=2.524s bwd=0.988s opt=0.003s] dispatch=3.309s [2026-10-19 16:06:25,487] loop.py:45 INFO TRAIN_STEP step=12300 loss=1.324904 Training: 68%|██████▊ | 12300/18018 [26:36:11<11:06:31, 6.99s/step, loss=1.325, acc=0.701, acc_len=1.69, thru=14.2, I=21.2, T=35.9, wait=3.3s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:06:32,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:06:33,911] loop.py:45 INFO TRAIN_STEP step=12301 loss=1.324322 Training: 68%|██████▊ | 12301/18018 [26:36:19<11:47:17, 7.42s/step, loss=1.324, acc=0.700, acc_len=1.68, thru=19.1, I=17.9, T=28.1, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:06:33,906] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12300: forward=3355.6ms backward=1157.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:06:34,095] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:06:41,500] loop.py:45 INFO TRAIN_STEP step=12302 loss=1.369063 Training: 68%|██████▊ | 12302/18018 [26:36:27<11:51:52, 7.47s/step, loss=1.369, acc=0.687, acc_len=1.64, thru=12.1, I=15.1, T=34.0, wait=3.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:06:52,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:06:52,813] loop.py:45 INFO TRAIN_STEP step=12303 loss=1.294853 Training: 68%|██████▊ | 12303/18018 [26:36:38<13:41:29, 8.62s/step, loss=1.295, acc=0.704, acc_len=1.69, thru=15.8, I=12.6, T=32.5, wait=7.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:06:53,157] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:06:59,362] loop.py:45 INFO TRAIN_STEP step=12304 loss=1.234004 Training: 68%|██████▊ | 12304/18018 [26:36:45<12:42:01, 8.00s/step, loss=1.234, acc=0.711, acc_len=1.72, thru=16.3, I=16.0, T=34.8, wait=2.9s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:07:06,826] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:07:08,585] loop.py:380 INFO TIMING step=12305: step=3.770s data=1.474s compute=3.736s [fwd=2.659s bwd=1.064s opt=0.003s] dispatch=5.445s [2026-10-19 16:07:08,586] loop.py:45 INFO TRAIN_STEP step=12305 loss=1.459531 Training: 68%|██████▊ | 12305/18018 [26:36:54<13:16:48, 8.37s/step, loss=1.460, acc=0.671, acc_len=1.58, thru=17.3, I=18.2, T=34.0, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:07:08,837] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:07:14,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:07:14,723] loop.py:45 INFO TRAIN_STEP step=12306 loss=1.293478 Training: 68%|██████▊ | 12306/18018 [26:37:00<12:12:56, 7.70s/step, loss=1.293, acc=0.698, acc_len=1.67, thru=19.7, I=25.5, T=36.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:07:14,719] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12305: forward=2424.8ms backward=989.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:07:15,121] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:07:23,092] loop.py:45 INFO TRAIN_STEP step=12307 loss=1.419478 Training: 68%|██████▊ | 12307/18018 [26:37:09<12:32:00, 7.90s/step, loss=1.419, acc=0.679, acc_len=1.61, thru=10.6, I=14.2, T=34.3, wait=4.6s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:07:33,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:07:34,575] loop.py:45 INFO TRAIN_STEP step=12308 loss=1.262158 Training: 68%|██████▊ | 12308/18018 [26:37:20<14:14:05, 8.97s/step, loss=1.262, acc=0.709, acc_len=1.71, thru=16.5, I=13.5, T=28.4, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:07:34,774] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:07:43,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:07:43,583] loop.py:45 INFO TRAIN_STEP step=12309 loss=1.378528 Training: 68%|██████▊ | 12309/18018 [26:37:29<14:14:55, 8.99s/step, loss=1.379, acc=0.683, acc_len=1.62, thru=13.4, I=14.1, T=28.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:07:43,789] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:07:51,069] loop.py:380 INFO TIMING step=12310: step=4.100s data=1.882s compute=4.072s [fwd=2.906s bwd=1.152s opt=0.003s] dispatch=3.376s [2026-10-19 16:07:51,070] loop.py:45 INFO TRAIN_STEP step=12310 loss=1.240119 Training: 68%|██████▊ | 12310/18018 [26:37:37<13:32:00, 8.54s/step, loss=1.240, acc=0.710, acc_len=1.72, thru=16.0, I=16.5, T=31.2, wait=3.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:07:58,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:08:00,039] loop.py:45 INFO TRAIN_STEP step=12311 loss=1.368589 Training: 68%|██████▊ | 12311/18018 [26:37:46<13:44:13, 8.67s/step, loss=1.369, acc=0.689, acc_len=1.64, thru=16.2, I=17.7, T=33.4, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:08:00,035] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12310: forward=2681.8ms backward=1096.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:08:00,273] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:08:04,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:08:05,411] loop.py:45 INFO TRAIN_STEP step=12312 loss=1.260090 Training: 68%|██████▊ | 12312/18018 [26:37:51<12:10:05, 7.68s/step, loss=1.260, acc=0.704, acc_len=1.69, thru=21.4, I=31.2, T=38.2, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:08:05,807] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:08:11,416] loop.py:45 INFO TRAIN_STEP step=12313 loss=1.251380 Training: 68%|██████▊ | 12313/18018 [26:37:57<11:22:14, 7.18s/step, loss=1.251, acc=0.706, acc_len=1.70, thru=19.8, I=24.9, T=36.5, wait=2.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:08:19,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:08:21,165] loop.py:45 INFO TRAIN_STEP step=12314 loss=1.232426 Training: 68%|██████▊ | 12314/18018 [26:38:07<12:35:34, 7.95s/step, loss=1.232, acc=0.712, acc_len=1.72, thru=15.5, I=15.9, T=27.1, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:08:21,295] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:08:26,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:08:27,475] loop.py:380 INFO TIMING step=12315: step=3.816s data=1.696s compute=3.774s [fwd=2.683s bwd=1.077s opt=0.003s] dispatch=2.486s [2026-10-19 16:08:27,476] loop.py:45 INFO TRAIN_STEP step=12315 loss=1.498950 Training: 68%|██████▊ | 12315/18018 [26:38:13<11:48:48, 7.46s/step, loss=1.499, acc=0.663, acc_len=1.56, thru=16.6, I=22.8, T=33.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:08:27,730] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:08:34,005] loop.py:45 INFO TRAIN_STEP step=12316 loss=1.271142 Training: 68%|██████▊ | 12316/18018 [26:38:20<11:22:10, 7.18s/step, loss=1.271, acc=0.708, acc_len=1.70, thru=17.0, I=22.7, T=34.7, wait=2.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:08:34,000] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12315: forward=2657.7ms backward=988.8ms [2026-10-19 16:08:42,391] loop.py:45 INFO TRAIN_STEP step=12317 loss=1.314073 Training: 68%|██████▊ | 12317/18018 [26:38:28<11:56:32, 7.54s/step, loss=1.314, acc=0.700, acc_len=1.68, thru=13.1, I=13.3, T=31.6, wait=4.3s, pool=37, epoch=2/2][2026-10-19 16:08:49,407] loop.py:45 INFO TRAIN_STEP step=12318 loss=1.257608 Training: 68%|██████▊ | 12318/18018 [26:38:35<11:41:23, 7.38s/step, loss=1.258, acc=0.707, acc_len=1.71, thru=19.6, I=17.0, T=32.6, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:08:58,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:09:00,032] loop.py:45 INFO TRAIN_STEP step=12319 loss=1.371872 Training: 68%|██████▊ | 12319/18018 [26:38:46<13:13:39, 8.36s/step, loss=1.372, acc=0.689, acc_len=1.65, thru=13.9, I=15.7, T=30.4, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:09:00,300] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:09:06,608] loop.py:380 INFO TIMING step=12320: step=3.846s data=1.765s compute=3.816s [fwd=2.689s bwd=1.111s opt=0.003s] dispatch=2.722s [2026-10-19 16:09:06,609] loop.py:45 INFO TRAIN_STEP step=12320 loss=1.234831 Training: 68%|██████▊ | 12320/18018 [26:38:52<12:22:52, 7.82s/step, loss=1.235, acc=0.711, acc_len=1.71, thru=16.9, I=21.9, T=33.3, wait=2.7s, pool=57, epoch=2/2][2026-10-19 16:09:16,505] loop.py:45 INFO TRAIN_STEP step=12321 loss=1.380333 Training: 68%|██████▊ | 12321/18018 [26:39:02<13:21:45, 8.44s/step, loss=1.380, acc=0.687, acc_len=1.64, thru=10.2, I=9.8, T=32.3, wait=5.9s, pool=25, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 16:09:16,500] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12320: forward=2820.0ms backward=1100.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:09:23,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:09:24,771] loop.py:45 INFO TRAIN_STEP step=12322 loss=1.150097 Training: 68%|██████▊ | 12322/18018 [26:39:10<13:16:34, 8.39s/step, loss=1.150, acc=0.729, acc_len=1.78, thru=19.8, I=17.1, T=34.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:09:25,087] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:09:33,197] loop.py:45 INFO TRAIN_STEP step=12323 loss=1.446520 Training: 68%|██████▊ | 12323/18018 [26:39:19<13:17:27, 8.40s/step, loss=1.447, acc=0.678, acc_len=1.60, thru=11.7, I=15.2, T=32.8, wait=4.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:09:42,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:09:42,965] loop.py:45 INFO TRAIN_STEP step=12324 loss=1.336358 Training: 68%|██████▊ | 12324/18018 [26:39:29<13:56:12, 8.81s/step, loss=1.336, acc=0.694, acc_len=1.67, thru=15.7, I=15.2, T=34.5, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:09:43,117] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:09:49,453] loop.py:380 INFO TIMING step=12325: step=3.481s data=1.389s compute=3.446s [fwd=2.442s bwd=0.990s opt=0.003s] dispatch=2.997s [2026-10-19 16:09:49,453] loop.py:45 INFO TRAIN_STEP step=12325 loss=1.286920 Training: 68%|██████▊ | 12325/18018 [26:39:35<12:49:55, 8.11s/step, loss=1.287, acc=0.704, acc_len=1.69, thru=17.8, I=19.1, T=36.8, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:09:55,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:09:56,490] loop.py:45 INFO TRAIN_STEP step=12326 loss=1.254202 Training: 68%|██████▊ | 12326/18018 [26:39:42<12:19:09, 7.79s/step, loss=1.254, acc=0.708, acc_len=1.70, thru=18.5, I=20.2, T=36.8, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:09:56,486] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12325: forward=2437.7ms backward=991.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:09:56,594] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 16:10:02,270] loop.py:45 INFO TRAIN_STEP step=12327 loss=1.238133 Training: 68%|██████▊ | 12327/18018 [26:39:48<11:21:45, 7.19s/step, loss=1.238, acc=0.715, acc_len=1.73, thru=20.8, I=24.4, T=37.6, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:10:09,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:10:10,129] loop.py:45 INFO TRAIN_STEP step=12328 loss=1.353777 Training: 68%|██████▊ | 12328/18018 [26:39:56<11:40:42, 7.39s/step, loss=1.354, acc=0.689, acc_len=1.64, thru=16.2, I=17.6, T=33.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:10:10,176] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:10:16,854] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:10:17,028] loop.py:45 INFO TRAIN_STEP step=12329 loss=1.359073 Training: 68%|██████▊ | 12329/18018 [26:40:03<11:26:37, 7.24s/step, loss=1.359, acc=0.689, acc_len=1.64, thru=16.9, I=18.9, T=37.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:10:17,356] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:10:22,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:10:23,358] loop.py:380 INFO TIMING step=12330: step=3.525s data=1.543s compute=3.494s [fwd=2.500s bwd=0.980s opt=0.003s] dispatch=2.798s [2026-10-19 16:10:23,358] loop.py:45 INFO TRAIN_STEP step=12330 loss=1.332741 Training: 68%|██████▊ | 12330/18018 [26:40:09<11:00:33, 6.97s/step, loss=1.333, acc=0.697, acc_len=1.67, thru=20.1, I=19.7, T=36.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:10:23,510] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 16:10:30,433] loop.py:45 INFO TRAIN_STEP step=12331 loss=1.322877 Training: 68%|██████▊ | 12331/18018 [26:40:16<11:03:32, 7.00s/step, loss=1.323, acc=0.696, acc_len=1.66, thru=13.7, I=16.2, T=33.7, wait=3.3s, pool=25, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:10:30,429] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12330: forward=2644.9ms backward=1109.9ms [2026-10-19 16:10:37,721] loop.py:45 INFO TRAIN_STEP step=12332 loss=1.331690 Training: 68%|██████▊ | 12332/18018 [26:40:23<11:11:33, 7.09s/step, loss=1.332, acc=0.696, acc_len=1.67, thru=18.4, I=15.0, T=35.9, wait=3.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:10:46,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:10:47,687] loop.py:45 INFO TRAIN_STEP step=12333 loss=1.377408 Training: 68%|██████▊ | 12333/18018 [26:40:33<12:33:21, 7.95s/step, loss=1.377, acc=0.687, acc_len=1.64, thru=15.8, I=15.3, T=31.8, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:10:47,929] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:10:53,724] loop.py:45 INFO TRAIN_STEP step=12334 loss=1.214777 Training: 68%|██████▊ | 12334/18018 [26:40:39<11:38:47, 7.38s/step, loss=1.215, acc=0.713, acc_len=1.73, thru=16.0, I=19.9, T=37.8, wait=2.6s, pool=33, epoch=2/2][2026-10-19 16:11:05,128] loop.py:380 INFO TIMING step=12335: step=4.052s data=1.730s compute=4.016s [fwd=2.845s bwd=1.157s opt=0.003s] dispatch=7.343s [2026-10-19 16:11:05,128] loop.py:45 INFO TRAIN_STEP step=12335 loss=1.252629 Training: 68%|██████▊ | 12335/18018 [26:40:51<13:33:07, 8.58s/step, loss=1.253, acc=0.710, acc_len=1.72, thru=12.8, I=14.1, T=31.6, wait=7.3s, pool=61, epoch=2/2][2026-10-19 16:11:11,961] loop.py:45 INFO TRAIN_STEP step=12336 loss=1.272841 Training: 68%|██████▊ | 12336/18018 [26:40:58<12:43:18, 8.06s/step, loss=1.273, acc=0.702, acc_len=1.68, thru=14.3, I=15.6, T=36.3, wait=3.3s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:11:11,957] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12335: forward=2504.4ms backward=985.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:22,009] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:11:22,638] loop.py:45 INFO TRAIN_STEP step=12337 loss=1.323738 Training: 68%|██████▊ | 12337/18018 [26:41:08<13:57:26, 8.84s/step, loss=1.324, acc=0.695, acc_len=1.66, thru=16.8, I=16.1, T=33.8, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:23,012] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:28,011] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:11:28,082] loop.py:45 INFO TRAIN_STEP step=12338 loss=1.225964 Training: 68%|██████▊ | 12338/18018 [26:41:14<12:20:38, 7.82s/step, loss=1.226, acc=0.710, acc_len=1.71, thru=21.7, I=25.5, T=36.4, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:28,518] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:32,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:11:34,345] loop.py:45 INFO TRAIN_STEP step=12339 loss=1.233587 Training: 68%|██████▊ | 12339/18018 [26:41:20<11:36:16, 7.36s/step, loss=1.234, acc=0.711, acc_len=1.72, thru=20.3, I=23.7, T=33.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:34,488] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:40,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:11:41,379] loop.py:380 INFO TIMING step=12340: step=3.743s data=1.683s compute=3.706s [fwd=2.624s bwd=1.067s opt=0.003s] dispatch=3.280s [2026-10-19 16:11:41,379] loop.py:45 INFO TRAIN_STEP step=12340 loss=1.285160 Training: 68%|██████▊ | 12340/18018 [26:41:27<11:27:00, 7.26s/step, loss=1.285, acc=0.705, acc_len=1.70, thru=17.6, I=22.0, T=34.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:41,448] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 16:11:48,707] loop.py:45 INFO TRAIN_STEP step=12341 loss=1.303527 Training: 68%|██████▊ | 12341/18018 [26:41:34<11:28:49, 7.28s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=15.6, I=16.4, T=36.0, wait=3.8s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:11:48,703] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12340: forward=2552.3ms backward=958.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:55,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:11:56,374] loop.py:45 INFO TRAIN_STEP step=12342 loss=1.278676 Training: 68%|██████▊ | 12342/18018 [26:41:42<11:39:39, 7.40s/step, loss=1.279, acc=0.705, acc_len=1.69, thru=18.6, I=19.4, T=34.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:11:56,584] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:12:02,191] loop.py:45 INFO TRAIN_STEP step=12343 loss=1.233034 Training: 69%|██████▊ | 12343/18018 [26:41:48<10:54:40, 6.92s/step, loss=1.233, acc=0.711, acc_len=1.72, thru=20.8, I=22.2, T=36.0, wait=2.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:09,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:12:09,952] loop.py:45 INFO TRAIN_STEP step=12344 loss=1.344223 Training: 69%|██████▊ | 12344/18018 [26:41:56<11:18:23, 7.17s/step, loss=1.344, acc=0.696, acc_len=1.67, thru=17.3, I=18.4, T=36.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:10,336] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:15,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:12:17,775] loop.py:380 INFO TIMING step=12345: step=3.896s data=1.705s compute=3.865s [fwd=2.763s bwd=1.087s opt=0.003s] dispatch=3.919s [2026-10-19 16:12:17,775] loop.py:45 INFO TRAIN_STEP step=12345 loss=1.275990 Training: 69%|██████▊ | 12345/18018 [26:42:03<11:36:43, 7.37s/step, loss=1.276, acc=0.700, acc_len=1.68, thru=16.2, I=20.7, T=32.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:18,008] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:12:25,601] loop.py:45 INFO TRAIN_STEP step=12346 loss=1.220891 Training: 69%|██████▊ | 12346/18018 [26:42:11<11:49:35, 7.51s/step, loss=1.221, acc=0.712, acc_len=1.72, thru=13.2, I=16.3, T=35.6, wait=4.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:12:25,597] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12345: forward=2551.1ms backward=1001.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:32,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:12:32,890] loop.py:45 INFO TRAIN_STEP step=12347 loss=1.273758 Training: 69%|██████▊ | 12347/18018 [26:42:18<11:43:16, 7.44s/step, loss=1.274, acc=0.706, acc_len=1.70, thru=18.5, I=18.3, T=37.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:33,016] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:37,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:12:38,536] loop.py:45 INFO TRAIN_STEP step=12348 loss=1.363769 Training: 69%|██████▊ | 12348/18018 [26:42:24<10:52:13, 6.90s/step, loss=1.364, acc=0.689, acc_len=1.64, thru=21.3, I=23.6, T=35.5, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:38,896] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:12:43,944] loop.py:45 INFO TRAIN_STEP step=12349 loss=1.172583 Training: 69%|██████▊ | 12349/18018 [26:42:30<10:09:47, 6.45s/step, loss=1.173, acc=0.726, acc_len=1.77, thru=19.3, I=26.0, T=37.5, wait=2.0s, pool=37, epoch=2/2][2026-10-19 16:12:53,136] loop.py:380 INFO TIMING step=12350: step=3.578s data=1.425s compute=3.548s [fwd=2.520s bwd=1.013s opt=0.003s] dispatch=5.605s [2026-10-19 16:12:53,137] loop.py:45 INFO TRAIN_STEP step=12350 loss=1.393846 Training: 69%|██████▊ | 12350/18018 [26:42:39<11:27:16, 7.28s/step, loss=1.394, acc=0.683, acc_len=1.62, thru=15.7, I=13.8, T=35.8, wait=5.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:58,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:12:59,016] loop.py:45 INFO TRAIN_STEP step=12351 loss=1.305123 (TrainerActor pid=9310) [2026-10-19 16:12:59,012] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12350: forward=2444.1ms backward=950.0ms Training: 69%|██████▊ | 12351/18018 [26:42:45<10:47:35, 6.86s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=20.6, I=21.2, T=37.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:12:59,486] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:07,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:13:07,400] loop.py:45 INFO TRAIN_STEP step=12352 loss=1.461782 Training: 69%|██████▊ | 12352/18018 [26:42:53<11:30:49, 7.32s/step, loss=1.462, acc=0.667, acc_len=1.56, thru=15.6, I=17.7, T=36.3, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:07,515] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 16:13:13,963] loop.py:45 INFO TRAIN_STEP step=12353 loss=1.349600 Training: 69%|██████▊ | 12353/18018 [26:43:00<11:09:23, 7.09s/step, loss=1.350, acc=0.690, acc_len=1.64, thru=15.7, I=18.9, T=37.2, wait=3.1s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:24,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:13:25,472] loop.py:45 INFO TRAIN_STEP step=12354 loss=1.317193 Training: 69%|██████▊ | 12354/18018 [26:43:11<13:14:22, 8.42s/step, loss=1.317, acc=0.701, acc_len=1.68, thru=13.8, I=14.0, T=32.6, wait=7.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:25,878] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:30,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:13:32,257] loop.py:380 INFO TIMING step=12355: step=4.235s data=2.027s compute=4.198s [fwd=2.982s bwd=1.201s opt=0.003s] dispatch=2.540s [2026-10-19 16:13:32,257] loop.py:45 INFO TRAIN_STEP step=12355 loss=1.324311 Training: 69%|██████▊ | 12355/18018 [26:43:18<12:28:04, 7.93s/step, loss=1.324, acc=0.691, acc_len=1.65, thru=18.5, I=20.9, T=30.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:32,519] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:38,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:13:39,946] loop.py:45 INFO TRAIN_STEP step=12356 loss=1.238658 Training: 69%|██████▊ | 12356/18018 [26:43:26<12:21:11, 7.85s/step, loss=1.239, acc=0.710, acc_len=1.71, thru=14.2, I=19.4, T=30.2, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:13:39,941] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12355: forward=3227.9ms backward=972.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:40,446] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:46,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:13:46,708] loop.py:45 INFO TRAIN_STEP step=12357 loss=1.271113 Training: 69%|██████▊ | 12357/18018 [26:43:32<11:50:13, 7.53s/step, loss=1.271, acc=0.703, acc_len=1.69, thru=17.7, I=20.0, T=36.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:46,938] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:51,999] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:13:52,774] loop.py:45 INFO TRAIN_STEP step=12358 loss=1.387538 Training: 69%|██████▊ | 12358/18018 [26:43:38<11:08:43, 7.09s/step, loss=1.388, acc=0.679, acc_len=1.61, thru=20.7, I=21.3, T=37.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:53,005] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:58,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:13:58,853] loop.py:45 INFO TRAIN_STEP step=12359 loss=1.262996 Training: 69%|██████▊ | 12359/18018 [26:43:44<10:39:59, 6.79s/step, loss=1.263, acc=0.707, acc_len=1.71, thru=21.0, I=22.9, T=38.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:13:59,299] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:14:06,681] loop.py:380 INFO TIMING step=12360: step=3.564s data=1.493s compute=3.537s [fwd=2.528s bwd=0.996s opt=0.003s] dispatch=4.256s [2026-10-19 16:14:06,681] loop.py:45 INFO TRAIN_STEP step=12360 loss=1.289050 Training: 69%|██████▊ | 12360/18018 [26:43:52<11:09:21, 7.10s/step, loss=1.289, acc=0.701, acc_len=1.68, thru=15.6, I=17.5, T=35.9, wait=4.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:15,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:14:16,429] loop.py:45 INFO TRAIN_STEP step=12361 loss=1.348936 Training: 69%|██████▊ | 12361/18018 [26:44:02<12:24:15, 7.89s/step, loss=1.349, acc=0.694, acc_len=1.66, thru=14.6, I=14.1, T=32.3, wait=5.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:14:16,424] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12360: forward=2829.9ms backward=1083.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:16,478] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:21,122] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:14:23,790] loop.py:45 INFO TRAIN_STEP step=12362 loss=1.344137 Training: 69%|██████▊ | 12362/18018 [26:44:09<12:09:04, 7.73s/step, loss=1.344, acc=0.690, acc_len=1.65, thru=17.1, I=24.0, T=29.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:24,142] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:30,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:14:30,960] loop.py:45 INFO TRAIN_STEP step=12363 loss=1.188613 Training: 69%|██████▊ | 12363/18018 [26:44:17<11:52:59, 7.56s/step, loss=1.189, acc=0.719, acc_len=1.74, thru=14.5, I=22.8, T=29.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:31,025] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:38,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:14:40,004] loop.py:45 INFO TRAIN_STEP step=12364 loss=1.274879 Training: 69%|██████▊ | 12364/18018 [26:44:26<12:34:34, 8.01s/step, loss=1.275, acc=0.708, acc_len=1.71, thru=13.2, I=17.3, T=32.4, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:40,414] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:47,340] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:14:47,719] loop.py:380 INFO TIMING step=12365: step=3.638s data=1.489s compute=3.611s [fwd=2.522s bwd=1.075s opt=0.003s] dispatch=4.070s [2026-10-19 16:14:47,719] loop.py:45 INFO TRAIN_STEP step=12365 loss=1.264446 Training: 69%|██████▊ | 12365/18018 [26:44:33<12:26:13, 7.92s/step, loss=1.264, acc=0.709, acc_len=1.71, thru=14.6, I=18.8, T=35.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:14:47,852] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 16:14:54,391] loop.py:45 INFO TRAIN_STEP step=12366 loss=1.333062 Training: 69%|██████▊ | 12366/18018 [26:44:40<11:50:47, 7.55s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=18.1, I=17.8, T=34.6, wait=3.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:14:54,387] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12365: forward=2582.7ms backward=1059.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:00,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:15:00,905] loop.py:45 INFO TRAIN_STEP step=12367 loss=1.294801 Training: 69%|██████▊ | 12367/18018 [26:44:46<11:21:37, 7.24s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=19.2, I=19.4, T=35.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:01,159] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:06,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:15:07,470] loop.py:45 INFO TRAIN_STEP step=12368 loss=1.283244 Training: 69%|██████▊ | 12368/18018 [26:44:53<11:02:27, 7.04s/step, loss=1.283, acc=0.707, acc_len=1.71, thru=19.7, I=21.4, T=37.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:07,971] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:14,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:15:14,585] loop.py:45 INFO TRAIN_STEP step=12369 loss=1.245523 Training: 69%|██████▊ | 12369/18018 [26:45:00<11:04:32, 7.06s/step, loss=1.246, acc=0.710, acc_len=1.72, thru=18.6, I=20.7, T=35.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:15,079] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:15:21,863] loop.py:380 INFO TIMING step=12370: step=3.772s data=1.603s compute=3.741s [fwd=2.645s bwd=1.082s opt=0.003s] dispatch=3.499s [2026-10-19 16:15:21,863] loop.py:45 INFO TRAIN_STEP step=12370 loss=1.160091 Training: 69%|██████▊ | 12370/18018 [26:45:07<11:10:47, 7.13s/step, loss=1.160, acc=0.727, acc_len=1.78, thru=13.4, I=17.2, T=33.9, wait=3.5s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:31,207] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:15:31,790] loop.py:45 INFO TRAIN_STEP step=12371 loss=1.374863 Training: 69%|██████▊ | 12371/18018 [26:45:17<12:29:40, 7.97s/step, loss=1.375, acc=0.684, acc_len=1.62, thru=16.6, I=15.7, T=30.4, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:15:31,786] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12370: forward=2962.8ms backward=1198.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:32,211] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:15:38,781] loop.py:45 INFO TRAIN_STEP step=12372 loss=1.268579 Training: 69%|██████▊ | 12372/18018 [26:45:24<12:02:02, 7.67s/step, loss=1.269, acc=0.709, acc_len=1.71, thru=14.6, I=16.5, T=32.3, wait=3.0s, pool=41, epoch=2/2][2026-10-19 16:15:46,809] loop.py:45 INFO TRAIN_STEP step=12373 loss=1.306516 Training: 69%|██████▊ | 12373/18018 [26:45:32<12:11:51, 7.78s/step, loss=1.307, acc=0.703, acc_len=1.69, thru=17.4, I=15.9, T=32.3, wait=4.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:52,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:15:54,030] loop.py:45 INFO TRAIN_STEP step=12374 loss=1.305272 Training: 69%|██████▊ | 12374/18018 [26:45:40<11:56:02, 7.61s/step, loss=1.305, acc=0.696, acc_len=1.67, thru=18.3, I=18.7, T=34.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:15:54,454] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:16:00,726] loop.py:380 INFO TIMING step=12375: step=3.309s data=1.308s compute=3.281s [fwd=2.361s bwd=0.906s opt=0.003s] dispatch=3.377s [2026-10-19 16:16:00,726] loop.py:45 INFO TRAIN_STEP step=12375 loss=1.321518 Training: 69%|██████▊ | 12375/18018 [26:45:46<11:30:06, 7.34s/step, loss=1.322, acc=0.694, acc_len=1.65, thru=14.2, I=22.3, T=38.7, wait=3.4s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:16:08,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:16:09,605] loop.py:45 INFO TRAIN_STEP step=12376 loss=1.310970 Training: 69%|██████▊ | 12376/18018 [26:45:55<12:13:23, 7.80s/step, loss=1.311, acc=0.697, acc_len=1.67, thru=18.9, I=16.2, T=34.4, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:16:09,600] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12375: forward=2637.0ms backward=1049.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:16:09,737] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:16:16,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:16:17,366] loop.py:45 INFO TRAIN_STEP step=12377 loss=1.335765 Training: 69%|██████▊ | 12377/18018 [26:46:03<12:12:14, 7.79s/step, loss=1.336, acc=0.691, acc_len=1.64, thru=15.7, I=17.5, T=32.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:16:17,406] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:16:23,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:16:24,991] loop.py:45 INFO TRAIN_STEP step=12378 loss=1.348065 Training: 69%|██████▊ | 12378/18018 [26:46:11<12:07:30, 7.74s/step, loss=1.348, acc=0.691, acc_len=1.65, thru=16.9, I=18.8, T=32.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:16:25,306] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:16:31,213] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:16:33,367] loop.py:45 INFO TRAIN_STEP step=12379 loss=1.216047 Training: 69%|██████▊ | 12379/18018 [26:46:19<12:25:18, 7.93s/step, loss=1.216, acc=0.715, acc_len=1.73, thru=14.1, I=20.4, T=29.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:16:33,728] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 16:16:39,820] loop.py:380 INFO TIMING step=12380: step=3.768s data=1.732s compute=3.725s [fwd=2.567s bwd=1.144s opt=0.003s] dispatch=2.676s [2026-10-19 16:16:39,820] loop.py:45 INFO TRAIN_STEP step=12380 loss=1.162581 Training: 69%|██████▊ | 12380/18018 [26:46:25<11:43:29, 7.49s/step, loss=1.163, acc=0.725, acc_len=1.77, thru=15.2, I=21.9, T=34.0, wait=2.7s, pool=41, epoch=2/2][2026-10-19 16:16:47,753] loop.py:45 INFO TRAIN_STEP step=12381 loss=1.309139 Training: 69%|██████▊ | 12381/18018 [26:46:33<11:55:55, 7.62s/step, loss=1.309, acc=0.693, acc_len=1.66, thru=16.7, I=17.3, T=34.7, wait=4.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:16:47,749] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12380: forward=2507.5ms backward=1105.5ms [2026-10-19 16:16:55,378] loop.py:45 INFO TRAIN_STEP step=12382 loss=1.283763 Training: 69%|██████▊ | 12382/18018 [26:46:41<11:55:56, 7.62s/step, loss=1.284, acc=0.707, acc_len=1.70, thru=15.7, I=16.7, T=34.5, wait=3.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:01,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:17:01,995] loop.py:45 INFO TRAIN_STEP step=12383 loss=1.294741 Training: 69%|██████▊ | 12383/18018 [26:46:48<11:27:29, 7.32s/step, loss=1.295, acc=0.704, acc_len=1.69, thru=19.0, I=19.8, T=35.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:02,428] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:07,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:17:08,631] loop.py:45 INFO TRAIN_STEP step=12384 loss=1.285354 Training: 69%|██████▊ | 12384/18018 [26:46:54<11:08:06, 7.12s/step, loss=1.285, acc=0.701, acc_len=1.68, thru=19.1, I=23.5, T=31.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:08,660] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:13,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:17:14,615] loop.py:380 INFO TIMING step=12385: step=3.979s data=1.673s compute=3.952s [fwd=2.794s bwd=1.144s opt=0.003s] dispatch=1.996s [2026-10-19 16:17:14,616] loop.py:45 INFO TRAIN_STEP step=12385 loss=1.288070 Training: 69%|██████▊ | 12385/18018 [26:47:00<10:36:07, 6.78s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=19.1, I=24.5, T=32.2, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:15,077] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:23,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:17:23,630] loop.py:45 INFO TRAIN_STEP step=12386 loss=1.327980 Training: 69%|██████▊ | 12386/18018 [26:47:09<11:39:08, 7.45s/step, loss=1.328, acc=0.695, acc_len=1.66, thru=15.0, I=15.1, T=29.2, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:17:23,626] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12385: forward=2970.4ms backward=1370.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:24,138] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:30,241] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:17:31,289] loop.py:45 INFO TRAIN_STEP step=12387 loss=1.257208 Training: 69%|██████▊ | 12387/18018 [26:47:17<11:44:52, 7.51s/step, loss=1.257, acc=0.706, acc_len=1.70, thru=17.3, I=19.4, T=30.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:31,750] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:37,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:17:37,885] loop.py:45 INFO TRAIN_STEP step=12388 loss=1.262540 Training: 69%|██████▉ | 12388/18018 [26:47:23<11:19:03, 7.24s/step, loss=1.263, acc=0.713, acc_len=1.73, thru=17.3, I=22.0, T=32.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:38,119] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:44,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:17:45,224] loop.py:45 INFO TRAIN_STEP step=12389 loss=1.339430 Training: 69%|██████▉ | 12389/18018 [26:47:31<11:21:45, 7.27s/step, loss=1.339, acc=0.695, acc_len=1.66, thru=17.3, I=21.0, T=34.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:45,607] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:17:52,566] loop.py:380 INFO TIMING step=12390: step=3.874s data=1.846s compute=3.845s [fwd=2.823s bwd=1.008s opt=0.003s] dispatch=3.459s [2026-10-19 16:17:52,566] loop.py:45 INFO TRAIN_STEP step=12390 loss=1.225845 Training: 69%|██████▉ | 12390/18018 [26:47:38<11:23:49, 7.29s/step, loss=1.226, acc=0.715, acc_len=1.73, thru=16.2, I=20.1, T=33.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:52,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:53,074] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:58,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:17:59,674] loop.py:45 INFO TRAIN_STEP step=12391 loss=1.344742 Training: 69%|██████▉ | 12391/18018 [26:47:45<11:18:32, 7.24s/step, loss=1.345, acc=0.691, acc_len=1.65, thru=17.6, I=19.0, T=37.5, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:17:59,669] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12390: forward=2430.5ms backward=930.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:17:59,940] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:18:08,038] loop.py:45 INFO TRAIN_STEP step=12392 loss=1.442769 Training: 69%|██████▉ | 12392/18018 [26:47:54<11:50:12, 7.57s/step, loss=1.443, acc=0.675, acc_len=1.59, thru=13.8, I=17.8, T=36.1, wait=4.8s, pool=49, epoch=2/2][2026-10-19 16:18:16,737] loop.py:45 INFO TRAIN_STEP step=12393 loss=1.353305 Training: 69%|██████▉ | 12393/18018 [26:48:02<12:21:38, 7.91s/step, loss=1.353, acc=0.690, acc_len=1.64, thru=13.5, I=15.1, T=34.0, wait=4.9s, pool=49, epoch=2/2][2026-10-19 16:18:27,358] loop.py:45 INFO TRAIN_STEP step=12394 loss=1.262617 Training: 69%|██████▉ | 12394/18018 [26:48:13<13:37:46, 8.72s/step, loss=1.263, acc=0.703, acc_len=1.69, thru=12.9, I=12.7, T=30.0, wait=6.3s, pool=49, epoch=2/2][2026-10-19 16:18:36,851] loop.py:380 INFO TIMING step=12395: step=4.039s data=1.756s compute=4.006s [fwd=2.821s bwd=1.170s opt=0.003s] dispatch=5.444s [2026-10-19 16:18:36,851] loop.py:45 INFO TRAIN_STEP step=12395 loss=1.487024 Training: 69%|██████▉ | 12395/18018 [26:48:22<13:59:12, 8.95s/step, loss=1.487, acc=0.673, acc_len=1.59, thru=14.6, I=12.6, T=31.7, wait=5.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:18:43,177] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:18:44,438] loop.py:45 INFO TRAIN_STEP step=12396 loss=1.302038 Training: 69%|██████▉ | 12396/18018 [26:48:30<13:20:39, 8.54s/step, loss=1.302, acc=0.699, acc_len=1.68, thru=17.6, I=18.4, T=33.5, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:18:44,433] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12395: forward=2620.2ms backward=1133.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:18:44,687] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:18:50,956] loop.py:45 INFO TRAIN_STEP step=12397 loss=1.252500 Training: 69%|██████▉ | 12397/18018 [26:48:37<12:23:27, 7.94s/step, loss=1.252, acc=0.711, acc_len=1.72, thru=18.1, I=21.8, T=34.7, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:18:59,076] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:19:00,594] loop.py:45 INFO TRAIN_STEP step=12398 loss=1.212577 Training: 69%|██████▉ | 12398/18018 [26:48:46<13:11:09, 8.45s/step, loss=1.213, acc=0.715, acc_len=1.73, thru=14.1, I=15.0, T=30.1, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:19:01,083] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:19:06,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:19:07,702] loop.py:45 INFO TRAIN_STEP step=12399 loss=1.374212 Training: 69%|██████▉ | 12399/18018 [26:48:53<12:33:24, 8.04s/step, loss=1.374, acc=0.685, acc_len=1.63, thru=16.1, I=21.7, T=31.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:19:07,877] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:19:13,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:19:13,918] loop.py:380 INFO TIMING step=12400: step=3.608s data=1.588s compute=3.575s [fwd=2.603s bwd=0.958s opt=0.003s] dispatch=2.601s [2026-10-19 16:19:13,919] loop.py:45 INFO TRAIN_STEP step=12400 loss=1.290296 Training: 69%|██████▉ | 12400/18018 [26:49:00<11:41:58, 7.50s/step, loss=1.290, acc=0.700, acc_len=1.68, thru=18.9, I=21.1, T=35.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:19:14,308] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:19:20,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:19:21,319] loop.py:45 INFO TRAIN_STEP step=12401 loss=1.200528 Training: 69%|██████▉ | 12401/18018 [26:49:07<11:39:10, 7.47s/step, loss=1.201, acc=0.718, acc_len=1.74, thru=16.0, I=19.4, T=36.6, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:19:21,314] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12400: forward=2526.1ms backward=923.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:19:21,392] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:19:29,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:19:29,711] loop.py:45 INFO TRAIN_STEP step=12402 loss=1.251559 Training: 69%|██████▉ | 12402/18018 [26:49:15<12:04:56, 7.75s/step, loss=1.252, acc=0.708, acc_len=1.71, thru=14.4, I=15.4, T=31.4, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:19:29,960] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:19:37,770] loop.py:45 INFO TRAIN_STEP step=12403 loss=1.377181 Training: 69%|██████▉ | 12403/18018 [26:49:23<12:13:39, 7.84s/step, loss=1.377, acc=0.687, acc_len=1.63, thru=15.8, I=15.3, T=35.6, wait=4.5s, pool=53, epoch=2/2][2026-10-19 16:19:46,819] loop.py:45 INFO TRAIN_STEP step=12404 loss=1.181755 Training: 69%|██████▉ | 12404/18018 [26:49:32<12:47:23, 8.20s/step, loss=1.182, acc=0.724, acc_len=1.76, thru=15.4, I=14.4, T=34.0, wait=5.3s, pool=53, epoch=2/2][2026-10-19 16:19:53,929] loop.py:380 INFO TIMING step=12405: step=3.885s data=1.726s compute=3.855s [fwd=2.763s bwd=1.078s opt=0.003s] dispatch=3.218s [2026-10-19 16:19:53,930] loop.py:45 INFO TRAIN_STEP step=12405 loss=1.261124 Training: 69%|██████▉ | 12405/18018 [26:49:40<12:16:39, 7.87s/step, loss=1.261, acc=0.709, acc_len=1.71, thru=17.8, I=18.9, T=33.0, wait=3.2s, pool=57, epoch=2/2][2026-10-19 16:20:01,192] loop.py:45 INFO TRAIN_STEP step=12406 loss=1.363400 Training: 69%|██████▉ | 12406/18018 [26:49:47<11:59:22, 7.69s/step, loss=1.363, acc=0.689, acc_len=1.65, thru=16.7, I=18.4, T=32.7, wait=3.3s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:20:01,187] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12405: forward=2687.8ms backward=1168.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:12,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:20:12,396] loop.py:45 INFO TRAIN_STEP step=12407 loss=1.231204 Training: 69%|██████▉ | 12407/18018 [26:49:58<13:37:48, 8.75s/step, loss=1.231, acc=0.715, acc_len=1.73, thru=13.2, I=10.9, T=30.2, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:12,666] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:18,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:20:18,847] loop.py:45 INFO TRAIN_STEP step=12408 loss=1.158357 Training: 69%|██████▉ | 12408/18018 [26:50:04<12:33:18, 8.06s/step, loss=1.158, acc=0.726, acc_len=1.77, thru=18.5, I=20.2, T=33.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:19,292] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:25,710] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:20:27,688] loop.py:45 INFO TRAIN_STEP step=12409 loss=1.200584 Training: 69%|██████▉ | 12409/18018 [26:50:13<12:55:11, 8.29s/step, loss=1.201, acc=0.720, acc_len=1.76, thru=14.2, I=17.0, T=32.0, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:27,725] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:33,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:20:34,278] loop.py:380 INFO TIMING step=12410: step=3.842s data=1.740s compute=3.812s [fwd=2.691s bwd=1.106s opt=0.003s] dispatch=2.738s [2026-10-19 16:20:34,278] loop.py:45 INFO TRAIN_STEP step=12410 loss=1.303593 Training: 69%|██████▉ | 12410/18018 [26:50:20<12:07:18, 7.78s/step, loss=1.304, acc=0.700, acc_len=1.68, thru=17.7, I=22.8, T=33.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:34,780] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:20:40,717] loop.py:45 INFO TRAIN_STEP step=12411 loss=1.300280 Training: 69%|██████▉ | 12411/18018 [26:50:26<11:29:28, 7.38s/step, loss=1.300, acc=0.701, acc_len=1.69, thru=16.4, I=20.2, T=36.2, wait=2.9s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:20:40,712] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12410: forward=2452.8ms backward=1025.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:51,094] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:20:51,317] loop.py:45 INFO TRAIN_STEP step=12412 loss=1.343443 Training: 69%|██████▉ | 12412/18018 [26:50:37<12:59:41, 8.34s/step, loss=1.343, acc=0.697, acc_len=1.66, thru=14.6, I=14.2, T=32.9, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:20:51,598] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:20:57,431] loop.py:45 INFO TRAIN_STEP step=12413 loss=1.286548 Training: 69%|██████▉ | 12413/18018 [26:50:43<11:57:01, 7.68s/step, loss=1.287, acc=0.702, acc_len=1.69, thru=17.2, I=21.2, T=36.7, wait=2.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:05,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:21:07,188] loop.py:45 INFO TRAIN_STEP step=12414 loss=1.238694 Training: 69%|██████▉ | 12414/18018 [26:50:53<12:55:17, 8.30s/step, loss=1.239, acc=0.709, acc_len=1.71, thru=15.4, I=15.6, T=26.5, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:07,227] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:14,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:21:14,258] loop.py:380 INFO TIMING step=12415: step=3.717s data=1.488s compute=3.690s [fwd=2.610s bwd=1.065s opt=0.003s] dispatch=3.343s [2026-10-19 16:21:14,258] loop.py:45 INFO TRAIN_STEP step=12415 loss=1.277698 Training: 69%|██████▉ | 12415/18018 [26:51:00<12:20:36, 7.93s/step, loss=1.278, acc=0.707, acc_len=1.70, thru=16.3, I=17.4, T=34.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:14,651] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:20,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:21:20,536] loop.py:45 INFO TRAIN_STEP step=12416 loss=1.250016 Training: 69%|██████▉ | 12416/18018 [26:51:06<11:34:10, 7.43s/step, loss=1.250, acc=0.710, acc_len=1.72, thru=20.4, I=19.7, T=34.5, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:21:20,532] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12415: forward=2629.7ms backward=1030.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:20,973] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:21:28,319] loop.py:45 INFO TRAIN_STEP step=12417 loss=1.216279 Training: 69%|██████▉ | 12417/18018 [26:51:14<11:43:56, 7.54s/step, loss=1.216, acc=0.715, acc_len=1.74, thru=15.0, I=18.2, T=35.4, wait=4.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:37,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:21:37,659] loop.py:45 INFO TRAIN_STEP step=12418 loss=1.214729 Training: 69%|██████▉ | 12418/18018 [26:51:23<12:34:04, 8.08s/step, loss=1.215, acc=0.716, acc_len=1.74, thru=16.1, I=14.8, T=35.9, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:38,024] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:43,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:21:43,475] loop.py:45 INFO TRAIN_STEP step=12419 loss=1.440521 Training: 69%|██████▉ | 12419/18018 [26:51:29<11:30:36, 7.40s/step, loss=1.441, acc=0.674, acc_len=1.59, thru=21.2, I=23.6, T=36.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:43,536] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:52,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:21:53,068] loop.py:380 INFO TIMING step=12420: step=3.602s data=1.476s compute=3.574s [fwd=2.533s bwd=1.027s opt=0.003s] dispatch=5.981s [2026-10-19 16:21:53,069] loop.py:45 INFO TRAIN_STEP step=12420 loss=1.149129 Training: 69%|██████▉ | 12420/18018 [26:51:39<12:31:49, 8.06s/step, loss=1.149, acc=0.731, acc_len=1.79, thru=13.9, I=14.5, T=35.5, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:53,256] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:58,576] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:21:58,961] loop.py:45 INFO TRAIN_STEP step=12421 loss=1.257338 Training: 69%|██████▉ | 12421/18018 [26:51:45<11:31:07, 7.41s/step, loss=1.257, acc=0.709, acc_len=1.71, thru=20.6, I=20.5, T=37.3, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:21:58,957] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12420: forward=2448.6ms backward=933.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:21:59,081] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 16:22:06,459] loop.py:45 INFO TRAIN_STEP step=12422 loss=1.243258 Training: 69%|██████▉ | 12422/18018 [26:51:52<11:33:25, 7.43s/step, loss=1.243, acc=0.710, acc_len=1.71, thru=16.0, I=16.6, T=32.5, wait=3.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:12,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:22:13,854] loop.py:45 INFO TRAIN_STEP step=12423 loss=1.304099 Training: 69%|██████▉ | 12423/18018 [26:51:59<11:32:14, 7.42s/step, loss=1.304, acc=0.698, acc_len=1.67, thru=18.0, I=19.4, T=34.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:14,345] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:21,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:22:22,059] loop.py:45 INFO TRAIN_STEP step=12424 loss=1.364875 Training: 69%|██████▉ | 12424/18018 [26:52:08<11:53:59, 7.66s/step, loss=1.365, acc=0.691, acc_len=1.65, thru=14.8, I=20.2, T=34.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:22,093] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:27,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:22:27,992] loop.py:380 INFO TIMING step=12425: step=3.707s data=1.620s compute=3.679s [fwd=2.629s bwd=1.034s opt=0.003s] dispatch=2.215s [2026-10-19 16:22:27,992] loop.py:45 INFO TRAIN_STEP step=12425 loss=1.337162 Training: 69%|██████▉ | 12425/18018 [26:52:14<11:05:34, 7.14s/step, loss=1.337, acc=0.690, acc_len=1.65, thru=20.5, I=24.3, T=34.5, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:28,154] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:33,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:22:34,120] loop.py:45 INFO TRAIN_STEP step=12426 loss=1.178334 Training: 69%|██████▉ | 12426/18018 [26:52:20<10:37:10, 6.84s/step, loss=1.178, acc=0.724, acc_len=1.76, thru=19.9, I=22.8, T=36.8, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:22:34,115] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12425: forward=2440.5ms backward=988.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:34,613] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:22:41,614] loop.py:45 INFO TRAIN_STEP step=12427 loss=1.290793 Training: 69%|██████▉ | 12427/18018 [26:52:27<10:55:27, 7.03s/step, loss=1.291, acc=0.704, acc_len=1.70, thru=15.8, I=17.2, T=35.0, wait=3.8s, pool=53, epoch=2/2][2026-10-19 16:22:48,692] loop.py:45 INFO TRAIN_STEP step=12428 loss=1.270414 Training: 69%|██████▉ | 12428/18018 [26:52:35<11:04:35, 7.13s/step, loss=1.270, acc=0.703, acc_len=1.69, thru=13.4, I=18.8, T=29.8, wait=2.8s, pool=13, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:57,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:22:58,004] loop.py:45 INFO TRAIN_STEP step=12429 loss=1.265441 Training: 69%|██████▉ | 12429/18018 [26:52:44<11:57:17, 7.70s/step, loss=1.265, acc=0.706, acc_len=1.71, thru=20.4, I=21.0, T=37.1, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:22:58,354] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:23:05,991] loop.py:380 INFO TIMING step=12430: step=3.594s data=1.534s compute=3.570s [fwd=2.552s bwd=1.005s opt=0.003s] dispatch=4.384s [2026-10-19 16:23:05,991] loop.py:45 INFO TRAIN_STEP step=12430 loss=1.284276 Training: 69%|██████▉ | 12430/18018 [26:52:52<12:05:10, 7.79s/step, loss=1.284, acc=0.703, acc_len=1.68, thru=13.8, I=16.7, T=35.6, wait=4.4s, pool=37, epoch=2/2][2026-10-19 16:23:14,469] loop.py:45 INFO TRAIN_STEP step=12431 loss=1.254734 Training: 69%|██████▉ | 12431/18018 [26:53:00<12:24:26, 7.99s/step, loss=1.255, acc=0.711, acc_len=1.72, thru=18.4, I=15.8, T=33.8, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:23:14,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 16:23:14,465] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12430: forward=2609.3ms backward=1125.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:23:14,977] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:23:21,042] loop.py:45 INFO TRAIN_STEP step=12432 loss=1.263099 Training: 69%|██████▉ | 12432/18018 [26:53:07<11:44:35, 7.57s/step, loss=1.263, acc=0.706, acc_len=1.71, thru=18.4, I=18.0, T=38.2, wait=3.2s, pool=61, epoch=2/2][2026-10-19 16:23:28,658] loop.py:45 INFO TRAIN_STEP step=12433 loss=1.311953 Training: 69%|██████▉ | 12433/18018 [26:53:14<11:45:44, 7.58s/step, loss=1.312, acc=0.700, acc_len=1.67, thru=17.2, I=17.8, T=35.2, wait=4.0s, pool=61, epoch=2/2][2026-10-19 16:23:37,376] loop.py:45 INFO TRAIN_STEP step=12434 loss=1.249568 Training: 69%|██████▉ | 12434/18018 [26:53:23<12:17:19, 7.92s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=14.1, I=14.6, T=35.8, wait=5.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:23:46,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:23:46,518] loop.py:380 INFO TIMING step=12435: step=3.626s data=1.493s compute=3.592s [fwd=2.568s bwd=1.011s opt=0.003s] dispatch=5.507s [2026-10-19 16:23:46,518] loop.py:45 INFO TRAIN_STEP step=12435 loss=1.297833 Training: 69%|██████▉ | 12435/18018 [26:53:32<12:51:11, 8.29s/step, loss=1.298, acc=0.702, acc_len=1.69, thru=15.5, I=15.1, T=35.3, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:23:46,967] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:23:53,538] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:23:54,115] loop.py:45 INFO TRAIN_STEP step=12436 loss=1.270759 Training: 69%|██████▉ | 12436/18018 [26:53:40<12:31:50, 8.08s/step, loss=1.271, acc=0.701, acc_len=1.68, thru=16.0, I=17.5, T=35.7, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:23:54,110] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12435: forward=2553.6ms backward=974.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:23:54,543] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:23:58,689] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:24:00,459] loop.py:45 INFO TRAIN_STEP step=12437 loss=1.261746 Training: 69%|██████▉ | 12437/18018 [26:53:46<11:43:15, 7.56s/step, loss=1.262, acc=0.703, acc_len=1.69, thru=19.7, I=26.1, T=39.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:00,705] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:07,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:24:07,617] loop.py:45 INFO TRAIN_STEP step=12438 loss=1.185947 Training: 69%|██████▉ | 12438/18018 [26:53:53<11:31:49, 7.44s/step, loss=1.186, acc=0.723, acc_len=1.76, thru=17.8, I=23.1, T=35.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:07,943] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:13,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:24:13,919] loop.py:45 INFO TRAIN_STEP step=12439 loss=1.328360 Training: 69%|██████▉ | 12439/18018 [26:54:00<11:00:01, 7.10s/step, loss=1.328, acc=0.699, acc_len=1.68, thru=19.8, I=19.0, T=35.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:14,145] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:24:20,626] loop.py:380 INFO TIMING step=12440: step=3.461s data=1.456s compute=3.431s [fwd=2.434s bwd=0.985s opt=0.003s] dispatch=3.237s [2026-10-19 16:24:20,626] loop.py:45 INFO TRAIN_STEP step=12440 loss=1.254110 Training: 69%|██████▉ | 12440/18018 [26:54:06<10:48:56, 6.98s/step, loss=1.254, acc=0.709, acc_len=1.71, thru=15.3, I=17.1, T=37.0, wait=3.2s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:29,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:24:29,603] loop.py:45 INFO TRAIN_STEP step=12441 loss=1.298417 Training: 69%|██████▉ | 12441/18018 [26:54:15<11:44:30, 7.58s/step, loss=1.298, acc=0.704, acc_len=1.69, thru=18.0, I=16.3, T=32.0, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:24:29,598] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12440: forward=2733.0ms backward=1209.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:29,810] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:35,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:24:35,923] loop.py:45 INFO TRAIN_STEP step=12442 loss=1.249700 Training: 69%|██████▉ | 12442/18018 [26:54:22<11:09:16, 7.20s/step, loss=1.250, acc=0.707, acc_len=1.70, thru=19.4, I=18.2, T=36.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:36,233] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:41,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:24:41,752] loop.py:45 INFO TRAIN_STEP step=12443 loss=1.079201 Training: 69%|██████▉ | 12443/18018 [26:54:27<10:30:53, 6.79s/step, loss=1.079, acc=0.740, acc_len=1.82, thru=20.8, I=22.2, T=37.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:41,959] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:49,689] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:24:50,005] loop.py:45 INFO TRAIN_STEP step=12444 loss=1.306251 Training: 69%|██████▉ | 12444/18018 [26:54:36<11:11:36, 7.23s/step, loss=1.306, acc=0.697, acc_len=1.67, thru=16.2, I=17.4, T=34.5, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:24:50,192] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:24:57,954] loop.py:380 INFO TIMING step=12445: step=3.744s data=1.613s compute=3.713s [fwd=2.664s bwd=1.034s opt=0.003s] dispatch=4.194s [2026-10-19 16:24:57,954] loop.py:45 INFO TRAIN_STEP step=12445 loss=1.343880 Training: 69%|██████▉ | 12445/18018 [26:54:44<11:31:28, 7.44s/step, loss=1.344, acc=0.693, acc_len=1.66, thru=13.7, I=14.0, T=34.2, wait=4.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:25:07,213] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:25:08,629] loop.py:45 INFO TRAIN_STEP step=12446 loss=1.323763 Training: 69%|██████▉ | 12446/18018 [26:54:54<13:01:26, 8.41s/step, loss=1.324, acc=0.695, acc_len=1.67, thru=14.3, I=16.0, T=31.7, wait=6.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:25:08,624] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12445: forward=2859.6ms backward=1133.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:25:08,736] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:25:13,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:25:14,263] loop.py:45 INFO TRAIN_STEP step=12447 loss=1.255719 Training: 69%|██████▉ | 12447/18018 [26:55:00<11:43:48, 7.58s/step, loss=1.256, acc=0.710, acc_len=1.71, thru=20.2, I=23.6, T=37.3, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:25:14,493] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:25:22,112] loop.py:45 INFO TRAIN_STEP step=12448 loss=1.261423 Training: 69%|██████▉ | 12448/18018 [26:55:08<11:51:12, 7.66s/step, loss=1.261, acc=0.705, acc_len=1.71, thru=16.2, I=19.6, T=34.0, wait=4.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:25:29,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:25:30,097] loop.py:45 INFO TRAIN_STEP step=12449 loss=1.222116 Training: 69%|██████▉ | 12449/18018 [26:55:16<12:00:07, 7.76s/step, loss=1.222, acc=0.715, acc_len=1.73, thru=15.8, I=17.5, T=31.8, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:25:30,547] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:25:35,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:25:37,939] loop.py:380 INFO TIMING step=12450: step=3.692s data=1.637s compute=3.652s [fwd=2.604s bwd=1.033s opt=0.003s] dispatch=4.138s [2026-10-19 16:25:37,940] loop.py:45 INFO TRAIN_STEP step=12450 loss=1.238762 Training: 69%|██████▉ | 12450/18018 [26:55:24<12:02:14, 7.78s/step, loss=1.239, acc=0.713, acc_len=1.72, thru=15.4, I=22.7, T=34.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:25:38,195] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-19 16:25:44,276] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12450: forward=2538.1ms backward=992.0ms [2026-10-19 16:25:44,281] loop.py:45 INFO TRAIN_STEP step=12451 loss=1.333294 Training: 69%|██████▉ | 12451/18018 [26:55:30<11:22:02, 7.35s/step, loss=1.333, acc=0.693, acc_len=1.65, thru=18.5, I=25.9, T=35.8, wait=2.8s, pool=57, epoch=2/2][2026-10-19 16:25:52,796] loop.py:45 INFO TRAIN_STEP step=12452 loss=1.410027 Training: 69%|██████▉ | 12452/18018 [26:55:38<11:54:18, 7.70s/step, loss=1.410, acc=0.682, acc_len=1.62, thru=14.5, I=15.0, T=36.1, wait=5.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:25:58,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:25:59,968] loop.py:45 INFO TRAIN_STEP step=12453 loss=1.285596 Training: 69%|██████▉ | 12453/18018 [26:55:46<11:39:29, 7.54s/step, loss=1.286, acc=0.706, acc_len=1.70, thru=17.7, I=20.7, T=33.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:00,365] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:26:07,900] loop.py:45 INFO TRAIN_STEP step=12454 loss=1.171463 Training: 69%|██████▉ | 12454/18018 [26:55:53<11:50:14, 7.66s/step, loss=1.171, acc=0.720, acc_len=1.76, thru=14.2, I=18.9, T=28.6, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:15,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:26:16,984] loop.py:380 INFO TIMING step=12455: step=3.798s data=1.520s compute=3.767s [fwd=2.677s bwd=1.076s opt=0.003s] dispatch=5.275s [2026-10-19 16:26:16,984] loop.py:45 INFO TRAIN_STEP step=12455 loss=1.279393 Training: 69%|██████▉ | 12455/18018 [26:56:03<12:29:41, 8.09s/step, loss=1.279, acc=0.701, acc_len=1.69, thru=14.4, I=16.0, T=33.7, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:17,293] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:22,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:26:23,131] loop.py:45 INFO TRAIN_STEP step=12456 loss=1.283687 Training: 69%|██████▉ | 12456/18018 [26:56:09<11:35:37, 7.50s/step, loss=1.284, acc=0.703, acc_len=1.69, thru=19.6, I=25.9, T=37.8, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:26:23,126] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12455: forward=2380.5ms backward=962.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:23,383] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:30,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:26:31,920] loop.py:45 INFO TRAIN_STEP step=12457 loss=1.146347 Training: 69%|██████▉ | 12457/18018 [26:56:18<12:11:17, 7.89s/step, loss=1.146, acc=0.726, acc_len=1.77, thru=14.6, I=16.2, T=33.8, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:32,012] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 16:26:38,490] loop.py:45 INFO TRAIN_STEP step=12458 loss=1.263342 Training: 69%|██████▉ | 12458/18018 [26:56:24<11:34:22, 7.49s/step, loss=1.263, acc=0.703, acc_len=1.69, thru=14.9, I=17.2, T=39.0, wait=3.3s, pool=25, epoch=2/2](TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 16:26:38,549] eagle_store.py:517 WARNING batch_get_buffer missing keys (k398664-a0b59976_hs, k398664-a0b59976_ids, k398664-a0b59976_lhs); sleeping 0.50s. [2026-10-19 16:26:47,453] loop.py:45 INFO TRAIN_STEP step=12459 loss=1.383607 Training: 69%|██████▉ | 12459/18018 [26:56:33<12:15:06, 7.93s/step, loss=1.384, acc=0.687, acc_len=1.64, thru=17.5, I=14.7, T=33.6, wait=5.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:52,002] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 16:26:38,653] eagle_store.py:517 WARNING batch_get_buffer missing keys (k398666-9d10631b_hs, k398666-9d10631b_ids, k398666-9d10631b_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 16:26:53,888] loop.py:380 INFO TIMING step=12460: step=3.846s data=1.660s compute=3.813s [fwd=2.634s bwd=1.163s opt=0.003s] dispatch=2.582s [2026-10-19 16:26:53,888] loop.py:45 INFO TRAIN_STEP step=12460 loss=1.208076 Training: 69%|██████▉ | 12460/18018 [26:56:39<11:33:23, 7.49s/step, loss=1.208, acc=0.718, acc_len=1.74, thru=18.9, I=22.9, T=33.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:54,019] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:26:59,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:26:59,936] loop.py:45 INFO TRAIN_STEP step=12461 loss=1.272327 Training: 69%|██████▉ | 12461/18018 [26:56:46<10:53:16, 7.05s/step, loss=1.272, acc=0.702, acc_len=1.69, thru=19.7, I=26.2, T=37.8, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:26:59,932] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12460: forward=2389.9ms backward=949.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:00,163] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:06,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:27:07,892] loop.py:45 INFO TRAIN_STEP step=12462 loss=1.196341 Training: 69%|██████▉ | 12462/18018 [26:56:53<11:18:17, 7.33s/step, loss=1.196, acc=0.719, acc_len=1.74, thru=17.2, I=20.6, T=34.8, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:08,105] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:27:13,343] loop.py:45 INFO TRAIN_STEP step=12463 loss=1.223343 Training: 69%|██████▉ | 12463/18018 [26:56:59<10:26:02, 6.76s/step, loss=1.223, acc=0.716, acc_len=1.73, thru=21.2, I=23.9, T=38.1, wait=2.1s, pool=57, epoch=2/2][2026-10-19 16:27:21,424] loop.py:45 INFO TRAIN_STEP step=12464 loss=1.271991 Training: 69%|██████▉ | 12464/18018 [26:57:07<11:02:36, 7.16s/step, loss=1.272, acc=0.705, acc_len=1.70, thru=16.6, I=15.9, T=33.0, wait=4.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:31,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:27:32,368] loop.py:380 INFO TIMING step=12465: step=3.785s data=1.625s compute=3.735s [fwd=2.583s bwd=1.137s opt=0.003s] dispatch=7.150s [2026-10-19 16:27:32,369] loop.py:45 INFO TRAIN_STEP step=12465 loss=1.277425 Training: 69%|██████▉ | 12465/18018 [26:57:18<12:47:35, 8.29s/step, loss=1.277, acc=0.707, acc_len=1.71, thru=13.2, I=13.2, T=33.8, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:32,763] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:27:39,608] loop.py:45 INFO TRAIN_STEP step=12466 loss=1.289123 Training: 69%|██████▉ | 12466/18018 [26:57:25<12:18:12, 7.98s/step, loss=1.289, acc=0.698, acc_len=1.67, thru=16.6, I=19.3, T=35.0, wait=3.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:27:39,604] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12465: forward=2584.5ms backward=1023.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:46,356] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:27:47,480] loop.py:45 INFO TRAIN_STEP step=12467 loss=1.380667 Training: 69%|██████▉ | 12467/18018 [26:57:33<12:15:11, 7.95s/step, loss=1.381, acc=0.682, acc_len=1.61, thru=18.4, I=18.8, T=34.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:47,864] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:53,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:27:53,304] loop.py:45 INFO TRAIN_STEP step=12468 loss=1.359926 Training: 69%|██████▉ | 12468/18018 [26:57:39<11:16:04, 7.31s/step, loss=1.360, acc=0.687, acc_len=1.63, thru=20.7, I=21.6, T=39.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:53,612] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:58,943] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:27:59,362] loop.py:45 INFO TRAIN_STEP step=12469 loss=1.322216 Training: 69%|██████▉ | 12469/18018 [26:57:45<10:41:15, 6.93s/step, loss=1.322, acc=0.695, acc_len=1.66, thru=20.9, I=22.5, T=36.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:27:59,458] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 16:28:06,559] loop.py:380 INFO TIMING step=12470: step=3.996s data=1.823s compute=3.955s [fwd=2.774s bwd=1.166s opt=0.003s] dispatch=3.191s [2026-10-19 16:28:06,559] loop.py:45 INFO TRAIN_STEP step=12470 loss=1.101806 Training: 69%|██████▉ | 12470/18018 [26:57:52<10:48:26, 7.01s/step, loss=1.102, acc=0.736, acc_len=1.81, thru=18.0, I=20.0, T=32.0, wait=3.2s, pool=61, epoch=2/2][2026-10-19 16:28:12,718] loop.py:45 INFO TRAIN_STEP step=12471 loss=1.190672 Training: 69%|██████▉ | 12471/18018 [26:57:58<10:24:42, 6.76s/step, loss=1.191, acc=0.720, acc_len=1.76, thru=18.4, I=19.1, T=36.3, wait=2.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:28:12,714] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12470: forward=2438.8ms backward=1026.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:28:18,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 16:28:19,478] loop.py:45 INFO TRAIN_STEP step=12472 loss=1.251692 Training: 69%|██████▉ | 12472/18018 [26:58:05<10:24:38, 6.76s/step, loss=1.252, acc=0.707, acc_len=1.71, thru=19.4, I=20.6, T=35.9, wait=3.2s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:28:19,743] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:28:25,325] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:28:25,986] loop.py:45 INFO TRAIN_STEP step=12473 loss=1.232352 Training: 69%|██████▉ | 12473/18018 [26:58:12<10:17:36, 6.68s/step, loss=1.232, acc=0.712, acc_len=1.73, thru=19.3, I=21.9, T=35.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:28:26,333] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:28:31,298] loop.py:45 INFO TRAIN_STEP step=12474 loss=1.318438 Training: 69%|██████▉ | 12474/18018 [26:58:17<9:39:26, 6.27s/step, loss=1.318, acc=0.697, acc_len=1.67, thru=21.1, I=24.5, T=38.0, wait=1.9s, pool=53, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:28:38,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:28:40,357] loop.py:380 INFO TIMING step=12475: step=3.934s data=1.687s compute=3.909s [fwd=2.840s bwd=1.054s opt=0.003s] dispatch=5.116s [2026-10-19 16:28:40,357] loop.py:45 INFO TRAIN_STEP step=12475 loss=1.386378 Training: 69%|██████▉ | 12475/18018 [26:58:26<10:56:34, 7.11s/step, loss=1.386, acc=0.686, acc_len=1.63, thru=16.1, I=17.8, T=32.5, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:28:40,806] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:28:47,479] loop.py:45 INFO TRAIN_STEP step=12476 loss=1.222680 Training: 69%|██████▉ | 12476/18018 [26:58:33<10:56:53, 7.11s/step, loss=1.223, acc=0.715, acc_len=1.73, thru=12.9, I=15.2, T=32.9, wait=3.2s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:28:47,474] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12475: forward=2712.4ms backward=1133.6ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 16:28:47,553] eagle_store.py:517 WARNING batch_get_buffer missing keys (k399242-a1010025_hs, k399242-a1010025_ids, k399242-a1010025_lhs); sleeping 0.50s. [2026-10-19 16:28:55,683] loop.py:45 INFO TRAIN_STEP step=12477 loss=1.193032 Training: 69%|██████▉ | 12477/18018 [26:58:41<11:27:05, 7.44s/step, loss=1.193, acc=0.718, acc_len=1.74, thru=18.7, I=15.3, T=37.0, wait=4.7s, pool=61, epoch=2/2][2026-10-19 16:29:03,598] loop.py:45 INFO TRAIN_STEP step=12478 loss=1.290379 Training: 69%|██████▉ | 12478/18018 [26:58:49<11:40:13, 7.58s/step, loss=1.290, acc=0.703, acc_len=1.69, thru=13.8, I=14.3, T=35.3, wait=4.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:29:10,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 16:28:47,544] eagle_store.py:517 WARNING batch_get_buffer missing keys (k399241-9d8675aa_hs, k399241-9d8675aa_ids, k399241-9d8675aa_lhs); sleeping 0.50s. [2026-10-19 16:29:11,310] loop.py:45 INFO TRAIN_STEP step=12479 loss=1.371817 Training: 69%|██████▉ | 12479/18018 [26:58:57<11:43:34, 7.62s/step, loss=1.372, acc=0.689, acc_len=1.64, thru=18.9, I=18.2, T=34.6, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:29:11,534] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:29:20,887] loop.py:380 INFO TIMING step=12480: step=3.689s data=1.554s compute=3.663s [fwd=2.642s bwd=1.007s opt=0.003s] dispatch=5.878s [2026-10-19 16:29:20,887] loop.py:45 INFO TRAIN_STEP step=12480 loss=1.287006 Training: 69%|██████▉ | 12480/18018 [26:59:06<12:37:39, 8.21s/step, loss=1.287, acc=0.703, acc_len=1.70, thru=13.3, I=13.7, T=34.7, wait=5.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:29:28,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:29:29,358] loop.py:45 INFO TRAIN_STEP step=12481 loss=1.362874 Training: 69%|██████▉ | 12481/18018 [26:59:15<12:44:38, 8.29s/step, loss=1.363, acc=0.690, acc_len=1.65, thru=15.0, I=16.8, T=35.5, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:29:29,353] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12480: forward=2498.0ms backward=1048.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:29:29,837] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:29:37,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:29:37,503] loop.py:45 INFO TRAIN_STEP step=12482 loss=1.340994 Training: 69%|██████▉ | 12482/18018 [26:59:23<12:40:38, 8.24s/step, loss=1.341, acc=0.698, acc_len=1.68, thru=14.6, I=17.7, T=37.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:29:37,661] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:29:44,574] loop.py:45 INFO TRAIN_STEP step=12483 loss=1.267564 Training: 69%|██████▉ | 12483/18018 [26:59:30<12:07:59, 7.89s/step, loss=1.268, acc=0.707, acc_len=1.70, thru=16.0, I=16.7, T=35.7, wait=3.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:29:52,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:29:53,147] loop.py:45 INFO TRAIN_STEP step=12484 loss=1.297704 Training: 69%|██████▉ | 12484/18018 [26:59:39<12:26:45, 8.10s/step, loss=1.298, acc=0.700, acc_len=1.68, thru=17.2, I=16.5, T=34.4, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:29:53,431] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:00,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:30:02,146] loop.py:380 INFO TIMING step=12485: step=3.823s data=1.677s compute=3.781s [fwd=2.656s bwd=1.110s opt=0.003s] dispatch=5.169s [2026-10-19 16:30:02,146] loop.py:45 INFO TRAIN_STEP step=12485 loss=1.338029 Training: 69%|██████▉ | 12485/18018 [26:59:48<12:51:33, 8.37s/step, loss=1.338, acc=0.695, acc_len=1.66, thru=14.0, I=16.9, T=33.5, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:02,641] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:08,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:30:09,419] loop.py:45 INFO TRAIN_STEP step=12486 loss=1.335384 Training: 69%|██████▉ | 12486/18018 [26:59:55<12:21:11, 8.04s/step, loss=1.335, acc=0.697, acc_len=1.67, thru=16.2, I=24.6, T=37.8, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:30:09,414] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12485: forward=2398.8ms backward=934.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:09,558] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:15,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:30:16,520] loop.py:45 INFO TRAIN_STEP step=12487 loss=1.219186 Training: 69%|██████▉ | 12487/18018 [27:00:02<11:55:06, 7.76s/step, loss=1.219, acc=0.717, acc_len=1.74, thru=17.2, I=22.6, T=37.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:16,741] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:30:23,213] loop.py:45 INFO TRAIN_STEP step=12488 loss=1.349421 Training: 69%|██████▉ | 12488/18018 [27:00:09<11:25:38, 7.44s/step, loss=1.349, acc=0.689, acc_len=1.64, thru=18.9, I=23.1, T=34.3, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:30,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:30:30,431] loop.py:45 INFO TRAIN_STEP step=12489 loss=1.322665 Training: 69%|██████▉ | 12489/18018 [27:00:16<11:19:21, 7.37s/step, loss=1.323, acc=0.694, acc_len=1.66, thru=16.8, I=18.0, T=35.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:30,578] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:35,968] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:30:36,438] loop.py:380 INFO TIMING step=12490: step=3.369s data=1.402s compute=3.334s [fwd=2.370s bwd=0.950s opt=0.003s] dispatch=2.628s [2026-10-19 16:30:36,439] loop.py:45 INFO TRAIN_STEP step=12490 loss=1.223849 Training: 69%|██████▉ | 12490/18018 [27:00:22<10:41:31, 6.96s/step, loss=1.224, acc=0.712, acc_len=1.72, thru=20.3, I=21.3, T=38.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:36,475] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:41,209] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:30:42,726] loop.py:45 INFO TRAIN_STEP step=12491 loss=1.347255 Training: 69%|██████▉ | 12491/18018 [27:00:28<10:22:39, 6.76s/step, loss=1.347, acc=0.690, acc_len=1.65, thru=20.9, I=24.4, T=35.8, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:30:42,721] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12490: forward=2525.5ms backward=1009.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:30:43,224] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:30:48,685] loop.py:45 INFO TRAIN_STEP step=12492 loss=1.296732 Training: 69%|██████▉ | 12492/18018 [27:00:34<10:00:31, 6.52s/step, loss=1.297, acc=0.697, acc_len=1.67, thru=19.0, I=27.4, T=38.8, wait=2.7s, pool=53, epoch=2/2][2026-10-19 16:30:56,967] loop.py:45 INFO TRAIN_STEP step=12493 loss=1.242990 Training: 69%|██████▉ | 12493/18018 [27:00:43<10:49:03, 7.05s/step, loss=1.243, acc=0.711, acc_len=1.72, thru=12.6, I=15.2, T=35.4, wait=4.7s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:31:04,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:31:06,687] loop.py:45 INFO TRAIN_STEP step=12494 loss=1.326501 Training: 69%|██████▉ | 12494/18018 [27:00:52<12:02:44, 7.85s/step, loss=1.327, acc=0.695, acc_len=1.67, thru=16.9, I=16.1, T=34.2, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:31:06,964] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:31:13,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:31:14,265] loop.py:380 INFO TIMING step=12495: step=3.590s data=1.479s compute=3.559s [fwd=2.560s bwd=0.984s opt=0.003s] dispatch=3.977s [2026-10-19 16:31:14,265] loop.py:45 INFO TRAIN_STEP step=12495 loss=1.262231 Training: 69%|██████▉ | 12495/18018 [27:01:00<11:55:01, 7.77s/step, loss=1.262, acc=0.709, acc_len=1.71, thru=15.8, I=21.4, T=35.7, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:31:14,689] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:31:21,327] loop.py:45 INFO TRAIN_STEP step=12496 loss=1.252899 Training: 69%|██████▉ | 12496/18018 [27:01:07<11:35:25, 7.56s/step, loss=1.253, acc=0.707, acc_len=1.71, thru=17.4, I=19.2, T=35.2, wait=3.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:31:21,322] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12495: forward=2551.5ms backward=1038.8ms [2026-10-19 16:31:29,217] loop.py:45 INFO TRAIN_STEP step=12497 loss=1.161905 Training: 69%|██████▉ | 12497/18018 [27:01:15<11:44:29, 7.66s/step, loss=1.162, acc=0.721, acc_len=1.76, thru=14.1, I=15.2, T=35.9, wait=4.3s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:31:42,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:31:43,495] loop.py:45 INFO TRAIN_STEP step=12498 loss=1.463821 Training: 69%|██████▉ | 12498/18018 [27:01:29<14:47:09, 9.64s/step, loss=1.464, acc=0.672, acc_len=1.59, thru=11.9, I=11.6, T=28.0, wait=9.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:31:43,920] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:31:48,581] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:31:49,509] loop.py:45 INFO TRAIN_STEP step=12499 loss=1.228850 Training: 69%|██████▉ | 12499/18018 [27:01:35<13:06:49, 8.55s/step, loss=1.229, acc=0.715, acc_len=1.73, thru=18.1, I=23.2, T=36.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:31:49,598] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:31:55,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:31:55,381] loop.py:380 INFO TIMING step=12500: step=3.479s data=1.521s compute=3.455s [fwd=2.492s bwd=0.949s opt=0.003s] dispatch=2.384s [2026-10-19 16:31:55,381] loop.py:45 INFO TRAIN_STEP step=12500 loss=1.322814 Training: 69%|██████▉ | 12500/18018 [27:01:41<11:52:41, 7.75s/step, loss=1.323, acc=0.695, acc_len=1.66, thru=22.0, I=25.7, T=36.8, wait=2.4s, pool=65, epoch=2/2][2026-10-19 16:31:55,384] loop.py:419 INFO Saving checkpoint at step 12500... (TrainerActor pid=9310) [2026-10-19 16:31:58,798] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0012501 (TrainerActor pid=9310) [2026-10-19 16:31:58,807] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0011751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:32:00,221] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:32:00,222] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=1/64 [2026-10-19 16:32:06,499] loop.py:45 INFO TRAIN_STEP step=12501 loss=1.164587 Training: 69%|██████▉ | 12501/18018 [27:01:52<13:25:28, 8.76s/step, loss=1.165, acc=0.725, acc_len=1.77, thru=18.5, I=17.7, T=32.2, wait=2.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:32:06,494] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12500: forward=2779.3ms backward=1138.6ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 16:31:58,799] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0012501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 16:32:00,019] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0011751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:32:12,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:32:13,854] loop.py:45 INFO TRAIN_STEP step=12502 loss=1.207908 Training: 69%|██████▉ | 12502/18018 [27:01:59<12:46:35, 8.34s/step, loss=1.208, acc=0.717, acc_len=1.74, thru=19.1, I=18.2, T=36.6, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:32:14,081] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:32:20,696] loop.py:45 INFO TRAIN_STEP step=12503 loss=1.222280 Training: 69%|██████▉ | 12503/18018 [27:02:06<12:05:10, 7.89s/step, loss=1.222, acc=0.712, acc_len=1.72, thru=17.4, I=21.9, T=35.9, wait=3.3s, pool=57, epoch=2/2][2026-10-19 16:32:28,775] loop.py:45 INFO TRAIN_STEP step=12504 loss=1.223372 Training: 69%|██████▉ | 12504/18018 [27:02:14<12:10:19, 7.95s/step, loss=1.223, acc=0.714, acc_len=1.73, thru=15.7, I=16.1, T=36.7, wait=4.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:32:39,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:32:40,212] loop.py:380 INFO TIMING step=12505: step=3.973s data=1.667s compute=3.927s [fwd=2.752s bwd=1.161s opt=0.003s] dispatch=7.454s [2026-10-19 16:32:40,212] loop.py:45 INFO TRAIN_STEP step=12505 loss=1.313597 Training: 69%|██████▉ | 12505/18018 [27:02:26<13:46:24, 8.99s/step, loss=1.314, acc=0.698, acc_len=1.67, thru=13.5, I=12.1, T=32.2, wait=7.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:32:40,574] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:32:47,075] loop.py:45 INFO TRAIN_STEP step=12506 loss=1.380397 Training: 69%|██████▉ | 12506/18018 [27:02:33<12:47:28, 8.35s/step, loss=1.380, acc=0.683, acc_len=1.62, thru=17.5, I=16.2, T=37.2, wait=3.4s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:32:47,071] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12505: forward=2431.1ms backward=935.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:32:52,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:32:54,491] loop.py:45 INFO TRAIN_STEP step=12507 loss=1.321404 Training: 69%|██████▉ | 12507/18018 [27:02:40<12:21:29, 8.07s/step, loss=1.321, acc=0.695, acc_len=1.67, thru=19.5, I=22.0, T=34.1, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:32:54,733] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:32:58,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:33:00,429] loop.py:45 INFO TRAIN_STEP step=12508 loss=1.232309 Training: 69%|██████▉ | 12508/18018 [27:02:46<11:22:31, 7.43s/step, loss=1.232, acc=0.712, acc_len=1.72, thru=20.4, I=31.6, T=35.5, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:00,786] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:06,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:33:07,686] loop.py:45 INFO TRAIN_STEP step=12509 loss=1.244483 Training: 69%|██████▉ | 12509/18018 [27:02:53<11:17:36, 7.38s/step, loss=1.244, acc=0.711, acc_len=1.72, thru=16.4, I=23.2, T=36.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:08,045] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:33:13,480] loop.py:380 INFO TIMING step=12510: step=3.416s data=1.364s compute=3.377s [fwd=2.378s bwd=0.984s opt=0.003s] dispatch=2.368s [2026-10-19 16:33:13,480] loop.py:45 INFO TRAIN_STEP step=12510 loss=1.362451 Training: 69%|██████▉ | 12510/18018 [27:02:59<10:33:50, 6.90s/step, loss=1.362, acc=0.690, acc_len=1.65, thru=19.8, I=20.7, T=37.5, wait=2.4s, pool=49, epoch=2/2][2026-10-19 16:33:21,737] loop.py:45 INFO TRAIN_STEP step=12511 loss=1.330902 Training: 69%|██████▉ | 12511/18018 [27:03:07<11:10:55, 7.31s/step, loss=1.331, acc=0.693, acc_len=1.66, thru=16.7, I=15.5, T=35.0, wait=4.6s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:33:21,733] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12510: forward=2562.4ms backward=1046.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:28,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:33:29,648] loop.py:45 INFO TRAIN_STEP step=12512 loss=1.296432 Training: 69%|██████▉ | 12512/18018 [27:03:15<11:27:22, 7.49s/step, loss=1.296, acc=0.701, acc_len=1.68, thru=17.0, I=18.1, T=34.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:29,901] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:34,639] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:33:35,167] loop.py:45 INFO TRAIN_STEP step=12513 loss=1.279853 Training: 69%|██████▉ | 12513/18018 [27:03:21<10:32:55, 6.90s/step, loss=1.280, acc=0.703, acc_len=1.70, thru=20.8, I=25.8, T=38.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:35,643] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:33:42,711] loop.py:45 INFO TRAIN_STEP step=12514 loss=1.196941 Training: 69%|██████▉ | 12514/18018 [27:03:28<10:50:35, 7.09s/step, loss=1.197, acc=0.718, acc_len=1.75, thru=17.0, I=18.4, T=33.1, wait=3.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:48,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:33:49,982] loop.py:380 INFO TIMING step=12515: step=3.597s data=1.468s compute=3.567s [fwd=2.512s bwd=1.041s opt=0.003s] dispatch=3.666s [2026-10-19 16:33:49,982] loop.py:45 INFO TRAIN_STEP step=12515 loss=1.297617 Training: 69%|██████▉ | 12515/18018 [27:03:36<10:55:25, 7.15s/step, loss=1.298, acc=0.702, acc_len=1.68, thru=18.9, I=18.8, T=35.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:50,400] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:54,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:33:56,020] loop.py:45 INFO TRAIN_STEP step=12516 loss=1.348163 Training: 69%|██████▉ | 12516/18018 [27:03:42<10:24:48, 6.81s/step, loss=1.348, acc=0.687, acc_len=1.63, thru=20.3, I=28.6, T=38.2, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:33:56,014] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12515: forward=2373.4ms backward=933.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:33:56,429] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:34:05,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:34:05,769] loop.py:45 INFO TRAIN_STEP step=12517 loss=1.329117 Training: 69%|██████▉ | 12517/18018 [27:03:51<11:45:27, 7.69s/step, loss=1.329, acc=0.691, acc_len=1.64, thru=13.7, I=14.3, T=31.3, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:34:06,071] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:34:15,748] loop.py:45 INFO TRAIN_STEP step=12518 loss=1.399993 Training: 69%|██████▉ | 12518/18018 [27:04:01<12:48:11, 8.38s/step, loss=1.400, acc=0.684, acc_len=1.63, thru=12.8, I=11.9, T=31.2, wait=5.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:34:20,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:34:21,519] loop.py:45 INFO TRAIN_STEP step=12519 loss=1.250110 Training: 69%|██████▉ | 12519/18018 [27:04:07<11:36:17, 7.60s/step, loss=1.250, acc=0.709, acc_len=1.71, thru=20.2, I=21.6, T=37.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:34:21,554] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 16:34:27,667] loop.py:380 INFO TIMING step=12520: step=3.407s data=1.387s compute=3.367s [fwd=2.416s bwd=0.935s opt=0.003s] dispatch=2.730s [2026-10-19 16:34:27,667] loop.py:45 INFO TRAIN_STEP step=12520 loss=1.315390 Training: 69%|██████▉ | 12520/18018 [27:04:13<10:56:13, 7.16s/step, loss=1.315, acc=0.695, acc_len=1.67, thru=19.0, I=20.7, T=37.6, wait=2.7s, pool=49, epoch=2/2][2026-10-19 16:34:34,449] loop.py:45 INFO TRAIN_STEP step=12521 loss=1.423706 Training: 69%|██████▉ | 12521/18018 [27:04:20<10:45:42, 7.05s/step, loss=1.424, acc=0.681, acc_len=1.62, thru=19.1, I=19.0, T=34.6, wait=3.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:34:34,444] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12520: forward=2548.7ms backward=1070.2ms [2026-10-19 16:34:43,761] loop.py:45 INFO TRAIN_STEP step=12522 loss=1.272844 Training: 69%|██████▉ | 12522/18018 [27:04:29<11:47:47, 7.73s/step, loss=1.273, acc=0.706, acc_len=1.71, thru=13.8, I=14.2, T=26.9, wait=4.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:34:50,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:34:51,151] loop.py:45 INFO TRAIN_STEP step=12523 loss=1.228737 Training: 70%|██████▉ | 12523/18018 [27:04:37<11:38:25, 7.63s/step, loss=1.229, acc=0.718, acc_len=1.74, thru=16.6, I=17.3, T=32.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:34:51,291] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:34:57,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:34:58,075] loop.py:45 INFO TRAIN_STEP step=12524 loss=1.313441 Training: 70%|██████▉ | 12524/18018 [27:04:44<11:18:59, 7.42s/step, loss=1.313, acc=0.700, acc_len=1.68, thru=17.5, I=19.8, T=35.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:34:58,557] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:03,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:35:04,915] loop.py:380 INFO TIMING step=12525: step=3.641s data=1.514s compute=3.608s [fwd=2.510s bwd=1.083s opt=0.003s] dispatch=3.191s [2026-10-19 16:35:04,915] loop.py:45 INFO TRAIN_STEP step=12525 loss=1.242373 Training: 70%|██████▉ | 12525/18018 [27:04:50<11:03:04, 7.24s/step, loss=1.242, acc=0.713, acc_len=1.73, thru=18.5, I=25.0, T=35.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:05,264] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:35:10,951] loop.py:45 INFO TRAIN_STEP step=12526 loss=1.191208 (TrainerActor pid=9310) [2026-10-19 16:35:10,947] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12525: forward=2435.2ms backward=942.1ms Training: 70%|██████▉ | 12526/18018 [27:04:57<10:29:48, 6.88s/step, loss=1.191, acc=0.717, acc_len=1.73, thru=19.9, I=21.8, T=37.5, wait=2.6s, pool=61, epoch=2/2][2026-10-19 16:35:16,657] loop.py:45 INFO TRAIN_STEP step=12527 loss=1.318961 Training: 70%|██████▉ | 12527/18018 [27:05:02<9:57:32, 6.53s/step, loss=1.319, acc=0.694, acc_len=1.66, thru=20.1, I=22.4, T=36.1, wait=2.2s, pool=53, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:24,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:35:24,868] loop.py:45 INFO TRAIN_STEP step=12528 loss=1.265363 Training: 70%|██████▉ | 12528/18018 [27:05:10<10:43:29, 7.03s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=16.2, I=15.4, T=32.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:25,264] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:35:31,911] loop.py:45 INFO TRAIN_STEP step=12529 loss=1.314693 Training: 70%|██████▉ | 12529/18018 [27:05:17<10:43:39, 7.04s/step, loss=1.315, acc=0.700, acc_len=1.68, thru=17.2, I=18.1, T=30.7, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:39,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:35:40,146] loop.py:380 INFO TIMING step=12530: step=4.291s data=2.015s compute=4.248s [fwd=3.013s bwd=1.220s opt=0.003s] dispatch=3.936s [2026-10-19 16:35:40,146] loop.py:45 INFO TRAIN_STEP step=12530 loss=1.304795 Training: 70%|██████▉ | 12530/18018 [27:05:26<11:16:27, 7.40s/step, loss=1.305, acc=0.702, acc_len=1.69, thru=17.1, I=16.7, T=29.8, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:40,592] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:45,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:35:46,616] loop.py:45 INFO TRAIN_STEP step=12531 loss=1.297239 Training: 70%|██████▉ | 12531/18018 [27:05:32<10:50:59, 7.12s/step, loss=1.297, acc=0.698, acc_len=1.68, thru=17.6, I=23.4, T=35.8, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:35:46,611] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12530: forward=2361.0ms backward=1162.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:47,032] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:53,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:35:54,280] loop.py:45 INFO TRAIN_STEP step=12532 loss=1.247464 Training: 70%|██████▉ | 12532/18018 [27:05:40<11:05:45, 7.28s/step, loss=1.247, acc=0.708, acc_len=1.70, thru=16.3, I=22.5, T=35.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:35:54,626] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:36:00,240] loop.py:45 INFO TRAIN_STEP step=12533 loss=1.177279 Training: 70%|██████▉ | 12533/18018 [27:05:46<10:29:29, 6.89s/step, loss=1.177, acc=0.723, acc_len=1.75, thru=20.0, I=22.3, T=35.7, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:05,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:36:06,413] loop.py:45 INFO TRAIN_STEP step=12534 loss=1.268163 Training: 70%|██████▉ | 12534/18018 [27:05:52<10:09:44, 6.67s/step, loss=1.268, acc=0.705, acc_len=1.71, thru=20.3, I=22.5, T=34.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:06,842] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:13,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:36:14,410] loop.py:380 INFO TIMING step=12535: step=3.848s data=1.670s compute=3.822s [fwd=2.738s bwd=1.070s opt=0.003s] dispatch=4.142s [2026-10-19 16:36:14,411] loop.py:45 INFO TRAIN_STEP step=12535 loss=1.250408 Training: 70%|██████▉ | 12535/18018 [27:06:00<10:46:00, 7.07s/step, loss=1.250, acc=0.710, acc_len=1.72, thru=15.9, I=20.2, T=33.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:14,455] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 16:36:21,625] loop.py:45 INFO TRAIN_STEP step=12536 loss=1.304601 Training: 70%|██████▉ | 12536/18018 [27:06:07<10:49:56, 7.11s/step, loss=1.305, acc=0.701, acc_len=1.69, thru=17.6, I=17.9, T=29.8, wait=2.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:36:21,620] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12535: forward=3196.2ms backward=1049.1ms [2026-10-19 16:36:29,599] loop.py:45 INFO TRAIN_STEP step=12537 loss=1.294085 Training: 70%|██████▉ | 12537/18018 [27:06:15<11:13:18, 7.37s/step, loss=1.294, acc=0.702, acc_len=1.70, thru=15.5, I=14.1, T=34.0, wait=4.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:37,181] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:36:38,631] loop.py:45 INFO TRAIN_STEP step=12538 loss=1.321068 Training: 70%|██████▉ | 12538/18018 [27:06:24<11:58:47, 7.87s/step, loss=1.321, acc=0.700, acc_len=1.68, thru=16.0, I=16.8, T=30.7, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:38,705] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:46,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:36:47,656] loop.py:45 INFO TRAIN_STEP step=12539 loss=1.251076 Training: 70%|██████▉ | 12539/18018 [27:06:33<12:30:18, 8.22s/step, loss=1.251, acc=0.712, acc_len=1.72, thru=14.3, I=14.6, T=35.3, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:47,810] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:53,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:36:54,341] loop.py:380 INFO TIMING step=12540: step=3.989s data=1.726s compute=3.946s [fwd=2.695s bwd=1.237s opt=0.003s] dispatch=2.686s [2026-10-19 16:36:54,342] loop.py:45 INFO TRAIN_STEP step=12540 loss=1.395690 Training: 70%|██████▉ | 12540/18018 [27:06:40<11:48:11, 7.76s/step, loss=1.396, acc=0.685, acc_len=1.63, thru=19.2, I=21.4, T=32.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:36:54,373] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 16:37:01,482] loop.py:45 INFO TRAIN_STEP step=12541 loss=1.297298 Training: 70%|██████▉ | 12541/18018 [27:06:47<11:31:10, 7.57s/step, loss=1.297, acc=0.703, acc_len=1.69, thru=15.7, I=16.5, T=37.5, wait=3.7s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:37:01,478] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12540: forward=2420.4ms backward=947.3ms [2026-10-19 16:37:08,335] loop.py:45 INFO TRAIN_STEP step=12542 loss=1.241527 Training: 70%|██████▉ | 12542/18018 [27:06:54<11:11:21, 7.36s/step, loss=1.242, acc=0.707, acc_len=1.70, thru=18.2, I=19.3, T=34.5, wait=3.1s, pool=57, epoch=2/2][2026-10-19 16:37:17,564] loop.py:45 INFO TRAIN_STEP step=12543 loss=1.365402 Training: 70%|██████▉ | 12543/18018 [27:07:03<12:02:34, 7.92s/step, loss=1.365, acc=0.693, acc_len=1.65, thru=14.0, I=13.6, T=30.1, wait=5.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:23,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:37:26,175] loop.py:45 INFO TRAIN_STEP step=12544 loss=1.209470 Training: 70%|██████▉ | 12544/18018 [27:07:12<12:21:19, 8.13s/step, loss=1.209, acc=0.715, acc_len=1.74, thru=15.5, I=18.7, T=24.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:26,275] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:33,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:37:34,817] loop.py:380 INFO TIMING step=12545: step=3.957s data=1.793s compute=3.928s [fwd=2.778s bwd=1.136s opt=0.003s] dispatch=4.678s [2026-10-19 16:37:34,817] loop.py:45 INFO TRAIN_STEP step=12545 loss=1.239575 Training: 70%|██████▉ | 12545/18018 [27:07:20<12:35:22, 8.28s/step, loss=1.240, acc=0.711, acc_len=1.72, thru=15.0, I=16.9, T=32.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:35,094] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:40,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:37:41,323] loop.py:45 INFO TRAIN_STEP step=12546 loss=1.319500 Training: 70%|██████▉ | 12546/18018 [27:07:27<11:46:37, 7.75s/step, loss=1.320, acc=0.695, acc_len=1.66, thru=17.4, I=22.5, T=34.1, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:37:41,320] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12545: forward=2602.2ms backward=1100.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:41,576] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:46,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:37:47,767] loop.py:45 INFO TRAIN_STEP step=12547 loss=1.300010 Training: 70%|██████▉ | 12547/18018 [27:07:33<11:10:52, 7.36s/step, loss=1.300, acc=0.699, acc_len=1.67, thru=19.0, I=25.3, T=32.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:48,115] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:52,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:37:53,874] loop.py:45 INFO TRAIN_STEP step=12548 loss=1.266548 Training: 70%|██████▉ | 12548/18018 [27:07:39<10:36:31, 6.98s/step, loss=1.267, acc=0.705, acc_len=1.70, thru=19.5, I=25.6, T=35.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:37:54,269] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:38:00,085] loop.py:45 INFO TRAIN_STEP step=12549 loss=1.171447 Training: 70%|██████▉ | 12549/18018 [27:07:46<10:15:17, 6.75s/step, loss=1.171, acc=0.722, acc_len=1.76, thru=19.3, I=22.4, T=37.8, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:06,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:38:08,062] loop.py:380 INFO TIMING step=12550: step=3.658s data=1.629s compute=3.615s [fwd=2.580s bwd=1.021s opt=0.003s] dispatch=4.311s [2026-10-19 16:38:08,063] loop.py:45 INFO TRAIN_STEP step=12550 loss=1.181472 Training: 70%|██████▉ | 12550/18018 [27:07:54<10:48:45, 7.12s/step, loss=1.181, acc=0.725, acc_len=1.77, thru=17.1, I=19.2, T=35.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:08,165] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:12,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:38:13,562] loop.py:45 INFO TRAIN_STEP step=12551 loss=1.290701 Training: 70%|██████▉ | 12551/18018 [27:07:59<10:04:23, 6.63s/step, loss=1.291, acc=0.700, acc_len=1.68, thru=22.0, I=27.0, T=38.7, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:38:13,558] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12550: forward=2321.0ms backward=937.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:13,968] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:18,722] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:38:19,494] loop.py:45 INFO TRAIN_STEP step=12552 loss=1.253800 Training: 70%|██████▉ | 12552/18018 [27:08:05<9:45:09, 6.42s/step, loss=1.254, acc=0.712, acc_len=1.72, thru=20.9, I=25.2, T=38.0, wait=2.6s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:19,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:38:28,978] loop.py:45 INFO TRAIN_STEP step=12553 loss=1.332925 Training: 70%|██████▉ | 12553/18018 [27:08:15<11:08:37, 7.34s/step, loss=1.333, acc=0.693, acc_len=1.66, thru=13.0, I=13.2, T=32.4, wait=5.5s, pool=57, epoch=2/2][2026-10-19 16:38:36,126] loop.py:45 INFO TRAIN_STEP step=12554 loss=1.229465 Training: 70%|██████▉ | 12554/18018 [27:08:22<11:03:16, 7.28s/step, loss=1.229, acc=0.715, acc_len=1.73, thru=17.8, I=16.1, T=37.3, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:44,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:38:46,336] loop.py:380 INFO TIMING step=12555: step=3.910s data=1.655s compute=3.884s [fwd=2.791s bwd=1.078s opt=0.003s] dispatch=6.290s [2026-10-19 16:38:46,336] loop.py:45 INFO TRAIN_STEP step=12555 loss=1.213305 Training: 70%|██████▉ | 12555/18018 [27:08:32<12:23:02, 8.16s/step, loss=1.213, acc=0.717, acc_len=1.74, thru=13.7, I=15.5, T=32.7, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:46,823] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:52,403] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:38:52,705] loop.py:45 INFO TRAIN_STEP step=12556 loss=1.327944 Training: 70%|██████▉ | 12556/18018 [27:08:38<11:33:58, 7.62s/step, loss=1.328, acc=0.698, acc_len=1.67, thru=18.5, I=23.2, T=38.8, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:38:52,700] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12555: forward=2331.4ms backward=924.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:38:52,905] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:38:59,753] loop.py:45 INFO TRAIN_STEP step=12557 loss=1.283584 Training: 70%|██████▉ | 12557/18018 [27:08:45<11:18:11, 7.45s/step, loss=1.284, acc=0.701, acc_len=1.69, thru=18.9, I=20.0, T=37.8, wait=3.7s, pool=61, epoch=2/2][2026-10-19 16:39:07,319] loop.py:45 INFO TRAIN_STEP step=12558 loss=1.316270 Training: 70%|██████▉ | 12558/18018 [27:08:53<11:21:07, 7.48s/step, loss=1.316, acc=0.697, acc_len=1.67, thru=16.0, I=17.1, T=36.4, wait=4.0s, pool=57, epoch=2/2][2026-10-19 16:39:15,260] loop.py:45 INFO TRAIN_STEP step=12559 loss=1.228557 Training: 70%|██████▉ | 12559/18018 [27:09:01<11:33:28, 7.62s/step, loss=1.229, acc=0.710, acc_len=1.71, thru=13.5, I=13.6, T=34.3, wait=4.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:39:22,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:39:23,391] loop.py:380 INFO TIMING step=12560: step=3.813s data=1.590s compute=3.776s [fwd=2.778s bwd=0.983s opt=0.003s] dispatch=4.309s [2026-10-19 16:39:23,392] loop.py:45 INFO TRAIN_STEP step=12560 loss=1.378482 Training: 70%|██████▉ | 12560/18018 [27:09:09<11:47:18, 7.78s/step, loss=1.378, acc=0.688, acc_len=1.64, thru=18.4, I=18.0, T=33.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:39:23,688] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:39:32,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:39:34,174] loop.py:45 INFO TRAIN_STEP step=12561 loss=1.280922 Training: 70%|██████▉ | 12561/18018 [27:09:20<13:09:11, 8.68s/step, loss=1.281, acc=0.705, acc_len=1.70, thru=12.4, I=13.2, T=30.5, wait=6.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:39:34,169] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12560: forward=2942.8ms backward=1205.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:39:34,347] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:39:39,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:39:39,977] loop.py:45 INFO TRAIN_STEP step=12562 loss=1.193372 Training: 70%|██████▉ | 12562/18018 [27:09:26<11:50:37, 7.81s/step, loss=1.193, acc=0.721, acc_len=1.75, thru=19.2, I=24.7, T=35.0, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:39:40,184] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:39:45,973] loop.py:45 INFO TRAIN_STEP step=12563 loss=1.191156 Training: 70%|██████▉ | 12563/18018 [27:09:32<11:00:51, 7.27s/step, loss=1.191, acc=0.719, acc_len=1.74, thru=19.6, I=23.4, T=36.8, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:39:54,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:39:55,409] loop.py:45 INFO TRAIN_STEP step=12564 loss=1.194473 Training: 70%|██████▉ | 12564/18018 [27:09:41<11:59:52, 7.92s/step, loss=1.194, acc=0.715, acc_len=1.73, thru=14.9, I=14.9, T=33.4, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:39:55,480] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:00,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:40:01,063] loop.py:380 INFO TIMING step=12565: step=3.598s data=1.506s compute=3.565s [fwd=2.525s bwd=1.025s opt=0.003s] dispatch=2.046s [2026-10-19 16:40:01,063] loop.py:45 INFO TRAIN_STEP step=12565 loss=1.365137 Training: 70%|██████▉ | 12565/18018 [27:09:47<10:58:00, 7.24s/step, loss=1.365, acc=0.690, acc_len=1.64, thru=20.6, I=22.2, T=35.6, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:01,213] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:08,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:40:08,610] loop.py:45 INFO TRAIN_STEP step=12566 loss=1.284760 Training: 70%|██████▉ | 12566/18018 [27:09:54<11:06:10, 7.33s/step, loss=1.285, acc=0.702, acc_len=1.70, thru=16.5, I=19.4, T=33.2, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:40:08,605] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12565: forward=2543.0ms backward=1090.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:09,060] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:13,755] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:40:14,824] loop.py:45 INFO TRAIN_STEP step=12567 loss=1.265274 Training: 70%|██████▉ | 12567/18018 [27:10:00<10:35:36, 7.00s/step, loss=1.265, acc=0.708, acc_len=1.71, thru=20.0, I=21.0, T=38.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:15,267] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:19,621] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:40:20,739] loop.py:45 INFO TRAIN_STEP step=12568 loss=1.255448 Training: 70%|██████▉ | 12568/18018 [27:10:06<10:06:03, 6.67s/step, loss=1.255, acc=0.710, acc_len=1.71, thru=21.3, I=27.2, T=38.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:21,127] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:27,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:40:28,683] loop.py:45 INFO TRAIN_STEP step=12569 loss=1.218449 Training: 70%|██████▉ | 12569/18018 [27:10:14<10:40:37, 7.05s/step, loss=1.218, acc=0.713, acc_len=1.72, thru=15.1, I=19.2, T=35.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:28,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:33,457] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:40:34,316] loop.py:380 INFO TIMING step=12570: step=3.438s data=1.445s compute=3.407s [fwd=2.460s bwd=0.932s opt=0.003s] dispatch=2.185s [2026-10-19 16:40:34,317] loop.py:45 INFO TRAIN_STEP step=12570 loss=1.329746 Training: 70%|██████▉ | 12570/18018 [27:10:20<10:01:46, 6.63s/step, loss=1.330, acc=0.694, acc_len=1.66, thru=20.6, I=24.2, T=37.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:34,463] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:40:39,758] loop.py:45 INFO TRAIN_STEP step=12571 loss=1.252670 Training: 70%|██████▉ | 12571/18018 [27:10:25<9:29:23, 6.27s/step, loss=1.253, acc=0.706, acc_len=1.70, thru=21.2, I=27.2, T=38.0, wait=2.1s, pool=41, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 16:40:39,754] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12570: forward=2412.8ms backward=917.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:47,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:40:48,975] loop.py:45 INFO TRAIN_STEP step=12572 loss=1.394631 Training: 70%|██████▉ | 12572/18018 [27:10:35<10:49:29, 7.16s/step, loss=1.395, acc=0.684, acc_len=1.63, thru=16.1, I=18.1, T=31.1, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:49,101] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:54,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:40:55,840] loop.py:45 INFO TRAIN_STEP step=12573 loss=1.343839 Training: 70%|██████▉ | 12573/18018 [27:10:41<10:41:24, 7.07s/step, loss=1.344, acc=0.694, acc_len=1.66, thru=17.3, I=23.1, T=35.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:40:55,969] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 16:41:02,138] loop.py:45 INFO TRAIN_STEP step=12574 loss=1.246692 Training: 70%|██████▉ | 12574/18018 [27:10:48<10:20:20, 6.84s/step, loss=1.247, acc=0.708, acc_len=1.71, thru=18.5, I=19.0, T=37.6, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:08,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:41:09,795] loop.py:380 INFO TIMING step=12575: step=3.524s data=1.426s compute=3.495s [fwd=2.524s bwd=0.957s opt=0.003s] dispatch=4.126s [2026-10-19 16:41:09,796] loop.py:45 INFO TRAIN_STEP step=12575 loss=1.292578 Training: 70%|██████▉ | 12575/18018 [27:10:55<10:42:36, 7.08s/step, loss=1.293, acc=0.703, acc_len=1.69, thru=18.4, I=20.3, T=36.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:10,064] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:41:15,328] loop.py:45 INFO TRAIN_STEP step=12576 loss=1.207294 Training: 70%|██████▉ | 12576/18018 [27:11:01<10:00:15, 6.62s/step, loss=1.207, acc=0.717, acc_len=1.73, thru=18.3, I=22.8, T=39.5, wait=2.3s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:41:15,324] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12575: forward=2304.2ms backward=896.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:23,492] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:41:24,667] loop.py:45 INFO TRAIN_STEP step=12577 loss=1.310566 Training: 70%|██████▉ | 12577/18018 [27:11:10<11:14:11, 7.43s/step, loss=1.311, acc=0.698, acc_len=1.67, thru=17.4, I=17.0, T=33.7, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:25,003] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:30,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:41:31,695] loop.py:45 INFO TRAIN_STEP step=12578 loss=1.329905 Training: 70%|██████▉ | 12578/18018 [27:11:17<11:03:01, 7.31s/step, loss=1.330, acc=0.693, acc_len=1.65, thru=17.8, I=23.8, T=33.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:32,054] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:37,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:41:37,970] loop.py:45 INFO TRAIN_STEP step=12579 loss=1.263971 Training: 70%|██████▉ | 12579/18018 [27:11:24<10:34:36, 7.00s/step, loss=1.264, acc=0.706, acc_len=1.70, thru=19.2, I=22.7, T=35.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:38,383] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:41:45,506] loop.py:380 INFO TIMING step=12580: step=3.437s data=1.434s compute=3.401s [fwd=2.406s bwd=0.981s opt=0.003s] dispatch=4.091s [2026-10-19 16:41:45,506] loop.py:45 INFO TRAIN_STEP step=12580 loss=1.296396 Training: 70%|██████▉ | 12580/18018 [27:11:31<10:49:06, 7.16s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=14.0, I=15.8, T=37.2, wait=4.1s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:55,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:41:56,324] loop.py:45 INFO TRAIN_STEP step=12581 loss=1.189573 Training: 70%|██████▉ | 12581/18018 [27:11:42<12:28:20, 8.26s/step, loss=1.190, acc=0.717, acc_len=1.74, thru=15.4, I=13.8, T=33.2, wait=7.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:41:56,320] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12580: forward=2444.5ms backward=1135.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:41:56,786] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:01,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:42:02,352] loop.py:45 INFO TRAIN_STEP step=12582 loss=1.245030 Training: 70%|██████▉ | 12582/18018 [27:11:48<11:27:35, 7.59s/step, loss=1.245, acc=0.705, acc_len=1.70, thru=20.1, I=23.8, T=37.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:02,561] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:08,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:42:08,263] loop.py:45 INFO TRAIN_STEP step=12583 loss=1.136508 Training: 70%|██████▉ | 12583/18018 [27:11:54<10:41:51, 7.09s/step, loss=1.137, acc=0.732, acc_len=1.79, thru=21.4, I=22.8, T=36.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:08,505] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:42:16,870] loop.py:45 INFO TRAIN_STEP step=12584 loss=1.241565 Training: 70%|██████▉ | 12584/18018 [27:12:02<11:23:07, 7.54s/step, loss=1.242, acc=0.714, acc_len=1.73, thru=12.6, I=16.8, T=34.4, wait=4.9s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:24,141] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:42:25,703] loop.py:380 INFO TIMING step=12585: step=4.011s data=1.766s compute=3.979s [fwd=2.839s bwd=1.125s opt=0.003s] dispatch=4.812s [2026-10-19 16:42:25,703] loop.py:45 INFO TRAIN_STEP step=12585 loss=1.337287 Training: 70%|██████▉ | 12585/18018 [27:12:11<11:58:01, 7.93s/step, loss=1.337, acc=0.692, acc_len=1.65, thru=17.0, I=17.1, T=31.9, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:26,154] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:42:32,950] loop.py:45 INFO TRAIN_STEP step=12586 loss=1.381610 Training: 70%|██████▉ | 12586/18018 [27:12:19<11:39:19, 7.72s/step, loss=1.382, acc=0.687, acc_len=1.64, thru=14.8, I=20.1, T=37.5, wait=3.8s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:42:32,945] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12585: forward=2407.0ms backward=952.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:40,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:42:41,563] loop.py:45 INFO TRAIN_STEP step=12587 loss=1.237213 Training: 70%|██████▉ | 12587/18018 [27:12:27<12:03:18, 7.99s/step, loss=1.237, acc=0.710, acc_len=1.71, thru=16.6, I=16.9, T=32.8, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:41,685] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 16:42:48,060] loop.py:45 INFO TRAIN_STEP step=12588 loss=1.195634 Training: 70%|██████▉ | 12588/18018 [27:12:34<11:22:36, 7.54s/step, loss=1.196, acc=0.718, acc_len=1.74, thru=18.1, I=21.8, T=38.4, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:53,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:42:54,576] loop.py:45 INFO TRAIN_STEP step=12589 loss=1.257632 Training: 70%|██████▉ | 12589/18018 [27:12:40<10:54:39, 7.24s/step, loss=1.258, acc=0.704, acc_len=1.70, thru=19.8, I=22.4, T=36.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:54,762] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:42:59,930] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:43:00,407] loop.py:380 INFO TIMING step=12590: step=3.311s data=1.367s compute=3.286s [fwd=2.338s bwd=0.933s opt=0.003s] dispatch=2.509s [2026-10-19 16:43:00,407] loop.py:45 INFO TRAIN_STEP step=12590 loss=1.175879 Training: 70%|██████▉ | 12590/18018 [27:12:46<10:16:22, 6.81s/step, loss=1.176, acc=0.723, acc_len=1.75, thru=21.4, I=26.3, T=38.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:00,436] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:06,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:43:07,937] loop.py:45 INFO TRAIN_STEP step=12591 loss=1.257074 Training: 70%|██████▉ | 12591/18018 [27:12:54<10:35:45, 7.03s/step, loss=1.257, acc=0.705, acc_len=1.70, thru=17.3, I=20.0, T=34.2, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:43:07,932] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12590: forward=2619.8ms backward=1079.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:08,220] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:43:15,102] loop.py:45 INFO TRAIN_STEP step=12592 loss=1.135812 Training: 70%|██████▉ | 12592/18018 [27:13:01<10:39:17, 7.07s/step, loss=1.136, acc=0.728, acc_len=1.78, thru=15.3, I=19.2, T=37.2, wait=3.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:22,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:43:22,362] loop.py:45 INFO TRAIN_STEP step=12593 loss=1.236051 Training: 70%|██████▉ | 12593/18018 [27:13:08<10:44:23, 7.13s/step, loss=1.236, acc=0.710, acc_len=1.72, thru=18.6, I=17.4, T=36.1, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:22,614] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:43:28,647] loop.py:45 INFO TRAIN_STEP step=12594 loss=1.282385 Training: 70%|██████▉ | 12594/18018 [27:13:14<10:21:28, 6.87s/step, loss=1.282, acc=0.703, acc_len=1.69, thru=18.9, I=22.1, T=35.0, wait=2.6s, pool=53, epoch=2/2][2026-10-19 16:43:35,261] loop.py:380 INFO TIMING step=12595: step=3.559s data=1.478s compute=3.513s [fwd=2.465s bwd=1.034s opt=0.003s] dispatch=3.044s [2026-10-19 16:43:35,261] loop.py:45 INFO TRAIN_STEP step=12595 loss=1.207767 Training: 70%|██████▉ | 12595/18018 [27:13:21<10:14:14, 6.80s/step, loss=1.208, acc=0.720, acc_len=1.75, thru=18.4, I=18.6, T=36.0, wait=3.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:41,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:43:41,970] loop.py:45 INFO TRAIN_STEP step=12596 loss=1.331884 Training: 70%|██████▉ | 12596/18018 [27:13:28<10:11:47, 6.77s/step, loss=1.332, acc=0.693, acc_len=1.65, thru=19.3, I=20.5, T=36.8, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:43:41,965] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12595: forward=2501.9ms backward=932.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:42,301] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:48,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:43:48,428] loop.py:45 INFO TRAIN_STEP step=12597 loss=1.259577 Training: 70%|██████▉ | 12597/18018 [27:13:34<10:03:09, 6.68s/step, loss=1.260, acc=0.709, acc_len=1.71, thru=19.6, I=23.0, T=38.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:48,802] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:54,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:43:55,000] loop.py:45 INFO TRAIN_STEP step=12598 loss=1.386566 Training: 70%|██████▉ | 12598/18018 [27:13:41<10:00:15, 6.64s/step, loss=1.387, acc=0.689, acc_len=1.64, thru=19.7, I=21.4, T=36.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:43:55,306] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:44:01,773] loop.py:45 INFO TRAIN_STEP step=12599 loss=1.194107 Training: 70%|██████▉ | 12599/18018 [27:13:47<10:03:35, 6.68s/step, loss=1.194, acc=0.722, acc_len=1.76, thru=17.3, I=19.4, T=38.2, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:07,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:44:09,057] loop.py:380 INFO TIMING step=12600: step=3.724s data=1.634s compute=3.697s [fwd=2.641s bwd=1.042s opt=0.003s] dispatch=3.551s [2026-10-19 16:44:09,057] loop.py:45 INFO TRAIN_STEP step=12600 loss=1.322249 Training: 70%|██████▉ | 12600/18018 [27:13:55<10:19:50, 6.86s/step, loss=1.322, acc=0.696, acc_len=1.66, thru=20.7, I=21.0, T=34.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:09,291] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:44:15,256] loop.py:45 INFO TRAIN_STEP step=12601 loss=1.414976 Training: 70%|██████▉ | 12601/18018 [27:14:01<10:01:38, 6.66s/step, loss=1.415, acc=0.679, acc_len=1.61, thru=16.0, I=23.0, T=37.5, wait=2.8s, pool=25, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:44:15,250] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12600: forward=2398.0ms backward=975.9ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 16:44:15,325] eagle_store.py:517 WARNING batch_get_buffer missing keys (k403234-8f4cffbb_hs, k403234-8f4cffbb_ids, k403234-8f4cffbb_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:24,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:44:26,235] loop.py:45 INFO TRAIN_STEP step=12602 loss=1.344059 Training: 70%|██████▉ | 12602/18018 [27:14:12<11:58:23, 7.96s/step, loss=1.344, acc=0.692, acc_len=1.66, thru=15.2, I=16.4, T=31.5, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:26,416] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:44:33,631] loop.py:45 INFO TRAIN_STEP step=12603 loss=1.345304 Training: 70%|██████▉ | 12603/18018 [27:14:19<11:43:00, 7.79s/step, loss=1.345, acc=0.692, acc_len=1.65, thru=16.1, I=20.1, T=34.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:33,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:34,080] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:41,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:44:41,129] loop.py:45 INFO TRAIN_STEP step=12604 loss=1.186239 Training: 70%|██████▉ | 12604/18018 [27:14:27<11:34:58, 7.70s/step, loss=1.186, acc=0.724, acc_len=1.76, thru=17.2, I=17.1, T=34.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:41,580] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:44:46,910] loop.py:380 INFO TIMING step=12605: step=3.453s data=1.314s compute=3.425s [fwd=2.430s bwd=0.981s opt=0.003s] dispatch=2.320s [2026-10-19 16:44:46,910] loop.py:45 INFO TRAIN_STEP step=12605 loss=1.167042 Training: 70%|██████▉ | 12605/18018 [27:14:32<10:42:53, 7.13s/step, loss=1.167, acc=0.724, acc_len=1.77, thru=20.2, I=22.8, T=37.1, wait=2.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:52,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 16:44:53,705] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12605: forward=2377.6ms backward=989.5ms [2026-10-19 16:44:53,710] loop.py:45 INFO TRAIN_STEP step=12606 loss=1.271093 Training: 70%|██████▉ | 12606/18018 [27:14:39<10:33:59, 7.03s/step, loss=1.271, acc=0.709, acc_len=1.71, thru=18.8, I=20.8, T=37.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:44:53,809] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:01,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:45:02,084] loop.py:45 INFO TRAIN_STEP step=12607 loss=1.196039 Training: 70%|██████▉ | 12607/18018 [27:14:48<11:10:15, 7.43s/step, loss=1.196, acc=0.721, acc_len=1.76, thru=14.5, I=16.7, T=35.3, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:02,157] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:08,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:45:09,343] loop.py:45 INFO TRAIN_STEP step=12608 loss=1.238644 Training: 70%|██████▉ | 12608/18018 [27:14:55<11:05:26, 7.38s/step, loss=1.239, acc=0.711, acc_len=1.72, thru=17.9, I=19.9, T=34.3, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:09,622] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:15,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:45:16,554] loop.py:45 INFO TRAIN_STEP step=12609 loss=1.263108 Training: 70%|██████▉ | 12609/18018 [27:15:02<11:00:45, 7.33s/step, loss=1.263, acc=0.707, acc_len=1.71, thru=16.6, I=20.4, T=36.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:16,861] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:21,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:45:23,678] loop.py:380 INFO TIMING step=12610: step=3.699s data=1.697s compute=3.663s [fwd=2.616s bwd=1.032s opt=0.003s] dispatch=3.415s [2026-10-19 16:45:23,679] loop.py:45 INFO TRAIN_STEP step=12610 loss=1.199777 Training: 70%|██████▉ | 12610/18018 [27:15:09<10:55:05, 7.27s/step, loss=1.200, acc=0.718, acc_len=1.74, thru=17.6, I=23.5, T=34.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:23,844] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:45:29,422] loop.py:45 INFO TRAIN_STEP step=12611 loss=1.163244 Training: 70%|██████▉ | 12611/18018 [27:15:15<10:13:43, 6.81s/step, loss=1.163, acc=0.727, acc_len=1.78, thru=20.2, I=24.9, T=38.1, wait=2.4s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:45:29,417] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12610: forward=2396.9ms backward=922.5ms [2026-10-19 16:45:36,421] loop.py:45 INFO TRAIN_STEP step=12612 loss=1.355679 Training: 70%|██████▉ | 12612/18018 [27:15:22<10:18:46, 6.87s/step, loss=1.356, acc=0.688, acc_len=1.64, thru=17.3, I=19.4, T=33.9, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:45,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:45:45,748] loop.py:45 INFO TRAIN_STEP step=12613 loss=1.287550 Training: 70%|███████ | 12613/18018 [27:15:31<11:25:05, 7.61s/step, loss=1.288, acc=0.701, acc_len=1.69, thru=14.9, I=14.3, T=30.1, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:46,115] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:50,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:45:52,247] loop.py:45 INFO TRAIN_STEP step=12614 loss=1.384742 Training: 70%|███████ | 12614/18018 [27:15:38<10:55:01, 7.27s/step, loss=1.385, acc=0.689, acc_len=1.65, thru=17.6, I=22.3, T=33.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:45:52,378] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 16:45:59,008] loop.py:380 INFO TIMING step=12615: step=3.614s data=1.470s compute=3.568s [fwd=2.499s bwd=1.055s opt=0.003s] dispatch=3.138s [2026-10-19 16:45:59,008] loop.py:45 INFO TRAIN_STEP step=12615 loss=1.351526 Training: 70%|███████ | 12615/18018 [27:15:45<10:41:04, 7.12s/step, loss=1.352, acc=0.689, acc_len=1.64, thru=16.5, I=20.6, T=35.4, wait=3.1s, pool=49, epoch=2/2][2026-10-19 16:46:06,017] loop.py:45 INFO TRAIN_STEP step=12616 loss=1.296584 Training: 70%|███████ | 12616/18018 [27:15:52<10:37:59, 7.09s/step, loss=1.297, acc=0.700, acc_len=1.68, thru=18.0, I=18.9, T=37.9, wait=3.6s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:46:06,012] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12615: forward=2400.0ms backward=931.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:13,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:46:13,166] loop.py:45 INFO TRAIN_STEP step=12617 loss=1.220254 Training: 70%|███████ | 12617/18018 [27:15:59<10:39:32, 7.10s/step, loss=1.220, acc=0.711, acc_len=1.72, thru=18.2, I=18.6, T=36.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:13,547] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:46:21,254] loop.py:45 INFO TRAIN_STEP step=12618 loss=1.242044 Training: 70%|███████ | 12618/18018 [27:16:07<11:05:57, 7.40s/step, loss=1.242, acc=0.710, acc_len=1.71, thru=12.4, I=13.3, T=37.5, wait=4.7s, pool=33, epoch=2/2][2026-10-19 16:46:29,107] loop.py:45 INFO TRAIN_STEP step=12619 loss=1.265846 Training: 70%|███████ | 12619/18018 [27:16:15<11:18:10, 7.54s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=17.3, I=16.1, T=37.2, wait=4.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:38,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:46:39,234] loop.py:380 INFO TIMING step=12620: step=4.246s data=1.928s compute=4.215s [fwd=3.027s bwd=1.173s opt=0.003s] dispatch=5.872s [2026-10-19 16:46:39,235] loop.py:45 INFO TRAIN_STEP step=12620 loss=1.326892 Training: 70%|███████ | 12620/18018 [27:16:25<12:27:52, 8.31s/step, loss=1.327, acc=0.696, acc_len=1.67, thru=14.5, I=15.2, T=30.1, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:39,540] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:43,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:46:45,107] loop.py:45 INFO TRAIN_STEP step=12621 loss=1.268236 Training: 70%|███████ | 12621/18018 [27:16:31<11:21:54, 7.58s/step, loss=1.268, acc=0.704, acc_len=1.69, thru=19.4, I=23.2, T=35.9, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:46:45,102] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12620: forward=2481.5ms backward=1036.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:45,434] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:51,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:46:52,720] loop.py:45 INFO TRAIN_STEP step=12622 loss=1.276842 Training: 70%|███████ | 12622/18018 [27:16:38<11:22:40, 7.59s/step, loss=1.277, acc=0.704, acc_len=1.69, thru=16.1, I=21.4, T=35.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:52,955] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:57,528] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:46:58,456] loop.py:45 INFO TRAIN_STEP step=12623 loss=1.311421 Training: 70%|███████ | 12623/18018 [27:16:44<10:32:34, 7.04s/step, loss=1.311, acc=0.695, acc_len=1.66, thru=20.1, I=25.9, T=36.6, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:46:58,551] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 16:47:04,107] loop.py:45 INFO TRAIN_STEP step=12624 loss=1.316668 Training: 70%|███████ | 12624/18018 [27:16:50<9:55:04, 6.62s/step, loss=1.317, acc=0.697, acc_len=1.67, thru=22.3, I=24.6, T=38.0, wait=2.3s, pool=61, epoch=2/2] [2026-10-19 16:47:11,029] loop.py:380 INFO TIMING step=12625: step=3.801s data=1.616s compute=3.753s [fwd=2.595s bwd=1.144s opt=0.003s] dispatch=3.108s [2026-10-19 16:47:11,029] loop.py:45 INFO TRAIN_STEP step=12625 loss=1.246024 Training: 70%|███████ | 12625/18018 [27:16:57<10:03:11, 6.71s/step, loss=1.246, acc=0.710, acc_len=1.71, thru=18.1, I=17.5, T=33.7, wait=3.1s, pool=45, epoch=2/2][2026-10-19 16:47:21,805] loop.py:45 INFO TRAIN_STEP step=12626 loss=1.288001 Training: 70%|███████ | 12626/18018 [27:17:07<11:52:34, 7.93s/step, loss=1.288, acc=0.700, acc_len=1.68, thru=12.7, I=11.8, T=35.8, wait=7.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:47:21,801] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12625: forward=2567.0ms backward=965.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:47:30,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:47:31,103] loop.py:45 INFO TRAIN_STEP step=12627 loss=1.320780 Training: 70%|███████ | 12627/18018 [27:17:17<12:29:24, 8.34s/step, loss=1.321, acc=0.701, acc_len=1.68, thru=16.3, I=17.4, T=30.7, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:47:31,171] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 16:47:36,743] loop.py:45 INFO TRAIN_STEP step=12628 loss=1.277690 Training: 70%|███████ | 12628/18018 [27:17:22<11:16:27, 7.53s/step, loss=1.278, acc=0.706, acc_len=1.70, thru=19.5, I=20.4, T=37.4, wait=2.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:47:41,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:47:42,333] loop.py:45 INFO TRAIN_STEP step=12629 loss=1.180878 Training: 70%|███████ | 12629/18018 [27:17:28<10:24:03, 6.95s/step, loss=1.181, acc=0.720, acc_len=1.75, thru=22.1, I=22.7, T=36.8, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:47:42,833] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:47:48,908] loop.py:380 INFO TIMING step=12630: step=3.721s data=1.573s compute=3.695s [fwd=2.636s bwd=1.045s opt=0.003s] dispatch=2.845s [2026-10-19 16:47:48,908] loop.py:45 INFO TRAIN_STEP step=12630 loss=1.229928 Training: 70%|███████ | 12630/18018 [27:17:34<10:13:52, 6.84s/step, loss=1.230, acc=0.712, acc_len=1.72, thru=16.3, I=19.8, T=34.4, wait=2.8s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:47:57,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:47:59,306] loop.py:45 INFO TRAIN_STEP step=12631 loss=1.315440 Training: 70%|███████ | 12631/18018 [27:17:45<11:49:42, 7.90s/step, loss=1.315, acc=0.698, acc_len=1.67, thru=15.8, I=17.4, T=32.4, wait=6.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:47:59,300] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12630: forward=2870.7ms backward=1026.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:47:59,412] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:48:04,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:48:05,289] loop.py:45 INFO TRAIN_STEP step=12632 loss=1.301924 Training: 70%|███████ | 12632/18018 [27:17:51<10:57:51, 7.33s/step, loss=1.302, acc=0.696, acc_len=1.67, thru=19.7, I=24.1, T=38.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:48:05,482] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:48:11,364] loop.py:45 INFO TRAIN_STEP step=12633 loss=1.243345 Training: 70%|███████ | 12633/18018 [27:17:57<10:23:54, 6.95s/step, loss=1.243, acc=0.708, acc_len=1.71, thru=19.8, I=24.1, T=36.9, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:48:20,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:48:21,608] loop.py:45 INFO TRAIN_STEP step=12634 loss=1.274950 Training: 70%|███████ | 12634/18018 [27:18:07<11:52:28, 7.94s/step, loss=1.275, acc=0.703, acc_len=1.69, thru=14.4, I=13.9, T=33.2, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:48:21,872] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:48:26,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:48:27,738] loop.py:380 INFO TIMING step=12635: step=3.992s data=1.844s compute=3.965s [fwd=2.769s bwd=1.181s opt=0.003s] dispatch=2.128s [2026-10-19 16:48:27,738] loop.py:45 INFO TRAIN_STEP step=12635 loss=1.283257 Training: 70%|███████ | 12635/18018 [27:18:13<11:03:34, 7.40s/step, loss=1.283, acc=0.702, acc_len=1.69, thru=19.8, I=24.5, T=32.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:48:28,233] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:48:33,326] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:48:33,783] loop.py:45 INFO TRAIN_STEP step=12636 loss=1.324774 Training: 70%|███████ | 12636/18018 [27:18:19<10:27:09, 6.99s/step, loss=1.325, acc=0.696, acc_len=1.66, thru=19.7, I=25.8, T=39.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:48:33,834] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (TrainerActor pid=9310) [2026-10-19 16:48:33,778] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12635: forward=2301.4ms backward=927.9ms [2026-10-19 16:48:42,421] loop.py:45 INFO TRAIN_STEP step=12637 loss=1.286581 Training: 70%|███████ | 12637/18018 [27:18:28<11:11:17, 7.49s/step, loss=1.287, acc=0.702, acc_len=1.69, thru=15.1, I=16.3, T=34.4, wait=4.9s, pool=61, epoch=2/2][2026-10-19 16:48:48,814] loop.py:45 INFO TRAIN_STEP step=12638 loss=1.287476 Training: 70%|███████ | 12638/18018 [27:18:34<10:41:47, 7.16s/step, loss=1.287, acc=0.707, acc_len=1.70, thru=18.1, I=17.6, T=35.7, wait=2.8s, pool=49, epoch=2/2][2026-10-19 16:48:57,830] loop.py:45 INFO TRAIN_STEP step=12639 loss=1.183356 Training: 70%|███████ | 12639/18018 [27:18:43<11:31:40, 7.72s/step, loss=1.183, acc=0.718, acc_len=1.74, thru=15.6, I=16.0, T=34.5, wait=5.3s, pool=61, epoch=2/2][2026-10-19 16:49:04,733] loop.py:380 INFO TIMING step=12640: step=3.830s data=1.751s compute=3.794s [fwd=2.689s bwd=1.091s opt=0.003s] dispatch=3.063s [2026-10-19 16:49:04,733] loop.py:45 INFO TRAIN_STEP step=12640 loss=1.249322 Training: 70%|███████ | 12640/18018 [27:18:50<11:09:38, 7.47s/step, loss=1.249, acc=0.706, acc_len=1.70, thru=17.2, I=17.4, T=33.4, wait=3.1s, pool=57, epoch=2/2][2026-10-19 16:49:11,761] loop.py:45 INFO TRAIN_STEP step=12641 loss=1.269452 Training: 70%|███████ | 12641/18018 [27:18:57<10:57:36, 7.34s/step, loss=1.269, acc=0.706, acc_len=1.70, thru=18.1, I=17.4, T=34.7, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:49:11,757] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12640: forward=2577.9ms backward=1055.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:49:11,756] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:49:12,257] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:49:17,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:49:18,868] loop.py:45 INFO TRAIN_STEP step=12642 loss=1.166367 Training: 70%|███████ | 12642/18018 [27:19:04<10:51:21, 7.27s/step, loss=1.166, acc=0.721, acc_len=1.75, thru=17.6, I=22.2, T=37.2, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:49:19,264] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 16:49:25,919] loop.py:45 INFO TRAIN_STEP step=12643 loss=1.182821 Training: 70%|███████ | 12643/18018 [27:19:12<10:45:17, 7.20s/step, loss=1.183, acc=0.721, acc_len=1.75, thru=17.6, I=22.5, T=38.7, wait=3.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:49:31,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:49:31,513] loop.py:45 INFO TRAIN_STEP step=12644 loss=1.240728 Training: 70%|███████ | 12644/18018 [27:19:17<10:01:59, 6.72s/step, loss=1.241, acc=0.709, acc_len=1.71, thru=21.2, I=21.8, T=39.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:49:31,824] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:49:38,190] loop.py:380 INFO TIMING step=12645: step=3.544s data=1.505s compute=3.514s [fwd=2.519s bwd=0.983s opt=0.003s] dispatch=3.123s [2026-10-19 16:49:38,191] loop.py:45 INFO TRAIN_STEP step=12645 loss=1.441219 Training: 70%|███████ | 12645/18018 [27:19:24<10:00:42, 6.71s/step, loss=1.441, acc=0.680, acc_len=1.61, thru=19.3, I=22.4, T=36.1, wait=3.1s, pool=61, epoch=2/2][2026-10-19 16:49:46,481] loop.py:45 INFO TRAIN_STEP step=12646 loss=1.251356 Training: 70%|███████ | 12646/18018 [27:19:32<10:43:02, 7.18s/step, loss=1.251, acc=0.706, acc_len=1.70, thru=14.3, I=14.1, T=35.9, wait=4.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:49:46,476] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12645: forward=2508.0ms backward=1017.6ms [2026-10-19 16:49:54,700] loop.py:45 INFO TRAIN_STEP step=12647 loss=1.287295 Training: 70%|███████ | 12647/18018 [27:19:40<11:10:49, 7.49s/step, loss=1.287, acc=0.703, acc_len=1.69, thru=15.7, I=15.3, T=32.9, wait=4.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:01,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:50:02,983] loop.py:45 INFO TRAIN_STEP step=12648 loss=1.305515 Training: 70%|███████ | 12648/18018 [27:19:49<11:31:49, 7.73s/step, loss=1.306, acc=0.704, acc_len=1.69, thru=16.7, I=17.6, T=29.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:03,307] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:50:10,643] loop.py:45 INFO TRAIN_STEP step=12649 loss=1.245760 Training: 70%|███████ | 12649/18018 [27:19:56<11:29:51, 7.71s/step, loss=1.246, acc=0.708, acc_len=1.70, thru=13.3, I=16.2, T=38.0, wait=4.3s, pool=41, epoch=2/2][2026-10-19 16:50:18,495] loop.py:380 INFO TIMING step=12650: step=3.558s data=1.444s compute=3.524s [fwd=2.542s bwd=0.968s opt=0.003s] dispatch=4.284s [2026-10-19 16:50:18,495] loop.py:45 INFO TRAIN_STEP step=12650 loss=1.285258 Training: 70%|███████ | 12650/18018 [27:20:04<11:33:34, 7.75s/step, loss=1.285, acc=0.703, acc_len=1.69, thru=15.7, I=16.3, T=36.0, wait=4.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:25,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:50:25,959] loop.py:45 INFO TRAIN_STEP step=12651 loss=1.289843 Training: 70%|███████ | 12651/18018 [27:20:12<11:25:43, 7.67s/step, loss=1.290, acc=0.704, acc_len=1.70, thru=18.3, I=19.5, T=35.8, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:50:25,955] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12650: forward=2544.2ms backward=988.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:26,217] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:32,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:50:33,385] loop.py:45 INFO TRAIN_STEP step=12652 loss=1.312528 Training: 70%|███████ | 12652/18018 [27:20:19<11:19:09, 7.59s/step, loss=1.313, acc=0.696, acc_len=1.66, thru=17.5, I=21.1, T=34.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:33,569] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 16:50:38,996] eagle_store.py:517 WARNING batch_get_buffer missing keys (k404894-87d00928_hs, k404894-87d00928_ids, k404894-87d00928_lhs); sleeping 0.50s. [2026-10-19 16:50:39,972] loop.py:45 INFO TRAIN_STEP step=12653 loss=1.325612 Training: 70%|███████ | 12653/18018 [27:20:26<10:51:56, 7.29s/step, loss=1.326, acc=0.695, acc_len=1.66, thru=16.9, I=19.3, T=39.0, wait=3.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:46,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 16:50:38,995] eagle_store.py:517 WARNING batch_get_buffer missing keys (k404893-81d5e4d8_hs, k404893-81d5e4d8_ids, k404893-81d5e4d8_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 16:50:46,427] loop.py:45 INFO TRAIN_STEP step=12654 loss=1.279486 Training: 70%|███████ | 12654/18018 [27:20:32<10:29:27, 7.04s/step, loss=1.279, acc=0.703, acc_len=1.69, thru=20.2, I=20.6, T=36.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:46,695] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:54,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:50:54,673] loop.py:380 INFO TIMING step=12655: step=4.038s data=1.932s compute=4.013s [fwd=2.997s bwd=1.003s opt=0.003s] dispatch=4.198s [2026-10-19 16:50:54,674] loop.py:45 INFO TRAIN_STEP step=12655 loss=1.416476 Training: 70%|███████ | 12655/18018 [27:20:40<11:01:36, 7.40s/step, loss=1.416, acc=0.676, acc_len=1.60, thru=15.2, I=15.7, T=31.7, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:50:55,108] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:51:02,176] loop.py:45 INFO TRAIN_STEP step=12656 loss=1.250072 Training: 70%|███████ | 12656/18018 [27:20:48<11:04:12, 7.43s/step, loss=1.250, acc=0.710, acc_len=1.72, thru=15.3, I=15.9, T=30.4, wait=3.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:51:02,172] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12655: forward=3109.5ms backward=1049.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:51:08,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:51:08,641] loop.py:45 INFO TRAIN_STEP step=12657 loss=1.401234 Training: 70%|███████ | 12657/18018 [27:20:54<10:38:10, 7.14s/step, loss=1.401, acc=0.681, acc_len=1.62, thru=18.1, I=18.2, T=36.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:51:08,888] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:51:17,265] loop.py:45 INFO TRAIN_STEP step=12658 loss=1.284918 Training: 70%|███████ | 12658/18018 [27:21:03<11:17:46, 7.59s/step, loss=1.285, acc=0.703, acc_len=1.69, thru=14.3, I=15.8, T=30.8, wait=4.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:51:24,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:51:26,224] loop.py:45 INFO TRAIN_STEP step=12659 loss=1.337332 Training: 70%|███████ | 12659/18018 [27:21:12<11:54:25, 8.00s/step, loss=1.337, acc=0.694, acc_len=1.66, thru=16.0, I=17.8, T=30.9, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:51:26,731] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 16:51:33,067] loop.py:380 INFO TIMING step=12660: step=3.920s data=1.950s compute=3.886s [fwd=2.872s bwd=1.000s opt=0.003s] dispatch=2.914s [2026-10-19 16:51:33,067] loop.py:45 INFO TRAIN_STEP step=12660 loss=1.262938 Training: 70%|███████ | 12660/18018 [27:21:19<11:23:16, 7.65s/step, loss=1.263, acc=0.708, acc_len=1.71, thru=16.1, I=24.7, T=32.7, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:51:42,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:51:43,269] loop.py:45 INFO TRAIN_STEP step=12661 loss=1.314781 Training: 70%|███████ | 12661/18018 [27:21:29<12:31:30, 8.42s/step, loss=1.315, acc=0.695, acc_len=1.67, thru=13.8, I=12.8, T=32.0, wait=6.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:51:43,265] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12660: forward=2895.7ms backward=1062.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:51:43,617] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:51:49,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:51:50,544] loop.py:45 INFO TRAIN_STEP step=12662 loss=1.198223 Training: 70%|███████ | 12662/18018 [27:21:36<12:00:42, 8.07s/step, loss=1.198, acc=0.721, acc_len=1.75, thru=17.7, I=20.1, T=37.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:51:50,918] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:51:56,484] loop.py:45 INFO TRAIN_STEP step=12663 loss=1.186085 Training: 70%|███████ | 12663/18018 [27:21:42<11:03:32, 7.43s/step, loss=1.186, acc=0.723, acc_len=1.76, thru=18.5, I=25.8, T=36.0, wait=2.4s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:52:03,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:52:04,678] loop.py:45 INFO TRAIN_STEP step=12664 loss=1.288317 Training: 70%|███████ | 12664/18018 [27:21:50<11:23:44, 7.66s/step, loss=1.288, acc=0.703, acc_len=1.70, thru=16.9, I=17.0, T=35.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:52:04,786] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:52:10,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:52:11,216] loop.py:380 INFO TIMING step=12665: step=3.489s data=1.523s compute=3.452s [fwd=2.438s bwd=1.000s opt=0.003s] dispatch=3.039s [2026-10-19 16:52:11,216] loop.py:45 INFO TRAIN_STEP step=12665 loss=1.251408 Training: 70%|███████ | 12665/18018 [27:21:57<10:53:29, 7.32s/step, loss=1.251, acc=0.710, acc_len=1.72, thru=19.2, I=22.0, T=36.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:52:11,363] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:52:19,449] loop.py:45 INFO TRAIN_STEP step=12666 loss=1.356263 Training: 70%|███████ | 12666/18018 [27:22:05<11:17:42, 7.60s/step, loss=1.356, acc=0.689, acc_len=1.64, thru=14.3, I=13.9, T=32.8, wait=4.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:52:19,445] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12665: forward=2811.8ms backward=1055.9ms [2026-10-19 16:52:27,265] loop.py:45 INFO TRAIN_STEP step=12667 loss=1.262304 Training: 70%|███████ | 12667/18018 [27:22:13<11:23:21, 7.66s/step, loss=1.262, acc=0.705, acc_len=1.70, thru=15.5, I=13.3, T=31.0, wait=3.7s, pool=37, epoch=2/2][2026-10-19 16:52:34,724] loop.py:45 INFO TRAIN_STEP step=12668 loss=1.310231 Training: 70%|███████ | 12668/18018 [27:22:20<11:17:44, 7.60s/step, loss=1.310, acc=0.703, acc_len=1.69, thru=18.4, I=15.6, T=34.5, wait=3.7s, pool=61, epoch=2/2][2026-10-19 16:52:44,020] loop.py:45 INFO TRAIN_STEP step=12669 loss=1.326505 Training: 70%|███████ | 12669/18018 [27:22:30<12:03:00, 8.11s/step, loss=1.327, acc=0.697, acc_len=1.67, thru=14.8, I=14.9, T=33.3, wait=5.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:52:51,883] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:52:52,821] loop.py:380 INFO TIMING step=12670: step=4.012s data=1.772s compute=3.982s [fwd=2.835s bwd=1.133s opt=0.003s] dispatch=4.780s [2026-10-19 16:52:52,821] loop.py:45 INFO TRAIN_STEP step=12670 loss=1.326152 Training: 70%|███████ | 12670/18018 [27:22:38<12:21:20, 8.32s/step, loss=1.326, acc=0.693, acc_len=1.66, thru=16.0, I=16.6, T=31.9, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:52:52,901] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:00,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:53:01,374] loop.py:45 INFO TRAIN_STEP step=12671 loss=1.282872 Training: 70%|███████ | 12671/18018 [27:22:47<12:27:32, 8.39s/step, loss=1.283, acc=0.702, acc_len=1.69, thru=14.3, I=17.2, T=33.3, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:53:01,369] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12670: forward=2700.3ms backward=1099.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:01,476] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:05,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:53:06,752] loop.py:45 INFO TRAIN_STEP step=12672 loss=1.313657 Training: 70%|███████ | 12672/18018 [27:22:52<11:06:56, 7.49s/step, loss=1.314, acc=0.699, acc_len=1.68, thru=21.4, I=24.4, T=35.6, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:06,866] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:12,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:53:13,170] loop.py:45 INFO TRAIN_STEP step=12673 loss=1.313732 Training: 70%|███████ | 12673/18018 [27:22:59<10:38:14, 7.16s/step, loss=1.314, acc=0.697, acc_len=1.66, thru=19.7, I=22.1, T=38.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:13,258] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:21,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:53:21,684] loop.py:45 INFO TRAIN_STEP step=12674 loss=1.252697 Training: 70%|███████ | 12674/18018 [27:23:07<11:14:11, 7.57s/step, loss=1.253, acc=0.711, acc_len=1.71, thru=16.0, I=16.5, T=36.4, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:22,111] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:29,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:53:30,340] loop.py:380 INFO TIMING step=12675: step=4.120s data=1.896s compute=4.087s [fwd=2.982s bwd=1.090s opt=0.003s] dispatch=4.526s [2026-10-19 16:53:30,340] loop.py:45 INFO TRAIN_STEP step=12675 loss=1.333725 Training: 70%|███████ | 12675/18018 [27:23:16<11:43:07, 7.90s/step, loss=1.334, acc=0.694, acc_len=1.66, thru=16.1, I=17.5, T=31.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:30,443] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:37,701] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:53:39,520] loop.py:45 INFO TRAIN_STEP step=12676 loss=1.262046 Training: 70%|███████ | 12676/18018 [27:23:25<12:17:17, 8.28s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=14.1, I=17.5, T=28.8, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:53:39,515] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12675: forward=3124.7ms backward=1282.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:39,717] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:44,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:53:45,763] loop.py:45 INFO TRAIN_STEP step=12677 loss=1.264296 Training: 70%|███████ | 12677/18018 [27:23:31<11:22:44, 7.67s/step, loss=1.264, acc=0.705, acc_len=1.70, thru=18.1, I=24.9, T=36.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:53:46,232] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:53:53,996] loop.py:45 INFO TRAIN_STEP step=12678 loss=1.340611 Training: 70%|███████ | 12678/18018 [27:23:40<11:37:35, 7.84s/step, loss=1.341, acc=0.692, acc_len=1.66, thru=11.7, I=16.8, T=35.5, wait=4.6s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:01,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:54:02,728] loop.py:45 INFO TRAIN_STEP step=12679 loss=1.247910 Training: 70%|███████ | 12679/18018 [27:23:48<12:01:20, 8.11s/step, loss=1.248, acc=0.708, acc_len=1.71, thru=17.6, I=17.2, T=34.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:03,099] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:54:09,679] loop.py:380 INFO TIMING step=12680: step=4.137s data=1.864s compute=4.104s [fwd=2.894s bwd=1.195s opt=0.003s] dispatch=2.805s [2026-10-19 16:54:09,679] loop.py:45 INFO TRAIN_STEP step=12680 loss=1.224087 Training: 70%|███████ | 12680/18018 [27:23:55<11:30:25, 7.76s/step, loss=1.224, acc=0.712, acc_len=1.72, thru=16.6, I=18.4, T=30.9, wait=2.8s, pool=53, epoch=2/2][2026-10-19 16:54:16,887] loop.py:45 INFO TRAIN_STEP step=12681 loss=1.318256 Training: 70%|███████ | 12681/18018 [27:24:02<11:15:28, 7.59s/step, loss=1.318, acc=0.694, acc_len=1.66, thru=15.8, I=15.4, T=37.0, wait=3.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:54:16,883] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12680: forward=2411.3ms backward=981.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:27,746] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:54:27,893] loop.py:45 INFO TRAIN_STEP step=12682 loss=1.350959 Training: 70%|███████ | 12682/18018 [27:24:13<12:46:25, 8.62s/step, loss=1.351, acc=0.691, acc_len=1.66, thru=14.1, I=13.3, T=33.4, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:28,248] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:33,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:54:34,789] loop.py:45 INFO TRAIN_STEP step=12683 loss=1.402493 Training: 70%|███████ | 12683/18018 [27:24:20<12:00:21, 8.10s/step, loss=1.402, acc=0.681, acc_len=1.61, thru=18.4, I=20.5, T=34.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:34,956] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:40,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:54:41,100] loop.py:45 INFO TRAIN_STEP step=12684 loss=1.192342 Training: 70%|███████ | 12684/18018 [27:24:27<11:12:25, 7.56s/step, loss=1.192, acc=0.716, acc_len=1.73, thru=19.1, I=22.4, T=36.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:41,582] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:54:46,730] loop.py:380 INFO TIMING step=12685: step=3.318s data=1.291s compute=3.283s [fwd=2.339s bwd=0.930s opt=0.003s] dispatch=2.304s [2026-10-19 16:54:46,731] loop.py:45 INFO TRAIN_STEP step=12685 loss=1.255118 Training: 70%|███████ | 12685/18018 [27:24:32<10:20:46, 6.98s/step, loss=1.255, acc=0.707, acc_len=1.71, thru=21.6, I=25.8, T=38.6, wait=2.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:52,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:54:53,285] loop.py:45 INFO TRAIN_STEP step=12686 loss=1.262543 Training: 70%|███████ | 12686/18018 [27:24:39<10:09:10, 6.86s/step, loss=1.263, acc=0.708, acc_len=1.70, thru=20.8, I=22.8, T=35.4, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:54:53,280] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12685: forward=2586.6ms backward=985.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:54:53,532] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:55:00,150] loop.py:45 INFO TRAIN_STEP step=12687 loss=1.337006 Training: 70%|███████ | 12687/18018 [27:24:46<10:09:22, 6.86s/step, loss=1.337, acc=0.697, acc_len=1.66, thru=16.8, I=18.5, T=37.0, wait=3.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:06,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:55:06,411] loop.py:45 INFO TRAIN_STEP step=12688 loss=1.308122 Training: 70%|███████ | 12688/18018 [27:24:52<9:53:18, 6.68s/step, loss=1.308, acc=0.698, acc_len=1.67, thru=20.1, I=19.4, T=35.5, wait=2.6s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:06,705] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:11,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:55:12,096] loop.py:45 INFO TRAIN_STEP step=12689 loss=1.268376 Training: 70%|███████ | 12689/18018 [27:24:58<9:26:41, 6.38s/step, loss=1.268, acc=0.706, acc_len=1.71, thru=21.6, I=24.3, T=39.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:12,558] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:17,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:55:18,321] loop.py:380 INFO TIMING step=12690: step=3.358s data=1.402s compute=3.331s [fwd=2.422s bwd=0.895s opt=0.003s] dispatch=2.860s [2026-10-19 16:55:18,321] loop.py:45 INFO TRAIN_STEP step=12690 loss=1.201407 Training: 70%|███████ | 12690/18018 [27:25:04<9:22:30, 6.33s/step, loss=1.201, acc=0.717, acc_len=1.74, thru=21.2, I=27.0, T=38.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:18,521] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:55:24,069] loop.py:45 INFO TRAIN_STEP step=12691 loss=1.204830 Training: 70%|███████ | 12691/18018 [27:25:10<9:06:47, 6.16s/step, loss=1.205, acc=0.717, acc_len=1.74, thru=20.7, I=25.5, T=35.9, wait=2.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:55:24,064] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12690: forward=2519.2ms backward=1005.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:30,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:55:30,861] loop.py:45 INFO TRAIN_STEP step=12692 loss=1.321576 Training: 70%|███████ | 12692/18018 [27:25:16<9:23:31, 6.35s/step, loss=1.322, acc=0.700, acc_len=1.67, thru=18.1, I=20.0, T=32.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:31,163] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:36,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:55:36,881] loop.py:45 INFO TRAIN_STEP step=12693 loss=1.211009 Training: 70%|███████ | 12693/18018 [27:25:22<9:14:38, 6.25s/step, loss=1.211, acc=0.716, acc_len=1.74, thru=19.3, I=22.2, T=37.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:37,237] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:55:47,060] loop.py:45 INFO TRAIN_STEP step=12694 loss=1.235379 Training: 70%|███████ | 12694/18018 [27:25:33<10:59:08, 7.43s/step, loss=1.235, acc=0.713, acc_len=1.73, thru=10.2, I=10.2, T=35.4, wait=6.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:55,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:55:56,820] loop.py:380 INFO TIMING step=12695: step=3.530s data=1.451s compute=3.499s [fwd=2.567s bwd=0.918s opt=0.003s] dispatch=6.222s [2026-10-19 16:55:56,820] loop.py:45 INFO TRAIN_STEP step=12695 loss=1.231346 Training: 70%|███████ | 12695/18018 [27:25:42<12:01:05, 8.13s/step, loss=1.231, acc=0.711, acc_len=1.72, thru=16.4, I=15.4, T=36.3, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:55:56,901] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:02,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:56:03,121] loop.py:45 INFO TRAIN_STEP step=12696 loss=1.185955 Training: 70%|███████ | 12696/18018 [27:25:49<11:12:21, 7.58s/step, loss=1.186, acc=0.721, acc_len=1.75, thru=20.2, I=22.9, T=36.0, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:56:03,116] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12695: forward=2535.0ms backward=978.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:03,362] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:56:09,182] loop.py:45 INFO TRAIN_STEP step=12697 loss=1.207853 Training: 70%|███████ | 12697/18018 [27:25:55<10:31:50, 7.12s/step, loss=1.208, acc=0.717, acc_len=1.73, thru=16.9, I=22.5, T=37.1, wait=2.6s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:21,600] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:56:22,014] loop.py:45 INFO TRAIN_STEP step=12698 loss=1.134719 Training: 70%|███████ | 12698/18018 [27:26:08<13:03:30, 8.84s/step, loss=1.135, acc=0.730, acc_len=1.79, thru=10.9, I=12.0, T=30.6, wait=8.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:22,109] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:29,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:56:29,601] loop.py:45 INFO TRAIN_STEP step=12699 loss=1.250817 Training: 70%|███████ | 12699/18018 [27:26:15<12:30:05, 8.46s/step, loss=1.251, acc=0.710, acc_len=1.71, thru=17.2, I=16.9, T=35.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:29,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 16:56:36,860] loop.py:380 INFO TIMING step=12700: step=3.730s data=1.620s compute=3.703s [fwd=2.650s bwd=1.039s opt=0.003s] dispatch=3.521s [2026-10-19 16:56:36,860] loop.py:45 INFO TRAIN_STEP step=12700 loss=1.231497 Training: 70%|███████ | 12700/18018 [27:26:22<11:58:02, 8.10s/step, loss=1.231, acc=0.711, acc_len=1.72, thru=15.1, I=18.7, T=34.3, wait=3.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:48,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:56:49,882] loop.py:45 INFO TRAIN_STEP step=12701 loss=1.296027 Training: 70%|███████ | 12701/18018 [27:26:35<14:08:41, 9.58s/step, loss=1.296, acc=0.701, acc_len=1.68, thru=15.6, I=12.5, T=31.4, wait=8.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:56:49,878] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12700: forward=2962.2ms backward=1070.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:50,082] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:56,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:56:56,448] loop.py:45 INFO TRAIN_STEP step=12702 loss=1.201691 Training: 70%|███████ | 12702/18018 [27:26:42<12:48:28, 8.67s/step, loss=1.202, acc=0.716, acc_len=1.74, thru=18.1, I=23.0, T=37.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:56:56,627] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:02,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:57:03,148] loop.py:45 INFO TRAIN_STEP step=12703 loss=1.289816 Training: 71%|███████ | 12703/18018 [27:26:49<11:55:55, 8.08s/step, loss=1.290, acc=0.702, acc_len=1.68, thru=19.2, I=21.8, T=37.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:03,369] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:09,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:57:09,997] loop.py:45 INFO TRAIN_STEP step=12704 loss=1.207009 Training: 71%|███████ | 12704/18018 [27:26:56<11:22:59, 7.71s/step, loss=1.207, acc=0.720, acc_len=1.75, thru=18.8, I=22.3, T=34.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:10,053] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 16:57:17,977] loop.py:380 INFO TIMING step=12705: step=3.737s data=1.638s compute=3.700s [fwd=2.592s bwd=1.093s opt=0.003s] dispatch=4.233s [2026-10-19 16:57:17,977] loop.py:45 INFO TRAIN_STEP step=12705 loss=1.069419 Training: 71%|███████ | 12705/18018 [27:27:04<11:29:56, 7.79s/step, loss=1.069, acc=0.740, acc_len=1.83, thru=15.3, I=19.0, T=34.2, wait=4.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:26,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:57:27,359] loop.py:45 INFO TRAIN_STEP step=12706 loss=1.283669 Training: 71%|███████ | 12706/18018 [27:27:13<12:12:09, 8.27s/step, loss=1.284, acc=0.701, acc_len=1.69, thru=15.0, I=15.2, T=34.6, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:57:27,354] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12705: forward=2569.1ms backward=1086.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:27,568] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:32,652] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:57:33,811] loop.py:45 INFO TRAIN_STEP step=12707 loss=1.337473 Training: 71%|███████ | 12707/18018 [27:27:19<11:23:43, 7.72s/step, loss=1.337, acc=0.694, acc_len=1.66, thru=18.9, I=21.7, T=35.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:34,163] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:39,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:57:39,720] loop.py:45 INFO TRAIN_STEP step=12708 loss=1.214766 Training: 71%|███████ | 12708/18018 [27:27:25<10:35:24, 7.18s/step, loss=1.215, acc=0.714, acc_len=1.73, thru=20.4, I=23.8, T=38.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:39,919] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:45,597] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:57:45,983] loop.py:45 INFO TRAIN_STEP step=12709 loss=1.291425 Training: 71%|███████ | 12709/18018 [27:27:32<10:10:57, 6.90s/step, loss=1.291, acc=0.703, acc_len=1.69, thru=21.1, I=21.8, T=37.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:46,109] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:52,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:57:53,149] loop.py:380 INFO TIMING step=12710: step=3.985s data=1.787s compute=3.953s [fwd=2.798s bwd=1.140s opt=0.003s] dispatch=3.171s [2026-10-19 16:57:53,149] loop.py:45 INFO TRAIN_STEP step=12710 loss=1.261649 Training: 71%|███████ | 12710/18018 [27:27:39<10:17:43, 6.98s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=17.9, I=19.0, T=32.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:57:53,589] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:58:00,062] loop.py:45 INFO TRAIN_STEP step=12711 loss=1.160237 Training: 71%|███████ | 12711/18018 [27:27:46<10:15:49, 6.96s/step, loss=1.160, acc=0.722, acc_len=1.76, thru=16.9, I=17.5, T=36.5, wait=3.4s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:58:00,058] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12710: forward=2529.3ms backward=929.2ms [2026-10-19 16:58:06,826] loop.py:45 INFO TRAIN_STEP step=12712 loss=1.354943 Training: 71%|███████ | 12712/18018 [27:27:52<10:10:24, 6.90s/step, loss=1.355, acc=0.685, acc_len=1.64, thru=19.6, I=18.6, T=36.7, wait=3.3s, pool=61, epoch=2/2][2026-10-19 16:58:13,029] loop.py:45 INFO TRAIN_STEP step=12713 loss=1.403989 Training: 71%|███████ | 12713/18018 [27:27:59<9:51:43, 6.69s/step, loss=1.404, acc=0.681, acc_len=1.62, thru=18.8, I=20.6, T=36.3, wait=2.7s, pool=49, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:58:19,883] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:58:21,227] loop.py:45 INFO TRAIN_STEP step=12714 loss=1.285910 Training: 71%|███████ | 12714/18018 [27:28:07<10:31:32, 7.14s/step, loss=1.286, acc=0.702, acc_len=1.69, thru=16.0, I=17.9, T=36.5, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:58:21,389] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:58:27,355] loop.py:380 INFO TIMING step=12715: step=3.284s data=1.333s compute=3.254s [fwd=2.336s bwd=0.904s opt=0.003s] dispatch=2.835s [2026-10-19 16:58:27,355] loop.py:45 INFO TRAIN_STEP step=12715 loss=1.216097 Training: 71%|███████ | 12715/18018 [27:28:13<10:04:27, 6.84s/step, loss=1.216, acc=0.715, acc_len=1.73, thru=19.9, I=20.4, T=39.0, wait=2.8s, pool=57, epoch=2/2][2026-10-19 16:58:33,536] loop.py:45 INFO TRAIN_STEP step=12716 loss=1.262484 Training: 71%|███████ | 12716/18018 [27:28:19<9:46:58, 6.64s/step, loss=1.262, acc=0.710, acc_len=1.72, thru=20.9, I=20.4, T=37.3, wait=2.7s, pool=61, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 16:58:33,531] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12715: forward=2421.2ms backward=968.9ms [2026-10-19 16:58:41,309] loop.py:45 INFO TRAIN_STEP step=12717 loss=1.351466 Training: 71%|███████ | 12717/18018 [27:28:27<10:16:45, 6.98s/step, loss=1.351, acc=0.686, acc_len=1.63, thru=16.7, I=17.7, T=35.1, wait=4.1s, pool=57, epoch=2/2][2026-10-19 16:58:50,000] loop.py:45 INFO TRAIN_STEP step=12718 loss=1.387213 Training: 71%|███████ | 12718/18018 [27:28:36<11:02:00, 7.49s/step, loss=1.387, acc=0.686, acc_len=1.63, thru=14.8, I=14.7, T=34.7, wait=5.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:58:58,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:58:59,622] loop.py:45 INFO TRAIN_STEP step=12719 loss=1.303469 Training: 71%|███████ | 12719/18018 [27:28:45<11:58:13, 8.13s/step, loss=1.303, acc=0.703, acc_len=1.69, thru=14.2, I=15.4, T=33.6, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:00,021] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:06,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:59:07,636] loop.py:380 INFO TIMING step=12720: step=3.535s data=1.496s compute=3.506s [fwd=2.521s bwd=0.970s opt=0.003s] dispatch=4.470s [2026-10-19 16:59:07,636] loop.py:45 INFO TRAIN_STEP step=12720 loss=1.183218 Training: 71%|███████ | 12720/18018 [27:28:53<11:54:56, 8.10s/step, loss=1.183, acc=0.724, acc_len=1.76, thru=15.0, I=19.0, T=36.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:07,913] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 16:59:15,038] loop.py:45 INFO TRAIN_STEP step=12721 loss=1.191786 Training: 71%|███████ | 12721/18018 [27:29:01<11:36:25, 7.89s/step, loss=1.192, acc=0.720, acc_len=1.75, thru=16.3, I=18.7, T=34.6, wait=3.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:59:15,034] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12720: forward=2466.1ms backward=1056.8ms [2026-10-19 16:59:22,560] loop.py:45 INFO TRAIN_STEP step=12722 loss=1.284037 Training: 71%|███████ | 12722/18018 [27:29:08<11:26:34, 7.78s/step, loss=1.284, acc=0.703, acc_len=1.69, thru=16.3, I=15.9, T=34.3, wait=3.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:30,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:59:31,123] loop.py:45 INFO TRAIN_STEP step=12723 loss=1.231814 Training: 71%|███████ | 12723/18018 [27:29:17<11:47:16, 8.01s/step, loss=1.232, acc=0.714, acc_len=1.74, thru=16.6, I=16.3, T=35.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:31,631] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:37,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:59:38,075] loop.py:45 INFO TRAIN_STEP step=12724 loss=1.240728 Training: 71%|███████ | 12724/18018 [27:29:24<11:18:59, 7.70s/step, loss=1.241, acc=0.706, acc_len=1.70, thru=19.5, I=19.4, T=37.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:38,467] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 16:59:44,349] loop.py:380 INFO TIMING step=12725: step=3.268s data=1.310s compute=3.235s [fwd=2.317s bwd=0.903s opt=0.003s] dispatch=2.996s [2026-10-19 16:59:44,349] loop.py:45 INFO TRAIN_STEP step=12725 loss=1.250458 Training: 71%|███████ | 12725/18018 [27:29:30<10:41:12, 7.27s/step, loss=1.250, acc=0.706, acc_len=1.70, thru=17.6, I=19.7, T=39.2, wait=3.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:51,558] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:59:52,394] loop.py:45 INFO TRAIN_STEP step=12726 loss=1.319017 Training: 71%|███████ | 12726/18018 [27:29:38<11:01:39, 7.50s/step, loss=1.319, acc=0.697, acc_len=1.66, thru=18.7, I=17.9, T=36.2, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 16:59:52,390] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12725: forward=2489.0ms backward=1004.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:52,570] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:58,179] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 16:59:58,664] loop.py:45 INFO TRAIN_STEP step=12727 loss=1.283448 Training: 71%|███████ | 12727/18018 [27:29:44<10:28:55, 7.13s/step, loss=1.283, acc=0.702, acc_len=1.69, thru=19.9, I=21.3, T=37.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 16:59:58,693] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:06,421] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:00:07,063] loop.py:45 INFO TRAIN_STEP step=12728 loss=1.347128 Training: 71%|███████ | 12728/18018 [27:29:53<11:02:20, 7.51s/step, loss=1.347, acc=0.692, acc_len=1.65, thru=15.9, I=17.3, T=34.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:07,425] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:12,377] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:00:13,947] loop.py:45 INFO TRAIN_STEP step=12729 loss=1.292812 Training: 71%|███████ | 12729/18018 [27:30:00<10:45:34, 7.32s/step, loss=1.293, acc=0.706, acc_len=1.69, thru=18.1, I=24.4, T=35.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:14,386] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:21,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:00:22,567] loop.py:380 INFO TIMING step=12730: step=3.409s data=1.396s compute=3.356s [fwd=2.411s bwd=0.929s opt=0.003s] dispatch=5.202s [2026-10-19 17:00:22,567] loop.py:45 INFO TRAIN_STEP step=12730 loss=1.258390 Training: 71%|███████ | 12730/18018 [27:30:08<11:19:43, 7.71s/step, loss=1.258, acc=0.708, acc_len=1.71, thru=15.7, I=18.7, T=37.5, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:22,845] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:00:30,206] loop.py:45 INFO TRAIN_STEP step=12731 loss=1.354039 Training: 71%|███████ | 12731/18018 [27:30:16<11:17:39, 7.69s/step, loss=1.354, acc=0.690, acc_len=1.64, thru=14.8, I=17.9, T=35.9, wait=4.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:00:30,202] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12730: forward=2527.2ms backward=973.0ms (TrainerActor pid=9370) [2026-10-19 17:00:30,259] eagle_store.py:517 WARNING batch_get_buffer missing keys (k407398-9d90ea64_hs, k407398-9d90ea64_ids, k407398-9d90ea64_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:36,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 17:00:30,341] eagle_store.py:517 WARNING batch_get_buffer missing keys (k407399-91262f9e_hs, k407399-91262f9e_ids, k407399-91262f9e_lhs); sleeping 0.50s. [2026-10-19 17:00:37,994] loop.py:45 INFO TRAIN_STEP step=12732 loss=1.237994 Training: 71%|███████ | 12732/18018 [27:30:24<11:20:09, 7.72s/step, loss=1.238, acc=0.716, acc_len=1.74, thru=18.4, I=18.2, T=37.7, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:38,111] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:43,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:00:45,164] loop.py:45 INFO TRAIN_STEP step=12733 loss=1.285356 Training: 71%|███████ | 12733/18018 [27:30:31<11:05:28, 7.56s/step, loss=1.285, acc=0.704, acc_len=1.70, thru=17.2, I=22.3, T=35.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:45,328] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:50,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:00:51,077] loop.py:45 INFO TRAIN_STEP step=12734 loss=1.221948 Training: 71%|███████ | 12734/18018 [27:30:37<10:21:56, 7.06s/step, loss=1.222, acc=0.717, acc_len=1.74, thru=20.5, I=26.7, T=37.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:00:51,289] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:01:00,574] loop.py:380 INFO TIMING step=12735: step=3.672s data=1.600s compute=3.635s [fwd=2.602s bwd=1.019s opt=0.003s] dispatch=5.815s [2026-10-19 17:01:00,574] loop.py:45 INFO TRAIN_STEP step=12735 loss=1.228951 Training: 71%|███████ | 12735/18018 [27:30:46<11:26:07, 7.79s/step, loss=1.229, acc=0.712, acc_len=1.72, thru=13.1, I=13.1, T=34.9, wait=5.8s, pool=61, epoch=2/2][2026-10-19 17:01:10,946] loop.py:45 INFO TRAIN_STEP step=12736 loss=1.226692 Training: 71%|███████ | 12736/18018 [27:30:57<12:34:07, 8.57s/step, loss=1.227, acc=0.716, acc_len=1.73, thru=12.1, I=11.6, T=32.8, wait=6.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:01:10,942] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12735: forward=2744.6ms backward=1115.4ms [2026-10-19 17:01:18,231] loop.py:45 INFO TRAIN_STEP step=12737 loss=1.241536 Training: 71%|███████ | 12737/18018 [27:31:04<12:00:08, 8.18s/step, loss=1.242, acc=0.712, acc_len=1.71, thru=17.8, I=17.1, T=36.7, wait=3.8s, pool=61, epoch=2/2][2026-10-19 17:01:24,410] loop.py:45 INFO TRAIN_STEP step=12738 loss=1.204334 Training: 71%|███████ | 12738/18018 [27:31:10<11:07:06, 7.58s/step, loss=1.204, acc=0.722, acc_len=1.76, thru=19.4, I=19.3, T=36.0, wait=2.6s, pool=57, epoch=2/2][2026-10-19 17:01:31,880] loop.py:45 INFO TRAIN_STEP step=12739 loss=1.324825 Training: 71%|███████ | 12739/18018 [27:31:17<11:04:07, 7.55s/step, loss=1.325, acc=0.695, acc_len=1.66, thru=16.9, I=17.2, T=36.4, wait=3.9s, pool=57, epoch=2/2][2026-10-19 17:01:43,035] loop.py:380 INFO TIMING step=12740: step=4.117s data=1.885s compute=4.057s [fwd=2.866s bwd=1.176s opt=0.003s] dispatch=7.027s [2026-10-19 17:01:43,035] loop.py:45 INFO TRAIN_STEP step=12740 loss=1.303008 Training: 71%|███████ | 12740/18018 [27:31:29<12:39:11, 8.63s/step, loss=1.303, acc=0.699, acc_len=1.67, thru=10.3, I=11.6, T=31.1, wait=7.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:01:50,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:01:51,589] loop.py:45 INFO TRAIN_STEP step=12741 loss=1.303239 Training: 71%|███████ | 12741/18018 [27:31:37<12:36:58, 8.61s/step, loss=1.303, acc=0.703, acc_len=1.69, thru=18.2, I=15.7, T=32.3, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:01:51,585] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12740: forward=2810.8ms backward=1117.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:01:51,863] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:01:57,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:01:58,083] loop.py:45 INFO TRAIN_STEP step=12742 loss=1.115993 Training: 71%|███████ | 12742/18018 [27:31:44<11:41:07, 7.97s/step, loss=1.116, acc=0.733, acc_len=1.80, thru=19.1, I=20.9, T=36.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:01:58,501] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:02:05,851] loop.py:45 INFO TRAIN_STEP step=12743 loss=1.255299 Training: 71%|███████ | 12743/18018 [27:31:51<11:35:35, 7.91s/step, loss=1.255, acc=0.711, acc_len=1.71, thru=16.3, I=17.4, T=36.9, wait=4.3s, pool=61, epoch=2/2][2026-10-19 17:02:12,538] loop.py:45 INFO TRAIN_STEP step=12744 loss=1.392265 Training: 71%|███████ | 12744/18018 [27:31:58<11:03:05, 7.54s/step, loss=1.392, acc=0.684, acc_len=1.62, thru=18.0, I=16.2, T=35.8, wait=3.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:02:18,451] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:02:18,983] loop.py:380 INFO TIMING step=12745: step=3.491s data=1.474s compute=3.454s [fwd=2.448s bwd=0.991s opt=0.003s] dispatch=2.947s [2026-10-19 17:02:18,983] loop.py:45 INFO TRAIN_STEP step=12745 loss=1.360909 Training: 71%|███████ | 12745/18018 [27:32:05<10:34:00, 7.21s/step, loss=1.361, acc=0.688, acc_len=1.64, thru=19.7, I=20.1, T=36.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:02:19,458] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:02:25,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:02:26,150] loop.py:45 INFO TRAIN_STEP step=12746 loss=1.067535 Training: 71%|███████ | 12746/18018 [27:32:12<10:32:42, 7.20s/step, loss=1.068, acc=0.743, acc_len=1.83, thru=18.4, I=20.8, T=37.1, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:02:26,146] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12745: forward=2460.7ms backward=945.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:02:26,231] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 17:02:34,082] loop.py:45 INFO TRAIN_STEP step=12747 loss=1.298967 Training: 71%|███████ | 12747/18018 [27:32:20<10:51:50, 7.42s/step, loss=1.299, acc=0.700, acc_len=1.67, thru=15.8, I=16.9, T=35.2, wait=4.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:02:42,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:02:43,449] loop.py:45 INFO TRAIN_STEP step=12748 loss=1.361724 Training: 71%|███████ | 12748/18018 [27:32:29<11:43:00, 8.00s/step, loss=1.362, acc=0.692, acc_len=1.65, thru=15.7, I=17.0, T=33.4, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:02:43,620] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:02:50,688] loop.py:45 INFO TRAIN_STEP step=12749 loss=1.355435 Training: 71%|███████ | 12749/18018 [27:32:36<11:22:40, 7.77s/step, loss=1.355, acc=0.689, acc_len=1.64, thru=16.9, I=19.6, T=38.2, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:02:58,308] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:02:59,114] loop.py:380 INFO TIMING step=12750: step=3.878s data=1.669s compute=3.837s [fwd=2.781s bwd=1.045s opt=0.003s] dispatch=4.541s [2026-10-19 17:02:59,114] loop.py:45 INFO TRAIN_STEP step=12750 loss=1.244194 Training: 71%|███████ | 12750/18018 [27:32:45<11:39:45, 7.97s/step, loss=1.244, acc=0.708, acc_len=1.70, thru=17.1, I=18.2, T=33.0, wait=4.5s, pool=65, epoch=2/2][2026-10-19 17:02:59,117] loop.py:419 INFO Saving checkpoint at step 12750... (TrainerActor pid=9310) [2026-10-19 17:03:03,225] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0012751 (TrainerActor pid=9310) [2026-10-19 17:03:03,236] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0012001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:03:03,339] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:03:04,365] inference_manager.py:408 INFO Pool capacity available after 6.1s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:03:11,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 17:03:03,227] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0012751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 17:03:04,326] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0012001 [2026-10-19 17:03:12,364] loop.py:45 INFO TRAIN_STEP step=12751 loss=1.236186 Training: 71%|███████ | 12751/18018 [27:32:58<13:58:42, 9.55s/step, loss=1.236, acc=0.713, acc_len=1.72, thru=16.0, I=17.2, T=34.8, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:03:12,360] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12750: forward=2535.3ms backward=1077.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:03:12,838] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:03:19,085] loop.py:45 INFO TRAIN_STEP step=12752 loss=1.274719 Training: 71%|███████ | 12752/18018 [27:33:05<12:43:54, 8.70s/step, loss=1.275, acc=0.701, acc_len=1.68, thru=17.3, I=19.3, T=34.5, wait=3.0s, pool=45, epoch=2/2][2026-10-19 17:03:27,976] loop.py:45 INFO TRAIN_STEP step=12753 loss=1.375628 Training: 71%|███████ | 12753/18018 [27:33:14<12:48:43, 8.76s/step, loss=1.376, acc=0.688, acc_len=1.63, thru=14.0, I=13.6, T=32.9, wait=5.0s, pool=45, epoch=2/2][2026-10-19 17:03:35,803] loop.py:45 INFO TRAIN_STEP step=12754 loss=1.273342 Training: 71%|███████ | 12754/18018 [27:33:21<12:24:03, 8.48s/step, loss=1.273, acc=0.705, acc_len=1.70, thru=18.0, I=16.1, T=35.2, wait=4.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:03:42,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:03:43,644] loop.py:380 INFO TIMING step=12755: step=3.696s data=1.546s compute=3.664s [fwd=2.584s bwd=1.065s opt=0.003s] dispatch=4.133s [2026-10-19 17:03:43,644] loop.py:45 INFO TRAIN_STEP step=12755 loss=1.327587 Training: 71%|███████ | 12755/18018 [27:33:29<12:07:02, 8.29s/step, loss=1.328, acc=0.695, acc_len=1.66, thru=16.0, I=17.2, T=34.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:03:43,717] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:03:51,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:03:52,455] loop.py:45 INFO TRAIN_STEP step=12756 loss=1.352995 Training: 71%|███████ | 12756/18018 [27:33:38<12:20:36, 8.44s/step, loss=1.353, acc=0.689, acc_len=1.65, thru=13.9, I=15.8, T=36.3, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:03:52,450] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12755: forward=2494.0ms backward=993.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:03:52,917] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:03:59,588] loop.py:45 INFO TRAIN_STEP step=12757 loss=1.295637 Training: 71%|███████ | 12757/18018 [27:33:45<11:46:03, 8.05s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=17.6, I=17.0, T=35.3, wait=3.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:04:05,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:04:06,887] loop.py:45 INFO TRAIN_STEP step=12758 loss=1.289733 Training: 71%|███████ | 12758/18018 [27:33:52<11:26:00, 7.83s/step, loss=1.290, acc=0.706, acc_len=1.70, thru=18.7, I=20.5, T=35.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:04:07,041] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:04:15,401] loop.py:45 INFO TRAIN_STEP step=12759 loss=1.241463 Training: 71%|███████ | 12759/18018 [27:34:01<11:44:00, 8.03s/step, loss=1.241, acc=0.710, acc_len=1.71, thru=12.2, I=14.1, T=33.3, wait=4.7s, pool=53, epoch=2/2][2026-10-19 17:04:22,862] loop.py:380 INFO TIMING step=12760: step=3.537s data=1.464s compute=3.499s [fwd=2.511s bwd=0.973s opt=0.003s] dispatch=3.915s [2026-10-19 17:04:22,862] loop.py:45 INFO TRAIN_STEP step=12760 loss=1.368251 Training: 71%|███████ | 12760/18018 [27:34:08<11:28:49, 7.86s/step, loss=1.368, acc=0.689, acc_len=1.65, thru=16.0, I=16.1, T=36.2, wait=3.9s, pool=49, epoch=2/2][2026-10-19 17:04:28,788] loop.py:45 INFO TRAIN_STEP step=12761 loss=1.216568 Training: 71%|███████ | 12761/18018 [27:34:14<10:37:54, 7.28s/step, loss=1.217, acc=0.715, acc_len=1.73, thru=20.0, I=20.7, T=37.6, wait=2.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:04:28,783] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12760: forward=2415.0ms backward=950.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:04:34,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:04:34,818] loop.py:45 INFO TRAIN_STEP step=12762 loss=1.216622 Training: 71%|███████ | 12762/18018 [27:34:20<10:04:53, 6.91s/step, loss=1.217, acc=0.714, acc_len=1.73, thru=20.8, I=22.0, T=37.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:04:34,963] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:04:41,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:04:42,351] loop.py:45 INFO TRAIN_STEP step=12763 loss=1.303113 Training: 71%|███████ | 12763/18018 [27:34:28<10:21:18, 7.09s/step, loss=1.303, acc=0.698, acc_len=1.68, thru=17.9, I=20.6, T=35.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:04:42,656] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:04:48,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:04:49,131] loop.py:45 INFO TRAIN_STEP step=12764 loss=1.316881 Training: 71%|███████ | 12764/18018 [27:34:35<10:12:53, 7.00s/step, loss=1.317, acc=0.700, acc_len=1.68, thru=19.5, I=24.2, T=39.2, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:04:49,256] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 17:04:57,294] loop.py:380 INFO TIMING step=12765: step=3.736s data=1.667s compute=3.694s [fwd=2.647s bwd=1.032s opt=0.003s] dispatch=4.419s [2026-10-19 17:04:57,295] loop.py:45 INFO TRAIN_STEP step=12765 loss=1.289049 Training: 71%|███████ | 12765/18018 [27:34:43<10:43:21, 7.35s/step, loss=1.289, acc=0.702, acc_len=1.69, thru=14.1, I=15.7, T=34.3, wait=4.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:04,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:05:04,430] loop.py:45 INFO TRAIN_STEP step=12766 loss=1.307862 Training: 71%|███████ | 12766/18018 [27:34:50<10:37:39, 7.28s/step, loss=1.308, acc=0.701, acc_len=1.68, thru=17.9, I=17.2, T=35.1, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:05:04,425] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12765: forward=2611.9ms backward=982.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:04,852] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:05:11,865] loop.py:45 INFO TRAIN_STEP step=12767 loss=1.288339 Training: 71%|███████ | 12767/18018 [27:34:57<10:41:30, 7.33s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=17.5, I=16.0, T=35.1, wait=3.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:18,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:05:19,528] loop.py:45 INFO TRAIN_STEP step=12768 loss=1.103258 Training: 71%|███████ | 12768/18018 [27:35:05<10:50:09, 7.43s/step, loss=1.103, acc=0.738, acc_len=1.82, thru=19.1, I=18.0, T=35.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:19,806] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:25,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:05:25,889] loop.py:45 INFO TRAIN_STEP step=12769 loss=1.326058 Training: 71%|███████ | 12769/18018 [27:35:11<10:21:54, 7.11s/step, loss=1.326, acc=0.696, acc_len=1.67, thru=19.2, I=23.9, T=31.9, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:26,122] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:32,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:05:32,426] loop.py:380 INFO TIMING step=12770: step=3.641s data=1.570s compute=3.614s [fwd=2.571s bwd=1.029s opt=0.003s] dispatch=2.887s [2026-10-19 17:05:32,426] loop.py:45 INFO TRAIN_STEP step=12770 loss=1.331851 Training: 71%|███████ | 12770/18018 [27:35:18<10:06:45, 6.94s/step, loss=1.332, acc=0.696, acc_len=1.67, thru=18.3, I=19.8, T=35.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:32,722] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:37,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:05:38,526] loop.py:45 INFO TRAIN_STEP step=12771 loss=1.262045 Training: 71%|███████ | 12771/18018 [27:35:24<9:44:41, 6.69s/step, loss=1.262, acc=0.710, acc_len=1.71, thru=20.1, I=21.8, T=32.8, wait=2.2s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:05:38,521] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12770: forward=2953.0ms backward=909.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:38,996] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:05:48,629] loop.py:45 INFO TRAIN_STEP step=12772 loss=1.314996 Training: 71%|███████ | 12772/18018 [27:35:34<11:14:11, 7.71s/step, loss=1.315, acc=0.697, acc_len=1.67, thru=12.9, I=12.4, T=31.1, wait=6.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:54,763] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:05:56,299] loop.py:45 INFO TRAIN_STEP step=12773 loss=1.136841 Training: 71%|███████ | 12773/18018 [27:35:42<11:12:59, 7.70s/step, loss=1.137, acc=0.728, acc_len=1.78, thru=17.4, I=19.6, T=34.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:05:56,773] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:03,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:06:03,627] loop.py:45 INFO TRAIN_STEP step=12774 loss=1.108966 Training: 71%|███████ | 12774/18018 [27:35:49<11:03:11, 7.59s/step, loss=1.109, acc=0.732, acc_len=1.79, thru=16.8, I=21.5, T=37.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:04,006] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:09,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:06:09,967] loop.py:380 INFO TIMING step=12775: step=3.545s data=1.477s compute=3.496s [fwd=2.388s bwd=1.094s opt=0.003s] dispatch=2.786s [2026-10-19 17:06:09,967] loop.py:45 INFO TRAIN_STEP step=12775 loss=1.404607 Training: 71%|███████ | 12775/18018 [27:35:56<10:30:21, 7.21s/step, loss=1.405, acc=0.682, acc_len=1.62, thru=20.1, I=19.2, T=36.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:10,053] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 17:06:16,875] loop.py:45 INFO TRAIN_STEP step=12776 loss=1.241626 Training: 71%|███████ | 12776/18018 [27:36:02<10:22:09, 7.12s/step, loss=1.242, acc=0.710, acc_len=1.71, thru=15.9, I=16.0, T=34.7, wait=3.2s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:06:16,871] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12775: forward=2700.0ms backward=938.6ms [2026-10-19 17:06:24,965] loop.py:45 INFO TRAIN_STEP step=12777 loss=1.223372 Training: 71%|███████ | 12777/18018 [27:36:11<10:47:27, 7.41s/step, loss=1.223, acc=0.713, acc_len=1.73, thru=16.3, I=16.1, T=34.0, wait=4.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:34,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:06:35,846] loop.py:45 INFO TRAIN_STEP step=12778 loss=1.285000 Training: 71%|███████ | 12778/18018 [27:36:21<12:18:15, 8.45s/step, loss=1.285, acc=0.705, acc_len=1.70, thru=14.2, I=14.8, T=30.4, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:36,124] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:06:41,955] loop.py:45 INFO TRAIN_STEP step=12779 loss=1.219354 Training: 71%|███████ | 12779/18018 [27:36:28<11:16:38, 7.75s/step, loss=1.219, acc=0.715, acc_len=1.73, thru=16.4, I=20.1, T=35.6, wait=2.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:48,011] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:06:49,184] loop.py:380 INFO TIMING step=12780: step=3.531s data=1.444s compute=3.500s [fwd=2.481s bwd=1.005s opt=0.003s] dispatch=3.691s [2026-10-19 17:06:49,185] loop.py:45 INFO TRAIN_STEP step=12780 loss=1.182123 Training: 71%|███████ | 12780/18018 [27:36:35<11:02:52, 7.59s/step, loss=1.182, acc=0.720, acc_len=1.76, thru=17.8, I=19.8, T=36.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:49,516] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:55,915] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:06:56,684] loop.py:45 INFO TRAIN_STEP step=12781 loss=1.359504 Training: 71%|███████ | 12781/18018 [27:36:42<11:00:20, 7.57s/step, loss=1.360, acc=0.689, acc_len=1.64, thru=17.0, I=20.8, T=36.0, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:06:56,679] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12780: forward=2504.5ms backward=1010.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:06:56,919] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:07:03,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:07:04,058] loop.py:45 INFO TRAIN_STEP step=12782 loss=1.308374 Training: 71%|███████ | 12782/18018 [27:36:50<10:55:15, 7.51s/step, loss=1.308, acc=0.696, acc_len=1.67, thru=17.4, I=21.6, T=35.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:07:04,095] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:07:10,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:07:11,301] loop.py:45 INFO TRAIN_STEP step=12783 loss=1.240482 Training: 71%|███████ | 12783/18018 [27:36:57<10:48:06, 7.43s/step, loss=1.240, acc=0.710, acc_len=1.71, thru=16.8, I=20.6, T=34.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:07:11,452] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:07:16,770] loop.py:45 INFO TRAIN_STEP step=12784 loss=1.322884 Training: 71%|███████ | 12784/18018 [27:37:02<9:56:44, 6.84s/step, loss=1.323, acc=0.697, acc_len=1.67, thru=20.2, I=22.1, T=35.8, wait=1.9s, pool=61, epoch=2/2] [2026-10-19 17:07:23,444] loop.py:380 INFO TIMING step=12785: step=3.576s data=1.438s compute=3.549s [fwd=2.524s bwd=1.011s opt=0.003s] dispatch=3.089s [2026-10-19 17:07:23,444] loop.py:45 INFO TRAIN_STEP step=12785 loss=1.308679 Training: 71%|███████ | 12785/18018 [27:37:09<9:52:14, 6.79s/step, loss=1.309, acc=0.702, acc_len=1.68, thru=18.8, I=19.8, T=35.8, wait=3.1s, pool=61, epoch=2/2][2026-10-19 17:07:30,980] loop.py:45 INFO TRAIN_STEP step=12786 loss=1.228163 Training: 71%|███████ | 12786/18018 [27:37:17<10:11:37, 7.01s/step, loss=1.228, acc=0.711, acc_len=1.72, thru=16.9, I=16.2, T=36.3, wait=4.0s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:07:30,976] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12785: forward=2482.7ms backward=992.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:07:38,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:07:38,279] loop.py:45 INFO TRAIN_STEP step=12787 loss=1.492837 Training: 71%|███████ | 12787/18018 [27:37:24<10:18:57, 7.10s/step, loss=1.493, acc=0.667, acc_len=1.56, thru=18.1, I=17.4, T=35.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:07:38,660] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:07:46,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:07:46,698] loop.py:45 INFO TRAIN_STEP step=12788 loss=1.385416 Training: 71%|███████ | 12788/18018 [27:37:32<10:53:21, 7.50s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=15.4, I=17.0, T=34.8, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:07:46,898] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:07:54,907] loop.py:45 INFO TRAIN_STEP step=12789 loss=1.187842 Training: 71%|███████ | 12789/18018 [27:37:40<11:11:53, 7.71s/step, loss=1.188, acc=0.721, acc_len=1.75, thru=13.6, I=15.8, T=33.2, wait=4.3s, pool=49, epoch=2/2][2026-10-19 17:08:04,083] loop.py:380 INFO TIMING step=12790: step=4.075s data=1.808s compute=4.045s [fwd=2.778s bwd=1.253s opt=0.003s] dispatch=5.093s [2026-10-19 17:08:04,084] loop.py:45 INFO TRAIN_STEP step=12790 loss=1.247327 Training: 71%|███████ | 12790/18018 [27:37:50<11:50:05, 8.15s/step, loss=1.247, acc=0.708, acc_len=1.70, thru=13.1, I=11.6, T=31.4, wait=5.1s, pool=41, epoch=2/2][2026-10-19 17:08:13,837] loop.py:45 INFO TRAIN_STEP step=12791 loss=1.350672 Training: 71%|███████ | 12791/18018 [27:37:59<12:31:52, 8.63s/step, loss=1.351, acc=0.691, acc_len=1.65, thru=13.1, I=13.4, T=32.6, wait=5.8s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:08:13,833] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12790: forward=2771.6ms backward=1097.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:21,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:08:22,174] loop.py:45 INFO TRAIN_STEP step=12792 loss=1.309374 Training: 71%|███████ | 12792/18018 [27:38:08<12:24:04, 8.54s/step, loss=1.309, acc=0.697, acc_len=1.67, thru=18.3, I=16.0, T=34.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:22,499] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:26,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:08:27,758] loop.py:45 INFO TRAIN_STEP step=12793 loss=1.187150 Training: 71%|███████ | 12793/18018 [27:38:13<11:06:38, 7.66s/step, loss=1.187, acc=0.719, acc_len=1.74, thru=21.2, I=25.0, T=38.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:28,114] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:35,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:08:36,093] loop.py:45 INFO TRAIN_STEP step=12794 loss=1.253402 Training: 71%|███████ | 12794/18018 [27:38:22<11:24:15, 7.86s/step, loss=1.253, acc=0.712, acc_len=1.72, thru=14.6, I=17.0, T=36.0, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:36,137] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:41,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:08:41,828] loop.py:380 INFO TIMING step=12795: step=3.646s data=1.620s compute=3.612s [fwd=2.575s bwd=1.023s opt=0.003s] dispatch=2.080s [2026-10-19 17:08:41,829] loop.py:45 INFO TRAIN_STEP step=12795 loss=1.249117 Training: 71%|███████ | 12795/18018 [27:38:27<10:28:40, 7.22s/step, loss=1.249, acc=0.709, acc_len=1.72, thru=21.3, I=23.7, T=35.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:42,224] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:49,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:08:50,092] loop.py:45 INFO TRAIN_STEP step=12796 loss=1.209889 Training: 71%|███████ | 12796/18018 [27:38:36<10:55:44, 7.53s/step, loss=1.210, acc=0.718, acc_len=1.73, thru=15.4, I=18.2, T=33.0, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:08:50,088] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12795: forward=2753.4ms backward=1078.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:08:50,419] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:08:56,091] loop.py:45 INFO TRAIN_STEP step=12797 loss=1.319846 Training: 71%|███████ | 12797/18018 [27:38:42<10:15:32, 7.07s/step, loss=1.320, acc=0.695, acc_len=1.66, thru=16.8, I=21.0, T=39.0, wait=2.7s, pool=33, epoch=2/2][2026-10-19 17:09:09,113] loop.py:45 INFO TRAIN_STEP step=12798 loss=1.281783 Training: 71%|███████ | 12798/18018 [27:38:55<12:50:38, 8.86s/step, loss=1.282, acc=0.707, acc_len=1.70, thru=11.5, I=10.1, T=32.0, wait=9.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:16,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:09:17,474] loop.py:45 INFO TRAIN_STEP step=12799 loss=1.388099 Training: 71%|███████ | 12799/18018 [27:39:03<12:37:35, 8.71s/step, loss=1.388, acc=0.684, acc_len=1.62, thru=18.6, I=16.3, T=35.3, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:17,973] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:09:24,420] loop.py:380 INFO TIMING step=12800: step=3.376s data=1.356s compute=3.345s [fwd=2.394s bwd=0.936s opt=0.003s] dispatch=3.560s [2026-10-19 17:09:24,420] loop.py:45 INFO TRAIN_STEP step=12800 loss=1.319411 Training: 71%|███████ | 12800/18018 [27:39:10<11:51:24, 8.18s/step, loss=1.319, acc=0.696, acc_len=1.66, thru=18.3, I=20.4, T=37.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:24,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:24,866] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:30,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:09:31,865] loop.py:45 INFO TRAIN_STEP step=12801 loss=1.376801 Training: 71%|███████ | 12801/18018 [27:39:17<11:32:05, 7.96s/step, loss=1.377, acc=0.685, acc_len=1.63, thru=17.6, I=21.4, T=33.4, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:09:31,861] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12800: forward=2706.6ms backward=1080.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:32,057] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 17:09:39,422] loop.py:45 INFO TRAIN_STEP step=12802 loss=1.255972 Training: 71%|███████ | 12802/18018 [27:39:25<11:21:30, 7.84s/step, loss=1.256, acc=0.709, acc_len=1.70, thru=12.9, I=19.8, T=35.9, wait=4.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:46,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:09:48,058] loop.py:45 INFO TRAIN_STEP step=12803 loss=1.322979 Training: 71%|███████ | 12803/18018 [27:39:34<11:42:08, 8.08s/step, loss=1.323, acc=0.695, acc_len=1.66, thru=17.6, I=17.3, T=34.9, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:48,233] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:53,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:09:53,954] loop.py:45 INFO TRAIN_STEP step=12804 loss=1.333417 Training: 71%|███████ | 12804/18018 [27:39:40<10:45:05, 7.42s/step, loss=1.333, acc=0.695, acc_len=1.66, thru=20.9, I=25.1, T=38.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:09:54,379] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:10:03,135] loop.py:380 INFO TIMING step=12805: step=3.672s data=1.506s compute=3.646s [fwd=2.601s bwd=1.030s opt=0.003s] dispatch=5.498s [2026-10-19 17:10:03,135] loop.py:45 INFO TRAIN_STEP step=12805 loss=1.218921 Training: 71%|███████ | 12805/18018 [27:39:49<11:30:51, 7.95s/step, loss=1.219, acc=0.713, acc_len=1.73, thru=13.1, I=13.9, T=34.9, wait=5.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:10:08,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:10:08,953] loop.py:45 INFO TRAIN_STEP step=12806 loss=1.270732 Training: 71%|███████ | 12806/18018 [27:39:55<10:35:00, 7.31s/step, loss=1.271, acc=0.706, acc_len=1.70, thru=20.1, I=20.8, T=37.2, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:10:08,948] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12805: forward=2432.8ms backward=962.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:10:09,153] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:10:15,099] loop.py:45 INFO TRAIN_STEP step=12807 loss=1.331248 Training: 71%|███████ | 12807/18018 [27:40:01<10:04:31, 6.96s/step, loss=1.331, acc=0.695, acc_len=1.66, thru=18.0, I=20.0, T=35.1, wait=2.5s, pool=37, epoch=2/2][2026-10-19 17:10:22,888] loop.py:45 INFO TRAIN_STEP step=12808 loss=1.340021 Training: 71%|███████ | 12808/18018 [27:40:08<10:26:04, 7.21s/step, loss=1.340, acc=0.692, acc_len=1.65, thru=18.4, I=18.2, T=35.7, wait=4.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:10:31,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:10:32,011] loop.py:45 INFO TRAIN_STEP step=12809 loss=1.300436 Training: 71%|███████ | 12809/18018 [27:40:18<11:15:46, 7.78s/step, loss=1.300, acc=0.698, acc_len=1.67, thru=15.7, I=14.7, T=35.3, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:10:32,048] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 17:10:39,681] loop.py:380 INFO TIMING step=12810: step=3.663s data=1.597s compute=3.634s [fwd=2.626s bwd=0.993s opt=0.003s] dispatch=3.995s [2026-10-19 17:10:39,681] loop.py:45 INFO TRAIN_STEP step=12810 loss=1.356637 Training: 71%|███████ | 12810/18018 [27:40:25<11:12:42, 7.75s/step, loss=1.357, acc=0.693, acc_len=1.66, thru=16.2, I=19.0, T=34.9, wait=4.0s, pool=61, epoch=2/2][2026-10-19 17:10:47,717] loop.py:45 INFO TRAIN_STEP step=12811 loss=1.329349 Training: 71%|███████ | 12811/18018 [27:40:33<11:19:56, 7.83s/step, loss=1.329, acc=0.696, acc_len=1.66, thru=12.7, I=14.4, T=35.0, wait=4.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:10:47,713] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12810: forward=2580.2ms backward=1030.6ms [2026-10-19 17:10:57,962] loop.py:45 INFO TRAIN_STEP step=12812 loss=1.431929 Training: 71%|███████ | 12812/18018 [27:40:44<12:22:30, 8.56s/step, loss=1.432, acc=0.678, acc_len=1.61, thru=14.7, I=13.0, T=35.5, wait=6.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:11:04,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:11:05,049] loop.py:45 INFO TRAIN_STEP step=12813 loss=1.188674 Training: 71%|███████ | 12813/18018 [27:40:51<11:44:10, 8.12s/step, loss=1.189, acc=0.722, acc_len=1.76, thru=19.0, I=16.9, T=35.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:11:05,306] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:11:10,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:11:11,282] loop.py:45 INFO TRAIN_STEP step=12814 loss=1.290944 Training: 71%|███████ | 12814/18018 [27:40:57<10:54:57, 7.55s/step, loss=1.291, acc=0.703, acc_len=1.69, thru=19.8, I=20.9, T=37.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:11:11,617] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:11:19,860] loop.py:380 INFO TIMING step=12815: step=3.485s data=1.442s compute=3.449s [fwd=2.473s bwd=0.961s opt=0.003s] dispatch=5.084s [2026-10-19 17:11:19,860] loop.py:45 INFO TRAIN_STEP step=12815 loss=1.403908 Training: 71%|███████ | 12815/18018 [27:41:05<11:21:31, 7.86s/step, loss=1.404, acc=0.684, acc_len=1.62, thru=14.0, I=15.3, T=36.7, wait=5.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:11:27,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:11:28,201] loop.py:45 INFO TRAIN_STEP step=12816 loss=1.311291 Training: 71%|███████ | 12816/18018 [27:41:14<11:33:59, 8.00s/step, loss=1.311, acc=0.698, acc_len=1.68, thru=17.4, I=18.8, T=35.7, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:11:28,196] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12815: forward=2538.5ms backward=991.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:11:28,626] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:11:35,314] loop.py:45 INFO TRAIN_STEP step=12817 loss=1.203096 Training: 71%|███████ | 12817/18018 [27:41:21<11:10:42, 7.74s/step, loss=1.203, acc=0.716, acc_len=1.74, thru=16.4, I=19.3, T=36.0, wait=3.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:11:42,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:11:44,106] loop.py:45 INFO TRAIN_STEP step=12818 loss=1.198858 Training: 71%|███████ | 12818/18018 [27:41:30<11:37:56, 8.05s/step, loss=1.199, acc=0.716, acc_len=1.73, thru=16.5, I=17.7, T=35.5, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:11:44,416] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:11:50,528] loop.py:45 INFO TRAIN_STEP step=12819 loss=1.267857 Training: 71%|███████ | 12819/18018 [27:41:36<10:55:24, 7.56s/step, loss=1.268, acc=0.703, acc_len=1.69, thru=17.3, I=20.2, T=37.8, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:12:00,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:12:00,874] loop.py:380 INFO TIMING step=12820: step=3.932s data=1.712s compute=3.905s [fwd=2.827s bwd=1.064s opt=0.003s] dispatch=6.404s [2026-10-19 17:12:00,875] loop.py:45 INFO TRAIN_STEP step=12820 loss=1.250408 Training: 71%|███████ | 12820/18018 [27:41:46<12:07:37, 8.40s/step, loss=1.250, acc=0.710, acc_len=1.72, thru=14.4, I=14.6, T=32.6, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:12:01,219] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:12:07,818] loop.py:45 INFO TRAIN_STEP step=12821 loss=1.260876 Training: 71%|███████ | 12821/18018 [27:41:53<11:29:36, 7.96s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=16.3, I=17.3, T=37.5, wait=3.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:12:07,814] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12820: forward=2425.6ms backward=937.5ms [2026-10-19 17:12:14,338] loop.py:45 INFO TRAIN_STEP step=12822 loss=1.293588 Training: 71%|███████ | 12822/18018 [27:42:00<10:52:01, 7.53s/step, loss=1.294, acc=0.700, acc_len=1.69, thru=19.3, I=20.6, T=37.4, wait=3.1s, pool=61, epoch=2/2][2026-10-19 17:12:23,017] loop.py:45 INFO TRAIN_STEP step=12823 loss=1.285532 Training: 71%|███████ | 12823/18018 [27:42:09<11:21:44, 7.87s/step, loss=1.286, acc=0.703, acc_len=1.69, thru=13.2, I=12.7, T=33.4, wait=4.8s, pool=49, epoch=2/2][2026-10-19 17:12:30,339] loop.py:45 INFO TRAIN_STEP step=12824 loss=1.226397 Training: 71%|███████ | 12824/18018 [27:42:16<11:07:17, 7.71s/step, loss=1.226, acc=0.711, acc_len=1.71, thru=16.9, I=16.3, T=33.9, wait=3.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:12:36,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:12:37,006] loop.py:380 INFO TIMING step=12825: step=3.534s data=1.503s compute=3.496s [fwd=2.464s bwd=1.017s opt=0.003s] dispatch=3.125s [2026-10-19 17:12:37,006] loop.py:45 INFO TRAIN_STEP step=12825 loss=1.303421 Training: 71%|███████ | 12825/18018 [27:42:23<10:40:08, 7.40s/step, loss=1.303, acc=0.701, acc_len=1.69, thru=18.6, I=19.9, T=36.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:12:37,069] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:12:41,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:12:42,523] loop.py:45 INFO TRAIN_STEP step=12826 loss=1.226914 Training: 71%|███████ | 12826/18018 [27:42:28<9:51:12, 6.83s/step, loss=1.227, acc=0.712, acc_len=1.72, thru=22.0, I=26.1, T=38.2, wait=2.2s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:12:42,518] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12825: forward=2364.5ms backward=938.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:12:42,968] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:12:50,727] loop.py:45 INFO TRAIN_STEP step=12827 loss=1.340268 Training: 71%|███████ | 12827/18018 [27:42:36<10:26:45, 7.24s/step, loss=1.340, acc=0.694, acc_len=1.66, thru=13.2, I=17.3, T=37.0, wait=4.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:12:57,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:12:58,907] loop.py:45 INFO TRAIN_STEP step=12828 loss=1.330346 Training: 71%|███████ | 12828/18018 [27:42:44<10:50:53, 7.52s/step, loss=1.330, acc=0.697, acc_len=1.67, thru=16.8, I=18.8, T=33.8, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:12:59,165] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:13:05,598] loop.py:45 INFO TRAIN_STEP step=12829 loss=1.287679 Training: 71%|███████ | 12829/18018 [27:42:51<10:29:06, 7.27s/step, loss=1.288, acc=0.699, acc_len=1.68, thru=16.0, I=19.2, T=35.3, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:13:12,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:13:13,029] loop.py:380 INFO TIMING step=12830: step=3.676s data=1.477s compute=3.652s [fwd=2.652s bwd=0.986s opt=0.003s] dispatch=3.748s [2026-10-19 17:13:13,030] loop.py:45 INFO TRAIN_STEP step=12830 loss=1.259601 Training: 71%|███████ | 12830/18018 [27:42:59<10:33:07, 7.32s/step, loss=1.260, acc=0.705, acc_len=1.69, thru=19.0, I=18.5, T=34.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:13:13,141] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 17:13:21,348] loop.py:45 INFO TRAIN_STEP step=12831 loss=1.282609 Training: 71%|███████ | 12831/18018 [27:43:07<10:58:47, 7.62s/step, loss=1.283, acc=0.704, acc_len=1.70, thru=13.0, I=13.8, T=35.2, wait=4.7s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:13:21,344] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12830: forward=2479.5ms backward=1098.6ms [2026-10-19 17:13:30,503] loop.py:45 INFO TRAIN_STEP step=12832 loss=1.414260 Training: 71%|███████ | 12832/18018 [27:43:16<11:38:25, 8.08s/step, loss=1.414, acc=0.685, acc_len=1.63, thru=16.3, I=14.8, T=35.1, wait=5.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:13:37,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:13:37,567] loop.py:45 INFO TRAIN_STEP step=12833 loss=1.158323 Training: 71%|███████ | 12833/18018 [27:43:23<11:11:57, 7.78s/step, loss=1.158, acc=0.725, acc_len=1.77, thru=19.3, I=18.8, T=36.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:13:37,645] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:13:45,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:13:45,364] loop.py:45 INFO TRAIN_STEP step=12834 loss=1.166327 Training: 71%|███████ | 12834/18018 [27:43:31<11:12:22, 7.78s/step, loss=1.166, acc=0.728, acc_len=1.78, thru=17.3, I=17.9, T=34.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:13:45,771] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:13:51,424] loop.py:380 INFO TIMING step=12835: step=3.495s data=1.449s compute=3.452s [fwd=2.432s bwd=1.006s opt=0.003s] dispatch=2.558s [2026-10-19 17:13:51,425] loop.py:45 INFO TRAIN_STEP step=12835 loss=1.227232 Training: 71%|███████ | 12835/18018 [27:43:37<10:27:38, 7.27s/step, loss=1.227, acc=0.713, acc_len=1.73, thru=18.2, I=18.8, T=36.6, wait=2.6s, pool=49, epoch=2/2][2026-10-19 17:13:58,090] loop.py:45 INFO TRAIN_STEP step=12836 loss=1.361892 Training: 71%|███████ | 12836/18018 [27:43:44<10:11:59, 7.09s/step, loss=1.362, acc=0.691, acc_len=1.64, thru=18.4, I=19.3, T=38.6, wait=3.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:13:58,085] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12835: forward=2381.3ms backward=894.2ms [2026-10-19 17:14:06,545] loop.py:45 INFO TRAIN_STEP step=12837 loss=1.147096 Training: 71%|███████ | 12837/18018 [27:43:52<10:47:18, 7.50s/step, loss=1.147, acc=0.725, acc_len=1.77, thru=14.9, I=14.8, T=37.3, wait=5.0s, pool=53, epoch=2/2][2026-10-19 17:14:13,145] loop.py:45 INFO TRAIN_STEP step=12838 loss=1.163376 Training: 71%|███████▏ | 12838/18018 [27:43:59<10:23:59, 7.23s/step, loss=1.163, acc=0.726, acc_len=1.78, thru=19.0, I=17.3, T=34.4, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:14:19,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:14:20,380] loop.py:45 INFO TRAIN_STEP step=12839 loss=1.344804 Training: 71%|███████▏ | 12839/18018 [27:44:06<10:24:06, 7.23s/step, loss=1.345, acc=0.695, acc_len=1.66, thru=16.2, I=17.3, T=34.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:14:20,717] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:14:27,729] loop.py:380 INFO TIMING step=12840: step=3.548s data=1.532s compute=3.503s [fwd=2.509s bwd=0.980s opt=0.003s] dispatch=3.790s [2026-10-19 17:14:27,729] loop.py:45 INFO TRAIN_STEP step=12840 loss=1.227701 Training: 71%|███████▏ | 12840/18018 [27:44:13<10:27:07, 7.27s/step, loss=1.228, acc=0.714, acc_len=1.72, thru=15.7, I=21.3, T=36.1, wait=3.8s, pool=45, epoch=2/2][2026-10-19 17:14:36,426] loop.py:45 INFO TRAIN_STEP step=12841 loss=1.204092 Training: 71%|███████▏ | 12841/18018 [27:44:22<11:03:50, 7.69s/step, loss=1.204, acc=0.719, acc_len=1.75, thru=14.6, I=12.2, T=32.0, wait=4.7s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:14:36,421] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12840: forward=2780.9ms backward=1167.2ms [2026-10-19 17:14:47,135] loop.py:45 INFO TRAIN_STEP step=12842 loss=1.291492 Training: 71%|███████▏ | 12842/18018 [27:44:33<12:21:50, 8.60s/step, loss=1.291, acc=0.701, acc_len=1.69, thru=13.5, I=13.8, T=30.0, wait=6.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:14:54,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:14:54,320] loop.py:45 INFO TRAIN_STEP step=12843 loss=1.278155 Training: 71%|███████▏ | 12843/18018 [27:44:40<11:45:04, 8.17s/step, loss=1.278, acc=0.704, acc_len=1.70, thru=16.7, I=18.5, T=35.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:14:54,532] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:15:00,943] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:15:01,728] loop.py:45 INFO TRAIN_STEP step=12844 loss=1.276676 Training: 71%|███████▏ | 12844/18018 [27:44:47<11:25:07, 7.94s/step, loss=1.277, acc=0.702, acc_len=1.69, thru=16.5, I=18.5, T=36.3, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:15:01,947] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:15:06,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:15:07,929] loop.py:380 INFO TIMING step=12845: step=3.444s data=1.390s compute=3.407s [fwd=2.452s bwd=0.940s opt=0.003s] dispatch=2.746s [2026-10-19 17:15:07,929] loop.py:45 INFO TRAIN_STEP step=12845 loss=1.319128 Training: 71%|███████▏ | 12845/18018 [27:44:54<10:39:53, 7.42s/step, loss=1.319, acc=0.695, acc_len=1.66, thru=20.4, I=24.0, T=37.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:15:07,969] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:15:13,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:15:13,861] loop.py:45 INFO TRAIN_STEP step=12846 loss=1.193626 Training: 71%|███████▏ | 12846/18018 [27:44:59<10:01:11, 6.97s/step, loss=1.194, acc=0.720, acc_len=1.75, thru=21.2, I=24.0, T=38.2, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:15:13,856] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12845: forward=2400.7ms backward=910.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:15:13,961] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 17:15:19,823] loop.py:45 INFO TRAIN_STEP step=12847 loss=1.204234 Training: 71%|███████▏ | 12847/18018 [27:45:05<9:34:53, 6.67s/step, loss=1.204, acc=0.719, acc_len=1.75, thru=18.2, I=22.4, T=37.2, wait=2.5s, pool=37, epoch=2/2] [2026-10-19 17:15:30,543] loop.py:45 INFO TRAIN_STEP step=12848 loss=1.186479 Training: 71%|███████▏ | 12848/18018 [27:45:16<11:19:27, 7.89s/step, loss=1.186, acc=0.720, acc_len=1.75, thru=11.9, I=9.8, T=33.5, wait=6.9s, pool=33, epoch=2/2][2026-10-19 17:15:39,486] loop.py:45 INFO TRAIN_STEP step=12849 loss=1.386903 Training: 71%|███████▏ | 12849/18018 [27:45:25<11:46:41, 8.20s/step, loss=1.387, acc=0.682, acc_len=1.61, thru=17.1, I=13.8, T=32.1, wait=5.0s, pool=53, epoch=2/2][2026-10-19 17:15:46,546] loop.py:380 INFO TIMING step=12850: step=3.695s data=1.565s compute=3.655s [fwd=2.585s bwd=1.056s opt=0.003s] dispatch=3.356s [2026-10-19 17:15:46,546] loop.py:45 INFO TRAIN_STEP step=12850 loss=1.211589 Training: 71%|███████▏ | 12850/18018 [27:45:32<11:16:59, 7.86s/step, loss=1.212, acc=0.713, acc_len=1.72, thru=17.8, I=18.4, T=34.6, wait=3.4s, pool=61, epoch=2/2][2026-10-19 17:15:52,870] loop.py:45 INFO TRAIN_STEP step=12851 loss=1.253001 Training: 71%|███████▏ | 12851/18018 [27:45:38<10:37:10, 7.40s/step, loss=1.253, acc=0.702, acc_len=1.68, thru=17.0, I=18.2, T=36.4, wait=2.8s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:15:52,866] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12850: forward=2444.5ms backward=1016.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:01,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:16:01,875] loop.py:45 INFO TRAIN_STEP step=12852 loss=1.362630 Training: 71%|███████▏ | 12852/18018 [27:45:47<11:18:35, 7.88s/step, loss=1.363, acc=0.685, acc_len=1.63, thru=17.1, I=16.5, T=32.2, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:02,137] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:16:07,766] loop.py:45 INFO TRAIN_STEP step=12853 loss=1.206304 Training: 71%|███████▏ | 12853/18018 [27:45:53<10:27:04, 7.28s/step, loss=1.206, acc=0.717, acc_len=1.74, thru=20.1, I=23.3, T=38.5, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:12,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:16:13,189] loop.py:45 INFO TRAIN_STEP step=12854 loss=1.216568 Training: 71%|███████▏ | 12854/18018 [27:45:59<9:38:48, 6.73s/step, loss=1.217, acc=0.714, acc_len=1.73, thru=23.0, I=24.9, T=38.1, wait=2.1s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:13,348] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:19,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:16:20,136] loop.py:380 INFO TIMING step=12855: step=3.650s data=1.559s compute=3.624s [fwd=2.555s bwd=1.054s opt=0.003s] dispatch=3.289s [2026-10-19 17:16:20,137] loop.py:45 INFO TRAIN_STEP step=12855 loss=1.239814 Training: 71%|███████▏ | 12855/18018 [27:46:06<9:44:26, 6.79s/step, loss=1.240, acc=0.710, acc_len=1.71, thru=18.5, I=22.7, T=35.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:20,519] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:16:26,831] loop.py:45 INFO TRAIN_STEP step=12856 loss=1.372022 Training: 71%|███████▏ | 12856/18018 [27:46:12<9:41:49, 6.76s/step, loss=1.372, acc=0.690, acc_len=1.65, thru=18.2, I=21.1, T=37.5, wait=3.3s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:16:26,827] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12855: forward=2396.6ms backward=962.5ms [2026-10-19 17:16:35,314] loop.py:45 INFO TRAIN_STEP step=12857 loss=1.209810 Training: 71%|███████▏ | 12857/18018 [27:46:21<10:26:12, 7.28s/step, loss=1.210, acc=0.718, acc_len=1.74, thru=16.0, I=18.5, T=33.1, wait=4.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:40,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:16:41,602] loop.py:45 INFO TRAIN_STEP step=12858 loss=1.186441 Training: 71%|███████▏ | 12858/18018 [27:46:27<10:00:26, 6.98s/step, loss=1.186, acc=0.721, acc_len=1.75, thru=18.7, I=22.1, T=35.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:41,668] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:46,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:16:47,753] loop.py:45 INFO TRAIN_STEP step=12859 loss=1.379460 Training: 71%|███████▏ | 12859/18018 [27:46:33<9:38:53, 6.73s/step, loss=1.379, acc=0.683, acc_len=1.62, thru=20.4, I=26.4, T=36.8, wait=2.7s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:48,142] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:54,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:16:54,927] loop.py:380 INFO TIMING step=12860: step=3.706s data=1.588s compute=3.671s [fwd=2.647s bwd=1.010s opt=0.003s] dispatch=3.458s [2026-10-19 17:16:54,927] loop.py:45 INFO TRAIN_STEP step=12860 loss=1.234486 Training: 71%|███████▏ | 12860/18018 [27:46:41<9:50:06, 6.86s/step, loss=1.234, acc=0.711, acc_len=1.71, thru=17.8, I=20.2, T=34.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:16:55,290] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:00,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:17:02,138] loop.py:45 INFO TRAIN_STEP step=12861 loss=1.314660 Training: 71%|███████▏ | 12861/18018 [27:46:48<9:58:57, 6.97s/step, loss=1.315, acc=0.692, acc_len=1.65, thru=18.7, I=19.4, T=36.0, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:17:02,134] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12860: forward=2517.1ms backward=964.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:02,350] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:06,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:17:07,852] loop.py:45 INFO TRAIN_STEP step=12862 loss=1.204530 Training: 71%|███████▏ | 12862/18018 [27:46:53<9:26:28, 6.59s/step, loss=1.205, acc=0.714, acc_len=1.73, thru=21.6, I=27.9, T=38.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:08,053] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:14,155] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:17:15,241] loop.py:45 INFO TRAIN_STEP step=12863 loss=1.154398 Training: 71%|███████▏ | 12863/18018 [27:47:01<9:46:54, 6.83s/step, loss=1.154, acc=0.724, acc_len=1.76, thru=17.4, I=22.6, T=35.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:15,662] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:22,583] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:17:23,815] loop.py:45 INFO TRAIN_STEP step=12864 loss=1.369154 Training: 71%|███████▏ | 12864/18018 [27:47:09<10:31:42, 7.35s/step, loss=1.369, acc=0.691, acc_len=1.65, thru=13.8, I=17.9, T=34.2, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:24,091] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:30,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:17:31,479] loop.py:380 INFO TIMING step=12865: step=3.470s data=1.397s compute=3.435s [fwd=2.424s bwd=0.996s opt=0.003s] dispatch=4.184s [2026-10-19 17:17:31,479] loop.py:45 INFO TRAIN_STEP step=12865 loss=1.247853 Training: 71%|███████▏ | 12865/18018 [27:47:17<10:39:34, 7.45s/step, loss=1.248, acc=0.707, acc_len=1.70, thru=16.6, I=20.6, T=36.9, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:31,832] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:17:37,302] loop.py:45 INFO TRAIN_STEP step=12866 loss=1.216617 Training: 71%|███████▏ | 12866/18018 [27:47:23<9:57:38, 6.96s/step, loss=1.217, acc=0.715, acc_len=1.74, thru=20.1, I=22.6, T=37.4, wait=2.4s, pool=57, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:17:37,298] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12865: forward=2415.1ms backward=956.1ms [2026-10-19 17:17:45,395] loop.py:45 INFO TRAIN_STEP step=12867 loss=1.301428 Training: 71%|███████▏ | 12867/18018 [27:47:31<10:26:38, 7.30s/step, loss=1.301, acc=0.702, acc_len=1.69, thru=16.1, I=15.8, T=35.4, wait=4.5s, pool=61, epoch=2/2][2026-10-19 17:17:54,844] loop.py:45 INFO TRAIN_STEP step=12868 loss=1.312400 Training: 71%|███████▏ | 12868/18018 [27:47:40<11:21:53, 7.94s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=14.9, I=14.6, T=33.6, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:54,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:17:55,312] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:18:00,562] loop.py:45 INFO TRAIN_STEP step=12869 loss=1.351921 Training: 71%|███████▏ | 12869/18018 [27:47:46<10:24:27, 7.28s/step, loss=1.352, acc=0.691, acc_len=1.65, thru=17.3, I=19.9, T=38.6, wait=2.4s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:18:06,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:18:07,276] loop.py:380 INFO TIMING step=12870: step=3.533s data=1.500s compute=3.493s [fwd=2.447s bwd=1.032s opt=0.003s] dispatch=3.173s [2026-10-19 17:18:07,277] loop.py:45 INFO TRAIN_STEP step=12870 loss=1.296197 Training: 71%|███████▏ | 12870/18018 [27:47:53<10:09:51, 7.11s/step, loss=1.296, acc=0.702, acc_len=1.68, thru=19.7, I=21.1, T=36.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:18:07,671] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:18:13,756] loop.py:45 INFO TRAIN_STEP step=12871 loss=1.251629 Training: 71%|███████▏ | 12871/18018 [27:47:59<9:53:37, 6.92s/step, loss=1.252, acc=0.710, acc_len=1.71, thru=18.8, I=21.1, T=38.0, wait=3.1s, pool=57, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:18:13,751] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12870: forward=2381.1ms backward=940.4ms [2026-10-19 17:18:19,784] loop.py:45 INFO TRAIN_STEP step=12872 loss=1.303020 Training: 71%|███████▏ | 12872/18018 [27:48:05<9:30:33, 6.65s/step, loss=1.303, acc=0.701, acc_len=1.69, thru=20.4, I=20.4, T=37.3, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:18:24,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:18:25,454] loop.py:45 INFO TRAIN_STEP step=12873 loss=1.336666 Training: 71%|███████▏ | 12873/18018 [27:48:11<9:05:09, 6.36s/step, loss=1.337, acc=0.690, acc_len=1.65, thru=21.8, I=23.7, T=39.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:18:25,755] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:18:31,547] loop.py:45 INFO TRAIN_STEP step=12874 loss=1.282817 Training: 71%|███████▏ | 12874/18018 [27:48:17<8:58:16, 6.28s/step, loss=1.283, acc=0.704, acc_len=1.69, thru=18.1, I=22.0, T=38.8, wait=2.8s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:18:42,403] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:18:42,815] loop.py:380 INFO TIMING step=12875: step=4.493s data=1.591s compute=4.463s [fwd=3.389s bwd=1.060s opt=0.003s] dispatch=6.764s [2026-10-19 17:18:42,815] loop.py:45 INFO TRAIN_STEP step=12875 loss=1.211511 Training: 71%|███████▏ | 12875/18018 [27:48:28<11:06:25, 7.77s/step, loss=1.212, acc=0.714, acc_len=1.73, thru=13.2, I=14.3, T=28.5, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:18:42,920] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:18:48,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:18:49,606] loop.py:45 INFO TRAIN_STEP step=12876 loss=1.355193 Training: 71%|███████▏ | 12876/18018 [27:48:35<10:41:00, 7.48s/step, loss=1.355, acc=0.690, acc_len=1.64, thru=16.7, I=18.2, T=33.1, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:18:49,602] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12875: forward=2566.0ms backward=1161.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:18:49,942] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:18:57,666] loop.py:45 INFO TRAIN_STEP step=12877 loss=1.217207 Training: 71%|███████▏ | 12877/18018 [27:48:43<10:55:46, 7.65s/step, loss=1.217, acc=0.717, acc_len=1.75, thru=14.2, I=15.2, T=35.6, wait=4.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:03,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:19:04,471] loop.py:45 INFO TRAIN_STEP step=12878 loss=1.321300 Training: 71%|███████▏ | 12878/18018 [27:48:50<10:33:50, 7.40s/step, loss=1.321, acc=0.695, acc_len=1.67, thru=19.3, I=19.4, T=36.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:04,883] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:19:10,363] loop.py:45 INFO TRAIN_STEP step=12879 loss=1.285831 Training: 71%|███████▏ | 12879/18018 [27:48:56<9:55:00, 6.95s/step, loss=1.286, acc=0.700, acc_len=1.68, thru=18.4, I=24.2, T=38.4, wait=2.6s, pool=41, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:18,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:19:18,558] loop.py:380 INFO TIMING step=12880: step=3.728s data=1.490s compute=3.700s [fwd=2.629s bwd=1.058s opt=0.003s] dispatch=4.459s [2026-10-19 17:19:18,558] loop.py:45 INFO TRAIN_STEP step=12880 loss=1.274839 Training: 71%|███████▏ | 12880/18018 [27:49:04<10:26:59, 7.32s/step, loss=1.275, acc=0.707, acc_len=1.71, thru=16.8, I=17.6, T=34.3, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:18,615] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:23,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:19:25,492] loop.py:45 INFO TRAIN_STEP step=12881 loss=1.266488 Training: 71%|███████▏ | 12881/18018 [27:49:11<10:16:54, 7.21s/step, loss=1.266, acc=0.705, acc_len=1.70, thru=18.4, I=21.8, T=31.5, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:19:25,488] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12880: forward=3066.8ms backward=955.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:25,644] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:31,605] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:19:32,454] loop.py:45 INFO TRAIN_STEP step=12882 loss=1.267607 Training: 71%|███████▏ | 12882/18018 [27:49:18<10:10:28, 7.13s/step, loss=1.268, acc=0.702, acc_len=1.69, thru=16.1, I=20.8, T=37.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:32,608] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:37,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:19:39,459] loop.py:45 INFO TRAIN_STEP step=12883 loss=1.296452 Training: 72%|███████▏ | 12883/18018 [27:49:25<10:07:07, 7.09s/step, loss=1.296, acc=0.701, acc_len=1.68, thru=18.5, I=22.2, T=31.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:39,868] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:45,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:19:45,905] loop.py:45 INFO TRAIN_STEP step=12884 loss=1.269288 Training: 72%|███████▏ | 12884/18018 [27:49:31<9:50:25, 6.90s/step, loss=1.269, acc=0.705, acc_len=1.70, thru=18.1, I=22.4, T=38.4, wait=3.1s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:46,195] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:53,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:19:53,557] loop.py:380 INFO TIMING step=12885: step=4.029s data=2.010s compute=3.994s [fwd=3.036s bwd=0.944s opt=0.003s] dispatch=3.612s [2026-10-19 17:19:53,557] loop.py:45 INFO TRAIN_STEP step=12885 loss=1.393805 Training: 72%|███████▏ | 12885/18018 [27:49:39<10:09:34, 7.13s/step, loss=1.394, acc=0.683, acc_len=1.62, thru=17.3, I=18.2, T=31.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:19:53,726] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:20:00,594] loop.py:45 INFO TRAIN_STEP step=12886 loss=1.218406 Training: 72%|███████▏ | 12886/18018 [27:49:46<10:07:09, 7.10s/step, loss=1.218, acc=0.715, acc_len=1.73, thru=15.5, I=15.1, T=37.9, wait=3.6s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:20:00,589] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12885: forward=2389.9ms backward=950.4ms [2026-10-19 17:20:12,122] loop.py:45 INFO TRAIN_STEP step=12887 loss=1.250284 Training: 72%|███████▏ | 12887/18018 [27:49:58<12:00:38, 8.43s/step, loss=1.250, acc=0.707, acc_len=1.70, thru=14.6, I=12.5, T=34.1, wait=7.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:12,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:12,604] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:18,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:20:19,285] loop.py:45 INFO TRAIN_STEP step=12888 loss=1.434732 Training: 72%|███████▏ | 12888/18018 [27:50:05<11:28:06, 8.05s/step, loss=1.435, acc=0.675, acc_len=1.60, thru=18.4, I=18.8, T=35.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:19,451] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:25,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:20:26,220] loop.py:45 INFO TRAIN_STEP step=12889 loss=1.173582 Training: 72%|███████▏ | 12889/18018 [27:50:12<10:59:27, 7.71s/step, loss=1.174, acc=0.724, acc_len=1.76, thru=19.0, I=21.0, T=36.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:26,579] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:34,371] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:20:34,555] loop.py:380 INFO TIMING step=12890: step=3.566s data=1.391s compute=3.538s [fwd=2.502s bwd=1.022s opt=0.003s] dispatch=4.760s [2026-10-19 17:20:34,556] loop.py:45 INFO TRAIN_STEP step=12890 loss=1.365764 Training: 72%|███████▏ | 12890/18018 [27:50:20<11:15:12, 7.90s/step, loss=1.366, acc=0.692, acc_len=1.65, thru=15.2, I=16.5, T=35.9, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:34,873] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:20:42,196] loop.py:45 INFO TRAIN_STEP step=12891 loss=1.304449 Training: 72%|███████▏ | 12891/18018 [27:50:28<11:08:27, 7.82s/step, loss=1.304, acc=0.699, acc_len=1.67, thru=15.7, I=17.5, T=36.0, wait=4.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:20:42,191] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12890: forward=2549.6ms backward=968.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:48,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:20:48,850] loop.py:45 INFO TRAIN_STEP step=12892 loss=1.418620 Training: 72%|███████▏ | 12892/18018 [27:50:34<10:38:20, 7.47s/step, loss=1.419, acc=0.677, acc_len=1.60, thru=18.9, I=19.3, T=36.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:49,267] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:55,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:20:56,420] loop.py:45 INFO TRAIN_STEP step=12893 loss=1.282593 Training: 72%|███████▏ | 12893/18018 [27:50:42<10:40:44, 7.50s/step, loss=1.283, acc=0.703, acc_len=1.69, thru=17.7, I=18.5, T=35.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:20:56,466] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:03,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:21:03,280] loop.py:45 INFO TRAIN_STEP step=12894 loss=1.152549 Training: 72%|███████▏ | 12894/18018 [27:50:49<10:24:11, 7.31s/step, loss=1.153, acc=0.729, acc_len=1.78, thru=18.0, I=19.0, T=37.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:03,506] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:09,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:21:09,864] loop.py:380 INFO TIMING step=12895: step=3.319s data=1.311s compute=3.286s [fwd=2.354s bwd=0.918s opt=0.003s] dispatch=3.257s [2026-10-19 17:21:09,865] loop.py:45 INFO TRAIN_STEP step=12895 loss=1.226015 Training: 72%|███████▏ | 12895/18018 [27:50:55<10:05:30, 7.09s/step, loss=1.226, acc=0.714, acc_len=1.73, thru=19.9, I=21.6, T=38.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:10,003] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:16,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:21:17,219] loop.py:45 INFO TRAIN_STEP step=12896 loss=1.295341 Training: 72%|███████▏ | 12896/18018 [27:51:03<10:12:09, 7.17s/step, loss=1.295, acc=0.702, acc_len=1.69, thru=17.3, I=18.2, T=34.1, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:21:17,215] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12895: forward=2736.0ms backward=968.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:17,397] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:21:23,162] loop.py:45 INFO TRAIN_STEP step=12897 loss=1.205124 Training: 72%|███████▏ | 12897/18018 [27:51:09<9:40:33, 6.80s/step, loss=1.205, acc=0.721, acc_len=1.75, thru=17.7, I=21.3, T=34.0, wait=2.2s, pool=41, epoch=2/2] [2026-10-19 17:21:31,410] loop.py:45 INFO TRAIN_STEP step=12898 loss=1.387327 Training: 72%|███████▏ | 12898/18018 [27:51:17<10:17:31, 7.24s/step, loss=1.387, acc=0.682, acc_len=1.62, thru=16.6, I=16.6, T=34.8, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:38,773] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:21:39,876] loop.py:45 INFO TRAIN_STEP step=12899 loss=1.221834 Training: 72%|███████▏ | 12899/18018 [27:51:25<10:48:48, 7.60s/step, loss=1.222, acc=0.714, acc_len=1.73, thru=14.8, I=15.9, T=34.9, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:40,282] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:44,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:21:45,940] loop.py:380 INFO TIMING step=12900: step=3.562s data=1.478s compute=3.531s [fwd=2.516s bwd=0.999s opt=0.003s] dispatch=2.494s [2026-10-19 17:21:45,940] loop.py:45 INFO TRAIN_STEP step=12900 loss=1.267963 Training: 72%|███████▏ | 12900/18018 [27:51:32<10:09:14, 7.14s/step, loss=1.268, acc=0.709, acc_len=1.71, thru=20.0, I=25.9, T=35.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:46,041] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:51,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:21:52,020] loop.py:45 INFO TRAIN_STEP step=12901 loss=1.316645 Training: 72%|███████▏ | 12901/18018 [27:51:38<9:41:59, 6.82s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=21.4, I=25.4, T=36.5, wait=2.6s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:21:52,016] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12900: forward=2507.6ms backward=962.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:52,413] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:56,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:21:58,611] loop.py:45 INFO TRAIN_STEP step=12902 loss=1.332555 Training: 72%|███████▏ | 12902/18018 [27:51:44<9:35:55, 6.75s/step, loss=1.333, acc=0.695, acc_len=1.66, thru=19.8, I=26.1, T=30.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:21:59,104] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:22:03,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:22:04,856] loop.py:45 INFO TRAIN_STEP step=12903 loss=1.193461 Training: 72%|███████▏ | 12903/18018 [27:51:50<9:22:46, 6.60s/step, loss=1.193, acc=0.723, acc_len=1.76, thru=17.4, I=25.9, T=38.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:22:05,316] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:22:10,701] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:22:11,146] loop.py:45 INFO TRAIN_STEP step=12904 loss=1.304035 Training: 72%|███████▏ | 12904/18018 [27:51:57<9:14:43, 6.51s/step, loss=1.304, acc=0.701, acc_len=1.68, thru=20.9, I=25.3, T=37.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:22:11,205] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 17:22:17,445] loop.py:380 INFO TIMING step=12905: step=3.697s data=1.627s compute=3.666s [fwd=2.691s bwd=0.959s opt=0.003s] dispatch=2.593s [2026-10-19 17:22:17,445] loop.py:45 INFO TRAIN_STEP step=12905 loss=1.178688 Training: 72%|███████▏ | 12905/18018 [27:52:03<9:09:12, 6.44s/step, loss=1.179, acc=0.720, acc_len=1.75, thru=17.3, I=20.8, T=34.6, wait=2.6s, pool=37, epoch=2/2][2026-10-19 17:22:27,818] loop.py:45 INFO TRAIN_STEP step=12906 loss=1.191958 Training: 72%|███████▏ | 12906/18018 [27:52:13<10:49:29, 7.62s/step, loss=1.192, acc=0.720, acc_len=1.75, thru=13.5, I=13.6, T=33.8, wait=6.6s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:22:27,814] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12905: forward=2685.1ms backward=1040.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:22:36,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:22:37,074] loop.py:45 INFO TRAIN_STEP step=12907 loss=1.325723 Training: 72%|███████▏ | 12907/18018 [27:52:23<11:31:05, 8.11s/step, loss=1.326, acc=0.698, acc_len=1.68, thru=16.2, I=15.3, T=34.8, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:22:37,396] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:22:43,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:22:44,468] loop.py:45 INFO TRAIN_STEP step=12908 loss=1.325713 Training: 72%|███████▏ | 12908/18018 [27:52:30<11:12:39, 7.90s/step, loss=1.326, acc=0.692, acc_len=1.66, thru=18.3, I=20.4, T=33.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:22:44,678] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:22:50,984] loop.py:45 INFO TRAIN_STEP step=12909 loss=1.237142 Training: 72%|███████▏ | 12909/18018 [27:52:37<10:37:07, 7.48s/step, loss=1.237, acc=0.712, acc_len=1.71, thru=18.3, I=21.7, T=37.2, wait=3.1s, pool=61, epoch=2/2][2026-10-19 17:22:59,163] loop.py:380 INFO TIMING step=12910: step=3.650s data=1.531s compute=3.581s [fwd=2.476s bwd=1.090s opt=0.003s] dispatch=4.521s [2026-10-19 17:22:59,163] loop.py:45 INFO TRAIN_STEP step=12910 loss=1.219074 Training: 72%|███████▏ | 12910/18018 [27:52:45<10:54:51, 7.69s/step, loss=1.219, acc=0.716, acc_len=1.74, thru=15.6, I=16.6, T=35.1, wait=4.5s, pool=61, epoch=2/2][2026-10-19 17:23:09,127] loop.py:45 INFO TRAIN_STEP step=12911 loss=1.349444 Training: 72%|███████▏ | 12911/18018 [27:52:55<11:52:39, 8.37s/step, loss=1.349, acc=0.695, acc_len=1.66, thru=13.4, I=12.3, T=29.9, wait=5.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:23:09,122] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12910: forward=3031.3ms backward=1196.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:23:15,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:23:15,687] loop.py:45 INFO TRAIN_STEP step=12912 loss=1.205378 Training: 72%|███████▏ | 12912/18018 [27:53:01<11:06:19, 7.83s/step, loss=1.205, acc=0.717, acc_len=1.74, thru=18.2, I=18.1, T=32.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:23:16,182] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:23:22,187] loop.py:45 INFO TRAIN_STEP step=12913 loss=1.241290 Training: 72%|███████▏ | 12913/18018 [27:53:08<10:32:15, 7.43s/step, loss=1.241, acc=0.709, acc_len=1.72, thru=15.7, I=19.1, T=35.9, wait=2.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:23:31,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:23:32,888] loop.py:45 INFO TRAIN_STEP step=12914 loss=1.360083 Training: 72%|███████▏ | 12914/18018 [27:53:18<11:55:34, 8.41s/step, loss=1.360, acc=0.690, acc_len=1.64, thru=14.5, I=15.5, T=32.7, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:23:33,004] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 17:23:38,733] loop.py:380 INFO TIMING step=12915: step=3.442s data=1.443s compute=3.400s [fwd=2.412s bwd=0.973s opt=0.003s] dispatch=2.393s [2026-10-19 17:23:38,733] loop.py:45 INFO TRAIN_STEP step=12915 loss=1.325938 Training: 72%|███████▏ | 12915/18018 [27:53:24<10:49:54, 7.64s/step, loss=1.326, acc=0.695, acc_len=1.67, thru=18.8, I=20.3, T=37.2, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:23:46,019] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:23:46,247] loop.py:45 INFO TRAIN_STEP step=12916 loss=1.273545 Training: 72%|███████▏ | 12916/18018 [27:53:32<10:46:30, 7.60s/step, loss=1.274, acc=0.703, acc_len=1.69, thru=18.0, I=18.2, T=36.6, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:23:46,243] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12915: forward=2466.8ms backward=979.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:23:46,521] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:23:53,907] loop.py:45 INFO TRAIN_STEP step=12917 loss=1.375970 Training: 72%|███████▏ | 12917/18018 [27:53:39<10:47:48, 7.62s/step, loss=1.376, acc=0.686, acc_len=1.64, thru=16.6, I=18.1, T=36.3, wait=4.1s, pool=57, epoch=2/2][2026-10-19 17:24:02,052] loop.py:45 INFO TRAIN_STEP step=12918 loss=1.367034 Training: 72%|███████▏ | 12918/18018 [27:53:48<11:01:04, 7.78s/step, loss=1.367, acc=0.685, acc_len=1.63, thru=15.4, I=15.3, T=32.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:24:02,019] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:24:02,520] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:24:09,111] loop.py:45 INFO TRAIN_STEP step=12919 loss=1.279741 Training: 72%|███████▏ | 12919/18018 [27:53:55<10:42:45, 7.56s/step, loss=1.280, acc=0.703, acc_len=1.69, thru=16.8, I=16.3, T=35.9, wait=3.5s, pool=53, epoch=2/2][2026-10-19 17:24:16,338] loop.py:380 INFO TIMING step=12920: step=3.666s data=1.572s compute=3.628s [fwd=2.549s bwd=1.065s opt=0.003s] dispatch=3.549s [2026-10-19 17:24:16,338] loop.py:45 INFO TRAIN_STEP step=12920 loss=1.242561 Training: 72%|███████▏ | 12920/18018 [27:54:02<10:34:05, 7.46s/step, loss=1.243, acc=0.714, acc_len=1.73, thru=18.5, I=17.3, T=34.9, wait=3.5s, pool=53, epoch=2/2][2026-10-19 17:24:22,844] loop.py:45 INFO TRAIN_STEP step=12921 loss=1.392234 Training: 72%|███████▏ | 12921/18018 [27:54:08<10:09:33, 7.18s/step, loss=1.392, acc=0.687, acc_len=1.63, thru=17.9, I=18.7, T=37.4, wait=3.1s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:24:22,840] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12920: forward=2427.9ms backward=950.2ms [2026-10-19 17:24:31,585] loop.py:45 INFO TRAIN_STEP step=12922 loss=1.347919 Training: 72%|███████▏ | 12922/18018 [27:54:17<10:49:12, 7.64s/step, loss=1.348, acc=0.689, acc_len=1.65, thru=15.4, I=14.7, T=35.6, wait=5.1s, pool=57, epoch=2/2][2026-10-19 17:24:39,269] loop.py:45 INFO TRAIN_STEP step=12923 loss=1.354801 Training: 72%|███████▏ | 12923/18018 [27:54:25<10:50:06, 7.66s/step, loss=1.355, acc=0.693, acc_len=1.66, thru=16.6, I=17.2, T=35.1, wait=4.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:24:46,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:24:48,518] loop.py:45 INFO TRAIN_STEP step=12924 loss=1.264552 Training: 72%|███████▏ | 12924/18018 [27:54:34<11:30:34, 8.13s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=15.9, I=16.0, T=32.2, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:24:48,736] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:24:54,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:24:55,772] loop.py:380 INFO TIMING step=12925: step=3.441s data=1.360s compute=3.410s [fwd=2.440s bwd=0.955s opt=0.003s] dispatch=3.805s [2026-10-19 17:24:55,773] loop.py:45 INFO TRAIN_STEP step=12925 loss=1.196139 Training: 72%|███████▏ | 12925/18018 [27:54:41<11:08:02, 7.87s/step, loss=1.196, acc=0.717, acc_len=1.74, thru=15.6, I=20.5, T=37.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:24:55,918] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:25:02,988] loop.py:45 INFO TRAIN_STEP step=12926 loss=1.199750 Training: 72%|███████▏ | 12926/18018 [27:54:49<10:51:13, 7.67s/step, loss=1.200, acc=0.717, acc_len=1.74, thru=17.3, I=19.8, T=36.5, wait=3.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:25:02,984] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12925: forward=2536.0ms backward=927.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:12,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:25:12,541] loop.py:45 INFO TRAIN_STEP step=12927 loss=1.397898 Training: 72%|███████▏ | 12927/18018 [27:54:58<11:38:59, 8.24s/step, loss=1.398, acc=0.688, acc_len=1.64, thru=14.6, I=14.6, T=35.2, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:12,664] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:18,680] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:25:19,166] loop.py:45 INFO TRAIN_STEP step=12928 loss=1.251954 Training: 72%|███████▏ | 12928/18018 [27:55:05<10:57:50, 7.75s/step, loss=1.252, acc=0.707, acc_len=1.70, thru=19.3, I=18.0, T=36.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:19,234] inference_manager.py:408 INFO Pool capacity available after 0.6s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:24,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:25:25,730] loop.py:45 INFO TRAIN_STEP step=12929 loss=1.220736 Training: 72%|███████▏ | 12929/18018 [27:55:11<10:27:22, 7.40s/step, loss=1.221, acc=0.714, acc_len=1.73, thru=19.1, I=21.1, T=37.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:25,759] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:31,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:25:31,377] loop.py:380 INFO TIMING step=12930: step=3.322s data=1.318s compute=3.289s [fwd=2.343s bwd=0.930s opt=0.003s] dispatch=2.316s [2026-10-19 17:25:31,378] loop.py:45 INFO TRAIN_STEP step=12930 loss=1.271489 Training: 72%|███████▏ | 12930/18018 [27:55:17<9:42:45, 6.87s/step, loss=1.271, acc=0.703, acc_len=1.69, thru=22.2, I=24.6, T=38.5, wait=2.3s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:31,576] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:25:37,553] loop.py:45 INFO TRAIN_STEP step=12931 loss=1.411743 Training: 72%|███████▏ | 12931/18018 [27:55:23<9:24:54, 6.66s/step, loss=1.412, acc=0.681, acc_len=1.61, thru=19.9, I=19.9, T=37.8, wait=2.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:25:37,549] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12930: forward=2411.9ms backward=929.6ms [2026-10-19 17:25:43,649] loop.py:45 INFO TRAIN_STEP step=12932 loss=1.218604 Training: 72%|███████▏ | 12932/18018 [27:55:29<9:10:21, 6.49s/step, loss=1.219, acc=0.717, acc_len=1.74, thru=19.7, I=19.1, T=36.5, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:49,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:25:49,922] loop.py:45 INFO TRAIN_STEP step=12933 loss=1.195240 Training: 72%|███████▏ | 12933/18018 [27:55:36<9:04:40, 6.43s/step, loss=1.195, acc=0.719, acc_len=1.75, thru=20.3, I=21.4, T=35.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:50,295] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:57,338] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:25:58,075] loop.py:45 INFO TRAIN_STEP step=12934 loss=1.363828 Training: 72%|███████▏ | 12934/18018 [27:55:44<9:48:27, 6.94s/step, loss=1.364, acc=0.691, acc_len=1.65, thru=15.8, I=16.9, T=35.3, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:25:58,341] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:26:03,143] loop.py:380 INFO TIMING step=12935: step=3.237s data=1.249s compute=3.207s [fwd=2.294s bwd=0.898s opt=0.003s] dispatch=1.822s [2026-10-19 17:26:03,143] loop.py:45 INFO TRAIN_STEP step=12935 loss=1.188078 Training: 72%|███████▏ | 12935/18018 [27:55:49<9:00:43, 6.38s/step, loss=1.188, acc=0.718, acc_len=1.74, thru=18.1, I=24.8, T=39.5, wait=1.8s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:26:11,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:26:12,090] loop.py:45 INFO TRAIN_STEP step=12936 loss=1.343103 Training: 72%|███████▏ | 12936/18018 [27:55:58<10:05:43, 7.15s/step, loss=1.343, acc=0.692, acc_len=1.66, thru=17.9, I=16.2, T=37.2, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:26:12,086] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12935: forward=2456.9ms backward=935.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:26:12,364] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:26:18,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:26:18,744] loop.py:45 INFO TRAIN_STEP step=12937 loss=1.230741 Training: 72%|███████▏ | 12937/18018 [27:56:04<9:52:57, 7.00s/step, loss=1.231, acc=0.712, acc_len=1.72, thru=19.5, I=21.5, T=35.2, wait=3.0s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:26:19,070] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:26:23,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:26:24,685] loop.py:45 INFO TRAIN_STEP step=12938 loss=1.287369 Training: 72%|███████▏ | 12938/18018 [27:56:10<9:25:53, 6.68s/step, loss=1.287, acc=0.701, acc_len=1.69, thru=20.9, I=25.9, T=39.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:26:24,895] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:26:30,190] loop.py:45 INFO TRAIN_STEP step=12939 loss=1.249730 Training: 72%|███████▏ | 12939/18018 [27:56:16<8:55:47, 6.33s/step, loss=1.250, acc=0.709, acc_len=1.71, thru=20.6, I=23.6, T=38.5, wait=2.2s, pool=45, epoch=2/2][2026-10-19 17:26:38,519] loop.py:380 INFO TIMING step=12940: step=3.579s data=1.503s compute=3.535s [fwd=2.534s bwd=0.985s opt=0.003s] dispatch=4.743s [2026-10-19 17:26:38,520] loop.py:45 INFO TRAIN_STEP step=12940 loss=1.277896 Training: 72%|███████▏ | 12940/18018 [27:56:24<9:46:32, 6.93s/step, loss=1.278, acc=0.706, acc_len=1.70, thru=15.8, I=17.9, T=35.8, wait=4.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:26:46,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:26:46,922] loop.py:45 INFO TRAIN_STEP step=12941 loss=1.307735 Training: 72%|███████▏ | 12941/18018 [27:56:33<10:23:46, 7.37s/step, loss=1.308, acc=0.701, acc_len=1.69, thru=16.0, I=15.8, T=35.3, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:26:46,917] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12940: forward=2576.2ms backward=1001.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:26:47,415] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:26:53,425] loop.py:45 INFO TRAIN_STEP step=12942 loss=1.292061 Training: 72%|███████▏ | 12942/18018 [27:56:39<10:01:32, 7.11s/step, loss=1.292, acc=0.700, acc_len=1.67, thru=17.3, I=18.2, T=38.8, wait=3.2s, pool=45, epoch=2/2][2026-10-19 17:27:03,666] loop.py:45 INFO TRAIN_STEP step=12943 loss=1.313467 Training: 72%|███████▏ | 12943/18018 [27:56:49<11:20:52, 8.05s/step, loss=1.313, acc=0.699, acc_len=1.67, thru=12.0, I=12.2, T=34.7, wait=6.5s, pool=41, epoch=2/2][2026-10-19 17:27:11,223] loop.py:45 INFO TRAIN_STEP step=12944 loss=1.169981 Training: 72%|███████▏ | 12944/18018 [27:56:57<11:08:13, 7.90s/step, loss=1.170, acc=0.724, acc_len=1.77, thru=18.2, I=17.0, T=30.7, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:17,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:27:18,573] loop.py:380 INFO TIMING step=12945: step=3.474s data=1.384s compute=3.443s [fwd=2.467s bwd=0.961s opt=0.003s] dispatch=3.868s [2026-10-19 17:27:18,573] loop.py:45 INFO TRAIN_STEP step=12945 loss=1.241782 Training: 72%|███████▏ | 12945/18018 [27:57:04<10:54:10, 7.74s/step, loss=1.242, acc=0.712, acc_len=1.72, thru=16.2, I=17.9, T=36.8, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:18,896] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:24,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:27:25,188] loop.py:45 INFO TRAIN_STEP step=12946 loss=1.229983 Training: 72%|███████▏ | 12946/18018 [27:57:11<10:25:32, 7.40s/step, loss=1.230, acc=0.716, acc_len=1.74, thru=19.1, I=23.2, T=36.9, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:27:25,184] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12945: forward=2461.2ms backward=967.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:25,652] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:27:32,655] loop.py:45 INFO TRAIN_STEP step=12947 loss=1.272110 Training: 72%|███████▏ | 12947/18018 [27:57:18<10:27:09, 7.42s/step, loss=1.272, acc=0.707, acc_len=1.70, thru=12.1, I=18.0, T=37.2, wait=4.0s, pool=13, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 17:27:32,856] eagle_store.py:517 WARNING batch_get_buffer missing keys (k414299-81347b9c_hs, k414299-81347b9c_ids, k414299-81347b9c_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:41,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 17:27:33,322] eagle_store.py:517 WARNING batch_get_buffer missing keys (k414312-af9d5bda_hs, k414312-af9d5bda_ids, k414312-af9d5bda_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 17:27:42,575] loop.py:45 INFO TRAIN_STEP step=12948 loss=1.212154 Training: 72%|███████▏ | 12948/18018 [27:57:28<11:30:20, 8.17s/step, loss=1.212, acc=0.715, acc_len=1.73, thru=18.4, I=16.3, T=35.8, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:42,620] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:48,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:27:48,563] loop.py:45 INFO TRAIN_STEP step=12949 loss=1.084763 Training: 72%|███████▏ | 12949/18018 [27:57:34<10:34:55, 7.52s/step, loss=1.085, acc=0.738, acc_len=1.81, thru=20.2, I=22.6, T=35.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:48,975] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:54,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:27:55,413] loop.py:380 INFO TIMING step=12950: step=3.655s data=1.661s compute=3.621s [fwd=2.596s bwd=1.011s opt=0.003s] dispatch=3.187s [2026-10-19 17:27:55,413] loop.py:45 INFO TRAIN_STEP step=12950 loss=1.252422 Training: 72%|███████▏ | 12950/18018 [27:57:41<10:17:55, 7.32s/step, loss=1.252, acc=0.709, acc_len=1.70, thru=18.0, I=21.0, T=35.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:27:55,717] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:28:01,320] loop.py:45 INFO TRAIN_STEP step=12951 loss=1.246717 Training: 72%|███████▏ | 12951/18018 [27:57:47<9:42:07, 6.89s/step, loss=1.247, acc=0.711, acc_len=1.71, thru=20.1, I=24.1, T=39.2, wait=2.6s, pool=57, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:28:01,315] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12950: forward=2324.6ms backward=887.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:28:09,346] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:28:09,430] loop.py:45 INFO TRAIN_STEP step=12952 loss=1.208072 Training: 72%|███████▏ | 12952/18018 [27:57:55<10:12:49, 7.26s/step, loss=1.208, acc=0.720, acc_len=1.75, thru=16.1, I=16.5, T=36.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:28:09,849] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:28:16,977] loop.py:45 INFO TRAIN_STEP step=12953 loss=1.185393 Training: 72%|███████▏ | 12953/18018 [27:58:03<10:20:01, 7.34s/step, loss=1.185, acc=0.723, acc_len=1.76, thru=18.1, I=17.3, T=36.1, wait=4.0s, pool=61, epoch=2/2][2026-10-19 17:28:23,883] loop.py:45 INFO TRAIN_STEP step=12954 loss=1.257520 Training: 72%|███████▏ | 12954/18018 [27:58:09<10:08:56, 7.22s/step, loss=1.258, acc=0.707, acc_len=1.71, thru=16.0, I=19.8, T=36.6, wait=3.4s, pool=37, epoch=2/2][2026-10-19 17:28:32,953] loop.py:380 INFO TIMING step=12955: step=3.631s data=1.623s compute=3.559s [fwd=2.466s bwd=1.078s opt=0.003s] dispatch=5.424s [2026-10-19 17:28:32,953] loop.py:45 INFO TRAIN_STEP step=12955 loss=1.239025 Training: 72%|███████▏ | 12955/18018 [27:58:19<10:55:38, 7.77s/step, loss=1.239, acc=0.712, acc_len=1.73, thru=15.9, I=14.9, T=35.2, wait=5.4s, pool=53, epoch=2/2][2026-10-19 17:28:43,533] loop.py:45 INFO TRAIN_STEP step=12956 loss=1.347377 Training: 72%|███████▏ | 12956/18018 [27:58:29<12:06:40, 8.61s/step, loss=1.347, acc=0.697, acc_len=1.67, thru=12.6, I=12.3, T=32.7, wait=6.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:28:43,529] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12955: forward=2690.6ms backward=1160.4ms [2026-10-19 17:28:51,347] loop.py:45 INFO TRAIN_STEP step=12957 loss=1.223755 Training: 72%|███████▏ | 12957/18018 [27:58:37<11:46:15, 8.37s/step, loss=1.224, acc=0.716, acc_len=1.73, thru=15.7, I=15.3, T=35.6, wait=4.2s, pool=53, epoch=2/2][2026-10-19 17:29:01,378] loop.py:45 INFO TRAIN_STEP step=12958 loss=1.198544 Training: 72%|███████▏ | 12958/18018 [27:58:47<12:28:02, 8.87s/step, loss=1.199, acc=0.721, acc_len=1.75, thru=13.1, I=13.1, T=35.8, wait=6.4s, pool=53, epoch=2/2][2026-10-19 17:29:10,085] loop.py:45 INFO TRAIN_STEP step=12959 loss=1.373526 Training: 72%|███████▏ | 12959/18018 [27:58:56<12:23:50, 8.82s/step, loss=1.374, acc=0.685, acc_len=1.63, thru=14.6, I=14.9, T=34.3, wait=5.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:29:18,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:29:18,880] loop.py:380 INFO TIMING step=12960: step=3.644s data=1.431s compute=3.621s [fwd=2.599s bwd=1.007s opt=0.003s] dispatch=5.141s [2026-10-19 17:29:18,880] loop.py:45 INFO TRAIN_STEP step=12960 loss=1.201572 Training: 72%|███████▏ | 12960/18018 [27:59:04<12:22:59, 8.81s/step, loss=1.202, acc=0.719, acc_len=1.75, thru=16.0, I=15.8, T=35.1, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:29:18,988] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:29:24,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:29:25,021] loop.py:45 INFO TRAIN_STEP step=12961 loss=1.309449 Training: 72%|███████▏ | 12961/18018 [27:59:11<11:15:17, 8.01s/step, loss=1.309, acc=0.692, acc_len=1.66, thru=19.9, I=21.9, T=38.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:29:25,055] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 17:29:25,016] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12960: forward=2357.3ms backward=922.5ms [2026-10-19 17:29:31,449] loop.py:45 INFO TRAIN_STEP step=12962 loss=1.274975 Training: 72%|███████▏ | 12962/18018 [27:59:17<10:35:03, 7.54s/step, loss=1.275, acc=0.707, acc_len=1.70, thru=16.9, I=19.7, T=38.7, wait=3.1s, pool=33, epoch=2/2][2026-10-19 17:29:39,128] loop.py:45 INFO TRAIN_STEP step=12963 loss=1.189151 Training: 72%|███████▏ | 12963/18018 [27:59:25<10:38:30, 7.58s/step, loss=1.189, acc=0.720, acc_len=1.75, thru=17.0, I=15.3, T=36.3, wait=4.1s, pool=45, epoch=2/2][2026-10-19 17:29:46,462] loop.py:45 INFO TRAIN_STEP step=12964 loss=1.252643 Training: 72%|███████▏ | 12964/18018 [27:59:32<10:32:13, 7.51s/step, loss=1.253, acc=0.710, acc_len=1.71, thru=16.2, I=15.4, T=36.7, wait=3.8s, pool=41, epoch=2/2][2026-10-19 17:29:54,976] loop.py:380 INFO TIMING step=12965: step=3.543s data=1.457s compute=3.489s [fwd=2.486s bwd=0.989s opt=0.003s] dispatch=4.963s [2026-10-19 17:29:54,977] loop.py:45 INFO TRAIN_STEP step=12965 loss=1.261330 Training: 72%|███████▏ | 12965/18018 [27:59:41<10:57:37, 7.81s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=15.7, I=14.8, T=36.1, wait=5.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:00,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 17:30:02,030] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12965: forward=2538.6ms backward=977.3ms [2026-10-19 17:30:02,034] loop.py:45 INFO TRAIN_STEP step=12966 loss=1.174464 Training: 72%|███████▏ | 12966/18018 [27:59:48<10:38:29, 7.58s/step, loss=1.174, acc=0.722, acc_len=1.76, thru=17.7, I=21.3, T=35.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:02,408] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:07,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:30:07,926] loop.py:45 INFO TRAIN_STEP step=12967 loss=1.221736 Training: 72%|███████▏ | 12967/18018 [27:59:54<9:55:39, 7.08s/step, loss=1.222, acc=0.715, acc_len=1.73, thru=20.8, I=27.6, T=38.1, wait=2.5s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:08,196] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:30:13,635] loop.py:45 INFO TRAIN_STEP step=12968 loss=1.191312 Training: 72%|███████▏ | 12968/18018 [27:59:59<9:21:07, 6.67s/step, loss=1.191, acc=0.721, acc_len=1.75, thru=19.6, I=23.1, T=38.2, wait=2.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:21,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:30:22,607] loop.py:45 INFO TRAIN_STEP step=12969 loss=1.433733 Training: 72%|███████▏ | 12969/18018 [28:00:08<10:19:08, 7.36s/step, loss=1.434, acc=0.679, acc_len=1.61, thru=16.9, I=17.8, T=35.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:22,673] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 17:30:28,165] loop.py:380 INFO TIMING step=12970: step=3.366s data=1.373s compute=3.339s [fwd=2.415s bwd=0.910s opt=0.003s] dispatch=2.181s [2026-10-19 17:30:28,166] loop.py:45 INFO TRAIN_STEP step=12970 loss=1.260137 Training: 72%|███████▏ | 12970/18018 [28:00:14<9:33:35, 6.82s/step, loss=1.260, acc=0.708, acc_len=1.71, thru=20.5, I=24.9, T=38.0, wait=2.2s, pool=57, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:39,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:30:39,795] loop.py:45 INFO TRAIN_STEP step=12971 loss=1.323645 Training: 72%|███████▏ | 12971/18018 [28:00:25<11:34:56, 8.26s/step, loss=1.324, acc=0.696, acc_len=1.67, thru=12.7, I=11.9, T=33.4, wait=7.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:30:39,790] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12970: forward=2739.7ms backward=1039.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:39,866] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:45,893] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:30:46,034] loop.py:45 INFO TRAIN_STEP step=12972 loss=1.270636 Training: 72%|███████▏ | 12972/18018 [28:00:32<10:43:45, 7.65s/step, loss=1.271, acc=0.708, acc_len=1.71, thru=19.6, I=17.5, T=36.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:46,394] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:51,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:30:51,913] loop.py:45 INFO TRAIN_STEP step=12973 loss=1.146435 Training: 72%|███████▏ | 12973/18018 [28:00:37<9:58:47, 7.12s/step, loss=1.146, acc=0.731, acc_len=1.80, thru=21.4, I=24.4, T=37.3, wait=2.4s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:52,238] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:30:59,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:30:59,764] loop.py:45 INFO TRAIN_STEP step=12974 loss=1.243120 Training: 72%|███████▏ | 12974/18018 [28:00:45<10:17:05, 7.34s/step, loss=1.243, acc=0.709, acc_len=1.71, thru=16.6, I=17.7, T=35.8, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:00,118] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:05,076] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:31:06,204] loop.py:380 INFO TIMING step=12975: step=3.591s data=1.451s compute=3.555s [fwd=2.534s bwd=1.008s opt=0.003s] dispatch=2.840s [2026-10-19 17:31:06,204] loop.py:45 INFO TRAIN_STEP step=12975 loss=1.188577 Training: 72%|███████▏ | 12975/18018 [28:00:52<9:54:14, 7.07s/step, loss=1.189, acc=0.718, acc_len=1.74, thru=19.1, I=21.9, T=35.6, wait=2.8s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:06,586] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:11,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:31:12,637] loop.py:45 INFO TRAIN_STEP step=12976 loss=1.243595 Training: 72%|███████▏ | 12976/18018 [28:00:58<9:38:06, 6.88s/step, loss=1.244, acc=0.712, acc_len=1.72, thru=19.7, I=25.8, T=38.2, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:31:12,633] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12975: forward=2362.0ms backward=941.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:12,957] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:31:20,301] loop.py:45 INFO TRAIN_STEP step=12977 loss=1.309449 Training: 72%|███████▏ | 12977/18018 [28:01:06<9:57:45, 7.11s/step, loss=1.309, acc=0.697, acc_len=1.67, thru=14.5, I=17.8, T=34.5, wait=3.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:26,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:31:27,136] loop.py:45 INFO TRAIN_STEP step=12978 loss=1.258901 Training: 72%|███████▏ | 12978/18018 [28:01:13<9:50:37, 7.03s/step, loss=1.259, acc=0.705, acc_len=1.70, thru=18.3, I=20.3, T=35.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:27,171] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:32,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:31:33,653] loop.py:45 INFO TRAIN_STEP step=12979 loss=1.366944 Training: 72%|███████▏ | 12979/18018 [28:01:19<9:37:29, 6.88s/step, loss=1.367, acc=0.686, acc_len=1.64, thru=18.7, I=22.9, T=32.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:33,858] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:31:40,852] loop.py:380 INFO TIMING step=12980: step=3.682s data=1.520s compute=3.646s [fwd=2.603s bwd=1.031s opt=0.003s] dispatch=3.508s [2026-10-19 17:31:40,852] loop.py:45 INFO TRAIN_STEP step=12980 loss=1.374721 Training: 72%|███████▏ | 12980/18018 [28:01:26<9:45:28, 6.97s/step, loss=1.375, acc=0.690, acc_len=1.65, thru=15.6, I=16.0, T=34.8, wait=3.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:48,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:31:48,809] loop.py:45 INFO TRAIN_STEP step=12981 loss=1.252294 Training: 72%|███████▏ | 12981/18018 [28:01:34<10:10:09, 7.27s/step, loss=1.252, acc=0.706, acc_len=1.70, thru=18.2, I=17.9, T=36.3, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:31:48,804] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12980: forward=2456.0ms backward=1022.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:31:49,294] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:31:56,983] loop.py:45 INFO TRAIN_STEP step=12982 loss=1.251293 Training: 72%|███████▏ | 12982/18018 [28:01:43<10:32:53, 7.54s/step, loss=1.251, acc=0.710, acc_len=1.72, thru=15.5, I=16.4, T=35.6, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:04,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:32:05,075] loop.py:45 INFO TRAIN_STEP step=12983 loss=1.257636 Training: 72%|███████▏ | 12983/18018 [28:01:51<10:46:40, 7.71s/step, loss=1.258, acc=0.709, acc_len=1.71, thru=16.6, I=17.7, T=34.6, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:05,326] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:11,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:32:11,283] loop.py:45 INFO TRAIN_STEP step=12984 loss=1.246643 Training: 72%|███████▏ | 12984/18018 [28:01:57<10:08:46, 7.26s/step, loss=1.247, acc=0.710, acc_len=1.71, thru=19.7, I=21.3, T=35.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:11,544] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:17,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:32:18,414] loop.py:380 INFO TIMING step=12985: step=3.698s data=1.542s compute=3.655s [fwd=2.612s bwd=1.028s opt=0.003s] dispatch=3.424s [2026-10-19 17:32:18,414] loop.py:45 INFO TRAIN_STEP step=12985 loss=1.263823 Training: 72%|███████▏ | 12985/18018 [28:02:04<10:05:30, 7.22s/step, loss=1.264, acc=0.706, acc_len=1.71, thru=17.4, I=19.4, T=34.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:18,537] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 17:32:25,215] loop.py:45 INFO TRAIN_STEP step=12986 loss=1.225741 Training: 72%|███████▏ | 12986/18018 [28:02:11<9:54:52, 7.09s/step, loss=1.226, acc=0.714, acc_len=1.73, thru=18.3, I=21.2, T=34.1, wait=3.0s, pool=57, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:32:25,210] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12985: forward=2654.2ms backward=1044.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:31,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:32:33,249] loop.py:45 INFO TRAIN_STEP step=12987 loss=1.293031 Training: 72%|███████▏ | 12987/18018 [28:02:19<10:18:30, 7.38s/step, loss=1.293, acc=0.702, acc_len=1.69, thru=15.8, I=17.6, T=35.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:33,484] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:38,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:32:39,513] loop.py:45 INFO TRAIN_STEP step=12988 loss=1.426058 Training: 72%|███████▏ | 12988/18018 [28:02:25<9:50:22, 7.04s/step, loss=1.426, acc=0.676, acc_len=1.60, thru=19.7, I=25.8, T=32.2, wait=2.3s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:39,564] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:46,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:32:47,087] loop.py:45 INFO TRAIN_STEP step=12989 loss=1.297863 Training: 72%|███████▏ | 12989/18018 [28:02:33<10:03:37, 7.20s/step, loss=1.298, acc=0.701, acc_len=1.68, thru=15.3, I=18.3, T=37.4, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:32:47,328] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:32:56,195] loop.py:380 INFO TIMING step=12990: step=3.723s data=1.502s compute=3.685s [fwd=2.666s bwd=1.004s opt=0.003s] dispatch=5.375s [2026-10-19 17:32:56,195] loop.py:45 INFO TRAIN_STEP step=12990 loss=1.356496 Training: 72%|███████▏ | 12990/18018 [28:02:42<10:51:26, 7.77s/step, loss=1.356, acc=0.689, acc_len=1.64, thru=12.6, I=15.6, T=34.4, wait=5.4s, pool=57, epoch=2/2][2026-10-19 17:33:03,740] loop.py:45 INFO TRAIN_STEP step=12991 loss=1.251692 Training: 72%|███████▏ | 12991/18018 [28:02:49<10:45:30, 7.70s/step, loss=1.252, acc=0.712, acc_len=1.72, thru=14.7, I=16.5, T=36.7, wait=4.0s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:33:03,736] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12990: forward=2425.6ms backward=1014.2ms [2026-10-19 17:33:13,393] loop.py:45 INFO TRAIN_STEP step=12992 loss=1.259716 Training: 72%|███████▏ | 12992/18018 [28:02:59<11:34:24, 8.29s/step, loss=1.260, acc=0.707, acc_len=1.71, thru=15.0, I=16.1, T=32.2, wait=5.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:20,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:33:21,058] loop.py:45 INFO TRAIN_STEP step=12993 loss=1.261269 Training: 72%|███████▏ | 12993/18018 [28:03:07<11:18:32, 8.10s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=16.5, I=16.7, T=35.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:21,217] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:27,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:33:27,863] loop.py:45 INFO TRAIN_STEP step=12994 loss=1.211984 Training: 72%|███████▏ | 12994/18018 [28:03:13<10:45:48, 7.71s/step, loss=1.212, acc=0.714, acc_len=1.73, thru=18.8, I=21.7, T=31.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:28,017] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:34,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:33:34,718] loop.py:380 INFO TIMING step=12995: step=3.454s data=1.417s compute=3.415s [fwd=2.429s bwd=0.972s opt=0.003s] dispatch=3.393s [2026-10-19 17:33:34,718] loop.py:45 INFO TRAIN_STEP step=12995 loss=1.413275 Training: 72%|███████▏ | 12995/18018 [28:03:20<10:24:09, 7.46s/step, loss=1.413, acc=0.681, acc_len=1.62, thru=17.4, I=18.3, T=37.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:34,808] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 17:33:40,827] loop.py:45 INFO TRAIN_STEP step=12996 loss=1.306289 Training: 72%|███████▏ | 12996/18018 [28:03:26<9:50:16, 7.05s/step, loss=1.306, acc=0.699, acc_len=1.69, thru=18.7, I=20.4, T=36.6, wait=2.6s, pool=45, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:33:40,823] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=12995: forward=2472.9ms backward=985.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:49,077] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:33:49,241] loop.py:45 INFO TRAIN_STEP step=12997 loss=1.337994 Training: 72%|███████▏ | 12997/18018 [28:03:35<10:24:16, 7.46s/step, loss=1.338, acc=0.693, acc_len=1.66, thru=17.4, I=17.0, T=34.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:49,579] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:56,378] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:33:57,125] loop.py:45 INFO TRAIN_STEP step=12998 loss=1.365471 Training: 72%|███████▏ | 12998/18018 [28:03:43<10:34:47, 7.59s/step, loss=1.365, acc=0.686, acc_len=1.64, thru=17.3, I=18.4, T=31.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:33:57,383] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:04,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:34:05,348] loop.py:45 INFO TRAIN_STEP step=12999 loss=1.293957 Training: 72%|███████▏ | 12999/18018 [28:03:51<10:50:37, 7.78s/step, loss=1.294, acc=0.702, acc_len=1.69, thru=15.1, I=18.7, T=32.8, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:05,465] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:12,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:34:12,947] loop.py:380 INFO TIMING step=13000: step=3.896s data=1.568s compute=3.865s [fwd=2.820s bwd=1.031s opt=0.003s] dispatch=3.696s [2026-10-19 17:34:12,948] loop.py:45 INFO TRAIN_STEP step=13000 loss=1.237347 Training: 72%|███████▏ | 13000/18018 [28:03:59<10:45:59, 7.72s/step, loss=1.237, acc=0.711, acc_len=1.72, thru=15.7, I=16.5, T=32.9, wait=3.7s, pool=65, epoch=2/2][2026-10-19 17:34:12,950] loop.py:419 INFO Saving checkpoint at step 13000... (TrainerActor pid=9310) [2026-10-19 17:34:16,395] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0013001 (TrainerActor pid=9310) [2026-10-19 17:34:16,403] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0012251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:17,422] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.1s): pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:17,422] inference_manager.py:408 INFO Pool capacity available after 5.1s, pool_size=57/64 [2026-10-19 17:34:22,202] loop.py:45 INFO TRAIN_STEP step=13001 loss=1.189994 Training: 72%|███████▏ | 13001/18018 [28:04:08<11:24:18, 8.18s/step, loss=1.190, acc=0.721, acc_len=1.75, thru=12.4, I=27.7, T=39.3, wait=1.6s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:34:22,198] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13000: forward=2288.7ms backward=917.2ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 17:34:16,397] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0013001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 17:34:17,385] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0012251 [2026-10-19 17:34:28,988] loop.py:45 INFO TRAIN_STEP step=13002 loss=1.193087 Training: 72%|███████▏ | 13002/18018 [28:04:15<10:49:03, 7.76s/step, loss=1.193, acc=0.721, acc_len=1.76, thru=19.6, I=19.8, T=37.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:28,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:29,459] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:34:37,821] loop.py:45 INFO TRAIN_STEP step=13003 loss=1.355040 Training: 72%|███████▏ | 13003/18018 [28:04:23<11:15:49, 8.09s/step, loss=1.355, acc=0.688, acc_len=1.64, thru=13.7, I=13.9, T=34.5, wait=5.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:43,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:34:44,508] loop.py:45 INFO TRAIN_STEP step=13004 loss=1.368590 Training: 72%|███████▏ | 13004/18018 [28:04:30<10:40:36, 7.67s/step, loss=1.369, acc=0.687, acc_len=1.63, thru=19.2, I=19.3, T=35.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:44,925] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:49,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:34:50,694] loop.py:380 INFO TIMING step=13005: step=3.391s data=1.332s compute=3.360s [fwd=2.434s bwd=0.911s opt=0.003s] dispatch=2.786s [2026-10-19 17:34:50,694] loop.py:45 INFO TRAIN_STEP step=13005 loss=1.341963 Training: 72%|███████▏ | 13005/18018 [28:04:36<10:03:22, 7.22s/step, loss=1.342, acc=0.693, acc_len=1.66, thru=19.8, I=23.5, T=37.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:50,745] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:57,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:34:58,651] loop.py:45 INFO TRAIN_STEP step=13006 loss=1.322882 Training: 72%|███████▏ | 13006/18018 [28:04:44<10:21:41, 7.44s/step, loss=1.323, acc=0.697, acc_len=1.67, thru=16.0, I=18.7, T=33.8, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:34:58,646] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13005: forward=2788.5ms backward=954.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:34:58,703] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:06,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:35:06,089] loop.py:45 INFO TRAIN_STEP step=13007 loss=1.287076 Training: 72%|███████▏ | 13007/18018 [28:04:52<10:21:26, 7.44s/step, loss=1.287, acc=0.704, acc_len=1.69, thru=17.4, I=19.5, T=35.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:06,516] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:35:12,452] loop.py:45 INFO TRAIN_STEP step=13008 loss=1.228017 Training: 72%|███████▏ | 13008/18018 [28:04:58<9:54:17, 7.12s/step, loss=1.228, acc=0.711, acc_len=1.72, thru=19.2, I=18.8, T=37.1, wait=2.9s, pool=61, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:17,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:35:19,210] loop.py:45 INFO TRAIN_STEP step=13009 loss=1.294904 Training: 72%|███████▏ | 13009/18018 [28:05:05<9:45:12, 7.01s/step, loss=1.295, acc=0.704, acc_len=1.70, thru=18.7, I=21.5, T=29.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:19,505] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:23,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:35:24,774] loop.py:380 INFO TIMING step=13010: step=3.215s data=1.296s compute=3.184s [fwd=2.298s bwd=0.871s opt=0.003s] dispatch=2.338s [2026-10-19 17:35:24,774] loop.py:45 INFO TRAIN_STEP step=13010 loss=1.245653 Training: 72%|███████▏ | 13010/18018 [28:05:10<9:08:48, 6.58s/step, loss=1.246, acc=0.710, acc_len=1.72, thru=19.8, I=27.6, T=39.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:24,891] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:31,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:35:31,736] loop.py:45 INFO TRAIN_STEP step=13011 loss=1.296917 Training: 72%|███████▏ | 13011/18018 [28:05:17<9:18:28, 6.69s/step, loss=1.297, acc=0.700, acc_len=1.68, thru=19.0, I=23.9, T=37.1, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:35:31,731] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13010: forward=2450.0ms backward=957.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:32,090] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:39,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:35:40,124] loop.py:45 INFO TRAIN_STEP step=13012 loss=1.163161 Training: 72%|███████▏ | 13012/18018 [28:05:26<10:00:44, 7.20s/step, loss=1.163, acc=0.725, acc_len=1.77, thru=15.3, I=16.9, T=33.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:40,476] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:47,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:35:47,709] loop.py:45 INFO TRAIN_STEP step=13013 loss=1.275458 Training: 72%|███████▏ | 13013/18018 [28:05:33<10:10:18, 7.32s/step, loss=1.275, acc=0.701, acc_len=1.68, thru=16.8, I=19.9, T=33.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:48,047] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:52,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:35:53,588] loop.py:45 INFO TRAIN_STEP step=13014 loss=1.360931 Training: 72%|███████▏ | 13014/18018 [28:05:39<9:34:10, 6.88s/step, loss=1.361, acc=0.689, acc_len=1.64, thru=20.6, I=23.8, T=37.2, wait=2.4s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:53,687] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:58,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:35:58,482] loop.py:380 INFO TIMING step=13015: step=3.320s data=1.397s compute=3.292s [fwd=2.385s bwd=0.893s opt=0.003s] dispatch=1.564s [2026-10-19 17:35:58,482] loop.py:45 INFO TRAIN_STEP step=13015 loss=1.272578 Training: 72%|███████▏ | 13015/18018 [28:05:44<8:44:16, 6.29s/step, loss=1.273, acc=0.706, acc_len=1.70, thru=24.3, I=26.7, T=38.6, wait=1.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:35:58,706] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:04,373] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:36:05,044] loop.py:45 INFO TRAIN_STEP step=13016 loss=1.212138 Training: 72%|███████▏ | 13016/18018 [28:05:51<8:51:00, 6.37s/step, loss=1.212, acc=0.715, acc_len=1.73, thru=20.1, I=22.3, T=32.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:36:05,038] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13015: forward=2766.5ms backward=1070.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:05,379] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:10,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:36:11,141] loop.py:45 INFO TRAIN_STEP step=13017 loss=1.277031 Training: 72%|███████▏ | 13017/18018 [28:05:57<8:44:07, 6.29s/step, loss=1.277, acc=0.706, acc_len=1.70, thru=19.2, I=21.2, T=36.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:11,333] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:18,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:36:18,769] loop.py:45 INFO TRAIN_STEP step=13018 loss=1.303555 Training: 72%|███████▏ | 13018/18018 [28:06:04<9:17:29, 6.69s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=18.0, I=18.1, T=32.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:18,962] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:24,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:36:25,589] loop.py:45 INFO TRAIN_STEP step=13019 loss=1.301286 Training: 72%|███████▏ | 13019/18018 [28:06:11<9:20:40, 6.73s/step, loss=1.301, acc=0.701, acc_len=1.68, thru=18.6, I=18.9, T=33.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:25,967] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:36:31,700] loop.py:380 INFO TIMING step=13020: step=3.518s data=1.456s compute=3.474s [fwd=2.548s bwd=0.910s opt=0.003s] dispatch=2.583s [2026-10-19 17:36:31,701] loop.py:45 INFO TRAIN_STEP step=13020 loss=1.275065 Training: 72%|███████▏ | 13020/18018 [28:06:17<9:05:04, 6.54s/step, loss=1.275, acc=0.704, acc_len=1.69, thru=18.5, I=21.4, T=36.4, wait=2.6s, pool=57, epoch=2/2][2026-10-19 17:36:39,192] loop.py:45 INFO TRAIN_STEP step=13021 loss=1.240742 Training: 72%|███████▏ | 13021/18018 [28:06:25<9:28:44, 6.83s/step, loss=1.241, acc=0.710, acc_len=1.71, thru=14.0, I=16.6, T=36.7, wait=4.0s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:36:39,188] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13020: forward=2420.3ms backward=1003.8ms [2026-10-19 17:36:48,080] loop.py:45 INFO TRAIN_STEP step=13022 loss=1.329927 Training: 72%|███████▏ | 13022/18018 [28:06:34<10:19:57, 7.45s/step, loss=1.330, acc=0.695, acc_len=1.66, thru=16.3, I=15.4, T=33.5, wait=5.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:53,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:36:54,610] loop.py:45 INFO TRAIN_STEP step=13023 loss=1.331466 Training: 72%|███████▏ | 13023/18018 [28:06:40<9:56:57, 7.17s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=18.6, I=20.6, T=35.6, wait=2.9s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:36:54,984] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:37:00,381] loop.py:45 INFO TRAIN_STEP step=13024 loss=1.237933 Training: 72%|███████▏ | 13024/18018 [28:06:46<9:21:58, 6.75s/step, loss=1.238, acc=0.710, acc_len=1.72, thru=17.8, I=24.9, T=39.1, wait=2.5s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:08,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:37:09,986] loop.py:380 INFO TIMING step=13025: step=4.178s data=1.902s compute=4.141s [fwd=3.018s bwd=1.109s opt=0.003s] dispatch=5.416s [2026-10-19 17:37:09,986] loop.py:45 INFO TRAIN_STEP step=13025 loss=1.325356 Training: 72%|███████▏ | 13025/18018 [28:06:56<10:33:04, 7.61s/step, loss=1.325, acc=0.697, acc_len=1.67, thru=17.1, I=16.8, T=30.6, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:10,022] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:15,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:37:16,035] loop.py:45 INFO TRAIN_STEP step=13026 loss=1.314720 Training: 72%|███████▏ | 13026/18018 [28:07:02<9:53:58, 7.14s/step, loss=1.315, acc=0.695, acc_len=1.66, thru=18.9, I=25.4, T=36.8, wait=2.6s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:37:16,031] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13025: forward=2416.7ms backward=1013.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:16,363] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:37:22,237] loop.py:45 INFO TRAIN_STEP step=13027 loss=1.277440 Training: 72%|███████▏ | 13027/18018 [28:07:08<9:30:30, 6.86s/step, loss=1.277, acc=0.706, acc_len=1.70, thru=18.9, I=23.7, T=37.3, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:28,637] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:37:29,086] loop.py:45 INFO TRAIN_STEP step=13028 loss=1.106472 Training: 72%|███████▏ | 13028/18018 [28:07:15<9:30:11, 6.86s/step, loss=1.106, acc=0.737, acc_len=1.81, thru=18.7, I=19.6, T=33.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:29,139] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:39,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:37:40,051] loop.py:45 INFO TRAIN_STEP step=13029 loss=1.376553 Training: 72%|███████▏ | 13029/18018 [28:07:26<11:12:35, 8.09s/step, loss=1.377, acc=0.690, acc_len=1.65, thru=11.1, I=12.1, T=31.8, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:40,384] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:37:47,552] loop.py:380 INFO TIMING step=13030: step=3.801s data=1.548s compute=3.759s [fwd=2.655s bwd=1.090s opt=0.003s] dispatch=3.690s [2026-10-19 17:37:47,552] loop.py:45 INFO TRAIN_STEP step=13030 loss=1.380592 Training: 72%|███████▏ | 13030/18018 [28:07:33<10:57:43, 7.91s/step, loss=1.381, acc=0.683, acc_len=1.62, thru=15.3, I=16.6, T=33.7, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:56,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:37:57,277] loop.py:45 INFO TRAIN_STEP step=13031 loss=1.277367 Training: 72%|███████▏ | 13031/18018 [28:07:43<11:42:50, 8.46s/step, loss=1.277, acc=0.701, acc_len=1.68, thru=14.8, I=16.3, T=34.3, wait=6.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:37:57,272] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13030: forward=2619.8ms backward=1037.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:37:57,572] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:38:04,630] loop.py:45 INFO TRAIN_STEP step=13032 loss=1.177172 Training: 72%|███████▏ | 13032/18018 [28:07:50<11:15:10, 8.12s/step, loss=1.177, acc=0.718, acc_len=1.74, thru=15.3, I=16.4, T=34.4, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:38:12,746] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:38:13,261] loop.py:45 INFO TRAIN_STEP step=13033 loss=1.169865 Training: 72%|███████▏ | 13033/18018 [28:07:59<11:27:39, 8.28s/step, loss=1.170, acc=0.723, acc_len=1.76, thru=17.7, I=16.3, T=34.1, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:38:13,752] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:38:20,239] loop.py:45 INFO TRAIN_STEP step=13034 loss=1.227710 Training: 72%|███████▏ | 13034/18018 [28:08:06<10:55:10, 7.89s/step, loss=1.228, acc=0.713, acc_len=1.72, thru=14.9, I=20.4, T=36.5, wait=3.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:38:31,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:38:31,993] loop.py:380 INFO TIMING step=13035: step=3.928s data=1.684s compute=3.880s [fwd=2.756s bwd=1.109s opt=0.003s] dispatch=7.817s [2026-10-19 17:38:31,993] loop.py:45 INFO TRAIN_STEP step=13035 loss=1.358918 Training: 72%|███████▏ | 13035/18018 [28:08:18<12:31:21, 9.05s/step, loss=1.359, acc=0.693, acc_len=1.65, thru=14.4, I=13.4, T=32.6, wait=7.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:38:32,282] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:38:38,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:38:38,661] loop.py:45 INFO TRAIN_STEP step=13036 loss=1.410216 Training: 72%|███████▏ | 13036/18018 [28:08:24<11:31:59, 8.33s/step, loss=1.410, acc=0.684, acc_len=1.62, thru=18.5, I=19.0, T=36.1, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:38:38,657] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13035: forward=2548.0ms backward=949.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:38:38,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 17:38:45,253] loop.py:45 INFO TRAIN_STEP step=13037 loss=1.309197 Training: 72%|███████▏ | 13037/18018 [28:08:31<10:48:28, 7.81s/step, loss=1.309, acc=0.701, acc_len=1.68, thru=18.8, I=19.2, T=36.1, wait=3.0s, pool=61, epoch=2/2][2026-10-19 17:38:52,920] loop.py:45 INFO TRAIN_STEP step=13038 loss=1.312099 Training: 72%|███████▏ | 13038/18018 [28:08:39<10:44:49, 7.77s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=17.1, I=16.3, T=35.5, wait=4.0s, pool=53, epoch=2/2][2026-10-19 17:38:59,616] loop.py:45 INFO TRAIN_STEP step=13039 loss=1.205949 Training: 72%|███████▏ | 13039/18018 [28:08:45<10:17:54, 7.45s/step, loss=1.206, acc=0.720, acc_len=1.75, thru=17.8, I=18.2, T=37.7, wait=3.3s, pool=57, epoch=2/2][2026-10-19 17:39:07,777] loop.py:380 INFO TIMING step=13040: step=3.724s data=1.559s compute=3.688s [fwd=2.640s bwd=1.034s opt=0.003s] dispatch=4.426s [2026-10-19 17:39:07,777] loop.py:45 INFO TRAIN_STEP step=13040 loss=1.343946 Training: 72%|███████▏ | 13040/18018 [28:08:53<10:35:37, 7.66s/step, loss=1.344, acc=0.691, acc_len=1.65, thru=15.2, I=15.6, T=34.4, wait=4.4s, pool=37, epoch=2/2][2026-10-19 17:39:20,760] loop.py:45 INFO TRAIN_STEP step=13041 loss=1.338355 Training: 72%|███████▏ | 13041/18018 [28:09:06<12:47:49, 9.26s/step, loss=1.338, acc=0.692, acc_len=1.66, thru=12.8, I=11.8, T=31.7, wait=8.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:39:20,756] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13040: forward=2784.0ms backward=1187.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:20,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:21,210] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:29,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:39:30,353] loop.py:45 INFO TRAIN_STEP step=13042 loss=1.375363 Training: 72%|███████▏ | 13042/18018 [28:09:16<12:56:02, 9.36s/step, loss=1.375, acc=0.687, acc_len=1.64, thru=13.7, I=14.1, T=31.6, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:30,798] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:36,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:39:36,940] loop.py:45 INFO TRAIN_STEP step=13043 loss=1.300898 Training: 72%|███████▏ | 13043/18018 [28:09:23<11:46:58, 8.53s/step, loss=1.301, acc=0.699, acc_len=1.67, thru=17.9, I=19.7, T=36.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:37,298] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:43,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:39:43,434] loop.py:45 INFO TRAIN_STEP step=13044 loss=1.221424 Training: 72%|███████▏ | 13044/18018 [28:09:29<10:56:17, 7.92s/step, loss=1.221, acc=0.717, acc_len=1.74, thru=19.8, I=21.3, T=36.5, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:43,804] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:49,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:39:50,330] loop.py:380 INFO TIMING step=13045: step=3.621s data=1.512s compute=3.575s [fwd=2.525s bwd=1.035s opt=0.003s] dispatch=3.267s [2026-10-19 17:39:50,331] loop.py:45 INFO TRAIN_STEP step=13045 loss=1.209373 Training: 72%|███████▏ | 13045/18018 [28:09:36<10:30:51, 7.61s/step, loss=1.209, acc=0.715, acc_len=1.73, thru=18.8, I=21.5, T=35.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:39:50,564] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:39:58,845] loop.py:45 INFO TRAIN_STEP step=13046 loss=1.174541 Training: 72%|███████▏ | 13046/18018 [28:09:44<10:53:07, 7.88s/step, loss=1.175, acc=0.722, acc_len=1.76, thru=11.4, I=13.3, T=32.7, wait=4.6s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:39:58,841] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13045: forward=2734.1ms backward=1107.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:08,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:40:08,833] loop.py:45 INFO TRAIN_STEP step=13047 loss=1.253424 Training: 72%|███████▏ | 13047/18018 [28:09:54<11:45:23, 8.51s/step, loss=1.253, acc=0.707, acc_len=1.71, thru=15.6, I=15.1, T=34.1, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:09,201] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:14,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:40:15,457] loop.py:45 INFO TRAIN_STEP step=13048 loss=1.220124 Training: 72%|███████▏ | 13048/18018 [28:10:01<10:58:17, 7.95s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=19.3, I=23.0, T=37.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:15,663] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:22,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:40:22,878] loop.py:45 INFO TRAIN_STEP step=13049 loss=1.291683 Training: 72%|███████▏ | 13049/18018 [28:10:08<10:45:04, 7.79s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=17.2, I=21.0, T=35.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:23,167] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:40:30,401] loop.py:380 INFO TIMING step=13050: step=3.437s data=1.422s compute=3.411s [fwd=2.456s bwd=0.941s opt=0.003s] dispatch=4.076s [2026-10-19 17:40:30,401] loop.py:45 INFO TRAIN_STEP step=13050 loss=1.211405 Training: 72%|███████▏ | 13050/18018 [28:10:16<10:38:16, 7.71s/step, loss=1.211, acc=0.715, acc_len=1.74, thru=15.4, I=18.3, T=37.2, wait=4.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:36,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:40:38,050] loop.py:45 INFO TRAIN_STEP step=13051 loss=1.238844 Training: 72%|███████▏ | 13051/18018 [28:10:24<10:36:40, 7.69s/step, loss=1.239, acc=0.709, acc_len=1.71, thru=17.6, I=20.1, T=35.1, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:40:38,046] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13050: forward=2584.5ms backward=1012.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:38,495] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:43,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:40:43,952] loop.py:45 INFO TRAIN_STEP step=13052 loss=1.280322 Training: 72%|███████▏ | 13052/18018 [28:10:30<9:52:09, 7.15s/step, loss=1.280, acc=0.703, acc_len=1.68, thru=21.3, I=27.3, T=36.1, wait=2.3s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:44,415] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:50,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:40:50,889] loop.py:45 INFO TRAIN_STEP step=13053 loss=1.265498 Training: 72%|███████▏ | 13053/18018 [28:10:36<9:46:36, 7.09s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=18.2, I=21.9, T=36.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:40:51,295] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:40:58,070] loop.py:45 INFO TRAIN_STEP step=13054 loss=1.302952 Training: 72%|███████▏ | 13054/18018 [28:10:44<9:48:46, 7.12s/step, loss=1.303, acc=0.700, acc_len=1.68, thru=15.0, I=19.4, T=38.3, wait=3.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:05,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:41:06,565] loop.py:380 INFO TIMING step=13055: step=3.700s data=1.533s compute=3.676s [fwd=2.665s bwd=0.997s opt=0.003s] dispatch=4.787s [2026-10-19 17:41:06,566] loop.py:45 INFO TRAIN_STEP step=13055 loss=1.267764 Training: 72%|███████▏ | 13055/18018 [28:10:52<10:22:55, 7.53s/step, loss=1.268, acc=0.704, acc_len=1.70, thru=17.6, I=17.6, T=34.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:06,786] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:41:15,974] loop.py:45 INFO TRAIN_STEP step=13056 loss=1.205108 Training: 72%|███████▏ | 13056/18018 [28:11:02<11:09:23, 8.09s/step, loss=1.205, acc=0.718, acc_len=1.74, thru=11.3, I=12.4, T=33.7, wait=5.6s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:41:15,970] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13055: forward=2734.7ms backward=1026.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:21,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:41:23,036] loop.py:45 INFO TRAIN_STEP step=13057 loss=1.279864 Training: 72%|███████▏ | 13057/18018 [28:11:09<10:43:35, 7.78s/step, loss=1.280, acc=0.706, acc_len=1.70, thru=18.7, I=20.8, T=35.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:23,514] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:30,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:41:30,638] loop.py:45 INFO TRAIN_STEP step=13058 loss=1.185058 Training: 72%|███████▏ | 13058/18018 [28:11:16<10:38:58, 7.73s/step, loss=1.185, acc=0.721, acc_len=1.75, thru=16.4, I=21.9, T=35.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:30,666] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 17:41:35,639] loop.py:45 INFO TRAIN_STEP step=13059 loss=1.244220 Training: 72%|███████▏ | 13059/18018 [28:11:21<9:31:13, 6.91s/step, loss=1.244, acc=0.713, acc_len=1.72, thru=19.1, I=21.0, T=38.9, wait=1.7s, pool=37, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:44,860] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:41:45,390] loop.py:380 INFO TIMING step=13060: step=3.722s data=1.570s compute=3.693s [fwd=2.655s bwd=1.023s opt=0.003s] dispatch=6.019s [2026-10-19 17:41:45,390] loop.py:45 INFO TRAIN_STEP step=13060 loss=1.372845 Training: 72%|███████▏ | 13060/18018 [28:11:31<10:41:25, 7.76s/step, loss=1.373, acc=0.689, acc_len=1.64, thru=16.5, I=15.3, T=34.4, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:45,865] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:51,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:41:51,506] loop.py:45 INFO TRAIN_STEP step=13061 loss=1.243012 Training: 72%|███████▏ | 13061/18018 [28:11:37<10:00:31, 7.27s/step, loss=1.243, acc=0.708, acc_len=1.70, thru=19.9, I=20.2, T=37.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:41:51,502] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13060: forward=2412.3ms backward=918.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:51,928] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:58,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:41:58,636] loop.py:45 INFO TRAIN_STEP step=13062 loss=1.273274 Training: 72%|███████▏ | 13062/18018 [28:11:44<9:56:59, 7.23s/step, loss=1.273, acc=0.703, acc_len=1.69, thru=18.5, I=19.4, T=34.4, wait=3.4s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:41:58,800] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:42:04,695] loop.py:45 INFO TRAIN_STEP step=13063 loss=1.263123 Training: 72%|███████▏ | 13063/18018 [28:11:50<9:27:53, 6.88s/step, loss=1.263, acc=0.710, acc_len=1.72, thru=19.6, I=19.4, T=36.5, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:11,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:42:11,432] loop.py:45 INFO TRAIN_STEP step=13064 loss=1.386311 Training: 73%|███████▎ | 13064/18018 [28:11:57<9:24:20, 6.84s/step, loss=1.386, acc=0.682, acc_len=1.62, thru=18.9, I=19.8, T=37.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:11,571] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 17:42:17,481] loop.py:380 INFO TIMING step=13065: step=3.466s data=1.410s compute=3.425s [fwd=2.420s bwd=0.991s opt=0.003s] dispatch=2.574s [2026-10-19 17:42:17,481] loop.py:45 INFO TRAIN_STEP step=13065 loss=1.308516 Training: 73%|███████▎ | 13065/18018 [28:12:03<9:04:44, 6.60s/step, loss=1.309, acc=0.697, acc_len=1.67, thru=18.8, I=19.4, T=36.9, wait=2.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:25,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:42:26,278] loop.py:45 INFO TRAIN_STEP step=13066 loss=1.215011 Training: 73%|███████▎ | 13066/18018 [28:12:12<9:59:05, 7.26s/step, loss=1.215, acc=0.717, acc_len=1.74, thru=16.2, I=15.6, T=33.6, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:42:26,273] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13065: forward=2686.1ms backward=1054.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:26,467] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:32,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:42:32,353] loop.py:45 INFO TRAIN_STEP step=13067 loss=1.322839 Training: 73%|███████▎ | 13067/18018 [28:12:18<9:29:38, 6.90s/step, loss=1.323, acc=0.695, acc_len=1.67, thru=19.8, I=19.2, T=37.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:32,730] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:42:38,597] loop.py:45 INFO TRAIN_STEP step=13068 loss=1.154772 Training: 73%|███████▎ | 13068/18018 [28:12:24<9:13:11, 6.71s/step, loss=1.155, acc=0.727, acc_len=1.77, thru=18.5, I=20.1, T=37.8, wait=2.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:45,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:42:46,462] loop.py:45 INFO TRAIN_STEP step=13069 loss=1.201856 Training: 73%|███████▎ | 13069/18018 [28:12:32<9:41:50, 7.05s/step, loss=1.202, acc=0.719, acc_len=1.74, thru=17.5, I=17.8, T=35.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:46,496] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:52,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:42:53,949] loop.py:380 INFO TIMING step=13070: step=3.656s data=1.537s compute=3.631s [fwd=2.632s bwd=0.985s opt=0.003s] dispatch=3.820s [2026-10-19 17:42:53,950] loop.py:45 INFO TRAIN_STEP step=13070 loss=1.338728 Training: 73%|███████▎ | 13070/18018 [28:12:40<9:52:27, 7.18s/step, loss=1.339, acc=0.690, acc_len=1.65, thru=17.2, I=19.1, T=35.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:54,204] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:42:58,822] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:43:00,310] loop.py:45 INFO TRAIN_STEP step=13071 loss=1.335349 Training: 73%|███████▎ | 13071/18018 [28:12:46<9:31:53, 6.94s/step, loss=1.335, acc=0.691, acc_len=1.65, thru=19.4, I=26.1, T=36.6, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:43:00,305] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13070: forward=2526.6ms backward=922.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:43:00,337] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 17:43:07,550] loop.py:45 INFO TRAIN_STEP step=13072 loss=1.310125 Training: 73%|███████▎ | 13072/18018 [28:12:53<9:39:19, 7.03s/step, loss=1.310, acc=0.698, acc_len=1.67, thru=16.4, I=19.2, T=37.3, wait=3.8s, pool=57, epoch=2/2][2026-10-19 17:43:15,002] loop.py:45 INFO TRAIN_STEP step=13073 loss=1.318268 Training: 73%|███████▎ | 13073/18018 [28:13:01<9:49:39, 7.15s/step, loss=1.318, acc=0.697, acc_len=1.66, thru=18.5, I=17.0, T=35.4, wait=3.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:43:20,942] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:43:22,273] loop.py:45 INFO TRAIN_STEP step=13074 loss=1.227388 Training: 73%|███████▎ | 13074/18018 [28:13:08<9:52:27, 7.19s/step, loss=1.227, acc=0.713, acc_len=1.73, thru=18.7, I=19.5, T=35.1, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:43:22,451] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:43:29,839] loop.py:380 INFO TIMING step=13075: step=3.536s data=1.474s compute=3.502s [fwd=2.514s bwd=0.972s opt=0.003s] dispatch=4.019s [2026-10-19 17:43:29,839] loop.py:45 INFO TRAIN_STEP step=13075 loss=1.223493 Training: 73%|███████▎ | 13075/18018 [28:13:15<10:01:36, 7.30s/step, loss=1.223, acc=0.712, acc_len=1.72, thru=14.1, I=16.5, T=36.2, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:43:38,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:43:38,374] loop.py:45 INFO TRAIN_STEP step=13076 loss=1.241951 Training: 73%|███████▎ | 13076/18018 [28:13:24<10:31:55, 7.67s/step, loss=1.242, acc=0.709, acc_len=1.71, thru=17.2, I=15.7, T=35.9, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:43:38,370] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13075: forward=2539.9ms backward=983.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:43:38,746] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:43:45,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:43:47,203] loop.py:45 INFO TRAIN_STEP step=13077 loss=1.256901 Training: 73%|███████▎ | 13077/18018 [28:13:33<11:00:25, 8.02s/step, loss=1.257, acc=0.709, acc_len=1.71, thru=14.6, I=16.8, T=34.8, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:43:47,258] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:43:54,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:43:54,871] loop.py:45 INFO TRAIN_STEP step=13078 loss=1.351771 Training: 73%|███████▎ | 13078/18018 [28:13:40<10:51:31, 7.91s/step, loss=1.352, acc=0.685, acc_len=1.63, thru=15.3, I=19.3, T=35.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:43:55,284] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:03,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:44:04,269] loop.py:45 INFO TRAIN_STEP step=13079 loss=1.343213 Training: 73%|███████▎ | 13079/18018 [28:13:50<11:28:07, 8.36s/step, loss=1.343, acc=0.699, acc_len=1.68, thru=14.2, I=14.3, T=34.1, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:04,352] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:10,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:44:10,887] loop.py:380 INFO TIMING step=13080: step=3.652s data=1.591s compute=3.619s [fwd=2.594s bwd=1.010s opt=0.003s] dispatch=2.956s [2026-10-19 17:44:10,888] loop.py:45 INFO TRAIN_STEP step=13080 loss=1.278148 Training: 73%|███████▎ | 13080/18018 [28:13:56<10:44:58, 7.84s/step, loss=1.278, acc=0.707, acc_len=1.71, thru=18.8, I=20.6, T=35.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:11,112] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:17,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 17:44:17,306] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13080: forward=2487.5ms backward=1002.0ms [2026-10-19 17:44:17,310] loop.py:45 INFO TRAIN_STEP step=13081 loss=1.148337 Training: 73%|███████▎ | 13081/18018 [28:14:03<10:09:58, 7.41s/step, loss=1.148, acc=0.727, acc_len=1.78, thru=19.3, I=22.1, T=36.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:17,564] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:44:24,880] loop.py:45 INFO TRAIN_STEP step=13082 loss=1.112357 Training: 73%|███████▎ | 13082/18018 [28:14:10<10:13:41, 7.46s/step, loss=1.112, acc=0.732, acc_len=1.79, thru=14.3, I=17.8, T=36.0, wait=4.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:32,492] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:44:34,108] loop.py:45 INFO TRAIN_STEP step=13083 loss=1.242289 Training: 73%|███████▎ | 13083/18018 [28:14:20<10:57:10, 7.99s/step, loss=1.242, acc=0.710, acc_len=1.72, thru=16.6, I=16.5, T=35.5, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:34,506] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:39,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:44:40,287] loop.py:45 INFO TRAIN_STEP step=13084 loss=1.329082 Training: 73%|███████▎ | 13084/18018 [28:14:26<10:12:21, 7.45s/step, loss=1.329, acc=0.696, acc_len=1.67, thru=19.9, I=25.9, T=37.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:40,720] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:46,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:44:47,237] loop.py:380 INFO TIMING step=13085: step=3.464s data=1.430s compute=3.437s [fwd=2.449s bwd=0.974s opt=0.003s] dispatch=3.476s [2026-10-19 17:44:47,237] loop.py:45 INFO TRAIN_STEP step=13085 loss=1.211352 Training: 73%|███████▎ | 13085/18018 [28:14:33<10:00:00, 7.30s/step, loss=1.211, acc=0.716, acc_len=1.73, thru=19.8, I=20.9, T=36.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:47,438] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:54,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:44:55,250] loop.py:45 INFO TRAIN_STEP step=13086 loss=1.410580 Training: 73%|███████▎ | 13086/18018 [28:14:41<10:17:32, 7.51s/step, loss=1.411, acc=0.681, acc_len=1.62, thru=17.2, I=17.7, T=36.5, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:44:55,246] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13085: forward=2485.0ms backward=980.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:44:55,317] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 17:45:04,041] loop.py:45 INFO TRAIN_STEP step=13087 loss=1.356945 Training: 73%|███████▎ | 13087/18018 [28:14:50<10:48:54, 7.90s/step, loss=1.357, acc=0.692, acc_len=1.66, thru=12.4, I=14.9, T=33.5, wait=5.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:11,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:45:12,956] loop.py:45 INFO TRAIN_STEP step=13088 loss=1.208876 Training: 73%|███████▎ | 13088/18018 [28:14:59<11:13:55, 8.20s/step, loss=1.209, acc=0.720, acc_len=1.75, thru=16.2, I=16.4, T=32.0, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:12,986] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:20,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:45:20,955] loop.py:45 INFO TRAIN_STEP step=13089 loss=1.333620 Training: 73%|███████▎ | 13089/18018 [28:15:07<11:08:45, 8.14s/step, loss=1.334, acc=0.697, acc_len=1.67, thru=16.0, I=18.9, T=32.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:21,273] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:28,705] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:45:29,128] loop.py:380 INFO TIMING step=13090: step=3.738s data=1.594s compute=3.687s [fwd=2.586s bwd=1.085s opt=0.003s] dispatch=4.427s [2026-10-19 17:45:29,129] loop.py:45 INFO TRAIN_STEP step=13090 loss=1.459428 Training: 73%|███████▎ | 13090/18018 [28:15:15<11:09:28, 8.15s/step, loss=1.459, acc=0.671, acc_len=1.58, thru=15.2, I=16.4, T=34.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:29,219] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 17:45:35,128] loop.py:45 INFO TRAIN_STEP step=13091 loss=1.439380 Training: 73%|███████▎ | 13091/18018 [28:15:21<10:16:17, 7.50s/step, loss=1.439, acc=0.674, acc_len=1.59, thru=20.1, I=19.0, T=37.5, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:45:35,124] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13090: forward=2444.2ms backward=928.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:35,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:35,624] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:40,854] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:45:41,612] loop.py:45 INFO TRAIN_STEP step=13092 loss=1.323737 Training: 73%|███████▎ | 13092/18018 [28:15:27<9:51:03, 7.20s/step, loss=1.324, acc=0.696, acc_len=1.67, thru=19.7, I=22.0, T=35.7, wait=2.9s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:41,860] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:47,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:45:48,374] loop.py:45 INFO TRAIN_STEP step=13093 loss=1.207728 Training: 73%|███████▎ | 13093/18018 [28:15:34<9:40:10, 7.07s/step, loss=1.208, acc=0.712, acc_len=1.72, thru=18.5, I=21.9, T=34.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:48,471] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:54,485] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:45:55,302] loop.py:45 INFO TRAIN_STEP step=13094 loss=1.334687 Training: 73%|███████▎ | 13094/18018 [28:15:41<9:36:33, 7.03s/step, loss=1.335, acc=0.693, acc_len=1.65, thru=18.8, I=21.4, T=36.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:55,490] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:45:59,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:46:01,135] loop.py:380 INFO TIMING step=13095: step=4.036s data=1.417s compute=3.997s [fwd=3.024s bwd=0.959s opt=0.003s] dispatch=1.789s [2026-10-19 17:46:01,135] loop.py:45 INFO TRAIN_STEP step=13095 loss=1.120328 Training: 73%|███████▎ | 13095/18018 [28:15:47<9:07:09, 6.67s/step, loss=1.120, acc=0.733, acc_len=1.80, thru=21.2, I=27.5, T=31.7, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:01,294] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:46:08,034] loop.py:45 INFO TRAIN_STEP step=13096 loss=1.319419 Training: 73%|███████▎ | 13096/18018 [28:15:54<9:12:39, 6.74s/step, loss=1.319, acc=0.694, acc_len=1.66, thru=16.8, I=20.1, T=36.0, wait=3.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:46:08,030] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13095: forward=2482.7ms backward=1019.7ms [2026-10-19 17:46:17,413] loop.py:45 INFO TRAIN_STEP step=13097 loss=1.301050 Training: 73%|███████▎ | 13097/18018 [28:16:03<10:17:33, 7.53s/step, loss=1.301, acc=0.705, acc_len=1.69, thru=13.3, I=14.7, T=33.0, wait=5.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:25,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:46:25,270] loop.py:45 INFO TRAIN_STEP step=13098 loss=1.353948 Training: 73%|███████▎ | 13098/18018 [28:16:11<10:25:30, 7.63s/step, loss=1.354, acc=0.691, acc_len=1.65, thru=17.9, I=18.0, T=31.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:25,539] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:31,797] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:46:32,503] loop.py:45 INFO TRAIN_STEP step=13099 loss=1.342200 Training: 73%|███████▎ | 13099/18018 [28:16:18<10:15:39, 7.51s/step, loss=1.342, acc=0.692, acc_len=1.66, thru=17.0, I=18.4, T=36.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:32,801] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:37,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:46:38,427] loop.py:380 INFO TIMING step=13100: step=3.679s data=1.586s compute=3.639s [fwd=2.577s bwd=1.049s opt=0.003s] dispatch=2.237s [2026-10-19 17:46:38,427] loop.py:45 INFO TRAIN_STEP step=13100 loss=1.149616 Training: 73%|███████▎ | 13100/18018 [28:16:24<9:36:31, 7.03s/step, loss=1.150, acc=0.725, acc_len=1.76, thru=21.3, I=25.2, T=34.8, wait=2.2s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:38,456] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:43,919] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:46:44,123] loop.py:45 INFO TRAIN_STEP step=13101 loss=1.150179 Training: 73%|███████▎ | 13101/18018 [28:16:30<9:03:30, 6.63s/step, loss=1.150, acc=0.725, acc_len=1.76, thru=21.5, I=23.2, T=34.7, wait=2.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:46:44,119] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13100: forward=2623.2ms backward=1022.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:44,419] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:49,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:46:50,436] loop.py:45 INFO TRAIN_STEP step=13102 loss=1.213779 Training: 73%|███████▎ | 13102/18018 [28:16:36<8:55:35, 6.54s/step, loss=1.214, acc=0.718, acc_len=1.74, thru=18.8, I=21.5, T=34.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:50,649] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:55,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:46:56,326] loop.py:45 INFO TRAIN_STEP step=13103 loss=1.394957 Training: 73%|███████▎ | 13103/18018 [28:16:42<8:39:34, 6.34s/step, loss=1.395, acc=0.683, acc_len=1.62, thru=20.8, I=24.3, T=36.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:46:56,669] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:47:02,391] loop.py:45 INFO TRAIN_STEP step=13104 loss=1.349422 Training: 73%|███████▎ | 13104/18018 [28:16:48<8:32:41, 6.26s/step, loss=1.349, acc=0.693, acc_len=1.66, thru=20.0, I=22.6, T=37.9, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:07,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:47:08,729] loop.py:380 INFO TIMING step=13105: step=3.510s data=1.398s compute=3.479s [fwd=2.499s bwd=0.965s opt=0.003s] dispatch=2.816s [2026-10-19 17:47:08,729] loop.py:45 INFO TRAIN_STEP step=13105 loss=1.294626 Training: 73%|███████▎ | 13105/18018 [28:16:54<8:34:28, 6.28s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=20.1, I=23.3, T=36.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:08,765] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 17:47:15,118] loop.py:45 INFO TRAIN_STEP step=13106 loss=1.172391 Training: 73%|███████▎ | 13106/18018 [28:17:01<8:36:57, 6.31s/step, loss=1.172, acc=0.723, acc_len=1.76, thru=16.5, I=20.8, T=36.4, wait=2.9s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:47:15,113] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13105: forward=2487.3ms backward=990.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:25,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:47:26,349] loop.py:45 INFO TRAIN_STEP step=13107 loss=1.347342 Training: 73%|███████▎ | 13107/18018 [28:17:12<10:37:36, 7.79s/step, loss=1.347, acc=0.690, acc_len=1.65, thru=14.3, I=15.6, T=31.2, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:26,390] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:32,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:47:32,774] loop.py:45 INFO TRAIN_STEP step=13108 loss=1.258531 Training: 73%|███████▎ | 13108/18018 [28:17:18<10:03:55, 7.38s/step, loss=1.259, acc=0.708, acc_len=1.71, thru=18.9, I=21.6, T=36.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:32,997] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:37,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:47:38,884] loop.py:45 INFO TRAIN_STEP step=13109 loss=1.192844 Training: 73%|███████▎ | 13109/18018 [28:17:24<9:32:36, 7.00s/step, loss=1.193, acc=0.718, acc_len=1.74, thru=20.4, I=24.1, T=36.6, wait=2.6s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:39,260] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:46,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:47:47,841] loop.py:380 INFO TIMING step=13110: step=3.812s data=1.664s compute=3.757s [fwd=2.644s bwd=1.098s opt=0.003s] dispatch=5.138s [2026-10-19 17:47:47,842] loop.py:45 INFO TRAIN_STEP step=13110 loss=1.432175 Training: 73%|███████▎ | 13110/18018 [28:17:33<10:20:33, 7.59s/step, loss=1.432, acc=0.679, acc_len=1.61, thru=14.9, I=18.3, T=33.6, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:48,157] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 17:47:54,317] loop.py:45 INFO TRAIN_STEP step=13111 loss=1.298102 Training: 73%|███████▎ | 13111/18018 [28:17:40<9:53:09, 7.25s/step, loss=1.298, acc=0.700, acc_len=1.68, thru=17.7, I=22.3, T=37.4, wait=3.0s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:47:54,313] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13110: forward=2434.5ms backward=945.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:54,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:54,791] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:47:59,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:48:00,762] loop.py:45 INFO TRAIN_STEP step=13112 loss=1.294323 Training: 73%|███████▎ | 13112/18018 [28:17:46<9:33:12, 7.01s/step, loss=1.294, acc=0.701, acc_len=1.69, thru=19.8, I=22.5, T=37.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:01,241] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 17:48:07,523] loop.py:45 INFO TRAIN_STEP step=13113 loss=1.396130 Training: 73%|███████▎ | 13113/18018 [28:17:53<9:27:01, 6.94s/step, loss=1.396, acc=0.684, acc_len=1.62, thru=18.3, I=24.4, T=36.3, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:16,953] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:48:17,146] loop.py:45 INFO TRAIN_STEP step=13114 loss=1.254599 Training: 73%|███████▎ | 13114/18018 [28:18:03<10:32:46, 7.74s/step, loss=1.255, acc=0.706, acc_len=1.70, thru=15.4, I=13.6, T=34.5, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:17,455] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:21,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:48:23,081] loop.py:380 INFO TIMING step=13115: step=3.556s data=1.513s compute=3.521s [fwd=2.534s bwd=0.973s opt=0.003s] dispatch=2.372s [2026-10-19 17:48:23,081] loop.py:45 INFO TRAIN_STEP step=13115 loss=1.338814 Training: 73%|███████▎ | 13115/18018 [28:18:09<9:48:23, 7.20s/step, loss=1.339, acc=0.695, acc_len=1.66, thru=20.2, I=21.2, T=36.0, wait=2.4s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:23,467] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:48:29,950] loop.py:45 INFO TRAIN_STEP step=13116 loss=1.352914 Training: 73%|███████▎ | 13116/18018 [28:18:16<9:40:07, 7.10s/step, loss=1.353, acc=0.692, acc_len=1.65, thru=16.6, I=19.4, T=35.3, wait=3.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:48:29,946] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13115: forward=2554.1ms backward=1024.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:35,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:48:36,476] loop.py:45 INFO TRAIN_STEP step=13117 loss=1.294693 Training: 73%|███████▎ | 13117/18018 [28:18:22<9:25:52, 6.93s/step, loss=1.295, acc=0.698, acc_len=1.67, thru=19.4, I=22.0, T=35.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:36,505] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:42,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:48:43,668] loop.py:45 INFO TRAIN_STEP step=13118 loss=1.330860 Training: 73%|███████▎ | 13118/18018 [28:18:29<9:32:19, 7.01s/step, loss=1.331, acc=0.692, acc_len=1.65, thru=17.7, I=22.6, T=28.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:43,999] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:48:51,113] loop.py:45 INFO TRAIN_STEP step=13119 loss=1.372019 Training: 73%|███████▎ | 13119/18018 [28:18:37<9:42:53, 7.14s/step, loss=1.372, acc=0.687, acc_len=1.64, thru=13.6, I=15.8, T=36.1, wait=3.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:48:59,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:49:01,041] loop.py:380 INFO TIMING step=13120: step=3.779s data=1.592s compute=3.750s [fwd=2.711s bwd=1.024s opt=0.003s] dispatch=6.138s [2026-10-19 17:49:01,041] loop.py:45 INFO TRAIN_STEP step=13120 loss=1.316490 Training: 73%|███████▎ | 13120/18018 [28:18:47<10:51:03, 7.98s/step, loss=1.316, acc=0.698, acc_len=1.67, thru=16.4, I=15.6, T=33.9, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:01,294] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:49:08,919] loop.py:45 INFO TRAIN_STEP step=13121 loss=1.272492 Training: 73%|███████▎ | 13121/18018 [28:18:55<10:48:35, 7.95s/step, loss=1.272, acc=0.707, acc_len=1.70, thru=13.3, I=16.1, T=30.3, wait=3.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:49:08,914] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13120: forward=3202.1ms backward=983.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:14,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:49:15,263] loop.py:45 INFO TRAIN_STEP step=13122 loss=1.264849 Training: 73%|███████▎ | 13122/18018 [28:19:01<10:09:11, 7.47s/step, loss=1.265, acc=0.707, acc_len=1.71, thru=18.5, I=19.7, T=35.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:15,633] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:21,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:49:24,059] loop.py:45 INFO TRAIN_STEP step=13123 loss=1.252539 Training: 73%|███████▎ | 13123/18018 [28:19:10<10:41:36, 7.86s/step, loss=1.253, acc=0.706, acc_len=1.70, thru=14.0, I=18.6, T=29.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:24,143] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:29,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:49:29,927] loop.py:45 INFO TRAIN_STEP step=13124 loss=1.194312 Training: 73%|███████▎ | 13124/18018 [28:19:16<9:52:37, 7.27s/step, loss=1.194, acc=0.716, acc_len=1.74, thru=18.0, I=25.5, T=38.6, wait=2.5s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:30,010] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:36,750] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:49:38,062] loop.py:380 INFO TIMING step=13125: step=4.025s data=1.754s compute=3.997s [fwd=2.860s bwd=1.123s opt=0.003s] dispatch=4.102s [2026-10-19 17:49:38,062] loop.py:45 INFO TRAIN_STEP step=13125 loss=1.279005 Training: 73%|███████▎ | 13125/18018 [28:19:24<10:13:48, 7.53s/step, loss=1.279, acc=0.706, acc_len=1.70, thru=16.9, I=19.7, T=31.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:38,258] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:42,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 17:49:44,144] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13125: forward=2810.4ms backward=1055.1ms [2026-10-19 17:49:44,149] loop.py:45 INFO TRAIN_STEP step=13126 loss=1.298205 Training: 73%|███████▎ | 13126/18018 [28:19:30<9:38:28, 7.09s/step, loss=1.298, acc=0.699, acc_len=1.67, thru=19.4, I=23.7, T=32.7, wait=2.2s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:44,655] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:50,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:49:51,919] loop.py:45 INFO TRAIN_STEP step=13127 loss=1.330880 Training: 73%|███████▎ | 13127/18018 [28:19:38<9:54:52, 7.30s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=14.8, I=20.9, T=36.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:52,121] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:58,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:49:59,462] loop.py:45 INFO TRAIN_STEP step=13128 loss=1.176142 Training: 73%|███████▎ | 13128/18018 [28:19:45<10:00:44, 7.37s/step, loss=1.176, acc=0.726, acc_len=1.77, thru=16.6, I=20.8, T=34.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:49:59,766] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:50:05,488] loop.py:45 INFO TRAIN_STEP step=13129 loss=1.240230 Training: 73%|███████▎ | 13129/18018 [28:19:51<9:27:40, 6.97s/step, loss=1.240, acc=0.710, acc_len=1.72, thru=19.6, I=20.1, T=35.2, wait=2.4s, pool=61, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:11,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:50:11,643] loop.py:380 INFO TIMING step=13130: step=3.392s data=1.343s compute=3.358s [fwd=2.399s bwd=0.944s opt=0.003s] dispatch=2.755s [2026-10-19 17:50:11,644] loop.py:45 INFO TRAIN_STEP step=13130 loss=1.231898 Training: 73%|███████▎ | 13130/18018 [28:19:57<9:07:44, 6.72s/step, loss=1.232, acc=0.712, acc_len=1.72, thru=20.5, I=20.6, T=37.7, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:11,942] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:50:18,806] loop.py:45 INFO TRAIN_STEP step=13131 loss=1.300520 (TrainerActor pid=9310) [2026-10-19 17:50:18,801] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13130: forward=2645.6ms backward=1083.9ms Training: 73%|███████▎ | 13131/18018 [28:20:04<9:18:20, 6.86s/step, loss=1.301, acc=0.698, acc_len=1.68, thru=18.0, I=19.5, T=33.8, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:25,982] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:50:26,733] loop.py:45 INFO TRAIN_STEP step=13132 loss=1.164062 Training: 73%|███████▎ | 13132/18018 [28:20:12<9:44:29, 7.18s/step, loss=1.164, acc=0.725, acc_len=1.77, thru=16.7, I=17.2, T=37.5, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:26,986] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:33,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:50:33,277] loop.py:45 INFO TRAIN_STEP step=13133 loss=1.337602 Training: 73%|███████▎ | 13133/18018 [28:20:19<9:28:50, 6.99s/step, loss=1.338, acc=0.691, acc_len=1.65, thru=19.6, I=21.6, T=38.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:33,632] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:39,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:50:39,535] loop.py:45 INFO TRAIN_STEP step=13134 loss=1.309038 Training: 73%|███████▎ | 13134/18018 [28:20:25<9:10:55, 6.77s/step, loss=1.309, acc=0.696, acc_len=1.67, thru=20.1, I=21.8, T=37.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:39,930] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:46,638] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:50:47,728] loop.py:380 INFO TIMING step=13135: step=3.741s data=1.603s compute=3.709s [fwd=2.676s bwd=1.019s opt=0.003s] dispatch=4.444s [2026-10-19 17:50:47,728] loop.py:45 INFO TRAIN_STEP step=13135 loss=1.364737 Training: 73%|███████▎ | 13135/18018 [28:20:33<9:45:37, 7.20s/step, loss=1.365, acc=0.688, acc_len=1.64, thru=16.6, I=18.8, T=34.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:50:48,146] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:50:53,952] loop.py:45 INFO TRAIN_STEP step=13136 loss=1.263082 Training: 73%|███████▎ | 13136/18018 [28:20:40<9:21:46, 6.90s/step, loss=1.263, acc=0.705, acc_len=1.69, thru=17.8, I=20.4, T=39.3, wait=3.0s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:50:53,947] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13135: forward=2305.2ms backward=905.1ms [2026-10-19 17:51:01,890] loop.py:45 INFO TRAIN_STEP step=13137 loss=1.351510 Training: 73%|███████▎ | 13137/18018 [28:20:47<9:46:52, 7.21s/step, loss=1.352, acc=0.696, acc_len=1.66, thru=16.9, I=15.6, T=35.1, wait=4.3s, pool=49, epoch=2/2][2026-10-19 17:51:09,143] loop.py:45 INFO TRAIN_STEP step=13138 loss=1.272707 Training: 73%|███████▎ | 13138/18018 [28:20:55<9:47:45, 7.23s/step, loss=1.273, acc=0.704, acc_len=1.69, thru=15.6, I=17.0, T=33.6, wait=3.4s, pool=41, epoch=2/2][2026-10-19 17:51:17,327] loop.py:45 INFO TRAIN_STEP step=13139 loss=1.286087 Training: 73%|███████▎ | 13139/18018 [28:21:03<10:10:57, 7.51s/step, loss=1.286, acc=0.702, acc_len=1.69, thru=17.3, I=17.8, T=31.4, wait=4.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:22,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:51:24,253] loop.py:380 INFO TIMING step=13140: step=3.637s data=1.515s compute=3.599s [fwd=2.562s bwd=1.022s opt=0.003s] dispatch=3.281s [2026-10-19 17:51:24,254] loop.py:45 INFO TRAIN_STEP step=13140 loss=1.206638 Training: 73%|███████▎ | 13140/18018 [28:21:10<9:56:29, 7.34s/step, loss=1.207, acc=0.715, acc_len=1.75, thru=17.7, I=21.3, T=35.2, wait=3.3s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:24,472] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:28,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:51:29,795] loop.py:45 INFO TRAIN_STEP step=13141 loss=1.281765 Training: 73%|███████▎ | 13141/18018 [28:21:15<9:12:38, 6.80s/step, loss=1.282, acc=0.702, acc_len=1.69, thru=21.3, I=28.6, T=37.1, wait=2.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:51:29,791] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13140: forward=2482.4ms backward=930.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:29,850] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:34,641] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:51:36,503] loop.py:45 INFO TRAIN_STEP step=13142 loss=1.267444 Training: 73%|███████▎ | 13142/18018 [28:21:22<9:10:17, 6.77s/step, loss=1.267, acc=0.705, acc_len=1.70, thru=18.9, I=24.4, T=34.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:36,662] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:42,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:51:43,061] loop.py:45 INFO TRAIN_STEP step=13143 loss=1.198853 Training: 73%|███████▎ | 13143/18018 [28:21:29<9:04:59, 6.71s/step, loss=1.199, acc=0.718, acc_len=1.74, thru=18.8, I=24.2, T=37.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:43,123] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:48,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:51:48,754] loop.py:45 INFO TRAIN_STEP step=13144 loss=1.278923 Training: 73%|███████▎ | 13144/18018 [28:21:34<8:40:06, 6.40s/step, loss=1.279, acc=0.706, acc_len=1.69, thru=21.2, I=23.5, T=37.5, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:48,957] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:55,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:51:56,925] loop.py:380 INFO TIMING step=13145: step=3.712s data=1.683s compute=3.677s [fwd=2.658s bwd=1.004s opt=0.003s] dispatch=4.450s [2026-10-19 17:51:56,925] loop.py:45 INFO TRAIN_STEP step=13145 loss=1.297670 Training: 73%|███████▎ | 13145/18018 [28:21:43<9:23:05, 6.93s/step, loss=1.298, acc=0.704, acc_len=1.70, thru=15.3, I=17.9, T=34.5, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:51:57,136] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:52:02,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:52:03,723] loop.py:45 INFO TRAIN_STEP step=13146 loss=1.313046 Training: 73%|███████▎ | 13146/18018 [28:21:49<9:19:43, 6.89s/step, loss=1.313, acc=0.701, acc_len=1.68, thru=18.0, I=23.0, T=34.6, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:52:03,718] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13145: forward=2620.3ms backward=1032.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:52:03,945] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:52:11,981] loop.py:45 INFO TRAIN_STEP step=13147 loss=1.254141 Training: 73%|███████▎ | 13147/18018 [28:21:58<9:52:50, 7.30s/step, loss=1.254, acc=0.707, acc_len=1.71, thru=14.1, I=16.5, T=34.3, wait=4.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:52:19,669] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:52:20,539] loop.py:45 INFO TRAIN_STEP step=13148 loss=1.147656 Training: 73%|███████▎ | 13148/18018 [28:22:06<10:23:15, 7.68s/step, loss=1.148, acc=0.724, acc_len=1.76, thru=15.7, I=16.8, T=34.5, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:52:20,682] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:52:27,017] loop.py:45 INFO TRAIN_STEP step=13149 loss=1.211493 Training: 73%|███████▎ | 13149/18018 [28:22:13<9:53:54, 7.32s/step, loss=1.211, acc=0.711, acc_len=1.71, thru=18.4, I=22.3, T=31.3, wait=2.4s, pool=61, epoch=2/2] [2026-10-19 17:52:33,610] loop.py:380 INFO TIMING step=13150: step=3.548s data=1.477s compute=3.521s [fwd=2.528s bwd=0.978s opt=0.003s] dispatch=3.037s [2026-10-19 17:52:33,610] loop.py:45 INFO TRAIN_STEP step=13150 loss=1.287823 Training: 73%|███████▎ | 13150/18018 [28:22:19<9:36:09, 7.10s/step, loss=1.288, acc=0.699, acc_len=1.68, thru=13.8, I=17.0, T=36.1, wait=3.0s, pool=25, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 17:52:33,701] eagle_store.py:517 WARNING batch_get_buffer missing keys (k420798-9b9bc106_hs, k420798-9b9bc106_ids, k420798-9b9bc106_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:52:40,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 17:52:33,690] eagle_store.py:517 WARNING batch_get_buffer missing keys (k420796-9a2d4bb8_hs, k420796-9a2d4bb8_ids, k420796-9a2d4bb8_lhs); sleeping 0.50s. [2026-10-19 17:52:41,760] loop.py:45 INFO TRAIN_STEP step=13151 loss=1.317186 Training: 73%|███████▎ | 13151/18018 [28:22:27<10:01:31, 7.42s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=18.4, I=18.3, T=32.6, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:52:41,755] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13150: forward=2791.6ms backward=1085.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:52:42,212] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 17:52:52,880] loop.py:45 INFO TRAIN_STEP step=13152 loss=1.059185 Training: 73%|███████▎ | 13152/18018 [28:22:38<11:31:36, 8.53s/step, loss=1.059, acc=0.744, acc_len=1.83, thru=11.1, I=14.1, T=35.4, wait=7.5s, pool=61, epoch=2/2][2026-10-19 17:52:59,210] loop.py:45 INFO TRAIN_STEP step=13153 loss=1.335208 Training: 73%|███████▎ | 13153/18018 [28:22:45<10:37:53, 7.87s/step, loss=1.335, acc=0.695, acc_len=1.66, thru=18.4, I=19.6, T=35.9, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:53:06,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:53:06,107] loop.py:45 INFO TRAIN_STEP step=13154 loss=1.261092 Training: 73%|███████▎ | 13154/18018 [28:22:52<10:14:09, 7.58s/step, loss=1.261, acc=0.709, acc_len=1.71, thru=19.4, I=21.3, T=33.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:53:06,503] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:53:13,320] loop.py:380 INFO TIMING step=13155: step=3.742s data=1.558s compute=3.716s [fwd=2.645s bwd=1.057s opt=0.003s] dispatch=3.464s [2026-10-19 17:53:13,321] loop.py:45 INFO TRAIN_STEP step=13155 loss=1.209079 Training: 73%|███████▎ | 13155/18018 [28:22:59<10:05:20, 7.47s/step, loss=1.209, acc=0.718, acc_len=1.74, thru=15.0, I=16.1, T=34.2, wait=3.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:53:22,357] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 17:53:22,614] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13155: forward=2646.1ms backward=1065.2ms [2026-10-19 17:53:22,618] loop.py:45 INFO TRAIN_STEP step=13156 loss=1.387582 Training: 73%|███████▎ | 13156/18018 [28:23:08<10:49:36, 8.02s/step, loss=1.388, acc=0.683, acc_len=1.63, thru=16.1, I=15.2, T=34.1, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:53:22,859] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:53:32,147] loop.py:45 INFO TRAIN_STEP step=13157 loss=1.233448 Training: 73%|███████▎ | 13157/18018 [28:23:18<11:26:10, 8.47s/step, loss=1.233, acc=0.714, acc_len=1.73, thru=12.0, I=11.8, T=33.9, wait=5.7s, pool=49, epoch=2/2][2026-10-19 17:53:39,636] loop.py:45 INFO TRAIN_STEP step=13158 loss=1.242619 Training: 73%|███████▎ | 13158/18018 [28:23:25<11:02:14, 8.18s/step, loss=1.243, acc=0.709, acc_len=1.71, thru=16.3, I=15.9, T=37.2, wait=4.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:53:49,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:53:50,022] loop.py:45 INFO TRAIN_STEP step=13159 loss=1.259052 Training: 73%|███████▎ | 13159/18018 [28:23:36<11:55:46, 8.84s/step, loss=1.259, acc=0.709, acc_len=1.71, thru=15.5, I=13.8, T=33.6, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:53:50,302] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:53:56,408] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:53:57,822] loop.py:380 INFO TIMING step=13160: step=3.748s data=1.574s compute=3.711s [fwd=2.662s bwd=1.035s opt=0.003s] dispatch=4.044s [2026-10-19 17:53:57,822] loop.py:45 INFO TRAIN_STEP step=13160 loss=1.309888 Training: 73%|███████▎ | 13160/18018 [28:23:43<11:30:26, 8.53s/step, loss=1.310, acc=0.697, acc_len=1.67, thru=16.0, I=19.3, T=34.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:53:57,930] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:03,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:54:03,408] loop.py:45 INFO TRAIN_STEP step=13161 loss=1.289943 Training: 73%|███████▎ | 13161/18018 [28:23:49<10:18:51, 7.64s/step, loss=1.290, acc=0.702, acc_len=1.69, thru=20.6, I=22.1, T=38.5, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:54:03,404] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13160: forward=2334.6ms backward=936.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:03,813] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:10,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:54:10,532] loop.py:45 INFO TRAIN_STEP step=13162 loss=1.352191 Training: 73%|███████▎ | 13162/18018 [28:23:56<10:06:04, 7.49s/step, loss=1.352, acc=0.689, acc_len=1.64, thru=18.0, I=19.4, T=35.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:10,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:15,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:54:16,684] loop.py:45 INFO TRAIN_STEP step=13163 loss=1.308175 Training: 73%|███████▎ | 13163/18018 [28:24:02<9:33:32, 7.09s/step, loss=1.308, acc=0.699, acc_len=1.67, thru=20.6, I=23.7, T=33.2, wait=2.3s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:16,970] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:25,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:54:25,645] loop.py:45 INFO TRAIN_STEP step=13164 loss=1.265766 Training: 73%|███████▎ | 13164/18018 [28:24:11<10:18:52, 7.65s/step, loss=1.266, acc=0.709, acc_len=1.71, thru=13.6, I=15.8, T=32.8, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:26,014] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:54:33,220] loop.py:380 INFO TIMING step=13165: step=3.750s data=1.645s compute=3.715s [fwd=2.702s bwd=0.999s opt=0.003s] dispatch=3.815s [2026-10-19 17:54:33,220] loop.py:45 INFO TRAIN_STEP step=13165 loss=1.212219 Training: 73%|███████▎ | 13165/18018 [28:24:19<10:16:51, 7.63s/step, loss=1.212, acc=0.714, acc_len=1.73, thru=16.1, I=16.2, T=34.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:33,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:33,722] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:37,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:54:39,112] loop.py:45 INFO TRAIN_STEP step=13166 loss=1.202235 Training: 73%|███████▎ | 13166/18018 [28:24:25<9:34:43, 7.11s/step, loss=1.202, acc=0.717, acc_len=1.74, thru=20.3, I=23.4, T=36.7, wait=2.4s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 17:54:39,108] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13165: forward=2453.8ms backward=994.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:39,467] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:44,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:54:44,818] loop.py:45 INFO TRAIN_STEP step=13167 loss=1.320619 Training: 73%|███████▎ | 13167/18018 [28:24:30<9:00:35, 6.69s/step, loss=1.321, acc=0.696, acc_len=1.67, thru=22.0, I=27.0, T=39.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:45,142] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:50,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:54:51,118] loop.py:45 INFO TRAIN_STEP step=13168 loss=1.263189 Training: 73%|███████▎ | 13168/18018 [28:24:37<8:51:08, 6.57s/step, loss=1.263, acc=0.704, acc_len=1.70, thru=20.5, I=23.6, T=38.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:54:51,389] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:54:56,983] loop.py:45 INFO TRAIN_STEP step=13169 loss=1.197365 Training: 73%|███████▎ | 13169/18018 [28:24:43<8:33:51, 6.36s/step, loss=1.197, acc=0.716, acc_len=1.73, thru=19.8, I=22.8, T=38.9, wait=2.6s, pool=45, epoch=2/2][2026-10-19 17:55:06,901] loop.py:380 INFO TIMING step=13170: step=3.914s data=1.572s compute=3.881s [fwd=2.740s bwd=1.126s opt=0.003s] dispatch=5.997s [2026-10-19 17:55:06,902] loop.py:45 INFO TRAIN_STEP step=13170 loss=1.284643 Training: 73%|███████▎ | 13170/18018 [28:24:52<10:00:06, 7.43s/step, loss=1.285, acc=0.706, acc_len=1.70, thru=14.4, I=14.0, T=32.7, wait=6.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:14,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 17:55:14,815] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13170: forward=2658.7ms backward=1098.9ms [2026-10-19 17:55:14,820] loop.py:45 INFO TRAIN_STEP step=13171 loss=1.173146 Training: 73%|███████▎ | 13171/18018 [28:25:00<10:11:49, 7.57s/step, loss=1.173, acc=0.725, acc_len=1.77, thru=16.3, I=15.7, T=33.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:15,174] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:55:20,961] loop.py:45 INFO TRAIN_STEP step=13172 loss=1.322219 Training: 73%|███████▎ | 13172/18018 [28:25:07<9:37:08, 7.15s/step, loss=1.322, acc=0.692, acc_len=1.66, thru=19.2, I=18.0, T=38.4, wait=2.8s, pool=61, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:29,954] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:55:30,916] loop.py:45 INFO TRAIN_STEP step=13173 loss=1.124107 Training: 73%|███████▎ | 13173/18018 [28:25:17<10:45:00, 7.99s/step, loss=1.124, acc=0.732, acc_len=1.79, thru=13.4, I=14.6, T=31.2, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:30,970] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:36,530] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:55:37,245] loop.py:45 INFO TRAIN_STEP step=13174 loss=1.215249 Training: 73%|███████▎ | 13174/18018 [28:25:23<10:04:40, 7.49s/step, loss=1.215, acc=0.715, acc_len=1.74, thru=18.2, I=22.2, T=36.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:37,537] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:42,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:55:43,805] loop.py:380 INFO TIMING step=13175: step=3.407s data=1.337s compute=3.375s [fwd=2.378s bwd=0.983s opt=0.003s] dispatch=3.143s [2026-10-19 17:55:43,805] loop.py:45 INFO TRAIN_STEP step=13175 loss=1.286633 Training: 73%|███████▎ | 13175/18018 [28:25:29<9:42:00, 7.21s/step, loss=1.287, acc=0.702, acc_len=1.69, thru=19.7, I=23.4, T=37.6, wait=3.1s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:44,034] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:55:49,238] loop.py:45 INFO TRAIN_STEP step=13176 loss=1.295257 Training: 73%|███████▎ | 13176/18018 [28:25:35<8:58:49, 6.68s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=23.4, I=26.3, T=40.0, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:55:49,234] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13175: forward=2296.7ms backward=863.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:49,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:49,719] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:54,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:55:55,501] loop.py:45 INFO TRAIN_STEP step=13177 loss=1.200459 Training: 73%|███████▎ | 13177/18018 [28:25:41<8:48:46, 6.55s/step, loss=1.200, acc=0.716, acc_len=1.74, thru=21.3, I=23.0, T=36.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:55:55,593] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 17:56:01,462] loop.py:45 INFO TRAIN_STEP step=13178 loss=1.328646 Training: 73%|███████▎ | 13178/18018 [28:25:47<8:34:15, 6.38s/step, loss=1.329, acc=0.693, acc_len=1.66, thru=19.9, I=22.5, T=39.2, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:56:06,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:56:07,287] loop.py:45 INFO TRAIN_STEP step=13179 loss=1.199828 Training: 73%|███████▎ | 13179/18018 [28:25:53<8:20:53, 6.21s/step, loss=1.200, acc=0.716, acc_len=1.73, thru=21.9, I=22.3, T=37.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:56:07,746] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:56:12,840] loop.py:380 INFO TIMING step=13180: step=3.330s data=1.321s compute=3.300s [fwd=2.364s bwd=0.921s opt=0.003s] dispatch=2.212s [2026-10-19 17:56:12,840] loop.py:45 INFO TRAIN_STEP step=13180 loss=1.209619 Training: 73%|███████▎ | 13180/18018 [28:25:58<8:04:50, 6.01s/step, loss=1.210, acc=0.716, acc_len=1.74, thru=22.0, I=24.5, T=38.4, wait=2.2s, pool=61, epoch=2/2][2026-10-19 17:56:22,801] loop.py:45 INFO TRAIN_STEP step=13181 loss=1.280076 Training: 73%|███████▎ | 13181/18018 [28:26:08<9:40:12, 7.20s/step, loss=1.280, acc=0.703, acc_len=1.69, thru=11.7, I=11.3, T=34.5, wait=6.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:56:22,797] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13180: forward=2635.2ms backward=1031.2ms [2026-10-19 17:56:29,711] loop.py:45 INFO TRAIN_STEP step=13182 loss=1.251876 Training: 73%|███████▎ | 13182/18018 [28:26:15<9:33:08, 7.11s/step, loss=1.252, acc=0.709, acc_len=1.71, thru=18.4, I=15.7, T=38.6, wait=3.6s, pool=53, epoch=2/2][2026-10-19 17:56:38,040] loop.py:45 INFO TRAIN_STEP step=13183 loss=1.399633 Training: 73%|███████▎ | 13183/18018 [28:26:24<10:02:32, 7.48s/step, loss=1.400, acc=0.683, acc_len=1.62, thru=15.4, I=16.5, T=34.8, wait=4.6s, pool=61, epoch=2/2][2026-10-19 17:56:45,890] loop.py:45 INFO TRAIN_STEP step=13184 loss=1.144813 Training: 73%|███████▎ | 13184/18018 [28:26:31<10:11:27, 7.59s/step, loss=1.145, acc=0.727, acc_len=1.78, thru=14.7, I=16.0, T=34.5, wait=4.1s, pool=45, epoch=2/2][2026-10-19 17:56:55,216] loop.py:380 INFO TIMING step=13185: step=3.637s data=1.518s compute=3.607s [fwd=2.576s bwd=1.017s opt=0.003s] dispatch=5.676s [2026-10-19 17:56:55,216] loop.py:45 INFO TRAIN_STEP step=13185 loss=1.242627 Training: 73%|███████▎ | 13185/18018 [28:26:41<10:53:11, 8.11s/step, loss=1.243, acc=0.711, acc_len=1.72, thru=12.4, I=13.4, T=35.2, wait=5.7s, pool=29, epoch=2/2][2026-10-19 17:57:11,941] loop.py:45 INFO TRAIN_STEP step=13186 loss=1.307912 Training: 73%|███████▎ | 13186/18018 [28:26:58<14:21:11, 10.69s/step, loss=1.308, acc=0.699, acc_len=1.68, thru=12.4, I=9.0, T=33.1, wait=12.8s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:57:11,936] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13185: forward=2676.2ms backward=1111.5ms [2026-10-19 17:57:21,237] loop.py:45 INFO TRAIN_STEP step=13187 loss=1.306488 Training: 73%|███████▎ | 13187/18018 [28:27:07<13:47:23, 10.28s/step, loss=1.306, acc=0.697, acc_len=1.67, thru=14.8, I=13.9, T=34.3, wait=5.6s, pool=61, epoch=2/2][2026-10-19 17:57:27,164] loop.py:45 INFO TRAIN_STEP step=13188 loss=1.327142 Training: 73%|███████▎ | 13188/18018 [28:27:13<12:02:05, 8.97s/step, loss=1.327, acc=0.695, acc_len=1.66, thru=18.9, I=17.0, T=37.3, wait=2.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:57:33,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:57:34,000] loop.py:45 INFO TRAIN_STEP step=13189 loss=1.181345 Training: 73%|███████▎ | 13189/18018 [28:27:20<11:10:25, 8.33s/step, loss=1.181, acc=0.722, acc_len=1.76, thru=18.9, I=19.6, T=37.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:57:34,451] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:57:40,209] loop.py:380 INFO TIMING step=13190: step=3.325s data=1.333s compute=3.286s [fwd=2.343s bwd=0.928s opt=0.003s] dispatch=2.876s [2026-10-19 17:57:40,210] loop.py:45 INFO TRAIN_STEP step=13190 loss=1.231831 Training: 73%|███████▎ | 13190/18018 [28:27:26<10:19:07, 7.69s/step, loss=1.232, acc=0.710, acc_len=1.72, thru=16.9, I=23.2, T=38.5, wait=2.9s, pool=33, epoch=2/2][2026-10-19 17:57:49,927] loop.py:45 INFO TRAIN_STEP step=13191 loss=1.253477 Training: 73%|███████▎ | 13191/18018 [28:27:36<11:07:49, 8.30s/step, loss=1.253, acc=0.709, acc_len=1.71, thru=15.8, I=15.7, T=34.7, wait=6.0s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:57:49,922] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13190: forward=2640.6ms backward=994.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:57:58,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:57:58,767] loop.py:45 INFO TRAIN_STEP step=13192 loss=1.283137 Training: 73%|███████▎ | 13192/18018 [28:27:44<11:20:40, 8.46s/step, loss=1.283, acc=0.703, acc_len=1.69, thru=14.5, I=15.2, T=34.6, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:57:59,100] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:04,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:58:04,728] loop.py:45 INFO TRAIN_STEP step=13193 loss=1.157384 Training: 73%|███████▎ | 13193/18018 [28:27:50<10:20:11, 7.71s/step, loss=1.157, acc=0.724, acc_len=1.76, thru=20.5, I=21.7, T=39.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:04,910] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:58:12,131] loop.py:45 INFO TRAIN_STEP step=13194 loss=1.385438 Training: 73%|███████▎ | 13194/18018 [28:27:58<10:12:32, 7.62s/step, loss=1.385, acc=0.688, acc_len=1.63, thru=17.8, I=18.0, T=34.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:12,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:12,621] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:19,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:58:20,475] loop.py:380 INFO TIMING step=13195: step=3.611s data=1.540s compute=3.543s [fwd=2.513s bwd=1.016s opt=0.003s] dispatch=4.727s [2026-10-19 17:58:20,476] loop.py:45 INFO TRAIN_STEP step=13195 loss=1.347096 Training: 73%|███████▎ | 13195/18018 [28:28:06<10:29:59, 7.84s/step, loss=1.347, acc=0.694, acc_len=1.66, thru=16.4, I=17.3, T=35.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:20,717] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:27,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:58:27,965] loop.py:45 INFO TRAIN_STEP step=13196 loss=1.272988 Training: 73%|███████▎ | 13196/18018 [28:28:14<10:21:28, 7.73s/step, loss=1.273, acc=0.705, acc_len=1.70, thru=16.4, I=17.8, T=35.8, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:58:27,961] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13195: forward=2574.7ms backward=955.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:28,107] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:33,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:58:34,553] loop.py:45 INFO TRAIN_STEP step=13197 loss=1.447817 Training: 73%|███████▎ | 13197/18018 [28:28:20<9:53:43, 7.39s/step, loss=1.448, acc=0.671, acc_len=1.58, thru=20.6, I=21.3, T=37.8, wait=3.2s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:34,863] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:58:40,620] loop.py:45 INFO TRAIN_STEP step=13198 loss=1.298623 Training: 73%|███████▎ | 13198/18018 [28:28:26<9:21:42, 6.99s/step, loss=1.299, acc=0.703, acc_len=1.68, thru=19.7, I=23.9, T=37.8, wait=2.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:46,968] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:58:47,686] loop.py:45 INFO TRAIN_STEP step=13199 loss=1.246013 Training: 73%|███████▎ | 13199/18018 [28:28:33<9:23:24, 7.01s/step, loss=1.246, acc=0.709, acc_len=1.71, thru=18.4, I=19.4, T=37.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:58:47,975] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 17:58:55,076] loop.py:380 INFO TIMING step=13200: step=3.624s data=1.414s compute=3.593s [fwd=2.521s bwd=1.057s opt=0.003s] dispatch=3.756s [2026-10-19 17:58:55,077] loop.py:45 INFO TRAIN_STEP step=13200 loss=1.289666 Training: 73%|███████▎ | 13200/18018 [28:28:41<9:32:20, 7.13s/step, loss=1.290, acc=0.704, acc_len=1.69, thru=16.6, I=19.1, T=35.3, wait=3.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:59:00,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:59:01,730] loop.py:45 INFO TRAIN_STEP step=13201 loss=1.346553 Training: 73%|███████▎ | 13201/18018 [28:28:47<9:20:45, 6.98s/step, loss=1.347, acc=0.695, acc_len=1.66, thru=19.3, I=21.3, T=36.3, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:59:01,726] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13200: forward=2487.1ms backward=994.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:59:02,229] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 17:59:08,771] loop.py:45 INFO TRAIN_STEP step=13202 loss=1.280585 Training: 73%|███████▎ | 13202/18018 [28:28:54<9:21:59, 7.00s/step, loss=1.281, acc=0.704, acc_len=1.69, thru=17.0, I=19.5, T=34.5, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:59:15,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:59:15,390] loop.py:45 INFO TRAIN_STEP step=13203 loss=1.235936 Training: 73%|███████▎ | 13203/18018 [28:29:01<9:12:41, 6.89s/step, loss=1.236, acc=0.713, acc_len=1.72, thru=19.4, I=18.2, T=37.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:59:15,702] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 17:59:23,366] loop.py:45 INFO TRAIN_STEP step=13204 loss=1.420462 Training: 73%|███████▎ | 13204/18018 [28:29:09<9:38:55, 7.22s/step, loss=1.420, acc=0.682, acc_len=1.62, thru=16.1, I=17.0, T=32.8, wait=4.1s, pool=53, epoch=2/2][2026-10-19 17:59:31,849] loop.py:380 INFO TIMING step=13205: step=3.728s data=1.638s compute=3.679s [fwd=2.626s bwd=1.038s opt=0.003s] dispatch=4.741s [2026-10-19 17:59:31,849] loop.py:45 INFO TRAIN_STEP step=13205 loss=1.303644 Training: 73%|███████▎ | 13205/18018 [28:29:17<10:09:07, 7.59s/step, loss=1.304, acc=0.700, acc_len=1.68, thru=14.8, I=15.6, T=34.3, wait=4.7s, pool=49, epoch=2/2][2026-10-19 17:59:39,757] loop.py:45 INFO TRAIN_STEP step=13206 loss=1.282758 Training: 73%|███████▎ | 13206/18018 [28:29:25<10:16:35, 7.69s/step, loss=1.283, acc=0.702, acc_len=1.69, thru=16.3, I=15.2, T=34.0, wait=4.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 17:59:39,753] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13205: forward=2636.5ms backward=1081.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:59:45,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:59:46,279] loop.py:45 INFO TRAIN_STEP step=13207 loss=1.255390 Training: 73%|███████▎ | 13207/18018 [28:29:32<9:48:25, 7.34s/step, loss=1.255, acc=0.708, acc_len=1.71, thru=19.3, I=19.4, T=36.1, wait=3.0s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:59:46,730] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:59:52,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 17:59:52,964] loop.py:45 INFO TRAIN_STEP step=13208 loss=1.307493 Training: 73%|███████▎ | 13208/18018 [28:29:39<9:32:35, 7.14s/step, loss=1.307, acc=0.698, acc_len=1.68, thru=18.6, I=22.5, T=37.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 17:59:53,459] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:00:01,324] loop.py:45 INFO TRAIN_STEP step=13209 loss=1.205770 Training: 73%|███████▎ | 13209/18018 [28:29:47<10:01:48, 7.51s/step, loss=1.206, acc=0.716, acc_len=1.74, thru=13.7, I=15.5, T=35.9, wait=4.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:08,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:00:09,556] loop.py:380 INFO TIMING step=13210: step=3.829s data=1.760s compute=3.767s [fwd=2.536s bwd=1.216s opt=0.003s] dispatch=4.392s [2026-10-19 18:00:09,557] loop.py:45 INFO TRAIN_STEP step=13210 loss=1.351186 Training: 73%|███████▎ | 13210/18018 [28:29:55<10:19:02, 7.73s/step, loss=1.351, acc=0.694, acc_len=1.66, thru=16.8, I=18.0, T=33.4, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:09,900] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:16,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:00:16,890] loop.py:45 INFO TRAIN_STEP step=13211 loss=1.318491 Training: 73%|███████▎ | 13211/18018 [28:30:02<10:09:30, 7.61s/step, loss=1.318, acc=0.699, acc_len=1.68, thru=16.9, I=20.7, T=36.0, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:00:16,886] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13210: forward=2483.6ms backward=1008.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:17,233] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:23,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:00:24,586] loop.py:45 INFO TRAIN_STEP step=13212 loss=1.208820 Training: 73%|███████▎ | 13212/18018 [28:30:10<10:11:32, 7.63s/step, loss=1.209, acc=0.716, acc_len=1.74, thru=16.6, I=19.7, T=35.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:24,942] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:31,544] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:00:31,749] loop.py:45 INFO TRAIN_STEP step=13213 loss=1.312929 Training: 73%|███████▎ | 13213/18018 [28:30:17<10:00:00, 7.49s/step, loss=1.313, acc=0.701, acc_len=1.68, thru=18.2, I=19.9, T=38.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:32,047] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:41,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:00:41,658] loop.py:45 INFO TRAIN_STEP step=13214 loss=1.240960 Training: 73%|███████▎ | 13214/18018 [28:30:27<10:57:55, 8.22s/step, loss=1.241, acc=0.712, acc_len=1.73, thru=13.3, I=14.1, T=34.6, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:42,018] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:47,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:00:47,773] loop.py:380 INFO TIMING step=13215: step=3.649s data=1.503s compute=3.620s [fwd=2.560s bwd=1.045s opt=0.003s] dispatch=2.457s [2026-10-19 18:00:47,773] loop.py:45 INFO TRAIN_STEP step=13215 loss=1.161521 Training: 73%|███████▎ | 13215/18018 [28:30:33<10:07:21, 7.59s/step, loss=1.162, acc=0.721, acc_len=1.76, thru=19.9, I=20.7, T=35.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:48,142] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:52,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:00:53,676] loop.py:45 INFO TRAIN_STEP step=13216 loss=1.242988 Training: 73%|███████▎ | 13216/18018 [28:30:39<9:26:48, 7.08s/step, loss=1.243, acc=0.708, acc_len=1.71, thru=20.9, I=25.8, T=37.2, wait=2.5s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:00:53,672] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13215: forward=2428.6ms backward=960.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:00:53,966] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:01:00,557] loop.py:45 INFO TRAIN_STEP step=13217 loss=1.287509 Training: 73%|███████▎ | 13217/18018 [28:30:46<9:21:48, 7.02s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=17.6, I=22.9, T=38.3, wait=3.5s, pool=61, epoch=2/2][2026-10-19 18:01:08,105] loop.py:45 INFO TRAIN_STEP step=13218 loss=1.238146 Training: 73%|███████▎ | 13218/18018 [28:30:54<9:34:23, 7.18s/step, loss=1.238, acc=0.711, acc_len=1.72, thru=14.9, I=16.9, T=37.6, wait=4.1s, pool=41, epoch=2/2][2026-10-19 18:01:16,294] loop.py:45 INFO TRAIN_STEP step=13219 loss=1.246250 Training: 73%|███████▎ | 13219/18018 [28:31:02<9:58:24, 7.48s/step, loss=1.246, acc=0.709, acc_len=1.72, thru=16.5, I=17.3, T=36.9, wait=4.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:01:23,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:01:24,080] loop.py:380 INFO TIMING step=13220: step=3.691s data=1.424s compute=3.663s [fwd=2.520s bwd=1.128s opt=0.003s] dispatch=4.086s [2026-10-19 18:01:24,080] loop.py:45 INFO TRAIN_STEP step=13220 loss=1.169099 Training: 73%|███████▎ | 13220/18018 [28:31:10<10:05:35, 7.57s/step, loss=1.169, acc=0.724, acc_len=1.76, thru=17.3, I=17.7, T=34.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:01:24,408] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:01:30,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 18:01:31,638] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13220: forward=2540.3ms backward=1024.2ms [2026-10-19 18:01:31,642] loop.py:45 INFO TRAIN_STEP step=13221 loss=1.235325 Training: 73%|███████▎ | 13221/18018 [28:31:17<10:05:15, 7.57s/step, loss=1.235, acc=0.714, acc_len=1.73, thru=17.1, I=18.8, T=35.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:01:31,727] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 18:01:37,252] loop.py:45 INFO TRAIN_STEP step=13222 loss=1.349822 Training: 73%|███████▎ | 13222/18018 [28:31:23<9:18:06, 6.98s/step, loss=1.350, acc=0.692, acc_len=1.65, thru=20.6, I=23.9, T=38.8, wait=2.3s, pool=61, epoch=2/2] [2026-10-19 18:01:42,949] loop.py:45 INFO TRAIN_STEP step=13223 loss=1.264848 Training: 73%|███████▎ | 13223/18018 [28:31:29<8:47:06, 6.60s/step, loss=1.265, acc=0.702, acc_len=1.68, thru=19.1, I=21.0, T=38.0, wait=2.3s, pool=41, epoch=2/2][2026-10-19 18:01:51,026] loop.py:45 INFO TRAIN_STEP step=13224 loss=1.155636 Training: 73%|███████▎ | 13224/18018 [28:31:37<9:22:36, 7.04s/step, loss=1.156, acc=0.727, acc_len=1.77, thru=16.5, I=15.9, T=36.3, wait=4.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:01:57,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:01:59,209] loop.py:380 INFO TIMING step=13225: step=3.852s data=1.675s compute=3.814s [fwd=2.690s bwd=1.109s opt=0.003s] dispatch=4.320s [2026-10-19 18:01:59,209] loop.py:45 INFO TRAIN_STEP step=13225 loss=1.314044 Training: 73%|███████▎ | 13225/18018 [28:31:45<9:49:49, 7.38s/step, loss=1.314, acc=0.694, acc_len=1.66, thru=17.5, I=18.8, T=33.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:01:59,418] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:03,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:02:05,091] loop.py:45 INFO TRAIN_STEP step=13226 loss=1.300204 Training: 73%|███████▎ | 13226/18018 [28:31:51<9:13:40, 6.93s/step, loss=1.300, acc=0.699, acc_len=1.68, thru=20.1, I=26.2, T=38.0, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:02:05,087] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13225: forward=2335.8ms backward=977.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:05,492] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:12,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:02:12,829] loop.py:45 INFO TRAIN_STEP step=13227 loss=1.214155 Training: 73%|███████▎ | 13227/18018 [28:31:58<9:32:53, 7.17s/step, loss=1.214, acc=0.713, acc_len=1.72, thru=17.4, I=22.3, T=37.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:13,252] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:19,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:02:20,217] loop.py:45 INFO TRAIN_STEP step=13228 loss=1.293417 Training: 73%|███████▎ | 13228/18018 [28:32:06<9:37:53, 7.24s/step, loss=1.293, acc=0.700, acc_len=1.68, thru=17.8, I=20.5, T=34.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:20,426] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:25,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:02:25,725] loop.py:45 INFO TRAIN_STEP step=13229 loss=1.260239 Training: 73%|███████▎ | 13229/18018 [28:32:11<8:56:19, 6.72s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=21.6, I=24.5, T=38.2, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:25,935] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:02:34,627] loop.py:380 INFO TIMING step=13230: step=3.466s data=1.348s compute=3.432s [fwd=2.437s bwd=0.979s opt=0.003s] dispatch=5.427s [2026-10-19 18:02:34,627] loop.py:45 INFO TRAIN_STEP step=13230 loss=1.284666 Training: 73%|███████▎ | 13230/18018 [28:32:20<9:48:24, 7.37s/step, loss=1.285, acc=0.701, acc_len=1.69, thru=14.0, I=14.3, T=36.9, wait=5.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:40,600] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:02:40,818] loop.py:45 INFO TRAIN_STEP step=13231 loss=1.178606 Training: 73%|███████▎ | 13231/18018 [28:32:26<9:19:59, 7.02s/step, loss=1.179, acc=0.721, acc_len=1.75, thru=20.2, I=20.0, T=37.7, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:02:40,813] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13230: forward=2390.8ms backward=964.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:41,102] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:46,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:02:46,206] loop.py:45 INFO TRAIN_STEP step=13232 loss=1.314850 Training: 73%|███████▎ | 13232/18018 [28:32:32<8:40:51, 6.53s/step, loss=1.315, acc=0.698, acc_len=1.67, thru=22.1, I=22.7, T=39.3, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:02:46,573] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:02:53,226] loop.py:45 INFO TRAIN_STEP step=13233 loss=1.294248 Training: 73%|███████▎ | 13233/18018 [28:32:39<8:52:30, 6.68s/step, loss=1.294, acc=0.698, acc_len=1.67, thru=18.7, I=18.7, T=35.6, wait=3.4s, pool=57, epoch=2/2][2026-10-19 18:02:59,609] loop.py:45 INFO TRAIN_STEP step=13234 loss=1.166831 Training: 73%|███████▎ | 13234/18018 [28:32:45<8:45:17, 6.59s/step, loss=1.167, acc=0.724, acc_len=1.76, thru=18.3, I=16.9, T=36.2, wait=2.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:03:05,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:03:06,131] loop.py:380 INFO TIMING step=13235: step=3.660s data=1.549s compute=3.621s [fwd=2.578s bwd=1.028s opt=0.003s] dispatch=2.854s [2026-10-19 18:03:06,131] loop.py:45 INFO TRAIN_STEP step=13235 loss=1.267299 Training: 73%|███████▎ | 13235/18018 [28:32:52<8:43:37, 6.57s/step, loss=1.267, acc=0.706, acc_len=1.69, thru=19.7, I=19.9, T=35.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:03:06,492] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:03:13,119] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:03:13,749] loop.py:45 INFO TRAIN_STEP step=13236 loss=1.324442 Training: 73%|███████▎ | 13236/18018 [28:32:59<9:08:37, 6.88s/step, loss=1.324, acc=0.694, acc_len=1.66, thru=16.5, I=17.7, T=35.8, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:03:13,745] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13235: forward=2537.1ms backward=991.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:03:14,124] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:03:21,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:03:23,097] loop.py:45 INFO TRAIN_STEP step=13237 loss=1.150731 Training: 73%|███████▎ | 13237/18018 [28:33:09<10:07:25, 7.62s/step, loss=1.151, acc=0.726, acc_len=1.76, thru=13.2, I=16.2, T=35.2, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:03:23,480] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:03:29,792] loop.py:45 INFO TRAIN_STEP step=13238 loss=1.199950 Training: 73%|███████▎ | 13238/18018 [28:33:15<9:45:06, 7.34s/step, loss=1.200, acc=0.720, acc_len=1.75, thru=16.1, I=19.7, T=36.8, wait=3.2s, pool=41, epoch=2/2] [2026-10-19 18:03:39,132] loop.py:45 INFO TRAIN_STEP step=13239 loss=1.371668 Training: 73%|███████▎ | 13239/18018 [28:33:25<10:32:39, 7.94s/step, loss=1.372, acc=0.687, acc_len=1.64, thru=13.7, I=12.8, T=35.0, wait=5.7s, pool=41, epoch=2/2][2026-10-19 18:03:46,997] loop.py:380 INFO TIMING step=13240: step=3.592s data=1.485s compute=3.564s [fwd=2.557s bwd=0.993s opt=0.003s] dispatch=4.265s [2026-10-19 18:03:46,997] loop.py:45 INFO TRAIN_STEP step=13240 loss=1.328346 Training: 73%|███████▎ | 13240/18018 [28:33:33<10:30:40, 7.92s/step, loss=1.328, acc=0.695, acc_len=1.66, thru=16.9, I=16.1, T=35.6, wait=4.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:03:53,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:03:53,515] loop.py:45 INFO TRAIN_STEP step=13241 loss=1.311805 Training: 73%|███████▎ | 13241/18018 [28:33:39<9:57:07, 7.50s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=19.2, I=19.3, T=37.5, wait=3.1s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:03:53,510] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13240: forward=2407.2ms backward=959.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:03:53,712] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 18:04:01,675] loop.py:45 INFO TRAIN_STEP step=13242 loss=1.202054 Training: 73%|███████▎ | 13242/18018 [28:33:47<10:12:40, 7.70s/step, loss=1.202, acc=0.721, acc_len=1.76, thru=14.8, I=14.8, T=36.8, wait=4.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:04:11,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:04:12,711] loop.py:45 INFO TRAIN_STEP step=13243 loss=1.281160 Training: 73%|███████▎ | 13243/18018 [28:33:58<11:32:19, 8.70s/step, loss=1.281, acc=0.705, acc_len=1.70, thru=13.8, I=15.2, T=35.2, wait=7.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:04:12,894] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:04:17,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:04:18,575] loop.py:45 INFO TRAIN_STEP step=13244 loss=1.306605 Training: 74%|███████▎ | 13244/18018 [28:34:04<10:24:27, 7.85s/step, loss=1.307, acc=0.698, acc_len=1.67, thru=20.0, I=28.8, T=34.4, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:04:18,650] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-19 18:04:24,370] loop.py:380 INFO TIMING step=13245: step=3.368s data=1.405s compute=3.323s [fwd=2.382s bwd=0.926s opt=0.003s] dispatch=2.418s [2026-10-19 18:04:24,370] loop.py:45 INFO TRAIN_STEP step=13245 loss=1.225797 Training: 74%|███████▎ | 13245/18018 [28:34:10<9:35:21, 7.23s/step, loss=1.226, acc=0.713, acc_len=1.73, thru=21.4, I=27.9, T=38.0, wait=2.4s, pool=61, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:04:29,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:04:30,994] loop.py:45 INFO TRAIN_STEP step=13246 loss=1.206109 Training: 74%|███████▎ | 13246/18018 [28:34:17<9:20:44, 7.05s/step, loss=1.206, acc=0.715, acc_len=1.72, thru=20.4, I=22.8, T=35.2, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:04:30,990] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13245: forward=2576.6ms backward=1009.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:04:31,199] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:04:38,131] loop.py:45 INFO TRAIN_STEP step=13247 loss=1.306053 Training: 74%|███████▎ | 13247/18018 [28:34:24<9:22:38, 7.08s/step, loss=1.306, acc=0.698, acc_len=1.67, thru=16.9, I=19.7, T=38.0, wait=3.8s, pool=57, epoch=2/2][2026-10-19 18:04:47,059] loop.py:45 INFO TRAIN_STEP step=13248 loss=1.201734 Training: 74%|███████▎ | 13248/18018 [28:34:33<10:06:39, 7.63s/step, loss=1.202, acc=0.718, acc_len=1.74, thru=16.2, I=15.3, T=36.2, wait=5.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:04:53,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:04:54,139] loop.py:45 INFO TRAIN_STEP step=13249 loss=1.263159 Training: 74%|███████▎ | 13249/18018 [28:34:40<9:53:23, 7.47s/step, loss=1.263, acc=0.705, acc_len=1.70, thru=19.1, I=19.0, T=35.2, wait=3.4s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:04:54,630] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:04:59,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:05:00,247] loop.py:380 INFO TIMING step=13250: step=3.416s data=1.426s compute=3.391s [fwd=2.411s bwd=0.966s opt=0.003s] dispatch=2.684s [2026-10-19 18:05:00,247] loop.py:45 INFO TRAIN_STEP step=13250 loss=1.294118 Training: 74%|███████▎ | 13250/18018 [28:34:46<9:20:55, 7.06s/step, loss=1.294, acc=0.701, acc_len=1.68, thru=20.1, I=24.2, T=37.5, wait=2.7s, pool=65, epoch=2/2][2026-10-19 18:05:00,250] loop.py:419 INFO Saving checkpoint at step 13250... (TrainerActor pid=9310) [2026-10-19 18:05:03,757] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0013251 (TrainerActor pid=9310) [2026-10-19 18:05:03,767] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0012501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:04,452] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:04,975] inference_manager.py:408 INFO Pool capacity available after 5.6s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:10,076] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 18:05:03,757] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0013251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 18:05:04,907] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0012501 [2026-10-19 18:05:10,818] loop.py:45 INFO TRAIN_STEP step=13251 loss=1.313228 (TrainerActor pid=9310) [2026-10-19 18:05:10,813] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13250: forward=2400.6ms backward=917.2ms Training: 74%|███████▎ | 13251/18018 [28:34:56<10:44:34, 8.11s/step, loss=1.313, acc=0.697, acc_len=1.67, thru=22.0, I=22.2, T=38.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:11,083] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:16,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:05:16,778] loop.py:45 INFO TRAIN_STEP step=13252 loss=1.262058 Training: 74%|███████▎ | 13252/18018 [28:35:02<9:53:05, 7.47s/step, loss=1.262, acc=0.705, acc_len=1.69, thru=21.4, I=25.0, T=38.8, wait=2.7s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:17,106] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:23,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:05:24,708] loop.py:45 INFO TRAIN_STEP step=13253 loss=1.289422 Training: 74%|███████▎ | 13253/18018 [28:35:10<10:04:03, 7.61s/step, loss=1.289, acc=0.702, acc_len=1.68, thru=17.1, I=19.6, T=34.9, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:24,992] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:05:30,971] loop.py:45 INFO TRAIN_STEP step=13254 loss=1.230290 Training: 74%|███████▎ | 13254/18018 [28:35:17<9:31:54, 7.20s/step, loss=1.230, acc=0.709, acc_len=1.71, thru=17.6, I=18.9, T=38.9, wait=3.0s, pool=45, epoch=2/2] (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 18:05:38,866] eagle_store.py:517 WARNING batch_get_buffer missing keys (k424162-9ca83a5b_hs, k424162-9ca83a5b_ids, k424162-9ca83a5b_lhs); sleeping 0.50s. [2026-10-19 18:05:39,937] loop.py:380 INFO TIMING step=13255: step=3.723s data=1.485s compute=3.697s [fwd=2.666s bwd=1.018s opt=0.003s] dispatch=5.233s [2026-10-19 18:05:39,937] loop.py:45 INFO TRAIN_STEP step=13255 loss=1.230803 Training: 74%|███████▎ | 13255/18018 [28:35:26<10:13:47, 7.73s/step, loss=1.231, acc=0.718, acc_len=1.74, thru=14.0, I=13.5, T=34.4, wait=5.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:47,374] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 18:05:38,866] eagle_store.py:517 WARNING batch_get_buffer missing keys (k424161-b142c8ff_hs, k424161-b142c8ff_ids, k424161-b142c8ff_lhs); sleeping 0.50s. [2026-10-19 18:05:48,961] loop.py:45 INFO TRAIN_STEP step=13256 loss=1.361231 Training: 74%|███████▎ | 13256/18018 [28:35:35<10:44:25, 8.12s/step, loss=1.361, acc=0.689, acc_len=1.64, thru=17.3, I=17.7, T=31.5, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:05:48,957] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13255: forward=2740.5ms backward=1231.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:49,390] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:53,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:05:54,619] loop.py:45 INFO TRAIN_STEP step=13257 loss=1.276155 Training: 74%|███████▎ | 13257/18018 [28:35:40<9:45:41, 7.38s/step, loss=1.276, acc=0.700, acc_len=1.68, thru=20.1, I=26.1, T=36.4, wait=2.1s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:05:54,878] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:06:00,874] loop.py:45 INFO TRAIN_STEP step=13258 loss=1.142605 Training: 74%|███████▎ | 13258/18018 [28:35:46<9:18:42, 7.04s/step, loss=1.143, acc=0.726, acc_len=1.77, thru=17.9, I=20.0, T=37.2, wait=2.8s, pool=45, epoch=2/2][2026-10-19 18:06:11,233] loop.py:45 INFO TRAIN_STEP step=13259 loss=1.307320 Training: 74%|███████▎ | 13259/18018 [28:35:57<10:37:35, 8.04s/step, loss=1.307, acc=0.698, acc_len=1.67, thru=13.7, I=14.3, T=32.1, wait=6.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:17,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:06:19,698] loop.py:380 INFO TIMING step=13260: step=4.057s data=1.493s compute=4.030s [fwd=3.007s bwd=1.009s opt=0.003s] dispatch=4.397s [2026-10-19 18:06:19,699] loop.py:45 INFO TRAIN_STEP step=13260 loss=1.371662 Training: 74%|███████▎ | 13260/18018 [28:36:05<10:47:33, 8.17s/step, loss=1.372, acc=0.688, acc_len=1.65, thru=15.7, I=18.0, T=31.6, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:20,003] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:25,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:06:26,525] loop.py:45 INFO TRAIN_STEP step=13261 loss=1.209376 Training: 74%|███████▎ | 13261/18018 [28:36:12<10:15:32, 7.76s/step, loss=1.209, acc=0.720, acc_len=1.75, thru=18.1, I=25.2, T=30.0, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:06:26,520] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13260: forward=3341.3ms backward=883.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:26,975] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:32,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:06:34,172] loop.py:45 INFO TRAIN_STEP step=13262 loss=1.333341 Training: 74%|███████▎ | 13262/18018 [28:36:20<10:12:38, 7.73s/step, loss=1.333, acc=0.693, acc_len=1.66, thru=14.7, I=20.1, T=25.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:34,636] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:40,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:06:41,191] loop.py:45 INFO TRAIN_STEP step=13263 loss=1.382654 Training: 74%|███████▎ | 13263/18018 [28:36:27<9:55:43, 7.52s/step, loss=1.383, acc=0.684, acc_len=1.63, thru=15.7, I=21.8, T=37.1, wait=3.6s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:41,441] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:47,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:06:48,601] loop.py:45 INFO TRAIN_STEP step=13264 loss=1.158155 Training: 74%|███████▎ | 13264/18018 [28:36:34<9:52:58, 7.48s/step, loss=1.158, acc=0.728, acc_len=1.78, thru=16.8, I=19.4, T=36.7, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:48,868] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:54,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:06:54,969] loop.py:380 INFO TIMING step=13265: step=3.356s data=1.396s compute=3.324s [fwd=2.378s bwd=0.933s opt=0.003s] dispatch=3.003s [2026-10-19 18:06:54,970] loop.py:45 INFO TRAIN_STEP step=13265 loss=1.207163 Training: 74%|███████▎ | 13265/18018 [28:36:41<9:26:22, 7.15s/step, loss=1.207, acc=0.719, acc_len=1.75, thru=19.8, I=21.9, T=38.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:06:55,159] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:07:01,303] loop.py:45 INFO TRAIN_STEP step=13266 loss=1.271129 Training: 74%|███████▎ | 13266/18018 [28:36:47<9:06:49, 6.90s/step, loss=1.271, acc=0.703, acc_len=1.69, thru=17.1, I=19.0, T=37.8, wait=2.9s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:07:01,299] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13265: forward=2347.6ms backward=980.8ms [2026-10-19 18:07:11,951] loop.py:45 INFO TRAIN_STEP step=13267 loss=1.180699 Training: 74%|███████▎ | 13267/18018 [28:36:58<10:35:40, 8.03s/step, loss=1.181, acc=0.721, acc_len=1.75, thru=13.9, I=14.2, T=32.7, wait=6.7s, pool=45, epoch=2/2][2026-10-19 18:07:20,223] loop.py:45 INFO TRAIN_STEP step=13268 loss=1.242539 Training: 74%|███████▎ | 13268/18018 [28:37:06<10:41:17, 8.10s/step, loss=1.243, acc=0.709, acc_len=1.71, thru=14.3, I=13.8, T=33.9, wait=4.5s, pool=41, epoch=2/2][2026-10-19 18:07:28,177] loop.py:45 INFO TRAIN_STEP step=13269 loss=1.309093 Training: 74%|███████▎ | 13269/18018 [28:37:14<10:37:45, 8.06s/step, loss=1.309, acc=0.700, acc_len=1.67, thru=16.1, I=13.7, T=36.5, wait=4.4s, pool=37, epoch=2/2][2026-10-19 18:07:36,545] loop.py:380 INFO TIMING step=13270: step=3.629s data=1.456s compute=3.596s [fwd=2.543s bwd=1.038s opt=0.003s] dispatch=4.729s [2026-10-19 18:07:36,545] loop.py:45 INFO TRAIN_STEP step=13270 loss=1.302460 Training: 74%|███████▎ | 13270/18018 [28:37:22<10:44:56, 8.15s/step, loss=1.302, acc=0.701, acc_len=1.68, thru=15.7, I=14.8, T=35.3, wait=4.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:07:43,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:07:44,581] loop.py:45 INFO TRAIN_STEP step=13271 loss=1.259360 Training: 74%|███████▎ | 13271/18018 [28:37:30<10:42:07, 8.12s/step, loss=1.259, acc=0.709, acc_len=1.71, thru=17.6, I=17.5, T=34.6, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:07:44,576] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13270: forward=2531.3ms backward=1114.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:07:44,811] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:07:51,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:07:51,885] loop.py:45 INFO TRAIN_STEP step=13272 loss=1.370847 Training: 74%|███████▎ | 13272/18018 [28:37:37<10:22:40, 7.87s/step, loss=1.371, acc=0.684, acc_len=1.63, thru=16.6, I=19.4, T=36.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:07:51,931] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:07:56,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:07:58,136] loop.py:45 INFO TRAIN_STEP step=13273 loss=1.283967 Training: 74%|███████▎ | 13273/18018 [28:37:44<9:44:05, 7.39s/step, loss=1.284, acc=0.704, acc_len=1.70, thru=20.1, I=24.7, T=36.8, wait=2.8s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:07:58,379] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 18:08:04,544] loop.py:45 INFO TRAIN_STEP step=13274 loss=1.134487 Training: 74%|███████▎ | 13274/18018 [28:37:50<9:20:46, 7.09s/step, loss=1.134, acc=0.736, acc_len=1.81, thru=17.4, I=24.1, T=39.7, wait=3.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:08:11,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:08:12,394] loop.py:380 INFO TIMING step=13275: step=3.733s data=1.635s compute=3.706s [fwd=2.645s bwd=1.048s opt=0.003s] dispatch=4.108s [2026-10-19 18:08:12,394] loop.py:45 INFO TRAIN_STEP step=13275 loss=1.231232 Training: 74%|███████▎ | 13275/18018 [28:37:58<9:38:39, 7.32s/step, loss=1.231, acc=0.713, acc_len=1.73, thru=17.4, I=17.9, T=34.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:08:12,712] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:08:20,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:08:20,549] loop.py:45 INFO TRAIN_STEP step=13276 loss=1.232493 Training: 74%|███████▎ | 13276/18018 [28:38:06<9:58:16, 7.57s/step, loss=1.232, acc=0.714, acc_len=1.73, thru=16.5, I=19.4, T=35.3, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:08:20,544] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13275: forward=2535.6ms backward=1038.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:08:20,860] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:08:27,452] loop.py:45 INFO TRAIN_STEP step=13277 loss=1.355379 Training: 74%|███████▎ | 13277/18018 [28:38:13<9:42:21, 7.37s/step, loss=1.355, acc=0.689, acc_len=1.64, thru=17.3, I=17.4, T=36.6, wait=3.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:08:33,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:08:34,641] loop.py:45 INFO TRAIN_STEP step=13278 loss=1.261268 Training: 74%|███████▎ | 13278/18018 [28:38:20<9:37:58, 7.32s/step, loss=1.261, acc=0.706, acc_len=1.70, thru=18.1, I=18.6, T=37.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:08:34,995] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:08:41,262] loop.py:45 INFO TRAIN_STEP step=13279 loss=1.271701 Training: 74%|███████▎ | 13279/18018 [28:38:27<9:21:20, 7.11s/step, loss=1.272, acc=0.705, acc_len=1.70, thru=18.4, I=21.2, T=38.1, wait=3.2s, pool=53, epoch=2/2][2026-10-19 18:08:47,775] loop.py:380 INFO TIMING step=13280: step=3.511s data=1.455s compute=3.472s [fwd=2.488s bwd=0.968s opt=0.003s] dispatch=2.994s [2026-10-19 18:08:47,775] loop.py:45 INFO TRAIN_STEP step=13280 loss=1.238281 Training: 74%|███████▎ | 13280/18018 [28:38:33<9:07:07, 6.93s/step, loss=1.238, acc=0.710, acc_len=1.72, thru=18.2, I=19.6, T=36.5, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:08:54,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 18:08:56,750] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13280: forward=2696.5ms backward=1057.7ms [2026-10-19 18:08:56,755] loop.py:45 INFO TRAIN_STEP step=13281 loss=1.356290 Training: 74%|███████▎ | 13281/18018 [28:38:42<9:55:38, 7.54s/step, loss=1.356, acc=0.693, acc_len=1.66, thru=16.3, I=17.9, T=33.5, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:08:56,950] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 18:09:02,964] loop.py:45 INFO TRAIN_STEP step=13282 loss=1.239373 Training: 74%|███████▎ | 13282/18018 [28:38:49<9:23:52, 7.14s/step, loss=1.239, acc=0.714, acc_len=1.72, thru=16.9, I=23.9, T=38.3, wait=2.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:10,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:09:11,196] loop.py:45 INFO TRAIN_STEP step=13283 loss=1.214824 Training: 74%|███████▎ | 13283/18018 [28:38:57<9:49:34, 7.47s/step, loss=1.215, acc=0.719, acc_len=1.75, thru=17.4, I=18.0, T=31.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:11,627] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:16,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:09:17,276] loop.py:45 INFO TRAIN_STEP step=13284 loss=1.316584 Training: 74%|███████▎ | 13284/18018 [28:39:03<9:16:31, 7.05s/step, loss=1.317, acc=0.696, acc_len=1.67, thru=19.2, I=22.9, T=35.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:17,308] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 18:09:22,765] loop.py:380 INFO TIMING step=13285: step=3.425s data=1.428s compute=3.395s [fwd=2.441s bwd=0.940s opt=0.003s] dispatch=2.055s [2026-10-19 18:09:22,766] loop.py:45 INFO TRAIN_STEP step=13285 loss=1.242124 Training: 74%|███████▎ | 13285/18018 [28:39:08<8:39:23, 6.58s/step, loss=1.242, acc=0.711, acc_len=1.72, thru=18.5, I=23.7, T=37.4, wait=2.1s, pool=25, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 18:09:22,818] eagle_store.py:517 WARNING batch_get_buffer missing keys (k425126-bff0ade3_hs, k425126-bff0ade3_ids, k425126-bff0ade3_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:30,955] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 18:09:22,862] eagle_store.py:517 WARNING batch_get_buffer missing keys (k425127-a6fceb51_hs, k425127-a6fceb51_ids, k425127-a6fceb51_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 18:09:32,387] loop.py:45 INFO TRAIN_STEP step=13286 loss=1.320656 Training: 74%|███████▎ | 13286/18018 [28:39:18<9:51:08, 7.50s/step, loss=1.321, acc=0.697, acc_len=1.67, thru=17.8, I=17.4, T=32.5, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:09:32,383] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13285: forward=2623.3ms backward=1223.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:32,476] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:37,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:09:38,279] loop.py:45 INFO TRAIN_STEP step=13287 loss=1.218062 Training: 74%|███████▎ | 13287/18018 [28:39:24<9:13:06, 7.01s/step, loss=1.218, acc=0.715, acc_len=1.73, thru=19.9, I=25.6, T=35.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:38,579] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:43,583] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:09:44,174] loop.py:45 INFO TRAIN_STEP step=13288 loss=1.287748 Training: 74%|███████▎ | 13288/18018 [28:39:30<8:46:26, 6.68s/step, loss=1.288, acc=0.699, acc_len=1.68, thru=20.5, I=22.3, T=37.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:44,588] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:09:50,241] loop.py:45 INFO TRAIN_STEP step=13289 loss=1.254017 Training: 74%|███████▍ | 13289/18018 [28:39:36<8:31:53, 6.49s/step, loss=1.254, acc=0.707, acc_len=1.70, thru=20.3, I=24.2, T=38.9, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:57,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:09:58,001] loop.py:380 INFO TIMING step=13290: step=3.670s data=1.558s compute=3.640s [fwd=2.628s bwd=0.997s opt=0.003s] dispatch=4.082s [2026-10-19 18:09:58,001] loop.py:45 INFO TRAIN_STEP step=13290 loss=1.348518 Training: 74%|███████▍ | 13290/18018 [28:39:44<9:01:42, 6.87s/step, loss=1.349, acc=0.694, acc_len=1.66, thru=17.5, I=18.0, T=34.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:09:58,304] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:10:05,023] loop.py:45 INFO TRAIN_STEP step=13291 loss=1.296196 Training: 74%|███████▍ | 13291/18018 [28:39:51<9:05:03, 6.92s/step, loss=1.296, acc=0.697, acc_len=1.68, thru=17.1, I=17.8, T=35.7, wait=3.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:10:05,019] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13290: forward=2508.2ms backward=1033.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:10:10,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:10:13,079] loop.py:45 INFO TRAIN_STEP step=13292 loss=1.374456 Training: 74%|███████▍ | 13292/18018 [28:39:59<9:31:49, 7.26s/step, loss=1.374, acc=0.687, acc_len=1.64, thru=17.1, I=20.3, T=34.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:10:13,244] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 18:10:18,602] loop.py:45 INFO TRAIN_STEP step=13293 loss=1.220341 Training: 74%|███████▍ | 13293/18018 [28:40:04<8:50:41, 6.74s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=19.5, I=25.6, T=37.5, wait=2.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:10:27,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:10:29,098] loop.py:45 INFO TRAIN_STEP step=13294 loss=1.301429 Training: 74%|███████▍ | 13294/18018 [28:40:15<10:19:18, 7.87s/step, loss=1.301, acc=0.699, acc_len=1.67, thru=13.0, I=14.6, T=32.6, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:10:29,338] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:10:37,614] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:10:38,828] loop.py:380 INFO TIMING step=13295: step=4.427s data=2.050s compute=4.113s [fwd=2.920s bwd=1.177s opt=0.003s] dispatch=5.294s [2026-10-19 18:10:38,828] loop.py:45 INFO TRAIN_STEP step=13295 loss=1.367718 Training: 74%|███████▍ | 13295/18018 [28:40:24<11:03:11, 8.43s/step, loss=1.368, acc=0.691, acc_len=1.65, thru=13.5, I=14.3, T=28.9, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:10:39,120] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:10:44,354] loop.py:45 INFO TRAIN_STEP step=13296 loss=1.321135 Training: 74%|███████▍ | 13296/18018 [28:40:30<9:54:38, 7.56s/step, loss=1.321, acc=0.697, acc_len=1.67, thru=16.8, I=19.7, T=39.2, wait=2.3s, pool=37, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:10:44,350] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13295: forward=2284.0ms backward=933.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:10:50,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:10:52,063] loop.py:45 INFO TRAIN_STEP step=13297 loss=1.373682 Training: 74%|███████▍ | 13297/18018 [28:40:38<9:58:06, 7.60s/step, loss=1.374, acc=0.690, acc_len=1.64, thru=18.8, I=19.8, T=36.0, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:10:52,410] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:10:59,147] loop.py:45 INFO TRAIN_STEP step=13298 loss=1.259916 Training: 74%|███████▍ | 13298/18018 [28:40:45<9:45:45, 7.45s/step, loss=1.260, acc=0.705, acc_len=1.70, thru=16.1, I=17.9, T=37.4, wait=3.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:06,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:11:06,347] loop.py:45 INFO TRAIN_STEP step=13299 loss=1.376082 Training: 74%|███████▍ | 13299/18018 [28:40:52<9:39:52, 7.37s/step, loss=1.376, acc=0.685, acc_len=1.63, thru=18.9, I=18.3, T=35.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:06,518] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:11,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:11:12,357] loop.py:380 INFO TIMING step=13300: step=3.530s data=1.441s compute=3.492s [fwd=2.522s bwd=0.956s opt=0.003s] dispatch=2.470s [2026-10-19 18:11:12,357] loop.py:45 INFO TRAIN_STEP step=13300 loss=1.346913 Training: 74%|███████▍ | 13300/18018 [28:40:58<9:07:35, 6.96s/step, loss=1.347, acc=0.692, acc_len=1.65, thru=20.2, I=22.4, T=36.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:12,554] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:11:18,920] loop.py:45 INFO TRAIN_STEP step=13301 loss=1.358445 Training: 74%|███████▍ | 13301/18018 [28:41:05<8:58:03, 6.84s/step, loss=1.358, acc=0.688, acc_len=1.64, thru=18.1, I=23.1, T=36.0, wait=3.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:11:18,915] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13300: forward=2530.7ms backward=975.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:24,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:11:25,547] loop.py:45 INFO TRAIN_STEP step=13302 loss=1.193236 Training: 74%|███████▍ | 13302/18018 [28:41:11<8:52:46, 6.78s/step, loss=1.193, acc=0.719, acc_len=1.75, thru=19.0, I=21.3, T=37.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:25,786] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:11:34,206] loop.py:45 INFO TRAIN_STEP step=13303 loss=1.385420 Training: 74%|███████▍ | 13303/18018 [28:41:20<9:36:56, 7.34s/step, loss=1.385, acc=0.682, acc_len=1.62, thru=10.9, I=12.8, T=35.0, wait=5.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:42,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:11:42,567] loop.py:45 INFO TRAIN_STEP step=13304 loss=1.220756 Training: 74%|███████▍ | 13304/18018 [28:41:28<10:00:55, 7.65s/step, loss=1.221, acc=0.715, acc_len=1.73, thru=17.7, I=16.2, T=36.4, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:42,800] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:50,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:11:51,744] loop.py:380 INFO TIMING step=13305: step=3.816s data=1.618s compute=3.783s [fwd=2.677s bwd=1.091s opt=0.003s] dispatch=5.350s [2026-10-19 18:11:51,744] loop.py:45 INFO TRAIN_STEP step=13305 loss=1.377514 Training: 74%|███████▍ | 13305/18018 [28:41:37<10:36:48, 8.11s/step, loss=1.378, acc=0.687, acc_len=1.63, thru=14.1, I=16.4, T=33.5, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:11:51,813] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 18:11:59,293] loop.py:45 INFO TRAIN_STEP step=13306 loss=1.133496 Training: 74%|███████▍ | 13306/18018 [28:41:45<10:23:29, 7.94s/step, loss=1.133, acc=0.730, acc_len=1.78, thru=15.7, I=17.4, T=35.8, wait=4.0s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:11:59,290] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13305: forward=2543.5ms backward=990.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:12:07,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:12:07,840] loop.py:45 INFO TRAIN_STEP step=13307 loss=1.406631 Training: 74%|███████▍ | 13307/18018 [28:41:53<10:37:39, 8.12s/step, loss=1.407, acc=0.680, acc_len=1.62, thru=15.2, I=15.1, T=33.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:12:08,003] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:12:14,233] loop.py:45 INFO TRAIN_STEP step=13308 loss=1.276075 Training: 74%|███████▍ | 13308/18018 [28:42:00<9:56:53, 7.60s/step, loss=1.276, acc=0.704, acc_len=1.70, thru=18.6, I=18.5, T=32.0, wait=2.4s, pool=53, epoch=2/2] [2026-10-19 18:12:21,028] loop.py:45 INFO TRAIN_STEP step=13309 loss=1.294999 Training: 74%|███████▍ | 13309/18018 [28:42:07<9:37:42, 7.36s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=15.8, I=17.0, T=36.1, wait=3.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:12:32,105] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:12:34,033] loop.py:380 INFO TIMING step=13310: step=4.238s data=1.626s compute=4.198s [fwd=2.944s bwd=1.237s opt=0.003s] dispatch=8.757s [2026-10-19 18:12:34,033] loop.py:45 INFO TRAIN_STEP step=13310 loss=1.261729 Training: 74%|███████▍ | 13310/18018 [28:42:20<11:50:23, 9.05s/step, loss=1.262, acc=0.706, acc_len=1.70, thru=12.2, I=12.8, T=30.2, wait=8.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:12:34,122] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 [2026-10-19 18:12:39,855] loop.py:45 INFO TRAIN_STEP step=13311 loss=1.264188 Training: 74%|███████▍ | 13311/18018 [28:42:25<10:34:10, 8.08s/step, loss=1.264, acc=0.705, acc_len=1.70, thru=18.3, I=22.5, T=38.9, wait=2.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:12:39,850] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13310: forward=2339.1ms backward=907.9ms [2026-10-19 18:12:47,078] loop.py:45 INFO TRAIN_STEP step=13312 loss=1.252635 Training: 74%|███████▍ | 13312/18018 [28:42:33<10:13:49, 7.83s/step, loss=1.253, acc=0.711, acc_len=1.71, thru=18.3, I=18.9, T=38.0, wait=3.8s, pool=57, epoch=2/2][2026-10-19 18:12:57,779] loop.py:45 INFO TRAIN_STEP step=13313 loss=1.345459 Training: 74%|███████▍ | 13313/18018 [28:42:43<11:21:21, 8.69s/step, loss=1.345, acc=0.693, acc_len=1.66, thru=11.3, I=11.8, T=33.5, wait=6.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:13:03,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:13:03,843] loop.py:45 INFO TRAIN_STEP step=13314 loss=1.212912 Training: 74%|███████▍ | 13314/18018 [28:42:49<10:19:24, 7.90s/step, loss=1.213, acc=0.715, acc_len=1.73, thru=19.8, I=19.1, T=37.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:13:04,241] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:13:12,263] loop.py:380 INFO TIMING step=13315: step=3.817s data=1.666s compute=3.788s [fwd=2.719s bwd=1.054s opt=0.003s] dispatch=4.595s [2026-10-19 18:13:12,264] loop.py:45 INFO TRAIN_STEP step=13315 loss=1.420782 Training: 74%|███████▍ | 13315/18018 [28:42:58<10:31:29, 8.06s/step, loss=1.421, acc=0.680, acc_len=1.61, thru=14.6, I=14.9, T=33.5, wait=4.6s, pool=53, epoch=2/2][2026-10-19 18:13:20,699] loop.py:45 INFO TRAIN_STEP step=13316 loss=1.356378 Training: 74%|███████▍ | 13316/18018 [28:43:06<10:40:27, 8.17s/step, loss=1.356, acc=0.689, acc_len=1.64, thru=15.6, I=15.4, T=34.8, wait=4.7s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:13:20,695] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13315: forward=2624.6ms backward=1007.6ms [2026-10-19 18:13:29,081] loop.py:45 INFO TRAIN_STEP step=13317 loss=1.328037 Training: 74%|███████▍ | 13317/18018 [28:43:15<10:45:05, 8.23s/step, loss=1.328, acc=0.698, acc_len=1.68, thru=15.7, I=14.1, T=34.0, wait=4.6s, pool=53, epoch=2/2][2026-10-19 18:13:35,630] loop.py:45 INFO TRAIN_STEP step=13318 loss=1.110340 Training: 74%|███████▍ | 13318/18018 [28:43:21<10:05:20, 7.73s/step, loss=1.110, acc=0.731, acc_len=1.79, thru=17.1, I=17.0, T=38.8, wait=3.2s, pool=41, epoch=2/2][2026-10-19 18:13:43,665] loop.py:45 INFO TRAIN_STEP step=13319 loss=1.313937 Training: 74%|███████▍ | 13319/18018 [28:43:29<10:12:30, 7.82s/step, loss=1.314, acc=0.698, acc_len=1.67, thru=14.4, I=16.4, T=36.6, wait=4.5s, pool=25, epoch=2/2][2026-10-19 18:13:55,404] loop.py:380 INFO TIMING step=13320: step=3.933s data=1.644s compute=3.900s [fwd=2.748s bwd=1.137s opt=0.003s] dispatch=7.795s [2026-10-19 18:13:55,405] loop.py:45 INFO TRAIN_STEP step=13320 loss=1.273011 Training: 74%|███████▍ | 13320/18018 [28:43:41<11:44:24, 9.00s/step, loss=1.273, acc=0.709, acc_len=1.71, thru=16.5, I=12.8, T=32.5, wait=7.8s, pool=61, epoch=2/2][2026-10-19 18:14:04,656] loop.py:45 INFO TRAIN_STEP step=13321 loss=1.238738 Training: 74%|███████▍ | 13321/18018 [28:43:50<11:50:11, 9.07s/step, loss=1.239, acc=0.710, acc_len=1.72, thru=12.0, I=12.4, T=35.8, wait=5.7s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:14:04,651] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13320: forward=2549.5ms backward=987.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:12,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:14:13,015] loop.py:45 INFO TRAIN_STEP step=13322 loss=1.309352 Training: 74%|███████▍ | 13322/18018 [28:43:59<11:33:19, 8.86s/step, loss=1.309, acc=0.698, acc_len=1.67, thru=17.8, I=15.9, T=35.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:13,130] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 18:14:18,138] loop.py:45 INFO TRAIN_STEP step=13323 loss=1.187625 Training: 74%|███████▍ | 13323/18018 [28:44:04<10:05:32, 7.74s/step, loss=1.188, acc=0.717, acc_len=1.74, thru=19.3, I=22.9, T=37.9, wait=1.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:26,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:14:26,173] loop.py:45 INFO TRAIN_STEP step=13324 loss=1.192223 Training: 74%|███████▍ | 13324/18018 [28:44:12<10:12:18, 7.83s/step, loss=1.192, acc=0.721, acc_len=1.75, thru=17.2, I=16.6, T=33.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:26,516] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:35,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:14:35,793] loop.py:380 INFO TIMING step=13325: step=3.656s data=1.515s compute=3.624s [fwd=2.577s bwd=1.032s opt=0.003s] dispatch=5.955s [2026-10-19 18:14:35,793] loop.py:45 INFO TRAIN_STEP step=13325 loss=1.193914 Training: 74%|███████▍ | 13325/18018 [28:44:21<10:54:18, 8.37s/step, loss=1.194, acc=0.717, acc_len=1.74, thru=13.8, I=14.6, T=35.0, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:35,828] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 18:14:41,846] loop.py:45 INFO TRAIN_STEP step=13326 loss=1.217890 Training: 74%|███████▍ | 13326/18018 [28:44:27<9:59:50, 7.67s/step, loss=1.218, acc=0.713, acc_len=1.73, thru=18.2, I=16.6, T=38.6, wait=2.7s, pool=45, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:14:41,842] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13325: forward=2392.9ms backward=881.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:49,092] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:14:49,626] loop.py:45 INFO TRAIN_STEP step=13327 loss=1.158222 Training: 74%|███████▍ | 13327/18018 [28:44:35<10:02:17, 7.70s/step, loss=1.158, acc=0.727, acc_len=1.78, thru=18.3, I=16.9, T=36.0, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:50,099] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:57,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:14:57,824] loop.py:45 INFO TRAIN_STEP step=13328 loss=1.094541 Training: 74%|███████▍ | 13328/18018 [28:44:43<10:13:49, 7.85s/step, loss=1.095, acc=0.736, acc_len=1.81, thru=16.5, I=18.6, T=32.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:14:57,908] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:02,314] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:15:03,204] loop.py:45 INFO TRAIN_STEP step=13329 loss=1.342041 Training: 74%|███████▍ | 13329/18018 [28:44:49<9:15:40, 7.11s/step, loss=1.342, acc=0.692, acc_len=1.65, thru=20.2, I=21.9, T=36.7, wait=1.9s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:03,318] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:09,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:15:11,364] loop.py:380 INFO TIMING step=13330: step=3.760s data=1.692s compute=3.719s [fwd=2.680s bwd=1.025s opt=0.003s] dispatch=4.391s [2026-10-19 18:15:11,365] loop.py:45 INFO TRAIN_STEP step=13330 loss=1.248145 Training: 74%|███████▍ | 13330/18018 [28:44:57<9:40:12, 7.43s/step, loss=1.248, acc=0.706, acc_len=1.70, thru=15.3, I=19.6, T=34.0, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:11,608] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:18,915] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:15:19,504] loop.py:45 INFO TRAIN_STEP step=13331 loss=1.221642 Training: 74%|███████▍ | 13331/18018 [28:45:05<9:56:45, 7.64s/step, loss=1.222, acc=0.715, acc_len=1.73, thru=13.7, I=19.3, T=33.0, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:15:19,500] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13330: forward=2612.9ms backward=1225.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:19,921] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:15:27,679] loop.py:45 INFO TRAIN_STEP step=13332 loss=1.444715 Training: 74%|███████▍ | 13332/18018 [28:45:13<10:09:17, 7.80s/step, loss=1.445, acc=0.677, acc_len=1.60, thru=14.6, I=15.8, T=33.3, wait=4.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:32,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:15:34,431] loop.py:45 INFO TRAIN_STEP step=13333 loss=1.266429 Training: 74%|███████▍ | 13333/18018 [28:45:20<9:44:29, 7.49s/step, loss=1.266, acc=0.705, acc_len=1.70, thru=18.9, I=21.3, T=36.2, wait=3.2s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:34,460] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 18:15:40,606] loop.py:45 INFO TRAIN_STEP step=13334 loss=1.312683 Training: 74%|███████▍ | 13334/18018 [28:45:26<9:13:39, 7.09s/step, loss=1.313, acc=0.695, acc_len=1.66, thru=16.8, I=23.0, T=39.7, wait=2.9s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:50,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:15:51,562] loop.py:380 INFO TIMING step=13335: step=4.138s data=1.758s compute=4.107s [fwd=3.017s bwd=1.077s opt=0.003s] dispatch=6.810s [2026-10-19 18:15:51,562] loop.py:45 INFO TRAIN_STEP step=13335 loss=1.291431 Training: 74%|███████▍ | 13335/18018 [28:45:37<10:44:03, 8.25s/step, loss=1.291, acc=0.699, acc_len=1.68, thru=14.4, I=14.9, T=30.9, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:51,997] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:15:59,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:16:00,953] loop.py:45 INFO TRAIN_STEP step=13336 loss=1.257935 Training: 74%|███████▍ | 13336/18018 [28:45:47<11:10:34, 8.59s/step, loss=1.258, acc=0.709, acc_len=1.71, thru=13.2, I=14.7, T=30.4, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:16:00,948] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13335: forward=3019.9ms backward=1129.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:16:01,088] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 18:16:07,885] loop.py:45 INFO TRAIN_STEP step=13337 loss=1.299288 Training: 74%|███████▍ | 13337/18018 [28:45:53<10:31:30, 8.09s/step, loss=1.299, acc=0.696, acc_len=1.67, thru=14.9, I=18.1, T=36.4, wait=3.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:16:15,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:16:15,479] loop.py:45 INFO TRAIN_STEP step=13338 loss=1.395401 Training: 74%|███████▍ | 13338/18018 [28:46:01<10:19:42, 7.94s/step, loss=1.395, acc=0.685, acc_len=1.63, thru=18.1, I=17.8, T=37.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:16:15,654] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:16:21,778] loop.py:45 INFO TRAIN_STEP step=13339 loss=1.363962 Training: 74%|███████▍ | 13339/18018 [28:46:07<9:41:01, 7.45s/step, loss=1.364, acc=0.687, acc_len=1.64, thru=19.8, I=20.5, T=33.8, wait=2.5s, pool=57, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:16:28,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:16:29,495] loop.py:380 INFO TIMING step=13340: step=3.831s data=1.601s compute=3.797s [fwd=2.746s bwd=1.036s opt=0.003s] dispatch=3.878s [2026-10-19 18:16:29,495] loop.py:45 INFO TRAIN_STEP step=13340 loss=1.391324 Training: 74%|███████▍ | 13340/18018 [28:46:15<9:47:08, 7.53s/step, loss=1.391, acc=0.684, acc_len=1.62, thru=17.1, I=18.2, T=33.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:16:29,907] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:16:36,635] loop.py:45 INFO TRAIN_STEP step=13341 loss=1.201440 Training: 74%|███████▍ | 13341/18018 [28:46:22<9:37:54, 7.41s/step, loss=1.201, acc=0.719, acc_len=1.75, thru=15.7, I=17.5, T=36.9, wait=3.7s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:16:36,630] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13340: forward=2409.7ms backward=1006.0ms [2026-10-19 18:16:47,225] loop.py:45 INFO TRAIN_STEP step=13342 loss=1.317615 Training: 74%|███████▍ | 13342/18018 [28:46:33<10:52:00, 8.37s/step, loss=1.318, acc=0.700, acc_len=1.68, thru=10.7, I=13.8, T=27.7, wait=6.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:00,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:17:03,161] loop.py:45 INFO TRAIN_STEP step=13343 loss=1.408491 Training: 74%|███████▍ | 13343/18018 [28:46:49<13:48:47, 10.64s/step, loss=1.408, acc=0.683, acc_len=1.62, thru=9.9, I=10.8, T=27.1, wait=11.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:03,363] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:08,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:17:09,052] loop.py:45 INFO TRAIN_STEP step=13344 loss=1.291506 Training: 74%|███████▍ | 13344/18018 [28:46:55<11:57:46, 9.21s/step, loss=1.292, acc=0.702, acc_len=1.68, thru=17.6, I=25.7, T=38.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:09,191] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:13,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:17:15,266] loop.py:380 INFO TIMING step=13345: step=3.516s data=1.530s compute=3.488s [fwd=2.516s bwd=0.957s opt=0.003s] dispatch=2.688s [2026-10-19 18:17:15,267] loop.py:45 INFO TRAIN_STEP step=13345 loss=1.269188 Training: 74%|███████▍ | 13345/18018 [28:47:01<10:47:29, 8.31s/step, loss=1.269, acc=0.703, acc_len=1.69, thru=20.4, I=28.4, T=36.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:15,475] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:20,173] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:17:20,914] loop.py:45 INFO TRAIN_STEP step=13346 loss=1.335438 Training: 74%|███████▍ | 13346/18018 [28:47:06<9:45:03, 7.51s/step, loss=1.335, acc=0.698, acc_len=1.67, thru=21.3, I=29.4, T=38.9, wait=2.3s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:17:20,909] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13345: forward=2335.5ms backward=915.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:21,178] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:27,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:17:28,070] loop.py:45 INFO TRAIN_STEP step=13347 loss=1.361373 Training: 74%|███████▍ | 13347/18018 [28:47:14<9:36:36, 7.41s/step, loss=1.361, acc=0.687, acc_len=1.63, thru=18.1, I=20.7, T=33.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:28,317] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:34,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:17:36,201] loop.py:45 INFO TRAIN_STEP step=13348 loss=1.269313 Training: 74%|███████▍ | 13348/18018 [28:47:22<9:53:22, 7.62s/step, loss=1.269, acc=0.705, acc_len=1.69, thru=16.4, I=20.3, T=35.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:36,691] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:41,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:17:42,518] loop.py:45 INFO TRAIN_STEP step=13349 loss=1.260370 Training: 74%|███████▍ | 13349/18018 [28:47:28<9:22:45, 7.23s/step, loss=1.260, acc=0.701, acc_len=1.68, thru=19.4, I=27.4, T=38.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:42,627] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:47,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:17:48,638] loop.py:380 INFO TIMING step=13350: step=3.591s data=1.554s compute=3.559s [fwd=2.510s bwd=1.034s opt=0.003s] dispatch=2.520s [2026-10-19 18:17:48,638] loop.py:45 INFO TRAIN_STEP step=13350 loss=1.361925 Training: 74%|███████▍ | 13350/18018 [28:47:34<8:56:43, 6.90s/step, loss=1.362, acc=0.688, acc_len=1.64, thru=21.1, I=26.4, T=35.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:17:48,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:17:55,299] loop.py:45 INFO TRAIN_STEP step=13351 loss=1.212788 Training: 74%|███████▍ | 13351/18018 [28:47:41<8:51:05, 6.83s/step, loss=1.213, acc=0.715, acc_len=1.73, thru=16.1, I=19.9, T=36.1, wait=3.1s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:17:55,294] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13350: forward=2510.2ms backward=979.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:18:02,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:18:03,554] loop.py:45 INFO TRAIN_STEP step=13352 loss=1.315948 Training: 74%|███████▍ | 13352/18018 [28:47:49<9:24:14, 7.26s/step, loss=1.316, acc=0.698, acc_len=1.67, thru=16.0, I=16.2, T=34.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:18:03,943] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:18:09,685] loop.py:45 INFO TRAIN_STEP step=13353 loss=1.196318 Training: 74%|███████▍ | 13353/18018 [28:47:55<8:57:51, 6.92s/step, loss=1.196, acc=0.716, acc_len=1.74, thru=18.2, I=21.9, T=37.0, wait=2.7s, pool=45, epoch=2/2][2026-10-19 18:18:17,936] loop.py:45 INFO TRAIN_STEP step=13354 loss=1.214900 Training: 74%|███████▍ | 13354/18018 [28:48:04<9:28:47, 7.32s/step, loss=1.215, acc=0.714, acc_len=1.73, thru=15.1, I=15.0, T=33.6, wait=4.4s, pool=49, epoch=2/2][2026-10-19 18:18:26,640] loop.py:380 INFO TIMING step=13355: step=3.806s data=1.560s compute=3.775s [fwd=2.620s bwd=1.140s opt=0.003s] dispatch=4.890s [2026-10-19 18:18:26,640] loop.py:45 INFO TRAIN_STEP step=13355 loss=1.368991 Training: 74%|███████▍ | 13355/18018 [28:48:12<10:01:01, 7.73s/step, loss=1.369, acc=0.686, acc_len=1.64, thru=14.5, I=12.9, T=33.6, wait=4.9s, pool=45, epoch=2/2][2026-10-19 18:18:33,679] loop.py:45 INFO TRAIN_STEP step=13356 loss=1.349758 Training: 74%|███████▍ | 13356/18018 [28:48:19<9:44:42, 7.53s/step, loss=1.350, acc=0.691, acc_len=1.65, thru=15.3, I=16.1, T=38.4, wait=3.7s, pool=33, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:18:33,674] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13355: forward=2381.6ms backward=905.1ms [2026-10-19 18:18:45,834] loop.py:45 INFO TRAIN_STEP step=13357 loss=1.314135 Training: 74%|███████▍ | 13357/18018 [28:48:31<11:32:29, 8.91s/step, loss=1.314, acc=0.699, acc_len=1.67, thru=18.7, I=11.2, T=33.4, wait=8.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:18:53,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:18:54,669] loop.py:45 INFO TRAIN_STEP step=13358 loss=1.376901 Training: 74%|███████▍ | 13358/18018 [28:48:40<11:30:30, 8.89s/step, loss=1.377, acc=0.685, acc_len=1.63, thru=16.2, I=17.9, T=34.6, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:18:54,763] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:00,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:19:01,281] loop.py:45 INFO TRAIN_STEP step=13359 loss=1.288353 Training: 74%|███████▍ | 13359/18018 [28:48:47<10:37:14, 8.21s/step, loss=1.288, acc=0.705, acc_len=1.69, thru=18.9, I=22.6, T=37.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:01,746] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:06,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:19:07,184] loop.py:380 INFO TIMING step=13360: step=3.413s data=1.437s compute=3.386s [fwd=2.431s bwd=0.941s opt=0.003s] dispatch=2.482s [2026-10-19 18:19:07,184] loop.py:45 INFO TRAIN_STEP step=13360 loss=1.089840 Training: 74%|███████▍ | 13360/18018 [28:48:53<9:43:27, 7.52s/step, loss=1.090, acc=0.741, acc_len=1.82, thru=20.9, I=27.8, T=37.5, wait=2.5s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:07,682] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:19:13,341] loop.py:45 INFO TRAIN_STEP step=13361 loss=1.179448 Training: 74%|███████▍ | 13361/18018 [28:48:59<9:11:41, 7.11s/step, loss=1.179, acc=0.720, acc_len=1.75, thru=18.9, I=21.8, T=38.9, wait=2.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:19:13,336] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13360: forward=2318.5ms backward=925.5ms [2026-10-19 18:19:23,037] loop.py:45 INFO TRAIN_STEP step=13362 loss=1.270564 Training: 74%|███████▍ | 13362/18018 [28:49:09<10:11:55, 7.89s/step, loss=1.271, acc=0.704, acc_len=1.69, thru=13.1, I=13.1, T=35.0, wait=6.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:33,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:19:35,457] loop.py:45 INFO TRAIN_STEP step=13363 loss=1.292493 Training: 74%|███████▍ | 13363/18018 [28:49:21<11:57:14, 9.24s/step, loss=1.292, acc=0.704, acc_len=1.70, thru=12.3, I=14.4, T=30.7, wait=8.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:35,523] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 [2026-10-19 18:19:42,077] loop.py:45 INFO TRAIN_STEP step=13364 loss=1.273448 Training: 74%|███████▍ | 13364/18018 [28:49:28<10:56:02, 8.46s/step, loss=1.273, acc=0.703, acc_len=1.70, thru=16.0, I=22.1, T=37.2, wait=3.2s, pool=53, epoch=2/2][2026-10-19 18:19:49,861] loop.py:380 INFO TIMING step=13365: step=3.691s data=1.592s compute=3.657s [fwd=2.617s bwd=1.026s opt=0.003s] dispatch=4.085s [2026-10-19 18:19:49,861] loop.py:45 INFO TRAIN_STEP step=13365 loss=1.315162 Training: 74%|███████▍ | 13365/18018 [28:49:35<10:40:10, 8.26s/step, loss=1.315, acc=0.691, acc_len=1.65, thru=16.7, I=16.6, T=34.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:49,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:50,312] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:57,110] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:19:57,212] loop.py:45 INFO TRAIN_STEP step=13366 loss=1.297866 Training: 74%|███████▍ | 13366/18018 [28:49:43<10:19:00, 7.98s/step, loss=1.298, acc=0.704, acc_len=1.69, thru=17.9, I=18.3, T=36.7, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:19:57,208] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13365: forward=2450.0ms backward=986.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:19:57,612] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:20:03,717] loop.py:45 INFO TRAIN_STEP step=13367 loss=1.378160 Training: 74%|███████▍ | 13367/18018 [28:49:49<9:44:29, 7.54s/step, loss=1.378, acc=0.685, acc_len=1.62, thru=18.2, I=19.1, T=36.4, wait=3.0s, pool=53, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:20:11,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:20:12,427] loop.py:45 INFO TRAIN_STEP step=13368 loss=1.237680 Training: 74%|███████▍ | 13368/18018 [28:49:58<10:11:32, 7.89s/step, loss=1.238, acc=0.711, acc_len=1.72, thru=17.0, I=17.1, T=34.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:20:12,621] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:20:18,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:20:18,876] loop.py:45 INFO TRAIN_STEP step=13369 loss=1.372406 Training: 74%|███████▍ | 13369/18018 [28:50:04<9:37:54, 7.46s/step, loss=1.372, acc=0.688, acc_len=1.64, thru=19.9, I=18.6, T=37.5, wait=3.0s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:20:19,326] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:20:25,526] loop.py:380 INFO TIMING step=13370: step=3.634s data=1.610s compute=3.603s [fwd=2.564s bwd=1.024s opt=0.003s] dispatch=3.009s [2026-10-19 18:20:25,526] loop.py:45 INFO TRAIN_STEP step=13370 loss=1.275497 Training: 74%|███████▍ | 13370/18018 [28:50:11<9:19:03, 7.22s/step, loss=1.275, acc=0.705, acc_len=1.70, thru=17.9, I=19.9, T=35.2, wait=3.0s, pool=49, epoch=2/2][2026-10-19 18:20:33,936] loop.py:45 INFO TRAIN_STEP step=13371 loss=1.314323 Training: 74%|███████▍ | 13371/18018 [28:50:20<9:46:40, 7.57s/step, loss=1.314, acc=0.696, acc_len=1.66, thru=14.8, I=16.1, T=35.8, wait=4.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:20:33,931] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13370: forward=2513.3ms backward=1010.3ms [2026-10-19 18:20:43,277] loop.py:45 INFO TRAIN_STEP step=13372 loss=1.273950 Training: 74%|███████▍ | 13372/18018 [28:50:29<10:27:30, 8.10s/step, loss=1.274, acc=0.703, acc_len=1.69, thru=11.8, I=10.7, T=33.8, wait=5.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:20:51,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:20:52,021] loop.py:45 INFO TRAIN_STEP step=13373 loss=1.340900 Training: 74%|███████▍ | 13373/18018 [28:50:38<10:42:15, 8.30s/step, loss=1.341, acc=0.696, acc_len=1.67, thru=17.8, I=15.6, T=28.8, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:20:52,248] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:20:59,840] loop.py:45 INFO TRAIN_STEP step=13374 loss=1.243354 Training: 74%|███████▍ | 13374/18018 [28:50:45<10:31:04, 8.15s/step, loss=1.243, acc=0.712, acc_len=1.72, thru=14.6, I=14.2, T=36.1, wait=4.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:06,122] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:21:07,718] loop.py:380 INFO TIMING step=13375: step=4.372s data=2.358s compute=4.340s [fwd=3.401s bwd=0.925s opt=0.003s] dispatch=3.496s [2026-10-19 18:21:07,718] loop.py:45 INFO TRAIN_STEP step=13375 loss=1.376291 Training: 74%|███████▍ | 13375/18018 [28:50:53<10:24:29, 8.07s/step, loss=1.376, acc=0.683, acc_len=1.62, thru=16.7, I=19.4, T=29.3, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:08,136] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:13,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 18:21:14,458] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13375: forward=3098.7ms backward=933.9ms [2026-10-19 18:21:14,463] loop.py:45 INFO TRAIN_STEP step=13376 loss=1.289499 Training: 74%|███████▍ | 13376/18018 [28:51:00<9:53:39, 7.67s/step, loss=1.289, acc=0.699, acc_len=1.68, thru=17.3, I=22.7, T=31.3, wait=2.7s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:14,658] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:21,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:21:23,167] loop.py:45 INFO TRAIN_STEP step=13377 loss=1.249985 Training: 74%|███████▍ | 13377/18018 [28:51:09<10:17:27, 7.98s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=13.8, I=16.8, T=34.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:23,263] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-19 18:21:29,586] loop.py:45 INFO TRAIN_STEP step=13378 loss=1.217380 Training: 74%|███████▍ | 13378/18018 [28:51:15<9:41:05, 7.51s/step, loss=1.217, acc=0.712, acc_len=1.72, thru=16.8, I=19.4, T=37.7, wait=3.0s, pool=45, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:35,954] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:21:36,306] loop.py:45 INFO TRAIN_STEP step=13379 loss=1.123698 Training: 74%|███████▍ | 13379/18018 [28:51:22<9:22:27, 7.27s/step, loss=1.124, acc=0.731, acc_len=1.79, thru=19.4, I=19.6, T=33.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:36,459] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:21:41,915] loop.py:380 INFO TIMING step=13380: step=3.446s data=1.484s compute=3.418s [fwd=2.454s bwd=0.951s opt=0.003s] dispatch=2.155s [2026-10-19 18:21:41,915] loop.py:45 INFO TRAIN_STEP step=13380 loss=1.186822 Training: 74%|███████▍ | 13380/18018 [28:51:27<8:43:41, 6.77s/step, loss=1.187, acc=0.723, acc_len=1.76, thru=20.4, I=22.9, T=37.1, wait=2.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:51,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:21:51,886] loop.py:45 INFO TRAIN_STEP step=13381 loss=1.211645 Training: 74%|███████▍ | 13381/18018 [28:51:37<9:57:40, 7.73s/step, loss=1.212, acc=0.717, acc_len=1.74, thru=15.4, I=14.2, T=34.2, wait=6.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:21:51,882] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13380: forward=2657.8ms backward=1034.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:21:52,253] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:22:03,826] loop.py:45 INFO TRAIN_STEP step=13382 loss=1.312131 Training: 74%|███████▍ | 13382/18018 [28:51:49<11:35:04, 9.00s/step, loss=1.312, acc=0.698, acc_len=1.67, thru=8.7, I=7.6, T=32.8, wait=8.0s, pool=37, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:14,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:22:14,501] loop.py:45 INFO TRAIN_STEP step=13383 loss=1.323928 Training: 74%|███████▍ | 13383/18018 [28:52:00<12:13:50, 9.50s/step, loss=1.324, acc=0.697, acc_len=1.67, thru=15.3, I=14.5, T=32.2, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:14,745] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:21,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:22:22,210] loop.py:45 INFO TRAIN_STEP step=13384 loss=1.262386 Training: 74%|███████▍ | 13384/18018 [28:52:08<11:32:10, 8.96s/step, loss=1.262, acc=0.708, acc_len=1.70, thru=18.7, I=17.3, T=35.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:22,420] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:28,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:22:29,439] loop.py:380 INFO TIMING step=13385: step=3.692s data=1.638s compute=3.643s [fwd=2.553s bwd=1.076s opt=0.003s] dispatch=3.530s [2026-10-19 18:22:29,439] loop.py:45 INFO TRAIN_STEP step=13385 loss=1.276737 Training: 74%|███████▍ | 13385/18018 [28:52:15<10:51:54, 8.44s/step, loss=1.277, acc=0.706, acc_len=1.70, thru=17.7, I=19.4, T=34.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:29,504] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:35,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:22:36,202] loop.py:45 INFO TRAIN_STEP step=13386 loss=1.196571 Training: 74%|███████▍ | 13386/18018 [28:52:22<10:12:54, 7.94s/step, loss=1.197, acc=0.721, acc_len=1.75, thru=18.2, I=20.6, T=36.9, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:22:36,197] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13385: forward=2441.3ms backward=969.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:36,309] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:43,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:22:44,035] loop.py:45 INFO TRAIN_STEP step=13387 loss=1.175063 Training: 74%|███████▍ | 13387/18018 [28:52:30<10:10:17, 7.91s/step, loss=1.175, acc=0.722, acc_len=1.76, thru=16.9, I=18.9, T=36.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:44,478] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:22:52,002] loop.py:45 INFO TRAIN_STEP step=13388 loss=1.291264 Training: 74%|███████▍ | 13388/18018 [28:52:38<10:11:32, 7.92s/step, loss=1.291, acc=0.697, acc_len=1.67, thru=16.4, I=17.4, T=36.4, wait=4.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:58,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:22:58,881] loop.py:45 INFO TRAIN_STEP step=13389 loss=1.300934 Training: 74%|███████▍ | 13389/18018 [28:52:44<9:47:13, 7.61s/step, loss=1.301, acc=0.699, acc_len=1.68, thru=19.0, I=20.3, T=36.9, wait=3.4s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:22:59,095] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:23:05,622] loop.py:380 INFO TIMING step=13390: step=3.643s data=1.591s compute=3.615s [fwd=2.545s bwd=1.056s opt=0.003s] dispatch=3.085s [2026-10-19 18:23:05,622] loop.py:45 INFO TRAIN_STEP step=13390 loss=1.243335 Training: 74%|███████▍ | 13390/18018 [28:52:51<9:26:56, 7.35s/step, loss=1.243, acc=0.708, acc_len=1.71, thru=17.4, I=19.8, T=35.1, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:23:15,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:23:15,939] loop.py:45 INFO TRAIN_STEP step=13391 loss=1.238053 Training: 74%|███████▍ | 13391/18018 [28:53:02<10:35:29, 8.24s/step, loss=1.238, acc=0.712, acc_len=1.72, thru=14.6, I=13.7, T=33.8, wait=6.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:23:15,933] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13390: forward=2688.4ms backward=1028.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:23:16,268] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:23:24,029] loop.py:45 INFO TRAIN_STEP step=13392 loss=1.222469 Training: 74%|███████▍ | 13392/18018 [28:53:10<10:31:49, 8.19s/step, loss=1.222, acc=0.714, acc_len=1.72, thru=14.5, I=16.0, T=29.7, wait=3.8s, pool=53, epoch=2/2][2026-10-19 18:23:32,548] loop.py:45 INFO TRAIN_STEP step=13393 loss=1.292518 Training: 74%|███████▍ | 13393/18018 [28:53:18<10:39:11, 8.29s/step, loss=1.293, acc=0.702, acc_len=1.68, thru=12.8, I=12.7, T=36.3, wait=5.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:23:39,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:23:40,528] loop.py:45 INFO TRAIN_STEP step=13394 loss=1.289528 Training: 74%|███████▍ | 13394/18018 [28:53:26<10:31:48, 8.20s/step, loss=1.290, acc=0.699, acc_len=1.68, thru=18.1, I=18.8, T=34.8, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:23:40,648] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:23:46,418] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:23:47,204] loop.py:380 INFO TIMING step=13395: step=3.468s data=1.464s compute=3.438s [fwd=2.475s bwd=0.949s opt=0.003s] dispatch=3.200s [2026-10-19 18:23:47,204] loop.py:45 INFO TRAIN_STEP step=13395 loss=1.262481 Training: 74%|███████▍ | 13395/18018 [28:53:33<9:56:29, 7.74s/step, loss=1.262, acc=0.708, acc_len=1.71, thru=18.4, I=21.6, T=36.9, wait=3.2s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:23:47,422] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:23:53,114] loop.py:45 INFO TRAIN_STEP step=13396 loss=1.204869 Training: 74%|███████▍ | 13396/18018 [28:53:39<9:14:03, 7.19s/step, loss=1.205, acc=0.719, acc_len=1.75, thru=20.0, I=20.6, T=37.0, wait=2.4s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:23:53,111] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13395: forward=2430.5ms backward=970.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:00,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:24:00,266] loop.py:45 INFO TRAIN_STEP step=13397 loss=1.423475 Training: 74%|███████▍ | 13397/18018 [28:53:46<9:12:58, 7.18s/step, loss=1.423, acc=0.681, acc_len=1.62, thru=18.4, I=18.7, T=35.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:00,632] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:07,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:24:07,696] loop.py:45 INFO TRAIN_STEP step=13398 loss=1.231387 Training: 74%|███████▍ | 13398/18018 [28:53:53<9:18:40, 7.26s/step, loss=1.231, acc=0.715, acc_len=1.73, thru=17.3, I=19.2, T=34.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:07,922] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:24:13,575] loop.py:45 INFO TRAIN_STEP step=13399 loss=1.242949 Training: 74%|███████▍ | 13399/18018 [28:53:59<8:46:43, 6.84s/step, loss=1.243, acc=0.709, acc_len=1.71, thru=19.3, I=20.1, T=39.5, wait=2.6s, pool=49, epoch=2/2][2026-10-19 18:24:22,069] loop.py:380 INFO TIMING step=13400: step=3.795s data=1.570s compute=3.736s [fwd=2.616s bwd=1.104s opt=0.003s] dispatch=4.690s [2026-10-19 18:24:22,069] loop.py:45 INFO TRAIN_STEP step=13400 loss=1.240957 Training: 74%|███████▍ | 13400/18018 [28:54:08<9:24:43, 7.34s/step, loss=1.241, acc=0.710, acc_len=1.71, thru=15.2, I=15.7, T=33.7, wait=4.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:29,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:24:30,319] loop.py:45 INFO TRAIN_STEP step=13401 loss=1.332910 Training: 74%|███████▍ | 13401/18018 [28:54:16<9:45:41, 7.61s/step, loss=1.333, acc=0.693, acc_len=1.66, thru=16.4, I=16.9, T=33.5, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:24:30,314] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13400: forward=2698.3ms backward=1072.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:30,479] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:34,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:24:35,777] loop.py:45 INFO TRAIN_STEP step=13402 loss=1.320769 Training: 74%|███████▍ | 13402/18018 [28:54:21<8:55:54, 6.97s/step, loss=1.321, acc=0.696, acc_len=1.66, thru=20.4, I=24.3, T=37.4, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:36,179] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:41,647] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:24:41,815] loop.py:45 INFO TRAIN_STEP step=13403 loss=1.224353 Training: 74%|███████▍ | 13403/18018 [28:54:27<8:34:19, 6.69s/step, loss=1.224, acc=0.706, acc_len=1.71, thru=21.1, I=26.3, T=38.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:42,149] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:24:47,927] loop.py:45 INFO TRAIN_STEP step=13404 loss=1.230946 Training: 74%|███████▍ | 13404/18018 [28:54:34<8:20:57, 6.51s/step, loss=1.231, acc=0.708, acc_len=1.70, thru=21.2, I=22.3, T=37.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:47,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:24:48,397] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:24:54,649] loop.py:380 INFO TIMING step=13405: step=3.461s data=1.437s compute=3.430s [fwd=2.484s bwd=0.932s opt=0.003s] dispatch=3.253s [2026-10-19 18:24:54,649] loop.py:45 INFO TRAIN_STEP step=13405 loss=1.269204 Training: 74%|███████▍ | 13405/18018 [28:54:40<8:25:38, 6.58s/step, loss=1.269, acc=0.706, acc_len=1.70, thru=17.3, I=18.1, T=37.0, wait=3.3s, pool=45, epoch=2/2][2026-10-19 18:25:03,125] loop.py:45 INFO TRAIN_STEP step=13406 loss=1.202615 Training: 74%|███████▍ | 13406/18018 [28:54:49<9:09:23, 7.15s/step, loss=1.203, acc=0.714, acc_len=1.72, thru=15.2, I=16.9, T=36.3, wait=4.9s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:25:03,120] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13405: forward=2508.3ms backward=977.2ms [2026-10-19 18:25:12,469] loop.py:45 INFO TRAIN_STEP step=13407 loss=1.337087 Training: 74%|███████▍ | 13407/18018 [28:54:58<9:59:51, 7.81s/step, loss=1.337, acc=0.689, acc_len=1.64, thru=14.1, I=12.2, T=32.9, wait=5.4s, pool=45, epoch=2/2][2026-10-19 18:25:22,233] loop.py:45 INFO TRAIN_STEP step=13408 loss=1.345502 Training: 74%|███████▍ | 13408/18018 [28:55:08<10:44:53, 8.39s/step, loss=1.346, acc=0.689, acc_len=1.64, thru=12.6, I=13.5, T=32.3, wait=5.8s, pool=37, epoch=2/2][2026-10-19 18:25:33,876] loop.py:45 INFO TRAIN_STEP step=13409 loss=1.213823 Training: 74%|███████▍ | 13409/18018 [28:55:19<11:59:36, 9.37s/step, loss=1.214, acc=0.718, acc_len=1.74, thru=11.3, I=10.9, T=30.4, wait=7.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:25:41,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:25:42,842] loop.py:380 INFO TIMING step=13410: step=4.376s data=1.976s compute=4.349s [fwd=3.110s bwd=1.224s opt=0.003s] dispatch=4.581s [2026-10-19 18:25:42,842] loop.py:45 INFO TRAIN_STEP step=13410 loss=1.294405 Training: 74%|███████▍ | 13410/18018 [28:55:28<11:50:12, 9.25s/step, loss=1.294, acc=0.702, acc_len=1.69, thru=16.4, I=16.6, T=29.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:25:43,190] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 18:25:49,332] loop.py:45 INFO TRAIN_STEP step=13411 loss=1.273840 Training: 74%|███████▍ | 13411/18018 [28:55:35<10:46:33, 8.42s/step, loss=1.274, acc=0.707, acc_len=1.70, thru=16.1, I=19.6, T=37.4, wait=3.1s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:25:49,327] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13410: forward=2412.4ms backward=959.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:25:56,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:25:57,965] loop.py:45 INFO TRAIN_STEP step=13412 loss=1.328393 Training: 74%|███████▍ | 13412/18018 [28:55:44<10:51:16, 8.48s/step, loss=1.328, acc=0.698, acc_len=1.68, thru=17.2, I=18.7, T=32.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:25:58,351] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 18:26:04,585] loop.py:45 INFO TRAIN_STEP step=13413 loss=1.273381 Training: 74%|███████▍ | 13413/18018 [28:55:50<10:08:13, 7.92s/step, loss=1.273, acc=0.706, acc_len=1.70, thru=16.3, I=20.6, T=35.8, wait=3.0s, pool=49, epoch=2/2][2026-10-19 18:26:14,443] loop.py:45 INFO TRAIN_STEP step=13414 loss=1.378689 Training: 74%|███████▍ | 13414/18018 [28:56:00<10:52:38, 8.51s/step, loss=1.379, acc=0.688, acc_len=1.64, thru=14.8, I=14.3, T=33.0, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:14,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:14,892] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:19,916] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:26:20,717] loop.py:380 INFO TIMING step=13415: step=3.449s data=1.414s compute=3.422s [fwd=2.475s bwd=0.932s opt=0.003s] dispatch=2.814s [2026-10-19 18:26:20,717] loop.py:45 INFO TRAIN_STEP step=13415 loss=1.424298 Training: 74%|███████▍ | 13415/18018 [28:56:06<10:01:08, 7.84s/step, loss=1.424, acc=0.682, acc_len=1.62, thru=19.8, I=21.1, T=37.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:20,920] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:26,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:26:27,723] loop.py:45 INFO TRAIN_STEP step=13416 loss=1.297325 Training: 74%|███████▍ | 13416/18018 [28:56:13<9:41:55, 7.59s/step, loss=1.297, acc=0.699, acc_len=1.68, thru=18.1, I=21.7, T=37.0, wait=3.5s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:26:27,719] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13415: forward=2468.8ms backward=941.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:27,796] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 18:26:33,643] loop.py:45 INFO TRAIN_STEP step=13417 loss=1.123141 Training: 74%|███████▍ | 13417/18018 [28:56:19<9:03:23, 7.09s/step, loss=1.123, acc=0.729, acc_len=1.78, thru=20.0, I=19.6, T=37.2, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:39,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:26:39,826] loop.py:45 INFO TRAIN_STEP step=13418 loss=1.225610 Training: 74%|███████▍ | 13418/18018 [28:56:25<8:42:31, 6.82s/step, loss=1.226, acc=0.715, acc_len=1.73, thru=20.8, I=20.9, T=33.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:40,270] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:47,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:26:48,234] loop.py:45 INFO TRAIN_STEP step=13419 loss=1.399015 Training: 74%|███████▍ | 13419/18018 [28:56:34<9:19:02, 7.29s/step, loss=1.399, acc=0.683, acc_len=1.62, thru=16.3, I=17.2, T=35.4, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:26:48,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:26:57,345] loop.py:380 INFO TIMING step=13420: step=3.549s data=1.463s compute=3.514s [fwd=2.508s bwd=0.991s opt=0.003s] dispatch=5.552s [2026-10-19 18:26:57,345] loop.py:45 INFO TRAIN_STEP step=13420 loss=1.211070 Training: 74%|███████▍ | 13420/18018 [28:56:43<10:00:40, 7.84s/step, loss=1.211, acc=0.716, acc_len=1.74, thru=10.8, I=12.2, T=36.1, wait=5.6s, pool=33, epoch=2/2][2026-10-19 18:27:06,083] loop.py:45 INFO TRAIN_STEP step=13421 loss=1.316128 Training: 74%|███████▍ | 13421/18018 [28:56:52<10:21:13, 8.11s/step, loss=1.316, acc=0.699, acc_len=1.67, thru=14.1, I=16.2, T=33.9, wait=5.0s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:27:06,078] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13420: forward=2713.0ms backward=1003.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:27:12,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:27:13,668] loop.py:45 INFO TRAIN_STEP step=13422 loss=1.161907 Training: 74%|███████▍ | 13422/18018 [28:56:59<10:09:05, 7.95s/step, loss=1.162, acc=0.724, acc_len=1.77, thru=18.1, I=18.0, T=35.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:27:13,815] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:27:19,972] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:27:20,409] loop.py:45 INFO TRAIN_STEP step=13423 loss=1.283557 Training: 74%|███████▍ | 13423/18018 [28:57:06<9:41:11, 7.59s/step, loss=1.284, acc=0.703, acc_len=1.69, thru=18.9, I=21.2, T=37.1, wait=3.3s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:27:20,484] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:27:29,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:27:29,099] loop.py:45 INFO TRAIN_STEP step=13424 loss=1.249152 Training: 75%|███████▍ | 13424/18018 [28:57:15<10:06:16, 7.92s/step, loss=1.249, acc=0.707, acc_len=1.71, thru=14.8, I=14.9, T=35.4, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:27:29,533] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:27:37,327] loop.py:380 INFO TIMING step=13425: step=3.662s data=1.525s compute=3.625s [fwd=2.583s bwd=1.027s opt=0.003s] dispatch=4.558s [2026-10-19 18:27:37,327] loop.py:45 INFO TRAIN_STEP step=13425 loss=1.233019 Training: 75%|███████▍ | 13425/18018 [28:57:23<10:13:18, 8.01s/step, loss=1.233, acc=0.711, acc_len=1.72, thru=14.9, I=16.5, T=35.0, wait=4.6s, pool=53, epoch=2/2][2026-10-19 18:27:45,781] loop.py:45 INFO TRAIN_STEP step=13426 loss=1.336747 Training: 75%|███████▍ | 13426/18018 [28:57:31<10:23:16, 8.14s/step, loss=1.337, acc=0.695, acc_len=1.67, thru=13.2, I=13.5, T=33.9, wait=4.7s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:27:45,776] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13425: forward=2669.7ms backward=1057.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:27:54,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:27:54,691] loop.py:45 INFO TRAIN_STEP step=13427 loss=1.291003 Training: 75%|███████▍ | 13427/18018 [28:57:40<10:40:44, 8.37s/step, loss=1.291, acc=0.703, acc_len=1.70, thru=16.8, I=17.3, T=34.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:27:54,982] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:28:07,199] loop.py:45 INFO TRAIN_STEP step=13428 loss=1.309086 Training: 75%|███████▍ | 13428/18018 [28:57:53<12:15:27, 9.61s/step, loss=1.309, acc=0.698, acc_len=1.68, thru=10.4, I=8.7, T=34.9, wait=8.8s, pool=57, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:28:14,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:28:15,664] loop.py:45 INFO TRAIN_STEP step=13429 loss=1.350554 Training: 75%|███████▍ | 13429/18018 [28:58:01<11:48:59, 9.27s/step, loss=1.351, acc=0.691, acc_len=1.65, thru=17.3, I=18.4, T=35.7, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:28:16,005] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:28:22,517] loop.py:380 INFO TIMING step=13430: step=3.794s data=1.737s compute=3.761s [fwd=2.718s bwd=1.029s opt=0.003s] dispatch=3.049s [2026-10-19 18:28:22,517] loop.py:45 INFO TRAIN_STEP step=13430 loss=1.352345 Training: 75%|███████▍ | 13430/18018 [28:58:08<10:53:20, 8.54s/step, loss=1.352, acc=0.689, acc_len=1.65, thru=17.6, I=24.0, T=33.7, wait=3.0s, pool=57, epoch=2/2][2026-10-19 18:28:29,478] loop.py:45 INFO TRAIN_STEP step=13431 loss=1.167773 Training: 75%|███████▍ | 13431/18018 [28:58:15<10:17:03, 8.07s/step, loss=1.168, acc=0.720, acc_len=1.75, thru=14.4, I=16.7, T=37.3, wait=3.5s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:28:29,473] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13430: forward=2431.7ms backward=954.4ms [2026-10-19 18:28:41,488] loop.py:45 INFO TRAIN_STEP step=13432 loss=1.327641 Training: 75%|███████▍ | 13432/18018 [28:58:27<11:47:05, 9.25s/step, loss=1.328, acc=0.697, acc_len=1.67, thru=13.7, I=11.4, T=33.8, wait=8.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:28:46,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:28:47,716] loop.py:45 INFO TRAIN_STEP step=13433 loss=1.276224 Training: 75%|███████▍ | 13433/18018 [28:58:33<10:37:37, 8.34s/step, loss=1.276, acc=0.701, acc_len=1.68, thru=19.2, I=21.4, T=37.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:28:48,087] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:28:56,102] loop.py:45 INFO TRAIN_STEP step=13434 loss=1.330300 Training: 75%|███████▍ | 13434/18018 [28:58:42<10:38:28, 8.36s/step, loss=1.330, acc=0.696, acc_len=1.67, thru=9.7, I=12.9, T=36.3, wait=4.8s, pool=21, epoch=2/2] (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 18:28:56,132] eagle_store.py:517 WARNING batch_get_buffer missing keys (k429886-8e78c2c0_hs, k429886-8e78c2c0_ids, k429886-8e78c2c0_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:03,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 18:28:56,135] eagle_store.py:517 WARNING batch_get_buffer missing keys (k429887-b94063ff_hs, k429887-b94063ff_ids, k429887-b94063ff_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 18:29:04,320] loop.py:380 INFO TIMING step=13435: step=3.646s data=1.570s compute=3.599s [fwd=2.552s bwd=1.034s opt=0.003s] dispatch=4.563s [2026-10-19 18:29:04,320] loop.py:45 INFO TRAIN_STEP step=13435 loss=1.270467 Training: 75%|███████▍ | 13435/18018 [28:58:50<10:35:07, 8.31s/step, loss=1.270, acc=0.705, acc_len=1.69, thru=20.1, I=17.0, T=35.1, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:04,743] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:10,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:29:10,974] loop.py:45 INFO TRAIN_STEP step=13436 loss=1.349796 Training: 75%|███████▍ | 13436/18018 [28:58:57<9:56:55, 7.82s/step, loss=1.350, acc=0.687, acc_len=1.64, thru=19.4, I=20.4, T=34.6, wait=2.9s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:29:10,969] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13435: forward=2639.3ms backward=1015.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:11,402] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:16,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:29:17,284] loop.py:45 INFO TRAIN_STEP step=13437 loss=1.236788 Training: 75%|███████▍ | 13437/18018 [28:59:03<9:22:19, 7.37s/step, loss=1.237, acc=0.713, acc_len=1.72, thru=19.8, I=21.6, T=36.7, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:17,509] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:29:23,441] loop.py:45 INFO TRAIN_STEP step=13438 loss=1.318143 Training: 75%|███████▍ | 13438/18018 [28:59:09<8:54:30, 7.00s/step, loss=1.318, acc=0.693, acc_len=1.65, thru=19.8, I=21.0, T=37.9, wait=2.8s, pool=61, epoch=2/2][2026-10-19 18:29:29,610] loop.py:45 INFO TRAIN_STEP step=13439 loss=1.322863 Training: 75%|███████▍ | 13439/18018 [28:59:15<8:35:19, 6.75s/step, loss=1.323, acc=0.696, acc_len=1.67, thru=18.8, I=22.2, T=37.6, wait=2.8s, pool=45, epoch=2/2][2026-10-19 18:29:37,856] loop.py:380 INFO TIMING step=13440: step=3.821s data=1.698s compute=3.788s [fwd=2.678s bwd=1.096s opt=0.003s] dispatch=4.417s [2026-10-19 18:29:37,857] loop.py:45 INFO TRAIN_STEP step=13440 loss=1.346852 Training: 75%|███████▍ | 13440/18018 [28:59:23<9:09:23, 7.20s/step, loss=1.347, acc=0.695, acc_len=1.66, thru=14.7, I=15.4, T=33.5, wait=4.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:43,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:29:44,718] loop.py:45 INFO TRAIN_STEP step=13441 loss=1.355147 Training: 75%|███████▍ | 13441/18018 [28:59:30<9:01:33, 7.10s/step, loss=1.355, acc=0.694, acc_len=1.66, thru=17.8, I=19.6, T=36.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:44,787] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (TrainerActor pid=9310) [2026-10-19 18:29:44,714] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13440: forward=2475.2ms backward=974.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:50,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:29:51,292] loop.py:45 INFO TRAIN_STEP step=13442 loss=1.248098 Training: 75%|███████▍ | 13442/18018 [28:59:37<8:49:22, 6.94s/step, loss=1.248, acc=0.711, acc_len=1.72, thru=19.3, I=21.9, T=34.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:51,476] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:55,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:29:56,624] loop.py:45 INFO TRAIN_STEP step=13443 loss=1.194401 Training: 75%|███████▍ | 13443/18018 [28:59:42<8:12:30, 6.46s/step, loss=1.194, acc=0.718, acc_len=1.74, thru=21.6, I=25.8, T=37.6, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:29:56,678] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 18:30:03,120] loop.py:45 INFO TRAIN_STEP step=13444 loss=1.183757 Training: 75%|███████▍ | 13444/18018 [28:59:49<8:13:13, 6.47s/step, loss=1.184, acc=0.719, acc_len=1.74, thru=19.3, I=24.1, T=38.2, wait=3.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:09,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:30:10,956] loop.py:380 INFO TIMING step=13445: step=3.630s data=1.518s compute=3.603s [fwd=2.548s bwd=1.041s opt=0.003s] dispatch=4.196s [2026-10-19 18:30:10,956] loop.py:45 INFO TRAIN_STEP step=13445 loss=1.366917 Training: 75%|███████▍ | 13445/18018 [28:59:57<8:44:21, 6.88s/step, loss=1.367, acc=0.683, acc_len=1.62, thru=17.3, I=19.7, T=35.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:11,313] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:16,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:30:17,079] loop.py:45 INFO TRAIN_STEP step=13446 loss=1.264798 Training: 75%|███████▍ | 13446/18018 [29:00:03<8:26:55, 6.65s/step, loss=1.265, acc=0.700, acc_len=1.67, thru=20.2, I=25.1, T=35.8, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:30:17,075] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13445: forward=2514.5ms backward=1018.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:17,413] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:30:23,355] loop.py:45 INFO TRAIN_STEP step=13447 loss=1.227286 Training: 75%|███████▍ | 13447/18018 [29:00:09<8:18:13, 6.54s/step, loss=1.227, acc=0.713, acc_len=1.73, thru=16.5, I=21.6, T=34.6, wait=2.6s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:33,917] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:30:34,278] loop.py:45 INFO TRAIN_STEP step=13448 loss=1.143572 Training: 75%|███████▍ | 13448/18018 [29:00:20<9:58:16, 7.85s/step, loss=1.144, acc=0.730, acc_len=1.79, thru=15.3, I=14.6, T=34.7, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:34,426] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:41,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:30:43,037] loop.py:45 INFO TRAIN_STEP step=13449 loss=1.324876 Training: 75%|███████▍ | 13449/18018 [29:00:29<10:18:44, 8.13s/step, loss=1.325, acc=0.694, acc_len=1.66, thru=15.1, I=17.2, T=32.7, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:43,443] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:49,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:30:49,730] loop.py:380 INFO TIMING step=13450: step=3.496s data=1.410s compute=3.444s [fwd=2.419s bwd=1.010s opt=0.003s] dispatch=3.190s [2026-10-19 18:30:49,730] loop.py:45 INFO TRAIN_STEP step=13450 loss=1.308154 Training: 75%|███████▍ | 13450/18018 [29:00:35<9:45:54, 7.70s/step, loss=1.308, acc=0.699, acc_len=1.68, thru=17.2, I=21.1, T=36.6, wait=3.2s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:30:49,817] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 18:30:58,003] loop.py:45 INFO TRAIN_STEP step=13451 loss=1.321921 Training: 75%|███████▍ | 13451/18018 [29:00:44<9:58:58, 7.87s/step, loss=1.322, acc=0.694, acc_len=1.66, thru=15.3, I=15.4, T=35.4, wait=4.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:30:57,999] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13450: forward=2547.4ms backward=1020.3ms [2026-10-19 18:31:04,516] loop.py:45 INFO TRAIN_STEP step=13452 loss=1.287229 Training: 75%|███████▍ | 13452/18018 [29:00:50<9:27:51, 7.46s/step, loss=1.287, acc=0.701, acc_len=1.68, thru=18.7, I=18.1, T=36.7, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:31:16,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:31:17,324] loop.py:45 INFO TRAIN_STEP step=13453 loss=1.326619 Training: 75%|███████▍ | 13453/18018 [29:01:03<11:29:48, 9.07s/step, loss=1.327, acc=0.692, acc_len=1.65, thru=12.3, I=10.7, T=32.3, wait=8.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:31:17,403] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:31:21,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:31:23,255] loop.py:45 INFO TRAIN_STEP step=13454 loss=1.255954 Training: 75%|███████▍ | 13454/18018 [29:01:09<10:18:06, 8.13s/step, loss=1.256, acc=0.711, acc_len=1.72, thru=19.7, I=26.0, T=36.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:31:23,597] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:31:28,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:31:29,122] loop.py:380 INFO TIMING step=13455: step=3.519s data=1.472s compute=3.490s [fwd=2.462s bwd=1.014s opt=0.003s] dispatch=2.338s [2026-10-19 18:31:29,123] loop.py:45 INFO TRAIN_STEP step=13455 loss=1.208143 Training: 75%|███████▍ | 13455/18018 [29:01:15<9:26:24, 7.45s/step, loss=1.208, acc=0.711, acc_len=1.71, thru=21.6, I=31.8, T=36.4, wait=2.3s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:31:29,155] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:31:35,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:31:36,460] loop.py:45 INFO TRAIN_STEP step=13456 loss=1.253153 (TrainerActor pid=9310) [2026-10-19 18:31:36,456] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13455: forward=2633.6ms backward=1099.6ms Training: 75%|███████▍ | 13456/18018 [29:01:22<9:23:46, 7.41s/step, loss=1.253, acc=0.706, acc_len=1.70, thru=17.7, I=19.5, T=33.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:31:36,563] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 18:31:42,902] loop.py:45 INFO TRAIN_STEP step=13457 loss=1.273698 Training: 75%|███████▍ | 13457/18018 [29:01:28<9:01:28, 7.12s/step, loss=1.274, acc=0.702, acc_len=1.69, thru=18.9, I=20.1, T=33.8, wait=2.6s, pool=61, epoch=2/2][2026-10-19 18:31:50,957] loop.py:45 INFO TRAIN_STEP step=13458 loss=1.322263 Training: 75%|███████▍ | 13458/18018 [29:01:37<9:22:38, 7.40s/step, loss=1.322, acc=0.701, acc_len=1.68, thru=11.9, I=15.2, T=33.9, wait=4.3s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:31:58,862] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:32:00,024] loop.py:45 INFO TRAIN_STEP step=13459 loss=1.331113 Training: 75%|███████▍ | 13459/18018 [29:01:46<10:00:25, 7.90s/step, loss=1.331, acc=0.695, acc_len=1.67, thru=19.0, I=16.5, T=35.5, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:32:00,370] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:32:08,145] loop.py:380 INFO TIMING step=13460: step=3.633s data=1.543s compute=3.603s [fwd=2.566s bwd=1.024s opt=0.003s] dispatch=4.479s [2026-10-19 18:32:08,146] loop.py:45 INFO TRAIN_STEP step=13460 loss=1.329975 Training: 75%|███████▍ | 13460/18018 [29:01:54<10:05:14, 7.97s/step, loss=1.330, acc=0.696, acc_len=1.66, thru=14.5, I=16.5, T=35.2, wait=4.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:32:12,705] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:32:14,193] loop.py:45 INFO TRAIN_STEP step=13461 loss=1.371205 Training: 75%|███████▍ | 13461/18018 [29:02:00<9:21:23, 7.39s/step, loss=1.371, acc=0.688, acc_len=1.63, thru=20.2, I=21.1, T=35.6, wait=2.4s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:32:14,189] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13460: forward=2549.1ms backward=1002.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:32:14,230] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 18:32:21,049] loop.py:45 INFO TRAIN_STEP step=13462 loss=1.273407 Training: 75%|███████▍ | 13462/18018 [29:02:07<9:09:04, 7.23s/step, loss=1.273, acc=0.702, acc_len=1.68, thru=15.3, I=17.7, T=37.8, wait=3.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:32:27,708] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:32:29,386] loop.py:45 INFO TRAIN_STEP step=13463 loss=1.413477 Training: 75%|███████▍ | 13463/18018 [29:02:15<9:34:09, 7.56s/step, loss=1.413, acc=0.683, acc_len=1.63, thru=18.8, I=18.9, T=32.7, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:32:29,720] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:32:35,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:32:36,635] loop.py:45 INFO TRAIN_STEP step=13464 loss=1.320562 Training: 75%|███████▍ | 13464/18018 [29:02:22<9:26:50, 7.47s/step, loss=1.321, acc=0.697, acc_len=1.67, thru=16.1, I=23.2, T=37.8, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:32:36,996] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:32:41,975] loop.py:380 INFO TIMING step=13465: step=3.263s data=1.349s compute=3.238s [fwd=2.322s bwd=0.902s opt=0.003s] dispatch=2.067s [2026-10-19 18:32:41,975] loop.py:45 INFO TRAIN_STEP step=13465 loss=1.139733 Training: 75%|███████▍ | 13465/18018 [29:02:28<8:38:17, 6.83s/step, loss=1.140, acc=0.727, acc_len=1.77, thru=21.1, I=22.9, T=39.2, wait=2.1s, pool=61, epoch=2/2][2026-10-19 18:32:47,817] loop.py:45 INFO TRAIN_STEP step=13466 loss=1.213481 Training: 75%|███████▍ | 13466/18018 [29:02:33<8:15:44, 6.53s/step, loss=1.213, acc=0.713, acc_len=1.72, thru=22.2, I=23.4, T=38.3, wait=2.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:32:47,812] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13465: forward=2390.0ms backward=913.9ms [2026-10-19 18:32:55,351] loop.py:45 INFO TRAIN_STEP step=13467 loss=1.216804 Training: 75%|███████▍ | 13467/18018 [29:02:41<8:38:17, 6.83s/step, loss=1.217, acc=0.710, acc_len=1.72, thru=17.2, I=17.2, T=34.7, wait=3.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:33:02,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:33:03,744] loop.py:45 INFO TRAIN_STEP step=13468 loss=1.346444 Training: 75%|███████▍ | 13468/18018 [29:02:49<9:13:41, 7.30s/step, loss=1.346, acc=0.693, acc_len=1.66, thru=15.6, I=17.1, T=33.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:33:04,129] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:33:09,977] loop.py:45 INFO TRAIN_STEP step=13469 loss=1.155881 Training: 75%|███████▍ | 13469/18018 [29:02:56<8:49:14, 6.98s/step, loss=1.156, acc=0.725, acc_len=1.77, thru=19.1, I=19.4, T=38.2, wait=2.9s, pool=57, epoch=2/2][2026-10-19 18:33:16,775] loop.py:380 INFO TIMING step=13470: step=3.627s data=1.466s compute=3.600s [fwd=2.613s bwd=0.971s opt=0.003s] dispatch=3.164s [2026-10-19 18:33:16,775] loop.py:45 INFO TRAIN_STEP step=13470 loss=1.211287 Training: 75%|███████▍ | 13470/18018 [29:03:02<8:44:57, 6.93s/step, loss=1.211, acc=0.716, acc_len=1.74, thru=18.4, I=18.2, T=35.3, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:33:23,954] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:33:26,148] loop.py:45 INFO TRAIN_STEP step=13471 loss=1.269701 Training: 75%|███████▍ | 13471/18018 [29:03:12<9:40:30, 7.66s/step, loss=1.270, acc=0.708, acc_len=1.70, thru=14.9, I=18.2, T=29.6, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:33:26,144] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13470: forward=3015.1ms backward=1249.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:33:26,470] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:33:32,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:33:32,668] loop.py:45 INFO TRAIN_STEP step=13472 loss=1.291067 Training: 75%|███████▍ | 13472/18018 [29:03:18<9:14:30, 7.32s/step, loss=1.291, acc=0.701, acc_len=1.68, thru=16.9, I=23.9, T=37.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:33:32,725] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 18:33:41,522] loop.py:45 INFO TRAIN_STEP step=13473 loss=1.253928 Training: 75%|███████▍ | 13473/18018 [29:03:27<9:49:15, 7.78s/step, loss=1.254, acc=0.707, acc_len=1.70, thru=14.4, I=14.9, T=35.9, wait=5.3s, pool=61, epoch=2/2][2026-10-19 18:33:49,468] loop.py:45 INFO TRAIN_STEP step=13474 loss=1.285012 Training: 75%|███████▍ | 13474/18018 [29:03:35<9:52:56, 7.83s/step, loss=1.285, acc=0.704, acc_len=1.69, thru=15.3, I=16.5, T=35.7, wait=4.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:33:55,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:33:56,717] loop.py:380 INFO TIMING step=13475: step=3.656s data=1.565s compute=3.627s [fwd=2.579s bwd=1.034s opt=0.003s] dispatch=3.582s [2026-10-19 18:33:56,717] loop.py:45 INFO TRAIN_STEP step=13475 loss=1.374643 Training: 75%|███████▍ | 13475/18018 [29:03:42<9:39:35, 7.65s/step, loss=1.375, acc=0.687, acc_len=1.64, thru=18.2, I=18.9, T=35.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:33:57,191] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:34:05,334] loop.py:45 INFO TRAIN_STEP step=13476 loss=1.172814 Training: 75%|███████▍ | 13476/18018 [29:03:51<10:01:18, 7.94s/step, loss=1.173, acc=0.722, acc_len=1.76, thru=13.7, I=14.8, T=37.2, wait=5.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:34:05,330] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13475: forward=2363.0ms backward=1032.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:12,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:34:13,344] loop.py:45 INFO TRAIN_STEP step=13477 loss=1.302930 Training: 75%|███████▍ | 13477/18018 [29:03:59<10:02:42, 7.96s/step, loss=1.303, acc=0.702, acc_len=1.68, thru=17.9, I=18.5, T=35.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:13,389] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 18:34:18,806] loop.py:45 INFO TRAIN_STEP step=13478 loss=1.278041 Training: 75%|███████▍ | 13478/18018 [29:04:04<9:05:46, 7.21s/step, loss=1.278, acc=0.704, acc_len=1.69, thru=19.9, I=22.5, T=37.3, wait=2.0s, pool=49, epoch=2/2] [2026-10-19 18:34:25,321] loop.py:45 INFO TRAIN_STEP step=13479 loss=1.271565 Training: 75%|███████▍ | 13479/18018 [29:04:11<8:49:52, 7.00s/step, loss=1.272, acc=0.707, acc_len=1.70, thru=16.8, I=18.8, T=36.6, wait=3.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:33,994] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:34:34,335] loop.py:380 INFO TIMING step=13480: step=3.776s data=1.560s compute=3.747s [fwd=2.675s bwd=1.058s opt=0.003s] dispatch=5.227s [2026-10-19 18:34:34,335] loop.py:45 INFO TRAIN_STEP step=13480 loss=1.231163 Training: 75%|███████▍ | 13480/18018 [29:04:20<9:35:18, 7.61s/step, loss=1.231, acc=0.711, acc_len=1.72, thru=17.1, I=15.6, T=33.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:34,496] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:42,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:34:42,944] loop.py:45 INFO TRAIN_STEP step=13481 loss=1.331259 Training: 75%|███████▍ | 13481/18018 [29:04:29<9:57:55, 7.91s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=15.7, I=17.0, T=35.1, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:34:42,940] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13480: forward=2596.7ms backward=995.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:43,127] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:50,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:34:50,380] loop.py:45 INFO TRAIN_STEP step=13482 loss=1.160898 Training: 75%|███████▍ | 13482/18018 [29:04:36<9:47:07, 7.77s/step, loss=1.161, acc=0.725, acc_len=1.77, thru=16.9, I=21.2, T=34.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:50,673] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:58,799] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:34:58,863] loop.py:45 INFO TRAIN_STEP step=13483 loss=1.167489 Training: 75%|███████▍ | 13483/18018 [29:04:44<10:03:12, 7.98s/step, loss=1.167, acc=0.722, acc_len=1.76, thru=15.0, I=14.6, T=37.2, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:34:59,302] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:35:04,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:35:05,189] loop.py:45 INFO TRAIN_STEP step=13484 loss=1.326916 Training: 75%|███████▍ | 13484/18018 [29:04:51<9:25:35, 7.48s/step, loss=1.327, acc=0.696, acc_len=1.66, thru=20.4, I=21.9, T=36.9, wait=2.9s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:35:05,453] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:35:10,857] loop.py:380 INFO TIMING step=13485: step=3.300s data=1.353s compute=3.261s [fwd=2.338s bwd=0.910s opt=0.003s] dispatch=2.359s [2026-10-19 18:35:10,857] loop.py:45 INFO TRAIN_STEP step=13485 loss=1.339174 Training: 75%|███████▍ | 13485/18018 [29:04:56<8:44:16, 6.94s/step, loss=1.339, acc=0.694, acc_len=1.66, thru=20.2, I=23.9, T=38.8, wait=2.4s, pool=49, epoch=2/2][2026-10-19 18:35:20,118] loop.py:45 INFO TRAIN_STEP step=13486 loss=1.312263 Training: 75%|███████▍ | 13486/18018 [29:05:06<9:36:48, 7.64s/step, loss=1.312, acc=0.693, acc_len=1.66, thru=14.2, I=14.4, T=30.8, wait=5.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:35:20,114] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13485: forward=3129.4ms backward=982.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:35:26,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:35:27,288] loop.py:45 INFO TRAIN_STEP step=13487 loss=1.240051 Training: 75%|███████▍ | 13487/18018 [29:05:13<9:26:06, 7.50s/step, loss=1.240, acc=0.712, acc_len=1.72, thru=17.4, I=18.4, T=36.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:35:27,321] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 18:35:33,642] loop.py:45 INFO TRAIN_STEP step=13488 loss=1.282859 Training: 75%|███████▍ | 13488/18018 [29:05:19<9:00:03, 7.15s/step, loss=1.283, acc=0.700, acc_len=1.67, thru=17.0, I=20.6, T=38.1, wait=3.0s, pool=37, epoch=2/2][2026-10-19 18:35:44,190] loop.py:45 INFO TRAIN_STEP step=13489 loss=1.343097 Training: 75%|███████▍ | 13489/18018 [29:05:30<10:16:47, 8.17s/step, loss=1.343, acc=0.692, acc_len=1.65, thru=15.3, I=13.9, T=31.6, wait=6.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:35:52,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:35:53,621] loop.py:380 INFO TIMING step=13490: step=3.729s data=1.501s compute=3.685s [fwd=2.590s bwd=1.080s opt=0.003s] dispatch=5.695s [2026-10-19 18:35:53,621] loop.py:45 INFO TRAIN_STEP step=13490 loss=1.337172 Training: 75%|███████▍ | 13490/18018 [29:05:39<10:45:12, 8.55s/step, loss=1.337, acc=0.692, acc_len=1.65, thru=14.4, I=14.2, T=34.3, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:35:53,740] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 18:35:59,385] loop.py:45 INFO TRAIN_STEP step=13491 loss=1.178947 Training: 75%|███████▍ | 13491/18018 [29:05:45<9:42:02, 7.71s/step, loss=1.179, acc=0.723, acc_len=1.76, thru=19.0, I=22.6, T=38.4, wait=2.4s, pool=45, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:35:59,380] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13490: forward=2395.6ms backward=890.2ms [2026-10-19 18:36:07,897] loop.py:45 INFO TRAIN_STEP step=13492 loss=1.272693 Training: 75%|███████▍ | 13492/18018 [29:05:53<9:59:57, 7.95s/step, loss=1.273, acc=0.704, acc_len=1.69, thru=15.3, I=15.6, T=35.7, wait=4.9s, pool=53, epoch=2/2][2026-10-19 18:36:16,971] loop.py:45 INFO TRAIN_STEP step=13493 loss=1.341821 Training: 75%|███████▍ | 13493/18018 [29:06:03<10:25:07, 8.29s/step, loss=1.342, acc=0.695, acc_len=1.66, thru=14.5, I=14.6, T=32.6, wait=5.1s, pool=57, epoch=2/2][2026-10-19 18:36:26,343] loop.py:45 INFO TRAIN_STEP step=13494 loss=1.190571 Training: 75%|███████▍ | 13494/18018 [29:06:12<10:49:30, 8.61s/step, loss=1.191, acc=0.725, acc_len=1.76, thru=10.8, I=14.1, T=31.7, wait=5.3s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:36:37,359] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:36:38,413] loop.py:380 INFO TIMING step=13495: step=4.239s data=1.873s compute=4.207s [fwd=2.985s bwd=1.207s opt=0.003s] dispatch=7.823s [2026-10-19 18:36:38,413] loop.py:45 INFO TRAIN_STEP step=13495 loss=1.331407 Training: 75%|███████▍ | 13495/18018 [29:06:24<12:07:31, 9.65s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=13.4, I=15.0, T=30.2, wait=7.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:36:38,867] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:36:43,847] loop.py:45 INFO TRAIN_STEP step=13496 loss=1.227250 Training: 75%|███████▍ | 13496/18018 [29:06:29<10:32:02, 8.39s/step, loss=1.227, acc=0.712, acc_len=1.72, thru=19.8, I=22.8, T=38.9, wait=2.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:36:43,843] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13495: forward=2338.1ms backward=911.9ms [2026-10-19 18:36:52,130] loop.py:45 INFO TRAIN_STEP step=13497 loss=1.191974 Training: 75%|███████▍ | 13497/18018 [29:06:38<10:29:35, 8.36s/step, loss=1.192, acc=0.720, acc_len=1.76, thru=15.7, I=16.6, T=36.0, wait=4.7s, pool=61, epoch=2/2][2026-10-19 18:36:58,025] loop.py:45 INFO TRAIN_STEP step=13498 loss=1.170140 Training: 75%|███████▍ | 13498/18018 [29:06:44<9:33:47, 7.62s/step, loss=1.170, acc=0.721, acc_len=1.75, thru=19.2, I=18.2, T=39.6, wait=2.7s, pool=53, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:05,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:37:06,508] loop.py:45 INFO TRAIN_STEP step=13499 loss=1.324485 Training: 75%|███████▍ | 13499/18018 [29:06:52<9:53:14, 7.88s/step, loss=1.324, acc=0.696, acc_len=1.66, thru=16.5, I=18.2, T=34.4, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:06,761] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:12,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:37:13,489] loop.py:380 INFO TIMING step=13500: step=3.477s data=1.482s compute=3.445s [fwd=2.463s bwd=0.968s opt=0.003s] dispatch=3.496s [2026-10-19 18:37:13,489] loop.py:45 INFO TRAIN_STEP step=13500 loss=1.247571 Training: 75%|███████▍ | 13500/18018 [29:06:59<9:32:54, 7.61s/step, loss=1.248, acc=0.711, acc_len=1.72, thru=18.2, I=22.8, T=36.8, wait=3.5s, pool=65, epoch=2/2][2026-10-19 18:37:13,494] loop.py:419 INFO Saving checkpoint at step 13500... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:17,504] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 18:37:17,682] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0013501 (TrainerActor pid=9310) [2026-10-19 18:37:17,689] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0012751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:19,019] inference_manager.py:408 INFO Pool capacity available after 6.6s, pool_size=33/64 [2026-10-19 18:37:25,623] loop.py:45 INFO TRAIN_STEP step=13501 loss=1.222003 Training: 75%|███████▍ | 13501/18018 [29:07:11<11:15:03, 8.97s/step, loss=1.222, acc=0.712, acc_len=1.72, thru=19.4, I=19.0, T=31.6, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:37:25,618] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13500: forward=3058.8ms backward=946.7ms (TrainerActor pid=9310) [2026-10-19 18:37:17,681] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0013501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 18:37:18,930] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0012751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:25,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:26,052] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:33,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:37:34,067] loop.py:45 INFO TRAIN_STEP step=13502 loss=1.245381 Training: 75%|███████▍ | 13502/18018 [29:07:20<11:02:59, 8.81s/step, loss=1.245, acc=0.709, acc_len=1.71, thru=16.5, I=16.3, T=35.8, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:34,488] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:37:41,430] loop.py:45 INFO TRAIN_STEP step=13503 loss=1.335390 Training: 75%|███████▍ | 13503/18018 [29:07:27<10:30:14, 8.38s/step, loss=1.335, acc=0.693, acc_len=1.65, thru=16.4, I=17.1, T=33.6, wait=3.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:51,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:37:52,150] loop.py:45 INFO TRAIN_STEP step=13504 loss=1.382734 Training: 75%|███████▍ | 13504/18018 [29:07:38<11:23:02, 9.08s/step, loss=1.383, acc=0.689, acc_len=1.64, thru=13.5, I=13.2, T=31.0, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:52,474] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:58,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:37:59,048] loop.py:380 INFO TIMING step=13505: step=3.466s data=1.374s compute=3.427s [fwd=2.437s bwd=0.976s opt=0.003s] dispatch=3.423s [2026-10-19 18:37:59,048] loop.py:45 INFO TRAIN_STEP step=13505 loss=1.197239 Training: 75%|███████▍ | 13505/18018 [29:07:45<10:33:40, 8.42s/step, loss=1.197, acc=0.719, acc_len=1.74, thru=17.5, I=21.2, T=36.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:37:59,297] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:38:05,480] loop.py:45 INFO TRAIN_STEP step=13506 loss=1.217438 Training: 75%|███████▍ | 13506/18018 [29:07:51<9:48:32, 7.83s/step, loss=1.217, acc=0.717, acc_len=1.74, thru=17.3, I=20.1, T=34.8, wait=2.7s, pool=41, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:38:05,475] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13505: forward=2567.2ms backward=1056.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:12,517] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:38:13,303] loop.py:45 INFO TRAIN_STEP step=13507 loss=1.241436 Training: 75%|███████▍ | 13507/18018 [29:07:59<9:48:23, 7.83s/step, loss=1.241, acc=0.708, acc_len=1.71, thru=18.0, I=18.7, T=35.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:13,525] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:19,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:38:20,357] loop.py:45 INFO TRAIN_STEP step=13508 loss=1.249869 Training: 75%|███████▍ | 13508/18018 [29:08:06<9:30:49, 7.59s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=17.9, I=20.9, T=36.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:20,618] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:26,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:38:27,305] loop.py:45 INFO TRAIN_STEP step=13509 loss=1.220365 Training: 75%|███████▍ | 13509/18018 [29:08:13<9:16:08, 7.40s/step, loss=1.220, acc=0.716, acc_len=1.74, thru=18.5, I=22.0, T=37.2, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:27,503] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:34,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:38:35,414] loop.py:380 INFO TIMING step=13510: step=3.797s data=1.655s compute=3.760s [fwd=2.667s bwd=1.079s opt=0.003s] dispatch=4.303s [2026-10-19 18:38:35,414] loop.py:45 INFO TRAIN_STEP step=13510 loss=1.386236 Training: 75%|███████▍ | 13510/18018 [29:08:21<9:31:59, 7.61s/step, loss=1.386, acc=0.682, acc_len=1.62, thru=16.9, I=19.2, T=33.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:35,539] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 18:38:42,007] loop.py:45 INFO TRAIN_STEP step=13511 loss=1.259510 Training: 75%|███████▍ | 13511/18018 [29:08:28<9:08:53, 7.31s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=16.4, I=22.8, T=37.5, wait=3.2s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:38:42,002] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13510: forward=2414.6ms backward=948.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:48,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:38:50,038] loop.py:45 INFO TRAIN_STEP step=13512 loss=1.294888 Training: 75%|███████▍ | 13512/18018 [29:08:36<9:25:01, 7.52s/step, loss=1.295, acc=0.702, acc_len=1.69, thru=16.3, I=18.8, T=34.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:50,473] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:56,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:38:57,266] loop.py:45 INFO TRAIN_STEP step=13513 loss=1.439381 Training: 75%|███████▍ | 13513/18018 [29:08:43<9:18:17, 7.44s/step, loss=1.439, acc=0.677, acc_len=1.60, thru=16.7, I=20.8, T=35.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:38:57,329] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:02,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:39:03,282] loop.py:45 INFO TRAIN_STEP step=13514 loss=1.291172 Training: 75%|███████▌ | 13514/18018 [29:08:49<8:46:11, 7.01s/step, loss=1.291, acc=0.700, acc_len=1.68, thru=20.5, I=23.7, T=38.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:03,322] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:08,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:39:09,418] loop.py:380 INFO TIMING step=13515: step=3.772s data=1.674s compute=3.744s [fwd=2.718s bwd=1.012s opt=0.003s] dispatch=2.354s [2026-10-19 18:39:09,419] loop.py:45 INFO TRAIN_STEP step=13515 loss=1.272813 Training: 75%|███████▌ | 13515/18018 [29:08:55<8:26:20, 6.75s/step, loss=1.273, acc=0.705, acc_len=1.69, thru=20.9, I=24.5, T=33.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:09,878] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:39:15,881] loop.py:45 INFO TRAIN_STEP step=13516 loss=1.291820 Training: 75%|███████▌ | 13516/18018 [29:09:01<8:19:50, 6.66s/step, loss=1.292, acc=0.704, acc_len=1.69, thru=19.0, I=21.0, T=37.8, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:39:15,877] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13515: forward=2405.2ms backward=937.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:15,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:16,308] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:39:22,574] loop.py:45 INFO TRAIN_STEP step=13517 loss=1.188724 Training: 75%|███████▌ | 13517/18018 [29:09:08<8:20:29, 6.67s/step, loss=1.189, acc=0.720, acc_len=1.75, thru=17.0, I=20.4, T=35.5, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:30,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:39:31,926] loop.py:45 INFO TRAIN_STEP step=13518 loss=1.346879 Training: 75%|███████▌ | 13518/18018 [29:09:18<9:20:38, 7.48s/step, loss=1.347, acc=0.693, acc_len=1.66, thru=16.4, I=15.9, T=33.5, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:32,407] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:39:40,169] loop.py:45 INFO TRAIN_STEP step=13519 loss=1.236727 Training: 75%|███████▌ | 13519/18018 [29:09:26<9:37:49, 7.71s/step, loss=1.237, acc=0.709, acc_len=1.71, thru=12.6, I=14.8, T=37.2, wait=4.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:45,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:39:46,888] loop.py:380 INFO TIMING step=13520: step=3.611s data=1.528s compute=3.577s [fwd=2.587s bwd=0.975s opt=0.003s] dispatch=3.097s [2026-10-19 18:39:46,888] loop.py:45 INFO TRAIN_STEP step=13520 loss=1.154331 Training: 75%|███████▌ | 13520/18018 [29:09:32<9:15:31, 7.41s/step, loss=1.154, acc=0.725, acc_len=1.76, thru=19.5, I=20.7, T=35.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:46,957] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:52,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:39:52,699] loop.py:45 INFO TRAIN_STEP step=13521 loss=1.263119 Training: 75%|███████▌ | 13521/18018 [29:09:38<8:39:21, 6.93s/step, loss=1.263, acc=0.704, acc_len=1.69, thru=21.4, I=25.0, T=35.1, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:39:52,694] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13520: forward=2561.7ms backward=1042.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:39:53,130] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:39:58,820] loop.py:45 INFO TRAIN_STEP step=13522 loss=1.352003 Training: 75%|███████▌ | 13522/18018 [29:09:44<8:21:05, 6.69s/step, loss=1.352, acc=0.693, acc_len=1.65, thru=18.8, I=20.6, T=38.3, wait=2.8s, pool=49, epoch=2/2][2026-10-19 18:40:07,468] loop.py:45 INFO TRAIN_STEP step=13523 loss=1.288708 Training: 75%|███████▌ | 13523/18018 [29:09:53<9:05:01, 7.28s/step, loss=1.289, acc=0.705, acc_len=1.70, thru=16.0, I=14.9, T=31.5, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:15,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:40:16,622] loop.py:45 INFO TRAIN_STEP step=13524 loss=1.398025 Training: 75%|███████▌ | 13524/18018 [29:10:02<9:47:11, 7.84s/step, loss=1.398, acc=0.682, acc_len=1.62, thru=16.2, I=17.8, T=34.2, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:16,793] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:40:24,445] loop.py:380 INFO TIMING step=13525: step=3.678s data=1.549s compute=3.613s [fwd=2.585s bwd=1.013s opt=0.003s] dispatch=4.134s [2026-10-19 18:40:24,445] loop.py:45 INFO TRAIN_STEP step=13525 loss=1.290928 Training: 75%|███████▌ | 13525/18018 [29:10:10<9:46:38, 7.83s/step, loss=1.291, acc=0.700, acc_len=1.68, thru=14.7, I=15.7, T=34.8, wait=4.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:31,316] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 18:40:32,415] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13525: forward=2710.3ms backward=1025.8ms [2026-10-19 18:40:32,420] loop.py:45 INFO TRAIN_STEP step=13526 loss=1.260131 Training: 75%|███████▌ | 13526/18018 [29:10:18<9:49:43, 7.88s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=17.2, I=19.2, T=33.9, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:32,826] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:39,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:40:39,567] loop.py:45 INFO TRAIN_STEP step=13527 loss=1.323143 Training: 75%|███████▌ | 13527/18018 [29:10:25<9:33:08, 7.66s/step, loss=1.323, acc=0.692, acc_len=1.65, thru=16.9, I=20.6, T=31.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:40,023] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:45,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:40:46,424] loop.py:45 INFO TRAIN_STEP step=13528 loss=1.293539 Training: 75%|███████▌ | 13528/18018 [29:10:32<9:15:05, 7.42s/step, loss=1.294, acc=0.699, acc_len=1.67, thru=16.5, I=20.8, T=32.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:46,745] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:51,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:40:52,863] loop.py:45 INFO TRAIN_STEP step=13529 loss=1.337951 Training: 75%|███████▌ | 13529/18018 [29:10:38<8:52:58, 7.12s/step, loss=1.338, acc=0.693, acc_len=1.66, thru=19.0, I=24.0, T=34.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:40:52,927] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 18:40:58,146] loop.py:380 INFO TIMING step=13530: step=3.249s data=1.331s compute=3.226s [fwd=2.309s bwd=0.903s opt=0.003s] dispatch=2.026s [2026-10-19 18:40:58,146] loop.py:45 INFO TRAIN_STEP step=13530 loss=1.253289 Training: 75%|███████▌ | 13530/18018 [29:10:44<8:11:32, 6.57s/step, loss=1.253, acc=0.704, acc_len=1.69, thru=20.9, I=24.8, T=39.4, wait=2.0s, pool=61, epoch=2/2][2026-10-19 18:41:04,467] loop.py:45 INFO TRAIN_STEP step=13531 loss=1.271067 Training: 75%|███████▌ | 13531/18018 [29:10:50<8:05:50, 6.50s/step, loss=1.271, acc=0.704, acc_len=1.69, thru=21.2, I=21.9, T=36.2, wait=2.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:41:04,462] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13530: forward=2540.7ms backward=945.9ms [2026-10-19 18:41:10,328] loop.py:45 INFO TRAIN_STEP step=13532 loss=1.300199 Training: 75%|███████▌ | 13532/18018 [29:10:56<7:51:26, 6.31s/step, loss=1.300, acc=0.700, acc_len=1.67, thru=19.7, I=20.1, T=35.5, wait=2.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:18,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:41:19,900] loop.py:45 INFO TRAIN_STEP step=13533 loss=1.211968 Training: 75%|███████▌ | 13533/18018 [29:11:05<9:04:35, 7.29s/step, loss=1.212, acc=0.716, acc_len=1.73, thru=15.6, I=17.4, T=31.4, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:20,162] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:27,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:41:28,153] loop.py:45 INFO TRAIN_STEP step=13534 loss=1.300206 Training: 75%|███████▌ | 13534/18018 [29:11:14<9:26:12, 7.58s/step, loss=1.300, acc=0.701, acc_len=1.68, thru=16.0, I=17.5, T=33.5, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:28,440] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:36,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:41:36,400] loop.py:380 INFO TIMING step=13535: step=3.765s data=1.583s compute=3.731s [fwd=2.609s bwd=1.107s opt=0.003s] dispatch=4.472s [2026-10-19 18:41:36,400] loop.py:45 INFO TRAIN_STEP step=13535 loss=1.118302 Training: 75%|███████▌ | 13535/18018 [29:11:22<9:41:03, 7.78s/step, loss=1.118, acc=0.730, acc_len=1.79, thru=16.0, I=16.5, T=34.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:36,777] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:41:43,176] loop.py:45 INFO TRAIN_STEP step=13536 loss=1.254591 Training: 75%|███████▌ | 13536/18018 [29:11:29<9:18:30, 7.48s/step, loss=1.255, acc=0.708, acc_len=1.71, thru=17.3, I=18.3, T=36.8, wait=3.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:41:43,173] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13535: forward=2461.1ms backward=973.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:49,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:41:51,205] loop.py:45 INFO TRAIN_STEP step=13537 loss=1.257755 Training: 75%|███████▌ | 13537/18018 [29:11:37<9:30:45, 7.64s/step, loss=1.258, acc=0.709, acc_len=1.71, thru=16.9, I=19.6, T=32.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:51,513] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:55,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:41:56,915] loop.py:45 INFO TRAIN_STEP step=13538 loss=1.273457 Training: 75%|███████▌ | 13538/18018 [29:11:42<8:47:21, 7.06s/step, loss=1.273, acc=0.702, acc_len=1.69, thru=20.2, I=26.5, T=35.5, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:41:56,993] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 18:42:02,093] loop.py:45 INFO TRAIN_STEP step=13539 loss=1.203083 Training: 75%|███████▌ | 13539/18018 [29:11:48<8:05:03, 6.50s/step, loss=1.203, acc=0.713, acc_len=1.72, thru=22.3, I=26.6, T=37.8, wait=1.8s, pool=49, epoch=2/2][2026-10-19 18:42:12,565] loop.py:380 INFO TIMING step=13540: step=3.807s data=1.586s compute=3.777s [fwd=2.699s bwd=1.062s opt=0.003s] dispatch=6.654s [2026-10-19 18:42:12,565] loop.py:45 INFO TRAIN_STEP step=13540 loss=1.283699 Training: 75%|███████▌ | 13540/18018 [29:11:58<9:33:54, 7.69s/step, loss=1.284, acc=0.701, acc_len=1.69, thru=12.3, I=12.1, T=33.6, wait=6.7s, pool=41, epoch=2/2][2026-10-19 18:42:22,452] loop.py:45 INFO TRAIN_STEP step=13541 loss=1.227923 Training: 75%|███████▌ | 13541/18018 [29:12:08<10:22:57, 8.35s/step, loss=1.228, acc=0.715, acc_len=1.74, thru=14.4, I=12.8, T=31.1, wait=5.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:42:22,447] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13540: forward=2926.3ms backward=1122.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:42:28,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:42:28,754] loop.py:45 INFO TRAIN_STEP step=13542 loss=1.260715 Training: 75%|███████▌ | 13542/18018 [29:12:14<9:37:01, 7.73s/step, loss=1.261, acc=0.710, acc_len=1.71, thru=18.9, I=18.3, T=37.2, wait=2.9s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:42:29,059] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:42:34,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:42:35,531] loop.py:45 INFO TRAIN_STEP step=13543 loss=1.236051 Training: 75%|███████▌ | 13543/18018 [29:12:21<9:15:26, 7.45s/step, loss=1.236, acc=0.708, acc_len=1.71, thru=19.1, I=21.0, T=36.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:42:36,035] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:42:42,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:42:43,215] loop.py:45 INFO TRAIN_STEP step=13544 loss=1.321418 Training: 75%|███████▌ | 13544/18018 [29:12:29<9:20:37, 7.52s/step, loss=1.321, acc=0.698, acc_len=1.67, thru=16.3, I=19.2, T=38.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:42:43,465] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:42:49,140] loop.py:380 INFO TIMING step=13545: step=3.585s data=1.455s compute=3.553s [fwd=2.461s bwd=1.077s opt=0.003s] dispatch=2.332s [2026-10-19 18:42:49,140] loop.py:45 INFO TRAIN_STEP step=13545 loss=1.278334 Training: 75%|███████▌ | 13545/18018 [29:12:35<8:44:50, 7.04s/step, loss=1.278, acc=0.705, acc_len=1.70, thru=21.2, I=20.1, T=35.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:42:49,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:42:49,587] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:42:55,967] loop.py:45 INFO TRAIN_STEP step=13546 loss=1.335321 Training: 75%|███████▌ | 13546/18018 [29:12:42<8:39:57, 6.98s/step, loss=1.335, acc=0.691, acc_len=1.65, thru=15.8, I=17.4, T=36.4, wait=3.3s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:42:55,962] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13545: forward=2529.7ms backward=949.4ms [2026-10-19 18:43:07,014] loop.py:45 INFO TRAIN_STEP step=13547 loss=1.296985 Training: 75%|███████▌ | 13547/18018 [29:12:53<10:10:52, 8.20s/step, loss=1.297, acc=0.705, acc_len=1.69, thru=13.6, I=13.0, T=32.2, wait=7.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:12,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:43:13,986] loop.py:45 INFO TRAIN_STEP step=13548 loss=1.287513 Training: 75%|███████▌ | 13548/18018 [29:13:00<9:43:20, 7.83s/step, loss=1.288, acc=0.703, acc_len=1.68, thru=17.3, I=19.6, T=31.6, wait=2.9s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:14,316] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:43:20,558] loop.py:45 INFO TRAIN_STEP step=13549 loss=1.185096 Training: 75%|███████▌ | 13549/18018 [29:13:06<9:15:07, 7.45s/step, loss=1.185, acc=0.721, acc_len=1.75, thru=17.7, I=19.5, T=39.4, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:26,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:43:27,925] loop.py:380 INFO TIMING step=13550: step=3.625s data=1.485s compute=3.579s [fwd=2.555s bwd=1.009s opt=0.003s] dispatch=3.734s [2026-10-19 18:43:27,926] loop.py:45 INFO TRAIN_STEP step=13550 loss=1.451027 Training: 75%|███████▌ | 13550/18018 [29:13:14<9:13:05, 7.43s/step, loss=1.451, acc=0.679, acc_len=1.61, thru=19.0, I=20.3, T=35.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:28,107] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:34,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:43:34,596] loop.py:45 INFO TRAIN_STEP step=13551 loss=1.286049 Training: 75%|███████▌ | 13551/18018 [29:13:20<8:55:59, 7.20s/step, loss=1.286, acc=0.703, acc_len=1.69, thru=18.9, I=22.8, T=35.5, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:43:34,592] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13550: forward=2568.2ms backward=992.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:35,069] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:40,185] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:43:41,761] loop.py:45 INFO TRAIN_STEP step=13552 loss=1.179031 Training: 75%|███████▌ | 13552/18018 [29:13:27<8:55:10, 7.19s/step, loss=1.179, acc=0.724, acc_len=1.76, thru=18.4, I=22.3, T=34.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:42,196] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:49,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:43:50,282] loop.py:45 INFO TRAIN_STEP step=13553 loss=1.326594 Training: 75%|███████▌ | 13553/18018 [29:13:36<9:24:45, 7.59s/step, loss=1.327, acc=0.695, acc_len=1.66, thru=15.8, I=17.2, T=34.5, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:43:50,633] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:43:55,685] loop.py:45 INFO TRAIN_STEP step=13554 loss=1.313030 Training: 75%|███████▌ | 13554/18018 [29:13:41<8:35:50, 6.93s/step, loss=1.313, acc=0.698, acc_len=1.67, thru=19.6, I=20.4, T=38.8, wait=2.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:04,767] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:44:05,485] loop.py:380 INFO TIMING step=13555: step=3.791s data=1.585s compute=3.758s [fwd=2.663s bwd=1.080s opt=0.003s] dispatch=6.000s [2026-10-19 18:44:05,486] loop.py:45 INFO TRAIN_STEP step=13555 loss=1.324026 Training: 75%|███████▌ | 13555/18018 [29:13:51<9:39:43, 7.79s/step, loss=1.324, acc=0.700, acc_len=1.68, thru=15.0, I=14.4, T=33.8, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:05,774] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:12,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:44:12,974] loop.py:45 INFO TRAIN_STEP step=13556 loss=1.245271 Training: 75%|███████▌ | 13556/18018 [29:13:59<9:32:46, 7.70s/step, loss=1.245, acc=0.712, acc_len=1.72, thru=17.2, I=20.4, T=36.3, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:44:12,970] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13555: forward=2522.3ms backward=958.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:13,355] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:18,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:44:19,248] loop.py:45 INFO TRAIN_STEP step=13557 loss=1.268638 Training: 75%|███████▌ | 13557/18018 [29:14:05<9:00:47, 7.27s/step, loss=1.269, acc=0.703, acc_len=1.69, thru=19.9, I=23.8, T=36.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:19,688] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:25,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:44:26,576] loop.py:45 INFO TRAIN_STEP step=13558 loss=1.118438 Training: 75%|███████▌ | 13558/18018 [29:14:12<9:01:51, 7.29s/step, loss=1.118, acc=0.731, acc_len=1.79, thru=16.8, I=19.5, T=35.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:26,951] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:44:34,594] loop.py:45 INFO TRAIN_STEP step=13559 loss=1.324280 Training: 75%|███████▌ | 13559/18018 [29:14:20<9:18:00, 7.51s/step, loss=1.324, acc=0.696, acc_len=1.66, thru=15.4, I=16.1, T=33.6, wait=4.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:39,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:44:40,413] loop.py:380 INFO TIMING step=13560: step=3.403s data=1.424s compute=3.362s [fwd=2.421s bwd=0.927s opt=0.003s] dispatch=2.406s [2026-10-19 18:44:40,413] loop.py:45 INFO TRAIN_STEP step=13560 loss=1.233704 Training: 75%|███████▌ | 13560/18018 [29:14:26<8:40:11, 7.00s/step, loss=1.234, acc=0.710, acc_len=1.72, thru=20.7, I=22.6, T=37.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:40,828] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:48,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:44:49,251] loop.py:45 INFO TRAIN_STEP step=13561 loss=1.255263 Training: 75%|███████▌ | 13561/18018 [29:14:35<9:21:02, 7.55s/step, loss=1.255, acc=0.708, acc_len=1.70, thru=13.3, I=16.3, T=35.4, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:44:49,246] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13560: forward=2556.0ms backward=1017.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:44:49,344] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 18:44:56,076] loop.py:45 INFO TRAIN_STEP step=13562 loss=1.167510 Training: 75%|███████▌ | 13562/18018 [29:14:42<9:04:40, 7.33s/step, loss=1.168, acc=0.726, acc_len=1.78, thru=16.4, I=21.9, T=37.8, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:05,706] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:45:07,094] loop.py:45 INFO TRAIN_STEP step=13563 loss=1.366657 Training: 75%|███████▌ | 13563/18018 [29:14:53<10:26:38, 8.44s/step, loss=1.367, acc=0.687, acc_len=1.63, thru=14.0, I=13.8, T=30.7, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:07,219] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 18:45:12,566] loop.py:45 INFO TRAIN_STEP step=13564 loss=1.263394 Training: 75%|███████▌ | 13564/18018 [29:14:58<9:20:23, 7.55s/step, loss=1.263, acc=0.710, acc_len=1.70, thru=18.7, I=21.2, T=38.1, wait=2.1s, pool=45, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:21,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:45:22,058] loop.py:380 INFO TIMING step=13565: step=3.589s data=1.551s compute=3.552s [fwd=2.541s bwd=0.997s opt=0.003s] dispatch=5.895s [2026-10-19 18:45:22,058] loop.py:45 INFO TRAIN_STEP step=13565 loss=1.162068 Training: 75%|███████▌ | 13565/18018 [29:15:08<10:03:33, 8.13s/step, loss=1.162, acc=0.726, acc_len=1.76, thru=16.3, I=16.5, T=35.7, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:22,092] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:27,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:45:28,270] loop.py:45 INFO TRAIN_STEP step=13566 loss=1.167554 Training: 75%|███████▌ | 13566/18018 [29:15:14<9:20:35, 7.56s/step, loss=1.168, acc=0.722, acc_len=1.76, thru=19.9, I=20.3, T=38.3, wait=2.9s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:45:28,266] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13565: forward=2364.5ms backward=916.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:28,490] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:34,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:45:34,710] loop.py:45 INFO TRAIN_STEP step=13567 loss=1.394182 Training: 75%|███████▌ | 13567/18018 [29:15:20<8:55:41, 7.22s/step, loss=1.394, acc=0.683, acc_len=1.62, thru=20.2, I=22.0, T=39.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:35,071] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:40,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:45:40,685] loop.py:45 INFO TRAIN_STEP step=13568 loss=1.213187 Training: 75%|███████▌ | 13568/18018 [29:15:26<8:27:52, 6.85s/step, loss=1.213, acc=0.712, acc_len=1.72, thru=20.9, I=23.8, T=39.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:40,800] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:46,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:45:46,857] loop.py:45 INFO TRAIN_STEP step=13569 loss=1.268863 Training: 75%|███████▌ | 13569/18018 [29:15:32<8:12:44, 6.65s/step, loss=1.269, acc=0.705, acc_len=1.70, thru=21.2, I=23.0, T=38.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:47,264] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:52,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:45:53,660] loop.py:380 INFO TIMING step=13570: step=3.405s data=1.391s compute=3.378s [fwd=2.419s bwd=0.946s opt=0.003s] dispatch=3.387s [2026-10-19 18:45:53,660] loop.py:45 INFO TRAIN_STEP step=13570 loss=1.265930 Training: 75%|███████▌ | 13570/18018 [29:15:39<8:16:05, 6.69s/step, loss=1.266, acc=0.705, acc_len=1.70, thru=19.0, I=20.7, T=37.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:45:53,783] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 18:46:01,844] loop.py:45 INFO TRAIN_STEP step=13571 loss=1.348474 Training: 75%|███████▌ | 13571/18018 [29:15:47<8:49:10, 7.14s/step, loss=1.348, acc=0.693, acc_len=1.65, thru=14.1, I=15.5, T=36.4, wait=4.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:46:01,840] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13570: forward=2419.1ms backward=1041.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:08,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:46:09,396] loop.py:45 INFO TRAIN_STEP step=13572 loss=1.283849 Training: 75%|███████▌ | 13572/18018 [29:15:55<8:58:14, 7.26s/step, loss=1.284, acc=0.703, acc_len=1.68, thru=17.9, I=19.5, T=34.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:09,543] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:14,893] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:46:15,787] loop.py:45 INFO TRAIN_STEP step=13573 loss=1.220060 Training: 75%|███████▌ | 13573/18018 [29:16:01<8:38:41, 7.00s/step, loss=1.220, acc=0.716, acc_len=1.73, thru=19.4, I=23.2, T=35.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:15,908] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:21,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:46:22,323] loop.py:45 INFO TRAIN_STEP step=13574 loss=1.332158 Training: 75%|███████▌ | 13574/18018 [29:16:08<8:28:15, 6.86s/step, loss=1.332, acc=0.692, acc_len=1.66, thru=18.6, I=21.7, T=35.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:22,432] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 18:46:30,796] loop.py:380 INFO TIMING step=13575: step=3.558s data=1.423s compute=3.527s [fwd=2.531s bwd=0.982s opt=0.003s] dispatch=4.904s [2026-10-19 18:46:30,796] loop.py:45 INFO TRAIN_STEP step=13575 loss=1.280298 Training: 75%|███████▌ | 13575/18018 [29:16:16<9:03:52, 7.34s/step, loss=1.280, acc=0.702, acc_len=1.69, thru=14.8, I=13.9, T=36.0, wait=4.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:36,461] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:46:37,733] loop.py:45 INFO TRAIN_STEP step=13576 loss=1.328996 Training: 75%|███████▌ | 13576/18018 [29:16:23<8:54:45, 7.22s/step, loss=1.329, acc=0.696, acc_len=1.67, thru=19.3, I=20.2, T=31.9, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:46:37,729] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13575: forward=2846.3ms backward=1117.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:37,972] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:44,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:46:44,850] loop.py:45 INFO TRAIN_STEP step=13577 loss=1.230369 Training: 75%|███████▌ | 13577/18018 [29:16:30<8:52:14, 7.19s/step, loss=1.230, acc=0.709, acc_len=1.71, thru=17.0, I=20.2, T=36.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:45,335] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:51,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:46:52,225] loop.py:45 INFO TRAIN_STEP step=13578 loss=1.322287 Training: 75%|███████▌ | 13578/18018 [29:16:38<8:56:14, 7.25s/step, loss=1.322, acc=0.696, acc_len=1.67, thru=18.7, I=21.9, T=30.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:52,639] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:46:59,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:47:00,644] loop.py:45 INFO TRAIN_STEP step=13579 loss=1.286538 Training: 75%|███████▌ | 13579/18018 [29:16:46<9:22:06, 7.60s/step, loss=1.287, acc=0.705, acc_len=1.69, thru=14.6, I=19.7, T=35.2, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:47:01,069] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 18:47:07,905] loop.py:380 INFO TIMING step=13580: step=3.649s data=1.602s compute=3.616s [fwd=2.587s bwd=1.014s opt=0.003s] dispatch=3.604s [2026-10-19 18:47:07,906] loop.py:45 INFO TRAIN_STEP step=13580 loss=1.287015 Training: 75%|███████▌ | 13580/18018 [29:16:53<9:14:31, 7.50s/step, loss=1.287, acc=0.698, acc_len=1.67, thru=13.7, I=17.3, T=35.1, wait=3.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:47:14,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:47:14,836] loop.py:45 INFO TRAIN_STEP step=13581 loss=1.317319 Training: 75%|███████▌ | 13581/18018 [29:17:00<9:01:49, 7.33s/step, loss=1.317, acc=0.693, acc_len=1.66, thru=19.3, I=17.6, T=36.0, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:47:14,832] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13580: forward=2487.5ms backward=1026.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:47:15,127] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:47:22,183] loop.py:45 INFO TRAIN_STEP step=13582 loss=1.209841 Training: 75%|███████▌ | 13582/18018 [29:17:08<9:02:07, 7.33s/step, loss=1.210, acc=0.716, acc_len=1.74, thru=16.9, I=16.3, T=34.1, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:47:28,184] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:47:28,365] loop.py:45 INFO TRAIN_STEP step=13583 loss=1.208263 Training: 75%|███████▌ | 13583/18018 [29:17:14<8:36:32, 6.99s/step, loss=1.208, acc=0.718, acc_len=1.75, thru=19.6, I=19.6, T=36.1, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:47:28,686] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:47:35,766] loop.py:45 INFO TRAIN_STEP step=13584 loss=1.287683 Training: 75%|███████▌ | 13584/18018 [29:17:21<8:45:32, 7.11s/step, loss=1.288, acc=0.702, acc_len=1.69, thru=12.7, I=13.9, T=37.2, wait=3.9s, pool=29, epoch=2/2][2026-10-19 18:47:44,323] loop.py:380 INFO TIMING step=13585: step=3.527s data=1.387s compute=3.491s [fwd=2.496s bwd=0.980s opt=0.003s] dispatch=5.022s [2026-10-19 18:47:44,324] loop.py:45 INFO TRAIN_STEP step=13585 loss=1.304502 Training: 75%|███████▌ | 13585/18018 [29:17:30<9:17:28, 7.55s/step, loss=1.305, acc=0.701, acc_len=1.69, thru=17.0, I=14.5, T=36.3, wait=5.0s, pool=53, epoch=2/2][2026-10-19 18:47:50,496] loop.py:45 INFO TRAIN_STEP step=13586 loss=1.215320 Training: 75%|███████▌ | 13586/18018 [29:17:36<8:46:56, 7.13s/step, loss=1.215, acc=0.713, acc_len=1.73, thru=18.1, I=17.8, T=35.8, wait=2.6s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:47:50,492] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13585: forward=2514.1ms backward=1002.8ms [2026-10-19 18:47:58,342] loop.py:45 INFO TRAIN_STEP step=13587 loss=1.261501 Training: 75%|███████▌ | 13587/18018 [29:17:44<9:02:39, 7.35s/step, loss=1.262, acc=0.708, acc_len=1.71, thru=13.7, I=17.2, T=34.5, wait=4.1s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:48:06,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:48:06,507] loop.py:45 INFO TRAIN_STEP step=13588 loss=1.273152 Training: 75%|███████▌ | 13588/18018 [29:17:52<9:20:32, 7.59s/step, loss=1.273, acc=0.705, acc_len=1.70, thru=20.7, I=16.5, T=35.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:48:06,596] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 18:48:12,338] loop.py:45 INFO TRAIN_STEP step=13589 loss=1.258373 Training: 75%|███████▌ | 13589/18018 [29:17:58<8:41:28, 7.06s/step, loss=1.258, acc=0.709, acc_len=1.71, thru=18.8, I=20.1, T=38.2, wait=2.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:48:22,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:48:23,480] loop.py:380 INFO TIMING step=13590: step=4.346s data=2.034s compute=4.117s [fwd=2.811s bwd=1.290s opt=0.003s] dispatch=6.786s [2026-10-19 18:48:23,480] loop.py:45 INFO TRAIN_STEP step=13590 loss=1.416017 Training: 75%|███████▌ | 13590/18018 [29:18:09<10:11:38, 8.29s/step, loss=1.416, acc=0.682, acc_len=1.62, thru=14.0, I=14.1, T=29.5, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:48:23,980] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:48:28,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:48:30,124] loop.py:45 INFO TRAIN_STEP step=13591 loss=1.253352 Training: 75%|███████▌ | 13591/18018 [29:18:16<9:35:07, 7.79s/step, loss=1.253, acc=0.708, acc_len=1.71, thru=17.6, I=22.3, T=31.8, wait=2.6s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 18:48:30,119] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13590: forward=2828.2ms backward=1146.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:48:30,277] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:48:36,812] loop.py:45 INFO TRAIN_STEP step=13592 loss=1.176527 Training: 75%|███████▌ | 13592/18018 [29:18:22<9:10:29, 7.46s/step, loss=1.177, acc=0.723, acc_len=1.76, thru=16.0, I=20.1, T=33.7, wait=2.9s, pool=49, epoch=2/2][2026-10-19 18:48:45,482] loop.py:45 INFO TRAIN_STEP step=13593 loss=1.405117 Training: 75%|███████▌ | 13593/18018 [29:18:31<9:37:02, 7.82s/step, loss=1.405, acc=0.684, acc_len=1.62, thru=16.3, I=15.4, T=37.0, wait=5.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:48:52,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:48:53,988] loop.py:45 INFO TRAIN_STEP step=13594 loss=1.290266 Training: 75%|███████▌ | 13594/18018 [29:18:40<9:52:03, 8.03s/step, loss=1.290, acc=0.699, acc_len=1.68, thru=16.1, I=18.0, T=33.8, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:48:54,365] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:49:00,712] loop.py:380 INFO TIMING step=13595: step=3.618s data=1.394s compute=3.582s [fwd=2.424s bwd=1.145s opt=0.003s] dispatch=3.095s [2026-10-19 18:49:00,712] loop.py:45 INFO TRAIN_STEP step=13595 loss=1.243922 Training: 75%|███████▌ | 13595/18018 [29:18:46<9:23:01, 7.64s/step, loss=1.244, acc=0.708, acc_len=1.71, thru=15.9, I=21.4, T=35.4, wait=3.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:06,694] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:49:08,574] loop.py:45 INFO TRAIN_STEP step=13596 loss=1.196248 Training: 75%|███████▌ | 13596/18018 [29:18:54<9:27:49, 7.70s/step, loss=1.196, acc=0.718, acc_len=1.75, thru=16.7, I=19.1, T=33.9, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:49:08,569] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13595: forward=2693.8ms backward=1045.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:08,722] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:14,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:49:14,712] loop.py:45 INFO TRAIN_STEP step=13597 loss=1.196312 Training: 75%|███████▌ | 13597/18018 [29:19:00<8:53:03, 7.23s/step, loss=1.196, acc=0.717, acc_len=1.74, thru=20.0, I=24.1, T=36.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:14,849] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:20,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:49:21,521] loop.py:45 INFO TRAIN_STEP step=13598 loss=1.231575 Training: 75%|███████▌ | 13598/18018 [29:19:07<8:43:34, 7.11s/step, loss=1.232, acc=0.714, acc_len=1.73, thru=19.2, I=20.8, T=37.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:21,757] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:49:29,930] loop.py:45 INFO TRAIN_STEP step=13599 loss=1.182715 Training: 75%|███████▌ | 13599/18018 [29:19:16<9:12:09, 7.50s/step, loss=1.183, acc=0.720, acc_len=1.75, thru=14.5, I=15.1, T=29.4, wait=4.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:36,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:49:37,016] loop.py:380 INFO TIMING step=13600: step=3.669s data=1.556s compute=3.627s [fwd=2.605s bwd=1.007s opt=0.003s] dispatch=3.409s [2026-10-19 18:49:37,016] loop.py:45 INFO TRAIN_STEP step=13600 loss=1.273235 Training: 75%|███████▌ | 13600/18018 [29:19:23<9:02:58, 7.37s/step, loss=1.273, acc=0.711, acc_len=1.72, thru=18.1, I=18.2, T=34.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:37,116] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:44,638] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:49:45,707] loop.py:45 INFO TRAIN_STEP step=13601 loss=1.347170 Training: 75%|███████▌ | 13601/18018 [29:19:31<9:31:55, 7.77s/step, loss=1.347, acc=0.689, acc_len=1.64, thru=15.3, I=18.2, T=32.1, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:49:45,702] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13600: forward=2835.4ms backward=1108.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:46,146] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:51,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:49:52,354] loop.py:45 INFO TRAIN_STEP step=13602 loss=1.240286 Training: 75%|███████▌ | 13602/18018 [29:19:38<9:07:05, 7.43s/step, loss=1.240, acc=0.712, acc_len=1.72, thru=18.5, I=24.4, T=30.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:52,514] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:57,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:49:58,264] loop.py:45 INFO TRAIN_STEP step=13603 loss=1.164244 Training: 75%|███████▌ | 13603/18018 [29:19:44<8:33:17, 6.98s/step, loss=1.164, acc=0.723, acc_len=1.77, thru=19.5, I=25.2, T=38.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:49:58,393] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 18:50:04,492] loop.py:45 INFO TRAIN_STEP step=13604 loss=1.135277 Training: 76%|███████▌ | 13604/18018 [29:19:50<8:16:39, 6.75s/step, loss=1.135, acc=0.728, acc_len=1.77, thru=19.7, I=24.1, T=37.2, wait=2.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:50:12,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:50:12,539] loop.py:380 INFO TIMING step=13605: step=3.491s data=1.523s compute=3.454s [fwd=2.460s bwd=0.980s opt=0.003s] dispatch=4.548s [2026-10-19 18:50:12,539] loop.py:45 INFO TRAIN_STEP step=13605 loss=1.367179 Training: 76%|███████▌ | 13605/18018 [29:19:58<8:45:09, 7.14s/step, loss=1.367, acc=0.687, acc_len=1.64, thru=17.4, I=16.7, T=36.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:50:13,009] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:50:18,589] loop.py:45 INFO TRAIN_STEP step=13606 loss=1.376858 Training: 76%|███████▌ | 13606/18018 [29:20:04<8:21:06, 6.81s/step, loss=1.377, acc=0.686, acc_len=1.63, thru=20.3, I=23.1, T=37.6, wait=2.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:50:18,584] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13605: forward=2435.9ms backward=917.2ms [2026-10-19 18:50:28,835] loop.py:45 INFO TRAIN_STEP step=13607 loss=1.182779 Training: 76%|███████▌ | 13607/18018 [29:20:14<9:36:31, 7.84s/step, loss=1.183, acc=0.718, acc_len=1.75, thru=13.3, I=11.7, T=33.4, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:50:28,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:50:29,285] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:50:33,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:50:35,794] loop.py:45 INFO TRAIN_STEP step=13608 loss=1.347711 Training: 76%|███████▌ | 13608/18018 [29:20:21<9:16:57, 7.58s/step, loss=1.348, acc=0.696, acc_len=1.66, thru=18.9, I=20.1, T=30.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:50:35,878] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=33/64 [2026-10-19 18:50:43,130] loop.py:45 INFO TRAIN_STEP step=13609 loss=1.237524 Training: 76%|███████▌ | 13609/18018 [29:20:29<9:11:28, 7.50s/step, loss=1.238, acc=0.709, acc_len=1.71, thru=13.0, I=17.2, T=37.3, wait=3.9s, pool=37, epoch=2/2][2026-10-19 18:50:53,386] loop.py:380 INFO TIMING step=13610: step=3.735s data=1.550s compute=3.701s [fwd=2.668s bwd=1.020s opt=0.003s] dispatch=6.513s [2026-10-19 18:50:53,386] loop.py:45 INFO TRAIN_STEP step=13610 loss=1.245043 Training: 76%|███████▌ | 13610/18018 [29:20:39<10:11:59, 8.33s/step, loss=1.245, acc=0.708, acc_len=1.71, thru=14.4, I=13.6, T=34.3, wait=6.5s, pool=53, epoch=2/2][2026-10-19 18:51:01,600] loop.py:45 INFO TRAIN_STEP step=13611 loss=1.273042 Training: 76%|███████▌ | 13611/18018 [29:20:47<10:09:21, 8.30s/step, loss=1.273, acc=0.702, acc_len=1.69, thru=15.0, I=15.1, T=34.7, wait=4.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:51:01,596] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13610: forward=2526.2ms backward=1069.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:07,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:51:09,210] loop.py:45 INFO TRAIN_STEP step=13612 loss=1.320050 Training: 76%|███████▌ | 13612/18018 [29:20:55<9:54:02, 8.09s/step, loss=1.320, acc=0.694, acc_len=1.66, thru=18.0, I=19.1, T=35.2, wait=4.0s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:09,464] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:14,927] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:51:15,227] loop.py:45 INFO TRAIN_STEP step=13613 loss=1.206697 Training: 76%|███████▌ | 13613/18018 [29:21:01<9:08:15, 7.47s/step, loss=1.207, acc=0.713, acc_len=1.72, thru=21.2, I=25.7, T=37.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:15,428] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:51:22,788] loop.py:45 INFO TRAIN_STEP step=13614 loss=1.260789 Training: 76%|███████▌ | 13614/18018 [29:21:08<9:10:13, 7.50s/step, loss=1.261, acc=0.707, acc_len=1.71, thru=17.3, I=18.6, T=32.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:22,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:23,294] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:28,669] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:51:30,078] loop.py:380 INFO TIMING step=13615: step=4.086s data=1.912s compute=4.049s [fwd=3.119s bwd=0.914s opt=0.003s] dispatch=3.196s [2026-10-19 18:51:30,079] loop.py:45 INFO TRAIN_STEP step=13615 loss=1.305734 Training: 76%|███████▌ | 13615/18018 [29:21:16<9:05:32, 7.43s/step, loss=1.306, acc=0.700, acc_len=1.68, thru=17.7, I=18.7, T=31.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:30,195] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:36,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:51:36,513] loop.py:45 INFO TRAIN_STEP step=13616 loss=1.271629 Training: 76%|███████▌ | 13616/18018 [29:21:22<8:43:25, 7.13s/step, loss=1.272, acc=0.705, acc_len=1.69, thru=19.0, I=21.0, T=37.5, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:51:36,508] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13615: forward=2417.0ms backward=946.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:36,923] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:51:43,743] loop.py:45 INFO TRAIN_STEP step=13617 loss=1.325591 Training: 76%|███████▌ | 13617/18018 [29:21:29<8:45:21, 7.16s/step, loss=1.326, acc=0.695, acc_len=1.66, thru=18.8, I=17.7, T=36.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:43,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:44,245] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:49,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:51:49,948] loop.py:45 INFO TRAIN_STEP step=13618 loss=1.272535 Training: 76%|███████▌ | 13618/18018 [29:21:36<8:24:14, 6.88s/step, loss=1.273, acc=0.705, acc_len=1.70, thru=20.1, I=22.1, T=37.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:51:50,120] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:51:56,378] loop.py:45 INFO TRAIN_STEP step=13619 loss=1.212987 Training: 76%|███████▌ | 13619/18018 [29:21:42<8:14:18, 6.74s/step, loss=1.213, acc=0.712, acc_len=1.72, thru=18.2, I=24.4, T=36.6, wait=2.9s, pool=49, epoch=2/2][2026-10-19 18:52:04,348] loop.py:380 INFO TIMING step=13620: step=3.605s data=1.486s compute=3.581s [fwd=2.543s bwd=1.025s opt=0.003s] dispatch=4.357s [2026-10-19 18:52:04,348] loop.py:45 INFO TRAIN_STEP step=13620 loss=1.255311 Training: 76%|███████▌ | 13620/18018 [29:21:50<8:41:11, 7.11s/step, loss=1.255, acc=0.706, acc_len=1.70, thru=15.5, I=15.6, T=35.5, wait=4.4s, pool=57, epoch=2/2][2026-10-19 18:52:11,425] loop.py:45 INFO TRAIN_STEP step=13621 loss=1.179919 Training: 76%|███████▌ | 13621/18018 [29:21:57<8:40:18, 7.10s/step, loss=1.180, acc=0.725, acc_len=1.77, thru=17.3, I=17.6, T=38.7, wait=3.8s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:52:11,421] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13620: forward=2356.1ms backward=913.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:52:19,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:52:19,623] loop.py:45 INFO TRAIN_STEP step=13622 loss=1.278346 Training: 76%|███████▌ | 13622/18018 [29:22:05<9:04:20, 7.43s/step, loss=1.278, acc=0.703, acc_len=1.69, thru=17.5, I=17.4, T=36.0, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:52:20,061] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:52:27,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:52:28,116] loop.py:45 INFO TRAIN_STEP step=13623 loss=1.200475 Training: 76%|███████▌ | 13623/18018 [29:22:14<9:27:38, 7.75s/step, loss=1.200, acc=0.722, acc_len=1.75, thru=15.5, I=17.5, T=35.4, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:52:28,592] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:52:34,857] loop.py:45 INFO TRAIN_STEP step=13624 loss=1.313928 Training: 76%|███████▌ | 13624/18018 [29:22:20<9:05:23, 7.45s/step, loss=1.314, acc=0.695, acc_len=1.66, thru=15.0, I=20.2, T=36.6, wait=3.2s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:52:41,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:52:42,404] loop.py:380 INFO TIMING step=13625: step=3.571s data=1.367s compute=3.543s [fwd=2.562s bwd=0.966s opt=0.003s] dispatch=3.964s [2026-10-19 18:52:42,405] loop.py:45 INFO TRAIN_STEP step=13625 loss=1.254432 Training: 76%|███████▌ | 13625/18018 [29:22:28<9:07:25, 7.48s/step, loss=1.254, acc=0.707, acc_len=1.71, thru=19.5, I=18.3, T=35.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:52:42,638] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:52:49,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:52:50,470] loop.py:45 INFO TRAIN_STEP step=13626 loss=1.170603 Training: 76%|███████▌ | 13626/18018 [29:22:36<9:20:13, 7.65s/step, loss=1.171, acc=0.723, acc_len=1.75, thru=15.3, I=17.6, T=36.6, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:52:50,466] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13625: forward=2481.9ms backward=971.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:52:50,579] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 18:52:56,484] loop.py:45 INFO TRAIN_STEP step=13627 loss=1.254911 Training: 76%|███████▌ | 13627/18018 [29:22:42<8:44:05, 7.16s/step, loss=1.255, acc=0.705, acc_len=1.69, thru=19.6, I=20.9, T=36.8, wait=2.5s, pool=57, epoch=2/2][2026-10-19 18:53:03,145] loop.py:45 INFO TRAIN_STEP step=13628 loss=1.265518 Training: 76%|███████▌ | 13628/18018 [29:22:49<8:32:58, 7.01s/step, loss=1.266, acc=0.710, acc_len=1.72, thru=17.1, I=17.7, T=36.0, wait=3.1s, pool=41, epoch=2/2][2026-10-19 18:53:09,891] loop.py:45 INFO TRAIN_STEP step=13629 loss=1.348578 Training: 76%|███████▌ | 13629/18018 [29:22:55<8:27:00, 6.93s/step, loss=1.349, acc=0.690, acc_len=1.64, thru=16.3, I=18.1, T=39.0, wait=3.5s, pool=37, epoch=2/2][2026-10-19 18:53:18,715] loop.py:380 INFO TIMING step=13630: step=3.493s data=1.377s compute=3.465s [fwd=2.467s bwd=0.984s opt=0.003s] dispatch=5.324s [2026-10-19 18:53:18,715] loop.py:45 INFO TRAIN_STEP step=13630 loss=1.270291 Training: 76%|███████▌ | 13630/18018 [29:23:04<9:08:29, 7.50s/step, loss=1.270, acc=0.708, acc_len=1.71, thru=17.6, I=15.2, T=36.6, wait=5.3s, pool=65, epoch=2/2][2026-10-19 18:53:27,336] loop.py:45 INFO TRAIN_STEP step=13631 loss=1.242310 Training: 76%|███████▌ | 13631/18018 [29:23:13<9:32:55, 7.84s/step, loss=1.242, acc=0.712, acc_len=1.72, thru=12.4, I=15.9, T=30.6, wait=4.4s, pool=25, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:53:27,332] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13630: forward=3093.8ms backward=1050.2ms [2026-10-19 18:53:37,571] loop.py:45 INFO TRAIN_STEP step=13632 loss=1.270739 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:53:37,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 76%|███████▌ | 13632/18018 [29:23:23<10:25:24, 8.56s/step, loss=1.271, acc=0.706, acc_len=1.70, thru=16.4, I=14.8, T=36.0, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:53:37,991] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:53:44,366] loop.py:45 INFO TRAIN_STEP step=13633 loss=1.197819 Training: 76%|███████▌ | 13633/18018 [29:23:30<9:46:40, 8.03s/step, loss=1.198, acc=0.716, acc_len=1.74, thru=17.6, I=18.1, T=37.5, wait=3.4s, pool=49, epoch=2/2] [2026-10-19 18:53:51,715] loop.py:45 INFO TRAIN_STEP step=13634 loss=1.209974 Training: 76%|███████▌ | 13634/18018 [29:23:37<9:31:38, 7.82s/step, loss=1.210, acc=0.715, acc_len=1.73, thru=15.6, I=16.4, T=37.7, wait=3.9s, pool=41, epoch=2/2][2026-10-19 18:53:59,020] loop.py:380 INFO TIMING step=13635: step=3.425s data=1.404s compute=3.390s [fwd=2.440s bwd=0.937s opt=0.003s] dispatch=3.873s [2026-10-19 18:53:59,020] loop.py:45 INFO TRAIN_STEP step=13635 loss=1.264296 Training: 76%|███████▌ | 13635/18018 [29:23:45<9:20:10, 7.67s/step, loss=1.264, acc=0.705, acc_len=1.70, thru=16.9, I=16.7, T=37.4, wait=3.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:04,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:54:05,731] loop.py:45 INFO TRAIN_STEP step=13636 loss=1.233909 Training: 76%|███████▌ | 13636/18018 [29:23:51<8:59:04, 7.38s/step, loss=1.234, acc=0.712, acc_len=1.73, thru=19.4, I=21.2, T=36.9, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:54:05,724] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13635: forward=2476.6ms backward=953.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:05,973] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:11,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:54:12,022] loop.py:45 INFO TRAIN_STEP step=13637 loss=1.254255 Training: 76%|███████▌ | 13637/18018 [29:23:58<8:35:03, 7.05s/step, loss=1.254, acc=0.706, acc_len=1.70, thru=20.8, I=22.8, T=37.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:12,224] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:18,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:54:18,776] loop.py:45 INFO TRAIN_STEP step=13638 loss=1.197703 Training: 76%|███████▌ | 13638/18018 [29:24:04<8:28:20, 6.96s/step, loss=1.198, acc=0.718, acc_len=1.75, thru=19.9, I=20.7, T=37.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:19,255] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:25,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:54:26,188] loop.py:45 INFO TRAIN_STEP step=13639 loss=1.486059 Training: 76%|███████▌ | 13639/18018 [29:24:12<8:38:04, 7.10s/step, loss=1.486, acc=0.665, acc_len=1.56, thru=18.2, I=21.2, T=35.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:26,439] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:33,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:54:33,513] loop.py:380 INFO TIMING step=13640: step=3.499s data=1.458s compute=3.460s [fwd=2.506s bwd=0.940s opt=0.003s] dispatch=3.816s [2026-10-19 18:54:33,513] loop.py:45 INFO TRAIN_STEP step=13640 loss=1.284417 Training: 76%|███████▌ | 13640/18018 [29:24:19<8:42:53, 7.17s/step, loss=1.284, acc=0.704, acc_len=1.69, thru=17.7, I=17.9, T=36.6, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:33,861] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:39,464] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:54:40,780] loop.py:45 INFO TRAIN_STEP step=13641 loss=1.327366 Training: 76%|███████▌ | 13641/18018 [29:24:26<8:44:59, 7.20s/step, loss=1.327, acc=0.698, acc_len=1.67, thru=18.8, I=20.4, T=35.0, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:54:40,776] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13640: forward=2594.9ms backward=1012.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:40,974] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:46,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:54:47,075] loop.py:45 INFO TRAIN_STEP step=13642 loss=1.281390 Training: 76%|███████▌ | 13642/18018 [29:24:33<8:25:08, 6.93s/step, loss=1.281, acc=0.705, acc_len=1.70, thru=20.0, I=24.8, T=36.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:47,227] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:52,064] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:54:53,007] loop.py:45 INFO TRAIN_STEP step=13643 loss=1.327021 Training: 76%|███████▌ | 13643/18018 [29:24:39<8:03:19, 6.63s/step, loss=1.327, acc=0.694, acc_len=1.66, thru=20.8, I=25.5, T=37.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:53,076] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:58,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:54:58,758] loop.py:45 INFO TRAIN_STEP step=13644 loss=1.245646 Training: 76%|███████▌ | 13644/18018 [29:24:44<7:43:58, 6.36s/step, loss=1.246, acc=0.708, acc_len=1.70, thru=20.8, I=23.9, T=38.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:54:59,196] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:06,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:55:07,776] loop.py:380 INFO TIMING step=13645: step=3.625s data=1.514s compute=3.599s [fwd=2.577s bwd=1.008s opt=0.003s] dispatch=5.386s [2026-10-19 18:55:07,776] loop.py:45 INFO TRAIN_STEP step=13645 loss=1.262546 Training: 76%|███████▌ | 13645/18018 [29:24:53<8:41:53, 7.16s/step, loss=1.263, acc=0.709, acc_len=1.71, thru=14.6, I=17.6, T=35.3, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:08,064] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:13,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:55:14,838] loop.py:45 INFO TRAIN_STEP step=13646 loss=1.255283 Training: 76%|███████▌ | 13646/18018 [29:25:00<8:39:38, 7.13s/step, loss=1.255, acc=0.704, acc_len=1.69, thru=16.9, I=21.7, T=33.1, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:55:14,834] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13645: forward=2820.1ms backward=1003.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:15,216] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:19,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:55:20,801] loop.py:45 INFO TRAIN_STEP step=13647 loss=1.209281 Training: 76%|███████▌ | 13647/18018 [29:25:06<8:13:57, 6.78s/step, loss=1.209, acc=0.712, acc_len=1.72, thru=19.9, I=26.3, T=38.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:21,141] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:27,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:55:27,768] loop.py:45 INFO TRAIN_STEP step=13648 loss=1.170029 Training: 76%|███████▌ | 13648/18018 [29:25:13<8:17:56, 6.84s/step, loss=1.170, acc=0.721, acc_len=1.76, thru=19.2, I=23.0, T=37.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:28,220] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:33,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:55:34,296] loop.py:45 INFO TRAIN_STEP step=13649 loss=1.256110 Training: 76%|███████▌ | 13649/18018 [29:25:20<8:11:05, 6.74s/step, loss=1.256, acc=0.707, acc_len=1.70, thru=19.3, I=23.6, T=38.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:34,512] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:39,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:55:40,049] loop.py:380 INFO TIMING step=13650: step=3.314s data=1.371s compute=3.291s [fwd=2.357s bwd=0.920s opt=0.003s] dispatch=2.428s [2026-10-19 18:55:40,049] loop.py:45 INFO TRAIN_STEP step=13650 loss=1.265872 Training: 76%|███████▌ | 13650/18018 [29:25:26<7:49:17, 6.45s/step, loss=1.266, acc=0.705, acc_len=1.70, thru=21.2, I=24.1, T=38.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:40,453] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:47,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:55:48,429] loop.py:45 INFO TRAIN_STEP step=13651 loss=1.344113 Training: 76%|███████▌ | 13651/18018 [29:25:34<8:31:24, 7.03s/step, loss=1.344, acc=0.696, acc_len=1.66, thru=16.1, I=17.1, T=35.2, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:55:48,424] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13650: forward=2595.0ms backward=998.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:55:48,456] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 18:55:54,256] loop.py:45 INFO TRAIN_STEP step=13652 loss=1.304256 Training: 76%|███████▌ | 13652/18018 [29:25:40<8:05:11, 6.67s/step, loss=1.304, acc=0.699, acc_len=1.67, thru=18.4, I=20.1, T=37.2, wait=2.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:04,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:56:05,246] loop.py:45 INFO TRAIN_STEP step=13653 loss=1.326818 Training: 76%|███████▌ | 13653/18018 [29:25:51<9:39:20, 7.96s/step, loss=1.327, acc=0.696, acc_len=1.67, thru=14.9, I=14.4, T=32.7, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:05,576] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:11,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:56:11,733] loop.py:45 INFO TRAIN_STEP step=13654 loss=1.199231 Training: 76%|███████▌ | 13654/18018 [29:25:57<9:07:01, 7.52s/step, loss=1.199, acc=0.716, acc_len=1.74, thru=18.5, I=21.6, T=37.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:11,769] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:18,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:56:19,418] loop.py:380 INFO TIMING step=13655: step=3.576s data=1.540s compute=3.543s [fwd=2.572s bwd=0.956s opt=0.003s] dispatch=4.098s [2026-10-19 18:56:19,418] loop.py:45 INFO TRAIN_STEP step=13655 loss=1.417444 Training: 76%|███████▌ | 13655/18018 [29:26:05<9:10:26, 7.57s/step, loss=1.417, acc=0.680, acc_len=1.61, thru=16.9, I=20.5, T=35.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:19,531] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 18:56:25,726] loop.py:45 INFO TRAIN_STEP step=13656 loss=1.321040 Training: 76%|███████▌ | 13656/18018 [29:26:11<8:42:55, 7.19s/step, loss=1.321, acc=0.697, acc_len=1.66, thru=17.2, I=20.2, T=39.5, wait=3.1s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:56:25,722] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13655: forward=2315.6ms backward=870.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:32,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:56:33,649] loop.py:45 INFO TRAIN_STEP step=13657 loss=1.261949 Training: 76%|███████▌ | 13657/18018 [29:26:19<8:58:35, 7.41s/step, loss=1.262, acc=0.707, acc_len=1.71, thru=18.8, I=17.6, T=35.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:34,149] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:56:40,950] loop.py:45 INFO TRAIN_STEP step=13658 loss=1.294888 Training: 76%|███████▌ | 13658/18018 [29:26:27<8:56:08, 7.38s/step, loss=1.295, acc=0.701, acc_len=1.69, thru=15.1, I=20.0, T=35.6, wait=3.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:49,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:56:49,621] loop.py:45 INFO TRAIN_STEP step=13659 loss=1.339890 Training: 76%|███████▌ | 13659/18018 [29:26:35<9:24:11, 7.77s/step, loss=1.340, acc=0.695, acc_len=1.66, thru=15.9, I=15.9, T=36.6, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:56:49,887] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 18:56:57,328] loop.py:380 INFO TIMING step=13660: step=3.845s data=1.562s compute=3.803s [fwd=2.695s bwd=1.093s opt=0.003s] dispatch=3.852s [2026-10-19 18:56:57,328] loop.py:45 INFO TRAIN_STEP step=13660 loss=1.295337 Training: 76%|███████▌ | 13660/18018 [29:26:43<9:22:43, 7.75s/step, loss=1.295, acc=0.699, acc_len=1.67, thru=15.6, I=15.0, T=33.3, wait=3.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:03,357] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:57:03,648] loop.py:45 INFO TRAIN_STEP step=13661 loss=1.406402 Training: 76%|███████▌ | 13661/18018 [29:26:49<8:51:30, 7.32s/step, loss=1.406, acc=0.680, acc_len=1.61, thru=19.8, I=18.9, T=36.4, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:57:03,644] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13660: forward=2496.9ms backward=970.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:03,859] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:08,822] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:57:09,456] loop.py:45 INFO TRAIN_STEP step=13662 loss=1.253949 Training: 76%|███████▌ | 13662/18018 [29:26:55<8:18:26, 6.87s/step, loss=1.254, acc=0.708, acc_len=1.71, thru=21.0, I=23.2, T=39.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:09,826] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:57:16,824] loop.py:45 INFO TRAIN_STEP step=13663 loss=1.121278 Training: 76%|███████▌ | 13663/18018 [29:27:02<8:29:20, 7.02s/step, loss=1.121, acc=0.728, acc_len=1.78, thru=16.5, I=21.3, T=35.0, wait=3.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:23,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:57:24,613] loop.py:45 INFO TRAIN_STEP step=13664 loss=1.219001 Training: 76%|███████▌ | 13664/18018 [29:27:10<8:46:01, 7.25s/step, loss=1.219, acc=0.714, acc_len=1.73, thru=17.2, I=17.7, T=33.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:24,701] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:30,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:57:30,836] loop.py:380 INFO TIMING step=13665: step=3.561s data=1.530s compute=3.536s [fwd=2.546s bwd=0.976s opt=0.003s] dispatch=2.653s [2026-10-19 18:57:30,836] loop.py:45 INFO TRAIN_STEP step=13665 loss=1.266953 Training: 76%|███████▌ | 13665/18018 [29:27:16<8:23:34, 6.94s/step, loss=1.267, acc=0.707, acc_len=1.70, thru=19.3, I=22.0, T=35.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:31,125] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:37,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:57:37,380] loop.py:45 INFO TRAIN_STEP step=13666 loss=1.240556 Training: 76%|███████▌ | 13666/18018 [29:27:23<8:14:48, 6.82s/step, loss=1.241, acc=0.713, acc_len=1.72, thru=19.3, I=22.3, T=38.3, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:57:37,376] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13665: forward=2393.0ms backward=899.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:57:37,607] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 18:57:43,496] loop.py:45 INFO TRAIN_STEP step=13667 loss=1.374721 Training: 76%|███████▌ | 13667/18018 [29:27:29<7:59:18, 6.61s/step, loss=1.375, acc=0.685, acc_len=1.64, thru=19.6, I=20.7, T=38.5, wait=2.8s, pool=53, epoch=2/2][2026-10-19 18:57:49,967] loop.py:45 INFO TRAIN_STEP step=13668 loss=1.261703 Training: 76%|███████▌ | 13668/18018 [29:27:36<7:56:15, 6.57s/step, loss=1.262, acc=0.706, acc_len=1.70, thru=19.8, I=22.8, T=36.9, wait=3.0s, pool=61, epoch=2/2][2026-10-19 18:57:57,843] loop.py:45 INFO TRAIN_STEP step=13669 loss=1.308876 Training: 76%|███████▌ | 13669/18018 [29:27:43<8:24:32, 6.96s/step, loss=1.309, acc=0.696, acc_len=1.66, thru=14.4, I=16.0, T=37.4, wait=4.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:07,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:58:08,223] loop.py:380 INFO TIMING step=13670: step=4.025s data=1.907s compute=3.630s [fwd=2.460s bwd=1.155s opt=0.003s] dispatch=6.344s [2026-10-19 18:58:08,223] loop.py:45 INFO TRAIN_STEP step=13670 loss=1.172175 Training: 76%|███████▌ | 13670/18018 [29:27:54<9:38:45, 7.99s/step, loss=1.172, acc=0.722, acc_len=1.76, thru=14.4, I=14.1, T=31.8, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:08,322] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 18:58:14,056] loop.py:45 INFO TRAIN_STEP step=13671 loss=1.267864 Training: 76%|███████▌ | 13671/18018 [29:28:00<8:51:50, 7.34s/step, loss=1.268, acc=0.706, acc_len=1.70, thru=18.0, I=20.0, T=36.8, wait=2.3s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:58:14,052] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13670: forward=2448.9ms backward=979.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:21,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:58:21,930] loop.py:45 INFO TRAIN_STEP step=13672 loss=1.260137 Training: 76%|███████▌ | 13672/18018 [29:28:08<9:03:16, 7.50s/step, loss=1.260, acc=0.705, acc_len=1.70, thru=18.4, I=17.1, T=37.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:22,019] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:27,265] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:58:28,150] loop.py:45 INFO TRAIN_STEP step=13673 loss=1.158626 Training: 76%|███████▌ | 13673/18018 [29:28:14<8:35:20, 7.12s/step, loss=1.159, acc=0.726, acc_len=1.77, thru=20.5, I=22.0, T=36.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:28,270] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 18:58:34,732] loop.py:45 INFO TRAIN_STEP step=13674 loss=1.277168 Training: 76%|███████▌ | 13674/18018 [29:28:20<8:23:37, 6.96s/step, loss=1.277, acc=0.707, acc_len=1.71, thru=15.4, I=20.3, T=37.3, wait=3.1s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:42,720] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:58:43,366] loop.py:380 INFO TIMING step=13675: step=3.477s data=1.462s compute=3.449s [fwd=2.466s bwd=0.968s opt=0.003s] dispatch=5.146s [2026-10-19 18:58:43,366] loop.py:45 INFO TRAIN_STEP step=13675 loss=1.324452 Training: 76%|███████▌ | 13675/18018 [29:28:29<8:59:52, 7.46s/step, loss=1.324, acc=0.694, acc_len=1.67, thru=18.3, I=15.8, T=36.8, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:43,726] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:51,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:58:52,154] loop.py:45 INFO TRAIN_STEP step=13676 loss=1.337248 Training: 76%|███████▌ | 13676/18018 [29:28:38<9:28:40, 7.86s/step, loss=1.337, acc=0.695, acc_len=1.67, thru=14.0, I=16.3, T=34.5, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:58:52,150] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13675: forward=2651.9ms backward=993.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:52,191] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:58:59,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:59:00,545] loop.py:45 INFO TRAIN_STEP step=13677 loss=1.275395 Training: 76%|███████▌ | 13677/18018 [29:28:46<9:40:03, 8.02s/step, loss=1.275, acc=0.706, acc_len=1.69, thru=15.2, I=18.5, T=34.0, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:59:01,017] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:59:08,367] loop.py:45 INFO TRAIN_STEP step=13678 loss=1.235274 Training: 76%|███████▌ | 13678/18018 [29:28:54<9:35:41, 7.96s/step, loss=1.235, acc=0.712, acc_len=1.72, thru=14.2, I=19.8, T=38.1, wait=4.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:59:18,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:59:19,771] loop.py:45 INFO TRAIN_STEP step=13679 loss=1.232684 Training: 76%|███████▌ | 13679/18018 [29:29:05<10:50:18, 8.99s/step, loss=1.233, acc=0.717, acc_len=1.74, thru=14.4, I=13.3, T=33.3, wait=7.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:59:19,889] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:59:26,039] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:59:26,614] loop.py:380 INFO TIMING step=13680: step=3.425s data=1.383s compute=3.388s [fwd=2.422s bwd=0.952s opt=0.003s] dispatch=3.410s [2026-10-19 18:59:26,615] loop.py:45 INFO TRAIN_STEP step=13680 loss=1.184866 Training: 76%|███████▌ | 13680/18018 [29:29:12<10:03:33, 8.35s/step, loss=1.185, acc=0.719, acc_len=1.75, thru=17.9, I=21.7, T=37.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:59:27,045] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 18:59:34,979] loop.py:45 INFO TRAIN_STEP step=13681 loss=1.197353 Training: 76%|███████▌ | 13681/18018 [29:29:21<10:03:44, 8.35s/step, loss=1.197, acc=0.713, acc_len=1.72, thru=12.3, I=12.5, T=35.8, wait=4.8s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 18:59:34,976] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13680: forward=2564.5ms backward=956.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:59:45,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 18:59:46,431] loop.py:45 INFO TRAIN_STEP step=13682 loss=1.275228 Training: 76%|███████▌ | 13682/18018 [29:29:32<11:10:49, 9.28s/step, loss=1.275, acc=0.707, acc_len=1.71, thru=13.3, I=15.6, T=32.5, wait=7.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 18:59:46,624] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 18:59:52,476] loop.py:45 INFO TRAIN_STEP step=13683 loss=1.243331 Training: 76%|███████▌ | 13683/18018 [29:29:38<10:00:29, 8.31s/step, loss=1.243, acc=0.710, acc_len=1.72, thru=16.8, I=19.9, T=39.3, wait=2.8s, pool=41, epoch=2/2][2026-10-19 19:00:00,304] loop.py:45 INFO TRAIN_STEP step=13684 loss=1.240722 Training: 76%|███████▌ | 13684/18018 [29:29:46<9:49:55, 8.17s/step, loss=1.241, acc=0.711, acc_len=1.72, thru=15.6, I=16.6, T=35.1, wait=4.2s, pool=33, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:08,615] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:00:09,421] loop.py:380 INFO TIMING step=13685: step=3.666s data=1.521s compute=3.632s [fwd=2.559s bwd=1.059s opt=0.003s] dispatch=5.442s [2026-10-19 19:00:09,422] loop.py:45 INFO TRAIN_STEP step=13685 loss=1.206095 Training: 76%|███████▌ | 13685/18018 [29:29:55<10:10:19, 8.45s/step, loss=1.206, acc=0.717, acc_len=1.73, thru=16.9, I=14.8, T=34.9, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:09,619] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:15,688] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:00:16,105] loop.py:45 INFO TRAIN_STEP step=13686 loss=1.229228 Training: 76%|███████▌ | 13686/18018 [29:30:02<9:31:56, 7.92s/step, loss=1.229, acc=0.713, acc_len=1.72, thru=18.1, I=19.9, T=35.8, wait=3.1s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 19:00:16,101] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13685: forward=2535.1ms backward=992.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:16,199] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:21,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:00:22,108] loop.py:45 INFO TRAIN_STEP step=13687 loss=1.284774 Training: 76%|███████▌ | 13687/18018 [29:30:08<8:50:13, 7.35s/step, loss=1.285, acc=0.700, acc_len=1.68, thru=20.6, I=22.5, T=33.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:22,574] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:27,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:00:28,292] loop.py:45 INFO TRAIN_STEP step=13688 loss=1.409242 Training: 76%|███████▌ | 13688/18018 [29:30:14<8:24:58, 7.00s/step, loss=1.409, acc=0.679, acc_len=1.61, thru=19.3, I=23.7, T=36.1, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:28,550] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:32,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:00:33,452] loop.py:45 INFO TRAIN_STEP step=13689 loss=1.220925 Training: 76%|███████▌ | 13689/18018 [29:30:19<7:45:04, 6.45s/step, loss=1.221, acc=0.713, acc_len=1.73, thru=22.8, I=29.4, T=38.3, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:33,558] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 19:00:39,315] loop.py:380 INFO TIMING step=13690: step=3.307s data=1.325s compute=3.279s [fwd=2.394s bwd=0.870s opt=0.003s] dispatch=2.547s [2026-10-19 19:00:39,315] loop.py:45 INFO TRAIN_STEP step=13690 loss=1.367583 Training: 76%|███████▌ | 13690/18018 [29:30:25<7:32:20, 6.27s/step, loss=1.368, acc=0.687, acc_len=1.64, thru=21.0, I=23.2, T=38.7, wait=2.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:45,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:00:46,231] loop.py:45 INFO TRAIN_STEP step=13691 loss=1.267004 Training: 76%|███████▌ | 13691/18018 [29:30:32<7:46:12, 6.46s/step, loss=1.267, acc=0.705, acc_len=1.70, thru=20.1, I=20.3, T=37.0, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:00:46,226] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13690: forward=2436.2ms backward=989.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:46,308] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:52,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:00:53,035] loop.py:45 INFO TRAIN_STEP step=13692 loss=1.307308 Training: 76%|███████▌ | 13692/18018 [29:30:39<7:53:27, 6.57s/step, loss=1.307, acc=0.702, acc_len=1.68, thru=19.2, I=22.7, T=36.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:00:53,244] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:01:02,446] loop.py:45 INFO TRAIN_STEP step=13693 loss=1.261040 Training: 76%|███████▌ | 13693/18018 [29:30:48<8:54:48, 7.42s/step, loss=1.261, acc=0.710, acc_len=1.72, thru=13.5, I=13.3, T=33.8, wait=5.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:01:07,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:01:09,496] loop.py:45 INFO TRAIN_STEP step=13694 loss=1.295047 Training: 76%|███████▌ | 13694/18018 [29:30:55<8:46:42, 7.31s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=18.8, I=19.2, T=36.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:01:09,923] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:01:15,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:01:16,017] loop.py:380 INFO TIMING step=13695: step=3.459s data=1.468s compute=3.427s [fwd=2.447s bwd=0.965s opt=0.003s] dispatch=3.055s [2026-10-19 19:01:16,018] loop.py:45 INFO TRAIN_STEP step=13695 loss=1.291185 Training: 76%|███████▌ | 13695/18018 [29:31:02<8:29:35, 7.07s/step, loss=1.291, acc=0.700, acc_len=1.68, thru=19.3, I=26.4, T=37.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:01:16,352] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:01:22,174] loop.py:45 INFO TRAIN_STEP step=13696 loss=1.170126 Training: 76%|███████▌ | 13696/18018 [29:31:08<8:09:39, 6.80s/step, loss=1.170, acc=0.724, acc_len=1.76, thru=18.1, I=23.2, T=35.5, wait=2.5s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:01:22,169] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13695: forward=2559.4ms backward=1001.4ms [2026-10-19 19:01:31,901] loop.py:45 INFO TRAIN_STEP step=13697 loss=1.272312 Training: 76%|███████▌ | 13697/18018 [29:31:17<9:12:53, 7.68s/step, loss=1.272, acc=0.703, acc_len=1.69, thru=15.0, I=13.2, T=35.3, wait=6.1s, pool=57, epoch=2/2][2026-10-19 19:01:43,456] loop.py:45 INFO TRAIN_STEP step=13698 loss=1.301027 Training: 76%|███████▌ | 13698/18018 [29:31:29<10:36:30, 8.84s/step, loss=1.301, acc=0.700, acc_len=1.68, thru=11.0, I=10.3, T=32.7, wait=7.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:01:51,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:01:51,338] loop.py:45 INFO TRAIN_STEP step=13699 loss=1.272486 Training: 76%|███████▌ | 13699/18018 [29:31:37<10:15:36, 8.55s/step, loss=1.272, acc=0.712, acc_len=1.72, thru=17.5, I=16.0, T=34.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:01:51,790] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:01:58,263] loop.py:380 INFO TIMING step=13700: step=3.747s data=1.613s compute=3.720s [fwd=2.662s bwd=1.044s opt=0.003s] dispatch=3.170s [2026-10-19 19:01:58,263] loop.py:45 INFO TRAIN_STEP step=13700 loss=1.274196 Training: 76%|███████▌ | 13700/18018 [29:31:44<9:40:20, 8.06s/step, loss=1.274, acc=0.703, acc_len=1.69, thru=17.9, I=19.3, T=34.2, wait=3.2s, pool=61, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:05,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:02:05,793] loop.py:45 INFO TRAIN_STEP step=13701 loss=1.256891 Training: 76%|███████▌ | 13701/18018 [29:31:51<9:28:42, 7.90s/step, loss=1.257, acc=0.702, acc_len=1.68, thru=17.2, I=17.1, T=36.2, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:02:05,788] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13700: forward=2451.0ms backward=1033.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:05,948] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:02:12,666] loop.py:45 INFO TRAIN_STEP step=13702 loss=1.252682 Training: 76%|███████▌ | 13702/18018 [29:31:58<9:06:18, 7.59s/step, loss=1.253, acc=0.710, acc_len=1.72, thru=18.8, I=19.7, T=35.0, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:18,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:02:19,001] loop.py:45 INFO TRAIN_STEP step=13703 loss=1.326467 Training: 76%|███████▌ | 13703/18018 [29:32:05<8:39:03, 7.22s/step, loss=1.326, acc=0.693, acc_len=1.66, thru=19.4, I=21.8, T=36.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:19,152] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:02:25,871] loop.py:45 INFO TRAIN_STEP step=13704 loss=1.168284 Training: 76%|███████▌ | 13704/18018 [29:32:11<8:31:25, 7.11s/step, loss=1.168, acc=0.722, acc_len=1.76, thru=16.9, I=18.4, T=36.3, wait=3.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:33,032] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:02:34,125] loop.py:380 INFO TIMING step=13705: step=3.694s data=1.591s compute=3.663s [fwd=2.650s bwd=1.001s opt=0.003s] dispatch=4.550s [2026-10-19 19:02:34,125] loop.py:45 INFO TRAIN_STEP step=13705 loss=1.266949 Training: 76%|███████▌ | 13705/18018 [29:32:20<8:55:52, 7.45s/step, loss=1.267, acc=0.706, acc_len=1.70, thru=17.5, I=19.0, T=34.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:34,540] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-19 19:02:41,113] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13705: forward=2364.3ms backward=992.4ms [2026-10-19 19:02:41,118] loop.py:45 INFO TRAIN_STEP step=13706 loss=1.221937 Training: 76%|███████▌ | 13706/18018 [29:32:27<8:45:48, 7.32s/step, loss=1.222, acc=0.713, acc_len=1.73, thru=16.0, I=22.9, T=37.6, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:48,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:02:48,913] loop.py:45 INFO TRAIN_STEP step=13707 loss=1.155600 Training: 76%|███████▌ | 13707/18018 [29:32:34<8:55:58, 7.46s/step, loss=1.156, acc=0.728, acc_len=1.78, thru=17.0, I=17.2, T=36.5, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:49,180] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:56,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:02:56,919] loop.py:45 INFO TRAIN_STEP step=13708 loss=1.288706 Training: 76%|███████▌ | 13708/18018 [29:32:43<9:07:38, 7.62s/step, loss=1.289, acc=0.705, acc_len=1.70, thru=17.0, I=18.0, T=35.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:02:57,367] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:01,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 19:03:02,505] loop.py:45 INFO TRAIN_STEP step=13709 loss=1.255208 Training: 76%|███████▌ | 13709/18018 [29:32:48<8:23:37, 7.01s/step, loss=1.255, acc=0.708, acc_len=1.70, thru=22.1, I=25.0, T=38.4, wait=2.2s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:02,649] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:08,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:03:09,691] loop.py:380 INFO TIMING step=13710: step=3.812s data=1.703s compute=3.782s [fwd=2.716s bwd=1.054s opt=0.003s] dispatch=3.365s [2026-10-19 19:03:09,691] loop.py:45 INFO TRAIN_STEP step=13710 loss=1.315623 Training: 76%|███████▌ | 13710/18018 [29:32:55<8:27:16, 7.07s/step, loss=1.316, acc=0.699, acc_len=1.68, thru=17.6, I=21.8, T=33.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:09,911] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:03:17,992] loop.py:45 INFO TRAIN_STEP step=13711 loss=1.248219 Training: 76%|███████▌ | 13711/18018 [29:33:04<8:53:43, 7.44s/step, loss=1.248, acc=0.706, acc_len=1.70, thru=12.9, I=14.2, T=34.6, wait=4.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:03:17,988] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13710: forward=2658.3ms backward=995.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:24,182] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:03:25,650] loop.py:45 INFO TRAIN_STEP step=13712 loss=1.293078 Training: 76%|███████▌ | 13712/18018 [29:33:11<8:58:25, 7.50s/step, loss=1.293, acc=0.700, acc_len=1.68, thru=18.4, I=18.2, T=33.8, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:25,699] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 19:03:31,914] loop.py:45 INFO TRAIN_STEP step=13713 loss=1.249991 Training: 76%|███████▌ | 13713/18018 [29:33:18<8:31:41, 7.13s/step, loss=1.250, acc=0.710, acc_len=1.72, thru=17.4, I=23.7, T=38.0, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:39,756] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:03:40,836] loop.py:45 INFO TRAIN_STEP step=13714 loss=1.278238 Training: 76%|███████▌ | 13714/18018 [29:33:26<9:10:01, 7.67s/step, loss=1.278, acc=0.700, acc_len=1.68, thru=14.9, I=15.9, T=33.1, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:41,264] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:46,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:03:48,574] loop.py:380 INFO TIMING step=13715: step=4.602s data=1.778s compute=4.564s [fwd=3.478s bwd=1.071s opt=0.003s] dispatch=3.127s [2026-10-19 19:03:48,574] loop.py:45 INFO TRAIN_STEP step=13715 loss=1.174538 Training: 76%|███████▌ | 13715/18018 [29:33:34<9:11:27, 7.69s/step, loss=1.175, acc=0.723, acc_len=1.76, thru=15.2, I=20.3, T=27.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:48,727] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:55,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:03:56,079] loop.py:45 INFO TRAIN_STEP step=13716 loss=1.319556 Training: 76%|███████▌ | 13716/18018 [29:33:42<9:07:18, 7.63s/step, loss=1.320, acc=0.695, acc_len=1.66, thru=14.4, I=19.0, T=36.4, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:03:56,073] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13715: forward=2490.5ms backward=976.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:03:56,516] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:04,155] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:04:04,425] loop.py:45 INFO TRAIN_STEP step=13717 loss=1.235247 Training: 76%|███████▌ | 13717/18018 [29:33:50<9:22:32, 7.85s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=15.2, I=16.8, T=34.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:04,658] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:09,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:04:10,522] loop.py:45 INFO TRAIN_STEP step=13718 loss=1.383737 Training: 76%|███████▌ | 13718/18018 [29:33:56<8:44:47, 7.32s/step, loss=1.384, acc=0.687, acc_len=1.64, thru=20.3, I=20.0, T=35.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:10,757] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:16,643] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:04:17,499] loop.py:45 INFO TRAIN_STEP step=13719 loss=1.243024 Training: 76%|███████▌ | 13719/18018 [29:34:03<8:37:15, 7.22s/step, loss=1.243, acc=0.708, acc_len=1.71, thru=19.3, I=20.6, T=35.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:17,653] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:04:25,205] loop.py:380 INFO TIMING step=13720: step=3.473s data=1.447s compute=3.435s [fwd=2.434s bwd=0.986s opt=0.003s] dispatch=4.221s [2026-10-19 19:04:25,205] loop.py:45 INFO TRAIN_STEP step=13720 loss=1.344984 Training: 76%|███████▌ | 13720/18018 [29:34:11<8:47:36, 7.37s/step, loss=1.345, acc=0.691, acc_len=1.65, thru=14.8, I=16.9, T=36.9, wait=4.2s, pool=53, epoch=2/2][2026-10-19 19:04:33,040] loop.py:45 INFO TRAIN_STEP step=13721 loss=1.380346 Training: 76%|███████▌ | 13721/18018 [29:34:19<8:57:30, 7.51s/step, loss=1.380, acc=0.684, acc_len=1.63, thru=16.9, I=18.7, T=34.9, wait=4.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:04:33,035] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13720: forward=2565.2ms backward=1048.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:41,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:04:42,035] loop.py:45 INFO TRAIN_STEP step=13722 loss=1.301285 Training: 76%|███████▌ | 13722/18018 [29:34:28<9:29:26, 7.95s/step, loss=1.301, acc=0.701, acc_len=1.68, thru=15.7, I=15.4, T=34.3, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:42,153] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:47,207] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:04:48,559] loop.py:45 INFO TRAIN_STEP step=13723 loss=1.216750 Training: 76%|███████▌ | 13723/18018 [29:34:34<8:58:36, 7.52s/step, loss=1.217, acc=0.715, acc_len=1.73, thru=19.3, I=21.4, T=36.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:04:48,720] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:04:55,382] loop.py:45 INFO TRAIN_STEP step=13724 loss=1.210306 Training: 76%|███████▌ | 13724/18018 [29:34:41<8:43:24, 7.31s/step, loss=1.210, acc=0.716, acc_len=1.73, thru=17.1, I=20.9, T=35.8, wait=3.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:01,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:05:02,495] loop.py:380 INFO TIMING step=13725: step=3.561s data=1.445s compute=3.529s [fwd=2.515s bwd=0.999s opt=0.003s] dispatch=3.543s [2026-10-19 19:05:02,495] loop.py:45 INFO TRAIN_STEP step=13725 loss=1.318813 Training: 76%|███████▌ | 13725/18018 [29:34:48<8:38:57, 7.25s/step, loss=1.319, acc=0.696, acc_len=1.67, thru=18.5, I=19.3, T=35.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:02,604] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:07,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:05:08,665] loop.py:45 INFO TRAIN_STEP step=13726 loss=1.296310 Training: 76%|███████▌ | 13726/18018 [29:34:54<8:15:38, 6.93s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=20.7, I=24.9, T=37.9, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:05:08,660] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13725: forward=2393.1ms backward=932.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:08,832] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:14,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:05:15,432] loop.py:45 INFO TRAIN_STEP step=13727 loss=1.347718 Training: 76%|███████▌ | 13727/18018 [29:35:01<8:12:03, 6.88s/step, loss=1.348, acc=0.692, acc_len=1.65, thru=19.3, I=22.5, T=37.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:15,463] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:21,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:05:22,359] loop.py:45 INFO TRAIN_STEP step=13728 loss=1.277006 Training: 76%|███████▌ | 13728/18018 [29:35:08<8:12:55, 6.89s/step, loss=1.277, acc=0.705, acc_len=1.70, thru=18.4, I=21.2, T=29.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:22,482] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:28,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:05:29,484] loop.py:45 INFO TRAIN_STEP step=13729 loss=1.249659 Training: 76%|███████▌ | 13729/18018 [29:35:15<8:17:46, 6.96s/step, loss=1.250, acc=0.708, acc_len=1.70, thru=16.5, I=20.4, T=34.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:29,676] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:05:36,713] loop.py:380 INFO TIMING step=13730: step=3.439s data=1.331s compute=3.410s [fwd=2.433s bwd=0.962s opt=0.003s] dispatch=3.780s [2026-10-19 19:05:36,713] loop.py:45 INFO TRAIN_STEP step=13730 loss=1.113724 Training: 76%|███████▌ | 13730/18018 [29:35:22<8:23:17, 7.04s/step, loss=1.114, acc=0.732, acc_len=1.79, thru=14.4, I=16.7, T=37.2, wait=3.8s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:44,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:05:44,972] loop.py:45 INFO TRAIN_STEP step=13731 loss=1.268551 Training: 76%|███████▌ | 13731/18018 [29:35:31<8:49:15, 7.41s/step, loss=1.269, acc=0.708, acc_len=1.70, thru=18.2, I=16.7, T=35.5, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:05:44,967] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13730: forward=2397.0ms backward=1131.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:05:45,002] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 19:05:52,341] loop.py:45 INFO TRAIN_STEP step=13732 loss=1.342050 Training: 76%|███████▌ | 13732/18018 [29:35:38<8:48:19, 7.40s/step, loss=1.342, acc=0.696, acc_len=1.66, thru=17.3, I=17.1, T=36.3, wait=3.8s, pool=61, epoch=2/2][2026-10-19 19:05:59,174] loop.py:45 INFO TRAIN_STEP step=13733 loss=1.393516 Training: 76%|███████▌ | 13733/18018 [29:35:45<8:36:08, 7.23s/step, loss=1.394, acc=0.682, acc_len=1.62, thru=17.5, I=16.0, T=33.6, wait=3.0s, pool=45, epoch=2/2][2026-10-19 19:06:06,314] loop.py:45 INFO TRAIN_STEP step=13734 loss=1.289905 Training: 76%|███████▌ | 13734/18018 [29:35:52<8:34:13, 7.20s/step, loss=1.290, acc=0.702, acc_len=1.69, thru=17.6, I=16.9, T=35.1, wait=3.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:13,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:06:14,600] loop.py:380 INFO TIMING step=13735: step=3.515s data=1.462s compute=3.473s [fwd=2.469s bwd=0.990s opt=0.003s] dispatch=4.760s [2026-10-19 19:06:14,600] loop.py:45 INFO TRAIN_STEP step=13735 loss=1.130096 Training: 76%|███████▌ | 13735/18018 [29:36:00<8:57:13, 7.53s/step, loss=1.130, acc=0.733, acc_len=1.80, thru=17.3, I=17.8, T=36.4, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:14,983] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:06:20,917] loop.py:45 INFO TRAIN_STEP step=13736 loss=1.283410 Training: 76%|███████▌ | 13736/18018 [29:36:07<8:31:13, 7.16s/step, loss=1.283, acc=0.700, acc_len=1.67, thru=19.1, I=22.9, T=37.7, wait=2.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:06:20,913] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13735: forward=2413.0ms backward=930.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:27,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:06:28,291] loop.py:45 INFO TRAIN_STEP step=13737 loss=1.209663 Training: 76%|███████▌ | 13737/18018 [29:36:14<8:35:39, 7.23s/step, loss=1.210, acc=0.714, acc_len=1.73, thru=18.1, I=18.9, T=28.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:28,793] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:34,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:06:36,495] loop.py:45 INFO TRAIN_STEP step=13738 loss=1.285960 Training: 76%|███████▌ | 13738/18018 [29:36:22<8:56:26, 7.52s/step, loss=1.286, acc=0.703, acc_len=1.69, thru=16.0, I=17.9, T=31.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:36,683] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 19:06:44,381] loop.py:45 INFO TRAIN_STEP step=13739 loss=1.293545 Training: 76%|███████▋ | 13739/18018 [29:36:30<9:04:05, 7.63s/step, loss=1.294, acc=0.701, acc_len=1.68, thru=12.4, I=13.6, T=35.1, wait=4.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:52,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:06:52,330] loop.py:380 INFO TIMING step=13740: step=3.635s data=1.454s compute=3.605s [fwd=2.566s bwd=1.025s opt=0.003s] dispatch=4.305s [2026-10-19 19:06:52,330] loop.py:45 INFO TRAIN_STEP step=13740 loss=1.209880 Training: 76%|███████▋ | 13740/18018 [29:36:38<9:10:48, 7.73s/step, loss=1.210, acc=0.719, acc_len=1.75, thru=18.6, I=15.8, T=35.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:52,672] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:58,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 19:06:59,345] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13740: forward=2475.8ms backward=1048.5ms [2026-10-19 19:06:59,349] loop.py:45 INFO TRAIN_STEP step=13741 loss=1.264735 Training: 76%|███████▋ | 13741/18018 [29:36:45<8:55:37, 7.51s/step, loss=1.265, acc=0.704, acc_len=1.69, thru=18.0, I=19.3, T=35.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:06:59,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:05,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:07:05,776] loop.py:45 INFO TRAIN_STEP step=13742 loss=1.149017 Training: 76%|███████▋ | 13742/18018 [29:36:51<8:32:14, 7.19s/step, loss=1.149, acc=0.724, acc_len=1.76, thru=19.7, I=21.5, T=38.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:06,095] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:11,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:07:11,808] loop.py:45 INFO TRAIN_STEP step=13743 loss=1.342978 Training: 76%|███████▋ | 13743/18018 [29:36:57<8:07:22, 6.84s/step, loss=1.343, acc=0.689, acc_len=1.64, thru=20.6, I=23.2, T=35.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:12,020] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:07:20,748] loop.py:45 INFO TRAIN_STEP step=13744 loss=1.259220 Training: 76%|███████▋ | 13744/18018 [29:37:06<8:52:07, 7.47s/step, loss=1.259, acc=0.706, acc_len=1.70, thru=12.7, I=15.6, T=34.1, wait=5.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:31,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:07:32,367] loop.py:380 INFO TIMING step=13745: step=3.824s data=1.520s compute=3.796s [fwd=2.706s bwd=1.075s opt=0.003s] dispatch=7.787s [2026-10-19 19:07:32,367] loop.py:45 INFO TRAIN_STEP step=13745 loss=1.293316 Training: 76%|███████▋ | 13745/18018 [29:37:18<10:20:43, 8.72s/step, loss=1.293, acc=0.703, acc_len=1.69, thru=13.0, I=11.6, T=33.5, wait=7.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:32,486] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:39,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:07:40,961] loop.py:45 INFO TRAIN_STEP step=13746 loss=1.322322 Training: 76%|███████▋ | 13746/18018 [29:37:27<10:17:56, 8.68s/step, loss=1.322, acc=0.697, acc_len=1.67, thru=15.1, I=17.8, T=36.0, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:07:40,956] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13745: forward=2532.8ms backward=976.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:41,192] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:46,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:07:47,388] loop.py:45 INFO TRAIN_STEP step=13747 loss=1.217505 Training: 76%|███████▋ | 13747/18018 [29:37:33<9:29:40, 8.00s/step, loss=1.218, acc=0.713, acc_len=1.72, thru=19.3, I=24.6, T=38.5, wait=3.1s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:47,886] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:54,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:07:55,130] loop.py:45 INFO TRAIN_STEP step=13748 loss=1.260038 Training: 76%|███████▋ | 13748/18018 [29:37:41<9:23:59, 7.93s/step, loss=1.260, acc=0.709, acc_len=1.71, thru=16.9, I=20.0, T=35.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:07:55,282] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:00,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:08:01,411] loop.py:45 INFO TRAIN_STEP step=13749 loss=1.129809 Training: 76%|███████▋ | 13749/18018 [29:37:47<8:48:48, 7.43s/step, loss=1.130, acc=0.732, acc_len=1.79, thru=19.7, I=21.1, T=38.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:01,526] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:07,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:08:08,504] loop.py:380 INFO TIMING step=13750: step=3.624s data=1.640s compute=3.588s [fwd=2.552s bwd=1.022s opt=0.003s] dispatch=3.458s [2026-10-19 19:08:08,504] loop.py:45 INFO TRAIN_STEP step=13750 loss=1.159886 Training: 76%|███████▋ | 13750/18018 [29:37:54<8:41:23, 7.33s/step, loss=1.160, acc=0.724, acc_len=1.76, thru=17.9, I=22.5, T=35.3, wait=3.5s, pool=65, epoch=2/2][2026-10-19 19:08:08,507] loop.py:419 INFO Saving checkpoint at step 13750... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 19:08:11,921] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0013751 (TrainerActor pid=9310) [2026-10-19 19:08:11,937] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0013001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:12,714] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:13,222] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=9/64 [2026-10-19 19:08:18,640] loop.py:45 INFO TRAIN_STEP step=13751 loss=1.226636 Training: 76%|███████▋ | 13751/18018 [29:38:04<9:41:10, 8.17s/step, loss=1.227, acc=0.713, acc_len=1.73, thru=19.3, I=25.5, T=37.5, wait=2.2s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:08:18,635] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13750: forward=2411.5ms backward=946.6ms (TrainerActor pid=9310) [2026-10-19 19:08:11,919] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0013751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 19:08:13,046] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0013001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:28,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:08:28,537] loop.py:45 INFO TRAIN_STEP step=13752 loss=1.392199 Training: 76%|███████▋ | 13752/18018 [29:38:14<10:17:47, 8.69s/step, loss=1.392, acc=0.682, acc_len=1.62, thru=15.4, I=14.2, T=34.7, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:28,777] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:35,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:08:35,511] loop.py:45 INFO TRAIN_STEP step=13753 loss=1.283047 Training: 76%|███████▋ | 13753/18018 [29:38:21<9:41:04, 8.17s/step, loss=1.283, acc=0.700, acc_len=1.68, thru=18.2, I=18.4, T=36.9, wait=3.5s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:35,856] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:42,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:08:42,607] loop.py:45 INFO TRAIN_STEP step=13754 loss=1.264831 Training: 76%|███████▋ | 13754/18018 [29:38:28<9:17:58, 7.85s/step, loss=1.265, acc=0.703, acc_len=1.70, thru=18.4, I=19.1, T=36.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:42,748] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:48,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:08:49,020] loop.py:380 INFO TIMING step=13755: step=3.616s data=1.511s compute=3.580s [fwd=2.583s bwd=0.982s opt=0.003s] dispatch=2.787s [2026-10-19 19:08:49,020] loop.py:45 INFO TRAIN_STEP step=13755 loss=1.390255 Training: 76%|███████▋ | 13755/18018 [29:38:35<8:47:10, 7.42s/step, loss=1.390, acc=0.687, acc_len=1.64, thru=19.4, I=19.4, T=35.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:49,150] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:55,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:08:56,265] loop.py:45 INFO TRAIN_STEP step=13756 loss=1.146809 Training: 76%|███████▋ | 13756/18018 [29:38:42<8:43:17, 7.37s/step, loss=1.147, acc=0.728, acc_len=1.77, thru=17.8, I=19.5, T=31.1, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:08:56,260] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13755: forward=3092.1ms backward=978.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:08:56,733] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:04,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:09:04,841] loop.py:45 INFO TRAIN_STEP step=13757 loss=1.304601 Training: 76%|███████▋ | 13757/18018 [29:38:50<9:08:58, 7.73s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=14.5, I=15.7, T=35.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:05,271] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:09:11,671] loop.py:45 INFO TRAIN_STEP step=13758 loss=1.252449 Training: 76%|███████▋ | 13758/18018 [29:38:57<8:49:36, 7.46s/step, loss=1.252, acc=0.712, acc_len=1.72, thru=18.9, I=19.9, T=35.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:11,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:12,164] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:17,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:09:18,720] loop.py:45 INFO TRAIN_STEP step=13759 loss=1.345931 Training: 76%|███████▋ | 13759/18018 [29:39:04<8:40:46, 7.34s/step, loss=1.346, acc=0.690, acc_len=1.64, thru=18.2, I=20.5, T=38.1, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:18,943] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:24,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:09:25,154] loop.py:380 INFO TIMING step=13760: step=3.306s data=1.343s compute=3.274s [fwd=2.352s bwd=0.908s opt=0.003s] dispatch=3.119s [2026-10-19 19:09:25,155] loop.py:45 INFO TRAIN_STEP step=13760 loss=1.220580 Training: 76%|███████▋ | 13760/18018 [29:39:11<8:21:29, 7.07s/step, loss=1.221, acc=0.714, acc_len=1.73, thru=20.1, I=23.9, T=38.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:25,214] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 19:09:31,847] loop.py:45 INFO TRAIN_STEP step=13761 loss=1.192686 Training: 76%|███████▋ | 13761/18018 [29:39:17<8:13:25, 6.95s/step, loss=1.193, acc=0.717, acc_len=1.73, thru=15.7, I=20.8, T=36.7, wait=3.2s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:09:31,841] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13760: forward=2451.8ms backward=991.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:40,054] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:09:40,274] loop.py:45 INFO TRAIN_STEP step=13762 loss=1.188019 Training: 76%|███████▋ | 13762/18018 [29:39:26<8:44:35, 7.40s/step, loss=1.188, acc=0.719, acc_len=1.74, thru=17.9, I=15.8, T=37.8, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:40,555] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:48,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:09:48,440] loop.py:45 INFO TRAIN_STEP step=13763 loss=1.309447 Training: 76%|███████▋ | 13763/18018 [29:39:34<9:00:53, 7.63s/step, loss=1.309, acc=0.699, acc_len=1.67, thru=16.6, I=17.2, T=35.5, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:48,861] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:55,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:09:55,404] loop.py:45 INFO TRAIN_STEP step=13764 loss=1.396693 Training: 76%|███████▋ | 13764/18018 [29:39:41<8:46:40, 7.43s/step, loss=1.397, acc=0.679, acc_len=1.61, thru=17.7, I=18.4, T=37.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:09:55,791] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:10:01,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:10:01,885] loop.py:380 INFO TIMING step=13765: step=3.365s data=1.346s compute=3.336s [fwd=2.418s bwd=0.903s opt=0.003s] dispatch=3.106s [2026-10-19 19:10:01,886] loop.py:45 INFO TRAIN_STEP step=13765 loss=1.395473 Training: 76%|███████▋ | 13765/18018 [29:39:47<8:26:21, 7.14s/step, loss=1.395, acc=0.683, acc_len=1.62, thru=19.6, I=20.6, T=38.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:10:02,255] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:10:06,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:10:08,531] loop.py:45 INFO TRAIN_STEP step=13766 loss=1.307593 Training: 76%|███████▋ | 13766/18018 [29:39:54<8:15:39, 6.99s/step, loss=1.308, acc=0.698, acc_len=1.68, thru=19.1, I=23.5, T=35.6, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:10:08,526] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13765: forward=2551.8ms backward=997.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:10:08,999] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 19:10:15,146] loop.py:45 INFO TRAIN_STEP step=13767 loss=1.350529 Training: 76%|███████▋ | 13767/18018 [29:40:01<8:07:31, 6.88s/step, loss=1.351, acc=0.691, acc_len=1.64, thru=17.8, I=23.2, T=38.7, wait=3.3s, pool=53, epoch=2/2][2026-10-19 19:10:22,148] loop.py:45 INFO TRAIN_STEP step=13768 loss=1.222932 Training: 76%|███████▋ | 13768/18018 [29:40:08<8:09:54, 6.92s/step, loss=1.223, acc=0.713, acc_len=1.73, thru=18.7, I=19.0, T=37.2, wait=3.6s, pool=57, epoch=2/2][2026-10-19 19:10:29,493] loop.py:45 INFO TRAIN_STEP step=13769 loss=1.330897 Training: 76%|███████▋ | 13769/18018 [29:40:15<8:18:54, 7.05s/step, loss=1.331, acc=0.696, acc_len=1.67, thru=18.4, I=17.7, T=34.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:10:29,441] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:10:29,942] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:10:38,666] loop.py:380 INFO TIMING step=13770: step=3.884s data=1.749s compute=3.837s [fwd=2.708s bwd=1.114s opt=0.003s] dispatch=5.282s [2026-10-19 19:10:38,666] loop.py:45 INFO TRAIN_STEP step=13770 loss=1.265701 Training: 76%|███████▋ | 13770/18018 [29:40:24<9:04:00, 7.68s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=12.7, I=14.6, T=33.0, wait=5.3s, pool=45, epoch=2/2][2026-10-19 19:10:47,228] loop.py:45 INFO TRAIN_STEP step=13771 loss=1.271668 Training: 76%|███████▋ | 13771/18018 [29:40:33<9:22:31, 7.95s/step, loss=1.272, acc=0.705, acc_len=1.69, thru=15.6, I=14.7, T=34.9, wait=4.9s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:10:47,223] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13770: forward=2608.4ms backward=1013.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:10:55,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:10:55,819] loop.py:45 INFO TRAIN_STEP step=13772 loss=1.175352 Training: 76%|███████▋ | 13772/18018 [29:40:41<9:36:05, 8.14s/step, loss=1.175, acc=0.725, acc_len=1.76, thru=17.2, I=16.1, T=36.2, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:10:56,300] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:03,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:11:03,561] loop.py:45 INFO TRAIN_STEP step=13773 loss=1.363554 Training: 76%|███████▋ | 13773/18018 [29:40:49<9:27:27, 8.02s/step, loss=1.364, acc=0.686, acc_len=1.63, thru=17.1, I=18.8, T=35.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:04,048] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:08,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:11:09,820] loop.py:45 INFO TRAIN_STEP step=13774 loss=1.265634 Training: 76%|███████▋ | 13774/18018 [29:40:55<8:49:59, 7.49s/step, loss=1.266, acc=0.703, acc_len=1.69, thru=19.7, I=23.6, T=37.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:10,164] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:11:15,637] loop.py:380 INFO TIMING step=13775: step=3.358s data=1.402s compute=3.330s [fwd=2.384s bwd=0.931s opt=0.003s] dispatch=2.450s [2026-10-19 19:11:15,638] loop.py:45 INFO TRAIN_STEP step=13775 loss=1.259209 Training: 76%|███████▋ | 13775/18018 [29:41:01<8:14:16, 6.99s/step, loss=1.259, acc=0.706, acc_len=1.70, thru=20.2, I=23.4, T=38.1, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:22,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:11:24,365] loop.py:45 INFO TRAIN_STEP step=13776 loss=1.310308 Training: 76%|███████▋ | 13776/18018 [29:41:10<8:51:01, 7.51s/step, loss=1.310, acc=0.701, acc_len=1.69, thru=16.9, I=20.4, T=32.6, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:11:24,360] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13775: forward=2656.4ms backward=1166.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:24,789] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 19:11:30,723] loop.py:45 INFO TRAIN_STEP step=13777 loss=1.209690 Training: 76%|███████▋ | 13777/18018 [29:41:16<8:26:27, 7.17s/step, loss=1.210, acc=0.717, acc_len=1.73, thru=17.6, I=25.3, T=37.6, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:36,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:11:37,671] loop.py:45 INFO TRAIN_STEP step=13778 loss=1.306851 Training: 76%|███████▋ | 13778/18018 [29:41:23<8:21:45, 7.10s/step, loss=1.307, acc=0.699, acc_len=1.67, thru=18.2, I=21.2, T=36.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:37,897] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:11:44,080] loop.py:45 INFO TRAIN_STEP step=13779 loss=1.145249 Training: 76%|███████▋ | 13779/18018 [29:41:30<8:07:00, 6.89s/step, loss=1.145, acc=0.725, acc_len=1.77, thru=18.2, I=26.4, T=34.8, wait=2.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:51,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:11:51,589] loop.py:380 INFO TIMING step=13780: step=3.656s data=1.539s compute=3.625s [fwd=2.621s bwd=0.989s opt=0.003s] dispatch=3.843s [2026-10-19 19:11:51,589] loop.py:45 INFO TRAIN_STEP step=13780 loss=1.366359 Training: 76%|███████▋ | 13780/18018 [29:41:37<8:19:53, 7.08s/step, loss=1.366, acc=0.685, acc_len=1.63, thru=16.9, I=18.2, T=35.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:51,847] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:58,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:11:59,450] loop.py:45 INFO TRAIN_STEP step=13781 loss=1.291320 Training: 76%|███████▋ | 13781/18018 [29:41:45<8:36:23, 7.31s/step, loss=1.291, acc=0.702, acc_len=1.68, thru=15.2, I=17.1, T=36.8, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:11:59,446] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13780: forward=2483.5ms backward=954.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:11:59,832] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:12:06,008] loop.py:45 INFO TRAIN_STEP step=13782 loss=1.204185 Training: 76%|███████▋ | 13782/18018 [29:41:52<8:20:23, 7.09s/step, loss=1.204, acc=0.716, acc_len=1.73, thru=17.6, I=19.0, T=34.7, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:11,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:12:12,831] loop.py:45 INFO TRAIN_STEP step=13783 loss=1.263156 Training: 76%|███████▋ | 13783/18018 [29:41:58<8:14:32, 7.01s/step, loss=1.263, acc=0.706, acc_len=1.70, thru=19.8, I=21.0, T=35.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:13,328] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:20,140] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:12:20,876] loop.py:45 INFO TRAIN_STEP step=13784 loss=1.192934 Training: 77%|███████▋ | 13784/18018 [29:42:06<8:36:25, 7.32s/step, loss=1.193, acc=0.722, acc_len=1.76, thru=14.5, I=19.5, T=36.5, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:21,146] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:12:27,582] loop.py:380 INFO TIMING step=13785: step=3.848s data=1.542s compute=3.807s [fwd=2.607s bwd=1.185s opt=0.003s] dispatch=2.849s [2026-10-19 19:12:27,582] loop.py:45 INFO TRAIN_STEP step=13785 loss=1.427469 Training: 77%|███████▋ | 13785/18018 [29:42:13<8:23:22, 7.14s/step, loss=1.427, acc=0.679, acc_len=1.61, thru=16.7, I=18.8, T=33.3, wait=2.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:34,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:12:34,213] loop.py:45 INFO TRAIN_STEP step=13786 loss=1.207700 Training: 77%|███████▋ | 13786/18018 [29:42:20<8:12:32, 6.98s/step, loss=1.208, acc=0.714, acc_len=1.73, thru=18.2, I=18.6, T=38.1, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:12:34,208] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13785: forward=2365.3ms backward=949.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:34,558] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:12:42,010] loop.py:45 INFO TRAIN_STEP step=13787 loss=1.277730 Training: 77%|███████▋ | 13787/18018 [29:42:28<8:29:39, 7.23s/step, loss=1.278, acc=0.702, acc_len=1.69, thru=16.3, I=17.7, T=36.2, wait=4.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:49,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:12:50,782] loop.py:45 INFO TRAIN_STEP step=13788 loss=1.252808 Training: 77%|███████▋ | 13788/18018 [29:42:36<9:02:12, 7.69s/step, loss=1.253, acc=0.712, acc_len=1.72, thru=15.3, I=17.1, T=32.2, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:50,811] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:56,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:12:57,646] loop.py:45 INFO TRAIN_STEP step=13789 loss=1.171632 Training: 77%|███████▋ | 13789/18018 [29:42:43<8:44:37, 7.44s/step, loss=1.172, acc=0.725, acc_len=1.77, thru=17.2, I=20.1, T=35.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:12:57,861] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:13:04,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:13:04,697] loop.py:380 INFO TIMING step=13790: step=3.522s data=1.486s compute=3.491s [fwd=2.515s bwd=0.962s opt=0.003s] dispatch=3.518s [2026-10-19 19:13:04,698] loop.py:45 INFO TRAIN_STEP step=13790 loss=1.411064 Training: 77%|███████▋ | 13790/18018 [29:42:50<8:36:14, 7.33s/step, loss=1.411, acc=0.679, acc_len=1.61, thru=17.7, I=20.0, T=36.3, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:13:05,065] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:13:10,696] loop.py:45 INFO TRAIN_STEP step=13791 loss=1.238772 Training: 77%|███████▋ | 13791/18018 [29:42:56<8:08:01, 6.93s/step, loss=1.239, acc=0.712, acc_len=1.72, thru=19.4, I=20.9, T=37.0, wait=2.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:13:10,692] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13790: forward=2459.7ms backward=960.4ms [2026-10-19 19:13:18,037] loop.py:45 INFO TRAIN_STEP step=13792 loss=1.269826 Training: 77%|███████▋ | 13792/18018 [29:43:04<8:16:39, 7.05s/step, loss=1.270, acc=0.704, acc_len=1.69, thru=16.1, I=16.4, T=36.0, wait=3.8s, pool=53, epoch=2/2][2026-10-19 19:13:24,413] loop.py:45 INFO TRAIN_STEP step=13793 loss=1.263642 Training: 77%|███████▋ | 13793/18018 [29:43:10<8:02:17, 6.85s/step, loss=1.264, acc=0.708, acc_len=1.71, thru=18.5, I=21.3, T=36.7, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:13:34,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:13:35,070] loop.py:45 INFO TRAIN_STEP step=13794 loss=1.157835 Training: 77%|███████▋ | 13794/18018 [29:43:21<9:22:34, 7.99s/step, loss=1.158, acc=0.730, acc_len=1.78, thru=14.9, I=12.9, T=32.3, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:13:35,227] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 19:13:43,852] loop.py:380 INFO TIMING step=13795: step=3.681s data=1.398s compute=3.646s [fwd=2.618s bwd=1.014s opt=0.003s] dispatch=5.091s [2026-10-19 19:13:43,852] loop.py:45 INFO TRAIN_STEP step=13795 loss=1.327597 Training: 77%|███████▋ | 13795/18018 [29:43:29<9:39:08, 8.23s/step, loss=1.328, acc=0.699, acc_len=1.68, thru=12.2, I=14.3, T=34.8, wait=5.1s, pool=49, epoch=2/2][2026-10-19 19:13:51,322] loop.py:45 INFO TRAIN_STEP step=13796 loss=1.187302 Training: 77%|███████▋ | 13796/18018 [29:43:37<9:23:05, 8.00s/step, loss=1.187, acc=0.721, acc_len=1.75, thru=14.5, I=18.5, T=35.4, wait=3.8s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:13:51,318] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13795: forward=2403.6ms backward=1073.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:13:58,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:13:59,776] loop.py:45 INFO TRAIN_STEP step=13797 loss=1.286339 Training: 77%|███████▋ | 13797/18018 [29:43:45<9:32:25, 8.14s/step, loss=1.286, acc=0.705, acc_len=1.69, thru=17.2, I=16.9, T=35.8, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:13:59,872] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:05,492] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:14:06,264] loop.py:45 INFO TRAIN_STEP step=13798 loss=1.151562 Training: 77%|███████▋ | 13798/18018 [29:43:52<8:57:27, 7.64s/step, loss=1.152, acc=0.725, acc_len=1.77, thru=19.3, I=24.7, T=37.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:06,499] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:12,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:14:13,611] loop.py:45 INFO TRAIN_STEP step=13799 loss=1.308023 Training: 77%|███████▋ | 13799/18018 [29:43:59<8:51:07, 7.55s/step, loss=1.308, acc=0.701, acc_len=1.68, thru=17.8, I=20.5, T=36.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:13,994] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:14:20,449] loop.py:380 INFO TIMING step=13800: step=3.281s data=1.299s compute=3.242s [fwd=2.322s bwd=0.906s opt=0.003s] dispatch=3.548s [2026-10-19 19:14:20,449] loop.py:45 INFO TRAIN_STEP step=13800 loss=1.233796 Training: 77%|███████▋ | 13800/18018 [29:44:06<8:35:53, 7.34s/step, loss=1.234, acc=0.712, acc_len=1.72, thru=15.4, I=19.0, T=39.0, wait=3.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:28,499] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:14:29,532] loop.py:45 INFO TRAIN_STEP step=13801 loss=1.288040 Training: 77%|███████▋ | 13801/18018 [29:44:15<9:12:33, 7.86s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=17.6, I=17.8, T=36.1, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:14:29,528] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13800: forward=2538.1ms backward=966.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:30,008] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:14:36,049] loop.py:45 INFO TRAIN_STEP step=13802 loss=1.237556 Training: 77%|███████▋ | 13802/18018 [29:44:22<8:44:03, 7.46s/step, loss=1.238, acc=0.709, acc_len=1.71, thru=18.6, I=21.5, T=37.5, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:42,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:14:42,619] loop.py:45 INFO TRAIN_STEP step=13803 loss=1.398571 Training: 77%|███████▋ | 13803/18018 [29:44:28<8:25:15, 7.19s/step, loss=1.399, acc=0.689, acc_len=1.64, thru=19.8, I=21.1, T=37.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:43,042] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:50,009] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:14:50,350] loop.py:45 INFO TRAIN_STEP step=13804 loss=1.243258 Training: 77%|███████▋ | 13804/18018 [29:44:36<8:36:30, 7.35s/step, loss=1.243, acc=0.709, acc_len=1.70, thru=16.5, I=18.4, T=35.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:50,511] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:56,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:14:56,732] loop.py:380 INFO TIMING step=13805: step=3.410s data=1.394s compute=3.381s [fwd=2.445s bwd=0.921s opt=0.003s] dispatch=2.962s [2026-10-19 19:14:56,732] loop.py:45 INFO TRAIN_STEP step=13805 loss=1.173365 Training: 77%|███████▋ | 13805/18018 [29:44:42<8:15:50, 7.06s/step, loss=1.173, acc=0.723, acc_len=1.76, thru=19.6, I=20.5, T=37.5, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:14:57,110] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:01,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:15:02,926] loop.py:45 INFO TRAIN_STEP step=13806 loss=1.261061 Training: 77%|███████▋ | 13806/18018 [29:44:49<7:57:30, 6.80s/step, loss=1.261, acc=0.702, acc_len=1.68, thru=20.4, I=24.9, T=35.7, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:15:02,921] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13805: forward=2546.6ms backward=994.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:02,960] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:08,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:15:09,192] loop.py:45 INFO TRAIN_STEP step=13807 loss=1.234847 Training: 77%|███████▋ | 13807/18018 [29:44:55<7:46:04, 6.64s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=19.8, I=25.3, T=33.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:09,562] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:13,955] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:15:14,919] loop.py:45 INFO TRAIN_STEP step=13808 loss=1.268426 Training: 77%|███████▋ | 13808/18018 [29:45:01<7:26:46, 6.37s/step, loss=1.268, acc=0.707, acc_len=1.70, thru=20.9, I=26.8, T=38.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:14,972] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:21,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:15:21,760] loop.py:45 INFO TRAIN_STEP step=13809 loss=1.420864 Training: 77%|███████▋ | 13809/18018 [29:45:07<7:36:37, 6.51s/step, loss=1.421, acc=0.680, acc_len=1.61, thru=19.1, I=22.5, T=36.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:22,065] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:27,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:15:28,057] loop.py:380 INFO TIMING step=13810: step=3.449s data=1.365s compute=3.418s [fwd=2.446s bwd=0.958s opt=0.003s] dispatch=2.838s [2026-10-19 19:15:28,057] loop.py:45 INFO TRAIN_STEP step=13810 loss=1.286692 Training: 77%|███████▋ | 13810/18018 [29:45:14<7:32:00, 6.44s/step, loss=1.287, acc=0.699, acc_len=1.68, thru=20.3, I=21.2, T=37.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:28,332] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:34,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:15:35,017] loop.py:45 INFO TRAIN_STEP step=13811 loss=1.337083 Training: 77%|███████▋ | 13811/18018 [29:45:21<7:42:45, 6.60s/step, loss=1.337, acc=0.694, acc_len=1.66, thru=18.5, I=18.8, T=35.3, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:15:35,012] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13810: forward=2616.4ms backward=964.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:35,291] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:39,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:15:40,896] loop.py:45 INFO TRAIN_STEP step=13812 loss=1.360816 Training: 77%|███████▋ | 13812/18018 [29:45:26<7:27:30, 6.38s/step, loss=1.361, acc=0.690, acc_len=1.64, thru=20.8, I=23.6, T=36.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:40,948] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:48,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:15:49,567] loop.py:45 INFO TRAIN_STEP step=13813 loss=1.157779 Training: 77%|███████▋ | 13813/18018 [29:45:35<8:15:29, 7.07s/step, loss=1.158, acc=0.726, acc_len=1.77, thru=13.7, I=15.8, T=34.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:15:49,732] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:15:57,605] loop.py:45 INFO TRAIN_STEP step=13814 loss=1.303046 Training: 77%|███████▋ | 13814/18018 [29:45:43<8:35:40, 7.36s/step, loss=1.303, acc=0.700, acc_len=1.68, thru=12.4, I=14.3, T=37.1, wait=4.6s, pool=29, epoch=2/2][2026-10-19 19:16:05,978] loop.py:380 INFO TIMING step=13815: step=3.617s data=1.576s compute=3.583s [fwd=2.525s bwd=1.044s opt=0.003s] dispatch=4.748s [2026-10-19 19:16:05,978] loop.py:45 INFO TRAIN_STEP step=13815 loss=1.307250 Training: 77%|███████▋ | 13815/18018 [29:45:52<8:56:51, 7.66s/step, loss=1.307, acc=0.698, acc_len=1.67, thru=18.0, I=14.3, T=35.4, wait=4.7s, pool=61, epoch=2/2][2026-10-19 19:16:14,418] loop.py:45 INFO TRAIN_STEP step=13816 loss=1.167312 Training: 77%|███████▋ | 13816/18018 [29:46:00<9:13:03, 7.90s/step, loss=1.167, acc=0.724, acc_len=1.76, thru=15.3, I=14.3, T=33.9, wait=4.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:16:14,413] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13815: forward=2718.9ms backward=1005.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:16:20,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:16:20,930] loop.py:45 INFO TRAIN_STEP step=13817 loss=1.215686 Training: 77%|███████▋ | 13817/18018 [29:46:07<8:43:47, 7.48s/step, loss=1.216, acc=0.718, acc_len=1.74, thru=20.1, I=19.8, T=36.5, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:16:21,375] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:16:26,874] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:16:27,675] loop.py:45 INFO TRAIN_STEP step=13818 loss=1.394609 Training: 77%|███████▋ | 13818/18018 [29:46:13<8:28:16, 7.26s/step, loss=1.395, acc=0.683, acc_len=1.62, thru=18.8, I=21.8, T=37.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:16:27,879] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:16:33,157] loop.py:45 INFO TRAIN_STEP step=13819 loss=1.196477 Training: 77%|███████▋ | 13819/18018 [29:46:19<7:50:46, 6.73s/step, loss=1.196, acc=0.716, acc_len=1.73, thru=20.9, I=26.4, T=39.8, wait=2.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:16:41,565] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:16:41,926] loop.py:380 INFO TIMING step=13820: step=3.684s data=1.558s compute=3.641s [fwd=2.536s bwd=1.091s opt=0.003s] dispatch=5.075s [2026-10-19 19:16:41,926] loop.py:45 INFO TRAIN_STEP step=13820 loss=1.224563 Training: 77%|███████▋ | 13820/18018 [29:46:28<8:33:31, 7.34s/step, loss=1.225, acc=0.716, acc_len=1.73, thru=14.9, I=14.9, T=34.7, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:16:42,078] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 19:16:47,973] loop.py:45 INFO TRAIN_STEP step=13821 loss=1.211232 Training: 77%|███████▋ | 13821/18018 [29:46:34<8:06:19, 6.95s/step, loss=1.211, acc=0.718, acc_len=1.75, thru=20.1, I=21.1, T=38.9, wait=2.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:16:47,969] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13820: forward=2338.8ms backward=903.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:16:55,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:16:55,583] loop.py:45 INFO TRAIN_STEP step=13822 loss=1.310130 Training: 77%|███████▋ | 13822/18018 [29:46:41<8:19:58, 7.15s/step, loss=1.310, acc=0.693, acc_len=1.65, thru=16.8, I=18.0, T=35.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:16:55,814] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:17:02,164] loop.py:45 INFO TRAIN_STEP step=13823 loss=1.314314 Training: 77%|███████▋ | 13823/18018 [29:46:48<8:07:53, 6.98s/step, loss=1.314, acc=0.696, acc_len=1.66, thru=18.7, I=19.4, T=37.1, wait=3.1s, pool=57, epoch=2/2][2026-10-19 19:17:07,723] loop.py:45 INFO TRAIN_STEP step=13824 loss=1.230140 Training: 77%|███████▋ | 13824/18018 [29:46:53<7:38:01, 6.55s/step, loss=1.230, acc=0.712, acc_len=1.73, thru=21.3, I=22.9, T=36.0, wait=2.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:17:13,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:17:14,487] loop.py:380 INFO TIMING step=13825: step=3.589s data=1.536s compute=3.548s [fwd=2.506s bwd=1.028s opt=0.003s] dispatch=3.167s [2026-10-19 19:17:14,487] loop.py:45 INFO TRAIN_STEP step=13825 loss=1.289415 Training: 77%|███████▋ | 13825/18018 [29:47:00<7:42:22, 6.62s/step, loss=1.289, acc=0.702, acc_len=1.68, thru=18.8, I=21.2, T=35.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:17:14,821] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:17:20,863] loop.py:45 INFO TRAIN_STEP step=13826 loss=1.339321 Training: 77%|███████▋ | 13826/18018 [29:47:06<7:37:13, 6.54s/step, loss=1.339, acc=0.690, acc_len=1.64, thru=18.8, I=21.2, T=36.5, wait=2.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:17:20,859] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13825: forward=2413.4ms backward=1040.0ms [2026-10-19 19:17:27,988] loop.py:45 INFO TRAIN_STEP step=13827 loss=1.258962 Training: 77%|███████▋ | 13827/18018 [29:47:14<7:49:16, 6.72s/step, loss=1.259, acc=0.709, acc_len=1.71, thru=18.5, I=18.6, T=36.1, wait=3.6s, pool=61, epoch=2/2][2026-10-19 19:17:35,680] loop.py:45 INFO TRAIN_STEP step=13828 loss=1.285804 Training: 77%|███████▋ | 13828/18018 [29:47:21<8:09:31, 7.01s/step, loss=1.286, acc=0.702, acc_len=1.69, thru=16.9, I=17.1, T=36.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:17:35,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:17:36,130] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:17:42,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:17:43,818] loop.py:45 INFO TRAIN_STEP step=13829 loss=1.279857 Training: 77%|███████▋ | 13829/18018 [29:47:29<8:33:03, 7.35s/step, loss=1.280, acc=0.700, acc_len=1.68, thru=16.3, I=18.5, T=36.0, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:17:44,225] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:17:49,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:17:50,041] loop.py:380 INFO TIMING step=13830: step=3.281s data=1.330s compute=3.244s [fwd=2.309s bwd=0.919s opt=0.003s] dispatch=2.933s [2026-10-19 19:17:50,041] loop.py:45 INFO TRAIN_STEP step=13830 loss=1.182758 Training: 77%|███████▋ | 13830/18018 [29:47:36<8:09:24, 7.01s/step, loss=1.183, acc=0.720, acc_len=1.74, thru=19.9, I=24.8, T=39.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:17:50,252] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:17:56,768] loop.py:45 INFO TRAIN_STEP step=13831 loss=1.295141 Training: 77%|███████▋ | 13831/18018 [29:47:42<8:03:18, 6.93s/step, loss=1.295, acc=0.702, acc_len=1.69, thru=18.5, I=22.5, T=35.5, wait=3.1s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:17:56,763] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13830: forward=2538.9ms backward=1026.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:02,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:18:03,391] loop.py:45 INFO TRAIN_STEP step=13832 loss=1.364559 Training: 77%|███████▋ | 13832/18018 [29:47:49<7:56:52, 6.84s/step, loss=1.365, acc=0.686, acc_len=1.63, thru=19.3, I=20.4, T=37.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:03,737] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:18:11,011] loop.py:45 INFO TRAIN_STEP step=13833 loss=1.301993 Training: 77%|███████▋ | 13833/18018 [29:47:57<8:13:07, 7.07s/step, loss=1.302, acc=0.697, acc_len=1.67, thru=13.9, I=14.8, T=37.3, wait=4.2s, pool=41, epoch=2/2][2026-10-19 19:18:19,703] loop.py:45 INFO TRAIN_STEP step=13834 loss=1.235248 Training: 77%|███████▋ | 13834/18018 [29:48:05<8:46:58, 7.56s/step, loss=1.235, acc=0.713, acc_len=1.74, thru=17.0, I=15.4, T=34.8, wait=5.0s, pool=57, epoch=2/2][2026-10-19 19:18:29,453] loop.py:380 INFO TIMING step=13835: step=3.613s data=1.461s compute=3.583s [fwd=2.543s bwd=1.026s opt=0.003s] dispatch=6.129s [2026-10-19 19:18:29,454] loop.py:45 INFO TRAIN_STEP step=13835 loss=1.367371 Training: 77%|███████▋ | 13835/18018 [29:48:15<9:32:42, 8.21s/step, loss=1.367, acc=0.685, acc_len=1.63, thru=13.3, I=13.3, T=35.4, wait=6.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:36,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:18:36,877] loop.py:45 INFO TRAIN_STEP step=13836 loss=1.294420 Training: 77%|███████▋ | 13836/18018 [29:48:22<9:16:04, 7.98s/step, loss=1.294, acc=0.701, acc_len=1.68, thru=18.4, I=17.2, T=35.4, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:18:36,871] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13835: forward=2550.7ms backward=1019.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:37,068] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:44,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:18:44,725] loop.py:45 INFO TRAIN_STEP step=13837 loss=1.420334 Training: 77%|███████▋ | 13837/18018 [29:48:30<9:13:10, 7.94s/step, loss=1.420, acc=0.676, acc_len=1.60, thru=16.1, I=18.7, T=36.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:45,015] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:50,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:18:50,947] loop.py:45 INFO TRAIN_STEP step=13838 loss=1.244887 Training: 77%|███████▋ | 13838/18018 [29:48:37<8:37:11, 7.42s/step, loss=1.245, acc=0.706, acc_len=1.69, thru=20.1, I=22.6, T=34.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:51,114] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:57,499] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:18:57,610] loop.py:45 INFO TRAIN_STEP step=13839 loss=1.237231 Training: 77%|███████▋ | 13839/18018 [29:48:43<8:21:08, 7.20s/step, loss=1.237, acc=0.709, acc_len=1.71, thru=18.2, I=20.3, T=36.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:18:58,001] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:19:02,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:19:04,237] loop.py:380 INFO TIMING step=13840: step=3.519s data=1.498s compute=3.484s [fwd=2.500s bwd=0.970s opt=0.003s] dispatch=3.099s [2026-10-19 19:19:04,237] loop.py:45 INFO TRAIN_STEP step=13840 loss=1.224263 Training: 77%|███████▋ | 13840/18018 [29:48:50<8:09:09, 7.02s/step, loss=1.224, acc=0.715, acc_len=1.73, thru=19.3, I=22.5, T=36.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:19:04,281] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 19:19:11,238] loop.py:45 INFO TRAIN_STEP step=13841 loss=1.187001 Training: 77%|███████▋ | 13841/18018 [29:48:57<8:08:31, 7.02s/step, loss=1.187, acc=0.721, acc_len=1.76, thru=15.9, I=18.8, T=34.1, wait=3.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:19:11,234] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13840: forward=2593.2ms backward=1105.8ms [2026-10-19 19:19:18,457] loop.py:45 INFO TRAIN_STEP step=13842 loss=1.237072 Training: 77%|███████▋ | 13842/18018 [29:49:04<8:12:36, 7.08s/step, loss=1.237, acc=0.710, acc_len=1.72, thru=15.9, I=14.9, T=38.0, wait=3.8s, pool=45, epoch=2/2][2026-10-19 19:19:26,092] loop.py:45 INFO TRAIN_STEP step=13843 loss=1.247960 Training: 77%|███████▋ | 13843/18018 [29:49:12<8:24:08, 7.25s/step, loss=1.248, acc=0.708, acc_len=1.71, thru=17.5, I=15.7, T=35.1, wait=4.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:19:32,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:19:32,609] loop.py:45 INFO TRAIN_STEP step=13844 loss=1.200264 Training: 77%|███████▋ | 13844/18018 [29:49:18<8:08:50, 7.03s/step, loss=1.200, acc=0.717, acc_len=1.74, thru=20.3, I=18.3, T=38.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:19:32,791] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:19:40,636] loop.py:380 INFO TIMING step=13845: step=3.699s data=1.588s compute=3.669s [fwd=2.633s bwd=1.022s opt=0.003s] dispatch=4.319s [2026-10-19 19:19:40,636] loop.py:45 INFO TRAIN_STEP step=13845 loss=1.264997 Training: 77%|███████▋ | 13845/18018 [29:49:26<8:29:39, 7.33s/step, loss=1.265, acc=0.708, acc_len=1.71, thru=12.5, I=14.6, T=34.6, wait=4.3s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:19:47,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:19:48,670] loop.py:45 INFO TRAIN_STEP step=13846 loss=1.274099 Training: 77%|███████▋ | 13846/18018 [29:49:34<8:44:14, 7.54s/step, loss=1.274, acc=0.706, acc_len=1.70, thru=17.6, I=16.4, T=37.6, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:19:48,666] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13845: forward=2377.0ms backward=975.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:19:48,782] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 19:19:55,553] loop.py:45 INFO TRAIN_STEP step=13847 loss=1.178836 Training: 77%|███████▋ | 13847/18018 [29:49:41<8:30:20, 7.34s/step, loss=1.179, acc=0.721, acc_len=1.75, thru=18.1, I=21.2, T=35.1, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:03,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:20:04,729] loop.py:45 INFO TRAIN_STEP step=13848 loss=1.193094 Training: 77%|███████▋ | 13848/18018 [29:49:50<9:08:29, 7.89s/step, loss=1.193, acc=0.721, acc_len=1.75, thru=14.4, I=16.3, T=32.9, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:05,213] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:09,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:20:10,285] loop.py:45 INFO TRAIN_STEP step=13849 loss=1.164201 Training: 77%|███████▋ | 13849/18018 [29:49:56<8:19:42, 7.19s/step, loss=1.164, acc=0.727, acc_len=1.78, thru=20.0, I=26.3, T=40.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:10,735] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:16,906] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:20:17,362] loop.py:380 INFO TIMING step=13850: step=3.367s data=1.365s compute=3.339s [fwd=2.377s bwd=0.948s opt=0.003s] dispatch=3.700s [2026-10-19 19:20:17,362] loop.py:45 INFO TRAIN_STEP step=13850 loss=1.349051 Training: 77%|███████▋ | 13850/18018 [29:50:03<8:17:12, 7.16s/step, loss=1.349, acc=0.690, acc_len=1.64, thru=18.9, I=22.2, T=38.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:17,416] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:23,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:20:23,684] loop.py:45 INFO TRAIN_STEP step=13851 loss=1.315565 Training: 77%|███████▋ | 13851/18018 [29:50:09<7:59:38, 6.91s/step, loss=1.316, acc=0.698, acc_len=1.67, thru=20.2, I=20.5, T=36.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:23,749] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (TrainerActor pid=9310) [2026-10-19 19:20:23,678] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13850: forward=2459.5ms backward=968.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:30,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:20:32,490] loop.py:45 INFO TRAIN_STEP step=13852 loss=1.206213 Training: 77%|███████▋ | 13852/18018 [29:50:18<8:39:07, 7.48s/step, loss=1.206, acc=0.717, acc_len=1.74, thru=14.8, I=17.6, T=33.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:32,881] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:39,975] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:20:41,301] loop.py:45 INFO TRAIN_STEP step=13853 loss=1.344100 Training: 77%|███████▋ | 13853/18018 [29:50:27<9:06:45, 7.88s/step, loss=1.344, acc=0.694, acc_len=1.66, thru=15.2, I=17.1, T=28.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:41,483] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:20:46,617] loop.py:45 INFO TRAIN_STEP step=13854 loss=1.334746 Training: 77%|███████▋ | 13854/18018 [29:50:32<8:13:17, 7.11s/step, loss=1.335, acc=0.695, acc_len=1.66, thru=18.2, I=20.4, T=39.2, wait=2.0s, pool=49, epoch=2/2][2026-10-19 19:20:54,070] loop.py:380 INFO TIMING step=13855: step=3.459s data=1.409s compute=3.432s [fwd=2.513s bwd=0.904s opt=0.003s] dispatch=3.984s [2026-10-19 19:20:54,070] loop.py:45 INFO TRAIN_STEP step=13855 loss=1.221053 Training: 77%|███████▋ | 13855/18018 [29:50:40<8:20:21, 7.21s/step, loss=1.221, acc=0.716, acc_len=1.74, thru=18.2, I=17.5, T=37.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:54,071] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:20:54,572] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:21:01,334] loop.py:45 INFO TRAIN_STEP step=13856 loss=1.244361 Training: 77%|███████▋ | 13856/18018 [29:50:47<8:21:20, 7.23s/step, loss=1.244, acc=0.708, acc_len=1.70, thru=18.6, I=17.9, T=36.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:01,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 19:21:01,330] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13855: forward=2521.7ms backward=981.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:01,794] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:07,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:21:08,084] loop.py:45 INFO TRAIN_STEP step=13857 loss=1.353372 Training: 77%|███████▋ | 13857/18018 [29:50:54<8:11:20, 7.08s/step, loss=1.353, acc=0.691, acc_len=1.65, thru=19.0, I=20.6, T=36.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:08,268] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:14,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:21:15,154] loop.py:45 INFO TRAIN_STEP step=13858 loss=1.196967 Training: 77%|███████▋ | 13858/18018 [29:51:01<8:10:54, 7.08s/step, loss=1.197, acc=0.719, acc_len=1.75, thru=18.2, I=22.3, T=36.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:15,354] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:21:23,773] loop.py:45 INFO TRAIN_STEP step=13859 loss=1.149875 Training: 77%|███████▋ | 13859/18018 [29:51:09<8:42:45, 7.54s/step, loss=1.150, acc=0.728, acc_len=1.78, thru=13.7, I=15.1, T=36.3, wait=5.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:29,744] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:21:30,613] loop.py:380 INFO TIMING step=13860: step=3.622s data=1.525s compute=3.591s [fwd=2.592s bwd=0.984s opt=0.003s] dispatch=3.209s [2026-10-19 19:21:30,613] loop.py:45 INFO TRAIN_STEP step=13860 loss=1.325442 Training: 77%|███████▋ | 13860/18018 [29:51:16<8:28:04, 7.33s/step, loss=1.325, acc=0.695, acc_len=1.67, thru=19.0, I=18.6, T=35.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:30,760] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:36,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:21:36,917] loop.py:45 INFO TRAIN_STEP step=13861 loss=1.339036 Training: 77%|███████▋ | 13861/18018 [29:51:23<8:06:33, 7.02s/step, loss=1.339, acc=0.696, acc_len=1.67, thru=19.7, I=24.2, T=32.7, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:21:36,913] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13860: forward=2960.7ms backward=911.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:37,304] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:21:45,338] loop.py:45 INFO TRAIN_STEP step=13862 loss=1.313659 Training: 77%|███████▋ | 13862/18018 [29:51:31<8:35:28, 7.44s/step, loss=1.314, acc=0.695, acc_len=1.66, thru=14.8, I=15.3, T=37.9, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:45,324] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:45,827] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:51,340] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:21:52,452] loop.py:45 INFO TRAIN_STEP step=13863 loss=1.181374 Training: 77%|███████▋ | 13863/18018 [29:51:38<8:28:33, 7.34s/step, loss=1.181, acc=0.725, acc_len=1.77, thru=18.4, I=20.5, T=36.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:52,846] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:59,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:21:59,498] loop.py:45 INFO TRAIN_STEP step=13864 loss=1.216048 Training: 77%|███████▋ | 13864/18018 [29:51:45<8:22:16, 7.25s/step, loss=1.216, acc=0.714, acc_len=1.73, thru=18.4, I=21.6, T=38.2, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:21:59,773] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:05,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:22:05,783] loop.py:380 INFO TIMING step=13865: step=3.471s data=1.443s compute=3.439s [fwd=2.505s bwd=0.919s opt=0.003s] dispatch=2.804s [2026-10-19 19:22:05,783] loop.py:45 INFO TRAIN_STEP step=13865 loss=1.352479 Training: 77%|███████▋ | 13865/18018 [29:51:51<8:01:58, 6.96s/step, loss=1.352, acc=0.693, acc_len=1.66, thru=20.5, I=22.0, T=36.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:06,236] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:12,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:22:12,826] loop.py:45 INFO TRAIN_STEP step=13866 loss=1.104535 Training: 77%|███████▋ | 13866/18018 [29:51:58<8:03:33, 6.99s/step, loss=1.105, acc=0.733, acc_len=1.80, thru=18.5, I=20.0, T=37.3, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:22:12,822] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13865: forward=2412.2ms backward=976.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:12,967] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:22:19,680] loop.py:45 INFO TRAIN_STEP step=13867 loss=1.216407 Training: 77%|███████▋ | 13867/18018 [29:52:05<8:00:37, 6.95s/step, loss=1.216, acc=0.712, acc_len=1.72, thru=18.0, I=21.4, T=36.8, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:26,660] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:22:27,777] loop.py:45 INFO TRAIN_STEP step=13868 loss=1.235230 Training: 77%|███████▋ | 13868/18018 [29:52:13<8:24:24, 7.29s/step, loss=1.235, acc=0.709, acc_len=1.71, thru=16.2, I=17.4, T=36.4, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:28,170] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:34,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:22:34,890] loop.py:45 INFO TRAIN_STEP step=13869 loss=1.213318 Training: 77%|███████▋ | 13869/18018 [29:52:20<8:20:34, 7.24s/step, loss=1.213, acc=0.720, acc_len=1.75, thru=17.8, I=21.1, T=37.1, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:35,205] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:22:40,609] loop.py:380 INFO TIMING step=13870: step=3.290s data=1.321s compute=3.261s [fwd=2.361s bwd=0.886s opt=0.003s] dispatch=2.418s [2026-10-19 19:22:40,609] loop.py:45 INFO TRAIN_STEP step=13870 loss=1.361066 Training: 77%|███████▋ | 13870/18018 [29:52:26<7:48:52, 6.78s/step, loss=1.361, acc=0.689, acc_len=1.64, thru=21.2, I=21.9, T=38.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:40,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:22:41,094] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:22:49,024] loop.py:45 INFO TRAIN_STEP step=13871 loss=1.292473 Training: 77%|███████▋ | 13871/18018 [29:52:35<8:22:39, 7.27s/step, loss=1.292, acc=0.700, acc_len=1.68, thru=11.3, I=15.6, T=28.9, wait=4.0s, pool=21, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:22:49,020] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13870: forward=3419.7ms backward=959.4ms [2026-10-19 19:22:56,877] loop.py:45 INFO TRAIN_STEP step=13872 loss=1.228942 Training: 77%|███████▋ | 13872/18018 [29:52:42<8:34:31, 7.45s/step, loss=1.229, acc=0.712, acc_len=1.72, thru=22.4, I=22.1, T=38.6, wait=4.5s, pool=57, epoch=2/2][2026-10-19 19:23:04,047] loop.py:45 INFO TRAIN_STEP step=13873 loss=1.302421 Training: 77%|███████▋ | 13873/18018 [29:52:50<8:28:41, 7.36s/step, loss=1.302, acc=0.700, acc_len=1.68, thru=16.8, I=17.0, T=38.2, wait=3.8s, pool=53, epoch=2/2][2026-10-19 19:23:11,403] loop.py:45 INFO TRAIN_STEP step=13874 loss=1.232391 Training: 77%|███████▋ | 13874/18018 [29:52:57<8:28:31, 7.36s/step, loss=1.232, acc=0.710, acc_len=1.72, thru=18.5, I=16.7, T=36.6, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:23:17,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:23:18,599] loop.py:380 INFO TIMING step=13875: step=3.914s data=1.783s compute=3.873s [fwd=2.751s bwd=1.108s opt=0.003s] dispatch=3.269s [2026-10-19 19:23:18,600] loop.py:45 INFO TRAIN_STEP step=13875 loss=1.256150 Training: 77%|███████▋ | 13875/18018 [29:53:04<8:24:53, 7.31s/step, loss=1.256, acc=0.706, acc_len=1.70, thru=19.5, I=18.9, T=32.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:23:18,881] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:23:23,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:23:24,591] loop.py:45 INFO TRAIN_STEP step=13876 loss=1.254098 Training: 77%|███████▋ | 13876/18018 [29:53:10<7:57:24, 6.92s/step, loss=1.254, acc=0.706, acc_len=1.70, thru=20.0, I=24.5, T=37.2, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:23:24,586] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13875: forward=2420.5ms backward=975.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:23:24,881] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:23:30,305] loop.py:45 INFO TRAIN_STEP step=13877 loss=1.265391 Training: 77%|███████▋ | 13877/18018 [29:53:16<7:32:24, 6.55s/step, loss=1.265, acc=0.711, acc_len=1.72, thru=21.5, I=24.7, T=38.2, wait=2.4s, pool=61, epoch=2/2][2026-10-19 19:23:37,361] loop.py:45 INFO TRAIN_STEP step=13878 loss=1.279440 Training: 77%|███████▋ | 13878/18018 [29:53:23<7:42:38, 6.70s/step, loss=1.279, acc=0.705, acc_len=1.70, thru=18.8, I=19.8, T=35.1, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:23:43,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:23:44,204] loop.py:45 INFO TRAIN_STEP step=13879 loss=1.293808 Training: 77%|███████▋ | 13879/18018 [29:53:30<7:45:23, 6.75s/step, loss=1.294, acc=0.696, acc_len=1.67, thru=18.2, I=19.2, T=37.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:23:44,384] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:23:50,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:23:51,300] loop.py:380 INFO TIMING step=13880: step=3.412s data=1.389s compute=3.384s [fwd=2.451s bwd=0.919s opt=0.003s] dispatch=3.676s [2026-10-19 19:23:51,300] loop.py:45 INFO TRAIN_STEP step=13880 loss=1.301969 Training: 77%|███████▋ | 13880/18018 [29:53:37<7:52:31, 6.85s/step, loss=1.302, acc=0.700, acc_len=1.68, thru=19.0, I=21.0, T=37.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:23:51,593] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:23:57,825] loop.py:45 INFO TRAIN_STEP step=13881 loss=1.242152 Training: 77%|███████▋ | 13881/18018 [29:53:43<7:45:40, 6.75s/step, loss=1.242, acc=0.710, acc_len=1.72, thru=18.9, I=22.2, T=38.5, wait=3.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:23:57,821] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13880: forward=2368.1ms backward=916.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:24:04,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:24:04,541] loop.py:45 INFO TRAIN_STEP step=13882 loss=1.299003 Training: 77%|███████▋ | 13882/18018 [29:53:50<7:44:45, 6.74s/step, loss=1.299, acc=0.698, acc_len=1.67, thru=19.3, I=20.9, T=37.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:24:04,732] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:24:12,094] loop.py:45 INFO TRAIN_STEP step=13883 loss=1.192783 Training: 77%|███████▋ | 13883/18018 [29:53:58<8:01:27, 6.99s/step, loss=1.193, acc=0.718, acc_len=1.74, thru=14.4, I=17.6, T=36.3, wait=4.0s, pool=45, epoch=2/2][2026-10-19 19:24:23,994] loop.py:45 INFO TRAIN_STEP step=13884 loss=1.252970 Training: 77%|███████▋ | 13884/18018 [29:54:10<9:42:52, 8.46s/step, loss=1.253, acc=0.707, acc_len=1.71, thru=11.4, I=10.0, T=35.7, wait=8.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:24:34,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:24:34,764] loop.py:380 INFO TIMING step=13885: step=4.329s data=1.413s compute=4.289s [fwd=3.141s bwd=1.132s opt=0.003s] dispatch=6.432s [2026-10-19 19:24:34,765] loop.py:45 INFO TRAIN_STEP step=13885 loss=1.270342 Training: 77%|███████▋ | 13885/18018 [29:54:20<10:30:29, 9.15s/step, loss=1.270, acc=0.709, acc_len=1.72, thru=14.2, I=14.7, T=29.6, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:24:34,854] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:24:41,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:24:43,782] loop.py:45 INFO TRAIN_STEP step=13886 loss=1.215209 Training: 77%|███████▋ | 13886/18018 [29:54:29<10:27:30, 9.11s/step, loss=1.215, acc=0.711, acc_len=1.72, thru=13.8, I=16.3, T=31.5, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:24:43,777] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13885: forward=2816.3ms backward=1196.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:24:44,183] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 19:24:50,262] loop.py:45 INFO TRAIN_STEP step=13887 loss=1.299928 Training: 77%|███████▋ | 13887/18018 [29:54:36<9:33:00, 8.32s/step, loss=1.300, acc=0.698, acc_len=1.67, thru=17.1, I=23.0, T=38.2, wait=3.1s, pool=57, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:24:57,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:24:57,508] loop.py:45 INFO TRAIN_STEP step=13888 loss=1.238680 Training: 77%|███████▋ | 13888/18018 [29:54:43<9:10:37, 8.00s/step, loss=1.239, acc=0.716, acc_len=1.73, thru=18.6, I=18.7, T=36.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:24:57,766] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:25:06,074] loop.py:45 INFO TRAIN_STEP step=13889 loss=1.328904 Training: 77%|███████▋ | 13889/18018 [29:54:52<9:22:13, 8.17s/step, loss=1.329, acc=0.694, acc_len=1.65, thru=14.9, I=15.5, T=33.1, wait=4.7s, pool=57, epoch=2/2][2026-10-19 19:25:13,184] loop.py:380 INFO TIMING step=13890: step=3.510s data=1.440s compute=3.478s [fwd=2.481s bwd=0.983s opt=0.003s] dispatch=3.591s [2026-10-19 19:25:13,185] loop.py:45 INFO TRAIN_STEP step=13890 loss=1.382141 Training: 77%|███████▋ | 13890/18018 [29:54:59<9:00:13, 7.85s/step, loss=1.382, acc=0.686, acc_len=1.63, thru=16.4, I=15.7, T=36.5, wait=3.6s, pool=41, epoch=2/2][2026-10-19 19:25:20,724] loop.py:45 INFO TRAIN_STEP step=13891 loss=1.210346 Training: 77%|███████▋ | 13891/18018 [29:55:06<8:53:40, 7.76s/step, loss=1.210, acc=0.720, acc_len=1.75, thru=16.3, I=18.1, T=35.6, wait=3.9s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:25:20,719] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13890: forward=2536.0ms backward=1002.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:25:30,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:25:30,682] loop.py:45 INFO TRAIN_STEP step=13892 loss=1.241543 Training: 77%|███████▋ | 13892/18018 [29:55:16<9:38:52, 8.42s/step, loss=1.242, acc=0.714, acc_len=1.72, thru=14.9, I=13.7, T=34.8, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:25:30,758] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:25:35,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:25:36,631] loop.py:45 INFO TRAIN_STEP step=13893 loss=1.252322 Training: 77%|███████▋ | 13893/18018 [29:55:22<8:47:49, 7.68s/step, loss=1.252, acc=0.706, acc_len=1.70, thru=20.3, I=19.3, T=36.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:25:37,062] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:25:43,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:25:43,377] loop.py:45 INFO TRAIN_STEP step=13894 loss=1.112416 Training: 77%|███████▋ | 13894/18018 [29:55:29<8:28:27, 7.40s/step, loss=1.112, acc=0.730, acc_len=1.79, thru=18.6, I=22.4, T=38.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:25:43,731] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:25:48,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:25:49,526] loop.py:380 INFO TIMING step=13895: step=3.466s data=1.443s compute=3.438s [fwd=2.494s bwd=0.930s opt=0.003s] dispatch=2.675s [2026-10-19 19:25:49,526] loop.py:45 INFO TRAIN_STEP step=13895 loss=1.424490 Training: 77%|███████▋ | 13895/18018 [29:55:35<8:02:38, 7.02s/step, loss=1.424, acc=0.679, acc_len=1.61, thru=20.4, I=23.3, T=36.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:25:49,586] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 19:25:56,933] loop.py:45 INFO TRAIN_STEP step=13896 loss=1.220483 (TrainerActor pid=9310) [2026-10-19 19:25:56,929] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13895: forward=2372.7ms backward=948.9ms Training: 77%|███████▋ | 13896/18018 [29:55:43<8:10:24, 7.14s/step, loss=1.220, acc=0.715, acc_len=1.73, thru=14.7, I=16.5, T=38.0, wait=4.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:04,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:26:05,033] loop.py:45 INFO TRAIN_STEP step=13897 loss=1.306047 Training: 77%|███████▋ | 13897/18018 [29:55:51<8:30:06, 7.43s/step, loss=1.306, acc=0.700, acc_len=1.68, thru=17.6, I=17.7, T=33.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:05,277] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:12,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:26:12,869] loop.py:45 INFO TRAIN_STEP step=13898 loss=1.226306 Training: 77%|███████▋ | 13898/18018 [29:55:58<8:38:23, 7.55s/step, loss=1.226, acc=0.716, acc_len=1.73, thru=16.0, I=17.8, T=33.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:13,193] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:26:19,085] loop.py:45 INFO TRAIN_STEP step=13899 loss=1.239437 Training: 77%|███████▋ | 13899/18018 [29:56:05<8:10:47, 7.15s/step, loss=1.239, acc=0.708, acc_len=1.71, thru=16.8, I=20.5, T=36.1, wait=2.7s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:26,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:26:27,205] loop.py:380 INFO TIMING step=13900: step=3.582s data=1.511s compute=3.545s [fwd=2.557s bwd=0.973s opt=0.003s] dispatch=4.530s [2026-10-19 19:26:27,205] loop.py:45 INFO TRAIN_STEP step=13900 loss=1.254196 Training: 77%|███████▋ | 13900/18018 [29:56:13<8:30:38, 7.44s/step, loss=1.254, acc=0.708, acc_len=1.71, thru=17.7, I=16.9, T=35.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:27,545] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:26:36,003] loop.py:45 INFO TRAIN_STEP step=13901 loss=1.373925 Training: 77%|███████▋ | 13901/18018 [29:56:22<8:58:28, 7.85s/step, loss=1.374, acc=0.686, acc_len=1.63, thru=11.5, I=12.6, T=36.1, wait=5.2s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:26:35,999] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13900: forward=2459.3ms backward=1026.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:43,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:26:43,779] loop.py:45 INFO TRAIN_STEP step=13902 loss=1.215466 Training: 77%|███████▋ | 13902/18018 [29:56:29<8:56:54, 7.83s/step, loss=1.215, acc=0.718, acc_len=1.74, thru=18.6, I=17.8, T=36.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:44,064] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:50,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:26:51,580] loop.py:45 INFO TRAIN_STEP step=13903 loss=1.341194 Training: 77%|███████▋ | 13903/18018 [29:56:37<8:56:15, 7.82s/step, loss=1.341, acc=0.689, acc_len=1.64, thru=16.1, I=18.4, T=34.1, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:51,702] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:56,542] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:26:57,450] loop.py:45 INFO TRAIN_STEP step=13904 loss=1.366572 Training: 77%|███████▋ | 13904/18018 [29:56:43<8:15:59, 7.23s/step, loss=1.367, acc=0.687, acc_len=1.64, thru=20.6, I=24.0, T=37.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:26:57,560] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 19:27:06,803] loop.py:380 INFO TIMING step=13905: step=3.725s data=1.449s compute=3.695s [fwd=2.653s bwd=1.027s opt=0.003s] dispatch=5.620s [2026-10-19 19:27:06,803] loop.py:45 INFO TRAIN_STEP step=13905 loss=1.307304 Training: 77%|███████▋ | 13905/18018 [29:56:52<8:59:26, 7.87s/step, loss=1.307, acc=0.698, acc_len=1.68, thru=13.0, I=12.4, T=34.4, wait=5.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:13,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:27:14,124] loop.py:45 INFO TRAIN_STEP step=13906 loss=1.302753 Training: 77%|███████▋ | 13906/18018 [29:57:00<8:48:05, 7.71s/step, loss=1.303, acc=0.698, acc_len=1.68, thru=18.9, I=18.8, T=35.3, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:27:14,119] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13905: forward=2594.8ms backward=987.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:14,238] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:20,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:27:20,733] loop.py:45 INFO TRAIN_STEP step=13907 loss=1.196295 Training: 77%|███████▋ | 13907/18018 [29:57:06<8:25:22, 7.38s/step, loss=1.196, acc=0.720, acc_len=1.76, thru=19.2, I=21.8, T=37.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:21,148] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:25,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:27:27,204] loop.py:45 INFO TRAIN_STEP step=13908 loss=1.260770 Training: 77%|███████▋ | 13908/18018 [29:57:13<8:06:41, 7.10s/step, loss=1.261, acc=0.707, acc_len=1.70, thru=20.1, I=24.2, T=35.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:27,470] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:27:33,250] loop.py:45 INFO TRAIN_STEP step=13909 loss=1.222537 Training: 77%|███████▋ | 13909/18018 [29:57:19<7:44:48, 6.79s/step, loss=1.223, acc=0.715, acc_len=1.73, thru=20.1, I=25.7, T=38.1, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:40,567] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:27:41,247] loop.py:380 INFO TIMING step=13910: step=3.562s data=1.411s compute=3.534s [fwd=2.558s bwd=0.962s opt=0.003s] dispatch=4.425s [2026-10-19 19:27:41,247] loop.py:45 INFO TRAIN_STEP step=13910 loss=1.388805 Training: 77%|███████▋ | 13910/18018 [29:57:27<8:09:30, 7.15s/step, loss=1.389, acc=0.684, acc_len=1.62, thru=16.9, I=18.7, T=35.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:41,573] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:46,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:27:47,911] loop.py:45 INFO TRAIN_STEP step=13911 loss=1.254207 Training: 77%|███████▋ | 13911/18018 [29:57:33<7:59:25, 7.00s/step, loss=1.254, acc=0.712, acc_len=1.71, thru=18.7, I=22.5, T=36.5, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:27:47,906] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13910: forward=2493.7ms backward=953.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:27:48,378] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:27:55,405] loop.py:45 INFO TRAIN_STEP step=13912 loss=1.239413 Training: 77%|███████▋ | 13912/18018 [29:57:41<8:09:22, 7.15s/step, loss=1.239, acc=0.711, acc_len=1.72, thru=14.2, I=16.9, T=36.8, wait=4.0s, pool=41, epoch=2/2][2026-10-19 19:28:02,763] loop.py:45 INFO TRAIN_STEP step=13913 loss=1.284770 Training: 77%|███████▋ | 13913/18018 [29:57:48<8:13:32, 7.21s/step, loss=1.285, acc=0.699, acc_len=1.68, thru=17.9, I=17.0, T=36.1, wait=3.8s, pool=57, epoch=2/2][2026-10-19 19:28:12,442] loop.py:45 INFO TRAIN_STEP step=13914 loss=1.212461 Training: 77%|███████▋ | 13914/18018 [29:57:58<9:03:56, 7.95s/step, loss=1.212, acc=0.716, acc_len=1.74, thru=10.9, I=13.9, T=35.7, wait=6.1s, pool=29, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 19:28:12,535] eagle_store.py:517 WARNING batch_get_buffer missing keys (k445246-b9453064_hs, k445246-b9453064_ids, k445246-b9453064_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:28:20,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 19:28:12,501] eagle_store.py:517 WARNING batch_get_buffer missing keys (k445245-b0341a28_hs, k445245-b0341a28_ids, k445245-b0341a28_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 19:28:20,918] loop.py:380 INFO TIMING step=13915: step=3.613s data=1.480s compute=3.581s [fwd=2.594s bwd=0.973s opt=0.003s] dispatch=4.856s [2026-10-19 19:28:20,918] loop.py:45 INFO TRAIN_STEP step=13915 loss=1.225693 Training: 77%|███████▋ | 13915/18018 [29:58:07<9:14:36, 8.11s/step, loss=1.226, acc=0.713, acc_len=1.73, thru=19.4, I=15.1, T=35.4, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:28:21,075] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 19:28:26,973] loop.py:45 INFO TRAIN_STEP step=13916 loss=1.235822 Training: 77%|███████▋ | 13916/18018 [29:58:13<8:32:20, 7.49s/step, loss=1.236, acc=0.714, acc_len=1.73, thru=18.0, I=23.1, T=38.3, wait=2.7s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:28:26,968] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13915: forward=2381.1ms backward=909.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:28:34,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:28:35,064] loop.py:45 INFO TRAIN_STEP step=13917 loss=1.200998 Training: 77%|███████▋ | 13917/18018 [29:58:21<8:44:25, 7.67s/step, loss=1.201, acc=0.717, acc_len=1.74, thru=16.2, I=17.5, T=36.8, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:28:35,224] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:28:40,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:28:40,695] loop.py:45 INFO TRAIN_STEP step=13918 loss=1.310301 Training: 77%|███████▋ | 13918/18018 [29:58:26<8:02:25, 7.06s/step, loss=1.310, acc=0.695, acc_len=1.66, thru=20.5, I=23.7, T=37.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:28:40,949] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:28:46,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:28:48,441] loop.py:45 INFO TRAIN_STEP step=13919 loss=1.233519 Training: 77%|███████▋ | 13919/18018 [29:58:34<8:16:24, 7.27s/step, loss=1.234, acc=0.715, acc_len=1.73, thru=16.8, I=20.6, T=36.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:28:48,476] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 19:28:54,883] loop.py:380 INFO TIMING step=13920: step=3.421s data=1.484s compute=3.392s [fwd=2.432s bwd=0.946s opt=0.003s] dispatch=3.011s [2026-10-19 19:28:54,883] loop.py:45 INFO TRAIN_STEP step=13920 loss=1.278579 Training: 77%|███████▋ | 13920/18018 [29:58:40<7:59:22, 7.02s/step, loss=1.279, acc=0.704, acc_len=1.69, thru=18.7, I=23.9, T=37.4, wait=3.0s, pool=57, epoch=2/2][2026-10-19 19:29:02,929] loop.py:45 INFO TRAIN_STEP step=13921 loss=1.189258 Training: 77%|███████▋ | 13921/18018 [29:58:49<8:20:16, 7.33s/step, loss=1.189, acc=0.720, acc_len=1.75, thru=17.0, I=16.7, T=34.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:02,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 19:29:02,924] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13920: forward=2637.1ms backward=987.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:03,427] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:08,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:29:08,836] loop.py:45 INFO TRAIN_STEP step=13922 loss=1.220281 Training: 77%|███████▋ | 13922/18018 [29:58:54<7:51:06, 6.90s/step, loss=1.220, acc=0.712, acc_len=1.72, thru=20.4, I=22.9, T=37.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:09,233] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:15,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:29:15,579] loop.py:45 INFO TRAIN_STEP step=13923 loss=1.184875 Training: 77%|███████▋ | 13923/18018 [29:59:01<7:47:44, 6.85s/step, loss=1.185, acc=0.721, acc_len=1.76, thru=18.8, I=20.3, T=38.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:16,075] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:29:22,137] loop.py:45 INFO TRAIN_STEP step=13924 loss=1.288711 Training: 77%|███████▋ | 13924/18018 [29:59:08<7:41:33, 6.76s/step, loss=1.289, acc=0.703, acc_len=1.69, thru=18.1, I=18.1, T=37.5, wait=3.1s, pool=49, epoch=2/2][2026-10-19 19:29:30,179] loop.py:380 INFO TIMING step=13925: step=3.514s data=1.445s compute=3.485s [fwd=2.467s bwd=1.003s opt=0.003s] dispatch=4.520s [2026-10-19 19:29:30,179] loop.py:45 INFO TRAIN_STEP step=13925 loss=1.274516 Training: 77%|███████▋ | 13925/18018 [29:59:16<8:07:37, 7.15s/step, loss=1.275, acc=0.706, acc_len=1.70, thru=15.0, I=15.3, T=36.4, wait=4.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:37,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 19:29:37,811] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13925: forward=2430.0ms backward=951.0ms [2026-10-19 19:29:37,816] loop.py:45 INFO TRAIN_STEP step=13926 loss=1.280182 Training: 77%|███████▋ | 13926/18018 [29:59:23<8:17:31, 7.30s/step, loss=1.280, acc=0.705, acc_len=1.70, thru=19.0, I=16.6, T=37.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:38,088] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:29:44,318] loop.py:45 INFO TRAIN_STEP step=13927 loss=1.350116 Training: 77%|███████▋ | 13927/18018 [29:59:30<8:01:08, 7.06s/step, loss=1.350, acc=0.691, acc_len=1.65, thru=19.7, I=20.0, T=37.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:44,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:44,811] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:29:51,305] loop.py:45 INFO TRAIN_STEP step=13928 loss=1.172418 Training: 77%|███████▋ | 13928/18018 [29:59:37<7:59:36, 7.04s/step, loss=1.172, acc=0.723, acc_len=1.76, thru=19.2, I=19.0, T=37.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:51,300] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:51,801] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:58,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:29:58,828] loop.py:45 INFO TRAIN_STEP step=13929 loss=1.357149 Training: 77%|███████▋ | 13929/18018 [29:59:44<8:09:29, 7.18s/step, loss=1.357, acc=0.690, acc_len=1.64, thru=17.9, I=18.4, T=35.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:29:58,989] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:04,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:30:04,916] loop.py:380 INFO TIMING step=13930: step=3.392s data=1.406s compute=3.366s [fwd=2.429s bwd=0.922s opt=0.003s] dispatch=2.686s [2026-10-19 19:30:04,917] loop.py:45 INFO TRAIN_STEP step=13930 loss=1.282579 Training: 77%|███████▋ | 13930/18018 [29:59:51<7:47:01, 6.85s/step, loss=1.283, acc=0.704, acc_len=1.70, thru=20.5, I=22.5, T=37.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:05,178] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:30:13,845] loop.py:45 INFO TRAIN_STEP step=13931 loss=1.293193 Training: 77%|███████▋ | 13931/18018 [29:59:59<8:29:15, 7.48s/step, loss=1.293, acc=0.701, acc_len=1.68, thru=12.9, I=13.8, T=33.3, wait=5.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:30:13,841] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13930: forward=2674.6ms backward=1125.8ms [2026-10-19 19:30:19,782] loop.py:45 INFO TRAIN_STEP step=13932 loss=1.263079 Training: 77%|███████▋ | 13932/18018 [30:00:05<7:57:42, 7.01s/step, loss=1.263, acc=0.707, acc_len=1.70, thru=19.6, I=19.9, T=38.8, wait=2.6s, pool=61, epoch=2/2][2026-10-19 19:30:26,965] loop.py:45 INFO TRAIN_STEP step=13933 loss=1.214698 Training: 77%|███████▋ | 13933/18018 [30:00:13<8:01:00, 7.07s/step, loss=1.215, acc=0.718, acc_len=1.75, thru=16.5, I=17.1, T=37.1, wait=3.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:36,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:30:36,889] loop.py:45 INFO TRAIN_STEP step=13934 loss=1.238333 Training: 77%|███████▋ | 13934/18018 [30:00:22<8:59:16, 7.92s/step, loss=1.238, acc=0.715, acc_len=1.73, thru=15.2, I=16.3, T=34.0, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:37,118] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:42,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:30:44,045] loop.py:380 INFO TIMING step=13935: step=3.549s data=1.435s compute=3.516s [fwd=2.525s bwd=0.976s opt=0.003s] dispatch=3.597s [2026-10-19 19:30:44,045] loop.py:45 INFO TRAIN_STEP step=13935 loss=1.223189 Training: 77%|███████▋ | 13935/18018 [30:00:30<8:43:30, 7.69s/step, loss=1.223, acc=0.714, acc_len=1.73, thru=17.0, I=20.1, T=36.1, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:44,116] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:50,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:30:50,650] loop.py:45 INFO TRAIN_STEP step=13936 loss=1.183482 Training: 77%|███████▋ | 13936/18018 [30:00:36<8:21:08, 7.37s/step, loss=1.183, acc=0.721, acc_len=1.75, thru=19.5, I=22.7, T=38.7, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:30:50,646] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13935: forward=2314.5ms backward=940.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:50,925] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:55,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:30:56,853] loop.py:45 INFO TRAIN_STEP step=13937 loss=1.358529 Training: 77%|███████▋ | 13937/18018 [30:00:42<7:57:17, 7.02s/step, loss=1.359, acc=0.688, acc_len=1.63, thru=20.6, I=22.4, T=36.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:30:56,883] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:03,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:31:03,543] loop.py:45 INFO TRAIN_STEP step=13938 loss=1.234709 Training: 77%|███████▋ | 13938/18018 [30:00:49<7:50:30, 6.92s/step, loss=1.235, acc=0.707, acc_len=1.70, thru=18.7, I=21.8, T=37.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:03,681] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:08,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:31:09,816] loop.py:45 INFO TRAIN_STEP step=13939 loss=1.382959 Training: 77%|███████▋ | 13939/18018 [30:00:55<7:37:14, 6.73s/step, loss=1.383, acc=0.681, acc_len=1.62, thru=20.3, I=21.0, T=32.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:09,852] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 19:31:15,861] loop.py:380 INFO TIMING step=13940: step=3.676s data=1.565s compute=3.646s [fwd=2.531s bwd=1.100s opt=0.003s] dispatch=2.359s [2026-10-19 19:31:15,861] loop.py:45 INFO TRAIN_STEP step=13940 loss=1.178831 Training: 77%|███████▋ | 13940/18018 [30:01:01<7:23:11, 6.52s/step, loss=1.179, acc=0.722, acc_len=1.76, thru=19.2, I=22.2, T=34.8, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:24,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:31:25,126] loop.py:45 INFO TRAIN_STEP step=13941 loss=1.236062 Training: 77%|███████▋ | 13941/18018 [30:01:11<8:19:03, 7.34s/step, loss=1.236, acc=0.709, acc_len=1.71, thru=14.7, I=15.9, T=34.3, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:31:25,122] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13940: forward=2542.1ms backward=1141.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:25,562] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:31:31,027] loop.py:45 INFO TRAIN_STEP step=13942 loss=1.133422 Training: 77%|███████▋ | 13942/18018 [30:01:17<7:49:32, 6.91s/step, loss=1.133, acc=0.730, acc_len=1.79, thru=17.4, I=17.9, T=36.3, wait=2.4s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:39,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:31:40,109] loop.py:45 INFO TRAIN_STEP step=13943 loss=1.168403 Training: 77%|███████▋ | 13943/18018 [30:01:26<8:33:35, 7.56s/step, loss=1.168, acc=0.721, acc_len=1.75, thru=16.8, I=16.4, T=36.4, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:40,386] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:31:47,354] loop.py:45 INFO TRAIN_STEP step=13944 loss=1.230796 Training: 77%|███████▋ | 13944/18018 [30:01:33<8:26:58, 7.47s/step, loss=1.231, acc=0.712, acc_len=1.73, thru=13.3, I=18.7, T=38.7, wait=3.9s, pool=17, epoch=2/2](TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 19:31:47,398] eagle_store.py:517 WARNING batch_get_buffer missing keys (k446212-bc618fdd_hs, k446212-bc618fdd_ids, k446212-bc618fdd_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:55,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9371) [2026-10-19 19:31:47,648] eagle_store.py:517 WARNING batch_get_buffer missing keys (k446202-8af444b9_hs, k446202-8af444b9_ids, k446202-8af444b9_lhs); sleeping 0.50s. [repeated 4x across cluster] [2026-10-19 19:31:56,796] loop.py:380 INFO TIMING step=13945: step=3.642s data=1.536s compute=3.601s [fwd=2.529s bwd=1.058s opt=0.003s] dispatch=5.794s [2026-10-19 19:31:56,796] loop.py:45 INFO TRAIN_STEP step=13945 loss=1.227602 Training: 77%|███████▋ | 13945/18018 [30:01:42<9:07:08, 8.06s/step, loss=1.228, acc=0.713, acc_len=1.73, thru=18.8, I=17.3, T=35.1, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:31:57,120] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:32:04,396] loop.py:45 INFO TRAIN_STEP step=13946 loss=1.287991 Training: 77%|███████▋ | 13946/18018 [30:01:50<8:57:40, 7.92s/step, loss=1.288, acc=0.701, acc_len=1.68, thru=15.1, I=20.0, T=35.9, wait=4.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:32:04,392] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13945: forward=2461.9ms backward=1047.5ms [2026-10-19 19:32:13,856] loop.py:45 INFO TRAIN_STEP step=13947 loss=1.340866 Training: 77%|███████▋ | 13947/18018 [30:01:59<9:28:44, 8.38s/step, loss=1.341, acc=0.695, acc_len=1.67, thru=13.2, I=13.0, T=35.2, wait=5.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:32:22,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:32:23,498] loop.py:45 INFO TRAIN_STEP step=13948 loss=1.190474 Training: 77%|███████▋ | 13948/18018 [30:02:09<9:54:18, 8.76s/step, loss=1.190, acc=0.720, acc_len=1.75, thru=15.5, I=15.8, T=33.9, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:32:23,831] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:32:30,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:32:31,288] loop.py:45 INFO TRAIN_STEP step=13949 loss=1.267309 Training: 77%|███████▋ | 13949/18018 [30:02:17<9:34:21, 8.47s/step, loss=1.267, acc=0.703, acc_len=1.68, thru=16.5, I=22.0, T=37.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:32:31,706] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:32:36,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:32:37,078] loop.py:380 INFO TIMING step=13950: step=3.454s data=1.417s compute=3.426s [fwd=2.425s bwd=0.987s opt=0.003s] dispatch=2.328s [2026-10-19 19:32:37,078] loop.py:45 INFO TRAIN_STEP step=13950 loss=1.232991 Training: 77%|███████▋ | 13950/18018 [30:02:23<8:39:43, 7.67s/step, loss=1.233, acc=0.713, acc_len=1.72, thru=21.9, I=26.5, T=37.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:32:37,543] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-19 19:32:44,091] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13950: forward=2493.1ms backward=1040.6ms [2026-10-19 19:32:44,096] loop.py:45 INFO TRAIN_STEP step=13951 loss=1.329773 Training: 77%|███████▋ | 13951/18018 [30:02:30<8:26:25, 7.47s/step, loss=1.330, acc=0.691, acc_len=1.65, thru=17.8, I=18.9, T=35.8, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:32:49,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:32:50,395] loop.py:45 INFO TRAIN_STEP step=13952 loss=1.218754 Training: 77%|███████▋ | 13952/18018 [30:02:36<8:02:30, 7.12s/step, loss=1.219, acc=0.713, acc_len=1.72, thru=19.8, I=20.8, T=36.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:32:50,752] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:32:57,024] loop.py:45 INFO TRAIN_STEP step=13953 loss=1.234374 Training: 77%|███████▋ | 13953/18018 [30:02:43<7:52:24, 6.97s/step, loss=1.234, acc=0.711, acc_len=1.72, thru=18.4, I=21.1, T=38.8, wait=3.3s, pool=61, epoch=2/2][2026-10-19 19:33:03,094] loop.py:45 INFO TRAIN_STEP step=13954 loss=1.235156 Training: 77%|███████▋ | 13954/18018 [30:02:49<7:33:57, 6.70s/step, loss=1.235, acc=0.711, acc_len=1.71, thru=19.0, I=19.1, T=38.6, wait=2.7s, pool=49, epoch=2/2][2026-10-19 19:33:10,602] loop.py:380 INFO TIMING step=13955: step=3.796s data=1.677s compute=3.765s [fwd=2.714s bwd=1.036s opt=0.003s] dispatch=3.702s [2026-10-19 19:33:10,602] loop.py:45 INFO TRAIN_STEP step=13955 loss=1.268265 Training: 77%|███████▋ | 13955/18018 [30:02:56<7:50:09, 6.94s/step, loss=1.268, acc=0.701, acc_len=1.69, thru=16.8, I=17.7, T=33.7, wait=3.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:33:17,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:33:18,761] loop.py:45 INFO TRAIN_STEP step=13956 loss=1.178236 Training: 77%|███████▋ | 13956/18018 [30:03:04<8:14:45, 7.31s/step, loss=1.178, acc=0.721, acc_len=1.76, thru=17.4, I=19.0, T=32.3, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:33:18,757] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13955: forward=2749.6ms backward=1160.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:33:19,057] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 19:33:26,591] loop.py:45 INFO TRAIN_STEP step=13957 loss=1.301887 Training: 77%|███████▋ | 13957/18018 [30:03:12<8:25:14, 7.46s/step, loss=1.302, acc=0.700, acc_len=1.68, thru=12.5, I=17.3, T=32.9, wait=3.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:33:32,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:33:33,531] loop.py:45 INFO TRAIN_STEP step=13958 loss=1.227524 Training: 77%|███████▋ | 13958/18018 [30:03:19<8:14:30, 7.31s/step, loss=1.228, acc=0.713, acc_len=1.72, thru=17.6, I=20.3, T=35.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:33:33,636] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:33:39,316] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:33:39,371] loop.py:45 INFO TRAIN_STEP step=13959 loss=1.098862 Training: 77%|███████▋ | 13959/18018 [30:03:25<7:44:31, 6.87s/step, loss=1.099, acc=0.736, acc_len=1.80, thru=21.5, I=24.5, T=37.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:33:39,818] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:33:46,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:33:46,897] loop.py:380 INFO TIMING step=13960: step=3.336s data=1.347s compute=3.297s [fwd=2.376s bwd=0.906s opt=0.003s] dispatch=4.182s [2026-10-19 19:33:46,898] loop.py:45 INFO TRAIN_STEP step=13960 loss=1.338315 Training: 77%|███████▋ | 13960/18018 [30:03:32<7:57:50, 7.07s/step, loss=1.338, acc=0.688, acc_len=1.64, thru=18.0, I=19.0, T=38.4, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:33:46,929] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 19:33:53,341] loop.py:45 INFO TRAIN_STEP step=13961 loss=1.322574 Training: 77%|███████▋ | 13961/18018 [30:03:39<7:45:06, 6.88s/step, loss=1.323, acc=0.696, acc_len=1.67, thru=19.2, I=19.2, T=33.3, wait=2.6s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:33:53,336] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13960: forward=2655.9ms backward=1138.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:01,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:34:01,835] loop.py:45 INFO TRAIN_STEP step=13962 loss=1.236259 Training: 77%|███████▋ | 13962/18018 [30:03:47<8:17:44, 7.36s/step, loss=1.236, acc=0.712, acc_len=1.73, thru=16.2, I=16.5, T=33.4, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:01,863] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 19:34:07,279] loop.py:45 INFO TRAIN_STEP step=13963 loss=1.297190 Training: 77%|███████▋ | 13963/18018 [30:03:53<7:38:42, 6.79s/step, loss=1.297, acc=0.705, acc_len=1.70, thru=19.8, I=20.3, T=37.9, wait=2.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:14,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:34:14,846] loop.py:45 INFO TRAIN_STEP step=13964 loss=1.351472 Training: 78%|███████▊ | 13964/18018 [30:04:00<7:54:22, 7.02s/step, loss=1.351, acc=0.690, acc_len=1.65, thru=17.4, I=18.1, T=36.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:15,260] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:20,066] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:34:21,042] loop.py:380 INFO TIMING step=13965: step=3.411s data=1.414s compute=3.381s [fwd=2.412s bwd=0.954s opt=0.003s] dispatch=2.776s [2026-10-19 19:34:21,042] loop.py:45 INFO TRAIN_STEP step=13965 loss=1.233156 Training: 78%|███████▊ | 13965/18018 [30:04:07<7:37:35, 6.77s/step, loss=1.233, acc=0.709, acc_len=1.71, thru=20.4, I=24.9, T=37.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:21,075] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:26,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:34:27,807] loop.py:45 INFO TRAIN_STEP step=13966 loss=1.174047 Training: 78%|███████▊ | 13966/18018 [30:04:13<7:37:18, 6.77s/step, loss=1.174, acc=0.722, acc_len=1.75, thru=18.9, I=23.9, T=34.9, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:34:27,802] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13965: forward=2631.4ms backward=990.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:27,841] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 19:34:34,445] loop.py:45 INFO TRAIN_STEP step=13967 loss=1.214198 Training: 78%|███████▊ | 13967/18018 [30:04:20<7:34:24, 6.73s/step, loss=1.214, acc=0.716, acc_len=1.74, thru=15.8, I=21.0, T=36.2, wait=3.1s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:43,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:34:45,008] loop.py:45 INFO TRAIN_STEP step=13968 loss=1.384446 Training: 78%|███████▊ | 13968/18018 [30:04:31<8:51:58, 7.88s/step, loss=1.384, acc=0.683, acc_len=1.62, thru=16.6, I=15.3, T=34.7, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:34:45,133] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 19:34:52,244] loop.py:45 INFO TRAIN_STEP step=13969 loss=1.207308 Training: 78%|███████▊ | 13969/18018 [30:04:38<8:38:45, 7.69s/step, loss=1.207, acc=0.715, acc_len=1.73, thru=16.7, I=19.4, T=36.2, wait=3.7s, pool=61, epoch=2/2][2026-10-19 19:34:58,478] loop.py:380 INFO TIMING step=13970: step=3.455s data=1.407s compute=3.426s [fwd=2.495s bwd=0.917s opt=0.003s] dispatch=2.770s [2026-10-19 19:34:58,478] loop.py:45 INFO TRAIN_STEP step=13970 loss=1.269452 Training: 78%|███████▊ | 13970/18018 [30:04:44<8:09:11, 7.25s/step, loss=1.269, acc=0.703, acc_len=1.69, thru=18.8, I=18.4, T=37.0, wait=2.8s, pool=53, epoch=2/2][2026-10-19 19:35:08,220] loop.py:45 INFO TRAIN_STEP step=13971 loss=1.331962 Training: 78%|███████▊ | 13971/18018 [30:04:54<8:59:30, 8.00s/step, loss=1.332, acc=0.695, acc_len=1.66, thru=14.3, I=14.4, T=35.3, wait=6.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:35:08,215] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13970: forward=2544.3ms backward=1026.1ms [2026-10-19 19:35:16,881] loop.py:45 INFO TRAIN_STEP step=13972 loss=1.221590 Training: 78%|███████▊ | 13972/18018 [30:05:02<9:12:45, 8.20s/step, loss=1.222, acc=0.715, acc_len=1.73, thru=13.6, I=13.2, T=33.9, wait=4.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:35:24,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:35:25,072] loop.py:45 INFO TRAIN_STEP step=13973 loss=1.273763 Training: 78%|███████▊ | 13973/18018 [30:05:11<9:12:32, 8.20s/step, loss=1.274, acc=0.706, acc_len=1.70, thru=17.8, I=17.8, T=33.4, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:35:25,285] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:35:30,913] loop.py:45 INFO TRAIN_STEP step=13974 loss=1.356931 Training: 78%|███████▊ | 13974/18018 [30:05:16<8:24:43, 7.49s/step, loss=1.357, acc=0.690, acc_len=1.64, thru=18.6, I=20.5, T=38.1, wait=2.5s, pool=49, epoch=2/2][2026-10-19 19:35:40,900] loop.py:380 INFO TIMING step=13975: step=3.831s data=1.589s compute=3.795s [fwd=2.731s bwd=1.049s opt=0.003s] dispatch=6.148s [2026-10-19 19:35:40,900] loop.py:45 INFO TRAIN_STEP step=13975 loss=1.283824 Training: 78%|███████▊ | 13975/18018 [30:05:27<9:15:28, 8.24s/step, loss=1.284, acc=0.706, acc_len=1.69, thru=14.2, I=14.1, T=33.4, wait=6.1s, pool=61, epoch=2/2][2026-10-19 19:35:49,046] loop.py:45 INFO TRAIN_STEP step=13976 loss=1.326411 Training: 78%|███████▊ | 13976/18018 [30:05:35<9:13:02, 8.21s/step, loss=1.326, acc=0.700, acc_len=1.68, thru=16.0, I=14.6, T=37.2, wait=4.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:35:49,042] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13975: forward=2437.7ms backward=950.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:35:54,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:35:55,713] loop.py:45 INFO TRAIN_STEP step=13977 loss=1.349054 Training: 78%|███████▊ | 13977/18018 [30:05:41<8:41:44, 7.75s/step, loss=1.349, acc=0.691, acc_len=1.65, thru=19.9, I=19.1, T=37.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:35:55,862] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 19:36:03,611] loop.py:45 INFO TRAIN_STEP step=13978 loss=1.189812 Training: 78%|███████▊ | 13978/18018 [30:05:49<8:44:41, 7.79s/step, loss=1.190, acc=0.722, acc_len=1.75, thru=14.7, I=17.9, T=36.7, wait=4.4s, pool=49, epoch=2/2][2026-10-19 19:36:12,985] loop.py:45 INFO TRAIN_STEP step=13979 loss=1.275080 Training: 78%|███████▊ | 13979/18018 [30:05:59<9:16:27, 8.27s/step, loss=1.275, acc=0.707, acc_len=1.70, thru=13.4, I=13.1, T=35.9, wait=5.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:20,978] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:36:21,824] loop.py:380 INFO TIMING step=13980: step=3.729s data=1.573s compute=3.702s [fwd=2.673s bwd=1.016s opt=0.003s] dispatch=5.102s [2026-10-19 19:36:21,824] loop.py:45 INFO TRAIN_STEP step=13980 loss=1.299577 Training: 78%|███████▊ | 13980/18018 [30:06:07<9:27:53, 8.44s/step, loss=1.300, acc=0.698, acc_len=1.68, thru=16.2, I=16.4, T=34.3, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:21,983] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:27,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:36:28,909] loop.py:45 INFO TRAIN_STEP step=13981 loss=1.166418 Training: 78%|███████▊ | 13981/18018 [30:06:14<9:00:26, 8.03s/step, loss=1.166, acc=0.728, acc_len=1.78, thru=17.0, I=19.3, T=34.5, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:36:28,905] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13980: forward=2565.2ms backward=1095.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:28,964] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 19:36:34,674] loop.py:45 INFO TRAIN_STEP step=13982 loss=1.209345 Training: 78%|███████▊ | 13982/18018 [30:06:20<8:14:31, 7.35s/step, loss=1.209, acc=0.716, acc_len=1.73, thru=20.4, I=22.0, T=37.8, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:40,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:36:41,731] loop.py:45 INFO TRAIN_STEP step=13983 loss=1.244126 Training: 78%|███████▊ | 13983/18018 [30:06:27<8:08:29, 7.26s/step, loss=1.244, acc=0.711, acc_len=1.72, thru=18.8, I=21.6, T=29.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:41,929] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:48,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:36:49,749] loop.py:45 INFO TRAIN_STEP step=13984 loss=1.293660 Training: 78%|███████▊ | 13984/18018 [30:06:35<8:23:35, 7.49s/step, loss=1.294, acc=0.702, acc_len=1.69, thru=14.5, I=16.9, T=35.6, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:49,858] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:55,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:36:55,559] loop.py:380 INFO TIMING step=13985: step=3.352s data=1.415s compute=3.322s [fwd=2.409s bwd=0.900s opt=0.003s] dispatch=2.448s [2026-10-19 19:36:55,559] loop.py:45 INFO TRAIN_STEP step=13985 loss=1.226228 Training: 78%|███████▊ | 13985/18018 [30:06:41<7:49:32, 6.99s/step, loss=1.226, acc=0.713, acc_len=1.72, thru=21.4, I=24.3, T=38.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:36:55,984] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:37:03,311] loop.py:45 INFO TRAIN_STEP step=13986 loss=1.291881 Training: 78%|███████▊ | 13986/18018 [30:06:49<8:04:52, 7.22s/step, loss=1.292, acc=0.704, acc_len=1.69, thru=16.1, I=17.3, T=36.0, wait=4.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:37:03,307] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13985: forward=2521.5ms backward=989.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:37:12,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:37:12,172] loop.py:45 INFO TRAIN_STEP step=13987 loss=1.292820 Training: 78%|███████▊ | 13987/18018 [30:06:58<8:37:56, 7.71s/step, loss=1.293, acc=0.701, acc_len=1.68, thru=14.8, I=14.7, T=29.3, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:37:12,512] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:37:21,840] loop.py:45 INFO TRAIN_STEP step=13988 loss=1.291338 Training: 78%|███████▊ | 13988/18018 [30:07:07<9:17:18, 8.30s/step, loss=1.291, acc=0.707, acc_len=1.70, thru=12.6, I=14.2, T=33.5, wait=5.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:37:30,079] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:37:30,288] loop.py:45 INFO TRAIN_STEP step=13989 loss=1.228121 Training: 78%|███████▊ | 13989/18018 [30:07:16<9:20:09, 8.34s/step, loss=1.228, acc=0.715, acc_len=1.73, thru=16.2, I=14.8, T=33.8, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:37:30,581] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:37:36,575] loop.py:380 INFO TIMING step=13990: step=3.362s data=1.348s compute=3.332s [fwd=2.414s bwd=0.904s opt=0.003s] dispatch=2.916s [2026-10-19 19:37:36,575] loop.py:45 INFO TRAIN_STEP step=13990 loss=1.303792 Training: 78%|███████▊ | 13990/18018 [30:07:22<8:38:37, 7.73s/step, loss=1.304, acc=0.699, acc_len=1.67, thru=18.6, I=20.8, T=38.1, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:37:45,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:37:46,366] loop.py:45 INFO TRAIN_STEP step=13991 loss=1.311932 Training: 78%|███████▊ | 13991/18018 [30:07:32<9:20:08, 8.35s/step, loss=1.312, acc=0.701, acc_len=1.69, thru=14.2, I=15.1, T=33.6, wait=6.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:37:46,362] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13990: forward=2753.0ms backward=1013.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:37:46,658] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:37:56,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:37:57,596] loop.py:45 INFO TRAIN_STEP step=13992 loss=1.212841 Training: 78%|███████▊ | 13992/18018 [30:07:43<10:18:03, 9.21s/step, loss=1.213, acc=0.716, acc_len=1.74, thru=11.8, I=12.7, T=32.6, wait=7.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:37:57,817] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:02,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:38:03,516] loop.py:45 INFO TRAIN_STEP step=13993 loss=1.216374 Training: 78%|███████▊ | 13993/18018 [30:07:49<9:11:39, 8.22s/step, loss=1.216, acc=0.712, acc_len=1.72, thru=20.1, I=26.4, T=38.0, wait=2.5s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:03,795] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:38:09,531] loop.py:45 INFO TRAIN_STEP step=13994 loss=1.252447 Training: 78%|███████▊ | 13994/18018 [30:07:55<8:27:02, 7.56s/step, loss=1.252, acc=0.708, acc_len=1.71, thru=20.8, I=26.6, T=36.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:09,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:10,010] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:16,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:38:16,552] loop.py:380 INFO TIMING step=13995: step=3.536s data=1.454s compute=3.505s [fwd=2.517s bwd=0.973s opt=0.003s] dispatch=3.478s [2026-10-19 19:38:16,552] loop.py:45 INFO TRAIN_STEP step=13995 loss=1.247845 Training: 78%|███████▊ | 13995/18018 [30:08:02<8:16:07, 7.40s/step, loss=1.248, acc=0.707, acc_len=1.70, thru=17.7, I=19.0, T=36.2, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:16,685] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:24,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:38:24,329] loop.py:45 INFO TRAIN_STEP step=13996 loss=1.310318 Training: 78%|███████▊ | 13996/18018 [30:08:10<8:23:32, 7.51s/step, loss=1.310, acc=0.695, acc_len=1.66, thru=16.6, I=18.7, T=35.3, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:38:24,324] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=13995: forward=2606.2ms backward=974.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:24,604] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:38:31,825] loop.py:45 INFO TRAIN_STEP step=13997 loss=1.184318 Training: 78%|███████▊ | 13997/18018 [30:08:17<8:23:09, 7.51s/step, loss=1.184, acc=0.721, acc_len=1.75, thru=14.4, I=15.1, T=37.0, wait=4.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:39,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 19:38:41,019] loop.py:45 INFO TRAIN_STEP step=13998 loss=1.183640 Training: 78%|███████▊ | 13998/18018 [30:08:27<8:56:52, 8.01s/step, loss=1.184, acc=0.719, acc_len=1.75, thru=17.4, I=18.3, T=29.2, wait=4.8s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:41,275] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:49,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:38:49,650] loop.py:45 INFO TRAIN_STEP step=13999 loss=1.232723 Training: 78%|███████▊ | 13999/18018 [30:08:35<9:09:09, 8.20s/step, loss=1.233, acc=0.710, acc_len=1.71, thru=14.8, I=15.6, T=35.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:49,977] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:38:59,020] loop.py:380 INFO TIMING step=14000: step=4.236s data=2.150s compute=4.178s [fwd=3.037s bwd=1.127s opt=0.003s] dispatch=5.126s [2026-10-19 19:38:59,020] loop.py:45 INFO TRAIN_STEP step=14000 loss=1.288222 Training: 78%|███████▊ | 14000/18018 [30:08:45<9:32:33, 8.55s/step, loss=1.288, acc=0.704, acc_len=1.69, thru=12.2, I=14.7, T=30.2, wait=5.1s, pool=49, epoch=2/2][2026-10-19 19:38:59,023] loop.py:419 INFO Saving checkpoint at step 14000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:38:59,892] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 19:39:02,446] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0014001 (TrainerActor pid=9310) [2026-10-19 19:39:02,453] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0013251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:39:03,914] inference_manager.py:408 INFO Pool capacity available after 4.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:39:09,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 19:39:02,448] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0014001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 19:39:03,443] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0013251 [2026-10-19 19:39:10,171] loop.py:45 INFO TRAIN_STEP step=14001 loss=1.259130 Training: 78%|███████▊ | 14001/18018 [30:08:56<10:24:38, 9.33s/step, loss=1.259, acc=0.710, acc_len=1.72, thru=20.5, I=17.2, T=34.5, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:39:10,166] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14000: forward=2633.4ms backward=1029.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:39:10,569] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:39:16,360] loop.py:45 INFO TRAIN_STEP step=14002 loss=1.189694 Training: 78%|███████▊ | 14002/18018 [30:09:02<9:21:27, 8.39s/step, loss=1.190, acc=0.717, acc_len=1.74, thru=16.9, I=21.7, T=38.7, wait=2.9s, pool=37, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:39:24,981] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:39:26,100] loop.py:45 INFO TRAIN_STEP step=14003 loss=1.224808 Training: 78%|███████▊ | 14003/18018 [30:09:12<9:48:25, 8.79s/step, loss=1.225, acc=0.716, acc_len=1.73, thru=16.2, I=15.9, T=32.0, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:39:26,489] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:39:32,011] loop.py:45 INFO TRAIN_STEP step=14004 loss=1.159504 Training: 78%|███████▊ | 14004/18018 [30:09:18<8:50:24, 7.93s/step, loss=1.160, acc=0.724, acc_len=1.76, thru=19.7, I=22.9, T=38.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:39:32,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:39:32,506] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:39:39,256] loop.py:380 INFO TIMING step=14005: step=3.496s data=1.458s compute=3.464s [fwd=2.489s bwd=0.960s opt=0.003s] dispatch=3.742s [2026-10-19 19:39:39,256] loop.py:45 INFO TRAIN_STEP step=14005 loss=1.309953 Training: 78%|███████▊ | 14005/18018 [30:09:25<8:36:38, 7.72s/step, loss=1.310, acc=0.698, acc_len=1.67, thru=16.4, I=18.3, T=36.6, wait=3.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:39:46,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:39:47,208] loop.py:45 INFO TRAIN_STEP step=14006 loss=1.267073 Training: 78%|███████▊ | 14006/18018 [30:09:33<8:41:04, 7.79s/step, loss=1.267, acc=0.707, acc_len=1.70, thru=17.2, I=18.7, T=36.7, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:39:47,204] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14005: forward=2450.9ms backward=992.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:39:47,665] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:39:54,470] loop.py:45 INFO TRAIN_STEP step=14007 loss=1.309308 Training: 78%|███████▊ | 14007/18018 [30:09:40<8:30:14, 7.63s/step, loss=1.309, acc=0.697, acc_len=1.67, thru=13.4, I=16.4, T=37.5, wait=3.8s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:02,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:40:04,241] loop.py:45 INFO TRAIN_STEP step=14008 loss=1.368142 Training: 78%|███████▊ | 14008/18018 [30:09:50<9:13:00, 8.27s/step, loss=1.368, acc=0.687, acc_len=1.63, thru=16.5, I=17.0, T=32.1, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:04,691] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:10,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:40:10,668] loop.py:45 INFO TRAIN_STEP step=14009 loss=1.296880 Training: 78%|███████▊ | 14009/18018 [30:09:56<8:35:50, 7.72s/step, loss=1.297, acc=0.700, acc_len=1.68, thru=18.5, I=24.1, T=36.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:10,879] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:40:16,453] loop.py:380 INFO TIMING step=14010: step=3.355s data=1.340s compute=3.323s [fwd=2.382s bwd=0.926s opt=0.003s] dispatch=2.421s [2026-10-19 19:40:16,453] loop.py:45 INFO TRAIN_STEP step=14010 loss=1.226846 Training: 78%|███████▊ | 14010/18018 [30:10:02<7:56:54, 7.14s/step, loss=1.227, acc=0.713, acc_len=1.72, thru=19.7, I=21.7, T=38.2, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:24,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:40:25,491] loop.py:45 INFO TRAIN_STEP step=14011 loss=1.186189 Training: 78%|███████▊ | 14011/18018 [30:10:11<8:34:53, 7.71s/step, loss=1.186, acc=0.716, acc_len=1.74, thru=16.2, I=16.6, T=37.5, wait=5.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:40:25,486] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14010: forward=2438.2ms backward=935.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:25,669] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:31,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:40:32,073] loop.py:45 INFO TRAIN_STEP step=14012 loss=1.335215 Training: 78%|███████▊ | 14012/18018 [30:10:18<8:12:08, 7.37s/step, loss=1.335, acc=0.693, acc_len=1.65, thru=19.3, I=23.5, T=37.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:32,249] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:40:37,891] loop.py:45 INFO TRAIN_STEP step=14013 loss=1.219654 Training: 78%|███████▊ | 14013/18018 [30:10:23<7:40:55, 6.91s/step, loss=1.220, acc=0.713, acc_len=1.72, thru=19.7, I=20.3, T=39.8, wait=2.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:45,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:40:45,939] loop.py:45 INFO TRAIN_STEP step=14014 loss=1.286186 Training: 78%|███████▊ | 14014/18018 [30:10:32<8:03:40, 7.25s/step, loss=1.286, acc=0.699, acc_len=1.68, thru=18.1, I=18.7, T=36.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:46,061] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:52,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:40:53,360] loop.py:380 INFO TIMING step=14015: step=3.547s data=1.489s compute=3.515s [fwd=2.573s bwd=0.928s opt=0.003s] dispatch=3.865s [2026-10-19 19:40:53,361] loop.py:45 INFO TRAIN_STEP step=14015 loss=1.274672 Training: 78%|███████▊ | 14015/18018 [30:10:39<8:07:03, 7.30s/step, loss=1.275, acc=0.701, acc_len=1.69, thru=16.4, I=19.0, T=36.1, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:53,624] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:59,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:40:59,874] loop.py:45 INFO TRAIN_STEP step=14016 loss=1.203576 Training: 78%|███████▊ | 14016/18018 [30:10:45<7:51:10, 7.06s/step, loss=1.204, acc=0.716, acc_len=1.73, thru=19.4, I=23.5, T=38.8, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:40:59,869] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14015: forward=2359.9ms backward=904.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:40:59,903] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 19:41:06,057] loop.py:45 INFO TRAIN_STEP step=14017 loss=1.263059 Training: 78%|███████▊ | 14017/18018 [30:10:52<7:33:25, 6.80s/step, loss=1.263, acc=0.704, acc_len=1.69, thru=20.5, I=20.6, T=37.9, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:12,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:41:12,839] loop.py:45 INFO TRAIN_STEP step=14018 loss=1.293285 Training: 78%|███████▊ | 14018/18018 [30:10:58<7:32:56, 6.79s/step, loss=1.293, acc=0.703, acc_len=1.70, thru=19.4, I=19.9, T=38.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:13,158] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:18,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:41:19,299] loop.py:45 INFO TRAIN_STEP step=14019 loss=1.334368 Training: 78%|███████▊ | 14019/18018 [30:11:05<7:26:09, 6.69s/step, loss=1.334, acc=0.695, acc_len=1.66, thru=20.2, I=22.6, T=37.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:19,328] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:24,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:41:25,019] loop.py:380 INFO TIMING step=14020: step=3.326s data=1.352s compute=3.295s [fwd=2.363s bwd=0.918s opt=0.003s] dispatch=2.386s [2026-10-19 19:41:25,019] loop.py:45 INFO TRAIN_STEP step=14020 loss=1.306581 Training: 78%|███████▊ | 14020/18018 [30:11:11<7:06:33, 6.40s/step, loss=1.307, acc=0.700, acc_len=1.68, thru=21.9, I=22.9, T=38.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:25,142] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:33,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:41:33,686] loop.py:45 INFO TRAIN_STEP step=14021 loss=1.286801 Training: 78%|███████▊ | 14021/18018 [30:11:19<7:51:46, 7.08s/step, loss=1.287, acc=0.704, acc_len=1.70, thru=15.3, I=16.4, T=36.3, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:41:33,682] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14020: forward=2481.0ms backward=1001.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:33,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:38,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:41:39,427] loop.py:45 INFO TRAIN_STEP step=14022 loss=1.093625 Training: 78%|███████▊ | 14022/18018 [30:11:25<7:24:51, 6.68s/step, loss=1.094, acc=0.737, acc_len=1.81, thru=21.1, I=20.4, T=39.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:39,719] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:43,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:41:44,494] loop.py:45 INFO TRAIN_STEP step=14023 loss=1.348380 Training: 78%|███████▊ | 14023/18018 [30:11:30<6:52:30, 6.20s/step, loss=1.348, acc=0.692, acc_len=1.65, thru=23.3, I=31.0, T=40.4, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:41:44,798] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:41:51,564] loop.py:45 INFO TRAIN_STEP step=14024 loss=1.320394 Training: 78%|███████▊ | 14024/18018 [30:11:37<7:09:54, 6.46s/step, loss=1.320, acc=0.694, acc_len=1.66, thru=17.5, I=23.3, T=39.0, wait=3.8s, pool=49, epoch=2/2][2026-10-19 19:41:59,851] loop.py:380 INFO TIMING step=14025: step=3.653s data=1.520s compute=3.621s [fwd=2.577s bwd=1.029s opt=0.003s] dispatch=4.623s [2026-10-19 19:41:59,851] loop.py:45 INFO TRAIN_STEP step=14025 loss=1.271371 Training: 78%|███████▊ | 14025/18018 [30:11:45<7:46:21, 7.01s/step, loss=1.271, acc=0.707, acc_len=1.71, thru=15.0, I=14.9, T=35.0, wait=4.6s, pool=53, epoch=2/2][2026-10-19 19:42:08,747] loop.py:45 INFO TRAIN_STEP step=14026 loss=1.258183 Training: 78%|███████▊ | 14026/18018 [30:11:54<8:23:49, 7.57s/step, loss=1.258, acc=0.706, acc_len=1.70, thru=14.9, I=14.8, T=35.0, wait=5.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:42:08,742] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14025: forward=2458.6ms backward=993.5ms [2026-10-19 19:42:15,904] loop.py:45 INFO TRAIN_STEP step=14027 loss=1.162639 Training: 78%|███████▊ | 14027/18018 [30:12:01<8:15:25, 7.45s/step, loss=1.163, acc=0.724, acc_len=1.76, thru=18.9, I=16.8, T=36.9, wait=3.7s, pool=61, epoch=2/2][2026-10-19 19:42:22,934] loop.py:45 INFO TRAIN_STEP step=14028 loss=1.229841 Training: 78%|███████▊ | 14028/18018 [30:12:09<8:06:57, 7.32s/step, loss=1.230, acc=0.711, acc_len=1.72, thru=17.8, I=17.5, T=37.0, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:42:28,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:42:30,220] loop.py:45 INFO TRAIN_STEP step=14029 loss=1.287186 Training: 78%|███████▊ | 14029/18018 [30:12:16<8:06:09, 7.31s/step, loss=1.287, acc=0.702, acc_len=1.69, thru=18.2, I=20.2, T=36.2, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:42:30,258] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:42:37,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:42:38,727] loop.py:380 INFO TIMING step=14030: step=3.665s data=1.552s compute=3.634s [fwd=2.635s bwd=0.985s opt=0.003s] dispatch=4.832s [2026-10-19 19:42:38,727] loop.py:45 INFO TRAIN_STEP step=14030 loss=1.194148 Training: 78%|███████▊ | 14030/18018 [30:12:24<8:29:50, 7.67s/step, loss=1.194, acc=0.717, acc_len=1.74, thru=13.2, I=19.5, T=34.9, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:42:39,023] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 19:42:44,327] loop.py:45 INFO TRAIN_STEP step=14031 loss=1.145964 Training: 78%|███████▊ | 14031/18018 [30:12:30<7:48:24, 7.05s/step, loss=1.146, acc=0.726, acc_len=1.77, thru=20.8, I=22.6, T=39.7, wait=2.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:42:44,323] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14030: forward=2279.3ms backward=907.3ms [2026-10-19 19:42:52,642] loop.py:45 INFO TRAIN_STEP step=14032 loss=1.257826 Training: 78%|███████▊ | 14032/18018 [30:12:38<8:13:33, 7.43s/step, loss=1.258, acc=0.707, acc_len=1.70, thru=16.0, I=15.7, T=36.3, wait=4.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:42:59,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:43:00,183] loop.py:45 INFO TRAIN_STEP step=14033 loss=1.191917 Training: 78%|███████▊ | 14033/18018 [30:12:46<8:15:35, 7.46s/step, loss=1.192, acc=0.719, acc_len=1.74, thru=18.2, I=17.8, T=35.7, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:43:00,632] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:43:08,592] loop.py:45 INFO TRAIN_STEP step=14034 loss=1.189866 Training: 78%|███████▊ | 14034/18018 [30:12:54<8:34:21, 7.75s/step, loss=1.190, acc=0.719, acc_len=1.75, thru=14.0, I=15.0, T=36.9, wait=4.9s, pool=57, epoch=2/2][2026-10-19 19:43:16,238] loop.py:380 INFO TIMING step=14035: step=3.452s data=1.426s compute=3.410s [fwd=2.415s bwd=0.980s opt=0.003s] dispatch=4.185s [2026-10-19 19:43:16,238] loop.py:45 INFO TRAIN_STEP step=14035 loss=1.236023 Training: 78%|███████▊ | 14035/18018 [30:13:02<8:32:16, 7.72s/step, loss=1.236, acc=0.708, acc_len=1.71, thru=14.1, I=14.9, T=37.1, wait=4.2s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:43:24,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:43:25,196] loop.py:45 INFO TRAIN_STEP step=14036 loss=1.297260 Training: 78%|███████▊ | 14036/18018 [30:13:11<8:56:49, 8.09s/step, loss=1.297, acc=0.701, acc_len=1.68, thru=17.0, I=16.6, T=35.4, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:43:25,191] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14035: forward=2483.6ms backward=1074.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:43:25,545] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:43:32,219] loop.py:45 INFO TRAIN_STEP step=14037 loss=1.227594 Training: 78%|███████▊ | 14037/18018 [30:13:18<8:35:30, 7.77s/step, loss=1.228, acc=0.713, acc_len=1.73, thru=18.3, I=18.4, T=37.4, wait=3.6s, pool=61, epoch=2/2][2026-10-19 19:43:44,013] loop.py:45 INFO TRAIN_STEP step=14038 loss=1.412981 Training: 78%|███████▊ | 14038/18018 [30:13:30<9:55:24, 8.98s/step, loss=1.413, acc=0.681, acc_len=1.62, thru=10.6, I=9.2, T=33.4, wait=7.9s, pool=41, epoch=2/2] [2026-10-19 19:43:51,979] loop.py:45 INFO TRAIN_STEP step=14039 loss=1.422732 Training: 78%|███████▊ | 14039/18018 [30:13:38<9:35:11, 8.67s/step, loss=1.423, acc=0.678, acc_len=1.61, thru=15.8, I=14.9, T=35.4, wait=4.3s, pool=45, epoch=2/2][2026-10-19 19:44:00,067] loop.py:380 INFO TIMING step=14040: step=3.435s data=1.347s compute=3.404s [fwd=2.450s bwd=0.939s opt=0.003s] dispatch=4.645s [2026-10-19 19:44:00,067] loop.py:45 INFO TRAIN_STEP step=14040 loss=1.290772 Training: 78%|███████▊ | 14040/18018 [30:13:46<9:23:23, 8.50s/step, loss=1.291, acc=0.702, acc_len=1.69, thru=16.4, I=16.7, T=37.3, wait=4.6s, pool=57, epoch=2/2][2026-10-19 19:44:09,712] loop.py:45 INFO TRAIN_STEP step=14041 loss=1.353343 Training: 78%|███████▊ | 14041/18018 [30:13:55<9:46:03, 8.84s/step, loss=1.353, acc=0.690, acc_len=1.65, thru=14.1, I=13.0, T=35.0, wait=6.0s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:44:09,708] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14040: forward=2392.2ms backward=1014.8ms [2026-10-19 19:44:16,163] loop.py:45 INFO TRAIN_STEP step=14042 loss=1.308803 Training: 78%|███████▊ | 14042/18018 [30:14:02<8:58:24, 8.12s/step, loss=1.309, acc=0.696, acc_len=1.66, thru=16.9, I=16.7, T=38.0, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:44:26,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:44:27,053] loop.py:45 INFO TRAIN_STEP step=14043 loss=1.451085 Training: 78%|███████▊ | 14043/18018 [30:14:13<9:53:14, 8.95s/step, loss=1.451, acc=0.671, acc_len=1.58, thru=14.5, I=13.3, T=33.5, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:44:27,397] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:44:32,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:44:34,616] loop.py:45 INFO TRAIN_STEP step=14044 loss=1.308500 Training: 78%|███████▊ | 14044/18018 [30:14:20<9:25:27, 8.54s/step, loss=1.308, acc=0.699, acc_len=1.67, thru=17.5, I=20.7, T=34.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:44:34,875] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 19:44:41,555] loop.py:380 INFO TIMING step=14045: step=3.427s data=1.475s compute=3.400s [fwd=2.459s bwd=0.927s opt=0.003s] dispatch=3.501s [2026-10-19 19:44:41,555] loop.py:45 INFO TRAIN_STEP step=14045 loss=1.331298 Training: 78%|███████▊ | 14045/18018 [30:14:27<8:53:31, 8.06s/step, loss=1.331, acc=0.690, acc_len=1.64, thru=16.5, I=21.7, T=37.3, wait=3.5s, pool=53, epoch=2/2][2026-10-19 19:44:49,613] loop.py:45 INFO TRAIN_STEP step=14046 loss=1.237639 Training: 78%|███████▊ | 14046/18018 [30:14:35<8:53:23, 8.06s/step, loss=1.238, acc=0.711, acc_len=1.72, thru=17.3, I=16.6, T=35.8, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:44:49,609] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14045: forward=2549.1ms backward=978.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:44:49,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:44:50,089] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:44:56,050] loop.py:45 INFO TRAIN_STEP step=14047 loss=1.263052 Training: 78%|███████▊ | 14047/18018 [30:14:42<8:21:07, 7.57s/step, loss=1.263, acc=0.707, acc_len=1.70, thru=17.3, I=19.8, T=38.4, wait=3.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:05,419] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:45:05,805] loop.py:45 INFO TRAIN_STEP step=14048 loss=1.318225 Training: 78%|███████▊ | 14048/18018 [30:14:51<9:04:16, 8.23s/step, loss=1.318, acc=0.693, acc_len=1.65, thru=15.2, I=14.4, T=36.3, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:05,926] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:11,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:45:11,449] loop.py:45 INFO TRAIN_STEP step=14049 loss=1.334044 Training: 78%|███████▊ | 14049/18018 [30:14:57<8:12:56, 7.45s/step, loss=1.334, acc=0.696, acc_len=1.67, thru=21.2, I=22.5, T=38.3, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:11,713] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:45:16,844] loop.py:380 INFO TIMING step=14050: step=3.264s data=1.285s compute=3.234s [fwd=2.332s bwd=0.887s opt=0.003s] dispatch=2.121s [2026-10-19 19:45:16,844] loop.py:45 INFO TRAIN_STEP step=14050 loss=1.280067 Training: 78%|███████▊ | 14050/18018 [30:15:02<7:32:03, 6.84s/step, loss=1.280, acc=0.706, acc_len=1.70, thru=22.5, I=23.9, T=39.2, wait=2.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:24,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:45:25,192] loop.py:45 INFO TRAIN_STEP step=14051 loss=1.281281 Training: 78%|███████▊ | 14051/18018 [30:15:11<8:01:55, 7.29s/step, loss=1.281, acc=0.702, acc_len=1.68, thru=16.8, I=16.7, T=36.3, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:45:25,188] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14050: forward=2543.6ms backward=940.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:25,232] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:32,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:45:34,142] loop.py:45 INFO TRAIN_STEP step=14052 loss=1.293790 Training: 78%|███████▊ | 14052/18018 [30:15:20<8:34:42, 7.79s/step, loss=1.294, acc=0.702, acc_len=1.68, thru=14.1, I=17.9, T=34.7, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:34,303] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:39,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:45:40,391] loop.py:45 INFO TRAIN_STEP step=14053 loss=1.327566 Training: 78%|███████▊ | 14053/18018 [30:15:26<8:04:05, 7.33s/step, loss=1.328, acc=0.693, acc_len=1.65, thru=19.7, I=25.8, T=38.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:40,470] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 19:45:47,036] loop.py:45 INFO TRAIN_STEP step=14054 loss=1.388685 Training: 78%|███████▊ | 14054/18018 [30:15:33<7:50:28, 7.12s/step, loss=1.389, acc=0.681, acc_len=1.61, thru=16.0, I=17.1, T=38.8, wait=3.3s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:57,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:45:57,740] loop.py:380 INFO TIMING step=14055: step=3.846s data=1.632s compute=3.804s [fwd=2.707s bwd=1.083s opt=0.003s] dispatch=6.850s [2026-10-19 19:45:57,740] loop.py:45 INFO TRAIN_STEP step=14055 loss=1.351886 Training: 78%|███████▊ | 14055/18018 [30:15:43<9:01:22, 8.20s/step, loss=1.352, acc=0.690, acc_len=1.64, thru=15.2, I=15.7, T=33.3, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:45:58,049] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:46:04,936] loop.py:45 INFO TRAIN_STEP step=14056 loss=1.345178 Training: 78%|███████▊ | 14056/18018 [30:15:51<8:41:26, 7.90s/step, loss=1.345, acc=0.691, acc_len=1.65, thru=15.9, I=17.7, T=37.7, wait=3.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:46:04,932] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14055: forward=2450.0ms backward=910.6ms [2026-10-19 19:46:13,624] loop.py:45 INFO TRAIN_STEP step=14057 loss=1.293352 Training: 78%|███████▊ | 14057/18018 [30:15:59<8:57:01, 8.13s/step, loss=1.293, acc=0.700, acc_len=1.68, thru=15.2, I=16.0, T=34.6, wait=5.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:46:21,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:46:21,408] loop.py:45 INFO TRAIN_STEP step=14058 loss=1.298623 Training: 78%|███████▊ | 14058/18018 [30:16:07<8:49:52, 8.03s/step, loss=1.299, acc=0.703, acc_len=1.69, thru=16.4, I=16.5, T=35.7, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:46:21,635] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:46:27,395] loop.py:45 INFO TRAIN_STEP step=14059 loss=1.316837 Training: 78%|███████▊ | 14059/18018 [30:16:13<8:09:22, 7.42s/step, loss=1.317, acc=0.696, acc_len=1.67, thru=20.1, I=21.0, T=38.0, wait=2.6s, pool=61, epoch=2/2][2026-10-19 19:46:34,521] loop.py:380 INFO TIMING step=14060: step=3.401s data=1.377s compute=3.369s [fwd=2.431s bwd=0.924s opt=0.003s] dispatch=3.714s [2026-10-19 19:46:34,522] loop.py:45 INFO TRAIN_STEP step=14060 loss=1.200152 Training: 78%|███████▊ | 14060/18018 [30:16:20<8:03:26, 7.33s/step, loss=1.200, acc=0.719, acc_len=1.75, thru=17.3, I=17.2, T=37.6, wait=3.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:46:42,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:46:43,742] loop.py:45 INFO TRAIN_STEP step=14061 loss=1.234881 Training: 78%|███████▊ | 14061/18018 [30:16:29<8:40:45, 7.90s/step, loss=1.235, acc=0.714, acc_len=1.73, thru=16.0, I=18.2, T=35.9, wait=5.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:46:43,738] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14060: forward=2538.8ms backward=984.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:46:44,246] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 19:46:49,876] loop.py:45 INFO TRAIN_STEP step=14062 loss=1.324274 Training: 78%|███████▊ | 14062/18018 [30:16:35<8:05:49, 7.37s/step, loss=1.324, acc=0.695, acc_len=1.66, thru=20.1, I=24.2, T=39.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:46:49,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:46:50,352] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:46:57,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:46:58,147] loop.py:45 INFO TRAIN_STEP step=14063 loss=1.292376 Training: 78%|███████▊ | 14063/18018 [30:16:44<8:23:29, 7.64s/step, loss=1.292, acc=0.702, acc_len=1.69, thru=17.0, I=19.1, T=37.2, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:46:58,318] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:47:05,846] loop.py:45 INFO TRAIN_STEP step=14064 loss=1.298746 Training: 78%|███████▊ | 14064/18018 [30:16:51<8:24:35, 7.66s/step, loss=1.299, acc=0.698, acc_len=1.67, thru=14.1, I=15.6, T=37.5, wait=4.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:13,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:47:14,747] loop.py:380 INFO TIMING step=14065: step=3.773s data=1.567s compute=3.733s [fwd=2.636s bwd=1.083s opt=0.003s] dispatch=5.120s [2026-10-19 19:47:14,747] loop.py:45 INFO TRAIN_STEP step=14065 loss=1.295034 Training: 78%|███████▊ | 14065/18018 [30:17:00<8:49:05, 8.03s/step, loss=1.295, acc=0.703, acc_len=1.70, thru=16.8, I=18.4, T=33.9, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:14,851] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (TrainerActor pid=9310) [2026-10-19 19:47:21,158] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14065: forward=2333.5ms backward=898.4ms [2026-10-19 19:47:21,162] loop.py:45 INFO TRAIN_STEP step=14066 loss=1.327282 Training: 78%|███████▊ | 14066/18018 [30:17:07<8:16:59, 7.55s/step, loss=1.327, acc=0.696, acc_len=1.66, thru=19.0, I=21.5, T=38.9, wait=3.1s, pool=57, epoch=2/2][2026-10-19 19:47:29,073] loop.py:45 INFO TRAIN_STEP step=14067 loss=1.345278 Training: 78%|███████▊ | 14067/18018 [30:17:15<8:24:08, 7.66s/step, loss=1.345, acc=0.692, acc_len=1.65, thru=14.5, I=16.5, T=34.2, wait=4.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:39,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:47:39,688] loop.py:45 INFO TRAIN_STEP step=14068 loss=1.291666 Training: 78%|███████▊ | 14068/18018 [30:17:25<9:22:24, 8.54s/step, loss=1.292, acc=0.699, acc_len=1.67, thru=15.1, I=14.0, T=32.6, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:39,901] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:45,949] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:47:46,882] loop.py:45 INFO TRAIN_STEP step=14069 loss=1.249826 Training: 78%|███████▊ | 14069/18018 [30:17:32<8:55:36, 8.14s/step, loss=1.250, acc=0.709, acc_len=1.70, thru=18.8, I=18.9, T=37.1, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:46,963] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:51,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:47:52,065] loop.py:380 INFO TIMING step=14070: step=3.305s data=1.321s compute=3.279s [fwd=2.323s bwd=0.942s opt=0.003s] dispatch=1.870s [2026-10-19 19:47:52,065] loop.py:45 INFO TRAIN_STEP step=14070 loss=1.152542 Training: 78%|███████▊ | 14070/18018 [30:17:38<7:57:08, 7.25s/step, loss=1.153, acc=0.725, acc_len=1.77, thru=21.6, I=26.2, T=38.7, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:52,134] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:58,128] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:47:59,449] loop.py:45 INFO TRAIN_STEP step=14071 loss=1.394841 Training: 78%|███████▊ | 14071/18018 [30:17:45<7:59:41, 7.29s/step, loss=1.395, acc=0.684, acc_len=1.63, thru=18.3, I=21.0, T=36.7, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:47:59,444] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14070: forward=2463.2ms backward=978.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:47:59,638] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:48:05,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:48:05,405] loop.py:45 INFO TRAIN_STEP step=14072 loss=1.309882 Training: 78%|███████▊ | 14072/18018 [30:17:51<7:33:10, 6.89s/step, loss=1.310, acc=0.695, acc_len=1.66, thru=21.4, I=24.8, T=39.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:48:05,650] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:48:12,999] loop.py:45 INFO TRAIN_STEP step=14073 loss=1.134247 Training: 78%|███████▊ | 14073/18018 [30:17:59<7:46:56, 7.10s/step, loss=1.134, acc=0.727, acc_len=1.77, thru=17.8, I=18.0, T=37.4, wait=4.2s, pool=61, epoch=2/2][2026-10-19 19:48:20,592] loop.py:45 INFO TRAIN_STEP step=14074 loss=1.269019 Training: 78%|███████▊ | 14074/18018 [30:18:06<7:56:32, 7.25s/step, loss=1.269, acc=0.711, acc_len=1.71, thru=15.4, I=17.5, T=37.1, wait=4.1s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:48:27,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:48:28,437] loop.py:380 INFO TIMING step=14075: step=3.466s data=1.491s compute=3.402s [fwd=2.329s bwd=1.058s opt=0.003s] dispatch=4.367s [2026-10-19 19:48:28,438] loop.py:45 INFO TRAIN_STEP step=14075 loss=1.187378 Training: 78%|███████▊ | 14075/18018 [30:18:14<8:08:08, 7.43s/step, loss=1.187, acc=0.717, acc_len=1.75, thru=18.1, I=17.8, T=36.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:48:28,807] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:48:34,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:48:35,143] loop.py:45 INFO TRAIN_STEP step=14076 loss=1.356199 Training: 78%|███████▊ | 14076/18018 [30:18:21<7:53:46, 7.21s/step, loss=1.356, acc=0.686, acc_len=1.63, thru=19.4, I=22.0, T=37.8, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:48:35,135] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14075: forward=2398.1ms backward=914.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:48:35,285] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:48:41,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:48:41,305] loop.py:45 INFO TRAIN_STEP step=14077 loss=1.282093 Training: 78%|███████▊ | 14077/18018 [30:18:27<7:32:56, 6.90s/step, loss=1.282, acc=0.702, acc_len=1.69, thru=20.1, I=19.5, T=36.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:48:41,668] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:48:48,686] loop.py:45 INFO TRAIN_STEP step=14078 loss=1.248247 Training: 78%|███████▊ | 14078/18018 [30:18:34<7:42:23, 7.04s/step, loss=1.248, acc=0.706, acc_len=1.70, thru=17.9, I=17.6, T=37.0, wait=3.9s, pool=61, epoch=2/2][2026-10-19 19:48:54,715] loop.py:45 INFO TRAIN_STEP step=14079 loss=1.253530 Training: 78%|███████▊ | 14079/18018 [30:18:40<7:22:20, 6.74s/step, loss=1.254, acc=0.709, acc_len=1.71, thru=19.8, I=21.2, T=37.2, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:49:03,314] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:49:03,574] loop.py:380 INFO TIMING step=14080: step=3.548s data=1.432s compute=3.520s [fwd=2.554s bwd=0.951s opt=0.003s] dispatch=5.302s [2026-10-19 19:49:03,574] loop.py:45 INFO TRAIN_STEP step=14080 loss=1.368603 Training: 78%|███████▊ | 14080/18018 [30:18:49<8:04:00, 7.37s/step, loss=1.369, acc=0.684, acc_len=1.63, thru=15.2, I=15.5, T=36.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:49:03,815] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:49:11,105] loop.py:45 INFO TRAIN_STEP step=14081 loss=1.292527 Training: 78%|███████▊ | 14081/18018 [30:18:57<8:06:58, 7.42s/step, loss=1.293, acc=0.698, acc_len=1.67, thru=15.7, I=18.2, T=35.7, wait=3.9s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:49:11,100] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14080: forward=2492.9ms backward=1035.3ms [2026-10-19 19:49:18,967] loop.py:45 INFO TRAIN_STEP step=14082 loss=1.265159 Training: 78%|███████▊ | 14082/18018 [30:19:05<8:15:28, 7.55s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=14.8, I=16.4, T=36.4, wait=4.3s, pool=37, epoch=2/2][2026-10-19 19:49:28,983] loop.py:45 INFO TRAIN_STEP step=14083 loss=1.311426 Training: 78%|███████▊ | 14083/18018 [30:19:15<9:03:50, 8.29s/step, loss=1.311, acc=0.701, acc_len=1.69, thru=14.1, I=12.4, T=35.6, wait=6.4s, pool=45, epoch=2/2][2026-10-19 19:49:35,368] loop.py:45 INFO TRAIN_STEP step=14084 loss=1.164180 Training: 78%|███████▊ | 14084/18018 [30:19:21<8:26:10, 7.72s/step, loss=1.164, acc=0.724, acc_len=1.77, thru=18.6, I=18.3, T=37.4, wait=3.0s, pool=53, epoch=2/2][2026-10-19 19:49:44,517] loop.py:380 INFO TIMING step=14085: step=3.669s data=1.479s compute=3.642s [fwd=2.608s bwd=1.019s opt=0.003s] dispatch=5.471s [2026-10-19 19:49:44,517] loop.py:45 INFO TRAIN_STEP step=14085 loss=1.233444 Training: 78%|███████▊ | 14085/18018 [30:19:30<8:54:07, 8.15s/step, loss=1.233, acc=0.709, acc_len=1.71, thru=15.3, I=14.3, T=34.9, wait=5.5s, pool=57, epoch=2/2][2026-10-19 19:49:52,406] loop.py:45 INFO TRAIN_STEP step=14086 loss=1.308270 Training: 78%|███████▊ | 14086/18018 [30:19:38<8:48:56, 8.07s/step, loss=1.308, acc=0.698, acc_len=1.68, thru=16.1, I=17.0, T=33.3, wait=4.0s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:49:52,401] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14085: forward=2696.6ms backward=1089.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:49:58,797] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:49:59,974] loop.py:45 INFO TRAIN_STEP step=14087 loss=1.154455 Training: 78%|███████▊ | 14087/18018 [30:19:46<8:38:53, 7.92s/step, loss=1.154, acc=0.722, acc_len=1.75, thru=16.8, I=18.7, T=33.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:50:00,308] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:50:05,721] loop.py:45 INFO TRAIN_STEP step=14088 loss=1.233388 Training: 78%|███████▊ | 14088/18018 [30:19:51<7:56:03, 7.27s/step, loss=1.233, acc=0.711, acc_len=1.71, thru=18.5, I=23.2, T=38.5, wait=2.4s, pool=41, epoch=2/2][2026-10-19 19:50:13,685] loop.py:45 INFO TRAIN_STEP step=14089 loss=1.223550 Training: 78%|███████▊ | 14089/18018 [30:19:59<8:09:34, 7.48s/step, loss=1.224, acc=0.712, acc_len=1.72, thru=16.5, I=15.1, T=37.7, wait=4.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:50:19,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:50:20,794] loop.py:380 INFO TIMING step=14090: step=4.007s data=1.306s compute=3.974s [fwd=3.006s bwd=0.952s opt=0.003s] dispatch=3.095s [2026-10-19 19:50:20,794] loop.py:45 INFO TRAIN_STEP step=14090 loss=1.247732 Training: 78%|███████▊ | 14090/18018 [30:20:06<8:02:17, 7.37s/step, loss=1.248, acc=0.711, acc_len=1.72, thru=18.9, I=20.4, T=31.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:50:20,826] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 19:50:27,594] loop.py:45 INFO TRAIN_STEP step=14091 loss=1.247083 Training: 78%|███████▊ | 14091/18018 [30:20:13<7:50:59, 7.20s/step, loss=1.247, acc=0.709, acc_len=1.71, thru=16.6, I=18.2, T=34.7, wait=3.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:50:27,589] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14090: forward=2676.6ms backward=954.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:50:32,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:50:33,806] loop.py:45 INFO TRAIN_STEP step=14092 loss=1.219190 Training: 78%|███████▊ | 14092/18018 [30:20:19<7:31:36, 6.90s/step, loss=1.219, acc=0.715, acc_len=1.73, thru=19.7, I=20.3, T=37.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:50:33,917] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:50:39,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:50:40,520] loop.py:45 INFO TRAIN_STEP step=14093 loss=1.177598 Training: 78%|███████▊ | 14093/18018 [30:20:26<7:27:47, 6.85s/step, loss=1.178, acc=0.719, acc_len=1.74, thru=19.2, I=23.2, T=33.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:50:40,553] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 19:50:46,159] loop.py:45 INFO TRAIN_STEP step=14094 loss=1.277627 Training: 78%|███████▊ | 14094/18018 [30:20:32<7:04:04, 6.48s/step, loss=1.278, acc=0.700, acc_len=1.67, thru=20.9, I=24.5, T=39.4, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:50:53,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:50:55,054] loop.py:380 INFO TIMING step=14095: step=3.636s data=1.584s compute=3.597s [fwd=2.557s bwd=1.026s opt=0.003s] dispatch=5.246s [2026-10-19 19:50:55,054] loop.py:45 INFO TRAIN_STEP step=14095 loss=1.305498 Training: 78%|███████▊ | 14095/18018 [30:20:41<7:51:09, 7.21s/step, loss=1.305, acc=0.697, acc_len=1.67, thru=15.7, I=17.4, T=35.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:50:55,502] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:51:02,206] loop.py:45 INFO TRAIN_STEP step=14096 loss=1.209454 Training: 78%|███████▊ | 14096/18018 [30:20:48<7:50:00, 7.19s/step, loss=1.209, acc=0.714, acc_len=1.73, thru=13.3, I=20.2, T=38.3, wait=3.8s, pool=25, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:51:02,202] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14095: forward=2354.2ms backward=942.6ms [2026-10-19 19:51:11,203] loop.py:45 INFO TRAIN_STEP step=14097 loss=1.392604 Training: 78%|███████▊ | 14097/18018 [30:20:57<8:25:16, 7.73s/step, loss=1.393, acc=0.684, acc_len=1.63, thru=17.8, I=15.0, T=38.1, wait=5.6s, pool=61, epoch=2/2][2026-10-19 19:51:18,016] loop.py:45 INFO TRAIN_STEP step=14098 loss=1.186959 Training: 78%|███████▊ | 14098/18018 [30:21:04<8:07:07, 7.46s/step, loss=1.187, acc=0.721, acc_len=1.75, thru=17.8, I=19.5, T=37.0, wait=3.3s, pool=49, epoch=2/2][2026-10-19 19:51:27,409] loop.py:45 INFO TRAIN_STEP step=14099 loss=1.242612 Training: 78%|███████▊ | 14099/18018 [30:21:13<8:45:00, 8.04s/step, loss=1.243, acc=0.711, acc_len=1.72, thru=14.2, I=13.0, T=35.3, wait=5.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:51:33,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:51:34,316] loop.py:380 INFO TIMING step=14100: step=3.478s data=1.379s compute=3.451s [fwd=2.484s bwd=0.953s opt=0.003s] dispatch=3.418s [2026-10-19 19:51:34,317] loop.py:45 INFO TRAIN_STEP step=14100 loss=1.193002 Training: 78%|███████▊ | 14100/18018 [30:21:20<8:22:41, 7.70s/step, loss=1.193, acc=0.721, acc_len=1.75, thru=19.7, I=17.6, T=36.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:51:34,469] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 19:51:41,393] loop.py:45 INFO TRAIN_STEP step=14101 loss=1.166039 Training: 78%|███████▊ | 14101/18018 [30:21:27<8:10:23, 7.51s/step, loss=1.166, acc=0.724, acc_len=1.77, thru=14.1, I=17.4, T=35.6, wait=3.5s, pool=21, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:51:41,388] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14100: forward=2530.6ms backward=1009.2ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 19:51:41,466] eagle_store.py:517 WARNING batch_get_buffer missing keys (k451238-9282a0b3_hs, k451238-9282a0b3_ids, k451238-9282a0b3_lhs); sleeping 0.50s. [2026-10-19 19:51:51,951] loop.py:45 INFO TRAIN_STEP step=14102 loss=1.389346 Training: 78%|███████▊ | 14102/18018 [30:21:38<9:09:52, 8.43s/step, loss=1.389, acc=0.685, acc_len=1.63, thru=16.5, I=16.8, T=34.7, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:51:51,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 19:51:41,481] eagle_store.py:517 WARNING batch_get_buffer missing keys (k451237-96e4fd6a_hs, k451237-96e4fd6a_ids, k451237-96e4fd6a_lhs); sleeping 0.50s. [repeated 2x across cluster] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:51:52,412] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:51:59,534] loop.py:45 INFO TRAIN_STEP step=14103 loss=1.252484 Training: 78%|███████▊ | 14103/18018 [30:21:45<8:53:20, 8.17s/step, loss=1.252, acc=0.704, acc_len=1.69, thru=13.0, I=15.2, T=35.5, wait=4.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:07,345] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:52:07,471] loop.py:45 INFO TRAIN_STEP step=14104 loss=1.264581 Training: 78%|███████▊ | 14104/18018 [30:21:53<8:48:30, 8.10s/step, loss=1.265, acc=0.707, acc_len=1.71, thru=17.4, I=16.3, T=34.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:07,848] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:52:13,429] loop.py:380 INFO TIMING step=14105: step=3.301s data=1.291s compute=3.274s [fwd=2.368s bwd=0.892s opt=0.003s] dispatch=2.648s [2026-10-19 19:52:13,429] loop.py:45 INFO TRAIN_STEP step=14105 loss=1.244274 Training: 78%|███████▊ | 14105/18018 [30:21:59<8:06:27, 7.46s/step, loss=1.244, acc=0.713, acc_len=1.73, thru=19.0, I=19.3, T=38.8, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:20,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:52:21,981] loop.py:45 INFO TRAIN_STEP step=14106 loss=1.401538 Training: 78%|███████▊ | 14106/18018 [30:22:08<8:27:41, 7.79s/step, loss=1.402, acc=0.681, acc_len=1.61, thru=15.9, I=19.1, T=34.5, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:52:21,976] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14105: forward=2630.1ms backward=1032.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:22,324] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:27,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:52:28,366] loop.py:45 INFO TRAIN_STEP step=14107 loss=1.237953 Training: 78%|███████▊ | 14107/18018 [30:22:14<8:00:10, 7.37s/step, loss=1.238, acc=0.710, acc_len=1.72, thru=19.3, I=24.8, T=31.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:28,420] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:33,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:52:34,637] loop.py:45 INFO TRAIN_STEP step=14108 loss=1.254736 Training: 78%|███████▊ | 14108/18018 [30:22:20<7:38:35, 7.04s/step, loss=1.255, acc=0.709, acc_len=1.72, thru=19.4, I=22.8, T=36.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:35,097] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:41,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:52:42,075] loop.py:45 INFO TRAIN_STEP step=14109 loss=1.193914 Training: 78%|███████▊ | 14109/18018 [30:22:28<7:46:21, 7.16s/step, loss=1.194, acc=0.720, acc_len=1.75, thru=16.4, I=19.9, T=29.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:42,270] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:48,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:52:48,837] loop.py:380 INFO TIMING step=14110: step=3.490s data=1.312s compute=3.460s [fwd=2.463s bwd=0.983s opt=0.003s] dispatch=3.261s [2026-10-19 19:52:48,838] loop.py:45 INFO TRAIN_STEP step=14110 loss=1.275149 Training: 78%|███████▊ | 14110/18018 [30:22:34<7:38:29, 7.04s/step, loss=1.275, acc=0.708, acc_len=1.71, thru=17.8, I=19.5, T=36.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:49,222] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:54,893] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:52:55,176] loop.py:45 INFO TRAIN_STEP step=14111 loss=1.325500 Training: 78%|███████▊ | 14111/18018 [30:22:41<7:24:40, 6.83s/step, loss=1.326, acc=0.692, acc_len=1.65, thru=20.1, I=21.5, T=38.9, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:52:55,171] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14110: forward=2343.5ms backward=901.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:52:55,394] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:53:03,146] loop.py:45 INFO TRAIN_STEP step=14112 loss=1.254630 Training: 78%|███████▊ | 14112/18018 [30:22:49<7:46:49, 7.17s/step, loss=1.255, acc=0.710, acc_len=1.71, thru=17.3, I=17.0, T=36.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:53:03,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:53:03,577] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:53:08,865] loop.py:45 INFO TRAIN_STEP step=14113 loss=1.292890 Training: 78%|███████▊ | 14113/18018 [30:22:54<7:18:20, 6.73s/step, loss=1.293, acc=0.700, acc_len=1.68, thru=19.2, I=17.9, T=39.5, wait=2.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:53:17,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:53:19,280] loop.py:45 INFO TRAIN_STEP step=14114 loss=1.328891 Training: 78%|███████▊ | 14114/18018 [30:23:05<8:30:05, 7.84s/step, loss=1.329, acc=0.697, acc_len=1.67, thru=14.1, I=16.5, T=29.6, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:53:19,538] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 19:53:26,859] loop.py:380 INFO TIMING step=14115: step=3.538s data=1.480s compute=3.507s [fwd=2.587s bwd=0.906s opt=0.003s] dispatch=4.033s [2026-10-19 19:53:26,859] loop.py:45 INFO TRAIN_STEP step=14115 loss=1.196083 Training: 78%|███████▊ | 14115/18018 [30:23:12<8:24:53, 7.76s/step, loss=1.196, acc=0.718, acc_len=1.74, thru=13.7, I=17.2, T=36.2, wait=4.0s, pool=49, epoch=2/2][2026-10-19 19:53:34,204] loop.py:45 INFO TRAIN_STEP step=14116 loss=1.258754 Training: 78%|███████▊ | 14116/18018 [30:23:20<8:16:36, 7.64s/step, loss=1.259, acc=0.708, acc_len=1.70, thru=16.2, I=17.1, T=35.8, wait=3.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:53:34,199] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14115: forward=2511.9ms backward=1013.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:53:41,241] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:53:42,643] loop.py:45 INFO TRAIN_STEP step=14117 loss=1.323009 Training: 78%|███████▊ | 14117/18018 [30:23:28<8:32:10, 7.88s/step, loss=1.323, acc=0.695, acc_len=1.66, thru=16.4, I=18.4, T=34.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:53:42,767] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 19:53:50,624] loop.py:45 INFO TRAIN_STEP step=14118 loss=1.292535 Training: 78%|███████▊ | 14118/18018 [30:23:36<8:34:05, 7.91s/step, loss=1.293, acc=0.705, acc_len=1.69, thru=14.2, I=18.3, T=38.0, wait=4.6s, pool=53, epoch=2/2][2026-10-19 19:53:57,863] loop.py:45 INFO TRAIN_STEP step=14119 loss=1.277302 Training: 78%|███████▊ | 14119/18018 [30:23:43<8:20:49, 7.71s/step, loss=1.277, acc=0.705, acc_len=1.70, thru=18.3, I=18.8, T=36.4, wait=3.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:54:07,079] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:54:08,232] loop.py:380 INFO TIMING step=14120: step=3.924s data=1.902s compute=3.593s [fwd=2.435s bwd=1.144s opt=0.003s] dispatch=6.436s [2026-10-19 19:54:08,232] loop.py:45 INFO TRAIN_STEP step=14120 loss=1.309134 Training: 78%|███████▊ | 14120/18018 [30:23:54<9:12:35, 8.51s/step, loss=1.309, acc=0.700, acc_len=1.68, thru=12.6, I=13.6, T=32.6, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:54:08,588] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:54:13,861] loop.py:45 INFO TRAIN_STEP step=14121 loss=1.196580 Training: 78%|███████▊ | 14121/18018 [30:23:59<8:16:23, 7.64s/step, loss=1.197, acc=0.718, acc_len=1.74, thru=17.3, I=19.5, T=37.8, wait=2.2s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:54:13,857] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14120: forward=2388.6ms backward=956.9ms [2026-10-19 19:54:25,637] loop.py:45 INFO TRAIN_STEP step=14122 loss=1.260410 Training: 78%|███████▊ | 14122/18018 [30:24:11<9:36:47, 8.88s/step, loss=1.260, acc=0.708, acc_len=1.71, thru=10.8, I=11.4, T=31.6, wait=7.7s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:54:33,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:54:34,846] loop.py:45 INFO TRAIN_STEP step=14123 loss=1.290119 Training: 78%|███████▊ | 14123/18018 [30:24:20<9:43:00, 8.98s/step, loss=1.290, acc=0.699, acc_len=1.68, thru=16.5, I=16.3, T=35.4, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:54:35,082] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:54:41,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:54:41,992] loop.py:45 INFO TRAIN_STEP step=14124 loss=1.385257 Training: 78%|███████▊ | 14124/18018 [30:24:28<9:07:08, 8.43s/step, loss=1.385, acc=0.685, acc_len=1.63, thru=17.6, I=22.1, T=36.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:54:42,246] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:54:48,677] loop.py:380 INFO TIMING step=14125: step=3.510s data=1.545s compute=3.477s [fwd=2.520s bwd=0.941s opt=0.003s] dispatch=3.164s [2026-10-19 19:54:48,677] loop.py:45 INFO TRAIN_STEP step=14125 loss=1.333386 Training: 78%|███████▊ | 14125/18018 [30:24:34<8:32:59, 7.91s/step, loss=1.333, acc=0.696, acc_len=1.67, thru=17.1, I=19.4, T=36.5, wait=3.2s, pool=49, epoch=2/2][2026-10-19 19:54:57,819] loop.py:45 INFO TRAIN_STEP step=14126 loss=1.302845 Training: 78%|███████▊ | 14126/18018 [30:24:43<8:56:55, 8.28s/step, loss=1.303, acc=0.698, acc_len=1.67, thru=15.0, I=16.7, T=35.8, wait=5.6s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:54:57,815] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14125: forward=2511.2ms backward=1016.1ms [2026-10-19 19:55:06,041] loop.py:45 INFO TRAIN_STEP step=14127 loss=1.231365 Training: 78%|███████▊ | 14127/18018 [30:24:52<8:55:41, 8.26s/step, loss=1.231, acc=0.713, acc_len=1.72, thru=14.2, I=14.4, T=36.1, wait=4.7s, pool=49, epoch=2/2][2026-10-19 19:55:14,169] loop.py:45 INFO TRAIN_STEP step=14128 loss=1.322606 Training: 78%|███████▊ | 14128/18018 [30:25:00<8:52:59, 8.22s/step, loss=1.323, acc=0.693, acc_len=1.65, thru=16.0, I=16.0, T=34.3, wait=4.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:21,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:55:21,521] loop.py:45 INFO TRAIN_STEP step=14129 loss=1.321313 Training: 78%|███████▊ | 14129/18018 [30:25:07<8:35:57, 7.96s/step, loss=1.321, acc=0.693, acc_len=1.65, thru=17.9, I=18.3, T=36.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:21,604] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:29,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:55:30,677] loop.py:380 INFO TIMING step=14130: step=3.491s data=1.365s compute=3.456s [fwd=2.474s bwd=0.967s opt=0.003s] dispatch=5.656s [2026-10-19 19:55:30,678] loop.py:45 INFO TRAIN_STEP step=14130 loss=1.169666 Training: 78%|███████▊ | 14130/18018 [30:25:16<8:59:02, 8.32s/step, loss=1.170, acc=0.723, acc_len=1.77, thru=13.8, I=14.9, T=36.7, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:30,749] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 19:55:36,715] loop.py:45 INFO TRAIN_STEP step=14131 loss=1.380895 Training: 78%|███████▊ | 14131/18018 [30:25:22<8:14:36, 7.63s/step, loss=1.381, acc=0.684, acc_len=1.62, thru=19.6, I=20.2, T=38.4, wait=2.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:55:36,711] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14130: forward=2372.4ms backward=903.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:42,555] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:55:43,201] loop.py:45 INFO TRAIN_STEP step=14132 loss=1.290582 Training: 78%|███████▊ | 14132/18018 [30:25:29<7:52:08, 7.29s/step, loss=1.291, acc=0.702, acc_len=1.69, thru=20.2, I=21.3, T=38.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:43,560] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:49,983] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:55:50,486] loop.py:45 INFO TRAIN_STEP step=14133 loss=1.250006 Training: 78%|███████▊ | 14133/18018 [30:25:36<7:51:56, 7.29s/step, loss=1.250, acc=0.712, acc_len=1.72, thru=18.2, I=20.8, T=37.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:50,992] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:55,099] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:55:56,098] loop.py:45 INFO TRAIN_STEP step=14134 loss=1.276905 Training: 78%|███████▊ | 14134/18018 [30:25:42<7:19:16, 6.79s/step, loss=1.277, acc=0.705, acc_len=1.69, thru=20.8, I=25.1, T=38.9, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:55:56,130] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 19:56:03,770] loop.py:380 INFO TIMING step=14135: step=3.387s data=1.367s compute=3.339s [fwd=2.385s bwd=0.940s opt=0.003s] dispatch=4.275s [2026-10-19 19:56:03,770] loop.py:45 INFO TRAIN_STEP step=14135 loss=1.312754 Training: 78%|███████▊ | 14135/18018 [30:25:49<7:36:19, 7.05s/step, loss=1.313, acc=0.696, acc_len=1.66, thru=14.5, I=18.3, T=37.8, wait=4.3s, pool=37, epoch=2/2][2026-10-19 19:56:12,712] loop.py:45 INFO TRAIN_STEP step=14136 loss=1.382116 Training: 78%|███████▊ | 14136/18018 [30:25:58<8:12:54, 7.62s/step, loss=1.382, acc=0.689, acc_len=1.64, thru=17.7, I=16.1, T=35.7, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:56:12,707] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14135: forward=2539.8ms backward=989.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:12,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:13,205] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:19,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:56:19,623] loop.py:45 INFO TRAIN_STEP step=14137 loss=1.223184 Training: 78%|███████▊ | 14137/18018 [30:26:05<7:59:05, 7.41s/step, loss=1.223, acc=0.717, acc_len=1.74, thru=18.3, I=19.8, T=37.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:19,891] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 19:56:26,070] loop.py:45 INFO TRAIN_STEP step=14138 loss=1.326076 Training: 78%|███████▊ | 14138/18018 [30:26:12<7:40:19, 7.12s/step, loss=1.326, acc=0.697, acc_len=1.67, thru=19.2, I=21.3, T=37.3, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:31,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:56:32,272] loop.py:45 INFO TRAIN_STEP step=14139 loss=1.310665 Training: 78%|███████▊ | 14139/18018 [30:26:18<7:22:25, 6.84s/step, loss=1.311, acc=0.700, acc_len=1.68, thru=21.1, I=22.0, T=36.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:32,619] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:39,648] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:56:39,804] loop.py:380 INFO TIMING step=14140: step=3.435s data=1.411s compute=3.407s [fwd=2.444s bwd=0.949s opt=0.003s] dispatch=4.088s [2026-10-19 19:56:39,804] loop.py:45 INFO TRAIN_STEP step=14140 loss=1.108292 Training: 78%|███████▊ | 14140/18018 [30:26:25<7:35:39, 7.05s/step, loss=1.108, acc=0.734, acc_len=1.80, thru=16.9, I=18.7, T=37.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:40,151] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:47,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:56:48,171] loop.py:45 INFO TRAIN_STEP step=14141 loss=1.321524 Training: 78%|███████▊ | 14141/18018 [30:26:34<8:01:04, 7.45s/step, loss=1.322, acc=0.693, acc_len=1.66, thru=16.4, I=17.1, T=35.5, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:56:48,166] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14140: forward=2587.1ms backward=975.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:48,642] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:53,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:56:55,178] loop.py:45 INFO TRAIN_STEP step=14142 loss=1.298639 Training: 78%|███████▊ | 14142/18018 [30:26:41<7:52:30, 7.31s/step, loss=1.299, acc=0.699, acc_len=1.68, thru=18.9, I=22.4, T=36.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:56:55,213] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:57:00,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:57:01,188] loop.py:45 INFO TRAIN_STEP step=14143 loss=1.299031 Training: 78%|███████▊ | 14143/18018 [30:26:47<7:27:06, 6.92s/step, loss=1.299, acc=0.699, acc_len=1.68, thru=20.4, I=26.3, T=36.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:57:01,254] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 19:57:07,972] loop.py:45 INFO TRAIN_STEP step=14144 loss=1.107543 Training: 78%|███████▊ | 14144/18018 [30:26:54<7:24:17, 6.88s/step, loss=1.108, acc=0.731, acc_len=1.79, thru=18.9, I=20.6, T=38.4, wait=3.4s, pool=61, epoch=2/2][2026-10-19 19:57:14,082] loop.py:380 INFO TIMING step=14145: step=3.259s data=1.303s compute=3.224s [fwd=2.308s bwd=0.902s opt=0.003s] dispatch=2.842s [2026-10-19 19:57:14,082] loop.py:45 INFO TRAIN_STEP step=14145 loss=1.250177 Training: 79%|███████▊ | 14145/18018 [30:27:00<7:09:13, 6.65s/step, loss=1.250, acc=0.704, acc_len=1.69, thru=17.9, I=17.7, T=39.3, wait=2.8s, pool=33, epoch=2/2][2026-10-19 19:57:22,292] loop.py:45 INFO TRAIN_STEP step=14146 loss=1.313458 Training: 79%|███████▊ | 14146/18018 [30:27:08<7:39:18, 7.12s/step, loss=1.313, acc=0.699, acc_len=1.68, thru=16.1, I=18.1, T=34.7, wait=4.5s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:57:22,288] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14145: forward=2646.9ms backward=989.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:57:29,132] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:57:29,956] loop.py:45 INFO TRAIN_STEP step=14147 loss=1.259985 Training: 79%|███████▊ | 14147/18018 [30:27:16<7:49:48, 7.28s/step, loss=1.260, acc=0.703, acc_len=1.70, thru=17.7, I=18.1, T=36.4, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:57:30,140] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:57:37,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:57:38,801] loop.py:45 INFO TRAIN_STEP step=14148 loss=1.339507 Training: 79%|███████▊ | 14148/18018 [30:27:24<8:19:56, 7.75s/step, loss=1.340, acc=0.693, acc_len=1.66, thru=13.6, I=17.5, T=35.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:57:39,110] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:57:44,496] loop.py:45 INFO TRAIN_STEP step=14149 loss=1.193051 Training: 79%|███████▊ | 14149/18018 [30:27:30<7:39:59, 7.13s/step, loss=1.193, acc=0.717, acc_len=1.74, thru=20.6, I=24.0, T=39.7, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:57:49,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:57:51,104] loop.py:380 INFO TIMING step=14150: step=3.268s data=1.335s compute=3.223s [fwd=2.300s bwd=0.909s opt=0.003s] dispatch=3.331s [2026-10-19 19:57:51,104] loop.py:45 INFO TRAIN_STEP step=14150 loss=1.314331 Training: 79%|███████▊ | 14150/18018 [30:27:37<7:29:44, 6.98s/step, loss=1.314, acc=0.694, acc_len=1.66, thru=19.9, I=22.7, T=39.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:57:51,471] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:57:58,289] loop.py:45 INFO TRAIN_STEP step=14151 loss=1.218057 Training: 79%|███████▊ | 14151/18018 [30:27:44<7:33:36, 7.04s/step, loss=1.218, acc=0.712, acc_len=1.72, thru=17.7, I=21.2, T=36.7, wait=3.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:57:58,284] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14150: forward=2436.1ms backward=996.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:58:07,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:58:08,541] loop.py:45 INFO TRAIN_STEP step=14152 loss=1.177818 Training: 79%|███████▊ | 14152/18018 [30:27:54<8:35:40, 8.00s/step, loss=1.178, acc=0.724, acc_len=1.77, thru=14.3, I=14.7, T=34.3, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:58:08,866] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:58:15,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:58:16,144] loop.py:45 INFO TRAIN_STEP step=14153 loss=1.197685 Training: 79%|███████▊ | 14153/18018 [30:28:02<8:27:48, 7.88s/step, loss=1.198, acc=0.721, acc_len=1.75, thru=16.4, I=20.8, T=38.4, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:58:16,187] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 19:58:25,815] loop.py:45 INFO TRAIN_STEP step=14154 loss=1.156188 Training: 79%|███████▊ | 14154/18018 [30:28:11<9:02:13, 8.42s/step, loss=1.156, acc=0.728, acc_len=1.78, thru=13.3, I=12.8, T=33.6, wait=5.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:58:35,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:58:36,003] loop.py:380 INFO TIMING step=14155: step=3.729s data=1.594s compute=3.698s [fwd=2.646s bwd=1.037s opt=0.003s] dispatch=6.449s [2026-10-19 19:58:36,003] loop.py:45 INFO TRAIN_STEP step=14155 loss=1.394150 Training: 79%|███████▊ | 14155/18018 [30:28:22<9:36:12, 8.95s/step, loss=1.394, acc=0.686, acc_len=1.63, thru=13.2, I=13.7, T=34.3, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:58:36,259] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:58:43,305] loop.py:45 INFO TRAIN_STEP step=14156 loss=1.107474 Training: 79%|███████▊ | 14156/18018 [30:28:29<9:04:12, 8.45s/step, loss=1.107, acc=0.734, acc_len=1.80, thru=15.9, I=16.8, T=37.3, wait=3.9s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:58:43,301] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14155: forward=2416.0ms backward=964.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:58:51,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:58:52,311] loop.py:45 INFO TRAIN_STEP step=14157 loss=1.419926 Training: 79%|███████▊ | 14157/18018 [30:28:38<9:14:43, 8.62s/step, loss=1.420, acc=0.681, acc_len=1.62, thru=17.4, I=17.8, T=34.3, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:58:52,677] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 19:58:59,037] loop.py:45 INFO TRAIN_STEP step=14158 loss=1.176256 Training: 79%|███████▊ | 14158/18018 [30:28:45<8:38:02, 8.05s/step, loss=1.176, acc=0.722, acc_len=1.76, thru=18.1, I=23.4, T=37.8, wait=3.3s, pool=61, epoch=2/2][2026-10-19 19:59:11,187] loop.py:45 INFO TRAIN_STEP step=14159 loss=1.297646 Training: 79%|███████▊ | 14159/18018 [30:28:57<9:56:56, 9.28s/step, loss=1.298, acc=0.701, acc_len=1.69, thru=10.8, I=9.0, T=31.6, wait=8.1s, pool=49, epoch=2/2] [2026-10-19 19:59:20,945] loop.py:380 INFO TIMING step=14160: step=3.880s data=1.623s compute=3.785s [fwd=2.656s bwd=1.114s opt=0.003s] dispatch=5.869s [2026-10-19 19:59:20,945] loop.py:45 INFO TRAIN_STEP step=14160 loss=1.311636 Training: 79%|███████▊ | 14160/18018 [30:29:07<10:06:09, 9.43s/step, loss=1.312, acc=0.699, acc_len=1.67, thru=13.0, I=12.5, T=33.0, wait=5.9s, pool=45, epoch=2/2][2026-10-19 19:59:30,272] loop.py:45 INFO TRAIN_STEP step=14161 loss=1.253688 Training: 79%|███████▊ | 14161/18018 [30:29:16<10:03:54, 9.39s/step, loss=1.254, acc=0.707, acc_len=1.70, thru=16.1, I=14.4, T=36.1, wait=5.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 19:59:30,268] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14160: forward=2514.7ms backward=981.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:59:34,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:59:36,046] loop.py:45 INFO TRAIN_STEP step=14162 loss=1.150225 Training: 79%|███████▊ | 14162/18018 [30:29:22<8:53:58, 8.31s/step, loss=1.150, acc=0.728, acc_len=1.78, thru=20.5, I=20.3, T=37.0, wait=2.3s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:59:36,406] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:59:41,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:59:42,830] loop.py:45 INFO TRAIN_STEP step=14163 loss=1.229413 Training: 79%|███████▊ | 14163/18018 [30:29:28<8:24:26, 7.85s/step, loss=1.229, acc=0.712, acc_len=1.72, thru=18.8, I=23.3, T=37.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:59:43,243] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:59:48,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 19:59:49,077] loop.py:45 INFO TRAIN_STEP step=14164 loss=1.200700 Training: 79%|███████▊ | 14164/18018 [30:29:35<7:53:24, 7.37s/step, loss=1.201, acc=0.716, acc_len=1.74, thru=20.3, I=22.9, T=37.7, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 19:59:49,514] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 19:59:54,915] loop.py:380 INFO TIMING step=14165: step=3.264s data=1.334s compute=3.234s [fwd=2.329s bwd=0.891s opt=0.003s] dispatch=2.563s [2026-10-19 19:59:54,915] loop.py:45 INFO TRAIN_STEP step=14165 loss=1.281395 Training: 79%|███████▊ | 14165/18018 [30:29:40<7:23:43, 6.91s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=19.7, I=21.4, T=39.2, wait=2.6s, pool=53, epoch=2/2][2026-10-19 20:00:03,589] loop.py:45 INFO TRAIN_STEP step=14166 loss=1.403889 Training: 79%|███████▊ | 14166/18018 [30:29:49<7:57:39, 7.44s/step, loss=1.404, acc=0.684, acc_len=1.63, thru=15.7, I=14.9, T=34.9, wait=5.0s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:00:03,584] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14165: forward=2534.4ms backward=1076.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:09,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:00:09,467] loop.py:45 INFO TRAIN_STEP step=14167 loss=1.243853 Training: 79%|███████▊ | 14167/18018 [30:29:55<7:27:23, 6.97s/step, loss=1.244, acc=0.708, acc_len=1.71, thru=20.8, I=21.1, T=38.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:09,565] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:17,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:00:18,515] loop.py:45 INFO TRAIN_STEP step=14168 loss=1.185715 Training: 79%|███████▊ | 14168/18018 [30:30:04<8:07:18, 7.59s/step, loss=1.186, acc=0.720, acc_len=1.74, thru=13.8, I=15.3, T=35.2, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:18,673] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:23,286] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:00:24,101] loop.py:45 INFO TRAIN_STEP step=14169 loss=1.168985 Training: 79%|███████▊ | 14169/18018 [30:30:10<7:28:31, 6.99s/step, loss=1.169, acc=0.724, acc_len=1.76, thru=20.8, I=20.1, T=39.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:24,293] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:29,558] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:00:31,216] loop.py:380 INFO TIMING step=14170: step=3.463s data=1.462s compute=3.413s [fwd=2.429s bwd=0.970s opt=0.003s] dispatch=3.641s [2026-10-19 20:00:31,216] loop.py:45 INFO TRAIN_STEP step=14170 loss=1.342437 Training: 79%|███████▊ | 14170/18018 [30:30:17<7:30:46, 7.03s/step, loss=1.342, acc=0.692, acc_len=1.65, thru=18.3, I=23.0, T=37.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:31,574] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 20:00:37,072] loop.py:45 INFO TRAIN_STEP step=14171 loss=1.304771 Training: 79%|███████▊ | 14171/18018 [30:30:23<7:08:04, 6.68s/step, loss=1.305, acc=0.699, acc_len=1.67, thru=19.5, I=27.1, T=39.6, wait=2.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:00:37,067] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14170: forward=2311.5ms backward=877.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:45,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:00:47,250] loop.py:45 INFO TRAIN_STEP step=14172 loss=1.176885 Training: 79%|███████▊ | 14172/18018 [30:30:33<8:15:17, 7.73s/step, loss=1.177, acc=0.720, acc_len=1.75, thru=14.6, I=15.0, T=34.7, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:00:47,505] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:00:53,003] loop.py:45 INFO TRAIN_STEP step=14173 loss=1.319215 Training: 79%|███████▊ | 14173/18018 [30:30:39<7:37:16, 7.14s/step, loss=1.319, acc=0.694, acc_len=1.66, thru=19.7, I=26.3, T=38.7, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:01,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:01:02,279] loop.py:45 INFO TRAIN_STEP step=14174 loss=1.183567 Training: 79%|███████▊ | 14174/18018 [30:30:48<8:18:14, 7.78s/step, loss=1.184, acc=0.723, acc_len=1.76, thru=14.9, I=14.2, T=36.0, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:02,763] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:01:08,401] loop.py:380 INFO TIMING step=14175: step=3.302s data=1.336s compute=3.271s [fwd=2.346s bwd=0.910s opt=0.003s] dispatch=2.811s [2026-10-19 20:01:08,401] loop.py:45 INFO TRAIN_STEP step=14175 loss=1.240106 Training: 79%|███████▊ | 14175/18018 [30:30:54<7:46:19, 7.28s/step, loss=1.240, acc=0.710, acc_len=1.72, thru=18.1, I=20.0, T=38.8, wait=2.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:14,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:01:15,847] loop.py:45 INFO TRAIN_STEP step=14176 loss=1.203928 Training: 79%|███████▊ | 14176/18018 [30:31:01<7:49:24, 7.33s/step, loss=1.204, acc=0.719, acc_len=1.74, thru=17.7, I=19.3, T=35.8, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:01:15,842] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14175: forward=2523.2ms backward=1001.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:16,172] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:21,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:01:21,679] loop.py:45 INFO TRAIN_STEP step=14177 loss=1.196500 Training: 79%|███████▊ | 14177/18018 [30:31:07<7:20:28, 6.88s/step, loss=1.196, acc=0.718, acc_len=1.74, thru=20.3, I=24.1, T=38.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:22,162] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:27,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:01:28,543] loop.py:45 INFO TRAIN_STEP step=14178 loss=1.169440 Training: 79%|███████▊ | 14178/18018 [30:31:14<7:20:03, 6.88s/step, loss=1.169, acc=0.722, acc_len=1.75, thru=18.9, I=22.1, T=38.3, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:28,763] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:01:34,289] loop.py:45 INFO TRAIN_STEP step=14179 loss=1.304232 Training: 79%|███████▊ | 14179/18018 [30:31:20<6:58:14, 6.54s/step, loss=1.304, acc=0.699, acc_len=1.67, thru=20.2, I=22.8, T=38.3, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:41,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:01:41,152] loop.py:380 INFO TIMING step=14180: step=3.399s data=1.426s compute=3.371s [fwd=2.426s bwd=0.931s opt=0.003s] dispatch=3.455s [2026-10-19 20:01:41,152] loop.py:45 INFO TRAIN_STEP step=14180 loss=1.369976 Training: 79%|███████▊ | 14180/18018 [30:31:27<7:04:21, 6.63s/step, loss=1.370, acc=0.683, acc_len=1.62, thru=19.5, I=20.2, T=37.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:41,595] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:48,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:01:48,651] loop.py:45 INFO TRAIN_STEP step=14181 loss=1.314518 Training: 79%|███████▊ | 14181/18018 [30:31:34<7:20:51, 6.89s/step, loss=1.315, acc=0.698, acc_len=1.67, thru=17.7, I=17.8, T=37.4, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:01:48,646] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14180: forward=2427.6ms backward=946.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:01:48,729] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 20:01:53,746] loop.py:45 INFO TRAIN_STEP step=14182 loss=1.300436 Training: 79%|███████▊ | 14182/18018 [30:31:39<6:46:14, 6.35s/step, loss=1.300, acc=0.701, acc_len=1.68, thru=19.6, I=21.2, T=39.0, wait=1.8s, pool=37, epoch=2/2][2026-10-19 20:02:02,030] loop.py:45 INFO TRAIN_STEP step=14183 loss=1.361130 Training: 79%|███████▊ | 14183/18018 [30:31:48<7:23:09, 6.93s/step, loss=1.361, acc=0.686, acc_len=1.63, thru=15.3, I=15.3, T=36.2, wait=4.7s, pool=45, epoch=2/2][2026-10-19 20:02:08,954] loop.py:45 INFO TRAIN_STEP step=14184 loss=1.285309 Training: 79%|███████▊ | 14184/18018 [30:31:55<7:22:50, 6.93s/step, loss=1.285, acc=0.699, acc_len=1.68, thru=18.2, I=15.9, T=36.0, wait=3.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:02:15,330] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:02:15,966] loop.py:380 INFO TIMING step=14185: step=3.378s data=1.406s compute=3.347s [fwd=2.414s bwd=0.919s opt=0.003s] dispatch=3.625s [2026-10-19 20:02:15,967] loop.py:45 INFO TRAIN_STEP step=14185 loss=1.139987 Training: 79%|███████▊ | 14185/18018 [30:32:02<7:24:23, 6.96s/step, loss=1.140, acc=0.728, acc_len=1.78, thru=19.1, I=19.8, T=37.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:02:16,335] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:02:21,145] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:02:22,322] loop.py:45 INFO TRAIN_STEP step=14186 loss=1.355966 Training: 79%|███████▊ | 14186/18018 [30:32:08<7:12:41, 6.77s/step, loss=1.356, acc=0.689, acc_len=1.65, thru=20.1, I=24.7, T=38.0, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:02:22,317] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14185: forward=2418.1ms backward=905.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:02:22,655] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:02:28,953] loop.py:45 INFO TRAIN_STEP step=14187 loss=1.360017 Training: 79%|███████▊ | 14187/18018 [30:32:15<7:09:51, 6.73s/step, loss=1.360, acc=0.687, acc_len=1.64, thru=17.0, I=22.7, T=36.6, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:02:37,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:02:37,982] loop.py:45 INFO TRAIN_STEP step=14188 loss=1.372089 Training: 79%|███████▊ | 14188/18018 [30:32:24<7:53:42, 7.42s/step, loss=1.372, acc=0.684, acc_len=1.62, thru=15.7, I=16.4, T=35.0, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:02:38,285] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:02:45,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:02:45,932] loop.py:45 INFO TRAIN_STEP step=14189 loss=1.210229 Training: 79%|███████▊ | 14189/18018 [30:32:32<8:03:40, 7.58s/step, loss=1.210, acc=0.716, acc_len=1.73, thru=16.5, I=17.8, T=37.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:02:46,298] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:02:52,784] loop.py:380 INFO TIMING step=14190: step=3.290s data=1.340s compute=3.259s [fwd=2.357s bwd=0.887s opt=0.003s] dispatch=3.554s [2026-10-19 20:02:52,784] loop.py:45 INFO TRAIN_STEP step=14190 loss=1.194105 Training: 79%|███████▉ | 14190/18018 [30:32:38<7:49:38, 7.36s/step, loss=1.194, acc=0.719, acc_len=1.74, thru=18.7, I=19.9, T=38.9, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:03:00,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:03:00,747] loop.py:45 INFO TRAIN_STEP step=14191 loss=1.299591 Training: 79%|███████▉ | 14191/18018 [30:32:46<8:01:00, 7.54s/step, loss=1.300, acc=0.701, acc_len=1.68, thru=18.3, I=17.1, T=33.9, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:03:00,742] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14190: forward=2672.2ms backward=1048.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:03:01,173] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:03:07,646] loop.py:45 INFO TRAIN_STEP step=14192 loss=1.348267 Training: 79%|███████▉ | 14192/18018 [30:32:53<7:48:38, 7.35s/step, loss=1.348, acc=0.694, acc_len=1.66, thru=15.0, I=18.5, T=35.7, wait=3.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:03:18,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:03:18,931] loop.py:45 INFO TRAIN_STEP step=14193 loss=1.269556 Training: 79%|███████▉ | 14193/18018 [30:33:05<9:03:48, 8.53s/step, loss=1.270, acc=0.704, acc_len=1.69, thru=18.6, I=12.3, T=35.2, wait=7.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:03:19,027] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 20:03:24,832] loop.py:45 INFO TRAIN_STEP step=14194 loss=1.186578 Training: 79%|███████▉ | 14194/18018 [30:33:10<8:13:21, 7.74s/step, loss=1.187, acc=0.718, acc_len=1.74, thru=19.7, I=19.7, T=38.1, wait=2.5s, pool=53, epoch=2/2][2026-10-19 20:03:32,597] loop.py:380 INFO TIMING step=14195: step=3.416s data=1.384s compute=3.384s [fwd=2.432s bwd=0.937s opt=0.003s] dispatch=4.341s [2026-10-19 20:03:32,597] loop.py:45 INFO TRAIN_STEP step=14195 loss=1.272243 Training: 79%|███████▉ | 14195/18018 [30:33:18<8:13:41, 7.75s/step, loss=1.272, acc=0.706, acc_len=1.69, thru=16.1, I=15.3, T=37.5, wait=4.3s, pool=41, epoch=2/2][2026-10-19 20:03:41,323] loop.py:45 INFO TRAIN_STEP step=14196 loss=1.414901 Training: 79%|███████▉ | 14196/18018 [30:33:27<8:32:12, 8.04s/step, loss=1.415, acc=0.681, acc_len=1.62, thru=13.7, I=13.9, T=35.9, wait=5.2s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:03:41,319] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14195: forward=2548.9ms backward=963.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:03:51,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:03:53,189] loop.py:45 INFO TRAIN_STEP step=14197 loss=1.404685 Training: 79%|███████▉ | 14197/18018 [30:33:39<9:45:13, 9.19s/step, loss=1.405, acc=0.682, acc_len=1.62, thru=14.6, I=16.6, T=32.7, wait=7.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:03:53,455] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 20:03:59,501] loop.py:45 INFO TRAIN_STEP step=14198 loss=1.241862 Training: 79%|███████▉ | 14198/18018 [30:33:45<8:50:04, 8.33s/step, loss=1.242, acc=0.709, acc_len=1.70, thru=16.6, I=22.3, T=37.4, wait=2.9s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:04:11,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:04:12,050] loop.py:45 INFO TRAIN_STEP step=14199 loss=1.304492 Training: 79%|███████▉ | 14199/18018 [30:33:58<10:10:33, 9.59s/step, loss=1.304, acc=0.701, acc_len=1.69, thru=13.8, I=12.7, T=34.0, wait=8.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:04:12,349] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:04:18,166] loop.py:380 INFO TIMING step=14200: step=3.356s data=1.390s compute=3.323s [fwd=2.372s bwd=0.936s opt=0.003s] dispatch=2.752s [2026-10-19 20:04:18,167] loop.py:45 INFO TRAIN_STEP step=14200 loss=1.278937 Training: 79%|███████▉ | 14200/18018 [30:34:04<9:04:02, 8.55s/step, loss=1.279, acc=0.703, acc_len=1.69, thru=19.8, I=20.3, T=38.1, wait=2.8s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:04:18,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:04:18,669] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:04:28,596] loop.py:45 INFO TRAIN_STEP step=14201 loss=1.302493 Training: 79%|███████▉ | 14201/18018 [30:34:14<9:39:49, 9.11s/step, loss=1.302, acc=0.699, acc_len=1.68, thru=11.3, I=11.3, T=32.9, wait=6.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:04:28,591] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14200: forward=2687.1ms backward=1147.1ms [2026-10-19 20:04:35,759] loop.py:45 INFO TRAIN_STEP step=14202 loss=1.344990 Training: 79%|███████▉ | 14202/18018 [30:34:21<9:02:26, 8.53s/step, loss=1.345, acc=0.692, acc_len=1.65, thru=16.9, I=16.1, T=37.7, wait=3.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:04:41,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:04:42,206] loop.py:45 INFO TRAIN_STEP step=14203 loss=1.206865 Training: 79%|███████▉ | 14203/18018 [30:34:28<8:22:33, 7.90s/step, loss=1.207, acc=0.715, acc_len=1.73, thru=19.9, I=20.6, T=34.7, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:04:42,478] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:04:52,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:04:52,579] loop.py:45 INFO TRAIN_STEP step=14204 loss=1.319354 Training: 79%|███████▉ | 14204/18018 [30:34:38<9:09:32, 8.65s/step, loss=1.319, acc=0.702, acc_len=1.70, thru=13.0, I=12.8, T=34.4, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:04:52,729] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:04:59,278] loop.py:380 INFO TIMING step=14205: step=3.497s data=1.424s compute=3.468s [fwd=2.500s bwd=0.953s opt=0.003s] dispatch=3.192s [2026-10-19 20:04:59,279] loop.py:45 INFO TRAIN_STEP step=14205 loss=1.226396 Training: 79%|███████▉ | 14205/18018 [30:34:45<8:32:24, 8.06s/step, loss=1.226, acc=0.713, acc_len=1.72, thru=16.6, I=17.0, T=36.6, wait=3.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:06,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:05:06,806] loop.py:45 INFO TRAIN_STEP step=14206 loss=1.365208 Training: 79%|███████▉ | 14206/18018 [30:34:52<8:21:55, 7.90s/step, loss=1.365, acc=0.686, acc_len=1.63, thru=18.5, I=17.3, T=32.6, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:05:06,801] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14205: forward=2916.3ms backward=968.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:07,067] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:05:12,812] loop.py:45 INFO TRAIN_STEP step=14207 loss=1.341235 Training: 79%|███████▉ | 14207/18018 [30:34:58<7:45:45, 7.33s/step, loss=1.341, acc=0.691, acc_len=1.64, thru=18.1, I=21.4, T=38.6, wait=2.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:20,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:05:21,615] loop.py:45 INFO TRAIN_STEP step=14208 loss=1.326909 Training: 79%|███████▉ | 14208/18018 [30:35:07<8:13:36, 7.77s/step, loss=1.327, acc=0.697, acc_len=1.67, thru=16.1, I=17.4, T=36.4, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:21,767] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:26,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:05:27,581] loop.py:45 INFO TRAIN_STEP step=14209 loss=1.146230 Training: 79%|███████▉ | 14209/18018 [30:35:13<7:39:03, 7.23s/step, loss=1.146, acc=0.728, acc_len=1.77, thru=20.6, I=25.9, T=38.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:28,033] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:34,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:05:34,378] loop.py:380 INFO TIMING step=14210: step=3.320s data=1.282s compute=3.293s [fwd=2.382s bwd=0.895s opt=0.003s] dispatch=3.468s [2026-10-19 20:05:34,378] loop.py:45 INFO TRAIN_STEP step=14210 loss=1.233207 Training: 79%|███████▉ | 14210/18018 [30:35:20<7:30:41, 7.10s/step, loss=1.233, acc=0.708, acc_len=1.71, thru=18.6, I=22.6, T=38.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:34,570] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:05:41,459] loop.py:45 INFO TRAIN_STEP step=14211 loss=1.235921 Training: 79%|███████▉ | 14211/18018 [30:35:27<7:30:10, 7.10s/step, loss=1.236, acc=0.711, acc_len=1.72, thru=14.5, I=21.0, T=34.6, wait=3.4s, pool=25, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:05:41,454] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14210: forward=2359.3ms backward=1072.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:49,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:05:49,917] loop.py:45 INFO TRAIN_STEP step=14212 loss=1.265853 Training: 79%|███████▉ | 14212/18018 [30:35:36<7:56:01, 7.50s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=19.9, I=16.4, T=37.1, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:05:49,952] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 20:05:57,244] loop.py:45 INFO TRAIN_STEP step=14213 loss=1.159135 Training: 79%|███████▉ | 14213/18018 [30:35:43<7:52:29, 7.45s/step, loss=1.159, acc=0.726, acc_len=1.78, thru=18.4, I=17.3, T=37.5, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:02,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:06:03,668] loop.py:45 INFO TRAIN_STEP step=14214 loss=1.293408 Training: 79%|███████▉ | 14214/18018 [30:35:49<7:32:49, 7.14s/step, loss=1.293, acc=0.698, acc_len=1.67, thru=19.9, I=20.4, T=38.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:04,095] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:09,357] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:06:10,702] loop.py:380 INFO TIMING step=14215: step=3.434s data=1.386s compute=3.406s [fwd=2.457s bwd=0.936s opt=0.003s] dispatch=3.592s [2026-10-19 20:06:10,703] loop.py:45 INFO TRAIN_STEP step=14215 loss=1.259140 Training: 79%|███████▉ | 14215/18018 [30:35:56<7:30:39, 7.11s/step, loss=1.259, acc=0.708, acc_len=1.71, thru=19.0, I=25.4, T=37.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:10,865] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:06:18,170] loop.py:45 INFO TRAIN_STEP step=14216 loss=1.141038 Training: 79%|███████▉ | 14216/18018 [30:36:04<7:37:24, 7.22s/step, loss=1.141, acc=0.725, acc_len=1.77, thru=17.0, I=21.8, T=39.2, wait=4.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:06:18,165] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14215: forward=2307.7ms backward=917.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:24,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:06:25,675] loop.py:45 INFO TRAIN_STEP step=14217 loss=1.259195 Training: 79%|███████▉ | 14217/18018 [30:36:11<7:42:40, 7.30s/step, loss=1.259, acc=0.706, acc_len=1.70, thru=18.3, I=18.7, T=35.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:25,795] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:33,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:06:34,298] loop.py:45 INFO TRAIN_STEP step=14218 loss=1.300650 Training: 79%|███████▉ | 14218/18018 [30:36:20<8:07:37, 7.70s/step, loss=1.301, acc=0.699, acc_len=1.68, thru=14.6, I=15.7, T=36.6, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:34,788] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:39,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:06:41,903] loop.py:45 INFO TRAIN_STEP step=14219 loss=1.222096 Training: 79%|███████▉ | 14219/18018 [30:36:27<8:05:43, 7.67s/step, loss=1.222, acc=0.712, acc_len=1.72, thru=18.0, I=26.8, T=27.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:42,189] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:46,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:06:47,789] loop.py:380 INFO TIMING step=14220: step=3.232s data=1.302s compute=3.203s [fwd=2.288s bwd=0.901s opt=0.003s] dispatch=2.643s [2026-10-19 20:06:47,789] loop.py:45 INFO TRAIN_STEP step=14220 loss=1.098865 Training: 79%|███████▉ | 14220/18018 [30:36:33<7:31:41, 7.14s/step, loss=1.099, acc=0.736, acc_len=1.81, thru=17.7, I=28.5, T=39.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:06:47,993] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:06:53,456] loop.py:45 INFO TRAIN_STEP step=14221 loss=1.322781 Training: 79%|███████▉ | 14221/18018 [30:36:39<7:03:44, 6.70s/step, loss=1.323, acc=0.694, acc_len=1.65, thru=20.6, I=23.7, T=40.0, wait=2.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:06:53,451] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14220: forward=2269.3ms backward=887.4ms [2026-10-19 20:07:02,066] loop.py:45 INFO TRAIN_STEP step=14222 loss=1.300864 Training: 79%|███████▉ | 14222/18018 [30:36:48<7:39:54, 7.27s/step, loss=1.301, acc=0.705, acc_len=1.70, thru=15.2, I=17.0, T=28.6, wait=4.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:10,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:07:11,017] loop.py:45 INFO TRAIN_STEP step=14223 loss=1.279958 Training: 79%|███████▉ | 14223/18018 [30:36:57<8:11:40, 7.77s/step, loss=1.280, acc=0.706, acc_len=1.70, thru=14.6, I=14.6, T=32.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:11,178] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:17,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:07:18,565] loop.py:45 INFO TRAIN_STEP step=14224 loss=1.265500 Training: 79%|███████▉ | 14224/18018 [30:37:04<8:07:15, 7.71s/step, loss=1.266, acc=0.706, acc_len=1.71, thru=16.4, I=17.5, T=37.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:18,903] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:25,140] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:07:25,455] loop.py:380 INFO TIMING step=14225: step=3.473s data=1.370s compute=3.441s [fwd=2.484s bwd=0.943s opt=0.003s] dispatch=3.409s [2026-10-19 20:07:25,456] loop.py:45 INFO TRAIN_STEP step=14225 loss=1.232153 Training: 79%|███████▉ | 14225/18018 [30:37:11<7:51:40, 7.46s/step, loss=1.232, acc=0.712, acc_len=1.72, thru=18.9, I=19.9, T=36.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:25,642] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:30,780] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:07:31,485] loop.py:45 INFO TRAIN_STEP step=14226 loss=1.334586 Training: 79%|███████▉ | 14226/18018 [30:37:17<7:24:26, 7.03s/step, loss=1.335, acc=0.693, acc_len=1.66, thru=20.6, I=21.5, T=38.6, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:07:31,481] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14225: forward=2382.5ms backward=889.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:31,787] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:07:40,749] loop.py:45 INFO TRAIN_STEP step=14227 loss=1.239870 Training: 79%|███████▉ | 14227/18018 [30:37:26<8:06:35, 7.70s/step, loss=1.240, acc=0.712, acc_len=1.72, thru=11.2, I=12.8, T=35.0, wait=5.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:47,034] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:07:48,866] loop.py:45 INFO TRAIN_STEP step=14228 loss=1.177820 Training: 79%|███████▉ | 14228/18018 [30:37:34<8:14:21, 7.83s/step, loss=1.178, acc=0.722, acc_len=1.76, thru=16.7, I=19.6, T=30.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:49,050] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:56,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:07:56,921] loop.py:45 INFO TRAIN_STEP step=14229 loss=1.243282 Training: 79%|███████▉ | 14229/18018 [30:37:43<8:18:34, 7.90s/step, loss=1.243, acc=0.708, acc_len=1.70, thru=13.3, I=17.1, T=36.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:07:57,076] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:08:03,839] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:08:04,296] loop.py:380 INFO TIMING step=14230: step=3.551s data=1.528s compute=3.505s [fwd=2.481s bwd=1.011s opt=0.003s] dispatch=3.814s [2026-10-19 20:08:04,297] loop.py:45 INFO TRAIN_STEP step=14230 loss=1.205680 Training: 79%|███████▉ | 14230/18018 [30:37:50<8:08:35, 7.74s/step, loss=1.206, acc=0.710, acc_len=1.71, thru=18.9, I=18.8, T=36.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:08:04,349] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 20:08:10,132] loop.py:45 INFO TRAIN_STEP step=14231 loss=1.220063 Training: 79%|███████▉ | 14231/18018 [30:37:56<7:32:21, 7.17s/step, loss=1.220, acc=0.716, acc_len=1.74, thru=19.3, I=20.7, T=38.8, wait=2.5s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:08:10,127] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14230: forward=2366.0ms backward=894.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:08:17,916] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:08:18,132] loop.py:45 INFO TRAIN_STEP step=14232 loss=1.317339 Training: 79%|███████▉ | 14232/18018 [30:38:04<7:48:03, 7.42s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=18.0, I=17.2, T=35.5, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:08:18,417] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:08:23,891] loop.py:45 INFO TRAIN_STEP step=14233 loss=1.293825 Training: 79%|███████▉ | 14233/18018 [30:38:09<7:16:31, 6.92s/step, loss=1.294, acc=0.700, acc_len=1.68, thru=19.7, I=21.4, T=39.1, wait=2.5s, pool=49, epoch=2/2][2026-10-19 20:08:32,156] loop.py:45 INFO TRAIN_STEP step=14234 loss=1.143257 Training: 79%|███████▉ | 14234/18018 [30:38:18<7:41:50, 7.32s/step, loss=1.143, acc=0.730, acc_len=1.79, thru=15.7, I=16.0, T=36.3, wait=4.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:08:38,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:08:39,947] loop.py:380 INFO TIMING step=14235: step=3.682s data=1.583s compute=3.648s [fwd=2.572s bwd=1.061s opt=0.003s] dispatch=4.101s [2026-10-19 20:08:39,948] loop.py:45 INFO TRAIN_STEP step=14235 loss=1.196951 Training: 79%|███████▉ | 14235/18018 [30:38:26<7:50:35, 7.46s/step, loss=1.197, acc=0.717, acc_len=1.74, thru=17.7, I=19.7, T=34.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:08:39,995] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 20:08:45,298] loop.py:45 INFO TRAIN_STEP step=14236 loss=1.282376 Training: 79%|███████▉ | 14236/18018 [30:38:31<7:10:30, 6.83s/step, loss=1.282, acc=0.701, acc_len=1.68, thru=22.1, I=27.6, T=39.8, wait=2.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:08:45,293] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14235: forward=2295.1ms backward=874.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:08:52,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:08:52,643] loop.py:45 INFO TRAIN_STEP step=14237 loss=1.298939 Training: 79%|███████▉ | 14237/18018 [30:38:38<7:20:07, 6.98s/step, loss=1.299, acc=0.703, acc_len=1.69, thru=18.0, I=19.8, T=36.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:08:52,805] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:09:01,296] loop.py:45 INFO TRAIN_STEP step=14238 loss=1.248441 Training: 79%|███████▉ | 14238/18018 [30:38:47<7:51:34, 7.49s/step, loss=1.248, acc=0.712, acc_len=1.73, thru=13.3, I=12.8, T=37.9, wait=5.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:08,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:09:09,374] loop.py:45 INFO TRAIN_STEP step=14239 loss=1.335450 Training: 79%|███████▉ | 14239/18018 [30:38:55<8:02:39, 7.66s/step, loss=1.335, acc=0.697, acc_len=1.67, thru=18.3, I=18.0, T=37.5, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:09,623] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:15,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:09:16,202] loop.py:380 INFO TIMING step=14240: step=3.601s data=1.537s compute=3.573s [fwd=2.565s bwd=0.994s opt=0.003s] dispatch=3.217s [2026-10-19 20:09:16,202] loop.py:45 INFO TRAIN_STEP step=14240 loss=1.241508 Training: 79%|███████▉ | 14240/18018 [30:39:02<7:46:45, 7.41s/step, loss=1.242, acc=0.711, acc_len=1.73, thru=18.7, I=23.9, T=35.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:16,273] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:21,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:09:21,709] loop.py:45 INFO TRAIN_STEP step=14241 loss=1.321424 Training: 79%|███████▉ | 14241/18018 [30:39:07<7:10:37, 6.84s/step, loss=1.321, acc=0.693, acc_len=1.65, thru=22.2, I=27.3, T=39.8, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:09:21,704] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14240: forward=2296.8ms backward=875.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:21,806] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:30,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:09:31,617] loop.py:45 INFO TRAIN_STEP step=14242 loss=1.248117 Training: 79%|███████▉ | 14242/18018 [30:39:17<8:08:25, 7.76s/step, loss=1.248, acc=0.711, acc_len=1.72, thru=13.4, I=13.5, T=34.2, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:32,114] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:38,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:09:38,416] loop.py:45 INFO TRAIN_STEP step=14243 loss=1.173857 Training: 79%|███████▉ | 14243/18018 [30:39:24<7:50:05, 7.47s/step, loss=1.174, acc=0.723, acc_len=1.76, thru=18.9, I=20.8, T=39.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:38,566] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:09:46,233] loop.py:45 INFO TRAIN_STEP step=14244 loss=1.238614 Training: 79%|███████▉ | 14244/18018 [30:39:32<7:56:32, 7.58s/step, loss=1.239, acc=0.713, acc_len=1.72, thru=15.3, I=15.1, T=35.3, wait=4.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:55,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:09:56,541] loop.py:380 INFO TIMING step=14245: step=4.037s data=1.553s compute=4.008s [fwd=2.964s bwd=1.030s opt=0.003s] dispatch=6.261s [2026-10-19 20:09:56,541] loop.py:45 INFO TRAIN_STEP step=14245 loss=1.270528 Training: 79%|███████▉ | 14245/18018 [30:39:42<8:47:54, 8.40s/step, loss=1.271, acc=0.707, acc_len=1.71, thru=15.0, I=16.2, T=31.7, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:09:56,893] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:02,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:10:02,681] loop.py:45 INFO TRAIN_STEP step=14246 loss=1.306509 Training: 79%|███████▉ | 14246/18018 [30:39:48<8:05:16, 7.72s/step, loss=1.307, acc=0.693, acc_len=1.65, thru=19.6, I=24.0, T=36.3, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:10:02,676] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14245: forward=2514.2ms backward=970.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:02,955] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:11,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:10:12,069] loop.py:45 INFO TRAIN_STEP step=14247 loss=1.237802 Training: 79%|███████▉ | 14247/18018 [30:39:58<8:36:36, 8.22s/step, loss=1.238, acc=0.711, acc_len=1.72, thru=14.0, I=14.7, T=34.6, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:12,481] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:18,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:10:18,742] loop.py:45 INFO TRAIN_STEP step=14248 loss=1.288928 Training: 79%|███████▉ | 14248/18018 [30:40:04<8:07:15, 7.75s/step, loss=1.289, acc=0.706, acc_len=1.70, thru=18.9, I=18.5, T=36.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:19,124] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:10:26,527] loop.py:45 INFO TRAIN_STEP step=14249 loss=1.321729 Training: 79%|███████▉ | 14249/18018 [30:40:12<8:07:43, 7.76s/step, loss=1.322, acc=0.692, acc_len=1.66, thru=17.0, I=17.1, T=37.9, wait=4.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:33,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:10:34,507] loop.py:380 INFO TIMING step=14250: step=3.382s data=1.337s compute=3.352s [fwd=2.396s bwd=0.941s opt=0.003s] dispatch=4.589s [2026-10-19 20:10:34,507] loop.py:45 INFO TRAIN_STEP step=14250 loss=1.281963 Training: 79%|███████▉ | 14250/18018 [30:40:20<8:11:39, 7.83s/step, loss=1.282, acc=0.707, acc_len=1.70, thru=18.1, I=16.7, T=37.8, wait=4.6s, pool=65, epoch=2/2][2026-10-19 20:10:34,510] loop.py:419 INFO Saving checkpoint at step 14250... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 20:10:37,932] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0014251 (TrainerActor pid=9310) [2026-10-19 20:10:37,985] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0013501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:38,663] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:39,175] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=33/64 [2026-10-19 20:10:47,202] loop.py:45 INFO TRAIN_STEP step=14251 loss=1.168742 Training: 79%|███████▉ | 14251/18018 [30:40:33<9:43:12, 9.29s/step, loss=1.169, acc=0.722, acc_len=1.76, thru=13.5, I=12.9, T=34.8, wait=4.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:10:47,197] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14250: forward=2595.3ms backward=1033.5ms (TrainerActor pid=9310) [2026-10-19 20:10:37,931] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0014251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 20:10:39,081] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0013501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:53,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:10:53,582] loop.py:45 INFO TRAIN_STEP step=14252 loss=1.176478 Training: 79%|███████▉ | 14252/18018 [30:40:39<8:48:14, 8.42s/step, loss=1.176, acc=0.726, acc_len=1.78, thru=20.7, I=18.5, T=37.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:53,917] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:10:59,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:11:00,710] loop.py:45 INFO TRAIN_STEP step=14253 loss=1.261049 Training: 79%|███████▉ | 14253/18018 [30:40:46<8:23:54, 8.03s/step, loss=1.261, acc=0.707, acc_len=1.70, thru=17.6, I=20.1, T=38.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:00,859] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:05,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:11:06,366] loop.py:45 INFO TRAIN_STEP step=14254 loss=1.197182 Training: 79%|███████▉ | 14254/18018 [30:40:52<7:39:02, 7.32s/step, loss=1.197, acc=0.717, acc_len=1.73, thru=22.1, I=26.6, T=39.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:06,738] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:10,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:11:12,208] loop.py:380 INFO TIMING step=14255: step=3.387s data=1.354s compute=3.361s [fwd=2.441s bwd=0.905s opt=0.003s] dispatch=2.446s [2026-10-19 20:11:12,208] loop.py:45 INFO TRAIN_STEP step=14255 loss=1.253501 Training: 79%|███████▉ | 14255/18018 [30:40:58<7:11:09, 6.87s/step, loss=1.254, acc=0.712, acc_len=1.72, thru=22.0, I=29.1, T=37.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:12,447] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:18,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:11:18,402] loop.py:45 INFO TRAIN_STEP step=14256 loss=1.279358 Training: 79%|███████▉ | 14256/18018 [30:41:04<6:58:15, 6.67s/step, loss=1.279, acc=0.707, acc_len=1.70, thru=20.7, I=24.9, T=39.0, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:11:18,397] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14255: forward=2314.6ms backward=911.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:18,607] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:24,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:11:25,421] loop.py:45 INFO TRAIN_STEP step=14257 loss=1.308331 Training: 79%|███████▉ | 14257/18018 [30:41:11<7:04:42, 6.78s/step, loss=1.308, acc=0.696, acc_len=1.67, thru=18.5, I=21.4, T=38.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:25,689] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:11:31,538] loop.py:45 INFO TRAIN_STEP step=14258 loss=1.288514 Training: 79%|███████▉ | 14258/18018 [30:41:17<6:52:13, 6.58s/step, loss=1.289, acc=0.701, acc_len=1.68, thru=18.9, I=22.7, T=39.1, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:39,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:11:40,728] loop.py:45 INFO TRAIN_STEP step=14259 loss=1.200382 Training: 79%|███████▉ | 14259/18018 [30:41:26<7:41:14, 7.36s/step, loss=1.200, acc=0.718, acc_len=1.74, thru=15.1, I=15.6, T=33.4, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:40,955] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:46,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:11:47,026] loop.py:380 INFO TIMING step=14260: step=3.387s data=1.378s compute=3.351s [fwd=2.427s bwd=0.910s opt=0.003s] dispatch=2.899s [2026-10-19 20:11:47,027] loop.py:45 INFO TRAIN_STEP step=14260 loss=1.342070 Training: 79%|███████▉ | 14260/18018 [30:41:33<7:21:05, 7.04s/step, loss=1.342, acc=0.693, acc_len=1.65, thru=19.4, I=22.9, T=37.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:47,212] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:11:52,267] loop.py:45 INFO TRAIN_STEP step=14261 loss=1.181659 Training: 79%|███████▉ | 14261/18018 [30:41:38<6:47:07, 6.50s/step, loss=1.182, acc=0.721, acc_len=1.75, thru=21.6, I=26.0, T=38.6, wait=1.9s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:11:52,262] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14260: forward=2370.2ms backward=898.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:11:59,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:12:00,514] loop.py:45 INFO TRAIN_STEP step=14262 loss=1.226509 Training: 79%|███████▉ | 14262/18018 [30:41:46<7:19:47, 7.03s/step, loss=1.227, acc=0.714, acc_len=1.73, thru=15.9, I=17.3, T=35.4, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:00,796] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:06,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:12:06,998] loop.py:45 INFO TRAIN_STEP step=14263 loss=1.269598 Training: 79%|███████▉ | 14263/18018 [30:41:53<7:09:29, 6.86s/step, loss=1.270, acc=0.701, acc_len=1.69, thru=19.0, I=23.4, T=38.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:07,414] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:12,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:12:13,293] loop.py:45 INFO TRAIN_STEP step=14264 loss=1.278516 Training: 79%|███████▉ | 14264/18018 [30:41:59<6:58:44, 6.69s/step, loss=1.279, acc=0.702, acc_len=1.68, thru=20.2, I=23.2, T=37.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:13,730] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:18,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:12:18,911] loop.py:380 INFO TIMING step=14265: step=3.270s data=1.352s compute=3.241s [fwd=2.332s bwd=0.894s opt=0.003s] dispatch=2.338s [2026-10-19 20:12:18,911] loop.py:45 INFO TRAIN_STEP step=14265 loss=1.257893 Training: 79%|███████▉ | 14265/18018 [30:42:04<6:38:24, 6.37s/step, loss=1.258, acc=0.707, acc_len=1.70, thru=22.5, I=26.0, T=39.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:19,338] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:12:24,630] loop.py:45 INFO TRAIN_STEP step=14266 loss=1.320796 Training: 79%|███████▉ | 14266/18018 [30:42:10<6:26:12, 6.18s/step, loss=1.321, acc=0.693, acc_len=1.65, thru=22.1, I=22.3, T=38.0, wait=2.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:12:24,625] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14265: forward=2397.3ms backward=918.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:31,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:12:32,127] loop.py:45 INFO TRAIN_STEP step=14267 loss=1.291485 Training: 79%|███████▉ | 14267/18018 [30:42:18<6:50:47, 6.57s/step, loss=1.291, acc=0.699, acc_len=1.68, thru=17.2, I=18.1, T=36.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:32,172] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 20:12:37,383] loop.py:45 INFO TRAIN_STEP step=14268 loss=1.196923 Training: 79%|███████▉ | 14268/18018 [30:42:23<6:26:02, 6.18s/step, loss=1.197, acc=0.716, acc_len=1.74, thru=20.5, I=21.9, T=39.0, wait=2.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:42,806] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:12:43,891] loop.py:45 INFO TRAIN_STEP step=14269 loss=1.243729 Training: 79%|███████▉ | 14269/18018 [30:42:29<6:32:06, 6.28s/step, loss=1.244, acc=0.711, acc_len=1.72, thru=20.4, I=22.6, T=38.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:44,316] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:48,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:12:49,810] loop.py:380 INFO TIMING step=14270: step=3.446s data=1.403s compute=3.420s [fwd=2.439s bwd=0.967s opt=0.003s] dispatch=2.466s [2026-10-19 20:12:49,810] loop.py:45 INFO TRAIN_STEP step=14270 loss=1.206947 Training: 79%|███████▉ | 14270/18018 [30:42:35<6:25:19, 6.17s/step, loss=1.207, acc=0.714, acc_len=1.72, thru=20.5, I=25.7, T=37.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:12:50,140] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:12:56,503] loop.py:45 INFO TRAIN_STEP step=14271 loss=1.225538 Training: 79%|███████▉ | 14271/18018 [30:42:42<6:35:05, 6.33s/step, loss=1.226, acc=0.712, acc_len=1.73, thru=15.8, I=23.3, T=38.0, wait=3.3s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:12:56,498] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14270: forward=2404.9ms backward=919.1ms [2026-10-19 20:13:06,146] loop.py:45 INFO TRAIN_STEP step=14272 loss=1.315627 Training: 79%|███████▉ | 14272/18018 [30:42:52<7:37:03, 7.32s/step, loss=1.316, acc=0.698, acc_len=1.67, thru=16.2, I=12.8, T=34.2, wait=5.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:13,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:13:14,009] loop.py:45 INFO TRAIN_STEP step=14273 loss=1.227486 Training: 79%|███████▉ | 14273/18018 [30:43:00<7:47:06, 7.48s/step, loss=1.227, acc=0.712, acc_len=1.72, thru=17.8, I=18.7, T=34.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:14,322] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:13:20,258] loop.py:45 INFO TRAIN_STEP step=14274 loss=1.261528 Training: 79%|███████▉ | 14274/18018 [30:43:06<7:23:54, 7.11s/step, loss=1.262, acc=0.707, acc_len=1.71, thru=18.4, I=19.2, T=37.8, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:27,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:13:28,079] loop.py:380 INFO TIMING step=14275: step=3.576s data=1.492s compute=3.540s [fwd=2.519s bwd=1.006s opt=0.003s] dispatch=4.235s [2026-10-19 20:13:28,079] loop.py:45 INFO TRAIN_STEP step=14275 loss=1.305904 Training: 79%|███████▉ | 14275/18018 [30:43:14<7:36:58, 7.33s/step, loss=1.306, acc=0.698, acc_len=1.67, thru=18.1, I=18.8, T=35.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:28,383] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:37,543] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:13:38,552] loop.py:45 INFO TRAIN_STEP step=14276 loss=1.205928 Training: 79%|███████▉ | 14276/18018 [30:43:24<8:35:44, 8.27s/step, loss=1.206, acc=0.716, acc_len=1.74, thru=12.0, I=13.6, T=31.5, wait=6.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:13:38,548] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14275: forward=2578.2ms backward=1077.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:39,050] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:43,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:13:44,469] loop.py:45 INFO TRAIN_STEP step=14277 loss=1.241230 Training: 79%|███████▉ | 14277/18018 [30:43:30<7:51:39, 7.56s/step, loss=1.241, acc=0.709, acc_len=1.70, thru=19.2, I=23.4, T=39.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:44,660] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:49,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:13:50,017] loop.py:45 INFO TRAIN_STEP step=14278 loss=1.196152 Training: 79%|███████▉ | 14278/18018 [30:43:36<7:13:46, 6.96s/step, loss=1.196, acc=0.717, acc_len=1.74, thru=22.6, I=27.1, T=37.6, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:50,205] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:55,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:13:56,037] loop.py:45 INFO TRAIN_STEP step=14279 loss=1.290372 Training: 79%|███████▉ | 14279/18018 [30:43:42<6:56:05, 6.68s/step, loss=1.290, acc=0.702, acc_len=1.69, thru=21.2, I=24.3, T=38.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:13:56,538] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:03,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:14:03,610] loop.py:380 INFO TIMING step=14280: step=3.366s data=1.371s compute=3.332s [fwd=2.393s bwd=0.924s opt=0.003s] dispatch=4.200s [2026-10-19 20:14:03,610] loop.py:45 INFO TRAIN_STEP step=14280 loss=1.169669 Training: 79%|███████▉ | 14280/18018 [30:43:49<7:12:44, 6.95s/step, loss=1.170, acc=0.720, acc_len=1.75, thru=17.2, I=19.8, T=38.0, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:03,654] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:09,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:14:09,579] loop.py:45 INFO TRAIN_STEP step=14281 loss=1.337332 Training: 79%|███████▉ | 14281/18018 [30:43:55<6:54:23, 6.65s/step, loss=1.337, acc=0.692, acc_len=1.65, thru=20.7, I=20.8, T=38.0, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:14:09,574] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14280: forward=2375.7ms backward=944.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:09,769] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:16,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:14:17,481] loop.py:45 INFO TRAIN_STEP step=14282 loss=1.425308 Training: 79%|███████▉ | 14282/18018 [30:44:03<7:17:35, 7.03s/step, loss=1.425, acc=0.680, acc_len=1.61, thru=16.5, I=19.8, T=36.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:17,642] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:14:24,023] loop.py:45 INFO TRAIN_STEP step=14283 loss=1.408354 Training: 79%|███████▉ | 14283/18018 [30:44:10<7:08:25, 6.88s/step, loss=1.408, acc=0.680, acc_len=1.61, thru=16.4, I=21.2, T=37.8, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:32,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:14:33,225] loop.py:45 INFO TRAIN_STEP step=14284 loss=1.363461 Training: 79%|███████▉ | 14284/18018 [30:44:19<7:51:37, 7.58s/step, loss=1.363, acc=0.685, acc_len=1.64, thru=16.1, I=16.9, T=36.4, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:33,473] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:38,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:14:39,081] loop.py:380 INFO TIMING step=14285: step=3.398s data=1.368s compute=3.371s [fwd=2.411s bwd=0.946s opt=0.003s] dispatch=2.448s [2026-10-19 20:14:39,081] loop.py:45 INFO TRAIN_STEP step=14285 loss=1.230892 Training: 79%|███████▉ | 14285/18018 [30:44:25<7:19:20, 7.06s/step, loss=1.231, acc=0.711, acc_len=1.72, thru=20.8, I=21.6, T=37.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:39,244] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:43,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:14:44,619] loop.py:45 INFO TRAIN_STEP step=14286 loss=1.196219 Training: 79%|███████▉ | 14286/18018 [30:44:30<6:50:48, 6.60s/step, loss=1.196, acc=0.719, acc_len=1.74, thru=22.9, I=25.7, T=39.0, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:14:44,615] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14285: forward=2348.7ms backward=893.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:44,811] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:52,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:14:53,628] loop.py:45 INFO TRAIN_STEP step=14287 loss=1.275964 Training: 79%|███████▉ | 14287/18018 [30:44:39<7:35:32, 7.33s/step, loss=1.276, acc=0.706, acc_len=1.71, thru=14.5, I=16.6, T=35.1, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:14:54,017] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:15:00,995] loop.py:45 INFO TRAIN_STEP step=14288 loss=1.262049 Training: 79%|███████▉ | 14288/18018 [30:44:47<7:36:08, 7.34s/step, loss=1.262, acc=0.706, acc_len=1.70, thru=16.2, I=19.2, T=35.0, wait=3.7s, pool=61, epoch=2/2][2026-10-19 20:15:07,186] loop.py:45 INFO TRAIN_STEP step=14289 loss=1.347757 Training: 79%|███████▉ | 14289/18018 [30:44:53<7:14:39, 6.99s/step, loss=1.348, acc=0.695, acc_len=1.66, thru=17.4, I=21.4, T=36.1, wait=2.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:15:14,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:15:15,414] loop.py:380 INFO TIMING step=14290: step=3.753s data=1.591s compute=3.713s [fwd=2.587s bwd=1.111s opt=0.003s] dispatch=4.468s [2026-10-19 20:15:15,414] loop.py:45 INFO TRAIN_STEP step=14290 loss=1.241965 Training: 79%|███████▉ | 14290/18018 [30:45:01<7:37:33, 7.36s/step, loss=1.242, acc=0.710, acc_len=1.72, thru=16.9, I=16.9, T=34.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:15:15,631] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:15:21,125] loop.py:45 INFO TRAIN_STEP step=14291 loss=1.319967 Training: 79%|███████▉ | 14291/18018 [30:45:07<7:06:36, 6.87s/step, loss=1.320, acc=0.693, acc_len=1.65, thru=17.8, I=21.9, T=38.7, wait=2.4s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:15:21,120] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14290: forward=2362.5ms backward=901.8ms [2026-10-19 20:15:30,665] loop.py:45 INFO TRAIN_STEP step=14292 loss=1.329074 Training: 79%|███████▉ | 14292/18018 [30:45:16<7:56:16, 7.67s/step, loss=1.329, acc=0.697, acc_len=1.67, thru=13.7, I=13.0, T=36.5, wait=6.0s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:15:39,460] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:15:40,984] loop.py:45 INFO TRAIN_STEP step=14293 loss=1.235679 Training: 79%|███████▉ | 14293/18018 [30:45:27<8:45:30, 8.46s/step, loss=1.236, acc=0.710, acc_len=1.72, thru=16.2, I=18.0, T=34.9, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:15:41,472] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 20:15:47,874] loop.py:45 INFO TRAIN_STEP step=14294 loss=1.241062 Training: 79%|███████▉ | 14294/18018 [30:45:33<8:16:03, 7.99s/step, loss=1.241, acc=0.712, acc_len=1.72, thru=17.3, I=21.5, T=38.0, wait=3.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:15:53,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:15:54,084] loop.py:380 INFO TIMING step=14295: step=3.742s data=1.630s compute=3.707s [fwd=2.618s bwd=1.075s opt=0.003s] dispatch=2.459s [2026-10-19 20:15:54,084] loop.py:45 INFO TRAIN_STEP step=14295 loss=1.313371 Training: 79%|███████▉ | 14295/18018 [30:45:40<7:42:48, 7.46s/step, loss=1.313, acc=0.699, acc_len=1.68, thru=20.8, I=22.1, T=34.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:15:54,362] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:15:59,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:16:01,128] loop.py:45 INFO TRAIN_STEP step=14296 loss=1.241826 Training: 79%|███████▉ | 14296/18018 [30:45:47<7:34:56, 7.33s/step, loss=1.242, acc=0.710, acc_len=1.72, thru=17.2, I=23.0, T=35.4, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:16:01,124] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14295: forward=2582.5ms backward=985.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:01,563] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:06,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:16:07,000] loop.py:45 INFO TRAIN_STEP step=14297 loss=1.193685 Training: 79%|███████▉ | 14297/18018 [30:45:53<7:07:36, 6.90s/step, loss=1.194, acc=0.722, acc_len=1.76, thru=20.6, I=27.7, T=38.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:07,298] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:16:13,388] loop.py:45 INFO TRAIN_STEP step=14298 loss=1.379908 Training: 79%|███████▉ | 14298/18018 [30:45:59<6:58:02, 6.74s/step, loss=1.380, acc=0.688, acc_len=1.63, thru=20.3, I=22.9, T=37.7, wait=3.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:20,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:16:20,693] loop.py:45 INFO TRAIN_STEP step=14299 loss=1.152661 Training: 79%|███████▉ | 14299/18018 [30:46:06<7:08:22, 6.91s/step, loss=1.153, acc=0.724, acc_len=1.77, thru=19.1, I=20.0, T=36.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:20,993] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:16:27,343] loop.py:380 INFO TIMING step=14300: step=3.349s data=1.324s compute=3.315s [fwd=2.369s bwd=0.932s opt=0.003s] dispatch=3.292s [2026-10-19 20:16:27,343] loop.py:45 INFO TRAIN_STEP step=14300 loss=1.209316 Training: 79%|███████▉ | 14300/18018 [30:46:13<7:03:25, 6.83s/step, loss=1.209, acc=0.717, acc_len=1.74, thru=18.8, I=19.3, T=38.2, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:33,535] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:16:33,936] loop.py:45 INFO TRAIN_STEP step=14301 loss=1.253291 Training: 79%|███████▉ | 14301/18018 [30:46:20<6:58:49, 6.76s/step, loss=1.253, acc=0.705, acc_len=1.70, thru=19.7, I=20.3, T=39.1, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:16:33,931] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14300: forward=2312.7ms backward=909.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:34,045] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:38,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-19 20:16:40,171] loop.py:45 INFO TRAIN_STEP step=14302 loss=1.079672 Training: 79%|███████▉ | 14302/18018 [30:46:26<6:48:57, 6.60s/step, loss=1.080, acc=0.741, acc_len=1.82, thru=20.9, I=23.3, T=36.6, wait=2.7s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:40,331] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:45,535] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:16:45,627] loop.py:45 INFO TRAIN_STEP step=14303 loss=1.168808 Training: 79%|███████▉ | 14303/18018 [30:46:31<6:27:31, 6.26s/step, loss=1.169, acc=0.723, acc_len=1.76, thru=22.7, I=25.8, T=38.4, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:46,038] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:53,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:16:54,518] loop.py:45 INFO TRAIN_STEP step=14304 loss=1.265582 Training: 79%|███████▉ | 14304/18018 [30:46:40<7:16:20, 7.05s/step, loss=1.266, acc=0.704, acc_len=1.69, thru=15.8, I=17.8, T=35.5, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:16:54,778] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:17:02,025] loop.py:380 INFO TIMING step=14305: step=3.511s data=1.423s compute=3.480s [fwd=2.491s bwd=0.974s opt=0.003s] dispatch=3.985s [2026-10-19 20:17:02,025] loop.py:45 INFO TRAIN_STEP step=14305 loss=1.295530 Training: 79%|███████▉ | 14305/18018 [30:46:48<7:24:39, 7.19s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=13.6, I=17.2, T=36.5, wait=4.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:10,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:17:10,721] loop.py:45 INFO TRAIN_STEP step=14306 loss=1.333463 Training: 79%|███████▉ | 14306/18018 [30:46:56<7:52:37, 7.64s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=17.6, I=16.4, T=34.8, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:17:10,717] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14305: forward=2591.2ms backward=1029.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:10,796] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:17,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:17:18,045] loop.py:45 INFO TRAIN_STEP step=14307 loss=1.126395 Training: 79%|███████▉ | 14307/18018 [30:47:04<7:46:36, 7.54s/step, loss=1.126, acc=0.730, acc_len=1.79, thru=16.8, I=17.1, T=34.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:18,446] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:25,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:17:26,167] loop.py:45 INFO TRAIN_STEP step=14308 loss=1.379221 Training: 79%|███████▉ | 14308/18018 [30:47:12<7:57:14, 7.72s/step, loss=1.379, acc=0.685, acc_len=1.63, thru=16.3, I=17.7, T=33.8, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:26,386] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:31,532] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:17:32,154] loop.py:45 INFO TRAIN_STEP step=14309 loss=1.369776 Training: 79%|███████▉ | 14309/18018 [30:47:18<7:24:58, 7.20s/step, loss=1.370, acc=0.685, acc_len=1.63, thru=20.3, I=21.3, T=38.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:32,536] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:38,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:17:39,414] loop.py:380 INFO TIMING step=14310: step=3.290s data=1.302s compute=3.251s [fwd=2.302s bwd=0.935s opt=0.003s] dispatch=3.963s [2026-10-19 20:17:39,415] loop.py:45 INFO TRAIN_STEP step=14310 loss=1.260235 Training: 79%|███████▉ | 14310/18018 [30:47:25<7:26:02, 7.22s/step, loss=1.260, acc=0.710, acc_len=1.71, thru=18.1, I=21.0, T=38.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:39,736] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-19 20:17:45,898] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14310: forward=2588.7ms backward=1056.2ms [2026-10-19 20:17:45,903] loop.py:45 INFO TRAIN_STEP step=14311 loss=1.299711 Training: 79%|███████▉ | 14311/18018 [30:47:31<7:12:22, 7.00s/step, loss=1.300, acc=0.698, acc_len=1.67, thru=19.2, I=19.8, T=34.7, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:52,814] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:17:54,674] loop.py:45 INFO TRAIN_STEP step=14312 loss=1.191909 Training: 79%|███████▉ | 14312/18018 [30:47:40<7:45:08, 7.53s/step, loss=1.192, acc=0.716, acc_len=1.74, thru=15.3, I=16.9, T=30.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:17:54,835] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:00,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:18:00,375] loop.py:45 INFO TRAIN_STEP step=14313 loss=1.333022 Training: 79%|███████▉ | 14313/18018 [30:47:46<7:11:06, 6.98s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=19.6, I=22.4, T=37.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:00,614] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:18:08,752] loop.py:45 INFO TRAIN_STEP step=14314 loss=1.196182 Training: 79%|███████▉ | 14314/18018 [30:47:54<7:36:49, 7.40s/step, loss=1.196, acc=0.719, acc_len=1.74, thru=14.1, I=14.2, T=36.9, wait=4.9s, pool=45, epoch=2/2][2026-10-19 20:18:16,934] loop.py:380 INFO TIMING step=14315: step=3.792s data=1.534s compute=3.747s [fwd=2.686s bwd=1.047s opt=0.003s] dispatch=4.382s [2026-10-19 20:18:16,934] loop.py:45 INFO TRAIN_STEP step=14315 loss=1.239484 Training: 79%|███████▉ | 14315/18018 [30:48:03<7:51:16, 7.64s/step, loss=1.239, acc=0.711, acc_len=1.72, thru=16.9, I=17.8, T=33.8, wait=4.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:23,883] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:18:24,971] loop.py:45 INFO TRAIN_STEP step=14316 loss=1.273075 Training: 79%|███████▉ | 14316/18018 [30:48:11<7:58:28, 7.75s/step, loss=1.273, acc=0.705, acc_len=1.70, thru=16.5, I=17.3, T=35.1, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:18:24,967] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14315: forward=2610.6ms backward=996.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:25,393] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:31,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:18:32,355] loop.py:45 INFO TRAIN_STEP step=14317 loss=1.193945 Training: 79%|███████▉ | 14317/18018 [30:48:18<7:51:28, 7.64s/step, loss=1.194, acc=0.713, acc_len=1.73, thru=17.2, I=22.3, T=36.7, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:32,777] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:18:37,895] loop.py:45 INFO TRAIN_STEP step=14318 loss=1.290807 Training: 79%|███████▉ | 14318/18018 [30:48:23<7:12:25, 7.01s/step, loss=1.291, acc=0.697, acc_len=1.67, thru=20.5, I=23.2, T=38.9, wait=2.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:49,614] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:18:49,871] loop.py:45 INFO TRAIN_STEP step=14319 loss=1.218947 Training: 79%|███████▉ | 14319/18018 [30:48:35<8:44:06, 8.50s/step, loss=1.219, acc=0.719, acc_len=1.75, thru=12.5, I=12.0, T=34.3, wait=8.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:50,117] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:56,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:18:57,845] loop.py:380 INFO TIMING step=14320: step=3.690s data=1.643s compute=3.620s [fwd=2.565s bwd=1.041s opt=0.003s] dispatch=4.275s [2026-10-19 20:18:57,845] loop.py:45 INFO TRAIN_STEP step=14320 loss=1.197481 Training: 79%|███████▉ | 14320/18018 [30:48:43<8:34:14, 8.34s/step, loss=1.197, acc=0.716, acc_len=1.73, thru=17.0, I=18.2, T=34.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:18:58,075] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:19:05,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:19:05,371] loop.py:45 INFO TRAIN_STEP step=14321 loss=1.415100 Training: 79%|███████▉ | 14321/18018 [30:48:51<8:18:58, 8.10s/step, loss=1.415, acc=0.681, acc_len=1.62, thru=16.2, I=20.1, T=37.5, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:19:05,366] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14320: forward=2449.2ms backward=924.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:19:05,588] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:19:11,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:19:11,988] loop.py:45 INFO TRAIN_STEP step=14322 loss=1.217686 Training: 79%|███████▉ | 14322/18018 [30:48:58<7:51:29, 7.65s/step, loss=1.218, acc=0.716, acc_len=1.73, thru=19.6, I=19.0, T=37.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:19:12,146] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:19:17,745] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:19:18,646] loop.py:45 INFO TRAIN_STEP step=14323 loss=1.155078 Training: 79%|███████▉ | 14323/18018 [30:49:04<7:32:58, 7.36s/step, loss=1.155, acc=0.726, acc_len=1.77, thru=19.6, I=20.2, T=36.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:19:18,752] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 20:19:24,805] loop.py:45 INFO TRAIN_STEP step=14324 loss=1.263886 Training: 79%|███████▉ | 14324/18018 [30:49:10<7:10:42, 7.00s/step, loss=1.264, acc=0.703, acc_len=1.70, thru=19.5, I=21.3, T=38.6, wait=2.8s, pool=57, epoch=2/2][2026-10-19 20:19:32,340] loop.py:380 INFO TIMING step=14325: step=3.845s data=1.800s compute=3.811s [fwd=2.808s bwd=0.990s opt=0.003s] dispatch=3.681s [2026-10-19 20:19:32,340] loop.py:45 INFO TRAIN_STEP step=14325 loss=1.180352 Training: 80%|███████▉ | 14325/18018 [30:49:18<7:20:34, 7.16s/step, loss=1.180, acc=0.721, acc_len=1.76, thru=17.4, I=18.7, T=33.3, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:19:39,498] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:19:39,598] loop.py:45 INFO TRAIN_STEP step=14326 loss=1.261619 Training: 80%|███████▉ | 14326/18018 [30:49:25<7:22:15, 7.19s/step, loss=1.262, acc=0.709, acc_len=1.72, thru=17.4, I=17.3, T=36.3, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:19:39,594] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14325: forward=2514.6ms backward=970.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:19:40,001] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:19:45,951] loop.py:45 INFO TRAIN_STEP step=14327 loss=1.167103 Training: 80%|███████▉ | 14327/18018 [30:49:32<7:06:47, 6.94s/step, loss=1.167, acc=0.720, acc_len=1.75, thru=17.3, I=19.9, T=36.6, wait=2.8s, pool=41, epoch=2/2][2026-10-19 20:19:55,551] loop.py:45 INFO TRAIN_STEP step=14328 loss=1.180144 Training: 80%|███████▉ | 14328/18018 [30:49:41<7:55:45, 7.74s/step, loss=1.180, acc=0.723, acc_len=1.76, thru=15.2, I=15.4, T=34.4, wait=5.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:03,583] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:20:03,942] loop.py:45 INFO TRAIN_STEP step=14329 loss=1.229609 Training: 80%|███████▉ | 14329/18018 [30:49:50<8:07:43, 7.93s/step, loss=1.230, acc=0.713, acc_len=1.73, thru=15.4, I=15.6, T=36.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:04,092] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:20:10,060] loop.py:380 INFO TIMING step=14330: step=3.389s data=1.405s compute=3.360s [fwd=2.448s bwd=0.897s opt=0.003s] dispatch=2.720s [2026-10-19 20:20:10,060] loop.py:45 INFO TRAIN_STEP step=14330 loss=1.348407 Training: 80%|███████▉ | 14330/18018 [30:49:56<7:34:07, 7.39s/step, loss=1.348, acc=0.688, acc_len=1.64, thru=16.9, I=18.5, T=37.8, wait=2.7s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:19,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:20:20,717] loop.py:45 INFO TRAIN_STEP step=14331 loss=1.203753 Training: 80%|███████▉ | 14331/18018 [30:50:06<8:34:13, 8.37s/step, loss=1.204, acc=0.715, acc_len=1.73, thru=15.9, I=15.3, T=29.0, wait=6.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:20:20,713] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14330: forward=3315.5ms backward=1043.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:20,844] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:25,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:20:26,110] loop.py:45 INFO TRAIN_STEP step=14332 loss=1.181783 Training: 80%|███████▉ | 14332/18018 [30:50:12<7:39:16, 7.48s/step, loss=1.182, acc=0.719, acc_len=1.75, thru=20.2, I=24.8, T=37.8, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:26,361] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:20:35,158] loop.py:45 INFO TRAIN_STEP step=14333 loss=1.214632 Training: 80%|███████▉ | 14333/18018 [30:50:21<8:08:07, 7.95s/step, loss=1.215, acc=0.714, acc_len=1.73, thru=13.8, I=13.7, T=36.9, wait=5.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:44,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:20:45,127] loop.py:45 INFO TRAIN_STEP step=14334 loss=1.259052 Training: 80%|███████▉ | 14334/18018 [30:50:31<8:45:13, 8.55s/step, loss=1.259, acc=0.707, acc_len=1.70, thru=13.4, I=15.2, T=32.8, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:45,231] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:51,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:20:51,283] loop.py:380 INFO TIMING step=14335: step=3.701s data=1.736s compute=3.661s [fwd=2.704s bwd=0.943s opt=0.003s] dispatch=2.446s [2026-10-19 20:20:51,284] loop.py:45 INFO TRAIN_STEP step=14335 loss=1.236321 Training: 80%|███████▉ | 14335/18018 [30:50:37<8:00:56, 7.84s/step, loss=1.236, acc=0.713, acc_len=1.72, thru=19.7, I=22.0, T=34.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:20:51,616] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:00,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:21:00,170] loop.py:45 INFO TRAIN_STEP step=14336 loss=1.281168 Training: 80%|███████▉ | 14336/18018 [30:50:46<8:20:09, 8.15s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=15.0, I=15.0, T=33.7, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:21:00,166] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14335: forward=2695.9ms backward=1063.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:00,598] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:07,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:21:07,725] loop.py:45 INFO TRAIN_STEP step=14337 loss=1.129442 Training: 80%|███████▉ | 14337/18018 [30:50:53<8:09:03, 7.97s/step, loss=1.129, acc=0.732, acc_len=1.79, thru=16.9, I=17.6, T=36.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:07,754] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:13,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:21:13,627] loop.py:45 INFO TRAIN_STEP step=14338 loss=1.298072 Training: 80%|███████▉ | 14338/18018 [30:50:59<7:30:49, 7.35s/step, loss=1.298, acc=0.702, acc_len=1.68, thru=20.8, I=22.0, T=37.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:14,114] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:22,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:21:22,976] loop.py:45 INFO TRAIN_STEP step=14339 loss=1.281353 Training: 80%|███████▉ | 14339/18018 [30:51:09<8:07:29, 7.95s/step, loss=1.281, acc=0.707, acc_len=1.70, thru=14.3, I=15.7, T=34.2, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:23,394] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:21:29,413] loop.py:380 INFO TIMING step=14340: step=3.480s data=1.445s compute=3.448s [fwd=2.478s bwd=0.955s opt=0.003s] dispatch=2.949s [2026-10-19 20:21:29,414] loop.py:45 INFO TRAIN_STEP step=14340 loss=1.176934 Training: 80%|███████▉ | 14340/18018 [30:51:15<7:39:31, 7.50s/step, loss=1.177, acc=0.721, acc_len=1.75, thru=16.7, I=19.3, T=36.8, wait=2.9s, pool=41, epoch=2/2][2026-10-19 20:21:37,205] loop.py:45 INFO TRAIN_STEP step=14341 loss=1.301014 Training: 80%|███████▉ | 14341/18018 [30:51:23<7:44:48, 7.58s/step, loss=1.301, acc=0.703, acc_len=1.69, thru=15.9, I=14.0, T=35.5, wait=4.2s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:21:37,201] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14340: forward=2537.5ms backward=1016.8ms [2026-10-19 20:21:45,996] loop.py:45 INFO TRAIN_STEP step=14342 loss=1.317608 Training: 80%|███████▉ | 14342/18018 [30:51:32<8:06:51, 7.95s/step, loss=1.318, acc=0.700, acc_len=1.68, thru=17.1, I=14.0, T=36.2, wait=5.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:51,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:21:53,517] loop.py:45 INFO TRAIN_STEP step=14343 loss=1.285589 Training: 80%|███████▉ | 14343/18018 [30:51:39<7:58:57, 7.82s/step, loss=1.286, acc=0.703, acc_len=1.69, thru=18.9, I=19.7, T=30.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:21:53,896] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 20:21:59,952] loop.py:45 INFO TRAIN_STEP step=14344 loss=1.323427 Training: 80%|███████▉ | 14344/18018 [30:51:46<7:33:21, 7.40s/step, loss=1.323, acc=0.694, acc_len=1.66, thru=17.3, I=20.7, T=35.1, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:05,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:22:06,830] loop.py:380 INFO TIMING step=14345: step=3.663s data=1.572s compute=3.622s [fwd=2.537s bwd=1.070s opt=0.003s] dispatch=3.206s [2026-10-19 20:22:06,831] loop.py:45 INFO TRAIN_STEP step=14345 loss=1.227587 Training: 80%|███████▉ | 14345/18018 [30:51:52<7:23:37, 7.25s/step, loss=1.228, acc=0.715, acc_len=1.74, thru=18.6, I=21.6, T=34.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:07,202] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:13,581] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:22:13,797] loop.py:45 INFO TRAIN_STEP step=14346 loss=1.219849 Training: 80%|███████▉ | 14346/18018 [30:51:59<7:18:21, 7.16s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=16.9, I=21.9, T=35.9, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:22:13,793] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14345: forward=2535.4ms backward=987.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:14,083] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:22:21,158] loop.py:45 INFO TRAIN_STEP step=14347 loss=1.152291 Training: 80%|███████▉ | 14347/18018 [30:52:07<7:21:52, 7.22s/step, loss=1.152, acc=0.725, acc_len=1.77, thru=18.4, I=17.6, T=37.5, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:27,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:22:28,726] loop.py:45 INFO TRAIN_STEP step=14348 loss=1.321253 Training: 80%|███████▉ | 14348/18018 [30:52:14<7:28:06, 7.33s/step, loss=1.321, acc=0.695, acc_len=1.67, thru=18.0, I=19.4, T=36.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:28,892] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:33,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:22:34,330] loop.py:45 INFO TRAIN_STEP step=14349 loss=1.277603 Training: 80%|███████▉ | 14349/18018 [30:52:20<6:56:21, 6.81s/step, loss=1.278, acc=0.705, acc_len=1.69, thru=21.2, I=26.6, T=38.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:34,622] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:39,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:22:41,373] loop.py:380 INFO TIMING step=14350: step=3.966s data=1.977s compute=3.920s [fwd=2.991s bwd=0.914s opt=0.003s] dispatch=3.069s [2026-10-19 20:22:41,373] loop.py:45 INFO TRAIN_STEP step=14350 loss=1.274043 Training: 80%|███████▉ | 14350/18018 [30:52:27<7:00:34, 6.88s/step, loss=1.274, acc=0.705, acc_len=1.70, thru=19.2, I=24.1, T=32.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:41,406] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:49,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:22:49,505] loop.py:45 INFO TRAIN_STEP step=14351 loss=1.275726 Training: 80%|███████▉ | 14351/18018 [30:52:35<7:23:23, 7.25s/step, loss=1.276, acc=0.706, acc_len=1.70, thru=14.3, I=17.4, T=34.9, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:22:49,502] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14350: forward=2534.3ms backward=1075.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:22:49,947] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:22:55,667] loop.py:45 INFO TRAIN_STEP step=14352 loss=1.209402 Training: 80%|███████▉ | 14352/18018 [30:52:41<7:03:12, 6.93s/step, loss=1.209, acc=0.715, acc_len=1.73, thru=19.6, I=20.0, T=37.6, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:01,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:23:02,659] loop.py:45 INFO TRAIN_STEP step=14353 loss=1.274848 Training: 80%|███████▉ | 14353/18018 [30:52:48<7:04:21, 6.95s/step, loss=1.275, acc=0.706, acc_len=1.70, thru=18.6, I=21.0, T=37.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:03,001] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:23:09,897] loop.py:45 INFO TRAIN_STEP step=14354 loss=1.281825 Training: 80%|███████▉ | 14354/18018 [30:52:55<7:09:35, 7.03s/step, loss=1.282, acc=0.704, acc_len=1.69, thru=16.8, I=22.1, T=37.5, wait=3.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:15,108] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:23:15,811] loop.py:380 INFO TIMING step=14355: step=3.538s data=1.552s compute=3.498s [fwd=2.450s bwd=1.034s opt=0.003s] dispatch=2.365s [2026-10-19 20:23:15,811] loop.py:45 INFO TRAIN_STEP step=14355 loss=1.284668 Training: 80%|███████▉ | 14355/18018 [30:53:01<6:48:54, 6.70s/step, loss=1.285, acc=0.699, acc_len=1.68, thru=21.0, I=21.7, T=36.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:16,112] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:22,326] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:23:22,719] loop.py:45 INFO TRAIN_STEP step=14356 loss=1.186066 Training: 80%|███████▉ | 14356/18018 [30:53:08<6:52:37, 6.76s/step, loss=1.186, acc=0.719, acc_len=1.75, thru=18.6, I=21.0, T=37.7, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:23:22,715] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14355: forward=2417.0ms backward=931.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:22,833] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:32,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:23:32,722] loop.py:45 INFO TRAIN_STEP step=14357 loss=1.455203 Training: 80%|███████▉ | 14357/18018 [30:53:18<7:51:53, 7.73s/step, loss=1.455, acc=0.673, acc_len=1.59, thru=13.0, I=13.2, T=35.2, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:33,009] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:23:38,585] loop.py:45 INFO TRAIN_STEP step=14358 loss=1.272178 Training: 80%|███████▉ | 14358/18018 [30:53:24<7:17:30, 7.17s/step, loss=1.272, acc=0.700, acc_len=1.68, thru=20.1, I=22.0, T=38.9, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:44,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:23:45,587] loop.py:45 INFO TRAIN_STEP step=14359 loss=1.294491 Training: 80%|███████▉ | 14359/18018 [30:53:31<7:14:15, 7.12s/step, loss=1.294, acc=0.703, acc_len=1.69, thru=18.4, I=21.7, T=37.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:45,653] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 20:23:51,718] loop.py:380 INFO TIMING step=14360: step=3.452s data=1.485s compute=3.421s [fwd=2.473s bwd=0.934s opt=0.003s] dispatch=2.671s [2026-10-19 20:23:51,718] loop.py:45 INFO TRAIN_STEP step=14360 loss=1.272609 Training: 80%|███████▉ | 14360/18018 [30:53:37<6:56:03, 6.82s/step, loss=1.273, acc=0.704, acc_len=1.70, thru=21.0, I=24.3, T=37.1, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:58,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:23:59,310] loop.py:45 INFO TRAIN_STEP step=14361 loss=1.263864 Training: 80%|███████▉ | 14361/18018 [30:53:45<7:10:00, 7.06s/step, loss=1.264, acc=0.707, acc_len=1.71, thru=18.1, I=18.3, T=38.5, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:23:59,306] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14360: forward=2339.6ms backward=926.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:23:59,480] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 20:24:08,624] loop.py:45 INFO TRAIN_STEP step=14362 loss=1.313811 Training: 80%|███████▉ | 14362/18018 [30:53:54<7:51:08, 7.73s/step, loss=1.314, acc=0.694, acc_len=1.66, thru=14.1, I=14.2, T=33.8, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:24:08,619] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:24:09,122] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:24:14,923] loop.py:45 INFO TRAIN_STEP step=14363 loss=1.312424 Training: 80%|███████▉ | 14363/18018 [30:54:01<7:24:48, 7.30s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=18.5, I=16.9, T=36.8, wait=2.8s, pool=49, epoch=2/2][2026-10-19 20:24:22,081] loop.py:45 INFO TRAIN_STEP step=14364 loss=1.285474 Training: 80%|███████▉ | 14364/18018 [30:54:08<7:22:05, 7.26s/step, loss=1.285, acc=0.699, acc_len=1.68, thru=17.4, I=19.2, T=35.9, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:24:29,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:24:30,522] loop.py:380 INFO TIMING step=14365: step=3.696s data=1.562s compute=3.662s [fwd=2.672s bwd=0.975s opt=0.003s] dispatch=4.736s [2026-10-19 20:24:30,523] loop.py:45 INFO TRAIN_STEP step=14365 loss=1.294950 Training: 80%|███████▉ | 14365/18018 [30:54:16<7:43:33, 7.61s/step, loss=1.295, acc=0.704, acc_len=1.69, thru=15.4, I=16.5, T=34.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:24:30,968] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:24:36,726] loop.py:45 INFO TRAIN_STEP step=14366 loss=1.217198 Training: 80%|███████▉ | 14366/18018 [30:54:22<7:17:50, 7.19s/step, loss=1.217, acc=0.715, acc_len=1.73, thru=19.3, I=21.4, T=38.1, wait=2.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:24:36,721] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14365: forward=2442.4ms backward=870.5ms [2026-10-19 20:24:44,447] loop.py:45 INFO TRAIN_STEP step=14367 loss=1.314326 Training: 80%|███████▉ | 14367/18018 [30:54:30<7:27:12, 7.35s/step, loss=1.314, acc=0.697, acc_len=1.67, thru=15.8, I=17.0, T=36.6, wait=4.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:24:52,989] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:24:53,909] loop.py:45 INFO TRAIN_STEP step=14368 loss=1.312495 Training: 80%|███████▉ | 14368/18018 [30:54:39<8:05:38, 7.98s/step, loss=1.312, acc=0.697, acc_len=1.67, thru=14.8, I=14.7, T=34.6, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:24:53,995] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:24:59,955] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:25:00,353] loop.py:45 INFO TRAIN_STEP step=14369 loss=1.248138 Training: 80%|███████▉ | 14369/18018 [30:54:46<7:37:23, 7.52s/step, loss=1.248, acc=0.710, acc_len=1.72, thru=18.8, I=21.4, T=37.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:00,468] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 20:25:07,184] loop.py:380 INFO TIMING step=14370: step=3.388s data=1.395s compute=3.362s [fwd=2.430s bwd=0.918s opt=0.003s] dispatch=3.436s [2026-10-19 20:25:07,185] loop.py:45 INFO TRAIN_STEP step=14370 loss=1.259287 Training: 80%|███████▉ | 14370/18018 [30:54:53<7:24:40, 7.31s/step, loss=1.259, acc=0.707, acc_len=1.71, thru=17.1, I=18.3, T=37.8, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:15,740] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:25:16,402] loop.py:45 INFO TRAIN_STEP step=14371 loss=1.329072 Training: 80%|███████▉ | 14371/18018 [30:55:02<7:59:18, 7.89s/step, loss=1.329, acc=0.693, acc_len=1.65, thru=16.4, I=16.4, T=36.8, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:25:16,398] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14370: forward=2478.7ms backward=940.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:16,747] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:25:22,582] loop.py:45 INFO TRAIN_STEP step=14372 loss=1.203817 Training: 80%|███████▉ | 14372/18018 [30:55:08<7:28:05, 7.37s/step, loss=1.204, acc=0.717, acc_len=1.74, thru=19.0, I=21.3, T=38.5, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:29,274] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:25:29,670] loop.py:45 INFO TRAIN_STEP step=14373 loss=1.282054 Training: 80%|███████▉ | 14373/18018 [30:55:15<7:22:44, 7.29s/step, loss=1.282, acc=0.704, acc_len=1.69, thru=19.2, I=19.2, T=33.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:29,780] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:35,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:25:35,898] loop.py:45 INFO TRAIN_STEP step=14374 loss=1.358422 Training: 80%|███████▉ | 14374/18018 [30:55:21<7:03:20, 6.97s/step, loss=1.358, acc=0.687, acc_len=1.65, thru=19.7, I=21.1, T=37.7, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:36,296] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:43,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:25:43,792] loop.py:380 INFO TIMING step=14375: step=3.492s data=1.394s compute=3.448s [fwd=2.473s bwd=0.961s opt=0.003s] dispatch=4.391s [2026-10-19 20:25:43,792] loop.py:45 INFO TRAIN_STEP step=14375 loss=1.142360 Training: 80%|███████▉ | 14375/18018 [30:55:29<7:20:01, 7.25s/step, loss=1.142, acc=0.727, acc_len=1.78, thru=16.2, I=17.4, T=36.7, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:43,888] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:50,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:25:50,145] loop.py:45 INFO TRAIN_STEP step=14376 loss=1.279528 Training: 80%|███████▉ | 14376/18018 [30:55:36<7:03:34, 6.98s/step, loss=1.280, acc=0.703, acc_len=1.69, thru=19.9, I=20.9, T=37.3, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:25:50,141] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14375: forward=2454.5ms backward=923.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:50,557] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:56,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:25:57,253] loop.py:45 INFO TRAIN_STEP step=14377 loss=1.280767 Training: 80%|███████▉ | 14377/18018 [30:55:43<7:05:51, 7.02s/step, loss=1.281, acc=0.701, acc_len=1.69, thru=17.6, I=19.1, T=38.2, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:25:57,450] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:26:03,958] loop.py:45 INFO TRAIN_STEP step=14378 loss=1.150022 Training: 80%|███████▉ | 14378/18018 [30:55:50<7:00:03, 6.92s/step, loss=1.150, acc=0.728, acc_len=1.78, thru=17.1, I=19.5, T=36.9, wait=3.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:09,444] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:26:10,885] loop.py:45 INFO TRAIN_STEP step=14379 loss=1.242097 Training: 80%|███████▉ | 14379/18018 [30:55:56<6:59:59, 6.92s/step, loss=1.242, acc=0.710, acc_len=1.72, thru=19.6, I=21.3, T=34.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:10,955] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 20:26:16,595] loop.py:380 INFO TIMING step=14380: step=3.248s data=1.299s compute=3.215s [fwd=2.320s bwd=0.880s opt=0.003s] dispatch=2.453s [2026-10-19 20:26:16,595] loop.py:45 INFO TRAIN_STEP step=14380 loss=1.338107 Training: 80%|███████▉ | 14380/18018 [30:56:02<6:37:45, 6.56s/step, loss=1.338, acc=0.691, acc_len=1.65, thru=21.0, I=24.9, T=39.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:16,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:17,034] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:21,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:26:22,950] loop.py:45 INFO TRAIN_STEP step=14381 loss=1.306292 Training: 80%|███████▉ | 14381/18018 [30:56:09<6:33:54, 6.50s/step, loss=1.306, acc=0.697, acc_len=1.67, thru=20.2, I=22.5, T=39.4, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:26:22,946] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14380: forward=2323.5ms backward=884.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:23,357] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:26:31,404] loop.py:45 INFO TRAIN_STEP step=14382 loss=1.283437 Training: 80%|███████▉ | 14382/18018 [30:56:17<7:09:28, 7.09s/step, loss=1.283, acc=0.704, acc_len=1.69, thru=12.4, I=16.1, T=35.6, wait=4.8s, pool=45, epoch=2/2][2026-10-19 20:26:39,965] loop.py:45 INFO TRAIN_STEP step=14383 loss=1.291681 Training: 80%|███████▉ | 14383/18018 [30:56:26<7:36:00, 7.53s/step, loss=1.292, acc=0.699, acc_len=1.68, thru=15.2, I=14.2, T=36.2, wait=5.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:47,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:26:47,523] loop.py:45 INFO TRAIN_STEP step=14384 loss=1.263420 Training: 80%|███████▉ | 14384/18018 [30:56:33<7:36:28, 7.54s/step, loss=1.263, acc=0.707, acc_len=1.70, thru=18.4, I=18.3, T=36.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:47,608] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:52,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:26:53,718] loop.py:380 INFO TIMING step=14385: step=3.393s data=1.398s compute=3.365s [fwd=2.424s bwd=0.927s opt=0.003s] dispatch=2.793s [2026-10-19 20:26:53,718] loop.py:45 INFO TRAIN_STEP step=14385 loss=1.282006 Training: 80%|███████▉ | 14385/18018 [30:56:39<7:11:59, 7.13s/step, loss=1.282, acc=0.703, acc_len=1.69, thru=20.1, I=23.7, T=37.7, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:26:54,224] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:27:02,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:27:04,399] loop.py:45 INFO TRAIN_STEP step=14386 loss=1.303592 Training: 80%|███████▉ | 14386/18018 [30:56:50<8:16:14, 8.20s/step, loss=1.304, acc=0.698, acc_len=1.68, thru=12.6, I=13.6, T=32.3, wait=6.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:27:04,394] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14385: forward=2798.9ms backward=1109.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:27:04,486] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:27:10,434] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:27:11,680] loop.py:45 INFO TRAIN_STEP step=14387 loss=1.324142 Training: 80%|███████▉ | 14387/18018 [30:56:57<7:59:28, 7.92s/step, loss=1.324, acc=0.696, acc_len=1.67, thru=15.7, I=19.5, T=34.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:27:11,941] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:27:17,914] loop.py:45 INFO TRAIN_STEP step=14388 loss=1.402727 Training: 80%|███████▉ | 14388/18018 [30:57:03<7:28:41, 7.42s/step, loss=1.403, acc=0.682, acc_len=1.62, thru=18.5, I=24.1, T=39.0, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:27:23,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:27:24,817] loop.py:45 INFO TRAIN_STEP step=14389 loss=1.170594 Training: 80%|███████▉ | 14389/18018 [30:57:10<7:19:16, 7.26s/step, loss=1.171, acc=0.723, acc_len=1.76, thru=19.3, I=21.0, T=37.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:27:24,849] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 20:27:30,778] loop.py:380 INFO TIMING step=14390: step=3.431s data=1.443s compute=3.406s [fwd=2.421s bwd=0.970s opt=0.003s] dispatch=2.520s [2026-10-19 20:27:30,778] loop.py:45 INFO TRAIN_STEP step=14390 loss=1.255445 Training: 80%|███████▉ | 14390/18018 [30:57:16<6:55:31, 6.87s/step, loss=1.255, acc=0.711, acc_len=1.72, thru=19.3, I=22.6, T=37.3, wait=2.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:27:39,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:27:40,396] loop.py:45 INFO TRAIN_STEP step=14391 loss=1.283655 Training: 80%|███████▉ | 14391/18018 [30:57:26<7:45:14, 7.70s/step, loss=1.284, acc=0.704, acc_len=1.69, thru=15.9, I=15.3, T=33.1, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:27:40,391] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14390: forward=2777.3ms backward=1039.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:27:40,561] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:27:46,522] loop.py:45 INFO TRAIN_STEP step=14392 loss=1.313956 Training: 80%|███████▉ | 14392/18018 [30:57:32<7:16:38, 7.23s/step, loss=1.314, acc=0.694, acc_len=1.66, thru=18.2, I=19.0, T=38.3, wait=2.8s, pool=53, epoch=2/2][2026-10-19 20:27:55,881] loop.py:45 INFO TRAIN_STEP step=14393 loss=1.238798 Training: 80%|███████▉ | 14393/18018 [30:57:41<7:55:08, 7.86s/step, loss=1.239, acc=0.712, acc_len=1.72, thru=14.1, I=13.3, T=34.7, wait=5.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:01,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:28:02,109] loop.py:45 INFO TRAIN_STEP step=14394 loss=1.333142 Training: 80%|███████▉ | 14394/18018 [30:57:48<7:25:23, 7.37s/step, loss=1.333, acc=0.692, acc_len=1.65, thru=19.7, I=19.6, T=38.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:02,450] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:28:09,327] loop.py:380 INFO TIMING step=14395: step=3.312s data=1.360s compute=3.280s [fwd=2.368s bwd=0.898s opt=0.003s] dispatch=3.896s [2026-10-19 20:28:09,327] loop.py:45 INFO TRAIN_STEP step=14395 loss=1.260609 Training: 80%|███████▉ | 14395/18018 [30:57:55<7:22:24, 7.33s/step, loss=1.261, acc=0.709, acc_len=1.70, thru=16.1, I=17.6, T=38.6, wait=3.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:15,617] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:28:17,063] loop.py:45 INFO TRAIN_STEP step=14396 loss=1.223935 Training: 80%|███████▉ | 14396/18018 [30:58:03<7:29:42, 7.45s/step, loss=1.224, acc=0.712, acc_len=1.73, thru=18.8, I=19.7, T=35.3, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:28:17,059] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14395: forward=2607.9ms backward=974.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:17,129] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 20:28:22,534] loop.py:45 INFO TRAIN_STEP step=14397 loss=1.258483 Training: 80%|███████▉ | 14397/18018 [30:58:08<6:53:44, 6.86s/step, loss=1.258, acc=0.704, acc_len=1.69, thru=20.5, I=24.9, T=38.6, wait=2.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:28,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:28:29,930] loop.py:45 INFO TRAIN_STEP step=14398 loss=1.225222 Training: 80%|███████▉ | 14398/18018 [30:58:16<7:03:25, 7.02s/step, loss=1.225, acc=0.713, acc_len=1.72, thru=17.9, I=21.0, T=36.8, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:30,047] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:34,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:28:35,484] loop.py:45 INFO TRAIN_STEP step=14399 loss=1.397486 Training: 80%|███████▉ | 14399/18018 [30:58:21<6:36:51, 6.58s/step, loss=1.397, acc=0.684, acc_len=1.62, thru=22.1, I=28.4, T=39.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:35,704] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:39,917] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:28:41,505] loop.py:380 INFO TIMING step=14400: step=3.444s data=1.344s compute=3.413s [fwd=2.408s bwd=0.991s opt=0.003s] dispatch=2.567s [2026-10-19 20:28:41,505] loop.py:45 INFO TRAIN_STEP step=14400 loss=1.217060 Training: 80%|███████▉ | 14400/18018 [30:58:27<6:26:35, 6.41s/step, loss=1.217, acc=0.716, acc_len=1.73, thru=21.3, I=28.4, T=37.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:28:41,929] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 20:28:49,557] loop.py:45 INFO TRAIN_STEP step=14401 loss=1.164215 Training: 80%|███████▉ | 14401/18018 [30:58:35<6:56:12, 6.90s/step, loss=1.164, acc=0.726, acc_len=1.78, thru=11.6, I=20.3, T=36.7, wait=4.6s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:28:49,553] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14400: forward=2459.1ms backward=979.1ms [2026-10-19 20:28:59,410] loop.py:45 INFO TRAIN_STEP step=14402 loss=1.226014 Training: 80%|███████▉ | 14402/18018 [30:58:45<7:49:25, 7.79s/step, loss=1.226, acc=0.714, acc_len=1.73, thru=15.1, I=12.7, T=34.9, wait=6.2s, pool=37, epoch=2/2][2026-10-19 20:29:06,714] loop.py:45 INFO TRAIN_STEP step=14403 loss=1.265439 Training: 80%|███████▉ | 14403/18018 [30:58:52<7:40:27, 7.64s/step, loss=1.265, acc=0.708, acc_len=1.70, thru=17.7, I=17.6, T=38.6, wait=4.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:29:13,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:29:14,960] loop.py:45 INFO TRAIN_STEP step=14404 loss=1.235803 Training: 80%|███████▉ | 14404/18018 [30:59:01<7:51:18, 7.82s/step, loss=1.236, acc=0.714, acc_len=1.73, thru=16.9, I=19.9, T=35.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:29:15,123] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:29:21,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:29:22,139] loop.py:380 INFO TIMING step=14405: step=3.459s data=1.463s compute=3.409s [fwd=2.400s bwd=0.997s opt=0.003s] dispatch=3.709s [2026-10-19 20:29:22,139] loop.py:45 INFO TRAIN_STEP step=14405 loss=1.305410 Training: 80%|███████▉ | 14405/18018 [30:59:08<7:39:25, 7.63s/step, loss=1.305, acc=0.701, acc_len=1.68, thru=16.1, I=20.4, T=37.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:29:22,389] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:29:27,646] loop.py:45 INFO TRAIN_STEP step=14406 loss=1.156644 Training: 80%|███████▉ | 14406/18018 [30:59:13<7:01:00, 6.99s/step, loss=1.157, acc=0.720, acc_len=1.75, thru=20.6, I=21.7, T=37.3, wait=2.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:29:27,642] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14405: forward=2504.3ms backward=884.2ms [2026-10-19 20:29:33,687] loop.py:45 INFO TRAIN_STEP step=14407 loss=1.285982 Training: 80%|███████▉ | 14407/18018 [30:59:19<6:43:39, 6.71s/step, loss=1.286, acc=0.702, acc_len=1.69, thru=20.0, I=21.7, T=38.5, wait=2.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:29:41,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:29:42,668] loop.py:45 INFO TRAIN_STEP step=14408 loss=1.140487 Training: 80%|███████▉ | 14408/18018 [30:59:28<7:24:36, 7.39s/step, loss=1.140, acc=0.727, acc_len=1.78, thru=16.6, I=16.3, T=37.1, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:29:42,877] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:29:49,222] loop.py:45 INFO TRAIN_STEP step=14409 loss=1.298099 Training: 80%|███████▉ | 14409/18018 [30:59:35<7:09:29, 7.14s/step, loss=1.298, acc=0.699, acc_len=1.68, thru=17.3, I=19.1, T=37.1, wait=3.1s, pool=45, epoch=2/2][2026-10-19 20:30:00,828] loop.py:380 INFO TIMING step=14410: step=3.698s data=1.550s compute=3.659s [fwd=2.646s bwd=1.001s opt=0.003s] dispatch=7.895s [2026-10-19 20:30:00,828] loop.py:45 INFO TRAIN_STEP step=14410 loss=1.188037 Training: 80%|███████▉ | 14410/18018 [30:59:46<8:29:53, 8.48s/step, loss=1.188, acc=0.724, acc_len=1.76, thru=10.1, I=12.1, T=34.6, wait=7.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:30:11,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:30:12,301] loop.py:45 INFO TRAIN_STEP step=14411 loss=1.247896 Training: 80%|███████▉ | 14411/18018 [30:59:58<9:23:40, 9.38s/step, loss=1.248, acc=0.712, acc_len=1.72, thru=17.3, I=12.0, T=34.9, wait=7.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:30:12,297] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14410: forward=2522.6ms backward=1085.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:30:12,769] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:30:22,225] loop.py:45 INFO TRAIN_STEP step=14412 loss=1.369695 Training: 80%|███████▉ | 14412/18018 [31:00:08<9:33:25, 9.54s/step, loss=1.370, acc=0.690, acc_len=1.64, thru=11.8, I=11.2, T=34.1, wait=6.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:30:30,905] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:30:31,139] loop.py:45 INFO TRAIN_STEP step=14413 loss=1.241549 Training: 80%|███████▉ | 14413/18018 [31:00:17<9:21:57, 9.35s/step, loss=1.242, acc=0.712, acc_len=1.72, thru=16.5, I=17.5, T=34.8, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:30:31,407] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:30:41,296] loop.py:45 INFO TRAIN_STEP step=14414 loss=1.313369 Training: 80%|███████▉ | 14414/18018 [31:00:27<9:36:16, 9.59s/step, loss=1.313, acc=0.702, acc_len=1.68, thru=12.5, I=12.2, T=36.2, wait=6.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:30:48,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:30:48,786] loop.py:380 INFO TIMING step=14415: step=3.469s data=1.436s compute=3.441s [fwd=2.466s bwd=0.961s opt=0.003s] dispatch=4.013s [2026-10-19 20:30:48,787] loop.py:45 INFO TRAIN_STEP step=14415 loss=1.359611 Training: 80%|████████ | 14415/18018 [31:00:34<8:58:13, 8.96s/step, loss=1.360, acc=0.689, acc_len=1.65, thru=18.5, I=17.5, T=36.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:30:48,863] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:30:57,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:30:58,852] loop.py:45 INFO TRAIN_STEP step=14416 loss=1.176609 Training: 80%|████████ | 14416/18018 [31:00:44<9:17:56, 9.29s/step, loss=1.177, acc=0.720, acc_len=1.75, thru=13.0, I=14.5, T=34.8, wait=6.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:30:58,847] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14415: forward=2597.1ms backward=1031.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:30:59,333] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:05,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:31:05,634] loop.py:45 INFO TRAIN_STEP step=14417 loss=1.252210 Training: 80%|████████ | 14417/18018 [31:00:51<8:32:35, 8.54s/step, loss=1.252, acc=0.710, acc_len=1.71, thru=18.4, I=24.7, T=35.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:05,767] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:11,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:31:12,480] loop.py:45 INFO TRAIN_STEP step=14418 loss=1.218492 Training: 80%|████████ | 14418/18018 [31:00:58<8:01:56, 8.03s/step, loss=1.218, acc=0.712, acc_len=1.72, thru=17.7, I=20.1, T=36.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:12,709] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:31:18,829] loop.py:45 INFO TRAIN_STEP step=14419 loss=1.366275 Training: 80%|████████ | 14419/18018 [31:01:04<7:31:28, 7.53s/step, loss=1.366, acc=0.687, acc_len=1.63, thru=18.9, I=20.2, T=36.8, wait=2.9s, pool=49, epoch=2/2][2026-10-19 20:31:26,033] loop.py:380 INFO TIMING step=14420: step=3.599s data=1.507s compute=3.537s [fwd=2.441s bwd=1.083s opt=0.003s] dispatch=3.597s [2026-10-19 20:31:26,033] loop.py:45 INFO TRAIN_STEP step=14420 loss=1.283687 Training: 80%|████████ | 14420/18018 [31:01:12<7:25:33, 7.43s/step, loss=1.284, acc=0.705, acc_len=1.70, thru=16.8, I=17.2, T=35.6, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:33,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:31:34,301] loop.py:45 INFO TRAIN_STEP step=14421 loss=1.364420 Training: 80%|████████ | 14421/18018 [31:01:20<7:40:31, 7.68s/step, loss=1.364, acc=0.687, acc_len=1.64, thru=17.4, I=16.8, T=33.1, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:31:34,297] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14420: forward=2874.8ms backward=952.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:34,516] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:39,581] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:31:40,051] loop.py:45 INFO TRAIN_STEP step=14422 loss=1.229106 Training: 80%|████████ | 14422/18018 [31:01:26<7:05:37, 7.10s/step, loss=1.229, acc=0.710, acc_len=1.71, thru=20.4, I=21.9, T=37.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:40,099] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:47,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:31:47,677] loop.py:45 INFO TRAIN_STEP step=14423 loss=1.281521 Training: 80%|████████ | 14423/18018 [31:01:33<7:14:58, 7.26s/step, loss=1.282, acc=0.704, acc_len=1.69, thru=17.6, I=18.2, T=37.7, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:47,837] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:54,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:31:56,335] loop.py:45 INFO TRAIN_STEP step=14424 loss=1.204482 Training: 80%|████████ | 14424/18018 [31:01:42<7:39:59, 7.68s/step, loss=1.204, acc=0.717, acc_len=1.74, thru=15.0, I=19.1, T=30.0, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:31:56,607] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:01,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:32:02,333] loop.py:380 INFO TIMING step=14425: step=3.363s data=1.386s compute=3.333s [fwd=2.383s bwd=0.937s opt=0.003s] dispatch=2.624s [2026-10-19 20:32:02,334] loop.py:45 INFO TRAIN_STEP step=14425 loss=1.248018 Training: 80%|████████ | 14425/18018 [31:01:48<7:09:39, 7.17s/step, loss=1.248, acc=0.707, acc_len=1.70, thru=18.8, I=26.8, T=38.1, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:02,673] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:32:10,434] loop.py:45 INFO TRAIN_STEP step=14426 loss=1.317109 Training: 80%|████████ | 14426/18018 [31:01:56<7:26:07, 7.45s/step, loss=1.317, acc=0.695, acc_len=1.66, thru=14.8, I=17.8, T=35.7, wait=4.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:32:10,430] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14425: forward=2551.9ms backward=986.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:17,198] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:32:18,016] loop.py:45 INFO TRAIN_STEP step=14427 loss=1.182696 Training: 80%|████████ | 14427/18018 [31:02:04<7:28:23, 7.49s/step, loss=1.183, acc=0.721, acc_len=1.75, thru=17.4, I=18.8, T=36.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:18,223] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:22,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:32:23,985] loop.py:45 INFO TRAIN_STEP step=14428 loss=1.261512 Training: 80%|████████ | 14428/18018 [31:02:10<7:00:53, 7.03s/step, loss=1.262, acc=0.708, acc_len=1.70, thru=20.9, I=25.9, T=37.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:24,446] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:32:31,346] loop.py:45 INFO TRAIN_STEP step=14429 loss=1.238941 Training: 80%|████████ | 14429/18018 [31:02:17<7:06:38, 7.13s/step, loss=1.239, acc=0.708, acc_len=1.70, thru=16.1, I=19.8, T=38.5, wait=4.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:39,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:32:39,782] loop.py:380 INFO TIMING step=14430: step=3.702s data=1.581s compute=3.666s [fwd=2.579s bwd=1.074s opt=0.003s] dispatch=4.725s [2026-10-19 20:32:39,782] loop.py:45 INFO TRAIN_STEP step=14430 loss=1.428859 Training: 80%|████████ | 14430/18018 [31:02:25<7:29:53, 7.52s/step, loss=1.429, acc=0.675, acc_len=1.60, thru=17.2, I=16.5, T=34.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:40,230] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:46,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:32:46,891] loop.py:45 INFO TRAIN_STEP step=14431 loss=1.327762 Training: 80%|████████ | 14431/18018 [31:02:32<7:22:22, 7.40s/step, loss=1.328, acc=0.694, acc_len=1.66, thru=18.0, I=19.9, T=37.3, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:32:46,887] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14430: forward=2418.9ms backward=946.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:47,224] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:51,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:32:52,713] loop.py:45 INFO TRAIN_STEP step=14432 loss=1.250243 Training: 80%|████████ | 14432/18018 [31:02:38<6:53:56, 6.93s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=21.7, I=25.5, T=38.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:52,949] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:58,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:32:58,641] loop.py:45 INFO TRAIN_STEP step=14433 loss=1.329748 Training: 80%|████████ | 14433/18018 [31:02:44<6:35:55, 6.63s/step, loss=1.330, acc=0.696, acc_len=1.66, thru=22.6, I=25.0, T=38.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:32:58,751] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:33:03,162] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:33:03,917] loop.py:45 INFO TRAIN_STEP step=14434 loss=1.178196 Training: 80%|████████ | 14434/18018 [31:02:50<6:11:37, 6.22s/step, loss=1.178, acc=0.721, acc_len=1.75, thru=23.3, I=25.8, T=37.8, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:33:04,169] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:33:08,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:33:09,809] loop.py:380 INFO TIMING step=14435: step=3.289s data=1.376s compute=3.261s [fwd=2.328s bwd=0.919s opt=0.003s] dispatch=2.593s [2026-10-19 20:33:09,809] loop.py:45 INFO TRAIN_STEP step=14435 loss=1.178131 Training: 80%|████████ | 14435/18018 [31:02:55<6:05:37, 6.12s/step, loss=1.178, acc=0.720, acc_len=1.75, thru=21.7, I=26.5, T=38.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:33:09,883] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (TrainerActor pid=9310) [2026-10-19 20:33:15,620] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14435: forward=2373.7ms backward=903.6ms [2026-10-19 20:33:15,625] loop.py:45 INFO TRAIN_STEP step=14436 loss=1.297337 Training: 80%|████████ | 14436/18018 [31:03:01<6:00:01, 6.03s/step, loss=1.297, acc=0.700, acc_len=1.68, thru=18.6, I=23.4, T=38.6, wait=2.5s, pool=33, epoch=2/2][2026-10-19 20:33:25,941] loop.py:45 INFO TRAIN_STEP step=14437 loss=1.297469 Training: 80%|████████ | 14437/18018 [31:03:12<7:16:40, 7.32s/step, loss=1.297, acc=0.701, acc_len=1.68, thru=15.1, I=16.4, T=34.4, wait=6.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:33:33,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:33:34,941] loop.py:45 INFO TRAIN_STEP step=14438 loss=1.350351 Training: 80%|████████ | 14438/18018 [31:03:21<7:46:38, 7.82s/step, loss=1.350, acc=0.694, acc_len=1.66, thru=14.3, I=15.3, T=34.4, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:33:35,271] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:33:41,906] loop.py:45 INFO TRAIN_STEP step=14439 loss=1.172940 Training: 80%|████████ | 14439/18018 [31:03:27<7:31:13, 7.56s/step, loss=1.173, acc=0.722, acc_len=1.75, thru=15.5, I=18.0, T=36.8, wait=3.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:33:49,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:33:49,324] loop.py:380 INFO TIMING step=14440: step=3.423s data=1.442s compute=3.392s [fwd=2.452s bwd=0.926s opt=0.003s] dispatch=3.986s [2026-10-19 20:33:49,324] loop.py:45 INFO TRAIN_STEP step=14440 loss=1.294315 Training: 80%|████████ | 14440/18018 [31:03:35<7:28:25, 7.52s/step, loss=1.294, acc=0.701, acc_len=1.68, thru=19.6, I=17.3, T=37.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:33:49,715] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:33:55,441] loop.py:45 INFO TRAIN_STEP step=14441 loss=1.306799 Training: 80%|████████ | 14441/18018 [31:03:41<7:03:15, 7.10s/step, loss=1.307, acc=0.697, acc_len=1.67, thru=18.9, I=19.3, T=38.4, wait=2.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:33:55,437] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14440: forward=2376.8ms backward=920.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:02,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:34:02,559] loop.py:45 INFO TRAIN_STEP step=14442 loss=1.291846 Training: 80%|████████ | 14442/18018 [31:03:48<7:03:27, 7.10s/step, loss=1.292, acc=0.698, acc_len=1.68, thru=18.5, I=19.0, T=36.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:02,683] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:07,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:34:08,012] loop.py:45 INFO TRAIN_STEP step=14443 loss=1.362582 Training: 80%|████████ | 14443/18018 [31:03:54<6:33:48, 6.61s/step, loss=1.363, acc=0.691, acc_len=1.65, thru=22.2, I=25.3, T=39.0, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:08,261] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:14,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:34:14,773] loop.py:45 INFO TRAIN_STEP step=14444 loss=1.301008 Training: 80%|████████ | 14444/18018 [31:04:00<6:36:24, 6.65s/step, loss=1.301, acc=0.701, acc_len=1.67, thru=19.4, I=21.1, T=38.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:14,880] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:21,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:34:22,182] loop.py:380 INFO TIMING step=14445: step=3.457s data=1.413s compute=3.431s [fwd=2.474s bwd=0.942s opt=0.003s] dispatch=3.942s [2026-10-19 20:34:22,182] loop.py:45 INFO TRAIN_STEP step=14445 loss=1.398872 Training: 80%|████████ | 14445/18018 [31:04:08<6:49:47, 6.88s/step, loss=1.399, acc=0.682, acc_len=1.62, thru=18.2, I=19.6, T=37.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:22,292] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:27,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:34:27,875] loop.py:45 INFO TRAIN_STEP step=14446 loss=1.274811 Training: 80%|████████ | 14446/18018 [31:04:13<6:28:25, 6.52s/step, loss=1.275, acc=0.704, acc_len=1.69, thru=20.9, I=23.5, T=39.4, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:34:27,870] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14445: forward=2312.3ms backward=891.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:28,178] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:34:35,357] loop.py:45 INFO TRAIN_STEP step=14447 loss=1.255482 Training: 80%|████████ | 14447/18018 [31:04:21<6:45:24, 6.81s/step, loss=1.255, acc=0.704, acc_len=1.69, thru=16.1, I=18.7, T=36.5, wait=4.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:44,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:34:44,376] loop.py:45 INFO TRAIN_STEP step=14448 loss=1.303761 Training: 80%|████████ | 14448/18018 [31:04:30<7:24:40, 7.47s/step, loss=1.304, acc=0.695, acc_len=1.66, thru=15.8, I=16.8, T=32.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:44,758] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:51,254] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:34:51,875] loop.py:45 INFO TRAIN_STEP step=14449 loss=1.218182 Training: 80%|████████ | 14449/18018 [31:04:37<7:25:00, 7.48s/step, loss=1.218, acc=0.715, acc_len=1.73, thru=16.7, I=17.7, T=37.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:52,259] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:58,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:34:58,823] loop.py:380 INFO TIMING step=14450: step=3.493s data=1.446s compute=3.462s [fwd=2.509s bwd=0.939s opt=0.003s] dispatch=3.447s [2026-10-19 20:34:58,823] loop.py:45 INFO TRAIN_STEP step=14450 loss=1.356618 Training: 80%|████████ | 14450/18018 [31:04:44<7:15:23, 7.32s/step, loss=1.357, acc=0.690, acc_len=1.65, thru=18.9, I=20.4, T=36.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:34:58,996] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:03,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:35:04,786] loop.py:45 INFO TRAIN_STEP step=14451 loss=1.290664 Training: 80%|████████ | 14451/18018 [31:04:50<6:51:01, 6.91s/step, loss=1.291, acc=0.701, acc_len=1.69, thru=21.1, I=23.0, T=38.3, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:35:04,781] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14450: forward=2368.7ms backward=930.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:04,851] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:10,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:35:11,096] loop.py:45 INFO TRAIN_STEP step=14452 loss=1.274300 Training: 80%|████████ | 14452/18018 [31:04:57<6:40:08, 6.73s/step, loss=1.274, acc=0.705, acc_len=1.70, thru=20.1, I=23.5, T=37.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:11,231] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 20:35:17,635] loop.py:45 INFO TRAIN_STEP step=14453 loss=1.240502 Training: 80%|████████ | 14453/18018 [31:05:03<6:36:41, 6.68s/step, loss=1.241, acc=0.714, acc_len=1.72, thru=15.2, I=22.2, T=38.3, wait=3.2s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 20:35:17,942] eagle_store.py:517 WARNING batch_get_buffer missing keys (k462506-b46f3bf5_hs, k462506-b46f3bf5_ids, k462506-b46f3bf5_lhs); sleeping 0.50s. [2026-10-19 20:35:26,810] loop.py:45 INFO TRAIN_STEP step=14454 loss=1.356290 Training: 80%|████████ | 14454/18018 [31:05:12<7:20:59, 7.42s/step, loss=1.356, acc=0.688, acc_len=1.63, thru=15.8, I=12.3, T=36.8, wait=5.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:36,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 20:35:18,006] eagle_store.py:517 WARNING batch_get_buffer missing keys (k462504-a354a809_hs, k462504-a354a809_ids, k462504-a354a809_lhs); sleeping 0.50s. [2026-10-19 20:35:36,473] loop.py:380 INFO TIMING step=14455: step=3.659s data=1.550s compute=3.612s [fwd=2.517s bwd=1.080s opt=0.003s] dispatch=5.996s [2026-10-19 20:35:36,473] loop.py:45 INFO TRAIN_STEP step=14455 loss=1.292347 Training: 80%|████████ | 14455/18018 [31:05:22<8:00:45, 8.10s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=16.1, I=16.2, T=35.0, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:36,598] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:42,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:35:43,310] loop.py:45 INFO TRAIN_STEP step=14456 loss=1.218950 Training: 80%|████████ | 14456/18018 [31:05:29<7:38:11, 7.72s/step, loss=1.219, acc=0.715, acc_len=1.73, thru=18.7, I=19.9, T=29.7, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:35:43,305] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14455: forward=3369.4ms backward=887.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:43,447] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:50,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:35:52,926] loop.py:45 INFO TRAIN_STEP step=14457 loss=1.300816 Training: 80%|████████ | 14457/18018 [31:05:39<8:11:53, 8.29s/step, loss=1.301, acc=0.695, acc_len=1.66, thru=13.5, I=17.1, T=32.5, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:35:53,118] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 20:35:59,639] loop.py:45 INFO TRAIN_STEP step=14458 loss=1.361313 Training: 80%|████████ | 14458/18018 [31:05:45<7:43:42, 7.82s/step, loss=1.361, acc=0.687, acc_len=1.63, thru=16.0, I=24.6, T=37.2, wait=3.3s, pool=57, epoch=2/2][2026-10-19 20:36:08,112] loop.py:45 INFO TRAIN_STEP step=14459 loss=1.188604 Training: 80%|████████ | 14459/18018 [31:05:54<7:55:15, 8.01s/step, loss=1.189, acc=0.720, acc_len=1.76, thru=14.6, I=13.9, T=37.1, wait=5.0s, pool=53, epoch=2/2][2026-10-19 20:36:14,927] loop.py:380 INFO TIMING step=14460: step=3.519s data=1.424s compute=3.491s [fwd=2.510s bwd=0.966s opt=0.003s] dispatch=3.288s [2026-10-19 20:36:14,927] loop.py:45 INFO TRAIN_STEP step=14460 loss=1.407384 Training: 80%|████████ | 14460/18018 [31:06:01<7:33:52, 7.65s/step, loss=1.407, acc=0.682, acc_len=1.62, thru=17.6, I=19.7, T=36.4, wait=3.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:21,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 20:36:22,599] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14460: forward=3014.2ms backward=948.0ms [2026-10-19 20:36:22,604] loop.py:45 INFO TRAIN_STEP step=14461 loss=1.210729 Training: 80%|████████ | 14461/18018 [31:06:08<7:34:06, 7.66s/step, loss=1.211, acc=0.715, acc_len=1.73, thru=16.7, I=19.4, T=31.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:22,633] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:29,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:36:30,897] loop.py:45 INFO TRAIN_STEP step=14462 loss=1.284828 Training: 80%|████████ | 14462/18018 [31:06:16<7:45:16, 7.85s/step, loss=1.285, acc=0.706, acc_len=1.70, thru=14.0, I=19.5, T=34.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:31,095] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:37,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:36:38,310] loop.py:45 INFO TRAIN_STEP step=14463 loss=1.199442 Training: 80%|████████ | 14463/18018 [31:06:24<7:37:21, 7.72s/step, loss=1.199, acc=0.716, acc_len=1.74, thru=17.0, I=22.2, T=30.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:38,673] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:43,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:36:44,467] loop.py:45 INFO TRAIN_STEP step=14464 loss=1.235817 Training: 80%|████████ | 14464/18018 [31:06:30<7:09:25, 7.25s/step, loss=1.236, acc=0.711, acc_len=1.72, thru=19.6, I=22.5, T=38.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:44,892] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:36:50,695] loop.py:380 INFO TIMING step=14465: step=3.279s data=1.302s compute=3.245s [fwd=2.325s bwd=0.905s opt=0.003s] dispatch=2.941s [2026-10-19 20:36:50,695] loop.py:45 INFO TRAIN_STEP step=14465 loss=1.272642 Training: 80%|████████ | 14465/18018 [31:06:36<6:51:12, 6.94s/step, loss=1.273, acc=0.706, acc_len=1.70, thru=20.0, I=24.8, T=39.0, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:56,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:36:57,024] loop.py:45 INFO TRAIN_STEP step=14466 loss=1.175324 Training: 80%|████████ | 14466/18018 [31:06:43<6:40:06, 6.76s/step, loss=1.175, acc=0.722, acc_len=1.76, thru=20.1, I=21.7, T=35.7, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:36:57,020] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14465: forward=2568.0ms backward=967.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:36:57,515] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:37:02,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:37:02,896] loop.py:45 INFO TRAIN_STEP step=14467 loss=1.341791 Training: 80%|████████ | 14467/18018 [31:06:48<6:24:17, 6.49s/step, loss=1.342, acc=0.689, acc_len=1.64, thru=20.5, I=25.7, T=37.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:37:03,212] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:37:08,927] loop.py:45 INFO TRAIN_STEP step=14468 loss=1.179397 Training: 80%|████████ | 14468/18018 [31:06:55<6:15:59, 6.35s/step, loss=1.179, acc=0.721, acc_len=1.76, thru=19.8, I=23.5, T=38.9, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:37:17,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:37:17,904] loop.py:45 INFO TRAIN_STEP step=14469 loss=1.283902 Training: 80%|████████ | 14469/18018 [31:07:03<7:02:23, 7.14s/step, loss=1.284, acc=0.704, acc_len=1.70, thru=15.2, I=14.1, T=36.2, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:37:18,199] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:37:24,098] loop.py:380 INFO TIMING step=14470: step=3.270s data=1.307s compute=3.232s [fwd=2.324s bwd=0.894s opt=0.003s] dispatch=2.916s [2026-10-19 20:37:24,099] loop.py:45 INFO TRAIN_STEP step=14470 loss=1.312243 Training: 80%|████████ | 14470/18018 [31:07:10<6:45:28, 6.86s/step, loss=1.312, acc=0.698, acc_len=1.68, thru=19.2, I=19.4, T=39.1, wait=2.9s, pool=57, epoch=2/2][2026-10-19 20:37:32,413] loop.py:45 INFO TRAIN_STEP step=14471 loss=1.396897 Training: 80%|████████ | 14471/18018 [31:07:18<7:11:18, 7.30s/step, loss=1.397, acc=0.685, acc_len=1.63, thru=15.9, I=15.4, T=36.7, wait=4.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:37:32,408] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14470: forward=2493.4ms backward=948.9ms [2026-10-19 20:37:39,208] loop.py:45 INFO TRAIN_STEP step=14472 loss=1.313900 Training: 80%|████████ | 14472/18018 [31:07:25<7:02:16, 7.15s/step, loss=1.314, acc=0.698, acc_len=1.68, thru=18.1, I=17.9, T=38.3, wait=3.4s, pool=53, epoch=2/2][2026-10-19 20:37:46,620] loop.py:45 INFO TRAIN_STEP step=14473 loss=1.249624 Training: 80%|████████ | 14473/18018 [31:07:32<7:06:48, 7.22s/step, loss=1.250, acc=0.709, acc_len=1.71, thru=18.1, I=19.1, T=35.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:37:46,587] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:37:47,088] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:37:52,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:37:52,932] loop.py:45 INFO TRAIN_STEP step=14474 loss=1.295581 Training: 80%|████████ | 14474/18018 [31:07:39<6:50:33, 6.95s/step, loss=1.296, acc=0.699, acc_len=1.67, thru=19.6, I=20.8, T=38.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:37:53,065] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:38:00,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:38:00,948] loop.py:380 INFO TIMING step=14475: step=3.195s data=1.282s compute=3.168s [fwd=2.280s bwd=0.874s opt=0.003s] dispatch=4.812s [2026-10-19 20:38:00,948] loop.py:45 INFO TRAIN_STEP step=14475 loss=1.273340 Training: 80%|████████ | 14475/18018 [31:07:47<7:09:18, 7.27s/step, loss=1.273, acc=0.705, acc_len=1.70, thru=17.3, I=18.0, T=40.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:38:01,444] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:38:07,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:38:08,261] loop.py:45 INFO TRAIN_STEP step=14476 loss=1.197117 Training: 80%|████████ | 14476/18018 [31:07:54<7:09:57, 7.28s/step, loss=1.197, acc=0.721, acc_len=1.75, thru=19.2, I=18.2, T=36.1, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:38:08,257] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14475: forward=2508.9ms backward=988.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:38:08,488] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:38:18,526] loop.py:45 INFO TRAIN_STEP step=14477 loss=1.232247 Training: 80%|████████ | 14477/18018 [31:08:04<8:02:37, 8.18s/step, loss=1.232, acc=0.715, acc_len=1.73, thru=11.5, I=11.6, T=35.2, wait=6.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:38:27,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:38:28,170] loop.py:45 INFO TRAIN_STEP step=14478 loss=1.175410 Training: 80%|████████ | 14478/18018 [31:08:14<8:28:26, 8.62s/step, loss=1.175, acc=0.726, acc_len=1.77, thru=14.8, I=15.7, T=33.9, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:38:28,319] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:38:34,313] loop.py:45 INFO TRAIN_STEP step=14479 loss=1.266738 Training: 80%|████████ | 14479/18018 [31:08:20<7:44:31, 7.88s/step, loss=1.267, acc=0.707, acc_len=1.71, thru=16.9, I=21.7, T=38.5, wait=2.8s, pool=37, epoch=2/2][2026-10-19 20:38:42,749] loop.py:380 INFO TIMING step=14480: step=3.447s data=1.353s compute=3.412s [fwd=2.462s bwd=0.936s opt=0.003s] dispatch=4.977s [2026-10-19 20:38:42,749] loop.py:45 INFO TRAIN_STEP step=14480 loss=1.155846 Training: 80%|████████ | 14480/18018 [31:08:28<7:54:15, 8.04s/step, loss=1.156, acc=0.728, acc_len=1.78, thru=15.3, I=14.7, T=37.1, wait=5.0s, pool=49, epoch=2/2][2026-10-19 20:38:49,804] loop.py:45 INFO TRAIN_STEP step=14481 loss=1.209860 Training: 80%|████████ | 14481/18018 [31:08:35<7:36:42, 7.75s/step, loss=1.210, acc=0.715, acc_len=1.73, thru=16.5, I=17.1, T=37.6, wait=3.6s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:38:49,799] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14480: forward=2432.8ms backward=930.4ms [2026-10-19 20:38:56,623] loop.py:45 INFO TRAIN_STEP step=14482 loss=1.450969 Training: 80%|████████ | 14482/18018 [31:08:42<7:20:07, 7.47s/step, loss=1.451, acc=0.673, acc_len=1.59, thru=19.2, I=19.9, T=35.9, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:39:04,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:39:05,570] loop.py:45 INFO TRAIN_STEP step=14483 loss=1.244225 Training: 80%|████████ | 14483/18018 [31:08:51<7:46:10, 7.91s/step, loss=1.244, acc=0.711, acc_len=1.72, thru=14.4, I=15.2, T=36.3, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:39:05,805] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:39:13,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:39:14,148] loop.py:45 INFO TRAIN_STEP step=14484 loss=1.173179 Training: 80%|████████ | 14484/18018 [31:09:00<7:57:47, 8.11s/step, loss=1.173, acc=0.721, acc_len=1.75, thru=15.4, I=18.1, T=36.8, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:39:14,627] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:39:24,067] loop.py:380 INFO TIMING step=14485: step=3.807s data=1.559s compute=3.746s [fwd=2.626s bwd=1.104s opt=0.003s] dispatch=6.102s [2026-10-19 20:39:24,067] loop.py:45 INFO TRAIN_STEP step=14485 loss=1.218154 Training: 80%|████████ | 14485/18018 [31:09:10<8:29:35, 8.65s/step, loss=1.218, acc=0.714, acc_len=1.73, thru=10.2, I=10.8, T=33.6, wait=6.1s, pool=33, epoch=2/2][2026-10-19 20:39:34,639] loop.py:45 INFO TRAIN_STEP step=14486 loss=1.255565 Training: 80%|████████ | 14486/18018 [31:09:20<9:03:19, 9.23s/step, loss=1.256, acc=0.709, acc_len=1.71, thru=11.1, I=12.8, T=33.2, wait=6.7s, pool=21, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:39:34,634] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14485: forward=2758.5ms backward=1043.1ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 20:39:34,698] eagle_store.py:517 WARNING batch_get_buffer missing keys (k463550-91f4c008_hs, k463550-91f4c008_ids, k463550-91f4c008_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:39:41,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 20:39:34,889] eagle_store.py:517 WARNING batch_get_buffer missing keys (k463561-8df4de3a_hs, k463561-8df4de3a_ids, k463561-8df4de3a_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 20:39:42,702] loop.py:45 INFO TRAIN_STEP step=14487 loss=1.189620 Training: 80%|████████ | 14487/18018 [31:09:28<8:42:31, 8.88s/step, loss=1.190, acc=0.724, acc_len=1.77, thru=21.0, I=17.5, T=34.7, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:39:43,191] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:39:48,076] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:39:48,267] loop.py:45 INFO TRAIN_STEP step=14488 loss=1.140825 Training: 80%|████████ | 14488/18018 [31:09:34<7:43:53, 7.88s/step, loss=1.141, acc=0.727, acc_len=1.77, thru=21.1, I=25.4, T=38.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:39:48,578] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:39:54,158] loop.py:45 INFO TRAIN_STEP step=14489 loss=1.282279 Training: 80%|████████ | 14489/18018 [31:09:40<7:08:35, 7.29s/step, loss=1.282, acc=0.701, acc_len=1.68, thru=20.8, I=23.7, T=39.2, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:40:01,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:40:01,933] loop.py:380 INFO TIMING step=14490: step=3.500s data=1.437s compute=3.455s [fwd=2.403s bwd=1.037s opt=0.003s] dispatch=4.266s [2026-10-19 20:40:01,934] loop.py:45 INFO TRAIN_STEP step=14490 loss=1.246281 Training: 80%|████████ | 14490/18018 [31:09:48<7:17:04, 7.43s/step, loss=1.246, acc=0.707, acc_len=1.71, thru=17.4, I=18.1, T=36.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:40:02,135] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:40:10,709] loop.py:45 INFO TRAIN_STEP step=14491 loss=1.237588 Training: 80%|████████ | 14491/18018 [31:09:56<7:40:39, 7.84s/step, loss=1.238, acc=0.707, acc_len=1.71, thru=13.4, I=15.0, T=36.0, wait=5.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:40:10,705] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14490: forward=2475.8ms backward=1022.4ms [2026-10-19 20:40:17,729] loop.py:45 INFO TRAIN_STEP step=14492 loss=1.181950 Training: 80%|████████ | 14492/18018 [31:10:03<7:26:05, 7.59s/step, loss=1.182, acc=0.723, acc_len=1.76, thru=16.0, I=16.6, T=37.7, wait=3.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:40:24,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:40:24,998] loop.py:45 INFO TRAIN_STEP step=14493 loss=1.291801 Training: 80%|████████ | 14493/18018 [31:10:11<7:20:16, 7.49s/step, loss=1.292, acc=0.701, acc_len=1.68, thru=19.1, I=19.4, T=37.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:40:25,026] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 20:40:31,339] loop.py:45 INFO TRAIN_STEP step=14494 loss=1.209373 Training: 80%|████████ | 14494/18018 [31:10:17<6:59:54, 7.15s/step, loss=1.209, acc=0.713, acc_len=1.73, thru=17.2, I=19.8, T=37.0, wait=2.9s, pool=37, epoch=2/2][2026-10-19 20:40:39,729] loop.py:380 INFO TIMING step=14495: step=3.559s data=1.506s compute=3.497s [fwd=2.443s bwd=1.038s opt=0.003s] dispatch=4.818s [2026-10-19 20:40:39,729] loop.py:45 INFO TRAIN_STEP step=14495 loss=1.228506 Training: 80%|████████ | 14495/18018 [31:10:25<7:21:34, 7.52s/step, loss=1.229, acc=0.710, acc_len=1.71, thru=15.9, I=13.8, T=36.0, wait=4.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:40:49,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:40:50,222] loop.py:45 INFO TRAIN_STEP step=14496 loss=1.274574 Training: 80%|████████ | 14496/18018 [31:10:36<8:13:49, 8.41s/step, loss=1.275, acc=0.706, acc_len=1.70, thru=15.3, I=15.0, T=34.4, wait=6.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:40:50,217] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14495: forward=2515.8ms backward=1137.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:40:50,415] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:40:57,233] loop.py:45 INFO TRAIN_STEP step=14497 loss=1.393119 Training: 80%|████████ | 14497/18018 [31:10:43<7:48:58, 7.99s/step, loss=1.393, acc=0.682, acc_len=1.62, thru=14.1, I=20.0, T=37.3, wait=3.6s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:08,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:41:08,957] loop.py:45 INFO TRAIN_STEP step=14498 loss=1.320217 Training: 80%|████████ | 14498/18018 [31:10:55<8:54:34, 9.11s/step, loss=1.320, acc=0.697, acc_len=1.67, thru=16.8, I=14.3, T=33.4, wait=7.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:09,100] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 20:41:14,880] loop.py:45 INFO TRAIN_STEP step=14499 loss=1.306164 Training: 80%|████████ | 14499/18018 [31:11:00<7:58:17, 8.16s/step, loss=1.306, acc=0.697, acc_len=1.67, thru=19.2, I=22.3, T=38.3, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:26,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:41:26,631] loop.py:380 INFO TIMING step=14500: step=3.819s data=1.512s compute=3.777s [fwd=2.668s bwd=1.094s opt=0.003s] dispatch=7.923s [2026-10-19 20:41:26,631] loop.py:45 INFO TRAIN_STEP step=14500 loss=1.303265 Training: 80%|████████ | 14500/18018 [31:11:12<9:01:24, 9.23s/step, loss=1.303, acc=0.700, acc_len=1.68, thru=12.7, I=11.8, T=33.5, wait=7.9s, pool=65, epoch=2/2][2026-10-19 20:41:26,635] loop.py:419 INFO Saving checkpoint at step 14500... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 20:41:30,010] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0014501 (TrainerActor pid=9310) [2026-10-19 20:41:30,022] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0013751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:31,392] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:31,392] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:36,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 20:41:30,009] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0014501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 20:41:31,269] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0013751 [2026-10-19 20:41:37,927] loop.py:45 INFO TRAIN_STEP step=14501 loss=1.242953 Training: 80%|████████ | 14501/18018 [31:11:24<9:37:29, 9.85s/step, loss=1.243, acc=0.711, acc_len=1.72, thru=19.6, I=23.4, T=36.5, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:41:37,923] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14500: forward=2484.9ms backward=969.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:38,406] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:44,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:41:44,660] loop.py:45 INFO TRAIN_STEP step=14502 loss=1.201221 Training: 80%|████████ | 14502/18018 [31:11:30<8:42:29, 8.92s/step, loss=1.201, acc=0.717, acc_len=1.74, thru=18.1, I=22.5, T=38.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:44,735] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:49,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:41:50,241] loop.py:45 INFO TRAIN_STEP step=14503 loss=1.290099 Training: 80%|████████ | 14503/18018 [31:11:36<7:43:46, 7.92s/step, loss=1.290, acc=0.704, acc_len=1.69, thru=21.7, I=23.7, T=38.5, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:50,662] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:57,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:41:57,697] loop.py:45 INFO TRAIN_STEP step=14504 loss=1.278128 Training: 80%|████████ | 14504/18018 [31:11:43<7:35:30, 7.78s/step, loss=1.278, acc=0.702, acc_len=1.69, thru=17.6, I=19.8, T=37.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:41:57,804] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 20:42:04,034] loop.py:380 INFO TIMING step=14505: step=3.306s data=1.335s compute=3.282s [fwd=2.377s bwd=0.890s opt=0.003s] dispatch=3.023s [2026-10-19 20:42:04,034] loop.py:45 INFO TRAIN_STEP step=14505 loss=1.315674 Training: 81%|████████ | 14505/18018 [31:11:50<7:10:03, 7.35s/step, loss=1.316, acc=0.699, acc_len=1.68, thru=19.7, I=20.6, T=38.7, wait=3.0s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:42:11,983] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14505: forward=2378.6ms backward=925.4ms [2026-10-19 20:42:11,988] loop.py:45 INFO TRAIN_STEP step=14506 loss=1.112151 Training: 81%|████████ | 14506/18018 [31:11:58<7:20:37, 7.53s/step, loss=1.112, acc=0.733, acc_len=1.79, thru=15.5, I=15.6, T=38.2, wait=4.6s, pool=41, epoch=2/2][2026-10-19 20:42:20,965] loop.py:45 INFO TRAIN_STEP step=14507 loss=1.177427 Training: 81%|████████ | 14507/18018 [31:12:07<7:45:56, 7.96s/step, loss=1.177, acc=0.719, acc_len=1.75, thru=15.5, I=14.2, T=36.1, wait=5.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:42:29,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:42:29,905] loop.py:45 INFO TRAIN_STEP step=14508 loss=1.211514 Training: 81%|████████ | 14508/18018 [31:12:15<8:02:59, 8.26s/step, loss=1.212, acc=0.719, acc_len=1.74, thru=15.8, I=16.0, T=35.5, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:42:30,214] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:42:34,933] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:42:35,676] loop.py:45 INFO TRAIN_STEP step=14509 loss=1.125199 Training: 81%|████████ | 14509/18018 [31:12:21<7:19:16, 7.51s/step, loss=1.125, acc=0.729, acc_len=1.79, thru=21.4, I=24.4, T=38.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:42:35,938] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:42:42,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:42:42,563] loop.py:380 INFO TIMING step=14510: step=3.375s data=1.367s compute=3.342s [fwd=2.389s bwd=0.938s opt=0.003s] dispatch=3.502s [2026-10-19 20:42:42,563] loop.py:45 INFO TRAIN_STEP step=14510 loss=1.339964 Training: 81%|████████ | 14510/18018 [31:12:28<7:08:09, 7.32s/step, loss=1.340, acc=0.692, acc_len=1.65, thru=18.6, I=20.7, T=37.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:42:42,945] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:42:48,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:42:48,939] loop.py:45 INFO TRAIN_STEP step=14511 loss=1.183369 Training: 81%|████████ | 14511/18018 [31:12:35<6:51:26, 7.04s/step, loss=1.183, acc=0.718, acc_len=1.74, thru=20.0, I=21.8, T=38.8, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:42:48,935] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14510: forward=2359.8ms backward=892.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:42:49,021] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 20:42:55,825] loop.py:45 INFO TRAIN_STEP step=14512 loss=1.243909 Training: 81%|████████ | 14512/18018 [31:12:41<6:48:38, 6.99s/step, loss=1.244, acc=0.711, acc_len=1.73, thru=15.9, I=18.6, T=37.9, wait=3.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:03,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:43:05,037] loop.py:45 INFO TRAIN_STEP step=14513 loss=1.289605 Training: 81%|████████ | 14513/18018 [31:12:51<7:27:24, 7.66s/step, loss=1.290, acc=0.701, acc_len=1.69, thru=16.5, I=18.5, T=33.3, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:05,064] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=57/64 [2026-10-19 20:43:11,128] loop.py:45 INFO TRAIN_STEP step=14514 loss=1.290503 Training: 81%|████████ | 14514/18018 [31:12:57<6:59:48, 7.19s/step, loss=1.291, acc=0.702, acc_len=1.69, thru=19.2, I=23.6, T=38.6, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:17,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:43:17,823] loop.py:380 INFO TIMING step=14515: step=3.372s data=1.319s compute=3.346s [fwd=2.386s bwd=0.945s opt=0.003s] dispatch=3.314s [2026-10-19 20:43:17,823] loop.py:45 INFO TRAIN_STEP step=14515 loss=1.224490 Training: 81%|████████ | 14515/18018 [31:13:03<6:51:04, 7.04s/step, loss=1.224, acc=0.713, acc_len=1.72, thru=19.2, I=19.3, T=38.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:17,892] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:22,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:43:23,056] loop.py:45 INFO TRAIN_STEP step=14516 loss=1.193221 Training: 81%|████████ | 14516/18018 [31:13:09<6:19:16, 6.50s/step, loss=1.193, acc=0.721, acc_len=1.75, thru=21.5, I=25.8, T=37.8, wait=1.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:43:23,052] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14515: forward=2401.7ms backward=941.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:23,524] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:29,020] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:43:29,695] loop.py:45 INFO TRAIN_STEP step=14517 loss=1.392045 Training: 81%|████████ | 14517/18018 [31:13:15<6:21:39, 6.54s/step, loss=1.392, acc=0.682, acc_len=1.61, thru=19.3, I=25.1, T=39.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:30,025] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:43:35,621] loop.py:45 INFO TRAIN_STEP step=14518 loss=1.261107 Training: 81%|████████ | 14518/18018 [31:13:21<6:10:47, 6.36s/step, loss=1.261, acc=0.706, acc_len=1.70, thru=18.5, I=22.5, T=39.8, wait=2.7s, pool=41, epoch=2/2][2026-10-19 20:43:43,536] loop.py:45 INFO TRAIN_STEP step=14519 loss=1.295529 Training: 81%|████████ | 14519/18018 [31:13:29<6:37:55, 6.82s/step, loss=1.296, acc=0.699, acc_len=1.67, thru=15.8, I=17.4, T=37.3, wait=4.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:49,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:43:51,267] loop.py:380 INFO TIMING step=14520: step=4.150s data=1.382s compute=4.117s [fwd=3.137s bwd=0.965s opt=0.003s] dispatch=3.572s [2026-10-19 20:43:51,268] loop.py:45 INFO TRAIN_STEP step=14520 loss=1.408770 Training: 81%|████████ | 14520/18018 [31:13:37<6:53:42, 7.10s/step, loss=1.409, acc=0.683, acc_len=1.62, thru=16.8, I=20.5, T=30.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:43:51,564] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 20:43:56,890] loop.py:45 INFO TRAIN_STEP step=14521 loss=1.209378 Training: 81%|████████ | 14521/18018 [31:13:42<6:27:49, 6.65s/step, loss=1.209, acc=0.717, acc_len=1.74, thru=18.8, I=29.5, T=38.1, wait=2.3s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:43:56,885] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14520: forward=2346.5ms backward=966.7ms [2026-10-19 20:44:06,148] loop.py:45 INFO TRAIN_STEP step=14522 loss=1.188191 Training: 81%|████████ | 14522/18018 [31:13:52<7:13:12, 7.43s/step, loss=1.188, acc=0.721, acc_len=1.75, thru=12.9, I=13.2, T=36.1, wait=5.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:14,539] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:44:14,842] loop.py:45 INFO TRAIN_STEP step=14523 loss=1.274950 Training: 81%|████████ | 14523/18018 [31:14:00<7:35:06, 7.81s/step, loss=1.275, acc=0.704, acc_len=1.68, thru=16.6, I=16.0, T=33.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:15,040] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:20,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:44:21,620] loop.py:45 INFO TRAIN_STEP step=14524 loss=1.286691 Training: 81%|████████ | 14524/18018 [31:14:07<7:16:54, 7.50s/step, loss=1.287, acc=0.703, acc_len=1.69, thru=18.1, I=20.5, T=36.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:21,696] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 20:44:28,638] loop.py:380 INFO TIMING step=14525: step=3.481s data=1.341s compute=3.444s [fwd=2.410s bwd=1.020s opt=0.003s] dispatch=3.527s [2026-10-19 20:44:28,638] loop.py:45 INFO TRAIN_STEP step=14525 loss=1.240464 Training: 81%|████████ | 14525/18018 [31:14:14<7:08:15, 7.36s/step, loss=1.240, acc=0.708, acc_len=1.71, thru=17.8, I=19.2, T=36.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:28,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:29,077] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:35,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:44:35,777] loop.py:45 INFO TRAIN_STEP step=14526 loss=1.281716 Training: 81%|████████ | 14526/18018 [31:14:21<7:04:23, 7.29s/step, loss=1.282, acc=0.703, acc_len=1.68, thru=19.3, I=18.8, T=38.9, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:44:35,772] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14525: forward=2342.6ms backward=908.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:35,898] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:42,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:44:43,440] loop.py:45 INFO TRAIN_STEP step=14527 loss=1.222680 Training: 81%|████████ | 14527/18018 [31:14:29<7:10:45, 7.40s/step, loss=1.223, acc=0.715, acc_len=1.73, thru=18.1, I=20.4, T=35.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:43,479] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 20:44:49,728] loop.py:45 INFO TRAIN_STEP step=14528 loss=1.263999 Training: 81%|████████ | 14528/18018 [31:14:35<6:51:08, 7.07s/step, loss=1.264, acc=0.707, acc_len=1.71, thru=17.0, I=17.3, T=38.6, wait=3.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:44:59,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:45:00,175] loop.py:45 INFO TRAIN_STEP step=14529 loss=1.267687 Training: 81%|████████ | 14529/18018 [31:14:46<7:49:57, 8.08s/step, loss=1.268, acc=0.706, acc_len=1.70, thru=15.2, I=15.8, T=34.4, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:45:00,442] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:45:05,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:45:05,889] loop.py:380 INFO TIMING step=14530: step=3.380s data=1.391s compute=3.351s [fwd=2.401s bwd=0.937s opt=0.003s] dispatch=2.326s [2026-10-19 20:45:05,889] loop.py:45 INFO TRAIN_STEP step=14530 loss=1.205746 Training: 81%|████████ | 14530/18018 [31:14:51<7:08:31, 7.37s/step, loss=1.206, acc=0.714, acc_len=1.73, thru=21.0, I=23.4, T=37.9, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:45:05,917] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:45:11,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:45:11,900] loop.py:45 INFO TRAIN_STEP step=14531 loss=1.067229 Training: 81%|████████ | 14531/18018 [31:14:57<6:44:39, 6.96s/step, loss=1.067, acc=0.740, acc_len=1.82, thru=21.5, I=23.2, T=38.8, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:45:11,895] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14530: forward=2388.6ms backward=867.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:45:11,927] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:45:19,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:45:19,850] loop.py:45 INFO TRAIN_STEP step=14532 loss=1.235798 Training: 81%|████████ | 14532/18018 [31:15:05<7:01:44, 7.26s/step, loss=1.236, acc=0.708, acc_len=1.70, thru=16.6, I=17.6, T=36.8, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:45:20,159] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:45:26,960] loop.py:45 INFO TRAIN_STEP step=14533 loss=1.300892 Training: 81%|████████ | 14533/18018 [31:15:13<6:59:03, 7.21s/step, loss=1.301, acc=0.701, acc_len=1.69, thru=16.9, I=19.6, T=37.2, wait=3.7s, pool=49, epoch=2/2][2026-10-19 20:45:36,994] loop.py:45 INFO TRAIN_STEP step=14534 loss=1.354550 Training: 81%|████████ | 14534/18018 [31:15:23<7:48:03, 8.06s/step, loss=1.355, acc=0.688, acc_len=1.64, thru=13.1, I=12.0, T=34.3, wait=6.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:45:45,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:45:46,209] loop.py:380 INFO TIMING step=14535: step=3.532s data=1.439s compute=3.499s [fwd=2.511s bwd=0.974s opt=0.003s] dispatch=5.673s [2026-10-19 20:45:46,209] loop.py:45 INFO TRAIN_STEP step=14535 loss=1.308758 Training: 81%|████████ | 14535/18018 [31:15:32<8:08:03, 8.41s/step, loss=1.309, acc=0.698, acc_len=1.67, thru=15.9, I=16.5, T=36.2, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:45:46,249] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 20:45:52,689] loop.py:45 INFO TRAIN_STEP step=14536 loss=1.220844 Training: 81%|████████ | 14536/18018 [31:15:38<7:34:17, 7.83s/step, loss=1.221, acc=0.715, acc_len=1.73, thru=17.6, I=19.8, T=38.9, wait=3.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:45:52,685] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14535: forward=2369.8ms backward=879.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:00,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:46:01,110] loop.py:45 INFO TRAIN_STEP step=14537 loss=1.307926 Training: 81%|████████ | 14537/18018 [31:15:47<7:44:35, 8.01s/step, loss=1.308, acc=0.699, acc_len=1.68, thru=17.7, I=17.1, T=36.3, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:01,338] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:09,725] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:46:09,790] loop.py:45 INFO TRAIN_STEP step=14538 loss=1.180350 Training: 81%|████████ | 14538/18018 [31:15:55<7:56:03, 8.21s/step, loss=1.180, acc=0.723, acc_len=1.77, thru=14.1, I=14.8, T=33.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:10,228] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:46:20,050] loop.py:45 INFO TRAIN_STEP step=14539 loss=1.321899 Training: 81%|████████ | 14539/18018 [31:16:06<8:31:37, 8.82s/step, loss=1.322, acc=0.695, acc_len=1.67, thru=11.4, I=11.2, T=33.6, wait=6.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:26,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:46:27,384] loop.py:380 INFO TIMING step=14540: step=3.534s data=1.406s compute=3.501s [fwd=2.511s bwd=0.975s opt=0.003s] dispatch=3.791s [2026-10-19 20:46:27,385] loop.py:45 INFO TRAIN_STEP step=14540 loss=1.297098 Training: 81%|████████ | 14540/18018 [31:16:13<8:05:35, 8.38s/step, loss=1.297, acc=0.702, acc_len=1.69, thru=18.0, I=19.6, T=36.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:27,578] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:32,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:46:33,386] loop.py:45 INFO TRAIN_STEP step=14541 loss=1.208095 Training: 81%|████████ | 14541/18018 [31:16:19<7:24:10, 7.66s/step, loss=1.208, acc=0.717, acc_len=1.74, thru=20.7, I=26.8, T=38.0, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:46:33,381] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14540: forward=2434.4ms backward=890.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:33,746] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:46:39,899] loop.py:45 INFO TRAIN_STEP step=14542 loss=1.185704 Training: 81%|████████ | 14542/18018 [31:16:25<7:03:58, 7.32s/step, loss=1.186, acc=0.722, acc_len=1.76, thru=18.1, I=21.1, T=36.1, wait=3.0s, pool=49, epoch=2/2][2026-10-19 20:46:47,683] loop.py:45 INFO TRAIN_STEP step=14543 loss=1.256390 Training: 81%|████████ | 14543/18018 [31:16:33<7:11:58, 7.46s/step, loss=1.256, acc=0.708, acc_len=1.71, thru=12.1, I=17.2, T=35.9, wait=4.2s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 20:46:47,977] eagle_store.py:517 WARNING batch_get_buffer missing keys (k465374-bc07672c_hs, k465374-bc07672c_ids, k465374-bc07672c_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:57,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 20:46:47,975] eagle_store.py:517 WARNING batch_get_buffer missing keys (k465373-829942d3_hs, k465373-829942d3_ids, k465373-829942d3_lhs); sleeping 0.50s. [2026-10-19 20:46:58,164] loop.py:45 INFO TRAIN_STEP step=14544 loss=1.362575 Training: 81%|████████ | 14544/18018 [31:16:44<8:04:21, 8.37s/step, loss=1.363, acc=0.689, acc_len=1.64, thru=17.6, I=15.2, T=36.3, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:46:58,656] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:47:05,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:47:05,953] loop.py:380 INFO TIMING step=14545: step=3.467s data=1.392s compute=3.425s [fwd=2.407s bwd=1.003s opt=0.003s] dispatch=4.313s [2026-10-19 20:47:05,953] loop.py:45 INFO TRAIN_STEP step=14545 loss=1.277919 Training: 81%|████████ | 14545/18018 [31:16:52<7:54:11, 8.19s/step, loss=1.278, acc=0.703, acc_len=1.69, thru=16.5, I=20.5, T=36.9, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:47:06,246] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:47:12,147] loop.py:45 INFO TRAIN_STEP step=14546 loss=1.216403 Training: 81%|████████ | 14546/18018 [31:16:58<7:19:23, 7.59s/step, loss=1.216, acc=0.714, acc_len=1.73, thru=18.1, I=20.7, T=40.0, wait=3.0s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:47:12,142] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14545: forward=2284.4ms backward=873.7ms [2026-10-19 20:47:20,255] loop.py:45 INFO TRAIN_STEP step=14547 loss=1.385263 Training: 81%|████████ | 14547/18018 [31:17:06<7:28:10, 7.75s/step, loss=1.385, acc=0.686, acc_len=1.63, thru=16.7, I=16.7, T=37.6, wait=4.7s, pool=57, epoch=2/2][2026-10-19 20:47:29,372] loop.py:45 INFO TRAIN_STEP step=14548 loss=1.261349 Training: 81%|████████ | 14548/18018 [31:17:15<7:51:49, 8.16s/step, loss=1.261, acc=0.704, acc_len=1.69, thru=13.3, I=13.2, T=33.9, wait=5.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:47:34,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:47:35,694] loop.py:45 INFO TRAIN_STEP step=14549 loss=1.227601 Training: 81%|████████ | 14549/18018 [31:17:21<7:19:49, 7.61s/step, loss=1.228, acc=0.714, acc_len=1.72, thru=19.6, I=20.6, T=39.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:47:35,758] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 20:47:42,770] loop.py:380 INFO TIMING step=14550: step=3.424s data=1.323s compute=3.394s [fwd=2.416s bwd=0.963s opt=0.003s] dispatch=3.645s [2026-10-19 20:47:42,770] loop.py:45 INFO TRAIN_STEP step=14550 loss=1.250705 Training: 81%|████████ | 14550/18018 [31:17:28<7:10:30, 7.45s/step, loss=1.251, acc=0.711, acc_len=1.71, thru=17.2, I=18.7, T=37.4, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:47:48,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:47:49,029] loop.py:45 INFO TRAIN_STEP step=14551 loss=1.376384 Training: 81%|████████ | 14551/18018 [31:17:35<6:49:47, 7.09s/step, loss=1.376, acc=0.686, acc_len=1.63, thru=20.7, I=21.5, T=38.1, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:47:49,025] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14550: forward=2391.2ms backward=921.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:47:49,225] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:47:54,648] loop.py:45 INFO TRAIN_STEP step=14552 loss=1.254875 Training: 81%|████████ | 14552/18018 [31:17:40<6:24:09, 6.65s/step, loss=1.255, acc=0.708, acc_len=1.71, thru=19.9, I=22.9, T=40.0, wait=2.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:01,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:48:01,859] loop.py:45 INFO TRAIN_STEP step=14553 loss=1.254577 Training: 81%|████████ | 14553/18018 [31:17:47<6:33:43, 6.82s/step, loss=1.255, acc=0.711, acc_len=1.72, thru=19.1, I=19.8, T=36.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:02,104] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:07,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:48:07,886] loop.py:45 INFO TRAIN_STEP step=14554 loss=1.315939 Training: 81%|████████ | 14554/18018 [31:17:53<6:19:55, 6.58s/step, loss=1.316, acc=0.697, acc_len=1.67, thru=20.6, I=23.0, T=37.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:08,325] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:48:14,863] loop.py:380 INFO TIMING step=14555: step=3.391s data=1.435s compute=3.346s [fwd=2.350s bwd=0.982s opt=0.003s] dispatch=3.577s [2026-10-19 20:48:14,863] loop.py:45 INFO TRAIN_STEP step=14555 loss=1.266233 Training: 81%|████████ | 14555/18018 [31:18:00<6:26:40, 6.70s/step, loss=1.266, acc=0.704, acc_len=1.69, thru=18.4, I=19.2, T=37.7, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:20,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:48:21,437] loop.py:45 INFO TRAIN_STEP step=14556 loss=1.162530 Training: 81%|████████ | 14556/18018 [31:18:07<6:24:22, 6.66s/step, loss=1.163, acc=0.722, acc_len=1.76, thru=19.6, I=21.3, T=38.6, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:48:21,433] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14555: forward=2389.1ms backward=886.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:21,843] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:26,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:48:27,432] loop.py:45 INFO TRAIN_STEP step=14557 loss=1.229985 Training: 81%|████████ | 14557/18018 [31:18:13<6:12:43, 6.46s/step, loss=1.230, acc=0.710, acc_len=1.71, thru=21.1, I=24.1, T=38.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:27,482] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 20:48:34,053] loop.py:45 INFO TRAIN_STEP step=14558 loss=1.360093 Training: 81%|████████ | 14558/18018 [31:18:20<6:15:22, 6.51s/step, loss=1.360, acc=0.685, acc_len=1.63, thru=18.4, I=19.0, T=38.1, wait=3.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:41,351] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:48:42,240] loop.py:45 INFO TRAIN_STEP step=14559 loss=1.217448 Training: 81%|████████ | 14559/18018 [31:18:28<6:44:18, 7.01s/step, loss=1.217, acc=0.714, acc_len=1.73, thru=17.5, I=17.8, T=36.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:42,361] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:47,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:48:47,828] loop.py:380 INFO TIMING step=14560: step=3.287s data=1.378s compute=3.243s [fwd=2.302s bwd=0.927s opt=0.003s] dispatch=2.290s [2026-10-19 20:48:47,828] loop.py:45 INFO TRAIN_STEP step=14560 loss=1.146049 Training: 81%|████████ | 14560/18018 [31:18:33<6:19:30, 6.58s/step, loss=1.146, acc=0.729, acc_len=1.78, thru=21.1, I=24.7, T=38.9, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:48,240] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:54,373] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:48:55,328] loop.py:45 INFO TRAIN_STEP step=14561 loss=1.235704 Training: 81%|████████ | 14561/18018 [31:18:41<6:35:15, 6.86s/step, loss=1.236, acc=0.712, acc_len=1.72, thru=16.9, I=19.0, T=36.2, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:48:55,323] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14560: forward=2507.7ms backward=987.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:48:55,390] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 20:49:03,092] loop.py:45 INFO TRAIN_STEP step=14562 loss=1.198407 Training: 81%|████████ | 14562/18018 [31:18:49<6:50:45, 7.13s/step, loss=1.198, acc=0.719, acc_len=1.74, thru=12.5, I=17.1, T=37.9, wait=4.4s, pool=25, epoch=2/2][2026-10-19 20:49:14,301] loop.py:45 INFO TRAIN_STEP step=14563 loss=1.214248 Training: 81%|████████ | 14563/18018 [31:19:00<8:01:02, 8.35s/step, loss=1.214, acc=0.717, acc_len=1.75, thru=12.5, I=12.8, T=33.5, wait=7.4s, pool=41, epoch=2/2][2026-10-19 20:49:22,949] loop.py:45 INFO TRAIN_STEP step=14564 loss=1.250035 Training: 81%|████████ | 14564/18018 [31:19:09<8:06:00, 8.44s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=15.3, I=14.5, T=34.9, wait=5.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:49:33,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:49:34,010] loop.py:380 INFO TIMING step=14565: step=3.851s data=1.563s compute=3.824s [fwd=2.662s bwd=1.146s opt=0.003s] dispatch=7.202s [2026-10-19 20:49:34,010] loop.py:45 INFO TRAIN_STEP step=14565 loss=1.262942 Training: 81%|████████ | 14565/18018 [31:19:20<8:51:05, 9.23s/step, loss=1.263, acc=0.706, acc_len=1.71, thru=15.9, I=13.0, T=33.2, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:49:34,434] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:49:42,381] loop.py:45 INFO TRAIN_STEP step=14566 loss=1.288346 Training: 81%|████████ | 14566/18018 [31:19:28<8:36:07, 8.97s/step, loss=1.288, acc=0.706, acc_len=1.70, thru=15.0, I=15.5, T=32.4, wait=4.4s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:49:42,377] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14565: forward=2899.0ms backward=1004.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:49:48,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:49:49,768] loop.py:45 INFO TRAIN_STEP step=14567 loss=1.271106 Training: 81%|████████ | 14567/18018 [31:19:35<8:08:39, 8.50s/step, loss=1.271, acc=0.706, acc_len=1.70, thru=18.5, I=20.1, T=33.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:49:50,265] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:49:55,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:49:56,941] loop.py:45 INFO TRAIN_STEP step=14568 loss=1.253429 Training: 81%|████████ | 14568/18018 [31:19:43<7:45:40, 8.10s/step, loss=1.253, acc=0.709, acc_len=1.71, thru=16.6, I=22.0, T=36.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:49:57,411] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:50:03,528] loop.py:45 INFO TRAIN_STEP step=14569 loss=1.181865 Training: 81%|████████ | 14569/18018 [31:19:49<7:19:29, 7.65s/step, loss=1.182, acc=0.723, acc_len=1.76, thru=18.4, I=22.7, T=38.7, wait=3.3s, pool=57, epoch=2/2][2026-10-19 20:50:10,246] loop.py:380 INFO TIMING step=14570: step=3.412s data=1.361s compute=3.367s [fwd=2.357s bwd=0.996s opt=0.003s] dispatch=3.296s [2026-10-19 20:50:10,247] loop.py:45 INFO TRAIN_STEP step=14570 loss=1.231020 Training: 81%|████████ | 14570/18018 [31:19:56<7:03:19, 7.37s/step, loss=1.231, acc=0.712, acc_len=1.72, thru=18.4, I=19.2, T=37.5, wait=3.3s, pool=57, epoch=2/2][2026-10-19 20:50:19,511] loop.py:45 INFO TRAIN_STEP step=14571 loss=1.287460 Training: 81%|████████ | 14571/18018 [31:20:05<7:35:58, 7.94s/step, loss=1.287, acc=0.704, acc_len=1.70, thru=13.9, I=13.3, T=34.8, wait=5.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:50:19,506] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14570: forward=2526.6ms backward=1113.2ms [2026-10-19 20:50:28,870] loop.py:45 INFO TRAIN_STEP step=14572 loss=1.279659 Training: 81%|████████ | 14572/18018 [31:20:14<8:00:17, 8.36s/step, loss=1.280, acc=0.704, acc_len=1.70, thru=15.2, I=14.9, T=33.1, wait=5.5s, pool=61, epoch=2/2][2026-10-19 20:50:36,099] loop.py:45 INFO TRAIN_STEP step=14573 loss=1.254612 Training: 81%|████████ | 14573/18018 [31:20:22<7:40:40, 8.02s/step, loss=1.255, acc=0.710, acc_len=1.71, thru=16.8, I=17.5, T=34.9, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:50:44,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:50:45,364] loop.py:45 INFO TRAIN_STEP step=14574 loss=1.347147 Training: 81%|████████ | 14574/18018 [31:20:31<8:01:54, 8.40s/step, loss=1.347, acc=0.690, acc_len=1.65, thru=15.3, I=15.2, T=35.5, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:50:45,599] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:50:51,502] loop.py:380 INFO TIMING step=14575: step=3.900s data=1.966s compute=3.869s [fwd=2.938s bwd=0.916s opt=0.003s] dispatch=2.228s [2026-10-19 20:50:51,502] loop.py:45 INFO TRAIN_STEP step=14575 loss=1.202117 Training: 81%|████████ | 14575/18018 [31:20:37<7:22:53, 7.72s/step, loss=1.202, acc=0.715, acc_len=1.73, thru=19.6, I=22.5, T=32.8, wait=2.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:50:58,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:50:59,470] loop.py:45 INFO TRAIN_STEP step=14576 loss=1.334658 Training: 81%|████████ | 14576/18018 [31:20:45<7:27:03, 7.79s/step, loss=1.335, acc=0.692, acc_len=1.66, thru=16.4, I=17.9, T=32.9, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:50:59,465] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14575: forward=2865.4ms backward=982.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:50:59,961] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:51:05,455] loop.py:45 INFO TRAIN_STEP step=14577 loss=1.250126 Training: 81%|████████ | 14577/18018 [31:20:51<6:55:48, 7.25s/step, loss=1.250, acc=0.710, acc_len=1.71, thru=18.8, I=23.6, T=38.5, wait=2.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:51:12,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:51:13,526] loop.py:45 INFO TRAIN_STEP step=14578 loss=1.211550 Training: 81%|████████ | 14578/18018 [31:20:59<7:09:48, 7.50s/step, loss=1.212, acc=0.714, acc_len=1.73, thru=16.6, I=18.0, T=37.0, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:51:13,952] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:51:23,604] loop.py:45 INFO TRAIN_STEP step=14579 loss=1.192633 Training: 81%|████████ | 14579/18018 [31:21:09<7:54:04, 8.27s/step, loss=1.193, acc=0.719, acc_len=1.75, thru=11.2, I=12.5, T=30.9, wait=5.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:51:31,274] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:51:32,965] loop.py:380 INFO TIMING step=14580: step=3.746s data=1.512s compute=3.712s [fwd=2.663s bwd=1.034s opt=0.003s] dispatch=5.607s [2026-10-19 20:51:32,965] loop.py:45 INFO TRAIN_STEP step=14580 loss=1.267005 Training: 81%|████████ | 14580/18018 [31:21:19<8:12:40, 8.60s/step, loss=1.267, acc=0.702, acc_len=1.69, thru=15.4, I=16.7, T=34.2, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:51:33,284] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 20:51:40,249] loop.py:45 INFO TRAIN_STEP step=14581 loss=1.344299 Training: 81%|████████ | 14581/18018 [31:21:26<7:49:55, 8.20s/step, loss=1.344, acc=0.690, acc_len=1.64, thru=15.8, I=21.0, T=33.1, wait=3.4s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:51:40,244] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14580: forward=2925.2ms backward=895.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:51:47,500] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:51:47,844] loop.py:45 INFO TRAIN_STEP step=14582 loss=1.263365 Training: 81%|████████ | 14582/18018 [31:21:33<7:39:23, 8.02s/step, loss=1.263, acc=0.708, acc_len=1.71, thru=16.6, I=16.6, T=37.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:51:48,002] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:51:53,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:51:54,141] loop.py:45 INFO TRAIN_STEP step=14583 loss=1.232324 Training: 81%|████████ | 14583/18018 [31:21:40<7:09:37, 7.50s/step, loss=1.232, acc=0.711, acc_len=1.72, thru=20.3, I=23.0, T=37.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:51:54,459] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:52:00,349] loop.py:45 INFO TRAIN_STEP step=14584 loss=1.229943 Training: 81%|████████ | 14584/18018 [31:21:46<6:47:15, 7.12s/step, loss=1.230, acc=0.713, acc_len=1.73, thru=17.7, I=21.9, T=37.4, wait=2.8s, pool=37, epoch=2/2][2026-10-19 20:52:07,997] loop.py:380 INFO TIMING step=14585: step=3.582s data=1.385s compute=3.551s [fwd=2.580s bwd=0.957s opt=0.003s] dispatch=4.055s [2026-10-19 20:52:07,997] loop.py:45 INFO TRAIN_STEP step=14585 loss=1.318150 Training: 81%|████████ | 14585/18018 [31:21:54<6:56:12, 7.27s/step, loss=1.318, acc=0.698, acc_len=1.67, thru=17.2, I=16.4, T=35.7, wait=4.1s, pool=53, epoch=2/2][2026-10-19 20:52:16,643] loop.py:45 INFO TRAIN_STEP step=14586 loss=1.298460 Training: 81%|████████ | 14586/18018 [31:22:02<7:19:39, 7.69s/step, loss=1.298, acc=0.698, acc_len=1.67, thru=16.7, I=15.8, T=34.7, wait=4.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:52:16,638] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14585: forward=2666.7ms backward=985.7ms [2026-10-19 20:52:22,452] loop.py:45 INFO TRAIN_STEP step=14587 loss=1.198477 Training: 81%|████████ | 14587/18018 [31:22:08<6:47:18, 7.12s/step, loss=1.198, acc=0.717, acc_len=1.73, thru=17.4, I=17.1, T=38.5, wait=2.5s, pool=37, epoch=2/2][2026-10-19 20:52:30,751] loop.py:45 INFO TRAIN_STEP step=14588 loss=1.130969 Training: 81%|████████ | 14588/18018 [31:22:16<7:07:23, 7.48s/step, loss=1.131, acc=0.731, acc_len=1.80, thru=13.7, I=14.7, T=36.3, wait=4.8s, pool=29, epoch=2/2][2026-10-19 20:52:40,976] loop.py:45 INFO TRAIN_STEP step=14589 loss=1.246941 Training: 81%|████████ | 14589/18018 [31:22:27<7:54:21, 8.30s/step, loss=1.247, acc=0.712, acc_len=1.71, thru=14.4, I=13.7, T=35.1, wait=6.6s, pool=49, epoch=2/2][2026-10-19 20:52:52,683] loop.py:380 INFO TIMING step=14590: step=3.406s data=1.347s compute=3.370s [fwd=2.417s bwd=0.939s opt=0.003s] dispatch=8.292s [2026-10-19 20:52:52,683] loop.py:45 INFO TRAIN_STEP step=14590 loss=1.268692 Training: 81%|████████ | 14590/18018 [31:22:38<8:52:36, 9.32s/step, loss=1.269, acc=0.704, acc_len=1.69, thru=10.8, I=18.0, T=37.6, wait=8.3s, pool=57, epoch=2/2][2026-10-19 20:53:00,387] loop.py:45 INFO TRAIN_STEP step=14591 loss=1.256407 Training: 81%|████████ | 14591/18018 [31:22:46<8:24:43, 8.84s/step, loss=1.256, acc=0.708, acc_len=1.71, thru=16.8, I=18.9, T=35.8, wait=4.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:53:00,384] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14590: forward=2492.7ms backward=1021.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:53:08,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:53:08,886] loop.py:45 INFO TRAIN_STEP step=14592 loss=1.224611 Training: 81%|████████ | 14592/18018 [31:22:54<8:18:48, 8.74s/step, loss=1.225, acc=0.714, acc_len=1.72, thru=14.9, I=15.7, T=34.9, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:53:09,254] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:53:17,877] loop.py:45 INFO TRAIN_STEP step=14593 loss=1.339977 Training: 81%|████████ | 14593/18018 [31:23:03<8:23:03, 8.81s/step, loss=1.340, acc=0.689, acc_len=1.64, thru=14.0, I=12.6, T=33.8, wait=5.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:53:22,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:53:23,320] loop.py:45 INFO TRAIN_STEP step=14594 loss=1.245395 Training: 81%|████████ | 14594/18018 [31:23:09<7:25:12, 7.80s/step, loss=1.245, acc=0.713, acc_len=1.72, thru=21.4, I=22.5, T=38.8, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:53:23,765] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:53:30,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:53:32,169] loop.py:380 INFO TIMING step=14595: step=3.591s data=1.567s compute=3.563s [fwd=2.533s bwd=1.016s opt=0.003s] dispatch=5.248s [2026-10-19 20:53:32,169] loop.py:45 INFO TRAIN_STEP step=14595 loss=1.239897 Training: 81%|████████ | 14595/18018 [31:23:18<7:42:59, 8.12s/step, loss=1.240, acc=0.710, acc_len=1.71, thru=13.6, I=17.5, T=35.6, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:53:32,200] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:53:37,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:53:37,773] loop.py:45 INFO TRAIN_STEP step=14596 loss=1.196046 Training: 81%|████████ | 14596/18018 [31:23:23<6:59:52, 7.36s/step, loss=1.196, acc=0.717, acc_len=1.74, thru=20.6, I=26.0, T=40.1, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:53:37,769] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14595: forward=2290.7ms backward=867.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:53:38,206] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:53:43,475] loop.py:45 INFO TRAIN_STEP step=14597 loss=1.295591 Training: 81%|████████ | 14597/18018 [31:23:29<6:31:21, 6.86s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=20.4, I=23.7, T=38.8, wait=2.4s, pool=49, epoch=2/2][2026-10-19 20:53:50,916] loop.py:45 INFO TRAIN_STEP step=14598 loss=1.185205 Training: 81%|████████ | 14598/18018 [31:23:36<6:41:06, 7.04s/step, loss=1.185, acc=0.716, acc_len=1.74, thru=16.7, I=16.8, T=37.8, wait=4.0s, pool=49, epoch=2/2][2026-10-19 20:53:58,008] loop.py:45 INFO TRAIN_STEP step=14599 loss=1.205963 Training: 81%|████████ | 14599/18018 [31:23:44<6:41:55, 7.05s/step, loss=1.206, acc=0.713, acc_len=1.73, thru=16.9, I=16.1, T=39.2, wait=3.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:06,971] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:54:08,540] loop.py:380 INFO TIMING step=14600: step=3.723s data=1.648s compute=3.466s [fwd=2.370s bwd=1.082s opt=0.003s] dispatch=6.802s [2026-10-19 20:54:08,541] loop.py:45 INFO TRAIN_STEP step=14600 loss=1.189009 Training: 81%|████████ | 14600/18018 [31:23:54<7:41:18, 8.10s/step, loss=1.189, acc=0.720, acc_len=1.75, thru=13.5, I=16.4, T=34.4, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:08,981] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:15,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:54:15,713] loop.py:45 INFO TRAIN_STEP step=14601 loss=1.187528 Training: 81%|████████ | 14601/18018 [31:24:01<7:25:21, 7.82s/step, loss=1.188, acc=0.719, acc_len=1.74, thru=17.2, I=21.3, T=37.6, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:54:15,709] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14600: forward=2450.8ms backward=919.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:16,039] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:21,493] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:54:22,027] loop.py:45 INFO TRAIN_STEP step=14602 loss=1.327132 Training: 81%|████████ | 14602/18018 [31:24:08<6:59:28, 7.37s/step, loss=1.327, acc=0.692, acc_len=1.65, thru=20.3, I=22.8, T=37.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:22,500] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:54:27,973] loop.py:45 INFO TRAIN_STEP step=14603 loss=1.182758 Training: 81%|████████ | 14603/18018 [31:24:14<6:35:04, 6.94s/step, loss=1.183, acc=0.721, acc_len=1.75, thru=19.6, I=21.0, T=39.0, wait=2.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:34,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:54:35,930] loop.py:45 INFO TRAIN_STEP step=14604 loss=1.208317 Training: 81%|████████ | 14604/18018 [31:24:22<6:52:18, 7.25s/step, loss=1.208, acc=0.714, acc_len=1.73, thru=17.6, I=19.4, T=36.8, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:36,345] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:54:41,631] loop.py:380 INFO TIMING step=14605: step=3.284s data=1.301s compute=3.255s [fwd=2.342s bwd=0.898s opt=0.003s] dispatch=2.407s [2026-10-19 20:54:41,631] loop.py:45 INFO TRAIN_STEP step=14605 loss=1.277505 Training: 81%|████████ | 14605/18018 [31:24:27<6:25:47, 6.78s/step, loss=1.278, acc=0.706, acc_len=1.70, thru=18.1, I=20.8, T=39.0, wait=2.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:48,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:54:48,782] loop.py:45 INFO TRAIN_STEP step=14606 loss=1.263316 Training: 81%|████████ | 14606/18018 [31:24:34<6:31:59, 6.89s/step, loss=1.263, acc=0.707, acc_len=1.70, thru=19.0, I=18.8, T=36.4, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:54:48,778] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14605: forward=2504.5ms backward=965.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:49,246] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:54,571] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:54:55,395] loop.py:45 INFO TRAIN_STEP step=14607 loss=1.363280 Training: 81%|████████ | 14607/18018 [31:24:41<6:27:07, 6.81s/step, loss=1.363, acc=0.687, acc_len=1.63, thru=18.9, I=21.4, T=38.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:54:55,578] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:55:00,613] loop.py:45 INFO TRAIN_STEP step=14608 loss=1.179368 Training: 81%|████████ | 14608/18018 [31:24:46<5:59:50, 6.33s/step, loss=1.179, acc=0.721, acc_len=1.75, thru=20.9, I=25.3, T=39.6, wait=2.0s, pool=49, epoch=2/2][2026-10-19 20:55:12,670] loop.py:45 INFO TRAIN_STEP step=14609 loss=1.314105 Training: 81%|████████ | 14609/18018 [31:24:58<7:37:21, 8.05s/step, loss=1.314, acc=0.699, acc_len=1.68, thru=13.0, I=11.3, T=31.5, wait=8.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:20,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:55:20,952] loop.py:380 INFO TIMING step=14610: step=3.620s data=1.563s compute=3.562s [fwd=2.510s bwd=1.038s opt=0.003s] dispatch=4.651s [2026-10-19 20:55:20,952] loop.py:45 INFO TRAIN_STEP step=14610 loss=1.328884 Training: 81%|████████ | 14610/18018 [31:25:07<7:41:10, 8.12s/step, loss=1.329, acc=0.695, acc_len=1.66, thru=15.1, I=15.7, T=35.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:21,277] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 20:55:28,097] loop.py:45 INFO TRAIN_STEP step=14611 loss=1.217913 (TrainerActor pid=9310) [2026-10-19 20:55:28,093] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14610: forward=2417.7ms backward=958.1ms Training: 81%|████████ | 14611/18018 [31:25:14<7:24:30, 7.83s/step, loss=1.218, acc=0.714, acc_len=1.72, thru=17.5, I=19.9, T=37.4, wait=3.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:36,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:55:36,490] loop.py:45 INFO TRAIN_STEP step=14612 loss=1.253692 Training: 81%|████████ | 14612/18018 [31:25:22<7:33:54, 8.00s/step, loss=1.254, acc=0.707, acc_len=1.71, thru=16.5, I=16.2, T=36.6, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:36,818] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:41,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:55:42,811] loop.py:45 INFO TRAIN_STEP step=14613 loss=1.190438 Training: 81%|████████ | 14613/18018 [31:25:28<7:05:17, 7.49s/step, loss=1.190, acc=0.721, acc_len=1.75, thru=19.9, I=22.1, T=37.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:42,978] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:48,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:55:49,099] loop.py:45 INFO TRAIN_STEP step=14614 loss=1.267041 Training: 81%|████████ | 14614/18018 [31:25:35<6:44:36, 7.13s/step, loss=1.267, acc=0.706, acc_len=1.71, thru=20.1, I=24.5, T=39.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:49,163] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:56,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:55:56,932] loop.py:380 INFO TIMING step=14615: step=3.629s data=1.539s compute=3.579s [fwd=2.568s bwd=0.997s opt=0.003s] dispatch=4.195s [2026-10-19 20:55:56,932] loop.py:45 INFO TRAIN_STEP step=14615 loss=1.249405 Training: 81%|████████ | 14615/18018 [31:25:43<6:56:26, 7.34s/step, loss=1.249, acc=0.709, acc_len=1.72, thru=15.6, I=18.1, T=35.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:55:57,105] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:56:05,404] loop.py:45 INFO TRAIN_STEP step=14616 loss=1.136914 Training: 81%|████████ | 14616/18018 [31:25:51<7:15:30, 7.68s/step, loss=1.137, acc=0.727, acc_len=1.78, thru=12.9, I=13.8, T=35.8, wait=4.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:56:05,399] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14615: forward=2533.3ms backward=992.3ms [2026-10-19 20:56:14,173] loop.py:45 INFO TRAIN_STEP step=14617 loss=1.212759 Training: 81%|████████ | 14617/18018 [31:26:00<7:33:51, 8.01s/step, loss=1.213, acc=0.715, acc_len=1.74, thru=16.9, I=15.8, T=35.4, wait=5.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:56:19,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:56:19,741] loop.py:45 INFO TRAIN_STEP step=14618 loss=1.218219 Training: 81%|████████ | 14618/18018 [31:26:05<6:52:18, 7.28s/step, loss=1.218, acc=0.714, acc_len=1.73, thru=21.7, I=22.6, T=38.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:56:19,813] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 20:56:27,717] loop.py:45 INFO TRAIN_STEP step=14619 loss=1.341029 Training: 81%|████████ | 14619/18018 [31:26:13<7:04:03, 7.49s/step, loss=1.341, acc=0.690, acc_len=1.65, thru=15.0, I=16.2, T=36.0, wait=4.4s, pool=45, epoch=2/2][2026-10-19 20:56:34,632] loop.py:380 INFO TIMING step=14620: step=3.399s data=1.388s compute=3.353s [fwd=2.388s bwd=0.951s opt=0.003s] dispatch=3.507s [2026-10-19 20:56:34,632] loop.py:45 INFO TRAIN_STEP step=14620 loss=1.236120 Training: 81%|████████ | 14620/18018 [31:26:20<6:54:14, 7.31s/step, loss=1.236, acc=0.711, acc_len=1.72, thru=18.1, I=17.4, T=37.7, wait=3.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:56:39,498] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:56:40,348] loop.py:45 INFO TRAIN_STEP step=14621 loss=1.341001 Training: 81%|████████ | 14621/18018 [31:26:26<6:27:00, 6.84s/step, loss=1.341, acc=0.693, acc_len=1.66, thru=21.6, I=23.1, T=37.8, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:56:40,342] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14620: forward=2358.1ms backward=970.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:56:40,506] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:56:46,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:56:46,865] loop.py:45 INFO TRAIN_STEP step=14622 loss=1.260133 Training: 81%|████████ | 14622/18018 [31:26:32<6:21:26, 6.74s/step, loss=1.260, acc=0.705, acc_len=1.69, thru=19.7, I=22.0, T=36.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:56:47,307] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:56:53,388] loop.py:45 INFO TRAIN_STEP step=14623 loss=1.258786 Training: 81%|████████ | 14623/18018 [31:26:39<6:17:42, 6.68s/step, loss=1.259, acc=0.709, acc_len=1.71, thru=18.6, I=19.4, T=38.2, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:00,963] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:57:01,389] loop.py:45 INFO TRAIN_STEP step=14624 loss=1.281656 Training: 81%|████████ | 14624/18018 [31:26:47<6:40:04, 7.07s/step, loss=1.282, acc=0.704, acc_len=1.69, thru=17.4, I=18.0, T=36.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:01,479] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:08,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:57:08,666] loop.py:380 INFO TIMING step=14625: step=3.511s data=1.382s compute=3.476s [fwd=2.493s bwd=0.969s opt=0.003s] dispatch=3.757s [2026-10-19 20:57:08,666] loop.py:45 INFO TRAIN_STEP step=14625 loss=1.127658 Training: 81%|████████ | 14625/18018 [31:26:54<6:43:25, 7.13s/step, loss=1.128, acc=0.730, acc_len=1.79, thru=17.2, I=19.2, T=36.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:09,127] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:57:15,990] loop.py:45 INFO TRAIN_STEP step=14626 loss=1.312256 Training: 81%|████████ | 14626/18018 [31:27:02<6:46:29, 7.19s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=16.0, I=16.9, T=36.3, wait=3.8s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:57:15,985] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14625: forward=2486.9ms backward=998.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:20,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:57:21,680] loop.py:45 INFO TRAIN_STEP step=14627 loss=1.245879 Training: 81%|████████ | 14627/18018 [31:27:07<6:20:57, 6.74s/step, loss=1.246, acc=0.705, acc_len=1.69, thru=21.2, I=23.1, T=39.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:22,145] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:57:27,545] loop.py:45 INFO TRAIN_STEP step=14628 loss=1.293627 Training: 81%|████████ | 14628/18018 [31:27:13<6:05:59, 6.48s/step, loss=1.294, acc=0.702, acc_len=1.68, thru=18.2, I=25.1, T=38.8, wait=2.6s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:35,333] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:57:36,722] loop.py:45 INFO TRAIN_STEP step=14629 loss=1.418285 Training: 81%|████████ | 14629/18018 [31:27:22<6:51:36, 7.29s/step, loss=1.418, acc=0.680, acc_len=1.61, thru=17.6, I=17.0, T=36.5, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:36,851] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 20:57:43,493] loop.py:380 INFO TIMING step=14630: step=3.321s data=1.344s compute=3.293s [fwd=2.356s bwd=0.923s opt=0.003s] dispatch=3.443s [2026-10-19 20:57:43,494] loop.py:45 INFO TRAIN_STEP step=14630 loss=1.203691 Training: 81%|████████ | 14630/18018 [31:27:29<6:42:45, 7.13s/step, loss=1.204, acc=0.721, acc_len=1.75, thru=17.9, I=19.4, T=38.5, wait=3.4s, pool=57, epoch=2/2][2026-10-19 20:57:50,382] loop.py:45 INFO TRAIN_STEP step=14631 loss=1.316774 Training: 81%|████████ | 14631/18018 [31:27:36<6:38:30, 7.06s/step, loss=1.317, acc=0.695, acc_len=1.67, thru=17.5, I=18.3, T=34.7, wait=3.2s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:57:50,377] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14630: forward=2592.9ms backward=1048.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:58,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:57:59,160] loop.py:45 INFO TRAIN_STEP step=14632 loss=1.299494 Training: 81%|████████ | 14632/18018 [31:27:45<7:07:29, 7.58s/step, loss=1.299, acc=0.697, acc_len=1.67, thru=17.1, I=15.8, T=36.4, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:57:59,221] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:58:07,418] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:58:08,471] loop.py:45 INFO TRAIN_STEP step=14633 loss=1.306578 Training: 81%|████████ | 14633/18018 [31:27:54<7:36:46, 8.10s/step, loss=1.307, acc=0.698, acc_len=1.67, thru=13.5, I=14.8, T=35.9, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:58:08,927] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 20:58:14,605] loop.py:45 INFO TRAIN_STEP step=14634 loss=1.223675 Training: 81%|████████ | 14634/18018 [31:28:00<7:03:23, 7.51s/step, loss=1.224, acc=0.712, acc_len=1.72, thru=17.7, I=20.2, T=40.0, wait=2.9s, pool=41, epoch=2/2][2026-10-19 20:58:23,847] loop.py:380 INFO TIMING step=14635: step=3.528s data=1.466s compute=3.490s [fwd=2.509s bwd=0.967s opt=0.003s] dispatch=5.706s [2026-10-19 20:58:23,847] loop.py:45 INFO TRAIN_STEP step=14635 loss=1.357675 Training: 81%|████████ | 14635/18018 [31:28:09<7:32:37, 8.03s/step, loss=1.358, acc=0.691, acc_len=1.65, thru=15.8, I=14.3, T=36.3, wait=5.7s, pool=49, epoch=2/2][2026-10-19 20:58:34,248] loop.py:45 INFO TRAIN_STEP step=14636 loss=1.360390 Training: 81%|████████ | 14636/18018 [31:28:20<8:12:36, 8.74s/step, loss=1.360, acc=0.692, acc_len=1.65, thru=13.6, I=13.4, T=33.0, wait=6.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:58:34,244] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14635: forward=2867.0ms backward=964.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:58:41,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:58:41,689] loop.py:45 INFO TRAIN_STEP step=14637 loss=1.280765 Training: 81%|████████ | 14637/18018 [31:28:27<7:50:31, 8.35s/step, loss=1.281, acc=0.703, acc_len=1.68, thru=16.9, I=18.6, T=36.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:58:42,101] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:58:50,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:58:50,875] loop.py:45 INFO TRAIN_STEP step=14638 loss=1.325562 Training: 81%|████████ | 14638/18018 [31:28:36<8:04:32, 8.60s/step, loss=1.326, acc=0.698, acc_len=1.67, thru=13.8, I=14.6, T=34.4, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:58:50,938] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 20:58:57,666] loop.py:45 INFO TRAIN_STEP step=14639 loss=1.287420 Training: 81%|████████ | 14639/18018 [31:28:43<7:33:47, 8.06s/step, loss=1.287, acc=0.704, acc_len=1.70, thru=16.8, I=18.2, T=37.4, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:03,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:59:04,985] loop.py:380 INFO TIMING step=14640: step=3.458s data=1.407s compute=3.430s [fwd=2.454s bwd=0.962s opt=0.003s] dispatch=3.853s [2026-10-19 20:59:04,986] loop.py:45 INFO TRAIN_STEP step=14640 loss=1.118541 Training: 81%|████████▏ | 14640/18018 [31:28:51<7:21:12, 7.84s/step, loss=1.119, acc=0.731, acc_len=1.79, thru=18.5, I=19.8, T=37.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:05,390] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:09,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:59:10,875] loop.py:45 INFO TRAIN_STEP step=14641 loss=1.187681 Training: 81%|████████▏ | 14641/18018 [31:28:56<6:48:12, 7.25s/step, loss=1.188, acc=0.719, acc_len=1.75, thru=21.2, I=28.3, T=39.5, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:59:10,871] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14640: forward=2316.6ms backward=884.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:11,273] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:17,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:59:17,462] loop.py:45 INFO TRAIN_STEP step=14642 loss=1.289459 Training: 81%|████████▏ | 14642/18018 [31:29:03<6:36:48, 7.05s/step, loss=1.289, acc=0.704, acc_len=1.69, thru=20.1, I=24.8, T=37.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:17,522] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:25,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:59:25,405] loop.py:45 INFO TRAIN_STEP step=14643 loss=1.273796 Training: 81%|████████▏ | 14643/18018 [31:29:11<6:51:44, 7.32s/step, loss=1.274, acc=0.703, acc_len=1.69, thru=16.2, I=16.4, T=34.9, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:25,808] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 20:59:34,349] loop.py:45 INFO TRAIN_STEP step=14644 loss=1.248222 Training: 81%|████████▏ | 14644/18018 [31:29:20<7:18:59, 7.81s/step, loss=1.248, acc=0.711, acc_len=1.72, thru=14.9, I=14.9, T=36.2, wait=5.4s, pool=61, epoch=2/2][2026-10-19 20:59:39,961] loop.py:380 INFO TIMING step=14645: step=3.398s data=1.426s compute=3.360s [fwd=2.363s bwd=0.981s opt=0.003s] dispatch=2.207s [2026-10-19 20:59:39,961] loop.py:45 INFO TRAIN_STEP step=14645 loss=1.322513 Training: 81%|████████▏ | 14645/18018 [31:29:26<6:41:51, 7.15s/step, loss=1.323, acc=0.696, acc_len=1.66, thru=19.2, I=18.2, T=37.7, wait=2.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:47,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:59:48,493] loop.py:45 INFO TRAIN_STEP step=14646 loss=1.269601 Training: 81%|████████▏ | 14646/18018 [31:29:34<7:05:05, 7.56s/step, loss=1.270, acc=0.704, acc_len=1.69, thru=17.0, I=17.1, T=36.8, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 20:59:48,488] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14645: forward=2443.7ms backward=980.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:48,862] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:54,506] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 20:59:55,624] loop.py:45 INFO TRAIN_STEP step=14647 loss=1.257188 Training: 81%|████████▏ | 14647/18018 [31:29:41<6:57:38, 7.43s/step, loss=1.257, acc=0.706, acc_len=1.70, thru=19.0, I=22.3, T=37.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 20:59:56,016] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:00:01,691] loop.py:45 INFO TRAIN_STEP step=14648 loss=1.234507 Training: 81%|████████▏ | 14648/18018 [31:29:47<6:34:29, 7.02s/step, loss=1.235, acc=0.709, acc_len=1.71, thru=18.2, I=23.2, T=40.1, wait=2.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:09,107] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:00:09,522] loop.py:45 INFO TRAIN_STEP step=14649 loss=1.202061 Training: 81%|████████▏ | 14649/18018 [31:29:55<6:48:01, 7.27s/step, loss=1.202, acc=0.719, acc_len=1.75, thru=18.1, I=18.0, T=37.5, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:09,611] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:16,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:00:17,235] loop.py:380 INFO TIMING step=14650: step=3.455s data=1.382s compute=3.424s [fwd=2.460s bwd=0.950s opt=0.003s] dispatch=4.248s [2026-10-19 21:00:17,236] loop.py:45 INFO TRAIN_STEP step=14650 loss=1.187491 Training: 81%|████████▏ | 14650/18018 [31:30:03<6:55:24, 7.40s/step, loss=1.187, acc=0.722, acc_len=1.75, thru=16.9, I=19.4, T=37.0, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:17,738] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:23,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:00:23,460] loop.py:45 INFO TRAIN_STEP step=14651 loss=1.368273 Training: 81%|████████▏ | 14651/18018 [31:30:09<6:35:29, 7.05s/step, loss=1.368, acc=0.687, acc_len=1.64, thru=20.2, I=22.0, T=38.0, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:00:23,456] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14650: forward=2367.4ms backward=961.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:23,770] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:00:30,317] loop.py:45 INFO TRAIN_STEP step=14652 loss=1.290473 Training: 81%|████████▏ | 14652/18018 [31:30:16<6:32:08, 6.99s/step, loss=1.290, acc=0.700, acc_len=1.68, thru=16.7, I=19.4, T=36.8, wait=3.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:38,097] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:00:38,943] loop.py:45 INFO TRAIN_STEP step=14653 loss=1.311909 Training: 81%|████████▏ | 14653/18018 [31:30:25<6:59:33, 7.48s/step, loss=1.312, acc=0.700, acc_len=1.68, thru=16.8, I=16.4, T=35.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:39,101] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:47,275] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:00:47,733] loop.py:45 INFO TRAIN_STEP step=14654 loss=1.172735 Training: 81%|████████▏ | 14654/18018 [31:30:33<7:21:29, 7.87s/step, loss=1.173, acc=0.728, acc_len=1.78, thru=13.8, I=16.2, T=35.3, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:47,785] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 21:00:55,740] loop.py:380 INFO TIMING step=14655: step=3.575s data=1.483s compute=3.541s [fwd=2.528s bwd=0.998s opt=0.003s] dispatch=4.421s [2026-10-19 21:00:55,740] loop.py:45 INFO TRAIN_STEP step=14655 loss=1.288522 Training: 81%|████████▏ | 14655/18018 [31:30:41<7:23:33, 7.91s/step, loss=1.289, acc=0.701, acc_len=1.68, thru=17.0, I=16.4, T=35.8, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:55,719] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:00:56,220] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:01,412] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:01:02,091] loop.py:45 INFO TRAIN_STEP step=14656 loss=1.169917 Training: 81%|████████▏ | 14656/18018 [31:30:48<6:57:08, 7.44s/step, loss=1.170, acc=0.722, acc_len=1.76, thru=19.7, I=18.9, T=38.6, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:01:02,087] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14655: forward=2357.4ms backward=911.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:02,416] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:08,260] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:01:09,586] loop.py:45 INFO TRAIN_STEP step=14657 loss=1.347932 Training: 81%|████████▏ | 14657/18018 [31:30:55<6:57:54, 7.46s/step, loss=1.348, acc=0.695, acc_len=1.66, thru=18.4, I=22.5, T=37.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:09,769] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:01:16,260] loop.py:45 INFO TRAIN_STEP step=14658 loss=1.259953 Training: 81%|████████▏ | 14658/18018 [31:31:02<6:44:36, 7.23s/step, loss=1.260, acc=0.709, acc_len=1.71, thru=18.9, I=21.0, T=38.4, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:25,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:01:25,971] loop.py:45 INFO TRAIN_STEP step=14659 loss=1.217016 Training: 81%|████████▏ | 14659/18018 [31:31:12<7:26:10, 7.97s/step, loss=1.217, acc=0.711, acc_len=1.72, thru=16.6, I=13.5, T=34.6, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:26,220] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:31,011] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:01:31,801] loop.py:380 INFO TIMING step=14660: step=3.373s data=1.381s compute=3.337s [fwd=2.385s bwd=0.938s opt=0.003s] dispatch=2.447s [2026-10-19 21:01:31,802] loop.py:45 INFO TRAIN_STEP step=14660 loss=1.318996 Training: 81%|████████▏ | 14660/18018 [31:31:17<6:50:08, 7.33s/step, loss=1.319, acc=0.698, acc_len=1.67, thru=20.9, I=23.0, T=37.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:32,017] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:36,991] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:01:37,989] loop.py:45 INFO TRAIN_STEP step=14661 loss=1.233500 Training: 81%|████████▏ | 14661/18018 [31:31:24<6:30:52, 6.99s/step, loss=1.234, acc=0.712, acc_len=1.72, thru=20.9, I=25.3, T=39.3, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:01:37,985] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14660: forward=2316.1ms backward=897.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:38,025] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:43,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:01:43,514] loop.py:45 INFO TRAIN_STEP step=14662 loss=1.153944 Training: 81%|████████▏ | 14662/18018 [31:31:29<6:06:14, 6.55s/step, loss=1.154, acc=0.724, acc_len=1.76, thru=22.2, I=25.5, T=39.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:43,695] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:49,403] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:01:49,862] loop.py:45 INFO TRAIN_STEP step=14663 loss=1.222595 Training: 81%|████████▏ | 14663/18018 [31:31:35<6:02:46, 6.49s/step, loss=1.223, acc=0.712, acc_len=1.72, thru=21.4, I=22.3, T=38.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:49,917] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:54,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:01:55,495] loop.py:45 INFO TRAIN_STEP step=14664 loss=1.220530 Training: 81%|████████▏ | 14664/18018 [31:31:41<5:48:20, 6.23s/step, loss=1.221, acc=0.711, acc_len=1.72, thru=22.6, I=25.5, T=35.4, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:01:55,743] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:02,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:02:02,455] loop.py:380 INFO TIMING step=14665: step=3.426s data=1.350s compute=3.397s [fwd=2.431s bwd=0.952s opt=0.003s] dispatch=3.522s [2026-10-19 21:02:02,455] loop.py:45 INFO TRAIN_STEP step=14665 loss=1.283121 Training: 81%|████████▏ | 14665/18018 [31:31:48<6:00:24, 6.45s/step, loss=1.283, acc=0.703, acc_len=1.69, thru=18.0, I=20.7, T=37.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:02,546] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 21:02:08,380] loop.py:45 INFO TRAIN_STEP step=14666 loss=1.266371 Training: 81%|████████▏ | 14666/18018 [31:31:54<5:51:32, 6.29s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=19.8, I=21.4, T=38.7, wait=2.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:02:08,376] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14665: forward=2369.6ms backward=883.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:17,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:02:17,883] loop.py:45 INFO TRAIN_STEP step=14667 loss=1.311348 Training: 81%|████████▏ | 14667/18018 [31:32:03<6:45:11, 7.26s/step, loss=1.311, acc=0.700, acc_len=1.68, thru=15.1, I=15.8, T=36.2, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:18,218] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:25,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:02:26,660] loop.py:45 INFO TRAIN_STEP step=14668 loss=1.243324 Training: 81%|████████▏ | 14668/18018 [31:32:12<7:10:34, 7.71s/step, loss=1.243, acc=0.710, acc_len=1.71, thru=13.8, I=16.6, T=37.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:26,949] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:32,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:02:33,535] loop.py:45 INFO TRAIN_STEP step=14669 loss=1.307044 Training: 81%|████████▏ | 14669/18018 [31:32:19<6:56:27, 7.46s/step, loss=1.307, acc=0.695, acc_len=1.66, thru=18.9, I=21.0, T=36.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:33,960] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:02:41,792] loop.py:380 INFO TIMING step=14670: step=3.498s data=1.464s compute=3.461s [fwd=2.464s bwd=0.982s opt=0.003s] dispatch=4.748s [2026-10-19 21:02:41,792] loop.py:45 INFO TRAIN_STEP step=14670 loss=1.219088 Training: 81%|████████▏ | 14670/18018 [31:32:27<7:09:39, 7.70s/step, loss=1.219, acc=0.712, acc_len=1.71, thru=13.2, I=17.2, T=36.6, wait=4.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:51,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:02:52,551] loop.py:45 INFO TRAIN_STEP step=14671 loss=1.337381 Training: 81%|████████▏ | 14671/18018 [31:32:38<8:00:41, 8.62s/step, loss=1.337, acc=0.693, acc_len=1.66, thru=13.8, I=14.6, T=33.8, wait=7.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:02:52,547] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14670: forward=2669.9ms backward=1066.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:02:52,954] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:02:59,208] loop.py:45 INFO TRAIN_STEP step=14672 loss=1.333869 Training: 81%|████████▏ | 14672/18018 [31:32:45<7:27:44, 8.03s/step, loss=1.334, acc=0.692, acc_len=1.65, thru=17.2, I=20.0, T=39.2, wait=3.4s, pool=49, epoch=2/2][2026-10-19 21:03:06,366] loop.py:45 INFO TRAIN_STEP step=14673 loss=1.116599 Training: 81%|████████▏ | 14673/18018 [31:32:52<7:13:03, 7.77s/step, loss=1.117, acc=0.734, acc_len=1.80, thru=16.3, I=15.9, T=38.1, wait=3.8s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:17,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:03:18,150] loop.py:45 INFO TRAIN_STEP step=14674 loss=1.366926 Training: 81%|████████▏ | 14674/18018 [31:33:04<8:20:04, 8.97s/step, loss=1.367, acc=0.689, acc_len=1.64, thru=12.5, I=12.3, T=34.6, wait=8.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:18,298] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:23,698] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:03:24,405] loop.py:380 INFO TIMING step=14675: step=3.442s data=1.486s compute=3.403s [fwd=2.458s bwd=0.930s opt=0.003s] dispatch=2.806s [2026-10-19 21:03:24,406] loop.py:45 INFO TRAIN_STEP step=14675 loss=1.246484 Training: 81%|████████▏ | 14675/18018 [31:33:10<7:34:30, 8.16s/step, loss=1.246, acc=0.710, acc_len=1.72, thru=19.6, I=18.9, T=37.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:24,702] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:32,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:03:32,334] loop.py:45 INFO TRAIN_STEP step=14676 loss=1.201908 Training: 81%|████████▏ | 14676/18018 [31:33:18<7:30:34, 8.09s/step, loss=1.202, acc=0.717, acc_len=1.74, thru=17.1, I=19.2, T=35.9, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:03:32,330] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14675: forward=2519.0ms backward=999.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:32,581] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:40,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:03:41,092] loop.py:45 INFO TRAIN_STEP step=14677 loss=1.269505 Training: 81%|████████▏ | 14677/18018 [31:33:27<7:41:36, 8.29s/step, loss=1.270, acc=0.707, acc_len=1.71, thru=15.5, I=16.2, T=37.6, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:41,209] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:47,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:03:47,987] loop.py:45 INFO TRAIN_STEP step=14678 loss=1.315314 Training: 81%|████████▏ | 14678/18018 [31:33:34<7:18:09, 7.87s/step, loss=1.315, acc=0.692, acc_len=1.65, thru=19.2, I=19.1, T=36.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:48,319] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:53,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:03:54,833] loop.py:45 INFO TRAIN_STEP step=14679 loss=1.301798 Training: 81%|████████▏ | 14679/18018 [31:33:40<7:00:56, 7.56s/step, loss=1.302, acc=0.699, acc_len=1.68, thru=19.5, I=22.0, T=39.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:03:55,168] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:04:03,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:04:03,298] loop.py:380 INFO TIMING step=14680: step=3.329s data=1.326s compute=3.305s [fwd=2.363s bwd=0.927s opt=0.003s] dispatch=5.126s [2026-10-19 21:04:03,299] loop.py:45 INFO TRAIN_STEP step=14680 loss=1.204651 Training: 81%|████████▏ | 14680/18018 [31:33:49<7:15:51, 7.83s/step, loss=1.205, acc=0.719, acc_len=1.75, thru=14.8, I=18.5, T=38.5, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:04:03,627] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:04:10,146] loop.py:45 INFO TRAIN_STEP step=14681 loss=1.120345 Training: 81%|████████▏ | 14681/18018 [31:33:56<6:59:15, 7.54s/step, loss=1.120, acc=0.731, acc_len=1.79, thru=16.0, I=17.1, T=38.3, wait=3.5s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:04:10,143] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14680: forward=2373.0ms backward=923.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:04:20,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:04:21,203] loop.py:45 INFO TRAIN_STEP step=14682 loss=1.189060 Training: 81%|████████▏ | 14682/18018 [31:34:07<7:57:49, 8.59s/step, loss=1.189, acc=0.719, acc_len=1.75, thru=15.4, I=16.2, T=35.5, wait=7.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:04:21,516] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:04:26,352] loop.py:45 INFO TRAIN_STEP step=14683 loss=1.183519 Training: 81%|████████▏ | 14683/18018 [31:34:12<7:00:12, 7.56s/step, loss=1.184, acc=0.720, acc_len=1.75, thru=20.4, I=24.0, T=39.6, wait=1.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:04:33,614] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:04:33,743] loop.py:45 INFO TRAIN_STEP step=14684 loss=1.248278 Training: 81%|████████▏ | 14684/18018 [31:34:19<6:57:16, 7.51s/step, loss=1.248, acc=0.708, acc_len=1.71, thru=17.5, I=17.8, T=36.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:04:34,115] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:04:40,020] loop.py:380 INFO TIMING step=14685: step=3.321s data=1.332s compute=3.283s [fwd=2.374s bwd=0.894s opt=0.003s] dispatch=2.947s [2026-10-19 21:04:40,020] loop.py:45 INFO TRAIN_STEP step=14685 loss=1.322800 Training: 82%|████████▏ | 14685/18018 [31:34:26<6:36:38, 7.14s/step, loss=1.323, acc=0.695, acc_len=1.66, thru=20.1, I=22.5, T=38.5, wait=2.9s, pool=61, epoch=2/2][2026-10-19 21:04:47,587] loop.py:45 INFO TRAIN_STEP step=14686 loss=1.362902 Training: 82%|████████▏ | 14686/18018 [31:34:33<6:43:36, 7.27s/step, loss=1.363, acc=0.688, acc_len=1.63, thru=16.6, I=16.5, T=36.9, wait=4.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:04:47,582] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14685: forward=2457.9ms backward=953.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:04:54,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:04:55,531] loop.py:45 INFO TRAIN_STEP step=14687 loss=1.361572 Training: 82%|████████▏ | 14687/18018 [31:34:41<6:54:44, 7.47s/step, loss=1.362, acc=0.689, acc_len=1.64, thru=18.5, I=18.2, T=35.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:04:55,693] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:05:01,492] loop.py:45 INFO TRAIN_STEP step=14688 loss=1.266910 Training: 82%|████████▏ | 14688/18018 [31:34:47<6:29:30, 7.02s/step, loss=1.267, acc=0.706, acc_len=1.70, thru=18.8, I=21.9, T=40.0, wait=2.7s, pool=45, epoch=2/2][2026-10-19 21:05:08,569] loop.py:45 INFO TRAIN_STEP step=14689 loss=1.275266 Training: 82%|████████▏ | 14689/18018 [31:34:54<6:30:18, 7.03s/step, loss=1.275, acc=0.702, acc_len=1.69, thru=17.0, I=18.2, T=37.3, wait=3.6s, pool=45, epoch=2/2][2026-10-19 21:05:17,458] loop.py:380 INFO TIMING step=14690: step=3.687s data=1.440s compute=3.632s [fwd=2.626s bwd=0.990s opt=0.003s] dispatch=5.194s [2026-10-19 21:05:17,458] loop.py:45 INFO TRAIN_STEP step=14690 loss=1.210832 Training: 82%|████████▏ | 14690/18018 [31:35:03<7:01:04, 7.59s/step, loss=1.211, acc=0.718, acc_len=1.74, thru=15.5, I=14.3, T=34.7, wait=5.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:05:25,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:05:26,225] loop.py:45 INFO TRAIN_STEP step=14691 loss=1.272466 Training: 82%|████████▏ | 14691/18018 [31:35:12<7:20:30, 7.94s/step, loss=1.272, acc=0.703, acc_len=1.68, thru=16.5, I=17.4, T=36.2, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:05:26,220] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14690: forward=2489.7ms backward=1001.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:05:26,437] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:05:33,818] loop.py:45 INFO TRAIN_STEP step=14692 loss=1.219250 Training: 82%|████████▏ | 14692/18018 [31:35:19<7:14:33, 7.84s/step, loss=1.219, acc=0.716, acc_len=1.73, thru=15.6, I=19.2, T=37.4, wait=4.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:05:40,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:05:40,991] loop.py:45 INFO TRAIN_STEP step=14693 loss=1.412727 Training: 82%|████████▏ | 14693/18018 [31:35:27<7:03:20, 7.64s/step, loss=1.413, acc=0.680, acc_len=1.61, thru=18.7, I=19.6, T=37.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:05:41,319] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:05:45,927] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:05:46,533] loop.py:45 INFO TRAIN_STEP step=14694 loss=1.217360 Training: 82%|████████▏ | 14694/18018 [31:35:32<6:28:20, 7.01s/step, loss=1.217, acc=0.716, acc_len=1.73, thru=21.8, I=26.5, T=39.5, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:05:46,931] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:05:53,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:05:54,387] loop.py:380 INFO TIMING step=14695: step=3.394s data=1.372s compute=3.365s [fwd=2.400s bwd=0.951s opt=0.003s] dispatch=4.450s [2026-10-19 21:05:54,387] loop.py:45 INFO TRAIN_STEP step=14695 loss=1.128217 Training: 82%|████████▏ | 14695/18018 [31:35:40<6:42:15, 7.26s/step, loss=1.128, acc=0.731, acc_len=1.79, thru=17.1, I=19.3, T=37.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:05:54,445] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:00,537] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:06:00,682] loop.py:45 INFO TRAIN_STEP step=14696 loss=1.260214 Training: 82%|████████▏ | 14696/18018 [31:35:46<6:26:02, 6.97s/step, loss=1.260, acc=0.705, acc_len=1.69, thru=20.1, I=19.3, T=38.2, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:06:00,677] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14695: forward=2340.0ms backward=953.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:01,039] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:11,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:06:11,963] loop.py:45 INFO TRAIN_STEP step=14697 loss=1.369069 Training: 82%|████████▏ | 14697/18018 [31:35:58<7:37:28, 8.27s/step, loss=1.369, acc=0.690, acc_len=1.65, thru=11.8, I=11.5, T=33.6, wait=7.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:12,319] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:18,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:06:18,927] loop.py:45 INFO TRAIN_STEP step=14698 loss=1.124248 Training: 82%|████████▏ | 14698/18018 [31:36:05<7:15:43, 7.87s/step, loss=1.124, acc=0.731, acc_len=1.79, thru=18.6, I=17.1, T=38.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:19,294] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:06:26,073] loop.py:45 INFO TRAIN_STEP step=14699 loss=1.196779 Training: 82%|████████▏ | 14699/18018 [31:36:12<7:03:32, 7.66s/step, loss=1.197, acc=0.724, acc_len=1.76, thru=19.0, I=20.0, T=38.1, wait=3.8s, pool=61, epoch=2/2][2026-10-19 21:06:32,287] loop.py:380 INFO TIMING step=14700: step=3.338s data=1.338s compute=3.304s [fwd=2.381s bwd=0.909s opt=0.003s] dispatch=2.866s [2026-10-19 21:06:32,287] loop.py:45 INFO TRAIN_STEP step=14700 loss=1.320053 Training: 82%|████████▏ | 14700/18018 [31:36:18<6:39:26, 7.22s/step, loss=1.320, acc=0.698, acc_len=1.67, thru=20.4, I=20.3, T=38.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:32,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:32,763] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:39,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:06:39,586] loop.py:45 INFO TRAIN_STEP step=14701 loss=1.229881 Training: 82%|████████▏ | 14701/18018 [31:36:25<6:40:33, 7.25s/step, loss=1.230, acc=0.711, acc_len=1.71, thru=18.1, I=19.1, T=37.6, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:06:39,582] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14700: forward=2411.3ms backward=941.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:39,942] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:46,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:06:47,296] loop.py:45 INFO TRAIN_STEP step=14702 loss=1.206961 Training: 82%|████████▏ | 14702/18018 [31:36:33<6:48:09, 7.39s/step, loss=1.207, acc=0.715, acc_len=1.74, thru=16.5, I=18.5, T=25.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:47,442] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:54,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:06:55,864] loop.py:45 INFO TRAIN_STEP step=14703 loss=1.189542 Training: 82%|████████▏ | 14703/18018 [31:36:41<7:07:38, 7.74s/step, loss=1.190, acc=0.722, acc_len=1.75, thru=14.5, I=16.0, T=30.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:06:56,063] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:07:01,767] loop.py:45 INFO TRAIN_STEP step=14704 loss=1.302326 Training: 82%|████████▏ | 14704/18018 [31:36:47<6:37:04, 7.19s/step, loss=1.302, acc=0.697, acc_len=1.67, thru=17.0, I=20.3, T=39.0, wait=2.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:11,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:07:12,897] loop.py:380 INFO TIMING step=14705: step=3.883s data=1.556s compute=3.849s [fwd=2.719s bwd=1.116s opt=0.003s] dispatch=7.239s [2026-10-19 21:07:12,897] loop.py:45 INFO TRAIN_STEP step=14705 loss=1.265376 Training: 82%|████████▏ | 14705/18018 [31:36:58<7:42:15, 8.37s/step, loss=1.265, acc=0.708, acc_len=1.71, thru=14.5, I=16.1, T=33.0, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:13,145] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 21:07:18,708] loop.py:45 INFO TRAIN_STEP step=14706 loss=1.245841 Training: 82%|████████▏ | 14706/18018 [31:37:04<6:59:42, 7.60s/step, loss=1.246, acc=0.711, acc_len=1.71, thru=20.0, I=25.6, T=39.0, wait=2.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:07:18,704] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14705: forward=2335.3ms backward=894.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:26,141] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:07:27,130] loop.py:45 INFO TRAIN_STEP step=14707 loss=1.352363 Training: 82%|████████▏ | 14707/18018 [31:37:13<7:13:08, 7.85s/step, loss=1.352, acc=0.691, acc_len=1.65, thru=17.0, I=17.4, T=37.4, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:27,162] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:32,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:07:33,754] loop.py:45 INFO TRAIN_STEP step=14708 loss=1.293789 Training: 82%|████████▏ | 14708/18018 [31:37:19<6:52:44, 7.48s/step, loss=1.294, acc=0.698, acc_len=1.68, thru=19.3, I=22.1, T=38.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:33,900] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:39,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:07:39,981] loop.py:45 INFO TRAIN_STEP step=14709 loss=1.241544 Training: 82%|████████▏ | 14709/18018 [31:37:26<6:31:48, 7.10s/step, loss=1.242, acc=0.712, acc_len=1.72, thru=20.5, I=21.6, T=38.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:40,231] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:07:46,294] loop.py:380 INFO TIMING step=14710: step=3.376s data=1.437s compute=3.344s [fwd=2.417s bwd=0.912s opt=0.003s] dispatch=2.928s [2026-10-19 21:07:46,294] loop.py:45 INFO TRAIN_STEP step=14710 loss=1.270202 Training: 82%|████████▏ | 14710/18018 [31:37:32<6:18:36, 6.87s/step, loss=1.270, acc=0.703, acc_len=1.69, thru=19.7, I=20.8, T=37.9, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:52,904] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:07:53,266] loop.py:45 INFO TRAIN_STEP step=14711 loss=1.247162 Training: 82%|████████▏ | 14711/18018 [31:37:39<6:20:13, 6.90s/step, loss=1.247, acc=0.709, acc_len=1.71, thru=18.6, I=18.9, T=36.9, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:07:53,262] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14710: forward=2474.9ms backward=943.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:07:53,411] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 21:07:59,212] loop.py:45 INFO TRAIN_STEP step=14712 loss=1.268057 Training: 82%|████████▏ | 14712/18018 [31:37:45<6:04:20, 6.61s/step, loss=1.268, acc=0.704, acc_len=1.69, thru=20.1, I=20.4, T=38.6, wait=2.6s, pool=57, epoch=2/2][2026-10-19 21:08:07,495] loop.py:45 INFO TRAIN_STEP step=14713 loss=1.352511 Training: 82%|████████▏ | 14713/18018 [31:37:53<6:31:56, 7.12s/step, loss=1.353, acc=0.690, acc_len=1.65, thru=15.7, I=15.6, T=36.2, wait=4.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:13,355] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:08:14,181] loop.py:45 INFO TRAIN_STEP step=14714 loss=1.246361 Training: 82%|████████▏ | 14714/18018 [31:38:00<6:24:41, 6.99s/step, loss=1.246, acc=0.708, acc_len=1.71, thru=19.4, I=20.0, T=36.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:14,361] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:20,797] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:08:20,959] loop.py:380 INFO TIMING step=14715: step=3.372s data=1.351s compute=3.342s [fwd=2.399s bwd=0.929s opt=0.003s] dispatch=3.396s [2026-10-19 21:08:20,959] loop.py:45 INFO TRAIN_STEP step=14715 loss=1.186384 Training: 82%|████████▏ | 14715/18018 [31:38:07<6:21:07, 6.92s/step, loss=1.186, acc=0.719, acc_len=1.75, thru=18.8, I=21.3, T=38.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:21,298] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:08:28,001] loop.py:45 INFO TRAIN_STEP step=14716 loss=1.367181 Training: 82%|████████▏ | 14716/18018 [31:38:14<6:22:55, 6.96s/step, loss=1.367, acc=0.688, acc_len=1.64, thru=18.7, I=19.4, T=37.1, wait=3.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:08:27,998] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14715: forward=2447.0ms backward=951.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:37,907] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:08:38,521] loop.py:45 INFO TRAIN_STEP step=14717 loss=1.192228 Training: 82%|████████▏ | 14717/18018 [31:38:24<7:21:38, 8.03s/step, loss=1.192, acc=0.718, acc_len=1.74, thru=13.2, I=12.2, T=36.7, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:38,911] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:44,128] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:08:44,177] loop.py:45 INFO TRAIN_STEP step=14718 loss=1.266568 Training: 82%|████████▏ | 14718/18018 [31:38:30<6:42:20, 7.32s/step, loss=1.267, acc=0.707, acc_len=1.71, thru=21.6, I=24.0, T=39.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:44,630] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:51,986] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:08:52,092] loop.py:45 INFO TRAIN_STEP step=14719 loss=1.298140 Training: 82%|████████▏ | 14719/18018 [31:38:38<6:52:07, 7.50s/step, loss=1.298, acc=0.705, acc_len=1.69, thru=16.1, I=16.2, T=37.0, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:08:52,487] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:09:00,235] loop.py:380 INFO TIMING step=14720: step=3.393s data=1.357s compute=3.359s [fwd=2.431s bwd=0.914s opt=0.003s] dispatch=4.742s [2026-10-19 21:09:00,236] loop.py:45 INFO TRAIN_STEP step=14720 loss=1.361714 Training: 82%|████████▏ | 14720/18018 [31:38:46<7:02:41, 7.69s/step, loss=1.362, acc=0.687, acc_len=1.64, thru=14.2, I=15.4, T=37.7, wait=4.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:09,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:09:10,849] loop.py:45 INFO TRAIN_STEP step=14721 loss=1.328077 Training: 82%|████████▏ | 14721/18018 [31:38:56<7:50:47, 8.57s/step, loss=1.328, acc=0.698, acc_len=1.68, thru=14.9, I=15.1, T=33.6, wait=6.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:09:10,846] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14720: forward=2644.7ms backward=1118.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:10,925] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:18,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:09:18,721] loop.py:45 INFO TRAIN_STEP step=14722 loss=1.244411 Training: 82%|████████▏ | 14722/18018 [31:39:04<7:39:11, 8.36s/step, loss=1.244, acc=0.708, acc_len=1.70, thru=16.1, I=18.1, T=38.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:18,938] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:24,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:09:25,259] loop.py:45 INFO TRAIN_STEP step=14723 loss=1.362854 Training: 82%|████████▏ | 14723/18018 [31:39:11<7:09:00, 7.81s/step, loss=1.363, acc=0.689, acc_len=1.64, thru=19.7, I=19.4, T=38.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:25,632] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:30,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:09:31,532] loop.py:45 INFO TRAIN_STEP step=14724 loss=1.179226 Training: 82%|████████▏ | 14724/18018 [31:39:17<6:43:31, 7.35s/step, loss=1.179, acc=0.724, acc_len=1.77, thru=20.5, I=24.0, T=38.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:31,706] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:09:38,503] loop.py:380 INFO TIMING step=14725: step=3.251s data=1.302s compute=3.218s [fwd=2.323s bwd=0.882s opt=0.003s] dispatch=3.712s [2026-10-19 21:09:38,503] loop.py:45 INFO TRAIN_STEP step=14725 loss=1.388188 Training: 82%|████████▏ | 14725/18018 [31:39:24<6:37:11, 7.24s/step, loss=1.388, acc=0.681, acc_len=1.62, thru=17.8, I=19.5, T=39.4, wait=3.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:47,341] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:09:47,595] loop.py:45 INFO TRAIN_STEP step=14726 loss=1.335189 Training: 82%|████████▏ | 14726/18018 [31:39:33<7:07:33, 7.79s/step, loss=1.335, acc=0.696, acc_len=1.67, thru=15.8, I=15.9, T=34.7, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:09:47,590] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14725: forward=2621.9ms backward=1020.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:09:47,842] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:09:54,330] loop.py:45 INFO TRAIN_STEP step=14727 loss=1.301561 Training: 82%|████████▏ | 14727/18018 [31:39:40<6:50:02, 7.48s/step, loss=1.302, acc=0.698, acc_len=1.68, thru=19.1, I=18.3, T=38.2, wait=3.4s, pool=61, epoch=2/2][2026-10-19 21:10:02,058] loop.py:45 INFO TRAIN_STEP step=14728 loss=1.217968 Training: 82%|████████▏ | 14728/18018 [31:39:48<6:54:04, 7.55s/step, loss=1.218, acc=0.715, acc_len=1.73, thru=16.2, I=14.6, T=35.4, wait=4.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:08,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:10:09,229] loop.py:45 INFO TRAIN_STEP step=14729 loss=1.285663 Training: 82%|████████▏ | 14729/18018 [31:39:55<6:47:42, 7.44s/step, loss=1.286, acc=0.704, acc_len=1.69, thru=18.9, I=18.1, T=37.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:09,487] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:14,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:10:14,971] loop.py:380 INFO TIMING step=14730: step=3.217s data=1.279s compute=3.185s [fwd=2.302s bwd=0.869s opt=0.003s] dispatch=2.515s [2026-10-19 21:10:14,971] loop.py:45 INFO TRAIN_STEP step=14730 loss=1.250178 Training: 82%|████████▏ | 14730/18018 [31:40:01<6:19:42, 6.93s/step, loss=1.250, acc=0.709, acc_len=1.72, thru=21.4, I=24.1, T=39.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:15,195] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:10:21,746] loop.py:45 INFO TRAIN_STEP step=14731 loss=1.353858 Training: 82%|████████▏ | 14731/18018 [31:40:07<6:17:05, 6.88s/step, loss=1.354, acc=0.689, acc_len=1.65, thru=17.8, I=20.4, T=38.8, wait=3.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:10:21,741] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14730: forward=2342.5ms backward=902.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:29,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:10:30,880] loop.py:45 INFO TRAIN_STEP step=14732 loss=1.321794 Training: 82%|████████▏ | 14732/18018 [31:40:16<6:53:54, 7.56s/step, loss=1.322, acc=0.695, acc_len=1.67, thru=15.4, I=17.8, T=36.3, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:30,909] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:35,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:10:36,322] loop.py:45 INFO TRAIN_STEP step=14733 loss=1.239137 Training: 82%|████████▏ | 14733/18018 [31:40:22<6:19:03, 6.92s/step, loss=1.239, acc=0.708, acc_len=1.71, thru=22.1, I=28.6, T=41.1, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:36,498] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:42,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:10:43,598] loop.py:45 INFO TRAIN_STEP step=14734 loss=1.241995 Training: 82%|████████▏ | 14734/18018 [31:40:29<6:24:43, 7.03s/step, loss=1.242, acc=0.712, acc_len=1.73, thru=18.8, I=21.8, T=36.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:44,049] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:10:49,561] loop.py:380 INFO TIMING step=14735: step=3.147s data=1.247s compute=3.123s [fwd=2.239s bwd=0.871s opt=0.003s] dispatch=2.806s [2026-10-19 21:10:49,561] loop.py:45 INFO TRAIN_STEP step=14735 loss=1.281213 Training: 82%|████████▏ | 14735/18018 [31:40:35<6:07:06, 6.71s/step, loss=1.281, acc=0.701, acc_len=1.68, thru=18.4, I=20.2, T=40.7, wait=2.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:56,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:10:57,668] loop.py:45 INFO TRAIN_STEP step=14736 loss=1.207210 Training: 82%|████████▏ | 14736/18018 [31:40:43<6:29:56, 7.13s/step, loss=1.207, acc=0.717, acc_len=1.74, thru=18.6, I=19.1, T=37.7, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:10:57,664] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14735: forward=2389.9ms backward=946.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:10:57,977] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:03,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:11:04,317] loop.py:45 INFO TRAIN_STEP step=14737 loss=1.170811 Training: 82%|████████▏ | 14737/18018 [31:40:50<6:21:56, 6.98s/step, loss=1.171, acc=0.722, acc_len=1.75, thru=19.3, I=24.5, T=38.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:04,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:08,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:11:09,869] loop.py:45 INFO TRAIN_STEP step=14738 loss=1.165003 Training: 82%|████████▏ | 14738/18018 [31:40:55<5:58:20, 6.56s/step, loss=1.165, acc=0.724, acc_len=1.76, thru=22.2, I=28.0, T=39.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:09,902] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 21:11:16,246] loop.py:45 INFO TRAIN_STEP step=14739 loss=1.177141 Training: 82%|████████▏ | 14739/18018 [31:41:02<5:55:15, 6.50s/step, loss=1.177, acc=0.720, acc_len=1.75, thru=18.3, I=22.6, T=39.8, wait=3.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:23,241] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:11:23,947] loop.py:380 INFO TIMING step=14740: step=3.340s data=1.323s compute=3.314s [fwd=2.378s bwd=0.922s opt=0.003s] dispatch=4.353s [2026-10-19 21:11:23,947] loop.py:45 INFO TRAIN_STEP step=14740 loss=1.193771 Training: 82%|████████▏ | 14740/18018 [31:41:10<6:14:50, 6.86s/step, loss=1.194, acc=0.717, acc_len=1.74, thru=17.8, I=18.7, T=38.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:24,245] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:28,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:11:30,048] loop.py:45 INFO TRAIN_STEP step=14741 loss=1.260485 Training: 82%|████████▏ | 14741/18018 [31:41:16<6:02:16, 6.63s/step, loss=1.260, acc=0.707, acc_len=1.70, thru=20.8, I=25.7, T=39.1, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:11:30,043] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14740: forward=2325.3ms backward=901.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:30,305] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:11:36,397] loop.py:45 INFO TRAIN_STEP step=14742 loss=1.240726 Training: 82%|████████▏ | 14742/18018 [31:41:22<5:57:30, 6.55s/step, loss=1.241, acc=0.707, acc_len=1.71, thru=19.6, I=23.6, T=39.7, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:42,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:11:42,991] loop.py:45 INFO TRAIN_STEP step=14743 loss=1.342078 Training: 82%|████████▏ | 14743/18018 [31:41:29<5:58:11, 6.56s/step, loss=1.342, acc=0.690, acc_len=1.65, thru=20.1, I=21.5, T=36.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:43,293] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:48,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:11:49,384] loop.py:45 INFO TRAIN_STEP step=14744 loss=1.366820 Training: 82%|████████▏ | 14744/18018 [31:41:35<5:55:16, 6.51s/step, loss=1.367, acc=0.691, acc_len=1.64, thru=19.8, I=24.4, T=39.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:49,703] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:55,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:11:55,977] loop.py:380 INFO TIMING step=14745: step=3.354s data=1.378s compute=3.315s [fwd=2.353s bwd=0.946s opt=0.003s] dispatch=3.230s [2026-10-19 21:11:55,978] loop.py:45 INFO TRAIN_STEP step=14745 loss=1.215789 Training: 82%|████████▏ | 14745/18018 [31:41:42<5:56:31, 6.54s/step, loss=1.216, acc=0.715, acc_len=1.73, thru=19.9, I=22.6, T=38.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:11:56,050] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 21:12:04,076] loop.py:45 INFO TRAIN_STEP step=14746 loss=1.341260 Training: 82%|████████▏ | 14746/18018 [31:41:50<6:22:01, 7.01s/step, loss=1.341, acc=0.696, acc_len=1.66, thru=16.1, I=18.2, T=36.0, wait=4.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:12:04,071] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14745: forward=2516.0ms backward=1001.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:12:14,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:12:14,558] loop.py:45 INFO TRAIN_STEP step=14747 loss=1.347697 Training: 82%|████████▏ | 14747/18018 [31:42:00<7:18:43, 8.05s/step, loss=1.348, acc=0.692, acc_len=1.65, thru=13.0, I=12.1, T=34.1, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:12:14,824] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:12:22,939] loop.py:45 INFO TRAIN_STEP step=14748 loss=1.190541 Training: 82%|████████▏ | 14748/18018 [31:42:09<7:24:03, 8.15s/step, loss=1.191, acc=0.717, acc_len=1.74, thru=12.7, I=13.4, T=35.1, wait=4.7s, pool=45, epoch=2/2][2026-10-19 21:12:31,916] loop.py:45 INFO TRAIN_STEP step=14749 loss=1.358063 Training: 82%|████████▏ | 14749/18018 [31:42:18<7:37:27, 8.40s/step, loss=1.358, acc=0.695, acc_len=1.66, thru=15.3, I=16.7, T=33.7, wait=5.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:12:40,402] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:12:40,969] loop.py:380 INFO TIMING step=14750: step=3.557s data=1.438s compute=3.526s [fwd=2.479s bwd=1.032s opt=0.003s] dispatch=5.487s [2026-10-19 21:12:40,969] loop.py:45 INFO TRAIN_STEP step=14750 loss=1.316666 Training: 82%|████████▏ | 14750/18018 [31:42:27<7:48:03, 8.59s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=14.4, I=14.1, T=36.0, wait=5.5s, pool=65, epoch=2/2][2026-10-19 21:12:40,972] loop.py:419 INFO Saving checkpoint at step 14750... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 21:12:44,371] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0014751 (TrainerActor pid=9310) [2026-10-19 21:12:44,430] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0014001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:12:45,457] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.1s): pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:12:45,458] inference_manager.py:408 INFO Pool capacity available after 5.1s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:12:50,538] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 21:12:44,370] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0014751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 21:12:45,425] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0014001 [2026-10-19 21:12:51,338] loop.py:45 INFO TRAIN_STEP step=14751 loss=1.239794 Training: 82%|████████▏ | 14751/18018 [31:42:37<8:16:56, 9.13s/step, loss=1.240, acc=0.712, acc_len=1.72, thru=21.8, I=24.5, T=37.7, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:12:51,333] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14750: forward=2424.7ms backward=934.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:12:51,543] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:12:57,285] loop.py:45 INFO TRAIN_STEP step=14752 loss=1.332527 Training: 82%|████████▏ | 14752/18018 [31:42:43<7:24:49, 8.17s/step, loss=1.333, acc=0.689, acc_len=1.64, thru=19.4, I=21.5, T=39.3, wait=2.7s, pool=49, epoch=2/2][2026-10-19 21:13:03,719] loop.py:45 INFO TRAIN_STEP step=14753 loss=1.346451 Training: 82%|████████▏ | 14753/18018 [31:42:49<6:56:21, 7.65s/step, loss=1.346, acc=0.700, acc_len=1.68, thru=20.0, I=20.9, T=37.6, wait=3.0s, pool=57, epoch=2/2][2026-10-19 21:13:13,088] loop.py:45 INFO TRAIN_STEP step=14754 loss=1.263809 Training: 82%|████████▏ | 14754/18018 [31:42:59<7:24:14, 8.17s/step, loss=1.264, acc=0.705, acc_len=1.70, thru=13.9, I=14.1, T=35.9, wait=5.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:13:20,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:13:22,142] loop.py:380 INFO TIMING step=14755: step=3.652s data=1.464s compute=3.625s [fwd=2.577s bwd=1.033s opt=0.003s] dispatch=5.392s [2026-10-19 21:13:22,142] loop.py:45 INFO TRAIN_STEP step=14755 loss=1.323802 Training: 82%|████████▏ | 14755/18018 [31:43:08<7:38:34, 8.43s/step, loss=1.324, acc=0.692, acc_len=1.65, thru=14.7, I=16.6, T=35.0, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:13:22,246] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-19 21:13:30,068] loop.py:45 INFO TRAIN_STEP step=14756 loss=1.222278 Training: 82%|████████▏ | 14756/18018 [31:43:16<7:30:10, 8.28s/step, loss=1.222, acc=0.714, acc_len=1.72, thru=11.3, I=14.1, T=35.3, wait=4.3s, pool=25, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:13:30,064] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14755: forward=2505.2ms backward=1061.5ms (TrainerActor pid=9370) [2026-10-19 21:13:30,126] eagle_store.py:517 WARNING batch_get_buffer missing keys (k472194-9fb9a945_hs, k472194-9fb9a945_ids, k472194-9fb9a945_lhs); sleeping 0.50s. [2026-10-19 21:13:38,321] loop.py:45 INFO TRAIN_STEP step=14757 loss=1.359173 Training: 82%|████████▏ | 14757/18018 [31:43:24<7:29:36, 8.27s/step, loss=1.359, acc=0.687, acc_len=1.64, thru=18.8, I=15.7, T=36.8, wait=4.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:13:44,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9368) [2026-10-19 21:13:30,167] eagle_store.py:517 WARNING batch_get_buffer missing keys (k472192-9e652eaf_hs, k472192-9e652eaf_ids, k472192-9e652eaf_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-19 21:13:44,558] loop.py:45 INFO TRAIN_STEP step=14758 loss=1.309440 Training: 82%|████████▏ | 14758/18018 [31:43:30<6:56:15, 7.66s/step, loss=1.309, acc=0.698, acc_len=1.67, thru=20.1, I=21.1, T=37.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:13:45,031] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:13:50,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:13:50,855] loop.py:45 INFO TRAIN_STEP step=14759 loss=1.262603 Training: 82%|████████▏ | 14759/18018 [31:43:36<6:33:56, 7.25s/step, loss=1.263, acc=0.709, acc_len=1.71, thru=20.6, I=24.3, T=38.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:13:51,036] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:13:55,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:13:56,677] loop.py:380 INFO TIMING step=14760: step=3.364s data=1.335s compute=3.328s [fwd=2.326s bwd=0.987s opt=0.003s] dispatch=2.448s [2026-10-19 21:13:56,677] loop.py:45 INFO TRAIN_STEP step=14760 loss=1.299624 Training: 82%|████████▏ | 14760/18018 [31:43:42<6:10:29, 6.82s/step, loss=1.300, acc=0.698, acc_len=1.67, thru=21.3, I=25.7, T=38.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:13:57,108] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:14:03,871] loop.py:45 INFO TRAIN_STEP step=14761 loss=1.245141 Training: 82%|████████▏ | 14761/18018 [31:43:49<6:16:26, 6.93s/step, loss=1.245, acc=0.711, acc_len=1.72, thru=16.7, I=22.8, T=38.2, wait=3.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:14:03,866] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14760: forward=2343.6ms backward=949.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:11,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:14:11,579] loop.py:45 INFO TRAIN_STEP step=14762 loss=1.279064 Training: 82%|████████▏ | 14762/18018 [31:43:57<6:28:53, 7.17s/step, loss=1.279, acc=0.707, acc_len=1.70, thru=17.3, I=17.6, T=35.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:11,881] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:19,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:14:19,242] loop.py:45 INFO TRAIN_STEP step=14763 loss=1.199551 Training: 82%|████████▏ | 14763/18018 [31:44:05<6:36:51, 7.32s/step, loss=1.200, acc=0.719, acc_len=1.74, thru=16.9, I=16.8, T=36.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:19,708] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:26,804] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:14:26,955] loop.py:45 INFO TRAIN_STEP step=14764 loss=1.103076 Training: 82%|████████▏ | 14764/18018 [31:44:13<6:43:13, 7.43s/step, loss=1.103, acc=0.734, acc_len=1.80, thru=16.7, I=16.7, T=35.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:27,306] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:32,601] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:14:33,264] loop.py:380 INFO TIMING step=14765: step=3.254s data=1.292s compute=3.219s [fwd=2.311s bwd=0.893s opt=0.003s] dispatch=3.045s [2026-10-19 21:14:33,264] loop.py:45 INFO TRAIN_STEP step=14765 loss=1.419245 Training: 82%|████████▏ | 14765/18018 [31:44:19<6:24:47, 7.10s/step, loss=1.419, acc=0.676, acc_len=1.60, thru=19.6, I=21.5, T=39.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:33,610] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:39,357] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:14:40,037] loop.py:45 INFO TRAIN_STEP step=14766 loss=1.294997 Training: 82%|████████▏ | 14766/18018 [31:44:26<6:19:24, 7.00s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=19.2, I=23.7, T=39.1, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:14:40,032] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14765: forward=2349.0ms backward=878.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:14:40,365] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:14:45,668] loop.py:45 INFO TRAIN_STEP step=14767 loss=1.128222 Training: 82%|████████▏ | 14767/18018 [31:44:31<5:57:02, 6.59s/step, loss=1.128, acc=0.731, acc_len=1.79, thru=18.9, I=23.5, T=38.6, wait=2.3s, pool=41, epoch=2/2][2026-10-19 21:14:59,418] loop.py:45 INFO TRAIN_STEP step=14768 loss=1.199722 Training: 82%|████████▏ | 14768/18018 [31:44:45<7:53:16, 8.74s/step, loss=1.200, acc=0.717, acc_len=1.74, thru=10.5, I=10.4, T=34.0, wait=10.0s, pool=53, epoch=2/2][2026-10-19 21:15:08,447] loop.py:45 INFO TRAIN_STEP step=14769 loss=1.222877 Training: 82%|████████▏ | 14769/18018 [31:44:54<7:57:51, 8.82s/step, loss=1.223, acc=0.713, acc_len=1.72, thru=15.2, I=13.5, T=32.6, wait=5.1s, pool=61, epoch=2/2] [2026-10-19 21:15:18,545] loop.py:380 INFO TIMING step=14770: step=3.658s data=1.390s compute=3.616s [fwd=2.578s bwd=1.023s opt=0.003s] dispatch=6.431s [2026-10-19 21:15:18,545] loop.py:45 INFO TRAIN_STEP step=14770 loss=1.217775 Training: 82%|████████▏ | 14770/18018 [31:45:04<8:18:22, 9.21s/step, loss=1.218, acc=0.712, acc_len=1.72, thru=11.8, I=12.6, T=35.0, wait=6.4s, pool=49, epoch=2/2][2026-10-19 21:15:28,610] loop.py:45 INFO TRAIN_STEP step=14771 loss=1.311806 Training: 82%|████████▏ | 14771/18018 [31:45:14<8:32:10, 9.46s/step, loss=1.312, acc=0.699, acc_len=1.67, thru=14.5, I=12.5, T=35.9, wait=6.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:15:28,605] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14770: forward=2560.8ms backward=964.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:15:42,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:15:43,927] loop.py:45 INFO TRAIN_STEP step=14772 loss=1.220329 Training: 82%|████████▏ | 14772/18018 [31:45:30<10:07:01, 11.22s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=10.3, I=9.2, T=33.1, wait=11.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:15:44,284] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:15:49,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:15:50,857] loop.py:45 INFO TRAIN_STEP step=14773 loss=1.215573 Training: 82%|████████▏ | 14773/18018 [31:45:36<8:57:13, 9.93s/step, loss=1.216, acc=0.715, acc_len=1.73, thru=18.6, I=19.6, T=36.8, wait=3.4s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:15:51,302] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:15:57,453] loop.py:45 INFO TRAIN_STEP step=14774 loss=1.296069 Training: 82%|████████▏ | 14774/18018 [31:45:43<8:02:53, 8.93s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=17.6, I=20.7, T=38.2, wait=3.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:06,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:16:07,955] loop.py:380 INFO TIMING step=14775: step=3.711s data=1.422s compute=3.667s [fwd=2.616s bwd=1.036s opt=0.003s] dispatch=6.783s [2026-10-19 21:16:07,955] loop.py:45 INFO TRAIN_STEP step=14775 loss=1.274975 Training: 82%|████████▏ | 14775/18018 [31:45:54<8:28:14, 9.40s/step, loss=1.275, acc=0.705, acc_len=1.69, thru=14.0, I=15.7, T=34.5, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:08,178] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:16:13,413] loop.py:45 INFO TRAIN_STEP step=14776 loss=1.239730 Training: 82%|████████▏ | 14776/18018 [31:45:59<7:24:07, 8.22s/step, loss=1.240, acc=0.708, acc_len=1.72, thru=20.0, I=21.9, T=38.5, wait=2.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:16:13,409] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14775: forward=2334.8ms backward=934.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:19,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:16:19,355] loop.py:45 INFO TRAIN_STEP step=14777 loss=1.211721 Training: 82%|████████▏ | 14777/18018 [31:46:05<6:47:06, 7.54s/step, loss=1.212, acc=0.717, acc_len=1.74, thru=22.3, I=22.3, T=38.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:19,627] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:16:26,423] loop.py:45 INFO TRAIN_STEP step=14778 loss=1.247713 Training: 82%|████████▏ | 14778/18018 [31:46:12<6:39:23, 7.40s/step, loss=1.248, acc=0.713, acc_len=1.73, thru=18.2, I=18.7, T=38.0, wait=3.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:31,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:16:32,276] loop.py:45 INFO TRAIN_STEP step=14779 loss=1.350933 Training: 82%|████████▏ | 14779/18018 [31:46:18<6:14:15, 6.93s/step, loss=1.351, acc=0.690, acc_len=1.65, thru=20.9, I=21.7, T=38.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:32,384] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:37,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:16:37,991] loop.py:380 INFO TIMING step=14780: step=3.302s data=1.360s compute=3.274s [fwd=2.352s bwd=0.907s opt=0.003s] dispatch=2.403s [2026-10-19 21:16:37,991] loop.py:45 INFO TRAIN_STEP step=14780 loss=1.368368 Training: 82%|████████▏ | 14780/18018 [31:46:24<5:54:25, 6.57s/step, loss=1.368, acc=0.692, acc_len=1.65, thru=22.8, I=22.7, T=38.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:38,250] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:44,277] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:16:45,401] loop.py:45 INFO TRAIN_STEP step=14781 loss=1.299748 Training: 82%|████████▏ | 14781/18018 [31:46:31<6:07:57, 6.82s/step, loss=1.300, acc=0.700, acc_len=1.68, thru=18.5, I=20.7, T=36.9, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:16:45,397] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14780: forward=2428.3ms backward=989.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:16:45,788] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:16:53,284] loop.py:45 INFO TRAIN_STEP step=14782 loss=1.298313 Training: 82%|████████▏ | 14782/18018 [31:46:39<6:25:00, 7.14s/step, loss=1.298, acc=0.701, acc_len=1.68, thru=15.8, I=19.7, T=35.5, wait=4.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:17:00,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:17:00,898] loop.py:45 INFO TRAIN_STEP step=14783 loss=1.180181 Training: 82%|████████▏ | 14783/18018 [31:46:46<6:32:34, 7.28s/step, loss=1.180, acc=0.719, acc_len=1.74, thru=16.4, I=17.8, T=36.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:17:01,020] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 21:17:06,897] loop.py:45 INFO TRAIN_STEP step=14784 loss=1.213077 Training: 82%|████████▏ | 14784/18018 [31:46:52<6:11:44, 6.90s/step, loss=1.213, acc=0.715, acc_len=1.73, thru=16.9, I=18.9, T=39.3, wait=2.7s, pool=29, epoch=2/2][2026-10-19 21:17:14,534] loop.py:380 INFO TIMING step=14785: step=3.563s data=1.426s compute=3.521s [fwd=2.503s bwd=1.004s opt=0.003s] dispatch=4.065s [2026-10-19 21:17:14,535] loop.py:45 INFO TRAIN_STEP step=14785 loss=1.347956 Training: 82%|████████▏ | 14785/18018 [31:47:00<6:23:34, 7.12s/step, loss=1.348, acc=0.695, acc_len=1.66, thru=18.8, I=17.9, T=35.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:17:14,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:17:14,993] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:17:21,680] loop.py:45 INFO TRAIN_STEP step=14786 loss=1.176645 Training: 82%|████████▏ | 14786/18018 [31:47:07<6:23:54, 7.13s/step, loss=1.177, acc=0.723, acc_len=1.76, thru=16.8, I=17.7, T=36.6, wait=3.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:17:21,675] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14785: forward=2424.0ms backward=1007.0ms [2026-10-19 21:17:31,618] loop.py:45 INFO TRAIN_STEP step=14787 loss=1.311424 Training: 82%|████████▏ | 14787/18018 [31:47:17<7:09:11, 7.97s/step, loss=1.311, acc=0.698, acc_len=1.67, thru=14.1, I=13.7, T=35.9, wait=6.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:17:40,114] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:17:41,104] loop.py:45 INFO TRAIN_STEP step=14788 loss=1.339091 Training: 82%|████████▏ | 14788/18018 [31:47:27<7:33:33, 8.43s/step, loss=1.339, acc=0.691, acc_len=1.65, thru=14.5, I=15.1, T=36.0, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:17:41,137] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 21:17:46,246] loop.py:45 INFO TRAIN_STEP step=14789 loss=1.193070 Training: 82%|████████▏ | 14789/18018 [31:47:32<6:40:25, 7.44s/step, loss=1.193, acc=0.717, acc_len=1.74, thru=21.1, I=21.5, T=39.0, wait=1.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:17:52,199] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:17:53,462] loop.py:380 INFO TIMING step=14790: step=3.432s data=1.411s compute=3.401s [fwd=2.438s bwd=0.949s opt=0.003s] dispatch=3.775s [2026-10-19 21:17:53,462] loop.py:45 INFO TRAIN_STEP step=14790 loss=1.253523 Training: 82%|████████▏ | 14790/18018 [31:47:39<6:36:41, 7.37s/step, loss=1.254, acc=0.708, acc_len=1.71, thru=18.2, I=20.9, T=37.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:17:53,706] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:18:00,718] loop.py:45 INFO TRAIN_STEP step=14791 loss=1.278822 Training: 82%|████████▏ | 14791/18018 [31:47:46<6:34:43, 7.34s/step, loss=1.279, acc=0.703, acc_len=1.70, thru=16.1, I=19.0, T=39.4, wait=4.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:18:00,714] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14790: forward=2290.5ms backward=915.1ms [2026-10-19 21:18:08,730] loop.py:45 INFO TRAIN_STEP step=14792 loss=1.262873 Training: 82%|████████▏ | 14792/18018 [31:47:54<6:45:23, 7.54s/step, loss=1.263, acc=0.708, acc_len=1.71, thru=13.2, I=14.2, T=36.7, wait=4.5s, pool=17, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 21:18:09,202] eagle_store.py:517 WARNING batch_get_buffer missing keys (k473354-81a02065_hs, k473354-81a02065_ids, k473354-81a02065_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:18:16,741] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 21:18:09,199] eagle_store.py:517 WARNING batch_get_buffer missing keys (k473353-b495d4c1_hs, k473353-b495d4c1_ids, k473353-b495d4c1_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 21:18:17,367] loop.py:45 INFO TRAIN_STEP step=14793 loss=1.176237 Training: 82%|████████▏ | 14793/18018 [31:48:03<7:02:56, 7.87s/step, loss=1.176, acc=0.722, acc_len=1.75, thru=19.5, I=16.1, T=36.6, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:18:17,747] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:18:25,281] loop.py:45 INFO TRAIN_STEP step=14794 loss=1.265705 Training: 82%|████████▏ | 14794/18018 [31:48:11<7:03:30, 7.88s/step, loss=1.266, acc=0.705, acc_len=1.70, thru=15.4, I=18.7, T=37.5, wait=4.5s, pool=57, epoch=2/2][2026-10-19 21:18:36,714] loop.py:380 INFO TIMING step=14795: step=3.668s data=1.533s compute=3.637s [fwd=2.562s bwd=1.060s opt=0.003s] dispatch=7.755s [2026-10-19 21:18:36,714] loop.py:45 INFO TRAIN_STEP step=14795 loss=1.277707 Training: 82%|████████▏ | 14795/18018 [31:48:22<8:00:37, 8.95s/step, loss=1.278, acc=0.700, acc_len=1.68, thru=12.4, I=10.7, T=34.9, wait=7.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:18:43,695] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:18:44,219] loop.py:45 INFO TRAIN_STEP step=14796 loss=1.298417 (TrainerActor pid=9310) [2026-10-19 21:18:44,215] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14795: forward=2620.4ms backward=941.9ms Training: 82%|████████▏ | 14796/18018 [31:48:30<7:37:12, 8.51s/step, loss=1.298, acc=0.703, acc_len=1.69, thru=17.8, I=18.1, T=35.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:18:44,700] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:18:51,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:18:51,682] loop.py:45 INFO TRAIN_STEP step=14797 loss=1.259041 Training: 82%|████████▏ | 14797/18018 [31:48:37<7:20:11, 8.20s/step, loss=1.259, acc=0.707, acc_len=1.70, thru=17.5, I=18.1, T=38.0, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:18:52,096] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:18:58,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:18:58,782] loop.py:45 INFO TRAIN_STEP step=14798 loss=1.183259 Training: 82%|████████▏ | 14798/18018 [31:48:44<7:02:20, 7.87s/step, loss=1.183, acc=0.721, acc_len=1.76, thru=18.9, I=19.6, T=37.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:18:59,048] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:05,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:19:06,205] loop.py:45 INFO TRAIN_STEP step=14799 loss=1.284211 Training: 82%|████████▏ | 14799/18018 [31:48:52<6:55:01, 7.74s/step, loss=1.284, acc=0.703, acc_len=1.69, thru=17.3, I=19.4, T=35.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:06,256] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:11,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:19:12,227] loop.py:380 INFO TIMING step=14800: step=3.299s data=1.358s compute=3.272s [fwd=2.387s bwd=0.871s opt=0.003s] dispatch=2.712s [2026-10-19 21:19:12,227] loop.py:45 INFO TRAIN_STEP step=14800 loss=1.268234 Training: 82%|████████▏ | 14800/18018 [31:48:58<6:27:16, 7.22s/step, loss=1.268, acc=0.701, acc_len=1.68, thru=20.3, I=22.4, T=38.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:12,601] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:19:21,795] loop.py:45 INFO TRAIN_STEP step=14801 loss=1.199622 Training: 82%|████████▏ | 14801/18018 [31:49:07<7:04:54, 7.92s/step, loss=1.200, acc=0.715, acc_len=1.74, thru=13.1, I=12.5, T=35.1, wait=5.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:19:21,790] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14800: forward=2578.6ms backward=1016.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:28,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:19:29,877] loop.py:45 INFO TRAIN_STEP step=14802 loss=1.424804 Training: 82%|████████▏ | 14802/18018 [31:49:15<7:07:18, 7.97s/step, loss=1.425, acc=0.679, acc_len=1.61, thru=17.3, I=18.3, T=36.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:29,940] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:37,610] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:19:38,721] loop.py:45 INFO TRAIN_STEP step=14803 loss=1.172327 Training: 82%|████████▏ | 14803/18018 [31:49:24<7:21:13, 8.23s/step, loss=1.172, acc=0.722, acc_len=1.75, thru=13.5, I=15.5, T=30.8, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:39,118] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:19:45,821] loop.py:45 INFO TRAIN_STEP step=14804 loss=1.217957 Training: 82%|████████▏ | 14804/18018 [31:49:31<7:02:52, 7.89s/step, loss=1.218, acc=0.712, acc_len=1.73, thru=15.4, I=18.3, T=36.6, wait=3.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:55,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:19:55,413] loop.py:380 INFO TIMING step=14805: step=3.664s data=1.475s compute=3.633s [fwd=2.594s bwd=1.025s opt=0.003s] dispatch=5.917s [2026-10-19 21:19:55,413] loop.py:45 INFO TRAIN_STEP step=14805 loss=1.275193 Training: 82%|████████▏ | 14805/18018 [31:49:41<7:29:58, 8.40s/step, loss=1.275, acc=0.704, acc_len=1.69, thru=15.9, I=14.7, T=34.9, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:19:55,728] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:20:00,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:20:00,981] loop.py:45 INFO TRAIN_STEP step=14806 loss=1.204997 Training: 82%|████████▏ | 14806/18018 [31:49:47<6:44:20, 7.55s/step, loss=1.205, acc=0.717, acc_len=1.74, thru=21.3, I=23.0, T=37.5, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:20:00,977] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14805: forward=2462.4ms backward=904.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:20:01,145] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:20:08,407] loop.py:45 INFO TRAIN_STEP step=14807 loss=1.419720 Training: 82%|████████▏ | 14807/18018 [31:49:54<6:42:07, 7.51s/step, loss=1.420, acc=0.680, acc_len=1.62, thru=15.1, I=16.1, T=37.3, wait=4.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:20:21,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:20:21,438] loop.py:45 INFO TRAIN_STEP step=14808 loss=1.186357 Training: 82%|████████▏ | 14808/18018 [31:50:07<8:10:33, 9.17s/step, loss=1.186, acc=0.718, acc_len=1.74, thru=10.8, I=10.9, T=31.3, wait=8.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:20:21,604] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:20:29,463] loop.py:45 INFO TRAIN_STEP step=14809 loss=1.314541 Training: 82%|████████▏ | 14809/18018 [31:50:15<7:52:04, 8.83s/step, loss=1.315, acc=0.700, acc_len=1.68, thru=16.0, I=16.7, T=35.5, wait=4.4s, pool=61, epoch=2/2][2026-10-19 21:20:39,515] loop.py:380 INFO TIMING step=14810: step=3.693s data=1.466s compute=3.642s [fwd=2.620s bwd=1.009s opt=0.003s] dispatch=6.349s [2026-10-19 21:20:39,515] loop.py:45 INFO TRAIN_STEP step=14810 loss=1.261989 Training: 82%|████████▏ | 14810/18018 [31:50:25<8:11:35, 9.19s/step, loss=1.262, acc=0.707, acc_len=1.71, thru=11.8, I=13.1, T=34.7, wait=6.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:20:47,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:20:48,609] loop.py:45 INFO TRAIN_STEP step=14811 loss=1.420840 Training: 82%|████████▏ | 14811/18018 [31:50:34<8:09:48, 9.16s/step, loss=1.421, acc=0.678, acc_len=1.60, thru=15.5, I=14.8, T=34.6, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:20:48,603] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14810: forward=2604.6ms backward=1052.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:20:48,725] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:20:54,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:20:55,554] loop.py:45 INFO TRAIN_STEP step=14812 loss=1.246648 Training: 82%|████████▏ | 14812/18018 [31:50:41<7:34:06, 8.50s/step, loss=1.247, acc=0.711, acc_len=1.72, thru=17.4, I=20.8, T=36.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:20:55,816] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:01,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:21:02,054] loop.py:45 INFO TRAIN_STEP step=14813 loss=1.264678 Training: 82%|████████▏ | 14813/18018 [31:50:48<7:01:55, 7.90s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=19.3, I=20.9, T=36.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:02,321] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:09,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:21:09,665] loop.py:45 INFO TRAIN_STEP step=14814 loss=1.218904 Training: 82%|████████▏ | 14814/18018 [31:50:55<6:57:10, 7.81s/step, loss=1.219, acc=0.714, acc_len=1.74, thru=16.6, I=16.1, T=35.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:10,017] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:16,296] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:21:16,545] loop.py:380 INFO TIMING step=14815: step=3.414s data=1.421s compute=3.374s [fwd=2.407s bwd=0.952s opt=0.003s] dispatch=3.457s [2026-10-19 21:21:16,545] loop.py:45 INFO TRAIN_STEP step=14815 loss=1.262466 Training: 82%|████████▏ | 14815/18018 [31:51:02<6:42:08, 7.53s/step, loss=1.262, acc=0.707, acc_len=1.71, thru=18.9, I=19.0, T=37.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:16,797] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:21:23,578] loop.py:45 INFO TRAIN_STEP step=14816 loss=1.398462 Training: 82%|████████▏ | 14816/18018 [31:51:09<6:34:00, 7.38s/step, loss=1.398, acc=0.681, acc_len=1.62, thru=13.6, I=19.8, T=37.9, wait=3.6s, pool=13, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:21:23,573] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14815: forward=2354.0ms backward=958.6ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 21:21:24,479] eagle_store.py:517 WARNING batch_get_buffer missing keys (k474110-b1bbcc94_hs, k474110-b1bbcc94_ids, k474110-b1bbcc94_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:33,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 21:21:25,110] eagle_store.py:517 WARNING batch_get_buffer missing keys (k474112-b22f3de7_hs, k474112-b22f3de7_ids, k474112-b22f3de7_lhs); sleeping 0.50s. [repeated 7x across cluster] [2026-10-19 21:21:34,244] loop.py:45 INFO TRAIN_STEP step=14817 loss=1.332855 Training: 82%|████████▏ | 14817/18018 [31:51:20<7:26:23, 8.37s/step, loss=1.333, acc=0.692, acc_len=1.65, thru=18.4, I=14.4, T=35.5, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:34,271] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:39,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:21:39,488] loop.py:45 INFO TRAIN_STEP step=14818 loss=1.186088 Training: 82%|████████▏ | 14818/18018 [31:51:25<6:36:17, 7.43s/step, loss=1.186, acc=0.716, acc_len=1.73, thru=21.2, I=26.3, T=39.4, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:39,534] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:45,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:21:46,240] loop.py:45 INFO TRAIN_STEP step=14819 loss=1.204870 Training: 82%|████████▏ | 14819/18018 [31:51:32<6:25:18, 7.23s/step, loss=1.205, acc=0.714, acc_len=1.72, thru=19.8, I=20.7, T=36.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:46,340] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:53,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:21:55,471] loop.py:380 INFO TIMING step=14820: step=5.023s data=2.989s compute=4.991s [fwd=3.991s bwd=0.984s opt=0.003s] dispatch=4.199s [2026-10-19 21:21:55,472] loop.py:45 INFO TRAIN_STEP step=14820 loss=1.260266 Training: 82%|████████▏ | 14820/18018 [31:51:41<6:57:16, 7.83s/step, loss=1.260, acc=0.707, acc_len=1.70, thru=14.1, I=18.7, T=25.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:21:55,883] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:01,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:22:04,310] loop.py:45 INFO TRAIN_STEP step=14821 loss=1.279038 Training: 82%|████████▏ | 14821/18018 [31:51:50<7:13:14, 8.13s/step, loss=1.279, acc=0.704, acc_len=1.70, thru=15.2, I=20.1, T=27.1, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:22:04,305] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14820: forward=3698.8ms backward=970.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:04,427] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=17/64 [2026-10-19 21:22:09,658] loop.py:45 INFO TRAIN_STEP step=14822 loss=1.302032 Training: 82%|████████▏ | 14822/18018 [31:51:55<6:28:40, 7.30s/step, loss=1.302, acc=0.696, acc_len=1.66, thru=18.8, I=23.5, T=37.9, wait=2.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:16,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:22:16,741] loop.py:45 INFO TRAIN_STEP step=14823 loss=1.339429 Training: 82%|████████▏ | 14823/18018 [31:52:02<6:25:07, 7.23s/step, loss=1.339, acc=0.695, acc_len=1.66, thru=18.5, I=18.9, T=36.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:17,088] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:23,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:22:23,707] loop.py:45 INFO TRAIN_STEP step=14824 loss=1.316899 Training: 82%|████████▏ | 14824/18018 [31:52:09<6:20:45, 7.15s/step, loss=1.317, acc=0.697, acc_len=1.67, thru=19.2, I=20.7, T=38.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:23,740] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 21:22:29,612] loop.py:380 INFO TIMING step=14825: step=3.352s data=1.352s compute=3.327s [fwd=2.391s bwd=0.922s opt=0.003s] dispatch=2.543s [2026-10-19 21:22:29,613] loop.py:45 INFO TRAIN_STEP step=14825 loss=1.242507 Training: 82%|████████▏ | 14825/18018 [31:52:15<6:00:42, 6.78s/step, loss=1.243, acc=0.710, acc_len=1.71, thru=18.6, I=19.9, T=38.2, wait=2.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:36,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:22:38,178] loop.py:45 INFO TRAIN_STEP step=14826 loss=1.249710 Training: 82%|████████▏ | 14826/18018 [31:52:24<6:29:08, 7.31s/step, loss=1.250, acc=0.707, acc_len=1.70, thru=17.3, I=19.1, T=36.1, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:22:38,173] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14825: forward=2522.5ms backward=982.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:38,585] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:43,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:22:43,892] loop.py:45 INFO TRAIN_STEP step=14827 loss=1.167523 Training: 82%|████████▏ | 14827/18018 [31:52:29<6:03:28, 6.83s/step, loss=1.168, acc=0.720, acc_len=1.74, thru=21.0, I=28.4, T=39.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:44,244] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:49,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:22:49,886] loop.py:45 INFO TRAIN_STEP step=14828 loss=1.182519 Training: 82%|████████▏ | 14828/18018 [31:52:35<5:49:56, 6.58s/step, loss=1.183, acc=0.718, acc_len=1.74, thru=21.3, I=24.7, T=39.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:49,913] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:54,919] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:22:55,035] loop.py:45 INFO TRAIN_STEP step=14829 loss=1.190431 Training: 82%|████████▏ | 14829/18018 [31:52:41<5:26:57, 6.15s/step, loss=1.190, acc=0.721, acc_len=1.75, thru=23.3, I=26.3, T=38.5, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:22:55,422] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:23:02,068] loop.py:380 INFO TIMING step=14830: step=3.423s data=1.389s compute=3.395s [fwd=2.469s bwd=0.911s opt=0.003s] dispatch=3.603s [2026-10-19 21:23:02,068] loop.py:45 INFO TRAIN_STEP step=14830 loss=1.285478 Training: 82%|████████▏ | 14830/18018 [31:52:48<5:40:57, 6.42s/step, loss=1.285, acc=0.701, acc_len=1.68, thru=16.4, I=19.9, T=37.4, wait=3.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:23:12,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:23:13,375] loop.py:45 INFO TRAIN_STEP step=14831 loss=1.281551 Training: 82%|████████▏ | 14831/18018 [31:52:59<6:58:44, 7.88s/step, loss=1.282, acc=0.702, acc_len=1.69, thru=15.1, I=13.6, T=33.2, wait=7.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:23:13,371] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14830: forward=2648.2ms backward=1140.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:23:13,728] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:23:19,643] loop.py:45 INFO TRAIN_STEP step=14832 loss=1.224770 Training: 82%|████████▏ | 14832/18018 [31:53:05<6:32:53, 7.40s/step, loss=1.225, acc=0.712, acc_len=1.72, thru=17.7, I=17.5, T=39.9, wait=3.1s, pool=45, epoch=2/2][2026-10-19 21:23:28,286] loop.py:45 INFO TRAIN_STEP step=14833 loss=1.324746 Training: 82%|████████▏ | 14833/18018 [31:53:14<6:52:33, 7.77s/step, loss=1.325, acc=0.697, acc_len=1.67, thru=16.7, I=15.9, T=35.5, wait=5.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:23:34,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:23:35,359] loop.py:45 INFO TRAIN_STEP step=14834 loss=1.278650 Training: 82%|████████▏ | 14834/18018 [31:53:21<6:41:17, 7.56s/step, loss=1.279, acc=0.705, acc_len=1.69, thru=18.1, I=18.6, T=37.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:23:35,444] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 21:23:41,538] loop.py:380 INFO TIMING step=14835: step=3.395s data=1.403s compute=3.358s [fwd=2.391s bwd=0.952s opt=0.003s] dispatch=2.775s [2026-10-19 21:23:41,538] loop.py:45 INFO TRAIN_STEP step=14835 loss=1.203533 Training: 82%|████████▏ | 14835/18018 [31:53:27<6:19:10, 7.15s/step, loss=1.204, acc=0.715, acc_len=1.73, thru=18.9, I=20.1, T=37.7, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:23:48,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:23:48,797] loop.py:45 INFO TRAIN_STEP step=14836 loss=1.151979 Training: 82%|████████▏ | 14836/18018 [31:53:34<6:20:48, 7.18s/step, loss=1.152, acc=0.727, acc_len=1.78, thru=18.9, I=18.5, T=37.2, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:23:48,792] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14835: forward=2459.5ms backward=943.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:23:49,259] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:23:54,656] loop.py:45 INFO TRAIN_STEP step=14837 loss=1.263279 Training: 82%|████████▏ | 14837/18018 [31:53:40<5:59:41, 6.78s/step, loss=1.263, acc=0.704, acc_len=1.69, thru=19.2, I=23.2, T=40.0, wait=2.6s, pool=45, epoch=2/2][2026-10-19 21:24:04,315] loop.py:45 INFO TRAIN_STEP step=14838 loss=1.329956 Training: 82%|████████▏ | 14838/18018 [31:53:50<6:45:15, 7.65s/step, loss=1.330, acc=0.695, acc_len=1.66, thru=12.3, I=15.2, T=35.9, wait=6.1s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:24:14,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:24:14,489] loop.py:45 INFO TRAIN_STEP step=14839 loss=1.264020 Training: 82%|████████▏ | 14839/18018 [31:54:00<7:25:19, 8.41s/step, loss=1.264, acc=0.710, acc_len=1.71, thru=16.6, I=13.9, T=35.6, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:24:14,555] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:24:21,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:24:21,992] loop.py:380 INFO TIMING step=14840: step=3.462s data=1.382s compute=3.438s [fwd=2.470s bwd=0.954s opt=0.003s] dispatch=4.032s [2026-10-19 21:24:21,992] loop.py:45 INFO TRAIN_STEP step=14840 loss=1.234135 Training: 82%|████████▏ | 14840/18018 [31:54:08<7:10:53, 8.14s/step, loss=1.234, acc=0.712, acc_len=1.72, thru=17.3, I=18.8, T=37.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:24:22,121] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:24:27,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:24:27,625] loop.py:45 INFO TRAIN_STEP step=14841 loss=1.180988 Training: 82%|████████▏ | 14841/18018 [31:54:13<6:30:59, 7.38s/step, loss=1.181, acc=0.724, acc_len=1.76, thru=21.9, I=23.9, T=39.1, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:24:27,620] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14840: forward=2337.3ms backward=897.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:24:27,759] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 21:24:35,324] loop.py:45 INFO TRAIN_STEP step=14842 loss=1.269241 Training: 82%|████████▏ | 14842/18018 [31:54:21<6:35:50, 7.48s/step, loss=1.269, acc=0.710, acc_len=1.71, thru=17.1, I=18.7, T=39.0, wait=4.4s, pool=53, epoch=2/2][2026-10-19 21:24:43,532] loop.py:45 INFO TRAIN_STEP step=14843 loss=1.298997 Training: 82%|████████▏ | 14843/18018 [31:54:29<6:47:17, 7.70s/step, loss=1.299, acc=0.701, acc_len=1.69, thru=15.2, I=18.4, T=35.8, wait=4.6s, pool=57, epoch=2/2][2026-10-19 21:24:52,006] loop.py:45 INFO TRAIN_STEP step=14844 loss=1.299160 Training: 82%|████████▏ | 14844/18018 [31:54:38<6:59:31, 7.93s/step, loss=1.299, acc=0.700, acc_len=1.68, thru=11.8, I=12.8, T=36.9, wait=5.0s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:25:01,138] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:25:01,754] loop.py:380 INFO TIMING step=14845: step=3.727s data=1.495s compute=3.697s [fwd=2.640s bwd=1.042s opt=0.003s] dispatch=6.010s [2026-10-19 21:25:01,754] loop.py:45 INFO TRAIN_STEP step=14845 loss=1.229610 Training: 82%|████████▏ | 14845/18018 [31:54:47<7:28:12, 8.48s/step, loss=1.230, acc=0.713, acc_len=1.72, thru=16.7, I=15.8, T=34.3, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:25:02,143] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:25:07,121] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:25:07,760] loop.py:45 INFO TRAIN_STEP step=14846 loss=1.297346 Training: 82%|████████▏ | 14846/18018 [31:54:53<6:48:54, 7.73s/step, loss=1.297, acc=0.698, acc_len=1.67, thru=20.0, I=22.6, T=39.6, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:25:07,756] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14845: forward=2309.9ms backward=877.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:25:08,127] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:25:13,187] loop.py:45 INFO TRAIN_STEP step=14847 loss=1.187368 Training: 82%|████████▏ | 14847/18018 [31:54:59<6:12:10, 7.04s/step, loss=1.187, acc=0.721, acc_len=1.75, thru=22.7, I=26.8, T=39.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:25:13,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:25:13,640] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:25:21,153] loop.py:45 INFO TRAIN_STEP step=14848 loss=1.148999 Training: 82%|████████▏ | 14848/18018 [31:55:07<6:26:43, 7.32s/step, loss=1.149, acc=0.726, acc_len=1.77, thru=16.0, I=17.9, T=36.0, wait=4.4s, pool=61, epoch=2/2][2026-10-19 21:25:29,354] loop.py:45 INFO TRAIN_STEP step=14849 loss=1.322315 Training: 82%|████████▏ | 14849/18018 [31:55:15<6:40:32, 7.58s/step, loss=1.322, acc=0.694, acc_len=1.65, thru=13.9, I=14.3, T=36.3, wait=4.7s, pool=49, epoch=2/2][2026-10-19 21:25:38,370] loop.py:380 INFO TIMING step=14850: step=3.663s data=1.429s compute=3.638s [fwd=2.614s bwd=1.010s opt=0.003s] dispatch=5.345s [2026-10-19 21:25:38,370] loop.py:45 INFO TRAIN_STEP step=14850 loss=1.370931 Training: 82%|████████▏ | 14850/18018 [31:55:24<7:03:06, 8.01s/step, loss=1.371, acc=0.687, acc_len=1.62, thru=14.6, I=14.3, T=34.9, wait=5.3s, pool=53, epoch=2/2][2026-10-19 21:25:47,573] loop.py:45 INFO TRAIN_STEP step=14851 loss=1.332241 Training: 82%|████████▏ | 14851/18018 [31:55:33<7:21:49, 8.37s/step, loss=1.332, acc=0.694, acc_len=1.66, thru=15.1, I=16.0, T=35.1, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:25:47,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 21:25:47,568] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14850: forward=2568.3ms backward=1031.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:25:48,024] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:25:56,268] loop.py:45 INFO TRAIN_STEP step=14852 loss=1.311823 Training: 82%|████████▏ | 14852/18018 [31:55:42<7:26:49, 8.47s/step, loss=1.312, acc=0.697, acc_len=1.66, thru=14.6, I=15.3, T=36.5, wait=5.2s, pool=57, epoch=2/2][2026-10-19 21:26:04,982] loop.py:45 INFO TRAIN_STEP step=14853 loss=1.231743 Training: 82%|████████▏ | 14853/18018 [31:55:51<7:30:36, 8.54s/step, loss=1.232, acc=0.710, acc_len=1.71, thru=13.9, I=15.4, T=36.2, wait=5.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:26:14,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:26:14,845] loop.py:45 INFO TRAIN_STEP step=14854 loss=1.262482 Training: 82%|████████▏ | 14854/18018 [31:56:00<7:51:21, 8.94s/step, loss=1.262, acc=0.706, acc_len=1.71, thru=15.0, I=13.7, T=33.6, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:26:14,956] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:26:22,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:26:23,121] loop.py:380 INFO TIMING step=14855: step=3.562s data=1.459s compute=3.523s [fwd=2.492s bwd=1.017s opt=0.003s] dispatch=4.704s [2026-10-19 21:26:23,122] loop.py:45 INFO TRAIN_STEP step=14855 loss=1.319310 Training: 82%|████████▏ | 14855/18018 [31:56:09<7:40:42, 8.74s/step, loss=1.319, acc=0.700, acc_len=1.68, thru=16.2, I=17.0, T=35.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:26:23,235] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 21:26:29,726] loop.py:45 INFO TRAIN_STEP step=14856 loss=1.248587 Training: 82%|████████▏ | 14856/18018 [31:56:15<7:06:48, 8.10s/step, loss=1.249, acc=0.709, acc_len=1.71, thru=18.4, I=19.9, T=37.5, wait=3.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:26:29,721] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14855: forward=2445.3ms backward=921.3ms [2026-10-19 21:26:38,417] loop.py:45 INFO TRAIN_STEP step=14857 loss=1.327128 Training: 82%|████████▏ | 14857/18018 [31:56:24<7:16:00, 8.28s/step, loss=1.327, acc=0.696, acc_len=1.66, thru=16.1, I=15.1, T=36.8, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:26:38,415] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:26:38,917] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:26:44,310] loop.py:45 INFO TRAIN_STEP step=14858 loss=1.260711 Training: 82%|████████▏ | 14858/18018 [31:56:30<6:38:18, 7.56s/step, loss=1.261, acc=0.703, acc_len=1.69, thru=21.1, I=21.8, T=39.5, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:26:50,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:26:50,359] loop.py:45 INFO TRAIN_STEP step=14859 loss=1.321968 Training: 82%|████████▏ | 14859/18018 [31:56:36<6:14:12, 7.11s/step, loss=1.322, acc=0.695, acc_len=1.66, thru=21.3, I=21.6, T=39.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:26:50,719] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:26:58,183] loop.py:380 INFO TIMING step=14860: step=3.473s data=1.440s compute=3.415s [fwd=2.397s bwd=1.004s opt=0.003s] dispatch=4.343s [2026-10-19 21:26:58,183] loop.py:45 INFO TRAIN_STEP step=14860 loss=1.319473 Training: 82%|████████▏ | 14860/18018 [31:56:44<6:25:24, 7.32s/step, loss=1.319, acc=0.695, acc_len=1.67, thru=13.2, I=14.7, T=36.9, wait=4.3s, pool=25, epoch=2/2][2026-10-19 21:27:08,532] loop.py:45 INFO TRAIN_STEP step=14861 loss=1.280303 Training: 82%|████████▏ | 14861/18018 [31:56:54<7:13:02, 8.23s/step, loss=1.280, acc=0.703, acc_len=1.69, thru=14.1, I=13.6, T=35.3, wait=6.7s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:27:08,527] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14860: forward=2571.2ms backward=1005.5ms [2026-10-19 21:27:16,613] loop.py:45 INFO TRAIN_STEP step=14862 loss=1.371362 Training: 82%|████████▏ | 14862/18018 [31:57:02<7:10:32, 8.19s/step, loss=1.371, acc=0.683, acc_len=1.62, thru=18.0, I=16.0, T=36.1, wait=4.5s, pool=57, epoch=2/2][2026-10-19 21:27:23,239] loop.py:45 INFO TRAIN_STEP step=14863 loss=1.354596 Training: 82%|████████▏ | 14863/18018 [31:57:09<6:45:52, 7.72s/step, loss=1.355, acc=0.693, acc_len=1.66, thru=18.2, I=17.0, T=37.0, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:27:29,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:27:30,072] loop.py:45 INFO TRAIN_STEP step=14864 loss=1.362143 Training: 82%|████████▏ | 14864/18018 [31:57:16<6:31:45, 7.45s/step, loss=1.362, acc=0.686, acc_len=1.63, thru=18.4, I=18.6, T=37.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:27:30,349] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:27:37,731] loop.py:380 INFO TIMING step=14865: step=3.344s data=1.320s compute=3.314s [fwd=2.380s bwd=0.919s opt=0.003s] dispatch=4.304s [2026-10-19 21:27:37,731] loop.py:45 INFO TRAIN_STEP step=14865 loss=1.264120 Training: 83%|████████▎ | 14865/18018 [31:57:23<6:34:56, 7.52s/step, loss=1.264, acc=0.708, acc_len=1.71, thru=16.4, I=16.8, T=38.3, wait=4.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:27:46,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:27:46,564] loop.py:45 INFO TRAIN_STEP step=14866 loss=1.316550 Training: 83%|████████▎ | 14866/18018 [31:57:32<6:55:31, 7.91s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=16.6, I=15.9, T=35.3, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:27:46,560] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14865: forward=2550.0ms backward=1033.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:27:46,751] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:27:53,367] loop.py:45 INFO TRAIN_STEP step=14867 loss=1.244876 Training: 83%|████████▎ | 14867/18018 [31:57:39<6:37:56, 7.58s/step, loss=1.245, acc=0.709, acc_len=1.71, thru=16.4, I=16.2, T=37.9, wait=3.4s, pool=37, epoch=2/2][2026-10-19 21:28:00,568] loop.py:45 INFO TRAIN_STEP step=14868 loss=1.305837 Training: 83%|████████▎ | 14868/18018 [31:57:46<6:31:53, 7.46s/step, loss=1.306, acc=0.696, acc_len=1.67, thru=17.7, I=16.3, T=38.3, wait=3.9s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:09,722] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:28:10,456] loop.py:45 INFO TRAIN_STEP step=14869 loss=1.158256 Training: 83%|████████▎ | 14869/18018 [31:57:56<7:09:56, 8.19s/step, loss=1.158, acc=0.723, acc_len=1.76, thru=16.8, I=15.8, T=34.4, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:10,729] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:28:15,970] loop.py:380 INFO TIMING step=14870: step=3.231s data=1.304s compute=3.199s [fwd=2.295s bwd=0.890s opt=0.003s] dispatch=2.272s [2026-10-19 21:28:15,970] loop.py:45 INFO TRAIN_STEP step=14870 loss=1.350410 Training: 83%|████████▎ | 14870/18018 [31:58:02<6:27:38, 7.39s/step, loss=1.350, acc=0.691, acc_len=1.65, thru=19.4, I=21.9, T=39.6, wait=2.3s, pool=49, epoch=2/2][2026-10-19 21:28:23,122] loop.py:45 INFO TRAIN_STEP step=14871 loss=1.221348 Training: 83%|████████▎ | 14871/18018 [31:58:09<6:23:47, 7.32s/step, loss=1.221, acc=0.713, acc_len=1.73, thru=16.1, I=17.5, T=38.5, wait=3.8s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:28:23,117] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14870: forward=2346.3ms backward=919.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:30,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:28:31,434] loop.py:45 INFO TRAIN_STEP step=14872 loss=1.243226 Training: 83%|████████▎ | 14872/18018 [31:58:17<6:39:20, 7.62s/step, loss=1.243, acc=0.712, acc_len=1.71, thru=17.4, I=19.8, T=35.8, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:31,588] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:37,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:28:38,204] loop.py:45 INFO TRAIN_STEP step=14873 loss=1.168215 Training: 83%|████████▎ | 14873/18018 [31:58:24<6:25:53, 7.36s/step, loss=1.168, acc=0.722, acc_len=1.76, thru=18.3, I=23.9, T=38.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:38,578] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:43,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:28:44,904] loop.py:45 INFO TRAIN_STEP step=14874 loss=1.174705 Training: 83%|████████▎ | 14874/18018 [31:58:30<6:15:22, 7.16s/step, loss=1.175, acc=0.720, acc_len=1.75, thru=19.6, I=25.1, T=37.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:45,378] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:50,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:28:51,946] loop.py:380 INFO TIMING step=14875: step=3.392s data=1.423s compute=3.363s [fwd=2.443s bwd=0.906s opt=0.003s] dispatch=3.640s [2026-10-19 21:28:51,947] loop.py:45 INFO TRAIN_STEP step=14875 loss=1.293807 Training: 83%|████████▎ | 14875/18018 [31:58:38<6:13:20, 7.13s/step, loss=1.294, acc=0.701, acc_len=1.69, thru=17.6, I=25.2, T=37.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:28:52,215] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:28:58,529] loop.py:45 INFO TRAIN_STEP step=14876 loss=1.351104 Training: 83%|████████▎ | 14876/18018 [31:58:44<6:04:37, 6.96s/step, loss=1.351, acc=0.691, acc_len=1.65, thru=18.6, I=21.8, T=37.8, wait=3.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:28:58,524] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14875: forward=2402.3ms backward=928.1ms [2026-10-19 21:29:05,740] loop.py:45 INFO TRAIN_STEP step=14877 loss=1.273892 Training: 83%|████████▎ | 14877/18018 [31:58:51<6:08:24, 7.04s/step, loss=1.274, acc=0.702, acc_len=1.68, thru=17.8, I=18.1, T=36.4, wait=3.7s, pool=57, epoch=2/2][2026-10-19 21:29:14,752] loop.py:45 INFO TRAIN_STEP step=14878 loss=1.218961 Training: 83%|████████▎ | 14878/18018 [31:59:00<6:39:19, 7.63s/step, loss=1.219, acc=0.715, acc_len=1.73, thru=10.5, I=15.7, T=38.1, wait=5.6s, pool=17, epoch=2/2][2026-10-19 21:29:24,965] loop.py:45 INFO TRAIN_STEP step=14879 loss=1.287399 Training: 83%|████████▎ | 14879/18018 [31:59:11<7:19:42, 8.40s/step, loss=1.287, acc=0.704, acc_len=1.70, thru=18.9, I=13.1, T=36.4, wait=6.7s, pool=53, epoch=2/2][2026-10-19 21:29:30,981] loop.py:380 INFO TIMING step=14880: step=3.378s data=1.311s compute=3.349s [fwd=2.422s bwd=0.913s opt=0.003s] dispatch=2.629s [2026-10-19 21:29:30,981] loop.py:45 INFO TRAIN_STEP step=14880 loss=1.279331 Training: 83%|████████▎ | 14880/18018 [31:59:17<6:42:05, 7.69s/step, loss=1.279, acc=0.703, acc_len=1.69, thru=18.0, I=20.4, T=37.9, wait=2.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:29:37,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:29:38,025] loop.py:45 INFO TRAIN_STEP step=14881 loss=1.244717 Training: 83%|████████▎ | 14881/18018 [31:59:24<6:31:53, 7.50s/step, loss=1.245, acc=0.709, acc_len=1.72, thru=18.0, I=20.3, T=37.0, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:29:38,021] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14880: forward=2451.6ms backward=959.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:29:38,153] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:29:43,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:29:44,079] loop.py:45 INFO TRAIN_STEP step=14882 loss=1.273317 Training: 83%|████████▎ | 14882/18018 [31:59:30<6:09:09, 7.06s/step, loss=1.273, acc=0.706, acc_len=1.70, thru=20.5, I=22.2, T=38.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:29:44,285] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:29:52,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:29:52,431] loop.py:45 INFO TRAIN_STEP step=14883 loss=1.354058 Training: 83%|████████▎ | 14883/18018 [31:59:38<6:29:13, 7.45s/step, loss=1.354, acc=0.689, acc_len=1.64, thru=16.2, I=17.6, T=35.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:29:52,627] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:29:59,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:30:01,127] loop.py:45 INFO TRAIN_STEP step=14884 loss=1.264491 Training: 83%|████████▎ | 14884/18018 [31:59:47<6:48:38, 7.82s/step, loss=1.264, acc=0.704, acc_len=1.70, thru=15.1, I=18.1, T=35.7, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:01,469] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:06,418] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:30:07,334] loop.py:380 INFO TIMING step=14885: step=3.245s data=1.296s compute=3.222s [fwd=2.298s bwd=0.910s opt=0.003s] dispatch=2.953s [2026-10-19 21:30:07,335] loop.py:45 INFO TRAIN_STEP step=14885 loss=1.227947 Training: 83%|████████▎ | 14885/18018 [31:59:53<6:23:11, 7.34s/step, loss=1.228, acc=0.712, acc_len=1.72, thru=19.9, I=26.9, T=39.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:07,439] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 21:30:14,336] loop.py:45 INFO TRAIN_STEP step=14886 loss=1.100422 Training: 83%|████████▎ | 14886/18018 [32:00:00<6:17:48, 7.24s/step, loss=1.100, acc=0.735, acc_len=1.81, thru=16.8, I=18.5, T=37.3, wait=3.6s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:30:14,332] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14885: forward=2475.0ms backward=915.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:21,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:30:22,221] loop.py:45 INFO TRAIN_STEP step=14887 loss=1.257075 Training: 83%|████████▎ | 14887/18018 [32:00:08<6:27:47, 7.43s/step, loss=1.257, acc=0.707, acc_len=1.70, thru=18.7, I=19.0, T=35.4, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:22,609] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:26,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:30:27,888] loop.py:45 INFO TRAIN_STEP step=14888 loss=1.219252 Training: 83%|████████▎ | 14888/18018 [32:00:13<6:00:04, 6.90s/step, loss=1.219, acc=0.715, acc_len=1.73, thru=21.2, I=28.6, T=38.3, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:28,354] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:30:36,367] loop.py:45 INFO TRAIN_STEP step=14889 loss=1.201423 Training: 83%|████████▎ | 14889/18018 [32:00:22<6:24:36, 7.38s/step, loss=1.201, acc=0.719, acc_len=1.75, thru=11.7, I=15.0, T=35.9, wait=4.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:46,379] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:30:47,315] loop.py:380 INFO TIMING step=14890: step=3.771s data=1.507s compute=3.736s [fwd=2.658s bwd=1.063s opt=0.003s] dispatch=7.169s [2026-10-19 21:30:47,316] loop.py:45 INFO TRAIN_STEP step=14890 loss=1.215096 Training: 83%|████████▎ | 14890/18018 [32:00:33<7:20:24, 8.45s/step, loss=1.215, acc=0.720, acc_len=1.75, thru=15.2, I=14.5, T=33.9, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:47,389] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:51,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:30:52,909] loop.py:45 INFO TRAIN_STEP step=14891 loss=1.169025 Training: 83%|████████▎ | 14891/18018 [32:00:38<6:35:36, 7.59s/step, loss=1.169, acc=0.722, acc_len=1.76, thru=21.3, I=25.5, T=37.2, wait=2.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:30:52,904] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14890: forward=2449.3ms backward=942.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:53,382] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:58,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:30:58,542] loop.py:45 INFO TRAIN_STEP step=14892 loss=1.197334 Training: 83%|████████▎ | 14892/18018 [32:00:44<6:04:53, 7.00s/step, loss=1.197, acc=0.717, acc_len=1.74, thru=22.0, I=26.3, T=39.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:30:58,806] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:31:04,077] loop.py:45 INFO TRAIN_STEP step=14893 loss=1.197342 Training: 83%|████████▎ | 14893/18018 [32:00:50<5:41:50, 6.56s/step, loss=1.197, acc=0.713, acc_len=1.73, thru=21.7, I=24.0, T=38.4, wait=2.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:31:12,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:31:12,477] loop.py:45 INFO TRAIN_STEP step=14894 loss=1.275907 Training: 83%|████████▎ | 14894/18018 [32:00:58<6:10:23, 7.11s/step, loss=1.276, acc=0.704, acc_len=1.69, thru=16.6, I=16.3, T=35.3, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:31:12,763] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:31:18,971] loop.py:380 INFO TIMING step=14895: step=3.481s data=1.434s compute=3.454s [fwd=2.502s bwd=0.938s opt=0.003s] dispatch=3.007s [2026-10-19 21:31:18,972] loop.py:45 INFO TRAIN_STEP step=14895 loss=1.255513 Training: 83%|████████▎ | 14895/18018 [32:01:05<6:00:35, 6.93s/step, loss=1.256, acc=0.707, acc_len=1.71, thru=18.1, I=18.5, T=36.8, wait=3.0s, pool=57, epoch=2/2][2026-10-19 21:31:28,130] loop.py:45 INFO TRAIN_STEP step=14896 loss=1.284643 Training: 83%|████████▎ | 14896/18018 [32:01:14<6:35:21, 7.60s/step, loss=1.285, acc=0.704, acc_len=1.69, thru=14.4, I=14.5, T=36.3, wait=5.6s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:31:28,125] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14895: forward=2522.7ms backward=949.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:31:36,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:31:37,408] loop.py:45 INFO TRAIN_STEP step=14897 loss=1.349768 Training: 83%|████████▎ | 14897/18018 [32:01:23<7:01:25, 8.10s/step, loss=1.350, acc=0.695, acc_len=1.66, thru=15.3, I=15.7, T=34.5, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:31:37,656] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:31:42,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:31:43,277] loop.py:45 INFO TRAIN_STEP step=14898 loss=1.219834 Training: 83%|████████▎ | 14898/18018 [32:01:29<6:26:28, 7.43s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=20.4, I=22.5, T=38.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:31:43,512] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:31:49,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:31:50,394] loop.py:45 INFO TRAIN_STEP step=14899 loss=1.263297 Training: 83%|████████▎ | 14899/18018 [32:01:36<6:21:26, 7.34s/step, loss=1.263, acc=0.703, acc_len=1.70, thru=18.4, I=22.3, T=39.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:31:50,481] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 21:31:56,906] loop.py:380 INFO TIMING step=14900: step=3.289s data=1.332s compute=3.247s [fwd=2.326s bwd=0.907s opt=0.003s] dispatch=3.212s [2026-10-19 21:31:56,906] loop.py:45 INFO TRAIN_STEP step=14900 loss=1.326131 Training: 83%|████████▎ | 14900/18018 [32:01:42<6:08:24, 7.09s/step, loss=1.326, acc=0.697, acc_len=1.67, thru=18.7, I=19.1, T=38.9, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:01,927] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:32:02,669] loop.py:45 INFO TRAIN_STEP step=14901 loss=1.257420 Training: 83%|████████▎ | 14901/18018 [32:01:48<5:47:38, 6.69s/step, loss=1.257, acc=0.708, acc_len=1.71, thru=22.0, I=23.4, T=39.4, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:32:02,664] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14900: forward=2320.8ms backward=881.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:02,932] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:07,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:32:08,341] loop.py:45 INFO TRAIN_STEP step=14902 loss=1.233599 Training: 83%|████████▎ | 14902/18018 [32:01:54<5:31:40, 6.39s/step, loss=1.234, acc=0.710, acc_len=1.72, thru=22.5, I=26.8, T=38.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:08,778] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:32:14,924] loop.py:45 INFO TRAIN_STEP step=14903 loss=1.268683 Training: 83%|████████▎ | 14903/18018 [32:02:01<5:34:37, 6.45s/step, loss=1.269, acc=0.703, acc_len=1.68, thru=17.7, I=21.8, T=38.5, wait=3.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:23,192] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:32:23,941] loop.py:45 INFO TRAIN_STEP step=14904 loss=1.211236 Training: 83%|████████▎ | 14904/18018 [32:02:10<6:14:31, 7.22s/step, loss=1.211, acc=0.713, acc_len=1.73, thru=16.7, I=17.3, T=35.8, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:24,199] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:32:29,518] loop.py:380 INFO TIMING step=14905: step=3.241s data=1.246s compute=3.213s [fwd=2.318s bwd=0.879s opt=0.003s] dispatch=2.326s [2026-10-19 21:32:29,518] loop.py:45 INFO TRAIN_STEP step=14905 loss=1.192351 Training: 83%|████████▎ | 14905/18018 [32:02:15<5:48:52, 6.72s/step, loss=1.192, acc=0.719, acc_len=1.75, thru=21.4, I=23.6, T=39.5, wait=2.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:36,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:32:37,191] loop.py:45 INFO TRAIN_STEP step=14906 loss=1.324147 Training: 83%|████████▎ | 14906/18018 [32:02:23<6:03:31, 7.01s/step, loss=1.324, acc=0.694, acc_len=1.66, thru=17.3, I=17.6, T=36.1, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:32:37,186] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14905: forward=2526.3ms backward=978.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:37,304] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:43,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:32:43,787] loop.py:45 INFO TRAIN_STEP step=14907 loss=1.263620 Training: 83%|████████▎ | 14907/18018 [32:02:29<5:56:59, 6.89s/step, loss=1.264, acc=0.704, acc_len=1.69, thru=19.3, I=21.0, T=37.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:44,006] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:51,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:32:52,418] loop.py:45 INFO TRAIN_STEP step=14908 loss=1.295959 Training: 83%|████████▎ | 14908/18018 [32:02:38<6:23:59, 7.41s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=15.7, I=18.3, T=36.0, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:52,734] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:58,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:32:59,313] loop.py:45 INFO TRAIN_STEP step=14909 loss=1.332890 Training: 83%|████████▎ | 14909/18018 [32:02:45<6:15:55, 7.25s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=17.6, I=21.5, T=38.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:32:59,345] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:07,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:33:07,671] loop.py:380 INFO TIMING step=14910: step=3.320s data=1.343s compute=3.283s [fwd=2.367s bwd=0.902s opt=0.003s] dispatch=5.029s [2026-10-19 21:33:07,672] loop.py:45 INFO TRAIN_STEP step=14910 loss=1.297276 Training: 83%|████████▎ | 14910/18018 [32:02:53<6:32:55, 7.59s/step, loss=1.297, acc=0.704, acc_len=1.69, thru=15.4, I=15.1, T=38.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:08,033] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:12,908] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:33:13,252] loop.py:45 INFO TRAIN_STEP step=14911 loss=1.377501 Training: 83%|████████▎ | 14911/18018 [32:02:59<6:01:41, 6.98s/step, loss=1.378, acc=0.688, acc_len=1.64, thru=20.0, I=18.6, T=40.2, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:33:13,247] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14910: forward=2266.9ms backward=876.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:13,423] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:20,954] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:33:21,199] loop.py:45 INFO TRAIN_STEP step=14912 loss=1.322803 Training: 83%|████████▎ | 14912/18018 [32:03:07<6:16:29, 7.27s/step, loss=1.323, acc=0.700, acc_len=1.68, thru=17.0, I=17.4, T=37.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:21,455] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:27,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:33:28,453] loop.py:45 INFO TRAIN_STEP step=14913 loss=1.370875 Training: 83%|████████▎ | 14913/18018 [32:03:14<6:16:04, 7.27s/step, loss=1.371, acc=0.689, acc_len=1.65, thru=18.2, I=19.0, T=36.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:28,653] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:34,008] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:33:34,682] loop.py:45 INFO TRAIN_STEP step=14914 loss=1.355277 Training: 83%|████████▎ | 14914/18018 [32:03:20<5:59:51, 6.96s/step, loss=1.355, acc=0.690, acc_len=1.65, thru=20.3, I=23.6, T=36.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:35,014] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:39,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:33:40,464] loop.py:380 INFO TIMING step=14915: step=3.272s data=1.302s compute=3.238s [fwd=2.329s bwd=0.894s opt=0.003s] dispatch=2.499s [2026-10-19 21:33:40,464] loop.py:45 INFO TRAIN_STEP step=14915 loss=1.298775 Training: 83%|████████▎ | 14915/18018 [32:03:26<5:41:30, 6.60s/step, loss=1.299, acc=0.697, acc_len=1.67, thru=22.1, I=24.9, T=39.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:40,786] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:33:45,707] loop.py:45 INFO TRAIN_STEP step=14916 loss=1.166056 Training: 83%|████████▎ | 14916/18018 [32:03:31<5:20:17, 6.20s/step, loss=1.166, acc=0.725, acc_len=1.77, thru=22.4, I=27.6, T=40.3, wait=2.1s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:33:45,703] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14915: forward=2259.1ms backward=877.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:53,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:33:53,790] loop.py:45 INFO TRAIN_STEP step=14917 loss=1.185177 Training: 83%|████████▎ | 14917/18018 [32:03:39<5:49:28, 6.76s/step, loss=1.185, acc=0.715, acc_len=1.73, thru=17.5, I=18.0, T=36.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:33:54,209] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:00,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:34:00,604] loop.py:45 INFO TRAIN_STEP step=14918 loss=1.311250 Training: 83%|████████▎ | 14918/18018 [32:03:46<5:50:09, 6.78s/step, loss=1.311, acc=0.693, acc_len=1.66, thru=18.9, I=21.0, T=37.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:00,802] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:34:06,876] loop.py:45 INFO TRAIN_STEP step=14919 loss=1.231450 Training: 83%|████████▎ | 14919/18018 [32:03:52<5:42:13, 6.63s/step, loss=1.231, acc=0.712, acc_len=1.72, thru=20.3, I=22.5, T=38.3, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:14,391] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:34:14,996] loop.py:380 INFO TIMING step=14920: step=3.485s data=1.418s compute=3.449s [fwd=2.435s bwd=1.000s opt=0.003s] dispatch=4.627s [2026-10-19 21:34:14,997] loop.py:45 INFO TRAIN_STEP step=14920 loss=1.369207 Training: 83%|████████▎ | 14920/18018 [32:04:01<6:05:15, 7.07s/step, loss=1.369, acc=0.687, acc_len=1.63, thru=15.7, I=16.5, T=36.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:15,397] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:34:20,950] loop.py:45 INFO TRAIN_STEP step=14921 loss=1.425536 Training: 83%|████████▎ | 14921/18018 [32:04:07<5:47:47, 6.74s/step, loss=1.426, acc=0.677, acc_len=1.60, thru=18.7, I=22.9, T=36.4, wait=2.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:34:20,945] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14920: forward=2489.1ms backward=986.2ms [2026-10-19 21:34:27,475] loop.py:45 INFO TRAIN_STEP step=14922 loss=1.155411 Training: 83%|████████▎ | 14922/18018 [32:04:13<5:44:24, 6.67s/step, loss=1.155, acc=0.725, acc_len=1.76, thru=18.9, I=19.7, T=38.2, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:32,712] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:34:33,716] loop.py:45 INFO TRAIN_STEP step=14923 loss=1.255951 Training: 83%|████████▎ | 14923/18018 [32:04:19<5:37:33, 6.54s/step, loss=1.256, acc=0.708, acc_len=1.71, thru=21.3, I=22.5, T=38.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:34,219] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:39,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:34:39,462] loop.py:45 INFO TRAIN_STEP step=14924 loss=1.176224 Training: 83%|████████▎ | 14924/18018 [32:04:25<5:25:08, 6.31s/step, loss=1.176, acc=0.718, acc_len=1.74, thru=21.4, I=25.3, T=40.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:39,550] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:44,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:34:45,853] loop.py:380 INFO TIMING step=14925: step=3.323s data=1.304s compute=3.294s [fwd=2.393s bwd=0.886s opt=0.003s] dispatch=3.058s [2026-10-19 21:34:45,853] loop.py:45 INFO TRAIN_STEP step=14925 loss=1.279353 Training: 83%|████████▎ | 14925/18018 [32:04:31<5:26:19, 6.33s/step, loss=1.279, acc=0.705, acc_len=1.70, thru=20.6, I=22.7, T=38.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:45,974] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:50,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:34:51,508] loop.py:45 INFO TRAIN_STEP step=14926 loss=1.293397 Training: 83%|████████▎ | 14926/18018 [32:04:37<5:15:46, 6.13s/step, loss=1.293, acc=0.705, acc_len=1.70, thru=21.3, I=24.1, T=39.6, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:34:51,504] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14925: forward=2308.6ms backward=876.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:51,697] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:56,783] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:34:56,846] loop.py:45 INFO TRAIN_STEP step=14927 loss=1.194994 Training: 83%|████████▎ | 14927/18018 [32:04:42<5:03:28, 5.89s/step, loss=1.195, acc=0.719, acc_len=1.74, thru=23.1, I=25.8, T=39.9, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:34:57,285] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:03,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:35:04,246] loop.py:45 INFO TRAIN_STEP step=14928 loss=1.316987 Training: 83%|████████▎ | 14928/18018 [32:04:50<5:26:42, 6.34s/step, loss=1.317, acc=0.698, acc_len=1.67, thru=18.1, I=19.6, T=38.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:04,337] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:08,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:35:10,154] loop.py:45 INFO TRAIN_STEP step=14929 loss=1.204567 Training: 83%|████████▎ | 14929/18018 [32:04:56<5:19:53, 6.21s/step, loss=1.205, acc=0.718, acc_len=1.74, thru=20.9, I=25.5, T=38.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:10,326] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:16,576] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:35:17,172] loop.py:380 INFO TIMING step=14930: step=3.421s data=1.407s compute=3.384s [fwd=2.401s bwd=0.969s opt=0.003s] dispatch=3.586s [2026-10-19 21:35:17,172] loop.py:45 INFO TRAIN_STEP step=14930 loss=1.267804 Training: 83%|████████▎ | 14930/18018 [32:05:03<5:32:12, 6.45s/step, loss=1.268, acc=0.702, acc_len=1.69, thru=18.2, I=23.0, T=37.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:17,580] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:23,179] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:35:23,920] loop.py:45 INFO TRAIN_STEP step=14931 loss=1.300130 Training: 83%|████████▎ | 14931/18018 [32:05:10<5:36:35, 6.54s/step, loss=1.300, acc=0.699, acc_len=1.67, thru=19.0, I=21.5, T=37.7, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:35:23,916] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14930: forward=2396.4ms backward=951.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:24,184] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:30,704] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:35:31,498] loop.py:45 INFO TRAIN_STEP step=14932 loss=1.166227 Training: 83%|████████▎ | 14932/18018 [32:05:17<5:52:29, 6.85s/step, loss=1.166, acc=0.721, acc_len=1.75, thru=17.0, I=19.6, T=31.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:31,709] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:37,917] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:35:38,995] loop.py:45 INFO TRAIN_STEP step=14933 loss=1.355500 Training: 83%|████████▎ | 14933/18018 [32:05:25<6:02:15, 7.05s/step, loss=1.356, acc=0.692, acc_len=1.65, thru=15.8, I=18.3, T=35.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:39,421] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:45,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:35:47,834] loop.py:45 INFO TRAIN_STEP step=14934 loss=1.143679 Training: 83%|████████▎ | 14934/18018 [32:05:33<6:29:48, 7.58s/step, loss=1.144, acc=0.729, acc_len=1.78, thru=13.7, I=18.6, T=24.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:48,300] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:53,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:35:54,405] loop.py:380 INFO TIMING step=14935: step=3.932s data=1.915s compute=3.904s [fwd=2.991s bwd=0.899s opt=0.003s] dispatch=2.631s [2026-10-19 21:35:54,405] loop.py:45 INFO TRAIN_STEP step=14935 loss=1.365453 Training: 83%|████████▎ | 14935/18018 [32:05:40<6:14:05, 7.28s/step, loss=1.365, acc=0.682, acc_len=1.62, thru=15.6, I=22.6, T=32.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:35:54,802] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:36:00,785] loop.py:45 INFO TRAIN_STEP step=14936 loss=1.272566 Training: 83%|████████▎ | 14936/18018 [32:05:46<6:00:08, 7.01s/step, loss=1.273, acc=0.705, acc_len=1.69, thru=15.8, I=19.4, T=39.1, wait=3.1s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:36:00,781] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14935: forward=2309.7ms backward=910.5ms [2026-10-19 21:36:07,505] loop.py:45 INFO TRAIN_STEP step=14937 loss=1.168357 Training: 83%|████████▎ | 14937/18018 [32:05:53<5:55:33, 6.92s/step, loss=1.168, acc=0.724, acc_len=1.77, thru=16.5, I=18.6, T=37.5, wait=3.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:36:15,857] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:36:16,270] loop.py:45 INFO TRAIN_STEP step=14938 loss=1.271856 Training: 83%|████████▎ | 14938/18018 [32:06:02<6:23:44, 7.48s/step, loss=1.272, acc=0.703, acc_len=1.69, thru=17.0, I=16.2, T=35.0, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:36:16,359] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 21:36:23,047] loop.py:45 INFO TRAIN_STEP step=14939 loss=1.114927 Training: 83%|████████▎ | 14939/18018 [32:06:09<6:12:49, 7.27s/step, loss=1.115, acc=0.736, acc_len=1.81, thru=17.9, I=18.8, T=38.3, wait=3.4s, pool=57, epoch=2/2][2026-10-19 21:36:29,839] loop.py:380 INFO TIMING step=14940: step=3.303s data=1.306s compute=3.270s [fwd=2.343s bwd=0.912s opt=0.003s] dispatch=3.481s [2026-10-19 21:36:29,839] loop.py:45 INFO TRAIN_STEP step=14940 loss=1.397138 Training: 83%|████████▎ | 14940/18018 [32:06:15<6:05:27, 7.12s/step, loss=1.397, acc=0.686, acc_len=1.63, thru=16.1, I=16.7, T=38.8, wait=3.5s, pool=25, epoch=2/2][2026-10-19 21:36:38,101] loop.py:45 INFO TRAIN_STEP step=14941 loss=1.356230 Training: 83%|████████▎ | 14941/18018 [32:06:24<6:22:56, 7.47s/step, loss=1.356, acc=0.691, acc_len=1.65, thru=18.4, I=15.6, T=37.6, wait=4.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:36:38,097] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14940: forward=2391.4ms backward=964.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:36:44,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:36:45,409] loop.py:45 INFO TRAIN_STEP step=14942 loss=1.168473 Training: 83%|████████▎ | 14942/18018 [32:06:31<6:20:16, 7.42s/step, loss=1.168, acc=0.722, acc_len=1.76, thru=18.6, I=18.8, T=38.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:36:45,813] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:36:53,935] loop.py:45 INFO TRAIN_STEP step=14943 loss=1.167480 Training: 83%|████████▎ | 14943/18018 [32:06:40<6:37:10, 7.75s/step, loss=1.167, acc=0.723, acc_len=1.76, thru=14.1, I=15.6, T=35.8, wait=4.9s, pool=49, epoch=2/2][2026-10-19 21:37:01,743] loop.py:45 INFO TRAIN_STEP step=14944 loss=1.277008 Training: 83%|████████▎ | 14944/18018 [32:06:47<6:38:01, 7.77s/step, loss=1.277, acc=0.703, acc_len=1.70, thru=17.8, I=17.2, T=37.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:01,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:02,253] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:08,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:37:09,767] loop.py:380 INFO TIMING step=14945: step=3.409s data=1.368s compute=3.374s [fwd=2.422s bwd=0.937s opt=0.003s] dispatch=4.601s [2026-10-19 21:37:09,768] loop.py:45 INFO TRAIN_STEP step=14945 loss=1.208559 Training: 83%|████████▎ | 14945/18018 [32:06:55<6:41:43, 7.84s/step, loss=1.209, acc=0.718, acc_len=1.74, thru=16.9, I=19.4, T=37.5, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:09,983] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:15,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 21:37:15,977] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14945: forward=2358.6ms backward=884.9ms [2026-10-19 21:37:15,982] loop.py:45 INFO TRAIN_STEP step=14946 loss=1.287029 Training: 83%|████████▎ | 14946/18018 [32:07:02<6:16:33, 7.35s/step, loss=1.287, acc=0.703, acc_len=1.69, thru=20.3, I=22.9, T=39.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:16,149] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:22,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:37:23,928] loop.py:45 INFO TRAIN_STEP step=14947 loss=1.238447 Training: 83%|████████▎ | 14947/18018 [32:07:10<6:25:32, 7.53s/step, loss=1.238, acc=0.709, acc_len=1.71, thru=16.1, I=18.8, T=36.6, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:24,082] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:37:29,558] loop.py:45 INFO TRAIN_STEP step=14948 loss=1.180674 Training: 83%|████████▎ | 14948/18018 [32:07:15<5:56:12, 6.96s/step, loss=1.181, acc=0.720, acc_len=1.75, thru=21.4, I=21.3, T=39.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:29,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:30,062] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:37:37,016] loop.py:45 INFO TRAIN_STEP step=14949 loss=1.340676 Training: 83%|████████▎ | 14949/18018 [32:07:23<6:03:44, 7.11s/step, loss=1.341, acc=0.692, acc_len=1.65, thru=17.2, I=18.3, T=37.2, wait=4.0s, pool=53, epoch=2/2][2026-10-19 21:37:43,198] loop.py:380 INFO TIMING step=14950: step=3.435s data=1.335s compute=3.410s [fwd=2.430s bwd=0.966s opt=0.003s] dispatch=2.738s [2026-10-19 21:37:43,198] loop.py:45 INFO TRAIN_STEP step=14950 loss=1.222420 Training: 83%|████████▎ | 14950/18018 [32:07:29<5:49:20, 6.83s/step, loss=1.222, acc=0.714, acc_len=1.72, thru=19.0, I=20.5, T=37.3, wait=2.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:51,742] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:37:52,790] loop.py:45 INFO TRAIN_STEP step=14951 loss=1.180879 Training: 83%|████████▎ | 14951/18018 [32:07:38<6:31:32, 7.66s/step, loss=1.181, acc=0.722, acc_len=1.76, thru=14.8, I=14.4, T=31.7, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:37:52,785] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14950: forward=2915.2ms backward=1074.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:53,249] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:37:59,855] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:38:00,504] loop.py:45 INFO TRAIN_STEP step=14952 loss=1.237014 Training: 83%|████████▎ | 14952/18018 [32:07:46<6:32:15, 7.68s/step, loss=1.237, acc=0.711, acc_len=1.72, thru=15.3, I=17.9, T=34.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:00,859] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:05,852] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:38:07,089] loop.py:45 INFO TRAIN_STEP step=14953 loss=1.213319 Training: 83%|████████▎ | 14953/18018 [32:07:53<6:15:25, 7.35s/step, loss=1.213, acc=0.717, acc_len=1.73, thru=19.0, I=23.1, T=34.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:07,361] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:14,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:38:14,935] loop.py:45 INFO TRAIN_STEP step=14954 loss=1.249836 Training: 83%|████████▎ | 14954/18018 [32:08:01<6:22:52, 7.50s/step, loss=1.250, acc=0.710, acc_len=1.71, thru=15.0, I=18.0, T=39.0, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:15,340] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:20,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:38:20,954] loop.py:380 INFO TIMING step=14955: step=3.382s data=1.351s compute=3.354s [fwd=2.406s bwd=0.935s opt=0.003s] dispatch=2.630s [2026-10-19 21:38:20,954] loop.py:45 INFO TRAIN_STEP step=14955 loss=1.231116 Training: 83%|████████▎ | 14955/18018 [32:08:07<6:00:07, 7.05s/step, loss=1.231, acc=0.709, acc_len=1.71, thru=20.3, I=19.9, T=37.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:21,184] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:38:27,748] loop.py:45 INFO TRAIN_STEP step=14956 loss=1.148624 Training: 83%|████████▎ | 14956/18018 [32:08:13<5:55:59, 6.98s/step, loss=1.149, acc=0.725, acc_len=1.76, thru=16.4, I=17.2, T=38.1, wait=3.4s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:38:27,743] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14955: forward=2410.3ms backward=900.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:34,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:38:34,961] loop.py:45 INFO TRAIN_STEP step=14957 loss=1.148808 Training: 83%|████████▎ | 14957/18018 [32:08:21<5:59:31, 7.05s/step, loss=1.149, acc=0.729, acc_len=1.78, thru=18.9, I=18.3, T=37.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:35,339] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:38:43,583] loop.py:45 INFO TRAIN_STEP step=14958 loss=1.290757 Training: 83%|████████▎ | 14958/18018 [32:08:29<6:23:32, 7.52s/step, loss=1.291, acc=0.699, acc_len=1.68, thru=14.6, I=16.0, T=34.8, wait=4.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:50,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:38:51,056] loop.py:45 INFO TRAIN_STEP step=14959 loss=1.286907 Training: 83%|████████▎ | 14959/18018 [32:08:37<6:22:39, 7.51s/step, loss=1.287, acc=0.702, acc_len=1.68, thru=18.0, I=18.8, T=35.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:38:51,300] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:38:56,656] loop.py:380 INFO TIMING step=14960: step=3.321s data=1.302s compute=3.284s [fwd=2.368s bwd=0.902s opt=0.003s] dispatch=2.271s [2026-10-19 21:38:56,657] loop.py:45 INFO TRAIN_STEP step=14960 loss=1.141079 Training: 83%|████████▎ | 14960/18018 [32:08:42<5:53:24, 6.93s/step, loss=1.141, acc=0.725, acc_len=1.77, thru=20.1, I=22.8, T=38.5, wait=2.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:39:01,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:39:02,522] loop.py:45 INFO TRAIN_STEP step=14961 loss=1.239691 Training: 83%|████████▎ | 14961/18018 [32:08:48<5:36:59, 6.61s/step, loss=1.240, acc=0.710, acc_len=1.71, thru=21.9, I=23.8, T=39.2, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:39:02,517] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14960: forward=2315.6ms backward=902.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:39:02,641] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 21:39:09,340] loop.py:45 INFO TRAIN_STEP step=14962 loss=1.336854 Training: 83%|████████▎ | 14962/18018 [32:08:55<5:39:58, 6.67s/step, loss=1.337, acc=0.699, acc_len=1.68, thru=18.4, I=22.5, T=37.9, wait=3.4s, pool=57, epoch=2/2][2026-10-19 21:39:17,394] loop.py:45 INFO TRAIN_STEP step=14963 loss=1.350180 Training: 83%|████████▎ | 14963/18018 [32:09:03<6:00:55, 7.09s/step, loss=1.350, acc=0.690, acc_len=1.64, thru=15.0, I=14.7, T=36.8, wait=4.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:39:25,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:39:26,403] loop.py:45 INFO TRAIN_STEP step=14964 loss=1.227810 Training: 83%|████████▎ | 14964/18018 [32:09:12<6:30:09, 7.67s/step, loss=1.228, acc=0.712, acc_len=1.72, thru=15.8, I=18.0, T=36.1, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:39:26,489] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 21:39:33,589] loop.py:380 INFO TIMING step=14965: step=3.312s data=1.339s compute=3.276s [fwd=2.334s bwd=0.928s opt=0.003s] dispatch=3.865s [2026-10-19 21:39:33,589] loop.py:45 INFO TRAIN_STEP step=14965 loss=1.126899 Training: 83%|████████▎ | 14965/18018 [32:09:19<6:22:42, 7.52s/step, loss=1.127, acc=0.733, acc_len=1.79, thru=17.7, I=18.0, T=38.7, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:39:44,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:39:44,548] loop.py:45 INFO TRAIN_STEP step=14966 loss=1.234398 Training: 83%|████████▎ | 14966/18018 [32:09:30<7:15:02, 8.55s/step, loss=1.234, acc=0.711, acc_len=1.72, thru=12.1, I=13.2, T=32.9, wait=7.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:39:44,544] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14965: forward=2676.6ms backward=1159.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:39:45,050] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:39:50,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:39:50,611] loop.py:45 INFO TRAIN_STEP step=14967 loss=1.270784 Training: 83%|████████▎ | 14967/18018 [32:09:36<6:36:53, 7.81s/step, loss=1.271, acc=0.701, acc_len=1.69, thru=19.6, I=19.0, T=38.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:39:50,764] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:39:57,331] loop.py:45 INFO TRAIN_STEP step=14968 loss=1.163861 Training: 83%|████████▎ | 14968/18018 [32:09:43<6:20:13, 7.48s/step, loss=1.164, acc=0.723, acc_len=1.76, thru=18.6, I=21.5, T=38.1, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:04,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:40:04,613] loop.py:45 INFO TRAIN_STEP step=14969 loss=1.217987 Training: 83%|████████▎ | 14969/18018 [32:09:50<6:17:04, 7.42s/step, loss=1.218, acc=0.714, acc_len=1.72, thru=17.4, I=17.8, T=36.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:04,800] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:08,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:40:09,802] loop.py:380 INFO TIMING step=14970: step=3.388s data=1.410s compute=3.355s [fwd=2.405s bwd=0.936s opt=0.003s] dispatch=1.792s [2026-10-19 21:40:09,802] loop.py:45 INFO TRAIN_STEP step=14970 loss=1.119097 Training: 83%|████████▎ | 14970/18018 [32:09:55<5:42:59, 6.75s/step, loss=1.119, acc=0.729, acc_len=1.78, thru=21.5, I=24.9, T=37.8, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:10,232] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:17,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 21:40:18,046] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14970: forward=2360.1ms backward=990.6ms [2026-10-19 21:40:18,052] loop.py:45 INFO TRAIN_STEP step=14971 loss=1.323444 Training: 83%|████████▎ | 14971/18018 [32:10:04<6:05:40, 7.20s/step, loss=1.323, acc=0.697, acc_len=1.67, thru=15.0, I=16.7, T=37.5, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:18,270] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:40:26,230] loop.py:45 INFO TRAIN_STEP step=14972 loss=1.234164 Training: 83%|████████▎ | 14972/18018 [32:10:12<6:20:27, 7.49s/step, loss=1.234, acc=0.709, acc_len=1.70, thru=16.4, I=16.2, T=34.4, wait=4.4s, pool=57, epoch=2/2][2026-10-19 21:40:32,866] loop.py:45 INFO TRAIN_STEP step=14973 loss=1.256457 Training: 83%|████████▎ | 14973/18018 [32:10:18<6:07:14, 7.24s/step, loss=1.256, acc=0.711, acc_len=1.72, thru=18.8, I=19.8, T=38.8, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:39,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:40:39,588] loop.py:45 INFO TRAIN_STEP step=14974 loss=1.167553 Training: 83%|████████▎ | 14974/18018 [32:10:25<5:59:18, 7.08s/step, loss=1.168, acc=0.722, acc_len=1.76, thru=19.4, I=20.8, T=37.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:39,786] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:47,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:40:48,140] loop.py:380 INFO TIMING step=14975: step=3.507s data=1.412s compute=3.481s [fwd=2.472s bwd=0.995s opt=0.003s] dispatch=5.035s [2026-10-19 21:40:48,140] loop.py:45 INFO TRAIN_STEP step=14975 loss=1.238304 Training: 83%|████████▎ | 14975/18018 [32:10:34<6:21:33, 7.52s/step, loss=1.238, acc=0.712, acc_len=1.72, thru=14.8, I=15.8, T=36.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:40:48,304] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:40:54,336] loop.py:45 INFO TRAIN_STEP step=14976 loss=1.173568 Training: 83%|████████▎ | 14976/18018 [32:10:40<6:01:15, 7.13s/step, loss=1.174, acc=0.721, acc_len=1.75, thru=16.8, I=18.7, T=38.1, wait=2.8s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:40:54,329] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14975: forward=2388.0ms backward=923.7ms [2026-10-19 21:41:04,423] loop.py:45 INFO TRAIN_STEP step=14977 loss=1.272117 Training: 83%|████████▎ | 14977/18018 [32:10:50<6:46:08, 8.01s/step, loss=1.272, acc=0.706, acc_len=1.70, thru=15.7, I=13.1, T=35.0, wait=6.4s, pool=61, epoch=2/2][2026-10-19 21:41:11,575] loop.py:45 INFO TRAIN_STEP step=14978 loss=1.283621 Training: 83%|████████▎ | 14978/18018 [32:10:57<6:32:56, 7.76s/step, loss=1.284, acc=0.704, acc_len=1.69, thru=18.0, I=16.8, T=37.0, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:41:19,107] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:41:19,545] loop.py:45 INFO TRAIN_STEP step=14979 loss=1.232326 Training: 83%|████████▎ | 14979/18018 [32:11:05<6:36:02, 7.82s/step, loss=1.232, acc=0.712, acc_len=1.71, thru=17.4, I=18.4, T=36.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:41:19,610] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 21:41:26,352] loop.py:380 INFO TIMING step=14980: step=3.332s data=1.362s compute=3.301s [fwd=2.365s bwd=0.922s opt=0.003s] dispatch=3.464s [2026-10-19 21:41:26,352] loop.py:45 INFO TRAIN_STEP step=14980 loss=1.262549 Training: 83%|████████▎ | 14980/18018 [32:11:12<6:20:34, 7.52s/step, loss=1.263, acc=0.701, acc_len=1.68, thru=17.6, I=18.1, T=38.4, wait=3.5s, pool=57, epoch=2/2][2026-10-19 21:41:33,490] loop.py:45 INFO TRAIN_STEP step=14981 loss=1.282191 Training: 83%|████████▎ | 14981/18018 [32:11:19<6:14:41, 7.40s/step, loss=1.282, acc=0.702, acc_len=1.69, thru=15.5, I=16.9, T=38.0, wait=3.8s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:41:33,485] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14980: forward=2407.0ms backward=915.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:41:42,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:41:44,008] loop.py:45 INFO TRAIN_STEP step=14982 loss=1.315304 Training: 83%|████████▎ | 14982/18018 [32:11:30<7:01:50, 8.34s/step, loss=1.315, acc=0.695, acc_len=1.66, thru=15.8, I=17.0, T=34.9, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:41:44,416] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 21:41:50,214] loop.py:45 INFO TRAIN_STEP step=14983 loss=1.174413 Training: 83%|████████▎ | 14983/18018 [32:11:36<6:29:21, 7.70s/step, loss=1.174, acc=0.725, acc_len=1.77, thru=19.8, I=24.8, T=38.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:41:50,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:41:50,679] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:41:56,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:41:57,899] loop.py:45 INFO TRAIN_STEP step=14984 loss=1.323923 Training: 83%|████████▎ | 14984/18018 [32:11:43<6:29:04, 7.69s/step, loss=1.324, acc=0.693, acc_len=1.66, thru=17.8, I=21.0, T=36.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:41:58,015] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 21:42:03,703] loop.py:380 INFO TIMING step=14985: step=3.309s data=1.268s compute=3.284s [fwd=2.386s bwd=0.883s opt=0.003s] dispatch=2.484s [2026-10-19 21:42:03,703] loop.py:45 INFO TRAIN_STEP step=14985 loss=1.262118 Training: 83%|████████▎ | 14985/18018 [32:11:49<6:00:14, 7.13s/step, loss=1.262, acc=0.704, acc_len=1.69, thru=19.2, I=23.1, T=38.7, wait=2.5s, pool=53, epoch=2/2][2026-10-19 21:42:11,768] loop.py:45 INFO TRAIN_STEP step=14986 loss=1.306079 Training: 83%|████████▎ | 14986/18018 [32:11:57<6:14:22, 7.41s/step, loss=1.306, acc=0.695, acc_len=1.66, thru=16.4, I=17.0, T=35.7, wait=4.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:42:11,763] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14985: forward=2559.3ms backward=975.7ms [2026-10-19 21:42:18,854] loop.py:45 INFO TRAIN_STEP step=14987 loss=1.158714 Training: 83%|████████▎ | 14987/18018 [32:12:04<6:09:21, 7.31s/step, loss=1.159, acc=0.729, acc_len=1.79, thru=18.3, I=18.2, T=38.8, wait=3.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:42:24,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:42:25,088] loop.py:45 INFO TRAIN_STEP step=14988 loss=1.253743 Training: 83%|████████▎ | 14988/18018 [32:12:11<5:52:54, 6.99s/step, loss=1.254, acc=0.706, acc_len=1.70, thru=20.5, I=21.5, T=38.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:42:25,454] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:42:31,923] loop.py:45 INFO TRAIN_STEP step=14989 loss=1.366582 Training: 83%|████████▎ | 14989/18018 [32:12:18<5:50:27, 6.94s/step, loss=1.367, acc=0.689, acc_len=1.64, thru=16.8, I=19.5, T=36.7, wait=3.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:42:39,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:42:40,627] loop.py:380 INFO TIMING step=14990: step=3.517s data=1.447s compute=3.481s [fwd=2.484s bwd=0.982s opt=0.003s] dispatch=5.178s [2026-10-19 21:42:40,627] loop.py:45 INFO TRAIN_STEP step=14990 loss=1.349490 Training: 83%|████████▎ | 14990/18018 [32:12:26<6:17:01, 7.47s/step, loss=1.349, acc=0.689, acc_len=1.65, thru=17.1, I=17.7, T=36.4, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:42:41,040] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 21:42:46,544] loop.py:45 INFO TRAIN_STEP step=14991 loss=1.170703 Training: 83%|████████▎ | 14991/18018 [32:12:32<5:53:23, 7.00s/step, loss=1.171, acc=0.725, acc_len=1.76, thru=18.0, I=20.5, T=40.1, wait=2.7s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:42:46,539] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14990: forward=2256.8ms backward=888.4ms [2026-10-19 21:42:53,438] loop.py:45 INFO TRAIN_STEP step=14992 loss=1.205987 Training: 83%|████████▎ | 14992/18018 [32:12:39<5:51:34, 6.97s/step, loss=1.206, acc=0.716, acc_len=1.73, thru=18.8, I=19.1, T=38.4, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:04,207] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:43:04,557] loop.py:45 INFO TRAIN_STEP step=14993 loss=1.306940 Training: 83%|████████▎ | 14993/18018 [32:12:50<6:54:12, 8.22s/step, loss=1.307, acc=0.700, acc_len=1.68, thru=13.3, I=12.3, T=35.6, wait=7.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:04,714] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:09,974] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:43:10,715] loop.py:45 INFO TRAIN_STEP step=14994 loss=1.304593 Training: 83%|████████▎ | 14994/18018 [32:12:56<6:22:59, 7.60s/step, loss=1.305, acc=0.695, acc_len=1.66, thru=20.6, I=22.2, T=38.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:10,978] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:15,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:43:16,698] loop.py:380 INFO TIMING step=14995: step=3.316s data=1.339s compute=3.291s [fwd=2.371s bwd=0.906s opt=0.003s] dispatch=2.657s [2026-10-19 21:43:16,699] loop.py:45 INFO TRAIN_STEP step=14995 loss=1.178605 Training: 83%|████████▎ | 14995/18018 [32:13:02<5:58:26, 7.11s/step, loss=1.179, acc=0.721, acc_len=1.75, thru=20.8, I=24.8, T=38.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:16,784] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 21:43:23,087] loop.py:45 INFO TRAIN_STEP step=14996 loss=1.174073 Training: 83%|████████▎ | 14996/18018 [32:13:09<5:47:19, 6.90s/step, loss=1.174, acc=0.722, acc_len=1.76, thru=19.2, I=20.9, T=38.3, wait=3.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:43:23,083] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=14995: forward=2390.3ms backward=912.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:30,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:43:31,050] loop.py:45 INFO TRAIN_STEP step=14997 loss=1.257183 Training: 83%|████████▎ | 14997/18018 [32:13:17<6:03:19, 7.22s/step, loss=1.257, acc=0.707, acc_len=1.70, thru=17.4, I=16.8, T=36.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:31,379] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:43:37,586] loop.py:45 INFO TRAIN_STEP step=14998 loss=1.266532 Training: 83%|████████▎ | 14998/18018 [32:13:23<5:52:55, 7.01s/step, loss=1.267, acc=0.703, acc_len=1.70, thru=18.6, I=19.5, T=38.1, wait=3.2s, pool=57, epoch=2/2][2026-10-19 21:43:44,874] loop.py:45 INFO TRAIN_STEP step=14999 loss=1.282307 Training: 83%|████████▎ | 14999/18018 [32:13:30<5:57:00, 7.10s/step, loss=1.282, acc=0.706, acc_len=1.70, thru=17.3, I=18.3, T=36.8, wait=3.8s, pool=57, epoch=2/2][2026-10-19 21:43:53,067] loop.py:380 INFO TIMING step=15000: step=3.708s data=1.587s compute=3.680s [fwd=2.670s bwd=0.995s opt=0.003s] dispatch=4.476s [2026-10-19 21:43:53,067] loop.py:45 INFO TRAIN_STEP step=15000 loss=1.198967 Training: 83%|████████▎ | 15000/18018 [32:13:39<6:13:29, 7.43s/step, loss=1.199, acc=0.718, acc_len=1.75, thru=13.5, I=13.7, T=34.5, wait=4.5s, pool=37, epoch=2/2][2026-10-19 21:43:53,074] loop.py:419 INFO Saving checkpoint at step 15000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:53,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 21:43:56,488] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0015001 (TrainerActor pid=9310) [2026-10-19 21:43:56,586] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0014251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:43:57,974] inference_manager.py:408 INFO Pool capacity available after 4.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:44:02,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 21:43:56,487] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0015001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 21:43:57,488] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0014251 [2026-10-19 21:44:04,640] loop.py:45 INFO TRAIN_STEP step=15001 loss=1.180664 Training: 83%|████████▎ | 15001/18018 [32:13:50<7:15:54, 8.67s/step, loss=1.181, acc=0.719, acc_len=1.74, thru=19.2, I=26.0, T=35.4, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:44:04,635] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15000: forward=2603.6ms backward=957.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:44:04,707] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=41/64 [2026-10-19 21:44:09,987] loop.py:45 INFO TRAIN_STEP step=15002 loss=1.245345 Training: 83%|████████▎ | 15002/18018 [32:13:56<6:25:42, 7.67s/step, loss=1.245, acc=0.708, acc_len=1.70, thru=16.9, I=23.0, T=39.7, wait=2.1s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 21:44:10,401] eagle_store.py:517 WARNING batch_get_buffer missing keys (k480070-9601e8dc_hs, k480070-9601e8dc_ids, k480070-9601e8dc_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:44:18,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 21:44:10,338] eagle_store.py:517 WARNING batch_get_buffer missing keys (k480069-b410c91b_hs, k480069-b410c91b_ids, k480069-b410c91b_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 21:44:19,161] loop.py:45 INFO TRAIN_STEP step=15003 loss=1.398877 Training: 83%|████████▎ | 15003/18018 [32:14:05<6:48:08, 8.12s/step, loss=1.399, acc=0.683, acc_len=1.63, thru=18.1, I=18.3, T=36.6, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:44:19,609] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:44:25,628] loop.py:45 INFO TRAIN_STEP step=15004 loss=1.211726 Training: 83%|████████▎ | 15004/18018 [32:14:11<6:23:03, 7.63s/step, loss=1.212, acc=0.714, acc_len=1.73, thru=17.0, I=21.8, T=38.6, wait=3.1s, pool=41, epoch=2/2][2026-10-19 21:44:36,344] loop.py:380 INFO TIMING step=15005: step=3.935s data=1.557s compute=3.893s [fwd=2.840s bwd=1.038s opt=0.003s] dispatch=6.773s [2026-10-19 21:44:36,344] loop.py:45 INFO TRAIN_STEP step=15005 loss=1.259664 Training: 83%|████████▎ | 15005/18018 [32:14:22<7:09:29, 8.55s/step, loss=1.260, acc=0.710, acc_len=1.71, thru=13.7, I=12.1, T=32.5, wait=6.8s, pool=53, epoch=2/2][2026-10-19 21:44:44,607] loop.py:45 INFO TRAIN_STEP step=15006 loss=1.235887 Training: 83%|████████▎ | 15006/18018 [32:14:30<7:04:58, 8.47s/step, loss=1.236, acc=0.715, acc_len=1.73, thru=14.9, I=13.8, T=35.7, wait=4.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:44:44,603] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15005: forward=2555.4ms backward=981.0ms [2026-10-19 21:44:52,195] loop.py:45 INFO TRAIN_STEP step=15007 loss=1.079232 Training: 83%|████████▎ | 15007/18018 [32:14:38<6:51:40, 8.20s/step, loss=1.079, acc=0.738, acc_len=1.82, thru=15.7, I=16.9, T=35.5, wait=4.0s, pool=45, epoch=2/2][2026-10-19 21:45:00,564] loop.py:45 INFO TRAIN_STEP step=15008 loss=1.312837 Training: 83%|████████▎ | 15008/18018 [32:14:46<6:53:58, 8.25s/step, loss=1.313, acc=0.698, acc_len=1.67, thru=16.5, I=15.7, T=35.5, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:00,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:01,055] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:08,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:45:08,472] loop.py:45 INFO TRAIN_STEP step=15009 loss=1.227719 Training: 83%|████████▎ | 15009/18018 [32:14:54<6:48:41, 8.15s/step, loss=1.228, acc=0.713, acc_len=1.72, thru=17.5, I=16.4, T=37.8, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:08,644] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:45:14,802] loop.py:380 INFO TIMING step=15010: step=3.296s data=1.294s compute=3.261s [fwd=2.356s bwd=0.890s opt=0.003s] dispatch=3.026s [2026-10-19 21:45:14,802] loop.py:45 INFO TRAIN_STEP step=15010 loss=1.395643 Training: 83%|████████▎ | 15010/18018 [32:15:00<6:21:11, 7.60s/step, loss=1.396, acc=0.684, acc_len=1.63, thru=19.3, I=17.0, T=38.8, wait=3.0s, pool=57, epoch=2/2][2026-10-19 21:45:22,574] loop.py:45 INFO TRAIN_STEP step=15011 loss=1.191025 Training: 83%|████████▎ | 15011/18018 [32:15:08<6:23:38, 7.65s/step, loss=1.191, acc=0.718, acc_len=1.74, thru=16.9, I=17.1, T=36.4, wait=4.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:45:22,570] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15010: forward=2408.0ms backward=1037.5ms [2026-10-19 21:45:29,044] loop.py:45 INFO TRAIN_STEP step=15012 loss=1.242475 Training: 83%|████████▎ | 15012/18018 [32:15:15<6:05:38, 7.30s/step, loss=1.242, acc=0.712, acc_len=1.72, thru=17.7, I=17.7, T=38.4, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:35,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:45:35,830] loop.py:45 INFO TRAIN_STEP step=15013 loss=1.354827 Training: 83%|████████▎ | 15013/18018 [32:15:21<5:57:52, 7.15s/step, loss=1.355, acc=0.690, acc_len=1.65, thru=19.5, I=19.7, T=36.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:35,869] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:43,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:45:44,167] loop.py:45 INFO TRAIN_STEP step=15014 loss=1.310553 Training: 83%|████████▎ | 15014/18018 [32:15:30<6:15:36, 7.50s/step, loss=1.311, acc=0.696, acc_len=1.67, thru=16.1, I=17.8, T=36.7, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:44,654] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:50,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:45:50,818] loop.py:380 INFO TIMING step=15015: step=3.291s data=1.323s compute=3.264s [fwd=2.374s bwd=0.875s opt=0.003s] dispatch=3.351s [2026-10-19 21:45:50,818] loop.py:45 INFO TRAIN_STEP step=15015 loss=1.302492 Training: 83%|████████▎ | 15015/18018 [32:15:36<6:02:43, 7.25s/step, loss=1.302, acc=0.699, acc_len=1.68, thru=19.8, I=24.6, T=38.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:51,030] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:57,461] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:45:57,852] loop.py:45 INFO TRAIN_STEP step=15016 loss=1.207182 Training: 83%|████████▎ | 15016/18018 [32:15:43<5:59:25, 7.18s/step, loss=1.207, acc=0.716, acc_len=1.73, thru=18.5, I=21.3, T=38.2, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:45:57,848] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15015: forward=2388.1ms backward=911.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:45:57,963] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 21:46:03,139] loop.py:45 INFO TRAIN_STEP step=15017 loss=1.239224 Training: 83%|████████▎ | 15017/18018 [32:15:49<5:30:48, 6.61s/step, loss=1.239, acc=0.711, acc_len=1.72, thru=18.5, I=22.3, T=39.0, wait=2.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:46:11,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:46:11,561] loop.py:45 INFO TRAIN_STEP step=15018 loss=1.257908 Training: 83%|████████▎ | 15018/18018 [32:15:57<5:57:48, 7.16s/step, loss=1.258, acc=0.707, acc_len=1.70, thru=16.8, I=16.2, T=36.4, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:46:11,713] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:46:16,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:46:18,106] loop.py:45 INFO TRAIN_STEP step=15019 loss=1.373006 Training: 83%|████████▎ | 15019/18018 [32:16:04<5:48:33, 6.97s/step, loss=1.373, acc=0.687, acc_len=1.63, thru=19.4, I=22.8, T=37.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:46:18,244] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 21:46:24,753] loop.py:380 INFO TIMING step=15020: step=3.275s data=1.347s compute=3.247s [fwd=2.336s bwd=0.896s opt=0.003s] dispatch=3.362s [2026-10-19 21:46:24,753] loop.py:45 INFO TRAIN_STEP step=15020 loss=1.342397 Training: 83%|████████▎ | 15020/18018 [32:16:10<5:43:30, 6.87s/step, loss=1.342, acc=0.690, acc_len=1.65, thru=16.0, I=21.9, T=39.1, wait=3.4s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:46:32,615] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:46:33,111] loop.py:45 INFO TRAIN_STEP step=15021 loss=1.253800 Training: 83%|████████▎ | 15021/18018 [32:16:19<6:05:38, 7.32s/step, loss=1.254, acc=0.706, acc_len=1.70, thru=18.5, I=16.9, T=36.6, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:46:33,106] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15020: forward=2476.1ms backward=977.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:46:33,144] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:46:38,741] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:46:40,185] loop.py:45 INFO TRAIN_STEP step=15022 loss=1.209764 Training: 83%|████████▎ | 15022/18018 [32:16:26<6:01:49, 7.25s/step, loss=1.210, acc=0.715, acc_len=1.73, thru=18.4, I=21.7, T=37.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:46:40,250] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 21:46:47,021] loop.py:45 INFO TRAIN_STEP step=15023 loss=1.270797 Training: 83%|████████▎ | 15023/18018 [32:16:33<5:55:33, 7.12s/step, loss=1.271, acc=0.705, acc_len=1.70, thru=17.4, I=21.2, T=37.6, wait=3.4s, pool=49, epoch=2/2][2026-10-19 21:46:53,872] loop.py:45 INFO TRAIN_STEP step=15024 loss=1.260667 Training: 83%|████████▎ | 15024/18018 [32:16:39<5:51:22, 7.04s/step, loss=1.261, acc=0.707, acc_len=1.71, thru=18.9, I=18.7, T=37.5, wait=3.4s, pool=61, epoch=2/2][2026-10-19 21:47:02,198] loop.py:380 INFO TIMING step=15025: step=3.708s data=1.313s compute=3.676s [fwd=2.696s bwd=0.963s opt=0.003s] dispatch=4.608s [2026-10-19 21:47:02,198] loop.py:45 INFO TRAIN_STEP step=15025 loss=1.387246 Training: 83%|████████▎ | 15025/18018 [32:16:48<6:10:30, 7.43s/step, loss=1.387, acc=0.683, acc_len=1.62, thru=14.1, I=15.0, T=34.5, wait=4.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:47:10,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:47:10,724] loop.py:45 INFO TRAIN_STEP step=15026 loss=1.199382 Training: 83%|████████▎ | 15026/18018 [32:16:56<6:26:46, 7.76s/step, loss=1.199, acc=0.716, acc_len=1.74, thru=17.3, I=16.1, T=35.5, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:47:10,720] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15025: forward=2523.3ms backward=1020.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:47:10,862] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:47:16,675] loop.py:45 INFO TRAIN_STEP step=15027 loss=1.315548 Training: 83%|████████▎ | 15027/18018 [32:17:02<5:59:42, 7.22s/step, loss=1.316, acc=0.697, acc_len=1.67, thru=18.8, I=20.4, T=37.1, wait=2.5s, pool=49, epoch=2/2][2026-10-19 21:47:26,464] loop.py:45 INFO TRAIN_STEP step=15028 loss=1.303967 Training: 83%|████████▎ | 15028/18018 [32:17:12<6:37:59, 7.99s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=14.4, I=13.6, T=36.3, wait=6.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:47:33,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:47:33,932] loop.py:45 INFO TRAIN_STEP step=15029 loss=1.246078 Training: 83%|████████▎ | 15029/18018 [32:17:20<6:30:06, 7.83s/step, loss=1.246, acc=0.709, acc_len=1.71, thru=19.8, I=17.3, T=36.1, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:47:34,403] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:47:40,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:47:41,786] loop.py:380 INFO TIMING step=15030: step=3.466s data=1.397s compute=3.419s [fwd=2.439s bwd=0.965s opt=0.003s] dispatch=4.380s [2026-10-19 21:47:41,786] loop.py:45 INFO TRAIN_STEP step=15030 loss=1.235220 Training: 83%|████████▎ | 15030/18018 [32:17:27<6:30:22, 7.84s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=17.3, I=19.2, T=36.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:47:41,931] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:47:47,335] loop.py:45 INFO TRAIN_STEP step=15031 loss=1.381673 Training: 83%|████████▎ | 15031/18018 [32:17:33<5:56:02, 7.15s/step, loss=1.382, acc=0.684, acc_len=1.62, thru=20.9, I=24.1, T=39.0, wait=2.3s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:47:47,329] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15030: forward=2328.3ms backward=901.7ms [2026-10-19 21:47:54,761] loop.py:45 INFO TRAIN_STEP step=15032 loss=1.244833 Training: 83%|████████▎ | 15032/18018 [32:17:40<5:59:58, 7.23s/step, loss=1.245, acc=0.708, acc_len=1.71, thru=16.1, I=16.5, T=37.6, wait=4.0s, pool=49, epoch=2/2][2026-10-19 21:48:01,906] loop.py:45 INFO TRAIN_STEP step=15033 loss=1.253054 Training: 83%|████████▎ | 15033/18018 [32:17:47<5:58:34, 7.21s/step, loss=1.253, acc=0.709, acc_len=1.72, thru=14.9, I=17.2, T=35.4, wait=3.5s, pool=33, epoch=2/2][2026-10-19 21:48:10,491] loop.py:45 INFO TRAIN_STEP step=15034 loss=1.215543 Training: 83%|████████▎ | 15034/18018 [32:17:56<6:19:02, 7.62s/step, loss=1.216, acc=0.715, acc_len=1.73, thru=16.7, I=15.1, T=35.1, wait=4.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:48:16,683] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:48:17,233] loop.py:380 INFO TIMING step=15035: step=4.037s data=1.342s compute=3.995s [fwd=3.010s bwd=0.970s opt=0.003s] dispatch=2.693s [2026-10-19 21:48:17,234] loop.py:45 INFO TRAIN_STEP step=15035 loss=1.314268 Training: 83%|████████▎ | 15035/18018 [32:18:03<6:05:46, 7.36s/step, loss=1.314, acc=0.697, acc_len=1.67, thru=18.8, I=19.6, T=31.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:48:17,686] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:48:23,565] loop.py:45 INFO TRAIN_STEP step=15036 loss=1.303365 Training: 83%|████████▎ | 15036/18018 [32:18:09<5:50:20, 7.05s/step, loss=1.303, acc=0.697, acc_len=1.67, thru=17.0, I=20.8, T=38.6, wait=3.0s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:48:23,560] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15035: forward=2369.4ms backward=905.8ms [2026-10-19 21:48:33,543] loop.py:45 INFO TRAIN_STEP step=15037 loss=1.206989 Training: 83%|████████▎ | 15037/18018 [32:18:19<6:33:53, 7.93s/step, loss=1.207, acc=0.718, acc_len=1.74, thru=13.3, I=14.1, T=36.0, wait=6.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:48:42,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:48:42,558] loop.py:45 INFO TRAIN_STEP step=15038 loss=1.278107 Training: 83%|████████▎ | 15038/18018 [32:18:28<6:49:57, 8.25s/step, loss=1.278, acc=0.701, acc_len=1.68, thru=15.1, I=14.9, T=34.6, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:48:42,643] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:48:48,022] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:48:49,025] loop.py:45 INFO TRAIN_STEP step=15039 loss=1.143545 Training: 83%|████████▎ | 15039/18018 [32:18:35<6:23:09, 7.72s/step, loss=1.144, acc=0.725, acc_len=1.77, thru=19.4, I=20.2, T=36.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:48:49,529] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:48:58,044] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:48:58,333] loop.py:380 INFO TIMING step=15040: step=3.496s data=1.430s compute=3.465s [fwd=2.478s bwd=0.973s opt=0.003s] dispatch=5.805s [2026-10-19 21:48:58,333] loop.py:45 INFO TRAIN_STEP step=15040 loss=1.299066 Training: 83%|████████▎ | 15040/18018 [32:18:44<6:46:44, 8.19s/step, loss=1.299, acc=0.701, acc_len=1.68, thru=14.5, I=14.6, T=36.6, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:48:58,546] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:49:04,647] loop.py:45 INFO TRAIN_STEP step=15041 loss=1.186823 Training: 83%|████████▎ | 15041/18018 [32:18:50<6:18:41, 7.63s/step, loss=1.187, acc=0.721, acc_len=1.75, thru=17.2, I=20.5, T=38.3, wait=3.0s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:49:04,642] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15040: forward=2363.6ms backward=929.4ms [2026-10-19 21:49:12,637] loop.py:45 INFO TRAIN_STEP step=15042 loss=1.182539 Training: 83%|████████▎ | 15042/18018 [32:18:58<6:23:47, 7.74s/step, loss=1.183, acc=0.718, acc_len=1.74, thru=14.8, I=14.4, T=37.5, wait=4.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:49:21,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:49:22,708] loop.py:45 INFO TRAIN_STEP step=15043 loss=1.291411 Training: 83%|████████▎ | 15043/18018 [32:19:08<6:58:22, 8.44s/step, loss=1.291, acc=0.698, acc_len=1.68, thru=15.5, I=16.0, T=35.7, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:49:22,966] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:49:28,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:49:28,787] loop.py:45 INFO TRAIN_STEP step=15044 loss=1.237786 Training: 83%|████████▎ | 15044/18018 [32:19:14<6:23:08, 7.73s/step, loss=1.238, acc=0.712, acc_len=1.72, thru=20.4, I=21.7, T=39.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:49:29,179] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:49:34,617] loop.py:380 INFO TIMING step=15045: step=3.349s data=1.376s compute=3.314s [fwd=2.415s bwd=0.885s opt=0.003s] dispatch=2.472s [2026-10-19 21:49:34,617] loop.py:45 INFO TRAIN_STEP step=15045 loss=1.269129 Training: 83%|████████▎ | 15045/18018 [32:19:20<5:54:47, 7.16s/step, loss=1.269, acc=0.705, acc_len=1.69, thru=20.5, I=22.1, T=38.2, wait=2.5s, pool=57, epoch=2/2][2026-10-19 21:49:43,842] loop.py:45 INFO TRAIN_STEP step=15046 loss=1.304829 Training: 84%|████████▎ | 15046/18018 [32:19:29<6:25:19, 7.78s/step, loss=1.305, acc=0.701, acc_len=1.68, thru=14.0, I=13.9, T=36.0, wait=5.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:49:43,837] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15045: forward=2528.7ms backward=965.9ms [2026-10-19 21:49:54,014] loop.py:45 INFO TRAIN_STEP step=15047 loss=1.251786 Training: 84%|████████▎ | 15047/18018 [32:19:40<7:00:44, 8.50s/step, loss=1.252, acc=0.708, acc_len=1.71, thru=12.1, I=11.7, T=29.5, wait=5.8s, pool=49, epoch=2/2][2026-10-19 21:50:02,650] loop.py:45 INFO TRAIN_STEP step=15048 loss=1.270779 Training: 84%|████████▎ | 15048/18018 [32:19:48<7:02:40, 8.54s/step, loss=1.271, acc=0.704, acc_len=1.70, thru=14.9, I=13.0, T=37.2, wait=5.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:50:10,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:50:10,651] loop.py:45 INFO TRAIN_STEP step=15049 loss=1.202339 Training: 84%|████████▎ | 15049/18018 [32:19:56<6:54:33, 8.38s/step, loss=1.202, acc=0.716, acc_len=1.74, thru=17.8, I=17.9, T=36.5, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:50:11,143] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:50:16,918] loop.py:380 INFO TIMING step=15050: step=3.265s data=1.278s compute=3.231s [fwd=2.310s bwd=0.907s opt=0.003s] dispatch=2.995s [2026-10-19 21:50:16,919] loop.py:45 INFO TRAIN_STEP step=15050 loss=1.146148 Training: 84%|████████▎ | 15050/18018 [32:20:03<6:23:05, 7.74s/step, loss=1.146, acc=0.725, acc_len=1.77, thru=16.3, I=17.3, T=39.2, wait=3.0s, pool=29, epoch=2/2](TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 21:50:16,984] eagle_store.py:517 WARNING batch_get_buffer missing keys (k481604-a1c5da74_hs, k481604-a1c5da74_ids, k481604-a1c5da74_lhs); sleeping 0.50s. [2026-10-19 21:50:27,711] loop.py:45 INFO TRAIN_STEP step=15051 loss=1.140543 Training: 84%|████████▎ | 15051/18018 [32:20:13<7:08:11, 8.66s/step, loss=1.141, acc=0.728, acc_len=1.78, thru=14.0, I=14.0, T=29.4, wait=6.4s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:50:27,706] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15050: forward=3245.3ms backward=1062.5ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 21:50:16,990] eagle_store.py:517 WARNING batch_get_buffer missing keys (k481606-8513093c_hs, k481606-8513093c_ids, k481606-8513093c_lhs); sleeping 0.50s. [2026-10-19 21:50:37,404] loop.py:45 INFO TRAIN_STEP step=15052 loss=1.284181 Training: 84%|████████▎ | 15052/18018 [32:20:23<7:23:24, 8.97s/step, loss=1.284, acc=0.707, acc_len=1.71, thru=12.5, I=13.1, T=34.3, wait=6.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:50:44,744] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:50:46,010] loop.py:45 INFO TRAIN_STEP step=15053 loss=1.357336 Training: 84%|████████▎ | 15053/18018 [32:20:32<7:17:51, 8.86s/step, loss=1.357, acc=0.692, acc_len=1.65, thru=17.9, I=16.6, T=33.1, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:50:46,252] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:50:52,494] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:50:53,817] loop.py:45 INFO TRAIN_STEP step=15054 loss=1.207511 Training: 84%|████████▎ | 15054/18018 [32:20:39<7:02:05, 8.54s/step, loss=1.208, acc=0.717, acc_len=1.75, thru=15.0, I=19.9, T=31.4, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:50:54,001] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:50:59,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:51:01,359] loop.py:380 INFO TIMING step=15055: step=3.399s data=1.392s compute=3.365s [fwd=2.419s bwd=0.933s opt=0.003s] dispatch=4.133s [2026-10-19 21:51:01,359] loop.py:45 INFO TRAIN_STEP step=15055 loss=1.299440 Training: 84%|████████▎ | 15055/18018 [32:20:47<6:47:05, 8.24s/step, loss=1.299, acc=0.699, acc_len=1.68, thru=15.8, I=22.1, T=37.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:51:01,392] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:51:07,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:51:08,631] loop.py:45 INFO TRAIN_STEP step=15056 loss=1.291425 Training: 84%|████████▎ | 15056/18018 [32:20:54<6:32:32, 7.95s/step, loss=1.291, acc=0.701, acc_len=1.68, thru=18.1, I=25.6, T=37.6, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:51:08,626] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15055: forward=2437.7ms backward=923.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:51:09,038] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 21:51:14,882] loop.py:45 INFO TRAIN_STEP step=15057 loss=1.282129 Training: 84%|████████▎ | 15057/18018 [32:21:00<6:07:15, 7.44s/step, loss=1.282, acc=0.706, acc_len=1.70, thru=17.7, I=21.3, T=38.2, wait=2.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:51:24,778] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:51:26,228] loop.py:45 INFO TRAIN_STEP step=15058 loss=1.356063 Training: 84%|████████▎ | 15058/18018 [32:21:12<7:04:54, 8.61s/step, loss=1.356, acc=0.692, acc_len=1.66, thru=14.9, I=14.3, T=32.8, wait=7.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:51:26,294] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-19 21:51:33,197] loop.py:45 INFO TRAIN_STEP step=15059 loss=1.259456 Training: 84%|████████▎ | 15059/18018 [32:21:19<6:40:25, 8.12s/step, loss=1.259, acc=0.710, acc_len=1.71, thru=17.1, I=19.9, T=37.1, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:51:37,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:51:38,485] loop.py:380 INFO TIMING step=15060: step=3.245s data=1.255s compute=3.220s [fwd=2.319s bwd=0.887s opt=0.003s] dispatch=2.034s [2026-10-19 21:51:38,485] loop.py:45 INFO TRAIN_STEP step=15060 loss=1.169375 Training: 84%|████████▎ | 15060/18018 [32:21:24<5:58:27, 7.27s/step, loss=1.169, acc=0.720, acc_len=1.75, thru=22.3, I=24.2, T=39.4, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:51:38,699] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:51:44,605] loop.py:45 INFO TRAIN_STEP step=15061 loss=1.278663 Training: 84%|████████▎ | 15061/18018 [32:21:30<5:41:17, 6.92s/step, loss=1.279, acc=0.704, acc_len=1.69, thru=20.4, I=23.3, T=37.4, wait=2.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:51:44,600] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15060: forward=2458.1ms backward=926.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:51:50,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:51:51,715] loop.py:45 INFO TRAIN_STEP step=15062 loss=1.294020 Training: 84%|████████▎ | 15062/18018 [32:21:37<5:43:54, 6.98s/step, loss=1.294, acc=0.701, acc_len=1.68, thru=19.1, I=20.0, T=36.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:51:51,787] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 21:51:57,403] loop.py:45 INFO TRAIN_STEP step=15063 loss=1.323450 Training: 84%|████████▎ | 15063/18018 [32:21:43<5:24:41, 6.59s/step, loss=1.323, acc=0.693, acc_len=1.66, thru=20.0, I=23.0, T=40.4, wait=2.5s, pool=53, epoch=2/2][2026-10-19 21:52:05,543] loop.py:45 INFO TRAIN_STEP step=15064 loss=1.308449 Training: 84%|████████▎ | 15064/18018 [32:21:51<5:47:25, 7.06s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=14.9, I=15.5, T=36.3, wait=4.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:12,066] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:52:13,012] loop.py:380 INFO TIMING step=15065: step=3.519s data=1.452s compute=3.493s [fwd=2.538s bwd=0.941s opt=0.003s] dispatch=3.942s [2026-10-19 21:52:13,012] loop.py:45 INFO TRAIN_STEP step=15065 loss=1.169625 Training: 84%|████████▎ | 15065/18018 [32:21:59<5:53:26, 7.18s/step, loss=1.170, acc=0.730, acc_len=1.78, thru=19.1, I=19.3, T=36.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:13,090] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (TrainerActor pid=9310) [2026-10-19 21:52:21,543] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15065: forward=2571.7ms backward=921.8ms [2026-10-19 21:52:21,548] loop.py:45 INFO TRAIN_STEP step=15066 loss=1.274678 Training: 84%|████████▎ | 15066/18018 [32:22:07<6:13:16, 7.59s/step, loss=1.275, acc=0.709, acc_len=1.71, thru=13.8, I=15.7, T=36.2, wait=5.0s, pool=61, epoch=2/2][2026-10-19 21:52:30,057] loop.py:45 INFO TRAIN_STEP step=15067 loss=1.101970 Training: 84%|████████▎ | 15067/18018 [32:22:16<6:26:46, 7.86s/step, loss=1.102, acc=0.737, acc_len=1.81, thru=15.5, I=15.0, T=36.9, wait=5.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:35,097] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:52:36,477] loop.py:45 INFO TRAIN_STEP step=15068 loss=1.278059 Training: 84%|████████▎ | 15068/18018 [32:22:22<6:05:21, 7.43s/step, loss=1.278, acc=0.703, acc_len=1.69, thru=19.8, I=21.6, T=38.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:36,626] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:41,209] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:52:42,549] loop.py:45 INFO TRAIN_STEP step=15069 loss=1.278118 Training: 84%|████████▎ | 15069/18018 [32:22:28<5:45:11, 7.02s/step, loss=1.278, acc=0.704, acc_len=1.69, thru=21.4, I=29.0, T=35.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:42,720] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:49,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:52:49,277] loop.py:380 INFO TIMING step=15070: step=3.607s data=1.674s compute=3.571s [fwd=2.674s bwd=0.882s opt=0.003s] dispatch=3.110s [2026-10-19 21:52:49,278] loop.py:45 INFO TRAIN_STEP step=15070 loss=1.304775 Training: 84%|████████▎ | 15070/18018 [32:22:35<5:40:42, 6.93s/step, loss=1.305, acc=0.697, acc_len=1.67, thru=18.9, I=23.4, T=35.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:49,581] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:55,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:52:55,936] loop.py:45 INFO TRAIN_STEP step=15071 loss=1.369781 Training: 84%|████████▎ | 15071/18018 [32:22:42<5:36:32, 6.85s/step, loss=1.370, acc=0.687, acc_len=1.64, thru=19.8, I=20.2, T=35.5, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:52:55,931] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15070: forward=2653.1ms backward=906.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:52:56,198] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:53:01,502] loop.py:45 INFO TRAIN_STEP step=15072 loss=1.232034 Training: 84%|████████▎ | 15072/18018 [32:22:47<5:17:29, 6.47s/step, loss=1.232, acc=0.714, acc_len=1.73, thru=20.4, I=23.0, T=39.9, wait=2.3s, pool=57, epoch=2/2][2026-10-19 21:53:10,190] loop.py:45 INFO TRAIN_STEP step=15073 loss=1.236700 Training: 84%|████████▎ | 15073/18018 [32:22:56<5:50:03, 7.13s/step, loss=1.237, acc=0.712, acc_len=1.72, thru=14.8, I=14.8, T=33.2, wait=4.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:53:17,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:53:18,941] loop.py:45 INFO TRAIN_STEP step=15074 loss=1.249675 Training: 84%|████████▎ | 15074/18018 [32:23:05<6:13:46, 7.62s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=16.1, I=17.2, T=35.8, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:53:19,442] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:53:25,013] loop.py:380 INFO TIMING step=15075: step=3.212s data=1.288s compute=3.181s [fwd=2.288s bwd=0.880s opt=0.003s] dispatch=2.852s [2026-10-19 21:53:25,013] loop.py:45 INFO TRAIN_STEP step=15075 loss=1.243465 Training: 84%|████████▎ | 15075/18018 [32:23:11<5:50:54, 7.15s/step, loss=1.243, acc=0.708, acc_len=1.70, thru=19.0, I=22.3, T=39.9, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:53:33,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:53:34,503] loop.py:45 INFO TRAIN_STEP step=15076 loss=1.288705 Training: 84%|████████▎ | 15076/18018 [32:23:20<6:25:09, 7.85s/step, loss=1.289, acc=0.701, acc_len=1.69, thru=14.7, I=15.7, T=34.5, wait=5.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:53:34,498] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15075: forward=2687.8ms backward=971.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:53:34,842] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:53:40,983] loop.py:45 INFO TRAIN_STEP step=15077 loss=1.262821 Training: 84%|████████▎ | 15077/18018 [32:23:27<6:04:50, 7.44s/step, loss=1.263, acc=0.709, acc_len=1.71, thru=15.8, I=18.4, T=39.0, wait=3.2s, pool=37, epoch=2/2][2026-10-19 21:53:48,801] loop.py:45 INFO TRAIN_STEP step=15078 loss=1.130831 Training: 84%|████████▎ | 15078/18018 [32:23:34<6:10:13, 7.56s/step, loss=1.131, acc=0.731, acc_len=1.80, thru=18.1, I=17.6, T=36.8, wait=4.3s, pool=53, epoch=2/2][2026-10-19 21:53:57,208] loop.py:45 INFO TRAIN_STEP step=15079 loss=1.337256 Training: 84%|████████▎ | 15079/18018 [32:23:43<6:22:34, 7.81s/step, loss=1.337, acc=0.693, acc_len=1.65, thru=14.3, I=15.2, T=34.5, wait=4.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:54:05,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:54:06,485] loop.py:380 INFO TIMING step=15080: step=3.706s data=1.511s compute=3.671s [fwd=2.681s bwd=0.976s opt=0.003s] dispatch=5.562s [2026-10-19 21:54:06,486] loop.py:45 INFO TRAIN_STEP step=15080 loss=1.125299 Training: 84%|████████▎ | 15080/18018 [32:23:52<6:44:01, 8.25s/step, loss=1.125, acc=0.735, acc_len=1.80, thru=16.1, I=15.1, T=34.5, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:54:06,519] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:54:12,929] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:54:14,098] loop.py:45 INFO TRAIN_STEP step=15081 loss=1.203906 Training: 84%|████████▎ | 15081/18018 [32:24:00<6:34:28, 8.06s/step, loss=1.204, acc=0.715, acc_len=1.73, thru=17.3, I=19.4, T=36.9, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:54:14,093] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15080: forward=2443.3ms backward=973.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:54:14,439] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:54:20,654] loop.py:45 INFO TRAIN_STEP step=15082 loss=1.220093 Training: 84%|████████▎ | 15082/18018 [32:24:06<6:12:18, 7.61s/step, loss=1.220, acc=0.713, acc_len=1.72, thru=18.6, I=20.8, T=37.6, wait=3.1s, pool=53, epoch=2/2][2026-10-19 21:54:28,811] loop.py:45 INFO TRAIN_STEP step=15083 loss=1.335192 Training: 84%|████████▎ | 15083/18018 [32:24:14<6:20:12, 7.77s/step, loss=1.335, acc=0.694, acc_len=1.65, thru=16.1, I=16.8, T=35.9, wait=4.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:54:37,865] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:54:38,754] loop.py:45 INFO TRAIN_STEP step=15084 loss=1.253383 Training: 84%|████████▎ | 15084/18018 [32:24:24<6:51:54, 8.42s/step, loss=1.253, acc=0.709, acc_len=1.70, thru=14.1, I=13.6, T=35.0, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:54:38,871] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 21:54:44,867] loop.py:380 INFO TIMING step=15085: step=3.401s data=1.373s compute=3.370s [fwd=2.413s bwd=0.943s opt=0.003s] dispatch=2.703s [2026-10-19 21:54:44,867] loop.py:45 INFO TRAIN_STEP step=15085 loss=1.209682 Training: 84%|████████▎ | 15085/18018 [32:24:30<6:17:52, 7.73s/step, loss=1.210, acc=0.716, acc_len=1.73, thru=19.9, I=19.7, T=37.6, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:54:53,499] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:54:54,205] loop.py:45 INFO TRAIN_STEP step=15086 loss=1.247374 Training: 84%|████████▎ | 15086/18018 [32:24:40<6:41:21, 8.21s/step, loss=1.247, acc=0.711, acc_len=1.71, thru=14.2, I=14.6, T=37.3, wait=5.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:54:54,200] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15085: forward=2460.7ms backward=935.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:54:54,505] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:55:03,605] loop.py:45 INFO TRAIN_STEP step=15087 loss=1.249586 Training: 84%|████████▎ | 15087/18018 [32:24:49<6:58:37, 8.57s/step, loss=1.250, acc=0.709, acc_len=1.71, thru=10.1, I=11.5, T=33.0, wait=5.5s, pool=29, epoch=2/2][2026-10-19 21:55:11,917] loop.py:45 INFO TRAIN_STEP step=15088 loss=1.203400 Training: 84%|████████▎ | 15088/18018 [32:24:58<6:54:40, 8.49s/step, loss=1.203, acc=0.714, acc_len=1.73, thru=18.4, I=15.1, T=37.5, wait=4.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:18,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:55:18,385] loop.py:45 INFO TRAIN_STEP step=15089 loss=1.295627 Training: 84%|████████▎ | 15089/18018 [32:25:04<6:24:54, 7.88s/step, loss=1.296, acc=0.702, acc_len=1.68, thru=19.8, I=20.2, T=38.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:18,527] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:23,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:55:24,990] loop.py:380 INFO TIMING step=15090: step=3.339s data=1.354s compute=3.303s [fwd=2.380s bwd=0.909s opt=0.003s] dispatch=3.258s [2026-10-19 21:55:24,990] loop.py:45 INFO TRAIN_STEP step=15090 loss=1.313096 Training: 84%|████████▎ | 15090/18018 [32:25:11<6:06:02, 7.50s/step, loss=1.313, acc=0.695, acc_len=1.66, thru=19.7, I=23.1, T=38.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:25,300] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:55:32,927] loop.py:45 INFO TRAIN_STEP step=15091 loss=1.254498 Training: 84%|████████▍ | 15091/18018 [32:25:19<6:12:17, 7.63s/step, loss=1.254, acc=0.708, acc_len=1.70, thru=15.4, I=18.9, T=39.0, wait=4.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:55:32,922] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15090: forward=2340.8ms backward=897.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:38,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:55:39,709] loop.py:45 INFO TRAIN_STEP step=15092 loss=1.365492 Training: 84%|████████▍ | 15092/18018 [32:25:25<5:59:45, 7.38s/step, loss=1.365, acc=0.687, acc_len=1.63, thru=19.2, I=19.4, T=36.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:39,743] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:46,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:55:46,648] loop.py:45 INFO TRAIN_STEP step=15093 loss=1.324962 Training: 84%|████████▍ | 15093/18018 [32:25:32<5:53:11, 7.25s/step, loss=1.325, acc=0.695, acc_len=1.66, thru=18.5, I=21.3, T=38.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:46,747] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:51,946] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:55:52,891] loop.py:45 INFO TRAIN_STEP step=15094 loss=1.212590 Training: 84%|████████▍ | 15094/18018 [32:25:38<5:38:27, 6.95s/step, loss=1.213, acc=0.712, acc_len=1.72, thru=20.3, I=22.7, T=38.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:55:52,961] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 21:55:58,964] loop.py:380 INFO TIMING step=15095: step=3.299s data=1.336s compute=3.271s [fwd=2.356s bwd=0.900s opt=0.003s] dispatch=2.764s [2026-10-19 21:55:58,964] loop.py:45 INFO TRAIN_STEP step=15095 loss=1.185126 Training: 84%|████████▍ | 15095/18018 [32:25:45<5:25:36, 6.68s/step, loss=1.185, acc=0.719, acc_len=1.74, thru=21.0, I=24.6, T=38.8, wait=2.8s, pool=57, epoch=2/2][2026-10-19 21:56:05,999] loop.py:45 INFO TRAIN_STEP step=15096 loss=1.271538 Training: 84%|████████▍ | 15096/18018 [32:25:52<5:30:36, 6.79s/step, loss=1.272, acc=0.708, acc_len=1.71, thru=16.4, I=16.6, T=37.0, wait=3.6s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:56:05,995] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15095: forward=2470.1ms backward=938.9ms [2026-10-19 21:56:14,266] loop.py:45 INFO TRAIN_STEP step=15097 loss=1.244436 Training: 84%|████████▍ | 15097/18018 [32:26:00<5:52:03, 7.23s/step, loss=1.244, acc=0.708, acc_len=1.71, thru=17.1, I=16.2, T=35.9, wait=4.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:56:21,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:56:22,101] loop.py:45 INFO TRAIN_STEP step=15098 loss=1.269974 Training: 84%|████████▍ | 15098/18018 [32:26:08<6:00:46, 7.41s/step, loss=1.270, acc=0.706, acc_len=1.70, thru=16.1, I=17.7, T=35.9, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:56:22,521] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:56:29,039] loop.py:45 INFO TRAIN_STEP step=15099 loss=1.204271 Training: 84%|████████▍ | 15099/18018 [32:26:15<5:53:42, 7.27s/step, loss=1.204, acc=0.715, acc_len=1.73, thru=16.3, I=19.9, T=38.2, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:56:35,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:56:36,367] loop.py:380 INFO TIMING step=15100: step=3.349s data=1.384s compute=3.306s [fwd=2.348s bwd=0.943s opt=0.003s] dispatch=3.970s [2026-10-19 21:56:36,367] loop.py:45 INFO TRAIN_STEP step=15100 loss=1.331127 Training: 84%|████████▍ | 15100/18018 [32:26:22<5:54:24, 7.29s/step, loss=1.331, acc=0.697, acc_len=1.67, thru=18.4, I=19.9, T=38.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:56:36,777] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:56:43,538] loop.py:45 INFO TRAIN_STEP step=15101 loss=1.239463 Training: 84%|████████▍ | 15101/18018 [32:26:29<5:52:36, 7.25s/step, loss=1.239, acc=0.714, acc_len=1.73, thru=17.0, I=22.7, T=38.0, wait=3.8s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:56:43,534] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15100: forward=2341.0ms backward=964.8ms [2026-10-19 21:56:52,080] loop.py:45 INFO TRAIN_STEP step=15102 loss=1.316126 Training: 84%|████████▍ | 15102/18018 [32:26:38<6:11:17, 7.64s/step, loss=1.316, acc=0.700, acc_len=1.67, thru=12.9, I=13.5, T=35.5, wait=4.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:56:58,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:56:59,240] loop.py:45 INFO TRAIN_STEP step=15103 loss=1.239815 Training: 84%|████████▍ | 15103/18018 [32:26:45<6:04:09, 7.50s/step, loss=1.240, acc=0.712, acc_len=1.72, thru=18.6, I=19.7, T=37.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:56:59,332] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:57:06,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:57:07,306] loop.py:45 INFO TRAIN_STEP step=15104 loss=1.282035 Training: 84%|████████▍ | 15104/18018 [32:26:53<6:12:20, 7.67s/step, loss=1.282, acc=0.706, acc_len=1.70, thru=14.9, I=18.2, T=36.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:57:07,568] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:57:15,532] loop.py:380 INFO TIMING step=15105: step=3.563s data=1.380s compute=3.536s [fwd=2.460s bwd=1.061s opt=0.003s] dispatch=4.655s [2026-10-19 21:57:15,532] loop.py:45 INFO TRAIN_STEP step=15105 loss=1.286622 Training: 84%|████████▍ | 15105/18018 [32:27:01<6:20:22, 7.83s/step, loss=1.287, acc=0.700, acc_len=1.68, thru=12.5, I=13.1, T=35.9, wait=4.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:57:23,485] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:57:24,776] loop.py:45 INFO TRAIN_STEP step=15106 loss=1.254321 Training: 84%|████████▍ | 15106/18018 [32:27:10<6:40:44, 8.26s/step, loss=1.254, acc=0.711, acc_len=1.72, thru=16.4, I=18.4, T=33.5, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:57:24,771] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15105: forward=2463.4ms backward=1109.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:57:24,988] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:57:30,952] loop.py:45 INFO TRAIN_STEP step=15107 loss=1.287010 Training: 84%|████████▍ | 15107/18018 [32:27:17<6:10:20, 7.63s/step, loss=1.287, acc=0.703, acc_len=1.69, thru=18.5, I=22.0, T=38.8, wait=2.9s, pool=53, epoch=2/2][2026-10-19 21:57:38,952] loop.py:45 INFO TRAIN_STEP step=15108 loss=1.281738 Training: 84%|████████▍ | 15108/18018 [32:27:25<6:15:31, 7.74s/step, loss=1.282, acc=0.703, acc_len=1.69, thru=14.2, I=14.2, T=36.4, wait=4.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:57:45,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:57:46,329] loop.py:45 INFO TRAIN_STEP step=15109 loss=1.284518 Training: 84%|████████▍ | 15109/18018 [32:27:32<6:10:03, 7.63s/step, loss=1.285, acc=0.702, acc_len=1.69, thru=18.9, I=18.5, T=37.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:57:46,399] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:57:52,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:57:52,821] loop.py:380 INFO TIMING step=15110: step=3.415s data=1.390s compute=3.382s [fwd=2.427s bwd=0.941s opt=0.003s] dispatch=3.070s [2026-10-19 21:57:52,821] loop.py:45 INFO TRAIN_STEP step=15110 loss=1.261552 Training: 84%|████████▍ | 15110/18018 [32:27:38<5:53:22, 7.29s/step, loss=1.262, acc=0.709, acc_len=1.71, thru=19.7, I=21.9, T=37.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:57:53,077] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:57:58,875] loop.py:45 INFO TRAIN_STEP step=15111 loss=1.241865 Training: 84%|████████▍ | 15111/18018 [32:27:44<5:35:17, 6.92s/step, loss=1.242, acc=0.708, acc_len=1.71, thru=18.3, I=19.6, T=38.3, wait=2.7s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:57:58,871] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15110: forward=2369.4ms backward=928.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:07,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:58:07,453] loop.py:45 INFO TRAIN_STEP step=15112 loss=1.202588 Training: 84%|████████▍ | 15112/18018 [32:27:53<5:59:14, 7.42s/step, loss=1.203, acc=0.717, acc_len=1.74, thru=16.8, I=17.1, T=36.7, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:07,555] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:11,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:58:12,641] loop.py:45 INFO TRAIN_STEP step=15113 loss=1.316103 Training: 84%|████████▍ | 15113/18018 [32:27:58<5:26:45, 6.75s/step, loss=1.316, acc=0.697, acc_len=1.66, thru=23.1, I=28.5, T=39.1, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:12,793] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:18,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:58:18,895] loop.py:45 INFO TRAIN_STEP step=15114 loss=1.301749 Training: 84%|████████▍ | 15114/18018 [32:28:04<5:19:28, 6.60s/step, loss=1.302, acc=0.702, acc_len=1.68, thru=20.9, I=26.4, T=36.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:19,109] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 21:58:24,730] loop.py:380 INFO TIMING step=15115: step=3.302s data=1.360s compute=3.270s [fwd=2.327s bwd=0.929s opt=0.003s] dispatch=2.523s [2026-10-19 21:58:24,730] loop.py:45 INFO TRAIN_STEP step=15115 loss=1.283890 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:24,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 84%|████████▍ | 15115/18018 [32:28:10<5:08:12, 6.37s/step, loss=1.284, acc=0.699, acc_len=1.68, thru=21.0, I=21.9, T=38.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:25,223] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 21:58:30,512] loop.py:45 INFO TRAIN_STEP step=15116 loss=1.406524 (TrainerActor pid=9310) [2026-10-19 21:58:30,508] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15115: forward=2302.6ms backward=878.7ms Training: 84%|████████▍ | 15116/18018 [32:28:16<4:59:34, 6.19s/step, loss=1.407, acc=0.681, acc_len=1.61, thru=19.8, I=22.3, T=39.7, wait=2.5s, pool=41, epoch=2/2][2026-10-19 21:58:39,031] loop.py:45 INFO TRAIN_STEP step=15117 loss=1.174580 Training: 84%|████████▍ | 15117/18018 [32:28:25<5:33:14, 6.89s/step, loss=1.175, acc=0.722, acc_len=1.75, thru=16.6, I=16.2, T=37.6, wait=5.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:45,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:58:47,073] loop.py:45 INFO TRAIN_STEP step=15118 loss=1.309500 Training: 84%|████████▍ | 15118/18018 [32:28:33<5:49:47, 7.24s/step, loss=1.309, acc=0.697, acc_len=1.67, thru=15.9, I=17.7, T=37.1, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:58:47,478] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:58:54,018] loop.py:45 INFO TRAIN_STEP step=15119 loss=1.325099 Training: 84%|████████▍ | 15119/18018 [32:28:40<5:45:25, 7.15s/step, loss=1.325, acc=0.693, acc_len=1.66, thru=16.9, I=20.2, T=39.5, wait=3.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:59:03,507] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:59:04,998] loop.py:380 INFO TIMING step=15120: step=3.727s data=1.494s compute=3.685s [fwd=2.611s bwd=1.059s opt=0.003s] dispatch=7.243s [2026-10-19 21:59:04,998] loop.py:45 INFO TRAIN_STEP step=15120 loss=1.243616 Training: 84%|████████▍ | 15120/18018 [32:28:51<6:40:48, 8.30s/step, loss=1.244, acc=0.711, acc_len=1.72, thru=15.9, I=15.1, T=34.3, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:59:05,031] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 21:59:11,218] loop.py:45 INFO TRAIN_STEP step=15121 loss=1.258693 Training: 84%|████████▍ | 15121/18018 [32:28:57<6:10:34, 7.67s/step, loss=1.259, acc=0.707, acc_len=1.70, thru=18.4, I=22.1, T=38.0, wait=2.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:59:11,213] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15120: forward=2363.4ms backward=941.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:59:18,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:59:19,460] loop.py:45 INFO TRAIN_STEP step=15122 loss=1.205595 Training: 84%|████████▍ | 15122/18018 [32:29:05<6:18:37, 7.84s/step, loss=1.206, acc=0.716, acc_len=1.74, thru=16.6, I=17.9, T=37.0, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:59:19,607] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:59:25,686] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:59:26,942] loop.py:45 INFO TRAIN_STEP step=15123 loss=1.254367 Training: 84%|████████▍ | 15123/18018 [32:29:13<6:13:15, 7.74s/step, loss=1.254, acc=0.705, acc_len=1.69, thru=17.1, I=21.4, T=36.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:59:27,194] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 21:59:34,576] loop.py:45 INFO TRAIN_STEP step=15124 loss=1.272720 Training: 84%|████████▍ | 15124/18018 [32:29:20<6:11:38, 7.71s/step, loss=1.273, acc=0.704, acc_len=1.69, thru=15.0, I=17.4, T=38.5, wait=4.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:59:40,647] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 21:59:41,058] loop.py:380 INFO TIMING step=15125: step=3.368s data=1.323s compute=3.338s [fwd=2.421s bwd=0.902s opt=0.003s] dispatch=3.106s [2026-10-19 21:59:41,058] loop.py:45 INFO TRAIN_STEP step=15125 loss=1.310903 Training: 84%|████████▍ | 15125/18018 [32:29:27<5:53:51, 7.34s/step, loss=1.311, acc=0.696, acc_len=1.66, thru=20.4, I=22.7, T=38.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 21:59:41,155] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 21:59:47,147] loop.py:45 INFO TRAIN_STEP step=15126 loss=1.140451 Training: 84%|████████▍ | 15126/18018 [32:29:33<5:35:37, 6.96s/step, loss=1.140, acc=0.728, acc_len=1.78, thru=19.6, I=19.4, T=39.8, wait=2.9s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 21:59:47,143] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15125: forward=2273.5ms backward=894.0ms [2026-10-19 21:59:56,037] loop.py:45 INFO TRAIN_STEP step=15127 loss=1.307361 Training: 84%|████████▍ | 15127/18018 [32:29:42<6:03:22, 7.54s/step, loss=1.307, acc=0.699, acc_len=1.68, thru=15.6, I=15.6, T=35.0, wait=5.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:00:06,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:00:07,204] loop.py:45 INFO TRAIN_STEP step=15128 loss=1.164586 Training: 84%|████████▍ | 15128/18018 [32:29:53<6:55:39, 8.63s/step, loss=1.165, acc=0.730, acc_len=1.78, thru=13.0, I=13.6, T=33.8, wait=7.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:00:07,372] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:00:14,408] loop.py:45 INFO TRAIN_STEP step=15129 loss=1.236182 Training: 84%|████████▍ | 15129/18018 [32:30:00<6:34:52, 8.20s/step, loss=1.236, acc=0.710, acc_len=1.71, thru=15.3, I=17.6, T=36.7, wait=3.7s, pool=41, epoch=2/2][2026-10-19 22:00:22,264] loop.py:380 INFO TIMING step=15130: step=3.409s data=1.416s compute=3.376s [fwd=2.433s bwd=0.930s opt=0.003s] dispatch=4.440s [2026-10-19 22:00:22,264] loop.py:45 INFO TRAIN_STEP step=15130 loss=1.189223 Training: 84%|████████▍ | 15130/18018 [32:30:08<6:29:48, 8.10s/step, loss=1.189, acc=0.718, acc_len=1.74, thru=15.3, I=15.6, T=37.5, wait=4.4s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:00:29,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:00:29,788] loop.py:45 INFO TRAIN_STEP step=15131 loss=1.359754 Training: 84%|████████▍ | 15131/18018 [32:30:15<6:21:21, 7.93s/step, loss=1.360, acc=0.689, acc_len=1.64, thru=17.6, I=17.6, T=37.0, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:00:29,784] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15130: forward=2430.6ms backward=971.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:00:30,138] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:00:37,269] loop.py:45 INFO TRAIN_STEP step=15132 loss=1.193591 Training: 84%|████████▍ | 15132/18018 [32:30:23<6:14:47, 7.79s/step, loss=1.194, acc=0.720, acc_len=1.75, thru=15.8, I=16.3, T=38.1, wait=4.1s, pool=45, epoch=2/2][2026-10-19 22:00:44,245] loop.py:45 INFO TRAIN_STEP step=15133 loss=1.387593 Training: 84%|████████▍ | 15133/18018 [32:30:30<6:02:54, 7.55s/step, loss=1.388, acc=0.682, acc_len=1.62, thru=17.1, I=18.9, T=35.9, wait=3.4s, pool=49, epoch=2/2][2026-10-19 22:00:50,596] loop.py:45 INFO TRAIN_STEP step=15134 loss=1.268504 Training: 84%|████████▍ | 15134/18018 [32:30:36<5:45:30, 7.19s/step, loss=1.269, acc=0.706, acc_len=1.70, thru=18.9, I=19.4, T=38.1, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:00:56,432] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:00:57,507] loop.py:380 INFO TIMING step=15135: step=3.400s data=1.355s compute=3.371s [fwd=2.427s bwd=0.930s opt=0.003s] dispatch=3.503s [2026-10-19 22:00:57,507] loop.py:45 INFO TRAIN_STEP step=15135 loss=1.294531 Training: 84%|████████▍ | 15135/18018 [32:30:43<5:41:24, 7.11s/step, loss=1.295, acc=0.702, acc_len=1.68, thru=19.4, I=21.1, T=37.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:00:57,939] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:01:04,882] loop.py:45 INFO TRAIN_STEP step=15136 loss=1.300540 Training: 84%|████████▍ | 15136/18018 [32:30:50<5:45:10, 7.19s/step, loss=1.301, acc=0.696, acc_len=1.67, thru=16.4, I=21.2, T=39.1, wait=4.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:01:04,877] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15135: forward=2350.3ms backward=880.0ms [2026-10-19 22:01:12,025] loop.py:45 INFO TRAIN_STEP step=15137 loss=1.292258 Training: 84%|████████▍ | 15137/18018 [32:30:58<5:44:26, 7.17s/step, loss=1.292, acc=0.700, acc_len=1.68, thru=18.5, I=17.9, T=37.1, wait=3.7s, pool=61, epoch=2/2][2026-10-19 22:01:19,205] loop.py:45 INFO TRAIN_STEP step=15138 loss=1.157414 Training: 84%|████████▍ | 15138/18018 [32:31:05<5:44:24, 7.18s/step, loss=1.157, acc=0.724, acc_len=1.76, thru=15.1, I=14.8, T=37.6, wait=3.8s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:01:26,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:01:27,921] loop.py:45 INFO TRAIN_STEP step=15139 loss=1.257545 Training: 84%|████████▍ | 15139/18018 [32:31:14<6:06:29, 7.64s/step, loss=1.258, acc=0.707, acc_len=1.71, thru=18.2, I=19.1, T=35.1, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:01:28,299] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:01:33,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:01:34,316] loop.py:380 INFO TIMING step=15140: step=3.301s data=1.308s compute=3.268s [fwd=2.377s bwd=0.876s opt=0.003s] dispatch=3.084s [2026-10-19 22:01:34,316] loop.py:45 INFO TRAIN_STEP step=15140 loss=1.245951 Training: 84%|████████▍ | 15140/18018 [32:31:20<5:48:27, 7.26s/step, loss=1.246, acc=0.706, acc_len=1.70, thru=19.4, I=26.3, T=38.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:01:34,665] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:01:39,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 22:01:40,943] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15140: forward=2409.2ms backward=954.0ms [2026-10-19 22:01:40,947] loop.py:45 INFO TRAIN_STEP step=15141 loss=1.235616 Training: 84%|████████▍ | 15141/18018 [32:31:27<5:39:12, 7.07s/step, loss=1.236, acc=0.712, acc_len=1.72, thru=19.7, I=26.0, T=37.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:01:41,379] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:01:46,511] loop.py:45 INFO TRAIN_STEP step=15142 loss=1.392103 Training: 84%|████████▍ | 15142/18018 [32:31:32<5:17:24, 6.62s/step, loss=1.392, acc=0.683, acc_len=1.62, thru=19.0, I=27.3, T=39.7, wait=2.3s, pool=41, epoch=2/2][2026-10-19 22:01:55,156] loop.py:45 INFO TRAIN_STEP step=15143 loss=1.290932 Training: 84%|████████▍ | 15143/18018 [32:31:41<5:46:22, 7.23s/step, loss=1.291, acc=0.701, acc_len=1.68, thru=15.9, I=15.0, T=35.3, wait=5.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:02:02,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:02:03,999] loop.py:45 INFO TRAIN_STEP step=15144 loss=1.415435 Training: 84%|████████▍ | 15144/18018 [32:31:50<6:09:27, 7.71s/step, loss=1.415, acc=0.678, acc_len=1.60, thru=15.6, I=17.5, T=36.5, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:02:04,032] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 22:02:10,575] loop.py:380 INFO TIMING step=15145: step=3.555s data=1.399s compute=3.528s [fwd=2.504s bwd=1.011s opt=0.003s] dispatch=3.011s [2026-10-19 22:02:10,575] loop.py:45 INFO TRAIN_STEP step=15145 loss=1.176650 Training: 84%|████████▍ | 15145/18018 [32:31:56<5:52:57, 7.37s/step, loss=1.177, acc=0.723, acc_len=1.76, thru=18.1, I=20.8, T=36.0, wait=3.0s, pool=57, epoch=2/2][2026-10-19 22:02:16,029] loop.py:45 INFO TRAIN_STEP step=15146 loss=1.344258 Training: 84%|████████▍ | 15146/18018 [32:32:02<5:25:18, 6.80s/step, loss=1.344, acc=0.695, acc_len=1.66, thru=20.1, I=19.4, T=39.6, wait=2.2s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:02:16,025] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15145: forward=2317.7ms backward=876.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:02:23,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:02:23,983] loop.py:45 INFO TRAIN_STEP step=15147 loss=1.187897 Training: 84%|████████▍ | 15147/18018 [32:32:10<5:41:51, 7.14s/step, loss=1.188, acc=0.721, acc_len=1.75, thru=18.6, I=17.8, T=37.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:02:24,016] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 22:02:31,546] loop.py:45 INFO TRAIN_STEP step=15148 loss=1.077419 Training: 84%|████████▍ | 15148/18018 [32:32:17<5:47:42, 7.27s/step, loss=1.077, acc=0.736, acc_len=1.81, thru=15.1, I=16.7, T=32.5, wait=3.6s, pool=49, epoch=2/2][2026-10-19 22:02:40,683] loop.py:45 INFO TRAIN_STEP step=15149 loss=1.336547 Training: 84%|████████▍ | 15149/18018 [32:32:26<6:14:22, 7.83s/step, loss=1.337, acc=0.693, acc_len=1.66, thru=15.2, I=13.8, T=35.3, wait=5.5s, pool=57, epoch=2/2][2026-10-19 22:02:46,249] loop.py:380 INFO TIMING step=15150: step=3.339s data=1.341s compute=3.312s [fwd=2.373s bwd=0.925s opt=0.003s] dispatch=2.220s [2026-10-19 22:02:46,249] loop.py:45 INFO TRAIN_STEP step=15150 loss=1.308705 Training: 84%|████████▍ | 15150/18018 [32:32:32<5:41:49, 7.15s/step, loss=1.309, acc=0.700, acc_len=1.68, thru=19.6, I=19.5, T=38.3, wait=2.2s, pool=57, epoch=2/2][2026-10-19 22:02:53,410] loop.py:45 INFO TRAIN_STEP step=15151 loss=1.225830 Training: 84%|████████▍ | 15151/18018 [32:32:39<5:41:50, 7.15s/step, loss=1.226, acc=0.712, acc_len=1.72, thru=18.5, I=18.4, T=38.7, wait=3.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:02:53,405] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15150: forward=2356.7ms backward=904.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:02:59,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:02:59,873] loop.py:45 INFO TRAIN_STEP step=15152 loss=1.285680 Training: 84%|████████▍ | 15152/18018 [32:32:45<5:31:48, 6.95s/step, loss=1.286, acc=0.703, acc_len=1.69, thru=20.0, I=20.6, T=39.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:03:00,119] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:03:06,807] loop.py:45 INFO TRAIN_STEP step=15153 loss=1.416297 Training: 84%|████████▍ | 15153/18018 [32:32:52<5:31:31, 6.94s/step, loss=1.416, acc=0.680, acc_len=1.61, thru=17.4, I=20.3, T=38.1, wait=3.6s, pool=53, epoch=2/2][2026-10-19 22:03:14,556] loop.py:45 INFO TRAIN_STEP step=15154 loss=1.354651 Training: 84%|████████▍ | 15154/18018 [32:33:00<5:42:55, 7.18s/step, loss=1.355, acc=0.687, acc_len=1.64, thru=14.7, I=16.3, T=36.5, wait=4.2s, pool=37, epoch=2/2][2026-10-19 22:03:23,375] loop.py:380 INFO TIMING step=15155: step=3.379s data=1.349s compute=3.348s [fwd=2.393s bwd=0.941s opt=0.003s] dispatch=5.433s [2026-10-19 22:03:23,375] loop.py:45 INFO TRAIN_STEP step=15155 loss=1.201589 Training: 84%|████████▍ | 15155/18018 [32:33:09<6:06:11, 7.67s/step, loss=1.202, acc=0.715, acc_len=1.73, thru=16.5, I=14.6, T=37.9, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:03:23,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:03:23,868] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:03:29,673] loop.py:45 INFO TRAIN_STEP step=15156 loss=1.311828 Training: 84%|████████▍ | 15156/18018 [32:33:15<5:46:23, 7.26s/step, loss=1.312, acc=0.696, acc_len=1.67, thru=20.2, I=20.5, T=37.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:03:29,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 22:03:29,669] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15155: forward=2416.5ms backward=929.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:03:30,159] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:03:40,316] loop.py:45 INFO TRAIN_STEP step=15157 loss=1.366835 Training: 84%|████████▍ | 15157/18018 [32:33:26<6:34:38, 8.28s/step, loss=1.367, acc=0.687, acc_len=1.63, thru=11.3, I=11.4, T=35.1, wait=7.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:03:48,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:03:49,537] loop.py:45 INFO TRAIN_STEP step=15158 loss=1.248999 Training: 84%|████████▍ | 15158/18018 [32:33:35<6:48:02, 8.56s/step, loss=1.249, acc=0.715, acc_len=1.73, thru=16.1, I=17.7, T=33.7, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:03:49,606] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:03:54,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:03:54,616] loop.py:45 INFO TRAIN_STEP step=15159 loss=1.254565 Training: 84%|████████▍ | 15159/18018 [32:33:40<5:58:05, 7.52s/step, loss=1.255, acc=0.707, acc_len=1.70, thru=21.7, I=27.1, T=40.3, wait=1.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:03:55,022] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:04:00,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:04:00,568] loop.py:380 INFO TIMING step=15160: step=3.282s data=1.323s compute=3.251s [fwd=2.352s bwd=0.885s opt=0.003s] dispatch=2.663s [2026-10-19 22:04:00,568] loop.py:45 INFO TRAIN_STEP step=15160 loss=1.261909 Training: 84%|████████▍ | 15160/18018 [32:33:46<5:35:39, 7.05s/step, loss=1.262, acc=0.705, acc_len=1.70, thru=22.2, I=23.4, T=39.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:04:01,031] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:04:06,293] loop.py:45 INFO TRAIN_STEP step=15161 loss=1.213376 Training: 84%|████████▍ | 15161/18018 [32:33:52<5:16:40, 6.65s/step, loss=1.213, acc=0.715, acc_len=1.73, thru=20.2, I=23.9, T=38.5, wait=2.4s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:04:06,288] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15160: forward=2377.6ms backward=906.0ms [2026-10-19 22:04:13,666] loop.py:45 INFO TRAIN_STEP step=15162 loss=1.369325 Training: 84%|████████▍ | 15162/18018 [32:33:59<5:26:54, 6.87s/step, loss=1.369, acc=0.692, acc_len=1.65, thru=16.4, I=17.4, T=37.4, wait=3.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:04:21,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:04:22,630] loop.py:45 INFO TRAIN_STEP step=15163 loss=1.141435 Training: 84%|████████▍ | 15163/18018 [32:34:08<5:56:39, 7.50s/step, loss=1.141, acc=0.727, acc_len=1.77, thru=15.7, I=16.3, T=35.0, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:04:22,898] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:04:31,277] loop.py:45 INFO TRAIN_STEP step=15164 loss=1.308434 Training: 84%|████████▍ | 15164/18018 [32:34:17<6:12:57, 7.84s/step, loss=1.308, acc=0.696, acc_len=1.67, thru=11.8, I=12.9, T=35.1, wait=5.0s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:04:39,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:04:39,913] loop.py:380 INFO TIMING step=15165: step=4.039s data=2.043s compute=4.003s [fwd=3.071s bwd=0.918s opt=0.003s] dispatch=4.589s [2026-10-19 22:04:39,913] loop.py:45 INFO TRAIN_STEP step=15165 loss=1.247294 Training: 84%|████████▍ | 15165/18018 [32:34:25<6:24:12, 8.08s/step, loss=1.247, acc=0.710, acc_len=1.71, thru=18.5, I=15.0, T=31.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:04:40,136] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:04:47,120] loop.py:45 INFO TRAIN_STEP step=15166 loss=1.320005 Training: 84%|████████▍ | 15166/18018 [32:34:33<6:11:36, 7.82s/step, loss=1.320, acc=0.693, acc_len=1.66, thru=16.2, I=16.3, T=37.4, wait=3.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:04:47,116] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15165: forward=2446.4ms backward=931.4ms [2026-10-19 22:04:54,179] loop.py:45 INFO TRAIN_STEP step=15167 loss=1.369617 Training: 84%|████████▍ | 15167/18018 [32:34:40<6:00:38, 7.59s/step, loss=1.370, acc=0.688, acc_len=1.63, thru=18.7, I=17.3, T=35.4, wait=3.4s, pool=61, epoch=2/2][2026-10-19 22:05:00,835] loop.py:45 INFO TRAIN_STEP step=15168 loss=1.409491 Training: 84%|████████▍ | 15168/18018 [32:34:46<5:47:16, 7.31s/step, loss=1.409, acc=0.682, acc_len=1.62, thru=14.5, I=19.5, T=32.2, wait=2.7s, pool=13, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 22:05:02,101] eagle_store.py:517 WARNING batch_get_buffer missing keys (k485378-a963e3bc_hs, k485378-a963e3bc_ids, k485378-a963e3bc_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:08,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9371) [2026-10-19 22:05:02,085] eagle_store.py:517 WARNING batch_get_buffer missing keys (k485381-8681586a_hs, k485381-8681586a_ids, k485381-8681586a_lhs); sleeping 0.50s. [repeated 4x across cluster] [2026-10-19 22:05:09,456] loop.py:45 INFO TRAIN_STEP step=15169 loss=1.323963 Training: 84%|████████▍ | 15169/18018 [32:34:55<6:05:47, 7.70s/step, loss=1.324, acc=0.699, acc_len=1.67, thru=23.4, I=16.0, T=35.2, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:09,611] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:14,706] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:05:15,136] loop.py:380 INFO TIMING step=15170: step=3.237s data=1.299s compute=3.209s [fwd=2.309s bwd=0.887s opt=0.003s] dispatch=2.433s [2026-10-19 22:05:15,136] loop.py:45 INFO TRAIN_STEP step=15170 loss=1.220482 Training: 84%|████████▍ | 15170/18018 [32:35:01<5:36:48, 7.10s/step, loss=1.220, acc=0.716, acc_len=1.74, thru=21.8, I=24.1, T=39.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:15,219] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (TrainerActor pid=9310) [2026-10-19 22:05:21,235] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15170: forward=2346.6ms backward=922.0ms [2026-10-19 22:05:21,239] loop.py:45 INFO TRAIN_STEP step=15171 loss=1.226259 Training: 84%|████████▍ | 15171/18018 [32:35:07<5:22:33, 6.80s/step, loss=1.226, acc=0.715, acc_len=1.73, thru=19.9, I=20.8, T=38.6, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:27,929] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:05:28,621] loop.py:45 INFO TRAIN_STEP step=15172 loss=1.349172 Training: 84%|████████▍ | 15172/18018 [32:35:14<5:30:46, 6.97s/step, loss=1.349, acc=0.689, acc_len=1.64, thru=18.9, I=19.0, T=36.8, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:28,936] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:05:34,424] loop.py:45 INFO TRAIN_STEP step=15173 loss=1.216869 Training: 84%|████████▍ | 15173/18018 [32:35:20<5:13:59, 6.62s/step, loss=1.217, acc=0.715, acc_len=1.73, thru=18.6, I=23.2, T=39.4, wait=2.5s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:42,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:05:43,903] loop.py:45 INFO TRAIN_STEP step=15174 loss=1.217070 Training: 84%|████████▍ | 15174/18018 [32:35:29<5:54:31, 7.48s/step, loss=1.217, acc=0.714, acc_len=1.73, thru=17.5, I=16.8, T=36.0, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:44,235] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:05:50,787] loop.py:380 INFO TIMING step=15175: step=3.237s data=1.274s compute=3.211s [fwd=2.262s bwd=0.934s opt=0.003s] dispatch=3.639s [2026-10-19 22:05:50,788] loop.py:45 INFO TRAIN_STEP step=15175 loss=1.169440 Training: 84%|████████▍ | 15175/18018 [32:35:36<5:45:55, 7.30s/step, loss=1.169, acc=0.724, acc_len=1.76, thru=16.0, I=20.6, T=39.5, wait=3.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:56,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:05:57,785] loop.py:45 INFO TRAIN_STEP step=15176 loss=1.293370 Training: 84%|████████▍ | 15176/18018 [32:35:43<5:41:29, 7.21s/step, loss=1.293, acc=0.699, acc_len=1.68, thru=19.1, I=21.1, T=34.9, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:05:57,781] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15175: forward=2627.1ms backward=990.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:05:58,244] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:06:03,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:06:04,195] loop.py:45 INFO TRAIN_STEP step=15177 loss=1.219376 Training: 84%|████████▍ | 15177/18018 [32:35:50<5:30:03, 6.97s/step, loss=1.219, acc=0.717, acc_len=1.74, thru=18.9, I=23.9, T=38.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:06:04,507] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:06:12,905] loop.py:45 INFO TRAIN_STEP step=15178 loss=1.287572 Training: 84%|████████▍ | 15178/18018 [32:35:58<5:54:37, 7.49s/step, loss=1.288, acc=0.702, acc_len=1.69, thru=10.8, I=15.4, T=31.8, wait=4.7s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:06:23,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:06:24,281] loop.py:45 INFO TRAIN_STEP step=15179 loss=1.328152 Training: 84%|████████▍ | 15179/18018 [32:36:10<6:49:38, 8.66s/step, loss=1.328, acc=0.695, acc_len=1.67, thru=16.5, I=14.1, T=32.3, wait=7.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:06:24,722] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:06:32,421] loop.py:380 INFO TIMING step=15180: step=3.459s data=1.379s compute=3.425s [fwd=2.450s bwd=0.960s opt=0.003s] dispatch=4.672s [2026-10-19 22:06:32,422] loop.py:45 INFO TRAIN_STEP step=15180 loss=1.137068 Training: 84%|████████▍ | 15180/18018 [32:36:18<6:42:08, 8.50s/step, loss=1.137, acc=0.728, acc_len=1.77, thru=11.3, I=14.1, T=37.0, wait=4.7s, pool=33, epoch=2/2][2026-10-19 22:06:41,148] loop.py:45 INFO TRAIN_STEP step=15181 loss=1.362342 Training: 84%|████████▍ | 15181/18018 [32:36:27<6:45:11, 8.57s/step, loss=1.362, acc=0.692, acc_len=1.65, thru=16.7, I=14.7, T=37.3, wait=5.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:06:41,143] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15180: forward=2450.3ms backward=934.2ms [2026-10-19 22:06:48,415] loop.py:45 INFO TRAIN_STEP step=15182 loss=1.235254 Training: 84%|████████▍ | 15182/18018 [32:36:34<6:26:34, 8.18s/step, loss=1.235, acc=0.712, acc_len=1.72, thru=16.8, I=17.6, T=35.4, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:06:54,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:06:55,706] loop.py:45 INFO TRAIN_STEP step=15183 loss=1.103762 Training: 84%|████████▍ | 15183/18018 [32:36:41<6:13:51, 7.91s/step, loss=1.104, acc=0.736, acc_len=1.81, thru=18.0, I=18.4, T=38.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:06:55,963] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:00,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:07:01,694] loop.py:45 INFO TRAIN_STEP step=15184 loss=1.244237 Training: 84%|████████▍ | 15184/18018 [32:36:47<5:46:27, 7.33s/step, loss=1.244, acc=0.707, acc_len=1.70, thru=21.8, I=24.6, T=36.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:01,820] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:07,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:07:08,368] loop.py:380 INFO TIMING step=15185: step=3.699s data=1.687s compute=3.643s [fwd=2.680s bwd=0.948s opt=0.003s] dispatch=2.966s [2026-10-19 22:07:08,369] loop.py:45 INFO TRAIN_STEP step=15185 loss=1.172745 Training: 84%|████████▍ | 15185/18018 [32:36:54<5:36:58, 7.14s/step, loss=1.173, acc=0.721, acc_len=1.76, thru=19.4, I=23.1, T=34.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:08,847] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:07:14,922] loop.py:45 INFO TRAIN_STEP step=15186 loss=1.249378 Training: 84%|████████▍ | 15186/18018 [32:37:01<5:28:39, 6.96s/step, loss=1.249, acc=0.706, acc_len=1.70, thru=16.4, I=19.5, T=38.4, wait=3.2s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:07:14,918] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15185: forward=2364.9ms backward=920.1ms [2026-10-19 22:07:22,174] loop.py:45 INFO TRAIN_STEP step=15187 loss=1.213488 Training: 84%|████████▍ | 15187/18018 [32:37:08<5:32:34, 7.05s/step, loss=1.213, acc=0.713, acc_len=1.73, thru=16.4, I=15.6, T=38.3, wait=3.9s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:31,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:07:32,146] loop.py:45 INFO TRAIN_STEP step=15188 loss=1.269218 Training: 84%|████████▍ | 15188/18018 [32:37:18<6:13:49, 7.93s/step, loss=1.269, acc=0.704, acc_len=1.69, thru=16.9, I=17.1, T=35.1, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:32,293] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:07:38,819] loop.py:45 INFO TRAIN_STEP step=15189 loss=1.369447 Training: 84%|████████▍ | 15189/18018 [32:37:24<5:55:58, 7.55s/step, loss=1.369, acc=0.689, acc_len=1.65, thru=17.6, I=17.8, T=36.3, wait=3.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:44,540] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:07:44,616] loop.py:380 INFO TIMING step=15190: step=3.318s data=1.328s compute=3.284s [fwd=2.360s bwd=0.910s opt=0.003s] dispatch=2.469s [2026-10-19 22:07:44,616] loop.py:45 INFO TRAIN_STEP step=15190 loss=1.174212 Training: 84%|████████▍ | 15190/18018 [32:37:30<5:31:02, 7.02s/step, loss=1.174, acc=0.722, acc_len=1.76, thru=21.3, I=21.0, T=38.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:45,042] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:51,074] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:07:51,303] loop.py:45 INFO TRAIN_STEP step=15191 loss=1.305042 Training: 84%|████████▍ | 15191/18018 [32:37:37<5:26:11, 6.92s/step, loss=1.305, acc=0.698, acc_len=1.68, thru=19.0, I=20.9, T=38.0, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:07:51,298] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15190: forward=2410.2ms backward=918.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:07:51,576] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:07:56,973] loop.py:45 INFO TRAIN_STEP step=15192 loss=1.328090 Training: 84%|████████▍ | 15192/18018 [32:37:43<5:08:22, 6.55s/step, loss=1.328, acc=0.699, acc_len=1.68, thru=19.6, I=21.6, T=39.4, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:04,181] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:08:05,285] loop.py:45 INFO TRAIN_STEP step=15193 loss=1.254715 Training: 84%|████████▍ | 15193/18018 [32:37:51<5:33:12, 7.08s/step, loss=1.255, acc=0.706, acc_len=1.70, thru=17.0, I=17.8, T=37.5, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:05,689] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:08:11,497] loop.py:45 INFO TRAIN_STEP step=15194 loss=1.321148 Training: 84%|████████▍ | 15194/18018 [32:37:57<5:20:51, 6.82s/step, loss=1.321, acc=0.696, acc_len=1.66, thru=19.4, I=26.4, T=39.0, wait=2.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:18,885] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:08:19,965] loop.py:380 INFO TIMING step=15195: step=3.649s data=1.596s compute=3.614s [fwd=2.666s bwd=0.934s opt=0.003s] dispatch=4.811s [2026-10-19 22:08:19,965] loop.py:45 INFO TRAIN_STEP step=15195 loss=1.302183 Training: 84%|████████▍ | 15195/18018 [32:38:06<5:44:02, 7.31s/step, loss=1.302, acc=0.702, acc_len=1.68, thru=14.9, I=16.2, T=35.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:20,394] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:08:25,835] loop.py:45 INFO TRAIN_STEP step=15196 loss=1.263357 Training: 84%|████████▍ | 15196/18018 [32:38:11<5:23:34, 6.88s/step, loss=1.263, acc=0.702, acc_len=1.69, thru=18.0, I=22.8, T=39.5, wait=2.6s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:08:25,831] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15195: forward=2291.4ms backward=902.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:33,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:08:34,494] loop.py:45 INFO TRAIN_STEP step=15197 loss=1.307803 Training: 84%|████████▍ | 15197/18018 [32:38:20<5:48:34, 7.41s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=18.1, I=16.6, T=37.5, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:34,562] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:40,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:08:40,371] loop.py:45 INFO TRAIN_STEP step=15198 loss=1.280885 Training: 84%|████████▍ | 15198/18018 [32:38:26<5:26:48, 6.95s/step, loss=1.281, acc=0.704, acc_len=1.69, thru=20.9, I=22.8, T=40.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:40,625] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:46,107] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:08:46,480] loop.py:45 INFO TRAIN_STEP step=15199 loss=1.292922 Training: 84%|████████▍ | 15199/18018 [32:38:32<5:14:44, 6.70s/step, loss=1.293, acc=0.699, acc_len=1.67, thru=21.3, I=22.5, T=38.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:46,609] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:52,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:08:53,822] loop.py:380 INFO TIMING step=15200: step=3.597s data=1.497s compute=3.569s [fwd=2.560s bwd=0.995s opt=0.003s] dispatch=3.736s [2026-10-19 22:08:53,822] loop.py:45 INFO TRAIN_STEP step=15200 loss=1.320961 Training: 84%|████████▍ | 15200/18018 [32:38:39<5:23:42, 6.89s/step, loss=1.321, acc=0.697, acc_len=1.67, thru=18.3, I=21.0, T=35.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:54,074] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:08:59,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:09:00,261] loop.py:45 INFO TRAIN_STEP step=15201 loss=1.243073 Training: 84%|████████▍ | 15201/18018 [32:38:46<5:17:11, 6.76s/step, loss=1.243, acc=0.710, acc_len=1.71, thru=19.7, I=21.5, T=34.5, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:09:00,257] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15200: forward=2700.8ms backward=955.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:00,459] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:05,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:09:06,853] loop.py:45 INFO TRAIN_STEP step=15202 loss=1.285635 Training: 84%|████████▍ | 15202/18018 [32:38:52<5:14:46, 6.71s/step, loss=1.286, acc=0.702, acc_len=1.69, thru=19.0, I=21.8, T=35.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:07,162] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:13,408] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:09:13,554] loop.py:45 INFO TRAIN_STEP step=15203 loss=1.261673 Training: 84%|████████▍ | 15203/18018 [32:38:59<5:14:35, 6.71s/step, loss=1.262, acc=0.707, acc_len=1.71, thru=18.8, I=22.6, T=38.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:13,910] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:20,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:09:20,320] loop.py:45 INFO TRAIN_STEP step=15204 loss=1.388092 Training: 84%|████████▍ | 15204/18018 [32:39:06<5:15:21, 6.72s/step, loss=1.388, acc=0.690, acc_len=1.64, thru=18.8, I=20.7, T=36.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:20,644] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:27,517] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:09:28,722] loop.py:380 INFO TIMING step=15205: step=3.681s data=1.406s compute=3.648s [fwd=2.644s bwd=0.988s opt=0.003s] dispatch=4.709s [2026-10-19 22:09:28,722] loop.py:45 INFO TRAIN_STEP step=15205 loss=1.267001 Training: 84%|████████▍ | 15205/18018 [32:39:14<5:38:48, 7.23s/step, loss=1.267, acc=0.709, acc_len=1.71, thru=15.6, I=17.0, T=34.8, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:29,025] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:09:36,207] loop.py:45 INFO TRAIN_STEP step=15206 loss=1.299114 Training: 84%|████████▍ | 15206/18018 [32:39:22<5:42:19, 7.30s/step, loss=1.299, acc=0.697, acc_len=1.67, thru=14.1, I=16.8, T=37.8, wait=4.1s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:09:36,203] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15205: forward=2413.7ms backward=926.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:42,748] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:09:43,621] loop.py:45 INFO TRAIN_STEP step=15207 loss=1.297152 Training: 84%|████████▍ | 15207/18018 [32:39:29<5:43:44, 7.34s/step, loss=1.297, acc=0.701, acc_len=1.68, thru=18.4, I=17.7, T=38.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:09:43,766] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:09:49,724] loop.py:45 INFO TRAIN_STEP step=15208 loss=1.319054 Training: 84%|████████▍ | 15208/18018 [32:39:35<5:26:17, 6.97s/step, loss=1.319, acc=0.697, acc_len=1.66, thru=19.7, I=23.3, T=37.7, wait=2.7s, pool=57, epoch=2/2][2026-10-19 22:10:00,357] loop.py:45 INFO TRAIN_STEP step=15209 loss=1.354513 Training: 84%|████████▍ | 15209/18018 [32:39:46<6:17:38, 8.07s/step, loss=1.355, acc=0.687, acc_len=1.63, thru=12.4, I=12.0, T=33.8, wait=6.8s, pool=57, epoch=2/2][2026-10-19 22:10:10,247] loop.py:380 INFO TIMING step=15210: step=3.565s data=1.443s compute=3.541s [fwd=2.548s bwd=0.979s opt=0.003s] dispatch=6.318s [2026-10-19 22:10:10,247] loop.py:45 INFO TRAIN_STEP step=15210 loss=1.208860 Training: 84%|████████▍ | 15210/18018 [32:39:56<6:43:06, 8.61s/step, loss=1.209, acc=0.715, acc_len=1.73, thru=13.6, I=12.2, T=35.9, wait=6.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:15,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:10:16,107] loop.py:45 INFO TRAIN_STEP step=15211 loss=1.274509 (TrainerActor pid=9310) [2026-10-19 22:10:16,103] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15210: forward=2419.4ms backward=983.1ms Training: 84%|████████▍ | 15211/18018 [32:40:02<6:04:21, 7.79s/step, loss=1.275, acc=0.706, acc_len=1.70, thru=20.8, I=20.7, T=37.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:16,262] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:22,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:10:22,709] loop.py:45 INFO TRAIN_STEP step=15212 loss=1.249024 Training: 84%|████████▍ | 15212/18018 [32:40:08<5:47:33, 7.43s/step, loss=1.249, acc=0.708, acc_len=1.70, thru=18.9, I=21.6, T=37.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:23,145] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:29,736] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:10:30,090] loop.py:45 INFO TRAIN_STEP step=15213 loss=1.450327 Training: 84%|████████▍ | 15213/18018 [32:40:16<5:46:43, 7.42s/step, loss=1.450, acc=0.671, acc_len=1.58, thru=17.6, I=18.8, T=37.1, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:30,248] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:10:36,068] loop.py:45 INFO TRAIN_STEP step=15214 loss=1.311545 Training: 84%|████████▍ | 15214/18018 [32:40:22<5:26:25, 6.98s/step, loss=1.312, acc=0.696, acc_len=1.66, thru=18.0, I=18.5, T=38.6, wait=2.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:43,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:10:44,626] loop.py:380 INFO TIMING step=15215: step=3.633s data=1.486s compute=3.600s [fwd=2.562s bwd=1.023s opt=0.003s] dispatch=4.917s [2026-10-19 22:10:44,626] loop.py:45 INFO TRAIN_STEP step=15215 loss=1.194842 Training: 84%|████████▍ | 15215/18018 [32:40:30<5:48:23, 7.46s/step, loss=1.195, acc=0.724, acc_len=1.77, thru=17.5, I=18.1, T=35.2, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:44,825] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:50,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:10:50,752] loop.py:45 INFO TRAIN_STEP step=15216 loss=1.270300 Training: 84%|████████▍ | 15216/18018 [32:40:36<5:29:34, 7.06s/step, loss=1.270, acc=0.702, acc_len=1.69, thru=20.1, I=23.9, T=39.2, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:10:50,748] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15215: forward=2337.8ms backward=885.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:10:51,061] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:10:56,680] loop.py:45 INFO TRAIN_STEP step=15217 loss=1.241376 Training: 84%|████████▍ | 15217/18018 [32:40:42<5:13:42, 6.72s/step, loss=1.241, acc=0.708, acc_len=1.71, thru=20.6, I=22.7, T=36.7, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:02,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:11:03,188] loop.py:45 INFO TRAIN_STEP step=15218 loss=1.217160 Training: 84%|████████▍ | 15218/18018 [32:40:49<5:10:34, 6.66s/step, loss=1.217, acc=0.715, acc_len=1.73, thru=19.5, I=20.8, T=37.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:03,238] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:09,404] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:11:10,197] loop.py:45 INFO TRAIN_STEP step=15219 loss=1.305575 Training: 84%|████████▍ | 15219/18018 [32:40:56<5:15:24, 6.76s/step, loss=1.306, acc=0.698, acc_len=1.66, thru=18.1, I=20.4, T=37.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:10,411] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:16,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:11:17,147] loop.py:380 INFO TIMING step=15220: step=3.340s data=1.367s compute=3.305s [fwd=2.363s bwd=0.928s opt=0.003s] dispatch=3.601s [2026-10-19 22:11:17,147] loop.py:45 INFO TRAIN_STEP step=15220 loss=1.235169 Training: 84%|████████▍ | 15220/18018 [32:41:03<5:17:58, 6.82s/step, loss=1.235, acc=0.710, acc_len=1.72, thru=18.6, I=21.5, T=38.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:17,228] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:21,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:11:22,930] loop.py:45 INFO TRAIN_STEP step=15221 loss=1.133315 Training: 84%|████████▍ | 15221/18018 [32:41:09<5:03:19, 6.51s/step, loss=1.133, acc=0.729, acc_len=1.78, thru=21.2, I=28.1, T=38.6, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:11:22,926] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15220: forward=2361.2ms backward=911.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:23,434] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:28,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:11:28,675] loop.py:45 INFO TRAIN_STEP step=15222 loss=1.149867 Training: 84%|████████▍ | 15222/18018 [32:41:14<4:52:34, 6.28s/step, loss=1.150, acc=0.728, acc_len=1.78, thru=21.5, I=29.6, T=40.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:28,724] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 22:11:36,415] loop.py:45 INFO TRAIN_STEP step=15223 loss=1.189392 Training: 84%|████████▍ | 15223/18018 [32:41:22<5:12:53, 6.72s/step, loss=1.189, acc=0.721, acc_len=1.75, thru=15.3, I=16.7, T=36.5, wait=4.2s, pool=49, epoch=2/2][2026-10-19 22:11:45,835] loop.py:45 INFO TRAIN_STEP step=15224 loss=1.259084 Training: 84%|████████▍ | 15224/18018 [32:41:31<5:50:34, 7.53s/step, loss=1.259, acc=0.708, acc_len=1.71, thru=15.1, I=14.4, T=36.5, wait=5.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:52,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:11:52,983] loop.py:380 INFO TIMING step=15225: step=3.571s data=1.529s compute=3.486s [fwd=2.437s bwd=1.035s opt=0.003s] dispatch=3.567s [2026-10-19 22:11:52,984] loop.py:45 INFO TRAIN_STEP step=15225 loss=1.302325 Training: 84%|████████▍ | 15225/18018 [32:41:39<5:45:09, 7.41s/step, loss=1.302, acc=0.696, acc_len=1.68, thru=18.5, I=18.9, T=35.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:11:53,075] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 22:11:59,254] loop.py:45 INFO TRAIN_STEP step=15226 loss=1.266546 Training: 85%|████████▍ | 15226/18018 [32:41:45<5:29:03, 7.07s/step, loss=1.267, acc=0.706, acc_len=1.70, thru=19.4, I=23.6, T=38.5, wait=2.9s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:11:59,250] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15225: forward=2371.4ms backward=907.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:07,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:12:07,613] loop.py:45 INFO TRAIN_STEP step=15227 loss=1.285718 Training: 85%|████████▍ | 15227/18018 [32:41:53<5:46:53, 7.46s/step, loss=1.286, acc=0.700, acc_len=1.68, thru=16.6, I=16.8, T=36.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:07,839] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:12:13,242] loop.py:45 INFO TRAIN_STEP step=15228 loss=1.119731 Training: 85%|████████▍ | 15228/18018 [32:41:59<5:21:15, 6.91s/step, loss=1.120, acc=0.735, acc_len=1.81, thru=20.8, I=22.8, T=38.7, wait=2.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:22,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:12:22,637] loop.py:45 INFO TRAIN_STEP step=15229 loss=1.327990 Training: 85%|████████▍ | 15229/18018 [32:42:08<5:55:47, 7.65s/step, loss=1.328, acc=0.695, acc_len=1.66, thru=14.5, I=14.9, T=34.8, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:23,076] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:29,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:12:29,966] loop.py:380 INFO TIMING step=15230: step=3.473s data=1.449s compute=3.431s [fwd=2.468s bwd=0.948s opt=0.003s] dispatch=3.849s [2026-10-19 22:12:29,966] loop.py:45 INFO TRAIN_STEP step=15230 loss=1.379176 Training: 85%|████████▍ | 15230/18018 [32:42:16<5:51:07, 7.56s/step, loss=1.379, acc=0.691, acc_len=1.65, thru=17.8, I=18.8, T=36.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:30,374] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:35,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:12:35,779] loop.py:45 INFO TRAIN_STEP step=15231 loss=1.189151 Training: 85%|████████▍ | 15231/18018 [32:42:21<5:26:44, 7.03s/step, loss=1.189, acc=0.720, acc_len=1.75, thru=21.3, I=23.8, T=38.5, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:12:35,774] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15230: forward=2327.4ms backward=934.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:36,069] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:41,218] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:12:41,841] loop.py:45 INFO TRAIN_STEP step=15232 loss=1.068056 Training: 85%|████████▍ | 15232/18018 [32:42:27<5:13:03, 6.74s/step, loss=1.068, acc=0.743, acc_len=1.83, thru=21.2, I=23.8, T=38.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:42,223] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:47,974] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:12:49,083] loop.py:45 INFO TRAIN_STEP step=15233 loss=1.269682 Training: 85%|████████▍ | 15233/18018 [32:42:35<5:19:54, 6.89s/step, loss=1.270, acc=0.704, acc_len=1.69, thru=17.9, I=20.8, T=36.2, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:12:49,482] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:12:54,516] loop.py:45 INFO TRAIN_STEP step=15234 loss=1.257858 Training: 85%|████████▍ | 15234/18018 [32:42:40<4:59:28, 6.45s/step, loss=1.258, acc=0.708, acc_len=1.71, thru=21.6, I=29.7, T=40.0, wait=2.2s, pool=57, epoch=2/2][2026-10-19 22:13:02,189] loop.py:380 INFO TIMING step=15235: step=3.491s data=1.384s compute=3.454s [fwd=2.470s bwd=0.970s opt=0.003s] dispatch=4.175s [2026-10-19 22:13:02,189] loop.py:45 INFO TRAIN_STEP step=15235 loss=1.254654 Training: 85%|████████▍ | 15235/18018 [32:42:48<5:16:19, 6.82s/step, loss=1.255, acc=0.709, acc_len=1.71, thru=14.6, I=15.3, T=36.7, wait=4.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:13:09,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:13:10,738] loop.py:45 INFO TRAIN_STEP step=15236 loss=1.330337 Training: 85%|████████▍ | 15236/18018 [32:42:56<5:40:16, 7.34s/step, loss=1.330, acc=0.692, acc_len=1.66, thru=17.5, I=17.6, T=36.3, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:13:10,734] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15235: forward=2499.2ms backward=974.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:13:11,132] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:13:16,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:13:17,065] loop.py:45 INFO TRAIN_STEP step=15237 loss=1.277512 Training: 85%|████████▍ | 15237/18018 [32:43:03<5:26:06, 7.04s/step, loss=1.278, acc=0.704, acc_len=1.69, thru=19.8, I=24.4, T=36.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:13:17,437] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:13:24,450] loop.py:45 INFO TRAIN_STEP step=15238 loss=1.250187 Training: 85%|████████▍ | 15238/18018 [32:43:10<5:30:48, 7.14s/step, loss=1.250, acc=0.710, acc_len=1.72, thru=16.3, I=18.8, T=38.8, wait=4.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:13:34,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:13:35,326] loop.py:45 INFO TRAIN_STEP step=15239 loss=1.140874 Training: 85%|████████▍ | 15239/18018 [32:43:21<6:22:37, 8.26s/step, loss=1.141, acc=0.727, acc_len=1.77, thru=14.0, I=12.8, T=34.5, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:13:35,466] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 22:13:44,660] loop.py:380 INFO TIMING step=15240: step=3.770s data=1.591s compute=3.735s [fwd=2.608s bwd=1.113s opt=0.003s] dispatch=5.554s [2026-10-19 22:13:44,661] loop.py:45 INFO TRAIN_STEP step=15240 loss=1.341886 Training: 85%|████████▍ | 15240/18018 [32:43:30<6:37:25, 8.58s/step, loss=1.342, acc=0.689, acc_len=1.64, thru=11.7, I=15.2, T=34.0, wait=5.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:13:52,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:13:52,998] loop.py:45 INFO TRAIN_STEP step=15241 loss=1.296144 Training: 85%|████████▍ | 15241/18018 [32:43:39<6:33:49, 8.51s/step, loss=1.296, acc=0.699, acc_len=1.68, thru=16.2, I=16.4, T=35.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:13:53,032] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 22:13:52,994] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15240: forward=2546.3ms backward=1024.1ms [2026-10-19 22:14:00,426] loop.py:45 INFO TRAIN_STEP step=15242 loss=1.267908 Training: 85%|████████▍ | 15242/18018 [32:43:46<6:18:41, 8.18s/step, loss=1.268, acc=0.706, acc_len=1.70, thru=16.8, I=18.5, T=37.6, wait=4.0s, pool=57, epoch=2/2][2026-10-19 22:14:07,152] loop.py:45 INFO TRAIN_STEP step=15243 loss=1.224420 Training: 85%|████████▍ | 15243/18018 [32:43:53<5:58:19, 7.75s/step, loss=1.224, acc=0.715, acc_len=1.73, thru=17.0, I=16.7, T=37.9, wait=3.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:14:15,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:14:15,896] loop.py:45 INFO TRAIN_STEP step=15244 loss=1.242126 Training: 85%|████████▍ | 15244/18018 [32:44:01<6:11:59, 8.05s/step, loss=1.242, acc=0.711, acc_len=1.72, thru=17.0, I=17.6, T=37.8, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:14:16,061] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:14:22,819] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:14:23,475] loop.py:380 INFO TIMING step=15245: step=3.446s data=1.445s compute=3.401s [fwd=2.402s bwd=0.985s opt=0.003s] dispatch=4.125s [2026-10-19 22:14:23,475] loop.py:45 INFO TRAIN_STEP step=15245 loss=1.237768 Training: 85%|████████▍ | 15245/18018 [32:44:09<6:05:23, 7.91s/step, loss=1.238, acc=0.708, acc_len=1.71, thru=17.3, I=19.7, T=37.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:14:23,823] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:14:30,317] loop.py:45 INFO TRAIN_STEP step=15246 loss=1.272552 Training: 85%|████████▍ | 15246/18018 [32:44:16<5:50:31, 7.59s/step, loss=1.273, acc=0.702, acc_len=1.68, thru=15.8, I=19.2, T=36.5, wait=3.3s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:14:30,313] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15245: forward=2471.4ms backward=969.6ms [2026-10-19 22:14:38,667] loop.py:45 INFO TRAIN_STEP step=15247 loss=1.318002 Training: 85%|████████▍ | 15247/18018 [32:44:24<6:00:56, 7.82s/step, loss=1.318, acc=0.698, acc_len=1.68, thru=15.5, I=16.0, T=35.7, wait=4.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:14:46,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:14:47,306] loop.py:45 INFO TRAIN_STEP step=15248 loss=1.251887 Training: 85%|████████▍ | 15248/18018 [32:44:33<6:12:14, 8.06s/step, loss=1.252, acc=0.709, acc_len=1.71, thru=16.4, I=15.8, T=36.1, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:14:47,376] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:14:53,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:14:53,216] loop.py:45 INFO TRAIN_STEP step=15249 loss=1.238892 Training: 85%|████████▍ | 15249/18018 [32:44:39<5:42:15, 7.42s/step, loss=1.239, acc=0.711, acc_len=1.72, thru=20.6, I=21.6, T=38.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:14:53,542] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:15:00,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:15:00,577] loop.py:380 INFO TIMING step=15250: step=3.465s data=1.491s compute=3.420s [fwd=2.428s bwd=0.977s opt=0.003s] dispatch=3.889s [2026-10-19 22:15:00,577] loop.py:45 INFO TRAIN_STEP step=15250 loss=1.382048 Training: 85%|████████▍ | 15250/18018 [32:44:46<5:41:22, 7.40s/step, loss=1.382, acc=0.686, acc_len=1.64, thru=17.8, I=18.7, T=36.9, wait=3.9s, pool=65, epoch=2/2][2026-10-19 22:15:00,580] loop.py:419 INFO Saving checkpoint at step 15250... (TrainerActor pid=9310) [2026-10-19 22:15:04,015] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0015251 (TrainerActor pid=9310) [2026-10-19 22:15:04,153] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0014501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:15:05,442] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:15:05,442] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=1/64 [2026-10-19 22:15:11,639] loop.py:45 INFO TRAIN_STEP step=15251 loss=1.333510 Training: 85%|████████▍ | 15251/18018 [32:44:57<6:31:54, 8.50s/step, loss=1.334, acc=0.697, acc_len=1.67, thru=18.1, I=17.5, T=37.6, wait=3.0s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:15:11,635] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15250: forward=2410.8ms backward=952.5ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 22:15:04,017] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0015251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 22:15:05,186] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0014501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:15:19,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:15:19,606] loop.py:45 INFO TRAIN_STEP step=15252 loss=1.345435 Training: 85%|████████▍ | 15252/18018 [32:45:05<6:24:25, 8.34s/step, loss=1.345, acc=0.692, acc_len=1.66, thru=19.3, I=17.3, T=36.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:15:20,066] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:15:26,046] loop.py:45 INFO TRAIN_STEP step=15253 loss=1.249424 Training: 85%|████████▍ | 15253/18018 [32:45:12<5:58:04, 7.77s/step, loss=1.249, acc=0.706, acc_len=1.70, thru=16.2, I=21.3, T=37.5, wait=3.0s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:15:33,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:15:35,270] loop.py:45 INFO TRAIN_STEP step=15254 loss=1.288350 Training: 85%|████████▍ | 15254/18018 [32:45:21<6:18:01, 8.21s/step, loss=1.288, acc=0.700, acc_len=1.68, thru=16.9, I=16.6, T=30.8, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:15:35,320] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 22:15:41,581] loop.py:380 INFO TIMING step=15255: step=3.379s data=1.366s compute=3.340s [fwd=2.388s bwd=0.938s opt=0.003s] dispatch=2.922s [2026-10-19 22:15:41,581] loop.py:45 INFO TRAIN_STEP step=15255 loss=1.266805 Training: 85%|████████▍ | 15255/18018 [32:45:27<5:51:41, 7.64s/step, loss=1.267, acc=0.702, acc_len=1.69, thru=18.2, I=21.2, T=37.9, wait=2.9s, pool=57, epoch=2/2][2026-10-19 22:15:50,322] loop.py:45 INFO TRAIN_STEP step=15256 loss=1.272027 Training: 85%|████████▍ | 15256/18018 [32:45:36<6:06:47, 7.97s/step, loss=1.272, acc=0.705, acc_len=1.70, thru=13.6, I=15.3, T=36.1, wait=5.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:15:50,317] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15255: forward=2525.7ms backward=978.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:15:58,108] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:15:58,706] loop.py:45 INFO TRAIN_STEP step=15257 loss=1.226895 Training: 85%|████████▍ | 15257/18018 [32:45:44<6:12:24, 8.09s/step, loss=1.227, acc=0.713, acc_len=1.73, thru=17.2, I=16.2, T=33.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:15:59,112] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:16:06,040] loop.py:45 INFO TRAIN_STEP step=15258 loss=1.258543 Training: 85%|████████▍ | 15258/18018 [32:45:52<6:01:48, 7.87s/step, loss=1.259, acc=0.706, acc_len=1.70, thru=15.8, I=16.8, T=37.9, wait=3.9s, pool=45, epoch=2/2][2026-10-19 22:16:13,298] loop.py:45 INFO TRAIN_STEP step=15259 loss=1.266269 Training: 85%|████████▍ | 15259/18018 [32:45:59<5:53:18, 7.68s/step, loss=1.266, acc=0.705, acc_len=1.69, thru=18.6, I=17.5, T=38.1, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:16:19,340] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:16:19,634] loop.py:380 INFO TIMING step=15260: step=3.275s data=1.277s compute=3.249s [fwd=2.356s bwd=0.878s opt=0.003s] dispatch=3.052s [2026-10-19 22:16:19,634] loop.py:45 INFO TRAIN_STEP step=15260 loss=1.138893 Training: 85%|████████▍ | 15260/18018 [32:46:05<5:34:35, 7.28s/step, loss=1.139, acc=0.727, acc_len=1.77, thru=20.1, I=20.3, T=39.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:16:19,841] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:16:26,340] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:16:27,015] loop.py:45 INFO TRAIN_STEP step=15261 loss=1.286328 Training: 85%|████████▍ | 15261/18018 [32:46:13<5:35:53, 7.31s/step, loss=1.286, acc=0.701, acc_len=1.68, thru=18.1, I=20.2, T=37.5, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:16:27,011] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15260: forward=2435.1ms backward=931.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:16:27,344] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:16:32,967] loop.py:45 INFO TRAIN_STEP step=15262 loss=1.165551 Training: 85%|████████▍ | 15262/18018 [32:46:19<5:17:03, 6.90s/step, loss=1.166, acc=0.722, acc_len=1.75, thru=19.2, I=23.2, T=39.9, wait=2.7s, pool=37, epoch=2/2][2026-10-19 22:16:42,844] loop.py:45 INFO TRAIN_STEP step=15263 loss=1.314011 Training: 85%|████████▍ | 15263/18018 [32:46:28<5:57:55, 7.80s/step, loss=1.314, acc=0.699, acc_len=1.68, thru=16.0, I=14.4, T=35.3, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:16:42,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:16:43,351] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:16:49,433] loop.py:45 INFO TRAIN_STEP step=15264 loss=1.379222 Training: 85%|████████▍ | 15264/18018 [32:46:35<5:41:09, 7.43s/step, loss=1.379, acc=0.685, acc_len=1.62, thru=18.1, I=18.3, T=38.5, wait=3.3s, pool=53, epoch=2/2][2026-10-19 22:16:57,562] loop.py:380 INFO TIMING step=15265: step=3.386s data=1.338s compute=3.356s [fwd=2.420s bwd=0.921s opt=0.003s] dispatch=4.735s [2026-10-19 22:16:57,563] loop.py:45 INFO TRAIN_STEP step=15265 loss=1.247192 Training: 85%|████████▍ | 15265/18018 [32:46:43<5:50:39, 7.64s/step, loss=1.247, acc=0.709, acc_len=1.71, thru=12.6, I=12.0, T=37.8, wait=4.7s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:17:05,870] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:17:06,902] loop.py:45 INFO TRAIN_STEP step=15266 loss=1.228038 Training: 85%|████████▍ | 15266/18018 [32:46:52<6:13:51, 8.15s/step, loss=1.228, acc=0.711, acc_len=1.72, thru=18.8, I=16.2, T=33.6, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:17:06,897] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15265: forward=2682.2ms backward=1082.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:17:07,383] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:17:13,432] loop.py:45 INFO TRAIN_STEP step=15267 loss=1.251623 Training: 85%|████████▍ | 15267/18018 [32:46:59<5:51:27, 7.67s/step, loss=1.252, acc=0.709, acc_len=1.72, thru=18.0, I=24.4, T=38.3, wait=3.2s, pool=57, epoch=2/2][2026-10-19 22:17:20,080] loop.py:45 INFO TRAIN_STEP step=15268 loss=1.230042 Training: 85%|████████▍ | 15268/18018 [32:47:06<5:37:18, 7.36s/step, loss=1.230, acc=0.710, acc_len=1.72, thru=19.5, I=20.2, T=37.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:17:20,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:17:20,521] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:17:26,778] loop.py:45 INFO TRAIN_STEP step=15269 loss=1.231188 Training: 85%|████████▍ | 15269/18018 [32:47:12<5:28:08, 7.16s/step, loss=1.231, acc=0.711, acc_len=1.72, thru=16.0, I=19.4, T=38.4, wait=3.4s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:17:34,531] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:17:35,851] loop.py:380 INFO TIMING step=15270: step=3.697s data=1.553s compute=3.654s [fwd=2.606s bwd=1.034s opt=0.003s] dispatch=5.365s [2026-10-19 22:17:35,852] loop.py:45 INFO TRAIN_STEP step=15270 loss=1.322005 Training: 85%|████████▍ | 15270/18018 [32:47:21<5:54:14, 7.73s/step, loss=1.322, acc=0.697, acc_len=1.67, thru=17.0, I=17.9, T=34.6, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:17:36,039] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:17:41,522] loop.py:45 INFO TRAIN_STEP step=15271 loss=1.260784 Training: 85%|████████▍ | 15271/18018 [32:47:27<5:25:46, 7.12s/step, loss=1.261, acc=0.706, acc_len=1.70, thru=19.9, I=25.7, T=38.6, wait=2.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:17:41,517] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15270: forward=2372.7ms backward=903.1ms [2026-10-19 22:17:50,993] loop.py:45 INFO TRAIN_STEP step=15272 loss=1.264536 Training: 85%|████████▍ | 15272/18018 [32:47:37<5:57:58, 7.82s/step, loss=1.265, acc=0.705, acc_len=1.69, thru=11.2, I=11.7, T=34.7, wait=5.8s, pool=33, epoch=2/2][2026-10-19 22:17:59,201] loop.py:45 INFO TRAIN_STEP step=15273 loss=1.176040 Training: 85%|████████▍ | 15273/18018 [32:47:45<6:03:09, 7.94s/step, loss=1.176, acc=0.721, acc_len=1.76, thru=16.4, I=15.3, T=34.6, wait=4.5s, pool=49, epoch=2/2][2026-10-19 22:18:06,527] loop.py:45 INFO TRAIN_STEP step=15274 loss=1.298244 Training: 85%|████████▍ | 15274/18018 [32:47:52<5:54:39, 7.75s/step, loss=1.298, acc=0.700, acc_len=1.68, thru=16.3, I=16.2, T=36.6, wait=3.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:18:12,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:18:13,487] loop.py:380 INFO TIMING step=15275: step=3.371s data=1.392s compute=3.338s [fwd=2.399s bwd=0.926s opt=0.003s] dispatch=3.580s [2026-10-19 22:18:13,487] loop.py:45 INFO TRAIN_STEP step=15275 loss=1.300066 Training: 85%|████████▍ | 15275/18018 [32:47:59<5:43:37, 7.52s/step, loss=1.300, acc=0.698, acc_len=1.67, thru=19.0, I=20.3, T=38.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:18:13,764] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:18:22,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:18:23,504] loop.py:45 INFO TRAIN_STEP step=15276 loss=1.218370 Training: 85%|████████▍ | 15276/18018 [32:48:09<6:17:46, 8.27s/step, loss=1.218, acc=0.714, acc_len=1.73, thru=13.8, I=13.4, T=33.6, wait=6.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:18:23,499] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15275: forward=2621.7ms backward=1131.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:18:23,882] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:18:30,983] loop.py:45 INFO TRAIN_STEP step=15277 loss=1.230918 Training: 85%|████████▍ | 15277/18018 [32:48:17<6:06:48, 8.03s/step, loss=1.231, acc=0.711, acc_len=1.72, thru=16.5, I=18.2, T=37.6, wait=4.1s, pool=57, epoch=2/2][2026-10-19 22:18:38,996] loop.py:45 INFO TRAIN_STEP step=15278 loss=1.251510 Training: 85%|████████▍ | 15278/18018 [32:48:25<6:06:27, 8.02s/step, loss=1.252, acc=0.705, acc_len=1.70, thru=16.5, I=14.7, T=37.4, wait=4.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:18:48,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:18:48,905] loop.py:45 INFO TRAIN_STEP step=15279 loss=1.330809 Training: 85%|████████▍ | 15279/18018 [32:48:34<6:32:09, 8.59s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=15.1, I=16.3, T=35.3, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:18:49,225] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:18:55,075] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:18:56,305] loop.py:380 INFO TIMING step=15280: step=3.715s data=1.724s compute=3.677s [fwd=2.710s bwd=0.952s opt=0.003s] dispatch=3.674s [2026-10-19 22:18:56,305] loop.py:45 INFO TRAIN_STEP step=15280 loss=1.298772 Training: 85%|████████▍ | 15280/18018 [32:48:42<6:15:43, 8.23s/step, loss=1.299, acc=0.696, acc_len=1.67, thru=17.1, I=20.0, T=34.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:18:56,586] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:19:05,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:19:05,696] loop.py:45 INFO TRAIN_STEP step=15281 loss=1.257631 Training: 85%|████████▍ | 15281/18018 [32:48:51<6:31:24, 8.58s/step, loss=1.258, acc=0.708, acc_len=1.71, thru=14.1, I=15.6, T=34.6, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:19:05,691] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15280: forward=2550.2ms backward=1107.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:19:06,036] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:19:13,367] loop.py:45 INFO TRAIN_STEP step=15282 loss=1.233552 Training: 85%|████████▍ | 15282/18018 [32:48:59<6:18:50, 8.31s/step, loss=1.234, acc=0.709, acc_len=1.71, thru=16.3, I=17.4, T=35.7, wait=4.1s, pool=57, epoch=2/2][2026-10-19 22:19:21,266] loop.py:45 INFO TRAIN_STEP step=15283 loss=1.217772 Training: 85%|████████▍ | 15283/18018 [32:49:07<6:13:04, 8.18s/step, loss=1.218, acc=0.714, acc_len=1.73, thru=15.3, I=16.7, T=37.2, wait=4.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:19:29,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:19:29,701] loop.py:45 INFO TRAIN_STEP step=15284 loss=1.153515 Training: 85%|████████▍ | 15284/18018 [32:49:15<6:16:22, 8.26s/step, loss=1.154, acc=0.725, acc_len=1.77, thru=17.4, I=15.7, T=35.2, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:19:29,878] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:19:35,624] loop.py:380 INFO TIMING step=15285: step=3.256s data=1.290s compute=3.221s [fwd=2.315s bwd=0.892s opt=0.003s] dispatch=2.659s [2026-10-19 22:19:35,625] loop.py:45 INFO TRAIN_STEP step=15285 loss=1.272604 Training: 85%|████████▍ | 15285/18018 [32:49:21<5:44:19, 7.56s/step, loss=1.273, acc=0.704, acc_len=1.70, thru=18.5, I=19.3, T=39.3, wait=2.7s, pool=45, epoch=2/2][2026-10-19 22:19:42,797] loop.py:45 INFO TRAIN_STEP step=15286 loss=1.296497 Training: 85%|████████▍ | 15286/18018 [32:49:28<5:38:55, 7.44s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=15.9, I=17.8, T=36.5, wait=3.7s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:19:42,793] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15285: forward=2438.2ms backward=1010.3ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 22:19:42,890] eagle_store.py:517 WARNING batch_get_buffer missing keys (k489150-af2c5d55_hs, k489150-af2c5d55_ids, k489150-af2c5d55_lhs); sleeping 0.50s. [2026-10-19 22:19:56,497] loop.py:45 INFO TRAIN_STEP step=15287 loss=1.313516 Training: 85%|████████▍ | 15287/18018 [32:49:42<7:04:13, 9.32s/step, loss=1.314, acc=0.701, acc_len=1.68, thru=11.8, I=11.7, T=28.2, wait=9.2s, pool=61, epoch=2/2][2026-10-19 22:20:02,783] loop.py:45 INFO TRAIN_STEP step=15288 loss=1.252840 Training: 85%|████████▍ | 15288/18018 [32:49:48<6:22:38, 8.41s/step, loss=1.253, acc=0.707, acc_len=1.71, thru=16.1, I=16.0, T=36.1, wait=2.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:08,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 22:19:42,876] eagle_store.py:517 WARNING batch_get_buffer missing keys (k489149-9dc2c328_hs, k489149-9dc2c328_ids, k489149-9dc2c328_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 22:20:09,752] loop.py:45 INFO TRAIN_STEP step=15289 loss=1.359700 Training: 85%|████████▍ | 15289/18018 [32:49:55<6:02:51, 7.98s/step, loss=1.360, acc=0.691, acc_len=1.65, thru=18.5, I=22.1, T=31.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:09,834] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:15,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:20:15,764] loop.py:380 INFO TIMING step=15290: step=3.394s data=1.343s compute=3.365s [fwd=2.413s bwd=0.938s opt=0.003s] dispatch=2.609s [2026-10-19 22:20:15,765] loop.py:45 INFO TRAIN_STEP step=15290 loss=1.230355 Training: 85%|████████▍ | 15290/18018 [32:50:01<5:35:55, 7.39s/step, loss=1.230, acc=0.710, acc_len=1.72, thru=19.3, I=22.6, T=37.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:16,074] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:22,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:20:24,457] loop.py:45 INFO TRAIN_STEP step=15291 loss=1.300940 Training: 85%|████████▍ | 15291/18018 [32:50:10<5:53:34, 7.78s/step, loss=1.301, acc=0.697, acc_len=1.67, thru=15.2, I=21.8, T=31.7, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:20:24,453] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15290: forward=3029.4ms backward=957.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:24,688] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-19 22:20:30,603] loop.py:45 INFO TRAIN_STEP step=15292 loss=1.187780 Training: 85%|████████▍ | 15292/18018 [32:50:16<5:31:09, 7.29s/step, loss=1.188, acc=0.721, acc_len=1.75, thru=17.3, I=23.9, T=39.2, wait=2.9s, pool=53, epoch=2/2][2026-10-19 22:20:36,895] loop.py:45 INFO TRAIN_STEP step=15293 loss=1.240421 Training: 85%|████████▍ | 15293/18018 [32:50:22<5:17:28, 6.99s/step, loss=1.240, acc=0.710, acc_len=1.72, thru=18.6, I=20.3, T=38.0, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:43,868] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:20:44,896] loop.py:45 INFO TRAIN_STEP step=15294 loss=1.266293 Training: 85%|████████▍ | 15294/18018 [32:50:30<5:31:06, 7.29s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=17.5, I=18.7, T=36.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:45,376] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:50,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:20:51,225] loop.py:380 INFO TIMING step=15295: step=3.307s data=1.344s compute=3.271s [fwd=2.353s bwd=0.903s opt=0.003s] dispatch=3.014s [2026-10-19 22:20:51,226] loop.py:45 INFO TRAIN_STEP step=15295 loss=1.208980 Training: 85%|████████▍ | 15295/18018 [32:50:37<5:17:53, 7.00s/step, loss=1.209, acc=0.717, acc_len=1.74, thru=19.9, I=24.1, T=38.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:51,254] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:20:59,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 22:21:00,307] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15295: forward=2703.3ms backward=1100.9ms [2026-10-19 22:21:00,311] loop.py:45 INFO TRAIN_STEP step=15296 loss=1.378996 Training: 85%|████████▍ | 15296/18018 [32:50:46<5:46:06, 7.63s/step, loss=1.379, acc=0.688, acc_len=1.64, thru=14.5, I=15.6, T=33.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:21:00,344] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 22:21:07,245] loop.py:45 INFO TRAIN_STEP step=15297 loss=1.125236 Training: 85%|████████▍ | 15297/18018 [32:50:53<5:36:29, 7.42s/step, loss=1.125, acc=0.733, acc_len=1.80, thru=14.6, I=21.5, T=34.5, wait=3.2s, pool=33, epoch=2/2][2026-10-19 22:21:16,348] loop.py:45 INFO TRAIN_STEP step=15298 loss=1.302637 Training: 85%|████████▍ | 15298/18018 [32:51:02<5:59:16, 7.93s/step, loss=1.303, acc=0.701, acc_len=1.68, thru=16.6, I=13.4, T=35.0, wait=5.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:21:24,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:21:24,476] loop.py:45 INFO TRAIN_STEP step=15299 loss=1.280741 Training: 85%|████████▍ | 15299/18018 [32:51:10<6:01:54, 7.99s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=16.6, I=16.8, T=36.2, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:21:24,565] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 22:21:30,982] loop.py:380 INFO TIMING step=15300: step=3.380s data=1.349s compute=3.349s [fwd=2.395s bwd=0.940s opt=0.003s] dispatch=3.117s [2026-10-19 22:21:30,982] loop.py:45 INFO TRAIN_STEP step=15300 loss=1.173979 Training: 85%|████████▍ | 15300/18018 [32:51:17<5:41:38, 7.54s/step, loss=1.174, acc=0.722, acc_len=1.75, thru=16.9, I=18.6, T=37.9, wait=3.1s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:21:37,751] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:21:38,464] loop.py:45 INFO TRAIN_STEP step=15301 loss=1.294101 Training: 85%|████████▍ | 15301/18018 [32:51:24<5:40:42, 7.52s/step, loss=1.294, acc=0.702, acc_len=1.69, thru=18.6, I=19.0, T=34.9, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:21:38,459] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15300: forward=2725.7ms backward=903.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:21:38,754] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:21:44,325] loop.py:45 INFO TRAIN_STEP step=15302 loss=1.137260 Training: 85%|████████▍ | 15302/18018 [32:51:30<5:18:00, 7.03s/step, loss=1.137, acc=0.726, acc_len=1.78, thru=19.1, I=23.3, T=38.7, wait=2.5s, pool=49, epoch=2/2][2026-10-19 22:21:54,051] loop.py:45 INFO TRAIN_STEP step=15303 loss=1.254393 Training: 85%|████████▍ | 15303/18018 [32:51:40<5:54:31, 7.83s/step, loss=1.254, acc=0.710, acc_len=1.72, thru=13.7, I=13.5, T=34.1, wait=6.0s, pool=53, epoch=2/2][2026-10-19 22:22:01,527] loop.py:45 INFO TRAIN_STEP step=15304 loss=1.282053 Training: 85%|████████▍ | 15304/18018 [32:51:47<5:49:35, 7.73s/step, loss=1.282, acc=0.703, acc_len=1.69, thru=15.8, I=16.9, T=37.2, wait=4.0s, pool=45, epoch=2/2][2026-10-19 22:22:10,984] loop.py:380 INFO TIMING step=15305: step=3.587s data=1.432s compute=3.559s [fwd=2.548s bwd=0.996s opt=0.003s] dispatch=5.858s [2026-10-19 22:22:10,984] loop.py:45 INFO TRAIN_STEP step=15305 loss=1.370560 Training: 85%|████████▍ | 15305/18018 [32:51:57<6:12:52, 8.25s/step, loss=1.371, acc=0.689, acc_len=1.64, thru=14.3, I=13.5, T=35.7, wait=5.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:22:18,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:22:18,713] loop.py:45 INFO TRAIN_STEP step=15306 loss=1.368579 Training: 85%|████████▍ | 15306/18018 [32:52:04<6:05:44, 8.09s/step, loss=1.369, acc=0.686, acc_len=1.63, thru=17.0, I=18.6, T=35.7, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:22:18,709] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15305: forward=2507.4ms backward=1022.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:22:18,745] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:22:25,498] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:22:25,844] loop.py:45 INFO TRAIN_STEP step=15307 loss=1.095567 Training: 85%|████████▍ | 15307/18018 [32:52:11<5:52:34, 7.80s/step, loss=1.096, acc=0.739, acc_len=1.82, thru=17.6, I=19.2, T=36.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:22:26,010] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:22:32,952] loop.py:45 INFO TRAIN_STEP step=15308 loss=1.152131 Training: 85%|████████▍ | 15308/18018 [32:52:19<5:43:00, 7.59s/step, loss=1.152, acc=0.726, acc_len=1.77, thru=17.2, I=17.2, T=37.3, wait=3.7s, pool=57, epoch=2/2][2026-10-19 22:22:39,961] loop.py:45 INFO TRAIN_STEP step=15309 loss=1.186136 Training: 85%|████████▍ | 15309/18018 [32:52:26<5:34:55, 7.42s/step, loss=1.186, acc=0.721, acc_len=1.75, thru=17.3, I=16.8, T=37.7, wait=3.6s, pool=53, epoch=2/2][2026-10-19 22:22:47,062] loop.py:380 INFO TIMING step=15310: step=4.147s data=1.282s compute=4.113s [fwd=2.725s bwd=1.374s opt=0.003s] dispatch=2.948s [2026-10-19 22:22:47,062] loop.py:45 INFO TRAIN_STEP step=15310 loss=1.222037 Training: 85%|████████▍ | 15310/18018 [32:52:34<5:53:27, 7.83s/step, loss=1.222, acc=0.716, acc_len=1.73, thru=15.2, I=19.7, T=30.9, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:22:55,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:22:55,306] loop.py:45 INFO TRAIN_STEP step=15311 loss=1.237322 Training: 85%|████████▍ | 15311/18018 [32:52:41<5:36:00, 7.45s/step, loss=1.237, acc=0.709, acc_len=1.72, thru=23.2, I=22.3, T=38.6, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:22:55,302] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15310: forward=2378.8ms backward=894.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:22:55,526] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:23:02,525] loop.py:45 INFO TRAIN_STEP step=15312 loss=1.233032 Training: 85%|████████▍ | 15312/18018 [32:52:48<5:32:45, 7.38s/step, loss=1.233, acc=0.709, acc_len=1.71, thru=16.1, I=20.4, T=37.5, wait=3.8s, pool=49, epoch=2/2][2026-10-19 22:23:12,596] loop.py:45 INFO TRAIN_STEP step=15313 loss=1.266006 Training: 85%|████████▍ | 15313/18018 [32:52:58<6:09:04, 8.19s/step, loss=1.266, acc=0.709, acc_len=1.70, thru=14.3, I=12.8, T=34.7, wait=6.4s, pool=61, epoch=2/2][2026-10-19 22:23:19,376] loop.py:45 INFO TRAIN_STEP step=15314 loss=1.227353 Training: 85%|████████▍ | 15314/18018 [32:53:05<5:49:55, 7.76s/step, loss=1.227, acc=0.713, acc_len=1.73, thru=18.2, I=18.6, T=37.4, wait=3.4s, pool=57, epoch=2/2][2026-10-19 22:23:26,402] loop.py:380 INFO TIMING step=15315: step=3.441s data=1.353s compute=3.400s [fwd=2.421s bwd=0.965s opt=0.003s] dispatch=3.576s [2026-10-19 22:23:26,402] loop.py:45 INFO TRAIN_STEP step=15315 loss=1.155756 Training: 85%|████████▍ | 15315/18018 [32:53:12<5:39:52, 7.54s/step, loss=1.156, acc=0.726, acc_len=1.76, thru=17.9, I=19.3, T=37.2, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:23:34,197] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:23:35,072] loop.py:45 INFO TRAIN_STEP step=15316 loss=1.192688 Training: 85%|████████▌ | 15316/18018 [32:53:21<5:54:54, 7.88s/step, loss=1.193, acc=0.720, acc_len=1.75, thru=15.7, I=15.6, T=33.4, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:23:35,068] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15315: forward=2662.4ms backward=1121.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:23:35,203] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 22:23:43,217] loop.py:45 INFO TRAIN_STEP step=15317 loss=1.275980 Training: 85%|████████▌ | 15317/18018 [32:53:29<5:58:20, 7.96s/step, loss=1.276, acc=0.704, acc_len=1.70, thru=12.8, I=16.3, T=37.1, wait=4.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:23:49,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:23:50,470] loop.py:45 INFO TRAIN_STEP step=15318 loss=1.190920 Training: 85%|████████▌ | 15318/18018 [32:53:36<5:48:38, 7.75s/step, loss=1.191, acc=0.716, acc_len=1.74, thru=17.9, I=19.2, T=36.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:23:50,949] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:23:58,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:23:59,073] loop.py:45 INFO TRAIN_STEP step=15319 loss=1.274216 Training: 85%|████████▌ | 15319/18018 [32:53:45<6:00:04, 8.00s/step, loss=1.274, acc=0.698, acc_len=1.67, thru=14.6, I=18.5, T=37.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:23:59,107] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 22:24:08,866] loop.py:380 INFO TIMING step=15320: step=3.632s data=1.482s compute=3.596s [fwd=2.551s bwd=1.032s opt=0.003s] dispatch=6.150s [2026-10-19 22:24:08,866] loop.py:45 INFO TRAIN_STEP step=15320 loss=1.203063 Training: 85%|████████▌ | 15320/18018 [32:53:54<6:24:01, 8.54s/step, loss=1.203, acc=0.718, acc_len=1.74, thru=13.1, I=12.6, T=35.2, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:24:08,825] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:24:09,328] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:24:15,552] loop.py:45 INFO TRAIN_STEP step=15321 loss=1.231872 Training: 85%|████████▌ | 15321/18018 [32:54:01<5:58:52, 7.98s/step, loss=1.232, acc=0.714, acc_len=1.73, thru=17.3, I=17.3, T=37.1, wait=3.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:24:15,548] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15320: forward=2443.4ms backward=960.7ms [2026-10-19 22:24:22,600] loop.py:45 INFO TRAIN_STEP step=15322 loss=1.262069 Training: 85%|████████▌ | 15322/18018 [32:54:08<5:46:07, 7.70s/step, loss=1.262, acc=0.708, acc_len=1.71, thru=16.0, I=19.2, T=36.9, wait=3.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:24:32,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:24:32,190] loop.py:45 INFO TRAIN_STEP step=15323 loss=1.103837 Training: 85%|████████▌ | 15323/18018 [32:54:18<6:11:25, 8.27s/step, loss=1.104, acc=0.735, acc_len=1.80, thru=15.7, I=14.2, T=33.1, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:24:32,544] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:24:38,782] loop.py:45 INFO TRAIN_STEP step=15324 loss=1.399081 Training: 85%|████████▌ | 15324/18018 [32:54:24<5:48:42, 7.77s/step, loss=1.399, acc=0.684, acc_len=1.62, thru=17.2, I=18.1, T=36.9, wait=3.1s, pool=57, epoch=2/2][2026-10-19 22:24:44,886] loop.py:380 INFO TIMING step=15325: step=3.316s data=1.347s compute=3.288s [fwd=2.366s bwd=0.908s opt=0.003s] dispatch=2.779s [2026-10-19 22:24:44,886] loop.py:45 INFO TRAIN_STEP step=15325 loss=1.281440 Training: 85%|████████▌ | 15325/18018 [32:54:30<5:26:11, 7.27s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=21.0, I=20.9, T=38.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:24:44,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:24:45,349] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:24:53,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:24:54,255] loop.py:45 INFO TRAIN_STEP step=15326 loss=1.320026 Training: 85%|████████▌ | 15326/18018 [32:54:40<5:54:21, 7.90s/step, loss=1.320, acc=0.695, acc_len=1.66, thru=14.4, I=15.7, T=33.4, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:24:54,250] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15325: forward=2740.2ms backward=1028.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:24:54,382] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 22:25:01,840] loop.py:45 INFO TRAIN_STEP step=15327 loss=1.383325 Training: 85%|████████▌ | 15327/18018 [32:54:47<5:50:01, 7.80s/step, loss=1.383, acc=0.686, acc_len=1.63, thru=15.5, I=17.8, T=37.2, wait=4.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:25:11,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:25:12,046] loop.py:45 INFO TRAIN_STEP step=15328 loss=1.212436 Training: 85%|████████▌ | 15328/18018 [32:54:58<6:22:11, 8.52s/step, loss=1.212, acc=0.716, acc_len=1.73, thru=14.9, I=13.5, T=33.1, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:25:12,201] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:25:20,031] loop.py:45 INFO TRAIN_STEP step=15329 loss=1.175246 Training: 85%|████████▌ | 15329/18018 [32:55:06<6:14:47, 8.36s/step, loss=1.175, acc=0.724, acc_len=1.77, thru=12.8, I=14.4, T=35.5, wait=4.4s, pool=33, epoch=2/2][2026-10-19 22:25:29,368] loop.py:380 INFO TIMING step=15330: step=4.043s data=1.770s compute=3.987s [fwd=2.753s bwd=1.220s opt=0.003s] dispatch=5.285s [2026-10-19 22:25:29,369] loop.py:45 INFO TRAIN_STEP step=15330 loss=1.125948 Training: 85%|████████▌ | 15330/18018 [32:55:15<6:27:45, 8.66s/step, loss=1.126, acc=0.733, acc_len=1.80, thru=16.1, I=15.2, T=31.7, wait=5.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:25:35,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:25:36,141] loop.py:45 INFO TRAIN_STEP step=15331 loss=1.257558 Training: 85%|████████▌ | 15331/18018 [32:55:22<6:02:18, 8.09s/step, loss=1.258, acc=0.709, acc_len=1.71, thru=17.6, I=19.0, T=36.4, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:25:36,137] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15330: forward=2530.4ms backward=935.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:25:36,171] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 22:25:42,270] loop.py:45 INFO TRAIN_STEP step=15332 loss=1.284336 Training: 85%|████████▌ | 15332/18018 [32:55:28<5:35:49, 7.50s/step, loss=1.284, acc=0.701, acc_len=1.68, thru=17.7, I=19.7, T=38.7, wait=2.8s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:25:51,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:25:51,658] loop.py:45 INFO TRAIN_STEP step=15333 loss=1.273802 Training: 85%|████████▌ | 15333/18018 [32:55:37<6:01:01, 8.07s/step, loss=1.274, acc=0.705, acc_len=1.70, thru=16.9, I=17.6, T=35.8, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:25:52,157] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:25:57,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:25:57,732] loop.py:45 INFO TRAIN_STEP step=15334 loss=1.316795 Training: 85%|████████▌ | 15334/18018 [32:55:43<5:34:08, 7.47s/step, loss=1.317, acc=0.696, acc_len=1.66, thru=20.2, I=22.9, T=38.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:25:58,059] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:26:03,599] loop.py:380 INFO TIMING step=15335: step=3.278s data=1.341s compute=3.248s [fwd=2.335s bwd=0.899s opt=0.003s] dispatch=2.581s [2026-10-19 22:26:03,599] loop.py:45 INFO TRAIN_STEP step=15335 loss=1.248229 Training: 85%|████████▌ | 15335/18018 [32:55:49<5:12:31, 6.99s/step, loss=1.248, acc=0.711, acc_len=1.72, thru=21.2, I=22.1, T=39.1, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:11,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:26:12,259] loop.py:45 INFO TRAIN_STEP step=15336 loss=1.271265 Training: 85%|████████▌ | 15336/18018 [32:55:58<5:34:49, 7.49s/step, loss=1.271, acc=0.708, acc_len=1.71, thru=16.8, I=16.4, T=36.0, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:26:12,255] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15335: forward=2554.1ms backward=950.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:12,459] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:17,320] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:26:18,394] loop.py:45 INFO TRAIN_STEP step=15337 loss=1.412492 Training: 85%|████████▌ | 15337/18018 [32:56:04<5:16:32, 7.08s/step, loss=1.412, acc=0.677, acc_len=1.61, thru=20.9, I=23.0, T=39.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:18,829] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:24,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:26:25,856] loop.py:45 INFO TRAIN_STEP step=15338 loss=1.186756 Training: 85%|████████▌ | 15338/18018 [32:56:11<5:21:29, 7.20s/step, loss=1.187, acc=0.722, acc_len=1.75, thru=18.4, I=22.8, T=38.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:25,962] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:32,323] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:26:32,365] loop.py:45 INFO TRAIN_STEP step=15339 loss=1.286293 Training: 85%|████████▌ | 15339/18018 [32:56:18<5:12:06, 6.99s/step, loss=1.286, acc=0.700, acc_len=1.68, thru=19.9, I=20.0, T=36.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:32,825] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:26:40,989] loop.py:380 INFO TIMING step=15340: step=3.365s data=1.366s compute=3.335s [fwd=2.367s bwd=0.954s opt=0.003s] dispatch=5.252s [2026-10-19 22:26:40,989] loop.py:45 INFO TRAIN_STEP step=15340 loss=1.263660 Training: 85%|████████▌ | 15340/18018 [32:56:27<5:33:52, 7.48s/step, loss=1.264, acc=0.708, acc_len=1.71, thru=14.6, I=14.8, T=38.0, wait=5.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:48,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:26:49,265] loop.py:45 INFO TRAIN_STEP step=15341 loss=1.176996 Training: 85%|████████▌ | 15341/18018 [32:56:35<5:44:25, 7.72s/step, loss=1.177, acc=0.723, acc_len=1.76, thru=17.2, I=18.5, T=33.7, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:26:49,260] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15340: forward=2750.6ms backward=1008.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:49,624] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:53,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:26:55,088] loop.py:45 INFO TRAIN_STEP step=15342 loss=1.245998 Training: 85%|████████▌ | 15342/18018 [32:56:41<5:18:54, 7.15s/step, loss=1.246, acc=0.705, acc_len=1.69, thru=20.3, I=24.8, T=39.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:26:55,551] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 22:27:01,673] loop.py:45 INFO TRAIN_STEP step=15343 loss=1.229453 Training: 85%|████████▌ | 15343/18018 [32:56:47<5:11:14, 6.98s/step, loss=1.229, acc=0.712, acc_len=1.72, thru=17.4, I=22.1, T=39.8, wait=3.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:27:08,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:27:10,091] loop.py:45 INFO TRAIN_STEP step=15344 loss=1.299947 Training: 85%|████████▌ | 15344/18018 [32:56:56<5:30:19, 7.41s/step, loss=1.300, acc=0.705, acc_len=1.70, thru=18.1, I=18.7, T=36.8, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:27:10,246] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 22:27:15,707] loop.py:380 INFO TIMING step=15345: step=3.245s data=1.303s compute=3.216s [fwd=2.299s bwd=0.903s opt=0.003s] dispatch=2.363s [2026-10-19 22:27:15,708] loop.py:45 INFO TRAIN_STEP step=15345 loss=1.037082 Training: 85%|████████▌ | 15345/18018 [32:57:01<5:06:13, 6.87s/step, loss=1.037, acc=0.748, acc_len=1.85, thru=17.6, I=22.1, T=39.4, wait=2.4s, pool=29, epoch=2/2][2026-10-19 22:27:26,790] loop.py:45 INFO TRAIN_STEP step=15346 loss=1.271464 Training: 85%|████████▌ | 15346/18018 [32:57:12<6:02:20, 8.14s/step, loss=1.271, acc=0.707, acc_len=1.70, thru=15.8, I=13.3, T=35.6, wait=7.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:27:26,785] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15345: forward=2554.1ms backward=993.9ms [2026-10-19 22:27:34,369] loop.py:45 INFO TRAIN_STEP step=15347 loss=1.332292 Training: 85%|████████▌ | 15347/18018 [32:57:20<5:54:43, 7.97s/step, loss=1.332, acc=0.691, acc_len=1.65, thru=17.1, I=16.2, T=35.8, wait=4.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:27:40,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:27:40,787] loop.py:45 INFO TRAIN_STEP step=15348 loss=1.095756 Training: 85%|████████▌ | 15348/18018 [32:57:26<5:33:55, 7.50s/step, loss=1.096, acc=0.734, acc_len=1.80, thru=19.7, I=20.3, T=37.5, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:27:41,114] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:27:46,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:27:46,361] loop.py:45 INFO TRAIN_STEP step=15349 loss=1.262120 Training: 85%|████████▌ | 15349/18018 [32:57:32<5:08:03, 6.93s/step, loss=1.262, acc=0.711, acc_len=1.71, thru=21.5, I=22.4, T=38.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:27:46,704] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:27:52,674] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:27:52,804] loop.py:380 INFO TIMING step=15350: step=3.251s data=1.294s compute=3.227s [fwd=2.343s bwd=0.870s opt=0.003s] dispatch=3.182s [2026-10-19 22:27:52,804] loop.py:45 INFO TRAIN_STEP step=15350 loss=1.243470 Training: 85%|████████▌ | 15350/18018 [32:57:38<5:01:28, 6.78s/step, loss=1.243, acc=0.711, acc_len=1.72, thru=20.5, I=22.0, T=39.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:27:53,175] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:28:00,179] loop.py:45 INFO TRAIN_STEP step=15351 loss=1.257771 Training: 85%|████████▌ | 15351/18018 [32:57:46<5:09:17, 6.96s/step, loss=1.258, acc=0.704, acc_len=1.69, thru=18.2, I=18.0, T=37.2, wait=3.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:28:00,173] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15350: forward=2466.2ms backward=924.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:05,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:28:06,833] loop.py:45 INFO TRAIN_STEP step=15352 loss=1.333653 Training: 85%|████████▌ | 15352/18018 [32:57:52<5:05:08, 6.87s/step, loss=1.334, acc=0.692, acc_len=1.65, thru=19.6, I=22.2, T=36.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:06,974] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:11,339] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:28:12,335] loop.py:45 INFO TRAIN_STEP step=15353 loss=1.146998 Training: 85%|████████▌ | 15353/18018 [32:57:58<4:46:50, 6.46s/step, loss=1.147, acc=0.726, acc_len=1.77, thru=21.3, I=27.7, T=39.3, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:12,370] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 22:28:20,123] loop.py:45 INFO TRAIN_STEP step=15354 loss=1.244073 Training: 85%|████████▌ | 15354/18018 [32:58:06<5:04:27, 6.86s/step, loss=1.244, acc=0.712, acc_len=1.72, thru=14.4, I=18.5, T=36.2, wait=4.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:31,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:28:32,322] loop.py:380 INFO TIMING step=15355: step=4.059s data=1.736s compute=3.710s [fwd=2.638s bwd=1.056s opt=0.003s] dispatch=8.130s [2026-10-19 22:28:32,323] loop.py:45 INFO TRAIN_STEP step=15355 loss=1.221549 Training: 85%|████████▌ | 15355/18018 [32:58:18<6:15:27, 8.46s/step, loss=1.222, acc=0.713, acc_len=1.72, thru=12.5, I=12.5, T=31.5, wait=8.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:32,550] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:41,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:28:42,198] loop.py:45 INFO TRAIN_STEP step=15356 loss=1.389668 Training: 85%|████████▌ | 15356/18018 [32:58:28<6:34:09, 8.88s/step, loss=1.390, acc=0.685, acc_len=1.62, thru=13.3, I=14.0, T=32.7, wait=6.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:28:42,194] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15355: forward=2799.6ms backward=1072.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:42,670] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:28:49,352] loop.py:45 INFO TRAIN_STEP step=15357 loss=1.160638 Training: 85%|████████▌ | 15357/18018 [32:58:35<6:10:58, 8.36s/step, loss=1.161, acc=0.722, acc_len=1.75, thru=14.5, I=16.2, T=38.6, wait=3.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:58,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:28:59,814] loop.py:45 INFO TRAIN_STEP step=15358 loss=1.253166 Training: 85%|████████▌ | 15358/18018 [32:58:45<6:38:44, 8.99s/step, loss=1.253, acc=0.709, acc_len=1.72, thru=14.0, I=15.3, T=35.2, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:28:59,857] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 22:29:05,359] loop.py:45 INFO TRAIN_STEP step=15359 loss=1.321095 Training: 85%|████████▌ | 15359/18018 [32:58:51<5:52:46, 7.96s/step, loss=1.321, acc=0.693, acc_len=1.65, thru=20.5, I=18.5, T=38.7, wait=2.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:29:10,711] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:29:11,977] loop.py:380 INFO TIMING step=15360: step=3.578s data=1.427s compute=3.544s [fwd=2.492s bwd=1.036s opt=0.003s] dispatch=3.028s [2026-10-19 22:29:11,977] loop.py:45 INFO TRAIN_STEP step=15360 loss=1.288051 Training: 85%|████████▌ | 15360/18018 [32:58:58<5:34:47, 7.56s/step, loss=1.288, acc=0.706, acc_len=1.70, thru=19.7, I=22.4, T=35.8, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:29:12,220] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:29:17,664] loop.py:45 INFO TRAIN_STEP step=15361 loss=1.364232 Training: 85%|████████▌ | 15361/18018 [32:59:03<5:09:47, 7.00s/step, loss=1.364, acc=0.686, acc_len=1.63, thru=19.6, I=23.6, T=39.5, wait=2.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:29:17,659] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15360: forward=2292.2ms backward=903.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:29:24,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:29:24,873] loop.py:45 INFO TRAIN_STEP step=15362 loss=1.320003 Training: 85%|████████▌ | 15362/18018 [32:59:10<5:12:30, 7.06s/step, loss=1.320, acc=0.697, acc_len=1.67, thru=19.5, I=18.0, T=37.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:29:25,272] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:29:30,502] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:29:31,370] loop.py:45 INFO TRAIN_STEP step=15363 loss=1.412742 Training: 85%|████████▌ | 15363/18018 [32:59:17<5:04:56, 6.89s/step, loss=1.413, acc=0.682, acc_len=1.62, thru=19.7, I=21.5, T=38.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:29:31,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 22:29:39,318] loop.py:45 INFO TRAIN_STEP step=15364 loss=1.216200 Training: 85%|████████▌ | 15364/18018 [32:59:25<5:18:49, 7.21s/step, loss=1.216, acc=0.715, acc_len=1.73, thru=14.6, I=16.7, T=35.5, wait=4.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:29:45,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:29:46,139] loop.py:380 INFO TIMING step=15365: step=3.538s data=1.481s compute=3.507s [fwd=2.491s bwd=1.002s opt=0.003s] dispatch=3.276s [2026-10-19 22:29:46,139] loop.py:45 INFO TRAIN_STEP step=15365 loss=1.365172 Training: 85%|████████▌ | 15365/18018 [32:59:32<5:13:36, 7.09s/step, loss=1.365, acc=0.689, acc_len=1.64, thru=18.9, I=18.9, T=36.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:29:46,420] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:29:53,289] loop.py:45 INFO TRAIN_STEP step=15366 loss=1.328348 Training: 85%|████████▌ | 15366/18018 [32:59:39<5:14:12, 7.11s/step, loss=1.328, acc=0.693, acc_len=1.65, thru=17.5, I=19.4, T=37.4, wait=3.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:29:53,285] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15365: forward=2412.6ms backward=962.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:00,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:30:00,867] loop.py:45 INFO TRAIN_STEP step=15367 loss=1.316993 Training: 85%|████████▌ | 15367/18018 [32:59:46<5:20:20, 7.25s/step, loss=1.317, acc=0.698, acc_len=1.68, thru=18.4, I=19.3, T=35.3, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:01,022] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:06,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:30:06,461] loop.py:45 INFO TRAIN_STEP step=15368 loss=1.224585 Training: 85%|████████▌ | 15368/18018 [32:59:52<4:58:14, 6.75s/step, loss=1.225, acc=0.715, acc_len=1.73, thru=21.0, I=22.3, T=40.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:06,912] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:12,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:30:12,519] loop.py:45 INFO TRAIN_STEP step=15369 loss=1.205258 Training: 85%|████████▌ | 15369/18018 [32:59:58<4:48:57, 6.54s/step, loss=1.205, acc=0.717, acc_len=1.74, thru=22.2, I=23.3, T=37.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:12,663] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:21,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:30:21,833] loop.py:380 INFO TIMING step=15370: step=3.846s data=1.533s compute=3.815s [fwd=2.721s bwd=1.079s opt=0.003s] dispatch=5.458s [2026-10-19 22:30:21,833] loop.py:45 INFO TRAIN_STEP step=15370 loss=1.380601 Training: 85%|████████▌ | 15370/18018 [33:00:07<5:25:29, 7.38s/step, loss=1.381, acc=0.685, acc_len=1.62, thru=13.6, I=13.9, T=33.3, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:22,273] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:30:28,323] loop.py:45 INFO TRAIN_STEP step=15371 loss=1.084429 Training: 85%|████████▌ | 15371/18018 [33:00:14<5:13:41, 7.11s/step, loss=1.084, acc=0.738, acc_len=1.81, thru=18.3, I=20.4, T=39.0, wait=3.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:30:28,319] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15370: forward=2314.0ms backward=919.7ms [2026-10-19 22:30:37,129] loop.py:45 INFO TRAIN_STEP step=15372 loss=1.150545 Training: 85%|████████▌ | 15372/18018 [33:00:23<5:35:58, 7.62s/step, loss=1.151, acc=0.727, acc_len=1.77, thru=12.9, I=12.6, T=37.5, wait=5.4s, pool=41, epoch=2/2][2026-10-19 22:30:48,655] loop.py:45 INFO TRAIN_STEP step=15373 loss=1.202364 Training: 85%|████████▌ | 15373/18018 [33:00:34<6:27:33, 8.79s/step, loss=1.202, acc=0.722, acc_len=1.75, thru=13.0, I=12.3, T=33.2, wait=7.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:56,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:30:57,628] loop.py:45 INFO TRAIN_STEP step=15374 loss=1.349199 Training: 85%|████████▌ | 15374/18018 [33:00:43<6:29:47, 8.85s/step, loss=1.349, acc=0.693, acc_len=1.65, thru=15.8, I=16.2, T=31.7, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:30:57,679] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 22:31:03,854] loop.py:380 INFO TIMING step=15375: step=3.295s data=1.366s compute=3.257s [fwd=2.352s bwd=0.891s opt=0.003s] dispatch=2.922s [2026-10-19 22:31:03,855] loop.py:45 INFO TRAIN_STEP step=15375 loss=1.263369 Training: 85%|████████▌ | 15375/18018 [33:00:49<5:55:01, 8.06s/step, loss=1.263, acc=0.701, acc_len=1.68, thru=18.1, I=20.8, T=38.8, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:31:11,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:31:11,988] loop.py:45 INFO TRAIN_STEP step=15376 loss=1.225564 Training: 85%|████████▌ | 15376/18018 [33:00:58<5:55:53, 8.08s/step, loss=1.226, acc=0.715, acc_len=1.73, thru=16.5, I=17.4, T=36.3, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:31:11,983] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15375: forward=2499.7ms backward=979.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:31:12,412] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:31:17,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:31:17,959] loop.py:45 INFO TRAIN_STEP step=15377 loss=1.232453 Training: 85%|████████▌ | 15377/18018 [33:01:04<5:27:53, 7.45s/step, loss=1.232, acc=0.707, acc_len=1.71, thru=20.5, I=25.4, T=38.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:31:18,021] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 22:31:24,176] loop.py:45 INFO TRAIN_STEP step=15378 loss=1.253945 Training: 85%|████████▌ | 15378/18018 [33:01:10<5:11:29, 7.08s/step, loss=1.254, acc=0.710, acc_len=1.71, thru=18.0, I=22.1, T=37.8, wait=2.8s, pool=41, epoch=2/2][2026-10-19 22:31:33,489] loop.py:45 INFO TRAIN_STEP step=15379 loss=1.434693 Training: 85%|████████▌ | 15379/18018 [33:01:19<5:40:49, 7.75s/step, loss=1.435, acc=0.680, acc_len=1.61, thru=14.5, I=13.6, T=32.9, wait=5.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:31:40,332] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:31:40,379] loop.py:380 INFO TIMING step=15380: step=3.660s data=1.486s compute=3.614s [fwd=2.519s bwd=1.079s opt=0.003s] dispatch=3.220s [2026-10-19 22:31:40,379] loop.py:45 INFO TRAIN_STEP step=15380 loss=1.367103 Training: 85%|████████▌ | 15380/18018 [33:01:26<5:29:20, 7.49s/step, loss=1.367, acc=0.686, acc_len=1.63, thru=19.4, I=17.6, T=35.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:31:40,835] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:31:47,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:31:47,454] loop.py:45 INFO TRAIN_STEP step=15381 loss=1.261239 Training: 85%|████████▌ | 15381/18018 [33:01:33<5:23:45, 7.37s/step, loss=1.261, acc=0.706, acc_len=1.69, thru=18.4, I=18.2, T=36.5, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:31:47,450] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15380: forward=2514.1ms backward=950.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:31:47,791] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:31:53,844] loop.py:45 INFO TRAIN_STEP step=15382 loss=1.215303 Training: 85%|████████▌ | 15382/18018 [33:01:39<5:10:45, 7.07s/step, loss=1.215, acc=0.716, acc_len=1.74, thru=18.9, I=20.0, T=37.0, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:31:59,673] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:32:00,264] loop.py:45 INFO TRAIN_STEP step=15383 loss=1.277023 Training: 85%|████████▌ | 15383/18018 [33:01:46<5:02:03, 6.88s/step, loss=1.277, acc=0.707, acc_len=1.71, thru=19.8, I=20.9, T=36.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:00,678] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:05,799] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:32:06,561] loop.py:45 INFO TRAIN_STEP step=15384 loss=1.282534 Training: 85%|████████▌ | 15384/18018 [33:01:52<4:54:16, 6.70s/step, loss=1.283, acc=0.702, acc_len=1.68, thru=19.9, I=22.4, T=37.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:06,805] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:32:14,857] loop.py:380 INFO TIMING step=15385: step=3.663s data=1.486s compute=3.619s [fwd=2.559s bwd=1.045s opt=0.003s] dispatch=4.624s [2026-10-19 22:32:14,857] loop.py:45 INFO TRAIN_STEP step=15385 loss=1.269034 Training: 85%|████████▌ | 15385/18018 [33:02:00<5:15:07, 7.18s/step, loss=1.269, acc=0.708, acc_len=1.71, thru=13.3, I=15.0, T=34.9, wait=4.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:22,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:32:22,445] loop.py:45 INFO TRAIN_STEP step=15386 loss=1.306508 Training: 85%|████████▌ | 15386/18018 [33:02:08<5:20:21, 7.30s/step, loss=1.307, acc=0.697, acc_len=1.67, thru=18.0, I=17.9, T=37.7, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:32:22,440] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15385: forward=2404.2ms backward=947.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:22,790] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:27,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:32:28,723] loop.py:45 INFO TRAIN_STEP step=15387 loss=1.239429 Training: 85%|████████▌ | 15387/18018 [33:02:14<5:06:45, 7.00s/step, loss=1.239, acc=0.709, acc_len=1.71, thru=20.4, I=23.3, T=38.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:28,966] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:32,954] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:32:34,037] loop.py:45 INFO TRAIN_STEP step=15388 loss=1.069419 Training: 85%|████████▌ | 15388/18018 [33:02:20<4:44:33, 6.49s/step, loss=1.069, acc=0.740, acc_len=1.82, thru=22.8, I=28.2, T=40.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:34,463] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:32:41,914] loop.py:45 INFO TRAIN_STEP step=15389 loss=1.174199 Training: 85%|████████▌ | 15389/18018 [33:02:27<5:02:38, 6.91s/step, loss=1.174, acc=0.719, acc_len=1.75, thru=14.5, I=17.1, T=37.6, wait=4.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:47,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:32:48,125] loop.py:380 INFO TIMING step=15390: step=3.387s data=1.415s compute=3.350s [fwd=2.420s bwd=0.914s opt=0.003s] dispatch=2.815s [2026-10-19 22:32:48,125] loop.py:45 INFO TRAIN_STEP step=15390 loss=1.275739 Training: 85%|████████▌ | 15390/18018 [33:02:34<4:53:21, 6.70s/step, loss=1.276, acc=0.703, acc_len=1.70, thru=20.6, I=19.7, T=37.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:48,600] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:54,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:32:55,133] loop.py:45 INFO TRAIN_STEP step=15391 loss=1.280236 Training: 85%|████████▌ | 15391/18018 [33:02:41<4:57:20, 6.79s/step, loss=1.280, acc=0.698, acc_len=1.68, thru=18.6, I=19.8, T=37.4, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:32:55,128] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15390: forward=2449.5ms backward=934.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:32:55,458] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:33:01,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:33:01,622] loop.py:45 INFO TRAIN_STEP step=15392 loss=1.234140 Training: 85%|████████▌ | 15392/18018 [33:02:47<4:53:15, 6.70s/step, loss=1.234, acc=0.708, acc_len=1.70, thru=19.7, I=20.3, T=38.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:33:02,044] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:33:07,010] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:33:07,622] loop.py:45 INFO TRAIN_STEP step=15393 loss=1.257138 Training: 85%|████████▌ | 15393/18018 [33:02:53<4:43:57, 6.49s/step, loss=1.257, acc=0.708, acc_len=1.71, thru=21.3, I=22.6, T=36.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:33:08,015] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:33:13,528] loop.py:45 INFO TRAIN_STEP step=15394 loss=1.254167 Training: 85%|████████▌ | 15394/18018 [33:02:59<4:36:17, 6.32s/step, loss=1.254, acc=0.706, acc_len=1.70, thru=19.6, I=22.9, T=39.2, wait=2.6s, pool=53, epoch=2/2][2026-10-19 22:33:22,097] loop.py:380 INFO TIMING step=15395: step=3.375s data=1.349s compute=3.342s [fwd=2.403s bwd=0.924s opt=0.003s] dispatch=5.177s [2026-10-19 22:33:22,097] loop.py:45 INFO TRAIN_STEP step=15395 loss=1.243418 Training: 85%|████████▌ | 15395/18018 [33:03:08<5:05:35, 6.99s/step, loss=1.243, acc=0.708, acc_len=1.71, thru=14.6, I=14.8, T=37.9, wait=5.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:33:30,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 22:33:31,620] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15395: forward=2512.3ms backward=979.1ms [2026-10-19 22:33:31,625] loop.py:45 INFO TRAIN_STEP step=15396 loss=1.254692 Training: 85%|████████▌ | 15396/18018 [33:03:17<5:38:45, 7.75s/step, loss=1.255, acc=0.711, acc_len=1.72, thru=15.8, I=15.5, T=36.1, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:33:31,723] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 22:33:38,656] loop.py:45 INFO TRAIN_STEP step=15397 loss=1.262983 Training: 85%|████████▌ | 15397/18018 [33:03:24<5:29:09, 7.54s/step, loss=1.263, acc=0.708, acc_len=1.71, thru=17.3, I=20.0, T=38.2, wait=3.7s, pool=53, epoch=2/2][2026-10-19 22:33:47,675] loop.py:45 INFO TRAIN_STEP step=15398 loss=1.245075 Training: 85%|████████▌ | 15398/18018 [33:03:33<5:48:29, 7.98s/step, loss=1.245, acc=0.707, acc_len=1.70, thru=13.6, I=14.1, T=35.9, wait=5.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:33:56,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:33:56,784] loop.py:45 INFO TRAIN_STEP step=15399 loss=1.358927 Training: 85%|████████▌ | 15399/18018 [33:03:42<6:03:06, 8.32s/step, loss=1.359, acc=0.692, acc_len=1.64, thru=16.1, I=17.0, T=34.4, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:33:57,174] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:34:02,356] loop.py:380 INFO TIMING step=15400: step=3.306s data=1.321s compute=3.276s [fwd=2.350s bwd=0.912s opt=0.003s] dispatch=2.258s [2026-10-19 22:34:02,356] loop.py:45 INFO TRAIN_STEP step=15400 loss=1.356338 Training: 85%|████████▌ | 15400/18018 [33:03:48<5:27:00, 7.49s/step, loss=1.356, acc=0.687, acc_len=1.63, thru=20.1, I=24.5, T=38.7, wait=2.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:08,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:34:09,773] loop.py:45 INFO TRAIN_STEP step=15401 loss=1.214458 Training: 85%|████████▌ | 15401/18018 [33:03:55<5:25:52, 7.47s/step, loss=1.214, acc=0.716, acc_len=1.73, thru=17.7, I=19.8, T=38.0, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:34:09,769] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15400: forward=2385.6ms backward=937.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:10,165] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:34:15,817] loop.py:45 INFO TRAIN_STEP step=15402 loss=1.113173 Training: 85%|████████▌ | 15402/18018 [33:04:01<5:07:05, 7.04s/step, loss=1.113, acc=0.735, acc_len=1.80, thru=19.7, I=24.0, T=39.6, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:23,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:34:24,244] loop.py:45 INFO TRAIN_STEP step=15403 loss=1.273581 Training: 85%|████████▌ | 15403/18018 [33:04:10<5:25:05, 7.46s/step, loss=1.274, acc=0.706, acc_len=1.71, thru=16.2, I=17.2, T=36.6, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:24,378] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:30,090] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:34:30,346] loop.py:45 INFO TRAIN_STEP step=15404 loss=1.282065 Training: 85%|████████▌ | 15404/18018 [33:04:16<5:07:13, 7.05s/step, loss=1.282, acc=0.706, acc_len=1.70, thru=20.1, I=23.1, T=37.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:30,592] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:34:35,932] loop.py:380 INFO TIMING step=15405: step=3.268s data=1.292s compute=3.238s [fwd=2.340s bwd=0.883s opt=0.003s] dispatch=2.308s [2026-10-19 22:34:35,932] loop.py:45 INFO TRAIN_STEP step=15405 loss=1.273943 Training: 85%|████████▌ | 15405/18018 [33:04:22<4:47:55, 6.61s/step, loss=1.274, acc=0.706, acc_len=1.70, thru=18.3, I=19.4, T=39.2, wait=2.3s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:42,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:34:43,614] loop.py:45 INFO TRAIN_STEP step=15406 loss=1.210414 Training: 86%|████████▌ | 15406/18018 [33:04:29<5:01:49, 6.93s/step, loss=1.210, acc=0.713, acc_len=1.73, thru=19.2, I=19.0, T=37.2, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:34:43,610] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15405: forward=2478.5ms backward=921.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:43,993] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:34:50,053] loop.py:45 INFO TRAIN_STEP step=15407 loss=1.174164 Training: 86%|████████▌ | 15407/18018 [33:04:36<4:55:14, 6.78s/step, loss=1.174, acc=0.722, acc_len=1.75, thru=19.0, I=22.9, T=38.0, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:56,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:34:57,183] loop.py:45 INFO TRAIN_STEP step=15408 loss=1.250976 Training: 86%|████████▌ | 15408/18018 [33:04:43<4:59:39, 6.89s/step, loss=1.251, acc=0.708, acc_len=1.71, thru=17.7, I=19.2, T=37.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:34:57,306] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:35:01,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:35:02,843] loop.py:45 INFO TRAIN_STEP step=15409 loss=1.392746 Training: 86%|████████▌ | 15409/18018 [33:04:48<4:43:28, 6.52s/step, loss=1.393, acc=0.682, acc_len=1.62, thru=21.6, I=25.0, T=36.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:35:03,343] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:35:09,421] loop.py:380 INFO TIMING step=15410: step=3.317s data=1.355s compute=3.284s [fwd=2.346s bwd=0.923s opt=0.003s] dispatch=3.253s [2026-10-19 22:35:09,421] loop.py:45 INFO TRAIN_STEP step=15410 loss=1.226810 Training: 86%|████████▌ | 15410/18018 [33:04:55<4:44:10, 6.54s/step, loss=1.227, acc=0.713, acc_len=1.72, thru=15.4, I=24.6, T=38.6, wait=3.3s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:35:18,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:35:19,322] loop.py:45 INFO TRAIN_STEP step=15411 loss=1.354050 Training: 86%|████████▌ | 15411/18018 [33:05:05<5:27:54, 7.55s/step, loss=1.354, acc=0.693, acc_len=1.65, thru=17.4, I=15.7, T=32.3, wait=5.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:35:19,318] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15410: forward=2816.7ms backward=1102.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:35:19,592] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:35:25,505] loop.py:45 INFO TRAIN_STEP step=15412 loss=1.426379 Training: 86%|████████▌ | 15412/18018 [33:05:11<5:10:00, 7.14s/step, loss=1.426, acc=0.679, acc_len=1.60, thru=18.9, I=22.0, T=38.9, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:35:31,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:35:31,771] loop.py:45 INFO TRAIN_STEP step=15413 loss=1.155204 Training: 86%|████████▌ | 15413/18018 [33:05:17<4:58:31, 6.88s/step, loss=1.155, acc=0.723, acc_len=1.76, thru=20.3, I=20.5, T=36.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:35:32,035] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:35:37,838] loop.py:45 INFO TRAIN_STEP step=15414 loss=1.264984 Training: 86%|████████▌ | 15414/18018 [33:05:23<4:47:52, 6.63s/step, loss=1.265, acc=0.705, acc_len=1.69, thru=20.0, I=21.7, T=39.1, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:35:51,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:35:52,189] loop.py:380 INFO TIMING step=15415: step=4.415s data=1.545s compute=4.388s [fwd=3.272s bwd=1.101s opt=0.003s] dispatch=9.928s [2026-10-19 22:35:52,189] loop.py:45 INFO TRAIN_STEP step=15415 loss=1.129021 Training: 86%|████████▌ | 15415/18018 [33:05:38<6:28:13, 8.95s/step, loss=1.129, acc=0.732, acc_len=1.80, thru=8.8, I=8.9, T=29.0, wait=9.9s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:35:52,328] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-19 22:35:57,907] loop.py:45 INFO TRAIN_STEP step=15416 loss=1.075319 Training: 86%|████████▌ | 15416/18018 [33:05:43<5:46:01, 7.98s/step, loss=1.075, acc=0.739, acc_len=1.82, thru=17.6, I=15.3, T=38.1, wait=2.4s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:35:57,903] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15415: forward=2418.7ms backward=894.0ms [2026-10-19 22:36:06,852] loop.py:45 INFO TRAIN_STEP step=15417 loss=1.263309 Training: 86%|████████▌ | 15417/18018 [33:05:52<5:58:26, 8.27s/step, loss=1.263, acc=0.705, acc_len=1.70, thru=15.7, I=14.8, T=37.3, wait=5.5s, pool=57, epoch=2/2][2026-10-19 22:36:13,410] loop.py:45 INFO TRAIN_STEP step=15418 loss=1.213022 Training: 86%|████████▌ | 15418/18018 [33:05:59<5:36:06, 7.76s/step, loss=1.213, acc=0.715, acc_len=1.73, thru=15.9, I=17.4, T=38.8, wait=3.3s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:36:22,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:36:22,860] loop.py:45 INFO TRAIN_STEP step=15419 loss=1.246104 Training: 86%|████████▌ | 15419/18018 [33:06:08<5:57:58, 8.26s/step, loss=1.246, acc=0.708, acc_len=1.71, thru=17.4, I=16.4, T=33.9, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:36:22,984] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:36:28,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:36:29,233] loop.py:380 INFO TIMING step=15420: step=3.272s data=1.305s compute=3.249s [fwd=2.342s bwd=0.893s opt=0.003s] dispatch=3.091s [2026-10-19 22:36:29,233] loop.py:45 INFO TRAIN_STEP step=15420 loss=1.238852 Training: 86%|████████▌ | 15420/18018 [33:06:15<5:33:14, 7.70s/step, loss=1.239, acc=0.709, acc_len=1.71, thru=19.2, I=20.0, T=39.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:36:29,695] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:36:37,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:36:37,714] loop.py:45 INFO TRAIN_STEP step=15421 loss=1.300049 Training: 86%|████████▌ | 15421/18018 [33:06:23<5:43:19, 7.93s/step, loss=1.300, acc=0.701, acc_len=1.68, thru=16.1, I=18.5, T=37.7, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:36:37,709] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15420: forward=2453.5ms backward=894.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:36:38,147] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:36:44,620] loop.py:45 INFO TRAIN_STEP step=15422 loss=1.325410 Training: 86%|████████▌ | 15422/18018 [33:06:30<5:29:52, 7.62s/step, loss=1.325, acc=0.696, acc_len=1.66, thru=18.6, I=19.9, T=38.7, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:36:50,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:36:52,374] loop.py:45 INFO TRAIN_STEP step=15423 loss=1.302414 Training: 86%|████████▌ | 15423/18018 [33:06:38<5:31:24, 7.66s/step, loss=1.302, acc=0.698, acc_len=1.68, thru=18.3, I=22.0, T=29.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:36:52,747] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:00,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:37:00,954] loop.py:45 INFO TRAIN_STEP step=15424 loss=1.291472 Training: 86%|████████▌ | 15424/18018 [33:06:47<5:43:12, 7.94s/step, loss=1.291, acc=0.704, acc_len=1.70, thru=15.6, I=16.2, T=30.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:01,076] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:06,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:37:06,491] loop.py:380 INFO TIMING step=15425: step=3.228s data=1.275s compute=3.196s [fwd=2.302s bwd=0.879s opt=0.003s] dispatch=2.299s [2026-10-19 22:37:06,491] loop.py:45 INFO TRAIN_STEP step=15425 loss=1.243281 Training: 86%|████████▌ | 15425/18018 [33:06:52<5:11:55, 7.22s/step, loss=1.243, acc=0.710, acc_len=1.71, thru=20.3, I=21.0, T=39.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:06,652] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:37:14,443] loop.py:45 INFO TRAIN_STEP step=15426 loss=1.219707 Training: 86%|████████▌ | 15426/18018 [33:07:00<5:21:18, 7.44s/step, loss=1.220, acc=0.715, acc_len=1.73, thru=14.1, I=16.4, T=37.7, wait=4.5s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:37:14,438] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15425: forward=2322.8ms backward=1022.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:22,473] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:37:23,800] loop.py:45 INFO TRAIN_STEP step=15427 loss=1.245105 Training: 86%|████████▌ | 15427/18018 [33:07:09<5:46:05, 8.01s/step, loss=1.245, acc=0.708, acc_len=1.71, thru=17.2, I=17.2, T=35.4, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:23,997] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:30,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:37:31,419] loop.py:45 INFO TRAIN_STEP step=15428 loss=1.331400 Training: 86%|████████▌ | 15428/18018 [33:07:17<5:40:50, 7.90s/step, loss=1.331, acc=0.693, acc_len=1.66, thru=15.9, I=22.6, T=37.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:31,475] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:36,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:37:37,007] loop.py:45 INFO TRAIN_STEP step=15429 loss=1.166089 Training: 86%|████████▌ | 15429/18018 [33:07:23<5:10:48, 7.20s/step, loss=1.166, acc=0.721, acc_len=1.75, thru=21.5, I=25.8, T=39.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:37,153] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:37:44,332] loop.py:380 INFO TIMING step=15430: step=3.402s data=1.402s compute=3.358s [fwd=2.341s bwd=1.002s opt=0.003s] dispatch=3.914s [2026-10-19 22:37:44,332] loop.py:45 INFO TRAIN_STEP step=15430 loss=1.203882 Training: 86%|████████▌ | 15430/18018 [33:07:30<5:12:15, 7.24s/step, loss=1.204, acc=0.715, acc_len=1.73, thru=18.6, I=19.5, T=37.6, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:51,759] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:37:52,285] loop.py:45 INFO TRAIN_STEP step=15431 loss=1.275568 Training: 86%|████████▌ | 15431/18018 [33:07:38<5:21:21, 7.45s/step, loss=1.276, acc=0.705, acc_len=1.70, thru=16.6, I=17.2, T=35.8, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:37:52,280] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15430: forward=2583.6ms backward=947.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:37:52,765] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:37:58,277] loop.py:45 INFO TRAIN_STEP step=15432 loss=1.243897 Training: 86%|████████▌ | 15432/18018 [33:07:44<5:02:20, 7.01s/step, loss=1.244, acc=0.710, acc_len=1.71, thru=20.0, I=19.9, T=38.3, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:05,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:38:06,195] loop.py:45 INFO TRAIN_STEP step=15433 loss=1.250381 Training: 86%|████████▌ | 15433/18018 [33:07:52<5:13:54, 7.29s/step, loss=1.250, acc=0.706, acc_len=1.70, thru=17.2, I=17.6, T=37.3, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:06,578] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:11,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:38:13,358] loop.py:45 INFO TRAIN_STEP step=15434 loss=1.215482 Training: 86%|████████▌ | 15434/18018 [33:07:59<5:12:12, 7.25s/step, loss=1.215, acc=0.712, acc_len=1.73, thru=18.1, I=23.3, T=37.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:13,724] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:18,885] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:38:19,198] loop.py:380 INFO TIMING step=15435: step=3.146s data=1.190s compute=3.113s [fwd=2.243s bwd=0.855s opt=0.003s] dispatch=2.685s [2026-10-19 22:38:19,198] loop.py:45 INFO TRAIN_STEP step=15435 loss=1.249732 Training: 86%|████████▌ | 15435/18018 [33:08:05<4:53:53, 6.83s/step, loss=1.250, acc=0.711, acc_len=1.72, thru=21.3, I=28.0, T=40.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:19,387] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:26,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:38:26,927] loop.py:45 INFO TRAIN_STEP step=15436 loss=1.247760 Training: 86%|████████▌ | 15436/18018 [33:08:13<5:05:24, 7.10s/step, loss=1.248, acc=0.709, acc_len=1.71, thru=17.5, I=19.4, T=37.2, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:38:26,922] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15435: forward=2431.5ms backward=964.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:27,283] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:38:32,905] loop.py:45 INFO TRAIN_STEP step=15437 loss=1.289192 Training: 86%|████████▌ | 15437/18018 [33:08:18<4:50:52, 6.76s/step, loss=1.289, acc=0.703, acc_len=1.70, thru=19.6, I=19.1, T=37.7, wait=2.6s, pool=57, epoch=2/2][2026-10-19 22:38:39,396] loop.py:45 INFO TRAIN_STEP step=15438 loss=1.261781 Training: 86%|████████▌ | 15438/18018 [33:08:25<4:47:15, 6.68s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=16.1, I=18.1, T=37.6, wait=3.1s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:49,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:38:50,923] loop.py:45 INFO TRAIN_STEP step=15439 loss=1.201790 Training: 86%|████████▌ | 15439/18018 [33:08:37<5:49:38, 8.13s/step, loss=1.202, acc=0.715, acc_len=1.73, thru=14.4, I=14.1, T=29.8, wait=7.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:50,968] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:55,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:38:58,536] loop.py:380 INFO TIMING step=15440: step=4.375s data=1.665s compute=4.336s [fwd=3.377s bwd=0.945s opt=0.003s] dispatch=3.228s [2026-10-19 22:38:58,536] loop.py:45 INFO TRAIN_STEP step=15440 loss=1.177332 Training: 86%|████████▌ | 15440/18018 [33:08:44<5:42:46, 7.98s/step, loss=1.177, acc=0.720, acc_len=1.75, thru=15.6, I=22.6, T=29.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:38:58,868] inference_manager.py:408 INFO Pool capacity available after 3.0s, pool_size=1/64 [2026-10-19 22:39:04,115] loop.py:45 INFO TRAIN_STEP step=15441 loss=1.161185 Training: 86%|████████▌ | 15441/18018 [33:08:50<5:11:43, 7.26s/step, loss=1.161, acc=0.722, acc_len=1.76, thru=16.4, I=22.9, T=40.7, wait=2.4s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:39:04,110] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15440: forward=2229.0ms backward=873.0ms [2026-10-19 22:39:11,438] loop.py:45 INFO TRAIN_STEP step=15442 loss=1.259982 Training: 86%|████████▌ | 15442/18018 [33:08:57<5:12:27, 7.28s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=18.0, I=18.8, T=38.5, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:19,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:39:19,643] loop.py:45 INFO TRAIN_STEP step=15443 loss=1.277246 Training: 86%|████████▌ | 15443/18018 [33:09:05<5:24:16, 7.56s/step, loss=1.277, acc=0.707, acc_len=1.70, thru=16.5, I=16.1, T=35.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:19,741] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:24,435] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:39:25,189] loop.py:45 INFO TRAIN_STEP step=15444 loss=1.338990 Training: 86%|████████▌ | 15444/18018 [33:09:11<4:58:18, 6.95s/step, loss=1.339, acc=0.692, acc_len=1.64, thru=21.4, I=23.3, T=38.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:25,441] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:30,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:39:31,134] loop.py:380 INFO TIMING step=15445: step=3.325s data=1.359s compute=3.296s [fwd=2.364s bwd=0.918s opt=0.003s] dispatch=2.609s [2026-10-19 22:39:31,134] loop.py:45 INFO TRAIN_STEP step=15445 loss=1.257149 Training: 86%|████████▌ | 15445/18018 [33:09:17<4:45:11, 6.65s/step, loss=1.257, acc=0.707, acc_len=1.70, thru=22.4, I=26.4, T=38.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:31,544] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:35,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:39:36,539] loop.py:45 INFO TRAIN_STEP step=15446 loss=1.332822 Training: 86%|████████▌ | 15446/18018 [33:09:22<4:29:05, 6.28s/step, loss=1.333, acc=0.693, acc_len=1.65, thru=22.9, I=32.1, T=40.1, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:39:36,535] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15445: forward=2279.0ms backward=866.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:36,702] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:42,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:39:43,083] loop.py:45 INFO TRAIN_STEP step=15447 loss=1.201465 Training: 86%|████████▌ | 15447/18018 [33:09:29<4:32:24, 6.36s/step, loss=1.201, acc=0.713, acc_len=1.72, thru=20.5, I=24.5, T=39.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:43,355] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:39:49,855] loop.py:45 INFO TRAIN_STEP step=15448 loss=1.254362 Training: 86%|████████▌ | 15448/18018 [33:09:35<4:37:37, 6.48s/step, loss=1.254, acc=0.711, acc_len=1.72, thru=18.9, I=20.1, T=39.1, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:57,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:39:58,058] loop.py:45 INFO TRAIN_STEP step=15449 loss=1.285352 Training: 86%|████████▌ | 15449/18018 [33:09:44<4:59:38, 7.00s/step, loss=1.285, acc=0.704, acc_len=1.69, thru=16.2, I=17.1, T=35.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:39:58,112] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 22:40:03,705] loop.py:380 INFO TIMING step=15450: step=3.267s data=1.326s compute=3.228s [fwd=2.312s bwd=0.903s opt=0.003s] dispatch=2.369s [2026-10-19 22:40:03,705] loop.py:45 INFO TRAIN_STEP step=15450 loss=1.039108 Training: 86%|████████▌ | 15450/18018 [33:09:49<4:42:09, 6.59s/step, loss=1.039, acc=0.746, acc_len=1.84, thru=16.9, I=18.2, T=39.2, wait=2.4s, pool=25, epoch=2/2][2026-10-19 22:40:11,090] loop.py:45 INFO TRAIN_STEP step=15451 loss=1.237105 Training: 86%|████████▌ | 15451/18018 [33:09:57<4:52:13, 6.83s/step, loss=1.237, acc=0.712, acc_len=1.72, thru=19.0, I=18.4, T=37.2, wait=3.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:40:11,086] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15450: forward=2436.2ms backward=950.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:40:17,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:40:17,941] loop.py:45 INFO TRAIN_STEP step=15452 loss=1.306842 Training: 86%|████████▌ | 15452/18018 [33:10:04<4:52:22, 6.84s/step, loss=1.307, acc=0.699, acc_len=1.68, thru=18.4, I=19.9, T=37.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:40:18,388] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:40:25,628] loop.py:45 INFO TRAIN_STEP step=15453 loss=1.270010 Training: 86%|████████▌ | 15453/18018 [33:10:11<5:03:08, 7.09s/step, loss=1.270, acc=0.708, acc_len=1.71, thru=16.2, I=17.2, T=37.3, wait=4.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:40:32,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:40:32,806] loop.py:45 INFO TRAIN_STEP step=15454 loss=1.373347 Training: 86%|████████▌ | 15454/18018 [33:10:18<5:04:08, 7.12s/step, loss=1.373, acc=0.686, acc_len=1.64, thru=19.3, I=19.2, T=36.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:40:33,174] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:40:40,869] loop.py:380 INFO TIMING step=15455: step=3.540s data=1.464s compute=3.511s [fwd=2.538s bwd=0.959s opt=0.003s] dispatch=4.514s [2026-10-19 22:40:40,870] loop.py:45 INFO TRAIN_STEP step=15455 loss=1.407084 Training: 86%|████████▌ | 15455/18018 [33:10:26<5:16:10, 7.40s/step, loss=1.407, acc=0.680, acc_len=1.61, thru=15.3, I=19.9, T=36.2, wait=4.5s, pool=57, epoch=2/2][2026-10-19 22:40:48,338] loop.py:45 INFO TRAIN_STEP step=15456 loss=1.240622 Training: 86%|████████▌ | 15456/18018 [33:10:34<5:16:52, 7.42s/step, loss=1.241, acc=0.710, acc_len=1.71, thru=15.4, I=16.1, T=36.3, wait=3.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:40:48,333] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15455: forward=2475.6ms backward=1000.6ms [2026-10-19 22:40:54,949] loop.py:45 INFO TRAIN_STEP step=15457 loss=1.314661 Training: 86%|████████▌ | 15457/18018 [33:10:41<5:06:23, 7.18s/step, loss=1.315, acc=0.696, acc_len=1.66, thru=17.6, I=18.5, T=37.6, wait=3.2s, pool=45, epoch=2/2][2026-10-19 22:41:03,397] loop.py:45 INFO TRAIN_STEP step=15458 loss=1.367621 Training: 86%|████████▌ | 15458/18018 [33:10:49<5:22:30, 7.56s/step, loss=1.368, acc=0.690, acc_len=1.64, thru=14.9, I=14.8, T=35.2, wait=4.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:41:10,652] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:41:11,186] loop.py:45 INFO TRAIN_STEP step=15459 loss=1.270809 Training: 86%|████████▌ | 15459/18018 [33:10:57<5:25:21, 7.63s/step, loss=1.271, acc=0.706, acc_len=1.70, thru=18.1, I=17.8, T=36.4, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:41:11,657] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:41:18,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:41:18,344] loop.py:380 INFO TIMING step=15460: step=3.344s data=1.372s compute=3.299s [fwd=2.351s bwd=0.934s opt=0.003s] dispatch=3.804s [2026-10-19 22:41:18,344] loop.py:45 INFO TRAIN_STEP step=15460 loss=1.321285 Training: 86%|████████▌ | 15460/18018 [33:11:04<5:19:10, 7.49s/step, loss=1.321, acc=0.697, acc_len=1.67, thru=18.4, I=19.9, T=38.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:41:18,746] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:41:23,696] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:41:24,406] loop.py:45 INFO TRAIN_STEP step=15461 loss=1.178649 Training: 86%|████████▌ | 15461/18018 [33:11:10<5:00:51, 7.06s/step, loss=1.179, acc=0.720, acc_len=1.74, thru=20.1, I=20.9, T=37.4, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:41:24,401] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15460: forward=2393.9ms backward=978.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:41:24,701] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:41:30,774] loop.py:45 INFO TRAIN_STEP step=15462 loss=1.294779 Training: 86%|████████▌ | 15462/18018 [33:11:16<4:51:55, 6.85s/step, loss=1.295, acc=0.698, acc_len=1.67, thru=19.8, I=22.3, T=39.0, wait=3.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:41:35,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:41:36,776] loop.py:45 INFO TRAIN_STEP step=15463 loss=1.286953 Training: 86%|████████▌ | 15463/18018 [33:11:22<4:40:55, 6.60s/step, loss=1.287, acc=0.700, acc_len=1.68, thru=20.5, I=22.9, T=37.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:41:37,012] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:41:43,919] loop.py:45 INFO TRAIN_STEP step=15464 loss=1.363000 Training: 86%|████████▌ | 15464/18018 [33:11:30<4:47:47, 6.76s/step, loss=1.363, acc=0.690, acc_len=1.65, thru=17.5, I=21.6, T=37.5, wait=3.7s, pool=53, epoch=2/2][2026-10-19 22:41:51,279] loop.py:380 INFO TIMING step=15465: step=3.407s data=1.355s compute=3.376s [fwd=2.424s bwd=0.939s opt=0.003s] dispatch=3.944s [2026-10-19 22:41:51,279] loop.py:45 INFO TRAIN_STEP step=15465 loss=1.327553 Training: 86%|████████▌ | 15465/18018 [33:11:37<4:55:20, 6.94s/step, loss=1.328, acc=0.694, acc_len=1.66, thru=16.2, I=18.4, T=37.6, wait=3.9s, pool=45, epoch=2/2][2026-10-19 22:41:59,829] loop.py:45 INFO TRAIN_STEP step=15466 loss=1.188368 Training: 86%|████████▌ | 15466/18018 [33:11:45<5:15:44, 7.42s/step, loss=1.188, acc=0.720, acc_len=1.75, thru=15.6, I=15.6, T=36.9, wait=5.1s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:41:59,824] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15465: forward=2400.9ms backward=963.5ms [2026-10-19 22:42:07,532] loop.py:45 INFO TRAIN_STEP step=15467 loss=1.236812 Training: 86%|████████▌ | 15467/18018 [33:11:53<5:19:09, 7.51s/step, loss=1.237, acc=0.710, acc_len=1.72, thru=17.3, I=17.1, T=37.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:07,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:07,982] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:13,048] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:42:14,173] loop.py:45 INFO TRAIN_STEP step=15468 loss=1.330627 Training: 86%|████████▌ | 15468/18018 [33:12:00<5:08:00, 7.25s/step, loss=1.331, acc=0.696, acc_len=1.66, thru=19.8, I=21.0, T=36.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:14,554] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:19,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:42:20,062] loop.py:45 INFO TRAIN_STEP step=15469 loss=1.266447 Training: 86%|████████▌ | 15469/18018 [33:12:06<4:50:35, 6.84s/step, loss=1.266, acc=0.704, acc_len=1.69, thru=20.8, I=24.9, T=40.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:20,240] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:26,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:42:27,592] loop.py:380 INFO TIMING step=15470: step=3.344s data=1.323s compute=3.314s [fwd=2.365s bwd=0.934s opt=0.003s] dispatch=4.178s [2026-10-19 22:42:27,593] loop.py:45 INFO TRAIN_STEP step=15470 loss=1.233833 Training: 86%|████████▌ | 15470/18018 [33:12:13<4:59:16, 7.05s/step, loss=1.234, acc=0.712, acc_len=1.72, thru=17.6, I=20.1, T=38.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:27,988] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:32,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:42:33,704] loop.py:45 INFO TRAIN_STEP step=15471 loss=1.181835 Training: 86%|████████▌ | 15471/18018 [33:12:19<4:47:14, 6.77s/step, loss=1.182, acc=0.718, acc_len=1.74, thru=20.8, I=26.4, T=39.9, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:42:33,700] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15470: forward=2279.3ms backward=879.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:33,809] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:38,796] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:42:39,567] loop.py:45 INFO TRAIN_STEP step=15472 loss=1.232670 Training: 86%|████████▌ | 15472/18018 [33:12:25<4:35:35, 6.49s/step, loss=1.233, acc=0.709, acc_len=1.71, thru=21.3, I=25.8, T=38.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:39,801] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:44,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:42:45,451] loop.py:45 INFO TRAIN_STEP step=15473 loss=1.263419 Training: 86%|████████▌ | 15473/18018 [33:12:31<4:27:43, 6.31s/step, loss=1.263, acc=0.711, acc_len=1.72, thru=21.2, I=26.3, T=39.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:42:45,528] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 22:42:51,494] loop.py:45 INFO TRAIN_STEP step=15474 loss=1.220737 Training: 86%|████████▌ | 15474/18018 [33:12:37<4:24:14, 6.23s/step, loss=1.221, acc=0.714, acc_len=1.73, thru=16.9, I=19.8, T=38.8, wait=2.7s, pool=29, epoch=2/2][2026-10-19 22:42:58,502] loop.py:380 INFO TIMING step=15475: step=3.337s data=1.343s compute=3.301s [fwd=2.371s bwd=0.916s opt=0.003s] dispatch=3.659s [2026-10-19 22:42:58,503] loop.py:45 INFO TRAIN_STEP step=15475 loss=1.245427 Training: 86%|████████▌ | 15475/18018 [33:12:44<4:33:59, 6.46s/step, loss=1.245, acc=0.708, acc_len=1.71, thru=18.1, I=18.3, T=38.4, wait=3.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:43:05,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:43:06,133] loop.py:45 INFO TRAIN_STEP step=15476 loss=1.364986 Training: 86%|████████▌ | 15476/18018 [33:12:52<4:48:42, 6.81s/step, loss=1.365, acc=0.686, acc_len=1.63, thru=17.7, I=19.0, T=38.0, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:43:06,129] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15475: forward=2404.9ms backward=915.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:43:06,395] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:43:13,173] loop.py:45 INFO TRAIN_STEP step=15477 loss=1.248967 Training: 86%|████████▌ | 15477/18018 [33:12:59<4:51:25, 6.88s/step, loss=1.249, acc=0.708, acc_len=1.71, thru=16.5, I=18.5, T=39.8, wait=3.8s, pool=41, epoch=2/2][2026-10-19 22:43:20,568] loop.py:45 INFO TRAIN_STEP step=15478 loss=1.176415 Training: 86%|████████▌ | 15478/18018 [33:13:06<4:57:49, 7.04s/step, loss=1.176, acc=0.721, acc_len=1.76, thru=16.8, I=17.2, T=36.6, wait=3.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:43:27,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:43:28,064] loop.py:45 INFO TRAIN_STEP step=15479 loss=1.230803 Training: 86%|████████▌ | 15479/18018 [33:13:14<5:03:34, 7.17s/step, loss=1.231, acc=0.711, acc_len=1.72, thru=17.6, I=17.2, T=38.4, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:43:28,471] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:43:36,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:43:36,640] loop.py:380 INFO TIMING step=15480: step=3.483s data=1.395s compute=3.454s [fwd=2.483s bwd=0.958s opt=0.003s] dispatch=5.084s [2026-10-19 22:43:36,640] loop.py:45 INFO TRAIN_STEP step=15480 loss=1.297716 Training: 86%|████████▌ | 15480/18018 [33:13:22<5:21:14, 7.59s/step, loss=1.298, acc=0.699, acc_len=1.67, thru=16.1, I=16.6, T=36.7, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:43:36,836] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:43:42,343] loop.py:45 INFO TRAIN_STEP step=15481 loss=1.237178 Training: 86%|████████▌ | 15481/18018 [33:13:28<4:57:07, 7.03s/step, loss=1.237, acc=0.706, acc_len=1.70, thru=20.3, I=19.3, T=38.9, wait=2.4s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:43:42,339] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15480: forward=2339.6ms backward=904.1ms [2026-10-19 22:43:48,810] loop.py:45 INFO TRAIN_STEP step=15482 loss=1.115005 Training: 86%|████████▌ | 15482/18018 [33:13:34<4:49:53, 6.86s/step, loss=1.115, acc=0.736, acc_len=1.80, thru=18.7, I=19.5, T=38.2, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:43:54,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:43:54,909] loop.py:45 INFO TRAIN_STEP step=15483 loss=1.201728 Training: 86%|████████▌ | 15483/18018 [33:13:40<4:40:11, 6.63s/step, loss=1.202, acc=0.718, acc_len=1.74, thru=20.6, I=22.1, T=37.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:43:55,086] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:00,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:44:01,297] loop.py:45 INFO TRAIN_STEP step=15484 loss=1.289486 Training: 86%|████████▌ | 15484/18018 [33:13:47<4:36:58, 6.56s/step, loss=1.289, acc=0.702, acc_len=1.68, thru=20.0, I=21.3, T=37.5, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:01,371] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 22:44:07,886] loop.py:380 INFO TIMING step=15485: step=3.357s data=1.344s compute=3.329s [fwd=2.390s bwd=0.925s opt=0.003s] dispatch=3.223s [2026-10-19 22:44:07,886] loop.py:45 INFO TRAIN_STEP step=15485 loss=1.205719 Training: 86%|████████▌ | 15485/18018 [33:13:53<4:37:14, 6.57s/step, loss=1.206, acc=0.716, acc_len=1.74, thru=19.0, I=19.8, T=38.1, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:14,431] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:44:15,506] loop.py:45 INFO TRAIN_STEP step=15486 loss=1.192261 Training: 86%|████████▌ | 15486/18018 [33:14:01<4:50:28, 6.88s/step, loss=1.192, acc=0.717, acc_len=1.74, thru=17.9, I=18.9, T=38.2, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:44:15,502] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15485: forward=2381.6ms backward=920.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:15,938] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:20,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:44:21,651] loop.py:45 INFO TRAIN_STEP step=15487 loss=1.234726 Training: 86%|████████▌ | 15487/18018 [33:14:07<4:41:01, 6.66s/step, loss=1.235, acc=0.712, acc_len=1.72, thru=20.9, I=27.1, T=40.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:22,140] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:26,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:44:27,671] loop.py:45 INFO TRAIN_STEP step=15488 loss=1.297222 Training: 86%|████████▌ | 15488/18018 [33:14:13<4:32:45, 6.47s/step, loss=1.297, acc=0.701, acc_len=1.69, thru=20.8, I=26.7, T=39.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:27,789] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:34,008] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:44:34,587] loop.py:45 INFO TRAIN_STEP step=15489 loss=1.263862 Training: 86%|████████▌ | 15489/18018 [33:14:20<4:38:20, 6.60s/step, loss=1.264, acc=0.708, acc_len=1.71, thru=19.1, I=21.7, T=38.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:35,012] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:44:41,954] loop.py:380 INFO TIMING step=15490: step=3.387s data=1.392s compute=3.352s [fwd=2.417s bwd=0.921s opt=0.003s] dispatch=3.970s [2026-10-19 22:44:41,954] loop.py:45 INFO TRAIN_STEP step=15490 loss=1.140455 Training: 86%|████████▌ | 15490/18018 [33:14:28<4:47:51, 6.83s/step, loss=1.140, acc=0.729, acc_len=1.78, thru=17.0, I=18.4, T=37.8, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:49,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:44:50,844] loop.py:45 INFO TRAIN_STEP step=15491 loss=1.224415 Training: 86%|████████▌ | 15491/18018 [33:14:36<5:13:46, 7.45s/step, loss=1.224, acc=0.712, acc_len=1.73, thru=15.9, I=18.0, T=36.1, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:44:50,840] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15490: forward=2444.2ms backward=1045.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:44:51,176] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 22:44:56,857] loop.py:45 INFO TRAIN_STEP step=15492 loss=1.178163 Training: 86%|████████▌ | 15492/18018 [33:14:42<4:55:30, 7.02s/step, loss=1.178, acc=0.720, acc_len=1.75, thru=18.9, I=22.7, T=40.5, wait=2.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:45:06,219] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:45:06,496] loop.py:45 INFO TRAIN_STEP step=15493 loss=1.308111 Training: 86%|████████▌ | 15493/18018 [33:14:52<5:28:27, 7.81s/step, loss=1.308, acc=0.698, acc_len=1.68, thru=15.2, I=15.3, T=34.4, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:45:06,720] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:45:12,561] loop.py:45 INFO TRAIN_STEP step=15494 loss=1.187307 Training: 86%|████████▌ | 15494/18018 [33:14:58<5:06:22, 7.28s/step, loss=1.187, acc=0.721, acc_len=1.76, thru=16.7, I=19.1, T=38.7, wait=2.7s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:45:21,331] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:45:22,220] loop.py:380 INFO TIMING step=15495: step=3.594s data=1.468s compute=3.568s [fwd=2.555s bwd=0.999s opt=0.003s] dispatch=6.054s [2026-10-19 22:45:22,221] loop.py:45 INFO TRAIN_STEP step=15495 loss=1.335408 Training: 86%|████████▌ | 15495/18018 [33:15:08<5:36:13, 8.00s/step, loss=1.335, acc=0.693, acc_len=1.66, thru=16.6, I=15.8, T=35.6, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:45:22,336] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 22:45:28,666] loop.py:45 INFO TRAIN_STEP step=15496 loss=1.231008 Training: 86%|████████▌ | 15496/18018 [33:15:14<5:16:33, 7.53s/step, loss=1.231, acc=0.712, acc_len=1.72, thru=16.8, I=20.4, T=39.2, wait=3.2s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:45:28,662] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15495: forward=2288.1ms backward=932.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:45:37,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:45:38,953] loop.py:45 INFO TRAIN_STEP step=15497 loss=1.179820 Training: 86%|████████▌ | 15497/18018 [33:15:25<5:51:08, 8.36s/step, loss=1.180, acc=0.720, acc_len=1.75, thru=18.3, I=15.5, T=33.2, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:45:39,092] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:45:47,222] loop.py:45 INFO TRAIN_STEP step=15498 loss=1.335823 Training: 86%|████████▌ | 15498/18018 [33:15:33<5:49:52, 8.33s/step, loss=1.336, acc=0.696, acc_len=1.67, thru=13.5, I=15.2, T=38.6, wait=4.9s, pool=61, epoch=2/2][2026-10-19 22:45:53,455] loop.py:45 INFO TRAIN_STEP step=15499 loss=1.299492 Training: 86%|████████▌ | 15499/18018 [33:15:39<5:23:20, 7.70s/step, loss=1.299, acc=0.703, acc_len=1.68, thru=17.7, I=16.5, T=37.3, wait=2.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:01,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:46:02,557] loop.py:380 INFO TIMING step=15500: step=3.532s data=1.395s compute=3.494s [fwd=2.475s bwd=1.004s opt=0.003s] dispatch=5.560s [2026-10-19 22:46:02,557] loop.py:45 INFO TRAIN_STEP step=15500 loss=1.329654 Training: 86%|████████▌ | 15500/18018 [33:15:48<5:40:50, 8.12s/step, loss=1.330, acc=0.693, acc_len=1.66, thru=16.5, I=16.0, T=36.2, wait=5.6s, pool=65, epoch=2/2][2026-10-19 22:46:02,560] loop.py:419 INFO Saving checkpoint at step 15500... (TrainerActor pid=9310) [2026-10-19 22:46:05,840] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0015501 (TrainerActor pid=9310) [2026-10-19 22:46:05,852] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0014751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:06,805] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:07,307] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:12,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 22:46:05,841] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0015501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 22:46:07,095] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0014751 [2026-10-19 22:46:13,028] loop.py:45 INFO TRAIN_STEP step=15501 loss=1.143691 Training: 86%|████████▌ | 15501/18018 [33:15:59<6:10:15, 8.83s/step, loss=1.144, acc=0.728, acc_len=1.77, thru=22.4, I=24.3, T=39.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:13,056] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 22:46:13,023] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15500: forward=2310.1ms backward=908.3ms [2026-10-19 22:46:18,817] loop.py:45 INFO TRAIN_STEP step=15502 loss=1.244529 Training: 86%|████████▌ | 15502/18018 [33:16:04<5:31:54, 7.91s/step, loss=1.245, acc=0.711, acc_len=1.72, thru=20.9, I=22.6, T=39.7, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:26,856] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:46:27,651] loop.py:45 INFO TRAIN_STEP step=15503 loss=1.154593 Training: 86%|████████▌ | 15503/18018 [33:16:13<5:43:21, 8.19s/step, loss=1.155, acc=0.726, acc_len=1.77, thru=15.3, I=16.7, T=36.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:27,862] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:32,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:46:33,737] loop.py:45 INFO TRAIN_STEP step=15504 loss=1.290463 Training: 86%|████████▌ | 15504/18018 [33:16:19<5:16:43, 7.56s/step, loss=1.290, acc=0.703, acc_len=1.69, thru=20.1, I=22.9, T=39.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:33,858] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 22:46:39,618] loop.py:380 INFO TIMING step=15505: step=3.278s data=1.330s compute=3.245s [fwd=2.341s bwd=0.890s opt=0.003s] dispatch=2.595s [2026-10-19 22:46:39,618] loop.py:45 INFO TRAIN_STEP step=15505 loss=1.244093 Training: 86%|████████▌ | 15505/18018 [33:16:25<4:55:30, 7.06s/step, loss=1.244, acc=0.710, acc_len=1.72, thru=20.7, I=21.8, T=39.0, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:46,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:46:47,028] loop.py:45 INFO TRAIN_STEP step=15506 loss=1.280213 Training: 86%|████████▌ | 15506/18018 [33:16:33<4:59:50, 7.16s/step, loss=1.280, acc=0.705, acc_len=1.69, thru=18.1, I=18.5, T=37.4, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:46:47,023] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15505: forward=2418.6ms backward=950.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:46:47,521] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:46:54,089] loop.py:45 INFO TRAIN_STEP step=15507 loss=1.262109 Training: 86%|████████▌ | 15507/18018 [33:16:40<4:58:29, 7.13s/step, loss=1.262, acc=0.706, acc_len=1.70, thru=15.3, I=18.9, T=38.5, wait=3.7s, pool=33, epoch=2/2][2026-10-19 22:47:02,871] loop.py:45 INFO TRAIN_STEP step=15508 loss=1.414428 Training: 86%|████████▌ | 15508/18018 [33:16:48<5:19:04, 7.63s/step, loss=1.414, acc=0.680, acc_len=1.61, thru=16.9, I=15.0, T=36.2, wait=5.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:47:09,636] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:47:10,337] loop.py:45 INFO TRAIN_STEP step=15509 loss=1.281990 Training: 86%|████████▌ | 15509/18018 [33:16:56<5:16:55, 7.58s/step, loss=1.282, acc=0.702, acc_len=1.69, thru=18.5, I=18.8, T=37.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:47:10,642] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:47:18,339] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:47:18,890] loop.py:380 INFO TIMING step=15510: step=3.658s data=1.422s compute=3.630s [fwd=2.594s bwd=1.021s opt=0.003s] dispatch=4.886s [2026-10-19 22:47:18,890] loop.py:45 INFO TRAIN_STEP step=15510 loss=1.379113 Training: 86%|████████▌ | 15510/18018 [33:17:04<5:28:59, 7.87s/step, loss=1.379, acc=0.688, acc_len=1.64, thru=15.3, I=17.5, T=35.0, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:47:19,343] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:47:24,660] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:47:24,965] loop.py:45 INFO TRAIN_STEP step=15511 loss=1.343256 Training: 86%|████████▌ | 15511/18018 [33:17:11<5:06:22, 7.33s/step, loss=1.343, acc=0.688, acc_len=1.64, thru=20.1, I=22.9, T=39.2, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:47:24,960] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15510: forward=2356.5ms backward=868.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:47:25,162] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:47:33,751] loop.py:45 INFO TRAIN_STEP step=15512 loss=1.311540 Training: 86%|████████▌ | 15512/18018 [33:17:19<5:24:28, 7.77s/step, loss=1.312, acc=0.697, acc_len=1.68, thru=14.5, I=14.6, T=34.7, wait=5.1s, pool=61, epoch=2/2][2026-10-19 22:47:39,119] loop.py:45 INFO TRAIN_STEP step=15513 loss=1.192019 Training: 86%|████████▌ | 15513/18018 [33:17:25<4:54:14, 7.05s/step, loss=1.192, acc=0.716, acc_len=1.74, thru=20.7, I=19.8, T=38.9, wait=2.1s, pool=57, epoch=2/2][2026-10-19 22:47:46,848] loop.py:45 INFO TRAIN_STEP step=15514 loss=1.288245 Training: 86%|████████▌ | 15514/18018 [33:17:32<5:02:38, 7.25s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=16.8, I=16.4, T=37.0, wait=4.3s, pool=61, epoch=2/2][2026-10-19 22:47:56,313] loop.py:380 INFO TIMING step=15515: step=3.603s data=1.509s compute=3.558s [fwd=2.536s bwd=1.007s opt=0.003s] dispatch=5.854s [2026-10-19 22:47:56,313] loop.py:45 INFO TRAIN_STEP step=15515 loss=1.118214 Training: 86%|████████▌ | 15515/18018 [33:17:42<5:30:17, 7.92s/step, loss=1.118, acc=0.736, acc_len=1.80, thru=14.2, I=14.1, T=35.5, wait=5.9s, pool=61, epoch=2/2][2026-10-19 22:48:02,834] loop.py:45 INFO TRAIN_STEP step=15516 loss=1.263275 Training: 86%|████████▌ | 15516/18018 [33:17:48<5:12:37, 7.50s/step, loss=1.263, acc=0.707, acc_len=1.71, thru=20.2, I=19.0, T=38.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:48:02,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 22:48:02,830] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15515: forward=2390.1ms backward=904.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:48:03,292] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:48:10,019] loop.py:45 INFO TRAIN_STEP step=15517 loss=1.326398 Training: 86%|████████▌ | 15517/18018 [33:17:56<5:08:35, 7.40s/step, loss=1.326, acc=0.694, acc_len=1.66, thru=17.5, I=17.2, T=38.3, wait=3.8s, pool=49, epoch=2/2][2026-10-19 22:48:17,921] loop.py:45 INFO TRAIN_STEP step=15518 loss=1.439626 Training: 86%|████████▌ | 15518/18018 [33:18:04<5:14:43, 7.55s/step, loss=1.440, acc=0.677, acc_len=1.60, thru=17.0, I=16.6, T=36.9, wait=4.4s, pool=53, epoch=2/2][2026-10-19 22:48:25,275] loop.py:45 INFO TRAIN_STEP step=15519 loss=1.174869 Training: 86%|████████▌ | 15519/18018 [33:18:11<5:12:06, 7.49s/step, loss=1.175, acc=0.722, acc_len=1.75, thru=18.5, I=19.0, T=37.6, wait=3.9s, pool=61, epoch=2/2][2026-10-19 22:48:33,157] loop.py:380 INFO TIMING step=15520: step=3.598s data=1.418s compute=3.568s [fwd=2.592s bwd=0.960s opt=0.003s] dispatch=4.275s [2026-10-19 22:48:33,157] loop.py:45 INFO TRAIN_STEP step=15520 loss=1.245352 Training: 86%|████████▌ | 15520/18018 [33:18:19<5:16:49, 7.61s/step, loss=1.245, acc=0.710, acc_len=1.72, thru=13.4, I=14.8, T=35.6, wait=4.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:48:41,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:48:41,912] loop.py:45 INFO TRAIN_STEP step=15521 loss=1.270400 Training: 86%|████████▌ | 15521/18018 [33:18:27<5:30:59, 7.95s/step, loss=1.270, acc=0.703, acc_len=1.69, thru=17.2, I=15.5, T=36.7, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:48:41,955] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 22:48:41,908] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15520: forward=2438.4ms backward=998.7ms [2026-10-19 22:48:51,454] loop.py:45 INFO TRAIN_STEP step=15522 loss=1.341263 Training: 86%|████████▌ | 15522/18018 [33:18:37<5:50:43, 8.43s/step, loss=1.341, acc=0.694, acc_len=1.67, thru=13.1, I=13.0, T=35.9, wait=6.0s, pool=61, epoch=2/2][2026-10-19 22:48:58,052] loop.py:45 INFO TRAIN_STEP step=15523 loss=1.355293 Training: 86%|████████▌ | 15523/18018 [33:18:44<5:27:42, 7.88s/step, loss=1.355, acc=0.690, acc_len=1.65, thru=17.9, I=19.2, T=38.1, wait=3.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:04,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:49:04,780] loop.py:45 INFO TRAIN_STEP step=15524 loss=1.309526 Training: 86%|████████▌ | 15524/18018 [33:18:50<5:13:09, 7.53s/step, loss=1.310, acc=0.697, acc_len=1.67, thru=19.5, I=20.3, T=38.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:05,269] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:10,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:49:11,359] loop.py:380 INFO TIMING step=15525: step=3.396s data=1.401s compute=3.349s [fwd=2.356s bwd=0.978s opt=0.003s] dispatch=3.176s [2026-10-19 22:49:11,359] loop.py:45 INFO TRAIN_STEP step=15525 loss=1.183834 Training: 86%|████████▌ | 15525/18018 [33:18:57<5:01:08, 7.25s/step, loss=1.184, acc=0.723, acc_len=1.76, thru=19.6, I=22.2, T=37.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:11,811] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:17,108] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:49:17,706] loop.py:45 INFO TRAIN_STEP step=15526 loss=1.218307 Training: 86%|████████▌ | 15526/18018 [33:19:03<4:49:48, 6.98s/step, loss=1.218, acc=0.711, acc_len=1.73, thru=20.1, I=23.3, T=38.3, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:49:17,701] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15525: forward=2386.4ms backward=907.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:18,111] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:49:24,141] loop.py:45 INFO TRAIN_STEP step=15527 loss=1.349485 Training: 86%|████████▌ | 15527/18018 [33:19:10<4:42:55, 6.81s/step, loss=1.349, acc=0.690, acc_len=1.65, thru=20.4, I=21.8, T=38.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:24,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:24,627] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:49:32,136] loop.py:45 INFO TRAIN_STEP step=15528 loss=1.371188 Training: 86%|████████▌ | 15528/18018 [33:19:18<4:57:33, 7.17s/step, loss=1.371, acc=0.683, acc_len=1.62, thru=16.2, I=16.7, T=37.2, wait=4.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:39,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:49:39,653] loop.py:45 INFO TRAIN_STEP step=15529 loss=1.293455 Training: 86%|████████▌ | 15529/18018 [33:19:25<5:01:41, 7.27s/step, loss=1.293, acc=0.703, acc_len=1.70, thru=18.9, I=17.0, T=38.4, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:40,126] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:44,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:49:45,466] loop.py:380 INFO TIMING step=15530: step=3.225s data=1.224s compute=3.192s [fwd=2.298s bwd=0.880s opt=0.003s] dispatch=2.581s [2026-10-19 22:49:45,467] loop.py:45 INFO TRAIN_STEP step=15530 loss=1.236869 Training: 86%|████████▌ | 15530/18018 [33:19:31<4:43:27, 6.84s/step, loss=1.237, acc=0.708, acc_len=1.71, thru=20.5, I=22.1, T=39.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:49:45,811] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:49:51,676] loop.py:45 INFO TRAIN_STEP step=15531 loss=1.278444 Training: 86%|████████▌ | 15531/18018 [33:19:37<4:35:31, 6.65s/step, loss=1.278, acc=0.702, acc_len=1.69, thru=20.9, I=21.9, T=38.5, wait=2.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:49:51,671] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15530: forward=2362.6ms backward=903.1ms [2026-10-19 22:49:58,326] loop.py:45 INFO TRAIN_STEP step=15532 loss=1.307739 Training: 86%|████████▌ | 15532/18018 [33:19:44<4:35:27, 6.65s/step, loss=1.308, acc=0.701, acc_len=1.68, thru=19.1, I=18.7, T=38.7, wait=3.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:50:05,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:50:06,032] loop.py:45 INFO TRAIN_STEP step=15533 loss=1.254776 Training: 86%|████████▌ | 15533/18018 [33:19:52<4:48:31, 6.97s/step, loss=1.255, acc=0.710, acc_len=1.71, thru=17.6, I=19.0, T=36.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:50:06,066] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 22:50:13,596] loop.py:45 INFO TRAIN_STEP step=15534 loss=1.168812 Training: 86%|████████▌ | 15534/18018 [33:19:59<4:55:49, 7.15s/step, loss=1.169, acc=0.726, acc_len=1.77, thru=17.5, I=19.7, T=38.2, wait=4.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:50:21,415] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:50:22,073] loop.py:380 INFO TIMING step=15535: step=3.600s data=1.402s compute=3.565s [fwd=2.568s bwd=0.981s opt=0.003s] dispatch=4.867s [2026-10-19 22:50:22,074] loop.py:45 INFO TRAIN_STEP step=15535 loss=1.294118 Training: 86%|████████▌ | 15535/18018 [33:20:08<5:12:12, 7.54s/step, loss=1.294, acc=0.697, acc_len=1.67, thru=15.1, I=16.7, T=35.6, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:50:22,422] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:50:28,229] loop.py:45 INFO TRAIN_STEP step=15536 loss=1.166888 Training: 86%|████████▌ | 15536/18018 [33:20:14<4:54:50, 7.13s/step, loss=1.167, acc=0.724, acc_len=1.77, thru=18.1, I=21.5, T=35.2, wait=2.5s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:50:28,225] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15535: forward=2701.6ms backward=893.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:50:36,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:50:37,566] loop.py:45 INFO TRAIN_STEP step=15537 loss=1.238013 Training: 86%|████████▌ | 15537/18018 [33:20:23<5:22:10, 7.79s/step, loss=1.238, acc=0.714, acc_len=1.73, thru=16.9, I=17.0, T=26.5, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:50:37,768] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:50:47,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:50:48,615] loop.py:45 INFO TRAIN_STEP step=15538 loss=1.413728 Training: 86%|████████▌ | 15538/18018 [33:20:34<6:02:25, 8.77s/step, loss=1.414, acc=0.682, acc_len=1.61, thru=12.1, I=12.6, T=27.5, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:50:49,113] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 22:50:54,606] loop.py:45 INFO TRAIN_STEP step=15539 loss=1.244646 Training: 86%|████████▌ | 15539/18018 [33:20:40<5:27:51, 7.94s/step, loss=1.245, acc=0.712, acc_len=1.71, thru=17.3, I=18.5, T=40.0, wait=2.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:51:00,503] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:51:01,599] loop.py:380 INFO TIMING step=15540: step=3.400s data=1.373s compute=3.360s [fwd=2.383s bwd=0.962s opt=0.003s] dispatch=3.582s [2026-10-19 22:51:01,599] loop.py:45 INFO TRAIN_STEP step=15540 loss=1.340004 Training: 86%|████████▌ | 15540/18018 [33:20:47<5:16:03, 7.65s/step, loss=1.340, acc=0.695, acc_len=1.66, thru=19.1, I=21.6, T=37.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:51:02,011] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:51:08,361] loop.py:45 INFO TRAIN_STEP step=15541 loss=1.313999 Training: 86%|████████▋ | 15541/18018 [33:20:54<5:04:51, 7.38s/step, loss=1.314, acc=0.700, acc_len=1.68, thru=17.5, I=20.6, T=39.1, wait=3.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:51:08,357] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15540: forward=2331.9ms backward=895.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:51:17,060] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:51:17,360] loop.py:45 INFO TRAIN_STEP step=15542 loss=1.314912 Training: 86%|████████▋ | 15542/18018 [33:21:03<5:24:44, 7.87s/step, loss=1.315, acc=0.698, acc_len=1.67, thru=16.5, I=15.5, T=35.4, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:51:17,591] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:51:23,708] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:51:24,032] loop.py:45 INFO TRAIN_STEP step=15543 loss=1.333036 Training: 86%|████████▋ | 15543/18018 [33:21:10<5:09:47, 7.51s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=19.1, I=19.8, T=38.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:51:24,209] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:51:30,740] loop.py:45 INFO TRAIN_STEP step=15544 loss=1.392672 Training: 86%|████████▋ | 15544/18018 [33:21:16<4:59:44, 7.27s/step, loss=1.393, acc=0.682, acc_len=1.62, thru=18.6, I=20.7, T=38.0, wait=3.3s, pool=57, epoch=2/2][2026-10-19 22:51:38,651] loop.py:380 INFO TIMING step=15545: step=3.372s data=1.399s compute=3.329s [fwd=2.349s bwd=0.966s opt=0.003s] dispatch=4.531s [2026-10-19 22:51:38,651] loop.py:45 INFO TRAIN_STEP step=15545 loss=1.214502 Training: 86%|████████▋ | 15545/18018 [33:21:24<5:07:35, 7.46s/step, loss=1.215, acc=0.713, acc_len=1.73, thru=15.3, I=18.2, T=38.0, wait=4.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:51:45,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:51:46,691] loop.py:45 INFO TRAIN_STEP step=15546 loss=1.350655 Training: 86%|████████▋ | 15546/18018 [33:21:32<5:14:33, 7.64s/step, loss=1.351, acc=0.689, acc_len=1.64, thru=15.8, I=17.5, T=31.5, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:51:46,687] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15545: forward=3080.9ms backward=930.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:51:46,761] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 22:51:55,096] loop.py:45 INFO TRAIN_STEP step=15547 loss=1.234942 Training: 86%|████████▋ | 15547/18018 [33:21:41<5:23:58, 7.87s/step, loss=1.235, acc=0.711, acc_len=1.71, thru=13.6, I=15.6, T=33.1, wait=4.5s, pool=49, epoch=2/2][2026-10-19 22:52:02,580] loop.py:45 INFO TRAIN_STEP step=15548 loss=1.235685 Training: 86%|████████▋ | 15548/18018 [33:21:48<5:19:07, 7.75s/step, loss=1.236, acc=0.710, acc_len=1.71, thru=14.6, I=17.1, T=35.1, wait=3.8s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 22:52:03,856] eagle_store.py:517 WARNING batch_get_buffer missing keys (k497534-82173006_hs, k497534-82173006_ids, k497534-82173006_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:11,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 22:52:03,853] eagle_store.py:517 WARNING batch_get_buffer missing keys (k497533-90e083c1_hs, k497533-90e083c1_ids, k497533-90e083c1_lhs); sleeping 0.50s. [2026-10-19 22:52:12,250] loop.py:45 INFO TRAIN_STEP step=15549 loss=1.206622 Training: 86%|████████▋ | 15549/18018 [33:21:58<5:42:38, 8.33s/step, loss=1.207, acc=0.721, acc_len=1.75, thru=17.8, I=14.4, T=36.2, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:12,298] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 22:52:20,022] loop.py:380 INFO TIMING step=15550: step=3.382s data=1.365s compute=3.341s [fwd=2.390s bwd=0.936s opt=0.003s] dispatch=4.381s [2026-10-19 22:52:20,022] loop.py:45 INFO TRAIN_STEP step=15550 loss=1.362172 Training: 86%|████████▋ | 15550/18018 [33:22:06<5:35:38, 8.16s/step, loss=1.362, acc=0.688, acc_len=1.64, thru=15.6, I=15.4, T=37.9, wait=4.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:27,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:52:27,946] loop.py:45 INFO TRAIN_STEP step=15551 loss=1.313534 Training: 86%|████████▋ | 15551/18018 [33:22:14<5:32:37, 8.09s/step, loss=1.314, acc=0.694, acc_len=1.67, thru=19.2, I=18.9, T=33.8, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:52:27,942] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15550: forward=2726.3ms backward=1017.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:28,075] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:33,232] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:52:33,457] loop.py:45 INFO TRAIN_STEP step=15552 loss=1.352391 Training: 86%|████████▋ | 15552/18018 [33:22:19<5:00:41, 7.32s/step, loss=1.352, acc=0.692, acc_len=1.65, thru=20.8, I=23.4, T=38.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:33,735] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:38,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:52:39,462] loop.py:45 INFO TRAIN_STEP step=15553 loss=1.201981 Training: 86%|████████▋ | 15553/18018 [33:22:25<4:44:24, 6.92s/step, loss=1.202, acc=0.716, acc_len=1.74, thru=21.1, I=22.9, T=38.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:39,902] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:52:45,391] loop.py:45 INFO TRAIN_STEP step=15554 loss=1.313425 Training: 86%|████████▋ | 15554/18018 [33:22:31<4:32:02, 6.62s/step, loss=1.313, acc=0.700, acc_len=1.67, thru=21.2, I=23.2, T=39.0, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:53,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:52:54,587] loop.py:380 INFO TIMING step=15555: step=3.623s data=1.403s compute=3.594s [fwd=2.615s bwd=0.965s opt=0.003s] dispatch=5.564s [2026-10-19 22:52:54,588] loop.py:45 INFO TRAIN_STEP step=15555 loss=1.262817 Training: 86%|████████▋ | 15555/18018 [33:22:40<5:03:37, 7.40s/step, loss=1.263, acc=0.710, acc_len=1.72, thru=15.0, I=16.2, T=35.3, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:54,653] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:52:59,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:53:00,124] loop.py:45 INFO TRAIN_STEP step=15556 loss=1.235244 Training: 86%|████████▋ | 15556/18018 [33:22:46<4:40:36, 6.84s/step, loss=1.235, acc=0.710, acc_len=1.71, thru=21.7, I=24.7, T=39.7, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:53:00,119] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15555: forward=2270.8ms backward=899.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:53:00,581] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:53:05,840] loop.py:45 INFO TRAIN_STEP step=15557 loss=1.166542 Training: 86%|████████▋ | 15557/18018 [33:22:51<4:26:39, 6.50s/step, loss=1.167, acc=0.723, acc_len=1.76, thru=22.1, I=24.4, T=39.3, wait=2.4s, pool=61, epoch=2/2][2026-10-19 22:53:12,832] loop.py:45 INFO TRAIN_STEP step=15558 loss=1.292941 Training: 86%|████████▋ | 15558/18018 [33:22:58<4:32:34, 6.65s/step, loss=1.293, acc=0.699, acc_len=1.67, thru=16.9, I=17.5, T=37.7, wait=3.6s, pool=49, epoch=2/2][2026-10-19 22:53:21,672] loop.py:45 INFO TRAIN_STEP step=15559 loss=1.330289 Training: 86%|████████▋ | 15559/18018 [33:23:07<4:59:25, 7.31s/step, loss=1.330, acc=0.700, acc_len=1.68, thru=16.1, I=15.1, T=35.9, wait=5.3s, pool=57, epoch=2/2][2026-10-19 22:53:27,550] loop.py:380 INFO TIMING step=15560: step=3.393s data=1.405s compute=3.346s [fwd=2.364s bwd=0.968s opt=0.003s] dispatch=2.476s [2026-10-19 22:53:27,550] loop.py:45 INFO TRAIN_STEP step=15560 loss=1.167002 Training: 86%|████████▋ | 15560/18018 [33:23:13<4:41:45, 6.88s/step, loss=1.167, acc=0.722, acc_len=1.75, thru=19.7, I=22.1, T=37.7, wait=2.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:53:38,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:53:39,237] loop.py:45 INFO TRAIN_STEP step=15561 loss=1.404014 Training: 86%|████████▋ | 15561/18018 [33:23:25<5:40:44, 8.32s/step, loss=1.404, acc=0.682, acc_len=1.61, thru=11.3, I=12.2, T=33.5, wait=7.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:53:39,233] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15560: forward=2691.7ms backward=1078.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:53:39,479] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:53:47,713] loop.py:45 INFO TRAIN_STEP step=15562 loss=1.259260 Training: 86%|████████▋ | 15562/18018 [33:23:33<5:42:30, 8.37s/step, loss=1.259, acc=0.708, acc_len=1.71, thru=13.4, I=14.4, T=35.2, wait=4.8s, pool=57, epoch=2/2][2026-10-19 22:53:55,435] loop.py:45 INFO TRAIN_STEP step=15563 loss=1.269147 Training: 86%|████████▋ | 15563/18018 [33:23:41<5:34:26, 8.17s/step, loss=1.269, acc=0.707, acc_len=1.70, thru=14.9, I=15.6, T=37.7, wait=4.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:04,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:54:04,911] loop.py:45 INFO TRAIN_STEP step=15564 loss=1.321036 Training: 86%|████████▋ | 15564/18018 [33:23:50<5:50:16, 8.56s/step, loss=1.321, acc=0.697, acc_len=1.67, thru=16.6, I=15.8, T=34.6, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:05,261] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:54:12,627] loop.py:380 INFO TIMING step=15565: step=3.354s data=1.307s compute=3.326s [fwd=2.401s bwd=0.911s opt=0.003s] dispatch=4.351s [2026-10-19 22:54:12,627] loop.py:45 INFO TRAIN_STEP step=15565 loss=1.186237 Training: 86%|████████▋ | 15565/18018 [33:23:58<5:39:44, 8.31s/step, loss=1.186, acc=0.721, acc_len=1.75, thru=11.7, I=16.8, T=38.2, wait=4.4s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 22:54:12,962] eagle_store.py:517 WARNING batch_get_buffer missing keys (k498090-945b94bb_hs, k498090-945b94bb_ids, k498090-945b94bb_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:22,557] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 22:54:12,958] eagle_store.py:517 WARNING batch_get_buffer missing keys (k498089-8ee81e38_hs, k498089-8ee81e38_ids, k498089-8ee81e38_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 22:54:22,760] loop.py:45 INFO TRAIN_STEP step=15566 loss=1.308755 Training: 86%|████████▋ | 15566/18018 [33:24:08<6:01:57, 8.86s/step, loss=1.309, acc=0.701, acc_len=1.68, thru=17.5, I=14.2, T=34.2, wait=6.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:54:22,755] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15565: forward=2627.7ms backward=1066.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:23,058] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:54:31,594] loop.py:45 INFO TRAIN_STEP step=15567 loss=1.366156 Training: 86%|████████▋ | 15567/18018 [33:24:17<6:01:32, 8.85s/step, loss=1.366, acc=0.686, acc_len=1.63, thru=12.4, I=13.4, T=36.6, wait=5.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:38,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:54:39,906] loop.py:45 INFO TRAIN_STEP step=15568 loss=1.250585 Training: 86%|████████▋ | 15568/18018 [33:24:25<5:54:46, 8.69s/step, loss=1.251, acc=0.710, acc_len=1.71, thru=17.2, I=18.7, T=34.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:40,277] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:45,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:54:46,375] loop.py:45 INFO TRAIN_STEP step=15569 loss=1.358368 Training: 86%|████████▋ | 15569/18018 [33:24:32<5:27:27, 8.02s/step, loss=1.358, acc=0.685, acc_len=1.63, thru=19.0, I=23.2, T=39.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:46,409] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:50,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:54:52,106] loop.py:380 INFO TIMING step=15570: step=3.289s data=1.294s compute=3.262s [fwd=2.345s bwd=0.903s opt=0.003s] dispatch=2.434s [2026-10-19 22:54:52,107] loop.py:45 INFO TRAIN_STEP step=15570 loss=1.209322 Training: 86%|████████▋ | 15570/18018 [33:24:38<4:59:16, 7.34s/step, loss=1.209, acc=0.716, acc_len=1.74, thru=21.5, I=25.0, T=38.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:54:52,452] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:54:59,605] loop.py:45 INFO TRAIN_STEP step=15571 loss=1.274126 Training: 86%|████████▋ | 15571/18018 [33:24:45<5:01:08, 7.38s/step, loss=1.274, acc=0.705, acc_len=1.70, thru=16.4, I=19.3, T=37.6, wait=4.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:54:59,600] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15570: forward=2399.0ms backward=958.9ms [2026-10-19 22:55:07,486] loop.py:45 INFO TRAIN_STEP step=15572 loss=1.245226 Training: 86%|████████▋ | 15572/18018 [33:24:53<5:07:05, 7.53s/step, loss=1.245, acc=0.710, acc_len=1.72, thru=17.3, I=16.9, T=36.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:55:07,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:55:07,978] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:55:14,653] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:55:15,194] loop.py:45 INFO TRAIN_STEP step=15573 loss=1.377491 Training: 86%|████████▋ | 15573/18018 [33:25:01<5:09:06, 7.59s/step, loss=1.377, acc=0.689, acc_len=1.63, thru=17.0, I=18.5, T=36.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:55:15,657] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:55:21,417] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:55:22,099] loop.py:45 INFO TRAIN_STEP step=15574 loss=1.294214 Training: 86%|████████▋ | 15574/18018 [33:25:08<5:00:40, 7.38s/step, loss=1.294, acc=0.700, acc_len=1.67, thru=18.3, I=20.9, T=37.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:55:22,422] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:55:27,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:55:28,099] loop.py:380 INFO TIMING step=15575: step=3.371s data=1.344s compute=3.345s [fwd=2.394s bwd=0.937s opt=0.003s] dispatch=2.618s [2026-10-19 22:55:28,099] loop.py:45 INFO TRAIN_STEP step=15575 loss=1.209950 Training: 86%|████████▋ | 15575/18018 [33:25:14<4:43:40, 6.97s/step, loss=1.210, acc=0.716, acc_len=1.74, thru=21.2, I=22.8, T=38.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:55:28,385] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:55:35,234] loop.py:45 INFO TRAIN_STEP step=15576 loss=1.265278 Training: 86%|████████▋ | 15576/18018 [33:25:21<4:45:36, 7.02s/step, loss=1.265, acc=0.707, acc_len=1.70, thru=17.0, I=19.1, T=37.7, wait=3.7s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:55:35,230] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15575: forward=2407.5ms backward=939.8ms [2026-10-19 22:55:41,754] loop.py:45 INFO TRAIN_STEP step=15577 loss=1.101929 Training: 86%|████████▋ | 15577/18018 [33:25:27<4:39:26, 6.87s/step, loss=1.102, acc=0.732, acc_len=1.79, thru=18.5, I=19.2, T=37.5, wait=3.1s, pool=53, epoch=2/2][2026-10-19 22:55:50,638] loop.py:45 INFO TRAIN_STEP step=15578 loss=1.242164 Training: 86%|████████▋ | 15578/18018 [33:25:36<5:03:52, 7.47s/step, loss=1.242, acc=0.710, acc_len=1.72, thru=15.6, I=14.5, T=37.8, wait=5.5s, pool=57, epoch=2/2][2026-10-19 22:55:58,255] loop.py:45 INFO TRAIN_STEP step=15579 loss=1.275793 Training: 86%|████████▋ | 15579/18018 [33:25:44<5:05:31, 7.52s/step, loss=1.276, acc=0.707, acc_len=1.71, thru=16.1, I=16.5, T=38.5, wait=4.3s, pool=45, epoch=2/2][2026-10-19 22:56:07,203] loop.py:380 INFO TIMING step=15580: step=3.802s data=1.451s compute=3.772s [fwd=2.698s bwd=1.059s opt=0.003s] dispatch=5.138s [2026-10-19 22:56:07,204] loop.py:45 INFO TRAIN_STEP step=15580 loss=1.246479 Training: 86%|████████▋ | 15580/18018 [33:25:53<5:22:52, 7.95s/step, loss=1.246, acc=0.709, acc_len=1.71, thru=13.4, I=13.7, T=33.7, wait=5.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:56:17,425] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:56:18,249] loop.py:45 INFO TRAIN_STEP step=15581 loss=1.244631 Training: 86%|████████▋ | 15581/18018 [33:26:04<6:00:30, 8.88s/step, loss=1.245, acc=0.708, acc_len=1.70, thru=15.4, I=14.5, T=35.2, wait=7.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:56:18,245] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15580: forward=2576.8ms backward=1017.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:56:18,432] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:56:24,425] loop.py:45 INFO TRAIN_STEP step=15582 loss=1.203976 Training: 86%|████████▋ | 15582/18018 [33:26:10<5:27:30, 8.07s/step, loss=1.204, acc=0.720, acc_len=1.75, thru=17.9, I=18.4, T=40.2, wait=3.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:56:31,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:56:33,065] loop.py:45 INFO TRAIN_STEP step=15583 loss=1.436690 Training: 86%|████████▋ | 15583/18018 [33:26:19<5:34:18, 8.24s/step, loss=1.437, acc=0.672, acc_len=1.58, thru=18.3, I=18.7, T=36.8, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:56:33,297] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:56:38,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:56:39,350] loop.py:45 INFO TRAIN_STEP step=15584 loss=1.284113 Training: 86%|████████▋ | 15584/18018 [33:26:25<5:10:24, 7.65s/step, loss=1.284, acc=0.704, acc_len=1.70, thru=20.3, I=25.1, T=40.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:56:39,810] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:56:45,680] loop.py:380 INFO TIMING step=15585: step=3.290s data=1.327s compute=3.259s [fwd=2.356s bwd=0.890s opt=0.003s] dispatch=3.031s [2026-10-19 22:56:45,680] loop.py:45 INFO TRAIN_STEP step=15585 loss=1.224114 Training: 86%|████████▋ | 15585/18018 [33:26:31<4:54:11, 7.25s/step, loss=1.224, acc=0.712, acc_len=1.72, thru=20.2, I=22.2, T=38.9, wait=3.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:56:53,063] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:56:53,598] loop.py:45 INFO TRAIN_STEP step=15586 loss=1.209445 Training: 87%|████████▋ | 15586/18018 [33:26:39<5:02:09, 7.45s/step, loss=1.209, acc=0.716, acc_len=1.73, thru=17.7, I=18.6, T=36.3, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:56:53,593] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15585: forward=2481.2ms backward=1010.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:56:54,068] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:57:00,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:57:01,217] loop.py:45 INFO TRAIN_STEP step=15587 loss=1.270061 Training: 87%|████████▋ | 15587/18018 [33:26:47<5:04:01, 7.50s/step, loss=1.270, acc=0.706, acc_len=1.70, thru=16.9, I=19.6, T=38.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:57:01,622] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:57:06,939] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:57:07,645] loop.py:45 INFO TRAIN_STEP step=15588 loss=1.235030 Training: 87%|████████▋ | 15588/18018 [33:26:53<4:50:51, 7.18s/step, loss=1.235, acc=0.712, acc_len=1.72, thru=20.0, I=20.9, T=38.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:57:07,944] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:57:15,411] loop.py:45 INFO TRAIN_STEP step=15589 loss=1.275053 Training: 87%|████████▋ | 15589/18018 [33:27:01<4:57:50, 7.36s/step, loss=1.275, acc=0.703, acc_len=1.69, thru=14.1, I=18.3, T=37.1, wait=4.3s, pool=41, epoch=2/2][2026-10-19 22:57:21,772] loop.py:380 INFO TIMING step=15590: step=3.355s data=1.327s compute=3.324s [fwd=2.386s bwd=0.924s opt=0.003s] dispatch=2.995s [2026-10-19 22:57:21,772] loop.py:45 INFO TRAIN_STEP step=15590 loss=1.204399 Training: 87%|████████▋ | 15590/18018 [33:27:07<4:45:35, 7.06s/step, loss=1.204, acc=0.717, acc_len=1.74, thru=19.1, I=18.3, T=38.2, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:57:27,544] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:57:27,795] loop.py:45 INFO TRAIN_STEP step=15591 loss=1.260559 Training: 87%|████████▋ | 15591/18018 [33:27:13<4:32:56, 6.75s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=21.6, I=21.8, T=39.2, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:57:27,789] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15590: forward=2340.1ms backward=884.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:57:28,047] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:57:35,458] loop.py:45 INFO TRAIN_STEP step=15592 loss=1.364335 Training: 87%|████████▋ | 15592/18018 [33:27:21<4:43:56, 7.02s/step, loss=1.364, acc=0.690, acc_len=1.65, thru=15.8, I=16.1, T=37.6, wait=4.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:57:41,429] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:57:42,853] loop.py:45 INFO TRAIN_STEP step=15593 loss=1.280843 Training: 87%|████████▋ | 15593/18018 [33:27:28<4:48:20, 7.13s/step, loss=1.281, acc=0.702, acc_len=1.68, thru=18.3, I=20.2, T=37.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:57:42,942] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-19 22:57:49,388] loop.py:45 INFO TRAIN_STEP step=15594 loss=1.358817 Training: 87%|████████▋ | 15594/18018 [33:27:35<4:40:56, 6.95s/step, loss=1.359, acc=0.686, acc_len=1.63, thru=16.8, I=23.5, T=39.6, wait=3.3s, pool=37, epoch=2/2][2026-10-19 22:57:56,336] loop.py:380 INFO TIMING step=15595: step=3.300s data=1.291s compute=3.268s [fwd=2.349s bwd=0.903s opt=0.003s] dispatch=3.640s [2026-10-19 22:57:56,336] loop.py:45 INFO TRAIN_STEP step=15595 loss=1.291201 Training: 87%|████████▋ | 15595/18018 [33:27:42<4:40:44, 6.95s/step, loss=1.291, acc=0.698, acc_len=1.67, thru=18.8, I=18.7, T=38.8, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:01,062] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 22:58:02,488] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15595: forward=2370.8ms backward=912.8ms [2026-10-19 22:58:02,492] loop.py:45 INFO TRAIN_STEP step=15596 loss=1.366944 Training: 87%|████████▋ | 15596/18018 [33:27:48<4:31:00, 6.71s/step, loss=1.367, acc=0.684, acc_len=1.62, thru=21.8, I=23.8, T=38.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:02,576] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:10,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:58:10,762] loop.py:45 INFO TRAIN_STEP step=15597 loss=1.289564 Training: 87%|████████▋ | 15597/18018 [33:27:56<4:49:42, 7.18s/step, loss=1.290, acc=0.700, acc_len=1.67, thru=13.8, I=16.9, T=36.4, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:11,058] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 22:58:16,848] loop.py:45 INFO TRAIN_STEP step=15598 loss=1.287345 Training: 87%|████████▋ | 15598/18018 [33:28:02<4:36:22, 6.85s/step, loss=1.287, acc=0.705, acc_len=1.69, thru=19.5, I=17.6, T=39.1, wait=2.8s, pool=57, epoch=2/2][2026-10-19 22:58:23,793] loop.py:45 INFO TRAIN_STEP step=15599 loss=1.290513 Training: 87%|████████▋ | 15599/18018 [33:28:09<4:37:22, 6.88s/step, loss=1.291, acc=0.701, acc_len=1.68, thru=15.8, I=18.9, T=37.4, wait=3.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:34,469] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:58:34,920] loop.py:380 INFO TIMING step=15600: step=3.979s data=1.568s compute=3.949s [fwd=2.780s bwd=1.154s opt=0.003s] dispatch=7.140s [2026-10-19 22:58:34,921] loop.py:45 INFO TRAIN_STEP step=15600 loss=1.325978 Training: 87%|████████▋ | 15600/18018 [33:28:21<5:28:38, 8.15s/step, loss=1.326, acc=0.698, acc_len=1.67, thru=15.6, I=14.0, T=32.2, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:34,972] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:40,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:58:41,025] loop.py:45 INFO TRAIN_STEP step=15601 loss=1.295689 Training: 87%|████████▋ | 15601/18018 [33:28:27<5:03:42, 7.54s/step, loss=1.296, acc=0.700, acc_len=1.69, thru=19.1, I=21.2, T=38.0, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:58:41,021] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15600: forward=2388.4ms backward=928.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:41,267] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 22:58:46,635] loop.py:45 INFO TRAIN_STEP step=15602 loss=1.312578 Training: 87%|████████▋ | 15602/18018 [33:28:32<4:40:17, 6.96s/step, loss=1.313, acc=0.696, acc_len=1.67, thru=20.2, I=22.6, T=39.7, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:56,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:58:57,290] loop.py:45 INFO TRAIN_STEP step=15603 loss=1.187276 Training: 87%|████████▋ | 15603/18018 [33:28:43<5:24:45, 8.07s/step, loss=1.187, acc=0.721, acc_len=1.76, thru=14.1, I=14.8, T=35.4, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:58:57,687] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:59:03,261] loop.py:45 INFO TRAIN_STEP step=15604 loss=1.281995 Training: 87%|████████▋ | 15604/18018 [33:28:49<4:59:17, 7.44s/step, loss=1.282, acc=0.704, acc_len=1.69, thru=19.1, I=20.0, T=38.0, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:08,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:59:10,121] loop.py:380 INFO TIMING step=15605: step=3.455s data=1.440s compute=3.424s [fwd=2.435s bwd=0.974s opt=0.003s] dispatch=3.398s [2026-10-19 22:59:10,121] loop.py:45 INFO TRAIN_STEP step=15605 loss=1.278275 Training: 87%|████████▋ | 15605/18018 [33:28:56<4:52:11, 7.27s/step, loss=1.278, acc=0.703, acc_len=1.69, thru=19.5, I=21.8, T=37.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:10,339] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:59:15,260] loop.py:45 INFO TRAIN_STEP step=15606 loss=1.234983 Training: 87%|████████▋ | 15606/18018 [33:29:01<4:26:25, 6.63s/step, loss=1.235, acc=0.708, acc_len=1.70, thru=22.5, I=26.0, T=39.9, wait=1.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:59:15,256] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15605: forward=2285.7ms backward=886.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:15,250] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:15,752] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:20,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:59:21,190] loop.py:45 INFO TRAIN_STEP step=15607 loss=1.187782 Training: 87%|████████▋ | 15607/18018 [33:29:07<4:17:55, 6.42s/step, loss=1.188, acc=0.719, acc_len=1.75, thru=23.8, I=23.7, T=39.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:21,220] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:28,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:59:29,259] loop.py:45 INFO TRAIN_STEP step=15608 loss=1.211379 Training: 87%|████████▋ | 15608/18018 [33:29:15<4:37:41, 6.91s/step, loss=1.211, acc=0.714, acc_len=1.73, thru=17.3, I=18.6, T=36.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:29,346] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:33,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 22:59:34,554] loop.py:45 INFO TRAIN_STEP step=15609 loss=1.172147 Training: 87%|████████▋ | 15609/18018 [33:29:20<4:18:05, 6.43s/step, loss=1.172, acc=0.722, acc_len=1.76, thru=21.7, I=27.6, T=40.1, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 22:59:34,948] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 22:59:40,790] loop.py:380 INFO TIMING step=15610: step=3.237s data=1.287s compute=3.197s [fwd=2.286s bwd=0.897s opt=0.003s] dispatch=2.990s [2026-10-19 22:59:40,790] loop.py:45 INFO TRAIN_STEP step=15610 loss=1.248132 Training: 87%|████████▋ | 15610/18018 [33:29:26<4:15:40, 6.37s/step, loss=1.248, acc=0.706, acc_len=1.70, thru=18.4, I=25.3, T=39.5, wait=3.0s, pool=41, epoch=2/2][2026-10-19 22:59:49,119] loop.py:45 INFO TRAIN_STEP step=15611 loss=1.315157 Training: 87%|████████▋ | 15611/18018 [33:29:35<4:39:07, 6.96s/step, loss=1.315, acc=0.697, acc_len=1.66, thru=12.9, I=12.8, T=36.6, wait=4.8s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 22:59:49,115] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15610: forward=2403.8ms backward=1034.9ms [2026-10-19 22:59:57,530] loop.py:45 INFO TRAIN_STEP step=15612 loss=1.221907 Training: 87%|████████▋ | 15612/18018 [33:29:43<4:56:28, 7.39s/step, loss=1.222, acc=0.714, acc_len=1.73, thru=18.2, I=15.6, T=37.4, wait=5.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:02,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:00:03,236] loop.py:45 INFO TRAIN_STEP step=15613 loss=1.312818 Training: 87%|████████▋ | 15613/18018 [33:29:49<4:36:06, 6.89s/step, loss=1.313, acc=0.698, acc_len=1.67, thru=20.9, I=22.6, T=37.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:03,312] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:09,326] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:00:09,448] loop.py:45 INFO TRAIN_STEP step=15614 loss=1.269601 Training: 87%|████████▋ | 15614/18018 [33:29:55<4:27:49, 6.68s/step, loss=1.270, acc=0.707, acc_len=1.71, thru=21.0, I=23.3, T=38.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:09,829] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:00:15,307] loop.py:380 INFO TIMING step=15615: step=3.241s data=1.276s compute=3.213s [fwd=2.315s bwd=0.883s opt=0.003s] dispatch=2.610s [2026-10-19 23:00:15,307] loop.py:45 INFO TRAIN_STEP step=15615 loss=1.244494 Training: 87%|████████▋ | 15615/18018 [33:30:01<4:17:49, 6.44s/step, loss=1.244, acc=0.708, acc_len=1.71, thru=21.1, I=23.0, T=39.5, wait=2.6s, pool=61, epoch=2/2][2026-10-19 23:00:21,863] loop.py:45 INFO TRAIN_STEP step=15616 loss=1.048895 Training: 87%|████████▋ | 15616/18018 [33:30:07<4:19:07, 6.47s/step, loss=1.049, acc=0.743, acc_len=1.83, thru=20.1, I=21.9, T=37.5, wait=3.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:00:21,858] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15615: forward=2434.2ms backward=937.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:29,626] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:00:30,442] loop.py:45 INFO TRAIN_STEP step=15617 loss=1.095665 Training: 87%|████████▋ | 15617/18018 [33:30:16<4:44:17, 7.10s/step, loss=1.096, acc=0.738, acc_len=1.81, thru=15.1, I=15.5, T=36.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:30,632] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:37,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:00:38,456] loop.py:45 INFO TRAIN_STEP step=15618 loss=1.204014 Training: 87%|████████▋ | 15618/18018 [33:30:24<4:55:06, 7.38s/step, loss=1.204, acc=0.720, acc_len=1.75, thru=16.6, I=18.9, T=33.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:38,878] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:00:46,023] loop.py:45 INFO TRAIN_STEP step=15619 loss=1.158823 Training: 87%|████████▋ | 15619/18018 [33:30:32<4:57:15, 7.43s/step, loss=1.159, acc=0.724, acc_len=1.76, thru=16.0, I=17.2, T=38.2, wait=4.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:51,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:00:52,509] loop.py:380 INFO TIMING step=15620: step=3.396s data=1.380s compute=3.361s [fwd=2.410s bwd=0.937s opt=0.003s] dispatch=3.079s [2026-10-19 23:00:52,509] loop.py:45 INFO TRAIN_STEP step=15620 loss=1.226651 Training: 87%|████████▋ | 15620/18018 [33:30:38<4:45:45, 7.15s/step, loss=1.227, acc=0.711, acc_len=1.72, thru=19.9, I=18.9, T=37.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:00:52,819] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:00:59,207] loop.py:45 INFO TRAIN_STEP step=15621 loss=1.220301 Training: 87%|████████▋ | 15621/18018 [33:30:45<4:40:12, 7.01s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=16.9, I=22.7, T=38.8, wait=3.4s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:00:59,202] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15620: forward=2357.4ms backward=895.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:09,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:01:09,406] loop.py:45 INFO TRAIN_STEP step=15622 loss=1.400692 Training: 87%|████████▋ | 15622/18018 [33:30:55<5:18:14, 7.97s/step, loss=1.401, acc=0.681, acc_len=1.61, thru=15.9, I=13.3, T=34.4, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:09,844] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:01:16,372] loop.py:45 INFO TRAIN_STEP step=15623 loss=1.203504 Training: 87%|████████▋ | 15623/18018 [33:31:02<5:06:07, 7.67s/step, loss=1.204, acc=0.719, acc_len=1.75, thru=17.1, I=19.7, T=38.0, wait=3.6s, pool=57, epoch=2/2][2026-10-19 23:01:23,646] loop.py:45 INFO TRAIN_STEP step=15624 loss=1.227237 Training: 87%|████████▋ | 15624/18018 [33:31:09<5:01:16, 7.55s/step, loss=1.227, acc=0.712, acc_len=1.72, thru=17.7, I=18.3, T=38.2, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:30,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:01:30,709] loop.py:380 INFO TIMING step=15625: step=3.502s data=1.433s compute=3.476s [fwd=2.503s bwd=0.958s opt=0.003s] dispatch=3.551s [2026-10-19 23:01:30,709] loop.py:45 INFO TRAIN_STEP step=15625 loss=1.202445 Training: 87%|████████▋ | 15625/18018 [33:31:16<4:55:16, 7.40s/step, loss=1.202, acc=0.716, acc_len=1.74, thru=18.8, I=19.7, T=36.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:31,155] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:36,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:01:37,017] loop.py:45 INFO TRAIN_STEP step=15626 loss=1.228465 Training: 87%|████████▋ | 15626/18018 [33:31:23<4:42:04, 7.08s/step, loss=1.228, acc=0.711, acc_len=1.72, thru=20.3, I=22.4, T=38.7, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:01:37,013] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15625: forward=2359.6ms backward=905.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:37,160] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 23:01:43,453] loop.py:45 INFO TRAIN_STEP step=15627 loss=1.254738 Training: 87%|████████▋ | 15627/18018 [33:31:29<4:34:18, 6.88s/step, loss=1.255, acc=0.707, acc_len=1.71, thru=17.6, I=18.2, T=38.7, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:51,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:01:51,660] loop.py:45 INFO TRAIN_STEP step=15628 loss=1.343130 Training: 87%|████████▋ | 15628/18018 [33:31:37<4:50:00, 7.28s/step, loss=1.343, acc=0.690, acc_len=1.65, thru=18.1, I=16.5, T=37.7, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:51,981] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:56,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:01:57,065] loop.py:45 INFO TRAIN_STEP step=15629 loss=1.239117 Training: 87%|████████▋ | 15629/18018 [33:31:43<4:27:30, 6.72s/step, loss=1.239, acc=0.712, acc_len=1.72, thru=22.0, I=25.8, T=38.7, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:01:57,095] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 23:02:04,822] loop.py:380 INFO TIMING step=15630: step=3.438s data=1.442s compute=3.405s [fwd=2.451s bwd=0.939s opt=0.003s] dispatch=4.308s [2026-10-19 23:02:04,822] loop.py:45 INFO TRAIN_STEP step=15630 loss=1.364371 Training: 87%|████████▋ | 15630/18018 [33:31:50<4:39:48, 7.03s/step, loss=1.364, acc=0.684, acc_len=1.63, thru=16.1, I=18.5, T=37.2, wait=4.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:13,366] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:02:13,981] loop.py:45 INFO TRAIN_STEP step=15631 loss=1.319960 Training: 87%|████████▋ | 15631/18018 [33:32:00<5:05:03, 7.67s/step, loss=1.320, acc=0.697, acc_len=1.67, thru=14.6, I=15.2, T=34.9, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:02:13,977] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15630: forward=2597.7ms backward=1020.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:14,373] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:20,375] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:02:21,111] loop.py:45 INFO TRAIN_STEP step=15632 loss=1.368379 Training: 87%|████████▋ | 15632/18018 [33:32:07<4:58:31, 7.51s/step, loss=1.368, acc=0.689, acc_len=1.64, thru=18.1, I=21.1, T=38.1, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:21,381] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:26,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:02:26,896] loop.py:45 INFO TRAIN_STEP step=15633 loss=1.128217 Training: 87%|████████▋ | 15633/18018 [33:32:12<4:37:52, 6.99s/step, loss=1.128, acc=0.729, acc_len=1.77, thru=21.3, I=24.4, T=39.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:27,159] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:33,008] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:02:33,482] loop.py:45 INFO TRAIN_STEP step=15634 loss=1.190638 Training: 87%|████████▋ | 15634/18018 [33:32:19<4:32:54, 6.87s/step, loss=1.191, acc=0.719, acc_len=1.74, thru=19.9, I=21.0, T=38.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:33,516] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:41,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:02:41,921] loop.py:380 INFO TIMING step=15635: step=3.681s data=1.548s compute=3.653s [fwd=2.603s bwd=1.035s opt=0.003s] dispatch=4.750s [2026-10-19 23:02:41,921] loop.py:45 INFO TRAIN_STEP step=15635 loss=1.279117 Training: 87%|████████▋ | 15635/18018 [33:32:28<4:51:30, 7.34s/step, loss=1.279, acc=0.704, acc_len=1.69, thru=15.9, I=16.4, T=34.8, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:41,969] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:48,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:02:49,702] loop.py:45 INFO TRAIN_STEP step=15636 loss=1.283503 Training: 87%|████████▋ | 15636/18018 [33:32:35<4:56:40, 7.47s/step, loss=1.284, acc=0.700, acc_len=1.68, thru=17.3, I=19.6, T=38.3, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:02:49,697] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15635: forward=2366.9ms backward=922.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:49,885] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:54,799] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:02:55,535] loop.py:45 INFO TRAIN_STEP step=15637 loss=1.280713 Training: 87%|████████▋ | 15637/18018 [33:32:41<4:37:00, 6.98s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=21.0, I=26.3, T=39.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:02:55,803] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:03:02,346] loop.py:45 INFO TRAIN_STEP step=15638 loss=1.243857 Training: 87%|████████▋ | 15638/18018 [33:32:48<4:34:50, 6.93s/step, loss=1.244, acc=0.711, acc_len=1.71, thru=19.0, I=23.2, T=39.1, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:03:09,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:03:09,966] loop.py:45 INFO TRAIN_STEP step=15639 loss=1.195424 Training: 87%|████████▋ | 15639/18018 [33:32:56<4:42:57, 7.14s/step, loss=1.195, acc=0.717, acc_len=1.74, thru=17.6, I=17.6, T=36.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:03:10,336] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:03:18,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:03:18,484] loop.py:380 INFO TIMING step=15640: step=3.429s data=1.368s compute=3.390s [fwd=2.407s bwd=0.969s opt=0.003s] dispatch=5.083s [2026-10-19 23:03:18,485] loop.py:45 INFO TRAIN_STEP step=15640 loss=1.233664 Training: 87%|████████▋ | 15640/18018 [33:33:04<4:59:17, 7.55s/step, loss=1.234, acc=0.709, acc_len=1.72, thru=15.7, I=16.7, T=37.3, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:03:18,693] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:03:26,425] loop.py:45 INFO TRAIN_STEP step=15641 loss=1.301671 Training: 87%|████████▋ | 15641/18018 [33:33:12<5:03:46, 7.67s/step, loss=1.302, acc=0.697, acc_len=1.67, thru=15.0, I=15.1, T=37.0, wait=4.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:03:26,420] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15640: forward=2460.5ms backward=953.8ms [2026-10-19 23:03:32,414] loop.py:45 INFO TRAIN_STEP step=15642 loss=1.273332 Training: 87%|████████▋ | 15642/18018 [33:33:18<4:43:42, 7.16s/step, loss=1.273, acc=0.703, acc_len=1.69, thru=20.4, I=21.0, T=38.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:03:32,369] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:03:32,871] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:03:39,785] loop.py:45 INFO TRAIN_STEP step=15643 loss=1.281125 Training: 87%|████████▋ | 15643/18018 [33:33:25<4:46:03, 7.23s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=17.2, I=19.0, T=37.5, wait=4.0s, pool=53, epoch=2/2][2026-10-19 23:03:48,731] loop.py:45 INFO TRAIN_STEP step=15644 loss=1.257083 Training: 87%|████████▋ | 15644/18018 [33:33:34<5:06:20, 7.74s/step, loss=1.257, acc=0.706, acc_len=1.70, thru=14.5, I=14.4, T=36.9, wait=5.5s, pool=49, epoch=2/2][2026-10-19 23:03:55,710] loop.py:380 INFO TIMING step=15645: step=3.370s data=1.333s compute=3.347s [fwd=2.415s bwd=0.917s opt=0.003s] dispatch=3.599s [2026-10-19 23:03:55,710] loop.py:45 INFO TRAIN_STEP step=15645 loss=1.277681 Training: 87%|████████▋ | 15645/18018 [33:33:41<4:57:10, 7.51s/step, loss=1.278, acc=0.706, acc_len=1.71, thru=16.8, I=17.9, T=38.0, wait=3.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:04,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:04:04,892] loop.py:45 INFO TRAIN_STEP step=15646 loss=1.236708 Training: 87%|████████▋ | 15646/18018 [33:33:50<5:16:47, 8.01s/step, loss=1.237, acc=0.710, acc_len=1.72, thru=16.1, I=15.7, T=35.7, wait=5.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:04:04,887] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15645: forward=2475.4ms backward=1050.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:05,189] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:10,314] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:04:11,373] loop.py:45 INFO TRAIN_STEP step=15647 loss=1.320192 Training: 87%|████████▋ | 15647/18018 [33:33:57<4:58:30, 7.55s/step, loss=1.320, acc=0.696, acc_len=1.67, thru=18.6, I=21.0, T=36.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:11,824] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:17,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:04:18,242] loop.py:45 INFO TRAIN_STEP step=15648 loss=1.309896 Training: 87%|████████▋ | 15648/18018 [33:34:04<4:50:15, 7.35s/step, loss=1.310, acc=0.699, acc_len=1.67, thru=18.5, I=22.7, T=38.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:18,662] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:04:26,797] loop.py:45 INFO TRAIN_STEP step=15649 loss=1.267946 Training: 87%|████████▋ | 15649/18018 [33:34:12<5:04:24, 7.71s/step, loss=1.268, acc=0.704, acc_len=1.69, thru=14.6, I=14.9, T=35.4, wait=4.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:34,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:04:35,500] loop.py:380 INFO TIMING step=15650: step=4.768s data=2.112s compute=4.726s [fwd=3.720s bwd=0.992s opt=0.003s] dispatch=3.927s [2026-10-19 23:04:35,500] loop.py:45 INFO TRAIN_STEP step=15650 loss=1.306265 Training: 87%|████████▋ | 15650/18018 [33:34:21<5:16:02, 8.01s/step, loss=1.306, acc=0.702, acc_len=1.68, thru=16.4, I=18.1, T=26.8, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:35,704] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:04:44,593] loop.py:45 INFO TRAIN_STEP step=15651 loss=1.221116 Training: 87%|████████▋ | 15651/18018 [33:34:30<5:28:45, 8.33s/step, loss=1.221, acc=0.715, acc_len=1.74, thru=11.3, I=12.3, T=35.0, wait=5.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:04:44,590] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15650: forward=2579.3ms backward=1033.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:51,437] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:04:51,902] loop.py:45 INFO TRAIN_STEP step=15652 loss=1.275457 Training: 87%|████████▋ | 15652/18018 [33:34:37<5:16:30, 8.03s/step, loss=1.275, acc=0.703, acc_len=1.69, thru=18.9, I=17.6, T=37.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:04:51,949] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 23:04:56,753] loop.py:45 INFO TRAIN_STEP step=15653 loss=1.250641 Training: 87%|████████▋ | 15653/18018 [33:34:42<4:38:50, 7.07s/step, loss=1.251, acc=0.708, acc_len=1.70, thru=22.9, I=23.2, T=39.6, wait=1.6s, pool=53, epoch=2/2][2026-10-19 23:05:03,830] loop.py:45 INFO TRAIN_STEP step=15654 loss=1.285230 Training: 87%|████████▋ | 15654/18018 [33:34:49<4:38:45, 7.07s/step, loss=1.285, acc=0.700, acc_len=1.68, thru=16.4, I=19.4, T=36.6, wait=3.6s, pool=37, epoch=2/2][2026-10-19 23:05:12,519] loop.py:380 INFO TIMING step=15655: step=3.552s data=1.440s compute=3.528s [fwd=2.529s bwd=0.985s opt=0.003s] dispatch=5.127s [2026-10-19 23:05:12,519] loop.py:45 INFO TRAIN_STEP step=15655 loss=1.335076 Training: 87%|████████▋ | 15655/18018 [33:34:58<4:57:42, 7.56s/step, loss=1.335, acc=0.694, acc_len=1.66, thru=15.6, I=13.2, T=36.0, wait=5.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:05:21,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 23:05:22,185] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15655: forward=2616.5ms backward=1079.9ms [2026-10-19 23:05:22,189] loop.py:45 INFO TRAIN_STEP step=15656 loss=1.319953 Training: 87%|████████▋ | 15656/18018 [33:35:08<5:22:29, 8.19s/step, loss=1.320, acc=0.693, acc_len=1.65, thru=15.3, I=14.9, T=33.9, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:05:22,446] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:05:28,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:05:29,938] loop.py:45 INFO TRAIN_STEP step=15657 loss=1.225337 Training: 87%|████████▋ | 15657/18018 [33:35:16<5:17:07, 8.06s/step, loss=1.225, acc=0.717, acc_len=1.74, thru=16.9, I=18.8, T=36.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:05:30,243] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:05:38,226] loop.py:45 INFO TRAIN_STEP step=15658 loss=1.223178 Training: 87%|████████▋ | 15658/18018 [33:35:24<5:19:42, 8.13s/step, loss=1.223, acc=0.713, acc_len=1.73, thru=13.9, I=16.2, T=36.9, wait=4.8s, pool=61, epoch=2/2][2026-10-19 23:05:45,448] loop.py:45 INFO TRAIN_STEP step=15659 loss=1.182783 Training: 87%|████████▋ | 15659/18018 [33:35:31<5:08:51, 7.86s/step, loss=1.183, acc=0.722, acc_len=1.76, thru=18.7, I=16.6, T=37.9, wait=3.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:05:51,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:05:53,469] loop.py:380 INFO TIMING step=15660: step=4.048s data=2.054s compute=3.998s [fwd=3.026s bwd=0.957s opt=0.003s] dispatch=3.965s [2026-10-19 23:05:53,469] loop.py:45 INFO TRAIN_STEP step=15660 loss=1.285511 Training: 87%|████████▋ | 15660/18018 [33:35:39<5:10:42, 7.91s/step, loss=1.286, acc=0.702, acc_len=1.68, thru=16.8, I=19.0, T=31.6, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:05:53,976] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 23:06:01,723] loop.py:45 INFO TRAIN_STEP step=15661 loss=1.276113 Training: 87%|████████▋ | 15661/18018 [33:35:47<5:14:42, 8.01s/step, loss=1.276, acc=0.707, acc_len=1.70, thru=12.0, I=15.6, T=34.7, wait=4.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:06:01,719] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15660: forward=2710.8ms backward=937.4ms [2026-10-19 23:06:10,793] loop.py:45 INFO TRAIN_STEP step=15662 loss=1.172082 Training: 87%|████████▋ | 15662/18018 [33:35:56<5:26:58, 8.33s/step, loss=1.172, acc=0.725, acc_len=1.77, thru=13.3, I=12.5, T=34.8, wait=5.4s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:06:18,522] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:06:19,369] loop.py:45 INFO TRAIN_STEP step=15663 loss=1.210376 Training: 87%|████████▋ | 15663/18018 [33:36:05<5:29:48, 8.40s/step, loss=1.210, acc=0.717, acc_len=1.74, thru=18.1, I=17.5, T=37.2, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:06:19,530] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:06:24,536] loop.py:45 INFO TRAIN_STEP step=15664 loss=1.145194 Training: 87%|████████▋ | 15664/18018 [33:36:10<4:51:34, 7.43s/step, loss=1.145, acc=0.722, acc_len=1.76, thru=18.8, I=21.9, T=39.5, wait=1.9s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:06:38,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:06:38,702] loop.py:380 INFO TIMING step=15665: step=4.010s data=1.612s compute=3.982s [fwd=2.802s bwd=1.166s opt=0.003s] dispatch=10.147s [2026-10-19 23:06:38,702] loop.py:45 INFO TRAIN_STEP step=15665 loss=1.252616 Training: 87%|████████▋ | 15665/18018 [33:36:24<6:10:39, 9.45s/step, loss=1.253, acc=0.711, acc_len=1.72, thru=10.6, I=10.4, T=31.9, wait=10.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:06:39,112] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:06:44,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:06:45,495] loop.py:45 INFO TRAIN_STEP step=15666 loss=1.217177 Training: 87%|████████▋ | 15666/18018 [33:36:31<5:39:14, 8.65s/step, loss=1.217, acc=0.715, acc_len=1.73, thru=18.2, I=19.4, T=37.5, wait=3.4s, pool=65, epoch=2/2] (TrainerActor pid=9310) [2026-10-19 23:06:45,492] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15665: forward=2367.6ms backward=988.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:06:45,986] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:06:51,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:06:52,294] loop.py:45 INFO TRAIN_STEP step=15667 loss=1.191867 Training: 87%|████████▋ | 15667/18018 [33:36:38<5:17:18, 8.10s/step, loss=1.192, acc=0.716, acc_len=1.73, thru=19.1, I=20.7, T=38.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:06:52,684] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:06:59,980] loop.py:45 INFO TRAIN_STEP step=15668 loss=1.249922 Training: 87%|████████▋ | 15668/18018 [33:36:46<5:12:22, 7.98s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=16.4, I=18.7, T=38.6, wait=4.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:10,180] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:07:10,919] loop.py:45 INFO TRAIN_STEP step=15669 loss=1.273568 Training: 87%|████████▋ | 15669/18018 [33:36:57<5:47:00, 8.86s/step, loss=1.274, acc=0.707, acc_len=1.70, thru=14.8, I=13.2, T=32.0, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:11,188] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:16,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:07:16,988] loop.py:380 INFO TIMING step=15670: step=3.878s data=1.940s compute=3.847s [fwd=2.953s bwd=0.879s opt=0.003s] dispatch=2.182s [2026-10-19 23:07:16,989] loop.py:45 INFO TRAIN_STEP step=15670 loss=1.161145 Training: 87%|████████▋ | 15670/18018 [33:37:03<5:14:03, 8.03s/step, loss=1.161, acc=0.726, acc_len=1.77, thru=19.8, I=21.2, T=33.0, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:17,299] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:07:23,936] loop.py:45 INFO TRAIN_STEP step=15671 loss=1.148894 Training: 87%|████████▋ | 15671/18018 [33:37:10<5:01:16, 7.70s/step, loss=1.149, acc=0.725, acc_len=1.77, thru=16.1, I=16.0, T=38.5, wait=3.6s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:07:23,932] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15670: forward=2383.1ms backward=902.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:30,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:07:31,467] loop.py:45 INFO TRAIN_STEP step=15672 loss=1.170289 Training: 87%|████████▋ | 15672/18018 [33:37:17<4:59:08, 7.65s/step, loss=1.170, acc=0.720, acc_len=1.75, thru=19.6, I=18.6, T=37.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:31,505] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:37,787] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:07:38,279] loop.py:45 INFO TRAIN_STEP step=15673 loss=1.208203 Training: 87%|████████▋ | 15673/18018 [33:37:24<4:49:09, 7.40s/step, loss=1.208, acc=0.714, acc_len=1.73, thru=18.9, I=21.4, T=38.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:38,306] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 23:07:44,873] loop.py:45 INFO TRAIN_STEP step=15674 loss=1.255696 Training: 87%|████████▋ | 15674/18018 [33:37:30<4:39:38, 7.16s/step, loss=1.256, acc=0.712, acc_len=1.72, thru=17.8, I=17.7, T=38.6, wait=3.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:50,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:07:51,195] loop.py:380 INFO TIMING step=15675: step=3.420s data=1.385s compute=3.378s [fwd=2.408s bwd=0.956s opt=0.003s] dispatch=2.891s [2026-10-19 23:07:51,196] loop.py:45 INFO TRAIN_STEP step=15675 loss=1.243900 Training: 87%|████████▋ | 15675/18018 [33:37:37<4:29:43, 6.91s/step, loss=1.244, acc=0.705, acc_len=1.70, thru=20.4, I=21.3, T=37.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:07:51,625] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:07:57,504] loop.py:45 INFO TRAIN_STEP step=15676 loss=1.153209 Training: 87%|████████▋ | 15676/18018 [33:37:43<4:22:33, 6.73s/step, loss=1.153, acc=0.725, acc_len=1.77, thru=18.9, I=21.8, T=38.4, wait=3.0s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:07:57,499] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15675: forward=2335.2ms backward=946.2ms [2026-10-19 23:08:06,645] loop.py:45 INFO TRAIN_STEP step=15677 loss=1.281976 Training: 87%|████████▋ | 15677/18018 [33:37:52<4:50:42, 7.45s/step, loss=1.282, acc=0.704, acc_len=1.69, thru=14.6, I=13.9, T=30.4, wait=4.9s, pool=49, epoch=2/2][2026-10-19 23:08:17,266] loop.py:45 INFO TRAIN_STEP step=15678 loss=1.355128 Training: 87%|████████▋ | 15678/18018 [33:38:03<5:27:44, 8.40s/step, loss=1.355, acc=0.691, acc_len=1.65, thru=12.9, I=12.7, T=35.8, wait=7.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:23,179] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:08:23,846] loop.py:45 INFO TRAIN_STEP step=15679 loss=1.224262 Training: 87%|████████▋ | 15679/18018 [33:38:09<5:06:14, 7.86s/step, loss=1.224, acc=0.715, acc_len=1.73, thru=20.1, I=17.9, T=39.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:24,185] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:30,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:08:30,649] loop.py:380 INFO TIMING step=15680: step=3.460s data=1.474s compute=3.409s [fwd=2.370s bwd=1.025s opt=0.003s] dispatch=3.333s [2026-10-19 23:08:30,649] loop.py:45 INFO TRAIN_STEP step=15680 loss=1.281153 Training: 87%|████████▋ | 15680/18018 [33:38:16<4:53:47, 7.54s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=19.5, I=22.7, T=37.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:31,057] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:37,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:08:37,293] loop.py:45 INFO TRAIN_STEP step=15681 loss=1.341772 Training: 87%|████████▋ | 15681/18018 [33:38:23<4:43:10, 7.27s/step, loss=1.342, acc=0.690, acc_len=1.65, thru=19.3, I=21.2, T=37.5, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:08:37,289] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15680: forward=2348.5ms backward=993.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:37,727] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:44,102] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:08:44,221] loop.py:45 INFO TRAIN_STEP step=15682 loss=1.324479 Training: 87%|████████▋ | 15682/18018 [33:38:30<4:39:04, 7.17s/step, loss=1.324, acc=0.696, acc_len=1.67, thru=18.4, I=19.6, T=37.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:44,604] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:51,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:08:51,738] loop.py:45 INFO TRAIN_STEP step=15683 loss=1.270289 Training: 87%|████████▋ | 15683/18018 [33:38:37<4:43:02, 7.27s/step, loss=1.270, acc=0.707, acc_len=1.70, thru=17.5, I=18.3, T=38.0, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:51,890] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:56,505] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:08:57,694] loop.py:45 INFO TRAIN_STEP step=15684 loss=1.302410 Training: 87%|████████▋ | 15684/18018 [33:38:43<4:27:33, 6.88s/step, loss=1.302, acc=0.700, acc_len=1.68, thru=21.0, I=24.1, T=38.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:08:58,016] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:09:05,854] loop.py:380 INFO TIMING step=15685: step=3.543s data=1.439s compute=3.519s [fwd=2.491s bwd=1.014s opt=0.003s] dispatch=4.607s [2026-10-19 23:09:05,854] loop.py:45 INFO TRAIN_STEP step=15685 loss=1.233676 Training: 87%|████████▋ | 15685/18018 [33:38:51<4:42:22, 7.26s/step, loss=1.234, acc=0.714, acc_len=1.73, thru=15.2, I=18.4, T=36.1, wait=4.6s, pool=57, epoch=2/2][2026-10-19 23:09:14,489] loop.py:45 INFO TRAIN_STEP step=15686 loss=1.252243 Training: 87%|████████▋ | 15686/18018 [33:39:00<4:58:16, 7.67s/step, loss=1.252, acc=0.709, acc_len=1.70, thru=13.5, I=14.1, T=36.0, wait=5.1s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:09:14,484] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15685: forward=2431.0ms backward=1059.0ms [2026-10-19 23:09:23,545] loop.py:45 INFO TRAIN_STEP step=15687 loss=1.464485 Training: 87%|████████▋ | 15687/18018 [33:39:09<5:14:13, 8.09s/step, loss=1.464, acc=0.672, acc_len=1.58, thru=16.8, I=16.6, T=35.0, wait=5.4s, pool=57, epoch=2/2][2026-10-19 23:09:30,735] loop.py:45 INFO TRAIN_STEP step=15688 loss=1.322928 Training: 87%|████████▋ | 15688/18018 [33:39:16<5:03:38, 7.82s/step, loss=1.323, acc=0.696, acc_len=1.66, thru=16.0, I=17.9, T=37.0, wait=3.7s, pool=45, epoch=2/2][2026-10-19 23:09:39,646] loop.py:45 INFO TRAIN_STEP step=15689 loss=1.306906 Training: 87%|████████▋ | 15689/18018 [33:39:25<5:16:15, 8.15s/step, loss=1.307, acc=0.697, acc_len=1.67, thru=16.0, I=14.3, T=34.7, wait=5.2s, pool=61, epoch=2/2][2026-10-19 23:09:47,047] loop.py:380 INFO TIMING step=15690: step=3.480s data=1.443s compute=3.433s [fwd=2.420s bwd=0.998s opt=0.003s] dispatch=3.910s [2026-10-19 23:09:47,047] loop.py:45 INFO TRAIN_STEP step=15690 loss=1.328648 Training: 87%|████████▋ | 15690/18018 [33:39:33<5:07:25, 7.92s/step, loss=1.329, acc=0.696, acc_len=1.67, thru=16.9, I=18.1, T=36.8, wait=3.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:09:52,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:09:53,295] loop.py:45 INFO TRAIN_STEP step=15691 loss=1.313149 Training: 87%|████████▋ | 15691/18018 [33:39:39<4:47:46, 7.42s/step, loss=1.313, acc=0.697, acc_len=1.67, thru=20.1, I=22.1, T=38.7, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:09:53,290] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15690: forward=2355.6ms backward=909.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:09:53,515] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:09:59,343] loop.py:45 INFO TRAIN_STEP step=15692 loss=1.288809 Training: 87%|████████▋ | 15692/18018 [33:39:45<4:31:41, 7.01s/step, loss=1.289, acc=0.703, acc_len=1.69, thru=20.5, I=23.5, T=38.3, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:10:03,872] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:10:05,084] loop.py:45 INFO TRAIN_STEP step=15693 loss=1.256100 Training: 87%|████████▋ | 15693/18018 [33:39:51<4:16:50, 6.63s/step, loss=1.256, acc=0.706, acc_len=1.70, thru=21.7, I=24.2, T=38.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:10:05,379] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:10:12,759] loop.py:45 INFO TRAIN_STEP step=15694 loss=1.224495 Training: 87%|████████▋ | 15694/18018 [33:39:58<4:28:54, 6.94s/step, loss=1.224, acc=0.711, acc_len=1.72, thru=14.5, I=18.2, T=38.4, wait=4.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:10:18,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:10:19,208] loop.py:380 INFO TIMING step=15695: step=3.453s data=1.436s compute=3.408s [fwd=2.408s bwd=0.985s opt=0.003s] dispatch=2.988s [2026-10-19 23:10:19,208] loop.py:45 INFO TRAIN_STEP step=15695 loss=1.260642 Training: 87%|████████▋ | 15695/18018 [33:40:05<4:23:02, 6.79s/step, loss=1.261, acc=0.707, acc_len=1.70, thru=20.0, I=21.5, T=37.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:10:19,660] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:10:26,378] loop.py:45 INFO TRAIN_STEP step=15696 loss=1.134730 Training: 87%|████████▋ | 15696/18018 [33:40:12<4:27:19, 6.91s/step, loss=1.135, acc=0.729, acc_len=1.79, thru=15.2, I=17.6, T=37.8, wait=3.8s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:10:26,374] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15695: forward=2392.2ms backward=943.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:10:33,294] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:10:34,108] loop.py:45 INFO TRAIN_STEP step=15697 loss=1.250071 Training: 87%|████████▋ | 15697/18018 [33:40:20<4:36:44, 7.15s/step, loss=1.250, acc=0.709, acc_len=1.71, thru=18.1, I=18.3, T=36.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:10:34,299] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:10:39,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:10:40,971] loop.py:45 INFO TRAIN_STEP step=15698 loss=1.261658 Training: 87%|████████▋ | 15698/18018 [33:40:27<4:33:15, 7.07s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=18.5, I=23.1, T=37.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:10:41,116] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:10:48,364] loop.py:45 INFO TRAIN_STEP step=15699 loss=1.342103 Training: 87%|████████▋ | 15699/18018 [33:40:34<4:36:54, 7.16s/step, loss=1.342, acc=0.692, acc_len=1.65, thru=15.7, I=19.5, T=35.4, wait=3.8s, pool=57, epoch=2/2][2026-10-19 23:10:57,240] loop.py:380 INFO TIMING step=15700: step=3.536s data=1.440s compute=3.506s [fwd=2.481s bwd=1.011s opt=0.003s] dispatch=5.331s [2026-10-19 23:10:57,240] loop.py:45 INFO TRAIN_STEP step=15700 loss=1.276739 Training: 87%|████████▋ | 15700/18018 [33:40:43<4:56:38, 7.68s/step, loss=1.277, acc=0.703, acc_len=1.69, thru=13.0, I=13.4, T=36.2, wait=5.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:11:05,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:11:06,001] loop.py:45 INFO TRAIN_STEP step=15701 loss=1.365189 Training: 87%|████████▋ | 15701/18018 [33:40:52<5:09:01, 8.00s/step, loss=1.365, acc=0.686, acc_len=1.62, thru=17.1, I=17.7, T=35.4, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:11:05,996] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15700: forward=2485.7ms backward=1074.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:11:06,270] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:11:13,166] loop.py:45 INFO TRAIN_STEP step=15702 loss=1.220835 Training: 87%|████████▋ | 15702/18018 [33:40:59<4:59:13, 7.75s/step, loss=1.221, acc=0.713, acc_len=1.72, thru=14.1, I=17.1, T=37.3, wait=3.7s, pool=33, epoch=2/2][2026-10-19 23:11:21,529] loop.py:45 INFO TRAIN_STEP step=15703 loss=1.242591 Training: 87%|████████▋ | 15703/18018 [33:41:07<5:06:08, 7.93s/step, loss=1.243, acc=0.714, acc_len=1.74, thru=15.0, I=14.2, T=36.2, wait=4.8s, pool=33, epoch=2/2][2026-10-19 23:11:29,151] loop.py:45 INFO TRAIN_STEP step=15704 loss=1.274619 Training: 87%|████████▋ | 15704/18018 [33:41:15<5:02:25, 7.84s/step, loss=1.275, acc=0.700, acc_len=1.68, thru=20.3, I=17.5, T=37.3, wait=4.2s, pool=53, epoch=2/2][2026-10-19 23:11:38,022] loop.py:380 INFO TIMING step=15705: step=3.673s data=1.490s compute=3.644s [fwd=2.579s bwd=1.051s opt=0.003s] dispatch=5.189s [2026-10-19 23:11:38,023] loop.py:45 INFO TRAIN_STEP step=15705 loss=1.398098 Training: 87%|████████▋ | 15705/18018 [33:41:24<5:14:10, 8.15s/step, loss=1.398, acc=0.682, acc_len=1.62, thru=14.6, I=14.5, T=34.9, wait=5.2s, pool=61, epoch=2/2][2026-10-19 23:11:44,695] loop.py:45 INFO TRAIN_STEP step=15706 loss=1.363042 Training: 87%|████████▋ | 15706/18018 [33:41:30<4:56:59, 7.71s/step, loss=1.363, acc=0.684, acc_len=1.63, thru=17.2, I=19.4, T=35.9, wait=3.1s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:11:44,691] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15705: forward=2558.2ms backward=956.1ms [2026-10-19 23:11:52,815] loop.py:45 INFO TRAIN_STEP step=15707 loss=1.314581 Training: 87%|████████▋ | 15707/18018 [33:41:38<5:01:36, 7.83s/step, loss=1.315, acc=0.697, acc_len=1.67, thru=16.7, I=16.3, T=37.2, wait=4.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:11:58,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:11:59,094] loop.py:45 INFO TRAIN_STEP step=15708 loss=1.244289 Training: 87%|████████▋ | 15708/18018 [33:41:45<4:43:34, 7.37s/step, loss=1.244, acc=0.711, acc_len=1.72, thru=20.5, I=20.7, T=37.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:11:59,252] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:06,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:12:06,474] loop.py:45 INFO TRAIN_STEP step=15709 loss=1.291374 Training: 87%|████████▋ | 15709/18018 [33:41:52<4:43:36, 7.37s/step, loss=1.291, acc=0.701, acc_len=1.69, thru=17.5, I=18.6, T=38.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:06,550] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:10,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:12:12,043] loop.py:380 INFO TIMING step=15710: step=3.331s data=1.291s compute=3.300s [fwd=2.389s bwd=0.898s opt=0.003s] dispatch=2.228s [2026-10-19 23:12:12,043] loop.py:45 INFO TRAIN_STEP step=15710 loss=1.219865 Training: 87%|████████▋ | 15710/18018 [33:41:58<4:22:42, 6.83s/step, loss=1.220, acc=0.711, acc_len=1.72, thru=20.9, I=24.9, T=38.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:12,450] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:17,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:12:17,528] loop.py:45 INFO TRAIN_STEP step=15711 loss=1.335661 Training: 87%|████████▋ | 15711/18018 [33:42:03<4:07:04, 6.43s/step, loss=1.336, acc=0.695, acc_len=1.66, thru=23.5, I=28.0, T=40.1, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:12:17,524] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15710: forward=2257.7ms backward=885.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:17,708] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:23,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:12:23,899] loop.py:45 INFO TRAIN_STEP step=15712 loss=1.290222 Training: 87%|████████▋ | 15712/18018 [33:42:09<4:06:20, 6.41s/step, loss=1.290, acc=0.700, acc_len=1.68, thru=20.7, I=22.8, T=38.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:24,059] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:12:30,307] loop.py:45 INFO TRAIN_STEP step=15713 loss=1.210557 Training: 87%|████████▋ | 15713/18018 [33:42:16<4:06:11, 6.41s/step, loss=1.211, acc=0.709, acc_len=1.71, thru=17.6, I=21.5, T=38.6, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:40,887] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:12:42,150] loop.py:45 INFO TRAIN_STEP step=15714 loss=1.265016 Training: 87%|████████▋ | 15714/18018 [33:42:28<5:08:43, 8.04s/step, loss=1.265, acc=0.708, acc_len=1.71, thru=14.1, I=13.6, T=33.2, wait=8.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:42,394] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:12:48,026] loop.py:380 INFO TIMING step=15715: step=3.200s data=1.280s compute=3.172s [fwd=2.282s bwd=0.875s opt=0.003s] dispatch=2.665s [2026-10-19 23:12:48,026] loop.py:45 INFO TRAIN_STEP step=15715 loss=1.168718 Training: 87%|████████▋ | 15715/18018 [33:42:34<4:43:39, 7.39s/step, loss=1.169, acc=0.717, acc_len=1.74, thru=18.0, I=18.5, T=40.0, wait=2.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:54,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:12:56,464] loop.py:45 INFO TRAIN_STEP step=15716 loss=1.218372 Training: 87%|████████▋ | 15716/18018 [33:42:42<4:55:35, 7.70s/step, loss=1.218, acc=0.710, acc_len=1.72, thru=17.1, I=19.3, T=36.1, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:12:56,459] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15715: forward=2453.3ms backward=1002.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:12:56,921] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:03,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:13:03,161] loop.py:45 INFO TRAIN_STEP step=15717 loss=1.176004 Training: 87%|████████▋ | 15717/18018 [33:42:49<4:43:53, 7.40s/step, loss=1.176, acc=0.718, acc_len=1.74, thru=18.1, I=23.3, T=38.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:03,533] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:10,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:13:11,007] loop.py:45 INFO TRAIN_STEP step=15718 loss=1.201669 Training: 87%|████████▋ | 15718/18018 [33:42:57<4:48:51, 7.54s/step, loss=1.202, acc=0.716, acc_len=1.73, thru=17.3, I=20.2, T=36.8, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:11,318] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:13:16,829] loop.py:45 INFO TRAIN_STEP step=15719 loss=1.256062 Training: 87%|████████▋ | 15719/18018 [33:43:02<4:29:01, 7.02s/step, loss=1.256, acc=0.710, acc_len=1.71, thru=20.2, I=23.5, T=38.1, wait=2.5s, pool=57, epoch=2/2][2026-10-19 23:13:26,100] loop.py:380 INFO TIMING step=15720: step=3.677s data=1.506s compute=3.652s [fwd=2.592s bwd=1.045s opt=0.003s] dispatch=5.586s [2026-10-19 23:13:26,101] loop.py:45 INFO TRAIN_STEP step=15720 loss=1.287511 Training: 87%|████████▋ | 15720/18018 [33:43:12<4:54:48, 7.70s/step, loss=1.288, acc=0.699, acc_len=1.67, thru=12.6, I=13.7, T=34.8, wait=5.6s, pool=45, epoch=2/2][2026-10-19 23:13:33,765] loop.py:45 INFO TRAIN_STEP step=15721 loss=1.124901 Training: 87%|████████▋ | 15721/18018 [33:43:19<4:54:16, 7.69s/step, loss=1.125, acc=0.731, acc_len=1.79, thru=17.3, I=16.0, T=38.2, wait=4.3s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:13:33,760] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15720: forward=2372.5ms backward=933.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:39,305] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:13:40,056] loop.py:45 INFO TRAIN_STEP step=15722 loss=1.252483 Training: 87%|████████▋ | 15722/18018 [33:43:26<4:38:06, 7.27s/step, loss=1.252, acc=0.708, acc_len=1.72, thru=20.9, I=22.1, T=37.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:40,309] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:13:45,817] loop.py:45 INFO TRAIN_STEP step=15723 loss=1.240371 Training: 87%|████████▋ | 15723/18018 [33:43:31<4:20:41, 6.82s/step, loss=1.240, acc=0.707, acc_len=1.71, thru=22.1, I=24.1, T=39.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:45,815] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:46,317] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:50,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:13:51,611] loop.py:45 INFO TRAIN_STEP step=15724 loss=1.280914 Training: 87%|████████▋ | 15724/18018 [33:43:37<4:08:51, 6.51s/step, loss=1.281, acc=0.704, acc_len=1.69, thru=22.7, I=25.1, T=39.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:51,983] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:57,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:13:58,082] loop.py:380 INFO TIMING step=15725: step=3.336s data=1.353s compute=3.300s [fwd=2.362s bwd=0.923s opt=0.003s] dispatch=3.127s [2026-10-19 23:13:58,082] loop.py:45 INFO TRAIN_STEP step=15725 loss=1.234609 Training: 87%|████████▋ | 15725/18018 [33:43:44<4:08:19, 6.50s/step, loss=1.235, acc=0.716, acc_len=1.73, thru=20.0, I=23.7, T=38.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:13:58,400] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:14:04,404] loop.py:45 INFO TRAIN_STEP step=15726 loss=1.250114 Training: 87%|████████▋ | 15726/18018 [33:43:50<4:06:11, 6.44s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=18.9, I=20.2, T=38.3, wait=3.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:14:04,399] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15725: forward=2362.9ms backward=933.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:13,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:14:14,823] loop.py:45 INFO TRAIN_STEP step=15727 loss=1.171594 Training: 87%|████████▋ | 15727/18018 [33:44:00<4:51:36, 7.64s/step, loss=1.172, acc=0.724, acc_len=1.77, thru=14.2, I=14.9, T=36.1, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:15,321] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 23:14:21,326] loop.py:45 INFO TRAIN_STEP step=15728 loss=1.110903 Training: 87%|████████▋ | 15728/18018 [33:44:07<4:38:31, 7.30s/step, loss=1.111, acc=0.731, acc_len=1.79, thru=18.4, I=20.5, T=39.0, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:27,971] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:14:29,265] loop.py:45 INFO TRAIN_STEP step=15729 loss=1.264158 Training: 87%|████████▋ | 15729/18018 [33:44:15<4:45:43, 7.49s/step, loss=1.264, acc=0.704, acc_len=1.70, thru=17.2, I=18.8, T=36.9, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:29,479] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:14:35,326] loop.py:380 INFO TIMING step=15730: step=3.265s data=1.307s compute=3.236s [fwd=2.338s bwd=0.883s opt=0.003s] dispatch=2.787s [2026-10-19 23:14:35,326] loop.py:45 INFO TRAIN_STEP step=15730 loss=1.265068 Training: 87%|████████▋ | 15730/18018 [33:44:21<4:29:16, 7.06s/step, loss=1.265, acc=0.705, acc_len=1.71, thru=17.1, I=20.9, T=39.2, wait=2.8s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:40,462] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:14:41,736] loop.py:45 INFO TRAIN_STEP step=15731 loss=1.312889 Training: 87%|████████▋ | 15731/18018 [33:44:27<4:21:42, 6.87s/step, loss=1.313, acc=0.701, acc_len=1.68, thru=20.0, I=22.5, T=37.4, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:14:41,732] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15730: forward=2440.7ms backward=928.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:41,973] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:49,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:14:50,517] loop.py:45 INFO TRAIN_STEP step=15732 loss=1.233945 Training: 87%|████████▋ | 15732/18018 [33:44:36<4:43:27, 7.44s/step, loss=1.234, acc=0.709, acc_len=1.72, thru=15.0, I=15.3, T=36.4, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:50,996] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:57,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:14:57,788] loop.py:45 INFO TRAIN_STEP step=15733 loss=1.285017 Training: 87%|████████▋ | 15733/18018 [33:44:43<4:41:26, 7.39s/step, loss=1.285, acc=0.704, acc_len=1.69, thru=18.1, I=20.5, T=38.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:14:57,914] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 23:15:03,436] loop.py:45 INFO TRAIN_STEP step=15734 loss=1.200322 Training: 87%|████████▋ | 15734/18018 [33:44:49<4:21:27, 6.87s/step, loss=1.200, acc=0.718, acc_len=1.74, thru=19.1, I=20.7, T=37.8, wait=2.3s, pool=45, epoch=2/2][2026-10-19 23:15:12,199] loop.py:380 INFO TIMING step=15735: step=3.720s data=1.528s compute=3.684s [fwd=2.658s bwd=1.011s opt=0.003s] dispatch=5.029s [2026-10-19 23:15:12,199] loop.py:45 INFO TRAIN_STEP step=15735 loss=1.248734 Training: 87%|████████▋ | 15735/18018 [33:44:58<4:42:53, 7.43s/step, loss=1.249, acc=0.712, acc_len=1.72, thru=13.9, I=14.0, T=34.4, wait=5.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:15:20,018] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:15:21,127] loop.py:45 INFO TRAIN_STEP step=15736 loss=1.172860 Training: 87%|████████▋ | 15736/18018 [33:45:07<4:59:49, 7.88s/step, loss=1.173, acc=0.724, acc_len=1.77, thru=17.6, I=16.3, T=36.3, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:15:21,122] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15735: forward=2499.4ms backward=982.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:15:21,527] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:15:27,234] loop.py:45 INFO TRAIN_STEP step=15737 loss=1.216876 Training: 87%|████████▋ | 15737/18018 [33:45:13<4:39:27, 7.35s/step, loss=1.217, acc=0.715, acc_len=1.73, thru=18.2, I=19.6, T=39.2, wait=2.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:15:34,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:15:35,392] loop.py:45 INFO TRAIN_STEP step=15738 loss=1.362408 Training: 87%|████████▋ | 15738/18018 [33:45:21<4:48:32, 7.59s/step, loss=1.362, acc=0.696, acc_len=1.66, thru=17.4, I=19.4, T=36.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:15:35,603] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:15:40,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:15:41,440] loop.py:45 INFO TRAIN_STEP step=15739 loss=1.211393 Training: 87%|████████▋ | 15739/18018 [33:45:27<4:30:48, 7.13s/step, loss=1.211, acc=0.714, acc_len=1.73, thru=21.5, I=26.9, T=38.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:15:41,731] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:15:47,057] loop.py:380 INFO TIMING step=15740: step=3.184s data=1.247s compute=3.154s [fwd=2.278s bwd=0.862s opt=0.003s] dispatch=2.423s [2026-10-19 23:15:47,057] loop.py:45 INFO TRAIN_STEP step=15740 loss=1.236707 Training: 87%|████████▋ | 15740/18018 [33:45:33<4:13:26, 6.68s/step, loss=1.237, acc=0.708, acc_len=1.71, thru=22.1, I=27.1, T=40.2, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:15:54,990] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:15:55,930] loop.py:45 INFO TRAIN_STEP step=15741 loss=1.266932 Training: 87%|████████▋ | 15741/18018 [33:45:42<4:38:21, 7.33s/step, loss=1.267, acc=0.705, acc_len=1.70, thru=15.3, I=16.9, T=36.0, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:15:55,925] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15740: forward=2504.7ms backward=1000.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:15:56,016] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 23:16:02,100] loop.py:45 INFO TRAIN_STEP step=15742 loss=1.214005 Training: 87%|████████▋ | 15742/18018 [33:45:48<4:24:57, 6.98s/step, loss=1.214, acc=0.714, acc_len=1.73, thru=18.8, I=18.4, T=38.1, wait=2.8s, pool=53, epoch=2/2][2026-10-19 23:16:08,975] loop.py:45 INFO TRAIN_STEP step=15743 loss=1.141642 Training: 87%|████████▋ | 15743/18018 [33:45:55<4:23:35, 6.95s/step, loss=1.142, acc=0.729, acc_len=1.78, thru=18.0, I=19.9, T=36.1, wait=3.3s, pool=57, epoch=2/2][2026-10-19 23:16:15,855] loop.py:45 INFO TRAIN_STEP step=15744 loss=1.222842 Training: 87%|████████▋ | 15744/18018 [33:46:01<4:22:39, 6.93s/step, loss=1.223, acc=0.711, acc_len=1.72, thru=16.4, I=16.3, T=36.4, wait=3.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:16:22,794] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:16:24,271] loop.py:380 INFO TIMING step=15745: step=3.729s data=1.575s compute=3.690s [fwd=2.635s bwd=1.041s opt=0.003s] dispatch=4.679s [2026-10-19 23:16:24,271] loop.py:45 INFO TRAIN_STEP step=15745 loss=1.225782 Training: 87%|████████▋ | 15745/18018 [33:46:10<4:39:27, 7.38s/step, loss=1.226, acc=0.714, acc_len=1.73, thru=18.4, I=18.3, T=34.3, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:16:24,304] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 23:16:30,721] loop.py:45 INFO TRAIN_STEP step=15746 loss=1.155026 Training: 87%|████████▋ | 15746/18018 [33:46:16<4:28:45, 7.10s/step, loss=1.155, acc=0.725, acc_len=1.77, thru=18.2, I=20.9, T=39.0, wait=3.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:16:30,716] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15745: forward=2350.9ms backward=883.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:16:37,741] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:16:37,833] loop.py:45 INFO TRAIN_STEP step=15747 loss=1.262136 Training: 87%|████████▋ | 15747/18018 [33:46:23<4:28:48, 7.10s/step, loss=1.262, acc=0.704, acc_len=1.70, thru=18.5, I=19.0, T=37.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:16:38,244] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:16:45,931] loop.py:45 INFO TRAIN_STEP step=15748 loss=1.234675 Training: 87%|████████▋ | 15748/18018 [33:46:32<4:40:01, 7.40s/step, loss=1.235, acc=0.708, acc_len=1.71, thru=14.6, I=15.6, T=37.1, wait=4.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:16:53,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:16:54,474] loop.py:45 INFO TRAIN_STEP step=15749 loss=1.193824 Training: 87%|████████▋ | 15749/18018 [33:46:40<4:52:51, 7.74s/step, loss=1.194, acc=0.724, acc_len=1.76, thru=16.3, I=17.1, T=37.4, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:16:54,580] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 23:17:00,406] loop.py:380 INFO TIMING step=15750: step=3.220s data=1.265s compute=3.190s [fwd=2.300s bwd=0.876s opt=0.003s] dispatch=2.702s [2026-10-19 23:17:00,406] loop.py:45 INFO TRAIN_STEP step=15750 loss=1.124968 Training: 87%|████████▋ | 15750/18018 [33:46:46<4:32:08, 7.20s/step, loss=1.125, acc=0.732, acc_len=1.79, thru=18.8, I=21.3, T=39.8, wait=2.7s, pool=41, epoch=2/2][2026-10-19 23:17:00,408] loop.py:419 INFO Saving checkpoint at step 15750... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:01,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 (TrainerActor pid=9310) [2026-10-19 23:17:03,838] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0015751 (TrainerActor pid=9310) [2026-10-19 23:17:03,972] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0015001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:04,994] inference_manager.py:408 INFO Pool capacity available after 3.5s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:12,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 23:17:03,839] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0015751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 23:17:04,733] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0015001 [2026-10-19 23:17:12,985] loop.py:45 INFO TRAIN_STEP step=15751 loss=1.317831 Training: 87%|████████▋ | 15751/18018 [33:46:59<5:33:00, 8.81s/step, loss=1.318, acc=0.695, acc_len=1.66, thru=15.5, I=16.1, T=34.6, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:17:12,981] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15750: forward=2538.8ms backward=1105.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:13,310] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:20,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:17:20,131] loop.py:45 INFO TRAIN_STEP step=15752 loss=1.144205 Training: 87%|████████▋ | 15752/18018 [33:47:06<5:13:57, 8.31s/step, loss=1.144, acc=0.727, acc_len=1.77, thru=18.5, I=17.9, T=37.1, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:20,569] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:25,364] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:17:25,913] loop.py:45 INFO TRAIN_STEP step=15753 loss=1.192694 Training: 87%|████████▋ | 15753/18018 [33:47:11<4:45:10, 7.55s/step, loss=1.193, acc=0.721, acc_len=1.75, thru=21.3, I=22.0, T=39.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:26,370] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:17:31,898] loop.py:45 INFO TRAIN_STEP step=15754 loss=1.166854 Training: 87%|████████▋ | 15754/18018 [33:47:17<4:27:16, 7.08s/step, loss=1.167, acc=0.725, acc_len=1.77, thru=19.0, I=21.7, T=39.4, wait=2.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:38,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:17:39,370] loop.py:380 INFO TIMING step=15755: step=3.452s data=1.425s compute=3.398s [fwd=2.383s bwd=1.000s opt=0.003s] dispatch=4.013s [2026-10-19 23:17:39,371] loop.py:45 INFO TRAIN_STEP step=15755 loss=1.181870 Training: 87%|████████▋ | 15755/18018 [33:47:25<4:31:34, 7.20s/step, loss=1.182, acc=0.724, acc_len=1.77, thru=19.1, I=19.4, T=37.1, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:39,673] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:17:45,940] loop.py:45 INFO TRAIN_STEP step=15756 loss=1.245089 Training: 87%|████████▋ | 15756/18018 [33:47:32<4:24:18, 7.01s/step, loss=1.245, acc=0.712, acc_len=1.72, thru=17.6, I=21.6, T=38.2, wait=3.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:17:45,936] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15755: forward=2376.2ms backward=916.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:53,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:17:54,208] loop.py:45 INFO TRAIN_STEP step=15757 loss=1.283458 Training: 87%|████████▋ | 15757/18018 [33:47:40<4:38:25, 7.39s/step, loss=1.283, acc=0.705, acc_len=1.70, thru=17.4, I=17.7, T=37.0, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:54,610] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:17:59,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:18:00,184] loop.py:45 INFO TRAIN_STEP step=15758 loss=1.220644 Training: 87%|████████▋ | 15758/18018 [33:47:46<4:22:19, 6.96s/step, loss=1.221, acc=0.715, acc_len=1.74, thru=21.1, I=24.3, T=39.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:00,439] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:05,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:18:05,956] loop.py:45 INFO TRAIN_STEP step=15759 loss=1.226860 Training: 87%|████████▋ | 15759/18018 [33:47:52<4:08:43, 6.61s/step, loss=1.227, acc=0.713, acc_len=1.72, thru=22.9, I=23.7, T=35.6, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:06,121] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:13,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:18:13,870] loop.py:380 INFO TIMING step=15760: step=3.608s data=1.491s compute=3.555s [fwd=2.487s bwd=1.053s opt=0.003s] dispatch=4.296s [2026-10-19 23:18:13,870] loop.py:45 INFO TRAIN_STEP step=15760 loss=1.263371 Training: 87%|████████▋ | 15760/18018 [33:47:59<4:23:24, 7.00s/step, loss=1.263, acc=0.708, acc_len=1.70, thru=17.2, I=17.9, T=35.5, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:14,127] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:18,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:18:19,472] loop.py:45 INFO TRAIN_STEP step=15761 loss=1.228474 Training: 87%|████████▋ | 15761/18018 [33:48:05<4:07:31, 6.58s/step, loss=1.228, acc=0.712, acc_len=1.72, thru=21.4, I=24.6, T=39.0, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:18:19,468] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15760: forward=2354.3ms backward=878.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:19,692] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:18:26,013] loop.py:45 INFO TRAIN_STEP step=15762 loss=1.283641 Training: 87%|████████▋ | 15762/18018 [33:48:12<4:06:58, 6.57s/step, loss=1.284, acc=0.705, acc_len=1.69, thru=16.6, I=22.5, T=38.1, wait=3.2s, pool=33, epoch=2/2][2026-10-19 23:18:37,051] loop.py:45 INFO TRAIN_STEP step=15763 loss=1.249001 Training: 87%|████████▋ | 15763/18018 [33:48:23<4:57:13, 7.91s/step, loss=1.249, acc=0.714, acc_len=1.73, thru=14.0, I=12.5, T=35.0, wait=7.4s, pool=57, epoch=2/2][2026-10-19 23:18:43,811] loop.py:45 INFO TRAIN_STEP step=15764 loss=1.210576 Training: 87%|████████▋ | 15764/18018 [33:48:29<4:44:09, 7.56s/step, loss=1.211, acc=0.716, acc_len=1.74, thru=19.4, I=17.6, T=37.0, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:50,203] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:18:51,026] loop.py:380 INFO TIMING step=15765: step=3.498s data=1.390s compute=3.468s [fwd=2.487s bwd=0.967s opt=0.003s] dispatch=3.709s [2026-10-19 23:18:51,027] loop.py:45 INFO TRAIN_STEP step=15765 loss=1.293180 Training: 87%|████████▋ | 15765/18018 [33:48:37<4:40:07, 7.46s/step, loss=1.293, acc=0.699, acc_len=1.67, thru=17.2, I=19.1, T=36.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:51,218] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:57,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:18:58,353] loop.py:45 INFO TRAIN_STEP step=15766 loss=1.184921 Training: 88%|████████▊ | 15766/18018 [33:48:44<4:38:29, 7.42s/step, loss=1.185, acc=0.719, acc_len=1.74, thru=17.2, I=20.2, T=38.1, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:18:58,349] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15765: forward=2363.7ms backward=949.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:18:58,659] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:19:04,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:19:05,263] loop.py:45 INFO TRAIN_STEP step=15767 loss=1.286732 Training: 88%|████████▊ | 15767/18018 [33:48:51<4:32:37, 7.27s/step, loss=1.287, acc=0.705, acc_len=1.69, thru=18.9, I=20.5, T=38.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:19:05,433] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 23:19:11,362] loop.py:45 INFO TRAIN_STEP step=15768 loss=1.286698 Training: 88%|████████▊ | 15768/18018 [33:48:57<4:19:21, 6.92s/step, loss=1.287, acc=0.703, acc_len=1.68, thru=18.6, I=19.8, T=39.6, wait=2.9s, pool=41, epoch=2/2][2026-10-19 23:19:17,857] loop.py:45 INFO TRAIN_STEP step=15769 loss=1.221625 Training: 88%|████████▊ | 15769/18018 [33:49:03<4:14:31, 6.79s/step, loss=1.222, acc=0.714, acc_len=1.73, thru=18.4, I=17.5, T=38.4, wait=3.2s, pool=49, epoch=2/2][2026-10-19 23:19:25,931] loop.py:380 INFO TIMING step=15770: step=3.468s data=1.378s compute=3.443s [fwd=2.488s bwd=0.940s opt=0.003s] dispatch=4.597s [2026-10-19 23:19:25,931] loop.py:45 INFO TRAIN_STEP step=15770 loss=1.317171 Training: 88%|████████▊ | 15770/18018 [33:49:12<4:28:51, 7.18s/step, loss=1.317, acc=0.698, acc_len=1.67, thru=17.1, I=18.6, T=36.9, wait=4.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:19:33,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:19:35,390] loop.py:45 INFO TRAIN_STEP step=15771 loss=1.305148 Training: 88%|████████▊ | 15771/18018 [33:49:21<4:54:22, 7.86s/step, loss=1.305, acc=0.694, acc_len=1.66, thru=15.3, I=15.1, T=30.4, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:19:35,386] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15770: forward=3194.0ms backward=963.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:19:35,433] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 23:19:41,410] loop.py:45 INFO TRAIN_STEP step=15772 loss=1.216301 Training: 88%|████████▊ | 15772/18018 [33:49:27<4:33:35, 7.31s/step, loss=1.216, acc=0.715, acc_len=1.73, thru=17.0, I=25.9, T=31.9, wait=2.0s, pool=45, epoch=2/2][2026-10-19 23:19:48,995] loop.py:45 INFO TRAIN_STEP step=15773 loss=1.216755 Training: 88%|████████▊ | 15773/18018 [33:49:35<4:36:33, 7.39s/step, loss=1.217, acc=0.715, acc_len=1.74, thru=15.3, I=15.4, T=36.3, wait=4.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:19:56,612] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:19:58,096] loop.py:45 INFO TRAIN_STEP step=15774 loss=1.262147 Training: 88%|████████▊ | 15774/18018 [33:49:44<4:55:36, 7.90s/step, loss=1.262, acc=0.707, acc_len=1.71, thru=15.9, I=17.6, T=28.8, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:19:58,124] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:05,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:20:06,031] loop.py:380 INFO TIMING step=15775: step=3.539s data=1.288s compute=3.512s [fwd=2.578s bwd=0.921s opt=0.003s] dispatch=4.388s [2026-10-19 23:20:06,031] loop.py:45 INFO TRAIN_STEP step=15775 loss=1.281328 Training: 88%|████████▊ | 15775/18018 [33:49:52<4:55:48, 7.91s/step, loss=1.281, acc=0.700, acc_len=1.68, thru=14.9, I=16.7, T=36.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:06,283] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:11,043] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:20:11,900] loop.py:45 INFO TRAIN_STEP step=15776 loss=1.281853 Training: 88%|████████▊ | 15776/18018 [33:49:57<4:32:47, 7.30s/step, loss=1.282, acc=0.701, acc_len=1.68, thru=20.7, I=21.7, T=37.1, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:20:11,895] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15775: forward=2447.9ms backward=964.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:12,066] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:20:19,796] loop.py:45 INFO TRAIN_STEP step=15777 loss=1.252897 Training: 88%|████████▊ | 15777/18018 [33:50:05<4:39:19, 7.48s/step, loss=1.253, acc=0.707, acc_len=1.70, thru=16.6, I=17.3, T=38.5, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:26,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:20:26,160] loop.py:45 INFO TRAIN_STEP step=15778 loss=1.213696 Training: 88%|████████▊ | 15778/18018 [33:50:12<4:26:42, 7.14s/step, loss=1.214, acc=0.713, acc_len=1.72, thru=20.1, I=17.9, T=37.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:26,560] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:20:34,296] loop.py:45 INFO TRAIN_STEP step=15779 loss=1.360189 Training: 88%|████████▊ | 15779/18018 [33:50:20<4:37:42, 7.44s/step, loss=1.360, acc=0.686, acc_len=1.64, thru=16.3, I=16.9, T=36.4, wait=4.6s, pool=61, epoch=2/2][2026-10-19 23:20:39,736] loop.py:380 INFO TIMING step=15780: step=3.395s data=1.363s compute=3.364s [fwd=2.441s bwd=0.909s opt=0.003s] dispatch=2.036s [2026-10-19 23:20:39,736] loop.py:45 INFO TRAIN_STEP step=15780 loss=1.256449 Training: 88%|████████▊ | 15780/18018 [33:50:25<4:15:10, 6.84s/step, loss=1.256, acc=0.707, acc_len=1.70, thru=19.5, I=19.9, T=37.7, wait=2.0s, pool=41, epoch=2/2][2026-10-19 23:20:47,035] loop.py:45 INFO TRAIN_STEP step=15781 loss=1.263694 Training: 88%|████████▊ | 15781/18018 [33:50:33<4:20:12, 6.98s/step, loss=1.264, acc=0.708, acc_len=1.70, thru=17.6, I=18.7, T=35.0, wait=3.6s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:20:47,031] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15780: forward=2720.8ms backward=893.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:51,669] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:20:53,000] loop.py:45 INFO TRAIN_STEP step=15782 loss=1.196403 Training: 88%|████████▊ | 15782/18018 [33:50:39<4:08:44, 6.67s/step, loss=1.196, acc=0.716, acc_len=1.73, thru=20.4, I=23.5, T=34.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:53,178] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:58,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:20:59,019] loop.py:45 INFO TRAIN_STEP step=15783 loss=1.102198 Training: 88%|████████▊ | 15783/18018 [33:50:45<4:01:17, 6.48s/step, loss=1.102, acc=0.735, acc_len=1.80, thru=20.6, I=24.2, T=36.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:20:59,294] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:06,425] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:21:06,546] loop.py:45 INFO TRAIN_STEP step=15784 loss=1.234814 Training: 88%|████████▊ | 15784/18018 [33:50:52<4:12:53, 6.79s/step, loss=1.235, acc=0.715, acc_len=1.73, thru=16.9, I=18.9, T=37.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:06,926] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:13,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:21:13,931] loop.py:380 INFO TIMING step=15785: step=3.455s data=1.329s compute=3.419s [fwd=2.429s bwd=0.975s opt=0.003s] dispatch=3.922s [2026-10-19 23:21:13,932] loop.py:45 INFO TRAIN_STEP step=15785 loss=1.172525 Training: 88%|████████▊ | 15785/18018 [33:51:00<4:19:26, 6.97s/step, loss=1.173, acc=0.724, acc_len=1.77, thru=18.0, I=20.2, T=37.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:14,050] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 23:21:19,547] loop.py:45 INFO TRAIN_STEP step=15786 loss=1.281485 Training: 88%|████████▊ | 15786/18018 [33:51:05<4:04:09, 6.56s/step, loss=1.281, acc=0.699, acc_len=1.68, thru=20.3, I=20.9, T=39.0, wait=2.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:21:19,542] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15785: forward=2332.1ms backward=880.8ms [2026-10-19 23:21:26,801] loop.py:45 INFO TRAIN_STEP step=15787 loss=1.270904 Training: 88%|████████▊ | 15787/18018 [33:51:12<4:11:46, 6.77s/step, loss=1.271, acc=0.709, acc_len=1.71, thru=16.7, I=16.9, T=38.6, wait=3.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:33,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:21:35,097] loop.py:45 INFO TRAIN_STEP step=15788 loss=1.313122 Training: 88%|████████▊ | 15788/18018 [33:51:21<4:28:39, 7.23s/step, loss=1.313, acc=0.698, acc_len=1.68, thru=18.5, I=18.9, T=31.4, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:35,529] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:41,268] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:21:41,763] loop.py:45 INFO TRAIN_STEP step=15789 loss=1.258380 Training: 88%|████████▊ | 15789/18018 [33:51:27<4:22:16, 7.06s/step, loss=1.258, acc=0.707, acc_len=1.71, thru=18.1, I=24.2, T=38.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:41,797] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:48,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:21:49,206] loop.py:380 INFO TIMING step=15790: step=3.520s data=1.335s compute=3.490s [fwd=2.540s bwd=0.936s opt=0.003s] dispatch=3.913s [2026-10-19 23:21:49,206] loop.py:45 INFO TRAIN_STEP step=15790 loss=1.269788 Training: 88%|████████▊ | 15790/18018 [33:51:35<4:26:24, 7.17s/step, loss=1.270, acc=0.703, acc_len=1.69, thru=16.5, I=19.1, T=36.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:21:49,622] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:21:54,840] loop.py:45 INFO TRAIN_STEP step=15791 loss=1.259399 Training: 88%|████████▊ | 15791/18018 [33:51:40<4:09:08, 6.71s/step, loss=1.259, acc=0.708, acc_len=1.71, thru=19.6, I=23.4, T=39.9, wait=2.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:21:54,836] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15790: forward=2300.5ms backward=867.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:01,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:22:02,606] loop.py:45 INFO TRAIN_STEP step=15792 loss=1.298840 Training: 88%|████████▊ | 15792/18018 [33:51:48<4:20:46, 7.03s/step, loss=1.299, acc=0.700, acc_len=1.68, thru=18.0, I=19.8, T=36.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:02,827] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:10,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:22:11,235] loop.py:45 INFO TRAIN_STEP step=15793 loss=1.293646 Training: 88%|████████▊ | 15793/18018 [33:51:57<4:38:27, 7.51s/step, loss=1.294, acc=0.701, acc_len=1.69, thru=13.3, I=16.3, T=36.2, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:11,290] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:17,389] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:22:17,837] loop.py:45 INFO TRAIN_STEP step=15794 loss=1.266985 Training: 88%|████████▊ | 15794/18018 [33:52:03<4:28:13, 7.24s/step, loss=1.267, acc=0.707, acc_len=1.70, thru=18.8, I=21.5, T=37.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:17,910] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-19 23:22:24,552] loop.py:380 INFO TIMING step=15795: step=3.422s data=1.398s compute=3.388s [fwd=2.403s bwd=0.971s opt=0.003s] dispatch=3.284s [2026-10-19 23:22:24,553] loop.py:45 INFO TRAIN_STEP step=15795 loss=1.363484 Training: 88%|████████▊ | 15795/18018 [33:52:10<4:22:18, 7.08s/step, loss=1.363, acc=0.689, acc_len=1.64, thru=18.9, I=20.6, T=37.4, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:31,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:22:31,858] loop.py:45 INFO TRAIN_STEP step=15796 loss=1.294634 Training: 88%|████████▊ | 15796/18018 [33:52:17<4:24:42, 7.15s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=18.1, I=19.2, T=35.3, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:22:31,853] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15795: forward=2391.1ms backward=1168.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:32,159] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:39,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:22:39,234] loop.py:45 INFO TRAIN_STEP step=15797 loss=1.306850 Training: 88%|████████▊ | 15797/18018 [33:52:25<4:27:06, 7.22s/step, loss=1.307, acc=0.692, acc_len=1.65, thru=17.6, I=17.5, T=36.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:39,575] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:22:45,894] loop.py:45 INFO TRAIN_STEP step=15798 loss=1.339396 Training: 88%|████████▊ | 15798/18018 [33:52:31<4:20:49, 7.05s/step, loss=1.339, acc=0.693, acc_len=1.65, thru=17.3, I=16.6, T=37.8, wait=3.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:55,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:22:56,509] loop.py:45 INFO TRAIN_STEP step=15799 loss=1.214436 Training: 88%|████████▊ | 15799/18018 [33:52:42<5:00:17, 8.12s/step, loss=1.214, acc=0.716, acc_len=1.74, thru=14.2, I=14.8, T=32.0, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:22:56,794] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:01,838] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:23:02,736] loop.py:380 INFO TIMING step=15800: step=3.378s data=1.371s compute=3.348s [fwd=2.420s bwd=0.914s opt=0.003s] dispatch=2.838s [2026-10-19 23:23:02,736] loop.py:45 INFO TRAIN_STEP step=15800 loss=1.394511 Training: 88%|████████▊ | 15800/18018 [33:52:48<4:39:08, 7.55s/step, loss=1.395, acc=0.680, acc_len=1.62, thru=19.7, I=22.5, T=37.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:02,853] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:07,490] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:23:08,149] loop.py:45 INFO TRAIN_STEP step=15801 loss=1.321707 Training: 88%|████████▊ | 15801/18018 [33:52:54<4:15:19, 6.91s/step, loss=1.322, acc=0.695, acc_len=1.66, thru=22.6, I=26.6, T=40.0, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:23:08,144] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15800: forward=2289.7ms backward=860.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:08,497] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:23:13,951] loop.py:45 INFO TRAIN_STEP step=15802 loss=1.155274 Training: 88%|████████▊ | 15802/18018 [33:53:00<4:02:56, 6.58s/step, loss=1.155, acc=0.722, acc_len=1.76, thru=21.5, I=27.6, T=38.8, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:21,092] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:23:21,231] loop.py:45 INFO TRAIN_STEP step=15803 loss=1.304345 Training: 88%|████████▊ | 15803/18018 [33:53:07<4:10:35, 6.79s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=17.3, I=18.4, T=37.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:21,594] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:23:27,136] loop.py:45 INFO TRAIN_STEP step=15804 loss=1.163658 Training: 88%|████████▊ | 15804/18018 [33:53:13<4:00:42, 6.52s/step, loss=1.164, acc=0.725, acc_len=1.77, thru=20.1, I=21.2, T=37.6, wait=2.5s, pool=45, epoch=2/2][2026-10-19 23:23:34,928] loop.py:380 INFO TIMING step=15805: step=3.457s data=1.408s compute=3.420s [fwd=2.455s bwd=0.950s opt=0.003s] dispatch=4.325s [2026-10-19 23:23:34,928] loop.py:45 INFO TRAIN_STEP step=15805 loss=1.315970 Training: 88%|████████▊ | 15805/18018 [33:53:21<4:14:40, 6.90s/step, loss=1.316, acc=0.699, acc_len=1.68, thru=15.6, I=15.8, T=37.0, wait=4.3s, pool=45, epoch=2/2][2026-10-19 23:23:44,899] loop.py:45 INFO TRAIN_STEP step=15806 loss=1.291519 Training: 88%|████████▊ | 15806/18018 [33:53:30<4:48:25, 7.82s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=14.2, I=12.8, T=33.7, wait=6.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:23:44,895] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15805: forward=2741.4ms backward=1006.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:50,615] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:23:51,074] loop.py:45 INFO TRAIN_STEP step=15807 loss=1.182063 Training: 88%|████████▊ | 15807/18018 [33:53:37<4:30:04, 7.33s/step, loss=1.182, acc=0.720, acc_len=1.75, thru=19.3, I=19.9, T=37.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:51,119] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:58,582] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:23:59,553] loop.py:45 INFO TRAIN_STEP step=15808 loss=1.361682 Training: 88%|████████▊ | 15808/18018 [33:53:45<4:42:39, 7.67s/step, loss=1.362, acc=0.693, acc_len=1.66, thru=15.3, I=17.9, T=35.8, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:23:59,598] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:24:05,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:24:06,446] loop.py:45 INFO TRAIN_STEP step=15809 loss=1.304455 Training: 88%|████████▊ | 15809/18018 [33:53:52<4:33:54, 7.44s/step, loss=1.304, acc=0.696, acc_len=1.66, thru=18.8, I=21.9, T=38.2, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:24:06,477] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:24:11,044] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:24:12,202] loop.py:380 INFO TIMING step=15810: step=3.318s data=1.339s compute=3.290s [fwd=2.364s bwd=0.912s opt=0.003s] dispatch=2.429s [2026-10-19 23:24:12,202] loop.py:45 INFO TRAIN_STEP step=15810 loss=1.276859 Training: 88%|████████▊ | 15810/18018 [33:53:58<4:15:11, 6.93s/step, loss=1.277, acc=0.704, acc_len=1.70, thru=20.9, I=24.8, T=38.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:24:12,553] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:24:18,555] loop.py:45 INFO TRAIN_STEP step=15811 loss=1.287310 Training: 88%|████████▊ | 15811/18018 [33:54:04<4:08:40, 6.76s/step, loss=1.287, acc=0.701, acc_len=1.68, thru=19.5, I=23.2, T=37.6, wait=2.9s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:24:18,550] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15810: forward=2420.3ms backward=943.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:24:26,525] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:24:27,741] loop.py:45 INFO TRAIN_STEP step=15812 loss=1.342240 Training: 88%|████████▊ | 15812/18018 [33:54:13<4:35:19, 7.49s/step, loss=1.342, acc=0.692, acc_len=1.65, thru=15.4, I=16.6, T=34.4, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:24:28,036] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:24:37,022] loop.py:45 INFO TRAIN_STEP step=15813 loss=1.209535 Training: 88%|████████▊ | 15813/18018 [33:54:23<4:54:56, 8.03s/step, loss=1.210, acc=0.716, acc_len=1.74, thru=13.4, I=13.3, T=37.9, wait=5.9s, pool=57, epoch=2/2][2026-10-19 23:24:44,701] loop.py:45 INFO TRAIN_STEP step=15814 loss=1.311880 Training: 88%|████████▊ | 15814/18018 [33:54:30<4:50:59, 7.92s/step, loss=1.312, acc=0.699, acc_len=1.67, thru=16.4, I=17.1, T=38.0, wait=4.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:24:53,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:24:53,892] loop.py:380 INFO TIMING step=15815: step=3.601s data=1.473s compute=3.575s [fwd=2.523s bwd=1.038s opt=0.003s] dispatch=5.580s [2026-10-19 23:24:53,892] loop.py:45 INFO TRAIN_STEP step=15815 loss=1.134593 Training: 88%|████████▊ | 15815/18018 [33:54:39<5:04:49, 8.30s/step, loss=1.135, acc=0.729, acc_len=1.78, thru=14.7, I=14.7, T=35.5, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:24:54,069] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:01,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:25:01,685] loop.py:45 INFO TRAIN_STEP step=15816 loss=1.199285 Training: 88%|████████▊ | 15816/18018 [33:54:47<4:59:06, 8.15s/step, loss=1.199, acc=0.716, acc_len=1.73, thru=16.4, I=18.0, T=36.9, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:25:01,680] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15815: forward=2425.7ms backward=984.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:02,033] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:07,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:25:08,710] loop.py:45 INFO TRAIN_STEP step=15817 loss=1.358868 Training: 88%|████████▊ | 15817/18018 [33:54:54<4:46:34, 7.81s/step, loss=1.359, acc=0.689, acc_len=1.64, thru=18.8, I=20.5, T=37.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:08,928] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:13,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:25:14,140] loop.py:45 INFO TRAIN_STEP step=15818 loss=1.287581 Training: 88%|████████▊ | 15818/18018 [33:55:00<4:20:13, 7.10s/step, loss=1.288, acc=0.704, acc_len=1.69, thru=22.1, I=23.8, T=38.6, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:14,452] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:20,402] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:25:20,718] loop.py:45 INFO TRAIN_STEP step=15819 loss=1.262021 Training: 88%|████████▊ | 15819/18018 [33:55:06<4:14:25, 6.94s/step, loss=1.262, acc=0.704, acc_len=1.69, thru=19.7, I=22.8, T=36.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:20,904] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:30,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:25:30,835] loop.py:380 INFO TIMING step=15820: step=3.722s data=1.493s compute=3.692s [fwd=2.624s bwd=1.054s opt=0.003s] dispatch=6.384s [2026-10-19 23:25:30,835] loop.py:45 INFO TRAIN_STEP step=15820 loss=1.284485 Training: 88%|████████▊ | 15820/18018 [33:55:16<4:49:10, 7.89s/step, loss=1.284, acc=0.704, acc_len=1.69, thru=14.3, I=12.3, T=34.4, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:30,978] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:25:37,744] loop.py:45 INFO TRAIN_STEP step=15821 loss=1.202760 Training: 88%|████████▊ | 15821/18018 [33:55:23<4:38:13, 7.60s/step, loss=1.203, acc=0.718, acc_len=1.74, thru=17.6, I=17.8, T=37.5, wait=3.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:25:37,739] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15820: forward=2416.8ms backward=946.8ms [2026-10-19 23:25:45,086] loop.py:45 INFO TRAIN_STEP step=15822 loss=1.274082 Training: 88%|████████▊ | 15822/18018 [33:55:31<4:35:17, 7.52s/step, loss=1.274, acc=0.701, acc_len=1.69, thru=17.8, I=18.0, T=36.2, wait=3.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:53,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:25:53,697] loop.py:45 INFO TRAIN_STEP step=15823 loss=1.320037 Training: 88%|████████▊ | 15823/18018 [33:55:39<4:47:07, 7.85s/step, loss=1.320, acc=0.692, acc_len=1.65, thru=15.7, I=16.3, T=34.9, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:25:53,825] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 23:26:02,213] loop.py:45 INFO TRAIN_STEP step=15824 loss=1.217766 Training: 88%|████████▊ | 15824/18018 [33:55:48<4:54:18, 8.05s/step, loss=1.218, acc=0.715, acc_len=1.73, thru=13.7, I=16.5, T=37.0, wait=5.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:09,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:26:09,710] loop.py:380 INFO TIMING step=15825: step=3.722s data=1.538s compute=3.690s [fwd=2.595s bwd=1.081s opt=0.003s] dispatch=3.767s [2026-10-19 23:26:09,710] loop.py:45 INFO TRAIN_STEP step=15825 loss=1.244869 Training: 88%|████████▊ | 15825/18018 [33:55:55<4:48:09, 7.88s/step, loss=1.245, acc=0.713, acc_len=1.72, thru=18.7, I=17.0, T=34.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:09,956] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:15,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:26:16,117] loop.py:45 INFO TRAIN_STEP step=15826 loss=1.369973 Training: 88%|████████▊ | 15826/18018 [33:56:02<4:31:49, 7.44s/step, loss=1.370, acc=0.691, acc_len=1.65, thru=19.2, I=20.4, T=36.9, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:26:16,112] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15825: forward=2467.7ms backward=952.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:16,440] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:21,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:26:22,011] loop.py:45 INFO TRAIN_STEP step=15827 loss=1.259879 Training: 88%|████████▊ | 15827/18018 [33:56:08<4:14:45, 6.98s/step, loss=1.260, acc=0.708, acc_len=1.71, thru=21.1, I=22.8, T=38.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:22,333] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:28,104] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:26:29,037] loop.py:45 INFO TRAIN_STEP step=15828 loss=1.226921 Training: 88%|████████▊ | 15828/18018 [33:56:15<4:15:10, 6.99s/step, loss=1.227, acc=0.708, acc_len=1.71, thru=18.9, I=21.0, T=37.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:29,122] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-19 23:26:35,402] loop.py:45 INFO TRAIN_STEP step=15829 loss=1.165158 Training: 88%|████████▊ | 15829/18018 [33:56:21<4:08:12, 6.80s/step, loss=1.165, acc=0.722, acc_len=1.76, thru=15.8, I=20.8, T=38.9, wait=3.1s, pool=25, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 23:26:35,501] eagle_store.py:517 WARNING batch_get_buffer missing keys (k506526-9afcbeca_hs, k506526-9afcbeca_ids, k506526-9afcbeca_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:44,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 23:26:35,490] eagle_store.py:517 WARNING batch_get_buffer missing keys (k506525-94fa30cc_hs, k506525-94fa30cc_ids, k506525-94fa30cc_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 23:26:45,690] loop.py:380 INFO TIMING step=15830: step=3.686s data=1.439s compute=3.659s [fwd=2.620s bwd=1.024s opt=0.003s] dispatch=6.593s [2026-10-19 23:26:45,690] loop.py:45 INFO TRAIN_STEP step=15830 loss=1.384231 Training: 88%|████████▊ | 15830/18018 [33:56:31<4:46:13, 7.85s/step, loss=1.384, acc=0.684, acc_len=1.62, thru=18.7, I=15.6, T=34.7, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:26:45,852] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:26:51,809] loop.py:45 INFO TRAIN_STEP step=15831 loss=1.386298 Training: 88%|████████▊ | 15831/18018 [33:56:37<4:27:10, 7.33s/step, loss=1.386, acc=0.682, acc_len=1.62, thru=19.3, I=23.4, T=38.1, wait=2.8s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:26:51,805] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15830: forward=2420.0ms backward=886.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:00,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:27:01,166] loop.py:45 INFO TRAIN_STEP step=15832 loss=1.158850 Training: 88%|████████▊ | 15832/18018 [33:56:47<4:49:11, 7.94s/step, loss=1.159, acc=0.724, acc_len=1.76, thru=15.1, I=15.4, T=37.0, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:01,659] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:27:10,798] loop.py:45 INFO TRAIN_STEP step=15833 loss=1.428119 Training: 88%|████████▊ | 15833/18018 [33:56:56<5:07:35, 8.45s/step, loss=1.428, acc=0.678, acc_len=1.60, thru=10.1, I=10.8, T=36.1, wait=6.1s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:19,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:27:19,996] loop.py:45 INFO TRAIN_STEP step=15834 loss=1.245158 Training: 88%|████████▊ | 15834/18018 [33:57:06<5:15:39, 8.67s/step, loss=1.245, acc=0.709, acc_len=1.71, thru=17.9, I=15.8, T=36.5, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:20,283] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:24,095] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:27:25,784] loop.py:380 INFO TIMING step=15835: step=3.327s data=1.357s compute=3.298s [fwd=2.378s bwd=0.906s opt=0.003s] dispatch=2.451s [2026-10-19 23:27:25,784] loop.py:45 INFO TRAIN_STEP step=15835 loss=1.254354 Training: 88%|████████▊ | 15835/18018 [33:57:11<4:44:02, 7.81s/step, loss=1.254, acc=0.705, acc_len=1.69, thru=21.7, I=25.4, T=38.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:26,110] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:30,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:27:31,850] loop.py:45 INFO TRAIN_STEP step=15836 loss=1.244577 Training: 88%|████████▊ | 15836/18018 [33:57:17<4:24:54, 7.28s/step, loss=1.245, acc=0.713, acc_len=1.72, thru=21.7, I=30.7, T=39.1, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:27:31,845] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15835: forward=2314.7ms backward=906.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:32,329] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:37,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:27:37,479] loop.py:45 INFO TRAIN_STEP step=15837 loss=1.139258 Training: 88%|████████▊ | 15837/18018 [33:57:23<4:06:44, 6.79s/step, loss=1.139, acc=0.729, acc_len=1.78, thru=21.7, I=28.9, T=39.0, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:37,535] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 23:27:45,457] loop.py:45 INFO TRAIN_STEP step=15838 loss=1.207756 Training: 88%|████████▊ | 15838/18018 [33:57:31<4:19:35, 7.14s/step, loss=1.208, acc=0.716, acc_len=1.73, thru=15.4, I=17.4, T=37.9, wait=4.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:53,000] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:27:53,751] loop.py:45 INFO TRAIN_STEP step=15839 loss=1.198031 Training: 88%|████████▊ | 15839/18018 [33:57:39<4:32:00, 7.49s/step, loss=1.198, acc=0.721, acc_len=1.75, thru=16.0, I=16.5, T=36.1, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:54,007] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:27:58,679] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:27:59,921] loop.py:380 INFO TIMING step=15840: step=3.295s data=1.329s compute=3.262s [fwd=2.337s bwd=0.910s opt=0.003s] dispatch=2.865s [2026-10-19 23:27:59,922] loop.py:45 INFO TRAIN_STEP step=15840 loss=1.178602 Training: 88%|████████▊ | 15840/18018 [33:57:46<4:17:29, 7.09s/step, loss=1.179, acc=0.721, acc_len=1.76, thru=20.4, I=24.0, T=38.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:00,186] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:05,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 23:28:06,418] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15840: forward=2370.2ms backward=897.5ms [2026-10-19 23:28:06,423] loop.py:45 INFO TRAIN_STEP step=15841 loss=1.174386 Training: 88%|████████▊ | 15841/18018 [33:57:52<4:10:57, 6.92s/step, loss=1.174, acc=0.721, acc_len=1.76, thru=19.9, I=24.6, T=38.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:06,675] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:28:12,749] loop.py:45 INFO TRAIN_STEP step=15842 loss=1.216252 Training: 88%|████████▊ | 15842/18018 [33:57:58<4:04:23, 6.74s/step, loss=1.216, acc=0.712, acc_len=1.72, thru=18.6, I=20.0, T=38.3, wait=3.0s, pool=49, epoch=2/2][2026-10-19 23:28:20,541] loop.py:45 INFO TRAIN_STEP step=15843 loss=1.148952 Training: 88%|████████▊ | 15843/18018 [33:58:06<4:15:45, 7.06s/step, loss=1.149, acc=0.726, acc_len=1.77, thru=17.8, I=17.3, T=36.2, wait=4.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:26,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:28:27,361] loop.py:45 INFO TRAIN_STEP step=15844 loss=1.325810 Training: 88%|████████▊ | 15844/18018 [33:58:13<4:13:04, 6.98s/step, loss=1.326, acc=0.695, acc_len=1.66, thru=18.6, I=20.4, T=37.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:27,434] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:32,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:28:32,821] loop.py:380 INFO TIMING step=15845: step=3.252s data=1.269s compute=3.229s [fwd=2.353s bwd=0.862s opt=0.003s] dispatch=2.199s [2026-10-19 23:28:32,821] loop.py:45 INFO TRAIN_STEP step=15845 loss=1.272260 Training: 88%|████████▊ | 15845/18018 [33:58:18<3:56:24, 6.53s/step, loss=1.272, acc=0.704, acc_len=1.69, thru=22.7, I=25.0, T=39.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:33,223] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:28:40,224] loop.py:45 INFO TRAIN_STEP step=15846 loss=1.282763 Training: 88%|████████▊ | 15846/18018 [33:58:26<4:05:47, 6.79s/step, loss=1.283, acc=0.702, acc_len=1.69, thru=17.2, I=18.3, T=36.7, wait=3.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:28:40,219] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15845: forward=2465.2ms backward=976.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:47,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:28:48,086] loop.py:45 INFO TRAIN_STEP step=15847 loss=1.267296 Training: 88%|████████▊ | 15847/18018 [33:58:34<4:17:19, 7.11s/step, loss=1.267, acc=0.702, acc_len=1.69, thru=18.2, I=17.7, T=37.4, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:48,263] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:53,499] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:28:54,026] loop.py:45 INFO TRAIN_STEP step=15848 loss=1.222206 Training: 88%|████████▊ | 15848/18018 [33:58:40<4:04:28, 6.76s/step, loss=1.222, acc=0.711, acc_len=1.72, thru=21.2, I=22.2, T=38.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:28:54,505] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:28:59,619] loop.py:45 INFO TRAIN_STEP step=15849 loss=1.219945 Training: 88%|████████▊ | 15849/18018 [33:58:45<3:51:43, 6.41s/step, loss=1.220, acc=0.713, acc_len=1.72, thru=19.9, I=24.4, T=40.5, wait=2.4s, pool=41, epoch=2/2][2026-10-19 23:29:07,112] loop.py:380 INFO TIMING step=15850: step=3.562s data=1.516s compute=3.524s [fwd=2.529s bwd=0.981s opt=0.003s] dispatch=3.922s [2026-10-19 23:29:07,112] loop.py:45 INFO TRAIN_STEP step=15850 loss=1.325853 Training: 88%|████████▊ | 15850/18018 [33:58:53<4:03:21, 6.73s/step, loss=1.326, acc=0.694, acc_len=1.66, thru=18.8, I=18.6, T=35.9, wait=3.9s, pool=61, epoch=2/2][2026-10-19 23:29:13,257] loop.py:45 INFO TRAIN_STEP step=15851 loss=1.115535 Training: 88%|████████▊ | 15851/18018 [33:58:59<3:56:51, 6.56s/step, loss=1.116, acc=0.732, acc_len=1.79, thru=19.7, I=19.7, T=37.4, wait=2.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:29:13,253] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15850: forward=2422.2ms backward=954.6ms [2026-10-19 23:29:20,175] loop.py:45 INFO TRAIN_STEP step=15852 loss=1.335417 Training: 88%|████████▊ | 15852/18018 [33:59:06<4:00:38, 6.67s/step, loss=1.335, acc=0.694, acc_len=1.66, thru=18.4, I=19.4, T=36.7, wait=3.4s, pool=61, epoch=2/2][2026-10-19 23:29:28,697] loop.py:45 INFO TRAIN_STEP step=15853 loss=1.200911 Training: 88%|████████▊ | 15853/18018 [33:59:14<4:20:36, 7.22s/step, loss=1.201, acc=0.715, acc_len=1.73, thru=12.8, I=12.2, T=37.3, wait=5.1s, pool=37, epoch=2/2][2026-10-19 23:29:37,110] loop.py:45 INFO TRAIN_STEP step=15854 loss=1.275103 Training: 88%|████████▊ | 15854/18018 [33:59:23<4:33:21, 7.58s/step, loss=1.275, acc=0.702, acc_len=1.69, thru=14.3, I=12.9, T=37.2, wait=5.0s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:29:45,327] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:29:46,538] loop.py:380 INFO TIMING step=15855: step=3.556s data=1.435s compute=3.531s [fwd=2.544s bwd=0.973s opt=0.003s] dispatch=5.863s [2026-10-19 23:29:46,538] loop.py:45 INFO TRAIN_STEP step=15855 loss=1.295130 Training: 88%|████████▊ | 15855/18018 [33:59:32<4:53:14, 8.13s/step, loss=1.295, acc=0.706, acc_len=1.69, thru=18.1, I=17.2, T=36.0, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:29:46,836] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:29:54,978] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:29:55,574] loop.py:45 INFO TRAIN_STEP step=15856 loss=1.292307 Training: 88%|████████▊ | 15856/18018 [33:59:41<5:02:50, 8.40s/step, loss=1.292, acc=0.702, acc_len=1.68, thru=14.7, I=15.8, T=35.9, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:29:55,569] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15855: forward=2493.2ms backward=1026.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:29:55,984] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:30:01,455] loop.py:45 INFO TRAIN_STEP step=15857 loss=1.211759 Training: 88%|████████▊ | 15857/18018 [33:59:47<4:35:27, 7.65s/step, loss=1.212, acc=0.716, acc_len=1.73, thru=17.4, I=21.7, T=39.3, wait=2.6s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:10,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:30:11,444] loop.py:45 INFO TRAIN_STEP step=15858 loss=1.221495 Training: 88%|████████▊ | 15858/18018 [33:59:57<5:00:36, 8.35s/step, loss=1.221, acc=0.712, acc_len=1.72, thru=17.6, I=15.0, T=35.3, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:11,477] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:18,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:30:19,151] loop.py:45 INFO TRAIN_STEP step=15859 loss=1.253774 Training: 88%|████████▊ | 15859/18018 [34:00:05<4:53:32, 8.16s/step, loss=1.254, acc=0.708, acc_len=1.71, thru=16.8, I=18.6, T=37.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:19,252] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:25,257] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:30:26,578] loop.py:380 INFO TIMING step=15860: step=3.385s data=1.376s compute=3.354s [fwd=2.400s bwd=0.939s opt=0.003s] dispatch=4.031s [2026-10-19 23:30:26,578] loop.py:45 INFO TRAIN_STEP step=15860 loss=1.342292 Training: 88%|████████▊ | 15860/18018 [34:00:12<4:45:30, 7.94s/step, loss=1.342, acc=0.691, acc_len=1.65, thru=18.2, I=19.8, T=37.8, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:26,766] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:30:32,666] loop.py:45 INFO TRAIN_STEP step=15861 loss=1.313855 Training: 88%|████████▊ | 15861/18018 [34:00:18<4:25:24, 7.38s/step, loss=1.314, acc=0.698, acc_len=1.68, thru=15.8, I=23.6, T=37.3, wait=2.7s, pool=17, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:30:32,662] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15860: forward=2417.8ms backward=964.1ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 23:30:33,383] eagle_store.py:517 WARNING batch_get_buffer missing keys (k507554-8e90c2f4_hs, k507554-8e90c2f4_ids, k507554-8e90c2f4_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:42,193] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 23:30:33,380] eagle_store.py:517 WARNING batch_get_buffer missing keys (k507553-8801947e_hs, k507553-8801947e_ids, k507553-8801947e_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-19 23:30:43,266] loop.py:45 INFO TRAIN_STEP step=15862 loss=1.253839 Training: 88%|████████▊ | 15862/18018 [34:00:29<4:59:58, 8.35s/step, loss=1.254, acc=0.707, acc_len=1.70, thru=17.4, I=15.9, T=34.5, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:43,701] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:50,282] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:30:51,657] loop.py:45 INFO TRAIN_STEP step=15863 loss=1.254380 Training: 88%|████████▊ | 15863/18018 [34:00:37<5:00:18, 8.36s/step, loss=1.254, acc=0.706, acc_len=1.69, thru=13.6, I=18.4, T=37.2, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:51,800] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:58,087] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:30:58,524] loop.py:45 INFO TRAIN_STEP step=15864 loss=1.264665 Training: 88%|████████▊ | 15864/18018 [34:00:44<4:44:03, 7.91s/step, loss=1.265, acc=0.708, acc_len=1.70, thru=18.8, I=22.0, T=37.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:30:58,604] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-19 23:31:05,674] loop.py:380 INFO TIMING step=15865: step=3.415s data=1.341s compute=3.387s [fwd=2.418s bwd=0.955s opt=0.003s] dispatch=3.726s [2026-10-19 23:31:05,674] loop.py:45 INFO TRAIN_STEP step=15865 loss=1.247725 Training: 88%|████████▊ | 15865/18018 [34:00:51<4:35:43, 7.68s/step, loss=1.248, acc=0.709, acc_len=1.71, thru=17.6, I=18.6, T=37.5, wait=3.7s, pool=49, epoch=2/2][2026-10-19 23:31:13,445] loop.py:45 INFO TRAIN_STEP step=15866 loss=1.307439 Training: 88%|████████▊ | 15866/18018 [34:00:59<4:36:31, 7.71s/step, loss=1.307, acc=0.695, acc_len=1.67, thru=15.6, I=15.2, T=37.5, wait=4.3s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:31:13,441] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15865: forward=2423.8ms backward=952.8ms [2026-10-19 23:31:20,212] loop.py:45 INFO TRAIN_STEP step=15867 loss=1.109782 Training: 88%|████████▊ | 15867/18018 [34:01:06<4:26:17, 7.43s/step, loss=1.110, acc=0.732, acc_len=1.79, thru=17.4, I=17.8, T=38.2, wait=3.4s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:29,610] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:31:30,574] loop.py:45 INFO TRAIN_STEP step=15868 loss=1.249528 Training: 88%|████████▊ | 15868/18018 [34:01:16<4:57:41, 8.31s/step, loss=1.250, acc=0.710, acc_len=1.71, thru=15.1, I=16.1, T=34.5, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:30,631] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:35,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:31:36,702] loop.py:45 INFO TRAIN_STEP step=15869 loss=1.222738 Training: 88%|████████▊ | 15869/18018 [34:01:22<4:34:07, 7.65s/step, loss=1.223, acc=0.714, acc_len=1.72, thru=19.9, I=23.5, T=37.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:36,860] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:43,489] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:31:43,920] loop.py:380 INFO TIMING step=15870: step=3.393s data=1.392s compute=3.364s [fwd=2.400s bwd=0.950s opt=0.003s] dispatch=3.815s [2026-10-19 23:31:43,921] loop.py:45 INFO TRAIN_STEP step=15870 loss=1.181345 Training: 88%|████████▊ | 15870/18018 [34:01:30<4:29:19, 7.52s/step, loss=1.181, acc=0.723, acc_len=1.76, thru=18.0, I=21.3, T=37.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:43,994] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:49,616] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:31:50,005] loop.py:45 INFO TRAIN_STEP step=15871 loss=1.182527 (TrainerActor pid=9310) [2026-10-19 23:31:50,000] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15870: forward=2371.1ms backward=950.2ms Training: 88%|████████▊ | 15871/18018 [34:01:36<4:13:45, 7.09s/step, loss=1.183, acc=0.721, acc_len=1.75, thru=20.5, I=22.1, T=37.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:50,128] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:55,519] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:31:56,244] loop.py:45 INFO TRAIN_STEP step=15872 loss=1.160172 Training: 88%|████████▊ | 15872/18018 [34:01:42<4:04:29, 6.84s/step, loss=1.160, acc=0.725, acc_len=1.77, thru=20.5, I=22.2, T=38.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:31:56,525] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:00,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:32:01,950] loop.py:45 INFO TRAIN_STEP step=15873 loss=1.144359 Training: 88%|████████▊ | 15873/18018 [34:01:48<3:52:15, 6.50s/step, loss=1.144, acc=0.727, acc_len=1.77, thru=22.5, I=26.1, T=36.0, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:02,420] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:32:09,189] loop.py:45 INFO TRAIN_STEP step=15874 loss=1.294356 Training: 88%|████████▊ | 15874/18018 [34:01:55<4:00:05, 6.72s/step, loss=1.294, acc=0.704, acc_len=1.69, thru=16.3, I=18.7, T=36.6, wait=3.7s, pool=53, epoch=2/2][2026-10-19 23:32:16,011] loop.py:380 INFO TIMING step=15875: step=3.455s data=1.479s compute=3.406s [fwd=2.397s bwd=0.994s opt=0.003s] dispatch=3.360s [2026-10-19 23:32:16,012] loop.py:45 INFO TRAIN_STEP step=15875 loss=1.352777 Training: 88%|████████▊ | 15875/18018 [34:02:02<4:01:05, 6.75s/step, loss=1.353, acc=0.692, acc_len=1.65, thru=17.0, I=17.8, T=37.1, wait=3.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:23,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:32:24,977] loop.py:45 INFO TRAIN_STEP step=15876 loss=1.122009 Training: 88%|████████▊ | 15876/18018 [34:02:11<4:24:43, 7.42s/step, loss=1.122, acc=0.731, acc_len=1.79, thru=16.5, I=16.8, T=37.2, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:32:24,973] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15875: forward=2466.6ms backward=931.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:25,027] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-19 23:32:32,243] loop.py:45 INFO TRAIN_STEP step=15877 loss=1.199157 Training: 88%|████████▊ | 15877/18018 [34:02:18<4:22:58, 7.37s/step, loss=1.199, acc=0.724, acc_len=1.77, thru=14.6, I=17.0, T=36.8, wait=3.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:39,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:32:40,152] loop.py:45 INFO TRAIN_STEP step=15878 loss=1.197086 Training: 88%|████████▊ | 15878/18018 [34:02:26<4:28:39, 7.53s/step, loss=1.197, acc=0.716, acc_len=1.74, thru=17.8, I=18.4, T=36.6, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:40,185] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:44,701] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:32:45,656] loop.py:45 INFO TRAIN_STEP step=15879 loss=1.232290 Training: 88%|████████▊ | 15879/18018 [34:02:31<4:06:49, 6.92s/step, loss=1.232, acc=0.710, acc_len=1.72, thru=21.6, I=24.1, T=39.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:45,724] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-19 23:32:51,227] loop.py:380 INFO TIMING step=15880: step=3.267s data=1.316s compute=3.229s [fwd=2.298s bwd=0.919s opt=0.003s] dispatch=2.295s [2026-10-19 23:32:51,228] loop.py:45 INFO TRAIN_STEP step=15880 loss=1.265080 Training: 88%|████████▊ | 15880/18018 [34:02:37<3:52:15, 6.52s/step, loss=1.265, acc=0.705, acc_len=1.70, thru=21.8, I=26.8, T=39.2, wait=2.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:57,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:32:58,288] loop.py:45 INFO TRAIN_STEP step=15881 loss=1.213186 Training: 88%|████████▊ | 15881/18018 [34:02:44<3:57:57, 6.68s/step, loss=1.213, acc=0.712, acc_len=1.72, thru=18.8, I=20.3, T=38.1, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:32:58,283] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15880: forward=2371.2ms backward=947.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:32:58,614] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:33:06,606] loop.py:45 INFO TRAIN_STEP step=15882 loss=1.274580 Training: 88%|████████▊ | 15882/18018 [34:02:52<4:15:19, 7.17s/step, loss=1.275, acc=0.707, acc_len=1.71, thru=12.5, I=15.5, T=37.2, wait=4.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:13,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:33:15,377] loop.py:45 INFO TRAIN_STEP step=15883 loss=1.355962 Training: 88%|████████▊ | 15883/18018 [34:03:01<4:32:15, 7.65s/step, loss=1.356, acc=0.694, acc_len=1.66, thru=16.9, I=18.9, T=35.1, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:15,423] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=49/64 [2026-10-19 23:33:21,481] loop.py:45 INFO TRAIN_STEP step=15884 loss=1.186667 Training: 88%|████████▊ | 15884/18018 [34:03:07<4:15:37, 7.19s/step, loss=1.187, acc=0.718, acc_len=1.73, thru=19.6, I=26.5, T=39.2, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:29,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:33:29,718] loop.py:380 INFO TIMING step=15885: step=4.123s data=1.393s compute=4.090s [fwd=3.104s bwd=0.971s opt=0.003s] dispatch=4.105s [2026-10-19 23:33:29,719] loop.py:45 INFO TRAIN_STEP step=15885 loss=1.171476 Training: 88%|████████▊ | 15885/18018 [34:03:15<4:26:43, 7.50s/step, loss=1.171, acc=0.722, acc_len=1.75, thru=16.0, I=16.7, T=31.0, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:29,752] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:36,958] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:33:37,155] loop.py:45 INFO TRAIN_STEP step=15886 loss=1.294738 Training: 88%|████████▊ | 15886/18018 [34:03:23<4:25:52, 7.48s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=16.5, I=18.2, T=31.1, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:33:37,150] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15885: forward=3103.1ms backward=959.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:37,460] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:33:44,294] loop.py:45 INFO TRAIN_STEP step=15887 loss=1.327623 Training: 88%|████████▊ | 15887/18018 [34:03:30<4:22:04, 7.38s/step, loss=1.328, acc=0.699, acc_len=1.67, thru=17.3, I=17.0, T=34.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:44,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:44,790] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:51,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:33:52,120] loop.py:45 INFO TRAIN_STEP step=15888 loss=1.145635 Training: 88%|████████▊ | 15888/18018 [34:03:38<4:26:44, 7.51s/step, loss=1.146, acc=0.726, acc_len=1.77, thru=17.4, I=18.3, T=36.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:33:52,187] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 23:33:59,464] loop.py:45 INFO TRAIN_STEP step=15889 loss=1.314179 Training: 88%|████████▊ | 15889/18018 [34:03:45<4:24:46, 7.46s/step, loss=1.314, acc=0.700, acc_len=1.67, thru=13.6, I=16.7, T=36.0, wait=3.8s, pool=33, epoch=2/2][2026-10-19 23:34:09,605] loop.py:380 INFO TIMING step=15890: step=3.644s data=1.572s compute=3.585s [fwd=2.510s bwd=1.062s opt=0.003s] dispatch=6.490s [2026-10-19 23:34:09,606] loop.py:45 INFO TRAIN_STEP step=15890 loss=1.298859 Training: 88%|████████▊ | 15890/18018 [34:03:55<4:53:10, 8.27s/step, loss=1.299, acc=0.702, acc_len=1.69, thru=15.5, I=13.8, T=35.1, wait=6.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:34:15,705] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:34:17,360] loop.py:45 INFO TRAIN_STEP step=15891 loss=1.181119 Training: 88%|████████▊ | 15891/18018 [34:04:03<4:47:37, 8.11s/step, loss=1.181, acc=0.722, acc_len=1.76, thru=17.0, I=18.6, T=36.2, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:34:17,356] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15890: forward=2509.1ms backward=975.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:34:17,720] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:34:23,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:34:23,160] loop.py:45 INFO TRAIN_STEP step=15892 loss=1.248300 Training: 88%|████████▊ | 15892/18018 [34:04:09<4:22:51, 7.42s/step, loss=1.248, acc=0.711, acc_len=1.71, thru=20.9, I=24.5, T=39.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:34:23,597] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:34:29,868] loop.py:45 INFO TRAIN_STEP step=15893 loss=1.475603 Training: 88%|████████▊ | 15893/18018 [34:04:15<4:15:11, 7.21s/step, loss=1.476, acc=0.667, acc_len=1.56, thru=18.4, I=19.0, T=38.0, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:34:38,452] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:34:39,712] loop.py:45 INFO TRAIN_STEP step=15894 loss=1.264886 Training: 88%|████████▊ | 15894/18018 [34:04:25<4:43:06, 8.00s/step, loss=1.265, acc=0.704, acc_len=1.70, thru=14.1, I=16.2, T=33.8, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:34:39,961] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:34:45,834] loop.py:380 INFO TIMING step=15895: step=3.565s data=1.658s compute=3.537s [fwd=2.636s bwd=0.888s opt=0.003s] dispatch=2.547s [2026-10-19 23:34:45,834] loop.py:45 INFO TRAIN_STEP step=15895 loss=1.286630 Training: 88%|████████▊ | 15895/18018 [34:04:31<4:23:03, 7.43s/step, loss=1.287, acc=0.699, acc_len=1.67, thru=18.9, I=18.3, T=35.9, wait=2.5s, pool=57, epoch=2/2][2026-10-19 23:34:55,317] loop.py:45 INFO TRAIN_STEP step=15896 loss=1.215029 Training: 88%|████████▊ | 15896/18018 [34:04:41<4:44:40, 8.05s/step, loss=1.215, acc=0.715, acc_len=1.73, thru=12.5, I=12.3, T=34.9, wait=5.8s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:34:55,312] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15895: forward=2601.9ms backward=1030.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:02,443] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:35:03,387] loop.py:45 INFO TRAIN_STEP step=15897 loss=1.204843 Training: 88%|████████▊ | 15897/18018 [34:04:49<4:44:45, 8.06s/step, loss=1.205, acc=0.717, acc_len=1.74, thru=16.9, I=18.0, T=36.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:03,448] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 23:35:09,085] loop.py:45 INFO TRAIN_STEP step=15898 loss=1.137691 Training: 88%|████████▊ | 15898/18018 [34:04:55<4:19:37, 7.35s/step, loss=1.138, acc=0.728, acc_len=1.78, thru=20.9, I=22.2, T=39.5, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:16,974] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:35:17,881] loop.py:45 INFO TRAIN_STEP step=15899 loss=1.245167 Training: 88%|████████▊ | 15899/18018 [34:05:03<4:34:52, 7.78s/step, loss=1.245, acc=0.711, acc_len=1.72, thru=16.6, I=17.5, T=36.0, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:17,984] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-19 23:35:24,073] loop.py:380 INFO TIMING step=15900: step=3.305s data=1.292s compute=3.273s [fwd=2.352s bwd=0.906s opt=0.003s] dispatch=2.877s [2026-10-19 23:35:24,073] loop.py:45 INFO TRAIN_STEP step=15900 loss=1.244196 Training: 88%|████████▊ | 15900/18018 [34:05:10<4:17:53, 7.31s/step, loss=1.244, acc=0.711, acc_len=1.71, thru=17.0, I=20.9, T=38.7, wait=2.9s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:33,057] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:35:33,270] loop.py:45 INFO TRAIN_STEP step=15901 loss=1.171657 Training: 88%|████████▊ | 15901/18018 [34:05:19<4:37:47, 7.87s/step, loss=1.172, acc=0.726, acc_len=1.77, thru=16.7, I=16.5, T=35.8, wait=5.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:35:33,267] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15900: forward=2545.2ms backward=982.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:33,559] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:38,570] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:35:40,087] loop.py:45 INFO TRAIN_STEP step=15902 loss=1.269180 Training: 88%|████████▊ | 15902/18018 [34:05:26<4:26:27, 7.56s/step, loss=1.269, acc=0.704, acc_len=1.69, thru=18.1, I=22.0, T=28.2, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:40,580] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 23:35:46,594] loop.py:45 INFO TRAIN_STEP step=15903 loss=1.170409 Training: 88%|████████▊ | 15903/18018 [34:05:32<4:15:14, 7.24s/step, loss=1.170, acc=0.720, acc_len=1.75, thru=17.4, I=21.5, T=37.5, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:54,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:35:54,467] loop.py:45 INFO TRAIN_STEP step=15904 loss=1.234849 Training: 88%|████████▊ | 15904/18018 [34:05:40<4:21:48, 7.43s/step, loss=1.235, acc=0.709, acc_len=1.71, thru=18.0, I=18.1, T=36.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:35:54,668] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:36:02,744] loop.py:380 INFO TIMING step=15905: step=3.671s data=1.435s compute=3.629s [fwd=2.614s bwd=1.001s opt=0.003s] dispatch=4.599s [2026-10-19 23:36:02,744] loop.py:45 INFO TRAIN_STEP step=15905 loss=1.378335 Training: 88%|████████▊ | 15905/18018 [34:05:48<4:30:38, 7.69s/step, loss=1.378, acc=0.687, acc_len=1.63, thru=14.9, I=15.8, T=34.9, wait=4.6s, pool=53, epoch=2/2][2026-10-19 23:36:09,680] loop.py:45 INFO TRAIN_STEP step=15906 loss=1.250521 Training: 88%|████████▊ | 15906/18018 [34:05:55<4:22:35, 7.46s/step, loss=1.251, acc=0.709, acc_len=1.71, thru=17.3, I=18.5, T=37.6, wait=3.5s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:36:09,676] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15905: forward=2419.4ms backward=945.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:36:18,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:36:18,978] loop.py:45 INFO TRAIN_STEP step=15907 loss=1.331921 Training: 88%|████████▊ | 15907/18018 [34:06:05<4:41:52, 8.01s/step, loss=1.332, acc=0.694, acc_len=1.66, thru=15.3, I=15.2, T=36.0, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:36:19,239] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:36:25,016] loop.py:45 INFO TRAIN_STEP step=15908 loss=1.329549 Training: 88%|████████▊ | 15908/18018 [34:06:11<4:20:55, 7.42s/step, loss=1.330, acc=0.694, acc_len=1.66, thru=19.9, I=19.3, T=37.7, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:36:31,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:36:33,219] loop.py:45 INFO TRAIN_STEP step=15909 loss=1.323733 Training: 88%|████████▊ | 15909/18018 [34:06:19<4:29:04, 7.65s/step, loss=1.324, acc=0.692, acc_len=1.65, thru=17.1, I=18.3, T=32.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:36:33,362] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-19 23:36:41,507] loop.py:380 INFO TIMING step=15910: step=3.668s data=1.469s compute=3.632s [fwd=2.616s bwd=1.003s opt=0.003s] dispatch=4.611s [2026-10-19 23:36:41,507] loop.py:45 INFO TRAIN_STEP step=15910 loss=1.170544 Training: 88%|████████▊ | 15910/18018 [34:06:27<4:35:35, 7.84s/step, loss=1.171, acc=0.727, acc_len=1.78, thru=13.5, I=14.9, T=34.9, wait=4.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:36:47,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-19 23:36:48,993] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15910: forward=2421.1ms backward=955.4ms [2026-10-19 23:36:48,997] loop.py:45 INFO TRAIN_STEP step=15911 loss=1.375648 Training: 88%|████████▊ | 15911/18018 [34:06:35<4:31:43, 7.74s/step, loss=1.376, acc=0.685, acc_len=1.63, thru=19.1, I=18.2, T=37.4, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:36:49,306] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:36:56,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:36:56,736] loop.py:45 INFO TRAIN_STEP step=15912 loss=1.159489 Training: 88%|████████▊ | 15912/18018 [34:06:42<4:31:38, 7.74s/step, loss=1.159, acc=0.723, acc_len=1.76, thru=16.5, I=20.7, T=37.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:36:56,799] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-19 23:37:01,991] loop.py:45 INFO TRAIN_STEP step=15913 loss=1.179767 Training: 88%|████████▊ | 15913/18018 [34:06:48<4:05:21, 6.99s/step, loss=1.180, acc=0.721, acc_len=1.75, thru=19.0, I=19.6, T=40.0, wait=2.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:37:10,430] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:37:11,416] loop.py:45 INFO TRAIN_STEP step=15914 loss=1.287213 Training: 88%|████████▊ | 15914/18018 [34:06:57<4:30:49, 7.72s/step, loss=1.287, acc=0.703, acc_len=1.69, thru=18.4, I=16.7, T=35.5, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:37:11,449] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:37:15,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:37:16,629] loop.py:380 INFO TIMING step=15915: step=3.241s data=1.282s compute=3.214s [fwd=2.303s bwd=0.897s opt=0.003s] dispatch=1.961s [2026-10-19 23:37:16,629] loop.py:45 INFO TRAIN_STEP step=15915 loss=1.282924 Training: 88%|████████▊ | 15915/18018 [34:07:02<4:04:16, 6.97s/step, loss=1.283, acc=0.700, acc_len=1.68, thru=22.4, I=27.2, T=39.5, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:37:16,893] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:37:22,283] loop.py:45 INFO TRAIN_STEP step=15916 loss=1.283029 Training: 88%|████████▊ | 15916/18018 [34:07:08<3:50:20, 6.57s/step, loss=1.283, acc=0.706, acc_len=1.69, thru=21.7, I=25.3, T=40.4, wait=2.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:37:22,279] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15915: forward=2260.1ms backward=865.5ms [2026-10-19 23:37:34,620] loop.py:45 INFO TRAIN_STEP step=15917 loss=1.319955 Training: 88%|████████▊ | 15917/18018 [34:07:20<4:50:45, 8.30s/step, loss=1.320, acc=0.696, acc_len=1.67, thru=12.1, I=10.3, T=33.7, wait=8.5s, pool=45, epoch=2/2][2026-10-19 23:37:42,732] loop.py:45 INFO TRAIN_STEP step=15918 loss=1.285189 Training: 88%|████████▊ | 15918/18018 [34:07:28<4:48:37, 8.25s/step, loss=1.285, acc=0.703, acc_len=1.68, thru=13.2, I=13.4, T=36.8, wait=4.6s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:37:48,630] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:37:50,111] loop.py:45 INFO TRAIN_STEP step=15919 loss=1.218685 Training: 88%|████████▊ | 15919/18018 [34:07:36<4:39:23, 7.99s/step, loss=1.219, acc=0.714, acc_len=1.73, thru=19.9, I=18.6, T=34.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:37:50,146] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-19 23:37:55,499] loop.py:380 INFO TIMING step=15920: step=3.216s data=1.229s compute=3.188s [fwd=2.317s bwd=0.857s opt=0.003s] dispatch=2.162s [2026-10-19 23:37:55,499] loop.py:45 INFO TRAIN_STEP step=15920 loss=1.202692 Training: 88%|████████▊ | 15920/18018 [34:07:41<4:12:00, 7.21s/step, loss=1.203, acc=0.713, acc_len=1.72, thru=18.9, I=24.2, T=39.8, wait=2.2s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:38:03,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:38:04,277] loop.py:45 INFO TRAIN_STEP step=15921 loss=1.279407 Training: 88%|████████▊ | 15921/18018 [34:07:50<4:28:19, 7.68s/step, loss=1.279, acc=0.702, acc_len=1.69, thru=17.5, I=16.5, T=37.8, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:38:04,272] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15920: forward=2419.2ms backward=925.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:38:04,769] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:38:14,186] loop.py:45 INFO TRAIN_STEP step=15922 loss=1.274741 Training: 88%|████████▊ | 15922/18018 [34:08:00<4:51:36, 8.35s/step, loss=1.275, acc=0.703, acc_len=1.69, thru=12.3, I=11.9, T=34.0, wait=6.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:38:19,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:38:20,870] loop.py:45 INFO TRAIN_STEP step=15923 loss=1.275422 Training: 88%|████████▊ | 15923/18018 [34:08:06<4:34:02, 7.85s/step, loss=1.275, acc=0.703, acc_len=1.69, thru=19.6, I=20.2, T=36.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:38:21,312] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:38:28,449] loop.py:45 INFO TRAIN_STEP step=15924 loss=1.309923 Training: 88%|████████▊ | 15924/18018 [34:08:14<4:31:05, 7.77s/step, loss=1.310, acc=0.698, acc_len=1.67, thru=14.6, I=17.5, T=37.4, wait=4.1s, pool=45, epoch=2/2][2026-10-19 23:38:37,231] loop.py:380 INFO TIMING step=15925: step=3.518s data=1.438s compute=3.464s [fwd=2.458s bwd=0.992s opt=0.003s] dispatch=5.252s [2026-10-19 23:38:37,231] loop.py:45 INFO TRAIN_STEP step=15925 loss=1.428937 Training: 88%|████████▊ | 15925/18018 [34:08:23<4:41:33, 8.07s/step, loss=1.429, acc=0.678, acc_len=1.61, thru=15.5, I=15.9, T=36.4, wait=5.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:38:43,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:38:44,094] loop.py:45 INFO TRAIN_STEP step=15926 loss=1.296588 Training: 88%|████████▊ | 15926/18018 [34:08:30<4:28:48, 7.71s/step, loss=1.297, acc=0.699, acc_len=1.68, thru=19.2, I=20.4, T=36.2, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:38:44,090] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15925: forward=2467.0ms backward=1007.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:38:44,242] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:38:48,558] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:38:49,592] loop.py:45 INFO TRAIN_STEP step=15927 loss=1.255508 Training: 88%|████████▊ | 15927/18018 [34:08:35<4:05:32, 7.05s/step, loss=1.256, acc=0.709, acc_len=1.71, thru=22.0, I=26.3, T=37.8, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:38:50,068] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:38:57,900] loop.py:45 INFO TRAIN_STEP step=15928 loss=1.253377 Training: 88%|████████▊ | 15928/18018 [34:08:43<4:18:36, 7.42s/step, loss=1.253, acc=0.708, acc_len=1.71, thru=14.0, I=19.1, T=37.3, wait=4.9s, pool=49, epoch=2/2][2026-10-19 23:39:06,383] loop.py:45 INFO TRAIN_STEP step=15929 loss=1.332990 Training: 88%|████████▊ | 15929/18018 [34:08:52<4:29:32, 7.74s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=14.9, I=15.1, T=35.3, wait=4.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:39:16,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:39:18,097] loop.py:380 INFO TIMING step=15930: step=3.691s data=1.520s compute=3.662s [fwd=2.675s bwd=0.972s opt=0.003s] dispatch=8.015s [2026-10-19 23:39:18,098] loop.py:45 INFO TRAIN_STEP step=15930 loss=1.321802 Training: 88%|████████▊ | 15930/18018 [34:09:04<5:10:53, 8.93s/step, loss=1.322, acc=0.696, acc_len=1.67, thru=12.8, I=12.8, T=34.7, wait=8.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:39:18,194] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-19 23:39:24,965] loop.py:45 INFO TRAIN_STEP step=15931 loss=1.217012 Training: 88%|████████▊ | 15931/18018 [34:09:11<4:49:10, 8.31s/step, loss=1.217, acc=0.714, acc_len=1.73, thru=17.2, I=20.6, T=39.0, wait=3.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:39:24,962] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15930: forward=2332.8ms backward=901.9ms [2026-10-19 23:39:31,834] loop.py:45 INFO TRAIN_STEP step=15932 loss=1.262171 Training: 88%|████████▊ | 15932/18018 [34:09:17<4:33:58, 7.88s/step, loss=1.262, acc=0.710, acc_len=1.71, thru=18.7, I=19.0, T=37.2, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:39:37,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:39:38,486] loop.py:45 INFO TRAIN_STEP step=15933 loss=1.280107 Training: 88%|████████▊ | 15933/18018 [34:09:24<4:21:02, 7.51s/step, loss=1.280, acc=0.705, acc_len=1.70, thru=19.6, I=21.5, T=36.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:39:38,591] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:39:44,563] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:39:45,591] loop.py:45 INFO TRAIN_STEP step=15934 loss=1.320780 Training: 88%|████████▊ | 15934/18018 [34:09:31<4:16:41, 7.39s/step, loss=1.321, acc=0.697, acc_len=1.67, thru=17.4, I=20.6, T=37.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:39:46,073] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:39:52,121] loop.py:380 INFO TIMING step=15935: step=3.310s data=1.352s compute=3.282s [fwd=2.362s bwd=0.905s opt=0.003s] dispatch=3.210s [2026-10-19 23:39:52,121] loop.py:45 INFO TRAIN_STEP step=15935 loss=1.231189 Training: 88%|████████▊ | 15935/18018 [34:09:38<4:07:34, 7.13s/step, loss=1.231, acc=0.710, acc_len=1.71, thru=18.9, I=19.8, T=38.7, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:39:59,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:40:00,228] loop.py:45 INFO TRAIN_STEP step=15936 loss=1.188614 Training: 88%|████████▊ | 15936/18018 [34:09:46<4:17:38, 7.42s/step, loss=1.189, acc=0.720, acc_len=1.75, thru=17.6, I=17.2, T=34.3, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:40:00,223] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15935: forward=2442.8ms backward=1040.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:40:00,590] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:40:08,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:40:09,581] loop.py:45 INFO TRAIN_STEP step=15937 loss=1.261098 Training: 88%|████████▊ | 15937/18018 [34:09:55<4:37:35, 8.00s/step, loss=1.261, acc=0.707, acc_len=1.70, thru=14.2, I=15.2, T=34.2, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:40:09,712] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:40:15,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:40:16,033] loop.py:45 INFO TRAIN_STEP step=15938 loss=1.151204 Training: 88%|████████▊ | 15938/18018 [34:10:02<4:21:18, 7.54s/step, loss=1.151, acc=0.725, acc_len=1.77, thru=19.0, I=20.2, T=39.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:40:16,452] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:40:23,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:40:24,134] loop.py:45 INFO TRAIN_STEP step=15939 loss=1.254119 Training: 88%|████████▊ | 15939/18018 [34:10:10<4:27:02, 7.71s/step, loss=1.254, acc=0.713, acc_len=1.72, thru=16.4, I=18.7, T=33.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:40:24,292] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:40:30,481] loop.py:380 INFO TIMING step=15940: step=3.410s data=1.397s compute=3.371s [fwd=2.405s bwd=0.951s opt=0.003s] dispatch=2.927s [2026-10-19 23:40:30,481] loop.py:45 INFO TRAIN_STEP step=15940 loss=1.195976 Training: 88%|████████▊ | 15940/18018 [34:10:16<4:12:46, 7.30s/step, loss=1.196, acc=0.720, acc_len=1.75, thru=18.0, I=22.5, T=37.5, wait=2.9s, pool=57, epoch=2/2][2026-10-19 23:40:38,988] loop.py:45 INFO TRAIN_STEP step=15941 loss=1.404919 Training: 88%|████████▊ | 15941/18018 [34:10:25<4:25:12, 7.66s/step, loss=1.405, acc=0.680, acc_len=1.61, thru=14.8, I=15.9, T=38.0, wait=5.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:40:38,983] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15940: forward=2401.8ms backward=921.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:40:46,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:40:47,075] loop.py:45 INFO TRAIN_STEP step=15942 loss=1.246481 Training: 88%|████████▊ | 15942/18018 [34:10:33<4:29:30, 7.79s/step, loss=1.246, acc=0.710, acc_len=1.72, thru=16.4, I=16.6, T=37.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:40:47,239] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:40:53,620] loop.py:45 INFO TRAIN_STEP step=15943 loss=1.296874 Training: 88%|████████▊ | 15943/18018 [34:10:39<4:16:27, 7.42s/step, loss=1.297, acc=0.699, acc_len=1.68, thru=17.2, I=18.6, T=36.0, wait=3.0s, pool=45, epoch=2/2][2026-10-19 23:41:04,975] loop.py:45 INFO TRAIN_STEP step=15944 loss=1.270339 Training: 88%|████████▊ | 15944/18018 [34:10:51<4:57:09, 8.60s/step, loss=1.270, acc=0.707, acc_len=1.71, thru=16.0, I=11.8, T=34.3, wait=7.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:10,578] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:41:11,116] loop.py:380 INFO TIMING step=15945: step=3.302s data=1.296s compute=3.266s [fwd=2.334s bwd=0.918s opt=0.003s] dispatch=2.832s [2026-10-19 23:41:11,116] loop.py:45 INFO TRAIN_STEP step=15945 loss=1.270890 Training: 88%|████████▊ | 15945/18018 [34:10:57<4:31:35, 7.86s/step, loss=1.271, acc=0.707, acc_len=1.71, thru=19.8, I=20.3, T=38.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:11,583] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:17,148] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:41:17,727] loop.py:45 INFO TRAIN_STEP step=15946 loss=1.236940 Training: 89%|████████▊ | 15946/18018 [34:11:03<4:18:30, 7.49s/step, loss=1.237, acc=0.708, acc_len=1.71, thru=19.7, I=22.2, T=38.1, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:41:17,723] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15945: forward=2393.9ms backward=927.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:18,154] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:24,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:41:24,751] loop.py:45 INFO TRAIN_STEP step=15947 loss=1.331036 Training: 89%|████████▊ | 15947/18018 [34:11:10<4:13:35, 7.35s/step, loss=1.331, acc=0.693, acc_len=1.66, thru=18.6, I=20.1, T=37.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:25,016] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:41:30,347] loop.py:45 INFO TRAIN_STEP step=15948 loss=1.168239 Training: 89%|████████▊ | 15948/18018 [34:11:16<3:55:20, 6.82s/step, loss=1.168, acc=0.723, acc_len=1.77, thru=20.6, I=21.7, T=39.0, wait=2.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:37,139] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:41:38,046] loop.py:45 INFO TRAIN_STEP step=15949 loss=1.239388 Training: 89%|████████▊ | 15949/18018 [34:11:24<4:04:19, 7.09s/step, loss=1.239, acc=0.708, acc_len=1.71, thru=17.8, I=18.9, T=36.4, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:38,148] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:44,880] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:41:45,375] loop.py:380 INFO TIMING step=15950: step=3.419s data=1.334s compute=3.377s [fwd=2.376s bwd=0.986s opt=0.003s] dispatch=3.901s [2026-10-19 23:41:45,375] loop.py:45 INFO TRAIN_STEP step=15950 loss=1.287258 Training: 89%|████████▊ | 15950/18018 [34:11:31<4:06:43, 7.16s/step, loss=1.287, acc=0.699, acc_len=1.68, thru=18.4, I=19.9, T=37.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:45,410] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:50,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:41:50,521] loop.py:45 INFO TRAIN_STEP step=15951 loss=1.132317 Training: 89%|████████▊ | 15951/18018 [34:11:36<3:45:47, 6.55s/step, loss=1.132, acc=0.731, acc_len=1.79, thru=22.4, I=24.8, T=38.8, wait=1.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:41:50,516] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15950: forward=2365.6ms backward=896.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:50,746] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:58,672] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:41:59,403] loop.py:45 INFO TRAIN_STEP step=15952 loss=1.315742 Training: 89%|████████▊ | 15952/18018 [34:11:45<4:09:43, 7.25s/step, loss=1.316, acc=0.696, acc_len=1.66, thru=14.4, I=15.8, T=35.8, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:41:59,675] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:04,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:42:05,491] loop.py:45 INFO TRAIN_STEP step=15953 loss=1.208507 Training: 89%|████████▊ | 15953/18018 [34:11:51<3:57:35, 6.90s/step, loss=1.209, acc=0.715, acc_len=1.74, thru=20.3, I=23.3, T=38.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:05,518] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:14,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:42:14,652] loop.py:45 INFO TRAIN_STEP step=15954 loss=1.219342 Training: 89%|████████▊ | 15954/18018 [34:12:00<4:20:47, 7.58s/step, loss=1.219, acc=0.715, acc_len=1.73, thru=14.0, I=14.7, T=35.3, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:14,805] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:42:22,275] loop.py:380 INFO TIMING step=15955: step=3.387s data=1.387s compute=3.354s [fwd=2.357s bwd=0.983s opt=0.003s] dispatch=4.227s [2026-10-19 23:42:22,275] loop.py:45 INFO TRAIN_STEP step=15955 loss=1.222041 Training: 89%|████████▊ | 15955/18018 [34:12:08<4:21:04, 7.59s/step, loss=1.222, acc=0.714, acc_len=1.73, thru=12.9, I=15.6, T=37.8, wait=4.2s, pool=25, epoch=2/2](TrainerActor pid=9370) [2026-10-19 23:42:22,328] eagle_store.py:517 WARNING batch_get_buffer missing keys (k510558-9e01b512_hs, k510558-9e01b512_ids, k510558-9e01b512_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:33,256] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 23:42:22,607] eagle_store.py:517 WARNING batch_get_buffer missing keys (k510559-9fc74b7f_hs, k510559-9fc74b7f_ids, k510559-9fc74b7f_lhs); sleeping 0.50s. [repeated 6x across cluster] [2026-10-19 23:42:33,403] loop.py:45 INFO TRAIN_STEP step=15956 loss=1.216837 Training: 89%|████████▊ | 15956/18018 [34:12:19<4:57:23, 8.65s/step, loss=1.217, acc=0.719, acc_len=1.75, thru=16.5, I=13.2, T=35.6, wait=7.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:42:33,398] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15955: forward=2516.3ms backward=999.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:33,757] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:42:40,976] loop.py:45 INFO TRAIN_STEP step=15957 loss=1.233842 Training: 89%|████████▊ | 15957/18018 [34:12:27<4:46:08, 8.33s/step, loss=1.234, acc=0.714, acc_len=1.73, thru=17.1, I=18.7, T=36.5, wait=4.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:48,289] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:42:48,386] loop.py:45 INFO TRAIN_STEP step=15958 loss=1.216189 Training: 89%|████████▊ | 15958/18018 [34:12:34<4:36:29, 8.05s/step, loss=1.216, acc=0.714, acc_len=1.74, thru=18.0, I=17.9, T=37.9, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:48,790] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:54,357] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:42:55,399] loop.py:45 INFO TRAIN_STEP step=15959 loss=1.284322 Training: 89%|████████▊ | 15959/18018 [34:12:41<4:25:40, 7.74s/step, loss=1.284, acc=0.702, acc_len=1.69, thru=18.1, I=21.8, T=36.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:42:55,866] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:03,004] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:43:03,494] loop.py:380 INFO TIMING step=15960: step=3.528s data=1.453s compute=3.496s [fwd=2.527s bwd=0.954s opt=0.003s] dispatch=4.558s [2026-10-19 23:43:03,495] loop.py:45 INFO TRAIN_STEP step=15960 loss=1.147679 Training: 89%|████████▊ | 15960/18018 [34:12:49<4:29:11, 7.85s/step, loss=1.148, acc=0.728, acc_len=1.77, thru=15.5, I=18.1, T=36.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:03,526] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 23:43:09,168] loop.py:45 INFO TRAIN_STEP step=15961 loss=1.308027 Training: 89%|████████▊ | 15961/18018 [34:12:55<4:06:40, 7.20s/step, loss=1.308, acc=0.698, acc_len=1.67, thru=19.3, I=20.7, T=38.2, wait=2.3s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:43:09,164] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15960: forward=2408.3ms backward=897.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:16,798] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:43:17,749] loop.py:45 INFO TRAIN_STEP step=15962 loss=1.259508 Training: 89%|████████▊ | 15962/18018 [34:13:03<4:20:48, 7.61s/step, loss=1.260, acc=0.709, acc_len=1.71, thru=17.0, I=17.3, T=34.3, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:17,809] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-19 23:43:24,005] loop.py:45 INFO TRAIN_STEP step=15963 loss=1.216977 Training: 89%|████████▊ | 15963/18018 [34:13:10<4:06:45, 7.20s/step, loss=1.217, acc=0.713, acc_len=1.73, thru=19.2, I=18.8, T=38.1, wait=2.9s, pool=61, epoch=2/2][2026-10-19 23:43:31,013] loop.py:45 INFO TRAIN_STEP step=15964 loss=1.302914 Training: 89%|████████▊ | 15964/18018 [34:13:17<4:04:37, 7.15s/step, loss=1.303, acc=0.698, acc_len=1.67, thru=14.4, I=16.0, T=37.9, wait=3.6s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:38,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:43:39,924] loop.py:380 INFO TIMING step=15965: step=3.611s data=1.423s compute=3.579s [fwd=2.559s bwd=1.006s opt=0.003s] dispatch=5.291s [2026-10-19 23:43:39,924] loop.py:45 INFO TRAIN_STEP step=15965 loss=1.230412 Training: 89%|████████▊ | 15965/18018 [34:13:26<4:22:36, 7.67s/step, loss=1.230, acc=0.713, acc_len=1.72, thru=17.6, I=17.9, T=35.5, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:40,400] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:43:45,771] loop.py:45 INFO TRAIN_STEP step=15966 loss=1.272349 Training: 89%|████████▊ | 15966/18018 [34:13:31<4:03:43, 7.13s/step, loss=1.272, acc=0.702, acc_len=1.68, thru=18.4, I=21.9, T=40.1, wait=2.6s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:43:45,767] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15965: forward=2283.7ms backward=871.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:51,891] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:43:52,940] loop.py:45 INFO TRAIN_STEP step=15967 loss=1.044053 Training: 89%|████████▊ | 15967/18018 [34:13:39<4:04:02, 7.14s/step, loss=1.044, acc=0.743, acc_len=1.83, thru=18.6, I=19.7, T=36.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:53,401] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:58,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:43:58,914] loop.py:45 INFO TRAIN_STEP step=15968 loss=1.218555 Training: 89%|████████▊ | 15968/18018 [34:13:44<3:51:58, 6.79s/step, loss=1.219, acc=0.714, acc_len=1.73, thru=21.0, I=22.4, T=37.4, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:43:59,343] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:44:04,939] loop.py:45 INFO TRAIN_STEP step=15969 loss=1.261265 Training: 89%|████████▊ | 15969/18018 [34:13:51<3:44:03, 6.56s/step, loss=1.261, acc=0.706, acc_len=1.70, thru=20.7, I=21.4, T=39.0, wait=2.7s, pool=61, epoch=2/2][2026-10-19 23:44:11,271] loop.py:380 INFO TIMING step=15970: step=3.439s data=1.424s compute=3.409s [fwd=2.422s bwd=0.973s opt=0.003s] dispatch=2.884s [2026-10-19 23:44:11,272] loop.py:45 INFO TRAIN_STEP step=15970 loss=1.304084 Training: 89%|████████▊ | 15970/18018 [34:13:57<3:41:36, 6.49s/step, loss=1.304, acc=0.699, acc_len=1.67, thru=19.6, I=21.1, T=37.2, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:44:17,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:44:17,978] loop.py:45 INFO TRAIN_STEP step=15971 loss=1.256869 Training: 89%|████████▊ | 15971/18018 [34:14:04<3:43:40, 6.56s/step, loss=1.257, acc=0.707, acc_len=1.71, thru=19.5, I=19.8, T=37.1, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:44:17,973] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15970: forward=2444.5ms backward=955.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:44:18,100] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:44:23,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:44:24,148] loop.py:45 INFO TRAIN_STEP step=15972 loss=1.290037 Training: 89%|████████▊ | 15972/18018 [34:14:10<3:39:36, 6.44s/step, loss=1.290, acc=0.701, acc_len=1.68, thru=20.4, I=21.5, T=37.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:44:24,343] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:44:30,952] loop.py:45 INFO TRAIN_STEP step=15973 loss=1.155219 Training: 89%|████████▊ | 15973/18018 [34:14:17<3:43:13, 6.55s/step, loss=1.155, acc=0.726, acc_len=1.77, thru=16.5, I=17.3, T=37.8, wait=3.4s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:44:42,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:44:43,059] loop.py:45 INFO TRAIN_STEP step=15974 loss=1.388793 Training: 89%|████████▊ | 15974/18018 [34:14:29<4:39:55, 8.22s/step, loss=1.389, acc=0.682, acc_len=1.62, thru=14.2, I=14.7, T=33.6, wait=8.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:44:43,088] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 23:44:49,415] loop.py:380 INFO TIMING step=15975: step=3.248s data=1.324s compute=3.220s [fwd=2.326s bwd=0.880s opt=0.003s] dispatch=3.101s [2026-10-19 23:44:49,415] loop.py:45 INFO TRAIN_STEP step=15975 loss=1.151612 Training: 89%|████████▊ | 15975/18018 [34:14:35<4:20:47, 7.66s/step, loss=1.152, acc=0.729, acc_len=1.78, thru=16.9, I=19.8, T=39.4, wait=3.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:44:56,525] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:44:57,009] loop.py:45 INFO TRAIN_STEP step=15976 loss=1.346073 Training: 89%|████████▊ | 15976/18018 [34:14:43<4:20:00, 7.64s/step, loss=1.346, acc=0.688, acc_len=1.64, thru=18.6, I=19.2, T=37.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:44:57,042] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 23:44:57,005] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15975: forward=2432.1ms backward=935.6ms [2026-10-19 23:45:03,336] loop.py:45 INFO TRAIN_STEP step=15977 loss=1.159914 Training: 89%|████████▊ | 15977/18018 [34:14:49<4:06:27, 7.25s/step, loss=1.160, acc=0.723, acc_len=1.76, thru=19.1, I=19.9, T=38.3, wait=3.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:45:12,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:45:12,764] loop.py:45 INFO TRAIN_STEP step=15978 loss=1.360083 Training: 89%|████████▊ | 15978/18018 [34:14:58<4:28:37, 7.90s/step, loss=1.360, acc=0.689, acc_len=1.64, thru=15.7, I=15.8, T=34.3, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:45:13,106] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:45:19,269] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:45:19,998] loop.py:45 INFO TRAIN_STEP step=15979 loss=1.185879 Training: 89%|████████▊ | 15979/18018 [34:15:06<4:21:41, 7.70s/step, loss=1.186, acc=0.721, acc_len=1.75, thru=17.6, I=20.9, T=37.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:45:20,275] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:45:29,257] loop.py:380 INFO TIMING step=15980: step=3.427s data=1.340s compute=3.382s [fwd=2.418s bwd=0.948s opt=0.003s] dispatch=5.821s [2026-10-19 23:45:29,257] loop.py:45 INFO TRAIN_STEP step=15980 loss=1.138014 Training: 89%|████████▊ | 15980/18018 [34:15:15<4:37:26, 8.17s/step, loss=1.138, acc=0.730, acc_len=1.78, thru=11.6, I=12.7, T=37.3, wait=5.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:45:36,621] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:45:37,679] loop.py:45 INFO TRAIN_STEP step=15981 loss=1.225764 Training: 89%|████████▊ | 15981/18018 [34:15:23<4:39:53, 8.24s/step, loss=1.226, acc=0.714, acc_len=1.72, thru=17.4, I=17.8, T=36.6, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:45:37,675] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15980: forward=2531.9ms backward=921.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:45:38,129] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:45:45,009] loop.py:45 INFO TRAIN_STEP step=15982 loss=1.322486 Training: 89%|████████▊ | 15982/18018 [34:15:31<4:30:26, 7.97s/step, loss=1.322, acc=0.694, acc_len=1.66, thru=16.4, I=20.5, T=38.0, wait=4.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:45:50,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:45:51,676] loop.py:45 INFO TRAIN_STEP step=15983 loss=1.199145 Training: 89%|████████▊ | 15983/18018 [34:15:37<4:17:02, 7.58s/step, loss=1.199, acc=0.719, acc_len=1.74, thru=19.3, I=22.3, T=38.3, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:45:51,869] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:45:57,762] loop.py:45 INFO TRAIN_STEP step=15984 loss=1.239605 Training: 89%|████████▊ | 15984/18018 [34:15:43<4:01:45, 7.13s/step, loss=1.240, acc=0.708, acc_len=1.71, thru=20.2, I=23.8, T=39.2, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:05,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:46:06,210] loop.py:380 INFO TIMING step=15985: step=3.458s data=1.391s compute=3.425s [fwd=2.435s bwd=0.976s opt=0.003s] dispatch=4.980s [2026-10-19 23:46:06,211] loop.py:45 INFO TRAIN_STEP step=15985 loss=1.319364 Training: 89%|████████▊ | 15985/18018 [34:15:52<4:15:01, 7.53s/step, loss=1.319, acc=0.696, acc_len=1.67, thru=15.7, I=16.3, T=37.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:06,401] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:46:12,014] loop.py:45 INFO TRAIN_STEP step=15986 loss=1.202740 Training: 89%|████████▊ | 15986/18018 [34:15:58<3:57:23, 7.01s/step, loss=1.203, acc=0.717, acc_len=1.74, thru=18.1, I=19.5, T=35.4, wait=2.2s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:46:12,008] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15985: forward=2673.9ms backward=898.9ms [2026-10-19 23:46:24,044] loop.py:45 INFO TRAIN_STEP step=15987 loss=1.300350 Training: 89%|████████▊ | 15987/18018 [34:16:10<4:48:14, 8.52s/step, loss=1.300, acc=0.701, acc_len=1.69, thru=12.6, I=12.5, T=32.9, wait=8.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:30,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:46:30,758] loop.py:45 INFO TRAIN_STEP step=15988 loss=1.157840 Training: 89%|████████▊ | 15988/18018 [34:16:16<4:29:50, 7.98s/step, loss=1.158, acc=0.727, acc_len=1.77, thru=18.8, I=19.2, T=35.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:31,180] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:37,345] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:46:38,023] loop.py:45 INFO TRAIN_STEP step=15989 loss=1.193502 Training: 89%|████████▊ | 15989/18018 [34:16:24<4:22:29, 7.76s/step, loss=1.194, acc=0.716, acc_len=1.73, thru=18.4, I=21.1, T=37.6, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:38,349] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:44,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:46:45,217] loop.py:380 INFO TIMING step=15990: step=3.411s data=1.414s compute=3.379s [fwd=2.419s bwd=0.946s opt=0.003s] dispatch=3.773s [2026-10-19 23:46:45,217] loop.py:45 INFO TRAIN_STEP step=15990 loss=1.363863 Training: 89%|████████▊ | 15990/18018 [34:16:31<4:16:35, 7.59s/step, loss=1.364, acc=0.687, acc_len=1.63, thru=18.1, I=20.0, T=37.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:45,378] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:52,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:46:52,691] loop.py:45 INFO TRAIN_STEP step=15991 loss=1.257073 Training: 89%|████████▉ | 15991/18018 [34:16:38<4:15:16, 7.56s/step, loss=1.257, acc=0.703, acc_len=1.68, thru=18.6, I=19.8, T=37.1, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:46:52,687] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15990: forward=2432.4ms backward=975.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:46:53,157] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:46:59,047] loop.py:45 INFO TRAIN_STEP step=15992 loss=1.289105 Training: 89%|████████▉ | 15992/18018 [34:16:45<4:02:59, 7.20s/step, loss=1.289, acc=0.703, acc_len=1.69, thru=19.8, I=20.1, T=39.3, wait=3.1s, pool=61, epoch=2/2][2026-10-19 23:47:05,977] loop.py:45 INFO TRAIN_STEP step=15993 loss=1.300152 Training: 89%|████████▉ | 15993/18018 [34:16:52<4:00:10, 7.12s/step, loss=1.300, acc=0.699, acc_len=1.68, thru=18.6, I=19.2, T=38.5, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:47:11,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:47:12,961] loop.py:45 INFO TRAIN_STEP step=15994 loss=1.275391 Training: 89%|████████▉ | 15994/18018 [34:16:59<3:58:42, 7.08s/step, loss=1.275, acc=0.707, acc_len=1.70, thru=19.2, I=20.3, T=38.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:47:13,350] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:47:22,225] loop.py:380 INFO TIMING step=15995: step=3.742s data=1.528s compute=3.716s [fwd=2.628s bwd=1.073s opt=0.003s] dispatch=5.513s [2026-10-19 23:47:22,225] loop.py:45 INFO TRAIN_STEP step=15995 loss=1.232886 Training: 89%|████████▉ | 15995/18018 [34:17:08<4:20:44, 7.73s/step, loss=1.233, acc=0.713, acc_len=1.71, thru=10.8, I=12.3, T=34.2, wait=5.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:47:30,876] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:47:31,355] loop.py:45 INFO TRAIN_STEP step=15996 loss=1.262481 Training: 89%|████████▉ | 15996/18018 [34:17:17<4:34:44, 8.15s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=16.7, I=15.5, T=36.6, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:47:31,387] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-19 23:47:31,350] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=15995: forward=2435.3ms backward=996.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:47:38,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:47:38,664] loop.py:45 INFO TRAIN_STEP step=15997 loss=1.270853 Training: 89%|████████▉ | 15997/18018 [34:17:24<4:26:04, 7.90s/step, loss=1.271, acc=0.702, acc_len=1.69, thru=17.9, I=18.3, T=36.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:47:38,775] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-19 23:47:46,229] loop.py:45 INFO TRAIN_STEP step=15998 loss=1.332576 Training: 89%|████████▉ | 15998/18018 [34:17:32<4:22:33, 7.80s/step, loss=1.333, acc=0.695, acc_len=1.67, thru=13.2, I=14.6, T=36.9, wait=4.1s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 23:47:46,604] eagle_store.py:517 WARNING batch_get_buffer missing keys (k511938-95408efc_hs, k511938-95408efc_ids, k511938-95408efc_lhs); sleeping 0.50s. [2026-10-19 23:47:53,596] loop.py:45 INFO TRAIN_STEP step=15999 loss=1.270479 Training: 89%|████████▉ | 15999/18018 [34:17:39<4:18:03, 7.67s/step, loss=1.270, acc=0.709, acc_len=1.70, thru=18.3, I=16.6, T=38.2, wait=4.0s, pool=57, epoch=2/2][2026-10-19 23:48:01,278] loop.py:380 INFO TIMING step=16000: step=3.659s data=1.645s compute=3.626s [fwd=2.707s bwd=0.905s opt=0.003s] dispatch=4.015s [2026-10-19 23:48:01,279] loop.py:45 INFO TRAIN_STEP step=16000 loss=1.259182 Training: 89%|████████▉ | 16000/18018 [34:17:47<4:18:04, 7.67s/step, loss=1.259, acc=0.705, acc_len=1.70, thru=16.0, I=16.6, T=35.0, wait=4.0s, pool=53, epoch=2/2][2026-10-19 23:48:01,281] loop.py:419 INFO Saving checkpoint at step 16000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:01,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-19 23:47:46,729] eagle_store.py:517 WARNING batch_get_buffer missing keys (k511937-ad9ed262_hs, k511937-ad9ed262_ids, k511937-ad9ed262_lhs); sleeping 0.50s. [repeated 2x across cluster] (TrainerActor pid=9310) [2026-10-19 23:48:05,386] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0016001 (TrainerActor pid=9310) [2026-10-19 23:48:05,598] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0015251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:06,358] inference_manager.py:408 INFO Pool capacity available after 4.5s, pool_size=17/64 (TrainerActor pid=9310) [2026-10-19 23:48:13,397] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16000: forward=2557.1ms backward=1103.6ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 23:48:05,387] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0016001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-19 23:48:06,239] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0015251 [2026-10-19 23:48:13,401] loop.py:45 INFO TRAIN_STEP step=16001 loss=1.193724 Training: 89%|████████▉ | 16001/18018 [34:17:59<5:02:48, 9.01s/step, loss=1.194, acc=0.718, acc_len=1.74, thru=14.2, I=12.7, T=34.3, wait=3.4s, pool=37, epoch=2/2][2026-10-19 23:48:22,850] loop.py:45 INFO TRAIN_STEP step=16002 loss=1.200492 Training: 89%|████████▉ | 16002/18018 [34:18:08<5:07:06, 9.14s/step, loss=1.200, acc=0.719, acc_len=1.75, thru=17.0, I=14.3, T=34.2, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:22,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:23,330] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:48:30,476] loop.py:45 INFO TRAIN_STEP step=16003 loss=1.205207 Training: 89%|████████▉ | 16003/18018 [34:18:16<4:51:41, 8.69s/step, loss=1.205, acc=0.717, acc_len=1.74, thru=16.6, I=16.5, T=36.5, wait=4.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:36,529] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:48:37,544] loop.py:45 INFO TRAIN_STEP step=16004 loss=1.233329 Training: 89%|████████▉ | 16004/18018 [34:18:23<4:35:15, 8.20s/step, loss=1.233, acc=0.711, acc_len=1.72, thru=17.8, I=20.1, T=36.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:38,039] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:43,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:48:43,920] loop.py:380 INFO TIMING step=16005: step=3.552s data=1.620s compute=3.526s [fwd=2.632s bwd=0.880s opt=0.003s] dispatch=2.817s [2026-10-19 23:48:43,920] loop.py:45 INFO TRAIN_STEP step=16005 loss=1.283698 Training: 89%|████████▉ | 16005/18018 [34:18:30<4:16:47, 7.65s/step, loss=1.284, acc=0.706, acc_len=1.70, thru=20.1, I=23.9, T=36.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:44,159] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:48:49,970] loop.py:45 INFO TRAIN_STEP step=16006 loss=1.334378 Training: 89%|████████▉ | 16006/18018 [34:18:36<4:00:30, 7.17s/step, loss=1.334, acc=0.696, acc_len=1.67, thru=17.4, I=19.8, T=39.0, wait=2.8s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:48:49,966] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16005: forward=2347.5ms backward=895.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:55,979] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:48:57,575] loop.py:45 INFO TRAIN_STEP step=16007 loss=1.368410 Training: 89%|████████▉ | 16007/18018 [34:18:43<4:04:45, 7.30s/step, loss=1.368, acc=0.686, acc_len=1.63, thru=20.0, I=19.5, T=35.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:48:57,995] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 23:49:05,384] loop.py:45 INFO TRAIN_STEP step=16008 loss=1.178174 Training: 89%|████████▉ | 16008/18018 [34:18:51<4:09:44, 7.46s/step, loss=1.178, acc=0.717, acc_len=1.74, thru=11.9, I=15.6, T=39.2, wait=4.5s, pool=29, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-19 23:49:05,451] eagle_store.py:517 WARNING batch_get_buffer missing keys (k512258-ba74123d_hs, k512258-ba74123d_ids, k512258-ba74123d_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:49:14,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-19 23:49:05,458] eagle_store.py:517 WARNING batch_get_buffer missing keys (k512256-923948aa_hs, k512256-923948aa_ids, k512256-923948aa_lhs); sleeping 0.50s. [2026-10-19 23:49:15,291] loop.py:45 INFO TRAIN_STEP step=16009 loss=1.379100 Training: 89%|████████▉ | 16009/18018 [34:19:01<4:34:13, 8.19s/step, loss=1.379, acc=0.683, acc_len=1.62, thru=16.9, I=16.3, T=30.4, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:49:15,443] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-19 23:49:22,114] loop.py:380 INFO TIMING step=16010: step=3.385s data=1.348s compute=3.329s [fwd=2.390s bwd=0.926s opt=0.003s] dispatch=3.429s [2026-10-19 23:49:22,115] loop.py:45 INFO TRAIN_STEP step=16010 loss=1.094028 Training: 89%|████████▉ | 16010/18018 [34:19:08<4:20:22, 7.78s/step, loss=1.094, acc=0.737, acc_len=1.81, thru=15.9, I=17.5, T=37.8, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:49:28,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:49:29,229] loop.py:45 INFO TRAIN_STEP step=16011 loss=1.228185 Training: 89%|████████▉ | 16011/18018 [34:19:15<4:13:33, 7.58s/step, loss=1.228, acc=0.711, acc_len=1.72, thru=18.9, I=19.2, T=31.8, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:49:29,225] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16010: forward=3071.1ms backward=907.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:49:29,443] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:49:34,799] loop.py:45 INFO TRAIN_STEP step=16012 loss=1.221667 Training: 89%|████████▉ | 16012/18018 [34:19:20<3:53:14, 6.98s/step, loss=1.222, acc=0.719, acc_len=1.74, thru=20.7, I=22.5, T=39.2, wait=2.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:49:43,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:49:43,603] loop.py:45 INFO TRAIN_STEP step=16013 loss=1.300004 Training: 89%|████████▉ | 16013/18018 [34:19:29<4:11:28, 7.53s/step, loss=1.300, acc=0.701, acc_len=1.68, thru=16.1, I=16.3, T=34.7, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:49:43,967] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:49:50,299] loop.py:45 INFO TRAIN_STEP step=16014 loss=1.205322 Training: 89%|████████▉ | 16014/18018 [34:19:36<4:03:01, 7.28s/step, loss=1.205, acc=0.715, acc_len=1.73, thru=18.1, I=18.5, T=37.1, wait=3.2s, pool=61, epoch=2/2][2026-10-19 23:49:57,184] loop.py:380 INFO TIMING step=16015: step=3.428s data=1.409s compute=3.381s [fwd=2.399s bwd=0.969s opt=0.003s] dispatch=3.448s [2026-10-19 23:49:57,184] loop.py:45 INFO TRAIN_STEP step=16015 loss=1.218043 Training: 89%|████████▉ | 16015/18018 [34:19:43<3:59:00, 7.16s/step, loss=1.218, acc=0.713, acc_len=1.73, thru=18.4, I=17.8, T=37.3, wait=3.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:03,654] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:50:04,546] loop.py:45 INFO TRAIN_STEP step=16016 loss=1.362951 Training: 89%|████████▉ | 16016/18018 [34:19:50<4:00:54, 7.22s/step, loss=1.363, acc=0.691, acc_len=1.65, thru=18.1, I=19.5, T=36.2, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:50:04,541] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16015: forward=2519.8ms backward=974.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:04,666] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 23:50:10,833] loop.py:45 INFO TRAIN_STEP step=16017 loss=1.207806 Training: 89%|████████▉ | 16017/18018 [34:19:56<3:51:25, 6.94s/step, loss=1.208, acc=0.716, acc_len=1.74, thru=17.4, I=20.0, T=39.0, wait=3.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:19,797] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:50:20,803] loop.py:45 INFO TRAIN_STEP step=16018 loss=1.290609 Training: 89%|████████▉ | 16018/18018 [34:20:06<4:21:37, 7.85s/step, loss=1.291, acc=0.703, acc_len=1.69, thru=15.5, I=15.8, T=30.8, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:21,306] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:26,638] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:50:26,839] loop.py:45 INFO TRAIN_STEP step=16019 loss=1.197938 Training: 89%|████████▉ | 16019/18018 [34:20:12<4:03:23, 7.31s/step, loss=1.198, acc=0.718, acc_len=1.75, thru=19.2, I=22.1, T=38.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:27,140] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:32,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:50:33,365] loop.py:380 INFO TIMING step=16020: step=3.416s data=1.370s compute=3.370s [fwd=2.420s bwd=0.935s opt=0.003s] dispatch=3.101s [2026-10-19 23:50:33,366] loop.py:45 INFO TRAIN_STEP step=16020 loss=1.198077 Training: 89%|████████▉ | 16020/18018 [34:20:19<3:55:29, 7.07s/step, loss=1.198, acc=0.718, acc_len=1.74, thru=19.9, I=22.0, T=37.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:33,848] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-19 23:50:39,963] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16020: forward=2410.0ms backward=943.8ms [2026-10-19 23:50:39,967] loop.py:45 INFO TRAIN_STEP step=16021 loss=1.246094 Training: 89%|████████▉ | 16021/18018 [34:20:26<3:50:40, 6.93s/step, loss=1.246, acc=0.710, acc_len=1.71, thru=18.8, I=22.2, T=37.7, wait=3.2s, pool=57, epoch=2/2][2026-10-19 23:50:47,055] loop.py:45 INFO TRAIN_STEP step=16022 loss=1.188745 Training: 89%|████████▉ | 16022/18018 [34:20:33<3:52:09, 6.98s/step, loss=1.189, acc=0.720, acc_len=1.75, thru=18.0, I=20.3, T=31.2, wait=3.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:53,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:50:54,512] loop.py:45 INFO TRAIN_STEP step=16023 loss=1.362256 Training: 89%|████████▉ | 16023/18018 [34:20:40<3:56:47, 7.12s/step, loss=1.362, acc=0.691, acc_len=1.65, thru=15.7, I=17.4, T=35.7, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:50:54,691] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:51:02,135] loop.py:45 INFO TRAIN_STEP step=16024 loss=1.243037 Training: 89%|████████▉ | 16024/18018 [34:20:48<4:01:38, 7.27s/step, loss=1.243, acc=0.713, acc_len=1.72, thru=14.8, I=17.1, T=37.4, wait=4.2s, pool=49, epoch=2/2][2026-10-19 23:51:08,540] loop.py:380 INFO TIMING step=16025: step=3.318s data=1.300s compute=3.288s [fwd=2.385s bwd=0.889s opt=0.003s] dispatch=3.079s [2026-10-19 23:51:08,540] loop.py:45 INFO TRAIN_STEP step=16025 loss=1.294004 Training: 89%|████████▉ | 16025/18018 [34:20:54<3:52:54, 7.01s/step, loss=1.294, acc=0.699, acc_len=1.67, thru=19.1, I=18.9, T=38.6, wait=3.1s, pool=57, epoch=2/2][2026-10-19 23:51:15,483] loop.py:45 INFO TRAIN_STEP step=16026 loss=1.255478 Training: 89%|████████▉ | 16026/18018 [34:21:01<3:52:07, 6.99s/step, loss=1.255, acc=0.706, acc_len=1.70, thru=17.2, I=18.4, T=37.7, wait=3.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:51:15,479] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16025: forward=2429.3ms backward=906.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:51:22,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:51:24,407] loop.py:45 INFO TRAIN_STEP step=16027 loss=1.244776 Training: 89%|████████▉ | 16027/18018 [34:21:10<4:11:12, 7.57s/step, loss=1.245, acc=0.714, acc_len=1.72, thru=15.7, I=18.4, T=31.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:51:24,793] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-19 23:51:30,658] loop.py:45 INFO TRAIN_STEP step=16028 loss=1.246806 Training: 89%|████████▉ | 16028/18018 [34:21:16<3:57:58, 7.17s/step, loss=1.247, acc=0.708, acc_len=1.71, thru=17.8, I=21.4, T=38.0, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:51:37,938] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:51:38,424] loop.py:45 INFO TRAIN_STEP step=16029 loss=1.199037 Training: 89%|████████▉ | 16029/18018 [34:21:24<4:03:43, 7.35s/step, loss=1.199, acc=0.717, acc_len=1.74, thru=18.4, I=18.2, T=37.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:51:38,456] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-19 23:51:44,096] loop.py:380 INFO TIMING step=16030: step=3.263s data=1.349s compute=3.233s [fwd=2.312s bwd=0.907s opt=0.003s] dispatch=2.400s [2026-10-19 23:51:44,097] loop.py:45 INFO TRAIN_STEP step=16030 loss=1.198532 Training: 89%|████████▉ | 16030/18018 [34:21:30<3:46:53, 6.85s/step, loss=1.199, acc=0.718, acc_len=1.74, thru=20.5, I=20.7, T=39.2, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:51:50,249] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:51:51,248] loop.py:45 INFO TRAIN_STEP step=16031 loss=1.212739 Training: 89%|████████▉ | 16031/18018 [34:21:37<3:49:48, 6.94s/step, loss=1.213, acc=0.715, acc_len=1.73, thru=18.4, I=20.4, T=38.3, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:51:51,244] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16030: forward=2386.3ms backward=912.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:51:51,276] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:51:55,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:51:56,579] loop.py:45 INFO TRAIN_STEP step=16032 loss=1.299175 Training: 89%|████████▉ | 16032/18018 [34:21:42<3:33:43, 6.46s/step, loss=1.299, acc=0.700, acc_len=1.68, thru=22.8, I=29.3, T=39.3, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:51:56,897] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:52:02,030] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:52:03,043] loop.py:45 INFO TRAIN_STEP step=16033 loss=1.219185 Training: 89%|████████▉ | 16033/18018 [34:21:49<3:33:40, 6.46s/step, loss=1.219, acc=0.714, acc_len=1.73, thru=20.1, I=26.7, T=39.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:52:03,540] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:52:09,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:52:09,713] loop.py:45 INFO TRAIN_STEP step=16034 loss=1.336493 Training: 89%|████████▉ | 16034/18018 [34:21:55<3:35:41, 6.52s/step, loss=1.336, acc=0.695, acc_len=1.66, thru=19.6, I=24.7, T=38.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:52:10,017] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:52:16,040] loop.py:380 INFO TIMING step=16035: step=3.271s data=1.310s compute=3.242s [fwd=2.350s bwd=0.878s opt=0.003s] dispatch=3.047s [2026-10-19 23:52:16,040] loop.py:45 INFO TRAIN_STEP step=16035 loss=1.253041 Training: 89%|████████▉ | 16035/18018 [34:22:02<3:33:38, 6.46s/step, loss=1.253, acc=0.707, acc_len=1.70, thru=19.3, I=22.5, T=39.1, wait=3.0s, pool=53, epoch=2/2][2026-10-19 23:52:24,982] loop.py:45 INFO TRAIN_STEP step=16036 loss=1.293939 Training: 89%|████████▉ | 16036/18018 [34:22:11<3:58:04, 7.21s/step, loss=1.294, acc=0.697, acc_len=1.67, thru=14.5, I=13.1, T=35.7, wait=5.4s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:52:24,978] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16035: forward=2532.8ms backward=996.8ms [2026-10-19 23:52:31,340] loop.py:45 INFO TRAIN_STEP step=16037 loss=1.367510 Training: 89%|████████▉ | 16037/18018 [34:22:17<3:49:34, 6.95s/step, loss=1.368, acc=0.686, acc_len=1.63, thru=19.2, I=21.4, T=37.1, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:52:38,312] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:52:39,495] loop.py:45 INFO TRAIN_STEP step=16038 loss=1.328025 Training: 89%|████████▉ | 16038/18018 [34:22:25<4:01:20, 7.31s/step, loss=1.328, acc=0.693, acc_len=1.65, thru=15.5, I=17.0, T=34.9, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:52:39,820] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:52:49,135] loop.py:45 INFO TRAIN_STEP step=16039 loss=1.214836 Training: 89%|████████▉ | 16039/18018 [34:22:35<4:24:13, 8.01s/step, loss=1.215, acc=0.716, acc_len=1.73, thru=9.0, I=13.5, T=29.9, wait=5.3s, pool=21, epoch=2/2] [2026-10-19 23:52:58,408] loop.py:380 INFO TIMING step=16040: step=3.379s data=1.370s compute=3.345s [fwd=2.407s bwd=0.924s opt=0.003s] dispatch=5.885s [2026-10-19 23:52:58,408] loop.py:45 INFO TRAIN_STEP step=16040 loss=1.186563 Training: 89%|████████▉ | 16040/18018 [34:22:44<4:36:34, 8.39s/step, loss=1.187, acc=0.721, acc_len=1.75, thru=18.6, I=18.0, T=37.9, wait=5.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:04,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:53:04,965] loop.py:45 INFO TRAIN_STEP step=16041 loss=1.223998 Training: 89%|████████▉ | 16041/18018 [34:22:51<4:18:19, 7.84s/step, loss=1.224, acc=0.713, acc_len=1.73, thru=19.6, I=20.9, T=38.8, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:53:04,961] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16040: forward=2363.2ms backward=894.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:05,308] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:10,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:53:11,269] loop.py:45 INFO TRAIN_STEP step=16042 loss=1.257448 Training: 89%|████████▉ | 16042/18018 [34:22:57<4:03:01, 7.38s/step, loss=1.257, acc=0.705, acc_len=1.70, thru=20.4, I=24.8, T=38.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:11,722] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:53:18,129] loop.py:45 INFO TRAIN_STEP step=16043 loss=1.229117 Training: 89%|████████▉ | 16043/18018 [34:23:04<3:57:45, 7.22s/step, loss=1.229, acc=0.714, acc_len=1.73, thru=16.3, I=19.1, T=38.0, wait=3.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:25,035] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:53:25,416] loop.py:45 INFO TRAIN_STEP step=16044 loss=1.303190 Training: 89%|████████▉ | 16044/18018 [34:23:11<3:58:18, 7.24s/step, loss=1.303, acc=0.698, acc_len=1.68, thru=18.5, I=19.2, T=34.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:25,537] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:31,187] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:53:32,396] loop.py:380 INFO TIMING step=16045: step=3.575s data=1.557s compute=3.533s [fwd=2.525s bwd=0.993s opt=0.003s] dispatch=3.393s [2026-10-19 23:53:32,396] loop.py:45 INFO TRAIN_STEP step=16045 loss=1.349531 Training: 89%|████████▉ | 16045/18018 [34:23:18<3:55:34, 7.16s/step, loss=1.350, acc=0.688, acc_len=1.64, thru=17.8, I=20.8, T=35.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:32,697] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:53:39,545] loop.py:45 INFO TRAIN_STEP step=16046 loss=1.222658 Training: 89%|████████▉ | 16046/18018 [34:23:25<3:55:16, 7.16s/step, loss=1.223, acc=0.714, acc_len=1.72, thru=14.9, I=18.1, T=39.6, wait=3.9s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:53:39,541] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16045: forward=2303.4ms backward=875.0ms [2026-10-19 23:53:46,187] loop.py:45 INFO TRAIN_STEP step=16047 loss=1.133122 Training: 89%|████████▉ | 16047/18018 [34:23:32<3:50:04, 7.00s/step, loss=1.133, acc=0.733, acc_len=1.79, thru=16.1, I=17.0, T=37.4, wait=3.2s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:54,535] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:53:54,888] loop.py:45 INFO TRAIN_STEP step=16048 loss=1.443308 Training: 89%|████████▉ | 16048/18018 [34:23:40<4:06:41, 7.51s/step, loss=1.443, acc=0.677, acc_len=1.60, thru=17.8, I=16.5, T=36.4, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:53:55,037] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-19 23:54:01,292] loop.py:45 INFO TRAIN_STEP step=16049 loss=1.216199 Training: 89%|████████▉ | 16049/18018 [34:23:47<3:55:37, 7.18s/step, loss=1.216, acc=0.712, acc_len=1.72, thru=18.9, I=19.8, T=38.4, wait=3.1s, pool=57, epoch=2/2][2026-10-19 23:54:06,624] loop.py:380 INFO TIMING step=16050: step=3.224s data=1.257s compute=3.192s [fwd=2.301s bwd=0.878s opt=0.003s] dispatch=2.099s [2026-10-19 23:54:06,624] loop.py:45 INFO TRAIN_STEP step=16050 loss=1.221100 Training: 89%|████████▉ | 16050/18018 [34:23:52<3:37:18, 6.63s/step, loss=1.221, acc=0.714, acc_len=1.72, thru=19.2, I=20.2, T=39.7, wait=2.1s, pool=37, epoch=2/2][2026-10-19 23:54:14,161] loop.py:45 INFO TRAIN_STEP step=16051 loss=1.188867 Training: 89%|████████▉ | 16051/18018 [34:24:00<3:46:10, 6.90s/step, loss=1.189, acc=0.718, acc_len=1.74, thru=16.3, I=14.9, T=37.5, wait=4.1s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:54:14,157] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16050: forward=2429.1ms backward=947.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:54:22,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:54:22,780] loop.py:45 INFO TRAIN_STEP step=16052 loss=1.225199 Training: 89%|████████▉ | 16052/18018 [34:24:08<4:02:58, 7.42s/step, loss=1.225, acc=0.713, acc_len=1.72, thru=17.8, I=16.5, T=36.7, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:54:23,250] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:54:27,881] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:54:28,863] loop.py:45 INFO TRAIN_STEP step=16053 loss=1.282151 Training: 89%|████████▉ | 16053/18018 [34:24:14<3:49:46, 7.02s/step, loss=1.282, acc=0.701, acc_len=1.68, thru=22.0, I=26.2, T=36.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:54:28,895] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-19 23:54:34,296] loop.py:45 INFO TRAIN_STEP step=16054 loss=1.222671 Training: 89%|████████▉ | 16054/18018 [34:24:20<3:34:05, 6.54s/step, loss=1.223, acc=0.713, acc_len=1.73, thru=22.4, I=26.9, T=39.2, wait=2.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:54:41,825] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:54:42,135] loop.py:380 INFO TIMING step=16055: step=3.402s data=1.408s compute=3.367s [fwd=2.407s bwd=0.946s opt=0.003s] dispatch=4.428s [2026-10-19 23:54:42,135] loop.py:45 INFO TRAIN_STEP step=16055 loss=1.177725 Training: 89%|████████▉ | 16055/18018 [34:24:28<3:46:43, 6.93s/step, loss=1.178, acc=0.724, acc_len=1.76, thru=17.2, I=17.7, T=37.6, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:54:42,327] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:54:49,133] loop.py:45 INFO TRAIN_STEP step=16056 loss=1.113727 Training: 89%|████████▉ | 16056/18018 [34:24:35<3:47:16, 6.95s/step, loss=1.114, acc=0.732, acc_len=1.79, thru=15.8, I=16.9, T=39.2, wait=3.7s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:54:49,128] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16055: forward=2297.7ms backward=922.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:54:57,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:54:57,882] loop.py:45 INFO TRAIN_STEP step=16057 loss=1.253593 Training: 89%|████████▉ | 16057/18018 [34:24:43<4:04:48, 7.49s/step, loss=1.254, acc=0.707, acc_len=1.70, thru=17.9, I=16.7, T=35.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:54:58,164] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:55:07,662] loop.py:45 INFO TRAIN_STEP step=16058 loss=1.213637 Training: 89%|████████▉ | 16058/18018 [34:24:53<4:27:06, 8.18s/step, loss=1.214, acc=0.713, acc_len=1.72, thru=12.3, I=12.2, T=32.6, wait=5.8s, pool=49, epoch=2/2][2026-10-19 23:55:16,613] loop.py:45 INFO TRAIN_STEP step=16059 loss=1.344200 Training: 89%|████████▉ | 16059/18018 [34:25:02<4:34:33, 8.41s/step, loss=1.344, acc=0.688, acc_len=1.64, thru=13.4, I=12.4, T=34.8, wait=5.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:55:25,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:55:25,650] loop.py:380 INFO TIMING step=16060: step=3.498s data=1.408s compute=3.467s [fwd=2.493s bwd=0.960s opt=0.003s] dispatch=5.531s [2026-10-19 23:55:25,651] loop.py:45 INFO TRAIN_STEP step=16060 loss=1.365903 Training: 89%|████████▉ | 16060/18018 [34:25:11<4:40:34, 8.60s/step, loss=1.366, acc=0.686, acc_len=1.63, thru=16.9, I=15.1, T=36.6, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:55:26,064] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:55:30,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:55:31,667] loop.py:45 INFO TRAIN_STEP step=16061 loss=1.268406 Training: 89%|████████▉ | 16061/18018 [34:25:17<4:15:11, 7.82s/step, loss=1.268, acc=0.706, acc_len=1.70, thru=20.5, I=21.9, T=37.9, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:55:31,662] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16060: forward=2371.1ms backward=953.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:55:31,943] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:55:37,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:55:38,115] loop.py:45 INFO TRAIN_STEP step=16062 loss=1.196218 Training: 89%|████████▉ | 16062/18018 [34:25:24<4:01:34, 7.41s/step, loss=1.196, acc=0.717, acc_len=1.74, thru=19.9, I=22.7, T=39.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:55:38,478] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:55:45,289] loop.py:45 INFO TRAIN_STEP step=16063 loss=1.224353 Training: 89%|████████▉ | 16063/18018 [34:25:31<3:59:09, 7.34s/step, loss=1.224, acc=0.717, acc_len=1.74, thru=17.2, I=18.1, T=37.2, wait=3.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:55:51,553] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:55:52,371] loop.py:45 INFO TRAIN_STEP step=16064 loss=1.320547 Training: 89%|████████▉ | 16064/18018 [34:25:38<3:56:31, 7.26s/step, loss=1.321, acc=0.696, acc_len=1.67, thru=18.9, I=20.2, T=37.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:55:52,560] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:55:57,978] loop.py:380 INFO TIMING step=16065: step=3.198s data=1.287s compute=3.169s [fwd=2.281s bwd=0.875s opt=0.003s] dispatch=2.400s [2026-10-19 23:55:57,979] loop.py:45 INFO TRAIN_STEP step=16065 loss=1.251545 Training: 89%|████████▉ | 16065/18018 [34:25:44<3:40:12, 6.77s/step, loss=1.252, acc=0.706, acc_len=1.70, thru=20.0, I=23.4, T=40.0, wait=2.4s, pool=45, epoch=2/2][2026-10-19 23:56:04,492] loop.py:45 INFO TRAIN_STEP step=16066 loss=1.161797 Training: 89%|████████▉ | 16066/18018 [34:25:50<3:37:39, 6.69s/step, loss=1.162, acc=0.726, acc_len=1.77, thru=19.7, I=20.5, T=37.9, wait=3.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:56:04,487] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16065: forward=2386.1ms backward=942.0ms [2026-10-19 23:56:12,282] loop.py:45 INFO TRAIN_STEP step=16067 loss=1.254237 Training: 89%|████████▉ | 16067/18018 [34:25:58<3:48:15, 7.02s/step, loss=1.254, acc=0.707, acc_len=1.70, thru=16.2, I=16.5, T=36.4, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:12,264] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:12,766] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:16,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:56:17,959] loop.py:45 INFO TRAIN_STEP step=16068 loss=1.285256 Training: 89%|████████▉ | 16068/18018 [34:26:04<3:35:03, 6.62s/step, loss=1.285, acc=0.703, acc_len=1.68, thru=21.3, I=23.9, T=39.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:18,320] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:23,372] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:56:24,629] loop.py:45 INFO TRAIN_STEP step=16069 loss=1.231368 Training: 89%|████████▉ | 16069/18018 [34:26:10<3:35:28, 6.63s/step, loss=1.231, acc=0.713, acc_len=1.72, thru=19.8, I=25.4, T=37.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:24,880] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:56:31,307] loop.py:380 INFO TIMING step=16070: step=3.337s data=1.336s compute=3.306s [fwd=2.375s bwd=0.916s opt=0.003s] dispatch=3.331s [2026-10-19 23:56:31,307] loop.py:45 INFO TRAIN_STEP step=16070 loss=1.261520 Training: 89%|████████▉ | 16070/18018 [34:26:17<3:35:47, 6.65s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=18.5, I=21.0, T=38.4, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:38,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:56:39,376] loop.py:45 INFO TRAIN_STEP step=16071 loss=1.297597 Training: 89%|████████▉ | 16071/18018 [34:26:25<3:49:32, 7.07s/step, loss=1.298, acc=0.700, acc_len=1.67, thru=18.0, I=18.0, T=37.0, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:56:39,372] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16070: forward=2366.3ms backward=1036.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:39,510] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:45,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:56:46,062] loop.py:45 INFO TRAIN_STEP step=16072 loss=1.308202 Training: 89%|████████▉ | 16072/18018 [34:26:32<3:45:38, 6.96s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=19.2, I=20.3, T=35.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:46,566] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:52,270] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:56:53,623] loop.py:45 INFO TRAIN_STEP step=16073 loss=1.137106 Training: 89%|████████▉ | 16073/18018 [34:26:39<3:51:24, 7.14s/step, loss=1.137, acc=0.728, acc_len=1.78, thru=16.9, I=21.9, T=37.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:56:53,782] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-19 23:57:01,172] loop.py:45 INFO TRAIN_STEP step=16074 loss=1.248575 Training: 89%|████████▉ | 16074/18018 [34:26:47<3:55:18, 7.26s/step, loss=1.249, acc=0.709, acc_len=1.72, thru=13.7, I=20.0, T=38.1, wait=4.2s, pool=37, epoch=2/2][2026-10-19 23:57:12,466] loop.py:380 INFO TIMING step=16075: step=3.716s data=1.514s compute=3.686s [fwd=2.625s bwd=1.046s opt=0.003s] dispatch=7.563s [2026-10-19 23:57:12,466] loop.py:45 INFO TRAIN_STEP step=16075 loss=1.278983 Training: 89%|████████▉ | 16075/18018 [34:26:58<4:34:19, 8.47s/step, loss=1.279, acc=0.705, acc_len=1.70, thru=13.1, I=12.4, T=34.4, wait=7.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:19,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:57:20,223] loop.py:45 INFO TRAIN_STEP step=16076 loss=1.280748 Training: 89%|████████▉ | 16076/18018 [34:27:06<4:27:13, 8.26s/step, loss=1.281, acc=0.702, acc_len=1.68, thru=16.6, I=17.3, T=38.2, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:57:20,219] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16075: forward=2332.5ms backward=941.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:20,442] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:25,915] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:57:26,223] loop.py:45 INFO TRAIN_STEP step=16077 loss=1.265905 Training: 89%|████████▉ | 16077/18018 [34:27:12<4:05:12, 7.58s/step, loss=1.266, acc=0.707, acc_len=1.70, thru=21.4, I=23.6, T=39.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:26,417] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:31,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:57:33,243] loop.py:45 INFO TRAIN_STEP step=16078 loss=1.361813 Training: 89%|████████▉ | 16078/18018 [34:27:19<3:59:39, 7.41s/step, loss=1.362, acc=0.688, acc_len=1.64, thru=18.6, I=21.9, T=37.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:33,506] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:37,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:57:38,807] loop.py:45 INFO TRAIN_STEP step=16079 loss=1.209342 Training: 89%|████████▉ | 16079/18018 [34:27:24<3:41:36, 6.86s/step, loss=1.209, acc=0.718, acc_len=1.75, thru=22.3, I=29.3, T=39.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:38,926] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-19 23:57:44,267] loop.py:380 INFO TIMING step=16080: step=3.197s data=1.295s compute=3.168s [fwd=2.289s bwd=0.864s opt=0.003s] dispatch=2.253s [2026-10-19 23:57:44,268] loop.py:45 INFO TRAIN_STEP step=16080 loss=1.208182 Training: 89%|████████▉ | 16080/18018 [34:27:30<3:27:58, 6.44s/step, loss=1.208, acc=0.716, acc_len=1.74, thru=22.5, I=26.8, T=40.0, wait=2.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:51,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:57:52,204] loop.py:45 INFO TRAIN_STEP step=16081 loss=1.328811 Training: 89%|████████▉ | 16081/18018 [34:27:38<3:42:20, 6.89s/step, loss=1.329, acc=0.695, acc_len=1.66, thru=17.4, I=18.7, T=37.7, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:57:52,200] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16080: forward=2356.6ms backward=975.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:52,664] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:57:59,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:57:59,921] loop.py:45 INFO TRAIN_STEP step=16082 loss=1.167483 Training: 89%|████████▉ | 16082/18018 [34:27:46<3:50:15, 7.14s/step, loss=1.167, acc=0.723, acc_len=1.76, thru=16.6, I=19.3, T=38.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:00,288] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:05,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:58:06,625] loop.py:45 INFO TRAIN_STEP step=16083 loss=1.158193 Training: 89%|████████▉ | 16083/18018 [34:27:52<3:45:58, 7.01s/step, loss=1.158, acc=0.725, acc_len=1.77, thru=19.4, I=20.9, T=37.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:06,841] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:11,724] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:58:13,635] loop.py:45 INFO TRAIN_STEP step=16084 loss=1.206973 Training: 89%|████████▉ | 16084/18018 [34:27:59<3:45:51, 7.01s/step, loss=1.207, acc=0.714, acc_len=1.72, thru=18.6, I=25.6, T=37.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:13,728] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:20,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:58:20,844] loop.py:380 INFO TIMING step=16085: step=3.277s data=1.294s compute=3.251s [fwd=2.335s bwd=0.900s opt=0.003s] dispatch=3.924s [2026-10-19 23:58:20,844] loop.py:45 INFO TRAIN_STEP step=16085 loss=1.184485 Training: 89%|████████▉ | 16085/18018 [34:28:06<3:47:43, 7.07s/step, loss=1.184, acc=0.721, acc_len=1.75, thru=16.9, I=23.9, T=39.1, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:21,179] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:26,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:58:27,281] loop.py:45 INFO TRAIN_STEP step=16086 loss=1.202878 Training: 89%|████████▉ | 16086/18018 [34:28:13<3:41:29, 6.88s/step, loss=1.203, acc=0.717, acc_len=1.73, thru=20.1, I=23.4, T=38.6, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:58:27,277] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16085: forward=2362.9ms backward=906.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:27,608] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:58:35,814] loop.py:45 INFO TRAIN_STEP step=16087 loss=1.234433 Training: 89%|████████▉ | 16087/18018 [34:28:21<3:57:23, 7.38s/step, loss=1.234, acc=0.709, acc_len=1.71, thru=14.6, I=15.8, T=37.1, wait=5.1s, pool=61, epoch=2/2][2026-10-19 23:58:42,555] loop.py:45 INFO TRAIN_STEP step=16088 loss=1.255200 Training: 89%|████████▉ | 16088/18018 [34:28:28<3:51:05, 7.18s/step, loss=1.255, acc=0.708, acc_len=1.71, thru=15.8, I=17.1, T=35.9, wait=3.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:50,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:58:50,361] loop.py:45 INFO TRAIN_STEP step=16089 loss=1.176625 Training: 89%|████████▉ | 16089/18018 [34:28:36<3:56:58, 7.37s/step, loss=1.177, acc=0.720, acc_len=1.75, thru=18.5, I=17.3, T=37.5, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:58:50,697] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-19 23:58:56,406] loop.py:380 INFO TIMING step=16090: step=3.416s data=1.398s compute=3.386s [fwd=2.425s bwd=0.946s opt=0.003s] dispatch=2.620s [2026-10-19 23:58:56,406] loop.py:45 INFO TRAIN_STEP step=16090 loss=1.307591 Training: 89%|████████▉ | 16090/18018 [34:28:42<3:44:04, 6.97s/step, loss=1.308, acc=0.696, acc_len=1.67, thru=19.9, I=22.2, T=37.5, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:03,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:59:03,843] loop.py:45 INFO TRAIN_STEP step=16091 loss=1.237957 Training: 89%|████████▉ | 16091/18018 [34:28:49<3:48:28, 7.11s/step, loss=1.238, acc=0.709, acc_len=1.71, thru=17.0, I=18.1, T=36.6, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:59:03,837] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16090: forward=2442.4ms backward=993.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:03,958] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:09,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:59:09,989] loop.py:45 INFO TRAIN_STEP step=16092 loss=1.270695 Training: 89%|████████▉ | 16092/18018 [34:28:56<3:38:58, 6.82s/step, loss=1.271, acc=0.706, acc_len=1.71, thru=20.5, I=21.9, T=39.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:10,376] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:15,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:59:16,024] loop.py:45 INFO TRAIN_STEP step=16093 loss=1.191506 Training: 89%|████████▉ | 16093/18018 [34:29:02<3:31:18, 6.59s/step, loss=1.192, acc=0.720, acc_len=1.75, thru=20.7, I=22.9, T=38.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:16,200] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:21,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:59:22,311] loop.py:45 INFO TRAIN_STEP step=16094 loss=1.200007 Training: 89%|████████▉ | 16094/18018 [34:29:08<3:28:17, 6.50s/step, loss=1.200, acc=0.720, acc_len=1.74, thru=20.2, I=23.2, T=37.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:22,696] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-19 23:59:28,165] loop.py:380 INFO TIMING step=16095: step=3.249s data=1.258s compute=3.225s [fwd=2.341s bwd=0.869s opt=0.003s] dispatch=2.596s [2026-10-19 23:59:28,165] loop.py:45 INFO TRAIN_STEP step=16095 loss=1.283361 Training: 89%|████████▉ | 16095/18018 [34:29:14<3:22:02, 6.30s/step, loss=1.283, acc=0.705, acc_len=1.70, thru=20.8, I=22.5, T=39.4, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:35,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:59:36,410] loop.py:45 INFO TRAIN_STEP step=16096 loss=1.221399 Training: 89%|████████▉ | 16096/18018 [34:29:22<3:40:35, 6.89s/step, loss=1.221, acc=0.715, acc_len=1.74, thru=16.9, I=18.4, T=37.2, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-19 23:59:36,405] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16095: forward=2439.2ms backward=953.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:36,704] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:41,992] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:59:42,278] loop.py:45 INFO TRAIN_STEP step=16097 loss=1.208627 Training: 89%|████████▉ | 16097/18018 [34:29:28<3:30:41, 6.58s/step, loss=1.209, acc=0.713, acc_len=1.72, thru=21.2, I=25.0, T=39.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:42,493] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:48,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:59:48,763] loop.py:45 INFO TRAIN_STEP step=16098 loss=1.302716 Training: 89%|████████▉ | 16098/18018 [34:29:34<3:29:40, 6.55s/step, loss=1.303, acc=0.697, acc_len=1.67, thru=20.0, I=22.2, T=39.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:49,038] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:54,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-19 23:59:55,332] loop.py:45 INFO TRAIN_STEP step=16099 loss=1.222614 Training: 89%|████████▉ | 16099/18018 [34:29:41<3:29:42, 6.56s/step, loss=1.223, acc=0.715, acc_len=1.73, thru=19.8, I=22.0, T=37.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-19 23:59:55,489] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:00:01,501] loop.py:380 INFO TIMING step=16100: step=3.419s data=1.374s compute=3.381s [fwd=2.416s bwd=0.950s opt=0.003s] dispatch=2.742s [2026-10-20 00:00:01,501] loop.py:45 INFO TRAIN_STEP step=16100 loss=1.317712 Training: 89%|████████▉ | 16100/18018 [34:29:47<3:25:52, 6.44s/step, loss=1.318, acc=0.698, acc_len=1.67, thru=18.5, I=19.8, T=37.4, wait=2.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:09,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:00:10,816] loop.py:45 INFO TRAIN_STEP step=16101 loss=1.346765 Training: 89%|████████▉ | 16101/18018 [34:29:56<3:53:19, 7.30s/step, loss=1.347, acc=0.691, acc_len=1.65, thru=16.8, I=17.6, T=36.1, wait=5.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:00:10,811] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16100: forward=2525.1ms backward=982.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:11,304] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:15,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:00:16,615] loop.py:45 INFO TRAIN_STEP step=16102 loss=1.162632 Training: 89%|████████▉ | 16102/18018 [34:30:02<3:38:48, 6.85s/step, loss=1.163, acc=0.725, acc_len=1.77, thru=20.6, I=27.4, T=39.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:16,922] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:23,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:00:24,150] loop.py:45 INFO TRAIN_STEP step=16103 loss=1.411164 Training: 89%|████████▉ | 16103/18018 [34:30:10<3:45:14, 7.06s/step, loss=1.411, acc=0.676, acc_len=1.60, thru=17.8, I=21.4, T=37.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:24,294] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:00:31,813] loop.py:45 INFO TRAIN_STEP step=16104 loss=1.138238 Training: 89%|████████▉ | 16104/18018 [34:30:17<3:50:54, 7.24s/step, loss=1.138, acc=0.728, acc_len=1.78, thru=13.8, I=16.8, T=37.1, wait=4.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:39,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:00:40,658] loop.py:380 INFO TIMING step=16105: step=3.920s data=1.677s compute=3.894s [fwd=2.767s bwd=1.113s opt=0.003s] dispatch=4.914s [2026-10-20 00:00:40,658] loop.py:45 INFO TRAIN_STEP step=16105 loss=1.320142 Training: 89%|████████▉ | 16105/18018 [34:30:26<4:06:09, 7.72s/step, loss=1.320, acc=0.696, acc_len=1.66, thru=16.5, I=19.4, T=32.7, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:41,101] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 00:00:46,621] loop.py:45 INFO TRAIN_STEP step=16106 loss=1.096658 Training: 89%|████████▉ | 16106/18018 [34:30:32<3:49:12, 7.19s/step, loss=1.097, acc=0.738, acc_len=1.81, thru=19.0, I=25.2, T=38.2, wait=2.6s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:00:46,617] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16105: forward=2363.5ms backward=926.3ms [2026-10-20 00:00:53,386] loop.py:45 INFO TRAIN_STEP step=16107 loss=1.339467 Training: 89%|████████▉ | 16107/18018 [34:30:39<3:44:59, 7.06s/step, loss=1.339, acc=0.691, acc_len=1.65, thru=20.0, I=19.1, T=38.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:53,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:00:53,893] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:00:59,719] loop.py:45 INFO TRAIN_STEP step=16108 loss=1.348146 Training: 89%|████████▉ | 16108/18018 [34:30:45<3:37:56, 6.85s/step, loss=1.348, acc=0.693, acc_len=1.65, thru=19.0, I=19.5, T=38.8, wait=3.0s, pool=53, epoch=2/2][2026-10-20 00:01:06,446] loop.py:45 INFO TRAIN_STEP step=16109 loss=1.206177 Training: 89%|████████▉ | 16109/18018 [34:30:52<3:36:38, 6.81s/step, loss=1.206, acc=0.715, acc_len=1.73, thru=17.8, I=18.2, T=37.5, wait=3.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:01:11,964] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:01:13,271] loop.py:380 INFO TIMING step=16110: step=3.378s data=1.376s compute=3.345s [fwd=2.393s bwd=0.938s opt=0.003s] dispatch=3.440s [2026-10-20 00:01:13,271] loop.py:45 INFO TRAIN_STEP step=16110 loss=1.260065 Training: 89%|████████▉ | 16110/18018 [34:30:59<3:36:41, 6.81s/step, loss=1.260, acc=0.708, acc_len=1.71, thru=19.2, I=21.8, T=37.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:01:13,473] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:01:20,648] loop.py:45 INFO TRAIN_STEP step=16111 loss=1.267951 Training: 89%|████████▉ | 16111/18018 [34:31:06<3:41:57, 6.98s/step, loss=1.268, acc=0.709, acc_len=1.71, thru=15.8, I=21.0, T=36.9, wait=3.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:01:20,643] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16110: forward=2442.4ms backward=977.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:01:27,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:01:28,057] loop.py:45 INFO TRAIN_STEP step=16112 loss=1.151593 Training: 89%|████████▉ | 16112/18018 [34:31:14<3:45:52, 7.11s/step, loss=1.152, acc=0.726, acc_len=1.77, thru=17.7, I=18.7, T=37.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:01:28,437] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:01:34,268] loop.py:45 INFO TRAIN_STEP step=16113 loss=1.268366 Training: 89%|████████▉ | 16113/18018 [34:31:20<3:37:12, 6.84s/step, loss=1.268, acc=0.708, acc_len=1.70, thru=19.9, I=21.9, T=39.0, wait=2.9s, pool=61, epoch=2/2][2026-10-20 00:01:40,812] loop.py:45 INFO TRAIN_STEP step=16114 loss=1.229727 Training: 89%|████████▉ | 16114/18018 [34:31:26<3:34:14, 6.75s/step, loss=1.230, acc=0.710, acc_len=1.71, thru=19.7, I=19.4, T=37.8, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:01:48,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:01:50,028] loop.py:380 INFO TIMING step=16115: step=3.814s data=1.581s compute=3.766s [fwd=2.613s bwd=1.138s opt=0.003s] dispatch=5.393s [2026-10-20 00:01:50,028] loop.py:45 INFO TRAIN_STEP step=16115 loss=1.325185 Training: 89%|████████▉ | 16115/18018 [34:31:36<3:57:36, 7.49s/step, loss=1.325, acc=0.698, acc_len=1.67, thru=15.1, I=17.0, T=33.6, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:01:50,239] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:01:56,034] loop.py:45 INFO TRAIN_STEP step=16116 loss=1.286059 Training: 89%|████████▉ | 16116/18018 [34:31:42<3:43:20, 7.05s/step, loss=1.286, acc=0.703, acc_len=1.69, thru=18.6, I=17.3, T=38.8, wait=2.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:01:56,030] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16115: forward=2305.0ms backward=929.4ms [2026-10-20 00:02:03,866] loop.py:45 INFO TRAIN_STEP step=16117 loss=1.433536 Training: 89%|████████▉ | 16117/18018 [34:31:49<3:50:42, 7.28s/step, loss=1.434, acc=0.675, acc_len=1.59, thru=16.1, I=17.6, T=36.6, wait=4.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:10,198] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:02:10,665] loop.py:45 INFO TRAIN_STEP step=16118 loss=1.110871 Training: 89%|████████▉ | 16118/18018 [34:31:56<3:45:59, 7.14s/step, loss=1.111, acc=0.736, acc_len=1.81, thru=18.8, I=20.1, T=36.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:10,711] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:16,875] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:02:17,297] loop.py:45 INFO TRAIN_STEP step=16119 loss=1.215846 Training: 89%|████████▉ | 16119/18018 [34:32:03<3:41:05, 6.99s/step, loss=1.216, acc=0.714, acc_len=1.73, thru=19.4, I=21.7, T=38.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:17,380] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-20 00:02:23,410] loop.py:380 INFO TIMING step=16120: step=3.340s data=1.354s compute=3.303s [fwd=2.372s bwd=0.917s opt=0.003s] dispatch=2.764s [2026-10-20 00:02:23,411] loop.py:45 INFO TRAIN_STEP step=16120 loss=1.136646 Training: 89%|████████▉ | 16120/18018 [34:32:09<3:32:40, 6.72s/step, loss=1.137, acc=0.728, acc_len=1.78, thru=20.4, I=21.9, T=38.3, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:31,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:02:31,534] loop.py:45 INFO TRAIN_STEP step=16121 loss=1.334929 Training: 89%|████████▉ | 16121/18018 [34:32:17<3:45:50, 7.14s/step, loss=1.335, acc=0.694, acc_len=1.66, thru=17.2, I=17.4, T=37.4, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:02:31,529] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16120: forward=2430.6ms backward=939.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:32,037] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:38,313] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:02:38,639] loop.py:45 INFO TRAIN_STEP step=16122 loss=1.137686 Training: 89%|████████▉ | 16122/18018 [34:32:24<3:45:22, 7.13s/step, loss=1.138, acc=0.728, acc_len=1.77, thru=17.6, I=19.3, T=35.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:38,816] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:02:44,336] loop.py:45 INFO TRAIN_STEP step=16123 loss=1.144906 Training: 89%|████████▉ | 16123/18018 [34:32:30<3:31:39, 6.70s/step, loss=1.145, acc=0.727, acc_len=1.78, thru=20.7, I=24.4, T=39.2, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:53,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:02:53,802] loop.py:45 INFO TRAIN_STEP step=16124 loss=1.318412 Training: 89%|████████▉ | 16124/18018 [34:32:39<3:57:43, 7.53s/step, loss=1.318, acc=0.697, acc_len=1.66, thru=14.9, I=14.9, T=36.2, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:02:54,147] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:00,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:03:01,058] loop.py:380 INFO TIMING step=16125: step=3.371s data=1.369s compute=3.345s [fwd=2.393s bwd=0.937s opt=0.003s] dispatch=3.878s [2026-10-20 00:03:01,058] loop.py:45 INFO TRAIN_STEP step=16125 loss=1.349232 Training: 89%|████████▉ | 16125/18018 [34:32:47<3:54:59, 7.45s/step, loss=1.349, acc=0.692, acc_len=1.65, thru=17.8, I=19.6, T=38.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:01,461] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:08,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:03:08,782] loop.py:45 INFO TRAIN_STEP step=16126 loss=1.367504 Training: 89%|████████▉ | 16126/18018 [34:32:54<3:57:30, 7.53s/step, loss=1.368, acc=0.688, acc_len=1.63, thru=18.0, I=18.8, T=36.9, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:03:08,778] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16125: forward=2353.6ms backward=1033.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:08,949] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:13,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:03:14,895] loop.py:45 INFO TRAIN_STEP step=16127 loss=1.188807 Training: 90%|████████▉ | 16127/18018 [34:33:00<3:43:56, 7.11s/step, loss=1.189, acc=0.718, acc_len=1.74, thru=20.4, I=23.4, T=38.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:15,370] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:20,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:03:21,350] loop.py:45 INFO TRAIN_STEP step=16128 loss=1.187062 Training: 90%|████████▉ | 16128/18018 [34:33:07<3:37:41, 6.91s/step, loss=1.187, acc=0.717, acc_len=1.74, thru=20.0, I=27.3, T=38.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:21,700] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:03:27,261] loop.py:45 INFO TRAIN_STEP step=16129 loss=1.300375 Training: 90%|████████▉ | 16129/18018 [34:33:13<3:28:07, 6.61s/step, loss=1.300, acc=0.697, acc_len=1.66, thru=18.9, I=22.3, T=40.0, wait=2.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:33,731] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:03:34,452] loop.py:380 INFO TIMING step=16130: step=3.491s data=1.415s compute=3.448s [fwd=2.436s bwd=0.996s opt=0.003s] dispatch=3.691s [2026-10-20 00:03:34,452] loop.py:45 INFO TRAIN_STEP step=16130 loss=1.262695 Training: 90%|████████▉ | 16130/18018 [34:33:20<3:33:30, 6.79s/step, loss=1.263, acc=0.707, acc_len=1.70, thru=20.0, I=20.2, T=36.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:34,736] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:42,599] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:03:42,961] loop.py:45 INFO TRAIN_STEP step=16131 loss=1.196981 Training: 90%|████████▉ | 16131/18018 [34:33:29<3:49:38, 7.30s/step, loss=1.197, acc=0.716, acc_len=1.74, thru=15.7, I=18.1, T=36.0, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:03:42,956] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16130: forward=2375.2ms backward=1086.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:43,113] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:47,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:03:48,580] loop.py:45 INFO TRAIN_STEP step=16132 loss=1.299065 Training: 90%|████████▉ | 16132/18018 [34:33:34<3:33:39, 6.80s/step, loss=1.299, acc=0.696, acc_len=1.66, thru=20.9, I=24.5, T=38.3, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:48,834] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:53,710] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:03:54,124] loop.py:45 INFO TRAIN_STEP step=16133 loss=1.218487 Training: 90%|████████▉ | 16133/18018 [34:33:40<3:21:43, 6.42s/step, loss=1.218, acc=0.712, acc_len=1.72, thru=24.3, I=26.2, T=40.0, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:54,225] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:03:59,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:04:00,222] loop.py:45 INFO TRAIN_STEP step=16134 loss=1.230486 Training: 90%|████████▉ | 16134/18018 [34:33:46<3:18:34, 6.32s/step, loss=1.230, acc=0.715, acc_len=1.73, thru=22.0, I=23.9, T=39.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:00,690] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:06,919] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:04:07,768] loop.py:380 INFO TIMING step=16135: step=3.358s data=1.429s compute=3.310s [fwd=2.334s bwd=0.962s opt=0.003s] dispatch=4.180s [2026-10-20 00:04:07,768] loop.py:45 INFO TRAIN_STEP step=16135 loss=1.153215 Training: 90%|████████▉ | 16135/18018 [34:33:53<3:29:59, 6.69s/step, loss=1.153, acc=0.728, acc_len=1.78, thru=17.4, I=20.6, T=38.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:07,923] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:14,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:04:15,589] loop.py:45 INFO TRAIN_STEP step=16136 loss=1.332122 Training: 90%|████████▉ | 16136/18018 [34:34:01<3:40:30, 7.03s/step, loss=1.332, acc=0.695, acc_len=1.66, thru=16.5, I=19.5, T=39.4, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:04:15,584] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16135: forward=2289.9ms backward=911.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:15,808] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:23,770] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:04:25,188] loop.py:45 INFO TRAIN_STEP step=16137 loss=1.238245 Training: 90%|████████▉ | 16137/18018 [34:34:11<4:04:33, 7.80s/step, loss=1.238, acc=0.713, acc_len=1.73, thru=13.6, I=15.6, T=33.0, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:25,293] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:30,642] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:04:30,890] loop.py:45 INFO TRAIN_STEP step=16138 loss=1.238452 Training: 90%|████████▉ | 16138/18018 [34:34:16<3:44:41, 7.17s/step, loss=1.238, acc=0.712, acc_len=1.72, thru=20.4, I=23.5, T=38.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:31,145] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:04:38,887] loop.py:45 INFO TRAIN_STEP step=16139 loss=1.293269 Training: 90%|████████▉ | 16139/18018 [34:34:24<3:52:18, 7.42s/step, loss=1.293, acc=0.699, acc_len=1.68, thru=14.4, I=17.0, T=35.5, wait=4.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:45,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:04:46,427] loop.py:380 INFO TIMING step=16140: step=3.586s data=1.551s compute=3.540s [fwd=2.440s bwd=1.085s opt=0.003s] dispatch=3.945s [2026-10-20 00:04:46,428] loop.py:45 INFO TRAIN_STEP step=16140 loss=1.216212 Training: 90%|████████▉ | 16140/18018 [34:34:32<3:53:20, 7.46s/step, loss=1.216, acc=0.716, acc_len=1.74, thru=16.7, I=17.8, T=35.7, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:46,727] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:53,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:04:54,687] loop.py:45 INFO TRAIN_STEP step=16141 loss=1.321039 Training: 90%|████████▉ | 16141/18018 [34:34:40<4:00:47, 7.70s/step, loss=1.321, acc=0.698, acc_len=1.66, thru=15.9, I=18.9, T=35.0, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:04:54,682] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16140: forward=2594.6ms backward=1022.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:04:55,106] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:05:00,779] loop.py:45 INFO TRAIN_STEP step=16142 loss=1.241690 Training: 90%|████████▉ | 16142/18018 [34:34:46<3:45:36, 7.22s/step, loss=1.242, acc=0.713, acc_len=1.73, thru=16.8, I=21.5, T=39.7, wait=2.9s, pool=33, epoch=2/2][2026-10-20 00:05:09,354] loop.py:45 INFO TRAIN_STEP step=16143 loss=1.262107 Training: 90%|████████▉ | 16143/18018 [34:34:55<3:58:14, 7.62s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=16.9, I=15.1, T=36.4, wait=5.0s, pool=53, epoch=2/2][2026-10-20 00:05:19,252] loop.py:45 INFO TRAIN_STEP step=16144 loss=1.147955 Training: 90%|████████▉ | 16144/18018 [34:35:05<4:19:22, 8.30s/step, loss=1.148, acc=0.726, acc_len=1.77, thru=14.1, I=13.8, T=34.1, wait=6.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:05:24,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:05:24,737] loop.py:380 INFO TIMING step=16145: step=3.284s data=1.369s compute=3.250s [fwd=2.317s bwd=0.919s opt=0.003s] dispatch=2.194s [2026-10-20 00:05:24,737] loop.py:45 INFO TRAIN_STEP step=16145 loss=1.260947 Training: 90%|████████▉ | 16145/18018 [34:35:10<3:52:51, 7.46s/step, loss=1.261, acc=0.702, acc_len=1.68, thru=21.1, I=18.4, T=39.0, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:05:25,004] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:05:36,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:05:36,321] loop.py:45 INFO TRAIN_STEP step=16146 loss=1.376978 (TrainerActor pid=9310) [2026-10-20 00:05:36,316] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16145: forward=2527.2ms backward=1005.3ms Training: 90%|████████▉ | 16146/18018 [34:35:22<4:31:19, 8.70s/step, loss=1.377, acc=0.686, acc_len=1.63, thru=13.4, I=10.4, T=35.6, wait=8.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:05:36,778] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:05:43,549] loop.py:45 INFO TRAIN_STEP step=16147 loss=1.231841 Training: 90%|████████▉ | 16147/18018 [34:35:29<4:17:27, 8.26s/step, loss=1.232, acc=0.708, acc_len=1.70, thru=16.1, I=18.7, T=37.1, wait=3.8s, pool=53, epoch=2/2][2026-10-20 00:05:51,201] loop.py:45 INFO TRAIN_STEP step=16148 loss=1.285426 Training: 90%|████████▉ | 16148/18018 [34:35:37<4:11:39, 8.07s/step, loss=1.285, acc=0.702, acc_len=1.69, thru=16.9, I=17.1, T=36.0, wait=4.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:05:57,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:05:57,677] loop.py:45 INFO TRAIN_STEP step=16149 loss=1.088214 Training: 90%|████████▉ | 16149/18018 [34:35:43<3:56:36, 7.60s/step, loss=1.088, acc=0.736, acc_len=1.81, thru=20.3, I=19.8, T=38.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:05:57,883] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:02,495] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:06:03,904] loop.py:380 INFO TIMING step=16150: step=3.527s data=1.473s compute=3.498s [fwd=2.516s bwd=0.967s opt=0.003s] dispatch=2.690s [2026-10-20 00:06:03,904] loop.py:45 INFO TRAIN_STEP step=16150 loss=1.082587 Training: 90%|████████▉ | 16150/18018 [34:35:49<3:43:40, 7.18s/step, loss=1.083, acc=0.739, acc_len=1.82, thru=20.6, I=23.7, T=36.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:04,012] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-20 00:06:10,093] loop.py:45 INFO TRAIN_STEP step=16151 loss=1.189976 Training: 90%|████████▉ | 16151/18018 [34:35:56<3:34:16, 6.89s/step, loss=1.190, acc=0.720, acc_len=1.75, thru=19.7, I=22.5, T=38.0, wait=2.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:06:10,089] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16150: forward=2362.6ms backward=960.7ms [2026-10-20 00:06:17,058] loop.py:45 INFO TRAIN_STEP step=16152 loss=1.239992 Training: 90%|████████▉ | 16152/18018 [34:36:03<3:34:52, 6.91s/step, loss=1.240, acc=0.709, acc_len=1.71, thru=16.4, I=15.6, T=37.7, wait=3.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:24,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:06:24,640] loop.py:45 INFO TRAIN_STEP step=16153 loss=1.210811 Training: 90%|████████▉ | 16153/18018 [34:36:10<3:41:04, 7.11s/step, loss=1.211, acc=0.718, acc_len=1.74, thru=19.9, I=18.0, T=36.6, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:24,677] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:30,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:06:31,134] loop.py:45 INFO TRAIN_STEP step=16154 loss=1.117648 Training: 90%|████████▉ | 16154/18018 [34:36:17<3:35:11, 6.93s/step, loss=1.118, acc=0.729, acc_len=1.78, thru=19.9, I=20.3, T=37.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:31,347] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:35,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:06:36,885] loop.py:380 INFO TIMING step=16155: step=3.272s data=1.320s compute=3.242s [fwd=2.342s bwd=0.886s opt=0.003s] dispatch=2.468s [2026-10-20 00:06:36,885] loop.py:45 INFO TRAIN_STEP step=16155 loss=1.252839 Training: 90%|████████▉ | 16155/18018 [34:36:22<3:24:07, 6.57s/step, loss=1.253, acc=0.710, acc_len=1.72, thru=22.2, I=28.1, T=39.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:36,947] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:42,732] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:06:42,911] loop.py:45 INFO TRAIN_STEP step=16156 loss=1.175265 Training: 90%|████████▉ | 16156/18018 [34:36:28<3:18:52, 6.41s/step, loss=1.175, acc=0.722, acc_len=1.76, thru=21.9, I=25.7, T=38.7, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:06:42,906] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16155: forward=2335.4ms backward=925.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:43,233] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:48,903] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:06:49,836] loop.py:45 INFO TRAIN_STEP step=16157 loss=1.211095 Training: 90%|████████▉ | 16157/18018 [34:36:35<3:23:36, 6.56s/step, loss=1.211, acc=0.714, acc_len=1.74, thru=19.0, I=22.3, T=37.3, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:49,917] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:55,915] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:06:56,433] loop.py:45 INFO TRAIN_STEP step=16158 loss=1.100561 Training: 90%|████████▉ | 16158/18018 [34:36:42<3:23:46, 6.57s/step, loss=1.101, acc=0.730, acc_len=1.79, thru=19.5, I=21.5, T=37.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:06:56,923] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:07:03,245] loop.py:45 INFO TRAIN_STEP step=16159 loss=1.316405 Training: 90%|████████▉ | 16159/18018 [34:36:49<3:25:52, 6.64s/step, loss=1.316, acc=0.699, acc_len=1.67, thru=18.8, I=20.7, T=38.8, wait=3.5s, pool=61, epoch=2/2][2026-10-20 00:07:11,329] loop.py:380 INFO TIMING step=16160: step=3.561s data=1.422s compute=3.530s [fwd=2.533s bwd=0.982s opt=0.003s] dispatch=4.516s [2026-10-20 00:07:11,329] loop.py:45 INFO TRAIN_STEP step=16160 loss=1.257866 Training: 90%|████████▉ | 16160/18018 [34:36:57<3:39:08, 7.08s/step, loss=1.258, acc=0.711, acc_len=1.71, thru=15.5, I=14.4, T=35.9, wait=4.5s, pool=45, epoch=2/2][2026-10-20 00:07:17,718] loop.py:45 INFO TRAIN_STEP step=16161 loss=1.139525 Training: 90%|████████▉ | 16161/18018 [34:37:03<3:32:37, 6.87s/step, loss=1.140, acc=0.726, acc_len=1.77, thru=18.4, I=19.8, T=38.8, wait=3.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:07:17,713] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16160: forward=2346.9ms backward=906.7ms [2026-10-20 00:07:25,204] loop.py:45 INFO TRAIN_STEP step=16162 loss=1.181788 Training: 90%|████████▉ | 16162/18018 [34:37:11<3:38:15, 7.06s/step, loss=1.182, acc=0.721, acc_len=1.76, thru=15.9, I=15.8, T=36.8, wait=4.0s, pool=37, epoch=2/2][2026-10-20 00:07:31,953] loop.py:45 INFO TRAIN_STEP step=16163 loss=1.304310 Training: 90%|████████▉ | 16163/18018 [34:37:18<3:35:17, 6.96s/step, loss=1.304, acc=0.700, acc_len=1.68, thru=17.1, I=19.9, T=33.8, wait=2.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:07:38,970] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:07:39,752] loop.py:45 INFO TRAIN_STEP step=16164 loss=1.158775 Training: 90%|████████▉ | 16164/18018 [34:37:25<3:42:55, 7.21s/step, loss=1.159, acc=0.726, acc_len=1.77, thru=17.0, I=16.9, T=36.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:07:39,976] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:07:46,017] loop.py:380 INFO TIMING step=16165: step=3.382s data=1.406s compute=3.339s [fwd=2.374s bwd=0.950s opt=0.003s] dispatch=2.873s [2026-10-20 00:07:46,017] loop.py:45 INFO TRAIN_STEP step=16165 loss=1.202278 Training: 90%|████████▉ | 16165/18018 [34:37:32<3:33:58, 6.93s/step, loss=1.202, acc=0.715, acc_len=1.73, thru=20.1, I=21.7, T=37.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:07:45,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:07:46,457] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:07:52,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:07:53,532] loop.py:45 INFO TRAIN_STEP step=16166 loss=1.231193 Training: 90%|████████▉ | 16166/18018 [34:37:39<3:39:19, 7.11s/step, loss=1.231, acc=0.709, acc_len=1.71, thru=16.9, I=20.1, T=31.2, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:07:53,527] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16165: forward=3060.4ms backward=988.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:07:53,983] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:08:00,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:08:02,227] loop.py:45 INFO TRAIN_STEP step=16167 loss=1.221483 Training: 90%|████████▉ | 16167/18018 [34:37:48<3:53:54, 7.58s/step, loss=1.221, acc=0.712, acc_len=1.72, thru=13.5, I=16.7, T=28.8, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:08:02,733] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-20 00:08:08,855] loop.py:45 INFO TRAIN_STEP step=16168 loss=1.248146 Training: 90%|████████▉ | 16168/18018 [34:37:54<3:44:58, 7.30s/step, loss=1.248, acc=0.711, acc_len=1.71, thru=15.8, I=21.4, T=38.1, wait=3.3s, pool=57, epoch=2/2][2026-10-20 00:08:15,776] loop.py:45 INFO TRAIN_STEP step=16169 loss=1.272631 Training: 90%|████████▉ | 16169/18018 [34:38:01<3:41:22, 7.18s/step, loss=1.273, acc=0.704, acc_len=1.70, thru=16.5, I=18.9, T=35.2, wait=3.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:08:23,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:08:25,575] loop.py:380 INFO TIMING step=16170: step=4.855s data=2.081s compute=4.829s [fwd=3.848s bwd=0.967s opt=0.003s] dispatch=4.933s [2026-10-20 00:08:25,575] loop.py:45 INFO TRAIN_STEP step=16170 loss=1.324053 Training: 90%|████████▉ | 16170/18018 [34:38:11<4:05:23, 7.97s/step, loss=1.324, acc=0.691, acc_len=1.65, thru=15.7, I=16.2, T=26.4, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:08:25,900] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:08:32,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:08:32,943] loop.py:45 INFO TRAIN_STEP step=16171 loss=1.312012 Training: 90%|████████▉ | 16171/18018 [34:38:19<3:59:45, 7.79s/step, loss=1.312, acc=0.698, acc_len=1.68, thru=14.7, I=20.5, T=36.7, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:08:32,938] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16170: forward=2478.2ms backward=961.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:08:33,230] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:08:40,411] loop.py:45 INFO TRAIN_STEP step=16172 loss=1.274697 Training: 90%|████████▉ | 16172/18018 [34:38:26<3:56:38, 7.69s/step, loss=1.275, acc=0.704, acc_len=1.69, thru=12.6, I=13.6, T=36.9, wait=4.0s, pool=33, epoch=2/2][2026-10-20 00:08:50,426] loop.py:45 INFO TRAIN_STEP step=16173 loss=1.241585 Training: 90%|████████▉ | 16173/18018 [34:38:36<4:17:56, 8.39s/step, loss=1.242, acc=0.712, acc_len=1.72, thru=15.3, I=14.5, T=33.6, wait=6.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:08:56,993] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:08:57,970] loop.py:45 INFO TRAIN_STEP step=16174 loss=1.170053 Training: 90%|████████▉ | 16174/18018 [34:38:44<4:10:01, 8.14s/step, loss=1.170, acc=0.722, acc_len=1.75, thru=17.6, I=18.3, T=36.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:08:57,998] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:02,692] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:09:02,967] loop.py:380 INFO TIMING step=16175: step=3.286s data=1.279s compute=3.260s [fwd=2.338s bwd=0.909s opt=0.003s] dispatch=1.702s [2026-10-20 00:09:02,967] loop.py:45 INFO TRAIN_STEP step=16175 loss=1.181241 Training: 90%|████████▉ | 16175/18018 [34:38:49<3:40:58, 7.19s/step, loss=1.181, acc=0.720, acc_len=1.75, thru=22.1, I=27.0, T=38.9, wait=1.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:03,194] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:08,181] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:09:08,852] loop.py:45 INFO TRAIN_STEP step=16176 loss=1.187708 Training: 90%|████████▉ | 16176/18018 [34:38:54<3:28:48, 6.80s/step, loss=1.188, acc=0.721, acc_len=1.75, thru=22.9, I=23.7, T=39.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:09:08,847] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16175: forward=2292.2ms backward=873.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:09,187] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:09:15,351] loop.py:45 INFO TRAIN_STEP step=16177 loss=1.261742 Training: 90%|████████▉ | 16177/18018 [34:39:01<3:25:54, 6.71s/step, loss=1.262, acc=0.708, acc_len=1.71, thru=18.5, I=23.5, T=38.8, wait=3.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:25,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:09:25,900] loop.py:45 INFO TRAIN_STEP step=16178 loss=1.240804 Training: 90%|████████▉ | 16178/18018 [34:39:11<4:01:06, 7.86s/step, loss=1.241, acc=0.714, acc_len=1.73, thru=14.1, I=13.1, T=35.9, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:26,003] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-20 00:09:32,784] loop.py:45 INFO TRAIN_STEP step=16179 loss=1.229940 Training: 90%|████████▉ | 16179/18018 [34:39:18<3:51:59, 7.57s/step, loss=1.230, acc=0.714, acc_len=1.72, thru=15.6, I=17.6, T=36.5, wait=3.4s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:39,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:09:40,146] loop.py:380 INFO TIMING step=16180: step=3.489s data=1.424s compute=3.453s [fwd=2.456s bwd=0.983s opt=0.003s] dispatch=3.864s [2026-10-20 00:09:40,147] loop.py:45 INFO TRAIN_STEP step=16180 loss=1.282159 Training: 90%|████████▉ | 16180/18018 [34:39:26<3:49:57, 7.51s/step, loss=1.282, acc=0.701, acc_len=1.68, thru=18.5, I=19.8, T=36.7, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:40,300] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:09:46,785] loop.py:45 INFO TRAIN_STEP step=16181 loss=1.201443 Training: 90%|████████▉ | 16181/18018 [34:39:32<3:41:50, 7.25s/step, loss=1.201, acc=0.716, acc_len=1.73, thru=18.4, I=23.6, T=38.2, wait=3.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:09:46,780] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16180: forward=2402.3ms backward=907.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:54,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:09:55,312] loop.py:45 INFO TRAIN_STEP step=16182 loss=1.365373 Training: 90%|████████▉ | 16182/18018 [34:39:41<3:53:28, 7.63s/step, loss=1.365, acc=0.692, acc_len=1.65, thru=14.9, I=16.0, T=35.1, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:09:55,722] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:00,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:10:01,023] loop.py:45 INFO TRAIN_STEP step=16183 loss=1.223309 Training: 90%|████████▉ | 16183/18018 [34:39:47<3:35:45, 7.05s/step, loss=1.223, acc=0.712, acc_len=1.72, thru=21.1, I=25.9, T=39.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:01,361] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:10:07,340] loop.py:45 INFO TRAIN_STEP step=16184 loss=1.323043 Training: 90%|████████▉ | 16184/18018 [34:39:53<3:28:52, 6.83s/step, loss=1.323, acc=0.694, acc_len=1.66, thru=20.0, I=22.1, T=36.9, wait=2.8s, pool=57, epoch=2/2][2026-10-20 00:10:13,929] loop.py:380 INFO TIMING step=16185: step=3.401s data=1.325s compute=3.365s [fwd=2.435s bwd=0.915s opt=0.003s] dispatch=3.178s [2026-10-20 00:10:13,929] loop.py:45 INFO TRAIN_STEP step=16185 loss=1.107742 Training: 90%|████████▉ | 16185/18018 [34:40:00<3:26:30, 6.76s/step, loss=1.108, acc=0.733, acc_len=1.80, thru=19.1, I=19.3, T=37.6, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:20,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:10:21,395] loop.py:45 INFO TRAIN_STEP step=16186 loss=1.189301 Training: 90%|████████▉ | 16186/18018 [34:40:07<3:32:52, 6.97s/step, loss=1.189, acc=0.719, acc_len=1.74, thru=17.9, I=19.9, T=36.2, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:10:21,391] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16185: forward=2536.5ms backward=947.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:21,656] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:10:28,640] loop.py:45 INFO TRAIN_STEP step=16187 loss=1.345199 Training: 90%|████████▉ | 16187/18018 [34:40:14<3:35:15, 7.05s/step, loss=1.345, acc=0.689, acc_len=1.64, thru=16.1, I=17.9, T=36.8, wait=3.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:34,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:10:34,786] loop.py:45 INFO TRAIN_STEP step=16188 loss=1.175017 Training: 90%|████████▉ | 16188/18018 [34:40:20<3:26:50, 6.78s/step, loss=1.175, acc=0.724, acc_len=1.77, thru=20.0, I=21.9, T=38.4, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:34,968] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:10:41,449] loop.py:45 INFO TRAIN_STEP step=16189 loss=1.168197 Training: 90%|████████▉ | 16189/18018 [34:40:27<3:25:39, 6.75s/step, loss=1.168, acc=0.724, acc_len=1.76, thru=19.5, I=19.5, T=38.5, wait=3.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:47,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:10:48,508] loop.py:380 INFO TIMING step=16190: step=3.416s data=1.330s compute=3.385s [fwd=2.451s bwd=0.919s opt=0.003s] dispatch=3.633s [2026-10-20 00:10:48,508] loop.py:45 INFO TRAIN_STEP step=16190 loss=1.331775 Training: 90%|████████▉ | 16190/18018 [34:40:34<3:28:22, 6.84s/step, loss=1.332, acc=0.698, acc_len=1.67, thru=18.8, I=20.0, T=37.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:48,876] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:54,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:10:55,464] loop.py:45 INFO TRAIN_STEP step=16191 loss=1.257676 Training: 90%|████████▉ | 16191/18018 [34:40:41<3:29:19, 6.87s/step, loss=1.258, acc=0.711, acc_len=1.71, thru=19.0, I=21.6, T=36.8, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:10:55,460] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16190: forward=2475.6ms backward=949.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:10:55,534] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-20 00:11:02,855] loop.py:45 INFO TRAIN_STEP step=16192 loss=1.261039 Training: 90%|████████▉ | 16192/18018 [34:40:48<3:33:56, 7.03s/step, loss=1.261, acc=0.706, acc_len=1.71, thru=14.8, I=17.6, T=37.5, wait=4.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:11:10,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:11:11,900] loop.py:45 INFO TRAIN_STEP step=16193 loss=1.188985 Training: 90%|████████▉ | 16193/18018 [34:40:57<3:52:13, 7.63s/step, loss=1.189, acc=0.717, acc_len=1.74, thru=16.7, I=17.6, T=36.1, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:11:12,177] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:11:20,338] loop.py:45 INFO TRAIN_STEP step=16194 loss=1.358021 Training: 90%|████████▉ | 16194/18018 [34:41:06<3:59:24, 7.88s/step, loss=1.358, acc=0.685, acc_len=1.63, thru=10.8, I=14.2, T=37.0, wait=5.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:11:28,898] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:11:30,535] loop.py:380 INFO TIMING step=16195: step=3.673s data=1.471s compute=3.632s [fwd=2.518s bwd=1.099s opt=0.003s] dispatch=6.514s [2026-10-20 00:11:30,535] loop.py:45 INFO TRAIN_STEP step=16195 loss=1.342275 Training: 90%|████████▉ | 16195/18018 [34:41:16<4:20:26, 8.57s/step, loss=1.342, acc=0.692, acc_len=1.65, thru=16.0, I=17.0, T=34.9, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:11:30,915] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 00:11:36,504] loop.py:45 INFO TRAIN_STEP step=16196 loss=1.168820 Training: 90%|████████▉ | 16196/18018 [34:41:22<3:56:34, 7.79s/step, loss=1.169, acc=0.723, acc_len=1.76, thru=18.2, I=23.6, T=40.2, wait=2.8s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:11:36,499] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16195: forward=2262.8ms backward=876.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:11:43,442] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:11:44,395] loop.py:45 INFO TRAIN_STEP step=16197 loss=1.274608 Training: 90%|████████▉ | 16197/18018 [34:41:30<3:57:22, 7.82s/step, loss=1.275, acc=0.702, acc_len=1.68, thru=18.4, I=18.3, T=37.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:11:44,461] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:11:49,610] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:11:50,630] loop.py:45 INFO TRAIN_STEP step=16198 loss=1.238314 Training: 90%|████████▉ | 16198/18018 [34:41:36<3:42:47, 7.34s/step, loss=1.238, acc=0.708, acc_len=1.71, thru=20.1, I=25.0, T=39.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:11:51,123] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:11:57,572] loop.py:45 INFO TRAIN_STEP step=16199 loss=1.282042 Training: 90%|████████▉ | 16199/18018 [34:41:43<3:39:00, 7.22s/step, loss=1.282, acc=0.702, acc_len=1.68, thru=18.4, I=22.7, T=39.0, wait=3.7s, pool=61, epoch=2/2][2026-10-20 00:12:05,526] loop.py:380 INFO TIMING step=16200: step=3.468s data=1.406s compute=3.442s [fwd=2.466s bwd=0.962s opt=0.003s] dispatch=4.479s [2026-10-20 00:12:05,526] loop.py:45 INFO TRAIN_STEP step=16200 loss=1.210134 Training: 90%|████████▉ | 16200/18018 [34:41:51<3:45:31, 7.44s/step, loss=1.210, acc=0.715, acc_len=1.74, thru=15.7, I=17.3, T=36.9, wait=4.5s, pool=57, epoch=2/2][2026-10-20 00:12:11,866] loop.py:45 INFO TRAIN_STEP step=16201 loss=1.179759 Training: 90%|████████▉ | 16201/18018 [34:41:57<3:35:23, 7.11s/step, loss=1.180, acc=0.719, acc_len=1.75, thru=18.0, I=19.7, T=38.8, wait=3.0s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:12:11,861] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16200: forward=2328.5ms backward=922.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:22,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:12:22,858] loop.py:45 INFO TRAIN_STEP step=16202 loss=1.151216 Training: 90%|████████▉ | 16202/18018 [34:42:08<4:10:30, 8.28s/step, loss=1.151, acc=0.724, acc_len=1.76, thru=15.2, I=13.5, T=34.4, wait=7.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:22,916] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:28,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:12:28,765] loop.py:45 INFO TRAIN_STEP step=16203 loss=1.405249 Training: 90%|████████▉ | 16203/18018 [34:42:14<3:48:51, 7.57s/step, loss=1.405, acc=0.683, acc_len=1.62, thru=20.4, I=21.6, T=38.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:29,097] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:34,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:12:35,577] loop.py:45 INFO TRAIN_STEP step=16204 loss=1.172991 Training: 90%|████████▉ | 16204/18018 [34:42:21<3:41:54, 7.34s/step, loss=1.173, acc=0.720, acc_len=1.75, thru=19.1, I=22.0, T=34.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:35,804] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:12:42,872] loop.py:380 INFO TIMING step=16205: step=3.543s data=1.451s compute=3.518s [fwd=2.516s bwd=0.988s opt=0.003s] dispatch=3.740s [2026-10-20 00:12:42,872] loop.py:45 INFO TRAIN_STEP step=16205 loss=1.228052 Training: 90%|████████▉ | 16205/18018 [34:42:28<3:41:22, 7.33s/step, loss=1.228, acc=0.716, acc_len=1.74, thru=13.6, I=17.1, T=36.1, wait=3.7s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:52,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:12:53,441] loop.py:45 INFO TRAIN_STEP step=16206 loss=1.312084 Training: 90%|████████▉ | 16206/18018 [34:42:39<4:10:37, 8.30s/step, loss=1.312, acc=0.699, acc_len=1.67, thru=15.7, I=14.9, T=35.5, wait=6.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:12:53,436] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16205: forward=2535.6ms backward=1025.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:53,670] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:58,408] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:12:59,063] loop.py:45 INFO TRAIN_STEP step=16207 loss=1.264110 Training: 90%|████████▉ | 16207/18018 [34:42:45<3:46:14, 7.50s/step, loss=1.264, acc=0.702, acc_len=1.69, thru=21.2, I=23.3, T=39.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:12:59,413] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:07,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:13:07,945] loop.py:45 INFO TRAIN_STEP step=16208 loss=1.328843 Training: 90%|████████▉ | 16208/18018 [34:42:54<3:58:40, 7.91s/step, loss=1.329, acc=0.689, acc_len=1.64, thru=14.5, I=16.9, T=36.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:08,148] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:13,839] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:13:14,337] loop.py:45 INFO TRAIN_STEP step=16209 loss=1.300059 Training: 90%|████████▉ | 16209/18018 [34:43:00<3:44:47, 7.46s/step, loss=1.300, acc=0.699, acc_len=1.68, thru=19.6, I=20.2, T=38.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:14,843] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:19,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:13:20,677] loop.py:380 INFO TIMING step=16210: step=3.350s data=1.348s compute=3.315s [fwd=2.358s bwd=0.943s opt=0.003s] dispatch=2.979s [2026-10-20 00:13:20,677] loop.py:45 INFO TRAIN_STEP step=16210 loss=1.135464 Training: 90%|████████▉ | 16210/18018 [34:43:06<3:34:33, 7.12s/step, loss=1.135, acc=0.726, acc_len=1.77, thru=20.5, I=23.1, T=38.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:20,940] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:26,354] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:13:26,665] loop.py:45 INFO TRAIN_STEP step=16211 loss=1.260998 Training: 90%|████████▉ | 16211/18018 [34:43:12<3:24:13, 6.78s/step, loss=1.261, acc=0.706, acc_len=1.71, thru=21.1, I=23.2, T=39.2, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:13:26,660] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16210: forward=2352.9ms backward=860.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:26,856] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:13:32,151] loop.py:45 INFO TRAIN_STEP step=16212 loss=1.136329 Training: 90%|████████▉ | 16212/18018 [34:43:18<3:12:24, 6.39s/step, loss=1.136, acc=0.727, acc_len=1.77, thru=19.7, I=21.6, T=39.6, wait=2.2s, pool=49, epoch=2/2][2026-10-20 00:13:39,847] loop.py:45 INFO TRAIN_STEP step=16213 loss=1.415576 Training: 90%|████████▉ | 16213/18018 [34:43:25<3:24:04, 6.78s/step, loss=1.416, acc=0.678, acc_len=1.61, thru=15.7, I=17.8, T=35.8, wait=4.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:46,797] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:13:47,180] loop.py:45 INFO TRAIN_STEP step=16214 loss=1.357151 Training: 90%|████████▉ | 16214/18018 [34:43:33<3:28:55, 6.95s/step, loss=1.357, acc=0.692, acc_len=1.65, thru=17.7, I=18.4, T=35.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:47,299] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:51,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:13:52,578] loop.py:380 INFO TIMING step=16215: step=3.249s data=1.314s compute=3.224s [fwd=2.291s bwd=0.920s opt=0.003s] dispatch=2.138s [2026-10-20 00:13:52,578] loop.py:45 INFO TRAIN_STEP step=16215 loss=1.154982 Training: 90%|████████▉ | 16215/18018 [34:43:38<3:14:48, 6.48s/step, loss=1.155, acc=0.724, acc_len=1.76, thru=21.2, I=25.5, T=39.4, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:52,993] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:57,513] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:13:58,567] loop.py:45 INFO TRAIN_STEP step=16216 loss=1.199752 Training: 90%|████████▉ | 16216/18018 [34:43:44<3:10:14, 6.33s/step, loss=1.200, acc=0.719, acc_len=1.75, thru=21.7, I=27.1, T=38.8, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:13:58,562] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16215: forward=2332.1ms backward=918.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:13:59,021] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:14:04,389] loop.py:45 INFO TRAIN_STEP step=16217 loss=1.165296 Training: 90%|█████████ | 16217/18018 [34:43:50<3:05:32, 6.18s/step, loss=1.165, acc=0.721, acc_len=1.74, thru=20.9, I=25.0, T=39.5, wait=2.6s, pool=57, epoch=2/2][2026-10-20 00:14:12,682] loop.py:45 INFO TRAIN_STEP step=16218 loss=1.157136 Training: 90%|█████████ | 16218/18018 [34:43:58<3:24:26, 6.81s/step, loss=1.157, acc=0.730, acc_len=1.79, thru=15.9, I=16.4, T=36.8, wait=4.8s, pool=53, epoch=2/2][2026-10-20 00:14:20,176] loop.py:45 INFO TRAIN_STEP step=16219 loss=1.274338 Training: 90%|█████████ | 16219/18018 [34:44:06<3:30:25, 7.02s/step, loss=1.274, acc=0.704, acc_len=1.70, thru=17.2, I=16.8, T=36.6, wait=4.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:14:25,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:14:26,230] loop.py:380 INFO TIMING step=16220: step=3.339s data=1.355s compute=3.312s [fwd=2.380s bwd=0.918s opt=0.003s] dispatch=2.706s [2026-10-20 00:14:26,231] loop.py:45 INFO TRAIN_STEP step=16220 loss=1.201762 Training: 90%|█████████ | 16220/18018 [34:44:12<3:21:38, 6.73s/step, loss=1.202, acc=0.715, acc_len=1.73, thru=20.9, I=21.9, T=38.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:14:26,274] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-20 00:14:31,793] loop.py:45 INFO TRAIN_STEP step=16221 loss=1.299831 (TrainerActor pid=9310) [2026-10-20 00:14:31,788] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16220: forward=2370.1ms backward=890.8ms Training: 90%|█████████ | 16221/18018 [34:44:17<3:11:05, 6.38s/step, loss=1.300, acc=0.697, acc_len=1.67, thru=20.4, I=23.1, T=38.7, wait=2.2s, pool=45, epoch=2/2][2026-10-20 00:14:39,194] loop.py:45 INFO TRAIN_STEP step=16222 loss=1.248609 Training: 90%|█████████ | 16222/18018 [34:44:25<3:20:06, 6.69s/step, loss=1.249, acc=0.706, acc_len=1.71, thru=14.0, I=17.6, T=36.8, wait=3.9s, pool=25, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 00:14:39,286] eagle_store.py:517 WARNING batch_get_buffer missing keys (k519102-88fe8ab0_hs, k519102-88fe8ab0_ids, k519102-88fe8ab0_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:14:48,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-20 00:14:39,277] eagle_store.py:517 WARNING batch_get_buffer missing keys (k519101-93d537d8_hs, k519101-93d537d8_ids, k519101-93d537d8_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-20 00:14:49,289] loop.py:45 INFO TRAIN_STEP step=16223 loss=1.276190 Training: 90%|█████████ | 16223/18018 [34:44:35<3:50:36, 7.71s/step, loss=1.276, acc=0.701, acc_len=1.68, thru=17.5, I=15.2, T=34.6, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:14:49,445] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:14:56,316] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:14:56,669] loop.py:45 INFO TRAIN_STEP step=16224 loss=1.228805 Training: 90%|█████████ | 16224/18018 [34:44:42<3:47:32, 7.61s/step, loss=1.229, acc=0.715, acc_len=1.72, thru=17.1, I=19.2, T=37.2, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:14:56,831] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:15:02,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:15:02,960] loop.py:380 INFO TIMING step=16225: step=3.384s data=1.396s compute=3.346s [fwd=2.389s bwd=0.942s opt=0.003s] dispatch=2.898s [2026-10-20 00:15:02,961] loop.py:45 INFO TRAIN_STEP step=16225 loss=1.319636 Training: 90%|█████████ | 16225/18018 [34:44:49<3:35:36, 7.21s/step, loss=1.320, acc=0.695, acc_len=1.67, thru=20.1, I=20.0, T=37.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:15:03,152] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:15:10,081] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:15:11,141] loop.py:45 INFO TRAIN_STEP step=16226 loss=1.267877 Training: 90%|█████████ | 16226/18018 [34:44:57<3:44:07, 7.50s/step, loss=1.268, acc=0.703, acc_len=1.69, thru=16.5, I=18.2, T=36.3, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:15:11,137] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16225: forward=2446.8ms backward=1025.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:15:11,591] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:15:18,784] loop.py:45 INFO TRAIN_STEP step=16227 loss=1.185183 Training: 90%|█████████ | 16227/18018 [34:45:04<3:45:13, 7.55s/step, loss=1.185, acc=0.720, acc_len=1.75, thru=14.5, I=17.9, T=37.9, wait=4.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:15:26,301] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:15:26,949] loop.py:45 INFO TRAIN_STEP step=16228 loss=1.291975 Training: 90%|█████████ | 16228/18018 [34:45:13<3:50:39, 7.73s/step, loss=1.292, acc=0.702, acc_len=1.69, thru=15.3, I=16.1, T=34.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:15:27,308] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:15:35,350] loop.py:45 INFO TRAIN_STEP step=16229 loss=1.252936 Training: 90%|█████████ | 16229/18018 [34:45:21<3:56:30, 7.93s/step, loss=1.253, acc=0.704, acc_len=1.69, thru=13.7, I=14.9, T=36.7, wait=4.9s, pool=53, epoch=2/2][2026-10-20 00:15:42,492] loop.py:380 INFO TIMING step=16230: step=3.337s data=1.318s compute=3.301s [fwd=2.370s bwd=0.916s opt=0.003s] dispatch=3.797s [2026-10-20 00:15:42,492] loop.py:45 INFO TRAIN_STEP step=16230 loss=1.247090 Training: 90%|█████████ | 16230/18018 [34:45:28<3:49:18, 7.69s/step, loss=1.247, acc=0.707, acc_len=1.71, thru=18.1, I=16.8, T=38.4, wait=3.8s, pool=57, epoch=2/2][2026-10-20 00:15:50,646] loop.py:45 INFO TRAIN_STEP step=16231 loss=1.228884 Training: 90%|█████████ | 16231/18018 [34:45:36<3:53:18, 7.83s/step, loss=1.229, acc=0.712, acc_len=1.72, thru=15.4, I=14.1, T=36.3, wait=4.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:15:50,641] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16230: forward=2475.1ms backward=992.9ms [2026-10-20 00:15:59,725] loop.py:45 INFO TRAIN_STEP step=16232 loss=1.152163 Training: 90%|█████████ | 16232/18018 [34:45:45<4:04:17, 8.21s/step, loss=1.152, acc=0.725, acc_len=1.77, thru=15.7, I=15.5, T=35.2, wait=5.4s, pool=57, epoch=2/2][2026-10-20 00:16:07,234] loop.py:45 INFO TRAIN_STEP step=16233 loss=1.308599 Training: 90%|█████████ | 16233/18018 [34:45:53<3:57:55, 8.00s/step, loss=1.309, acc=0.697, acc_len=1.67, thru=15.9, I=16.4, T=37.7, wait=4.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:16:16,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:16:17,349] loop.py:45 INFO TRAIN_STEP step=16234 loss=1.327422 Training: 90%|█████████ | 16234/18018 [34:46:03<4:16:41, 8.63s/step, loss=1.327, acc=0.696, acc_len=1.66, thru=14.8, I=15.1, T=35.1, wait=6.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:16:17,808] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:16:22,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:16:23,393] loop.py:380 INFO TIMING step=16235: step=3.277s data=1.355s compute=3.246s [fwd=2.315s bwd=0.916s opt=0.003s] dispatch=2.758s [2026-10-20 00:16:23,393] loop.py:45 INFO TRAIN_STEP step=16235 loss=1.282740 Training: 90%|█████████ | 16235/18018 [34:46:09<3:53:27, 7.86s/step, loss=1.283, acc=0.701, acc_len=1.68, thru=20.9, I=25.9, T=39.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:16:23,439] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-20 00:16:31,540] loop.py:45 INFO TRAIN_STEP step=16236 loss=1.312330 Training: 90%|█████████ | 16236/18018 [34:46:17<3:55:56, 7.94s/step, loss=1.312, acc=0.697, acc_len=1.67, thru=12.9, I=14.2, T=36.7, wait=4.7s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:16:31,535] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16235: forward=2379.7ms backward=1018.7ms [2026-10-20 00:16:41,780] loop.py:45 INFO TRAIN_STEP step=16237 loss=1.167699 Training: 90%|█████████ | 16237/18018 [34:46:27<4:16:12, 8.63s/step, loss=1.168, acc=0.726, acc_len=1.77, thru=11.6, I=12.2, T=33.9, wait=6.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:16:50,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:16:51,038] loop.py:45 INFO TRAIN_STEP step=16238 loss=1.258437 Training: 90%|█████████ | 16238/18018 [34:46:37<4:21:39, 8.82s/step, loss=1.258, acc=0.707, acc_len=1.70, thru=17.6, I=15.8, T=34.5, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:16:51,107] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-20 00:16:58,792] loop.py:45 INFO TRAIN_STEP step=16239 loss=1.305368 Training: 90%|█████████ | 16239/18018 [34:46:44<4:12:02, 8.50s/step, loss=1.305, acc=0.697, acc_len=1.67, thru=15.9, I=16.0, T=36.6, wait=4.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:17:03,858] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:17:05,097] loop.py:380 INFO TIMING step=16240: step=3.408s data=1.381s compute=3.379s [fwd=2.414s bwd=0.951s opt=0.003s] dispatch=2.889s [2026-10-20 00:17:05,098] loop.py:45 INFO TRAIN_STEP step=16240 loss=1.295706 Training: 90%|█████████ | 16240/18018 [34:46:51<3:52:23, 7.84s/step, loss=1.296, acc=0.704, acc_len=1.69, thru=19.9, I=21.2, T=37.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:17:05,367] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:17:13,173] loop.py:45 INFO TRAIN_STEP step=16241 loss=1.188220 Training: 90%|█████████ | 16241/18018 [34:46:59<3:54:18, 7.91s/step, loss=1.188, acc=0.717, acc_len=1.74, thru=13.6, I=16.0, T=35.8, wait=4.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:17:13,168] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16240: forward=2473.3ms backward=1055.2ms [2026-10-20 00:17:19,224] loop.py:45 INFO TRAIN_STEP step=16242 loss=1.172395 Training: 90%|█████████ | 16242/18018 [34:47:05<3:37:39, 7.35s/step, loss=1.172, acc=0.723, acc_len=1.77, thru=19.9, I=17.7, T=39.8, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:17:26,592] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:17:26,967] loop.py:45 INFO TRAIN_STEP step=16243 loss=1.382640 Training: 90%|█████████ | 16243/18018 [34:47:13<3:41:00, 7.47s/step, loss=1.383, acc=0.685, acc_len=1.63, thru=17.3, I=18.4, T=36.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:17:27,104] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:17:32,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:17:33,011] loop.py:45 INFO TRAIN_STEP step=16244 loss=1.291083 Training: 90%|█████████ | 16244/18018 [34:47:19<3:28:12, 7.04s/step, loss=1.291, acc=0.699, acc_len=1.68, thru=20.2, I=22.8, T=38.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:17:33,493] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:17:38,175] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:17:38,795] loop.py:380 INFO TIMING step=16245: step=3.351s data=1.312s compute=3.316s [fwd=2.357s bwd=0.945s opt=0.003s] dispatch=2.426s [2026-10-20 00:17:38,796] loop.py:45 INFO TRAIN_STEP step=16245 loss=1.334091 Training: 90%|█████████ | 16245/18018 [34:47:24<3:16:57, 6.67s/step, loss=1.334, acc=0.692, acc_len=1.65, thru=22.0, I=27.4, T=38.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:17:39,179] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:17:46,020] loop.py:45 INFO TRAIN_STEP step=16246 loss=1.203395 Training: 90%|█████████ | 16246/18018 [34:47:32<3:21:49, 6.83s/step, loss=1.203, acc=0.713, acc_len=1.73, thru=13.4, I=19.3, T=37.5, wait=3.8s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:17:46,015] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16245: forward=2451.0ms backward=918.5ms [2026-10-20 00:17:58,923] loop.py:45 INFO TRAIN_STEP step=16247 loss=1.260988 Training: 90%|█████████ | 16247/18018 [34:47:45<4:15:25, 8.65s/step, loss=1.261, acc=0.707, acc_len=1.71, thru=12.8, I=9.8, T=35.4, wait=9.3s, pool=33, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:18:08,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:18:09,398] loop.py:45 INFO TRAIN_STEP step=16248 loss=1.422255 Training: 90%|█████████ | 16248/18018 [34:47:55<4:31:23, 9.20s/step, loss=1.422, acc=0.678, acc_len=1.60, thru=15.3, I=15.6, T=32.9, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:18:09,773] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:18:16,581] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:18:16,819] loop.py:45 INFO TRAIN_STEP step=16249 loss=1.312472 Training: 90%|█████████ | 16249/18018 [34:48:02<4:15:31, 8.67s/step, loss=1.312, acc=0.698, acc_len=1.68, thru=15.6, I=17.0, T=37.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:18:17,085] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:18:23,321] loop.py:380 INFO TIMING step=16250: step=3.532s data=1.484s compute=3.488s [fwd=2.448s bwd=1.025s opt=0.003s] dispatch=2.960s [2026-10-20 00:18:23,321] loop.py:45 INFO TRAIN_STEP step=16250 loss=1.365805 Training: 90%|█████████ | 16250/18018 [34:48:09<3:56:13, 8.02s/step, loss=1.366, acc=0.685, acc_len=1.63, thru=18.1, I=17.1, T=36.2, wait=3.0s, pool=49, epoch=2/2][2026-10-20 00:18:23,323] loop.py:419 INFO Saving checkpoint at step 16250... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:18:24,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 00:18:26,692] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0016251 (TrainerActor pid=9310) [2026-10-20 00:18:26,929] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0015501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:18:28,069] inference_manager.py:408 INFO Pool capacity available after 4.0s, pool_size=1/64 [2026-10-20 00:18:34,264] loop.py:45 INFO TRAIN_STEP step=16251 loss=1.353542 Training: 90%|█████████ | 16251/18018 [34:48:20<4:21:58, 8.90s/step, loss=1.354, acc=0.693, acc_len=1.66, thru=20.0, I=22.3, T=37.5, wait=3.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:18:34,259] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16250: forward=2419.9ms backward=950.6ms (TrainerActor pid=9310) [2026-10-20 00:18:26,691] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0016251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 00:18:27,633] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0015501 [2026-10-20 00:18:40,894] loop.py:45 INFO TRAIN_STEP step=16252 loss=1.162177 Training: 90%|█████████ | 16252/18018 [34:48:26<4:01:47, 8.22s/step, loss=1.162, acc=0.724, acc_len=1.76, thru=18.9, I=19.3, T=37.4, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:18:46,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:18:47,815] loop.py:45 INFO TRAIN_STEP step=16253 loss=1.132655 Training: 90%|█████████ | 16253/18018 [34:48:33<3:50:14, 7.83s/step, loss=1.133, acc=0.733, acc_len=1.79, thru=18.8, I=20.9, T=37.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:18:48,229] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:18:55,034] loop.py:45 INFO TRAIN_STEP step=16254 loss=1.181033 Training: 90%|█████████ | 16254/18018 [34:48:41<3:44:45, 7.65s/step, loss=1.181, acc=0.719, acc_len=1.75, thru=15.4, I=17.4, T=38.3, wait=3.9s, pool=45, epoch=2/2][2026-10-20 00:19:03,903] loop.py:380 INFO TIMING step=16255: step=3.575s data=1.448s compute=3.547s [fwd=2.543s bwd=0.989s opt=0.003s] dispatch=5.284s [2026-10-20 00:19:03,904] loop.py:45 INFO TRAIN_STEP step=16255 loss=1.272781 Training: 90%|█████████ | 16255/18018 [34:48:49<3:55:24, 8.01s/step, loss=1.273, acc=0.705, acc_len=1.70, thru=16.1, I=14.9, T=35.8, wait=5.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:10,147] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:19:10,298] loop.py:45 INFO TRAIN_STEP step=16256 loss=1.376714 Training: 90%|█████████ | 16256/18018 [34:48:56<3:41:02, 7.53s/step, loss=1.377, acc=0.681, acc_len=1.62, thru=19.3, I=19.6, T=37.5, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:19:10,294] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16255: forward=2456.7ms backward=909.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:10,649] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:17,661] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:19:18,284] loop.py:45 INFO TRAIN_STEP step=16257 loss=1.308507 Training: 90%|█████████ | 16257/18018 [34:49:04<3:44:56, 7.66s/step, loss=1.309, acc=0.699, acc_len=1.68, thru=17.0, I=18.1, T=37.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:18,665] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:19:26,289] loop.py:45 INFO TRAIN_STEP step=16258 loss=1.362668 Training: 90%|█████████ | 16258/18018 [34:49:12<3:47:49, 7.77s/step, loss=1.363, acc=0.689, acc_len=1.64, thru=14.5, I=16.3, T=37.9, wait=4.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:32,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:19:33,035] loop.py:45 INFO TRAIN_STEP step=16259 loss=1.244803 Training: 90%|█████████ | 16259/18018 [34:49:19<3:38:43, 7.46s/step, loss=1.245, acc=0.709, acc_len=1.71, thru=18.8, I=20.4, T=35.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:33,227] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:39,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:19:40,643] loop.py:380 INFO TIMING step=16260: step=3.491s data=1.457s compute=3.383s [fwd=2.404s bwd=0.964s opt=0.003s] dispatch=4.105s [2026-10-20 00:19:40,643] loop.py:45 INFO TRAIN_STEP step=16260 loss=1.248756 Training: 90%|█████████ | 16260/18018 [34:49:26<3:39:52, 7.50s/step, loss=1.249, acc=0.709, acc_len=1.71, thru=16.8, I=20.1, T=36.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:40,669] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:49,295] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:19:49,498] loop.py:45 INFO TRAIN_STEP step=16261 loss=1.211210 Training: 90%|█████████ | 16261/18018 [34:49:35<3:51:37, 7.91s/step, loss=1.211, acc=0.715, acc_len=1.72, thru=14.0, I=15.6, T=31.4, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:19:49,493] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16260: forward=3111.5ms backward=922.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:19:49,798] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:19:59,519] loop.py:45 INFO TRAIN_STEP step=16262 loss=1.296132 Training: 90%|█████████ | 16262/18018 [34:49:45<4:10:02, 8.54s/step, loss=1.296, acc=0.701, acc_len=1.68, thru=11.5, I=11.6, T=33.2, wait=6.2s, pool=49, epoch=2/2][2026-10-20 00:20:06,285] loop.py:45 INFO TRAIN_STEP step=16263 loss=1.206912 Training: 90%|█████████ | 16263/18018 [34:49:52<3:54:17, 8.01s/step, loss=1.207, acc=0.715, acc_len=1.73, thru=18.0, I=16.1, T=37.0, wait=3.3s, pool=61, epoch=2/2][2026-10-20 00:20:13,473] loop.py:45 INFO TRAIN_STEP step=16264 loss=1.323466 Training: 90%|█████████ | 16264/18018 [34:49:59<3:46:56, 7.76s/step, loss=1.323, acc=0.695, acc_len=1.66, thru=16.3, I=15.6, T=39.2, wait=3.9s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:21,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:20:21,824] loop.py:380 INFO TIMING step=16265: step=3.571s data=1.423s compute=3.538s [fwd=2.551s bwd=0.972s opt=0.003s] dispatch=4.771s [2026-10-20 00:20:21,824] loop.py:45 INFO TRAIN_STEP step=16265 loss=1.246253 Training: 90%|█████████ | 16265/18018 [34:50:07<3:51:58, 7.94s/step, loss=1.246, acc=0.706, acc_len=1.70, thru=18.8, I=16.4, T=35.8, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:21,920] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:27,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:20:28,225] loop.py:45 INFO TRAIN_STEP step=16266 loss=1.345024 Training: 90%|█████████ | 16266/18018 [34:50:14<3:38:22, 7.48s/step, loss=1.345, acc=0.690, acc_len=1.64, thru=19.5, I=20.5, T=37.8, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:20:28,222] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16265: forward=2415.2ms backward=922.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:28,392] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:20:34,371] loop.py:45 INFO TRAIN_STEP step=16267 loss=1.272047 Training: 90%|█████████ | 16267/18018 [34:50:20<3:26:34, 7.08s/step, loss=1.272, acc=0.705, acc_len=1.69, thru=20.4, I=21.1, T=39.5, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:41,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:20:42,308] loop.py:45 INFO TRAIN_STEP step=16268 loss=1.173576 Training: 90%|█████████ | 16268/18018 [34:50:28<3:33:57, 7.34s/step, loss=1.174, acc=0.724, acc_len=1.77, thru=18.2, I=18.6, T=35.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:42,678] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:48,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:20:48,567] loop.py:45 INFO TRAIN_STEP step=16269 loss=1.262055 Training: 90%|█████████ | 16269/18018 [34:50:34<3:24:25, 7.01s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=19.7, I=21.9, T=38.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:48,929] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:54,579] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:20:55,153] loop.py:380 INFO TIMING step=16270: step=3.331s data=1.310s compute=3.302s [fwd=2.346s bwd=0.941s opt=0.003s] dispatch=3.246s [2026-10-20 00:20:55,153] loop.py:45 INFO TRAIN_STEP step=16270 loss=1.268187 Training: 90%|█████████ | 16270/18018 [34:50:41<3:20:35, 6.89s/step, loss=1.268, acc=0.707, acc_len=1.70, thru=19.7, I=20.8, T=38.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:20:55,584] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:21:01,117] loop.py:45 INFO TRAIN_STEP step=16271 loss=1.159270 Training: 90%|█████████ | 16271/18018 [34:50:47<3:12:25, 6.61s/step, loss=1.159, acc=0.725, acc_len=1.77, thru=19.9, I=24.5, T=38.7, wait=2.6s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:21:01,111] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16270: forward=2350.3ms backward=913.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:21:08,209] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:21:09,185] loop.py:45 INFO TRAIN_STEP step=16272 loss=1.183317 Training: 90%|█████████ | 16272/18018 [34:50:55<3:25:02, 7.05s/step, loss=1.183, acc=0.716, acc_len=1.73, thru=16.2, I=17.8, T=36.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:21:09,212] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-20 00:21:16,025] loop.py:45 INFO TRAIN_STEP step=16273 loss=1.342651 Training: 90%|█████████ | 16273/18018 [34:51:02<3:23:07, 6.98s/step, loss=1.343, acc=0.692, acc_len=1.64, thru=16.0, I=20.0, T=38.0, wait=3.5s, pool=45, epoch=2/2][2026-10-20 00:21:23,267] loop.py:45 INFO TRAIN_STEP step=16274 loss=1.295478 Training: 90%|█████████ | 16274/18018 [34:51:09<3:25:15, 7.06s/step, loss=1.295, acc=0.697, acc_len=1.67, thru=17.3, I=17.3, T=38.5, wait=3.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:21:28,890] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:21:29,678] loop.py:380 INFO TIMING step=16275: step=3.430s data=1.352s compute=3.401s [fwd=2.456s bwd=0.931s opt=0.003s] dispatch=2.973s [2026-10-20 00:21:29,679] loop.py:45 INFO TRAIN_STEP step=16275 loss=1.229405 Training: 90%|█████████ | 16275/18018 [34:51:15<3:19:28, 6.87s/step, loss=1.229, acc=0.713, acc_len=1.72, thru=20.1, I=21.1, T=37.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:21:29,896] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:21:36,810] loop.py:45 INFO TRAIN_STEP step=16276 loss=1.302759 Training: 90%|█████████ | 16276/18018 [34:51:22<3:21:39, 6.95s/step, loss=1.303, acc=0.701, acc_len=1.68, thru=16.9, I=19.6, T=38.4, wait=3.8s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:21:36,806] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16275: forward=2385.8ms backward=905.0ms [2026-10-20 00:21:47,589] loop.py:45 INFO TRAIN_STEP step=16277 loss=1.228017 Training: 90%|█████████ | 16277/18018 [34:51:33<3:54:57, 8.10s/step, loss=1.228, acc=0.712, acc_len=1.72, thru=11.2, I=12.0, T=33.7, wait=7.0s, pool=45, epoch=2/2][2026-10-20 00:21:58,097] loop.py:45 INFO TRAIN_STEP step=16278 loss=1.286665 Training: 90%|█████████ | 16278/18018 [34:51:44<4:15:46, 8.82s/step, loss=1.287, acc=0.701, acc_len=1.68, thru=12.4, I=12.7, T=33.8, wait=6.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:05,666] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:22:07,722] loop.py:45 INFO TRAIN_STEP step=16279 loss=1.213168 Training: 90%|█████████ | 16279/18018 [34:51:53<4:22:36, 9.06s/step, loss=1.213, acc=0.714, acc_len=1.73, thru=15.5, I=16.9, T=27.4, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:08,184] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:13,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:22:14,136] loop.py:380 INFO TIMING step=16280: step=3.233s data=1.276s compute=3.208s [fwd=2.320s bwd=0.874s opt=0.003s] dispatch=3.173s [2026-10-20 00:22:14,136] loop.py:45 INFO TRAIN_STEP step=16280 loss=1.259999 Training: 90%|█████████ | 16280/18018 [34:52:00<3:59:27, 8.27s/step, loss=1.260, acc=0.706, acc_len=1.70, thru=17.5, I=22.6, T=39.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:14,434] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-20 00:22:20,182] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16280: forward=2377.2ms backward=888.9ms [2026-10-20 00:22:20,186] loop.py:45 INFO TRAIN_STEP step=16281 loss=1.218006 Training: 90%|█████████ | 16281/18018 [34:52:06<3:40:04, 7.60s/step, loss=1.218, acc=0.713, acc_len=1.72, thru=19.4, I=21.3, T=38.6, wait=2.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:25,923] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:22:26,996] loop.py:45 INFO TRAIN_STEP step=16282 loss=1.243307 Training: 90%|█████████ | 16282/18018 [34:52:13<3:33:04, 7.36s/step, loss=1.243, acc=0.710, acc_len=1.72, thru=19.5, I=21.0, T=37.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:27,432] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:32,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:22:32,863] loop.py:45 INFO TRAIN_STEP step=16283 loss=1.207255 Training: 90%|█████████ | 16283/18018 [34:52:18<3:19:57, 6.92s/step, loss=1.207, acc=0.716, acc_len=1.74, thru=21.7, I=26.5, T=35.9, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:33,127] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:40,623] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:22:41,834] loop.py:45 INFO TRAIN_STEP step=16284 loss=1.193193 Training: 90%|█████████ | 16284/18018 [34:52:27<3:37:40, 7.53s/step, loss=1.193, acc=0.719, acc_len=1.75, thru=13.8, I=16.6, T=29.5, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:42,132] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:46,324] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:22:47,995] loop.py:380 INFO TIMING step=16285: step=3.987s data=1.958s compute=3.955s [fwd=3.011s bwd=0.929s opt=0.003s] dispatch=2.167s [2026-10-20 00:22:47,995] loop.py:45 INFO TRAIN_STEP step=16285 loss=1.234202 Training: 90%|█████████ | 16285/18018 [34:52:34<3:25:39, 7.12s/step, loss=1.234, acc=0.709, acc_len=1.71, thru=18.6, I=25.1, T=32.1, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:48,338] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 00:22:54,562] loop.py:45 INFO TRAIN_STEP step=16286 loss=1.318365 Training: 90%|█████████ | 16286/18018 [34:52:40<3:20:44, 6.95s/step, loss=1.318, acc=0.695, acc_len=1.66, thru=18.0, I=22.7, T=38.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:54,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 00:22:54,558] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16285: forward=2344.2ms backward=923.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:22:55,055] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:23:01,689] loop.py:45 INFO TRAIN_STEP step=16287 loss=1.217956 Training: 90%|█████████ | 16287/18018 [34:52:47<3:22:08, 7.01s/step, loss=1.218, acc=0.714, acc_len=1.73, thru=16.1, I=18.9, T=34.0, wait=3.4s, pool=41, epoch=2/2][2026-10-20 00:23:10,845] loop.py:45 INFO TRAIN_STEP step=16288 loss=1.289908 Training: 90%|█████████ | 16288/18018 [34:52:56<3:40:36, 7.65s/step, loss=1.290, acc=0.702, acc_len=1.68, thru=15.6, I=14.4, T=35.3, wait=5.5s, pool=57, epoch=2/2][2026-10-20 00:23:17,946] loop.py:45 INFO TRAIN_STEP step=16289 loss=1.224903 Training: 90%|█████████ | 16289/18018 [34:53:04<3:35:43, 7.49s/step, loss=1.225, acc=0.711, acc_len=1.72, thru=17.0, I=18.5, T=37.4, wait=3.7s, pool=45, epoch=2/2][2026-10-20 00:23:29,749] loop.py:380 INFO TIMING step=16290: step=3.892s data=1.598s compute=3.856s [fwd=2.783s bwd=1.058s opt=0.003s] dispatch=7.904s [2026-10-20 00:23:29,750] loop.py:45 INFO TRAIN_STEP step=16290 loss=1.317389 Training: 90%|█████████ | 16290/18018 [34:53:15<4:12:53, 8.78s/step, loss=1.317, acc=0.696, acc_len=1.66, thru=12.2, I=9.1, T=32.9, wait=7.9s, pool=37, epoch=2/2] (TrainerActor pid=9310) [2026-10-20 00:23:40,349] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16290: forward=2659.7ms backward=1075.2ms [2026-10-20 00:23:40,354] loop.py:45 INFO TRAIN_STEP step=16291 loss=1.335697 Training: 90%|█████████ | 16291/18018 [34:53:26<4:28:29, 9.33s/step, loss=1.336, acc=0.693, acc_len=1.66, thru=12.5, I=14.5, T=33.8, wait=6.8s, pool=49, epoch=2/2][2026-10-20 00:23:48,111] loop.py:45 INFO TRAIN_STEP step=16292 loss=1.164969 Training: 90%|█████████ | 16292/18018 [34:53:34<4:14:46, 8.86s/step, loss=1.165, acc=0.727, acc_len=1.77, thru=16.2, I=15.5, T=36.8, wait=4.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:23:55,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:23:56,259] loop.py:45 INFO TRAIN_STEP step=16293 loss=1.398343 Training: 90%|█████████ | 16293/18018 [34:53:42<4:08:30, 8.64s/step, loss=1.398, acc=0.682, acc_len=1.62, thru=16.8, I=18.8, T=34.7, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:23:56,618] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:24:04,848] loop.py:45 INFO TRAIN_STEP step=16294 loss=1.279748 Training: 90%|█████████ | 16294/18018 [34:53:50<4:07:54, 8.63s/step, loss=1.280, acc=0.702, acc_len=1.69, thru=12.7, I=15.4, T=37.4, wait=5.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:11,866] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:24:13,474] loop.py:380 INFO TIMING step=16295: step=3.733s data=1.535s compute=3.698s [fwd=2.622s bwd=1.060s opt=0.003s] dispatch=4.886s [2026-10-20 00:24:13,474] loop.py:45 INFO TRAIN_STEP step=16295 loss=1.228639 Training: 90%|█████████ | 16295/18018 [34:53:59<4:07:44, 8.63s/step, loss=1.229, acc=0.713, acc_len=1.72, thru=15.8, I=18.4, T=34.3, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:13,876] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:19,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:24:20,568] loop.py:45 INFO TRAIN_STEP step=16296 loss=1.215996 Training: 90%|█████████ | 16296/18018 [34:54:06<3:54:24, 8.17s/step, loss=1.216, acc=0.714, acc_len=1.73, thru=17.0, I=22.8, T=36.8, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:24:20,563] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16295: forward=2469.2ms backward=956.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:20,923] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:26,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:24:27,101] loop.py:45 INFO TRAIN_STEP step=16297 loss=1.291176 Training: 90%|█████████ | 16297/18018 [34:54:13<3:40:11, 7.68s/step, loss=1.291, acc=0.698, acc_len=1.68, thru=19.8, I=23.0, T=39.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:27,550] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:24:34,084] loop.py:45 INFO TRAIN_STEP step=16298 loss=1.282807 Training: 90%|█████████ | 16298/18018 [34:54:20<3:34:06, 7.47s/step, loss=1.283, acc=0.701, acc_len=1.68, thru=18.1, I=22.2, T=38.5, wait=3.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:39,650] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:24:40,815] loop.py:45 INFO TRAIN_STEP step=16299 loss=1.249036 Training: 90%|█████████ | 16299/18018 [34:54:26<3:27:38, 7.25s/step, loss=1.249, acc=0.711, acc_len=1.72, thru=19.5, I=21.1, T=37.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:41,159] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:24:46,572] loop.py:380 INFO TIMING step=16300: step=3.197s data=1.257s compute=3.164s [fwd=2.257s bwd=0.893s opt=0.003s] dispatch=2.551s [2026-10-20 00:24:46,573] loop.py:45 INFO TRAIN_STEP step=16300 loss=1.209755 Training: 90%|█████████ | 16300/18018 [34:54:32<3:14:42, 6.80s/step, loss=1.210, acc=0.715, acc_len=1.74, thru=19.0, I=22.5, T=40.0, wait=2.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:54,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:24:55,603] loop.py:45 INFO TRAIN_STEP step=16301 loss=1.244154 Training: 90%|█████████ | 16301/18018 [34:54:41<3:33:45, 7.47s/step, loss=1.244, acc=0.711, acc_len=1.71, thru=17.5, I=17.5, T=34.1, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:24:55,599] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16300: forward=2441.9ms backward=1070.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:24:55,750] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:01,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:25:01,855] loop.py:45 INFO TRAIN_STEP step=16302 loss=1.265154 Training: 90%|█████████ | 16302/18018 [34:54:47<3:23:11, 7.10s/step, loss=1.265, acc=0.702, acc_len=1.68, thru=19.7, I=21.4, T=37.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:02,105] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:09,066] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:25:09,169] loop.py:45 INFO TRAIN_STEP step=16303 loss=1.377046 Training: 90%|█████████ | 16303/18018 [34:54:55<3:24:50, 7.17s/step, loss=1.377, acc=0.689, acc_len=1.64, thru=18.6, I=19.1, T=37.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:09,567] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:14,322] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:25:15,789] loop.py:45 INFO TRAIN_STEP step=16304 loss=1.247883 Training: 90%|█████████ | 16304/18018 [34:55:01<3:20:04, 7.00s/step, loss=1.248, acc=0.710, acc_len=1.71, thru=20.5, I=23.2, T=34.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:15,862] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:20,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:25:21,435] loop.py:380 INFO TIMING step=16305: step=3.311s data=1.214s compute=3.251s [fwd=2.292s bwd=0.945s opt=0.003s] dispatch=2.324s [2026-10-20 00:25:21,435] loop.py:45 INFO TRAIN_STEP step=16305 loss=1.166267 Training: 90%|█████████ | 16305/18018 [34:55:07<3:08:18, 6.60s/step, loss=1.166, acc=0.723, acc_len=1.76, thru=21.8, I=25.8, T=38.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:21,812] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:28,092] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:25:29,133] loop.py:45 INFO TRAIN_STEP step=16306 loss=1.192947 Training: 90%|█████████ | 16306/18018 [34:55:15<3:17:37, 6.93s/step, loss=1.193, acc=0.719, acc_len=1.74, thru=16.4, I=21.4, T=38.8, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:25:29,129] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16305: forward=2346.5ms backward=905.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:25:29,600] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:25:35,769] loop.py:45 INFO TRAIN_STEP step=16307 loss=1.302236 Training: 91%|█████████ | 16307/18018 [34:55:21<3:15:02, 6.84s/step, loss=1.302, acc=0.697, acc_len=1.67, thru=17.2, I=20.9, T=39.4, wait=3.4s, pool=41, epoch=2/2][2026-10-20 00:25:43,296] loop.py:45 INFO TRAIN_STEP step=16308 loss=1.219911 Training: 91%|█████████ | 16308/18018 [34:55:29<3:20:47, 7.05s/step, loss=1.220, acc=0.713, acc_len=1.72, thru=18.0, I=17.0, T=36.8, wait=4.0s, pool=53, epoch=2/2][2026-10-20 00:25:52,507] loop.py:45 INFO TRAIN_STEP step=16309 loss=1.311620 Training: 91%|█████████ | 16309/18018 [34:55:38<3:39:10, 7.69s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=14.0, I=14.6, T=37.4, wait=5.8s, pool=53, epoch=2/2][2026-10-20 00:25:59,537] loop.py:380 INFO TIMING step=16310: step=3.493s data=1.437s compute=3.464s [fwd=2.500s bwd=0.950s opt=0.003s] dispatch=3.529s [2026-10-20 00:25:59,537] loop.py:45 INFO TRAIN_STEP step=16310 loss=1.293818 Training: 91%|█████████ | 16310/18018 [34:55:45<3:33:21, 7.50s/step, loss=1.294, acc=0.700, acc_len=1.67, thru=18.2, I=18.5, T=36.6, wait=3.5s, pool=57, epoch=2/2][2026-10-20 00:26:08,764] loop.py:45 INFO TRAIN_STEP step=16311 loss=1.341455 Training: 91%|█████████ | 16311/18018 [34:55:54<3:48:02, 8.02s/step, loss=1.341, acc=0.698, acc_len=1.67, thru=14.7, I=14.3, T=35.6, wait=5.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:26:08,760] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16310: forward=2541.7ms backward=1003.5ms [2026-10-20 00:26:17,362] loop.py:45 INFO TRAIN_STEP step=16312 loss=1.357396 Training: 91%|█████████ | 16312/18018 [34:56:03<3:52:51, 8.19s/step, loss=1.357, acc=0.690, acc_len=1.64, thru=14.8, I=16.4, T=36.6, wait=5.1s, pool=57, epoch=2/2][2026-10-20 00:26:24,154] loop.py:45 INFO TRAIN_STEP step=16313 loss=1.259428 Training: 91%|█████████ | 16313/18018 [34:56:10<3:40:48, 7.77s/step, loss=1.259, acc=0.707, acc_len=1.70, thru=16.9, I=18.8, T=38.3, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:26:32,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:26:33,289] loop.py:45 INFO TRAIN_STEP step=16314 loss=1.379778 Training: 91%|█████████ | 16314/18018 [34:56:19<3:52:17, 8.18s/step, loss=1.380, acc=0.684, acc_len=1.62, thru=15.9, I=16.2, T=35.5, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:26:33,783] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:26:40,361] loop.py:380 INFO TIMING step=16315: step=3.401s data=1.446s compute=3.366s [fwd=2.398s bwd=0.953s opt=0.003s] dispatch=3.663s [2026-10-20 00:26:40,361] loop.py:45 INFO TRAIN_STEP step=16315 loss=1.295220 Training: 91%|█████████ | 16315/18018 [34:56:26<3:42:46, 7.85s/step, loss=1.295, acc=0.698, acc_len=1.68, thru=16.6, I=20.2, T=37.6, wait=3.7s, pool=53, epoch=2/2][2026-10-20 00:26:49,277] loop.py:45 INFO TRAIN_STEP step=16316 loss=1.183448 Training: 91%|█████████ | 16316/18018 [34:56:35<3:51:40, 8.17s/step, loss=1.183, acc=0.718, acc_len=1.74, thru=13.9, I=13.3, T=34.9, wait=5.2s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:26:49,273] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16315: forward=2607.6ms backward=1015.9ms [2026-10-20 00:26:58,699] loop.py:45 INFO TRAIN_STEP step=16317 loss=1.387897 Training: 91%|█████████ | 16317/18018 [34:56:44<4:02:13, 8.54s/step, loss=1.388, acc=0.686, acc_len=1.63, thru=15.8, I=15.0, T=35.6, wait=5.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:27:04,937] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:27:05,537] loop.py:45 INFO TRAIN_STEP step=16318 loss=1.206219 Training: 91%|█████████ | 16318/18018 [34:56:51<3:47:35, 8.03s/step, loss=1.206, acc=0.716, acc_len=1.74, thru=18.9, I=19.6, T=37.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:27:05,942] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:27:12,260] loop.py:45 INFO TRAIN_STEP step=16319 loss=1.199574 Training: 91%|█████████ | 16319/18018 [34:56:58<3:36:19, 7.64s/step, loss=1.200, acc=0.717, acc_len=1.74, thru=17.7, I=19.1, T=38.9, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:27:21,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:27:21,658] loop.py:380 INFO TIMING step=16320: step=3.508s data=1.394s compute=3.479s [fwd=2.504s bwd=0.960s opt=0.003s] dispatch=5.882s [2026-10-20 00:27:21,659] loop.py:45 INFO TRAIN_STEP step=16320 loss=1.181702 Training: 91%|█████████ | 16320/18018 [34:57:07<3:51:07, 8.17s/step, loss=1.182, acc=0.719, acc_len=1.75, thru=15.3, I=15.9, T=36.5, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:27:21,688] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:27:27,541] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:27:27,996] loop.py:45 INFO TRAIN_STEP step=16321 loss=1.363960 Training: 91%|█████████ | 16321/18018 [34:57:14<3:35:28, 7.62s/step, loss=1.364, acc=0.691, acc_len=1.65, thru=19.8, I=21.5, T=38.0, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:27:27,991] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16320: forward=2375.6ms backward=947.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:27:28,047] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-20 00:27:34,617] loop.py:45 INFO TRAIN_STEP step=16322 loss=1.248099 Training: 91%|█████████ | 16322/18018 [34:57:20<3:26:55, 7.32s/step, loss=1.248, acc=0.709, acc_len=1.72, thru=14.8, I=19.0, T=37.0, wait=3.2s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 00:27:34,700] eagle_store.py:517 WARNING batch_get_buffer missing keys (k522306-ae9e139e_hs, k522306-ae9e139e_ids, k522306-ae9e139e_lhs); sleeping 0.50s. [2026-10-20 00:27:44,061] loop.py:45 INFO TRAIN_STEP step=16323 loss=1.406899 Training: 91%|█████████ | 16323/18018 [34:57:30<3:44:46, 7.96s/step, loss=1.407, acc=0.681, acc_len=1.61, thru=15.6, I=14.4, T=36.9, wait=6.0s, pool=45, epoch=2/2][2026-10-20 00:27:51,872] loop.py:45 INFO TRAIN_STEP step=16324 loss=1.246949 Training: 91%|█████████ | 16324/18018 [34:57:37<3:43:24, 7.91s/step, loss=1.247, acc=0.708, acc_len=1.71, thru=16.3, I=15.8, T=37.1, wait=4.4s, pool=53, epoch=2/2][2026-10-20 00:28:00,330] loop.py:380 INFO TIMING step=16325: step=3.481s data=1.461s compute=3.449s [fwd=2.484s bwd=0.950s opt=0.003s] dispatch=4.967s [2026-10-20 00:28:00,330] loop.py:45 INFO TRAIN_STEP step=16325 loss=1.315697 Training: 91%|█████████ | 16325/18018 [34:57:46<3:47:53, 8.08s/step, loss=1.316, acc=0.695, acc_len=1.66, thru=14.2, I=16.1, T=36.8, wait=5.0s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:08,976] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-20 00:27:34,706] eagle_store.py:517 WARNING batch_get_buffer missing keys (k522305-a7b8d077_hs, k522305-a7b8d077_ids, k522305-a7b8d077_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-20 00:28:09,776] loop.py:45 INFO TRAIN_STEP step=16326 loss=1.252909 Training: 91%|█████████ | 16326/18018 [34:57:55<3:59:20, 8.49s/step, loss=1.253, acc=0.709, acc_len=1.71, thru=16.0, I=15.6, T=34.4, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:28:09,771] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16325: forward=2596.0ms backward=1071.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:09,981] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:17,079] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:28:17,900] loop.py:45 INFO TRAIN_STEP step=16327 loss=1.237147 Training: 91%|█████████ | 16327/18018 [34:58:03<3:56:08, 8.38s/step, loss=1.237, acc=0.711, acc_len=1.72, thru=15.2, I=18.4, T=34.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:18,084] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:23,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:28:23,839] loop.py:45 INFO TRAIN_STEP step=16328 loss=1.233646 Training: 91%|█████████ | 16328/18018 [34:58:09<3:35:21, 7.65s/step, loss=1.234, acc=0.712, acc_len=1.72, thru=20.4, I=22.4, T=37.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:23,897] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-20 00:28:29,455] loop.py:45 INFO TRAIN_STEP step=16329 loss=1.273116 Training: 91%|█████████ | 16329/18018 [34:58:15<3:18:08, 7.04s/step, loss=1.273, acc=0.706, acc_len=1.70, thru=21.5, I=22.7, T=38.7, wait=2.3s, pool=61, epoch=2/2][2026-10-20 00:28:37,149] loop.py:380 INFO TIMING step=16330: step=3.319s data=1.334s compute=3.279s [fwd=2.356s bwd=0.908s opt=0.003s] dispatch=4.361s [2026-10-20 00:28:37,149] loop.py:45 INFO TRAIN_STEP step=16330 loss=1.188334 Training: 91%|█████████ | 16330/18018 [34:58:23<3:23:30, 7.23s/step, loss=1.188, acc=0.723, acc_len=1.76, thru=17.7, I=17.0, T=38.6, wait=4.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:42,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:28:43,101] loop.py:45 INFO TRAIN_STEP step=16331 loss=1.115626 Training: 91%|█████████ | 16331/18018 [34:58:29<3:12:34, 6.85s/step, loss=1.116, acc=0.731, acc_len=1.79, thru=21.5, I=22.2, T=38.2, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:28:43,097] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16330: forward=2349.5ms backward=944.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:43,576] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:28:50,941] loop.py:45 INFO TRAIN_STEP step=16332 loss=1.212032 Training: 91%|█████████ | 16332/18018 [34:58:37<3:20:48, 7.15s/step, loss=1.212, acc=0.715, acc_len=1.74, thru=15.0, I=15.9, T=36.4, wait=4.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:57,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:28:58,634] loop.py:45 INFO TRAIN_STEP step=16333 loss=1.222527 Training: 91%|█████████ | 16333/18018 [34:58:44<3:25:19, 7.31s/step, loss=1.223, acc=0.716, acc_len=1.74, thru=18.2, I=19.0, T=36.6, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:28:58,769] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-20 00:29:05,188] loop.py:45 INFO TRAIN_STEP step=16334 loss=1.178102 Training: 91%|█████████ | 16334/18018 [34:58:51<3:18:47, 7.08s/step, loss=1.178, acc=0.721, acc_len=1.75, thru=19.6, I=20.3, T=38.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:29:05,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:29:05,657] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:29:12,303] loop.py:380 INFO TIMING step=16335: step=3.509s data=1.459s compute=3.479s [fwd=2.506s bwd=0.959s opt=0.003s] dispatch=3.598s [2026-10-20 00:29:12,303] loop.py:45 INFO TRAIN_STEP step=16335 loss=1.349219 Training: 91%|█████████ | 16335/18018 [34:58:58<3:18:57, 7.09s/step, loss=1.349, acc=0.694, acc_len=1.65, thru=18.9, I=18.9, T=36.5, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:29:19,805] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:29:20,171] loop.py:45 INFO TRAIN_STEP step=16336 loss=1.381246 Training: 91%|█████████ | 16336/18018 [34:59:06<3:25:22, 7.33s/step, loss=1.381, acc=0.679, acc_len=1.61, thru=17.6, I=17.5, T=37.4, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:29:20,166] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16335: forward=2449.5ms backward=927.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:29:20,319] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:29:28,567] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:29:29,552] loop.py:45 INFO TRAIN_STEP step=16337 loss=1.301891 Training: 91%|█████████ | 16337/18018 [34:59:15<3:42:31, 7.94s/step, loss=1.302, acc=0.701, acc_len=1.68, thru=13.9, I=15.0, T=36.5, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:29:29,583] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-20 00:29:36,456] loop.py:45 INFO TRAIN_STEP step=16338 loss=1.411816 Training: 91%|█████████ | 16338/18018 [34:59:22<3:33:38, 7.63s/step, loss=1.412, acc=0.678, acc_len=1.60, thru=17.4, I=20.4, T=38.3, wait=3.6s, pool=49, epoch=2/2][2026-10-20 00:29:43,385] loop.py:45 INFO TRAIN_STEP step=16339 loss=1.094451 Training: 91%|█████████ | 16339/18018 [34:59:29<3:27:39, 7.42s/step, loss=1.094, acc=0.736, acc_len=1.80, thru=15.7, I=20.0, T=36.2, wait=3.4s, pool=37, epoch=2/2][2026-10-20 00:29:53,552] loop.py:380 INFO TIMING step=16340: step=3.698s data=1.475s compute=3.667s [fwd=2.611s bwd=1.041s opt=0.003s] dispatch=6.458s [2026-10-20 00:29:53,553] loop.py:45 INFO TRAIN_STEP step=16340 loss=1.312434 Training: 91%|█████████ | 16340/18018 [34:59:39<3:50:33, 8.24s/step, loss=1.312, acc=0.695, acc_len=1.66, thru=12.9, I=11.7, T=34.6, wait=6.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:29:59,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-20 00:30:00,959] loop.py:45 INFO TRAIN_STEP step=16341 loss=1.151986 Training: 91%|█████████ | 16341/18018 [34:59:47<3:43:24, 7.99s/step, loss=1.152, acc=0.730, acc_len=1.78, thru=18.0, I=20.3, T=34.8, wait=3.7s, pool=69, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:30:00,955] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16340: forward=2642.2ms backward=994.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:01,155] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:06,658] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:30:07,151] loop.py:45 INFO TRAIN_STEP step=16342 loss=1.262319 Training: 91%|█████████ | 16342/18018 [34:59:53<3:28:10, 7.45s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=20.1, I=24.8, T=38.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:07,181] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:12,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:30:13,631] loop.py:45 INFO TRAIN_STEP step=16343 loss=1.365577 Training: 91%|█████████ | 16343/18018 [34:59:59<3:19:53, 7.16s/step, loss=1.366, acc=0.688, acc_len=1.64, thru=20.3, I=22.7, T=38.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:14,135] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:19,158] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:30:20,014] loop.py:45 INFO TRAIN_STEP step=16344 loss=1.228786 Training: 91%|█████████ | 16344/18018 [35:00:06<3:13:17, 6.93s/step, loss=1.229, acc=0.715, acc_len=1.73, thru=20.0, I=25.4, T=39.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:20,165] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:30:27,329] loop.py:380 INFO TIMING step=16345: step=3.398s data=1.403s compute=3.372s [fwd=2.412s bwd=0.946s opt=0.003s] dispatch=3.907s [2026-10-20 00:30:27,329] loop.py:45 INFO TRAIN_STEP step=16345 loss=1.220415 Training: 91%|█████████ | 16345/18018 [35:00:13<3:16:23, 7.04s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=17.6, I=19.2, T=37.7, wait=3.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:33,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:30:34,325] loop.py:45 INFO TRAIN_STEP step=16346 loss=1.290742 Training: 91%|█████████ | 16346/18018 [35:00:20<3:15:53, 7.03s/step, loss=1.291, acc=0.701, acc_len=1.69, thru=19.3, I=17.8, T=37.3, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:30:34,321] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16345: forward=2443.1ms backward=943.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:34,355] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:39,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:30:40,235] loop.py:45 INFO TRAIN_STEP step=16347 loss=1.381905 Training: 91%|█████████ | 16347/18018 [35:00:26<3:06:24, 6.69s/step, loss=1.382, acc=0.684, acc_len=1.63, thru=20.8, I=22.7, T=38.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:30:40,719] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:30:46,361] loop.py:45 INFO TRAIN_STEP step=16348 loss=1.307076 Training: 91%|█████████ | 16348/18018 [35:00:32<3:01:33, 6.52s/step, loss=1.307, acc=0.699, acc_len=1.67, thru=20.4, I=22.4, T=39.0, wait=2.8s, pool=61, epoch=2/2][2026-10-20 00:30:52,721] loop.py:45 INFO TRAIN_STEP step=16349 loss=1.139689 Training: 91%|█████████ | 16349/18018 [35:00:38<3:00:05, 6.47s/step, loss=1.140, acc=0.730, acc_len=1.78, thru=18.9, I=21.3, T=38.6, wait=3.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:01,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:31:02,066] loop.py:380 INFO TIMING step=16350: step=3.596s data=1.489s compute=3.564s [fwd=2.553s bwd=0.997s opt=0.003s] dispatch=5.740s [2026-10-20 00:31:02,066] loop.py:45 INFO TRAIN_STEP step=16350 loss=1.310310 Training: 91%|█████████ | 16350/18018 [35:00:48<3:23:56, 7.34s/step, loss=1.310, acc=0.697, acc_len=1.67, thru=15.3, I=15.3, T=35.6, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:02,226] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:07,910] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:31:08,303] loop.py:45 INFO TRAIN_STEP step=16351 loss=1.212927 Training: 91%|█████████ | 16351/18018 [35:00:54<3:14:39, 7.01s/step, loss=1.213, acc=0.717, acc_len=1.74, thru=19.9, I=21.6, T=37.6, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:31:08,298] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16350: forward=2378.6ms backward=972.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:08,415] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:13,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:31:14,429] loop.py:45 INFO TRAIN_STEP step=16352 loss=1.242118 Training: 91%|█████████ | 16352/18018 [35:01:00<3:07:12, 6.74s/step, loss=1.242, acc=0.707, acc_len=1.71, thru=20.9, I=22.3, T=37.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:14,460] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:22,291] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:31:23,191] loop.py:45 INFO TRAIN_STEP step=16353 loss=1.271623 Training: 91%|█████████ | 16353/18018 [35:01:09<3:23:54, 7.35s/step, loss=1.272, acc=0.705, acc_len=1.70, thru=14.5, I=16.6, T=36.3, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:23,298] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:29,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:31:30,571] loop.py:45 INFO TRAIN_STEP step=16354 loss=1.250877 Training: 91%|█████████ | 16354/18018 [35:01:16<3:24:02, 7.36s/step, loss=1.251, acc=0.707, acc_len=1.70, thru=16.9, I=19.5, T=36.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:30,948] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:36,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:31:36,851] loop.py:380 INFO TIMING step=16355: step=3.308s data=1.375s compute=3.274s [fwd=2.329s bwd=0.931s opt=0.003s] dispatch=2.963s [2026-10-20 00:31:36,851] loop.py:45 INFO TRAIN_STEP step=16355 loss=1.245551 Training: 91%|█████████ | 16355/18018 [35:01:22<3:14:58, 7.03s/step, loss=1.246, acc=0.704, acc_len=1.69, thru=20.0, I=22.5, T=38.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:37,174] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:31:43,437] loop.py:45 INFO TRAIN_STEP step=16356 loss=1.315028 Training: 91%|█████████ | 16356/18018 [35:01:29<3:11:07, 6.90s/step, loss=1.315, acc=0.698, acc_len=1.67, thru=18.9, I=22.0, T=38.5, wait=3.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:31:43,433] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16355: forward=2346.2ms backward=925.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:50,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:31:50,502] loop.py:45 INFO TRAIN_STEP step=16357 loss=1.251015 Training: 91%|█████████ | 16357/18018 [35:01:36<3:12:22, 6.95s/step, loss=1.251, acc=0.711, acc_len=1.72, thru=18.6, I=19.3, T=36.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:31:50,820] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:31:56,331] loop.py:45 INFO TRAIN_STEP step=16358 loss=1.114743 Training: 91%|█████████ | 16358/18018 [35:01:42<3:02:57, 6.61s/step, loss=1.115, acc=0.731, acc_len=1.79, thru=20.3, I=20.3, T=39.0, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:02,987] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:32:03,404] loop.py:45 INFO TRAIN_STEP step=16359 loss=1.266061 Training: 91%|█████████ | 16359/18018 [35:01:49<3:06:41, 6.75s/step, loss=1.266, acc=0.709, acc_len=1.71, thru=18.8, I=19.5, T=37.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:03,497] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:09,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:32:10,312] loop.py:380 INFO TIMING step=16360: step=3.366s data=1.361s compute=3.337s [fwd=2.397s bwd=0.926s opt=0.003s] dispatch=3.531s [2026-10-20 00:32:10,312] loop.py:45 INFO TRAIN_STEP step=16360 loss=1.359056 Training: 91%|█████████ | 16360/18018 [35:01:56<3:07:50, 6.80s/step, loss=1.359, acc=0.686, acc_len=1.63, thru=18.1, I=21.9, T=38.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:10,702] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:18,871] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:32:19,271] loop.py:45 INFO TRAIN_STEP step=16361 loss=1.197232 Training: 91%|█████████ | 16361/18018 [35:02:05<3:25:39, 7.45s/step, loss=1.197, acc=0.723, acc_len=1.76, thru=14.9, I=18.1, T=34.2, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:32:19,267] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16360: forward=2689.1ms backward=995.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:19,381] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-20 00:32:27,591] loop.py:45 INFO TRAIN_STEP step=16362 loss=1.259934 Training: 91%|█████████ | 16362/18018 [35:02:13<3:32:45, 7.71s/step, loss=1.260, acc=0.703, acc_len=1.69, thru=13.7, I=14.8, T=36.3, wait=4.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:34,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:32:35,200] loop.py:45 INFO TRAIN_STEP step=16363 loss=1.270133 Training: 91%|█████████ | 16363/18018 [35:02:21<3:31:47, 7.68s/step, loss=1.270, acc=0.707, acc_len=1.70, thru=16.5, I=17.7, T=37.6, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:35,687] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:32:40,866] loop.py:45 INFO TRAIN_STEP step=16364 loss=1.215550 Training: 91%|█████████ | 16364/18018 [35:02:26<3:15:05, 7.08s/step, loss=1.216, acc=0.711, acc_len=1.71, thru=21.3, I=22.4, T=39.5, wait=2.4s, pool=61, epoch=2/2][2026-10-20 00:32:48,304] loop.py:380 INFO TIMING step=16365: step=3.432s data=1.387s compute=3.398s [fwd=2.454s bwd=0.930s opt=0.003s] dispatch=3.991s [2026-10-20 00:32:48,305] loop.py:45 INFO TRAIN_STEP step=16365 loss=1.204995 Training: 91%|█████████ | 16365/18018 [35:02:34<3:17:53, 7.18s/step, loss=1.205, acc=0.716, acc_len=1.73, thru=18.7, I=19.0, T=37.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:48,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:48,754] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:54,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:32:54,317] loop.py:45 INFO TRAIN_STEP step=16366 loss=1.222057 Training: 91%|█████████ | 16366/18018 [35:02:40<3:08:06, 6.83s/step, loss=1.222, acc=0.712, acc_len=1.72, thru=20.7, I=20.9, T=39.6, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:32:54,312] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16365: forward=2310.3ms backward=878.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:32:54,571] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:33:00,450] loop.py:45 INFO TRAIN_STEP step=16367 loss=1.187130 Training: 91%|█████████ | 16367/18018 [35:02:46<3:02:13, 6.62s/step, loss=1.187, acc=0.718, acc_len=1.75, thru=20.6, I=20.7, T=38.4, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:33:07,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:33:07,628] loop.py:45 INFO TRAIN_STEP step=16368 loss=1.328205 Training: 91%|█████████ | 16368/18018 [35:02:53<3:06:42, 6.79s/step, loss=1.328, acc=0.697, acc_len=1.67, thru=18.7, I=18.5, T=37.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:33:07,990] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:33:13,516] loop.py:45 INFO TRAIN_STEP step=16369 loss=1.240141 Training: 91%|█████████ | 16369/18018 [35:02:59<2:59:09, 6.52s/step, loss=1.240, acc=0.709, acc_len=1.71, thru=18.1, I=19.6, T=38.9, wait=2.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:33:21,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:33:22,147] loop.py:380 INFO TIMING step=16370: step=3.667s data=1.501s compute=3.627s [fwd=2.556s bwd=1.057s opt=0.003s] dispatch=4.955s [2026-10-20 00:33:22,147] loop.py:45 INFO TRAIN_STEP step=16370 loss=1.288795 Training: 91%|█████████ | 16370/18018 [35:03:08<3:16:26, 7.15s/step, loss=1.289, acc=0.701, acc_len=1.68, thru=17.4, I=18.0, T=34.9, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:33:22,598] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:33:29,064] loop.py:45 INFO TRAIN_STEP step=16371 loss=1.338627 Training: 91%|█████████ | 16371/18018 [35:03:15<3:14:24, 7.08s/step, loss=1.339, acc=0.695, acc_len=1.66, thru=14.2, I=17.7, T=36.6, wait=3.4s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:33:29,060] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16370: forward=2446.6ms backward=985.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:33:37,957] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:33:38,003] loop.py:45 INFO TRAIN_STEP step=16372 loss=1.281911 Training: 91%|█████████ | 16372/18018 [35:03:24<3:29:33, 7.64s/step, loss=1.282, acc=0.699, acc_len=1.68, thru=17.8, I=15.7, T=38.8, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:33:38,460] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:33:44,955] loop.py:45 INFO TRAIN_STEP step=16373 loss=1.247114 Training: 91%|█████████ | 16373/18018 [35:03:31<3:23:48, 7.43s/step, loss=1.247, acc=0.708, acc_len=1.71, thru=17.7, I=19.2, T=38.6, wait=3.6s, pool=49, epoch=2/2][2026-10-20 00:33:55,264] loop.py:45 INFO TRAIN_STEP step=16374 loss=1.179929 Training: 91%|█████████ | 16374/18018 [35:03:41<3:47:18, 8.30s/step, loss=1.180, acc=0.721, acc_len=1.75, thru=12.0, I=11.7, T=36.0, wait=6.7s, pool=45, epoch=2/2][2026-10-20 00:34:01,544] loop.py:380 INFO TIMING step=16375: step=3.298s data=1.255s compute=3.272s [fwd=2.374s bwd=0.883s opt=0.003s] dispatch=2.973s [2026-10-20 00:34:01,545] loop.py:45 INFO TRAIN_STEP step=16375 loss=1.253332 Training: 91%|█████████ | 16375/18018 [35:03:47<3:30:36, 7.69s/step, loss=1.253, acc=0.705, acc_len=1.69, thru=16.6, I=17.1, T=38.8, wait=3.0s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:13,049] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:34:13,640] loop.py:45 INFO TRAIN_STEP step=16376 loss=1.306018 Training: 91%|█████████ | 16376/18018 [35:03:59<4:06:39, 9.01s/step, loss=1.306, acc=0.697, acc_len=1.67, thru=13.2, I=13.3, T=32.4, wait=8.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:34:13,635] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16375: forward=2724.7ms backward=1161.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:14,054] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:34:22,793] loop.py:45 INFO TRAIN_STEP step=16377 loss=1.319692 Training: 91%|█████████ | 16377/18018 [35:04:08<4:07:39, 9.06s/step, loss=1.320, acc=0.695, acc_len=1.67, thru=13.0, I=15.7, T=35.7, wait=5.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:30,728] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:34:30,984] loop.py:45 INFO TRAIN_STEP step=16378 loss=1.242414 Training: 91%|█████████ | 16378/18018 [35:04:17<4:00:23, 8.79s/step, loss=1.242, acc=0.713, acc_len=1.73, thru=16.3, I=16.2, T=35.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:31,231] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:36,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:34:36,797] loop.py:45 INFO TRAIN_STEP step=16379 loss=1.205243 Training: 91%|█████████ | 16379/18018 [35:04:22<3:35:49, 7.90s/step, loss=1.205, acc=0.716, acc_len=1.73, thru=20.4, I=23.1, T=39.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:37,029] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:42,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:34:43,305] loop.py:380 INFO TIMING step=16380: step=3.381s data=1.383s compute=3.342s [fwd=2.366s bwd=0.962s opt=0.003s] dispatch=3.119s [2026-10-20 00:34:43,306] loop.py:45 INFO TRAIN_STEP step=16380 loss=1.396514 Training: 91%|█████████ | 16380/18018 [35:04:29<3:24:17, 7.48s/step, loss=1.397, acc=0.685, acc_len=1.63, thru=20.0, I=24.4, T=37.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:43,800] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:34:50,027] loop.py:45 INFO TRAIN_STEP step=16381 loss=1.148438 Training: 91%|█████████ | 16381/18018 [35:04:36<3:17:55, 7.25s/step, loss=1.148, acc=0.725, acc_len=1.77, thru=18.9, I=21.2, T=39.2, wait=3.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:34:50,022] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16380: forward=2313.0ms backward=907.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:57,880] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:34:58,002] loop.py:45 INFO TRAIN_STEP step=16382 loss=1.214330 Training: 91%|█████████ | 16382/18018 [35:04:44<3:23:42, 7.47s/step, loss=1.214, acc=0.709, acc_len=1.71, thru=18.2, I=17.6, T=36.5, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:34:58,382] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:04,108] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:35:05,376] loop.py:45 INFO TRAIN_STEP step=16383 loss=1.226031 Training: 91%|█████████ | 16383/18018 [35:04:51<3:22:47, 7.44s/step, loss=1.226, acc=0.713, acc_len=1.72, thru=17.1, I=19.9, T=35.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:05,617] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:12,606] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:35:13,233] loop.py:45 INFO TRAIN_STEP step=16384 loss=1.256665 Training: 91%|█████████ | 16384/18018 [35:04:59<3:26:02, 7.57s/step, loss=1.257, acc=0.711, acc_len=1.71, thru=15.2, I=19.6, T=38.3, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:13,611] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:20,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:35:20,110] loop.py:380 INFO TIMING step=16385: step=3.400s data=1.389s compute=3.359s [fwd=2.418s bwd=0.926s opt=0.003s] dispatch=3.468s [2026-10-20 00:35:20,110] loop.py:45 INFO TRAIN_STEP step=16385 loss=1.158624 Training: 91%|█████████ | 16385/18018 [35:05:06<3:20:17, 7.36s/step, loss=1.159, acc=0.721, acc_len=1.75, thru=19.4, I=19.6, T=37.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:20,553] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:26,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:35:26,556] loop.py:45 INFO TRAIN_STEP step=16386 loss=1.289180 Training: 91%|█████████ | 16386/18018 [35:05:12<3:12:43, 7.09s/step, loss=1.289, acc=0.699, acc_len=1.67, thru=19.8, I=21.8, T=37.3, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:35:26,551] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16385: forward=2435.5ms backward=940.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:26,659] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:31,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:35:32,619] loop.py:45 INFO TRAIN_STEP step=16387 loss=1.115214 Training: 91%|█████████ | 16387/18018 [35:05:18<3:04:16, 6.78s/step, loss=1.115, acc=0.733, acc_len=1.79, thru=20.6, I=22.9, T=36.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:32,751] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:38,439] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:35:39,496] loop.py:45 INFO TRAIN_STEP step=16388 loss=1.190600 Training: 91%|█████████ | 16388/18018 [35:05:25<3:04:57, 6.81s/step, loss=1.191, acc=0.721, acc_len=1.75, thru=17.7, I=22.5, T=36.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:39,949] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:44,527] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:35:45,489] loop.py:45 INFO TRAIN_STEP step=16389 loss=1.171741 Training: 91%|█████████ | 16389/18018 [35:05:31<2:58:11, 6.56s/step, loss=1.172, acc=0.721, acc_len=1.75, thru=20.9, I=26.7, T=39.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:45,540] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:52,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:35:52,686] loop.py:380 INFO TIMING step=16390: step=3.866s data=1.270s compute=3.834s [fwd=2.956s bwd=0.866s opt=0.003s] dispatch=3.323s [2026-10-20 00:35:52,686] loop.py:45 INFO TRAIN_STEP step=16390 loss=1.250775 Training: 91%|█████████ | 16390/18018 [35:05:38<3:03:14, 6.75s/step, loss=1.251, acc=0.709, acc_len=1.71, thru=17.4, I=20.4, T=33.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:35:52,951] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:36:01,308] loop.py:45 INFO TRAIN_STEP step=16391 loss=1.281809 Training: 91%|█████████ | 16391/18018 [35:05:47<3:18:20, 7.31s/step, loss=1.282, acc=0.702, acc_len=1.69, thru=13.1, I=17.5, T=30.1, wait=4.4s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:36:01,303] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16390: forward=3187.7ms backward=1022.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:36:08,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:36:08,818] loop.py:45 INFO TRAIN_STEP step=16392 loss=1.169117 Training: 91%|█████████ | 16392/18018 [35:05:54<3:19:48, 7.37s/step, loss=1.169, acc=0.720, acc_len=1.75, thru=17.6, I=15.3, T=36.6, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:36:09,211] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:36:14,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:36:15,307] loop.py:45 INFO TRAIN_STEP step=16393 loss=1.289839 Training: 91%|█████████ | 16393/18018 [35:06:01<3:12:30, 7.11s/step, loss=1.290, acc=0.701, acc_len=1.69, thru=20.0, I=20.0, T=38.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:36:15,779] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:36:21,756] loop.py:45 INFO TRAIN_STEP step=16394 loss=1.265603 Training: 91%|█████████ | 16394/18018 [35:06:07<3:07:02, 6.91s/step, loss=1.266, acc=0.709, acc_len=1.71, thru=18.8, I=22.4, T=38.4, wait=3.1s, pool=53, epoch=2/2][2026-10-20 00:36:28,402] loop.py:380 INFO TIMING step=16395: step=3.274s data=1.305s compute=3.249s [fwd=2.339s bwd=0.896s opt=0.003s] dispatch=3.362s [2026-10-20 00:36:28,402] loop.py:45 INFO TRAIN_STEP step=16395 loss=1.181609 Training: 91%|█████████ | 16395/18018 [35:06:14<3:04:46, 6.83s/step, loss=1.182, acc=0.719, acc_len=1.75, thru=18.3, I=21.1, T=39.1, wait=3.4s, pool=57, epoch=2/2][2026-10-20 00:36:37,115] loop.py:45 INFO TRAIN_STEP step=16396 loss=1.279906 Training: 91%|█████████ | 16396/18018 [35:06:23<3:19:58, 7.40s/step, loss=1.280, acc=0.708, acc_len=1.71, thru=13.9, I=14.7, T=36.9, wait=5.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:36:37,110] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16395: forward=2466.6ms backward=954.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:36:49,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:36:49,951] loop.py:45 INFO TRAIN_STEP step=16397 loss=1.289975 Training: 91%|█████████ | 16397/18018 [35:06:36<4:03:52, 9.03s/step, loss=1.290, acc=0.701, acc_len=1.68, thru=11.0, I=10.5, T=29.6, wait=8.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:36:50,169] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:36:55,830] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:36:56,815] loop.py:45 INFO TRAIN_STEP step=16398 loss=1.335489 Training: 91%|█████████ | 16398/18018 [35:06:42<3:46:12, 8.38s/step, loss=1.335, acc=0.694, acc_len=1.66, thru=18.1, I=19.7, T=36.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:36:56,843] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:37:01,303] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:37:02,180] loop.py:45 INFO TRAIN_STEP step=16399 loss=1.158008 Training: 91%|█████████ | 16399/18018 [35:06:48<3:21:40, 7.47s/step, loss=1.158, acc=0.726, acc_len=1.78, thru=22.4, I=27.8, T=39.9, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:37:02,310] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-20 00:37:07,862] loop.py:380 INFO TIMING step=16400: step=3.236s data=1.325s compute=3.193s [fwd=2.285s bwd=0.895s opt=0.003s] dispatch=2.437s [2026-10-20 00:37:07,863] loop.py:45 INFO TRAIN_STEP step=16400 loss=1.187990 Training: 91%|█████████ | 16400/18018 [35:06:53<3:07:04, 6.94s/step, loss=1.188, acc=0.723, acc_len=1.76, thru=21.3, I=25.2, T=39.6, wait=2.4s, pool=53, epoch=2/2][2026-10-20 00:37:16,942] loop.py:45 INFO TRAIN_STEP step=16401 loss=1.242396 Training: 91%|█████████ | 16401/18018 [35:07:03<3:24:16, 7.58s/step, loss=1.242, acc=0.710, acc_len=1.72, thru=13.3, I=12.9, T=32.2, wait=5.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:37:16,938] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16400: forward=2934.3ms backward=982.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:37:24,082] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:37:24,758] loop.py:45 INFO TRAIN_STEP step=16402 loss=1.210436 Training: 91%|█████████ | 16402/18018 [35:07:10<3:26:03, 7.65s/step, loss=1.210, acc=0.716, acc_len=1.74, thru=17.7, I=17.4, T=36.3, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:37:25,088] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:37:31,111] loop.py:45 INFO TRAIN_STEP step=16403 loss=1.294651 Training: 91%|█████████ | 16403/18018 [35:07:17<3:15:26, 7.26s/step, loss=1.295, acc=0.702, acc_len=1.69, thru=16.9, I=21.3, T=38.2, wait=3.0s, pool=33, epoch=2/2](TrainerActor pid=9370) [2026-10-20 00:37:31,253] eagle_store.py:517 WARNING batch_get_buffer missing keys (k524894-8b68c7c4_hs, k524894-8b68c7c4_ids, k524894-8b68c7c4_lhs); sleeping 0.50s. [2026-10-20 00:37:40,321] loop.py:45 INFO TRAIN_STEP step=16404 loss=1.225748 Training: 91%|█████████ | 16404/18018 [35:07:26<3:31:02, 7.85s/step, loss=1.226, acc=0.715, acc_len=1.74, thru=16.5, I=14.7, T=35.4, wait=5.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:37:46,367] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 00:37:31,226] eagle_store.py:517 WARNING batch_get_buffer missing keys (k524895-ad3c88be_hs, k524895-ad3c88be_ids, k524895-ad3c88be_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-20 00:37:47,377] loop.py:380 INFO TIMING step=16405: step=3.501s data=1.503s compute=3.445s [fwd=2.392s bwd=1.039s opt=0.003s] dispatch=3.547s [2026-10-20 00:37:47,377] loop.py:45 INFO TRAIN_STEP step=16405 loss=1.251619 Training: 91%|█████████ | 16405/18018 [35:07:33<3:24:33, 7.61s/step, loss=1.252, acc=0.710, acc_len=1.71, thru=18.9, I=19.4, T=36.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:37:47,878] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:37:54,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:37:54,479] loop.py:45 INFO TRAIN_STEP step=16406 loss=1.313202 (TrainerActor pid=9310) [2026-10-20 00:37:54,475] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16405: forward=2475.6ms backward=944.1ms Training: 91%|█████████ | 16406/18018 [35:07:40<3:20:20, 7.46s/step, loss=1.313, acc=0.697, acc_len=1.67, thru=17.7, I=21.1, T=36.9, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:37:54,852] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:38:00,514] loop.py:45 INFO TRAIN_STEP step=16407 loss=1.281363 Training: 91%|█████████ | 16407/18018 [35:07:46<3:08:46, 7.03s/step, loss=1.281, acc=0.704, acc_len=1.69, thru=18.9, I=20.2, T=38.8, wait=2.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:38:06,422] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:38:07,264] loop.py:45 INFO TRAIN_STEP step=16408 loss=1.284142 Training: 91%|█████████ | 16408/18018 [35:07:53<3:06:23, 6.95s/step, loss=1.284, acc=0.704, acc_len=1.70, thru=19.7, I=21.3, T=34.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:38:07,427] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:38:13,241] loop.py:45 INFO TRAIN_STEP step=16409 loss=1.294117 Training: 91%|█████████ | 16409/18018 [35:07:59<2:58:27, 6.66s/step, loss=1.294, acc=0.700, acc_len=1.68, thru=19.9, I=20.5, T=38.9, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:38:18,302] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:38:19,111] loop.py:380 INFO TIMING step=16410: step=3.459s data=1.466s compute=3.412s [fwd=2.423s bwd=0.974s opt=0.003s] dispatch=2.403s [2026-10-20 00:38:19,111] loop.py:45 INFO TRAIN_STEP step=16410 loss=1.180934 Training: 91%|█████████ | 16410/18018 [35:08:05<2:52:02, 6.42s/step, loss=1.181, acc=0.724, acc_len=1.75, thru=21.5, I=22.6, T=37.0, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:38:19,306] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:38:26,511] loop.py:45 INFO TRAIN_STEP step=16411 loss=1.403750 Training: 91%|█████████ | 16411/18018 [35:08:12<2:59:50, 6.71s/step, loss=1.404, acc=0.681, acc_len=1.62, thru=14.6, I=18.9, T=37.3, wait=4.0s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:38:26,506] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16410: forward=2442.8ms backward=941.0ms [2026-10-20 00:38:36,487] loop.py:45 INFO TRAIN_STEP step=16412 loss=1.193275 Training: 91%|█████████ | 16412/18018 [35:08:22<3:25:53, 7.69s/step, loss=1.193, acc=0.720, acc_len=1.75, thru=14.8, I=13.1, T=35.8, wait=6.4s, pool=53, epoch=2/2][2026-10-20 00:38:44,830] loop.py:45 INFO TRAIN_STEP step=16413 loss=1.323516 Training: 91%|█████████ | 16413/18018 [35:08:30<3:30:59, 7.89s/step, loss=1.324, acc=0.692, acc_len=1.65, thru=16.4, I=16.1, T=36.7, wait=4.8s, pool=49, epoch=2/2][2026-10-20 00:38:55,079] loop.py:45 INFO TRAIN_STEP step=16414 loss=1.208106 Training: 91%|█████████ | 16414/18018 [35:08:41<3:49:48, 8.60s/step, loss=1.208, acc=0.717, acc_len=1.73, thru=14.8, I=12.2, T=32.7, wait=6.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:01,551] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:39:02,670] loop.py:380 INFO TIMING step=16415: step=3.584s data=1.421s compute=3.561s [fwd=2.579s bwd=0.967s opt=0.003s] dispatch=3.997s [2026-10-20 00:39:02,670] loop.py:45 INFO TRAIN_STEP step=16415 loss=1.251349 Training: 91%|█████████ | 16415/18018 [35:08:48<3:41:36, 8.29s/step, loss=1.251, acc=0.706, acc_len=1.70, thru=17.4, I=17.2, T=35.7, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:03,060] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:39:10,273] loop.py:45 INFO TRAIN_STEP step=16416 loss=1.182667 Training: 91%|█████████ | 16416/18018 [35:08:56<3:35:55, 8.09s/step, loss=1.183, acc=0.722, acc_len=1.75, thru=14.6, I=19.0, T=37.6, wait=4.2s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:39:10,269] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16415: forward=2392.8ms backward=967.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:18,649] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-20 00:39:18,848] loop.py:45 INFO TRAIN_STEP step=16417 loss=1.260508 Training: 91%|█████████ | 16417/18018 [35:09:04<3:39:41, 8.23s/step, loss=1.261, acc=0.703, acc_len=1.69, thru=17.4, I=16.6, T=30.5, wait=4.4s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:19,151] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:24,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:39:24,541] loop.py:45 INFO TRAIN_STEP step=16418 loss=1.258860 Training: 91%|█████████ | 16418/18018 [35:09:10<3:19:13, 7.47s/step, loss=1.259, acc=0.701, acc_len=1.68, thru=19.7, I=22.0, T=39.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:24,698] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:29,047] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:39:29,793] loop.py:45 INFO TRAIN_STEP step=16419 loss=1.183084 Training: 91%|█████████ | 16419/18018 [35:09:15<3:01:22, 6.81s/step, loss=1.183, acc=0.719, acc_len=1.74, thru=23.2, I=25.5, T=39.9, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:30,051] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:35,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:39:35,415] loop.py:380 INFO TIMING step=16420: step=3.379s data=1.402s compute=3.344s [fwd=2.381s bwd=0.949s opt=0.003s] dispatch=2.234s [2026-10-20 00:39:35,415] loop.py:45 INFO TRAIN_STEP step=16420 loss=1.185929 Training: 91%|█████████ | 16420/18018 [35:09:21<2:51:48, 6.45s/step, loss=1.186, acc=0.718, acc_len=1.74, thru=23.3, I=26.4, T=37.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:35,714] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:39:44,197] loop.py:45 INFO TRAIN_STEP step=16421 loss=1.303929 Training: 91%|█████████ | 16421/18018 [35:09:30<3:10:17, 7.15s/step, loss=1.304, acc=0.698, acc_len=1.67, thru=14.2, I=13.9, T=35.3, wait=5.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:39:44,193] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16420: forward=2550.1ms backward=1025.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:51,771] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:39:51,918] loop.py:45 INFO TRAIN_STEP step=16422 loss=1.236400 Training: 91%|█████████ | 16422/18018 [35:09:38<3:14:44, 7.32s/step, loss=1.236, acc=0.708, acc_len=1.71, thru=17.1, I=16.6, T=36.4, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:39:52,272] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:39:57,994] loop.py:45 INFO TRAIN_STEP step=16423 loss=1.302125 Training: 91%|█████████ | 16423/18018 [35:09:44<3:04:40, 6.95s/step, loss=1.302, acc=0.701, acc_len=1.68, thru=17.3, I=17.2, T=39.9, wait=2.9s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:40:05,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:40:05,678] loop.py:45 INFO TRAIN_STEP step=16424 loss=1.237726 Training: 91%|█████████ | 16424/18018 [35:09:51<3:10:26, 7.17s/step, loss=1.238, acc=0.714, acc_len=1.72, thru=20.9, I=18.3, T=35.3, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:40:05,972] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:40:12,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:40:13,890] loop.py:380 INFO TIMING step=16425: step=3.745s data=1.590s compute=3.711s [fwd=2.663s bwd=1.034s opt=0.003s] dispatch=4.461s [2026-10-20 00:40:13,890] loop.py:45 INFO TRAIN_STEP step=16425 loss=1.296734 Training: 91%|█████████ | 16425/18018 [35:09:59<3:18:39, 7.48s/step, loss=1.297, acc=0.703, acc_len=1.69, thru=15.1, I=17.2, T=34.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:40:14,361] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:40:19,880] loop.py:45 INFO TRAIN_STEP step=16426 loss=1.263413 Training: 91%|█████████ | 16426/18018 [35:10:05<3:06:38, 7.03s/step, loss=1.263, acc=0.705, acc_len=1.69, thru=18.8, I=22.8, T=38.8, wait=2.7s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:40:19,876] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16425: forward=2339.9ms backward=896.4ms [2026-10-20 00:40:29,428] loop.py:45 INFO TRAIN_STEP step=16427 loss=1.308513 Training: 91%|█████████ | 16427/18018 [35:10:15<3:26:31, 7.79s/step, loss=1.309, acc=0.698, acc_len=1.68, thru=13.9, I=13.7, T=35.0, wait=5.9s, pool=57, epoch=2/2][2026-10-20 00:40:35,897] loop.py:45 INFO TRAIN_STEP step=16428 loss=1.301102 Training: 91%|█████████ | 16428/18018 [35:10:21<3:15:53, 7.39s/step, loss=1.301, acc=0.698, acc_len=1.67, thru=17.8, I=18.0, T=38.8, wait=3.2s, pool=53, epoch=2/2][2026-10-20 00:40:46,570] loop.py:45 INFO TRAIN_STEP step=16429 loss=1.233961 Training: 91%|█████████ | 16429/18018 [35:10:32<3:41:51, 8.38s/step, loss=1.234, acc=0.709, acc_len=1.71, thru=12.5, I=12.2, T=35.7, wait=7.1s, pool=61, epoch=2/2][2026-10-20 00:40:55,885] loop.py:380 INFO TIMING step=16430: step=3.576s data=1.465s compute=3.518s [fwd=2.433s bwd=1.070s opt=0.003s] dispatch=5.728s [2026-10-20 00:40:55,885] loop.py:45 INFO TRAIN_STEP step=16430 loss=1.318860 Training: 91%|█████████ | 16430/18018 [35:10:41<3:49:10, 8.66s/step, loss=1.319, acc=0.698, acc_len=1.67, thru=12.7, I=13.9, T=35.8, wait=5.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:41:04,912] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:41:06,018] loop.py:45 INFO TRAIN_STEP step=16431 loss=1.362438 Training: 91%|█████████ | 16431/18018 [35:10:52<4:00:41, 9.10s/step, loss=1.362, acc=0.689, acc_len=1.64, thru=15.1, I=14.7, T=34.0, wait=6.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:41:06,013] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16430: forward=2644.7ms backward=1075.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:41:06,419] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:41:13,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:41:13,452] loop.py:45 INFO TRAIN_STEP step=16432 loss=1.269074 Training: 91%|█████████ | 16432/18018 [35:10:59<3:47:20, 8.60s/step, loss=1.269, acc=0.702, acc_len=1.69, thru=16.2, I=19.8, T=37.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:41:13,669] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:41:21,555] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:41:22,081] loop.py:45 INFO TRAIN_STEP step=16433 loss=1.224789 Training: 91%|█████████ | 16433/18018 [35:11:08<3:47:24, 8.61s/step, loss=1.225, acc=0.713, acc_len=1.72, thru=14.6, I=15.3, T=37.3, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:41:22,563] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:41:27,955] loop.py:45 INFO TRAIN_STEP step=16434 loss=1.220296 Training: 91%|█████████ | 16434/18018 [35:11:14<3:25:36, 7.79s/step, loss=1.220, acc=0.710, acc_len=1.72, thru=18.1, I=19.9, T=38.4, wait=2.5s, pool=45, epoch=2/2][2026-10-20 00:41:38,607] loop.py:380 INFO TIMING step=16435: step=3.647s data=1.490s compute=3.613s [fwd=2.548s bwd=1.050s opt=0.003s] dispatch=6.997s [2026-10-20 00:41:38,607] loop.py:45 INFO TRAIN_STEP step=16435 loss=1.288911 Training: 91%|█████████ | 16435/18018 [35:11:24<3:48:08, 8.65s/step, loss=1.289, acc=0.702, acc_len=1.69, thru=14.6, I=12.1, T=35.1, wait=7.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:41:45,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:41:45,713] loop.py:45 INFO TRAIN_STEP step=16436 loss=1.208352 Training: 91%|█████████ | 16436/18018 [35:11:31<3:35:49, 8.19s/step, loss=1.208, acc=0.714, acc_len=1.73, thru=18.3, I=17.8, T=37.2, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:41:45,708] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16435: forward=2462.7ms backward=940.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:41:46,015] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:41:52,651] loop.py:45 INFO TRAIN_STEP step=16437 loss=1.244952 Training: 91%|█████████ | 16437/18018 [35:11:38<3:25:49, 7.81s/step, loss=1.245, acc=0.709, acc_len=1.71, thru=16.7, I=19.9, T=37.1, wait=3.5s, pool=37, epoch=2/2][2026-10-20 00:42:00,727] loop.py:45 INFO TRAIN_STEP step=16438 loss=1.232671 Training: 91%|█████████ | 16438/18018 [35:11:46<3:27:46, 7.89s/step, loss=1.233, acc=0.709, acc_len=1.71, thru=16.1, I=15.5, T=37.2, wait=4.6s, pool=49, epoch=2/2][2026-10-20 00:42:08,178] loop.py:45 INFO TRAIN_STEP step=16439 loss=1.378661 Training: 91%|█████████ | 16439/18018 [35:11:54<3:24:11, 7.76s/step, loss=1.379, acc=0.691, acc_len=1.65, thru=16.2, I=15.9, T=37.9, wait=4.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:42:15,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:42:15,848] loop.py:380 INFO TIMING step=16440: step=3.534s data=1.455s compute=3.508s [fwd=2.536s bwd=0.958s opt=0.003s] dispatch=4.127s [2026-10-20 00:42:15,848] loop.py:45 INFO TRAIN_STEP step=16440 loss=1.180381 Training: 91%|█████████ | 16440/18018 [35:12:01<3:23:20, 7.73s/step, loss=1.180, acc=0.724, acc_len=1.76, thru=18.9, I=17.2, T=36.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:42:16,270] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:42:22,114] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:42:22,159] loop.py:45 INFO TRAIN_STEP step=16441 loss=1.269776 Training: 91%|█████████ | 16441/18018 [35:12:08<3:12:00, 7.31s/step, loss=1.270, acc=0.704, acc_len=1.69, thru=19.8, I=21.1, T=37.4, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:42:22,155] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16440: forward=2432.6ms backward=936.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:42:22,616] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:42:29,069] loop.py:45 INFO TRAIN_STEP step=16442 loss=1.313422 Training: 91%|█████████▏| 16442/18018 [35:12:15<3:08:48, 7.19s/step, loss=1.313, acc=0.692, acc_len=1.66, thru=17.8, I=18.9, T=37.8, wait=3.5s, pool=57, epoch=2/2][2026-10-20 00:42:34,980] loop.py:45 INFO TRAIN_STEP step=16443 loss=1.318944 Training: 91%|█████████▏| 16443/18018 [35:12:21<2:58:37, 6.80s/step, loss=1.319, acc=0.698, acc_len=1.68, thru=17.7, I=19.7, T=39.2, wait=2.6s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:42:42,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:42:42,723] loop.py:45 INFO TRAIN_STEP step=16444 loss=1.147572 Training: 91%|█████████▏| 16444/18018 [35:12:28<3:05:53, 7.09s/step, loss=1.148, acc=0.726, acc_len=1.78, thru=17.4, I=18.0, T=35.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:42:42,823] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-20 00:42:50,055] loop.py:380 INFO TIMING step=16445: step=3.424s data=1.348s compute=3.398s [fwd=2.458s bwd=0.925s opt=0.003s] dispatch=3.898s [2026-10-20 00:42:50,055] loop.py:45 INFO TRAIN_STEP step=16445 loss=1.249260 Training: 91%|█████████▏| 16445/18018 [35:12:36<3:07:42, 7.16s/step, loss=1.249, acc=0.713, acc_len=1.73, thru=16.1, I=18.8, T=37.4, wait=3.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:42:57,386] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:42:58,692] loop.py:45 INFO TRAIN_STEP step=16446 loss=1.158112 Training: 91%|█████████▏| 16446/18018 [35:12:44<3:19:11, 7.60s/step, loss=1.158, acc=0.726, acc_len=1.77, thru=15.1, I=17.4, T=33.3, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:42:58,689] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16445: forward=2738.9ms backward=1067.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:42:58,895] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:43:05,680] loop.py:45 INFO TRAIN_STEP step=16447 loss=1.320372 Training: 91%|█████████▏| 16447/18018 [35:12:51<3:14:14, 7.42s/step, loss=1.320, acc=0.696, acc_len=1.67, thru=14.9, I=19.9, T=37.7, wait=3.6s, pool=45, epoch=2/2][2026-10-20 00:43:14,884] loop.py:45 INFO TRAIN_STEP step=16448 loss=1.285446 Training: 91%|█████████▏| 16448/18018 [35:13:00<3:28:06, 7.95s/step, loss=1.285, acc=0.701, acc_len=1.68, thru=15.8, I=14.9, T=35.3, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:14,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:15,322] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:20,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:43:21,051] loop.py:45 INFO TRAIN_STEP step=16449 loss=1.284810 Training: 91%|█████████▏| 16449/18018 [35:13:07<3:13:57, 7.42s/step, loss=1.285, acc=0.702, acc_len=1.69, thru=20.0, I=22.4, T=37.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:21,080] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:27,573] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:43:28,045] loop.py:380 INFO TIMING step=16450: step=3.481s data=1.393s compute=3.445s [fwd=2.423s bwd=1.006s opt=0.003s] dispatch=3.506s [2026-10-20 00:43:28,045] loop.py:45 INFO TRAIN_STEP step=16450 loss=1.409159 Training: 91%|█████████▏| 16450/18018 [35:13:14<3:10:32, 7.29s/step, loss=1.409, acc=0.679, acc_len=1.60, thru=17.6, I=20.0, T=36.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:28,083] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:33,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:43:34,083] loop.py:45 INFO TRAIN_STEP step=16451 loss=1.242580 Training: 91%|█████████▏| 16451/18018 [35:13:20<3:00:35, 6.91s/step, loss=1.243, acc=0.708, acc_len=1.70, thru=20.5, I=20.1, T=37.6, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:43:34,078] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16450: forward=2406.5ms backward=945.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:34,423] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:43:42,060] loop.py:45 INFO TRAIN_STEP step=16452 loss=1.413132 Training: 91%|█████████▏| 16452/18018 [35:13:28<3:08:47, 7.23s/step, loss=1.413, acc=0.680, acc_len=1.60, thru=14.0, I=16.5, T=36.1, wait=4.4s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:48,142] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:43:49,335] loop.py:45 INFO TRAIN_STEP step=16453 loss=1.317184 Training: 91%|█████████▏| 16453/18018 [35:13:35<3:09:00, 7.25s/step, loss=1.317, acc=0.695, acc_len=1.66, thru=18.2, I=18.8, T=36.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:43:49,651] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:43:55,539] loop.py:45 INFO TRAIN_STEP step=16454 loss=1.230010 Training: 91%|█████████▏| 16454/18018 [35:13:41<3:00:44, 6.93s/step, loss=1.230, acc=0.712, acc_len=1.72, thru=18.4, I=23.0, T=38.9, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:44:02,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:44:03,376] loop.py:380 INFO TIMING step=16455: step=3.548s data=1.402s compute=3.504s [fwd=2.467s bwd=1.023s opt=0.003s] dispatch=4.277s [2026-10-20 00:44:03,376] loop.py:45 INFO TRAIN_STEP step=16455 loss=1.371908 Training: 91%|█████████▏| 16455/18018 [35:13:49<3:07:40, 7.20s/step, loss=1.372, acc=0.688, acc_len=1.64, thru=17.7, I=18.5, T=36.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:44:03,535] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:44:10,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:44:11,005] loop.py:45 INFO TRAIN_STEP step=16456 loss=1.239554 Training: 91%|█████████▏| 16456/18018 [35:13:57<3:10:51, 7.33s/step, loss=1.240, acc=0.710, acc_len=1.72, thru=15.7, I=16.9, T=35.6, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:44:11,001] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16455: forward=2531.9ms backward=1000.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:44:11,398] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:44:17,136] loop.py:45 INFO TRAIN_STEP step=16457 loss=1.249749 Training: 91%|█████████▏| 16457/18018 [35:14:03<3:01:22, 6.97s/step, loss=1.250, acc=0.716, acc_len=1.74, thru=17.5, I=19.5, T=39.1, wait=2.8s, pool=37, epoch=2/2][2026-10-20 00:44:28,610] loop.py:45 INFO TRAIN_STEP step=16458 loss=1.169089 Training: 91%|█████████▏| 16458/18018 [35:14:14<3:36:22, 8.32s/step, loss=1.169, acc=0.725, acc_len=1.77, thru=11.8, I=11.0, T=35.6, wait=7.9s, pool=41, epoch=2/2][2026-10-20 00:44:38,005] loop.py:45 INFO TRAIN_STEP step=16459 loss=1.246487 Training: 91%|█████████▏| 16459/18018 [35:14:24<3:44:36, 8.64s/step, loss=1.246, acc=0.707, acc_len=1.71, thru=15.7, I=14.4, T=35.9, wait=5.8s, pool=61, epoch=2/2][2026-10-20 00:44:44,232] loop.py:380 INFO TIMING step=16460: step=3.322s data=1.370s compute=3.275s [fwd=2.291s bwd=0.969s opt=0.003s] dispatch=2.894s [2026-10-20 00:44:44,232] loop.py:45 INFO TRAIN_STEP step=16460 loss=1.327023 Training: 91%|█████████▏| 16460/18018 [35:14:30<3:25:36, 7.92s/step, loss=1.327, acc=0.690, acc_len=1.64, thru=19.7, I=19.5, T=38.5, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:44:50,424] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:44:51,546] loop.py:45 INFO TRAIN_STEP step=16461 loss=1.314336 Training: 91%|█████████▏| 16461/18018 [35:14:37<3:20:47, 7.74s/step, loss=1.314, acc=0.695, acc_len=1.66, thru=18.1, I=20.0, T=37.7, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:44:51,542] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16460: forward=2406.4ms backward=936.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:44:51,931] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:44:58,294] loop.py:45 INFO TRAIN_STEP step=16462 loss=1.256251 Training: 91%|█████████▏| 16462/18018 [35:14:44<3:12:57, 7.44s/step, loss=1.256, acc=0.704, acc_len=1.69, thru=19.1, I=22.0, T=38.2, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:45:04,984] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:45:05,699] loop.py:45 INFO TRAIN_STEP step=16463 loss=1.282841 Training: 91%|█████████▏| 16463/18018 [35:14:51<3:12:34, 7.43s/step, loss=1.283, acc=0.705, acc_len=1.70, thru=18.4, I=18.9, T=37.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:45:05,989] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:45:11,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:45:11,932] loop.py:45 INFO TRAIN_STEP step=16464 loss=1.185119 Training: 91%|█████████▏| 16464/18018 [35:14:58<3:03:07, 7.07s/step, loss=1.185, acc=0.719, acc_len=1.74, thru=20.4, I=22.7, T=38.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:45:12,304] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:45:20,545] loop.py:380 INFO TIMING step=16465: step=3.670s data=1.473s compute=3.642s [fwd=2.618s bwd=1.009s opt=0.003s] dispatch=4.934s [2026-10-20 00:45:20,545] loop.py:45 INFO TRAIN_STEP step=16465 loss=1.222326 Training: 91%|█████████▏| 16465/18018 [35:15:06<3:14:59, 7.53s/step, loss=1.222, acc=0.711, acc_len=1.71, thru=15.3, I=16.1, T=34.9, wait=4.9s, pool=61, epoch=2/2][2026-10-20 00:45:27,070] loop.py:45 INFO TRAIN_STEP step=16466 loss=1.379408 Training: 91%|█████████▏| 16466/18018 [35:15:13<3:07:01, 7.23s/step, loss=1.379, acc=0.686, acc_len=1.63, thru=19.0, I=17.2, T=35.7, wait=2.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:45:27,065] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16465: forward=2537.4ms backward=995.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:45:34,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:45:34,692] loop.py:45 INFO TRAIN_STEP step=16467 loss=1.244972 Training: 91%|█████████▏| 16467/18018 [35:15:20<3:09:57, 7.35s/step, loss=1.245, acc=0.709, acc_len=1.71, thru=16.7, I=16.6, T=37.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:45:34,757] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 [2026-10-20 00:45:40,184] loop.py:45 INFO TRAIN_STEP step=16468 loss=1.245293 Training: 91%|█████████▏| 16468/18018 [35:15:26<2:55:26, 6.79s/step, loss=1.245, acc=0.710, acc_len=1.72, thru=19.7, I=20.1, T=37.5, wait=2.1s, pool=53, epoch=2/2][2026-10-20 00:45:46,085] loop.py:45 INFO TRAIN_STEP step=16469 loss=1.180779 Training: 91%|█████████▏| 16469/18018 [35:15:32<2:48:25, 6.52s/step, loss=1.181, acc=0.724, acc_len=1.76, thru=20.8, I=21.0, T=38.4, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:45:51,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:45:52,447] loop.py:380 INFO TIMING step=16470: step=3.432s data=1.306s compute=3.379s [fwd=2.366s bwd=0.998s opt=0.003s] dispatch=2.923s [2026-10-20 00:45:52,447] loop.py:45 INFO TRAIN_STEP step=16470 loss=1.197071 Training: 91%|█████████▏| 16470/18018 [35:15:38<2:47:04, 6.48s/step, loss=1.197, acc=0.716, acc_len=1.74, thru=20.8, I=22.8, T=37.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:45:52,900] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:45:59,417] loop.py:45 INFO TRAIN_STEP step=16471 loss=1.262097 Training: 91%|█████████▏| 16471/18018 [35:15:45<2:50:46, 6.62s/step, loss=1.262, acc=0.705, acc_len=1.70, thru=18.3, I=19.0, T=36.2, wait=3.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:45:59,412] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16470: forward=2382.7ms backward=1075.3ms [2026-10-20 00:46:06,766] loop.py:45 INFO TRAIN_STEP step=16472 loss=1.233174 Training: 91%|█████████▏| 16472/18018 [35:15:52<2:56:16, 6.84s/step, loss=1.233, acc=0.712, acc_len=1.72, thru=15.4, I=16.7, T=35.9, wait=3.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:46:14,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:46:15,570] loop.py:45 INFO TRAIN_STEP step=16473 loss=1.282833 Training: 91%|█████████▏| 16473/18018 [35:16:01<3:11:20, 7.43s/step, loss=1.283, acc=0.704, acc_len=1.70, thru=16.2, I=16.1, T=36.3, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:46:15,615] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:46:21,627] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:46:22,257] loop.py:45 INFO TRAIN_STEP step=16474 loss=1.296236 Training: 91%|█████████▏| 16474/18018 [35:16:08<3:05:29, 7.21s/step, loss=1.296, acc=0.701, acc_len=1.68, thru=18.9, I=21.7, T=36.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:46:22,631] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:46:28,991] loop.py:380 INFO TIMING step=16475: step=3.325s data=1.313s compute=3.295s [fwd=2.382s bwd=0.898s opt=0.003s] dispatch=3.396s [2026-10-20 00:46:28,992] loop.py:45 INFO TRAIN_STEP step=16475 loss=1.279875 Training: 91%|█████████▏| 16475/18018 [35:16:15<3:01:41, 7.06s/step, loss=1.280, acc=0.701, acc_len=1.69, thru=18.5, I=21.7, T=38.5, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:46:34,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:46:36,047] loop.py:45 INFO TRAIN_STEP step=16476 loss=1.174281 Training: 91%|█████████▏| 16476/18018 [35:16:22<3:01:30, 7.06s/step, loss=1.174, acc=0.719, acc_len=1.75, thru=19.2, I=20.4, T=37.9, wait=3.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:46:36,043] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16475: forward=2410.9ms backward=921.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:46:36,430] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:46:41,323] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:46:41,986] loop.py:45 INFO TRAIN_STEP step=16477 loss=1.192791 Training: 91%|█████████▏| 16477/18018 [35:16:28<2:52:44, 6.73s/step, loss=1.193, acc=0.716, acc_len=1.74, thru=21.0, I=26.5, T=40.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:46:42,328] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:46:48,739] loop.py:45 INFO TRAIN_STEP step=16478 loss=1.232684 Training: 91%|█████████▏| 16478/18018 [35:16:34<2:52:50, 6.73s/step, loss=1.233, acc=0.711, acc_len=1.72, thru=18.8, I=22.7, T=39.0, wait=3.5s, pool=57, epoch=2/2][2026-10-20 00:46:57,153] loop.py:45 INFO TRAIN_STEP step=16479 loss=1.234623 Training: 91%|█████████▏| 16479/18018 [35:16:43<3:05:39, 7.24s/step, loss=1.235, acc=0.709, acc_len=1.71, thru=15.6, I=16.7, T=35.6, wait=4.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:04,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:47:05,266] loop.py:380 INFO TIMING step=16480: step=3.733s data=1.485s compute=3.708s [fwd=2.635s bwd=1.058s opt=0.003s] dispatch=4.369s [2026-10-20 00:47:05,266] loop.py:45 INFO TRAIN_STEP step=16480 loss=1.296224 Training: 91%|█████████▏| 16480/18018 [35:16:51<3:12:15, 7.50s/step, loss=1.296, acc=0.701, acc_len=1.69, thru=16.0, I=17.7, T=34.3, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:05,540] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:47:11,234] loop.py:45 INFO TRAIN_STEP step=16481 loss=1.233031 Training: 91%|█████████▏| 16481/18018 [35:16:57<3:00:25, 7.04s/step, loss=1.233, acc=0.710, acc_len=1.71, thru=18.2, I=22.7, T=37.4, wait=2.5s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:47:11,229] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16480: forward=2390.7ms backward=978.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:17,774] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:47:19,588] loop.py:45 INFO TRAIN_STEP step=16482 loss=1.318802 Training: 91%|█████████▏| 16482/18018 [35:17:05<3:10:17, 7.43s/step, loss=1.319, acc=0.697, acc_len=1.67, thru=16.9, I=19.4, T=35.3, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:19,785] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:25,685] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:47:26,717] loop.py:45 INFO TRAIN_STEP step=16483 loss=1.265379 Training: 91%|█████████▏| 16483/18018 [35:17:12<3:07:49, 7.34s/step, loss=1.265, acc=0.707, acc_len=1.70, thru=16.1, I=22.8, T=36.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:27,197] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:32,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:47:33,261] loop.py:45 INFO TRAIN_STEP step=16484 loss=1.331548 Training: 91%|█████████▏| 16484/18018 [35:17:19<3:01:35, 7.10s/step, loss=1.332, acc=0.692, acc_len=1.65, thru=19.1, I=23.8, T=38.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:33,414] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:38,438] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:47:38,612] loop.py:380 INFO TIMING step=16485: step=3.225s data=1.264s compute=3.197s [fwd=2.290s bwd=0.893s opt=0.003s] dispatch=2.116s [2026-10-20 00:47:38,612] loop.py:45 INFO TRAIN_STEP step=16485 loss=1.248543 Training: 91%|█████████▏| 16485/18018 [35:17:24<2:48:03, 6.58s/step, loss=1.249, acc=0.708, acc_len=1.71, thru=21.7, I=24.9, T=39.7, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:38,940] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:45,678] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:47:46,431] loop.py:45 INFO TRAIN_STEP step=16486 loss=1.295279 Training: 91%|█████████▏| 16486/18018 [35:17:32<2:57:27, 6.95s/step, loss=1.295, acc=0.697, acc_len=1.66, thru=18.0, I=19.3, T=37.2, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:47:46,426] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16485: forward=2435.1ms backward=965.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:46,685] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:47:52,577] loop.py:45 INFO TRAIN_STEP step=16487 loss=1.277258 Training: 92%|█████████▏| 16487/18018 [35:17:38<2:51:11, 6.71s/step, loss=1.277, acc=0.705, acc_len=1.69, thru=19.6, I=22.7, T=38.9, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:58,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:47:58,199] loop.py:45 INFO TRAIN_STEP step=16488 loss=1.257930 Training: 92%|█████████▏| 16488/18018 [35:17:44<2:42:44, 6.38s/step, loss=1.258, acc=0.708, acc_len=1.70, thru=21.6, I=22.4, T=39.2, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:47:58,598] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:48:05,439] loop.py:45 INFO TRAIN_STEP step=16489 loss=1.234161 Training: 92%|█████████▏| 16489/18018 [35:17:51<2:49:12, 6.64s/step, loss=1.234, acc=0.713, acc_len=1.73, thru=17.2, I=21.3, T=37.0, wait=3.8s, pool=53, epoch=2/2][2026-10-20 00:48:11,912] loop.py:380 INFO TIMING step=16490: step=3.509s data=1.442s compute=3.480s [fwd=2.501s bwd=0.964s opt=0.003s] dispatch=2.954s [2026-10-20 00:48:11,912] loop.py:45 INFO TRAIN_STEP step=16490 loss=1.203078 Training: 92%|█████████▏| 16490/18018 [35:17:57<2:47:48, 6.59s/step, loss=1.203, acc=0.715, acc_len=1.74, thru=18.1, I=18.2, T=36.5, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:48:19,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:48:20,027] loop.py:45 INFO TRAIN_STEP step=16491 loss=1.264805 Training: 92%|█████████▏| 16491/18018 [35:18:06<2:59:20, 7.05s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=18.4, I=16.6, T=35.0, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:48:20,024] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16490: forward=2560.5ms backward=1039.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:48:20,260] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:48:25,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:48:26,554] loop.py:45 INFO TRAIN_STEP step=16492 loss=1.274319 Training: 92%|█████████▏| 16492/18018 [35:18:12<2:55:16, 6.89s/step, loss=1.274, acc=0.704, acc_len=1.69, thru=19.5, I=21.2, T=38.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:48:26,884] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:48:33,036] loop.py:45 INFO TRAIN_STEP step=16493 loss=1.270388 Training: 92%|█████████▏| 16493/18018 [35:18:19<2:52:00, 6.77s/step, loss=1.270, acc=0.705, acc_len=1.70, thru=17.4, I=20.1, T=37.9, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:48:41,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:48:42,411] loop.py:45 INFO TRAIN_STEP step=16494 loss=1.202973 Training: 92%|█████████▏| 16494/18018 [35:18:28<3:11:46, 7.55s/step, loss=1.203, acc=0.719, acc_len=1.75, thru=16.9, I=17.2, T=35.2, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:48:42,493] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-20 00:48:48,709] loop.py:380 INFO TIMING step=16495: step=3.563s data=1.400s compute=3.532s [fwd=2.502s bwd=1.015s opt=0.003s] dispatch=2.726s [2026-10-20 00:48:48,709] loop.py:45 INFO TRAIN_STEP step=16495 loss=1.266029 Training: 92%|█████████▏| 16495/18018 [35:18:34<3:02:07, 7.18s/step, loss=1.266, acc=0.708, acc_len=1.71, thru=19.2, I=20.5, T=35.9, wait=2.7s, pool=57, epoch=2/2][2026-10-20 00:48:55,139] loop.py:45 INFO TRAIN_STEP step=16496 loss=1.396505 Training: 92%|█████████▏| 16496/18018 [35:18:41<2:56:19, 6.95s/step, loss=1.397, acc=0.680, acc_len=1.62, thru=17.4, I=17.1, T=38.5, wait=3.1s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:48:55,135] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16495: forward=2371.1ms backward=906.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:03,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:49:03,737] loop.py:45 INFO TRAIN_STEP step=16497 loss=1.337042 Training: 92%|█████████▏| 16497/18018 [35:18:49<3:08:44, 7.45s/step, loss=1.337, acc=0.690, acc_len=1.65, thru=18.1, I=16.1, T=37.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:03,877] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:10,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:49:11,395] loop.py:45 INFO TRAIN_STEP step=16498 loss=1.249636 Training: 92%|█████████▏| 16498/18018 [35:18:57<3:10:13, 7.51s/step, loss=1.250, acc=0.709, acc_len=1.71, thru=16.6, I=18.4, T=35.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:11,458] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:16,873] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:49:17,658] loop.py:45 INFO TRAIN_STEP step=16499 loss=1.238175 Training: 92%|█████████▏| 16499/18018 [35:19:03<3:00:38, 7.14s/step, loss=1.238, acc=0.713, acc_len=1.72, thru=19.7, I=24.3, T=37.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:17,877] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:49:23,752] loop.py:380 INFO TIMING step=16500: step=3.294s data=1.341s compute=3.246s [fwd=2.302s bwd=0.929s opt=0.003s] dispatch=2.791s [2026-10-20 00:49:23,752] loop.py:45 INFO TRAIN_STEP step=16500 loss=1.473101 Training: 92%|█████████▏| 16500/18018 [35:19:09<2:52:36, 6.82s/step, loss=1.473, acc=0.670, acc_len=1.58, thru=20.6, I=24.1, T=38.9, wait=2.8s, pool=61, epoch=2/2][2026-10-20 00:49:23,755] loop.py:419 INFO Saving checkpoint at step 16500... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:23,822] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 00:49:27,134] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0016501 (TrainerActor pid=9310) [2026-10-20 00:49:27,352] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0015751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:28,852] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:28,853] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=1/64 [2026-10-20 00:49:35,185] loop.py:45 INFO TRAIN_STEP step=16501 loss=1.370340 Training: 92%|█████████▏| 16501/18018 [35:19:21<3:27:28, 8.21s/step, loss=1.370, acc=0.688, acc_len=1.64, thru=18.9, I=20.0, T=38.4, wait=3.4s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:49:35,181] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16500: forward=2361.5ms backward=923.9ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 00:49:27,135] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0016501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 00:49:28,421] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0015751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:40,761] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:49:41,789] loop.py:45 INFO TRAIN_STEP step=16502 loss=1.398311 Training: 92%|█████████▏| 16502/18018 [35:19:27<3:15:10, 7.72s/step, loss=1.398, acc=0.682, acc_len=1.62, thru=19.6, I=21.6, T=38.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:42,270] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:49:48,108] loop.py:45 INFO TRAIN_STEP step=16503 loss=1.164335 Training: 92%|█████████▏| 16503/18018 [35:19:34<3:04:26, 7.30s/step, loss=1.164, acc=0.723, acc_len=1.76, thru=15.6, I=25.1, T=39.7, wait=3.1s, pool=17, epoch=2/2](TrainerActor pid=9370) [2026-10-20 00:49:48,698] eagle_store.py:517 WARNING batch_get_buffer missing keys (k528098-aff9637b_hs, k528098-aff9637b_ids, k528098-aff9637b_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:49:58,781] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 00:49:48,701] eagle_store.py:517 WARNING batch_get_buffer missing keys (k528099-91f0d997_hs, k528099-91f0d997_ids, k528099-91f0d997_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-20 00:50:00,014] loop.py:45 INFO TRAIN_STEP step=16504 loss=1.304226 Training: 92%|█████████▏| 16504/18018 [35:19:46<3:39:07, 8.68s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=15.4, I=14.6, T=34.3, wait=8.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:50:00,291] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:50:05,533] loop.py:380 INFO TIMING step=16505: step=3.197s data=1.261s compute=3.169s [fwd=2.301s bwd=0.854s opt=0.003s] dispatch=2.314s [2026-10-20 00:50:05,533] loop.py:45 INFO TRAIN_STEP step=16505 loss=1.247132 Training: 92%|█████████▏| 16505/18018 [35:19:51<3:15:02, 7.73s/step, loss=1.247, acc=0.708, acc_len=1.70, thru=18.4, I=23.0, T=40.0, wait=2.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:50:12,921] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:50:13,316] loop.py:45 INFO TRAIN_STEP step=16506 loss=1.248074 Training: 92%|█████████▏| 16506/18018 [35:19:59<3:15:17, 7.75s/step, loss=1.248, acc=0.710, acc_len=1.72, thru=17.6, I=16.8, T=37.2, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:50:13,311] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16505: forward=2460.6ms backward=941.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:50:13,434] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-20 00:50:20,397] loop.py:45 INFO TRAIN_STEP step=16507 loss=1.231060 Training: 92%|█████████▏| 16507/18018 [35:20:06<3:10:05, 7.55s/step, loss=1.231, acc=0.711, acc_len=1.72, thru=17.0, I=17.1, T=37.1, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:50:28,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:50:29,111] loop.py:45 INFO TRAIN_STEP step=16508 loss=1.206380 Training: 92%|█████████▏| 16508/18018 [35:20:15<3:18:46, 7.90s/step, loss=1.206, acc=0.716, acc_len=1.74, thru=15.9, I=15.4, T=30.5, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:50:29,230] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:50:34,414] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:50:34,895] loop.py:45 INFO TRAIN_STEP step=16509 loss=1.251561 Training: 92%|█████████▏| 16509/18018 [35:20:20<3:02:41, 7.26s/step, loss=1.252, acc=0.710, acc_len=1.72, thru=19.5, I=21.1, T=39.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:50:34,924] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-20 00:50:41,907] loop.py:380 INFO TIMING step=16510: step=3.423s data=1.404s compute=3.381s [fwd=2.413s bwd=0.954s opt=0.003s] dispatch=3.580s [2026-10-20 00:50:41,907] loop.py:45 INFO TRAIN_STEP step=16510 loss=1.252674 Training: 92%|█████████▏| 16510/18018 [35:20:27<3:00:38, 7.19s/step, loss=1.253, acc=0.705, acc_len=1.70, thru=18.5, I=18.5, T=37.4, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:50:49,216] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:50:50,580] loop.py:45 INFO TRAIN_STEP step=16511 loss=1.283069 Training: 92%|█████████▏| 16511/18018 [35:20:36<3:11:45, 7.63s/step, loss=1.283, acc=0.703, acc_len=1.69, thru=16.6, I=17.4, T=33.0, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:50:50,576] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16510: forward=2852.6ms backward=976.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:50:50,738] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:50:56,203] loop.py:45 INFO TRAIN_STEP step=16512 loss=1.276811 Training: 92%|█████████▏| 16512/18018 [35:20:42<2:56:27, 7.03s/step, loss=1.277, acc=0.701, acc_len=1.68, thru=20.0, I=21.7, T=39.1, wait=2.3s, pool=61, epoch=2/2][2026-10-20 00:51:02,813] loop.py:45 INFO TRAIN_STEP step=16513 loss=1.205554 Training: 92%|█████████▏| 16513/18018 [35:20:48<2:53:10, 6.90s/step, loss=1.206, acc=0.715, acc_len=1.73, thru=18.8, I=19.0, T=38.6, wait=3.3s, pool=53, epoch=2/2][2026-10-20 00:51:09,831] loop.py:45 INFO TRAIN_STEP step=16514 loss=1.291804 Training: 92%|█████████▏| 16514/18018 [35:20:55<2:53:54, 6.94s/step, loss=1.292, acc=0.702, acc_len=1.68, thru=18.1, I=19.4, T=38.7, wait=3.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:16,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:51:17,187] loop.py:380 INFO TIMING step=16515: step=3.425s data=1.384s compute=3.389s [fwd=2.409s bwd=0.966s opt=0.003s] dispatch=3.924s [2026-10-20 00:51:17,187] loop.py:45 INFO TRAIN_STEP step=16515 loss=1.208019 Training: 92%|█████████▏| 16515/18018 [35:21:03<2:56:57, 7.06s/step, loss=1.208, acc=0.713, acc_len=1.72, thru=18.7, I=18.4, T=37.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:17,369] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:22,545] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:51:24,056] loop.py:45 INFO TRAIN_STEP step=16516 loss=1.261256 Training: 92%|█████████▏| 16516/18018 [35:21:10<2:55:21, 7.01s/step, loss=1.261, acc=0.705, acc_len=1.70, thru=18.8, I=21.7, T=32.1, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:51:24,051] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16515: forward=3003.7ms backward=924.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:24,553] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:29,625] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:51:30,611] loop.py:45 INFO TRAIN_STEP step=16517 loss=1.229434 Training: 92%|█████████▏| 16517/18018 [35:21:16<2:51:52, 6.87s/step, loss=1.229, acc=0.714, acc_len=1.74, thru=18.5, I=24.6, T=39.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:30,646] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:36,073] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:51:37,176] loop.py:45 INFO TRAIN_STEP step=16518 loss=1.211939 Training: 92%|█████████▏| 16518/18018 [35:21:23<2:49:27, 6.78s/step, loss=1.212, acc=0.716, acc_len=1.73, thru=20.0, I=23.5, T=38.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:37,581] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:42,786] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:51:43,300] loop.py:45 INFO TRAIN_STEP step=16519 loss=1.308134 Training: 92%|█████████▏| 16519/18018 [35:21:29<2:44:26, 6.58s/step, loss=1.308, acc=0.699, acc_len=1.68, thru=21.0, I=25.9, T=36.1, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:43,790] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:51:49,498] loop.py:380 INFO TIMING step=16520: step=3.350s data=1.294s compute=3.324s [fwd=2.419s bwd=0.891s opt=0.003s] dispatch=2.839s [2026-10-20 00:51:49,498] loop.py:45 INFO TRAIN_STEP step=16520 loss=1.265851 Training: 92%|█████████▏| 16520/18018 [35:21:35<2:41:27, 6.47s/step, loss=1.266, acc=0.708, acc_len=1.71, thru=18.7, I=20.8, T=38.2, wait=2.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:56,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:51:57,789] loop.py:45 INFO TRAIN_STEP step=16521 loss=1.295096 Training: 92%|█████████▏| 16521/18018 [35:21:43<2:55:00, 7.01s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=17.6, I=18.5, T=36.5, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:51:57,784] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16520: forward=2435.4ms backward=991.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:51:57,945] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-20 00:52:03,869] loop.py:45 INFO TRAIN_STEP step=16522 loss=1.374277 Training: 92%|█████████▏| 16522/18018 [35:21:49<2:47:53, 6.73s/step, loss=1.374, acc=0.684, acc_len=1.62, thru=19.5, I=22.4, T=37.4, wait=2.6s, pool=57, epoch=2/2][2026-10-20 00:52:09,986] loop.py:45 INFO TRAIN_STEP step=16523 loss=1.211060 Training: 92%|█████████▏| 16523/18018 [35:21:56<2:43:09, 6.55s/step, loss=1.211, acc=0.714, acc_len=1.73, thru=20.0, I=19.6, T=38.3, wait=2.8s, pool=61, epoch=2/2][2026-10-20 00:52:16,013] loop.py:45 INFO TRAIN_STEP step=16524 loss=1.209126 Training: 92%|█████████▏| 16524/18018 [35:22:02<2:39:09, 6.39s/step, loss=1.209, acc=0.717, acc_len=1.74, thru=20.1, I=20.2, T=37.7, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:52:22,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:52:23,145] loop.py:380 INFO TIMING step=16525: step=3.487s data=1.470s compute=3.440s [fwd=2.473s bwd=0.953s opt=0.003s] dispatch=3.638s [2026-10-20 00:52:23,146] loop.py:45 INFO TRAIN_STEP step=16525 loss=1.189804 Training: 92%|█████████▏| 16525/18018 [35:22:09<2:44:35, 6.61s/step, loss=1.190, acc=0.722, acc_len=1.75, thru=18.5, I=19.7, T=36.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:52:23,397] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:52:30,436] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:52:31,117] loop.py:45 INFO TRAIN_STEP step=16526 loss=1.158951 Training: 92%|█████████▏| 16526/18018 [35:22:17<2:54:36, 7.02s/step, loss=1.159, acc=0.726, acc_len=1.77, thru=16.4, I=20.4, T=35.8, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:52:31,112] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16525: forward=2565.9ms backward=964.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:52:31,442] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:52:39,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:52:40,070] loop.py:45 INFO TRAIN_STEP step=16527 loss=1.329991 Training: 92%|█████████▏| 16527/18018 [35:22:26<3:08:52, 7.60s/step, loss=1.330, acc=0.693, acc_len=1.66, thru=13.7, I=13.6, T=36.3, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:52:40,458] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:52:45,924] loop.py:45 INFO TRAIN_STEP step=16528 loss=1.177068 Training: 92%|█████████▏| 16528/18018 [35:22:32<2:55:47, 7.08s/step, loss=1.177, acc=0.723, acc_len=1.76, thru=17.5, I=20.1, T=38.5, wait=2.5s, pool=37, epoch=2/2][2026-10-20 00:52:55,256] loop.py:45 INFO TRAIN_STEP step=16529 loss=1.303746 Training: 92%|█████████▏| 16529/18018 [35:22:41<3:12:23, 7.75s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=18.5, I=15.0, T=35.2, wait=5.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:02,624] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:53:03,162] loop.py:380 INFO TIMING step=16530: step=3.485s data=1.416s compute=3.439s [fwd=2.473s bwd=0.952s opt=0.003s] dispatch=4.413s [2026-10-20 00:53:03,162] loop.py:45 INFO TRAIN_STEP step=16530 loss=1.203127 Training: 92%|█████████▏| 16530/18018 [35:22:49<3:13:24, 7.80s/step, loss=1.203, acc=0.720, acc_len=1.75, thru=16.7, I=17.3, T=36.7, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:03,629] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:12,652] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:53:14,229] loop.py:45 INFO TRAIN_STEP step=16531 loss=1.404359 Training: 92%|█████████▏| 16531/18018 [35:23:00<3:37:34, 8.78s/step, loss=1.404, acc=0.680, acc_len=1.61, thru=11.9, I=13.5, T=28.5, wait=6.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:53:14,225] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16530: forward=3174.1ms backward=1117.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:14,668] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 00:53:20,038] loop.py:45 INFO TRAIN_STEP step=16532 loss=1.134244 Training: 92%|█████████▏| 16532/18018 [35:23:06<3:15:21, 7.89s/step, loss=1.134, acc=0.731, acc_len=1.79, thru=16.2, I=20.9, T=38.4, wait=2.5s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:26,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:53:27,352] loop.py:45 INFO TRAIN_STEP step=16533 loss=1.174463 Training: 92%|█████████▏| 16533/18018 [35:23:13<3:10:58, 7.72s/step, loss=1.174, acc=0.722, acc_len=1.76, thru=19.1, I=20.1, T=36.0, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:27,386] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:33,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:53:33,877] loop.py:45 INFO TRAIN_STEP step=16534 loss=1.306138 Training: 92%|█████████▏| 16534/18018 [35:23:19<3:01:59, 7.36s/step, loss=1.306, acc=0.695, acc_len=1.66, thru=18.6, I=22.3, T=37.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:33,959] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:40,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:53:41,233] loop.py:380 INFO TIMING step=16535: step=3.473s data=1.454s compute=3.451s [fwd=2.489s bwd=0.948s opt=0.003s] dispatch=3.875s [2026-10-20 00:53:41,233] loop.py:45 INFO TRAIN_STEP step=16535 loss=1.199124 Training: 92%|█████████▏| 16535/18018 [35:23:27<3:01:51, 7.36s/step, loss=1.199, acc=0.721, acc_len=1.75, thru=17.2, I=19.2, T=36.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:41,321] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:45,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-20 00:53:46,471] loop.py:45 INFO TRAIN_STEP step=16536 loss=1.273737 Training: 92%|█████████▏| 16536/18018 [35:23:32<2:46:02, 6.72s/step, loss=1.274, acc=0.704, acc_len=1.69, thru=23.0, I=27.4, T=39.6, wait=2.0s, pool=69, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:53:46,467] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16535: forward=2285.3ms backward=902.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:46,795] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:52,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:53:53,170] loop.py:45 INFO TRAIN_STEP step=16537 loss=1.195279 Training: 92%|█████████▏| 16537/18018 [35:23:39<2:45:44, 6.71s/step, loss=1.195, acc=0.720, acc_len=1.75, thru=19.7, I=25.2, T=35.0, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:53:53,322] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:00,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:54:01,254] loop.py:45 INFO TRAIN_STEP step=16538 loss=1.294506 Training: 92%|█████████▏| 16538/18018 [35:23:47<2:55:46, 7.13s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=14.8, I=16.6, T=34.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:01,486] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:06,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:54:07,229] loop.py:45 INFO TRAIN_STEP step=16539 loss=1.220302 Training: 92%|█████████▏| 16539/18018 [35:23:53<2:47:08, 6.78s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=20.1, I=23.2, T=38.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:07,334] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-20 00:54:13,846] loop.py:380 INFO TIMING step=16540: step=3.320s data=1.272s compute=3.273s [fwd=2.339s bwd=0.918s opt=0.003s] dispatch=3.287s [2026-10-20 00:54:13,846] loop.py:45 INFO TRAIN_STEP step=16540 loss=1.235605 Training: 92%|█████████▏| 16540/18018 [35:23:59<2:45:48, 6.73s/step, loss=1.236, acc=0.712, acc_len=1.72, thru=15.6, I=18.8, T=38.6, wait=3.3s, pool=25, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 00:54:13,911] eagle_store.py:517 WARNING batch_get_buffer missing keys (k529279-9d03b589_hs, k529279-9d03b589_ids, k529279-9d03b589_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:22,208] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-20 00:54:13,974] eagle_store.py:517 WARNING batch_get_buffer missing keys (k529278-b82f392a_hs, k529278-b82f392a_ids, k529278-b82f392a_lhs); sleeping 0.50s. [2026-10-20 00:54:22,453] loop.py:45 INFO TRAIN_STEP step=16541 loss=1.156324 Training: 92%|█████████▏| 16541/18018 [35:24:08<2:59:33, 7.29s/step, loss=1.156, acc=0.725, acc_len=1.76, thru=17.2, I=16.8, T=34.8, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:54:22,448] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16540: forward=2398.8ms backward=1069.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:22,711] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:54:30,027] loop.py:45 INFO TRAIN_STEP step=16542 loss=1.303724 Training: 92%|█████████▏| 16542/18018 [35:24:16<3:01:30, 7.38s/step, loss=1.304, acc=0.701, acc_len=1.68, thru=16.4, I=18.1, T=35.4, wait=4.0s, pool=53, epoch=2/2][2026-10-20 00:54:39,300] loop.py:45 INFO TRAIN_STEP step=16543 loss=1.209623 Training: 92%|█████████▏| 16543/18018 [35:24:25<3:15:20, 7.95s/step, loss=1.210, acc=0.720, acc_len=1.75, thru=13.6, I=13.9, T=35.5, wait=5.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:48,377] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:54:49,448] loop.py:45 INFO TRAIN_STEP step=16544 loss=1.291489 Training: 92%|█████████▏| 16544/18018 [35:24:35<3:31:26, 8.61s/step, loss=1.291, acc=0.704, acc_len=1.70, thru=14.4, I=14.5, T=34.5, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:49,887] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:54:55,379] loop.py:380 INFO TIMING step=16545: step=3.578s data=1.587s compute=3.549s [fwd=2.547s bwd=0.988s opt=0.003s] dispatch=2.343s [2026-10-20 00:54:55,379] loop.py:45 INFO TRAIN_STEP step=16545 loss=1.330877 Training: 92%|█████████▏| 16545/18018 [35:24:41<3:11:34, 7.80s/step, loss=1.331, acc=0.697, acc_len=1.66, thru=20.5, I=22.2, T=35.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:55,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:54:55,882] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:02,637] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:55:03,050] loop.py:45 INFO TRAIN_STEP step=16546 loss=1.288362 Training: 92%|█████████▏| 16546/18018 [35:24:49<3:10:29, 7.76s/step, loss=1.288, acc=0.700, acc_len=1.68, thru=17.0, I=17.3, T=37.8, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:55:03,043] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16545: forward=2369.4ms backward=967.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:03,148] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:08,632] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:55:09,605] loop.py:45 INFO TRAIN_STEP step=16547 loss=1.272054 Training: 92%|█████████▏| 16547/18018 [35:24:55<3:01:28, 7.40s/step, loss=1.272, acc=0.705, acc_len=1.70, thru=19.6, I=21.6, T=37.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:09,640] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-20 00:55:16,185] loop.py:45 INFO TRAIN_STEP step=16548 loss=1.158396 Training: 92%|█████████▏| 16548/18018 [35:25:02<2:55:17, 7.15s/step, loss=1.158, acc=0.724, acc_len=1.76, thru=18.0, I=20.0, T=37.5, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:22,068] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:55:22,653] loop.py:45 INFO TRAIN_STEP step=16549 loss=1.177875 Training: 92%|█████████▏| 16549/18018 [35:25:08<2:50:07, 6.95s/step, loss=1.178, acc=0.723, acc_len=1.76, thru=19.8, I=19.9, T=37.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:23,074] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:27,998] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:55:28,156] loop.py:380 INFO TIMING step=16550: step=3.262s data=1.262s compute=3.234s [fwd=2.335s bwd=0.884s opt=0.003s] dispatch=2.233s [2026-10-20 00:55:28,157] loop.py:45 INFO TRAIN_STEP step=16550 loss=1.194567 Training: 92%|█████████▏| 16550/18018 [35:25:14<2:39:23, 6.51s/step, loss=1.195, acc=0.719, acc_len=1.75, thru=22.1, I=25.0, T=39.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:28,500] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:55:34,702] loop.py:45 INFO TRAIN_STEP step=16551 loss=1.316658 Training: 92%|█████████▏| 16551/18018 [35:25:20<2:39:31, 6.52s/step, loss=1.317, acc=0.699, acc_len=1.68, thru=16.6, I=19.8, T=38.8, wait=3.2s, pool=29, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:55:34,697] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16550: forward=2370.7ms backward=893.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:42,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:55:42,972] loop.py:45 INFO TRAIN_STEP step=16552 loss=1.284764 Training: 92%|█████████▏| 16552/18018 [35:25:29<2:52:12, 7.05s/step, loss=1.285, acc=0.703, acc_len=1.68, thru=18.4, I=16.5, T=34.8, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:43,005] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-20 00:55:48,851] loop.py:45 INFO TRAIN_STEP step=16553 loss=1.261002 Training: 92%|█████████▏| 16553/18018 [35:25:34<2:43:30, 6.70s/step, loss=1.261, acc=0.705, acc_len=1.70, thru=18.8, I=22.0, T=38.4, wait=2.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:54,755] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:55:55,899] loop.py:45 INFO TRAIN_STEP step=16554 loss=1.165555 Training: 92%|█████████▏| 16554/18018 [35:25:41<2:45:59, 6.80s/step, loss=1.166, acc=0.724, acc_len=1.76, thru=18.5, I=20.1, T=37.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:55:56,265] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 00:56:01,940] loop.py:380 INFO TIMING step=16555: step=3.228s data=1.274s compute=3.199s [fwd=2.328s bwd=0.856s opt=0.003s] dispatch=2.804s [2026-10-20 00:56:01,940] loop.py:45 INFO TRAIN_STEP step=16555 loss=1.327274 Training: 92%|█████████▏| 16555/18018 [35:25:48<2:40:19, 6.58s/step, loss=1.327, acc=0.692, acc_len=1.65, thru=17.6, I=25.2, T=39.7, wait=2.8s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:56:12,109] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16555: forward=2670.5ms backward=1026.7ms [2026-10-20 00:56:12,113] loop.py:45 INFO TRAIN_STEP step=16556 loss=1.209182 Training: 92%|█████████▏| 16556/18018 [35:25:58<3:06:28, 7.65s/step, loss=1.209, acc=0.716, acc_len=1.73, thru=15.7, I=12.4, T=34.1, wait=6.4s, pool=57, epoch=2/2][2026-10-20 00:56:17,831] loop.py:45 INFO TRAIN_STEP step=16557 loss=1.273828 Training: 92%|█████████▏| 16557/18018 [35:26:03<2:52:12, 7.07s/step, loss=1.274, acc=0.705, acc_len=1.70, thru=20.6, I=21.3, T=37.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:56:17,812] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:56:18,314] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:56:25,212] loop.py:45 INFO TRAIN_STEP step=16558 loss=1.242863 Training: 92%|█████████▏| 16558/18018 [35:26:11<2:54:21, 7.17s/step, loss=1.243, acc=0.711, acc_len=1.72, thru=16.7, I=18.6, T=37.2, wait=3.9s, pool=53, epoch=2/2][2026-10-20 00:56:33,285] loop.py:45 INFO TRAIN_STEP step=16559 loss=1.326071 Training: 92%|█████████▏| 16559/18018 [35:26:19<3:00:51, 7.44s/step, loss=1.326, acc=0.695, acc_len=1.67, thru=16.2, I=15.6, T=37.2, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:56:42,752] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:56:42,836] loop.py:380 INFO TIMING step=16560: step=3.493s data=1.405s compute=3.469s [fwd=2.491s bwd=0.964s opt=0.003s] dispatch=6.049s [2026-10-20 00:56:42,836] loop.py:45 INFO TRAIN_STEP step=16560 loss=1.311841 Training: 92%|█████████▏| 16560/18018 [35:26:28<3:16:08, 8.07s/step, loss=1.312, acc=0.697, acc_len=1.67, thru=14.5, I=15.3, T=36.6, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:56:43,255] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-20 00:56:49,832] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16560: forward=2413.7ms backward=960.1ms [2026-10-20 00:56:49,837] loop.py:45 INFO TRAIN_STEP step=16561 loss=1.292418 Training: 92%|█████████▏| 16561/18018 [35:26:35<3:08:13, 7.75s/step, loss=1.292, acc=0.703, acc_len=1.69, thru=18.7, I=20.2, T=37.3, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:56:59,009] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:57:00,400] loop.py:45 INFO TRAIN_STEP step=16562 loss=1.288319 Training: 92%|█████████▏| 16562/18018 [35:26:46<3:28:33, 8.59s/step, loss=1.288, acc=0.699, acc_len=1.68, thru=13.7, I=13.8, T=36.3, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:57:00,519] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-20 00:57:06,295] loop.py:45 INFO TRAIN_STEP step=16563 loss=1.251314 Training: 92%|█████████▏| 16563/18018 [35:26:52<3:08:45, 7.78s/step, loss=1.251, acc=0.710, acc_len=1.72, thru=19.7, I=21.8, T=39.5, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:57:14,864] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:57:15,386] loop.py:45 INFO TRAIN_STEP step=16564 loss=1.351202 Training: 92%|█████████▏| 16564/18018 [35:27:01<3:18:07, 8.18s/step, loss=1.351, acc=0.687, acc_len=1.64, thru=15.4, I=16.2, T=35.1, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:57:15,871] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:57:26,056] loop.py:380 INFO TIMING step=16565: step=3.962s data=1.684s compute=3.936s [fwd=2.820s bwd=1.101s opt=0.003s] dispatch=6.700s [2026-10-20 00:57:26,056] loop.py:45 INFO TRAIN_STEP step=16565 loss=1.179237 Training: 92%|█████████▏| 16565/18018 [35:27:12<3:36:07, 8.92s/step, loss=1.179, acc=0.724, acc_len=1.77, thru=11.9, I=11.1, T=32.3, wait=6.7s, pool=57, epoch=2/2][2026-10-20 00:57:32,997] loop.py:45 INFO TRAIN_STEP step=16566 loss=1.172238 Training: 92%|█████████▏| 16566/18018 [35:27:19<3:21:33, 8.33s/step, loss=1.172, acc=0.725, acc_len=1.76, thru=16.9, I=17.7, T=37.8, wait=3.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:57:32,993] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16565: forward=2388.6ms backward=937.8ms [2026-10-20 00:57:39,143] loop.py:45 INFO TRAIN_STEP step=16567 loss=1.356214 Training: 92%|█████████▏| 16567/18018 [35:27:25<3:05:34, 7.67s/step, loss=1.356, acc=0.693, acc_len=1.66, thru=18.4, I=19.0, T=36.9, wait=2.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:57:47,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:57:47,182] loop.py:45 INFO TRAIN_STEP step=16568 loss=1.183867 Training: 92%|█████████▏| 16568/18018 [35:27:33<3:08:07, 7.78s/step, loss=1.184, acc=0.721, acc_len=1.75, thru=16.7, I=16.3, T=35.9, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:57:47,613] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:57:52,862] loop.py:45 INFO TRAIN_STEP step=16569 loss=1.167928 Training: 92%|█████████▏| 16569/18018 [35:27:38<2:52:44, 7.15s/step, loss=1.168, acc=0.722, acc_len=1.75, thru=18.6, I=18.6, T=39.4, wait=2.4s, pool=49, epoch=2/2][2026-10-20 00:57:58,924] loop.py:380 INFO TIMING step=16570: step=3.288s data=1.302s compute=3.259s [fwd=2.344s bwd=0.900s opt=0.003s] dispatch=2.765s [2026-10-20 00:57:58,924] loop.py:45 INFO TRAIN_STEP step=16570 loss=1.165630 Training: 92%|█████████▏| 16570/18018 [35:27:45<2:44:43, 6.83s/step, loss=1.166, acc=0.723, acc_len=1.76, thru=21.4, I=22.7, T=38.9, wait=2.8s, pool=61, epoch=2/2][2026-10-20 00:58:06,842] loop.py:45 INFO TRAIN_STEP step=16571 loss=1.283355 Training: 92%|█████████▏| 16571/18018 [35:27:52<2:52:30, 7.15s/step, loss=1.283, acc=0.703, acc_len=1.69, thru=16.1, I=17.2, T=36.2, wait=4.4s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:58:06,837] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16570: forward=2537.9ms backward=954.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:58:14,318] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:58:14,378] loop.py:45 INFO TRAIN_STEP step=16572 loss=1.279067 Training: 92%|█████████▏| 16572/18018 [35:28:00<2:55:09, 7.27s/step, loss=1.279, acc=0.701, acc_len=1.68, thru=17.4, I=16.7, T=35.7, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:58:14,819] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:58:19,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:58:20,524] loop.py:45 INFO TRAIN_STEP step=16573 loss=1.254633 Training: 92%|█████████▏| 16573/18018 [35:28:06<2:46:55, 6.93s/step, loss=1.255, acc=0.708, acc_len=1.71, thru=20.1, I=21.9, T=37.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:58:20,793] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:58:26,620] loop.py:45 INFO TRAIN_STEP step=16574 loss=1.222516 Training: 92%|█████████▏| 16574/18018 [35:28:12<2:40:47, 6.68s/step, loss=1.223, acc=0.715, acc_len=1.73, thru=19.1, I=22.2, T=38.9, wait=2.8s, pool=49, epoch=2/2][2026-10-20 00:58:33,682] loop.py:380 INFO TIMING step=16575: step=3.437s data=1.400s compute=3.410s [fwd=2.509s bwd=0.887s opt=0.003s] dispatch=3.616s [2026-10-20 00:58:33,682] loop.py:45 INFO TRAIN_STEP step=16575 loss=1.323518 Training: 92%|█████████▏| 16575/18018 [35:28:19<2:43:25, 6.80s/step, loss=1.324, acc=0.689, acc_len=1.64, thru=16.0, I=20.1, T=37.2, wait=3.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:58:42,116] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:58:42,580] loop.py:45 INFO TRAIN_STEP step=16576 loss=1.266534 Training: 92%|█████████▏| 16576/18018 [35:28:28<2:58:28, 7.43s/step, loss=1.267, acc=0.706, acc_len=1.70, thru=17.0, I=15.3, T=35.0, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:58:42,576] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16575: forward=2565.0ms backward=1051.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:58:42,625] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:58:47,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:58:47,878] loop.py:45 INFO TRAIN_STEP step=16577 loss=1.161621 Training: 92%|█████████▏| 16577/18018 [35:28:33<2:43:00, 6.79s/step, loss=1.162, acc=0.726, acc_len=1.77, thru=21.0, I=24.1, T=39.6, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:58:48,288] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:58:53,966] loop.py:45 INFO TRAIN_STEP step=16578 loss=1.240097 Training: 92%|█████████▏| 16578/18018 [35:28:40<2:37:52, 6.58s/step, loss=1.240, acc=0.709, acc_len=1.71, thru=19.8, I=22.9, T=38.8, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:59:04,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:59:04,712] loop.py:45 INFO TRAIN_STEP step=16579 loss=1.308047 Training: 92%|█████████▏| 16579/18018 [35:28:50<3:07:45, 7.83s/step, loss=1.308, acc=0.698, acc_len=1.68, thru=14.1, I=13.2, T=35.3, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:59:05,056] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 00:59:11,286] loop.py:380 INFO TIMING step=16580: step=3.412s data=1.415s compute=3.376s [fwd=2.427s bwd=0.935s opt=0.003s] dispatch=3.151s [2026-10-20 00:59:11,286] loop.py:45 INFO TRAIN_STEP step=16580 loss=1.226855 Training: 92%|█████████▏| 16580/18018 [35:28:57<2:58:34, 7.45s/step, loss=1.227, acc=0.711, acc_len=1.72, thru=17.9, I=20.1, T=37.5, wait=3.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:59:19,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:59:20,159] loop.py:45 INFO TRAIN_STEP step=16581 loss=1.233702 Training: 92%|█████████▏| 16581/18018 [35:29:06<3:08:40, 7.88s/step, loss=1.234, acc=0.711, acc_len=1.72, thru=16.5, I=16.0, T=35.7, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:59:20,155] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16580: forward=2558.2ms backward=984.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:59:20,566] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:59:26,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 00:59:26,477] loop.py:45 INFO TRAIN_STEP step=16582 loss=1.292417 Training: 92%|█████████▏| 16582/18018 [35:29:12<2:57:22, 7.41s/step, loss=1.292, acc=0.700, acc_len=1.68, thru=19.7, I=21.3, T=38.2, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 00:59:26,849] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 00:59:32,076] loop.py:45 INFO TRAIN_STEP step=16583 loss=1.162484 Training: 92%|█████████▏| 16583/18018 [35:29:18<2:44:13, 6.87s/step, loss=1.162, acc=0.722, acc_len=1.75, thru=21.0, I=21.7, T=37.4, wait=2.2s, pool=57, epoch=2/2][2026-10-20 00:59:39,275] loop.py:45 INFO TRAIN_STEP step=16584 loss=1.282425 Training: 92%|█████████▏| 16584/18018 [35:29:25<2:46:29, 6.97s/step, loss=1.282, acc=0.704, acc_len=1.69, thru=17.2, I=19.3, T=38.4, wait=3.9s, pool=57, epoch=2/2][2026-10-20 00:59:46,600] loop.py:380 INFO TIMING step=16585: step=3.646s data=1.567s compute=3.618s [fwd=2.629s bwd=0.974s opt=0.003s] dispatch=3.671s [2026-10-20 00:59:46,600] loop.py:45 INFO TRAIN_STEP step=16585 loss=1.279828 Training: 92%|█████████▏| 16585/18018 [35:29:32<2:48:56, 7.07s/step, loss=1.280, acc=0.700, acc_len=1.68, thru=17.1, I=17.9, T=35.1, wait=3.7s, pool=53, epoch=2/2][2026-10-20 00:59:53,132] loop.py:45 INFO TRAIN_STEP step=16586 loss=1.228069 Training: 92%|█████████▏| 16586/18018 [35:29:39<2:44:57, 6.91s/step, loss=1.228, acc=0.713, acc_len=1.72, thru=17.0, I=19.6, T=37.7, wait=3.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 00:59:53,128] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16585: forward=2450.4ms backward=906.7ms [2026-10-20 01:00:00,709] loop.py:45 INFO TRAIN_STEP step=16587 loss=1.212078 Training: 92%|█████████▏| 16587/18018 [35:29:46<2:49:35, 7.11s/step, loss=1.212, acc=0.715, acc_len=1.73, thru=16.4, I=16.0, T=37.0, wait=4.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:00:08,886] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:00:09,966] loop.py:45 INFO TRAIN_STEP step=16588 loss=1.275004 Training: 92%|█████████▏| 16588/18018 [35:29:56<3:04:49, 7.75s/step, loss=1.275, acc=0.704, acc_len=1.70, thru=15.7, I=15.8, T=34.3, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:00:10,394] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:00:17,413] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:00:17,949] loop.py:45 INFO TRAIN_STEP step=16589 loss=1.240594 Training: 92%|█████████▏| 16589/18018 [35:30:04<3:06:19, 7.82s/step, loss=1.241, acc=0.712, acc_len=1.72, thru=15.4, I=18.0, T=36.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:00:18,419] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:00:24,266] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:00:24,356] loop.py:380 INFO TIMING step=16590: step=3.294s data=1.292s compute=3.267s [fwd=2.341s bwd=0.911s opt=0.003s] dispatch=3.104s [2026-10-20 01:00:24,357] loop.py:45 INFO TRAIN_STEP step=16590 loss=1.254965 Training: 92%|█████████▏| 16590/18018 [35:30:10<2:56:05, 7.40s/step, loss=1.255, acc=0.708, acc_len=1.71, thru=19.7, I=23.4, T=38.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:00:24,768] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:00:31,477] loop.py:45 INFO TRAIN_STEP step=16591 loss=1.193112 Training: 92%|█████████▏| 16591/18018 [35:30:17<2:53:59, 7.32s/step, loss=1.193, acc=0.720, acc_len=1.75, thru=18.0, I=18.9, T=36.8, wait=3.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:00:31,473] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16590: forward=2461.1ms backward=972.0ms [2026-10-20 01:00:38,953] loop.py:45 INFO TRAIN_STEP step=16592 loss=1.342090 Training: 92%|█████████▏| 16592/18018 [35:30:25<2:55:01, 7.36s/step, loss=1.342, acc=0.693, acc_len=1.65, thru=17.5, I=17.4, T=35.8, wait=3.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:00:44,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:00:45,058] loop.py:45 INFO TRAIN_STEP step=16593 loss=1.277901 Training: 92%|█████████▏| 16593/18018 [35:30:31<2:45:54, 6.99s/step, loss=1.278, acc=0.700, acc_len=1.68, thru=20.4, I=21.0, T=39.3, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:00:45,215] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:00:52,655] loop.py:45 INFO TRAIN_STEP step=16594 loss=1.216424 Training: 92%|█████████▏| 16594/18018 [35:30:38<2:50:10, 7.17s/step, loss=1.216, acc=0.713, acc_len=1.73, thru=16.6, I=20.6, T=35.8, wait=4.0s, pool=53, epoch=2/2][2026-10-20 01:00:59,637] loop.py:380 INFO TIMING step=16595: step=3.575s data=1.478s compute=3.551s [fwd=2.537s bwd=0.999s opt=0.003s] dispatch=3.395s [2026-10-20 01:00:59,638] loop.py:45 INFO TRAIN_STEP step=16595 loss=1.295306 Training: 92%|█████████▏| 16595/18018 [35:30:45<2:48:41, 7.11s/step, loss=1.295, acc=0.699, acc_len=1.67, thru=16.0, I=16.8, T=35.8, wait=3.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:05,149] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:01:06,572] loop.py:45 INFO TRAIN_STEP step=16596 loss=1.235741 Training: 92%|█████████▏| 16596/18018 [35:30:52<2:47:18, 7.06s/step, loss=1.236, acc=0.710, acc_len=1.72, thru=19.8, I=21.1, T=37.3, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:01:06,568] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16595: forward=2450.3ms backward=935.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:06,670] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:10,911] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:01:12,057] loop.py:45 INFO TRAIN_STEP step=16597 loss=1.193544 Training: 92%|█████████▏| 16597/18018 [35:30:58<2:35:59, 6.59s/step, loss=1.194, acc=0.718, acc_len=1.74, thru=22.3, I=29.1, T=38.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:12,418] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:01:18,556] loop.py:45 INFO TRAIN_STEP step=16598 loss=1.241253 Training: 92%|█████████▏| 16598/18018 [35:31:04<2:35:16, 6.56s/step, loss=1.241, acc=0.709, acc_len=1.71, thru=19.1, I=22.3, T=36.2, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:25,311] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:01:26,455] loop.py:45 INFO TRAIN_STEP step=16599 loss=1.357680 Training: 92%|█████████▏| 16599/18018 [35:31:12<2:44:39, 6.96s/step, loss=1.358, acc=0.688, acc_len=1.64, thru=17.4, I=18.8, T=36.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:26,820] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:01:32,502] loop.py:380 INFO TIMING step=16600: step=3.228s data=1.273s compute=3.196s [fwd=2.297s bwd=0.883s opt=0.003s] dispatch=2.810s [2026-10-20 01:01:32,502] loop.py:45 INFO TRAIN_STEP step=16600 loss=1.152333 Training: 92%|█████████▏| 16600/18018 [35:31:18<2:38:02, 6.69s/step, loss=1.152, acc=0.725, acc_len=1.77, thru=18.4, I=25.3, T=39.6, wait=2.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:41,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:01:41,697] loop.py:45 INFO TRAIN_STEP step=16601 loss=1.306832 Training: 92%|█████████▏| 16601/18018 [35:31:27<2:55:42, 7.44s/step, loss=1.307, acc=0.701, acc_len=1.68, thru=16.1, I=15.9, T=34.6, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:01:41,692] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16600: forward=2593.8ms backward=1053.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:42,030] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:48,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:01:49,305] loop.py:45 INFO TRAIN_STEP step=16602 loss=1.157172 Training: 92%|█████████▏| 16602/18018 [35:31:35<2:56:46, 7.49s/step, loss=1.157, acc=0.726, acc_len=1.77, thru=17.0, I=20.1, T=36.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:49,409] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:55,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:01:55,222] loop.py:45 INFO TRAIN_STEP step=16603 loss=1.273052 Training: 92%|█████████▏| 16603/18018 [35:31:41<2:45:30, 7.02s/step, loss=1.273, acc=0.704, acc_len=1.69, thru=20.8, I=20.3, T=39.1, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:01:55,530] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:02:02,626] loop.py:45 INFO TRAIN_STEP step=16604 loss=1.405077 Training: 92%|█████████▏| 16604/18018 [35:31:48<2:48:07, 7.13s/step, loss=1.405, acc=0.681, acc_len=1.62, thru=15.2, I=18.3, T=34.8, wait=3.7s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:02:10,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:02:11,555] loop.py:380 INFO TIMING step=16605: step=3.853s data=1.696s compute=3.811s [fwd=2.742s bwd=1.055s opt=0.003s] dispatch=5.063s [2026-10-20 01:02:11,555] loop.py:45 INFO TRAIN_STEP step=16605 loss=1.312665 Training: 92%|█████████▏| 16605/18018 [35:31:57<3:00:40, 7.67s/step, loss=1.313, acc=0.697, acc_len=1.66, thru=16.7, I=17.4, T=33.2, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:02:11,588] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-20 01:02:17,793] loop.py:45 INFO TRAIN_STEP step=16606 loss=1.273242 Training: 92%|█████████▏| 16606/18018 [35:32:03<2:50:25, 7.24s/step, loss=1.273, acc=0.702, acc_len=1.68, thru=17.1, I=24.2, T=38.0, wait=2.9s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:02:17,789] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16605: forward=2372.1ms backward=955.9ms [2026-10-20 01:02:25,973] loop.py:45 INFO TRAIN_STEP step=16607 loss=1.273625 Training: 92%|█████████▏| 16607/18018 [35:32:12<2:56:55, 7.52s/step, loss=1.274, acc=0.704, acc_len=1.69, thru=15.9, I=18.7, T=36.7, wait=4.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:02:32,836] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:02:33,925] loop.py:45 INFO TRAIN_STEP step=16608 loss=1.266030 Training: 92%|█████████▏| 16608/18018 [35:32:20<2:59:49, 7.65s/step, loss=1.266, acc=0.703, acc_len=1.69, thru=16.3, I=17.9, T=36.2, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:02:34,343] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:02:38,866] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:02:39,873] loop.py:45 INFO TRAIN_STEP step=16609 loss=1.167787 Training: 92%|█████████▏| 16609/18018 [35:32:25<2:47:41, 7.14s/step, loss=1.168, acc=0.725, acc_len=1.77, thru=20.8, I=27.4, T=38.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:02:40,373] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:02:46,081] loop.py:380 INFO TIMING step=16610: step=3.267s data=1.320s compute=3.238s [fwd=2.336s bwd=0.890s opt=0.003s] dispatch=2.931s [2026-10-20 01:02:46,081] loop.py:45 INFO TRAIN_STEP step=16610 loss=1.187755 Training: 92%|█████████▏| 16610/18018 [35:32:32<2:41:00, 6.86s/step, loss=1.188, acc=0.719, acc_len=1.75, thru=18.6, I=24.9, T=39.2, wait=2.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:02:55,186] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:02:55,305] loop.py:45 INFO TRAIN_STEP step=16611 loss=1.157766 Training: 92%|█████████▏| 16611/18018 [35:32:41<2:57:30, 7.57s/step, loss=1.158, acc=0.725, acc_len=1.77, thru=16.1, I=15.5, T=36.6, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:02:55,301] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16610: forward=2431.8ms backward=1009.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:02:55,691] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:03:02,803] loop.py:45 INFO TRAIN_STEP step=16612 loss=1.149690 Training: 92%|█████████▏| 16612/18018 [35:32:48<2:56:52, 7.55s/step, loss=1.150, acc=0.726, acc_len=1.77, thru=17.2, I=16.8, T=36.2, wait=4.0s, pool=61, epoch=2/2][2026-10-20 01:03:09,493] loop.py:45 INFO TRAIN_STEP step=16613 loss=1.266505 Training: 92%|█████████▏| 16613/18018 [35:32:55<2:50:43, 7.29s/step, loss=1.267, acc=0.706, acc_len=1.70, thru=20.0, I=19.1, T=32.4, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:16,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:03:17,889] loop.py:45 INFO TRAIN_STEP step=16614 loss=1.229844 Training: 92%|█████████▏| 16614/18018 [35:33:03<2:58:21, 7.62s/step, loss=1.230, acc=0.709, acc_len=1.71, thru=15.6, I=17.9, T=35.7, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:18,337] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:22,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:03:23,735] loop.py:380 INFO TIMING step=16615: step=3.290s data=1.318s compute=3.260s [fwd=2.350s bwd=0.896s opt=0.003s] dispatch=2.549s [2026-10-20 01:03:23,735] loop.py:45 INFO TRAIN_STEP step=16615 loss=1.265658 Training: 92%|█████████▏| 16615/18018 [35:33:09<2:45:46, 7.09s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=20.6, I=28.1, T=38.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:23,871] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:28,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:03:29,495] loop.py:45 INFO TRAIN_STEP step=16616 loss=1.215896 Training: 92%|█████████▏| 16616/18018 [35:33:15<2:36:21, 6.69s/step, loss=1.216, acc=0.717, acc_len=1.74, thru=22.1, I=30.0, T=38.2, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:03:29,490] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16615: forward=2348.0ms backward=954.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:29,675] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:03:35,931] loop.py:45 INFO TRAIN_STEP step=16617 loss=1.220691 Training: 92%|█████████▏| 16617/18018 [35:33:22<2:34:26, 6.61s/step, loss=1.221, acc=0.710, acc_len=1.71, thru=16.6, I=23.5, T=40.0, wait=3.2s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:43,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:03:44,704] loop.py:45 INFO TRAIN_STEP step=16618 loss=1.108229 Training: 92%|█████████▏| 16618/18018 [35:33:30<2:49:26, 7.26s/step, loss=1.108, acc=0.736, acc_len=1.81, thru=18.2, I=16.7, T=37.9, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:44,978] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:03:50,329] loop.py:45 INFO TRAIN_STEP step=16619 loss=1.224815 Training: 92%|█████████▏| 16619/18018 [35:33:36<2:37:52, 6.77s/step, loss=1.225, acc=0.712, acc_len=1.72, thru=21.1, I=25.0, T=38.3, wait=2.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:55,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:03:56,719] loop.py:380 INFO TIMING step=16620: step=3.343s data=1.362s compute=3.309s [fwd=2.370s bwd=0.925s opt=0.003s] dispatch=3.040s [2026-10-20 01:03:56,720] loop.py:45 INFO TRAIN_STEP step=16620 loss=1.326232 Training: 92%|█████████▏| 16620/18018 [35:33:42<2:35:06, 6.66s/step, loss=1.326, acc=0.694, acc_len=1.66, thru=20.4, I=22.0, T=38.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:03:56,884] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:04:04,555] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:04:06,367] loop.py:45 INFO TRAIN_STEP step=16621 loss=1.175932 Training: 92%|█████████▏| 16621/18018 [35:33:52<2:55:53, 7.55s/step, loss=1.176, acc=0.722, acc_len=1.76, thru=13.5, I=14.6, T=28.4, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:04:06,363] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16620: forward=3306.2ms backward=1122.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:04:06,565] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 01:04:12,077] loop.py:45 INFO TRAIN_STEP step=16622 loss=1.290916 Training: 92%|█████████▏| 16622/18018 [35:33:58<2:42:53, 7.00s/step, loss=1.291, acc=0.703, acc_len=1.69, thru=15.6, I=20.8, T=38.9, wait=2.4s, pool=29, epoch=2/2][2026-10-20 01:04:23,446] loop.py:45 INFO TRAIN_STEP step=16623 loss=1.288518 Training: 92%|█████████▏| 16623/18018 [35:34:09<3:13:13, 8.31s/step, loss=1.289, acc=0.704, acc_len=1.69, thru=16.4, I=12.3, T=34.7, wait=7.7s, pool=57, epoch=2/2][2026-10-20 01:04:30,422] loop.py:45 INFO TRAIN_STEP step=16624 loss=1.104857 Training: 92%|█████████▏| 16624/18018 [35:34:16<3:03:47, 7.91s/step, loss=1.105, acc=0.735, acc_len=1.80, thru=17.2, I=18.2, T=36.2, wait=3.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:04:41,932] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:04:42,935] loop.py:380 INFO TIMING step=16625: step=3.810s data=1.644s compute=3.758s [fwd=2.608s bwd=1.136s opt=0.003s] dispatch=8.694s [2026-10-20 01:04:42,935] loop.py:45 INFO TRAIN_STEP step=16625 loss=1.239942 Training: 92%|█████████▏| 16625/18018 [35:34:29<3:35:41, 9.29s/step, loss=1.240, acc=0.712, acc_len=1.73, thru=12.3, I=11.8, T=33.6, wait=8.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:04:43,438] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:04:48,589] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:04:49,661] loop.py:45 INFO TRAIN_STEP step=16626 loss=1.145668 Training: 92%|█████████▏| 16626/18018 [35:34:35<3:17:43, 8.52s/step, loss=1.146, acc=0.727, acc_len=1.78, thru=18.9, I=21.5, T=36.1, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:04:49,657] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16625: forward=2470.8ms backward=1016.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:04:50,100] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:04:55,879] loop.py:45 INFO TRAIN_STEP step=16627 loss=1.231308 Training: 92%|█████████▏| 16627/18018 [35:34:41<3:01:33, 7.83s/step, loss=1.231, acc=0.709, acc_len=1.71, thru=19.6, I=25.6, T=39.2, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:02,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:05:03,312] loop.py:45 INFO TRAIN_STEP step=16628 loss=1.261304 Training: 92%|█████████▏| 16628/18018 [35:34:49<2:58:38, 7.71s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=17.6, I=18.9, T=37.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:03,628] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:05:10,203] loop.py:45 INFO TRAIN_STEP step=16629 loss=1.308218 Training: 92%|█████████▏| 16629/18018 [35:34:56<2:52:48, 7.46s/step, loss=1.308, acc=0.699, acc_len=1.68, thru=18.4, I=19.8, T=37.1, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:16,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:05:16,954] loop.py:380 INFO TIMING step=16630: step=3.313s data=1.262s compute=3.285s [fwd=2.326s bwd=0.946s opt=0.003s] dispatch=3.430s [2026-10-20 01:05:16,954] loop.py:45 INFO TRAIN_STEP step=16630 loss=1.201627 Training: 92%|█████████▏| 16630/18018 [35:35:03<2:47:44, 7.25s/step, loss=1.202, acc=0.715, acc_len=1.73, thru=20.1, I=18.6, T=38.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:17,291] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:22,697] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 01:05:23,666] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16630: forward=2358.8ms backward=934.6ms [2026-10-20 01:05:23,671] loop.py:45 INFO TRAIN_STEP step=16631 loss=1.170001 Training: 92%|█████████▏| 16631/18018 [35:35:09<2:43:54, 7.09s/step, loss=1.170, acc=0.718, acc_len=1.74, thru=19.8, I=20.5, T=38.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:23,701] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:28,373] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:05:29,482] loop.py:45 INFO TRAIN_STEP step=16632 loss=1.259020 Training: 92%|█████████▏| 16632/18018 [35:35:15<2:34:56, 6.71s/step, loss=1.259, acc=0.707, acc_len=1.71, thru=21.4, I=26.7, T=35.8, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:29,882] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:36,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:05:37,108] loop.py:45 INFO TRAIN_STEP step=16633 loss=1.294627 Training: 92%|█████████▏| 16633/18018 [35:35:23<2:41:09, 6.98s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=16.9, I=20.6, T=40.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:37,555] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:05:43,844] loop.py:45 INFO TRAIN_STEP step=16634 loss=1.369974 Training: 92%|█████████▏| 16634/18018 [35:35:29<2:39:20, 6.91s/step, loss=1.370, acc=0.687, acc_len=1.64, thru=19.3, I=17.5, T=38.3, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:52,365] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:05:53,027] loop.py:380 INFO TIMING step=16635: step=3.474s data=1.406s compute=3.446s [fwd=2.465s bwd=0.968s opt=0.003s] dispatch=5.702s [2026-10-20 01:05:53,028] loop.py:45 INFO TRAIN_STEP step=16635 loss=1.146184 Training: 92%|█████████▏| 16635/18018 [35:35:39<2:54:58, 7.59s/step, loss=1.146, acc=0.730, acc_len=1.78, thru=15.5, I=16.7, T=36.8, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:05:53,372] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:06:00,052] loop.py:45 INFO TRAIN_STEP step=16636 loss=1.104632 Training: 92%|█████████▏| 16636/18018 [35:35:46<2:50:55, 7.42s/step, loss=1.105, acc=0.734, acc_len=1.80, thru=16.5, I=16.3, T=38.5, wait=3.7s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:06:00,048] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16635: forward=2345.9ms backward=934.3ms [2026-10-20 01:06:08,297] loop.py:45 INFO TRAIN_STEP step=16637 loss=1.275984 Training: 92%|█████████▏| 16637/18018 [35:35:54<2:56:29, 7.67s/step, loss=1.276, acc=0.706, acc_len=1.70, thru=17.4, I=16.8, T=37.3, wait=4.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:06:16,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:06:16,326] loop.py:45 INFO TRAIN_STEP step=16638 loss=1.187075 Training: 92%|█████████▏| 16638/18018 [35:36:02<2:58:51, 7.78s/step, loss=1.187, acc=0.717, acc_len=1.74, thru=16.5, I=17.4, T=35.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:06:16,581] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:06:23,147] loop.py:45 INFO TRAIN_STEP step=16639 loss=1.173044 Training: 92%|█████████▏| 16639/18018 [35:36:09<2:52:09, 7.49s/step, loss=1.173, acc=0.721, acc_len=1.75, thru=18.1, I=20.0, T=37.8, wait=3.4s, pool=61, epoch=2/2][2026-10-20 01:06:33,478] loop.py:380 INFO TIMING step=16640: step=3.606s data=1.464s compute=3.579s [fwd=2.575s bwd=0.990s opt=0.003s] dispatch=6.714s [2026-10-20 01:06:33,478] loop.py:45 INFO TRAIN_STEP step=16640 loss=1.319381 Training: 92%|█████████▏| 16640/18018 [35:36:19<3:11:35, 8.34s/step, loss=1.319, acc=0.695, acc_len=1.67, thru=12.8, I=11.1, T=35.5, wait=6.7s, pool=49, epoch=2/2][2026-10-20 01:06:41,017] loop.py:45 INFO TRAIN_STEP step=16641 loss=1.281230 Training: 92%|█████████▏| 16641/18018 [35:36:27<3:05:55, 8.10s/step, loss=1.281, acc=0.707, acc_len=1.70, thru=16.1, I=15.7, T=37.7, wait=4.1s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:06:41,012] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16640: forward=2425.2ms backward=929.5ms [2026-10-20 01:06:50,393] loop.py:45 INFO TRAIN_STEP step=16642 loss=1.222239 Training: 92%|█████████▏| 16642/18018 [35:36:36<3:14:33, 8.48s/step, loss=1.222, acc=0.714, acc_len=1.73, thru=15.0, I=15.5, T=32.4, wait=5.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:06:56,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:06:57,337] loop.py:45 INFO TRAIN_STEP step=16643 loss=1.352380 Training: 92%|█████████▏| 16643/18018 [35:36:43<3:03:49, 8.02s/step, loss=1.352, acc=0.691, acc_len=1.64, thru=18.1, I=18.9, T=36.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:06:57,709] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:02,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:07:03,005] loop.py:45 INFO TRAIN_STEP step=16644 loss=1.269067 Training: 92%|█████████▏| 16644/18018 [35:36:49<2:47:32, 7.32s/step, loss=1.269, acc=0.701, acc_len=1.68, thru=21.5, I=26.4, T=39.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:03,171] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:12,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:07:12,668] loop.py:380 INFO TIMING step=16645: step=4.069s data=1.990s compute=4.033s [fwd=3.048s bwd=0.970s opt=0.003s] dispatch=5.583s [2026-10-20 01:07:12,668] loop.py:45 INFO TRAIN_STEP step=16645 loss=1.278905 Training: 92%|█████████▏| 16645/18018 [35:36:58<3:03:30, 8.02s/step, loss=1.279, acc=0.705, acc_len=1.70, thru=13.5, I=13.8, T=31.5, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:12,728] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:17,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:07:19,114] loop.py:45 INFO TRAIN_STEP step=16646 loss=1.304983 Training: 92%|█████████▏| 16646/18018 [35:37:05<2:52:35, 7.55s/step, loss=1.305, acc=0.699, acc_len=1.68, thru=18.6, I=21.5, T=32.2, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:07:19,110] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16645: forward=3011.7ms backward=911.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:19,527] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:25,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:07:25,543] loop.py:45 INFO TRAIN_STEP step=16647 loss=1.318788 Training: 92%|█████████▏| 16647/18018 [35:37:11<2:44:47, 7.21s/step, loss=1.319, acc=0.696, acc_len=1.67, thru=18.6, I=24.2, T=38.7, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:25,591] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:30,135] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:07:30,516] loop.py:45 INFO TRAIN_STEP step=16648 loss=1.137106 Training: 92%|█████████▏| 16648/18018 [35:37:16<2:29:20, 6.54s/step, loss=1.137, acc=0.728, acc_len=1.78, thru=23.5, I=25.8, T=40.5, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:30,650] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:36,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:07:37,634] loop.py:45 INFO TRAIN_STEP step=16649 loss=1.315480 Training: 92%|█████████▏| 16649/18018 [35:37:23<2:33:10, 6.71s/step, loss=1.315, acc=0.693, acc_len=1.65, thru=18.8, I=22.8, T=35.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:38,063] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:43,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:07:43,961] loop.py:380 INFO TIMING step=16650: step=3.498s data=1.506s compute=3.469s [fwd=2.521s bwd=0.933s opt=0.003s] dispatch=2.822s [2026-10-20 01:07:43,962] loop.py:45 INFO TRAIN_STEP step=16650 loss=1.234425 Training: 92%|█████████▏| 16650/18018 [35:37:30<2:30:25, 6.60s/step, loss=1.234, acc=0.711, acc_len=1.72, thru=19.7, I=24.0, T=36.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:44,347] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-20 01:07:49,398] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16650: forward=2370.5ms backward=923.8ms [2026-10-20 01:07:49,403] loop.py:45 INFO TRAIN_STEP step=16651 loss=1.256125 Training: 92%|█████████▏| 16651/18018 [35:37:35<2:22:24, 6.25s/step, loss=1.256, acc=0.707, acc_len=1.70, thru=22.0, I=24.1, T=38.3, wait=2.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:55,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:07:56,224] loop.py:45 INFO TRAIN_STEP step=16652 loss=1.257853 Training: 92%|█████████▏| 16652/18018 [35:37:42<2:26:12, 6.42s/step, loss=1.258, acc=0.707, acc_len=1.71, thru=19.6, I=20.7, T=37.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:07:56,256] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:00,899] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:08:01,755] loop.py:45 INFO TRAIN_STEP step=16653 loss=1.371635 Training: 92%|█████████▏| 16653/18018 [35:37:47<2:20:01, 6.15s/step, loss=1.372, acc=0.686, acc_len=1.64, thru=22.4, I=25.3, T=38.7, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:01,911] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:09,509] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:08:09,617] loop.py:45 INFO TRAIN_STEP step=16654 loss=1.297438 Training: 92%|█████████▏| 16654/18018 [35:37:55<2:31:34, 6.67s/step, loss=1.297, acc=0.699, acc_len=1.68, thru=16.6, I=18.1, T=35.8, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:10,010] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:15,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:08:15,699] loop.py:380 INFO TIMING step=16655: step=3.391s data=1.457s compute=3.348s [fwd=2.316s bwd=1.017s opt=0.003s] dispatch=2.681s [2026-10-20 01:08:15,699] loop.py:45 INFO TRAIN_STEP step=16655 loss=1.305313 Training: 92%|█████████▏| 16655/18018 [35:38:01<2:27:28, 6.49s/step, loss=1.305, acc=0.697, acc_len=1.67, thru=20.1, I=21.5, T=37.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:15,997] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:08:23,556] loop.py:45 INFO TRAIN_STEP step=16656 loss=1.269440 Training: 92%|█████████▏| 16656/18018 [35:38:09<2:36:38, 6.90s/step, loss=1.269, acc=0.705, acc_len=1.69, thru=16.4, I=16.8, T=37.3, wait=4.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:08:23,552] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16655: forward=2444.1ms backward=940.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:28,037] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:08:29,806] loop.py:45 INFO TRAIN_STEP step=16657 loss=1.308116 Training: 92%|█████████▏| 16657/18018 [35:38:15<2:32:06, 6.71s/step, loss=1.308, acc=0.698, acc_len=1.67, thru=20.4, I=24.7, T=34.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:30,046] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:35,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:08:36,494] loop.py:45 INFO TRAIN_STEP step=16658 loss=1.301998 Training: 92%|█████████▏| 16658/18018 [35:38:22<2:31:52, 6.70s/step, loss=1.302, acc=0.701, acc_len=1.68, thru=18.2, I=25.5, T=40.2, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:36,929] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:43,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:08:44,168] loop.py:45 INFO TRAIN_STEP step=16659 loss=1.162569 Training: 92%|█████████▏| 16659/18018 [35:38:30<2:38:23, 6.99s/step, loss=1.163, acc=0.725, acc_len=1.77, thru=17.4, I=19.1, T=34.6, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:44,465] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:50,128] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:08:50,346] loop.py:380 INFO TIMING step=16660: step=3.328s data=1.342s compute=3.303s [fwd=2.373s bwd=0.916s opt=0.003s] dispatch=2.842s [2026-10-20 01:08:50,346] loop.py:45 INFO TRAIN_STEP step=16660 loss=1.443140 Training: 92%|█████████▏| 16660/18018 [35:38:36<2:32:43, 6.75s/step, loss=1.443, acc=0.676, acc_len=1.60, thru=20.0, I=20.3, T=38.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:50,630] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:56,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:08:56,901] loop.py:45 INFO TRAIN_STEP step=16661 loss=1.286636 Training: 92%|█████████▏| 16661/18018 [35:38:42<2:31:18, 6.69s/step, loss=1.287, acc=0.701, acc_len=1.69, thru=20.2, I=20.3, T=38.7, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:08:56,897] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16660: forward=2373.8ms backward=893.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:08:57,325] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:03,240] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:09:04,279] loop.py:45 INFO TRAIN_STEP step=16662 loss=1.366714 Training: 92%|█████████▏| 16662/18018 [35:38:50<2:35:51, 6.90s/step, loss=1.367, acc=0.688, acc_len=1.63, thru=18.1, I=20.1, T=36.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:04,748] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:09:10,378] loop.py:45 INFO TRAIN_STEP step=16663 loss=1.274030 Training: 92%|█████████▏| 16663/18018 [35:38:56<2:30:20, 6.66s/step, loss=1.274, acc=0.703, acc_len=1.69, thru=18.2, I=21.1, T=39.1, wait=2.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:18,635] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:09:19,320] loop.py:45 INFO TRAIN_STEP step=16664 loss=1.277029 Training: 92%|█████████▏| 16664/18018 [35:39:05<2:45:42, 7.34s/step, loss=1.277, acc=0.702, acc_len=1.69, thru=16.2, I=15.4, T=36.5, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:19,640] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:09:26,601] loop.py:380 INFO TIMING step=16665: step=3.406s data=1.357s compute=3.373s [fwd=2.440s bwd=0.918s opt=0.003s] dispatch=3.866s [2026-10-20 01:09:26,601] loop.py:45 INFO TRAIN_STEP step=16665 loss=1.229669 Training: 92%|█████████▏| 16665/18018 [35:39:12<2:45:09, 7.32s/step, loss=1.230, acc=0.711, acc_len=1.72, thru=16.4, I=17.1, T=37.6, wait=3.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:34,115] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:09:34,488] loop.py:45 INFO TRAIN_STEP step=16666 loss=1.186703 Training: 92%|█████████▏| 16666/18018 [35:39:20<2:48:50, 7.49s/step, loss=1.187, acc=0.718, acc_len=1.75, thru=18.2, I=17.6, T=36.6, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:09:34,483] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16665: forward=2407.3ms backward=1017.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:34,627] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:41,782] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-20 01:09:42,516] loop.py:45 INFO TRAIN_STEP step=16667 loss=1.250148 Training: 93%|█████████▎| 16667/18018 [35:39:28<2:52:19, 7.65s/step, loss=1.250, acc=0.707, acc_len=1.70, thru=17.2, I=18.2, T=34.1, wait=4.3s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:42,786] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=5/64 [2026-10-20 01:09:48,484] loop.py:45 INFO TRAIN_STEP step=16668 loss=1.189485 Training: 93%|█████████▎| 16668/18018 [35:39:34<2:40:49, 7.15s/step, loss=1.189, acc=0.718, acc_len=1.74, thru=16.9, I=21.9, T=38.6, wait=2.6s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:56,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:09:57,198] loop.py:45 INFO TRAIN_STEP step=16669 loss=1.317175 Training: 93%|█████████▎| 16669/18018 [35:39:43<2:51:16, 7.62s/step, loss=1.317, acc=0.697, acc_len=1.67, thru=17.3, I=16.8, T=36.5, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:09:57,580] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:10:04,021] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:10:04,596] loop.py:380 INFO TIMING step=16670: step=3.467s data=1.461s compute=3.421s [fwd=2.440s bwd=0.967s opt=0.003s] dispatch=3.922s [2026-10-20 01:10:04,596] loop.py:45 INFO TRAIN_STEP step=16670 loss=1.195547 Training: 93%|█████████▎| 16670/18018 [35:39:50<2:49:39, 7.55s/step, loss=1.196, acc=0.717, acc_len=1.74, thru=17.2, I=19.3, T=36.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:10:05,028] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:10:09,549] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:10:10,354] loop.py:45 INFO TRAIN_STEP step=16671 loss=1.190453 Training: 93%|█████████▎| 16671/18018 [35:39:56<2:37:27, 7.01s/step, loss=1.190, acc=0.718, acc_len=1.74, thru=20.9, I=24.6, T=39.2, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:10:10,349] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16670: forward=2307.2ms backward=915.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:10:10,558] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:10:18,454] loop.py:45 INFO TRAIN_STEP step=16672 loss=1.291422 Training: 93%|█████████▎| 16672/18018 [35:40:04<2:44:39, 7.34s/step, loss=1.291, acc=0.698, acc_len=1.67, thru=15.4, I=19.5, T=35.9, wait=4.5s, pool=57, epoch=2/2][2026-10-20 01:10:26,863] loop.py:45 INFO TRAIN_STEP step=16673 loss=1.252054 Training: 93%|█████████▎| 16673/18018 [35:40:12<2:51:43, 7.66s/step, loss=1.252, acc=0.707, acc_len=1.71, thru=12.4, I=13.2, T=35.1, wait=4.8s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:10:37,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:10:38,533] loop.py:45 INFO TRAIN_STEP step=16674 loss=1.317205 Training: 93%|█████████▎| 16674/18018 [35:40:24<3:18:32, 8.86s/step, loss=1.317, acc=0.697, acc_len=1.67, thru=15.0, I=12.8, T=33.9, wait=7.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:10:38,700] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:10:44,981] loop.py:380 INFO TIMING step=16675: step=3.282s data=1.351s compute=3.252s [fwd=2.334s bwd=0.904s opt=0.003s] dispatch=3.156s [2026-10-20 01:10:44,982] loop.py:45 INFO TRAIN_STEP step=16675 loss=1.157917 Training: 93%|█████████▎| 16675/18018 [35:40:31<3:02:10, 8.14s/step, loss=1.158, acc=0.725, acc_len=1.77, thru=15.8, I=20.9, T=39.0, wait=3.2s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:10:55,576] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:10:56,853] loop.py:45 INFO TRAIN_STEP step=16676 loss=1.319693 Training: 93%|█████████▎| 16676/18018 [35:40:42<3:27:04, 9.26s/step, loss=1.320, acc=0.698, acc_len=1.67, thru=14.6, I=13.7, T=31.2, wait=7.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:10:56,848] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16675: forward=2867.6ms backward=1176.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:10:57,085] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:11:03,958] loop.py:45 INFO TRAIN_STEP step=16677 loss=1.289221 Training: 93%|█████████▎| 16677/18018 [35:40:50<3:12:30, 8.61s/step, loss=1.289, acc=0.699, acc_len=1.67, thru=12.1, I=24.1, T=37.7, wait=3.7s, pool=13, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 01:11:04,542] eagle_store.py:517 WARNING batch_get_buffer missing keys (k533658-a0ceb376_hs, k533658-a0ceb376_ids, k533658-a0ceb376_lhs); sleeping 0.50s. [2026-10-20 01:11:14,155] loop.py:45 INFO TRAIN_STEP step=16678 loss=1.266727 Training: 93%|█████████▎| 16678/18018 [35:41:00<3:22:57, 9.09s/step, loss=1.267, acc=0.700, acc_len=1.68, thru=15.3, I=10.7, T=36.0, wait=6.6s, pool=33, epoch=2/2][2026-10-20 01:11:21,615] loop.py:45 INFO TRAIN_STEP step=16679 loss=1.312682 Training: 93%|█████████▎| 16679/18018 [35:41:07<3:11:54, 8.60s/step, loss=1.313, acc=0.695, acc_len=1.67, thru=18.6, I=15.4, T=37.0, wait=4.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:11:27,191] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:11:27,841] loop.py:380 INFO TIMING step=16680: step=3.412s data=1.344s compute=3.383s [fwd=2.477s bwd=0.892s opt=0.003s] dispatch=2.806s [2026-10-20 01:11:27,841] loop.py:45 INFO TRAIN_STEP step=16680 loss=1.313167 Training: 93%|█████████▎| 16680/18018 [35:41:13<2:55:53, 7.89s/step, loss=1.313, acc=0.697, acc_len=1.67, thru=20.0, I=20.8, T=37.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:11:28,198] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:11:33,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:11:35,139] loop.py:45 INFO TRAIN_STEP step=16681 loss=1.324771 Training: 93%|█████████▎| 16681/18018 [35:41:21<2:51:48, 7.71s/step, loss=1.325, acc=0.695, acc_len=1.67, thru=17.8, I=21.8, T=37.3, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:11:35,135] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16680: forward=2420.4ms backward=952.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:11:35,167] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-20 01:11:41,989] loop.py:45 INFO TRAIN_STEP step=16682 loss=1.223027 Training: 93%|█████████▎| 16682/18018 [35:41:28<2:45:56, 7.45s/step, loss=1.223, acc=0.714, acc_len=1.73, thru=18.2, I=21.3, T=38.8, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:11:49,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:11:50,338] loop.py:45 INFO TRAIN_STEP step=16683 loss=1.336847 Training: 93%|█████████▎| 16683/18018 [35:41:36<2:51:48, 7.72s/step, loss=1.337, acc=0.698, acc_len=1.67, thru=17.8, I=18.1, T=36.7, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:11:50,829] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:11:57,430] loop.py:45 INFO TRAIN_STEP step=16684 loss=1.308781 Training: 93%|█████████▎| 16684/18018 [35:41:43<2:47:27, 7.53s/step, loss=1.309, acc=0.698, acc_len=1.67, thru=15.2, I=17.1, T=37.8, wait=3.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:12:03,093] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:12:04,227] loop.py:380 INFO TIMING step=16685: step=3.338s data=1.314s compute=3.308s [fwd=2.373s bwd=0.921s opt=0.003s] dispatch=3.451s [2026-10-20 01:12:04,227] loop.py:45 INFO TRAIN_STEP step=16685 loss=1.223740 Training: 93%|█████████▎| 16685/18018 [35:41:50<2:42:26, 7.31s/step, loss=1.224, acc=0.712, acc_len=1.72, thru=19.6, I=21.2, T=38.3, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:12:04,601] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:12:10,934] loop.py:45 INFO TRAIN_STEP step=16686 loss=1.210567 Training: 93%|█████████▎| 16686/18018 [35:41:57<2:38:17, 7.13s/step, loss=1.211, acc=0.715, acc_len=1.73, thru=18.1, I=22.0, T=37.3, wait=3.3s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:12:10,930] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16685: forward=2423.7ms backward=966.7ms [2026-10-20 01:12:18,020] loop.py:45 INFO TRAIN_STEP step=16687 loss=1.267128 Training: 93%|█████████▎| 16687/18018 [35:42:04<2:37:53, 7.12s/step, loss=1.267, acc=0.704, acc_len=1.69, thru=18.5, I=18.9, T=36.7, wait=3.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:12:26,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:12:27,464] loop.py:45 INFO TRAIN_STEP step=16688 loss=1.157776 Training: 93%|█████████▎| 16688/18018 [35:42:13<2:53:14, 7.82s/step, loss=1.158, acc=0.724, acc_len=1.77, thru=16.4, I=14.3, T=36.4, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:12:27,927] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:12:34,805] loop.py:45 INFO TRAIN_STEP step=16689 loss=1.132243 Training: 93%|█████████▎| 16689/18018 [35:42:20<2:49:57, 7.67s/step, loss=1.132, acc=0.729, acc_len=1.78, thru=15.0, I=18.1, T=36.9, wait=3.9s, pool=37, epoch=2/2][2026-10-20 01:12:45,026] loop.py:380 INFO TIMING step=16690: step=3.804s data=1.694s compute=3.750s [fwd=2.648s bwd=1.088s opt=0.003s] dispatch=6.406s [2026-10-20 01:12:45,026] loop.py:45 INFO TRAIN_STEP step=16690 loss=1.300920 Training: 93%|█████████▎| 16690/18018 [35:42:31<3:06:44, 8.44s/step, loss=1.301, acc=0.702, acc_len=1.68, thru=14.6, I=12.0, T=33.6, wait=6.4s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:12:53,264] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16690: forward=2539.1ms backward=989.4ms [2026-10-20 01:12:53,269] loop.py:45 INFO TRAIN_STEP step=16691 loss=1.279044 Training: 93%|█████████▎| 16691/18018 [35:42:39<3:05:19, 8.38s/step, loss=1.279, acc=0.699, acc_len=1.68, thru=15.8, I=14.5, T=35.9, wait=4.7s, pool=45, epoch=2/2][2026-10-20 01:13:01,174] loop.py:45 INFO TRAIN_STEP step=16692 loss=1.395415 Training: 93%|█████████▎| 16692/18018 [35:42:47<3:02:01, 8.24s/step, loss=1.395, acc=0.685, acc_len=1.62, thru=14.9, I=14.9, T=36.5, wait=4.4s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:07,524] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:13:08,284] loop.py:45 INFO TRAIN_STEP step=16693 loss=1.216835 Training: 93%|█████████▎| 16693/18018 [35:42:54<2:54:25, 7.90s/step, loss=1.217, acc=0.713, acc_len=1.72, thru=18.6, I=19.7, T=36.1, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:08,528] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:13,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:13:14,787] loop.py:45 INFO TRAIN_STEP step=16694 loss=1.229991 Training: 93%|█████████▎| 16694/18018 [35:43:00<2:45:03, 7.48s/step, loss=1.230, acc=0.712, acc_len=1.73, thru=19.5, I=25.8, T=36.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:14,884] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:22,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:13:22,978] loop.py:380 INFO TIMING step=16695: step=3.455s data=1.421s compute=3.425s [fwd=2.460s bwd=0.951s opt=0.003s] dispatch=4.727s [2026-10-20 01:13:22,978] loop.py:45 INFO TRAIN_STEP step=16695 loss=1.280354 Training: 93%|█████████▎| 16695/18018 [35:43:09<2:49:38, 7.69s/step, loss=1.280, acc=0.705, acc_len=1.70, thru=14.1, I=18.2, T=37.1, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:23,369] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:28,243] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:13:29,022] loop.py:45 INFO TRAIN_STEP step=16696 loss=1.170599 Training: 93%|█████████▎| 16696/18018 [35:43:15<2:38:36, 7.20s/step, loss=1.171, acc=0.723, acc_len=1.76, thru=20.6, I=19.6, T=38.2, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:13:29,018] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16695: forward=2367.3ms backward=948.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:29,247] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:35,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:13:35,629] loop.py:45 INFO TRAIN_STEP step=16697 loss=1.196749 Training: 93%|█████████▎| 16697/18018 [35:43:21<2:34:34, 7.02s/step, loss=1.197, acc=0.719, acc_len=1.74, thru=19.6, I=20.9, T=38.1, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:13:36,111] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:13:41,526] loop.py:45 INFO TRAIN_STEP step=16698 loss=1.253262 Training: 93%|█████████▎| 16698/18018 [35:43:27<2:27:02, 6.68s/step, loss=1.253, acc=0.706, acc_len=1.70, thru=20.5, I=21.4, T=37.5, wait=2.5s, pool=57, epoch=2/2][2026-10-20 01:13:48,502] loop.py:45 INFO TRAIN_STEP step=16699 loss=1.310992 Training: 93%|█████████▎| 16699/18018 [35:43:34<2:28:51, 6.77s/step, loss=1.311, acc=0.698, acc_len=1.68, thru=17.4, I=19.0, T=37.3, wait=3.5s, pool=49, epoch=2/2][2026-10-20 01:13:56,175] loop.py:380 INFO TIMING step=16700: step=3.578s data=1.532s compute=3.541s [fwd=2.531s bwd=0.995s opt=0.003s] dispatch=4.088s [2026-10-20 01:13:56,175] loop.py:45 INFO TRAIN_STEP step=16700 loss=1.265871 Training: 93%|█████████▎| 16700/18018 [35:43:42<2:34:41, 7.04s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=15.0, I=17.3, T=35.8, wait=4.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:02,494] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:14:03,085] loop.py:45 INFO TRAIN_STEP step=16701 loss=1.213528 Training: 93%|█████████▎| 16701/18018 [35:43:49<2:33:41, 7.00s/step, loss=1.214, acc=0.711, acc_len=1.72, thru=17.8, I=19.3, T=36.2, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:14:03,080] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16700: forward=2484.6ms backward=999.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:03,497] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:09,335] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:14:10,238] loop.py:45 INFO TRAIN_STEP step=16702 loss=1.277523 Training: 93%|█████████▎| 16702/18018 [35:43:56<2:34:35, 7.05s/step, loss=1.278, acc=0.700, acc_len=1.68, thru=16.9, I=20.3, T=37.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:10,344] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:16,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:14:18,330] loop.py:45 INFO TRAIN_STEP step=16703 loss=1.346746 Training: 93%|█████████▎| 16703/18018 [35:44:04<2:41:18, 7.36s/step, loss=1.347, acc=0.690, acc_len=1.64, thru=16.0, I=19.5, T=36.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:18,749] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 01:14:24,273] loop.py:45 INFO TRAIN_STEP step=16704 loss=1.215604 Training: 93%|█████████▎| 16704/18018 [35:44:10<2:31:53, 6.94s/step, loss=1.216, acc=0.714, acc_len=1.73, thru=19.7, I=23.2, T=39.0, wait=2.7s, pool=53, epoch=2/2][2026-10-20 01:14:31,257] loop.py:380 INFO TIMING step=16705: step=3.522s data=1.478s compute=3.491s [fwd=2.502s bwd=0.975s opt=0.003s] dispatch=3.452s [2026-10-20 01:14:31,257] loop.py:45 INFO TRAIN_STEP step=16705 loss=1.333312 Training: 93%|█████████▎| 16705/18018 [35:44:17<2:32:04, 6.95s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=17.2, I=19.2, T=36.3, wait=3.5s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:39,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:14:40,628] loop.py:45 INFO TRAIN_STEP step=16706 loss=1.318167 Training: 93%|█████████▎| 16706/18018 [35:44:26<2:47:51, 7.68s/step, loss=1.318, acc=0.698, acc_len=1.67, thru=15.7, I=15.8, T=35.0, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:14:40,623] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16705: forward=2565.2ms backward=1038.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:40,956] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:45,447] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:14:45,633] loop.py:45 INFO TRAIN_STEP step=16707 loss=1.182236 Training: 93%|█████████▎| 16707/18018 [35:44:31<2:30:13, 6.87s/step, loss=1.182, acc=0.719, acc_len=1.75, thru=21.9, I=26.5, T=39.8, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:14:45,950] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:14:51,292] loop.py:45 INFO TRAIN_STEP step=16708 loss=1.243596 Training: 93%|█████████▎| 16708/18018 [35:44:37<2:22:08, 6.51s/step, loss=1.244, acc=0.711, acc_len=1.72, thru=20.9, I=25.5, T=39.3, wait=2.4s, pool=33, epoch=2/2][2026-10-20 01:15:01,447] loop.py:45 INFO TRAIN_STEP step=16709 loss=1.219813 Training: 93%|█████████▎| 16709/18018 [35:44:47<2:45:53, 7.60s/step, loss=1.220, acc=0.713, acc_len=1.73, thru=13.3, I=13.7, T=35.0, wait=6.5s, pool=33, epoch=2/2][2026-10-20 01:15:11,166] loop.py:380 INFO TIMING step=16710: step=3.715s data=1.514s compute=3.665s [fwd=2.570s bwd=1.080s opt=0.003s] dispatch=5.994s [2026-10-20 01:15:11,167] loop.py:45 INFO TRAIN_STEP step=16710 loss=1.259529 Training: 93%|█████████▎| 16710/18018 [35:44:57<2:59:35, 8.24s/step, loss=1.260, acc=0.709, acc_len=1.71, thru=16.5, I=13.9, T=34.5, wait=6.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:15:21,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:15:21,623] loop.py:45 INFO TRAIN_STEP step=16711 loss=1.337165 Training: 93%|█████████▎| 16711/18018 [35:45:07<3:13:56, 8.90s/step, loss=1.337, acc=0.690, acc_len=1.64, thru=13.3, I=13.1, T=33.7, wait=6.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:15:21,618] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16710: forward=2633.0ms backward=1108.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:15:22,035] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:15:27,547] loop.py:45 INFO TRAIN_STEP step=16712 loss=1.164135 Training: 93%|█████████▎| 16712/18018 [35:45:13<2:54:21, 8.01s/step, loss=1.164, acc=0.726, acc_len=1.78, thru=18.5, I=20.3, T=36.1, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:15:33,181] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:15:34,221] loop.py:45 INFO TRAIN_STEP step=16713 loss=1.217964 Training: 93%|█████████▎| 16713/18018 [35:45:20<2:45:29, 7.61s/step, loss=1.218, acc=0.714, acc_len=1.73, thru=19.2, I=21.0, T=38.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:15:34,692] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:15:40,758] loop.py:45 INFO TRAIN_STEP step=16714 loss=1.328957 Training: 93%|█████████▎| 16714/18018 [35:45:26<2:38:22, 7.29s/step, loss=1.329, acc=0.693, acc_len=1.66, thru=17.2, I=21.0, T=39.0, wait=3.2s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:15:49,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:15:50,126] loop.py:380 INFO TIMING step=16715: step=3.591s data=1.386s compute=3.553s [fwd=2.576s bwd=0.961s opt=0.003s] dispatch=5.768s [2026-10-20 01:15:50,126] loop.py:45 INFO TRAIN_STEP step=16715 loss=1.204352 Training: 93%|█████████▎| 16715/18018 [35:45:36<2:51:48, 7.91s/step, loss=1.204, acc=0.719, acc_len=1.75, thru=17.4, I=16.5, T=35.6, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:15:50,178] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:15:57,006] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:15:57,413] loop.py:45 INFO TRAIN_STEP step=16716 loss=1.266156 Training: 93%|█████████▎| 16716/18018 [35:45:43<2:47:37, 7.72s/step, loss=1.266, acc=0.705, acc_len=1.70, thru=17.5, I=19.8, T=36.8, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:15:57,407] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16715: forward=2488.8ms backward=941.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:15:57,508] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:16:02,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:16:02,976] loop.py:45 INFO TRAIN_STEP step=16717 loss=1.229429 Training: 93%|█████████▎| 16717/18018 [35:45:49<2:33:24, 7.08s/step, loss=1.229, acc=0.712, acc_len=1.72, thru=21.2, I=21.1, T=37.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:16:03,398] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:16:09,139] loop.py:45 INFO TRAIN_STEP step=16718 loss=1.214084 Training: 93%|█████████▎| 16718/18018 [35:45:55<2:27:22, 6.80s/step, loss=1.214, acc=0.715, acc_len=1.73, thru=19.1, I=21.1, T=37.4, wait=2.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:16:18,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:16:19,570] loop.py:45 INFO TRAIN_STEP step=16719 loss=1.266826 Training: 93%|█████████▎| 16719/18018 [35:46:05<2:50:50, 7.89s/step, loss=1.267, acc=0.706, acc_len=1.70, thru=15.4, I=15.0, T=33.5, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:16:19,604] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:16:26,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:16:27,100] loop.py:380 INFO TIMING step=16720: step=3.613s data=1.424s compute=3.586s [fwd=2.579s bwd=0.992s opt=0.003s] dispatch=3.906s [2026-10-20 01:16:27,100] loop.py:45 INFO TRAIN_STEP step=16720 loss=1.321902 Training: 93%|█████████▎| 16720/18018 [35:46:13<2:48:22, 7.78s/step, loss=1.322, acc=0.695, acc_len=1.67, thru=16.9, I=18.9, T=35.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:16:27,341] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-20 01:16:34,812] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16720: forward=2521.4ms backward=985.1ms [2026-10-20 01:16:34,817] loop.py:45 INFO TRAIN_STEP step=16721 loss=1.320647 Training: 93%|█████████▎| 16721/18018 [35:46:20<2:47:47, 7.76s/step, loss=1.321, acc=0.693, acc_len=1.66, thru=16.1, I=17.6, T=36.0, wait=4.2s, pool=57, epoch=2/2][2026-10-20 01:16:43,515] loop.py:45 INFO TRAIN_STEP step=16722 loss=1.301946 Training: 93%|█████████▎| 16722/18018 [35:46:29<2:53:45, 8.04s/step, loss=1.302, acc=0.699, acc_len=1.68, thru=14.8, I=16.3, T=36.8, wait=5.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:16:52,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:16:53,429] loop.py:45 INFO TRAIN_STEP step=16723 loss=1.278708 Training: 93%|█████████▎| 16723/18018 [35:46:39<3:05:42, 8.60s/step, loss=1.279, acc=0.703, acc_len=1.69, thru=14.6, I=13.5, T=33.3, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:16:53,926] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:16:59,255] loop.py:45 INFO TRAIN_STEP step=16724 loss=1.208693 Training: 93%|█████████▎| 16724/18018 [35:46:45<2:47:35, 7.77s/step, loss=1.209, acc=0.715, acc_len=1.73, thru=18.9, I=20.9, T=39.3, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:05,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:17:05,855] loop.py:380 INFO TIMING step=16725: step=3.412s data=1.338s compute=3.385s [fwd=2.436s bwd=0.934s opt=0.003s] dispatch=3.177s [2026-10-20 01:17:05,855] loop.py:45 INFO TRAIN_STEP step=16725 loss=1.330988 Training: 93%|█████████▎| 16725/18018 [35:46:51<2:39:52, 7.42s/step, loss=1.331, acc=0.693, acc_len=1.65, thru=19.8, I=20.8, T=37.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:05,883] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:11,909] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:17:12,137] loop.py:45 INFO TRAIN_STEP step=16726 loss=1.184957 Training: 93%|█████████▎| 16726/18018 [35:46:58<2:32:25, 7.08s/step, loss=1.185, acc=0.721, acc_len=1.75, thru=20.4, I=20.9, T=32.3, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:17:12,133] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16725: forward=2991.0ms backward=916.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:12,411] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:18,767] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:17:19,083] loop.py:45 INFO TRAIN_STEP step=16727 loss=1.205047 Training: 93%|█████████▎| 16727/18018 [35:47:05<2:31:25, 7.04s/step, loss=1.205, acc=0.719, acc_len=1.75, thru=18.2, I=18.5, T=36.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:19,269] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:24,534] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:17:25,491] loop.py:45 INFO TRAIN_STEP step=16728 loss=1.166196 Training: 93%|█████████▎| 16728/18018 [35:47:11<2:27:15, 6.85s/step, loss=1.166, acc=0.721, acc_len=1.75, thru=19.8, I=21.9, T=36.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:25,541] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-20 01:17:32,080] loop.py:45 INFO TRAIN_STEP step=16729 loss=1.273277 Training: 93%|█████████▎| 16729/18018 [35:47:18<2:25:27, 6.77s/step, loss=1.273, acc=0.705, acc_len=1.70, thru=18.5, I=20.2, T=37.9, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:37,785] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:17:38,573] loop.py:380 INFO TIMING step=16730: step=3.544s data=1.418s compute=3.512s [fwd=2.529s bwd=0.969s opt=0.003s] dispatch=2.941s [2026-10-20 01:17:38,573] loop.py:45 INFO TRAIN_STEP step=16730 loss=1.309361 Training: 93%|█████████▎| 16730/18018 [35:47:24<2:23:33, 6.69s/step, loss=1.309, acc=0.699, acc_len=1.68, thru=19.8, I=20.6, T=36.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:38,788] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:46,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:17:46,428] loop.py:45 INFO TRAIN_STEP step=16731 loss=1.229207 Training: 93%|█████████▎| 16731/18018 [35:47:32<2:30:57, 7.04s/step, loss=1.229, acc=0.708, acc_len=1.72, thru=16.0, I=17.3, T=37.0, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:17:46,424] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16730: forward=2434.8ms backward=962.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:46,844] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:52,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:17:53,169] loop.py:45 INFO TRAIN_STEP step=16732 loss=1.388579 Training: 93%|█████████▎| 16732/18018 [35:47:39<2:28:56, 6.95s/step, loss=1.389, acc=0.681, acc_len=1.62, thru=19.0, I=20.5, T=36.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:53,221] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:57,753] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:17:58,443] loop.py:45 INFO TRAIN_STEP step=16733 loss=1.120204 Training: 93%|█████████▎| 16733/18018 [35:47:44<2:18:03, 6.45s/step, loss=1.120, acc=0.732, acc_len=1.79, thru=22.5, I=28.5, T=39.6, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:17:58,757] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:18:05,679] loop.py:45 INFO TRAIN_STEP step=16734 loss=1.280698 Training: 93%|█████████▎| 16734/18018 [35:47:51<2:23:01, 6.68s/step, loss=1.281, acc=0.704, acc_len=1.69, thru=16.0, I=17.4, T=37.9, wait=3.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:18:11,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:18:12,525] loop.py:380 INFO TIMING step=16735: step=3.704s data=1.662s compute=3.677s [fwd=2.744s bwd=0.920s opt=0.003s] dispatch=3.133s [2026-10-20 01:18:12,526] loop.py:45 INFO TRAIN_STEP step=16735 loss=1.254078 Training: 93%|█████████▎| 16735/18018 [35:47:58<2:23:56, 6.73s/step, loss=1.254, acc=0.708, acc_len=1.71, thru=19.0, I=20.0, T=34.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:18:13,013] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-20 01:18:19,418] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16735: forward=2473.6ms backward=971.1ms [2026-10-20 01:18:19,422] loop.py:45 INFO TRAIN_STEP step=16736 loss=1.312350 Training: 93%|█████████▎| 16736/18018 [35:48:05<2:24:53, 6.78s/step, loss=1.312, acc=0.699, acc_len=1.68, thru=17.8, I=20.2, T=36.6, wait=3.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:18:27,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:18:29,088] loop.py:45 INFO TRAIN_STEP step=16737 loss=1.243782 Training: 93%|█████████▎| 16737/18018 [35:48:15<2:43:15, 7.65s/step, loss=1.244, acc=0.708, acc_len=1.72, thru=14.1, I=14.6, T=34.3, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:18:29,334] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:18:35,050] loop.py:45 INFO TRAIN_STEP step=16738 loss=1.115210 Training: 93%|█████████▎| 16738/18018 [35:48:21<2:32:21, 7.14s/step, loss=1.115, acc=0.731, acc_len=1.79, thru=19.2, I=21.7, T=38.5, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:18:44,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:18:44,662] loop.py:45 INFO TRAIN_STEP step=16739 loss=1.223938 Training: 93%|█████████▎| 16739/18018 [35:48:30<2:48:01, 7.88s/step, loss=1.224, acc=0.713, acc_len=1.72, thru=15.1, I=15.1, T=36.9, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:18:45,131] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:18:52,228] loop.py:380 INFO TIMING step=16740: step=3.428s data=1.330s compute=3.385s [fwd=2.439s bwd=0.931s opt=0.003s] dispatch=4.131s [2026-10-20 01:18:52,228] loop.py:45 INFO TRAIN_STEP step=16740 loss=1.225084 Training: 93%|█████████▎| 16740/18018 [35:48:38<2:45:53, 7.79s/step, loss=1.225, acc=0.715, acc_len=1.73, thru=14.6, I=15.9, T=37.3, wait=4.1s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:19:01,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:19:01,373] loop.py:45 INFO TRAIN_STEP step=16741 loss=1.353796 Training: 93%|█████████▎| 16741/18018 [35:48:47<2:54:24, 8.19s/step, loss=1.354, acc=0.692, acc_len=1.65, thru=17.4, I=15.3, T=34.8, wait=5.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:19:01,369] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16740: forward=2579.5ms backward=1041.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:19:01,714] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:19:09,268] loop.py:45 INFO TRAIN_STEP step=16742 loss=1.350498 Training: 93%|█████████▎| 16742/18018 [35:48:55<2:52:21, 8.10s/step, loss=1.350, acc=0.692, acc_len=1.65, thru=15.7, I=17.1, T=35.5, wait=4.3s, pool=53, epoch=2/2][2026-10-20 01:19:18,482] loop.py:45 INFO TRAIN_STEP step=16743 loss=1.264285 Training: 93%|█████████▎| 16743/18018 [35:49:04<2:59:17, 8.44s/step, loss=1.264, acc=0.704, acc_len=1.69, thru=14.2, I=13.2, T=35.8, wait=5.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:19:26,662] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:19:27,268] loop.py:45 INFO TRAIN_STEP step=16744 loss=1.287667 Training: 93%|█████████▎| 16744/18018 [35:49:13<3:01:23, 8.54s/step, loss=1.288, acc=0.704, acc_len=1.70, thru=16.0, I=16.0, T=30.9, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:19:27,670] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:19:36,383] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:19:36,986] loop.py:380 INFO TIMING step=16745: step=3.593s data=1.487s compute=3.563s [fwd=2.528s bwd=1.020s opt=0.003s] dispatch=6.113s [2026-10-20 01:19:36,986] loop.py:45 INFO TRAIN_STEP step=16745 loss=1.313602 Training: 93%|█████████▎| 16745/18018 [35:49:23<3:08:42, 8.89s/step, loss=1.314, acc=0.699, acc_len=1.67, thru=13.7, I=14.4, T=35.6, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:19:37,389] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:19:42,682] loop.py:45 INFO TRAIN_STEP step=16746 loss=1.269784 Training: 93%|█████████▎| 16746/18018 [35:49:28<2:48:13, 7.94s/step, loss=1.270, acc=0.702, acc_len=1.68, thru=17.6, I=18.9, T=37.9, wait=2.3s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:19:42,678] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16745: forward=2399.1ms backward=931.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:19:51,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:19:52,275] loop.py:45 INFO TRAIN_STEP step=16747 loss=1.182028 Training: 93%|█████████▎| 16747/18018 [35:49:38<2:58:37, 8.43s/step, loss=1.182, acc=0.721, acc_len=1.75, thru=16.0, I=14.8, T=36.2, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:19:52,722] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:20:00,812] loop.py:45 INFO TRAIN_STEP step=16748 loss=1.185915 Training: 93%|█████████▎| 16748/18018 [35:49:46<2:59:09, 8.46s/step, loss=1.186, acc=0.722, acc_len=1.76, thru=13.2, I=17.3, T=34.5, wait=4.8s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:07,854] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:20:08,186] loop.py:45 INFO TRAIN_STEP step=16749 loss=1.196659 Training: 93%|█████████▎| 16749/18018 [35:49:54<2:52:05, 8.14s/step, loss=1.197, acc=0.721, acc_len=1.76, thru=17.1, I=18.4, T=37.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:08,357] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:13,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:20:15,224] loop.py:380 INFO TIMING step=16750: step=4.228s data=2.209s compute=4.195s [fwd=3.191s bwd=0.991s opt=0.003s] dispatch=2.801s [2026-10-20 01:20:15,224] loop.py:45 INFO TRAIN_STEP step=16750 loss=1.136027 Training: 93%|█████████▎| 16750/18018 [35:50:01<2:44:59, 7.81s/step, loss=1.136, acc=0.730, acc_len=1.78, thru=18.0, I=22.0, T=30.3, wait=2.8s, pool=65, epoch=2/2][2026-10-20 01:20:15,227] loop.py:419 INFO Saving checkpoint at step 16750... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:18,486] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 01:20:18,647] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0016751 (TrainerActor pid=9310) [2026-10-20 01:20:18,806] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0016001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:19,995] inference_manager.py:408 INFO Pool capacity available after 6.6s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:25,647] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 01:20:18,649] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0016751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 01:20:19,815] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0016001 [2026-10-20 01:20:26,495] loop.py:45 INFO TRAIN_STEP step=16751 loss=1.288067 Training: 93%|█████████▎| 16751/18018 [35:50:12<3:06:48, 8.85s/step, loss=1.288, acc=0.700, acc_len=1.68, thru=19.7, I=22.3, T=38.5, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:20:26,491] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16750: forward=2360.4ms backward=917.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:26,659] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:31,775] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:20:32,708] loop.py:45 INFO TRAIN_STEP step=16752 loss=1.250700 Training: 93%|█████████▎| 16752/18018 [35:50:18<2:49:59, 8.06s/step, loss=1.251, acc=0.706, acc_len=1.71, thru=20.4, I=22.9, T=33.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:32,789] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-20 01:20:39,089] loop.py:45 INFO TRAIN_STEP step=16753 loss=1.336004 Training: 93%|█████████▎| 16753/18018 [35:50:25<2:39:16, 7.55s/step, loss=1.336, acc=0.696, acc_len=1.66, thru=18.1, I=19.9, T=35.7, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:47,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:20:47,413] loop.py:45 INFO TRAIN_STEP step=16754 loss=1.187597 Training: 93%|█████████▎| 16754/18018 [35:50:33<2:43:59, 7.78s/step, loss=1.188, acc=0.720, acc_len=1.75, thru=16.2, I=14.7, T=37.4, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:20:47,739] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:20:57,014] loop.py:380 INFO TIMING step=16755: step=4.291s data=1.494s compute=4.255s [fwd=3.222s bwd=1.019s opt=0.003s] dispatch=5.300s [2026-10-20 01:20:57,014] loop.py:45 INFO TRAIN_STEP step=16755 loss=1.152046 Training: 93%|█████████▎| 16755/18018 [35:50:43<2:55:19, 8.33s/step, loss=1.152, acc=0.729, acc_len=1.78, thru=13.8, I=14.3, T=29.8, wait=5.3s, pool=61, epoch=2/2][2026-10-20 01:21:03,929] loop.py:45 INFO TRAIN_STEP step=16756 loss=1.291279 Training: 93%|█████████▎| 16756/18018 [35:50:50<2:46:16, 7.91s/step, loss=1.291, acc=0.704, acc_len=1.69, thru=15.7, I=16.6, T=38.5, wait=3.6s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:21:03,925] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16755: forward=2367.4ms backward=914.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:21:12,550] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:21:13,914] loop.py:45 INFO TRAIN_STEP step=16757 loss=1.215986 Training: 93%|█████████▎| 16757/18018 [35:51:00<2:59:15, 8.53s/step, loss=1.216, acc=0.716, acc_len=1.73, thru=15.3, I=16.1, T=34.1, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:21:14,065] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:21:19,260] loop.py:45 INFO TRAIN_STEP step=16758 loss=1.222372 Training: 93%|█████████▎| 16758/18018 [35:51:05<2:39:02, 7.57s/step, loss=1.222, acc=0.716, acc_len=1.73, thru=19.3, I=21.3, T=39.5, wait=2.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:21:25,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:21:25,566] loop.py:45 INFO TRAIN_STEP step=16759 loss=1.162688 Training: 93%|█████████▎| 16759/18018 [35:51:11<2:30:56, 7.19s/step, loss=1.163, acc=0.721, acc_len=1.76, thru=21.5, I=22.4, T=37.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:21:25,661] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-20 01:21:31,503] loop.py:380 INFO TIMING step=16760: step=3.505s data=1.451s compute=3.475s [fwd=2.493s bwd=0.968s opt=0.003s] dispatch=2.424s [2026-10-20 01:21:31,503] loop.py:45 INFO TRAIN_STEP step=16760 loss=1.263016 Training: 93%|█████████▎| 16760/18018 [35:51:17<2:22:56, 6.82s/step, loss=1.263, acc=0.707, acc_len=1.71, thru=19.2, I=22.2, T=36.5, wait=2.4s, pool=41, epoch=2/2][2026-10-20 01:21:38,827] loop.py:45 INFO TRAIN_STEP step=16761 loss=1.259767 Training: 93%|█████████▎| 16761/18018 [35:51:24<2:25:58, 6.97s/step, loss=1.260, acc=0.707, acc_len=1.71, thru=16.4, I=16.0, T=37.2, wait=3.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:21:38,823] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16760: forward=2447.2ms backward=942.5ms [2026-10-20 01:21:48,749] loop.py:45 INFO TRAIN_STEP step=16762 loss=1.242355 Training: 93%|█████████▎| 16762/18018 [35:51:34<2:44:26, 7.86s/step, loss=1.242, acc=0.708, acc_len=1.71, thru=14.8, I=14.0, T=35.3, wait=6.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:21:53,700] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:21:54,282] loop.py:45 INFO TRAIN_STEP step=16763 loss=1.261556 Training: 93%|█████████▎| 16763/18018 [35:51:40<2:29:43, 7.16s/step, loss=1.262, acc=0.704, acc_len=1.69, thru=20.7, I=22.7, T=38.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:21:54,704] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:22:00,584] loop.py:45 INFO TRAIN_STEP step=16764 loss=1.291903 Training: 93%|█████████▎| 16764/18018 [35:51:46<2:24:14, 6.90s/step, loss=1.292, acc=0.700, acc_len=1.68, thru=18.0, I=21.9, T=37.8, wait=2.9s, pool=37, epoch=2/2][2026-10-20 01:22:09,546] loop.py:380 INFO TIMING step=16765: step=3.627s data=1.450s compute=3.594s [fwd=2.545s bwd=1.034s opt=0.003s] dispatch=5.326s [2026-10-20 01:22:09,547] loop.py:45 INFO TRAIN_STEP step=16765 loss=1.254007 Training: 93%|█████████▎| 16765/18018 [35:51:55<2:37:02, 7.52s/step, loss=1.254, acc=0.709, acc_len=1.71, thru=14.3, I=14.8, T=35.3, wait=5.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:18,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:22:19,522] loop.py:45 INFO TRAIN_STEP step=16766 loss=1.349622 Training: 93%|█████████▎| 16766/18018 [35:52:05<2:52:17, 8.26s/step, loss=1.350, acc=0.693, acc_len=1.66, thru=15.7, I=15.4, T=34.3, wait=6.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:22:19,518] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16765: forward=2647.5ms backward=1028.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:19,784] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:28,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:22:28,834] loop.py:45 INFO TRAIN_STEP step=16767 loss=1.110542 Training: 93%|█████████▎| 16767/18018 [35:52:14<2:58:44, 8.57s/step, loss=1.111, acc=0.732, acc_len=1.79, thru=14.2, I=14.1, T=34.7, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:29,123] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:33,410] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:22:34,435] loop.py:45 INFO TRAIN_STEP step=16768 loss=1.169236 Training: 93%|█████████▎| 16768/18018 [35:52:20<2:40:01, 7.68s/step, loss=1.169, acc=0.723, acc_len=1.76, thru=20.7, I=24.4, T=37.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:34,918] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:42,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:22:43,109] loop.py:45 INFO TRAIN_STEP step=16769 loss=1.353542 Training: 93%|█████████▎| 16769/18018 [35:52:29<2:46:05, 7.98s/step, loss=1.354, acc=0.692, acc_len=1.65, thru=13.6, I=17.0, T=36.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:43,481] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:52,124] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:22:52,503] loop.py:380 INFO TIMING step=16770: step=3.246s data=1.338s compute=3.221s [fwd=2.336s bwd=0.872s opt=0.003s] dispatch=6.139s [2026-10-20 01:22:52,504] loop.py:45 INFO TRAIN_STEP step=16770 loss=1.156774 Training: 93%|█████████▎| 16770/18018 [35:52:38<2:54:48, 8.40s/step, loss=1.157, acc=0.723, acc_len=1.76, thru=14.2, I=23.7, T=39.4, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:52,637] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:58,328] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:22:58,418] loop.py:45 INFO TRAIN_STEP step=16771 loss=1.182552 Training: 93%|█████████▎| 16771/18018 [35:52:44<2:39:08, 7.66s/step, loss=1.183, acc=0.719, acc_len=1.75, thru=21.7, I=22.8, T=38.5, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:22:58,414] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16770: forward=2358.3ms backward=922.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:22:58,831] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:23:05,411] loop.py:45 INFO TRAIN_STEP step=16772 loss=1.175488 Training: 93%|█████████▎| 16772/18018 [35:52:51<2:34:53, 7.46s/step, loss=1.175, acc=0.717, acc_len=1.75, thru=17.7, I=19.6, T=36.6, wait=3.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:13,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:23:13,403] loop.py:45 INFO TRAIN_STEP step=16773 loss=1.156074 Training: 93%|█████████▎| 16773/18018 [35:52:59<2:38:04, 7.62s/step, loss=1.156, acc=0.725, acc_len=1.76, thru=17.3, I=17.9, T=36.1, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:13,726] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:20,182] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:23:20,580] loop.py:45 INFO TRAIN_STEP step=16774 loss=1.441720 Training: 93%|█████████▎| 16774/18018 [35:53:06<2:35:12, 7.49s/step, loss=1.442, acc=0.678, acc_len=1.60, thru=17.3, I=18.6, T=35.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:20,692] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:25,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:23:26,607] loop.py:380 INFO TIMING step=16775: step=3.381s data=1.360s compute=3.354s [fwd=2.409s bwd=0.931s opt=0.003s] dispatch=2.637s [2026-10-20 01:23:26,607] loop.py:45 INFO TRAIN_STEP step=16775 loss=1.167860 Training: 93%|█████████▎| 16775/18018 [35:53:12<2:26:00, 7.05s/step, loss=1.168, acc=0.725, acc_len=1.77, thru=20.7, I=22.1, T=37.9, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:26,953] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:32,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:23:34,192] loop.py:45 INFO TRAIN_STEP step=16776 loss=1.340724 Training: 93%|█████████▎| 16776/18018 [35:53:20<2:29:13, 7.21s/step, loss=1.341, acc=0.692, acc_len=1.66, thru=17.2, I=21.2, T=37.0, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:23:34,188] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16775: forward=2473.2ms backward=934.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:34,483] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:23:40,692] loop.py:45 INFO TRAIN_STEP step=16777 loss=1.162805 Training: 93%|█████████▎| 16777/18018 [35:53:26<2:24:42, 7.00s/step, loss=1.163, acc=0.722, acc_len=1.76, thru=16.0, I=19.5, T=36.8, wait=3.0s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:50,174] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:23:51,382] loop.py:45 INFO TRAIN_STEP step=16778 loss=1.270933 Training: 93%|█████████▎| 16778/18018 [35:53:37<2:47:29, 8.10s/step, loss=1.271, acc=0.707, acc_len=1.70, thru=16.4, I=15.4, T=32.5, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:23:51,683] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:00,577] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:24:00,693] loop.py:45 INFO TRAIN_STEP step=16779 loss=1.276144 Training: 93%|█████████▎| 16779/18018 [35:53:46<2:54:50, 8.47s/step, loss=1.276, acc=0.700, acc_len=1.67, thru=14.2, I=13.7, T=35.4, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:01,078] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:06,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:24:07,454] loop.py:380 INFO TIMING step=16780: step=3.574s data=1.527s compute=3.523s [fwd=2.450s bwd=1.059s opt=0.003s] dispatch=3.177s [2026-10-20 01:24:07,454] loop.py:45 INFO TRAIN_STEP step=16780 loss=1.318026 Training: 93%|█████████▎| 16780/18018 [35:53:53<2:44:08, 7.96s/step, loss=1.318, acc=0.698, acc_len=1.68, thru=20.2, I=18.4, T=35.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:07,729] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:24:14,024] loop.py:45 INFO TRAIN_STEP step=16781 loss=1.333903 Training: 93%|█████████▎| 16781/18018 [35:54:00<2:35:25, 7.54s/step, loss=1.334, acc=0.695, acc_len=1.67, thru=18.4, I=20.7, T=37.7, wait=3.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:24:14,020] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16780: forward=2406.1ms backward=937.0ms [2026-10-20 01:24:19,778] loop.py:45 INFO TRAIN_STEP step=16782 loss=1.307329 Training: 93%|█████████▎| 16782/18018 [35:54:05<2:24:17, 7.00s/step, loss=1.307, acc=0.704, acc_len=1.69, thru=20.4, I=22.5, T=37.3, wait=2.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:26,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:24:27,776] loop.py:45 INFO TRAIN_STEP step=16783 loss=1.271212 Training: 93%|█████████▎| 16783/18018 [35:54:13<2:30:18, 7.30s/step, loss=1.271, acc=0.703, acc_len=1.68, thru=16.7, I=19.4, T=35.6, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:28,138] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:24:33,595] loop.py:45 INFO TRAIN_STEP step=16784 loss=1.176008 Training: 93%|█████████▎| 16784/18018 [35:54:19<2:21:02, 6.86s/step, loss=1.176, acc=0.720, acc_len=1.75, thru=18.9, I=23.5, T=37.1, wait=2.4s, pool=49, epoch=2/2][2026-10-20 01:24:42,074] loop.py:380 INFO TIMING step=16785: step=3.640s data=1.560s compute=3.604s [fwd=2.604s bwd=0.986s opt=0.003s] dispatch=4.826s [2026-10-20 01:24:42,074] loop.py:45 INFO TRAIN_STEP step=16785 loss=1.368726 Training: 93%|█████████▎| 16785/18018 [35:54:28<2:30:53, 7.34s/step, loss=1.369, acc=0.689, acc_len=1.64, thru=16.7, I=16.3, T=35.2, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:42,058] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:42,561] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:48,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:24:48,903] loop.py:45 INFO TRAIN_STEP step=16786 loss=1.401624 Training: 93%|█████████▎| 16786/18018 [35:54:34<2:27:36, 7.19s/step, loss=1.402, acc=0.680, acc_len=1.61, thru=18.2, I=20.2, T=37.8, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:24:48,899] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16785: forward=2434.2ms backward=911.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:49,267] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:24:54,914] loop.py:45 INFO TRAIN_STEP step=16787 loss=1.178094 Training: 93%|█████████▎| 16787/18018 [35:54:41<2:20:15, 6.84s/step, loss=1.178, acc=0.725, acc_len=1.77, thru=19.7, I=22.5, T=38.4, wait=2.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:24:59,518] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:25:01,238] loop.py:45 INFO TRAIN_STEP step=16788 loss=1.368238 Training: 93%|█████████▎| 16788/18018 [35:54:47<2:16:59, 6.68s/step, loss=1.368, acc=0.686, acc_len=1.63, thru=20.2, I=24.2, T=37.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:01,534] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 01:25:07,071] loop.py:45 INFO TRAIN_STEP step=16789 loss=1.127453 Training: 93%|█████████▎| 16789/18018 [35:54:53<2:11:39, 6.43s/step, loss=1.127, acc=0.729, acc_len=1.78, thru=20.1, I=29.7, T=40.5, wait=2.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:15,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:25:16,838] loop.py:380 INFO TIMING step=16790: step=3.693s data=1.468s compute=3.661s [fwd=2.620s bwd=1.027s opt=0.003s] dispatch=6.063s [2026-10-20 01:25:16,838] loop.py:45 INFO TRAIN_STEP step=16790 loss=1.324295 Training: 93%|█████████▎| 16790/18018 [35:55:02<2:32:03, 7.43s/step, loss=1.324, acc=0.697, acc_len=1.67, thru=15.8, I=15.6, T=34.7, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:16,883] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:22,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 01:25:23,243] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16790: forward=2412.5ms backward=916.4ms [2026-10-20 01:25:23,248] loop.py:45 INFO TRAIN_STEP step=16791 loss=1.396978 Training: 93%|█████████▎| 16791/18018 [35:55:09<2:25:39, 7.12s/step, loss=1.397, acc=0.680, acc_len=1.61, thru=19.5, I=22.7, T=37.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:23,744] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:30,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:25:31,210] loop.py:45 INFO TRAIN_STEP step=16792 loss=1.328173 Training: 93%|█████████▎| 16792/18018 [35:55:17<2:30:41, 7.38s/step, loss=1.328, acc=0.693, acc_len=1.66, thru=16.1, I=18.5, T=35.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:31,487] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:36,433] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:25:37,951] loop.py:45 INFO TRAIN_STEP step=16793 loss=1.248438 Training: 93%|█████████▎| 16793/18018 [35:55:24<2:26:40, 7.18s/step, loss=1.248, acc=0.706, acc_len=1.70, thru=18.1, I=21.8, T=35.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:38,444] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:43,454] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:25:43,942] loop.py:45 INFO TRAIN_STEP step=16794 loss=1.120919 Training: 93%|█████████▎| 16794/18018 [35:55:30<2:19:16, 6.83s/step, loss=1.121, acc=0.729, acc_len=1.78, thru=20.9, I=27.7, T=38.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:25:43,975] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-20 01:25:50,580] loop.py:380 INFO TIMING step=16795: step=3.291s data=1.336s compute=3.263s [fwd=2.337s bwd=0.912s opt=0.003s] dispatch=3.337s [2026-10-20 01:25:50,580] loop.py:45 INFO TRAIN_STEP step=16795 loss=1.186632 Training: 93%|█████████▎| 16795/18018 [35:55:36<2:17:59, 6.77s/step, loss=1.187, acc=0.719, acc_len=1.74, thru=18.9, I=20.7, T=38.9, wait=3.3s, pool=57, epoch=2/2][2026-10-20 01:25:57,851] loop.py:45 INFO TRAIN_STEP step=16796 loss=1.275603 Training: 93%|█████████▎| 16796/18018 [35:55:43<2:20:58, 6.92s/step, loss=1.276, acc=0.701, acc_len=1.69, thru=16.0, I=18.3, T=37.9, wait=3.9s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:25:57,847] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16795: forward=2424.3ms backward=912.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:26:06,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:26:06,285] loop.py:45 INFO TRAIN_STEP step=16797 loss=1.302129 Training: 93%|█████████▎| 16797/18018 [35:55:52<2:30:03, 7.37s/step, loss=1.302, acc=0.697, acc_len=1.67, thru=17.3, I=17.7, T=35.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:26:06,622] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:26:13,100] loop.py:45 INFO TRAIN_STEP step=16798 loss=1.251168 Training: 93%|█████████▎| 16798/18018 [35:55:59<2:26:31, 7.21s/step, loss=1.251, acc=0.710, acc_len=1.71, thru=16.2, I=19.3, T=36.5, wait=3.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:26:19,594] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:26:21,057] loop.py:45 INFO TRAIN_STEP step=16799 loss=1.359657 Training: 93%|█████████▎| 16799/18018 [35:56:07<2:30:58, 7.43s/step, loss=1.360, acc=0.690, acc_len=1.65, thru=15.7, I=18.7, T=35.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:26:21,102] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:26:27,478] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:26:28,221] loop.py:380 INFO TIMING step=16800: step=3.797s data=1.721s compute=3.767s [fwd=2.708s bwd=1.044s opt=0.003s] dispatch=3.358s [2026-10-20 01:26:28,222] loop.py:45 INFO TRAIN_STEP step=16800 loss=1.187615 Training: 93%|█████████▎| 16800/18018 [35:56:14<2:29:13, 7.35s/step, loss=1.188, acc=0.720, acc_len=1.74, thru=16.5, I=20.4, T=33.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:26:28,484] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-20 01:26:33,950] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16800: forward=2396.4ms backward=952.8ms [2026-10-20 01:26:33,955] loop.py:45 INFO TRAIN_STEP step=16801 loss=1.184423 Training: 93%|█████████▎| 16801/18018 [35:56:20<2:19:15, 6.87s/step, loss=1.184, acc=0.722, acc_len=1.76, thru=19.3, I=20.6, T=37.7, wait=2.3s, pool=49, epoch=2/2][2026-10-20 01:26:43,723] loop.py:45 INFO TRAIN_STEP step=16802 loss=1.367211 Training: 93%|█████████▎| 16802/18018 [35:56:29<2:36:47, 7.74s/step, loss=1.367, acc=0.688, acc_len=1.64, thru=12.8, I=12.8, T=33.8, wait=6.0s, pool=45, epoch=2/2][2026-10-20 01:26:50,391] loop.py:45 INFO TRAIN_STEP step=16803 loss=1.146322 Training: 93%|█████████▎| 16803/18018 [35:56:36<2:30:10, 7.42s/step, loss=1.146, acc=0.730, acc_len=1.79, thru=17.7, I=17.7, T=36.3, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:26:57,038] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:26:58,225] loop.py:45 INFO TRAIN_STEP step=16804 loss=1.305923 Training: 93%|█████████▎| 16804/18018 [35:56:44<2:32:36, 7.54s/step, loss=1.306, acc=0.702, acc_len=1.69, thru=17.7, I=18.0, T=35.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:26:58,548] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:27:06,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:27:07,249] loop.py:380 INFO TIMING step=16805: step=3.666s data=1.490s compute=3.626s [fwd=2.578s bwd=1.033s opt=0.003s] dispatch=5.347s [2026-10-20 01:27:07,249] loop.py:45 INFO TRAIN_STEP step=16805 loss=1.309573 Training: 93%|█████████▎| 16805/18018 [35:56:53<2:41:27, 7.99s/step, loss=1.310, acc=0.695, acc_len=1.67, thru=14.7, I=16.0, T=34.9, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:27:07,588] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:27:12,961] loop.py:45 INFO TRAIN_STEP step=16806 loss=1.270190 Training: 93%|█████████▎| 16806/18018 [35:56:59<2:27:31, 7.30s/step, loss=1.270, acc=0.706, acc_len=1.70, thru=19.7, I=21.9, T=37.9, wait=2.3s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:27:12,956] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16805: forward=2394.8ms backward=938.5ms [2026-10-20 01:27:20,641] loop.py:45 INFO TRAIN_STEP step=16807 loss=1.166181 Training: 93%|█████████▎| 16807/18018 [35:57:06<2:29:42, 7.42s/step, loss=1.166, acc=0.725, acc_len=1.76, thru=16.1, I=16.2, T=37.1, wait=4.2s, pool=57, epoch=2/2][2026-10-20 01:27:27,730] loop.py:45 INFO TRAIN_STEP step=16808 loss=1.288257 Training: 93%|█████████▎| 16808/18018 [35:57:13<2:27:35, 7.32s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=18.2, I=17.7, T=36.7, wait=3.6s, pool=57, epoch=2/2][2026-10-20 01:27:35,633] loop.py:45 INFO TRAIN_STEP step=16809 loss=1.318163 Training: 93%|█████████▎| 16809/18018 [35:57:21<2:31:00, 7.49s/step, loss=1.318, acc=0.695, acc_len=1.66, thru=16.8, I=15.7, T=32.2, wait=3.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:27:43,646] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:27:44,587] loop.py:380 INFO TIMING step=16810: step=3.662s data=1.388s compute=3.620s [fwd=2.599s bwd=1.007s opt=0.003s] dispatch=5.281s [2026-10-20 01:27:44,587] loop.py:45 INFO TRAIN_STEP step=16810 loss=1.150788 Training: 93%|█████████▎| 16810/18018 [35:57:30<2:39:41, 7.93s/step, loss=1.151, acc=0.725, acc_len=1.77, thru=16.8, I=15.6, T=35.0, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:27:44,657] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-20 01:27:50,455] loop.py:45 INFO TRAIN_STEP step=16811 loss=1.127164 Training: 93%|█████████▎| 16811/18018 [35:57:36<2:27:06, 7.31s/step, loss=1.127, acc=0.730, acc_len=1.79, thru=20.1, I=18.3, T=37.1, wait=2.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:27:50,451] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16810: forward=2445.9ms backward=952.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:27:58,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:27:59,222] loop.py:45 INFO TRAIN_STEP step=16812 loss=1.276878 Training: 93%|█████████▎| 16812/18018 [35:57:45<2:35:44, 7.75s/step, loss=1.277, acc=0.706, acc_len=1.70, thru=15.7, I=17.1, T=33.1, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:27:59,577] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:05,241] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:28:06,167] loop.py:45 INFO TRAIN_STEP step=16813 loss=1.248027 Training: 93%|█████████▎| 16813/18018 [35:57:52<2:30:46, 7.51s/step, loss=1.248, acc=0.705, acc_len=1.70, thru=17.6, I=22.2, T=38.2, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:06,249] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:11,403] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:28:12,198] loop.py:45 INFO TRAIN_STEP step=16814 loss=1.211560 Training: 93%|█████████▎| 16814/18018 [35:57:58<2:21:46, 7.06s/step, loss=1.212, acc=0.714, acc_len=1.72, thru=20.9, I=24.5, T=38.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:12,409] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:20,042] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:28:20,799] loop.py:380 INFO TIMING step=16815: step=3.600s data=1.441s compute=3.550s [fwd=2.456s bwd=1.079s opt=0.003s] dispatch=4.993s [2026-10-20 01:28:20,799] loop.py:45 INFO TRAIN_STEP step=16815 loss=1.298887 Training: 93%|█████████▎| 16815/18018 [35:58:06<2:30:54, 7.53s/step, loss=1.299, acc=0.697, acc_len=1.67, thru=15.2, I=18.1, T=35.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:21,048] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:27,263] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:28:27,574] loop.py:45 INFO TRAIN_STEP step=16816 loss=1.362276 Training: 93%|█████████▎| 16816/18018 [35:58:13<2:26:15, 7.30s/step, loss=1.362, acc=0.693, acc_len=1.66, thru=19.7, I=18.8, T=36.7, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:28:27,569] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16815: forward=2409.9ms backward=1025.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:27,765] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:28:33,289] loop.py:45 INFO TRAIN_STEP step=16817 loss=1.222217 Training: 93%|█████████▎| 16817/18018 [35:58:19<2:16:36, 6.82s/step, loss=1.222, acc=0.712, acc_len=1.72, thru=17.8, I=18.9, T=39.5, wait=2.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:42,467] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:28:43,966] loop.py:45 INFO TRAIN_STEP step=16818 loss=1.306753 Training: 93%|█████████▎| 16818/18018 [35:58:30<2:39:36, 7.98s/step, loss=1.307, acc=0.699, acc_len=1.68, thru=15.8, I=16.6, T=32.4, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:44,006] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-20 01:28:49,545] loop.py:45 INFO TRAIN_STEP step=16819 loss=1.143693 Training: 93%|█████████▎| 16819/18018 [35:58:35<2:25:04, 7.26s/step, loss=1.144, acc=0.727, acc_len=1.77, thru=18.7, I=23.3, T=36.8, wait=2.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:57,671] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:28:58,364] loop.py:380 INFO TIMING step=16820: step=3.689s data=1.581s compute=3.630s [fwd=2.555s bwd=1.061s opt=0.003s] dispatch=5.121s [2026-10-20 01:28:58,364] loop.py:45 INFO TRAIN_STEP step=16820 loss=1.325884 Training: 93%|█████████▎| 16820/18018 [35:58:44<2:34:17, 7.73s/step, loss=1.326, acc=0.695, acc_len=1.66, thru=16.8, I=16.4, T=34.7, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:28:58,678] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:29:04,693] loop.py:45 INFO TRAIN_STEP step=16821 loss=1.150786 Training: 93%|█████████▎| 16821/18018 [35:58:50<2:25:47, 7.31s/step, loss=1.151, acc=0.726, acc_len=1.77, thru=17.1, I=20.8, T=38.6, wait=3.0s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:29:04,688] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16820: forward=2338.0ms backward=925.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:14,807] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:29:15,363] loop.py:45 INFO TRAIN_STEP step=16822 loss=1.279298 Training: 93%|█████████▎| 16822/18018 [35:59:01<2:45:46, 8.32s/step, loss=1.279, acc=0.705, acc_len=1.70, thru=15.2, I=13.9, T=34.5, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:15,811] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:20,709] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:29:21,899] loop.py:45 INFO TRAIN_STEP step=16823 loss=1.266641 Training: 93%|█████████▎| 16823/18018 [35:59:07<2:34:59, 7.78s/step, loss=1.267, acc=0.705, acc_len=1.70, thru=19.7, I=21.1, T=37.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:22,216] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:28,945] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:29:29,803] loop.py:45 INFO TRAIN_STEP step=16824 loss=1.322082 Training: 93%|█████████▎| 16824/18018 [35:59:15<2:35:36, 7.82s/step, loss=1.322, acc=0.701, acc_len=1.68, thru=15.8, I=18.7, T=35.0, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:29,958] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-20 01:29:36,664] loop.py:380 INFO TIMING step=16825: step=3.387s data=1.354s compute=3.352s [fwd=2.356s bwd=0.981s opt=0.003s] dispatch=3.463s [2026-10-20 01:29:36,664] loop.py:45 INFO TRAIN_STEP step=16825 loss=1.223476 Training: 93%|█████████▎| 16825/18018 [35:59:22<2:29:44, 7.53s/step, loss=1.223, acc=0.712, acc_len=1.72, thru=15.0, I=18.9, T=37.8, wait=3.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:45,609] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:29:46,661] loop.py:45 INFO TRAIN_STEP step=16826 loss=1.158983 Training: 93%|█████████▎| 16826/18018 [35:59:32<2:44:19, 8.27s/step, loss=1.159, acc=0.726, acc_len=1.77, thru=16.6, I=15.2, T=35.8, wait=6.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:29:46,657] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16825: forward=2532.3ms backward=1002.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:47,118] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:53,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:29:55,034] loop.py:45 INFO TRAIN_STEP step=16827 loss=1.279391 Training: 93%|█████████▎| 16827/18018 [35:59:41<2:44:47, 8.30s/step, loss=1.279, acc=0.704, acc_len=1.69, thru=14.5, I=18.0, T=35.2, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:29:55,234] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:02,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:30:02,957] loop.py:45 INFO TRAIN_STEP step=16828 loss=1.384148 Training: 93%|█████████▎| 16828/18018 [35:59:49<2:42:23, 8.19s/step, loss=1.384, acc=0.683, acc_len=1.62, thru=14.8, I=18.0, T=34.6, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:03,347] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:11,012] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:30:11,733] loop.py:45 INFO TRAIN_STEP step=16829 loss=1.151723 Training: 93%|█████████▎| 16829/18018 [35:59:57<2:45:45, 8.36s/step, loss=1.152, acc=0.724, acc_len=1.77, thru=14.7, I=17.2, T=36.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:12,017] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:17,867] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:30:18,481] loop.py:380 INFO TIMING step=16830: step=3.534s data=1.323s compute=3.503s [fwd=2.508s bwd=0.980s opt=0.003s] dispatch=3.205s [2026-10-20 01:30:18,481] loop.py:45 INFO TRAIN_STEP step=16830 loss=1.358527 Training: 93%|█████████▎| 16830/18018 [36:00:04<2:36:00, 7.88s/step, loss=1.359, acc=0.689, acc_len=1.64, thru=19.3, I=19.4, T=36.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:18,869] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:26,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:30:27,044] loop.py:45 INFO TRAIN_STEP step=16831 loss=1.337851 Training: 93%|█████████▎| 16831/18018 [36:00:13<2:39:56, 8.08s/step, loss=1.338, acc=0.693, acc_len=1.65, thru=15.1, I=17.1, T=35.3, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:30:27,040] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16830: forward=2479.9ms backward=1069.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:27,299] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:31,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:30:32,871] loop.py:45 INFO TRAIN_STEP step=16832 loss=1.207764 Training: 93%|█████████▎| 16832/18018 [36:00:18<2:26:24, 7.41s/step, loss=1.208, acc=0.714, acc_len=1.72, thru=21.3, I=24.2, T=38.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:33,374] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:38,160] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:30:39,213] loop.py:45 INFO TRAIN_STEP step=16833 loss=1.231032 Training: 93%|█████████▎| 16833/18018 [36:00:25<2:19:59, 7.09s/step, loss=1.231, acc=0.713, acc_len=1.73, thru=21.0, I=27.0, T=38.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:39,668] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:30:45,680] loop.py:45 INFO TRAIN_STEP step=16834 loss=1.316002 Training: 93%|█████████▎| 16834/18018 [36:00:31<2:16:11, 6.90s/step, loss=1.316, acc=0.700, acc_len=1.68, thru=18.4, I=23.5, T=36.9, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:55,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:30:56,523] loop.py:380 INFO TIMING step=16835: step=3.748s data=1.574s compute=3.707s [fwd=2.601s bwd=1.092s opt=0.003s] dispatch=7.085s [2026-10-20 01:30:56,523] loop.py:45 INFO TRAIN_STEP step=16835 loss=1.199504 Training: 93%|█████████▎| 16835/18018 [36:00:42<2:39:23, 8.08s/step, loss=1.200, acc=0.722, acc_len=1.75, thru=13.7, I=14.5, T=34.2, wait=7.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:30:56,965] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:31:04,392] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:31:05,735] loop.py:45 INFO TRAIN_STEP step=16836 loss=1.245913 Training: 93%|█████████▎| 16836/18018 [36:00:51<2:45:55, 8.42s/step, loss=1.246, acc=0.703, acc_len=1.69, thru=14.8, I=15.0, T=32.6, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:31:05,730] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16835: forward=2759.1ms backward=1121.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:31:05,904] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:31:11,033] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:31:11,491] loop.py:45 INFO TRAIN_STEP step=16837 loss=1.225891 Training: 93%|█████████▎| 16837/18018 [36:00:57<2:30:01, 7.62s/step, loss=1.226, acc=0.708, acc_len=1.71, thru=20.2, I=22.3, T=40.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:31:11,544] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:31:17,449] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:31:18,319] loop.py:45 INFO TRAIN_STEP step=16838 loss=1.265807 Training: 93%|█████████▎| 16838/18018 [36:01:04<2:25:13, 7.38s/step, loss=1.266, acc=0.702, acc_len=1.69, thru=19.1, I=21.7, T=37.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:31:18,464] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:31:25,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:31:26,204] loop.py:45 INFO TRAIN_STEP step=16839 loss=1.240936 Training: 93%|█████████▎| 16839/18018 [36:01:12<2:28:03, 7.53s/step, loss=1.241, acc=0.713, acc_len=1.72, thru=15.7, I=17.7, T=35.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:31:26,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:31:32,390] loop.py:380 INFO TIMING step=16840: step=3.517s data=1.483s compute=3.487s [fwd=2.504s bwd=0.968s opt=0.003s] dispatch=2.659s [2026-10-20 01:31:32,391] loop.py:45 INFO TRAIN_STEP step=16840 loss=1.277928 Training: 93%|█████████▎| 16840/18018 [36:01:18<2:19:59, 7.13s/step, loss=1.278, acc=0.701, acc_len=1.69, thru=19.6, I=19.9, T=36.4, wait=2.7s, pool=61, epoch=2/2][2026-10-20 01:31:41,153] loop.py:45 INFO TRAIN_STEP step=16841 loss=1.263751 Training: 93%|█████████▎| 16841/18018 [36:01:27<2:29:27, 7.62s/step, loss=1.264, acc=0.705, acc_len=1.70, thru=15.3, I=14.2, T=34.5, wait=5.0s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:31:41,148] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16840: forward=2485.6ms backward=1032.9ms [2026-10-20 01:31:47,837] loop.py:45 INFO TRAIN_STEP step=16842 loss=1.221588 Training: 93%|█████████▎| 16842/18018 [36:01:33<2:23:50, 7.34s/step, loss=1.222, acc=0.712, acc_len=1.72, thru=17.2, I=16.5, T=33.8, wait=2.9s, pool=49, epoch=2/2][2026-10-20 01:31:55,196] loop.py:45 INFO TRAIN_STEP step=16843 loss=1.328313 Training: 93%|█████████▎| 16843/18018 [36:01:41<2:23:50, 7.35s/step, loss=1.328, acc=0.693, acc_len=1.66, thru=17.4, I=17.2, T=35.5, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:01,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:32:03,053] loop.py:45 INFO TRAIN_STEP step=16844 loss=1.285456 Training: 93%|█████████▎| 16844/18018 [36:01:49<2:26:44, 7.50s/step, loss=1.285, acc=0.703, acc_len=1.69, thru=16.8, I=17.8, T=35.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:03,224] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:32:08,922] loop.py:380 INFO TIMING step=16845: step=3.499s data=1.505s compute=3.472s [fwd=2.494s bwd=0.963s opt=0.003s] dispatch=2.360s [2026-10-20 01:32:08,922] loop.py:45 INFO TRAIN_STEP step=16845 loss=1.294676 Training: 93%|█████████▎| 16845/18018 [36:01:55<2:17:02, 7.01s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=20.4, I=21.8, T=36.6, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:16,738] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:32:17,120] loop.py:45 INFO TRAIN_STEP step=16846 loss=1.243979 (TrainerActor pid=9310) [2026-10-20 01:32:17,115] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16845: forward=2426.2ms backward=1035.7ms Training: 93%|█████████▎| 16846/18018 [36:02:03<2:23:53, 7.37s/step, loss=1.244, acc=0.705, acc_len=1.70, thru=16.0, I=15.6, T=34.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:17,241] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:24,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:32:25,979] loop.py:45 INFO TRAIN_STEP step=16847 loss=1.191236 Training: 94%|█████████▎| 16847/18018 [36:02:12<2:32:30, 7.81s/step, loss=1.191, acc=0.719, acc_len=1.75, thru=15.2, I=18.6, T=33.1, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:26,017] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:31,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:32:32,168] loop.py:45 INFO TRAIN_STEP step=16848 loss=1.281929 Training: 94%|█████████▎| 16848/18018 [36:02:18<2:22:52, 7.33s/step, loss=1.282, acc=0.702, acc_len=1.69, thru=20.1, I=24.6, T=38.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:32,530] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:38,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:32:38,615] loop.py:45 INFO TRAIN_STEP step=16849 loss=1.310614 Training: 94%|█████████▎| 16849/18018 [36:02:24<2:17:35, 7.06s/step, loss=1.311, acc=0.699, acc_len=1.67, thru=19.9, I=22.6, T=38.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:38,645] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:43,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:32:44,130] loop.py:380 INFO TIMING step=16850: step=3.310s data=1.293s compute=3.277s [fwd=2.324s bwd=0.939s opt=0.003s] dispatch=2.196s [2026-10-20 01:32:44,131] loop.py:45 INFO TRAIN_STEP step=16850 loss=1.270209 Training: 94%|█████████▎| 16850/18018 [36:02:30<2:08:26, 6.60s/step, loss=1.270, acc=0.707, acc_len=1.70, thru=22.8, I=25.9, T=38.7, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:44,346] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:52,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:32:52,858] loop.py:45 INFO TRAIN_STEP step=16851 loss=1.331840 Training: 94%|█████████▎| 16851/18018 [36:02:38<2:20:46, 7.24s/step, loss=1.332, acc=0.693, acc_len=1.65, thru=14.4, I=16.0, T=34.9, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:32:52,853] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16850: forward=2566.8ms backward=1047.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:32:52,981] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:00,137] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:33:01,304] loop.py:45 INFO TRAIN_STEP step=16852 loss=1.207161 Training: 94%|█████████▎| 16852/18018 [36:02:47<2:27:41, 7.60s/step, loss=1.207, acc=0.720, acc_len=1.76, thru=14.2, I=16.7, T=34.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:01,647] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:06,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:33:07,417] loop.py:45 INFO TRAIN_STEP step=16853 loss=1.223995 Training: 94%|█████████▎| 16853/18018 [36:02:53<2:18:53, 7.15s/step, loss=1.224, acc=0.712, acc_len=1.72, thru=20.1, I=24.8, T=37.8, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:07,729] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:14,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:33:15,187] loop.py:45 INFO TRAIN_STEP step=16854 loss=1.321802 Training: 94%|█████████▎| 16854/18018 [36:03:01<2:22:21, 7.34s/step, loss=1.322, acc=0.694, acc_len=1.66, thru=16.0, I=19.8, T=36.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:15,673] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:22,893] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:33:23,727] loop.py:380 INFO TIMING step=16855: step=3.508s data=1.435s compute=3.478s [fwd=2.485s bwd=0.977s opt=0.003s] dispatch=5.025s [2026-10-20 01:33:23,728] loop.py:45 INFO TRAIN_STEP step=16855 loss=1.262099 Training: 94%|█████████▎| 16855/18018 [36:03:09<2:29:14, 7.70s/step, loss=1.262, acc=0.706, acc_len=1.70, thru=15.3, I=18.2, T=36.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:23,897] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:33:30,356] loop.py:45 INFO TRAIN_STEP step=16856 loss=1.334756 Training: 94%|█████████▎| 16856/18018 [36:03:16<2:22:53, 7.38s/step, loss=1.335, acc=0.692, acc_len=1.65, thru=18.7, I=18.7, T=34.4, wait=2.9s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:33:30,351] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16855: forward=2576.8ms backward=1105.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:36,562] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:33:36,988] loop.py:45 INFO TRAIN_STEP step=16857 loss=1.274819 Training: 94%|█████████▎| 16857/18018 [36:03:23<2:18:25, 7.15s/step, loss=1.275, acc=0.702, acc_len=1.69, thru=18.9, I=20.1, T=36.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:37,067] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-20 01:33:44,330] loop.py:45 INFO TRAIN_STEP step=16858 loss=1.261927 Training: 94%|█████████▎| 16858/18018 [36:03:30<2:19:24, 7.21s/step, loss=1.262, acc=0.705, acc_len=1.70, thru=17.0, I=18.9, T=35.8, wait=3.8s, pool=53, epoch=2/2][2026-10-20 01:33:53,387] loop.py:45 INFO TRAIN_STEP step=16859 loss=1.309862 Training: 94%|█████████▎| 16859/18018 [36:03:39<2:29:58, 7.76s/step, loss=1.310, acc=0.700, acc_len=1.68, thru=14.4, I=13.2, T=33.6, wait=5.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:33:58,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:34:00,353] loop.py:380 INFO TIMING step=16860: step=3.725s data=1.580s compute=3.694s [fwd=2.664s bwd=1.016s opt=0.003s] dispatch=3.233s [2026-10-20 01:34:00,353] loop.py:45 INFO TRAIN_STEP step=16860 loss=1.315007 Training: 94%|█████████▎| 16860/18018 [36:03:46<2:25:14, 7.53s/step, loss=1.315, acc=0.703, acc_len=1.70, thru=19.3, I=19.9, T=34.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:00,386] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-20 01:34:05,934] loop.py:45 INFO TRAIN_STEP step=16861 loss=1.208076 Training: 94%|█████████▎| 16861/18018 [36:03:52<2:13:51, 6.94s/step, loss=1.208, acc=0.717, acc_len=1.74, thru=21.9, I=26.8, T=38.3, wait=2.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:34:05,930] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16860: forward=2399.7ms backward=900.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:11,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:34:12,030] loop.py:45 INFO TRAIN_STEP step=16862 loss=1.213871 Training: 94%|█████████▎| 16862/18018 [36:03:58<2:08:52, 6.69s/step, loss=1.214, acc=0.717, acc_len=1.74, thru=20.9, I=22.0, T=36.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:12,289] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:17,023] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:34:17,899] loop.py:45 INFO TRAIN_STEP step=16863 loss=1.252733 Training: 94%|█████████▎| 16863/18018 [36:04:03<2:04:00, 6.44s/step, loss=1.253, acc=0.706, acc_len=1.70, thru=21.5, I=24.6, T=38.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:18,031] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:23,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:34:23,690] loop.py:45 INFO TRAIN_STEP step=16864 loss=1.244120 Training: 94%|█████████▎| 16864/18018 [36:04:09<2:00:08, 6.25s/step, loss=1.244, acc=0.709, acc_len=1.71, thru=22.7, I=25.9, T=38.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:24,124] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:34:30,570] loop.py:380 INFO TIMING step=16865: step=3.482s data=1.419s compute=3.452s [fwd=2.478s bwd=0.960s opt=0.003s] dispatch=3.390s [2026-10-20 01:34:30,570] loop.py:45 INFO TRAIN_STEP step=16865 loss=1.088735 Training: 94%|█████████▎| 16865/18018 [36:04:16<2:03:41, 6.44s/step, loss=1.089, acc=0.737, acc_len=1.81, thru=17.3, I=19.7, T=36.8, wait=3.4s, pool=49, epoch=2/2][2026-10-20 01:34:39,506] loop.py:45 INFO TRAIN_STEP step=16866 loss=1.360284 Training: 94%|█████████▎| 16866/18018 [36:04:25<2:17:59, 7.19s/step, loss=1.360, acc=0.690, acc_len=1.64, thru=15.1, I=14.7, T=33.7, wait=5.1s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:34:39,502] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16865: forward=2651.0ms backward=1091.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:44,850] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:34:45,781] loop.py:45 INFO TRAIN_STEP step=16867 loss=1.333245 Training: 94%|█████████▎| 16867/18018 [36:04:31<2:12:37, 6.91s/step, loss=1.333, acc=0.697, acc_len=1.66, thru=19.5, I=20.9, T=35.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:45,860] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-20 01:34:52,569] loop.py:45 INFO TRAIN_STEP step=16868 loss=1.143912 Training: 94%|█████████▎| 16868/18018 [36:04:38<2:11:46, 6.88s/step, loss=1.144, acc=0.730, acc_len=1.78, thru=17.3, I=20.3, T=37.0, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:58,411] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:34:59,311] loop.py:45 INFO TRAIN_STEP step=16869 loss=1.192715 Training: 94%|█████████▎| 16869/18018 [36:04:45<2:10:54, 6.84s/step, loss=1.193, acc=0.717, acc_len=1.74, thru=19.6, I=19.8, T=39.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:34:59,421] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:35:04,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:35:05,157] loop.py:380 INFO TIMING step=16870: step=3.767s data=1.588s compute=3.735s [fwd=2.683s bwd=1.036s opt=0.003s] dispatch=2.071s [2026-10-20 01:35:05,158] loop.py:45 INFO TRAIN_STEP step=16870 loss=1.200215 Training: 94%|█████████▎| 16870/18018 [36:04:51<2:05:06, 6.54s/step, loss=1.200, acc=0.716, acc_len=1.74, thru=21.6, I=24.7, T=34.0, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:35:05,282] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:35:10,691] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:35:11,575] loop.py:45 INFO TRAIN_STEP step=16871 loss=1.307501 Training: 94%|█████████▎| 16871/18018 [36:04:57<2:04:18, 6.50s/step, loss=1.308, acc=0.696, acc_len=1.67, thru=19.8, I=21.3, T=36.3, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:35:11,571] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16870: forward=2556.2ms backward=926.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:35:11,707] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:35:17,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:35:18,291] loop.py:45 INFO TRAIN_STEP step=16872 loss=1.298435 Training: 94%|█████████▎| 16872/18018 [36:05:04<2:05:24, 6.57s/step, loss=1.298, acc=0.699, acc_len=1.68, thru=18.9, I=21.5, T=36.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:35:18,776] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:35:26,043] loop.py:45 INFO TRAIN_STEP step=16873 loss=1.189916 Training: 94%|█████████▎| 16873/18018 [36:05:12<2:12:05, 6.92s/step, loss=1.190, acc=0.716, acc_len=1.73, thru=16.1, I=15.6, T=38.3, wait=4.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:35:31,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:35:33,207] loop.py:45 INFO TRAIN_STEP step=16874 loss=1.300938 Training: 94%|█████████▎| 16874/18018 [36:05:19<2:13:22, 7.00s/step, loss=1.301, acc=0.700, acc_len=1.68, thru=20.1, I=19.7, T=35.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:35:33,305] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-20 01:35:38,351] loop.py:380 INFO TIMING step=16875: step=3.279s data=1.324s compute=3.249s [fwd=2.326s bwd=0.909s opt=0.003s] dispatch=1.855s [2026-10-20 01:35:38,351] loop.py:45 INFO TRAIN_STEP step=16875 loss=1.138671 Training: 94%|█████████▎| 16875/18018 [36:05:24<2:02:41, 6.44s/step, loss=1.139, acc=0.731, acc_len=1.80, thru=21.9, I=29.3, T=39.0, wait=1.9s, pool=45, epoch=2/2][2026-10-20 01:35:45,682] loop.py:45 INFO TRAIN_STEP step=16876 loss=1.188019 Training: 94%|█████████▎| 16876/18018 [36:05:31<2:07:39, 6.71s/step, loss=1.188, acc=0.717, acc_len=1.74, thru=16.8, I=17.1, T=36.7, wait=3.8s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:35:45,678] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16875: forward=2456.5ms backward=997.0ms [2026-10-20 01:35:53,803] loop.py:45 INFO TRAIN_STEP step=16877 loss=1.279044 Training: 94%|█████████▎| 16877/18018 [36:05:39<2:15:36, 7.13s/step, loss=1.279, acc=0.705, acc_len=1.70, thru=16.5, I=16.2, T=36.2, wait=4.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:35:59,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:36:00,850] loop.py:45 INFO TRAIN_STEP step=16878 loss=1.311422 Training: 94%|█████████▎| 16878/18018 [36:05:46<2:15:00, 7.11s/step, loss=1.311, acc=0.696, acc_len=1.66, thru=17.3, I=19.5, T=34.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:01,289] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:36:07,600] loop.py:45 INFO TRAIN_STEP step=16879 loss=1.196579 Training: 94%|█████████▎| 16879/18018 [36:05:53<2:12:52, 7.00s/step, loss=1.197, acc=0.718, acc_len=1.74, thru=16.5, I=21.8, T=38.0, wait=3.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:13,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:36:15,286] loop.py:380 INFO TIMING step=16880: step=3.776s data=1.590s compute=3.744s [fwd=2.694s bwd=1.038s opt=0.003s] dispatch=3.899s [2026-10-20 01:36:15,286] loop.py:45 INFO TRAIN_STEP step=16880 loss=1.268473 Training: 94%|█████████▎| 16880/18018 [36:06:01<2:16:39, 7.21s/step, loss=1.268, acc=0.708, acc_len=1.71, thru=17.2, I=19.5, T=33.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:15,457] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:20,518] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:36:21,011] loop.py:45 INFO TRAIN_STEP step=16881 loss=1.263856 Training: 94%|█████████▎| 16881/18018 [36:06:07<2:08:07, 6.76s/step, loss=1.264, acc=0.704, acc_len=1.69, thru=20.9, I=25.8, T=38.7, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:36:21,007] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16880: forward=2365.9ms backward=893.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:21,049] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:26,344] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:36:27,956] loop.py:45 INFO TRAIN_STEP step=16882 loss=1.261955 Training: 94%|█████████▎| 16882/18018 [36:06:14<2:09:03, 6.82s/step, loss=1.262, acc=0.710, acc_len=1.71, thru=18.8, I=23.8, T=27.8, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:28,356] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 01:36:34,793] loop.py:45 INFO TRAIN_STEP step=16883 loss=1.316324 Training: 94%|█████████▎| 16883/18018 [36:06:20<2:09:02, 6.82s/step, loss=1.316, acc=0.698, acc_len=1.67, thru=16.0, I=17.3, T=31.7, wait=2.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:42,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:36:43,703] loop.py:45 INFO TRAIN_STEP step=16884 loss=1.356255 Training: 94%|█████████▎| 16884/18018 [36:06:29<2:20:46, 7.45s/step, loss=1.356, acc=0.690, acc_len=1.64, thru=15.2, I=16.8, T=34.3, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:44,110] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:50,800] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:36:51,222] loop.py:380 INFO TIMING step=16885: step=3.387s data=1.309s compute=3.360s [fwd=2.395s bwd=0.951s opt=0.003s] dispatch=4.124s [2026-10-20 01:36:51,223] loop.py:45 INFO TRAIN_STEP step=16885 loss=1.204309 Training: 94%|█████████▎| 16885/18018 [36:06:37<2:21:03, 7.47s/step, loss=1.204, acc=0.714, acc_len=1.73, thru=16.4, I=19.9, T=37.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:51,315] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:57,136] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:36:57,289] loop.py:45 INFO TRAIN_STEP step=16886 loss=1.226300 Training: 94%|█████████▎| 16886/18018 [36:06:43<2:12:59, 7.05s/step, loss=1.226, acc=0.714, acc_len=1.73, thru=20.8, I=20.0, T=38.6, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:36:57,284] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16885: forward=2388.5ms backward=881.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:36:57,639] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:37:05,182] loop.py:45 INFO TRAIN_STEP step=16887 loss=1.273661 Training: 94%|█████████▎| 16887/18018 [36:06:51<2:17:39, 7.30s/step, loss=1.274, acc=0.700, acc_len=1.68, thru=14.4, I=17.2, T=36.9, wait=4.4s, pool=41, epoch=2/2][2026-10-20 01:37:12,999] loop.py:45 INFO TRAIN_STEP step=16888 loss=1.160911 Training: 94%|█████████▎| 16888/18018 [36:06:59<2:20:26, 7.46s/step, loss=1.161, acc=0.728, acc_len=1.78, thru=16.0, I=16.5, T=36.2, wait=4.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:37:20,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:37:21,424] loop.py:45 INFO TRAIN_STEP step=16889 loss=1.108595 Training: 94%|█████████▎| 16889/18018 [36:07:07<2:25:45, 7.75s/step, loss=1.109, acc=0.733, acc_len=1.80, thru=16.2, I=17.1, T=35.3, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:37:21,655] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:37:26,468] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:37:27,876] loop.py:380 INFO TIMING step=16890: step=3.898s data=1.773s compute=3.870s [fwd=2.729s bwd=1.127s opt=0.003s] dispatch=2.548s [2026-10-20 01:37:27,876] loop.py:45 INFO TRAIN_STEP step=16890 loss=1.323363 Training: 94%|█████████▎| 16890/18018 [36:07:13<2:18:20, 7.36s/step, loss=1.323, acc=0.695, acc_len=1.67, thru=19.6, I=25.8, T=32.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:37:27,977] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:37:33,889] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:37:33,968] loop.py:45 INFO TRAIN_STEP step=16891 loss=1.314715 Training: 94%|█████████▎| 16891/18018 [36:07:20<2:11:04, 6.98s/step, loss=1.315, acc=0.696, acc_len=1.66, thru=19.4, I=23.7, T=37.5, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:37:33,964] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16890: forward=2407.1ms backward=952.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:37:34,390] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:37:40,605] loop.py:45 INFO TRAIN_STEP step=16892 loss=1.303335 Training: 94%|█████████▍| 16892/18018 [36:07:26<2:09:02, 6.88s/step, loss=1.303, acc=0.700, acc_len=1.69, thru=18.2, I=18.3, T=37.8, wait=3.2s, pool=49, epoch=2/2][2026-10-20 01:37:47,861] loop.py:45 INFO TRAIN_STEP step=16893 loss=1.179422 Training: 94%|█████████▍| 16893/18018 [36:07:33<2:11:05, 6.99s/step, loss=1.179, acc=0.723, acc_len=1.76, thru=16.3, I=16.1, T=37.7, wait=3.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:37:55,246] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:37:55,754] loop.py:45 INFO TRAIN_STEP step=16894 loss=1.305551 Training: 94%|█████████▍| 16894/18018 [36:07:41<2:16:00, 7.26s/step, loss=1.306, acc=0.703, acc_len=1.69, thru=17.8, I=17.4, T=36.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:37:56,253] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:38:02,436] loop.py:380 INFO TIMING step=16895: step=3.318s data=1.330s compute=3.290s [fwd=2.358s bwd=0.918s opt=0.003s] dispatch=3.355s [2026-10-20 01:38:02,436] loop.py:45 INFO TRAIN_STEP step=16895 loss=1.278271 Training: 94%|█████████▍| 16895/18018 [36:07:48<2:12:38, 7.09s/step, loss=1.278, acc=0.705, acc_len=1.69, thru=17.9, I=19.4, T=38.6, wait=3.4s, pool=49, epoch=2/2][2026-10-20 01:38:10,996] loop.py:45 INFO TRAIN_STEP step=16896 loss=1.233609 Training: 94%|█████████▍| 16896/18018 [36:07:57<2:20:48, 7.53s/step, loss=1.234, acc=0.712, acc_len=1.72, thru=15.3, I=16.6, T=35.7, wait=5.0s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:38:10,991] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16895: forward=2529.2ms backward=1012.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:38:18,384] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:38:19,289] loop.py:45 INFO TRAIN_STEP step=16897 loss=1.190851 Training: 94%|█████████▍| 16897/18018 [36:08:05<2:24:57, 7.76s/step, loss=1.191, acc=0.721, acc_len=1.76, thru=16.1, I=16.7, T=36.4, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:38:19,400] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:38:27,841] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:38:27,983] loop.py:45 INFO TRAIN_STEP step=16898 loss=1.315883 Training: 94%|█████████▍| 16898/18018 [36:08:14<2:30:03, 8.04s/step, loss=1.316, acc=0.698, acc_len=1.68, thru=14.4, I=15.7, T=35.1, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:38:28,343] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:38:33,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:38:34,285] loop.py:45 INFO TRAIN_STEP step=16899 loss=1.312966 Training: 94%|█████████▍| 16899/18018 [36:08:20<2:20:12, 7.52s/step, loss=1.313, acc=0.697, acc_len=1.67, thru=19.6, I=21.6, T=35.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:38:34,359] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-20 01:38:41,191] loop.py:380 INFO TIMING step=16900: step=3.210s data=1.297s compute=3.184s [fwd=2.302s bwd=0.869s opt=0.003s] dispatch=3.685s [2026-10-20 01:38:41,191] loop.py:45 INFO TRAIN_STEP step=16900 loss=1.290961 Training: 94%|█████████▍| 16900/18018 [36:08:27<2:16:39, 7.33s/step, loss=1.291, acc=0.698, acc_len=1.68, thru=14.9, I=20.0, T=39.9, wait=3.7s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:38:51,795] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:38:51,969] loop.py:45 INFO TRAIN_STEP step=16901 loss=1.260817 Training: 94%|█████████▍| 16901/18018 [36:08:38<2:35:45, 8.37s/step, loss=1.261, acc=0.706, acc_len=1.70, thru=15.9, I=14.1, T=34.1, wait=7.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:38:51,964] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16900: forward=2678.7ms backward=1030.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:38:52,297] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:39:01,437] loop.py:45 INFO TRAIN_STEP step=16902 loss=1.240291 Training: 94%|█████████▍| 16902/18018 [36:08:47<2:41:47, 8.70s/step, loss=1.240, acc=0.713, acc_len=1.73, thru=13.3, I=13.1, T=32.5, wait=5.5s, pool=61, epoch=2/2][2026-10-20 01:39:08,714] loop.py:45 INFO TRAIN_STEP step=16903 loss=1.235241 Training: 94%|█████████▍| 16903/18018 [36:08:54<2:33:42, 8.27s/step, loss=1.235, acc=0.712, acc_len=1.72, thru=15.6, I=17.2, T=36.1, wait=3.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:15,693] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:39:17,321] loop.py:45 INFO TRAIN_STEP step=16904 loss=1.286940 Training: 94%|█████████▍| 16904/18018 [36:09:03<2:35:25, 8.37s/step, loss=1.287, acc=0.703, acc_len=1.69, thru=16.1, I=18.6, T=34.4, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:17,704] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 01:39:23,753] loop.py:380 INFO TIMING step=16905: step=3.530s data=1.560s compute=3.486s [fwd=2.449s bwd=1.022s opt=0.003s] dispatch=2.894s [2026-10-20 01:39:23,753] loop.py:45 INFO TRAIN_STEP step=16905 loss=1.252542 Training: 94%|█████████▍| 16905/18018 [36:09:09<2:24:31, 7.79s/step, loss=1.253, acc=0.711, acc_len=1.71, thru=16.7, I=27.1, T=36.3, wait=2.9s, pool=45, epoch=2/2][2026-10-20 01:39:31,822] loop.py:45 INFO TRAIN_STEP step=16906 loss=1.304777 (TrainerActor pid=9310) [2026-10-20 01:39:31,817] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16905: forward=2427.9ms backward=943.8ms Training: 94%|█████████▍| 16906/18018 [36:09:17<2:25:55, 7.87s/step, loss=1.305, acc=0.698, acc_len=1.67, thru=16.1, I=16.2, T=37.5, wait=4.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:38,699] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:39:39,955] loop.py:45 INFO TRAIN_STEP step=16907 loss=1.342761 Training: 94%|█████████▍| 16907/18018 [36:09:26<2:27:13, 7.95s/step, loss=1.343, acc=0.690, acc_len=1.64, thru=16.0, I=18.2, T=37.5, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:40,205] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:45,491] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:39:46,707] loop.py:45 INFO TRAIN_STEP step=16908 loss=1.270778 Training: 94%|█████████▍| 16908/18018 [36:09:32<2:20:26, 7.59s/step, loss=1.271, acc=0.706, acc_len=1.69, thru=18.8, I=24.0, T=35.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:47,000] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:52,265] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:39:53,222] loop.py:45 INFO TRAIN_STEP step=16909 loss=1.186067 Training: 94%|█████████▍| 16909/18018 [36:09:39<2:14:20, 7.27s/step, loss=1.186, acc=0.719, acc_len=1.75, thru=19.5, I=24.2, T=38.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:53,272] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:58,127] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:39:59,224] loop.py:380 INFO TIMING step=16910: step=3.574s data=1.526s compute=3.541s [fwd=2.531s bwd=0.995s opt=0.003s] dispatch=2.421s [2026-10-20 01:39:59,224] loop.py:45 INFO TRAIN_STEP step=16910 loss=1.250671 Training: 94%|█████████▍| 16910/18018 [36:09:45<2:07:12, 6.89s/step, loss=1.251, acc=0.709, acc_len=1.71, thru=21.0, I=26.2, T=35.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:39:59,634] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:04,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:40:05,638] loop.py:45 INFO TRAIN_STEP step=16911 loss=1.266468 Training: 94%|█████████▍| 16911/18018 [36:09:51<2:04:28, 6.75s/step, loss=1.266, acc=0.708, acc_len=1.71, thru=19.2, I=26.0, T=36.8, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:40:05,634] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16910: forward=2485.9ms backward=953.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:06,132] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:40:11,425] loop.py:45 INFO TRAIN_STEP step=16912 loss=1.229195 Training: 94%|█████████▍| 16912/18018 [36:09:57<1:59:02, 6.46s/step, loss=1.229, acc=0.712, acc_len=1.72, thru=20.5, I=23.4, T=36.7, wait=2.3s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:17,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:40:18,387] loop.py:45 INFO TRAIN_STEP step=16913 loss=1.179334 Training: 94%|█████████▍| 16913/18018 [36:10:04<2:01:43, 6.61s/step, loss=1.179, acc=0.722, acc_len=1.76, thru=18.9, I=20.1, T=36.8, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:18,609] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:23,640] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:40:24,463] loop.py:45 INFO TRAIN_STEP step=16914 loss=1.221975 Training: 94%|█████████▍| 16914/18018 [36:10:10<1:58:40, 6.45s/step, loss=1.222, acc=0.715, acc_len=1.74, thru=20.2, I=24.6, T=36.5, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:24,646] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:40:30,894] loop.py:380 INFO TIMING step=16915: step=3.454s data=1.406s compute=3.414s [fwd=2.466s bwd=0.934s opt=0.003s] dispatch=2.968s [2026-10-20 01:40:30,895] loop.py:45 INFO TRAIN_STEP step=16915 loss=1.349484 Training: 94%|█████████▍| 16915/18018 [36:10:16<1:58:28, 6.44s/step, loss=1.349, acc=0.693, acc_len=1.65, thru=17.6, I=19.3, T=37.1, wait=3.0s, pool=45, epoch=2/2][2026-10-20 01:40:38,012] loop.py:45 INFO TRAIN_STEP step=16916 loss=1.278877 Training: 94%|█████████▍| 16916/18018 [36:10:24<2:02:04, 6.65s/step, loss=1.279, acc=0.702, acc_len=1.69, thru=17.0, I=18.2, T=37.2, wait=3.7s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:40:38,007] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16915: forward=2466.0ms backward=935.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:44,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:40:45,074] loop.py:45 INFO TRAIN_STEP step=16917 loss=1.399090 Training: 94%|█████████▍| 16917/18018 [36:10:31<2:04:14, 6.77s/step, loss=1.399, acc=0.684, acc_len=1.62, thru=18.2, I=20.1, T=37.1, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:45,561] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:50,273] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:40:51,080] loop.py:45 INFO TRAIN_STEP step=16918 loss=1.227224 Training: 94%|█████████▍| 16918/18018 [36:10:37<1:59:55, 6.54s/step, loss=1.227, acc=0.714, acc_len=1.73, thru=20.4, I=24.5, T=38.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:51,280] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:58,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:40:58,603] loop.py:45 INFO TRAIN_STEP step=16919 loss=1.276506 Training: 94%|█████████▍| 16919/18018 [36:10:44<2:05:12, 6.84s/step, loss=1.277, acc=0.703, acc_len=1.69, thru=17.8, I=19.6, T=38.5, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:40:58,799] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:41:04,326] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:41:04,925] loop.py:380 INFO TIMING step=16920: step=3.451s data=1.443s compute=3.411s [fwd=2.373s bwd=1.024s opt=0.003s] dispatch=2.863s [2026-10-20 01:41:04,925] loop.py:45 INFO TRAIN_STEP step=16920 loss=1.080934 Training: 94%|█████████▍| 16920/18018 [36:10:51<2:02:16, 6.68s/step, loss=1.081, acc=0.738, acc_len=1.82, thru=20.2, I=20.5, T=37.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:41:05,332] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:41:11,466] loop.py:45 INFO TRAIN_STEP step=16921 loss=1.286217 Training: 94%|█████████▍| 16921/18018 [36:10:57<2:01:23, 6.64s/step, loss=1.286, acc=0.701, acc_len=1.68, thru=18.5, I=23.5, T=37.9, wait=3.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:41:11,461] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16920: forward=2408.0ms backward=928.1ms [2026-10-20 01:41:20,999] loop.py:45 INFO TRAIN_STEP step=16922 loss=1.295599 Training: 94%|█████████▍| 16922/18018 [36:11:07<2:17:08, 7.51s/step, loss=1.296, acc=0.700, acc_len=1.68, thru=12.9, I=12.5, T=34.2, wait=5.8s, pool=53, epoch=2/2][2026-10-20 01:41:30,215] loop.py:45 INFO TRAIN_STEP step=16923 loss=1.250069 Training: 94%|█████████▍| 16923/18018 [36:11:16<2:26:22, 8.02s/step, loss=1.250, acc=0.711, acc_len=1.71, thru=13.8, I=12.6, T=36.6, wait=5.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:41:39,734] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:41:40,185] loop.py:45 INFO TRAIN_STEP step=16924 loss=1.250830 Training: 94%|█████████▍| 16924/18018 [36:11:26<2:36:54, 8.61s/step, loss=1.251, acc=0.709, acc_len=1.71, thru=14.4, I=15.1, T=33.3, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:41:40,242] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-20 01:41:48,135] loop.py:380 INFO TIMING step=16925: step=3.606s data=1.415s compute=3.572s [fwd=2.515s bwd=1.043s opt=0.003s] dispatch=4.334s [2026-10-20 01:41:48,135] loop.py:45 INFO TRAIN_STEP step=16925 loss=1.333804 Training: 94%|█████████▍| 16925/18018 [36:11:34<2:33:10, 8.41s/step, loss=1.334, acc=0.695, acc_len=1.66, thru=12.7, I=15.4, T=35.5, wait=4.3s, pool=37, epoch=2/2][2026-10-20 01:41:56,808] loop.py:45 INFO TRAIN_STEP step=16926 loss=1.196445 (TrainerActor pid=9310) [2026-10-20 01:41:56,804] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16925: forward=2436.6ms backward=944.3ms Training: 94%|█████████▍| 16926/18018 [36:11:42<2:34:28, 8.49s/step, loss=1.196, acc=0.718, acc_len=1.74, thru=15.3, I=14.4, T=37.3, wait=5.2s, pool=45, epoch=2/2][2026-10-20 01:42:05,693] loop.py:45 INFO TRAIN_STEP step=16927 loss=1.233500 Training: 94%|█████████▍| 16927/18018 [36:11:51<2:36:30, 8.61s/step, loss=1.234, acc=0.711, acc_len=1.72, thru=14.7, I=14.3, T=33.5, wait=5.1s, pool=49, epoch=2/2][2026-10-20 01:42:14,212] loop.py:45 INFO TRAIN_STEP step=16928 loss=1.237289 Training: 94%|█████████▍| 16928/18018 [36:12:00<2:35:52, 8.58s/step, loss=1.237, acc=0.709, acc_len=1.72, thru=16.3, I=14.9, T=34.9, wait=4.8s, pool=57, epoch=2/2][2026-10-20 01:42:23,349] loop.py:45 INFO TRAIN_STEP step=16929 loss=1.216479 Training: 94%|█████████▍| 16929/18018 [36:12:09<2:38:47, 8.75s/step, loss=1.216, acc=0.715, acc_len=1.73, thru=14.1, I=14.8, T=35.0, wait=5.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:42:31,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:42:31,941] loop.py:380 INFO TIMING step=16930: step=3.590s data=1.465s compute=3.552s [fwd=2.538s bwd=0.999s opt=0.003s] dispatch=4.989s [2026-10-20 01:42:31,941] loop.py:45 INFO TRAIN_STEP step=16930 loss=1.210058 Training: 94%|█████████▍| 16930/18018 [36:12:18<2:37:46, 8.70s/step, loss=1.210, acc=0.716, acc_len=1.74, thru=16.6, I=15.6, T=35.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:42:32,194] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:42:37,087] loop.py:45 INFO TRAIN_STEP step=16931 loss=1.269150 (TrainerActor pid=9310) [2026-10-20 01:42:37,083] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16930: forward=2273.7ms backward=870.9ms Training: 94%|█████████▍| 16931/18018 [36:12:23<2:18:19, 7.63s/step, loss=1.269, acc=0.706, acc_len=1.70, thru=18.9, I=19.9, T=40.2, wait=2.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:42:44,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:42:46,527] loop.py:45 INFO TRAIN_STEP step=16932 loss=1.169134 Training: 94%|█████████▍| 16932/18018 [36:12:32<2:27:58, 8.18s/step, loss=1.169, acc=0.724, acc_len=1.77, thru=16.0, I=17.7, T=34.2, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:42:46,790] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 01:42:53,558] loop.py:45 INFO TRAIN_STEP step=16933 loss=1.256045 Training: 94%|█████████▍| 16933/18018 [36:12:39<2:21:38, 7.83s/step, loss=1.256, acc=0.706, acc_len=1.70, thru=16.5, I=21.5, T=36.3, wait=3.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:43:00,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:43:01,694] loop.py:45 INFO TRAIN_STEP step=16934 loss=1.281619 Training: 94%|█████████▍| 16934/18018 [36:12:47<2:23:08, 7.92s/step, loss=1.282, acc=0.704, acc_len=1.69, thru=16.9, I=18.6, T=34.0, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:43:01,738] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:43:09,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:43:09,486] loop.py:380 INFO TIMING step=16935: step=3.558s data=1.486s compute=3.533s [fwd=2.525s bwd=0.995s opt=0.003s] dispatch=4.226s [2026-10-20 01:43:09,486] loop.py:45 INFO TRAIN_STEP step=16935 loss=1.250142 Training: 94%|█████████▍| 16935/18018 [36:12:55<2:22:18, 7.88s/step, loss=1.250, acc=0.709, acc_len=1.71, thru=14.4, I=18.0, T=36.0, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:43:09,656] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:43:17,929] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:43:18,671] loop.py:45 INFO TRAIN_STEP step=16936 loss=1.275783 Training: 94%|█████████▍| 16936/18018 [36:13:04<2:29:13, 8.27s/step, loss=1.276, acc=0.700, acc_len=1.68, thru=14.0, I=15.3, T=33.4, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:43:18,666] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16935: forward=2483.4ms backward=1113.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:43:18,935] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:43:23,109] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:43:23,969] loop.py:45 INFO TRAIN_STEP step=16937 loss=1.289596 Training: 94%|█████████▍| 16937/18018 [36:13:10<2:12:59, 7.38s/step, loss=1.290, acc=0.698, acc_len=1.68, thru=21.3, I=24.9, T=38.7, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:43:24,115] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:43:30,397] loop.py:45 INFO TRAIN_STEP step=16938 loss=1.230037 Training: 94%|█████████▍| 16938/18018 [36:13:16<2:07:43, 7.10s/step, loss=1.230, acc=0.710, acc_len=1.71, thru=18.5, I=21.4, T=36.7, wait=2.9s, pool=45, epoch=2/2][2026-10-20 01:43:38,314] loop.py:45 INFO TRAIN_STEP step=16939 loss=1.198438 Training: 94%|█████████▍| 16939/18018 [36:13:24<2:12:01, 7.34s/step, loss=1.198, acc=0.718, acc_len=1.74, thru=16.7, I=15.6, T=34.6, wait=4.2s, pool=49, epoch=2/2][2026-10-20 01:43:46,389] loop.py:380 INFO TIMING step=16940: step=3.584s data=1.473s compute=3.543s [fwd=2.523s bwd=1.006s opt=0.003s] dispatch=4.483s [2026-10-20 01:43:46,389] loop.py:45 INFO TRAIN_STEP step=16940 loss=1.235913 Training: 94%|█████████▍| 16940/18018 [36:13:32<2:15:51, 7.56s/step, loss=1.236, acc=0.710, acc_len=1.72, thru=17.0, I=15.9, T=35.7, wait=4.5s, pool=61, epoch=2/2][2026-10-20 01:43:54,737] loop.py:45 INFO TRAIN_STEP step=16941 loss=1.217049 Training: 94%|█████████▍| 16941/18018 [36:13:40<2:19:57, 7.80s/step, loss=1.217, acc=0.710, acc_len=1.71, thru=15.8, I=16.3, T=35.6, wait=4.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:43:54,732] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16940: forward=2553.9ms backward=994.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:44:01,461] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:44:02,432] loop.py:45 INFO TRAIN_STEP step=16942 loss=1.076916 Training: 94%|█████████▍| 16942/18018 [36:13:48<2:19:17, 7.77s/step, loss=1.077, acc=0.737, acc_len=1.81, thru=16.6, I=18.3, T=36.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:44:02,478] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:44:07,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:44:07,963] loop.py:45 INFO TRAIN_STEP step=16943 loss=1.292212 Training: 94%|█████████▍| 16943/18018 [36:13:54<2:07:07, 7.10s/step, loss=1.292, acc=0.701, acc_len=1.69, thru=21.5, I=24.1, T=36.7, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:44:08,349] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:44:15,078] loop.py:45 INFO TRAIN_STEP step=16944 loss=1.306298 Training: 94%|█████████▍| 16944/18018 [36:14:01<2:07:07, 7.10s/step, loss=1.306, acc=0.696, acc_len=1.66, thru=15.9, I=16.4, T=36.9, wait=3.6s, pool=41, epoch=2/2][2026-10-20 01:44:22,711] loop.py:380 INFO TIMING step=16945: step=3.450s data=1.426s compute=3.408s [fwd=2.428s bwd=0.965s opt=0.003s] dispatch=4.174s [2026-10-20 01:44:22,712] loop.py:45 INFO TRAIN_STEP step=16945 loss=1.266160 Training: 94%|█████████▍| 16945/18018 [36:14:08<2:09:51, 7.26s/step, loss=1.266, acc=0.705, acc_len=1.70, thru=14.5, I=17.7, T=37.1, wait=4.2s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:44:29,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:44:30,220] loop.py:45 INFO TRAIN_STEP step=16946 loss=1.194090 Training: 94%|█████████▍| 16946/18018 [36:14:16<2:11:03, 7.34s/step, loss=1.194, acc=0.722, acc_len=1.75, thru=21.1, I=18.0, T=34.2, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:44:30,215] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16945: forward=2702.1ms backward=991.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:44:30,399] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:44:36,219] loop.py:45 INFO TRAIN_STEP step=16947 loss=1.310607 Training: 94%|█████████▍| 16947/18018 [36:14:22<2:03:47, 6.94s/step, loss=1.311, acc=0.698, acc_len=1.67, thru=17.4, I=20.6, T=37.7, wait=2.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:44:46,943] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:44:47,928] loop.py:45 INFO TRAIN_STEP step=16948 loss=1.419817 Training: 94%|█████████▍| 16948/18018 [36:14:34<2:29:12, 8.37s/step, loss=1.420, acc=0.675, acc_len=1.59, thru=12.0, I=13.1, T=33.8, wait=7.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:44:47,963] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-20 01:44:53,498] loop.py:45 INFO TRAIN_STEP step=16949 loss=1.267766 Training: 94%|█████████▍| 16949/18018 [36:14:39<2:14:07, 7.53s/step, loss=1.268, acc=0.708, acc_len=1.71, thru=18.6, I=16.5, T=38.2, wait=2.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:03,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:45:03,871] loop.py:380 INFO TIMING step=16950: step=3.643s data=1.525s compute=3.607s [fwd=2.552s bwd=1.040s opt=0.003s] dispatch=6.720s [2026-10-20 01:45:03,872] loop.py:45 INFO TRAIN_STEP step=16950 loss=1.294978 Training: 94%|█████████▍| 16950/18018 [36:14:49<2:29:11, 8.38s/step, loss=1.295, acc=0.700, acc_len=1.68, thru=14.3, I=13.7, T=35.1, wait=6.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:04,135] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:11,292] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:45:11,630] loop.py:45 INFO TRAIN_STEP step=16951 loss=1.312602 Training: 94%|█████████▍| 16951/18018 [36:14:57<2:25:43, 8.19s/step, loss=1.313, acc=0.697, acc_len=1.67, thru=16.4, I=17.3, T=35.7, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:45:11,625] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16950: forward=2463.3ms backward=1040.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:11,796] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:17,746] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:45:19,058] loop.py:45 INFO TRAIN_STEP step=16952 loss=1.237363 Training: 94%|█████████▍| 16952/18018 [36:15:05<2:21:29, 7.96s/step, loss=1.237, acc=0.709, acc_len=1.71, thru=18.1, I=19.3, T=35.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:19,258] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:26,470] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:45:27,147] loop.py:45 INFO TRAIN_STEP step=16953 loss=1.236894 Training: 94%|█████████▍| 16953/18018 [36:15:13<2:22:02, 8.00s/step, loss=1.237, acc=0.710, acc_len=1.72, thru=14.5, I=17.5, T=35.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:27,477] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:45:32,882] loop.py:45 INFO TRAIN_STEP step=16954 loss=1.234641 Training: 94%|█████████▍| 16954/18018 [36:15:18<2:09:49, 7.32s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=20.4, I=20.9, T=38.6, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:39,238] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:45:40,457] loop.py:380 INFO TIMING step=16955: step=3.554s data=1.459s compute=3.513s [fwd=2.510s bwd=0.989s opt=0.003s] dispatch=4.014s [2026-10-20 01:45:40,457] loop.py:45 INFO TRAIN_STEP step=16955 loss=1.278899 Training: 94%|█████████▍| 16955/18018 [36:15:26<2:11:04, 7.40s/step, loss=1.279, acc=0.704, acc_len=1.70, thru=17.6, I=20.0, T=36.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:45:40,749] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:45:46,270] loop.py:45 INFO TRAIN_STEP step=16956 loss=1.183857 Training: 94%|█████████▍| 16956/18018 [36:15:32<2:02:31, 6.92s/step, loss=1.184, acc=0.721, acc_len=1.75, thru=20.5, I=21.6, T=39.5, wait=2.6s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:45:46,265] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16955: forward=2320.7ms backward=880.9ms [2026-10-20 01:45:52,581] loop.py:45 INFO TRAIN_STEP step=16957 loss=1.293992 Training: 94%|█████████▍| 16957/18018 [36:15:38<1:59:09, 6.74s/step, loss=1.294, acc=0.702, acc_len=1.69, thru=20.1, I=22.3, T=38.6, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:46:00,784] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:46:01,547] loop.py:45 INFO TRAIN_STEP step=16958 loss=1.277972 Training: 94%|█████████▍| 16958/18018 [36:15:47<2:10:52, 7.41s/step, loss=1.278, acc=0.704, acc_len=1.69, thru=15.4, I=14.4, T=36.7, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:46:01,789] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:46:08,052] loop.py:45 INFO TRAIN_STEP step=16959 loss=1.205634 Training: 94%|█████████▍| 16959/18018 [36:15:54<2:05:57, 7.14s/step, loss=1.206, acc=0.717, acc_len=1.74, thru=17.4, I=22.0, T=37.5, wait=3.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:46:15,863] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:46:16,827] loop.py:380 INFO TIMING step=16960: step=3.543s data=1.401s compute=3.515s [fwd=2.528s bwd=0.973s opt=0.003s] dispatch=5.222s [2026-10-20 01:46:16,828] loop.py:45 INFO TRAIN_STEP step=16960 loss=1.292962 Training: 94%|█████████▍| 16960/18018 [36:16:02<2:14:30, 7.63s/step, loss=1.293, acc=0.702, acc_len=1.68, thru=16.8, I=15.7, T=36.1, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:46:16,883] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-20 01:46:22,494] loop.py:45 INFO TRAIN_STEP step=16961 loss=1.152417 Training: 94%|█████████▍| 16961/18018 [36:16:08<2:04:01, 7.04s/step, loss=1.152, acc=0.727, acc_len=1.78, thru=19.0, I=22.4, T=38.4, wait=2.3s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:46:22,489] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16960: forward=2349.7ms backward=930.9ms [2026-10-20 01:46:31,242] loop.py:45 INFO TRAIN_STEP step=16962 loss=1.250042 Training: 94%|█████████▍| 16962/18018 [36:16:17<2:12:54, 7.55s/step, loss=1.250, acc=0.710, acc_len=1.71, thru=14.8, I=14.9, T=32.6, wait=4.8s, pool=49, epoch=2/2][2026-10-20 01:46:39,657] loop.py:45 INFO TRAIN_STEP step=16963 loss=1.274789 Training: 94%|█████████▍| 16963/18018 [36:16:25<2:17:20, 7.81s/step, loss=1.275, acc=0.706, acc_len=1.70, thru=15.7, I=15.4, T=35.1, wait=4.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:46:44,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:46:45,535] loop.py:45 INFO TRAIN_STEP step=16964 loss=1.227733 Training: 94%|█████████▍| 16964/18018 [36:16:31<2:07:01, 7.23s/step, loss=1.228, acc=0.714, acc_len=1.72, thru=20.6, I=21.6, T=37.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:46:45,864] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:46:52,828] loop.py:380 INFO TIMING step=16965: step=3.370s data=1.364s compute=3.336s [fwd=2.393s bwd=0.929s opt=0.003s] dispatch=3.914s [2026-10-20 01:46:52,829] loop.py:45 INFO TRAIN_STEP step=16965 loss=1.175907 Training: 94%|█████████▍| 16965/18018 [36:16:38<2:07:14, 7.25s/step, loss=1.176, acc=0.719, acc_len=1.75, thru=17.2, I=20.0, T=38.0, wait=3.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:46:59,967] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16965: forward=2490.3ms backward=930.0ms [2026-10-20 01:46:59,971] loop.py:45 INFO TRAIN_STEP step=16966 loss=1.331072 Training: 94%|█████████▍| 16966/18018 [36:16:46<2:06:32, 7.22s/step, loss=1.331, acc=0.695, acc_len=1.67, thru=17.1, I=20.1, T=36.9, wait=3.7s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:05,827] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:47:06,491] loop.py:45 INFO TRAIN_STEP step=16967 loss=1.172815 Training: 94%|█████████▍| 16967/18018 [36:16:52<2:02:45, 7.01s/step, loss=1.173, acc=0.723, acc_len=1.76, thru=19.3, I=21.0, T=37.8, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:06,834] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:13,833] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:47:13,888] loop.py:45 INFO TRAIN_STEP step=16968 loss=1.205613 Training: 94%|█████████▍| 16968/18018 [36:16:59<2:04:40, 7.12s/step, loss=1.206, acc=0.719, acc_len=1.75, thru=17.9, I=19.7, T=37.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:14,334] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:47:22,303] loop.py:45 INFO TRAIN_STEP step=16969 loss=1.308618 Training: 94%|█████████▍| 16969/18018 [36:17:08<2:11:19, 7.51s/step, loss=1.309, acc=0.697, acc_len=1.67, thru=14.1, I=14.5, T=35.2, wait=4.8s, pool=49, epoch=2/2][2026-10-20 01:47:28,997] loop.py:380 INFO TIMING step=16970: step=3.303s data=1.361s compute=3.260s [fwd=2.322s bwd=0.924s opt=0.003s] dispatch=3.385s [2026-10-20 01:47:28,997] loop.py:45 INFO TRAIN_STEP step=16970 loss=1.199158 Training: 94%|█████████▍| 16970/18018 [36:17:15<2:06:55, 7.27s/step, loss=1.199, acc=0.721, acc_len=1.75, thru=18.6, I=16.9, T=38.8, wait=3.4s, pool=53, epoch=2/2][2026-10-20 01:47:37,165] loop.py:45 INFO TRAIN_STEP step=16971 loss=1.249449 Training: 94%|█████████▍| 16971/18018 [36:17:23<2:11:31, 7.54s/step, loss=1.249, acc=0.710, acc_len=1.72, thru=17.3, I=16.7, T=34.9, wait=4.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:47:37,161] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16970: forward=2612.1ms backward=1007.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:42,903] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:47:43,009] loop.py:45 INFO TRAIN_STEP step=16972 loss=1.155544 Training: 94%|█████████▍| 16972/18018 [36:17:29<2:02:32, 7.03s/step, loss=1.156, acc=0.725, acc_len=1.78, thru=20.8, I=19.8, T=37.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:43,406] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:48,931] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:47:49,488] loop.py:45 INFO TRAIN_STEP step=16973 loss=1.316454 Training: 94%|█████████▍| 16973/18018 [36:17:35<1:59:32, 6.86s/step, loss=1.316, acc=0.697, acc_len=1.67, thru=19.8, I=22.4, T=36.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:49,936] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:54,200] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:47:55,368] loop.py:45 INFO TRAIN_STEP step=16974 loss=1.248991 Training: 94%|█████████▍| 16974/18018 [36:17:41<1:54:17, 6.57s/step, loss=1.249, acc=0.708, acc_len=1.70, thru=21.3, I=26.4, T=38.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:47:55,707] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:48:02,133] loop.py:380 INFO TIMING step=16975: step=3.421s data=1.402s compute=3.385s [fwd=2.442s bwd=0.928s opt=0.003s] dispatch=3.336s [2026-10-20 01:48:02,133] loop.py:45 INFO TRAIN_STEP step=16975 loss=1.253953 Training: 94%|█████████▍| 16975/18018 [36:17:48<1:55:13, 6.63s/step, loss=1.254, acc=0.712, acc_len=1.72, thru=18.1, I=20.7, T=37.4, wait=3.3s, pool=53, epoch=2/2][2026-10-20 01:48:09,634] loop.py:45 INFO TRAIN_STEP step=16976 loss=1.225125 (TrainerActor pid=9310) [2026-10-20 01:48:09,629] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16975: forward=2426.8ms backward=950.1ms Training: 94%|█████████▍| 16976/18018 [36:17:55<1:59:38, 6.89s/step, loss=1.225, acc=0.713, acc_len=1.73, thru=17.7, I=18.3, T=37.4, wait=4.1s, pool=57, epoch=2/2][2026-10-20 01:48:18,479] loop.py:45 INFO TRAIN_STEP step=16977 loss=1.160493 Training: 94%|█████████▍| 16977/18018 [36:18:04<2:09:42, 7.48s/step, loss=1.160, acc=0.723, acc_len=1.76, thru=14.4, I=14.4, T=32.2, wait=4.9s, pool=53, epoch=2/2][2026-10-20 01:48:27,438] loop.py:45 INFO TRAIN_STEP step=16978 loss=1.155470 Training: 94%|█████████▍| 16978/18018 [36:18:13<2:17:17, 7.92s/step, loss=1.155, acc=0.725, acc_len=1.77, thru=15.0, I=13.3, T=37.1, wait=5.5s, pool=53, epoch=2/2][2026-10-20 01:48:35,573] loop.py:45 INFO TRAIN_STEP step=16979 loss=1.169466 Training: 94%|█████████▍| 16979/18018 [36:18:21<2:18:16, 7.99s/step, loss=1.169, acc=0.723, acc_len=1.76, thru=14.9, I=14.4, T=35.2, wait=4.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:48:44,210] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:48:44,323] loop.py:380 INFO TIMING step=16980: step=3.699s data=1.508s compute=3.671s [fwd=2.664s bwd=0.992s opt=0.003s] dispatch=5.043s [2026-10-20 01:48:44,323] loop.py:45 INFO TRAIN_STEP step=16980 loss=1.389475 Training: 94%|█████████▍| 16980/18018 [36:18:30<2:22:06, 8.21s/step, loss=1.389, acc=0.687, acc_len=1.63, thru=17.9, I=16.0, T=34.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:48:44,710] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:48:50,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:48:50,386] loop.py:45 INFO TRAIN_STEP step=16981 loss=1.246565 Training: 94%|█████████▍| 16981/18018 [36:18:36<2:10:48, 7.57s/step, loss=1.247, acc=0.705, acc_len=1.70, thru=20.5, I=20.9, T=37.6, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:48:50,382] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16980: forward=2401.4ms backward=944.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:48:50,811] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:48:57,330] loop.py:45 INFO TRAIN_STEP step=16982 loss=1.086164 Training: 94%|█████████▍| 16982/18018 [36:18:43<2:07:28, 7.38s/step, loss=1.086, acc=0.735, acc_len=1.80, thru=17.2, I=18.5, T=38.1, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:49:03,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:49:04,550] loop.py:45 INFO TRAIN_STEP step=16983 loss=1.235534 Training: 94%|█████████▍| 16983/18018 [36:18:50<2:06:30, 7.33s/step, loss=1.236, acc=0.712, acc_len=1.72, thru=19.0, I=18.9, T=35.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:49:04,902] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:49:11,346] loop.py:45 INFO TRAIN_STEP step=16984 loss=1.270347 Training: 94%|█████████▍| 16984/18018 [36:18:57<2:03:36, 7.17s/step, loss=1.270, acc=0.701, acc_len=1.68, thru=15.3, I=16.3, T=36.7, wait=3.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:49:19,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:49:20,571] loop.py:380 INFO TIMING step=16985: step=4.163s data=1.898s compute=4.113s [fwd=2.865s bwd=1.232s opt=0.003s] dispatch=5.052s [2026-10-20 01:49:20,571] loop.py:45 INFO TRAIN_STEP step=16985 loss=1.267712 Training: 94%|█████████▍| 16985/18018 [36:19:06<2:14:05, 7.79s/step, loss=1.268, acc=0.706, acc_len=1.70, thru=17.5, I=17.7, T=30.7, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:49:20,682] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-20 01:49:27,910] loop.py:45 INFO TRAIN_STEP step=16986 loss=1.116954 Training: 94%|█████████▍| 16986/18018 [36:19:13<2:11:38, 7.65s/step, loss=1.117, acc=0.734, acc_len=1.80, thru=12.9, I=15.1, T=38.0, wait=4.0s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:49:27,906] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16985: forward=2370.0ms backward=940.8ms (TrainerActor pid=9370) [2026-10-20 01:49:27,970] eagle_store.py:517 WARNING batch_get_buffer missing keys (k543558-aaaa0fd6_hs, k543558-aaaa0fd6_ids, k543558-aaaa0fd6_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:49:36,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 01:49:27,972] eagle_store.py:517 WARNING batch_get_buffer missing keys (k543559-a0f20086_hs, k543559-a0f20086_ids, k543559-a0f20086_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-20 01:49:36,630] loop.py:45 INFO TRAIN_STEP step=16987 loss=1.317732 Training: 94%|█████████▍| 16987/18018 [36:19:22<2:17:00, 7.97s/step, loss=1.318, acc=0.698, acc_len=1.67, thru=18.2, I=18.5, T=32.5, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:49:37,124] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:49:44,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:49:45,129] loop.py:45 INFO TRAIN_STEP step=16988 loss=1.285566 Training: 94%|█████████▍| 16988/18018 [36:19:31<2:19:34, 8.13s/step, loss=1.286, acc=0.703, acc_len=1.69, thru=14.5, I=15.2, T=37.1, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:49:45,516] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:49:51,082] loop.py:45 INFO TRAIN_STEP step=16989 loss=1.444587 Training: 94%|█████████▍| 16989/18018 [36:19:37<2:08:13, 7.48s/step, loss=1.445, acc=0.672, acc_len=1.59, thru=19.6, I=21.6, T=38.7, wait=2.6s, pool=49, epoch=2/2][2026-10-20 01:50:00,104] loop.py:380 INFO TIMING step=16990: step=3.630s data=1.533s compute=3.602s [fwd=2.590s bwd=0.998s opt=0.003s] dispatch=5.384s [2026-10-20 01:50:00,104] loop.py:45 INFO TRAIN_STEP step=16990 loss=1.241642 Training: 94%|█████████▍| 16990/18018 [36:19:46<2:16:02, 7.94s/step, loss=1.242, acc=0.713, acc_len=1.73, thru=15.3, I=15.5, T=35.3, wait=5.4s, pool=57, epoch=2/2][2026-10-20 01:50:07,720] loop.py:45 INFO TRAIN_STEP step=16991 loss=1.180237 Training: 94%|█████████▍| 16991/18018 [36:19:53<2:14:15, 7.84s/step, loss=1.180, acc=0.722, acc_len=1.75, thru=13.0, I=18.2, T=36.2, wait=4.1s, pool=13, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:50:07,716] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16990: forward=2505.9ms backward=984.7ms (TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 01:50:08,718] eagle_store.py:517 WARNING batch_get_buffer missing keys (k543706-9f01a7d7_hs, k543706-9f01a7d7_ids, k543706-9f01a7d7_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:16,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-20 01:50:09,158] eagle_store.py:517 WARNING batch_get_buffer missing keys (k543712-a8e8b0e8_hs, k543712-a8e8b0e8_ids, k543712-a8e8b0e8_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-20 01:50:17,028] loop.py:45 INFO TRAIN_STEP step=16992 loss=1.153365 Training: 94%|█████████▍| 16992/18018 [36:20:03<2:21:38, 8.28s/step, loss=1.153, acc=0.728, acc_len=1.77, thru=21.6, I=15.0, T=35.2, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:17,122] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:22,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:50:23,201] loop.py:45 INFO TRAIN_STEP step=16993 loss=1.151626 Training: 94%|█████████▍| 16993/18018 [36:20:09<2:10:41, 7.65s/step, loss=1.152, acc=0.728, acc_len=1.78, thru=20.1, I=23.0, T=38.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:23,262] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:28,848] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:50:29,252] loop.py:45 INFO TRAIN_STEP step=16994 loss=1.305177 Training: 94%|█████████▍| 16994/18018 [36:20:15<2:02:21, 7.17s/step, loss=1.305, acc=0.699, acc_len=1.67, thru=20.4, I=23.0, T=34.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:29,354] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:35,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:50:36,964] loop.py:380 INFO TIMING step=16995: step=4.187s data=2.065s compute=4.145s [fwd=3.122s bwd=1.008s opt=0.003s] dispatch=3.516s [2026-10-20 01:50:36,964] loop.py:45 INFO TRAIN_STEP step=16995 loss=1.326715 Training: 94%|█████████▍| 16995/18018 [36:20:23<2:05:01, 7.33s/step, loss=1.327, acc=0.694, acc_len=1.66, thru=17.0, I=20.1, T=30.6, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:37,064] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:43,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:50:43,446] loop.py:45 INFO TRAIN_STEP step=16996 loss=1.323999 Training: 94%|█████████▍| 16996/18018 [36:20:29<2:00:33, 7.08s/step, loss=1.324, acc=0.695, acc_len=1.67, thru=18.6, I=23.5, T=29.7, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:50:43,442] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=16995: forward=3362.6ms backward=899.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:43,822] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:50:50,971] loop.py:45 INFO TRAIN_STEP step=16997 loss=1.335884 Training: 94%|█████████▍| 16997/18018 [36:20:37<2:02:42, 7.21s/step, loss=1.336, acc=0.694, acc_len=1.65, thru=14.8, I=17.4, T=29.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:50,972] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:51,474] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:58,009] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:50:58,227] loop.py:45 INFO TRAIN_STEP step=16998 loss=1.223434 Training: 94%|█████████▍| 16998/18018 [36:20:44<2:02:49, 7.23s/step, loss=1.223, acc=0.714, acc_len=1.73, thru=16.8, I=17.3, T=33.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:50:58,511] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:51:04,962] loop.py:45 INFO TRAIN_STEP step=16999 loss=1.216230 Training: 94%|█████████▍| 16999/18018 [36:20:51<2:00:13, 7.08s/step, loss=1.216, acc=0.716, acc_len=1.73, thru=14.1, I=16.8, T=36.4, wait=3.2s, pool=29, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 01:51:05,048] eagle_store.py:517 WARNING batch_get_buffer missing keys (k543970-a3cad450_hs, k543970-a3cad450_ids, k543970-a3cad450_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:51:12,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-20 01:51:05,027] eagle_store.py:517 WARNING batch_get_buffer missing keys (k543969-bec84b27_hs, k543969-bec84b27_ids, k543969-bec84b27_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-20 01:51:14,090] loop.py:380 INFO TIMING step=17000: step=3.803s data=1.485s compute=3.771s [fwd=2.625s bwd=1.130s opt=0.003s] dispatch=5.313s [2026-10-20 01:51:14,090] loop.py:45 INFO TRAIN_STEP step=17000 loss=1.394075 Training: 94%|█████████▍| 17000/18018 [36:21:00<2:10:31, 7.69s/step, loss=1.394, acc=0.683, acc_len=1.62, thru=18.2, I=17.7, T=33.7, wait=5.3s, pool=65, epoch=2/2][2026-10-20 01:51:14,094] loop.py:419 INFO Saving checkpoint at step 17000... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:51:17,407] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 01:51:18,172] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0017001 (TrainerActor pid=9310) [2026-10-20 01:51:18,341] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0016251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:51:19,413] inference_manager.py:408 INFO Pool capacity available after 7.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:51:24,297] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 01:51:18,173] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0017001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 01:51:19,194] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0016251 [2026-10-20 01:51:25,448] loop.py:45 INFO TRAIN_STEP step=17001 loss=1.294674 Training: 94%|█████████▍| 17001/18018 [36:21:11<2:29:02, 8.79s/step, loss=1.295, acc=0.699, acc_len=1.67, thru=21.2, I=26.9, T=39.1, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:51:25,444] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17000: forward=2314.1ms backward=908.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:51:25,807] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:51:30,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:51:31,659] loop.py:45 INFO TRAIN_STEP step=17002 loss=1.231119 Training: 94%|█████████▍| 17002/18018 [36:21:17<2:15:46, 8.02s/step, loss=1.231, acc=0.713, acc_len=1.73, thru=20.8, I=27.1, T=38.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:51:32,068] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:51:38,628] loop.py:45 INFO TRAIN_STEP step=17003 loss=1.292269 Training: 94%|█████████▍| 17003/18018 [36:21:24<2:10:18, 7.70s/step, loss=1.292, acc=0.702, acc_len=1.68, thru=16.0, I=22.0, T=38.4, wait=3.6s, pool=37, epoch=2/2][2026-10-20 01:51:47,341] loop.py:45 INFO TRAIN_STEP step=17004 loss=1.219644 Training: 94%|█████████▍| 17004/18018 [36:21:33<2:15:17, 8.01s/step, loss=1.220, acc=0.714, acc_len=1.73, thru=16.2, I=13.7, T=36.5, wait=5.2s, pool=49, epoch=2/2][2026-10-20 01:51:55,680] loop.py:380 INFO TIMING step=17005: step=3.872s data=1.566s compute=3.843s [fwd=2.667s bwd=1.162s opt=0.003s] dispatch=4.458s [2026-10-20 01:51:55,680] loop.py:45 INFO TRAIN_STEP step=17005 loss=1.274615 Training: 94%|█████████▍| 17005/18018 [36:21:41<2:16:51, 8.11s/step, loss=1.275, acc=0.707, acc_len=1.71, thru=16.0, I=16.6, T=33.1, wait=4.5s, pool=49, epoch=2/2](TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 01:51:55,737] eagle_store.py:517 WARNING batch_get_buffer missing keys (k544163-8932a88c_hs, k544163-8932a88c_ids, k544163-8932a88c_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:05,839] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9370) [2026-10-20 01:51:55,740] eagle_store.py:517 WARNING batch_get_buffer missing keys (k544162-a1a02b99_hs, k544162-a1a02b99_ids, k544162-a1a02b99_lhs); sleeping 0.50s. [2026-10-20 01:52:08,102] loop.py:45 INFO TRAIN_STEP step=17006 loss=1.287696 Training: 94%|█████████▍| 17006/18018 [36:21:54<2:38:33, 9.40s/step, loss=1.288, acc=0.699, acc_len=1.68, thru=14.7, I=12.6, T=27.5, wait=7.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:52:08,097] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17005: forward=3494.6ms backward=1113.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:08,351] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-20 01:52:17,195] loop.py:45 INFO TRAIN_STEP step=17007 loss=1.150088 Training: 94%|█████████▍| 17007/18018 [36:22:03<2:36:51, 9.31s/step, loss=1.150, acc=0.726, acc_len=1.77, thru=13.1, I=14.1, T=36.8, wait=5.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:26,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:52:27,990] loop.py:45 INFO TRAIN_STEP step=17008 loss=1.204343 Training: 94%|█████████▍| 17008/18018 [36:22:14<2:44:11, 9.75s/step, loss=1.204, acc=0.716, acc_len=1.74, thru=13.4, I=13.4, T=31.8, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:28,471] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:32,811] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:52:33,775] loop.py:45 INFO TRAIN_STEP step=17009 loss=1.293180 Training: 94%|█████████▍| 17009/18018 [36:22:19<2:24:01, 8.56s/step, loss=1.293, acc=0.700, acc_len=1.68, thru=20.5, I=24.2, T=37.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:33,827] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-20 01:52:40,227] loop.py:380 INFO TIMING step=17010: step=3.414s data=1.419s compute=3.384s [fwd=2.455s bwd=0.915s opt=0.003s] dispatch=3.028s [2026-10-20 01:52:40,227] loop.py:45 INFO TRAIN_STEP step=17010 loss=1.275687 Training: 94%|█████████▍| 17010/18018 [36:22:26<2:13:13, 7.93s/step, loss=1.276, acc=0.703, acc_len=1.69, thru=18.9, I=21.8, T=37.5, wait=3.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:46,284] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:52:46,865] loop.py:45 INFO TRAIN_STEP step=17011 loss=1.236746 Training: 94%|█████████▍| 17011/18018 [36:22:32<2:06:34, 7.54s/step, loss=1.237, acc=0.709, acc_len=1.71, thru=19.5, I=20.6, T=35.0, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:52:46,860] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17010: forward=2622.3ms backward=990.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:47,288] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:52,511] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:52:52,570] loop.py:45 INFO TRAIN_STEP step=17012 loss=1.202527 Training: 94%|█████████▍| 17012/18018 [36:22:38<1:57:13, 6.99s/step, loss=1.203, acc=0.720, acc_len=1.75, thru=21.3, I=22.1, T=37.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:53,013] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:58,229] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:52:58,782] loop.py:45 INFO TRAIN_STEP step=17013 loss=1.283118 Training: 94%|█████████▍| 17013/18018 [36:22:44<1:53:11, 6.76s/step, loss=1.283, acc=0.699, acc_len=1.68, thru=20.5, I=21.8, T=37.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:52:59,233] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:03,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:53:04,238] loop.py:45 INFO TRAIN_STEP step=17014 loss=1.288050 Training: 94%|█████████▍| 17014/18018 [36:22:50<1:46:32, 6.37s/step, loss=1.288, acc=0.699, acc_len=1.68, thru=22.6, I=25.2, T=38.0, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:04,737] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:09,810] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:53:10,627] loop.py:380 INFO TIMING step=17015: step=3.268s data=1.292s compute=3.233s [fwd=2.321s bwd=0.898s opt=0.003s] dispatch=3.113s [2026-10-20 01:53:10,627] loop.py:45 INFO TRAIN_STEP step=17015 loss=1.257900 Training: 94%|█████████▍| 17015/18018 [36:22:56<1:46:33, 6.37s/step, loss=1.258, acc=0.706, acc_len=1.69, thru=20.4, I=24.9, T=39.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:10,814] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:19,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:53:19,760] loop.py:45 INFO TRAIN_STEP step=17016 loss=1.249225 Training: 94%|█████████▍| 17016/18018 [36:23:05<2:00:15, 7.20s/step, loss=1.249, acc=0.710, acc_len=1.71, thru=13.3, I=14.9, T=33.5, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:53:19,756] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17015: forward=2656.5ms backward=1118.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:19,925] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:24,150] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:53:25,507] loop.py:45 INFO TRAIN_STEP step=17017 loss=1.230022 Training: 94%|█████████▍| 17017/18018 [36:23:11<1:52:52, 6.77s/step, loss=1.230, acc=0.712, acc_len=1.72, thru=20.1, I=21.2, T=36.1, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:25,671] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 01:53:32,471] loop.py:45 INFO TRAIN_STEP step=17018 loss=1.155926 Training: 94%|█████████▍| 17018/18018 [36:23:18<1:53:44, 6.82s/step, loss=1.156, acc=0.726, acc_len=1.77, thru=14.8, I=19.2, T=37.6, wait=3.6s, pool=41, epoch=2/2][2026-10-20 01:53:40,200] loop.py:45 INFO TRAIN_STEP step=17019 loss=1.268401 Training: 94%|█████████▍| 17019/18018 [36:23:26<1:58:09, 7.10s/step, loss=1.268, acc=0.705, acc_len=1.70, thru=16.4, I=16.6, T=35.5, wait=4.1s, pool=53, epoch=2/2][2026-10-20 01:53:46,120] loop.py:380 INFO TIMING step=17020: step=3.410s data=1.408s compute=3.384s [fwd=2.402s bwd=0.968s opt=0.003s] dispatch=2.500s [2026-10-20 01:53:46,120] loop.py:45 INFO TRAIN_STEP step=17020 loss=1.218244 Training: 94%|█████████▍| 17020/18018 [36:23:32<1:52:09, 6.74s/step, loss=1.218, acc=0.713, acc_len=1.72, thru=16.0, I=17.6, T=37.5, wait=2.5s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:56,349] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:53:56,836] loop.py:45 INFO TRAIN_STEP step=17021 loss=1.350797 Training: 94%|█████████▍| 17021/18018 [36:23:42<2:11:51, 7.94s/step, loss=1.351, acc=0.690, acc_len=1.65, thru=16.2, I=15.3, T=34.7, wait=7.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:53:56,831] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17020: forward=2637.8ms backward=999.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:53:56,868] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:54:03,514] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:54:03,689] loop.py:45 INFO TRAIN_STEP step=17022 loss=1.211197 Training: 94%|█████████▍| 17022/18018 [36:23:49<2:06:20, 7.61s/step, loss=1.211, acc=0.718, acc_len=1.75, thru=18.0, I=18.6, T=37.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:54:04,016] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:54:11,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:54:12,648] loop.py:45 INFO TRAIN_STEP step=17023 loss=1.215781 Training: 94%|█████████▍| 17023/18018 [36:23:58<2:12:54, 8.01s/step, loss=1.216, acc=0.714, acc_len=1.73, thru=14.8, I=17.6, T=34.8, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:54:12,767] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-20 01:54:19,325] loop.py:45 INFO TRAIN_STEP step=17024 loss=1.274672 Training: 94%|█████████▍| 17024/18018 [36:24:05<2:06:07, 7.61s/step, loss=1.275, acc=0.705, acc_len=1.69, thru=16.8, I=23.5, T=37.7, wait=3.3s, pool=49, epoch=2/2][2026-10-20 01:54:29,450] loop.py:380 INFO TIMING step=17025: step=3.701s data=1.482s compute=3.677s [fwd=2.631s bwd=1.031s opt=0.003s] dispatch=6.415s [2026-10-20 01:54:29,451] loop.py:45 INFO TRAIN_STEP step=17025 loss=1.263798 Training: 94%|█████████▍| 17025/18018 [36:24:15<2:18:28, 8.37s/step, loss=1.264, acc=0.709, acc_len=1.71, thru=13.4, I=12.3, T=34.6, wait=6.4s, pool=57, epoch=2/2][2026-10-20 01:54:36,526] loop.py:45 INFO TRAIN_STEP step=17026 loss=1.249525 Training: 94%|█████████▍| 17026/18018 [36:24:22<2:11:55, 7.98s/step, loss=1.250, acc=0.707, acc_len=1.71, thru=18.1, I=17.9, T=35.9, wait=3.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:54:36,522] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17025: forward=2506.3ms backward=1001.0ms [2026-10-20 01:54:42,498] loop.py:45 INFO TRAIN_STEP step=17027 loss=1.311018 Training: 94%|█████████▍| 17027/18018 [36:24:28<2:01:50, 7.38s/step, loss=1.311, acc=0.698, acc_len=1.66, thru=20.9, I=21.0, T=36.2, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:54:42,496] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:54:42,997] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:54:48,981] loop.py:45 INFO TRAIN_STEP step=17028 loss=1.243612 Training: 95%|█████████▍| 17028/18018 [36:24:35<1:57:17, 7.11s/step, loss=1.244, acc=0.708, acc_len=1.71, thru=17.4, I=18.5, T=36.7, wait=3.0s, pool=49, epoch=2/2][2026-10-20 01:54:55,737] loop.py:45 INFO TRAIN_STEP step=17029 loss=1.299144 Training: 95%|█████████▍| 17029/18018 [36:24:41<1:55:27, 7.00s/step, loss=1.299, acc=0.700, acc_len=1.68, thru=18.0, I=19.2, T=36.6, wait=3.2s, pool=53, epoch=2/2][2026-10-20 01:55:05,362] loop.py:380 INFO TIMING step=17030: step=3.806s data=1.598s compute=3.769s [fwd=2.626s bwd=1.130s opt=0.003s] dispatch=5.807s [2026-10-20 01:55:05,362] loop.py:45 INFO TRAIN_STEP step=17030 loss=1.233642 Training: 95%|█████████▍| 17030/18018 [36:24:51<2:08:16, 7.79s/step, loss=1.234, acc=0.710, acc_len=1.71, thru=14.1, I=14.1, T=33.6, wait=5.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:12,278] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:55:13,932] loop.py:45 INFO TRAIN_STEP step=17031 loss=1.292786 Training: 95%|█████████▍| 17031/18018 [36:25:00<2:11:59, 8.02s/step, loss=1.293, acc=0.701, acc_len=1.69, thru=16.3, I=17.9, T=35.2, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:55:13,928] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17030: forward=2497.5ms backward=1077.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:14,291] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:18,262] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:55:19,060] loop.py:45 INFO TRAIN_STEP step=17032 loss=1.154312 Training: 95%|█████████▍| 17032/18018 [36:25:05<1:57:34, 7.16s/step, loss=1.154, acc=0.726, acc_len=1.77, thru=22.8, I=31.3, T=38.8, wait=1.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:19,267] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:24,086] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:55:25,015] loop.py:45 INFO TRAIN_STEP step=17033 loss=1.215613 Training: 95%|█████████▍| 17033/18018 [36:25:11<1:51:32, 6.79s/step, loss=1.216, acc=0.710, acc_len=1.72, thru=22.8, I=26.1, T=38.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:25,099] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-20 01:55:31,279] loop.py:45 INFO TRAIN_STEP step=17034 loss=1.242669 Training: 95%|█████████▍| 17034/18018 [36:25:17<1:48:49, 6.64s/step, loss=1.243, acc=0.709, acc_len=1.71, thru=17.6, I=20.9, T=36.6, wait=2.8s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:40,844] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:55:41,432] loop.py:380 INFO TIMING step=17035: step=3.702s data=1.458s compute=3.632s [fwd=2.557s bwd=1.060s opt=0.003s] dispatch=6.440s [2026-10-20 01:55:41,432] loop.py:45 INFO TRAIN_STEP step=17035 loss=1.372267 Training: 95%|█████████▍| 17035/18018 [36:25:27<2:06:01, 7.69s/step, loss=1.372, acc=0.687, acc_len=1.64, thru=16.0, I=14.9, T=34.6, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:41,849] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:47,455] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:55:47,808] loop.py:45 INFO TRAIN_STEP step=17036 loss=1.257391 Training: 95%|█████████▍| 17036/18018 [36:25:33<1:59:25, 7.30s/step, loss=1.257, acc=0.705, acc_len=1.70, thru=19.6, I=20.1, T=37.7, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:55:47,803] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17035: forward=2384.3ms backward=959.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:55:47,957] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-20 01:55:54,618] loop.py:45 INFO TRAIN_STEP step=17037 loss=1.281946 Training: 95%|█████████▍| 17037/18018 [36:25:40<1:56:53, 7.15s/step, loss=1.282, acc=0.702, acc_len=1.69, thru=18.4, I=18.2, T=37.0, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:00,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:56:01,273] loop.py:45 INFO TRAIN_STEP step=17038 loss=1.252715 Training: 95%|█████████▍| 17038/18018 [36:25:47<1:54:22, 7.00s/step, loss=1.253, acc=0.707, acc_len=1.71, thru=21.1, I=20.4, T=37.0, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:01,371] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:09,226] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:56:09,485] loop.py:45 INFO TRAIN_STEP step=17039 loss=1.197048 Training: 95%|█████████▍| 17039/18018 [36:25:55<2:00:10, 7.36s/step, loss=1.197, acc=0.716, acc_len=1.73, thru=14.8, I=17.4, T=33.8, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:09,728] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:56:16,164] loop.py:380 INFO TIMING step=17040: step=3.419s data=1.394s compute=3.383s [fwd=2.419s bwd=0.949s opt=0.003s] dispatch=3.251s [2026-10-20 01:56:16,164] loop.py:45 INFO TRAIN_STEP step=17040 loss=1.329821 Training: 95%|█████████▍| 17040/18018 [36:26:02<1:56:41, 7.16s/step, loss=1.330, acc=0.694, acc_len=1.66, thru=17.2, I=17.6, T=37.4, wait=3.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:23,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:56:24,111] loop.py:45 INFO TRAIN_STEP step=17041 loss=1.285142 Training: 95%|█████████▍| 17041/18018 [36:26:10<2:00:25, 7.40s/step, loss=1.285, acc=0.704, acc_len=1.70, thru=19.1, I=18.7, T=36.1, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:56:24,106] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17040: forward=2469.7ms backward=1029.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:24,383] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:30,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:56:30,190] loop.py:45 INFO TRAIN_STEP step=17042 loss=1.204128 Training: 95%|█████████▍| 17042/18018 [36:26:16<1:53:52, 7.00s/step, loss=1.204, acc=0.718, acc_len=1.74, thru=20.4, I=22.1, T=34.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:30,543] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:37,046] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:56:38,122] loop.py:45 INFO TRAIN_STEP step=17043 loss=1.325646 Training: 95%|█████████▍| 17043/18018 [36:26:24<1:58:17, 7.28s/step, loss=1.326, acc=0.700, acc_len=1.68, thru=16.0, I=18.2, T=33.5, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:38,558] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:43,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:56:43,950] loop.py:45 INFO TRAIN_STEP step=17044 loss=1.196391 Training: 95%|█████████▍| 17044/18018 [36:26:30<1:51:06, 6.84s/step, loss=1.196, acc=0.720, acc_len=1.74, thru=20.1, I=24.4, T=37.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:44,225] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:49,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:56:49,712] loop.py:380 INFO TIMING step=17045: step=3.259s data=1.337s compute=3.226s [fwd=2.312s bwd=0.900s opt=0.003s] dispatch=2.495s [2026-10-20 01:56:49,712] loop.py:45 INFO TRAIN_STEP step=17045 loss=1.268453 Training: 95%|█████████▍| 17045/18018 [36:26:35<1:45:44, 6.52s/step, loss=1.268, acc=0.705, acc_len=1.70, thru=22.2, I=23.0, T=39.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:56:50,157] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:56:55,870] loop.py:45 INFO TRAIN_STEP step=17046 loss=1.233324 Training: 95%|█████████▍| 17046/18018 [36:26:41<1:43:51, 6.41s/step, loss=1.233, acc=0.710, acc_len=1.71, thru=20.9, I=24.3, T=38.3, wait=2.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:56:55,865] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17045: forward=2343.6ms backward=942.2ms [2026-10-20 01:57:04,600] loop.py:45 INFO TRAIN_STEP step=17047 loss=1.275833 Training: 95%|█████████▍| 17047/18018 [36:26:50<1:55:00, 7.11s/step, loss=1.276, acc=0.703, acc_len=1.69, thru=12.8, I=15.5, T=35.9, wait=5.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:57:13,345] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:57:13,659] loop.py:45 INFO TRAIN_STEP step=17048 loss=1.158330 Training: 95%|█████████▍| 17048/18018 [36:26:59<2:04:21, 7.69s/step, loss=1.158, acc=0.723, acc_len=1.76, thru=15.8, I=14.9, T=32.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:57:13,847] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:57:19,249] loop.py:45 INFO TRAIN_STEP step=17049 loss=1.157643 Training: 95%|█████████▍| 17049/18018 [36:27:05<1:54:02, 7.06s/step, loss=1.158, acc=0.723, acc_len=1.76, thru=20.5, I=23.3, T=39.4, wait=2.3s, pool=61, epoch=2/2][2026-10-20 01:57:25,998] loop.py:380 INFO TIMING step=17050: step=3.452s data=1.399s compute=3.413s [fwd=2.465s bwd=0.933s opt=0.003s] dispatch=3.289s [2026-10-20 01:57:25,998] loop.py:45 INFO TRAIN_STEP step=17050 loss=1.305431 Training: 95%|█████████▍| 17050/18018 [36:27:12<1:52:27, 6.97s/step, loss=1.305, acc=0.698, acc_len=1.67, thru=17.6, I=18.5, T=37.1, wait=3.3s, pool=53, epoch=2/2][2026-10-20 01:57:33,581] loop.py:45 INFO TRAIN_STEP step=17051 loss=1.298295 Training: 95%|█████████▍| 17051/18018 [36:27:19<1:55:15, 7.15s/step, loss=1.298, acc=0.702, acc_len=1.69, thru=16.1, I=15.9, T=36.7, wait=4.1s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:57:33,576] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17050: forward=2472.7ms backward=968.3ms [2026-10-20 01:57:43,819] loop.py:45 INFO TRAIN_STEP step=17052 loss=1.247576 Training: 95%|█████████▍| 17052/18018 [36:27:29<2:10:02, 8.08s/step, loss=1.248, acc=0.712, acc_len=1.72, thru=14.2, I=11.1, T=33.1, wait=6.4s, pool=33, epoch=2/2][2026-10-20 01:57:53,607] loop.py:45 INFO TRAIN_STEP step=17053 loss=1.192536 Training: 95%|█████████▍| 17053/18018 [36:27:39<2:18:10, 8.59s/step, loss=1.193, acc=0.721, acc_len=1.76, thru=14.2, I=13.6, T=33.3, wait=5.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:57:59,710] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:58:00,374] loop.py:45 INFO TRAIN_STEP step=17054 loss=1.352278 Training: 95%|█████████▍| 17054/18018 [36:27:46<2:09:14, 8.04s/step, loss=1.352, acc=0.687, acc_len=1.64, thru=19.3, I=19.1, T=35.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:58:00,717] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:58:08,919] loop.py:380 INFO TIMING step=17055: step=3.700s data=1.582s compute=3.670s [fwd=2.653s bwd=1.002s opt=0.003s] dispatch=4.836s [2026-10-20 01:58:08,919] loop.py:45 INFO TRAIN_STEP step=17055 loss=1.234247 Training: 95%|█████████▍| 17055/18018 [36:27:55<2:11:31, 8.19s/step, loss=1.234, acc=0.710, acc_len=1.71, thru=14.2, I=16.2, T=34.6, wait=4.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:58:15,201] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:58:15,484] loop.py:45 INFO TRAIN_STEP step=17056 loss=1.196634 Training: 95%|█████████▍| 17056/18018 [36:28:01<2:03:32, 7.71s/step, loss=1.197, acc=0.718, acc_len=1.74, thru=18.7, I=20.5, T=37.6, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:58:15,479] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17055: forward=2332.8ms backward=998.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:58:15,704] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:58:21,783] loop.py:45 INFO TRAIN_STEP step=17057 loss=1.306380 Training: 95%|█████████▍| 17057/18018 [36:28:07<1:56:39, 7.28s/step, loss=1.306, acc=0.703, acc_len=1.69, thru=17.3, I=20.9, T=36.7, wait=2.8s, pool=33, epoch=2/2](TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 01:58:21,847] eagle_store.py:517 WARNING batch_get_buffer missing keys (k545824-9f87c983_hs, k545824-9f87c983_ids, k545824-9f87c983_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:58:30,195] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9368) [2026-10-20 01:58:21,966] eagle_store.py:517 WARNING batch_get_buffer missing keys (k545821-a3069391_hs, k545821-a3069391_ids, k545821-a3069391_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-20 01:58:31,998] loop.py:45 INFO TRAIN_STEP step=17058 loss=1.358089 Training: 95%|█████████▍| 17058/18018 [36:28:18<2:10:36, 8.16s/step, loss=1.358, acc=0.690, acc_len=1.65, thru=16.2, I=16.4, T=30.9, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:58:32,209] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 01:58:38,587] loop.py:45 INFO TRAIN_STEP step=17059 loss=1.076084 Training: 95%|█████████▍| 17059/18018 [36:28:24<2:02:55, 7.69s/step, loss=1.076, acc=0.740, acc_len=1.82, thru=15.4, I=19.5, T=38.4, wait=3.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:58:47,360] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:58:47,654] loop.py:380 INFO TIMING step=17060: step=3.611s data=1.496s compute=3.569s [fwd=2.531s bwd=1.024s opt=0.003s] dispatch=5.447s [2026-10-20 01:58:47,655] loop.py:45 INFO TRAIN_STEP step=17060 loss=1.304571 Training: 95%|█████████▍| 17060/18018 [36:28:33<2:09:23, 8.10s/step, loss=1.305, acc=0.701, acc_len=1.68, thru=17.0, I=17.1, T=35.4, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:58:47,862] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:58:54,011] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:58:54,914] loop.py:45 INFO TRAIN_STEP step=17061 loss=1.141461 Training: 95%|█████████▍| 17061/18018 [36:28:41<2:05:13, 7.85s/step, loss=1.141, acc=0.727, acc_len=1.77, thru=17.7, I=19.7, T=33.2, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:58:54,910] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17060: forward=2669.9ms backward=1145.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:58:55,016] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:59:02,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:59:02,393] loop.py:45 INFO TRAIN_STEP step=17062 loss=1.273257 Training: 95%|█████████▍| 17062/18018 [36:28:48<2:03:18, 7.74s/step, loss=1.273, acc=0.703, acc_len=1.70, thru=16.9, I=17.4, T=37.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:59:02,555] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:59:09,094] loop.py:45 INFO TRAIN_STEP step=17063 loss=1.217784 Training: 95%|█████████▍| 17063/18018 [36:28:55<1:58:12, 7.43s/step, loss=1.218, acc=0.712, acc_len=1.72, thru=18.1, I=18.1, T=35.2, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:59:15,726] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:59:15,944] loop.py:45 INFO TRAIN_STEP step=17064 loss=1.441672 Training: 95%|█████████▍| 17064/18018 [36:29:02<1:55:21, 7.25s/step, loss=1.442, acc=0.672, acc_len=1.59, thru=18.3, I=18.8, T=35.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:59:16,227] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 01:59:24,453] loop.py:380 INFO TIMING step=17065: step=3.657s data=1.511s compute=3.625s [fwd=2.604s bwd=1.007s opt=0.003s] dispatch=4.842s [2026-10-20 01:59:24,453] loop.py:45 INFO TRAIN_STEP step=17065 loss=1.176854 Training: 95%|█████████▍| 17065/18018 [36:29:10<2:01:11, 7.63s/step, loss=1.177, acc=0.718, acc_len=1.74, thru=13.9, I=15.1, T=35.0, wait=4.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:59:31,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:59:32,362] loop.py:45 INFO TRAIN_STEP step=17066 loss=1.191664 Training: 95%|█████████▍| 17066/18018 [36:29:18<2:02:23, 7.71s/step, loss=1.192, acc=0.720, acc_len=1.76, thru=17.0, I=17.5, T=33.7, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 01:59:32,357] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17065: forward=2697.9ms backward=1056.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:59:32,763] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:59:38,540] loop.py:45 INFO TRAIN_STEP step=17067 loss=1.292927 Training: 95%|█████████▍| 17067/18018 [36:29:24<1:54:59, 7.25s/step, loss=1.293, acc=0.701, acc_len=1.68, thru=16.5, I=19.2, T=37.1, wait=2.7s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:59:45,381] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 01:59:45,949] loop.py:45 INFO TRAIN_STEP step=17068 loss=1.236251 Training: 95%|█████████▍| 17068/18018 [36:29:32<1:55:34, 7.30s/step, loss=1.236, acc=0.713, acc_len=1.73, thru=19.3, I=18.4, T=37.3, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 01:59:46,385] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 01:59:51,561] loop.py:45 INFO TRAIN_STEP step=17069 loss=1.206736 Training: 95%|█████████▍| 17069/18018 [36:29:37<1:47:26, 6.79s/step, loss=1.207, acc=0.720, acc_len=1.75, thru=20.8, I=25.6, T=39.1, wait=2.3s, pool=49, epoch=2/2][2026-10-20 02:00:00,847] loop.py:380 INFO TIMING step=17070: step=3.605s data=1.475s compute=3.570s [fwd=2.575s bwd=0.979s opt=0.003s] dispatch=5.675s [2026-10-20 02:00:00,848] loop.py:45 INFO TRAIN_STEP step=17070 loss=1.201684 Training: 95%|█████████▍| 17070/18018 [36:29:46<1:59:09, 7.54s/step, loss=1.202, acc=0.714, acc_len=1.72, thru=15.2, I=14.3, T=35.5, wait=5.7s, pool=61, epoch=2/2][2026-10-20 02:00:08,949] loop.py:45 INFO TRAIN_STEP step=17071 loss=1.298103 Training: 95%|█████████▍| 17071/18018 [36:29:55<2:01:40, 7.71s/step, loss=1.298, acc=0.697, acc_len=1.67, thru=17.2, I=16.4, T=35.4, wait=4.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:00:08,944] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17070: forward=2554.4ms backward=1004.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:00:17,247] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:00:17,851] loop.py:45 INFO TRAIN_STEP step=17072 loss=1.394410 Training: 95%|█████████▍| 17072/18018 [36:30:03<2:07:11, 8.07s/step, loss=1.394, acc=0.685, acc_len=1.63, thru=15.2, I=15.4, T=34.7, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:00:18,255] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:00:24,226] loop.py:45 INFO TRAIN_STEP step=17073 loss=1.342732 Training: 95%|█████████▍| 17073/18018 [36:30:10<1:59:03, 7.56s/step, loss=1.343, acc=0.690, acc_len=1.64, thru=17.3, I=19.2, T=38.3, wait=3.0s, pool=45, epoch=2/2][2026-10-20 02:00:33,021] loop.py:45 INFO TRAIN_STEP step=17074 loss=1.285855 Training: 95%|█████████▍| 17074/18018 [36:30:19<2:04:46, 7.93s/step, loss=1.286, acc=0.698, acc_len=1.67, thru=16.5, I=17.0, T=36.3, wait=5.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:00:39,769] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:00:40,115] loop.py:380 INFO TIMING step=17075: step=3.530s data=1.505s compute=3.485s [fwd=2.426s bwd=1.045s opt=0.003s] dispatch=3.555s [2026-10-20 02:00:40,115] loop.py:45 INFO TRAIN_STEP step=17075 loss=1.215185 Training: 95%|█████████▍| 17075/18018 [36:30:26<2:00:41, 7.68s/step, loss=1.215, acc=0.714, acc_len=1.72, thru=19.1, I=18.5, T=36.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:00:40,271] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:00:46,256] loop.py:45 INFO TRAIN_STEP step=17076 loss=1.187350 Training: 95%|█████████▍| 17076/18018 [36:30:32<1:53:19, 7.22s/step, loss=1.187, acc=0.715, acc_len=1.73, thru=20.1, I=22.4, T=35.1, wait=2.5s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:00:46,252] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17075: forward=2578.3ms backward=1025.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:00:54,206] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:00:54,666] loop.py:45 INFO TRAIN_STEP step=17077 loss=1.287677 Training: 95%|█████████▍| 17077/18018 [36:30:40<1:58:48, 7.58s/step, loss=1.288, acc=0.703, acc_len=1.69, thru=15.0, I=15.9, T=33.2, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:00:54,719] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:01:00,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:01:01,669] loop.py:45 INFO TRAIN_STEP step=17078 loss=1.160850 Training: 95%|█████████▍| 17078/18018 [36:30:47<1:56:00, 7.40s/step, loss=1.161, acc=0.723, acc_len=1.76, thru=18.2, I=20.6, T=35.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:01:01,784] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-20 02:01:07,323] loop.py:45 INFO TRAIN_STEP step=17079 loss=1.116963 Training: 95%|█████████▍| 17079/18018 [36:30:53<1:47:38, 6.88s/step, loss=1.117, acc=0.733, acc_len=1.79, thru=20.6, I=23.0, T=39.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:01:07,304] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:01:07,805] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:01:14,879] loop.py:380 INFO TIMING step=17080: step=3.478s data=1.411s compute=3.450s [fwd=2.473s bwd=0.962s opt=0.003s] dispatch=4.072s [2026-10-20 02:01:14,880] loop.py:45 INFO TRAIN_STEP step=17080 loss=1.266363 Training: 95%|█████████▍| 17080/18018 [36:31:00<1:50:42, 7.08s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=17.0, I=15.7, T=36.8, wait=4.1s, pool=37, epoch=2/2][2026-10-20 02:01:24,051] loop.py:45 INFO TRAIN_STEP step=17081 loss=1.354630 Training: 95%|█████████▍| 17081/18018 [36:31:10<2:00:23, 7.71s/step, loss=1.355, acc=0.689, acc_len=1.64, thru=14.5, I=16.0, T=33.7, wait=5.4s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:01:24,047] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17080: forward=2738.5ms backward=1018.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:01:34,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:01:34,832] loop.py:45 INFO TRAIN_STEP step=17082 loss=1.369468 Training: 95%|█████████▍| 17082/18018 [36:31:20<2:14:38, 8.63s/step, loss=1.369, acc=0.683, acc_len=1.62, thru=13.9, I=13.9, T=33.5, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:01:35,241] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:01:41,240] loop.py:45 INFO TRAIN_STEP step=17083 loss=1.207582 Training: 95%|█████████▍| 17083/18018 [36:31:27<2:04:06, 7.96s/step, loss=1.208, acc=0.717, acc_len=1.74, thru=16.8, I=19.0, T=38.4, wait=3.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:01:51,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:01:51,663] loop.py:45 INFO TRAIN_STEP step=17084 loss=1.294678 Training: 95%|█████████▍| 17084/18018 [36:31:37<2:15:26, 8.70s/step, loss=1.295, acc=0.703, acc_len=1.69, thru=14.9, I=14.7, T=33.5, wait=6.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:01:51,693] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-20 02:01:57,040] loop.py:380 INFO TIMING step=17085: step=3.358s data=1.313s compute=3.331s [fwd=2.426s bwd=0.890s opt=0.003s] dispatch=2.010s [2026-10-20 02:01:57,040] loop.py:45 INFO TRAIN_STEP step=17085 loss=1.229930 Training: 95%|█████████▍| 17085/18018 [36:31:43<1:59:48, 7.70s/step, loss=1.230, acc=0.715, acc_len=1.73, thru=19.6, I=20.7, T=38.1, wait=2.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:02:03,348] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:02:04,798] loop.py:45 INFO TRAIN_STEP step=17086 loss=1.311126 Training: 95%|█████████▍| 17086/18018 [36:31:50<1:59:54, 7.72s/step, loss=1.311, acc=0.696, acc_len=1.66, thru=17.3, I=19.3, T=36.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:02:04,863] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (TrainerActor pid=9310) [2026-10-20 02:02:04,794] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17085: forward=2524.9ms backward=972.9ms [2026-10-20 02:02:11,956] loop.py:45 INFO TRAIN_STEP step=17087 loss=1.348757 Training: 95%|█████████▍| 17087/18018 [36:31:58<1:57:12, 7.55s/step, loss=1.349, acc=0.689, acc_len=1.64, thru=16.9, I=19.3, T=34.3, wait=3.4s, pool=61, epoch=2/2][2026-10-20 02:02:21,695] loop.py:45 INFO TRAIN_STEP step=17088 loss=1.333635 Training: 95%|█████████▍| 17088/18018 [36:32:07<2:07:13, 8.21s/step, loss=1.334, acc=0.690, acc_len=1.65, thru=13.0, I=13.0, T=33.1, wait=5.9s, pool=53, epoch=2/2][2026-10-20 02:02:32,744] loop.py:45 INFO TRAIN_STEP step=17089 loss=1.278544 Training: 95%|█████████▍| 17089/18018 [36:32:18<2:20:15, 9.06s/step, loss=1.279, acc=0.707, acc_len=1.71, thru=12.5, I=11.7, T=32.8, wait=7.1s, pool=61, epoch=2/2][2026-10-20 02:02:39,022] loop.py:380 INFO TIMING step=17090: step=3.632s data=1.485s compute=3.605s [fwd=2.571s bwd=1.019s opt=0.003s] dispatch=2.638s [2026-10-20 02:02:39,023] loop.py:45 INFO TRAIN_STEP step=17090 loss=1.228330 Training: 95%|█████████▍| 17090/18018 [36:32:25<2:07:12, 8.22s/step, loss=1.228, acc=0.711, acc_len=1.72, thru=17.3, I=16.2, T=35.2, wait=2.6s, pool=49, epoch=2/2][2026-10-20 02:02:47,748] loop.py:45 INFO TRAIN_STEP step=17091 loss=1.301243 (TrainerActor pid=9310) [2026-10-20 02:02:47,743] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17090: forward=2659.2ms backward=1062.4ms Training: 95%|█████████▍| 17091/18018 [36:32:33<2:09:24, 8.38s/step, loss=1.301, acc=0.701, acc_len=1.68, thru=13.7, I=13.4, T=33.9, wait=4.9s, pool=29, epoch=2/2][2026-10-20 02:02:55,109] loop.py:45 INFO TRAIN_STEP step=17092 loss=1.209467 Training: 95%|█████████▍| 17092/18018 [36:32:41<2:04:33, 8.07s/step, loss=1.209, acc=0.717, acc_len=1.74, thru=19.1, I=18.4, T=35.7, wait=3.8s, pool=49, epoch=2/2][2026-10-20 02:03:03,120] loop.py:45 INFO TRAIN_STEP step=17093 loss=1.319913 Training: 95%|█████████▍| 17093/18018 [36:32:49<2:04:09, 8.05s/step, loss=1.320, acc=0.696, acc_len=1.67, thru=16.1, I=16.3, T=34.0, wait=4.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:03:08,070] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:03:08,755] loop.py:45 INFO TRAIN_STEP step=17094 loss=1.183078 Training: 95%|█████████▍| 17094/18018 [36:32:54<1:52:50, 7.33s/step, loss=1.183, acc=0.718, acc_len=1.74, thru=20.8, I=20.8, T=39.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:03:09,076] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:03:15,334] loop.py:380 INFO TIMING step=17095: step=3.534s data=1.469s compute=3.497s [fwd=2.409s bwd=1.074s opt=0.003s] dispatch=3.037s [2026-10-20 02:03:15,334] loop.py:45 INFO TRAIN_STEP step=17095 loss=1.249460 Training: 95%|█████████▍| 17095/18018 [36:33:01<1:49:15, 7.10s/step, loss=1.249, acc=0.709, acc_len=1.71, thru=18.2, I=21.3, T=36.2, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:03:21,817] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:03:23,047] loop.py:45 INFO TRAIN_STEP step=17096 loss=1.321507 Training: 95%|█████████▍| 17096/18018 [36:33:09<1:51:58, 7.29s/step, loss=1.322, acc=0.692, acc_len=1.66, thru=16.3, I=18.5, T=34.7, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:03:23,042] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17095: forward=2635.9ms backward=1012.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:03:23,325] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:03:30,806] loop.py:45 INFO TRAIN_STEP step=17097 loss=1.251851 Training: 95%|█████████▍| 17097/18018 [36:33:16<1:54:01, 7.43s/step, loss=1.252, acc=0.710, acc_len=1.71, thru=10.9, I=16.4, T=37.1, wait=4.3s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 02:03:30,991] eagle_store.py:517 WARNING batch_get_buffer missing keys (k547114-ae079f09_hs, k547114-ae079f09_ids, k547114-ae079f09_lhs); sleeping 0.50s. [2026-10-20 02:03:40,926] loop.py:45 INFO TRAIN_STEP step=17098 loss=1.269413 Training: 95%|█████████▍| 17098/18018 [36:33:27<2:06:16, 8.24s/step, loss=1.269, acc=0.705, acc_len=1.70, thru=16.0, I=13.9, T=33.2, wait=6.3s, pool=53, epoch=2/2][2026-10-20 02:03:49,454] loop.py:45 INFO TRAIN_STEP step=17099 loss=1.260291 Training: 95%|█████████▍| 17099/18018 [36:33:35<2:07:28, 8.32s/step, loss=1.260, acc=0.705, acc_len=1.69, thru=16.8, I=16.0, T=36.6, wait=5.0s, pool=61, epoch=2/2][2026-10-20 02:03:56,583] loop.py:380 INFO TIMING step=17100: step=3.759s data=1.552s compute=3.722s [fwd=2.604s bwd=1.105s opt=0.003s] dispatch=3.361s [2026-10-20 02:03:56,583] loop.py:45 INFO TRAIN_STEP step=17100 loss=1.171627 Training: 95%|█████████▍| 17100/18018 [36:33:42<2:01:52, 7.97s/step, loss=1.172, acc=0.721, acc_len=1.75, thru=17.6, I=18.3, T=34.1, wait=3.4s, pool=57, epoch=2/2][2026-10-20 02:04:07,427] loop.py:45 INFO TRAIN_STEP step=17101 loss=1.287398 Training: 95%|█████████▍| 17101/18018 [36:33:53<2:14:55, 8.83s/step, loss=1.287, acc=0.704, acc_len=1.69, thru=12.6, I=13.1, T=33.3, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:04:07,423] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-20 02:03:30,989] eagle_store.py:517 WARNING batch_get_buffer missing keys (k547113-bb1527d8_hs, k547113-bb1527d8_ids, k547113-bb1527d8_lhs); sleeping 0.50s. [repeated 2x across cluster] (TrainerActor pid=9310) [2026-10-20 02:04:07,423] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17100: forward=2575.4ms backward=1197.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:04:07,926] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:04:14,678] loop.py:45 INFO TRAIN_STEP step=17102 loss=1.231365 Training: 95%|█████████▍| 17102/18018 [36:34:00<2:07:33, 8.36s/step, loss=1.231, acc=0.711, acc_len=1.72, thru=14.6, I=15.6, T=36.3, wait=3.7s, pool=53, epoch=2/2][2026-10-20 02:04:23,229] loop.py:45 INFO TRAIN_STEP step=17103 loss=1.238296 Training: 95%|█████████▍| 17103/18018 [36:34:09<2:08:18, 8.41s/step, loss=1.238, acc=0.710, acc_len=1.71, thru=16.7, I=16.5, T=34.2, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:04:23,168] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:04:23,670] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:04:29,744] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:04:30,377] loop.py:45 INFO TRAIN_STEP step=17104 loss=1.260673 Training: 95%|█████████▍| 17104/18018 [36:34:16<2:02:22, 8.03s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=18.7, I=19.2, T=36.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:04:30,749] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:04:38,055] loop.py:380 INFO TIMING step=17105: step=3.630s data=1.416s compute=3.592s [fwd=2.520s bwd=1.055s opt=0.003s] dispatch=4.041s [2026-10-20 02:04:38,055] loop.py:45 INFO TRAIN_STEP step=17105 loss=1.220218 Training: 95%|█████████▍| 17105/18018 [36:34:24<2:00:37, 7.93s/step, loss=1.220, acc=0.711, acc_len=1.72, thru=14.9, I=15.8, T=35.3, wait=4.0s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:04:44,986] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17105: forward=3038.1ms backward=917.1ms [2026-10-20 02:04:44,991] loop.py:45 INFO TRAIN_STEP step=17106 loss=1.185498 Training: 95%|█████████▍| 17106/18018 [36:34:31<1:55:59, 7.63s/step, loss=1.185, acc=0.716, acc_len=1.74, thru=17.1, I=21.1, T=31.9, wait=2.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:04:52,482] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:04:53,873] loop.py:45 INFO TRAIN_STEP step=17107 loss=1.333448 Training: 95%|█████████▍| 17107/18018 [36:34:39<2:01:32, 8.01s/step, loss=1.333, acc=0.697, acc_len=1.67, thru=15.2, I=15.3, T=25.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:04:54,002] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:04:59,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:05:00,476] loop.py:45 INFO TRAIN_STEP step=17108 loss=1.298909 Training: 95%|█████████▍| 17108/18018 [36:34:46<1:55:02, 7.58s/step, loss=1.299, acc=0.701, acc_len=1.68, thru=15.8, I=19.6, T=34.7, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:00,655] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:07,234] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:05:08,529] loop.py:45 INFO TRAIN_STEP step=17109 loss=1.241894 Training: 95%|█████████▍| 17109/18018 [36:34:54<1:57:02, 7.73s/step, loss=1.242, acc=0.710, acc_len=1.71, thru=14.5, I=18.2, T=33.0, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:08,746] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:14,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:05:15,207] loop.py:380 INFO TIMING step=17110: step=3.612s data=1.510s compute=3.579s [fwd=2.506s bwd=1.058s opt=0.003s] dispatch=3.057s [2026-10-20 02:05:15,207] loop.py:45 INFO TRAIN_STEP step=17110 loss=1.140149 Training: 95%|█████████▍| 17110/18018 [36:35:01<1:52:09, 7.41s/step, loss=1.140, acc=0.723, acc_len=1.76, thru=18.7, I=24.0, T=35.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:15,240] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:22,973] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 02:05:24,042] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17110: forward=2633.7ms backward=1118.2ms [2026-10-20 02:05:24,047] loop.py:45 INFO TRAIN_STEP step=17111 loss=1.261872 Training: 95%|█████████▍| 17111/18018 [36:35:10<1:58:30, 7.84s/step, loss=1.262, acc=0.710, acc_len=1.72, thru=13.4, I=15.9, T=33.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:24,482] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:31,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:05:31,574] loop.py:45 INFO TRAIN_STEP step=17112 loss=1.264493 Training: 95%|█████████▍| 17112/18018 [36:35:17<1:56:58, 7.75s/step, loss=1.264, acc=0.711, acc_len=1.72, thru=15.8, I=18.9, T=36.7, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:31,725] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:38,614] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:05:39,241] loop.py:45 INFO TRAIN_STEP step=17113 loss=1.266598 Training: 95%|█████████▍| 17113/18018 [36:35:25<1:56:28, 7.72s/step, loss=1.267, acc=0.705, acc_len=1.70, thru=17.0, I=18.5, T=35.2, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:05:39,621] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:05:46,140] loop.py:45 INFO TRAIN_STEP step=17114 loss=1.113153 Training: 95%|█████████▍| 17114/18018 [36:35:32<1:52:37, 7.47s/step, loss=1.113, acc=0.734, acc_len=1.80, thru=15.9, I=18.5, T=34.6, wait=3.2s, pool=41, epoch=2/2][2026-10-20 02:05:53,832] loop.py:380 INFO TIMING step=17115: step=3.948s data=1.772s compute=3.904s [fwd=2.711s bwd=1.179s opt=0.003s] dispatch=3.736s [2026-10-20 02:05:53,832] loop.py:45 INFO TRAIN_STEP step=17115 loss=1.262958 Training: 95%|█████████▍| 17115/18018 [36:35:39<1:53:29, 7.54s/step, loss=1.263, acc=0.704, acc_len=1.70, thru=16.8, I=17.1, T=32.4, wait=3.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:00,959] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:06:01,937] loop.py:45 INFO TRAIN_STEP step=17116 loss=1.154041 Training: 95%|█████████▍| 17116/18018 [36:35:48<1:55:54, 7.71s/step, loss=1.154, acc=0.725, acc_len=1.77, thru=17.3, I=17.8, T=35.3, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:06:01,932] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17115: forward=2586.2ms backward=989.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:01,975] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:07,388] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:06:08,114] loop.py:45 INFO TRAIN_STEP step=17117 loss=1.332694 Training: 95%|█████████▍| 17117/18018 [36:35:54<1:48:52, 7.25s/step, loss=1.333, acc=0.697, acc_len=1.67, thru=19.9, I=23.8, T=38.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:08,392] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:14,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:06:15,251] loop.py:45 INFO TRAIN_STEP step=17118 loss=1.151381 Training: 95%|█████████▌| 17118/18018 [36:36:01<1:48:14, 7.22s/step, loss=1.151, acc=0.729, acc_len=1.78, thru=17.9, I=23.0, T=31.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:15,550] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:25,791] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:06:26,522] loop.py:45 INFO TRAIN_STEP step=17119 loss=1.257780 Training: 95%|█████████▌| 17119/18018 [36:36:12<2:06:20, 8.43s/step, loss=1.258, acc=0.711, acc_len=1.72, thru=12.0, I=11.4, T=29.1, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:26,798] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:06:34,386] loop.py:380 INFO TIMING step=17120: step=3.952s data=1.823s compute=3.916s [fwd=2.850s bwd=1.052s opt=0.003s] dispatch=3.903s [2026-10-20 02:06:34,386] loop.py:45 INFO TRAIN_STEP step=17120 loss=1.390787 Training: 95%|█████████▌| 17120/18018 [36:36:20<2:03:39, 8.26s/step, loss=1.391, acc=0.680, acc_len=1.61, thru=14.5, I=16.7, T=32.4, wait=3.9s, pool=49, epoch=2/2][2026-10-20 02:06:43,798] loop.py:45 INFO TRAIN_STEP step=17121 loss=1.294398 Training: 95%|█████████▌| 17121/18018 [36:36:29<2:08:40, 8.61s/step, loss=1.294, acc=0.697, acc_len=1.67, thru=14.5, I=13.0, T=31.3, wait=5.3s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:06:43,793] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17120: forward=2941.9ms backward=1088.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:50,628] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:06:51,570] loop.py:45 INFO TRAIN_STEP step=17122 loss=1.246824 Training: 95%|█████████▌| 17122/18018 [36:36:37<2:04:47, 8.36s/step, loss=1.247, acc=0.709, acc_len=1.71, thru=16.5, I=17.8, T=32.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:51,643] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:56,677] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:06:58,210] loop.py:45 INFO TRAIN_STEP step=17123 loss=1.255214 Training: 95%|█████████▌| 17123/18018 [36:36:44<1:56:57, 7.84s/step, loss=1.255, acc=0.706, acc_len=1.70, thru=18.5, I=23.5, T=30.3, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:06:58,689] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 02:07:04,874] loop.py:45 INFO TRAIN_STEP step=17124 loss=1.317444 Training: 95%|█████████▌| 17124/18018 [36:36:50<1:51:35, 7.49s/step, loss=1.317, acc=0.695, acc_len=1.66, thru=15.6, I=21.0, T=36.7, wait=3.2s, pool=49, epoch=2/2][2026-10-20 02:07:15,035] loop.py:380 INFO TIMING step=17125: step=4.129s data=1.855s compute=4.104s [fwd=3.004s bwd=1.085s opt=0.003s] dispatch=6.020s [2026-10-20 02:07:15,036] loop.py:45 INFO TRAIN_STEP step=17125 loss=1.252762 Training: 95%|█████████▌| 17125/18018 [36:37:01<2:03:23, 8.29s/step, loss=1.253, acc=0.710, acc_len=1.71, thru=15.6, I=13.9, T=31.0, wait=6.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:07:22,608] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:07:23,432] loop.py:45 INFO TRAIN_STEP step=17126 loss=1.287402 Training: 95%|█████████▌| 17126/18018 [36:37:09<2:03:43, 8.32s/step, loss=1.287, acc=0.701, acc_len=1.68, thru=15.8, I=16.6, T=35.3, wait=4.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:07:23,428] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17125: forward=2567.0ms backward=1007.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:07:23,612] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:07:29,152] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:07:29,262] loop.py:45 INFO TRAIN_STEP step=17127 loss=1.208188 Training: 95%|█████████▌| 17127/18018 [36:37:15<1:52:28, 7.57s/step, loss=1.208, acc=0.714, acc_len=1.72, thru=20.7, I=21.8, T=37.9, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:07:29,653] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:07:36,789] loop.py:45 INFO TRAIN_STEP step=17128 loss=1.285098 Training: 95%|█████████▌| 17128/18018 [36:37:22<1:52:07, 7.56s/step, loss=1.285, acc=0.702, acc_len=1.69, thru=17.7, I=20.1, T=35.9, wait=3.9s, pool=61, epoch=2/2][2026-10-20 02:07:46,893] loop.py:45 INFO TRAIN_STEP step=17129 loss=1.290701 Training: 95%|█████████▌| 17129/18018 [36:37:32<2:03:20, 8.32s/step, loss=1.291, acc=0.702, acc_len=1.69, thru=13.1, I=12.3, T=32.7, wait=6.2s, pool=61, epoch=2/2][2026-10-20 02:07:54,590] loop.py:380 INFO TIMING step=17130: step=3.930s data=1.701s compute=3.887s [fwd=2.739s bwd=1.133s opt=0.003s] dispatch=3.755s [2026-10-20 02:07:54,590] loop.py:45 INFO TRAIN_STEP step=17130 loss=1.319037 Training: 95%|█████████▌| 17130/18018 [36:37:40<2:00:23, 8.13s/step, loss=1.319, acc=0.691, acc_len=1.65, thru=16.0, I=16.9, T=32.6, wait=3.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:01,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:08:02,088] loop.py:45 INFO TRAIN_STEP step=17131 loss=1.252242 Training: 95%|█████████▌| 17131/18018 [36:37:48<1:57:26, 7.94s/step, loss=1.252, acc=0.703, acc_len=1.69, thru=17.4, I=16.8, T=36.2, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:08:02,083] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17130: forward=2512.7ms backward=980.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:02,219] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:09,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:08:09,732] loop.py:45 INFO TRAIN_STEP step=17132 loss=1.304834 Training: 95%|█████████▌| 17132/18018 [36:37:55<1:55:58, 7.85s/step, loss=1.305, acc=0.700, acc_len=1.68, thru=17.7, I=19.1, T=34.3, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:10,208] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:14,501] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:08:15,706] loop.py:45 INFO TRAIN_STEP step=17133 loss=1.218899 Training: 95%|█████████▌| 17133/18018 [36:38:01<1:47:31, 7.29s/step, loss=1.219, acc=0.715, acc_len=1.73, thru=20.1, I=23.7, T=35.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:16,010] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:20,776] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:08:21,931] loop.py:45 INFO TRAIN_STEP step=17134 loss=1.237194 Training: 95%|█████████▌| 17134/18018 [36:38:08<1:42:41, 6.97s/step, loss=1.237, acc=0.709, acc_len=1.71, thru=20.0, I=25.8, T=35.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:22,284] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:26,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:08:27,643] loop.py:380 INFO TIMING step=17135: step=3.322s data=1.290s compute=3.289s [fwd=2.364s bwd=0.910s opt=0.003s] dispatch=2.382s [2026-10-20 02:08:27,643] loop.py:45 INFO TRAIN_STEP step=17135 loss=1.115082 Training: 95%|█████████▌| 17135/18018 [36:38:13<1:37:01, 6.59s/step, loss=1.115, acc=0.734, acc_len=1.80, thru=21.2, I=25.5, T=38.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:28,001] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:33,633] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:08:33,892] loop.py:45 INFO TRAIN_STEP step=17136 loss=1.333666 Training: 95%|█████████▌| 17136/18018 [36:38:19<1:35:23, 6.49s/step, loss=1.334, acc=0.694, acc_len=1.66, thru=21.2, I=24.0, T=35.1, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:08:33,888] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17135: forward=2570.8ms backward=1023.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:34,135] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:39,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:08:41,065] loop.py:45 INFO TRAIN_STEP step=17137 loss=1.279161 Training: 95%|█████████▌| 17137/18018 [36:38:27<1:38:18, 6.70s/step, loss=1.279, acc=0.702, acc_len=1.69, thru=17.8, I=20.0, T=35.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:41,255] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:46,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:08:47,378] loop.py:45 INFO TRAIN_STEP step=17138 loss=1.293384 Training: 95%|█████████▌| 17138/18018 [36:38:33<1:36:30, 6.58s/step, loss=1.293, acc=0.696, acc_len=1.67, thru=19.6, I=25.1, T=36.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:08:47,592] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:08:53,412] loop.py:45 INFO TRAIN_STEP step=17139 loss=1.105745 Training: 95%|█████████▌| 17139/18018 [36:38:39<1:33:59, 6.42s/step, loss=1.106, acc=0.730, acc_len=1.78, thru=18.4, I=23.0, T=38.5, wait=2.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:03,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:09:03,706] loop.py:380 INFO TIMING step=17140: step=4.007s data=1.625s compute=3.960s [fwd=2.768s bwd=1.177s opt=0.003s] dispatch=6.280s [2026-10-20 02:09:03,706] loop.py:45 INFO TRAIN_STEP step=17140 loss=1.230164 Training: 95%|█████████▌| 17140/18018 [36:38:49<1:50:54, 7.58s/step, loss=1.230, acc=0.713, acc_len=1.72, thru=14.8, I=15.1, T=31.9, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:03,845] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:08,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:09:10,462] loop.py:45 INFO TRAIN_STEP step=17141 loss=1.261622 Training: 95%|█████████▌| 17141/18018 [36:38:56<1:47:10, 7.33s/step, loss=1.262, acc=0.706, acc_len=1.69, thru=18.4, I=21.6, T=35.1, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:09:10,457] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17140: forward=2475.6ms backward=1106.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:10,491] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-20 02:09:15,733] loop.py:45 INFO TRAIN_STEP step=17142 loss=1.192123 Training: 95%|█████████▌| 17142/18018 [36:39:01<1:38:01, 6.71s/step, loss=1.192, acc=0.716, acc_len=1.73, thru=21.2, I=26.7, T=37.7, wait=1.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:23,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:09:24,089] loop.py:45 INFO TRAIN_STEP step=17143 loss=1.161126 Training: 95%|█████████▌| 17143/18018 [36:39:10<1:45:06, 7.21s/step, loss=1.161, acc=0.725, acc_len=1.77, thru=16.1, I=17.1, T=34.4, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:24,492] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:28,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:09:29,686] loop.py:45 INFO TRAIN_STEP step=17144 loss=1.216431 Training: 95%|█████████▌| 17144/18018 [36:39:15<1:37:56, 6.72s/step, loss=1.216, acc=0.713, acc_len=1.72, thru=21.0, I=23.1, T=38.1, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:29,816] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-20 02:09:36,683] loop.py:380 INFO TIMING step=17145: step=3.495s data=1.413s compute=3.464s [fwd=2.499s bwd=0.951s opt=0.003s] dispatch=3.494s [2026-10-20 02:09:36,683] loop.py:45 INFO TRAIN_STEP step=17145 loss=1.335416 Training: 95%|█████████▌| 17145/18018 [36:39:22<1:39:02, 6.81s/step, loss=1.335, acc=0.694, acc_len=1.66, thru=15.0, I=18.0, T=36.6, wait=3.5s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:46,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:09:46,539] loop.py:45 INFO TRAIN_STEP step=17146 loss=1.319813 Training: 95%|█████████▌| 17146/18018 [36:39:32<1:52:12, 7.72s/step, loss=1.320, acc=0.697, acc_len=1.67, thru=16.9, I=16.7, T=33.3, wait=6.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:09:46,534] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17145: forward=2768.7ms backward=1025.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:46,612] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:52,878] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:09:54,276] loop.py:45 INFO TRAIN_STEP step=17147 loss=1.378963 Training: 95%|█████████▌| 17147/18018 [36:39:40<1:52:09, 7.73s/step, loss=1.379, acc=0.687, acc_len=1.64, thru=16.4, I=18.7, T=34.9, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:09:54,398] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-20 02:10:01,820] loop.py:45 INFO TRAIN_STEP step=17148 loss=1.307723 Training: 95%|█████████▌| 17148/18018 [36:39:47<1:51:13, 7.67s/step, loss=1.308, acc=0.698, acc_len=1.67, thru=14.5, I=18.1, T=38.1, wait=4.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:09,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:10:10,309] loop.py:45 INFO TRAIN_STEP step=17149 loss=1.258429 Training: 95%|█████████▌| 17149/18018 [36:39:56<1:54:39, 7.92s/step, loss=1.258, acc=0.708, acc_len=1.71, thru=16.8, I=18.2, T=35.6, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:10,752] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:10:16,161] loop.py:380 INFO TIMING step=17150: step=3.465s data=1.462s compute=3.428s [fwd=2.454s bwd=0.958s opt=0.003s] dispatch=2.378s [2026-10-20 02:10:16,161] loop.py:45 INFO TRAIN_STEP step=17150 loss=1.219552 Training: 95%|█████████▌| 17150/18018 [36:40:02<1:45:34, 7.30s/step, loss=1.220, acc=0.713, acc_len=1.73, thru=19.2, I=21.3, T=36.9, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:26,029] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:10:26,685] loop.py:45 INFO TRAIN_STEP step=17151 loss=1.252610 Training: 95%|█████████▌| 17151/18018 [36:40:12<1:59:25, 8.26s/step, loss=1.253, acc=0.710, acc_len=1.72, thru=13.6, I=14.4, T=33.0, wait=6.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:10:26,680] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17150: forward=2724.8ms backward=1095.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:27,033] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:32,445] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:10:33,168] loop.py:45 INFO TRAIN_STEP step=17152 loss=1.287530 Training: 95%|█████████▌| 17152/18018 [36:40:19<1:51:35, 7.73s/step, loss=1.288, acc=0.701, acc_len=1.69, thru=18.6, I=22.7, T=36.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:33,451] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:38,705] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:10:39,500] loop.py:45 INFO TRAIN_STEP step=17153 loss=1.165035 Training: 95%|█████████▌| 17153/18018 [36:40:25<1:45:24, 7.31s/step, loss=1.165, acc=0.723, acc_len=1.76, thru=20.5, I=21.8, T=35.9, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:39,710] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:46,028] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:10:46,262] loop.py:45 INFO TRAIN_STEP step=17154 loss=1.161409 Training: 95%|█████████▌| 17154/18018 [36:40:32<1:42:54, 7.15s/step, loss=1.161, acc=0.726, acc_len=1.78, thru=19.2, I=20.3, T=38.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:10:46,530] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:10:52,052] loop.py:380 INFO TIMING step=17155: step=3.482s data=1.379s compute=3.452s [fwd=2.473s bwd=0.963s opt=0.003s] dispatch=2.298s [2026-10-20 02:10:52,052] loop.py:45 INFO TRAIN_STEP step=17155 loss=1.148163 Training: 95%|█████████▌| 17155/18018 [36:40:38<1:36:55, 6.74s/step, loss=1.148, acc=0.727, acc_len=1.78, thru=18.0, I=20.2, T=36.8, wait=2.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:11:01,362] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:11:01,872] loop.py:45 INFO TRAIN_STEP step=17156 loss=1.260514 Training: 95%|█████████▌| 17156/18018 [36:40:47<1:50:05, 7.66s/step, loss=1.261, acc=0.705, acc_len=1.70, thru=15.4, I=14.1, T=34.4, wait=6.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:11:01,868] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17155: forward=2646.6ms backward=1031.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:11:02,369] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:11:07,923] loop.py:45 INFO TRAIN_STEP step=17157 loss=1.370720 Training: 95%|█████████▌| 17157/18018 [36:40:54<1:43:01, 7.18s/step, loss=1.371, acc=0.687, acc_len=1.64, thru=18.4, I=19.7, T=36.7, wait=2.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:11:16,159] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:11:16,297] loop.py:45 INFO TRAIN_STEP step=17158 loss=1.205319 Training: 95%|█████████▌| 17158/18018 [36:41:02<1:48:02, 7.54s/step, loss=1.205, acc=0.713, acc_len=1.73, thru=16.3, I=16.7, T=35.1, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:11:16,660] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:11:23,834] loop.py:45 INFO TRAIN_STEP step=17159 loss=1.181819 Training: 95%|█████████▌| 17159/18018 [36:41:09<1:47:54, 7.54s/step, loss=1.182, acc=0.720, acc_len=1.75, thru=15.4, I=16.1, T=35.1, wait=3.9s, pool=53, epoch=2/2][2026-10-20 02:11:31,994] loop.py:380 INFO TIMING step=17160: step=3.656s data=1.505s compute=3.627s [fwd=2.645s bwd=0.967s opt=0.003s] dispatch=4.497s [2026-10-20 02:11:31,994] loop.py:45 INFO TRAIN_STEP step=17160 loss=1.298606 Training: 95%|█████████▌| 17160/18018 [36:41:18<1:50:27, 7.72s/step, loss=1.299, acc=0.699, acc_len=1.68, thru=17.2, I=16.7, T=35.0, wait=4.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:11:40,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:11:41,602] loop.py:45 INFO TRAIN_STEP step=17161 loss=1.188844 Training: 95%|█████████▌| 17161/18018 [36:41:27<1:58:24, 8.29s/step, loss=1.189, acc=0.719, acc_len=1.75, thru=13.9, I=15.3, T=32.7, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:11:41,597] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17160: forward=2792.6ms backward=1068.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:11:41,742] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-20 02:11:47,496] loop.py:45 INFO TRAIN_STEP step=17162 loss=1.210956 Training: 95%|█████████▌| 17162/18018 [36:41:33<1:48:00, 7.57s/step, loss=1.211, acc=0.714, acc_len=1.73, thru=18.3, I=21.3, T=37.4, wait=2.5s, pool=49, epoch=2/2][2026-10-20 02:11:57,034] loop.py:45 INFO TRAIN_STEP step=17163 loss=1.198505 Training: 95%|█████████▌| 17163/18018 [36:41:43<1:56:17, 8.16s/step, loss=1.199, acc=0.717, acc_len=1.74, thru=15.2, I=14.4, T=33.6, wait=5.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:12:03,988] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:12:04,455] loop.py:45 INFO TRAIN_STEP step=17164 loss=1.329137 Training: 95%|█████████▌| 17164/18018 [36:41:50<1:53:00, 7.94s/step, loss=1.329, acc=0.696, acc_len=1.67, thru=16.2, I=18.0, T=36.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:12:04,494] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-20 02:12:11,987] loop.py:380 INFO TIMING step=17165: step=3.893s data=1.758s compute=3.850s [fwd=2.779s bwd=1.056s opt=0.003s] dispatch=3.630s [2026-10-20 02:12:11,987] loop.py:45 INFO TRAIN_STEP step=17165 loss=1.255687 Training: 95%|█████████▌| 17165/18018 [36:41:58<1:51:08, 7.82s/step, loss=1.256, acc=0.705, acc_len=1.70, thru=13.6, I=16.1, T=32.9, wait=3.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:12:21,590] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:12:22,225] loop.py:45 INFO TRAIN_STEP step=17166 loss=1.335385 Training: 95%|█████████▌| 17166/18018 [36:42:08<2:01:18, 8.54s/step, loss=1.335, acc=0.691, acc_len=1.65, thru=15.8, I=15.1, T=31.5, wait=6.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:12:22,221] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17165: forward=2787.9ms backward=1201.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:12:22,595] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:12:29,947] loop.py:45 INFO TRAIN_STEP step=17167 loss=1.266993 Training: 95%|█████████▌| 17167/18018 [36:42:16<1:57:40, 8.30s/step, loss=1.267, acc=0.705, acc_len=1.70, thru=14.6, I=15.9, T=36.9, wait=4.2s, pool=41, epoch=2/2][2026-10-20 02:12:37,265] loop.py:45 INFO TRAIN_STEP step=17168 loss=1.253022 Training: 95%|█████████▌| 17168/18018 [36:42:23<1:53:22, 8.00s/step, loss=1.253, acc=0.707, acc_len=1.70, thru=18.4, I=17.6, T=35.0, wait=3.6s, pool=61, epoch=2/2][2026-10-20 02:12:44,276] loop.py:45 INFO TRAIN_STEP step=17169 loss=1.227489 Training: 95%|█████████▌| 17169/18018 [36:42:30<1:49:02, 7.71s/step, loss=1.227, acc=0.716, acc_len=1.73, thru=17.6, I=16.8, T=35.5, wait=3.4s, pool=49, epoch=2/2][2026-10-20 02:12:51,249] loop.py:380 INFO TIMING step=17170: step=3.677s data=1.527s compute=3.649s [fwd=2.565s bwd=1.070s opt=0.003s] dispatch=3.286s [2026-10-20 02:12:51,249] loop.py:45 INFO TRAIN_STEP step=17170 loss=1.205040 Training: 95%|█████████▌| 17170/18018 [36:42:37<1:45:48, 7.49s/step, loss=1.205, acc=0.718, acc_len=1.74, thru=16.5, I=17.5, T=34.8, wait=3.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:12:56,853] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:12:58,323] loop.py:45 INFO TRAIN_STEP step=17171 loss=1.274948 (TrainerActor pid=9310) [2026-10-20 02:12:58,318] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17170: forward=2579.0ms backward=978.5ms Training: 95%|█████████▌| 17171/18018 [36:42:44<1:43:55, 7.36s/step, loss=1.275, acc=0.706, acc_len=1.70, thru=19.0, I=19.9, T=35.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:12:58,378] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:03,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:13:03,887] loop.py:45 INFO TRAIN_STEP step=17172 loss=1.316215 Training: 95%|█████████▌| 17172/18018 [36:42:49<1:36:12, 6.82s/step, loss=1.316, acc=0.696, acc_len=1.67, thru=22.1, I=26.6, T=38.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:03,953] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:09,518] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:13:10,132] loop.py:45 INFO TRAIN_STEP step=17173 loss=1.143139 Training: 95%|█████████▌| 17173/18018 [36:42:56<1:33:38, 6.65s/step, loss=1.143, acc=0.726, acc_len=1.77, thru=21.1, I=23.7, T=37.0, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:10,524] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:16,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:13:17,896] loop.py:45 INFO TRAIN_STEP step=17174 loss=1.478721 Training: 95%|█████████▌| 17174/18018 [36:43:03<1:38:13, 6.98s/step, loss=1.479, acc=0.669, acc_len=1.57, thru=17.3, I=21.0, T=33.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:18,026] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-20 02:13:23,134] loop.py:380 INFO TIMING step=17175: step=3.422s data=1.496s compute=3.397s [fwd=2.425s bwd=0.957s opt=0.003s] dispatch=1.808s [2026-10-20 02:13:23,135] loop.py:45 INFO TRAIN_STEP step=17175 loss=1.110548 Training: 95%|█████████▌| 17175/18018 [36:43:09<1:30:45, 6.46s/step, loss=1.111, acc=0.731, acc_len=1.79, thru=18.2, I=23.1, T=37.4, wait=1.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:33,084] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:13:33,924] loop.py:45 INFO TRAIN_STEP step=17176 loss=1.262691 (TrainerActor pid=9310) [2026-10-20 02:13:33,920] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17175: forward=2738.6ms backward=1007.4ms Training: 95%|█████████▌| 17176/18018 [36:43:20<1:48:53, 7.76s/step, loss=1.263, acc=0.705, acc_len=1.70, thru=14.8, I=13.8, T=33.7, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:34,090] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:13:39,934] loop.py:45 INFO TRAIN_STEP step=17177 loss=1.161233 Training: 95%|█████████▌| 17177/18018 [36:43:26<1:41:23, 7.23s/step, loss=1.161, acc=0.726, acc_len=1.77, thru=16.7, I=19.6, T=36.6, wait=2.5s, pool=33, epoch=2/2][2026-10-20 02:13:47,965] loop.py:45 INFO TRAIN_STEP step=17178 loss=1.355150 Training: 95%|█████████▌| 17178/18018 [36:43:34<1:44:37, 7.47s/step, loss=1.355, acc=0.691, acc_len=1.65, thru=16.9, I=15.1, T=36.2, wait=4.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:54,228] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:13:55,039] loop.py:45 INFO TRAIN_STEP step=17179 loss=1.358316 Training: 95%|█████████▌| 17179/18018 [36:43:41<1:42:50, 7.35s/step, loss=1.358, acc=0.687, acc_len=1.64, thru=18.6, I=19.6, T=35.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:13:55,231] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:14:02,486] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:14:03,571] loop.py:380 INFO TIMING step=17180: step=3.670s data=1.529s compute=3.642s [fwd=2.594s bwd=1.033s opt=0.003s] dispatch=4.851s [2026-10-20 02:14:03,571] loop.py:45 INFO TRAIN_STEP step=17180 loss=1.218685 Training: 95%|█████████▌| 17180/18018 [36:43:49<1:47:38, 7.71s/step, loss=1.219, acc=0.714, acc_len=1.73, thru=14.5, I=16.4, T=34.9, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:14:03,996] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:14:10,120] loop.py:45 INFO TRAIN_STEP step=17181 loss=1.307963 Training: 95%|█████████▌| 17181/18018 [36:43:56<1:42:39, 7.36s/step, loss=1.308, acc=0.696, acc_len=1.67, thru=18.5, I=20.5, T=37.7, wait=3.1s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:14:10,116] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17180: forward=2358.6ms backward=978.8ms [2026-10-20 02:14:17,592] loop.py:45 INFO TRAIN_STEP step=17182 loss=1.261544 Training: 95%|█████████▌| 17182/18018 [36:44:03<1:43:00, 7.39s/step, loss=1.262, acc=0.705, acc_len=1.70, thru=17.8, I=17.3, T=36.5, wait=4.0s, pool=61, epoch=2/2][2026-10-20 02:14:25,317] loop.py:45 INFO TRAIN_STEP step=17183 loss=1.335402 Training: 95%|█████████▌| 17183/18018 [36:44:11<1:44:16, 7.49s/step, loss=1.335, acc=0.690, acc_len=1.64, thru=17.0, I=16.0, T=33.9, wait=3.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:14:33,426] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:14:34,053] loop.py:45 INFO TRAIN_STEP step=17184 loss=1.198388 Training: 95%|█████████▌| 17184/18018 [36:44:20<1:49:20, 7.87s/step, loss=1.198, acc=0.717, acc_len=1.74, thru=16.6, I=15.5, T=32.4, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:14:34,430] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:14:39,894] loop.py:380 INFO TIMING step=17185: step=3.533s data=1.474s compute=3.499s [fwd=2.468s bwd=1.016s opt=0.003s] dispatch=2.300s [2026-10-20 02:14:39,894] loop.py:45 INFO TRAIN_STEP step=17185 loss=1.397872 Training: 95%|█████████▌| 17185/18018 [36:44:25<1:40:45, 7.26s/step, loss=1.398, acc=0.681, acc_len=1.62, thru=18.3, I=21.8, T=36.2, wait=2.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:14:49,337] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:14:50,156] loop.py:45 INFO TRAIN_STEP step=17186 loss=1.229532 Training: 95%|█████████▌| 17186/18018 [36:44:36<1:53:09, 8.16s/step, loss=1.230, acc=0.712, acc_len=1.72, thru=16.2, I=14.3, T=33.0, wait=6.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:14:50,152] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17185: forward=2727.0ms backward=1094.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:14:50,343] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:14:56,398] loop.py:45 INFO TRAIN_STEP step=17187 loss=1.232419 Training: 95%|█████████▌| 17187/18018 [36:44:42<1:45:02, 7.58s/step, loss=1.232, acc=0.710, acc_len=1.71, thru=18.8, I=19.9, T=33.3, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:15:02,407] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:15:03,311] loop.py:45 INFO TRAIN_STEP step=17188 loss=1.215498 Training: 95%|█████████▌| 17188/18018 [36:44:49<1:42:08, 7.38s/step, loss=1.215, acc=0.715, acc_len=1.73, thru=19.1, I=19.9, T=36.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:15:03,427] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-20 02:15:10,790] loop.py:45 INFO TRAIN_STEP step=17189 loss=1.287562 Training: 95%|█████████▌| 17189/18018 [36:44:56<1:42:24, 7.41s/step, loss=1.288, acc=0.701, acc_len=1.68, thru=15.4, I=16.9, T=34.3, wait=3.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:15:19,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:15:19,331] loop.py:380 INFO TIMING step=17190: step=3.585s data=1.475s compute=3.541s [fwd=2.499s bwd=1.027s opt=0.003s] dispatch=4.946s [2026-10-20 02:15:19,332] loop.py:45 INFO TRAIN_STEP step=17190 loss=1.282690 Training: 95%|█████████▌| 17190/18018 [36:45:05<1:46:57, 7.75s/step, loss=1.283, acc=0.705, acc_len=1.70, thru=17.8, I=17.2, T=35.7, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:15:19,726] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:15:26,829] loop.py:45 INFO TRAIN_STEP step=17191 loss=1.225315 Training: 95%|█████████▌| 17191/18018 [36:45:12<1:45:46, 7.67s/step, loss=1.225, acc=0.713, acc_len=1.72, thru=15.6, I=15.8, T=36.6, wait=4.0s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:15:26,825] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17190: forward=2465.3ms backward=972.0ms [2026-10-20 02:15:35,654] loop.py:45 INFO TRAIN_STEP step=17192 loss=1.206508 Training: 95%|█████████▌| 17192/18018 [36:45:21<1:50:24, 8.02s/step, loss=1.207, acc=0.716, acc_len=1.73, thru=13.5, I=13.2, T=35.9, wait=5.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:15:43,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:15:43,164] loop.py:45 INFO TRAIN_STEP step=17193 loss=1.134183 Training: 95%|█████████▌| 17193/18018 [36:45:29<1:48:10, 7.87s/step, loss=1.134, acc=0.733, acc_len=1.80, thru=18.7, I=17.0, T=35.2, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:15:43,552] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:15:49,360] loop.py:45 INFO TRAIN_STEP step=17194 loss=1.232177 Training: 95%|█████████▌| 17194/18018 [36:45:35<1:41:09, 7.37s/step, loss=1.232, acc=0.709, acc_len=1.71, thru=17.8, I=22.7, T=36.6, wait=2.7s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:15:55,707] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:15:57,363] loop.py:380 INFO TIMING step=17195: step=3.947s data=1.602s compute=3.913s [fwd=2.774s bwd=1.124s opt=0.003s] dispatch=4.048s [2026-10-20 02:15:57,363] loop.py:45 INFO TRAIN_STEP step=17195 loss=1.375915 Training: 95%|█████████▌| 17195/18018 [36:45:43<1:43:38, 7.56s/step, loss=1.376, acc=0.690, acc_len=1.64, thru=16.9, I=19.1, T=32.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:15:57,711] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:01,967] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:16:03,112] loop.py:45 INFO TRAIN_STEP step=17196 loss=1.119449 Training: 95%|█████████▌| 17196/18018 [36:45:49<1:36:05, 7.01s/step, loss=1.119, acc=0.732, acc_len=1.79, thru=20.2, I=27.8, T=37.9, wait=2.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:16:03,107] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17195: forward=2384.3ms backward=949.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:03,475] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:16:12,257] loop.py:45 INFO TRAIN_STEP step=17197 loss=1.340527 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:12,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 95%|█████████▌| 17197/18018 [36:45:58<1:44:43, 7.65s/step, loss=1.341, acc=0.694, acc_len=1.67, thru=14.6, I=13.9, T=32.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:12,737] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:19,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:16:21,474] loop.py:45 INFO TRAIN_STEP step=17198 loss=1.283813 Training: 95%|█████████▌| 17198/18018 [36:46:07<1:51:00, 8.12s/step, loss=1.284, acc=0.706, acc_len=1.70, thru=14.5, I=17.8, T=33.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:21,934] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:27,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:16:28,212] loop.py:45 INFO TRAIN_STEP step=17199 loss=1.246621 Training: 95%|█████████▌| 17199/18018 [36:46:14<1:45:12, 7.71s/step, loss=1.247, acc=0.713, acc_len=1.72, thru=18.0, I=25.3, T=37.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:28,261] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:36,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:16:36,601] loop.py:380 INFO TIMING step=17200: step=3.697s data=1.529s compute=3.666s [fwd=2.618s bwd=1.036s opt=0.003s] dispatch=4.682s [2026-10-20 02:16:36,601] loop.py:45 INFO TRAIN_STEP step=17200 loss=1.403630 Training: 95%|█████████▌| 17200/18018 [36:46:22<1:47:52, 7.91s/step, loss=1.404, acc=0.678, acc_len=1.61, thru=14.6, I=17.1, T=34.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:16:36,728] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-20 02:16:43,976] loop.py:45 INFO TRAIN_STEP step=17201 loss=1.261537 Training: 95%|█████████▌| 17201/18018 [36:46:30<1:45:31, 7.75s/step, loss=1.262, acc=0.704, acc_len=1.69, thru=15.1, I=14.3, T=33.3, wait=3.5s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:16:43,971] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17200: forward=2710.0ms backward=1088.2ms [2026-10-20 02:16:51,939] loop.py:45 INFO TRAIN_STEP step=17202 loss=1.169563 Training: 95%|█████████▌| 17202/18018 [36:46:38<1:46:16, 7.81s/step, loss=1.170, acc=0.726, acc_len=1.78, thru=16.5, I=14.4, T=35.2, wait=4.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:00,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:17:01,238] loop.py:45 INFO TRAIN_STEP step=17203 loss=1.371370 Training: 95%|█████████▌| 17203/18018 [36:46:47<1:52:12, 8.26s/step, loss=1.371, acc=0.685, acc_len=1.63, thru=16.4, I=17.3, T=32.3, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:01,270] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:07,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:17:08,285] loop.py:45 INFO TRAIN_STEP step=17204 loss=1.282382 Training: 95%|█████████▌| 17204/18018 [36:46:54<1:47:07, 7.90s/step, loss=1.282, acc=0.703, acc_len=1.69, thru=17.0, I=18.8, T=34.2, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:08,731] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:17:14,631] loop.py:380 INFO TIMING step=17205: step=3.539s data=1.474s compute=3.513s [fwd=2.522s bwd=0.976s opt=0.003s] dispatch=2.799s [2026-10-20 02:17:14,631] loop.py:45 INFO TRAIN_STEP step=17205 loss=1.286897 Training: 95%|█████████▌| 17205/18018 [36:47:00<1:40:41, 7.43s/step, loss=1.287, acc=0.698, acc_len=1.67, thru=18.4, I=22.1, T=36.2, wait=2.8s, pool=57, epoch=2/2][2026-10-20 02:17:20,811] loop.py:45 INFO TRAIN_STEP step=17206 loss=1.252545 Training: 95%|█████████▌| 17206/18018 [36:47:06<1:35:29, 7.06s/step, loss=1.253, acc=0.709, acc_len=1.70, thru=17.6, I=19.5, T=35.5, wait=2.6s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:17:20,807] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17205: forward=2394.5ms backward=1130.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:28,917] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:17:29,823] loop.py:45 INFO TRAIN_STEP step=17207 loss=1.314085 Training: 95%|█████████▌| 17207/18018 [36:47:15<1:43:18, 7.64s/step, loss=1.314, acc=0.697, acc_len=1.67, thru=16.3, I=16.3, T=33.9, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:29,928] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:36,036] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:17:36,785] loop.py:45 INFO TRAIN_STEP step=17208 loss=1.252061 Training: 96%|█████████▌| 17208/18018 [36:47:22<1:40:25, 7.44s/step, loss=1.252, acc=0.711, acc_len=1.72, thru=18.4, I=20.1, T=32.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:37,039] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:43,602] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:17:43,830] loop.py:45 INFO TRAIN_STEP step=17209 loss=1.135229 Training: 96%|█████████▌| 17209/18018 [36:47:29<1:38:41, 7.32s/step, loss=1.135, acc=0.730, acc_len=1.79, thru=17.8, I=19.6, T=31.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:44,105] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:49,117] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:17:50,209] loop.py:380 INFO TIMING step=17210: step=3.422s data=1.366s compute=3.388s [fwd=2.440s bwd=0.934s opt=0.003s] dispatch=2.950s [2026-10-20 02:17:50,209] loop.py:45 INFO TRAIN_STEP step=17210 loss=1.217549 Training: 96%|█████████▌| 17210/18018 [36:47:36<1:34:46, 7.04s/step, loss=1.218, acc=0.711, acc_len=1.72, thru=18.5, I=21.4, T=37.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:17:50,626] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:17:57,180] loop.py:45 INFO TRAIN_STEP step=17211 loss=1.272960 Training: 96%|█████████▌| 17211/18018 [36:47:43<1:34:23, 7.02s/step, loss=1.273, acc=0.703, acc_len=1.68, thru=14.9, I=19.8, T=35.7, wait=3.4s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:17:57,175] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17210: forward=2545.9ms backward=999.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:02,567] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:18:04,018] loop.py:45 INFO TRAIN_STEP step=17212 loss=1.150489 Training: 96%|█████████▌| 17212/18018 [36:47:50<1:33:32, 6.96s/step, loss=1.150, acc=0.727, acc_len=1.78, thru=19.5, I=20.5, T=33.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:04,075] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:09,997] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:18:10,945] loop.py:45 INFO TRAIN_STEP step=17213 loss=1.310276 Training: 96%|█████████▌| 17213/18018 [36:47:57<1:33:17, 6.95s/step, loss=1.310, acc=0.695, acc_len=1.67, thru=17.8, I=23.0, T=36.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:11,019] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:17,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:18:18,180] loop.py:45 INFO TRAIN_STEP step=17214 loss=1.286788 Training: 96%|█████████▌| 17214/18018 [36:48:04<1:34:18, 7.04s/step, loss=1.287, acc=0.699, acc_len=1.68, thru=17.0, I=19.2, T=35.3, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:18,317] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-20 02:18:23,670] loop.py:380 INFO TIMING step=17215: step=3.295s data=1.313s compute=3.262s [fwd=2.323s bwd=0.926s opt=0.003s] dispatch=2.184s [2026-10-20 02:18:23,670] loop.py:45 INFO TRAIN_STEP step=17215 loss=1.256383 Training: 96%|█████████▌| 17215/18018 [36:48:09<1:27:58, 6.57s/step, loss=1.256, acc=0.703, acc_len=1.69, thru=19.7, I=19.3, T=38.8, wait=2.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:31,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:18:31,865] loop.py:45 INFO TRAIN_STEP step=17216 loss=1.257919 Training: 96%|█████████▌| 17216/18018 [36:48:17<1:34:21, 7.06s/step, loss=1.258, acc=0.706, acc_len=1.70, thru=17.6, I=16.6, T=33.0, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:18:31,861] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17215: forward=2724.2ms backward=1099.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:32,237] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:18:40,244] loop.py:45 INFO TRAIN_STEP step=17217 loss=1.368486 Training: 96%|█████████▌| 17217/18018 [36:48:26<1:39:31, 7.46s/step, loss=1.368, acc=0.688, acc_len=1.64, thru=15.3, I=15.4, T=37.2, wait=4.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:47,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:18:48,182] loop.py:45 INFO TRAIN_STEP step=17218 loss=1.302333 Training: 96%|█████████▌| 17218/18018 [36:48:34<1:41:20, 7.60s/step, loss=1.302, acc=0.698, acc_len=1.67, thru=17.0, I=17.8, T=35.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:48,268] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:53,900] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:18:56,049] loop.py:45 INFO TRAIN_STEP step=17219 loss=1.300317 Training: 96%|█████████▌| 17219/18018 [36:48:42<1:42:16, 7.68s/step, loss=1.300, acc=0.696, acc_len=1.67, thru=16.6, I=21.2, T=25.1, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:18:56,416] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-20 02:19:03,744] loop.py:380 INFO TIMING step=17220: step=3.585s data=1.529s compute=3.556s [fwd=2.528s bwd=1.015s opt=0.003s] dispatch=4.102s [2026-10-20 02:19:03,745] loop.py:45 INFO TRAIN_STEP step=17220 loss=1.287297 Training: 96%|█████████▌| 17220/18018 [36:48:49<1:42:12, 7.69s/step, loss=1.287, acc=0.701, acc_len=1.68, thru=12.6, I=17.5, T=35.7, wait=4.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:19:09,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:19:09,968] loop.py:45 INFO TRAIN_STEP step=17221 loss=1.353174 Training: 96%|█████████▌| 17221/18018 [36:48:56<1:36:15, 7.25s/step, loss=1.353, acc=0.689, acc_len=1.64, thru=19.9, I=18.9, T=38.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:19:09,995] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-20 02:19:09,963] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17220: forward=2354.7ms backward=905.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:19:15,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:19:15,863] loop.py:45 INFO TRAIN_STEP step=17222 loss=1.250800 Training: 96%|█████████▌| 17222/18018 [36:49:01<1:30:45, 6.84s/step, loss=1.251, acc=0.708, acc_len=1.71, thru=21.9, I=23.6, T=38.2, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:19:16,046] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:19:21,656] loop.py:45 INFO TRAIN_STEP step=17223 loss=1.278838 Training: 96%|█████████▌| 17223/18018 [36:49:07<1:26:28, 6.53s/step, loss=1.279, acc=0.706, acc_len=1.70, thru=19.6, I=21.3, T=37.9, wait=2.4s, pool=49, epoch=2/2][2026-10-20 02:19:30,498] loop.py:45 INFO TRAIN_STEP step=17224 loss=1.228893 Training: 96%|█████████▌| 17224/18018 [36:49:16<1:35:33, 7.22s/step, loss=1.229, acc=0.717, acc_len=1.74, thru=14.8, I=14.4, T=35.4, wait=5.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:19:38,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:19:39,362] loop.py:380 INFO TIMING step=17225: step=3.736s data=1.549s compute=3.704s [fwd=2.652s bwd=1.039s opt=0.003s] dispatch=5.120s [2026-10-20 02:19:39,362] loop.py:45 INFO TRAIN_STEP step=17225 loss=1.324374 Training: 96%|█████████▌| 17225/18018 [36:49:25<1:41:57, 7.71s/step, loss=1.324, acc=0.695, acc_len=1.66, thru=16.4, I=15.7, T=34.3, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:19:39,634] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:19:45,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:19:46,273] loop.py:45 INFO TRAIN_STEP step=17226 loss=1.164438 Training: 96%|█████████▌| 17226/18018 [36:49:32<1:38:38, 7.47s/step, loss=1.164, acc=0.724, acc_len=1.76, thru=17.4, I=21.6, T=38.5, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:19:46,268] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17225: forward=2354.2ms backward=924.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:19:46,531] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:19:53,066] loop.py:45 INFO TRAIN_STEP step=17227 loss=1.192068 Training: 96%|█████████▌| 17227/18018 [36:49:39<1:35:50, 7.27s/step, loss=1.192, acc=0.721, acc_len=1.75, thru=16.5, I=22.6, T=38.1, wait=3.4s, pool=41, epoch=2/2](TrainerActor pid=9370) [2026-10-20 02:19:53,124] eagle_store.py:517 WARNING batch_get_buffer missing keys (k551266-a2daa256_hs, k551266-a2daa256_ids, k551266-a2daa256_lhs); sleeping 0.50s. [2026-10-20 02:20:02,629] loop.py:45 INFO TRAIN_STEP step=17228 loss=1.322001 Training: 96%|█████████▌| 17228/18018 [36:49:48<1:44:46, 7.96s/step, loss=1.322, acc=0.700, acc_len=1.68, thru=13.4, I=13.8, T=32.1, wait=5.6s, pool=41, epoch=2/2][2026-10-20 02:20:10,579] loop.py:45 INFO TRAIN_STEP step=17229 loss=1.225847 Training: 96%|█████████▌| 17229/18018 [36:49:56<1:44:36, 7.95s/step, loss=1.226, acc=0.713, acc_len=1.72, thru=15.8, I=13.5, T=36.0, wait=4.4s, pool=45, epoch=2/2][2026-10-20 02:20:17,135] loop.py:380 INFO TIMING step=17230: step=3.352s data=1.319s compute=3.325s [fwd=2.407s bwd=0.904s opt=0.003s] dispatch=3.196s [2026-10-20 02:20:17,135] loop.py:45 INFO TRAIN_STEP step=17230 loss=1.254639 Training: 96%|█████████▌| 17230/18018 [36:50:03<1:38:57, 7.54s/step, loss=1.255, acc=0.707, acc_len=1.70, thru=19.1, I=19.3, T=38.2, wait=3.2s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:20:25,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 02:19:53,126] eagle_store.py:517 WARNING batch_get_buffer missing keys (k551267-801aaff9_hs, k551267-801aaff9_ids, k551267-801aaff9_lhs); sleeping 0.50s. [repeated 3x across cluster] [2026-10-20 02:20:27,425] loop.py:45 INFO TRAIN_STEP step=17231 loss=1.329009 Training: 96%|█████████▌| 17231/18018 [36:50:13<1:49:40, 8.36s/step, loss=1.329, acc=0.696, acc_len=1.66, thru=13.6, I=14.7, T=28.9, wait=5.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:20:27,421] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17230: forward=3243.9ms backward=1111.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:20:27,725] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:20:33,401] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:20:34,546] loop.py:45 INFO TRAIN_STEP step=17232 loss=1.252188 Training: 96%|█████████▌| 17232/18018 [36:50:20<1:44:39, 7.99s/step, loss=1.252, acc=0.709, acc_len=1.71, thru=15.0, I=20.3, T=37.7, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:20:34,907] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:20:40,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:20:40,307] loop.py:45 INFO TRAIN_STEP step=17233 loss=1.152550 Training: 96%|█████████▌| 17233/18018 [36:50:26<1:35:47, 7.32s/step, loss=1.153, acc=0.726, acc_len=1.77, thru=20.7, I=22.5, T=37.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:20:40,635] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:20:46,123] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:20:47,139] loop.py:45 INFO TRAIN_STEP step=17234 loss=1.298237 Training: 96%|█████████▌| 17234/18018 [36:50:33<1:33:44, 7.17s/step, loss=1.298, acc=0.698, acc_len=1.67, thru=19.8, I=21.7, T=33.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:20:47,634] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:20:53,884] loop.py:380 INFO TIMING step=17235: step=3.371s data=1.379s compute=3.332s [fwd=2.340s bwd=0.978s opt=0.003s] dispatch=3.364s [2026-10-20 02:20:53,884] loop.py:45 INFO TRAIN_STEP step=17235 loss=1.192269 Training: 96%|█████████▌| 17235/18018 [36:50:39<1:31:56, 7.05s/step, loss=1.192, acc=0.718, acc_len=1.74, thru=17.0, I=19.7, T=38.0, wait=3.4s, pool=57, epoch=2/2][2026-10-20 02:21:01,911] loop.py:45 INFO TRAIN_STEP step=17236 loss=1.194144 Training: 96%|█████████▌| 17236/18018 [36:50:47<1:35:39, 7.34s/step, loss=1.194, acc=0.719, acc_len=1.74, thru=15.8, I=17.3, T=36.9, wait=4.5s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:21:01,906] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17235: forward=2409.2ms backward=1009.2ms [2026-10-20 02:21:09,124] loop.py:45 INFO TRAIN_STEP step=17237 loss=1.245465 Training: 96%|█████████▌| 17237/18018 [36:50:55<1:35:03, 7.30s/step, loss=1.245, acc=0.709, acc_len=1.72, thru=17.7, I=20.5, T=36.1, wait=3.7s, pool=61, epoch=2/2][2026-10-20 02:21:16,963] loop.py:45 INFO TRAIN_STEP step=17238 loss=1.202872 Training: 96%|█████████▌| 17238/18018 [36:51:03<1:37:00, 7.46s/step, loss=1.203, acc=0.719, acc_len=1.74, thru=11.1, I=12.8, T=36.3, wait=4.3s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 02:21:17,014] eagle_store.py:517 WARNING batch_get_buffer missing keys (k551614-8521e89e_hs, k551614-8521e89e_ids, k551614-8521e89e_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:21:26,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9371) [2026-10-20 02:21:17,322] eagle_store.py:517 WARNING batch_get_buffer missing keys (k551621-9c163a7a_hs, k551621-9c163a7a_ids, k551621-9c163a7a_lhs); sleeping 0.50s. [repeated 5x across cluster] [2026-10-20 02:21:28,209] loop.py:45 INFO TRAIN_STEP step=17239 loss=1.229378 Training: 96%|█████████▌| 17239/18018 [36:51:14<1:51:37, 8.60s/step, loss=1.229, acc=0.713, acc_len=1.72, thru=16.6, I=15.5, T=32.4, wait=7.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:21:28,580] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 02:21:34,134] loop.py:380 INFO TIMING step=17240: step=3.409s data=1.321s compute=3.371s [fwd=2.429s bwd=0.927s opt=0.003s] dispatch=2.506s [2026-10-20 02:21:34,134] loop.py:45 INFO TRAIN_STEP step=17240 loss=1.258272 Training: 96%|█████████▌| 17240/18018 [36:51:20<1:41:05, 7.80s/step, loss=1.258, acc=0.706, acc_len=1.70, thru=19.4, I=27.2, T=37.5, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:21:41,165] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:21:41,919] loop.py:45 INFO TRAIN_STEP step=17241 loss=1.356488 Training: 96%|█████████▌| 17241/18018 [36:51:28<1:40:54, 7.79s/step, loss=1.356, acc=0.692, acc_len=1.64, thru=16.1, I=17.1, T=36.5, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:21:41,914] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17240: forward=2440.8ms backward=1003.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:21:42,172] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:21:47,640] loop.py:45 INFO TRAIN_STEP step=17242 loss=1.199209 Training: 96%|█████████▌| 17242/18018 [36:51:33<1:32:44, 7.17s/step, loss=1.199, acc=0.713, acc_len=1.73, thru=20.6, I=21.6, T=39.0, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:21:55,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:21:56,136] loop.py:45 INFO TRAIN_STEP step=17243 loss=1.271004 Training: 96%|█████████▌| 17243/18018 [36:51:42<1:37:45, 7.57s/step, loss=1.271, acc=0.705, acc_len=1.70, thru=16.1, I=16.9, T=36.0, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:21:56,414] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:22:03,390] loop.py:45 INFO TRAIN_STEP step=17244 loss=1.394524 Training: 96%|█████████▌| 17244/18018 [36:51:49<1:36:24, 7.47s/step, loss=1.395, acc=0.679, acc_len=1.61, thru=16.8, I=19.7, T=37.4, wait=3.8s, pool=61, epoch=2/2][2026-10-20 02:22:10,526] loop.py:380 INFO TIMING step=17245: step=3.563s data=1.472s compute=3.520s [fwd=2.520s bwd=0.986s opt=0.003s] dispatch=3.565s [2026-10-20 02:22:10,526] loop.py:45 INFO TRAIN_STEP step=17245 loss=1.326585 Training: 96%|█████████▌| 17245/18018 [36:51:56<1:34:58, 7.37s/step, loss=1.327, acc=0.694, acc_len=1.67, thru=17.7, I=17.5, T=35.9, wait=3.6s, pool=57, epoch=2/2][2026-10-20 02:22:17,979] loop.py:45 INFO TRAIN_STEP step=17246 loss=1.397687 Training: 96%|█████████▌| 17246/18018 [36:52:04<1:35:10, 7.40s/step, loss=1.398, acc=0.684, acc_len=1.63, thru=17.9, I=18.6, T=36.5, wait=3.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:22:17,974] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17245: forward=2491.7ms backward=965.1ms [2026-10-20 02:22:27,394] loop.py:45 INFO TRAIN_STEP step=17247 loss=1.336978 Training: 96%|█████████▌| 17247/18018 [36:52:13<1:42:49, 8.00s/step, loss=1.337, acc=0.698, acc_len=1.68, thru=13.5, I=12.6, T=33.2, wait=5.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:22:32,715] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:22:34,159] loop.py:45 INFO TRAIN_STEP step=17248 loss=1.267366 Training: 96%|█████████▌| 17248/18018 [36:52:20<1:37:55, 7.63s/step, loss=1.267, acc=0.705, acc_len=1.70, thru=19.3, I=19.0, T=35.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:22:34,222] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-20 02:22:40,177] loop.py:45 INFO TRAIN_STEP step=17249 loss=1.312028 Training: 96%|█████████▌| 17249/18018 [36:52:26<1:31:36, 7.15s/step, loss=1.312, acc=0.694, acc_len=1.65, thru=19.0, I=23.5, T=35.5, wait=2.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:22:48,101] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:22:49,012] loop.py:380 INFO TIMING step=17250: step=3.972s data=1.754s compute=3.941s [fwd=2.886s bwd=1.041s opt=0.003s] dispatch=4.855s [2026-10-20 02:22:49,012] loop.py:45 INFO TRAIN_STEP step=17250 loss=1.273791 Training: 96%|█████████▌| 17250/18018 [36:52:35<1:37:58, 7.65s/step, loss=1.274, acc=0.706, acc_len=1.70, thru=16.4, I=16.4, T=32.2, wait=4.9s, pool=65, epoch=2/2][2026-10-20 02:22:49,016] loop.py:419 INFO Saving checkpoint at step 17250... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:22:53,137] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 02:22:53,117] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0017251 (TrainerActor pid=9310) [2026-10-20 02:22:53,271] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0016501 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:22:54,644] inference_manager.py:408 INFO Pool capacity available after 6.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:00,202] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 02:22:53,119] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0017251 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 02:22:54,213] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0016501 [2026-10-20 02:23:01,038] loop.py:45 INFO TRAIN_STEP step=17251 loss=1.298320 Training: 96%|█████████▌| 17251/18018 [36:52:47<1:54:36, 8.97s/step, loss=1.298, acc=0.702, acc_len=1.69, thru=20.0, I=20.3, T=33.2, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:23:01,034] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17250: forward=2807.1ms backward=992.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:01,209] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:08,217] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:23:08,360] loop.py:45 INFO TRAIN_STEP step=17252 loss=1.264391 Training: 96%|█████████▌| 17252/18018 [36:52:54<1:48:09, 8.47s/step, loss=1.264, acc=0.711, acc_len=1.72, thru=16.2, I=17.6, T=37.5, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:08,718] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:17,319] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:23:18,291] loop.py:45 INFO TRAIN_STEP step=17253 loss=1.273912 Training: 96%|█████████▌| 17253/18018 [36:53:04<1:53:36, 8.91s/step, loss=1.274, acc=0.702, acc_len=1.69, thru=13.4, I=14.7, T=31.9, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:18,339] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:24,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:23:24,643] loop.py:45 INFO TRAIN_STEP step=17254 loss=1.294183 Training: 96%|█████████▌| 17254/18018 [36:53:10<1:43:40, 8.14s/step, loss=1.294, acc=0.703, acc_len=1.68, thru=18.9, I=19.5, T=37.6, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:25,009] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:30,395] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:23:31,274] loop.py:380 INFO TIMING step=17255: step=3.416s data=1.394s compute=3.387s [fwd=2.480s bwd=0.892s opt=0.003s] dispatch=3.208s [2026-10-20 02:23:31,274] loop.py:45 INFO TRAIN_STEP step=17255 loss=1.184029 Training: 96%|█████████▌| 17255/18018 [36:53:17<1:37:47, 7.69s/step, loss=1.184, acc=0.720, acc_len=1.75, thru=20.0, I=21.3, T=37.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:31,413] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:39,676] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:23:40,297] loop.py:45 INFO TRAIN_STEP step=17256 loss=1.243392 Training: 96%|█████████▌| 17256/18018 [36:53:26<1:42:43, 8.09s/step, loss=1.243, acc=0.709, acc_len=1.70, thru=13.3, I=15.0, T=35.1, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:23:40,293] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17255: forward=2629.3ms backward=979.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:23:40,681] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:23:47,014] loop.py:45 INFO TRAIN_STEP step=17257 loss=1.292734 Training: 96%|█████████▌| 17257/18018 [36:53:33<1:37:22, 7.68s/step, loss=1.293, acc=0.704, acc_len=1.70, thru=16.6, I=17.1, T=37.7, wait=3.3s, pool=37, epoch=2/2][2026-10-20 02:23:56,820] loop.py:45 INFO TRAIN_STEP step=17258 loss=1.168723 Training: 96%|█████████▌| 17258/18018 [36:53:42<1:45:20, 8.32s/step, loss=1.169, acc=0.723, acc_len=1.76, thru=15.0, I=14.4, T=32.5, wait=5.9s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:03,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:24:04,181] loop.py:45 INFO TRAIN_STEP step=17259 loss=1.124810 Training: 96%|█████████▌| 17259/18018 [36:53:50<1:41:34, 8.03s/step, loss=1.125, acc=0.728, acc_len=1.78, thru=17.1, I=18.7, T=37.3, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:04,590] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:10,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:24:11,154] loop.py:380 INFO TIMING step=17260: step=3.284s data=1.283s compute=3.254s [fwd=2.308s bwd=0.932s opt=0.003s] dispatch=3.682s [2026-10-20 02:24:11,154] loop.py:45 INFO TRAIN_STEP step=17260 loss=1.256265 Training: 96%|█████████▌| 17260/18018 [36:53:57<1:37:25, 7.71s/step, loss=1.256, acc=0.708, acc_len=1.71, thru=18.4, I=23.5, T=39.0, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:11,291] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-20 02:24:18,455] loop.py:45 INFO TRAIN_STEP step=17261 loss=1.272035 Training: 96%|█████████▌| 17261/18018 [36:54:04<1:35:44, 7.59s/step, loss=1.272, acc=0.704, acc_len=1.70, thru=18.1, I=20.8, T=36.7, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:24:18,450] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17260: forward=2431.7ms backward=993.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:18,440] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:18,942] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:25,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:24:26,334] loop.py:45 INFO TRAIN_STEP step=17262 loss=1.245677 Training: 96%|█████████▌| 17262/18018 [36:54:12<1:36:43, 7.68s/step, loss=1.246, acc=0.709, acc_len=1.71, thru=15.9, I=16.6, T=34.2, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:26,808] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:33,097] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:24:33,817] loop.py:45 INFO TRAIN_STEP step=17263 loss=1.283938 Training: 96%|█████████▌| 17263/18018 [36:54:19<1:35:51, 7.62s/step, loss=1.284, acc=0.703, acc_len=1.69, thru=18.1, I=19.2, T=34.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:34,101] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:40,213] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:24:40,417] loop.py:45 INFO TRAIN_STEP step=17264 loss=1.101147 Training: 96%|█████████▌| 17264/18018 [36:54:26<1:31:53, 7.31s/step, loss=1.101, acc=0.738, acc_len=1.82, thru=19.1, I=20.1, T=35.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:40,715] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:47,755] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:24:49,131] loop.py:380 INFO TIMING step=17265: step=4.233s data=1.952s compute=4.197s [fwd=2.937s bwd=1.244s opt=0.003s] dispatch=4.472s [2026-10-20 02:24:49,131] loop.py:45 INFO TRAIN_STEP step=17265 loss=1.247563 Training: 96%|█████████▌| 17265/18018 [36:54:35<1:37:03, 7.73s/step, loss=1.248, acc=0.707, acc_len=1.71, thru=16.4, I=19.6, T=30.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:24:49,267] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 [2026-10-20 02:24:55,025] loop.py:45 INFO TRAIN_STEP step=17266 loss=1.211888 Training: 96%|█████████▌| 17266/18018 [36:54:41<1:30:00, 7.18s/step, loss=1.212, acc=0.713, acc_len=1.73, thru=17.8, I=21.7, T=38.1, wait=2.5s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:24:55,021] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17265: forward=2341.9ms backward=957.9ms [2026-10-20 02:25:02,113] loop.py:45 INFO TRAIN_STEP step=17267 loss=1.238612 Training: 96%|█████████▌| 17267/18018 [36:54:48<1:29:31, 7.15s/step, loss=1.239, acc=0.707, acc_len=1.70, thru=18.4, I=18.9, T=37.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:02,069] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:02,571] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:11,591] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:25:12,153] loop.py:45 INFO TRAIN_STEP step=17268 loss=1.263510 Training: 96%|█████████▌| 17268/18018 [36:54:58<1:40:14, 8.02s/step, loss=1.264, acc=0.709, acc_len=1.71, thru=13.6, I=13.3, T=34.6, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:12,596] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:25:18,962] loop.py:45 INFO TRAIN_STEP step=17269 loss=1.148577 Training: 96%|█████████▌| 17269/18018 [36:55:05<1:35:34, 7.66s/step, loss=1.149, acc=0.728, acc_len=1.78, thru=15.6, I=22.0, T=34.9, wait=3.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:30,463] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:25:31,203] loop.py:380 INFO TIMING step=17270: step=4.314s data=1.976s compute=3.860s [fwd=2.662s bwd=1.183s opt=0.003s] dispatch=7.918s [2026-10-20 02:25:31,204] loop.py:45 INFO TRAIN_STEP step=17270 loss=1.326225 Training: 96%|█████████▌| 17270/18018 [36:55:17<1:52:36, 9.03s/step, loss=1.326, acc=0.695, acc_len=1.66, thru=15.2, I=11.3, T=29.7, wait=7.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:31,468] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:36,560] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 02:25:37,502] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17270: forward=2642.2ms backward=1115.5ms [2026-10-20 02:25:37,507] loop.py:45 INFO TRAIN_STEP step=17271 loss=1.316518 Training: 96%|█████████▌| 17271/18018 [36:55:23<1:42:15, 8.21s/step, loss=1.317, acc=0.695, acc_len=1.67, thru=19.4, I=20.9, T=33.6, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:37,581] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:41,995] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:25:43,456] loop.py:45 INFO TRAIN_STEP step=17272 loss=1.190476 Training: 96%|█████████▌| 17272/18018 [36:55:29<1:33:40, 7.53s/step, loss=1.190, acc=0.721, acc_len=1.75, thru=20.9, I=25.2, T=36.6, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:25:43,512] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-20 02:25:48,927] loop.py:45 INFO TRAIN_STEP step=17273 loss=1.190740 Training: 96%|█████████▌| 17273/18018 [36:55:35<1:25:51, 6.91s/step, loss=1.191, acc=0.716, acc_len=1.73, thru=22.0, I=26.8, T=39.3, wait=2.2s, pool=61, epoch=2/2][2026-10-20 02:25:56,435] loop.py:45 INFO TRAIN_STEP step=17274 loss=1.266698 Training: 96%|█████████▌| 17274/18018 [36:55:42<1:27:57, 7.09s/step, loss=1.267, acc=0.704, acc_len=1.69, thru=17.3, I=17.2, T=36.6, wait=4.0s, pool=53, epoch=2/2][2026-10-20 02:26:03,110] loop.py:380 INFO TIMING step=17275: step=3.289s data=1.343s compute=3.254s [fwd=2.335s bwd=0.905s opt=0.003s] dispatch=3.377s [2026-10-20 02:26:03,110] loop.py:45 INFO TRAIN_STEP step=17275 loss=1.362462 Training: 96%|█████████▌| 17275/18018 [36:55:49<1:26:16, 6.97s/step, loss=1.362, acc=0.688, acc_len=1.64, thru=17.8, I=16.3, T=38.9, wait=3.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:26:10,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:26:11,796] loop.py:45 INFO TRAIN_STEP step=17276 loss=1.376062 Training: 96%|█████████▌| 17276/18018 [36:55:57<1:32:32, 7.48s/step, loss=1.376, acc=0.685, acc_len=1.63, thru=16.7, I=18.1, T=35.2, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:26:11,791] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17275: forward=2487.5ms backward=1072.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:26:11,989] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:26:19,122] loop.py:45 INFO TRAIN_STEP step=17277 loss=1.258462 Training: 96%|█████████▌| 17277/18018 [36:56:05<1:31:49, 7.44s/step, loss=1.258, acc=0.709, acc_len=1.72, thru=15.7, I=19.8, T=35.2, wait=3.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:26:27,213] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:26:28,114] loop.py:45 INFO TRAIN_STEP step=17278 loss=1.195952 Training: 96%|█████████▌| 17278/18018 [36:56:14<1:37:28, 7.90s/step, loss=1.196, acc=0.717, acc_len=1.74, thru=15.3, I=16.1, T=32.1, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:26:28,225] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:26:35,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:26:37,035] loop.py:45 INFO TRAIN_STEP step=17279 loss=1.267570 Training: 96%|█████████▌| 17279/18018 [36:56:23<1:41:06, 8.21s/step, loss=1.268, acc=0.705, acc_len=1.70, thru=13.4, I=17.4, T=31.6, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:26:37,159] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:26:42,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:26:42,947] loop.py:380 INFO TIMING step=17280: step=3.601s data=1.505s compute=3.571s [fwd=2.587s bwd=0.969s opt=0.003s] dispatch=2.300s [2026-10-20 02:26:42,947] loop.py:45 INFO TRAIN_STEP step=17280 loss=1.169019 Training: 96%|█████████▌| 17280/18018 [36:56:29<1:32:29, 7.52s/step, loss=1.169, acc=0.724, acc_len=1.76, thru=19.6, I=23.2, T=35.5, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:26:43,320] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (TrainerActor pid=9310) [2026-10-20 02:26:49,266] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17280: forward=2556.1ms backward=1075.0ms [2026-10-20 02:26:49,271] loop.py:45 INFO TRAIN_STEP step=17281 loss=1.198173 Training: 96%|█████████▌| 17281/18018 [36:56:35<1:27:57, 7.16s/step, loss=1.198, acc=0.718, acc_len=1.74, thru=17.7, I=18.8, T=34.8, wait=2.6s, pool=45, epoch=2/2][2026-10-20 02:27:00,866] loop.py:45 INFO TRAIN_STEP step=17282 loss=1.233354 Training: 96%|█████████▌| 17282/18018 [36:56:46<1:44:09, 8.49s/step, loss=1.233, acc=0.716, acc_len=1.73, thru=12.6, I=11.3, T=29.7, wait=7.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:08,061] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:27:08,485] loop.py:45 INFO TRAIN_STEP step=17283 loss=1.368108 Training: 96%|█████████▌| 17283/18018 [36:56:54<1:40:48, 8.23s/step, loss=1.368, acc=0.686, acc_len=1.63, thru=18.9, I=16.4, T=36.1, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:08,570] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:14,645] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:27:15,365] loop.py:45 INFO TRAIN_STEP step=17284 loss=1.127957 Training: 96%|█████████▌| 17284/18018 [36:57:01<1:35:43, 7.82s/step, loss=1.128, acc=0.734, acc_len=1.80, thru=19.2, I=20.1, T=35.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:15,650] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:21,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:27:21,840] loop.py:380 INFO TIMING step=17285: step=3.311s data=1.307s compute=3.275s [fwd=2.339s bwd=0.922s opt=0.003s] dispatch=3.155s [2026-10-20 02:27:21,840] loop.py:45 INFO TRAIN_STEP step=17285 loss=1.237708 Training: 96%|█████████▌| 17285/18018 [36:57:07<1:30:38, 7.42s/step, loss=1.238, acc=0.711, acc_len=1.72, thru=19.5, I=20.6, T=38.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:22,177] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:27,088] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:27:28,248] loop.py:45 INFO TRAIN_STEP step=17286 loss=1.346708 Training: 96%|█████████▌| 17286/18018 [36:57:14<1:26:49, 7.12s/step, loss=1.347, acc=0.692, acc_len=1.66, thru=20.1, I=22.9, T=35.2, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:27:28,243] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17285: forward=2534.1ms backward=1053.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:28,597] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:27:34,356] loop.py:45 INFO TRAIN_STEP step=17287 loss=1.287067 Training: 96%|█████████▌| 17287/18018 [36:57:20<1:23:00, 6.81s/step, loss=1.287, acc=0.703, acc_len=1.68, thru=19.4, I=22.7, T=39.1, wait=2.8s, pool=57, epoch=2/2][2026-10-20 02:27:41,698] loop.py:45 INFO TRAIN_STEP step=17288 loss=1.240442 Training: 96%|█████████▌| 17288/18018 [36:57:27<1:24:49, 6.97s/step, loss=1.240, acc=0.710, acc_len=1.72, thru=15.3, I=16.5, T=36.4, wait=3.8s, pool=33, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 02:27:41,744] eagle_store.py:517 WARNING batch_get_buffer missing keys (k553222-bcceeece_hs, k553222-bcceeece_ids, k553222-bcceeece_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:50,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 02:27:41,789] eagle_store.py:517 WARNING batch_get_buffer missing keys (k553223-a68d842f_hs, k553223-a68d842f_ids, k553223-a68d842f_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-20 02:27:51,062] loop.py:45 INFO TRAIN_STEP step=17289 loss=1.225495 Training: 96%|█████████▌| 17289/18018 [36:57:37<1:33:25, 7.69s/step, loss=1.225, acc=0.711, acc_len=1.72, thru=17.3, I=17.4, T=34.1, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:27:51,088] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-20 02:27:56,863] loop.py:380 INFO TIMING step=17290: step=3.213s data=1.272s compute=3.178s [fwd=2.269s bwd=0.895s opt=0.003s] dispatch=2.580s [2026-10-20 02:27:56,863] loop.py:45 INFO TRAIN_STEP step=17290 loss=1.318577 Training: 96%|█████████▌| 17290/18018 [36:57:42<1:26:25, 7.12s/step, loss=1.319, acc=0.696, acc_len=1.66, thru=19.9, I=23.6, T=39.8, wait=2.6s, pool=57, epoch=2/2][2026-10-20 02:28:04,732] loop.py:45 INFO TRAIN_STEP step=17291 loss=1.290946 Training: 96%|█████████▌| 17291/18018 [36:57:50<1:29:01, 7.35s/step, loss=1.291, acc=0.699, acc_len=1.68, thru=15.0, I=16.1, T=35.0, wait=4.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:28:04,728] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17290: forward=2611.0ms backward=996.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:11,190] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:28:11,849] loop.py:45 INFO TRAIN_STEP step=17292 loss=1.312274 Training: 96%|█████████▌| 17292/18018 [36:57:57<1:28:03, 7.28s/step, loss=1.312, acc=0.700, acc_len=1.68, thru=18.4, I=19.2, T=34.3, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:12,196] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:16,902] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:28:17,644] loop.py:45 INFO TRAIN_STEP step=17293 loss=1.228462 Training: 96%|█████████▌| 17293/18018 [36:58:03<1:22:33, 6.83s/step, loss=1.228, acc=0.714, acc_len=1.73, thru=21.0, I=24.3, T=36.4, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:17,905] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:28:25,545] loop.py:45 INFO TRAIN_STEP step=17294 loss=1.261946 Training: 96%|█████████▌| 17294/18018 [36:58:11<1:26:18, 7.15s/step, loss=1.262, acc=0.708, acc_len=1.71, thru=12.5, I=15.1, T=37.0, wait=4.4s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:33,607] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:28:35,314] loop.py:380 INFO TIMING step=17295: step=3.715s data=1.566s compute=3.683s [fwd=2.606s bwd=1.061s opt=0.003s] dispatch=6.047s [2026-10-20 02:28:35,314] loop.py:45 INFO TRAIN_STEP step=17295 loss=1.215839 Training: 96%|█████████▌| 17295/18018 [36:58:21<1:35:39, 7.94s/step, loss=1.216, acc=0.718, acc_len=1.74, thru=17.7, I=16.6, T=34.5, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:35,619] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:39,789] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:28:41,431] loop.py:45 INFO TRAIN_STEP step=17296 loss=1.143434 Training: 96%|█████████▌| 17296/18018 [36:58:27<1:28:57, 7.39s/step, loss=1.143, acc=0.724, acc_len=1.76, thru=19.9, I=30.1, T=32.7, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:28:41,426] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17295: forward=2720.8ms backward=1153.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:41,802] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:50,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:28:50,966] loop.py:45 INFO TRAIN_STEP step=17297 loss=1.242000 Training: 96%|█████████▌| 17297/18018 [36:58:37<1:36:32, 8.03s/step, loss=1.242, acc=0.710, acc_len=1.71, thru=14.0, I=14.8, T=34.1, wait=5.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:51,276] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:56,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:28:57,731] loop.py:45 INFO TRAIN_STEP step=17298 loss=1.347036 Training: 96%|█████████▌| 17298/18018 [36:58:43<1:31:51, 7.65s/step, loss=1.347, acc=0.690, acc_len=1.64, thru=18.9, I=20.5, T=37.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:28:57,768] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:02,283] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:29:03,321] loop.py:45 INFO TRAIN_STEP step=17299 loss=1.211755 Training: 96%|█████████▌| 17299/18018 [36:58:49<1:24:17, 7.03s/step, loss=1.212, acc=0.716, acc_len=1.73, thru=21.2, I=27.5, T=39.2, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:03,796] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:11,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:29:12,300] loop.py:380 INFO TIMING step=17300: step=3.560s data=1.466s compute=3.533s [fwd=2.531s bwd=0.988s opt=0.003s] dispatch=5.412s [2026-10-20 02:29:12,300] loop.py:45 INFO TRAIN_STEP step=17300 loss=1.214231 Training: 96%|█████████▌| 17300/18018 [36:58:58<1:31:09, 7.62s/step, loss=1.214, acc=0.715, acc_len=1.73, thru=15.1, I=16.6, T=36.0, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:12,369] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:19,245] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:29:20,241] loop.py:45 INFO TRAIN_STEP step=17301 loss=1.284993 Training: 96%|█████████▌| 17301/18018 [36:59:06<1:32:11, 7.71s/step, loss=1.285, acc=0.707, acc_len=1.70, thru=15.9, I=19.7, T=35.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:20,268] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (TrainerActor pid=9310) [2026-10-20 02:29:20,236] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17300: forward=2578.3ms backward=1015.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:25,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:29:26,476] loop.py:45 INFO TRAIN_STEP step=17302 loss=1.262187 Training: 96%|█████████▌| 17302/18018 [36:59:12<1:26:45, 7.27s/step, loss=1.262, acc=0.707, acc_len=1.71, thru=20.1, I=24.3, T=38.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:26,739] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:32,425] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:29:33,171] loop.py:45 INFO TRAIN_STEP step=17303 loss=1.187556 Training: 96%|█████████▌| 17303/18018 [36:59:19<1:24:35, 7.10s/step, loss=1.188, acc=0.718, acc_len=1.74, thru=19.5, I=23.9, T=36.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:33,430] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:39,914] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:29:41,215] loop.py:45 INFO TRAIN_STEP step=17304 loss=1.164363 Training: 96%|█████████▌| 17304/18018 [36:59:27<1:27:50, 7.38s/step, loss=1.164, acc=0.723, acc_len=1.75, thru=16.1, I=19.8, T=34.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:41,420] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:46,743] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:29:47,382] loop.py:380 INFO TIMING step=17305: step=3.514s data=1.479s compute=3.490s [fwd=2.534s bwd=0.941s opt=0.003s] dispatch=2.645s [2026-10-20 02:29:47,383] loop.py:45 INFO TRAIN_STEP step=17305 loss=1.195889 Training: 96%|█████████▌| 17305/18018 [36:59:33<1:23:23, 7.02s/step, loss=1.196, acc=0.715, acc_len=1.72, thru=19.6, I=23.0, T=36.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:29:47,747] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:29:53,161] loop.py:45 INFO TRAIN_STEP step=17306 loss=1.272454 Training: 96%|█████████▌| 17306/18018 [36:59:39<1:18:51, 6.65s/step, loss=1.272, acc=0.702, acc_len=1.69, thru=18.4, I=22.0, T=38.1, wait=2.4s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:29:53,157] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17305: forward=2342.8ms backward=960.3ms [2026-10-20 02:30:01,949] loop.py:45 INFO TRAIN_STEP step=17307 loss=1.305150 Training: 96%|█████████▌| 17307/18018 [36:59:48<1:26:23, 7.29s/step, loss=1.305, acc=0.699, acc_len=1.68, thru=12.5, I=16.0, T=35.2, wait=5.1s, pool=21, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 02:30:02,030] eagle_store.py:517 WARNING batch_get_buffer missing keys (k553822-9d7c94d9_hs, k553822-9d7c94d9_ids, k553822-9d7c94d9_lhs); sleeping 0.50s. [2026-10-20 02:30:10,172] loop.py:45 INFO TRAIN_STEP step=17308 loss=1.280374 Training: 96%|█████████▌| 17308/18018 [36:59:56<1:29:34, 7.57s/step, loss=1.280, acc=0.707, acc_len=1.70, thru=16.4, I=15.0, T=37.2, wait=4.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:30:16,951] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:30:18,075] loop.py:45 INFO TRAIN_STEP step=17309 loss=1.373278 Training: 96%|█████████▌| 17309/18018 [37:00:04<1:30:36, 7.67s/step, loss=1.373, acc=0.686, acc_len=1.63, thru=17.2, I=18.8, T=34.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:30:18,456] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:30:23,052] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:30:24,934] loop.py:380 INFO TIMING step=17310: step=3.952s data=1.860s compute=3.912s [fwd=2.681s bwd=1.217s opt=0.003s] dispatch=2.900s [2026-10-20 02:30:24,934] loop.py:45 INFO TRAIN_STEP step=17310 loss=1.368818 Training: 96%|█████████▌| 17310/18018 [37:00:11<1:27:37, 7.43s/step, loss=1.369, acc=0.688, acc_len=1.64, thru=17.5, I=25.4, T=32.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:30:25,069] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 02:30:30,950] loop.py:45 INFO TRAIN_STEP step=17311 loss=1.207640 Training: 96%|█████████▌| 17311/18018 [37:00:17<1:22:31, 7.00s/step, loss=1.208, acc=0.717, acc_len=1.74, thru=16.7, I=22.8, T=36.4, wait=2.5s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:30:30,945] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17310: forward=2475.0ms backward=997.5ms [2026-10-20 02:30:38,485] loop.py:45 INFO TRAIN_STEP step=17312 loss=1.180729 Training: 96%|█████████▌| 17312/18018 [37:00:24<1:24:16, 7.16s/step, loss=1.181, acc=0.721, acc_len=1.75, thru=15.6, I=16.7, T=35.3, wait=3.9s, pool=41, epoch=2/2][2026-10-20 02:30:47,582] loop.py:45 INFO TRAIN_STEP step=17313 loss=1.257856 Training: 96%|█████████▌| 17313/18018 [37:00:33<1:30:59, 7.74s/step, loss=1.258, acc=0.707, acc_len=1.70, thru=14.2, I=12.4, T=33.9, wait=5.3s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:30:53,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:30:53,896] loop.py:45 INFO TRAIN_STEP step=17314 loss=1.235257 Training: 96%|█████████▌| 17314/18018 [37:00:39<1:25:48, 7.31s/step, loss=1.235, acc=0.713, acc_len=1.72, thru=19.5, I=19.6, T=36.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:30:54,090] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:31:03,002] loop.py:380 INFO TIMING step=17315: step=4.169s data=1.423s compute=4.131s [fwd=3.153s bwd=0.963s opt=0.003s] dispatch=4.929s [2026-10-20 02:31:03,002] loop.py:45 INFO TRAIN_STEP step=17315 loss=1.225296 Training: 96%|█████████▌| 17315/18018 [37:00:49<1:31:59, 7.85s/step, loss=1.225, acc=0.713, acc_len=1.73, thru=14.5, I=14.9, T=30.7, wait=4.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:08,427] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:31:09,443] loop.py:45 INFO TRAIN_STEP step=17316 loss=1.251864 Training: 96%|█████████▌| 17316/18018 [37:00:55<1:26:54, 7.43s/step, loss=1.252, acc=0.708, acc_len=1.71, thru=18.1, I=19.9, T=35.2, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:31:09,439] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17315: forward=2584.6ms backward=1009.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:09,935] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:14,682] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:31:15,416] loop.py:45 INFO TRAIN_STEP step=17317 loss=1.185855 Training: 96%|█████████▌| 17317/18018 [37:01:01<1:21:41, 6.99s/step, loss=1.186, acc=0.720, acc_len=1.75, thru=20.8, I=26.1, T=36.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:15,686] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:20,757] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:31:21,253] loop.py:45 INFO TRAIN_STEP step=17318 loss=1.210610 Training: 96%|█████████▌| 17318/18018 [37:01:07<1:17:32, 6.65s/step, loss=1.211, acc=0.714, acc_len=1.73, thru=22.0, I=26.7, T=36.6, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:21,760] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:26,209] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:31:28,358] loop.py:45 INFO TRAIN_STEP step=17319 loss=1.198831 Training: 96%|█████████▌| 17319/18018 [37:01:14<1:19:01, 6.78s/step, loss=1.199, acc=0.718, acc_len=1.74, thru=17.8, I=22.8, T=34.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:28,723] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:32,681] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:31:34,229] loop.py:380 INFO TIMING step=17320: step=3.687s data=1.619s compute=3.664s [fwd=2.630s bwd=1.020s opt=0.003s] dispatch=2.173s [2026-10-20 02:31:34,229] loop.py:45 INFO TRAIN_STEP step=17320 loss=1.283961 Training: 96%|█████████▌| 17320/18018 [37:01:20<1:15:43, 6.51s/step, loss=1.284, acc=0.704, acc_len=1.69, thru=20.1, I=29.7, T=34.7, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:34,690] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 02:31:41,998] loop.py:45 INFO TRAIN_STEP step=17321 loss=1.241066 Training: 96%|█████████▌| 17321/18018 [37:01:28<1:19:59, 6.89s/step, loss=1.241, acc=0.709, acc_len=1.71, thru=13.3, I=15.6, T=32.0, wait=3.8s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:31:41,992] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17320: forward=2859.2ms backward=1094.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:48,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:31:49,751] loop.py:45 INFO TRAIN_STEP step=17322 loss=1.224020 Training: 96%|█████████▌| 17322/18018 [37:01:35<1:22:54, 7.15s/step, loss=1.224, acc=0.709, acc_len=1.71, thru=18.7, I=17.5, T=34.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:50,120] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:55,024] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:31:55,385] loop.py:45 INFO TRAIN_STEP step=17323 loss=1.236708 Training: 96%|█████████▌| 17323/18018 [37:01:41<1:17:32, 6.69s/step, loss=1.237, acc=0.712, acc_len=1.72, thru=21.1, I=24.4, T=36.7, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:31:55,538] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-20 02:32:02,602] loop.py:45 INFO TRAIN_STEP step=17324 loss=1.217676 Training: 96%|█████████▌| 17324/18018 [37:01:48<1:19:14, 6.85s/step, loss=1.218, acc=0.713, acc_len=1.73, thru=15.0, I=19.0, T=37.5, wait=3.8s, pool=29, epoch=2/2][2026-10-20 02:32:11,991] loop.py:380 INFO TIMING step=17325: step=4.629s data=2.010s compute=4.596s [fwd=3.338s bwd=1.242s opt=0.003s] dispatch=4.750s [2026-10-20 02:32:11,991] loop.py:45 INFO TRAIN_STEP step=17325 loss=1.322454 Training: 96%|█████████▌| 17325/18018 [37:01:58<1:27:55, 7.61s/step, loss=1.322, acc=0.697, acc_len=1.67, thru=16.4, I=15.3, T=27.7, wait=4.7s, pool=53, epoch=2/2][2026-10-20 02:32:19,911] loop.py:45 INFO TRAIN_STEP step=17326 loss=1.261878 Training: 96%|█████████▌| 17326/18018 [37:02:05<1:28:51, 7.70s/step, loss=1.262, acc=0.709, acc_len=1.71, thru=14.4, I=14.2, T=34.6, wait=4.2s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:32:19,907] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17325: forward=2652.9ms backward=1002.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:32:25,233] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:32:26,424] loop.py:45 INFO TRAIN_STEP step=17327 loss=1.238952 Training: 96%|█████████▌| 17327/18018 [37:02:12<1:24:36, 7.35s/step, loss=1.239, acc=0.708, acc_len=1.71, thru=19.7, I=21.7, T=36.7, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:32:26,740] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:32:32,458] loop.py:45 INFO TRAIN_STEP step=17328 loss=1.216502 Training: 96%|█████████▌| 17328/18018 [37:02:18<1:19:57, 6.95s/step, loss=1.217, acc=0.713, acc_len=1.72, thru=20.5, I=23.8, T=37.3, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:32:41,267] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:32:41,979] loop.py:45 INFO TRAIN_STEP step=17329 loss=1.229692 Training: 96%|█████████▌| 17329/18018 [37:02:28<1:28:41, 7.72s/step, loss=1.230, acc=0.715, acc_len=1.73, thru=14.7, I=15.9, T=33.7, wait=5.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:32:42,273] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:32:49,164] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:32:50,572] loop.py:380 INFO TIMING step=17330: step=4.401s data=2.173s compute=4.370s [fwd=3.211s bwd=1.145s opt=0.003s] dispatch=4.182s [2026-10-20 02:32:50,573] loop.py:45 INFO TRAIN_STEP step=17330 loss=1.121135 Training: 96%|█████████▌| 17330/18018 [37:02:36<1:31:33, 7.98s/step, loss=1.121, acc=0.734, acc_len=1.80, thru=14.6, I=18.7, T=29.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:32:50,675] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:32:56,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:32:56,899] loop.py:45 INFO TRAIN_STEP step=17331 loss=1.208123 Training: 96%|█████████▌| 17331/18018 [37:02:42<1:25:43, 7.49s/step, loss=1.208, acc=0.716, acc_len=1.74, thru=18.1, I=23.0, T=38.0, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:32:56,894] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17330: forward=2373.7ms backward=942.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:32:57,116] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:33:02,701] loop.py:45 INFO TRAIN_STEP step=17332 loss=1.222699 Training: 96%|█████████▌| 17332/18018 [37:02:48<1:19:49, 6.98s/step, loss=1.223, acc=0.715, acc_len=1.73, thru=17.8, I=21.4, T=39.7, wait=2.6s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:10,793] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:33:10,925] loop.py:45 INFO TRAIN_STEP step=17333 loss=1.252575 Training: 96%|█████████▌| 17333/18018 [37:02:57<1:23:57, 7.35s/step, loss=1.253, acc=0.709, acc_len=1.71, thru=18.7, I=16.7, T=34.2, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:11,296] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:33:17,780] loop.py:45 INFO TRAIN_STEP step=17334 loss=1.277884 Training: 96%|█████████▌| 17334/18018 [37:03:03<1:22:07, 7.20s/step, loss=1.278, acc=0.702, acc_len=1.68, thru=17.1, I=17.3, T=31.3, wait=2.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:27,343] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:33:28,081] loop.py:380 INFO TIMING step=17335: step=4.044s data=1.795s compute=3.994s [fwd=2.896s bwd=1.083s opt=0.003s] dispatch=6.249s [2026-10-20 02:33:28,082] loop.py:45 INFO TRAIN_STEP step=17335 loss=1.231454 Training: 96%|█████████▌| 17335/18018 [37:03:14<1:32:35, 8.13s/step, loss=1.231, acc=0.712, acc_len=1.72, thru=14.2, I=13.9, T=31.7, wait=6.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:28,349] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:33:34,079] loop.py:45 INFO TRAIN_STEP step=17336 loss=1.270830 Training: 96%|█████████▌| 17336/18018 [37:03:20<1:25:10, 7.49s/step, loss=1.271, acc=0.706, acc_len=1.70, thru=17.6, I=18.4, T=36.9, wait=2.5s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:33:34,075] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17335: forward=2430.9ms backward=975.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:41,172] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:33:42,352] loop.py:45 INFO TRAIN_STEP step=17337 loss=1.274981 Training: 96%|█████████▌| 17337/18018 [37:03:28<1:27:42, 7.73s/step, loss=1.275, acc=0.703, acc_len=1.69, thru=17.2, I=17.9, T=36.2, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:42,685] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:47,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:33:48,395] loop.py:45 INFO TRAIN_STEP step=17338 loss=1.210498 Training: 96%|█████████▌| 17338/18018 [37:03:34<1:21:50, 7.22s/step, loss=1.210, acc=0.715, acc_len=1.74, thru=20.1, I=26.1, T=38.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:48,637] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:55,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:33:55,664] loop.py:45 INFO TRAIN_STEP step=17339 loss=1.357753 Training: 96%|█████████▌| 17339/18018 [37:03:41<1:21:52, 7.24s/step, loss=1.358, acc=0.690, acc_len=1.65, thru=18.4, I=20.2, T=36.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:33:55,851] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:34:00,618] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:34:01,135] loop.py:380 INFO TIMING step=17340: step=3.254s data=1.315s compute=3.224s [fwd=2.305s bwd=0.904s opt=0.003s] dispatch=2.209s [2026-10-20 02:34:01,136] loop.py:45 INFO TRAIN_STEP step=17340 loss=1.268366 Training: 96%|█████████▌| 17340/18018 [37:03:47<1:15:47, 6.71s/step, loss=1.268, acc=0.706, acc_len=1.70, thru=21.1, I=23.4, T=39.3, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:34:01,622] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:34:10,151] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:34:10,505] loop.py:45 INFO TRAIN_STEP step=17341 loss=1.179051 Training: 96%|█████████▌| 17341/18018 [37:03:56<1:24:41, 7.51s/step, loss=1.179, acc=0.722, acc_len=1.76, thru=13.3, I=14.2, T=30.5, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:34:10,500] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17340: forward=2909.2ms backward=1233.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:34:10,681] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:34:15,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:34:16,725] loop.py:45 INFO TRAIN_STEP step=17342 loss=1.246702 Training: 96%|█████████▌| 17342/18018 [37:04:02<1:20:12, 7.12s/step, loss=1.247, acc=0.711, acc_len=1.71, thru=19.1, I=20.9, T=36.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:34:16,846] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-20 02:34:23,661] loop.py:45 INFO TRAIN_STEP step=17343 loss=1.234730 Training: 96%|█████████▋| 17343/18018 [37:04:09<1:19:28, 7.06s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=18.4, I=18.6, T=36.8, wait=3.5s, pool=61, epoch=2/2][2026-10-20 02:34:30,540] loop.py:45 INFO TRAIN_STEP step=17344 loss=1.254504 Training: 96%|█████████▋| 17344/18018 [37:04:16<1:18:44, 7.01s/step, loss=1.255, acc=0.709, acc_len=1.71, thru=17.9, I=19.0, T=35.2, wait=3.2s, pool=45, epoch=2/2][2026-10-20 02:34:38,827] loop.py:380 INFO TIMING step=17345: step=3.396s data=1.336s compute=3.361s [fwd=2.411s bwd=0.935s opt=0.003s] dispatch=4.881s [2026-10-20 02:34:38,827] loop.py:45 INFO TRAIN_STEP step=17345 loss=1.253507 Training: 96%|█████████▋| 17345/18018 [37:04:24<1:22:55, 7.39s/step, loss=1.254, acc=0.707, acc_len=1.70, thru=14.5, I=16.2, T=37.7, wait=4.9s, pool=49, epoch=2/2][2026-10-20 02:34:47,748] loop.py:45 INFO TRAIN_STEP step=17346 loss=1.361706 Training: 96%|█████████▋| 17346/18018 [37:04:33<1:27:56, 7.85s/step, loss=1.362, acc=0.689, acc_len=1.64, thru=15.5, I=14.3, T=34.3, wait=5.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:34:47,744] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17345: forward=2555.5ms backward=1106.4ms [2026-10-20 02:34:54,744] loop.py:45 INFO TRAIN_STEP step=17347 loss=1.351129 Training: 96%|█████████▋| 17347/18018 [37:04:40<1:24:55, 7.59s/step, loss=1.351, acc=0.690, acc_len=1.64, thru=16.0, I=17.3, T=34.1, wait=3.2s, pool=49, epoch=2/2][2026-10-20 02:35:02,642] loop.py:45 INFO TRAIN_STEP step=17348 loss=1.310510 Training: 96%|█████████▋| 17348/18018 [37:04:48<1:25:48, 7.68s/step, loss=1.311, acc=0.701, acc_len=1.69, thru=16.8, I=16.6, T=33.8, wait=4.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:09,342] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:35:10,640] loop.py:45 INFO TRAIN_STEP step=17349 loss=1.207944 Training: 96%|█████████▋| 17349/18018 [37:04:56<1:26:44, 7.78s/step, loss=1.208, acc=0.720, acc_len=1.75, thru=17.1, I=18.6, T=34.1, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:10,852] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:16,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:35:18,059] loop.py:380 INFO TIMING step=17350: step=3.558s data=1.561s compute=3.533s [fwd=2.560s bwd=0.959s opt=0.003s] dispatch=3.850s [2026-10-20 02:35:18,059] loop.py:45 INFO TRAIN_STEP step=17350 loss=1.284773 Training: 96%|█████████▋| 17350/18018 [37:05:04<1:25:24, 7.67s/step, loss=1.285, acc=0.700, acc_len=1.68, thru=17.3, I=22.7, T=36.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:18,506] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:35:25,650] loop.py:45 INFO TRAIN_STEP step=17351 loss=1.190014 Training: 96%|█████████▋| 17351/18018 [37:05:11<1:25:00, 7.65s/step, loss=1.190, acc=0.721, acc_len=1.76, thru=13.8, I=16.4, T=35.6, wait=4.0s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:35:25,646] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17350: forward=2599.7ms backward=953.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:33,097] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:35:34,444] loop.py:45 INFO TRAIN_STEP step=17352 loss=1.302099 Training: 96%|█████████▋| 17352/18018 [37:05:20<1:28:42, 7.99s/step, loss=1.302, acc=0.700, acc_len=1.69, thru=16.6, I=18.2, T=32.1, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:34,606] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:40,111] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:35:40,514] loop.py:45 INFO TRAIN_STEP step=17353 loss=1.260003 Training: 96%|█████████▋| 17353/18018 [37:05:26<1:22:10, 7.41s/step, loss=1.260, acc=0.705, acc_len=1.70, thru=19.4, I=22.9, T=36.6, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:40,619] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:46,402] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:35:46,423] loop.py:45 INFO TRAIN_STEP step=17354 loss=1.189273 Training: 96%|█████████▋| 17354/18018 [37:05:32<1:17:03, 6.96s/step, loss=1.189, acc=0.718, acc_len=1.74, thru=21.0, I=21.4, T=37.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:46,904] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:52,575] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:35:54,061] loop.py:380 INFO TIMING step=17355: step=3.489s data=1.420s compute=3.463s [fwd=2.505s bwd=0.944s opt=0.003s] dispatch=4.142s [2026-10-20 02:35:54,061] loop.py:45 INFO TRAIN_STEP step=17355 loss=1.345989 Training: 96%|█████████▋| 17355/18018 [37:05:40<1:19:11, 7.17s/step, loss=1.346, acc=0.690, acc_len=1.64, thru=17.5, I=20.5, T=36.7, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:35:54,096] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:00,554] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:36:00,845] loop.py:45 INFO TRAIN_STEP step=17356 loss=1.322739 Training: 96%|█████████▋| 17356/18018 [37:05:46<1:17:47, 7.05s/step, loss=1.323, acc=0.695, acc_len=1.66, thru=18.6, I=21.9, T=31.1, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:36:00,841] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17355: forward=3109.5ms backward=965.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:01,057] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:07,718] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:36:07,922] loop.py:45 INFO TRAIN_STEP step=17357 loss=1.341222 Training: 96%|█████████▋| 17357/18018 [37:05:54<1:17:45, 7.06s/step, loss=1.341, acc=0.694, acc_len=1.65, thru=17.8, I=18.3, T=35.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:08,219] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:36:15,336] loop.py:45 INFO TRAIN_STEP step=17358 loss=1.314118 Training: 96%|█████████▋| 17358/18018 [37:06:01<1:18:49, 7.17s/step, loss=1.314, acc=0.698, acc_len=1.67, thru=17.7, I=17.7, T=35.1, wait=3.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:21,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:36:21,839] loop.py:45 INFO TRAIN_STEP step=17359 loss=1.256515 Training: 96%|█████████▋| 17359/18018 [37:06:07<1:16:31, 6.97s/step, loss=1.257, acc=0.704, acc_len=1.69, thru=18.9, I=18.8, T=35.4, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:21,905] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:26,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:36:28,045] loop.py:380 INFO TIMING step=17360: step=4.106s data=2.013s compute=4.079s [fwd=3.040s bwd=1.024s opt=0.003s] dispatch=2.090s [2026-10-20 02:36:28,046] loop.py:45 INFO TRAIN_STEP step=17360 loss=1.290287 Training: 96%|█████████▋| 17360/18018 [37:06:14<1:13:53, 6.74s/step, loss=1.290, acc=0.701, acc_len=1.68, thru=19.9, I=23.6, T=31.2, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:28,173] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 [2026-10-20 02:36:34,883] loop.py:45 INFO TRAIN_STEP step=17361 loss=1.217401 Training: 96%|█████████▋| 17361/18018 [37:06:20<1:14:06, 6.77s/step, loss=1.217, acc=0.711, acc_len=1.72, thru=14.1, I=21.1, T=34.5, wait=3.1s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:36:34,878] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17360: forward=2743.5ms backward=913.5ms [2026-10-20 02:36:42,387] loop.py:45 INFO TRAIN_STEP step=17362 loss=1.272336 Training: 96%|█████████▋| 17362/18018 [37:06:28<1:16:25, 6.99s/step, loss=1.272, acc=0.700, acc_len=1.68, thru=15.5, I=14.6, T=35.5, wait=3.9s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:49,396] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:36:49,512] loop.py:45 INFO TRAIN_STEP step=17363 loss=1.326296 Training: 96%|█████████▋| 17363/18018 [37:06:35<1:16:44, 7.03s/step, loss=1.326, acc=0.694, acc_len=1.66, thru=19.4, I=18.3, T=36.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:36:49,899] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:36:56,821] loop.py:45 INFO TRAIN_STEP step=17364 loss=1.319954 Training: 96%|█████████▋| 17364/18018 [37:06:42<1:17:31, 7.11s/step, loss=1.320, acc=0.700, acc_len=1.69, thru=15.7, I=15.8, T=34.7, wait=3.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:04,113] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:37:04,217] loop.py:380 INFO TIMING step=17365: step=3.763s data=1.535s compute=3.742s [fwd=2.647s bwd=1.081s opt=0.003s] dispatch=3.625s [2026-10-20 02:37:04,217] loop.py:45 INFO TRAIN_STEP step=17365 loss=1.209444 Training: 96%|█████████▋| 17365/18018 [37:06:50<1:18:20, 7.20s/step, loss=1.209, acc=0.714, acc_len=1.73, thru=20.6, I=18.1, T=34.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:04,615] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:37:10,902] loop.py:45 INFO TRAIN_STEP step=17366 loss=1.275661 Training: 96%|█████████▋| 17366/18018 [37:06:56<1:16:33, 7.04s/step, loss=1.276, acc=0.701, acc_len=1.68, thru=15.8, I=18.9, T=35.7, wait=3.1s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:37:10,897] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17365: forward=2487.8ms backward=1037.8ms [2026-10-20 02:37:22,698] loop.py:45 INFO TRAIN_STEP step=17367 loss=1.327433 Training: 96%|█████████▋| 17367/18018 [37:07:08<1:31:54, 8.47s/step, loss=1.327, acc=0.697, acc_len=1.67, thru=13.0, I=13.6, T=32.8, wait=7.9s, pool=57, epoch=2/2][2026-10-20 02:37:29,243] loop.py:45 INFO TRAIN_STEP step=17368 loss=1.257891 Training: 96%|█████████▋| 17368/18018 [37:07:15<1:25:30, 7.89s/step, loss=1.258, acc=0.705, acc_len=1.70, thru=16.4, I=19.5, T=38.5, wait=3.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:38,144] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:37:39,142] loop.py:45 INFO TRAIN_STEP step=17369 loss=1.102197 Training: 96%|█████████▋| 17369/18018 [37:07:25<1:31:52, 8.49s/step, loss=1.102, acc=0.738, acc_len=1.82, thru=15.5, I=14.5, T=33.9, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:39,171] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:43,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:37:44,592] loop.py:380 INFO TIMING step=17370: step=3.388s data=1.384s compute=3.357s [fwd=2.364s bwd=0.978s opt=0.003s] dispatch=2.054s [2026-10-20 02:37:44,592] loop.py:45 INFO TRAIN_STEP step=17370 loss=1.176177 Training: 96%|█████████▋| 17370/18018 [37:07:30<1:21:52, 7.58s/step, loss=1.176, acc=0.721, acc_len=1.76, thru=21.1, I=26.3, T=37.8, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:44,903] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:50,002] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:37:51,470] loop.py:45 INFO TRAIN_STEP step=17371 loss=1.318570 Training: 96%|█████████▋| 17371/18018 [37:07:37<1:19:28, 7.37s/step, loss=1.319, acc=0.696, acc_len=1.66, thru=19.2, I=25.4, T=38.7, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:37:51,465] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17370: forward=2357.2ms backward=905.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:51,521] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:56,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:37:57,786] loop.py:45 INFO TRAIN_STEP step=17372 loss=1.140188 Training: 96%|█████████▋| 17372/18018 [37:07:43<1:15:56, 7.05s/step, loss=1.140, acc=0.726, acc_len=1.77, thru=20.3, I=27.7, T=37.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:37:57,979] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:38:06,248] loop.py:45 INFO TRAIN_STEP step=17373 loss=1.226929 Training: 96%|█████████▋| 17373/18018 [37:07:52<1:20:22, 7.48s/step, loss=1.227, acc=0.711, acc_len=1.72, thru=11.8, I=14.5, T=33.0, wait=4.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:14,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:38:15,446] loop.py:45 INFO TRAIN_STEP step=17374 loss=1.301503 Training: 96%|█████████▋| 17374/18018 [37:08:01<1:25:47, 7.99s/step, loss=1.302, acc=0.699, acc_len=1.68, thru=16.9, I=16.9, T=33.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:15,549] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-20 02:38:21,284] loop.py:380 INFO TIMING step=17375: step=3.313s data=1.338s compute=3.279s [fwd=2.336s bwd=0.929s opt=0.003s] dispatch=2.518s [2026-10-20 02:38:21,285] loop.py:45 INFO TRAIN_STEP step=17375 loss=1.225300 Training: 96%|█████████▋| 17375/18018 [37:08:07<1:18:43, 7.35s/step, loss=1.225, acc=0.712, acc_len=1.73, thru=17.8, I=20.4, T=38.6, wait=2.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:28,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:38:29,400] loop.py:45 INFO TRAIN_STEP step=17376 loss=1.222398 Training: 96%|█████████▋| 17376/18018 [37:08:15<1:21:04, 7.58s/step, loss=1.222, acc=0.716, acc_len=1.74, thru=17.4, I=18.8, T=34.5, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:38:29,395] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17375: forward=2637.2ms backward=1030.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:29,664] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:36,130] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:38:36,875] loop.py:45 INFO TRAIN_STEP step=17377 loss=1.341144 Training: 96%|█████████▋| 17377/18018 [37:08:22<1:20:37, 7.55s/step, loss=1.341, acc=0.691, acc_len=1.65, thru=15.8, I=18.1, T=35.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:37,137] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:38:42,944] loop.py:45 INFO TRAIN_STEP step=17378 loss=1.215115 Training: 96%|█████████▋| 17378/18018 [37:08:29<1:15:45, 7.10s/step, loss=1.215, acc=0.715, acc_len=1.74, thru=20.2, I=21.7, T=33.5, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:42,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:43,446] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:49,083] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:38:49,619] loop.py:45 INFO TRAIN_STEP step=17379 loss=1.235371 Training: 96%|█████████▋| 17379/18018 [37:08:35<1:14:16, 6.97s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=18.7, I=19.8, T=34.5, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:50,088] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:54,837] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:38:55,400] loop.py:380 INFO TIMING step=17380: step=3.332s data=1.344s compute=3.303s [fwd=2.343s bwd=0.946s opt=0.003s] dispatch=2.443s [2026-10-20 02:38:55,400] loop.py:45 INFO TRAIN_STEP step=17380 loss=1.228204 Training: 96%|█████████▋| 17380/18018 [37:08:41<1:10:21, 6.62s/step, loss=1.228, acc=0.712, acc_len=1.72, thru=20.9, I=24.8, T=38.4, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:38:55,843] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:39:03,338] loop.py:45 INFO TRAIN_STEP step=17381 loss=1.276123 Training: 96%|█████████▋| 17381/18018 [37:08:49<1:14:27, 7.01s/step, loss=1.276, acc=0.706, acc_len=1.70, thru=12.8, I=16.5, T=31.2, wait=3.8s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:39:03,333] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17380: forward=3110.5ms backward=940.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:39:11,572] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:39:12,890] loop.py:45 INFO TRAIN_STEP step=17382 loss=1.251748 Training: 96%|█████████▋| 17382/18018 [37:08:58<1:22:24, 7.77s/step, loss=1.252, acc=0.706, acc_len=1.70, thru=16.5, I=15.9, T=35.0, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:39:13,080] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:39:18,382] loop.py:45 INFO TRAIN_STEP step=17383 loss=1.156650 Training: 96%|█████████▋| 17383/18018 [37:09:04<1:15:02, 7.09s/step, loss=1.157, acc=0.725, acc_len=1.77, thru=21.0, I=23.3, T=39.5, wait=2.2s, pool=61, epoch=2/2][2026-10-20 02:39:25,212] loop.py:45 INFO TRAIN_STEP step=17384 loss=1.201369 Training: 96%|█████████▋| 17384/18018 [37:09:11<1:14:05, 7.01s/step, loss=1.201, acc=0.715, acc_len=1.74, thru=17.4, I=18.3, T=36.2, wait=3.3s, pool=41, epoch=2/2][2026-10-20 02:39:32,335] loop.py:380 INFO TIMING step=17385: step=3.606s data=1.483s compute=3.570s [fwd=2.520s bwd=1.036s opt=0.003s] dispatch=3.508s [2026-10-20 02:39:32,335] loop.py:45 INFO TRAIN_STEP step=17385 loss=1.268267 Training: 96%|█████████▋| 17385/18018 [37:09:18<1:14:19, 7.05s/step, loss=1.268, acc=0.704, acc_len=1.70, thru=17.4, I=18.3, T=35.5, wait=3.5s, pool=49, epoch=2/2][2026-10-20 02:39:39,386] loop.py:45 INFO TRAIN_STEP step=17386 loss=1.110826 Training: 96%|█████████▋| 17386/18018 [37:09:25<1:14:13, 7.05s/step, loss=1.111, acc=0.731, acc_len=1.79, thru=17.2, I=18.3, T=33.9, wait=3.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:39:39,379] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17385: forward=2615.9ms backward=1088.8ms [2026-10-20 02:39:46,198] loop.py:45 INFO TRAIN_STEP step=17387 loss=1.112523 Training: 96%|█████████▋| 17387/18018 [37:09:32<1:13:22, 6.98s/step, loss=1.113, acc=0.732, acc_len=1.79, thru=17.4, I=18.7, T=37.6, wait=3.4s, pool=57, epoch=2/2][2026-10-20 02:39:54,461] loop.py:45 INFO TRAIN_STEP step=17388 loss=1.286957 Training: 97%|█████████▋| 17388/18018 [37:09:40<1:17:18, 7.36s/step, loss=1.287, acc=0.701, acc_len=1.68, thru=13.7, I=14.0, T=35.5, wait=4.6s, pool=45, epoch=2/2][2026-10-20 02:40:03,182] loop.py:45 INFO TRAIN_STEP step=17389 loss=1.395939 Training: 97%|█████████▋| 17389/18018 [37:09:49<1:21:27, 7.77s/step, loss=1.396, acc=0.689, acc_len=1.64, thru=16.7, I=16.1, T=35.8, wait=5.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:40:08,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:40:10,472] loop.py:380 INFO TIMING step=17390: step=3.915s data=1.732s compute=3.858s [fwd=2.682s bwd=1.161s opt=0.003s] dispatch=3.362s [2026-10-20 02:40:10,472] loop.py:45 INFO TRAIN_STEP step=17390 loss=1.209859 Training: 97%|█████████▋| 17390/18018 [37:09:56<1:19:48, 7.63s/step, loss=1.210, acc=0.716, acc_len=1.73, thru=16.9, I=20.0, T=32.7, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:40:10,971] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:40:16,285] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:40:16,979] loop.py:45 INFO TRAIN_STEP step=17391 loss=1.324633 Training: 97%|█████████▋| 17391/18018 [37:10:03<1:16:10, 7.29s/step, loss=1.325, acc=0.695, acc_len=1.66, thru=18.1, I=25.0, T=38.1, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:40:16,975] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17390: forward=2376.9ms backward=941.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:40:17,289] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:40:23,629] loop.py:45 INFO TRAIN_STEP step=17392 loss=1.189566 Training: 97%|█████████▋| 17392/18018 [37:10:09<1:14:03, 7.10s/step, loss=1.190, acc=0.718, acc_len=1.74, thru=18.6, I=19.3, T=36.7, wait=3.2s, pool=61, epoch=2/2][2026-10-20 02:40:31,422] loop.py:45 INFO TRAIN_STEP step=17393 loss=1.207758 Training: 97%|█████████▋| 17393/18018 [37:10:17<1:16:07, 7.31s/step, loss=1.208, acc=0.716, acc_len=1.74, thru=16.3, I=15.5, T=33.9, wait=4.0s, pool=45, epoch=2/2][2026-10-20 02:40:39,311] loop.py:45 INFO TRAIN_STEP step=17394 loss=1.259574 Training: 97%|█████████▋| 17394/18018 [37:10:25<1:17:47, 7.48s/step, loss=1.260, acc=0.708, acc_len=1.71, thru=17.2, I=16.2, T=34.6, wait=4.2s, pool=61, epoch=2/2][2026-10-20 02:40:45,274] loop.py:380 INFO TIMING step=17395: step=3.520s data=1.518s compute=3.488s [fwd=2.468s bwd=1.005s opt=0.003s] dispatch=2.436s [2026-10-20 02:40:45,274] loop.py:45 INFO TRAIN_STEP step=17395 loss=1.291836 Training: 97%|█████████▋| 17395/18018 [37:10:31<1:12:56, 7.03s/step, loss=1.292, acc=0.700, acc_len=1.68, thru=19.0, I=19.7, T=36.4, wait=2.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:40:54,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:40:55,041] loop.py:45 INFO TRAIN_STEP step=17396 loss=1.252980 Training: 97%|█████████▋| 17396/18018 [37:10:41<1:21:21, 7.85s/step, loss=1.253, acc=0.707, acc_len=1.71, thru=14.5, I=13.8, T=28.8, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:40:55,036] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17395: forward=3118.9ms backward=1275.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:40:55,330] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:41:01,926] loop.py:45 INFO TRAIN_STEP step=17397 loss=1.419122 Training: 97%|█████████▋| 17397/18018 [37:10:48<1:18:14, 7.56s/step, loss=1.419, acc=0.680, acc_len=1.61, thru=17.4, I=17.1, T=34.3, wait=3.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:10,928] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:41:11,521] loop.py:45 INFO TRAIN_STEP step=17398 loss=1.329428 Training: 97%|█████████▋| 17398/18018 [37:10:57<1:24:25, 8.17s/step, loss=1.329, acc=0.692, acc_len=1.66, thru=14.9, I=14.4, T=34.8, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:11,933] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:16,639] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:41:17,658] loop.py:45 INFO TRAIN_STEP step=17399 loss=1.232948 Training: 97%|█████████▋| 17399/18018 [37:11:03<1:17:59, 7.56s/step, loss=1.233, acc=0.711, acc_len=1.71, thru=19.9, I=24.4, T=38.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:18,150] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:41:24,264] loop.py:380 INFO TIMING step=17400: step=3.300s data=1.377s compute=3.264s [fwd=2.358s bwd=0.892s opt=0.003s] dispatch=3.297s [2026-10-20 02:41:24,264] loop.py:45 INFO TRAIN_STEP step=17400 loss=1.266804 Training: 97%|█████████▋| 17400/18018 [37:11:10<1:14:55, 7.27s/step, loss=1.267, acc=0.703, acc_len=1.69, thru=16.9, I=22.5, T=38.8, wait=3.3s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:32,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:41:33,141] loop.py:45 INFO TRAIN_STEP step=17401 loss=1.224108 Training: 97%|█████████▋| 17401/18018 [37:11:19<1:19:44, 7.76s/step, loss=1.224, acc=0.710, acc_len=1.71, thru=16.4, I=16.1, T=33.8, wait=5.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:41:33,137] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17400: forward=2744.1ms backward=1000.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:33,529] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:41:39,899] loop.py:45 INFO TRAIN_STEP step=17402 loss=1.144469 Training: 97%|█████████▋| 17402/18018 [37:11:25<1:16:32, 7.46s/step, loss=1.144, acc=0.727, acc_len=1.78, thru=16.6, I=19.6, T=35.9, wait=3.2s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:46,657] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:41:48,011] loop.py:45 INFO TRAIN_STEP step=17403 loss=1.255329 Training: 97%|█████████▋| 17403/18018 [37:11:34<1:18:26, 7.65s/step, loss=1.255, acc=0.709, acc_len=1.71, thru=16.5, I=18.1, T=33.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:48,170] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:53,055] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:41:53,763] loop.py:45 INFO TRAIN_STEP step=17404 loss=1.151311 Training: 97%|█████████▋| 17404/18018 [37:11:39<1:12:28, 7.08s/step, loss=1.151, acc=0.727, acc_len=1.77, thru=20.6, I=26.2, T=39.1, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:54,061] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:41:59,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:42:00,891] loop.py:380 INFO TIMING step=17405: step=4.001s data=1.871s compute=3.963s [fwd=2.781s bwd=1.170s opt=0.003s] dispatch=3.118s [2026-10-20 02:42:00,891] loop.py:45 INFO TRAIN_STEP step=17405 loss=1.198136 Training: 97%|█████████▋| 17405/18018 [37:11:46<1:12:29, 7.10s/step, loss=1.198, acc=0.717, acc_len=1.74, thru=18.8, I=22.9, T=32.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:42:01,105] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:42:07,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:42:07,390] loop.py:45 INFO TRAIN_STEP step=17406 loss=1.357542 Training: 97%|█████████▋| 17406/18018 [37:11:53<1:10:33, 6.92s/step, loss=1.358, acc=0.691, acc_len=1.65, thru=18.2, I=20.4, T=37.6, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:42:07,386] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17405: forward=2425.4ms backward=928.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:42:07,697] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:42:18,264] loop.py:45 INFO TRAIN_STEP step=17407 loss=1.195630 Training: 97%|█████████▋| 17407/18018 [37:12:04<1:22:31, 8.10s/step, loss=1.196, acc=0.720, acc_len=1.75, thru=9.7, I=9.5, T=32.2, wait=6.9s, pool=45, epoch=2/2] [2026-10-20 02:42:27,750] loop.py:45 INFO TRAIN_STEP step=17408 loss=1.337140 Training: 97%|█████████▋| 17408/18018 [37:12:13<1:26:36, 8.52s/step, loss=1.337, acc=0.693, acc_len=1.65, thru=13.3, I=15.1, T=32.5, wait=5.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:42:35,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:42:36,751] loop.py:45 INFO TRAIN_STEP step=17409 loss=1.337692 Training: 97%|█████████▋| 17409/18018 [37:12:22<1:27:55, 8.66s/step, loss=1.338, acc=0.691, acc_len=1.65, thru=19.0, I=16.0, T=34.5, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:42:37,249] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:42:42,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:42:43,650] loop.py:380 INFO TIMING step=17410: step=3.471s data=1.497s compute=3.438s [fwd=2.450s bwd=0.973s opt=0.003s] dispatch=3.416s [2026-10-20 02:42:43,650] loop.py:45 INFO TRAIN_STEP step=17410 loss=1.169073 Training: 97%|█████████▋| 17410/18018 [37:12:29<1:22:25, 8.13s/step, loss=1.169, acc=0.723, acc_len=1.76, thru=18.4, I=24.8, T=36.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:42:43,685] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-20 02:42:49,450] loop.py:45 INFO TRAIN_STEP step=17411 loss=1.187854 Training: 97%|█████████▋| 17411/18018 [37:12:35<1:15:12, 7.43s/step, loss=1.188, acc=0.718, acc_len=1.73, thru=21.2, I=22.9, T=35.5, wait=2.2s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:42:49,446] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17410: forward=2560.5ms backward=1002.2ms [2026-10-20 02:42:56,281] loop.py:45 INFO TRAIN_STEP step=17412 loss=1.236747 Training: 97%|█████████▋| 17412/18018 [37:12:42<1:13:15, 7.25s/step, loss=1.237, acc=0.709, acc_len=1.71, thru=18.4, I=19.2, T=35.3, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:02,045] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-20 02:43:03,342] loop.py:45 INFO TRAIN_STEP step=17413 loss=1.179050 Training: 97%|█████████▋| 17413/18018 [37:12:49<1:12:32, 7.19s/step, loss=1.179, acc=0.722, acc_len=1.76, thru=18.5, I=20.8, T=34.9, wait=3.4s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:03,555] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=5/64 [2026-10-20 02:43:09,211] loop.py:45 INFO TRAIN_STEP step=17414 loss=1.246184 Training: 97%|█████████▋| 17414/18018 [37:12:55<1:08:25, 6.80s/step, loss=1.246, acc=0.713, acc_len=1.73, thru=20.1, I=23.5, T=36.5, wait=2.4s, pool=53, epoch=2/2][2026-10-20 02:43:16,283] loop.py:380 INFO TIMING step=17415: step=3.406s data=1.334s compute=3.372s [fwd=2.441s bwd=0.918s opt=0.003s] dispatch=3.655s [2026-10-20 02:43:16,283] loop.py:45 INFO TRAIN_STEP step=17415 loss=1.273446 Training: 97%|█████████▋| 17415/18018 [37:13:02<1:09:08, 6.88s/step, loss=1.273, acc=0.704, acc_len=1.69, thru=15.8, I=15.8, T=37.6, wait=3.7s, pool=37, epoch=2/2][2026-10-20 02:43:24,252] loop.py:45 INFO TRAIN_STEP step=17416 loss=1.148776 Training: 97%|█████████▋| 17416/18018 [37:13:10<1:12:18, 7.21s/step, loss=1.149, acc=0.725, acc_len=1.77, thru=18.5, I=17.9, T=34.4, wait=4.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:43:24,247] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17415: forward=2667.4ms backward=1012.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:31,461] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:43:32,979] loop.py:45 INFO TRAIN_STEP step=17417 loss=1.233588 Training: 97%|█████████▋| 17417/18018 [37:13:19<1:16:45, 7.66s/step, loss=1.234, acc=0.710, acc_len=1.72, thru=15.7, I=16.7, T=36.2, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:33,475] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:38,479] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:43:39,668] loop.py:45 INFO TRAIN_STEP step=17418 loss=1.164252 Training: 97%|█████████▋| 17418/18018 [37:13:25<1:13:42, 7.37s/step, loss=1.164, acc=0.724, acc_len=1.77, thru=19.1, I=25.6, T=37.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:39,988] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:45,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=69/64 [2026-10-20 02:43:46,013] loop.py:45 INFO TRAIN_STEP step=17419 loss=1.243871 Training: 97%|█████████▋| 17419/18018 [37:13:32<1:10:30, 7.06s/step, loss=1.244, acc=0.709, acc_len=1.71, thru=20.4, I=23.1, T=35.9, wait=2.8s, pool=69, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:46,356] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=5/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:52,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:43:52,633] loop.py:380 INFO TIMING step=17420: step=3.412s data=1.398s compute=3.373s [fwd=2.432s bwd=0.928s opt=0.003s] dispatch=3.198s [2026-10-20 02:43:52,633] loop.py:45 INFO TRAIN_STEP step=17420 loss=1.273118 Training: 97%|█████████▋| 17420/18018 [37:13:38<1:09:03, 6.93s/step, loss=1.273, acc=0.704, acc_len=1.69, thru=18.9, I=22.2, T=37.5, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:53,116] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:58,385] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:43:59,857] loop.py:45 INFO TRAIN_STEP step=17421 loss=1.298704 Training: 97%|█████████▋| 17421/18018 [37:13:45<1:09:49, 7.02s/step, loss=1.299, acc=0.700, acc_len=1.69, thru=18.0, I=23.9, T=35.1, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:43:59,852] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17420: forward=2568.0ms backward=1034.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:43:59,904] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:44:08,188] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:44:08,232] loop.py:45 INFO TRAIN_STEP step=17422 loss=1.226189 Training: 97%|█████████▋| 17422/18018 [37:13:54<1:13:45, 7.42s/step, loss=1.226, acc=0.713, acc_len=1.72, thru=14.1, I=16.7, T=36.5, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:44:08,691] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:44:17,296] loop.py:45 INFO TRAIN_STEP step=17423 loss=1.482578 Training: 97%|█████████▋| 17423/18018 [37:14:03<1:18:30, 7.92s/step, loss=1.483, acc=0.665, acc_len=1.56, thru=13.5, I=14.7, T=31.2, wait=4.9s, pool=53, epoch=2/2][2026-10-20 02:44:27,421] loop.py:45 INFO TRAIN_STEP step=17424 loss=1.303635 Training: 97%|█████████▋| 17424/18018 [37:14:13<1:24:56, 8.58s/step, loss=1.304, acc=0.702, acc_len=1.68, thru=13.1, I=12.1, T=33.1, wait=6.3s, pool=53, epoch=2/2][2026-10-20 02:44:34,514] loop.py:380 INFO TIMING step=17425: step=3.720s data=1.612s compute=3.690s [fwd=2.669s bwd=1.006s opt=0.003s] dispatch=3.365s [2026-10-20 02:44:34,514] loop.py:45 INFO TRAIN_STEP step=17425 loss=1.284531 Training: 97%|█████████▋| 17425/18018 [37:14:20<1:20:23, 8.13s/step, loss=1.285, acc=0.703, acc_len=1.69, thru=16.6, I=18.8, T=34.4, wait=3.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:44:42,736] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:44:43,290] loop.py:45 INFO TRAIN_STEP step=17426 loss=1.207005 Training: 97%|█████████▋| 17426/18018 [37:14:29<1:22:09, 8.33s/step, loss=1.207, acc=0.716, acc_len=1.74, thru=16.2, I=15.0, T=33.9, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:44:43,285] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17425: forward=2597.2ms backward=1121.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:44:43,740] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:44:50,212] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:44:50,690] loop.py:45 INFO TRAIN_STEP step=17427 loss=1.228067 Training: 97%|█████████▋| 17427/18018 [37:14:36<1:19:16, 8.05s/step, loss=1.228, acc=0.710, acc_len=1.72, thru=17.8, I=20.2, T=32.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:44:50,717] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-20 02:44:56,758] loop.py:45 INFO TRAIN_STEP step=17428 loss=1.255991 Training: 97%|█████████▋| 17428/18018 [37:14:42<1:13:18, 7.45s/step, loss=1.256, acc=0.707, acc_len=1.69, thru=15.7, I=17.8, T=35.8, wait=2.5s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:06,016] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:45:07,257] loop.py:45 INFO TRAIN_STEP step=17429 loss=1.251982 Training: 97%|█████████▋| 17429/18018 [37:14:53<1:22:08, 8.37s/step, loss=1.252, acc=0.709, acc_len=1.71, thru=16.2, I=17.2, T=31.3, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:07,526] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:13,820] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:45:14,434] loop.py:380 INFO TIMING step=17430: step=3.545s data=1.454s compute=3.506s [fwd=2.498s bwd=0.993s opt=0.003s] dispatch=3.622s [2026-10-20 02:45:14,434] loop.py:45 INFO TRAIN_STEP step=17430 loss=1.269193 Training: 97%|█████████▋| 17430/18018 [37:15:00<1:18:29, 8.01s/step, loss=1.269, acc=0.706, acc_len=1.70, thru=17.0, I=20.5, T=36.1, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:14,826] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:45:22,140] loop.py:45 INFO TRAIN_STEP step=17431 loss=1.189585 Training: 97%|█████████▋| 17431/18018 [37:15:08<1:17:28, 7.92s/step, loss=1.190, acc=0.714, acc_len=1.72, thru=15.6, I=16.5, T=35.4, wait=4.1s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:45:22,136] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17430: forward=2491.7ms backward=1066.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:29,651] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:45:30,865] loop.py:45 INFO TRAIN_STEP step=17432 loss=1.235179 Training: 97%|█████████▋| 17432/18018 [37:15:16<1:19:42, 8.16s/step, loss=1.235, acc=0.710, acc_len=1.72, thru=17.3, I=19.1, T=34.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:31,159] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:38,934] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:45:39,070] loop.py:45 INFO TRAIN_STEP step=17433 loss=1.301392 Training: 97%|█████████▋| 17433/18018 [37:15:25<1:19:41, 8.17s/step, loss=1.301, acc=0.697, acc_len=1.67, thru=14.4, I=15.8, T=32.5, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:39,436] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:46,925] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:45:47,303] loop.py:45 INFO TRAIN_STEP step=17434 loss=1.317137 Training: 97%|█████████▋| 17434/18018 [37:15:33<1:19:44, 8.19s/step, loss=1.317, acc=0.696, acc_len=1.66, thru=16.8, I=16.3, T=34.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:45:47,429] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-20 02:45:55,347] loop.py:380 INFO TIMING step=17435: step=3.789s data=1.618s compute=3.756s [fwd=2.723s bwd=1.020s opt=0.003s] dispatch=4.247s [2026-10-20 02:45:55,347] loop.py:45 INFO TRAIN_STEP step=17435 loss=1.297881 Training: 97%|█████████▋| 17435/18018 [37:15:41<1:19:10, 8.15s/step, loss=1.298, acc=0.702, acc_len=1.68, thru=13.7, I=15.9, T=33.8, wait=4.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:46:04,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:46:04,529] loop.py:45 INFO TRAIN_STEP step=17436 loss=1.313676 Training: 97%|█████████▋| 17436/18018 [37:15:50<1:22:02, 8.46s/step, loss=1.314, acc=0.701, acc_len=1.68, thru=17.3, I=14.6, T=34.1, wait=5.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:46:04,525] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17435: forward=2681.0ms backward=1025.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:46:04,792] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:46:11,138] loop.py:45 INFO TRAIN_STEP step=17437 loss=1.185819 Training: 97%|█████████▋| 17437/18018 [37:15:57<1:16:31, 7.90s/step, loss=1.186, acc=0.721, acc_len=1.75, thru=18.2, I=18.2, T=36.9, wait=3.1s, pool=53, epoch=2/2][2026-10-20 02:46:19,199] loop.py:45 INFO TRAIN_STEP step=17438 loss=1.206333 Training: 97%|█████████▋| 17438/18018 [37:16:05<1:16:51, 7.95s/step, loss=1.206, acc=0.713, acc_len=1.72, thru=16.7, I=17.6, T=32.0, wait=4.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:46:28,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:46:28,608] loop.py:45 INFO TRAIN_STEP step=17439 loss=1.224300 Training: 97%|█████████▋| 17439/18018 [37:16:14<1:20:56, 8.39s/step, loss=1.224, acc=0.712, acc_len=1.72, thru=13.7, I=12.8, T=33.6, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:46:28,884] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:46:34,722] loop.py:380 INFO TIMING step=17440: step=3.636s data=1.572s compute=3.603s [fwd=2.570s bwd=1.019s opt=0.003s] dispatch=2.470s [2026-10-20 02:46:34,722] loop.py:45 INFO TRAIN_STEP step=17440 loss=1.227891 Training: 97%|█████████▋| 17440/18018 [37:16:20<1:14:14, 7.71s/step, loss=1.228, acc=0.711, acc_len=1.71, thru=18.1, I=20.6, T=35.2, wait=2.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:46:41,747] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:46:43,075] loop.py:45 INFO TRAIN_STEP step=17441 loss=1.273849 Training: 97%|█████████▋| 17441/18018 [37:16:29<1:15:58, 7.90s/step, loss=1.274, acc=0.704, acc_len=1.70, thru=16.0, I=18.2, T=33.9, wait=4.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:46:43,071] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17440: forward=2667.2ms backward=1065.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:46:43,256] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:46:48,965] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:46:49,514] loop.py:45 INFO TRAIN_STEP step=17442 loss=1.322890 Training: 97%|█████████▋| 17442/18018 [37:16:35<1:11:38, 7.46s/step, loss=1.323, acc=0.693, acc_len=1.66, thru=19.4, I=23.0, T=37.9, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:46:49,969] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:46:59,348] loop.py:45 INFO TRAIN_STEP step=17443 loss=1.263276 Training: 97%|█████████▋| 17443/18018 [37:16:45<1:18:19, 8.17s/step, loss=1.263, acc=0.706, acc_len=1.69, thru=13.2, I=13.1, T=31.3, wait=5.7s, pool=61, epoch=2/2][2026-10-20 02:47:05,822] loop.py:45 INFO TRAIN_STEP step=17444 loss=1.250218 Training: 97%|█████████▋| 17444/18018 [37:16:51<1:13:19, 7.66s/step, loss=1.250, acc=0.708, acc_len=1.71, thru=16.3, I=14.9, T=35.7, wait=2.9s, pool=37, epoch=2/2][2026-10-20 02:47:14,964] loop.py:380 INFO TIMING step=17445: step=3.622s data=1.599s compute=3.587s [fwd=2.584s bwd=0.990s opt=0.003s] dispatch=5.510s [2026-10-20 02:47:14,964] loop.py:45 INFO TRAIN_STEP step=17445 loss=1.236466 Training: 97%|█████████▋| 17445/18018 [37:17:01<1:17:25, 8.11s/step, loss=1.236, acc=0.709, acc_len=1.71, thru=13.5, I=14.5, T=35.3, wait=5.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:47:22,948] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:47:24,377] loop.py:45 INFO TRAIN_STEP step=17446 loss=1.153977 Training: 97%|█████████▋| 17446/18018 [37:17:10<1:21:01, 8.50s/step, loss=1.154, acc=0.726, acc_len=1.77, thru=17.6, I=16.8, T=30.9, wait=5.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:47:24,373] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17445: forward=3005.8ms backward=1090.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:47:24,462] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-20 02:47:30,877] loop.py:45 INFO TRAIN_STEP step=17447 loss=1.334426 Training: 97%|█████████▋| 17447/18018 [37:17:16<1:15:10, 7.90s/step, loss=1.334, acc=0.694, acc_len=1.66, thru=17.3, I=21.0, T=36.9, wait=3.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:47:37,242] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:47:38,268] loop.py:45 INFO TRAIN_STEP step=17448 loss=1.226927 Training: 97%|█████████▋| 17448/18018 [37:17:24<1:13:35, 7.75s/step, loss=1.227, acc=0.713, acc_len=1.72, thru=18.4, I=18.8, T=33.0, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:47:38,754] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:47:44,825] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:47:44,979] loop.py:45 INFO TRAIN_STEP step=17449 loss=1.044614 Training: 97%|█████████▋| 17449/18018 [37:17:31<1:10:30, 7.44s/step, loss=1.045, acc=0.745, acc_len=1.84, thru=17.5, I=20.2, T=38.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:47:45,327] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:47:52,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:47:53,592] loop.py:380 INFO TIMING step=17450: step=3.614s data=1.432s compute=3.577s [fwd=2.563s bwd=0.999s opt=0.003s] dispatch=4.991s [2026-10-20 02:47:53,592] loop.py:45 INFO TRAIN_STEP step=17450 loss=1.264654 Training: 97%|█████████▋| 17450/18018 [37:17:39<1:13:44, 7.79s/step, loss=1.265, acc=0.708, acc_len=1.70, thru=16.4, I=17.5, T=35.4, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:47:53,737] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:47:59,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:48:00,221] loop.py:45 INFO TRAIN_STEP step=17451 loss=1.261405 Training: 97%|█████████▋| 17451/18018 [37:17:46<1:10:19, 7.44s/step, loss=1.261, acc=0.708, acc_len=1.70, thru=18.9, I=21.3, T=37.1, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:48:00,216] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17450: forward=2452.1ms backward=947.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:48:00,707] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:48:07,064] loop.py:45 INFO TRAIN_STEP step=17452 loss=1.311439 Training: 97%|█████████▋| 17452/18018 [37:17:53<1:08:29, 7.26s/step, loss=1.311, acc=0.696, acc_len=1.67, thru=16.7, I=18.2, T=37.4, wait=3.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:48:12,146] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:48:13,266] loop.py:45 INFO TRAIN_STEP step=17453 loss=1.199243 Training: 97%|█████████▋| 17453/18018 [37:17:59<1:05:23, 6.94s/step, loss=1.199, acc=0.717, acc_len=1.74, thru=20.5, I=22.1, T=36.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:48:13,655] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:48:22,008] loop.py:45 INFO TRAIN_STEP step=17454 loss=1.234801 Training: 97%|█████████▋| 17454/18018 [37:18:08<1:10:20, 7.48s/step, loss=1.235, acc=0.710, acc_len=1.71, thru=12.6, I=14.5, T=32.1, wait=4.7s, pool=57, epoch=2/2][2026-10-20 02:48:30,257] loop.py:380 INFO TIMING step=17455: step=3.761s data=1.545s compute=3.735s [fwd=2.657s bwd=1.064s opt=0.003s] dispatch=4.480s [2026-10-20 02:48:30,257] loop.py:45 INFO TRAIN_STEP step=17455 loss=1.419806 Training: 97%|█████████▋| 17455/18018 [37:18:16<1:12:22, 7.71s/step, loss=1.420, acc=0.681, acc_len=1.62, thru=15.9, I=15.6, T=34.0, wait=4.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:48:35,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:48:36,812] loop.py:45 INFO TRAIN_STEP step=17456 loss=1.235332 Training: 97%|█████████▋| 17456/18018 [37:18:22<1:08:59, 7.37s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=18.0, I=19.4, T=35.4, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:48:36,808] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17455: forward=2523.8ms backward=1035.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:48:36,990] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:48:42,800] loop.py:45 INFO TRAIN_STEP step=17457 loss=1.177029 Training: 97%|█████████▋| 17457/18018 [37:18:28<1:05:00, 6.95s/step, loss=1.177, acc=0.720, acc_len=1.75, thru=19.2, I=22.8, T=36.8, wait=2.5s, pool=53, epoch=2/2][2026-10-20 02:48:51,791] loop.py:45 INFO TRAIN_STEP step=17458 loss=1.260968 Training: 97%|█████████▋| 17458/18018 [37:18:37<1:10:36, 7.56s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=13.1, I=14.9, T=32.5, wait=5.0s, pool=41, epoch=2/2][2026-10-20 02:49:01,741] loop.py:45 INFO TRAIN_STEP step=17459 loss=1.280567 Training: 97%|█████████▋| 17459/18018 [37:18:47<1:17:08, 8.28s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=15.6, I=12.4, T=33.3, wait=6.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:05,813] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:49:07,839] loop.py:380 INFO TIMING step=17460: step=3.689s data=1.596s compute=3.657s [fwd=2.620s bwd=1.021s opt=0.003s] dispatch=2.400s [2026-10-20 02:49:07,839] loop.py:45 INFO TRAIN_STEP step=17460 loss=1.229470 Training: 97%|█████████▋| 17460/18018 [37:18:53<1:10:54, 7.62s/step, loss=1.229, acc=0.709, acc_len=1.71, thru=19.6, I=24.1, T=34.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:08,326] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:13,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:49:14,639] loop.py:45 INFO TRAIN_STEP step=17461 loss=1.191465 Training: 97%|█████████▋| 17461/18018 [37:19:00<1:08:29, 7.38s/step, loss=1.191, acc=0.721, acc_len=1.75, thru=17.6, I=27.4, T=38.6, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:49:14,635] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17460: forward=2340.9ms backward=930.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:14,905] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:19,510] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:49:20,201] loop.py:45 INFO TRAIN_STEP step=17462 loss=1.308264 Training: 97%|█████████▋| 17462/18018 [37:19:06<1:03:18, 6.83s/step, loss=1.308, acc=0.698, acc_len=1.67, thru=21.6, I=27.5, T=37.8, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:20,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:49:26,357] loop.py:45 INFO TRAIN_STEP step=17463 loss=1.159123 Training: 97%|█████████▋| 17463/18018 [37:19:12<1:01:19, 6.63s/step, loss=1.159, acc=0.726, acc_len=1.77, thru=20.1, I=23.1, T=36.6, wait=2.6s, pool=61, epoch=2/2][2026-10-20 02:49:33,912] loop.py:45 INFO TRAIN_STEP step=17464 loss=1.257439 Training: 97%|█████████▋| 17464/18018 [37:19:19<1:03:46, 6.91s/step, loss=1.257, acc=0.709, acc_len=1.71, thru=15.0, I=15.6, T=35.5, wait=3.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:39,516] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:49:40,848] loop.py:380 INFO TIMING step=17465: step=3.601s data=1.569s compute=3.568s [fwd=2.584s bwd=0.970s opt=0.003s] dispatch=3.326s [2026-10-20 02:49:40,848] loop.py:45 INFO TRAIN_STEP step=17465 loss=1.295889 Training: 97%|█████████▋| 17465/18018 [37:19:26<1:03:44, 6.92s/step, loss=1.296, acc=0.700, acc_len=1.69, thru=18.1, I=21.1, T=35.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:41,025] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:49:47,591] loop.py:45 INFO TRAIN_STEP step=17466 loss=1.221063 Training: 97%|█████████▋| 17466/18018 [37:19:33<1:03:09, 6.87s/step, loss=1.221, acc=0.713, acc_len=1.72, thru=16.7, I=22.7, T=35.4, wait=3.1s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:49:47,586] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17465: forward=2544.0ms backward=1011.1ms [2026-10-20 02:49:58,950] loop.py:45 INFO TRAIN_STEP step=17467 loss=1.305144 Training: 97%|█████████▋| 17467/18018 [37:19:45<1:15:24, 8.21s/step, loss=1.305, acc=0.699, acc_len=1.68, thru=13.3, I=11.7, T=34.3, wait=7.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:58,924] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:49:59,427] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:05,163] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:50:05,964] loop.py:45 INFO TRAIN_STEP step=17468 loss=1.268902 Training: 97%|█████████▋| 17468/18018 [37:19:52<1:11:59, 7.85s/step, loss=1.269, acc=0.703, acc_len=1.69, thru=18.2, I=19.9, T=33.6, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:06,167] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:13,552] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:50:14,798] loop.py:45 INFO TRAIN_STEP step=17469 loss=1.225638 Training: 97%|█████████▋| 17469/18018 [37:20:00<1:14:32, 8.15s/step, loss=1.226, acc=0.714, acc_len=1.72, thru=13.7, I=16.6, T=33.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:15,061] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:20,595] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:50:21,217] loop.py:380 INFO TIMING step=17470: step=3.694s data=1.683s compute=3.663s [fwd=2.621s bwd=1.028s opt=0.003s] dispatch=2.715s [2026-10-20 02:50:21,217] loop.py:45 INFO TRAIN_STEP step=17470 loss=1.293831 Training: 97%|█████████▋| 17470/18018 [37:20:07<1:09:40, 7.63s/step, loss=1.294, acc=0.700, acc_len=1.68, thru=18.8, I=21.9, T=34.6, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:21,600] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:27,814] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:50:28,923] loop.py:45 INFO TRAIN_STEP step=17471 loss=1.158510 Training: 97%|█████████▋| 17471/18018 [37:20:15<1:09:45, 7.65s/step, loss=1.159, acc=0.724, acc_len=1.77, thru=17.2, I=20.9, T=37.4, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:50:28,919] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17470: forward=2435.5ms backward=947.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:29,324] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:36,394] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:50:36,756] loop.py:45 INFO TRAIN_STEP step=17472 loss=1.177972 Training: 97%|█████████▋| 17472/18018 [37:20:22<1:10:07, 7.71s/step, loss=1.178, acc=0.723, acc_len=1.76, thru=16.8, I=20.4, T=35.3, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:36,896] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-20 02:50:42,819] loop.py:45 INFO TRAIN_STEP step=17473 loss=1.164358 Training: 97%|█████████▋| 17473/18018 [37:20:28<1:05:31, 7.21s/step, loss=1.164, acc=0.724, acc_len=1.77, thru=19.0, I=18.8, T=35.5, wait=2.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:50,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:50:51,127] loop.py:45 INFO TRAIN_STEP step=17474 loss=1.261860 Training: 97%|█████████▋| 17474/18018 [37:20:37<1:08:22, 7.54s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=16.0, I=16.9, T=27.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:51,404] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:50:58,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:50:59,980] loop.py:380 INFO TIMING step=17475: step=4.580s data=2.142s compute=4.533s [fwd=3.216s bwd=1.303s opt=0.003s] dispatch=4.263s [2026-10-20 02:50:59,980] loop.py:45 INFO TRAIN_STEP step=17475 loss=1.110158 Training: 97%|█████████▋| 17475/18018 [37:20:46<1:11:48, 7.94s/step, loss=1.110, acc=0.734, acc_len=1.80, thru=14.5, I=16.6, T=28.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:00,298] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:05,809] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:51:06,212] loop.py:45 INFO TRAIN_STEP step=17476 loss=1.164316 Training: 97%|█████████▋| 17476/18018 [37:20:52<1:07:03, 7.42s/step, loss=1.164, acc=0.723, acc_len=1.76, thru=17.1, I=20.0, T=36.5, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:51:06,207] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17475: forward=2430.9ms backward=1007.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:06,326] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:11,448] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:51:12,571] loop.py:45 INFO TRAIN_STEP step=17477 loss=1.195103 Training: 97%|█████████▋| 17477/18018 [37:20:58<1:04:03, 7.10s/step, loss=1.195, acc=0.720, acc_len=1.75, thru=20.1, I=23.1, T=36.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:12,955] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:19,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:51:20,056] loop.py:45 INFO TRAIN_STEP step=17478 loss=1.160569 Training: 97%|█████████▋| 17478/18018 [37:21:06<1:04:58, 7.22s/step, loss=1.161, acc=0.728, acc_len=1.78, thru=16.8, I=20.6, T=36.1, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:20,350] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:25,880] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:51:26,756] loop.py:45 INFO TRAIN_STEP step=17479 loss=1.177532 Training: 97%|█████████▋| 17479/18018 [37:21:12<1:03:27, 7.06s/step, loss=1.178, acc=0.721, acc_len=1.75, thru=18.8, I=20.5, T=35.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:26,890] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-20 02:51:32,775] loop.py:380 INFO TIMING step=17480: step=3.342s data=1.356s compute=3.309s [fwd=2.370s bwd=0.924s opt=0.003s] dispatch=2.667s [2026-10-20 02:51:32,776] loop.py:45 INFO TRAIN_STEP step=17480 loss=1.319083 Training: 97%|█████████▋| 17480/18018 [37:21:18<1:00:31, 6.75s/step, loss=1.319, acc=0.694, acc_len=1.66, thru=18.2, I=19.4, T=38.3, wait=2.7s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:40,777] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 02:51:41,186] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17480: forward=2539.8ms backward=1000.2ms [2026-10-20 02:51:41,190] loop.py:45 INFO TRAIN_STEP step=17481 loss=1.356094 Training: 97%|█████████▋| 17481/18018 [37:21:27<1:04:52, 7.25s/step, loss=1.356, acc=0.688, acc_len=1.65, thru=16.9, I=16.8, T=35.7, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:51:41,293] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-20 02:51:49,362] loop.py:45 INFO TRAIN_STEP step=17482 loss=1.144671 Training: 97%|█████████▋| 17482/18018 [37:21:35<1:07:14, 7.53s/step, loss=1.145, acc=0.726, acc_len=1.78, thru=15.8, I=15.3, T=35.0, wait=4.5s, pool=61, epoch=2/2][2026-10-20 02:51:55,438] loop.py:45 INFO TRAIN_STEP step=17483 loss=1.354785 Training: 97%|█████████▋| 17483/18018 [37:21:41<1:03:13, 7.09s/step, loss=1.355, acc=0.688, acc_len=1.64, thru=17.7, I=16.8, T=35.2, wait=2.4s, pool=41, epoch=2/2][2026-10-20 02:52:04,066] loop.py:45 INFO TRAIN_STEP step=17484 loss=1.279098 Training: 97%|█████████▋| 17484/18018 [37:21:50<1:07:13, 7.55s/step, loss=1.279, acc=0.704, acc_len=1.69, thru=16.9, I=15.4, T=33.7, wait=4.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:52:09,574] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:52:10,383] loop.py:380 INFO TIMING step=17485: step=3.496s data=1.507s compute=3.470s [fwd=2.485s bwd=0.970s opt=0.003s] dispatch=2.809s [2026-10-20 02:52:10,383] loop.py:45 INFO TRAIN_STEP step=17485 loss=1.311793 Training: 97%|█████████▋| 17485/18018 [37:21:56<1:03:47, 7.18s/step, loss=1.312, acc=0.696, acc_len=1.67, thru=19.5, I=20.0, T=36.6, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:52:10,581] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:52:16,531] loop.py:45 INFO TRAIN_STEP step=17486 loss=1.187527 Training: 97%|█████████▋| 17486/18018 [37:22:02<1:00:55, 6.87s/step, loss=1.188, acc=0.718, acc_len=1.74, thru=18.6, I=23.6, T=34.2, wait=2.4s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:52:16,526] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17485: forward=2595.1ms backward=1097.7ms [2026-10-20 02:52:26,885] loop.py:45 INFO TRAIN_STEP step=17487 loss=1.247013 Training: 97%|█████████▋| 17487/18018 [37:22:12<1:10:03, 7.92s/step, loss=1.247, acc=0.708, acc_len=1.71, thru=13.6, I=12.2, T=26.6, wait=5.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:52:34,768] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:52:35,070] loop.py:45 INFO TRAIN_STEP step=17488 loss=1.247451 Training: 97%|█████████▋| 17488/18018 [37:22:21<1:10:38, 8.00s/step, loss=1.247, acc=0.707, acc_len=1.71, thru=16.9, I=15.5, T=28.6, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:52:35,271] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:52:42,860] loop.py:45 INFO TRAIN_STEP step=17489 loss=1.207852 Training: 97%|█████████▋| 17489/18018 [37:22:28<1:09:57, 7.94s/step, loss=1.208, acc=0.717, acc_len=1.74, thru=14.0, I=15.8, T=33.6, wait=4.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:52:52,485] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:52:53,973] loop.py:380 INFO TIMING step=17490: step=4.128s data=1.881s compute=4.096s [fwd=2.904s bwd=1.178s opt=0.003s] dispatch=6.973s [2026-10-20 02:52:53,973] loop.py:45 INFO TRAIN_STEP step=17490 loss=1.142150 Training: 97%|█████████▋| 17490/18018 [37:22:40<1:18:13, 8.89s/step, loss=1.142, acc=0.728, acc_len=1.78, thru=12.3, I=19.1, T=31.0, wait=7.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:52:54,006] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=57/64 [2026-10-20 02:52:59,893] loop.py:45 INFO TRAIN_STEP step=17491 loss=1.110273 Training: 97%|█████████▋| 17491/18018 [37:22:45<1:10:14, 8.00s/step, loss=1.110, acc=0.729, acc_len=1.78, thru=16.9, I=19.9, T=36.9, wait=2.4s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:52:59,888] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17490: forward=2417.3ms backward=1001.3ms [2026-10-20 02:53:10,699] loop.py:45 INFO TRAIN_STEP step=17492 loss=1.216466 Training: 97%|█████████▋| 17492/18018 [37:22:56<1:17:30, 8.84s/step, loss=1.216, acc=0.714, acc_len=1.73, thru=10.8, I=13.8, T=34.7, wait=7.1s, pool=37, epoch=2/2][2026-10-20 02:53:18,220] loop.py:45 INFO TRAIN_STEP step=17493 loss=1.306247 Training: 97%|█████████▋| 17493/18018 [37:23:04<1:13:53, 8.44s/step, loss=1.306, acc=0.699, acc_len=1.67, thru=15.7, I=16.2, T=36.6, wait=4.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:53:26,480] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:53:27,399] loop.py:45 INFO TRAIN_STEP step=17494 loss=1.211781 Training: 97%|█████████▋| 17494/18018 [37:23:13<1:15:40, 8.66s/step, loss=1.212, acc=0.711, acc_len=1.71, thru=16.0, I=15.9, T=33.2, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:53:27,497] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-20 02:53:33,770] loop.py:380 INFO TIMING step=17495: step=3.500s data=1.459s compute=3.459s [fwd=2.442s bwd=1.002s opt=0.003s] dispatch=2.863s [2026-10-20 02:53:33,771] loop.py:45 INFO TRAIN_STEP step=17495 loss=1.239745 Training: 97%|█████████▋| 17495/18018 [37:23:19<1:09:31, 7.98s/step, loss=1.240, acc=0.713, acc_len=1.72, thru=17.3, I=18.8, T=36.6, wait=2.9s, pool=49, epoch=2/2][2026-10-20 02:53:41,662] loop.py:45 INFO TRAIN_STEP step=17496 loss=1.284890 Training: 97%|█████████▋| 17496/18018 [37:23:27<1:09:10, 7.95s/step, loss=1.285, acc=0.705, acc_len=1.70, thru=15.8, I=15.1, T=35.0, wait=4.2s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:53:41,658] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17495: forward=2598.8ms backward=1005.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:53:49,713] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:53:50,875] loop.py:45 INFO TRAIN_STEP step=17497 loss=1.309528 Training: 97%|█████████▋| 17497/18018 [37:23:36<1:12:19, 8.33s/step, loss=1.310, acc=0.698, acc_len=1.67, thru=16.0, I=17.2, T=33.9, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:53:51,222] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:53:57,005] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:53:57,623] loop.py:45 INFO TRAIN_STEP step=17498 loss=1.213956 Training: 97%|█████████▋| 17498/18018 [37:23:43<1:08:04, 7.86s/step, loss=1.214, acc=0.715, acc_len=1.74, thru=17.9, I=22.3, T=37.0, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:53:58,008] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:03,244] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:54:04,263] loop.py:45 INFO TRAIN_STEP step=17499 loss=1.142027 Training: 97%|█████████▋| 17499/18018 [37:23:50<1:04:47, 7.49s/step, loss=1.142, acc=0.727, acc_len=1.77, thru=18.9, I=22.7, T=38.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:04,751] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:10,236] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:54:11,002] loop.py:380 INFO TIMING step=17500: step=3.529s data=1.552s compute=3.482s [fwd=2.497s bwd=0.971s opt=0.003s] dispatch=3.203s [2026-10-20 02:54:11,003] loop.py:45 INFO TRAIN_STEP step=17500 loss=1.307343 Training: 97%|█████████▋| 17500/18018 [37:23:57<1:02:43, 7.27s/step, loss=1.307, acc=0.699, acc_len=1.68, thru=18.9, I=23.2, T=36.3, wait=3.2s, pool=65, epoch=2/2][2026-10-20 02:54:11,007] loop.py:419 INFO Saving checkpoint at step 17500... (TrainerActor pid=9310) [2026-10-20 02:54:14,431] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0017501 (TrainerActor pid=9310) [2026-10-20 02:54:14,560] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0016751 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:15,273] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:15,792] inference_manager.py:408 INFO Pool capacity available after 5.6s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:20,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 02:54:14,432] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0017501 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 02:54:15,733] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0016751 [2026-10-20 02:54:22,202] loop.py:45 INFO TRAIN_STEP step=17501 loss=1.326721 Training: 97%|█████████▋| 17501/18018 [37:24:08<1:12:46, 8.45s/step, loss=1.327, acc=0.694, acc_len=1.66, thru=20.0, I=27.6, T=35.6, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:54:22,198] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17500: forward=2526.2ms backward=1020.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:22,536] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:27,971] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:54:28,707] loop.py:45 INFO TRAIN_STEP step=17502 loss=1.197023 Training: 97%|█████████▋| 17502/18018 [37:24:14<1:07:37, 7.86s/step, loss=1.197, acc=0.717, acc_len=1.74, thru=18.8, I=25.9, T=33.4, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:28,978] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:34,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:54:35,569] loop.py:45 INFO TRAIN_STEP step=17503 loss=1.120039 Training: 97%|█████████▋| 17503/18018 [37:24:21<1:04:55, 7.56s/step, loss=1.120, acc=0.734, acc_len=1.80, thru=17.7, I=19.1, T=36.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:35,954] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:54:42,031] loop.py:45 INFO TRAIN_STEP step=17504 loss=1.189802 Training: 97%|█████████▋| 17504/18018 [37:24:28<1:01:57, 7.23s/step, loss=1.190, acc=0.719, acc_len=1.75, thru=16.0, I=18.5, T=37.1, wait=3.0s, pool=29, epoch=2/2](TrainerActor pid=1753, ip=10.0.3.62) [2026-10-20 02:54:42,120] eagle_store.py:517 WARNING batch_get_buffer missing keys (k560134-a953ec8d_hs, k560134-a953ec8d_ids, k560134-a953ec8d_lhs); sleeping 0.50s. (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:51,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-20 02:54:42,095] eagle_store.py:517 WARNING batch_get_buffer missing keys (k560133-888aab37_hs, k560133-888aab37_ids, k560133-888aab37_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-20 02:54:51,590] loop.py:380 INFO TIMING step=17505: step=3.906s data=1.642s compute=3.866s [fwd=2.814s bwd=1.037s opt=0.003s] dispatch=5.645s [2026-10-20 02:54:51,590] loop.py:45 INFO TRAIN_STEP step=17505 loss=1.240400 Training: 97%|█████████▋| 17505/18018 [37:24:37<1:07:48, 7.93s/step, loss=1.240, acc=0.710, acc_len=1.71, thru=17.7, I=16.2, T=32.8, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:51,636] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:54:59,176] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:55:00,020] loop.py:45 INFO TRAIN_STEP step=17506 loss=1.232798 (TrainerActor pid=9310) [2026-10-20 02:55:00,016] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17505: forward=2628.2ms backward=987.8ms Training: 97%|█████████▋| 17506/18018 [37:24:46<1:08:57, 8.08s/step, loss=1.233, acc=0.714, acc_len=1.73, thru=15.7, I=17.5, T=35.0, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:55:00,182] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:55:06,829] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:55:07,217] loop.py:45 INFO TRAIN_STEP step=17507 loss=1.171746 Training: 97%|█████████▋| 17507/18018 [37:24:53<1:06:33, 7.82s/step, loss=1.172, acc=0.720, acc_len=1.75, thru=17.9, I=19.6, T=35.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:55:07,334] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:55:13,003] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:55:13,590] loop.py:45 INFO TRAIN_STEP step=17508 loss=1.273845 Training: 97%|█████████▋| 17508/18018 [37:24:59<1:02:44, 7.38s/step, loss=1.274, acc=0.701, acc_len=1.69, thru=19.9, I=19.8, T=35.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:55:14,007] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:55:19,941] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:55:20,306] loop.py:45 INFO TRAIN_STEP step=17509 loss=1.254961 Training: 97%|█████████▋| 17509/18018 [37:25:06<1:00:56, 7.18s/step, loss=1.255, acc=0.710, acc_len=1.72, thru=19.4, I=21.3, T=36.9, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:55:20,457] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-20 02:55:27,005] loop.py:380 INFO TIMING step=17510: step=3.557s data=1.536s compute=3.510s [fwd=2.445s bwd=1.050s opt=0.003s] dispatch=3.130s [2026-10-20 02:55:27,005] loop.py:45 INFO TRAIN_STEP step=17510 loss=1.231428 Training: 97%|█████████▋| 17510/18018 [37:25:13<59:34, 7.04s/step, loss=1.231, acc=0.712, acc_len=1.72, thru=16.8, I=20.1, T=36.0, wait=3.1s, pool=41, epoch=2/2] [2026-10-20 02:55:39,948] loop.py:45 INFO TRAIN_STEP step=17511 loss=1.315376 Training: 97%|█████████▋| 17511/18018 [37:25:26<1:14:26, 8.81s/step, loss=1.315, acc=0.695, acc_len=1.66, thru=11.4, I=10.4, T=33.6, wait=9.1s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:55:39,944] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17510: forward=2627.7ms backward=1104.0ms [2026-10-20 02:55:48,468] loop.py:45 INFO TRAIN_STEP step=17512 loss=1.165490 Training: 97%|█████████▋| 17512/18018 [37:25:34<1:13:33, 8.72s/step, loss=1.165, acc=0.728, acc_len=1.78, thru=15.3, I=13.5, T=35.5, wait=4.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:55:57,253] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:55:58,657] loop.py:45 INFO TRAIN_STEP step=17513 loss=1.323905 Training: 97%|█████████▋| 17513/18018 [37:25:44<1:17:06, 9.16s/step, loss=1.324, acc=0.694, acc_len=1.67, thru=14.1, I=16.9, T=31.0, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:55:58,770] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:03,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:56:04,583] loop.py:45 INFO TRAIN_STEP step=17514 loss=1.229482 Training: 97%|█████████▋| 17514/18018 [37:25:50<1:08:48, 8.19s/step, loss=1.229, acc=0.714, acc_len=1.73, thru=20.0, I=23.0, T=36.1, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:04,899] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:14,118] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:56:14,164] loop.py:380 INFO TIMING step=17515: step=4.258s data=2.079s compute=4.219s [fwd=2.938s bwd=1.266s opt=0.003s] dispatch=5.312s [2026-10-20 02:56:14,164] loop.py:45 INFO TRAIN_STEP step=17515 loss=1.278983 Training: 97%|█████████▋| 17515/18018 [37:26:00<1:12:09, 8.61s/step, loss=1.279, acc=0.705, acc_len=1.70, thru=13.8, I=13.9, T=30.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:14,620] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:56:23,573] loop.py:45 INFO TRAIN_STEP step=17516 loss=1.124531 Training: 97%|█████████▋| 17516/18018 [37:26:09<1:14:02, 8.85s/step, loss=1.125, acc=0.733, acc_len=1.79, thru=13.4, I=15.0, T=31.6, wait=5.3s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:56:23,568] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17515: forward=2818.8ms backward=1177.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:29,161] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:56:29,910] loop.py:45 INFO TRAIN_STEP step=17517 loss=1.189779 Training: 97%|█████████▋| 17517/18018 [37:26:15<1:07:35, 8.09s/step, loss=1.190, acc=0.721, acc_len=1.76, thru=19.1, I=19.4, T=33.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:30,167] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:36,999] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:56:38,702] loop.py:45 INFO TRAIN_STEP step=17518 loss=1.227393 Training: 97%|█████████▋| 17518/18018 [37:26:24<1:09:12, 8.30s/step, loss=1.227, acc=0.712, acc_len=1.72, thru=14.1, I=17.9, T=33.6, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:39,015] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:44,828] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:56:45,047] loop.py:45 INFO TRAIN_STEP step=17519 loss=1.314649 Training: 97%|█████████▋| 17519/18018 [37:26:31<1:04:10, 7.72s/step, loss=1.315, acc=0.696, acc_len=1.67, thru=18.9, I=24.6, T=36.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:45,330] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:51,488] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:56:53,148] loop.py:380 INFO TIMING step=17520: step=4.397s data=2.080s compute=4.363s [fwd=3.037s bwd=1.311s opt=0.003s] dispatch=3.696s [2026-10-20 02:56:53,148] loop.py:45 INFO TRAIN_STEP step=17520 loss=1.141518 Training: 97%|█████████▋| 17520/18018 [37:26:39<1:05:00, 7.83s/step, loss=1.142, acc=0.726, acc_len=1.77, thru=16.0, I=18.7, T=29.1, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:56:53,505] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 02:57:01,272] loop.py:45 INFO TRAIN_STEP step=17521 loss=1.165059 Training: 97%|█████████▋| 17521/18018 [37:26:47<1:05:35, 7.92s/step, loss=1.165, acc=0.724, acc_len=1.77, thru=13.1, I=15.2, T=35.0, wait=4.5s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:57:01,268] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17520: forward=2571.9ms backward=1049.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:08,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:57:09,553] loop.py:45 INFO TRAIN_STEP step=17522 loss=1.296853 Training: 97%|█████████▋| 17522/18018 [37:26:55<1:06:21, 8.03s/step, loss=1.297, acc=0.701, acc_len=1.68, thru=16.0, I=17.9, T=35.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:09,831] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:14,831] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:57:15,915] loop.py:45 INFO TRAIN_STEP step=17523 loss=1.222303 Training: 97%|█████████▋| 17523/18018 [37:27:01<1:02:06, 7.53s/step, loss=1.222, acc=0.713, acc_len=1.73, thru=19.6, I=26.4, T=37.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:16,341] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:57:24,326] loop.py:45 INFO TRAIN_STEP step=17524 loss=1.300150 Training: 97%|█████████▋| 17524/18018 [37:27:10<1:04:09, 7.79s/step, loss=1.300, acc=0.700, acc_len=1.68, thru=12.8, I=16.6, T=37.2, wait=5.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:32,481] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:57:32,787] loop.py:380 INFO TIMING step=17525: step=3.688s data=1.583s compute=3.652s [fwd=2.641s bwd=0.997s opt=0.003s] dispatch=4.764s [2026-10-20 02:57:32,787] loop.py:45 INFO TRAIN_STEP step=17525 loss=1.226619 Training: 97%|█████████▋| 17525/18018 [37:27:18<1:05:40, 7.99s/step, loss=1.227, acc=0.711, acc_len=1.72, thru=16.5, I=16.4, T=34.7, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:32,984] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:39,040] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:57:39,796] loop.py:45 INFO TRAIN_STEP step=17526 loss=1.340940 Training: 97%|█████████▋| 17526/18018 [37:27:25<1:03:07, 7.70s/step, loss=1.341, acc=0.693, acc_len=1.65, thru=18.2, I=19.8, T=35.1, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:57:39,792] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17525: forward=2600.8ms backward=992.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:40,046] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:57:46,465] loop.py:45 INFO TRAIN_STEP step=17527 loss=1.314346 Training: 97%|█████████▋| 17527/18018 [37:27:32<1:00:27, 7.39s/step, loss=1.314, acc=0.696, acc_len=1.67, thru=15.3, I=17.4, T=36.5, wait=3.1s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:53,977] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:57:54,149] loop.py:45 INFO TRAIN_STEP step=17528 loss=1.179997 Training: 97%|█████████▋| 17528/18018 [37:27:40<1:01:04, 7.48s/step, loss=1.180, acc=0.722, acc_len=1.76, thru=19.4, I=17.8, T=36.7, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:57:54,480] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:00,569] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:58:01,548] loop.py:45 INFO TRAIN_STEP step=17529 loss=1.225874 Training: 97%|█████████▋| 17529/18018 [37:27:47<1:00:45, 7.45s/step, loss=1.226, acc=0.714, acc_len=1.73, thru=17.2, I=19.6, T=34.8, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:01,583] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:07,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:58:08,228] loop.py:380 INFO TIMING step=17530: step=3.528s data=1.466s compute=3.486s [fwd=2.446s bwd=1.025s opt=0.003s] dispatch=3.143s [2026-10-20 02:58:08,228] loop.py:45 INFO TRAIN_STEP step=17530 loss=1.240955 Training: 97%|█████████▋| 17530/18018 [37:27:54<58:44, 7.22s/step, loss=1.241, acc=0.709, acc_len=1.71, thru=18.0, I=21.8, T=36.3, wait=3.1s, pool=65, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:08,345] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:13,224] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:58:14,044] loop.py:45 INFO TRAIN_STEP step=17531 loss=1.174781 Training: 97%|█████████▋| 17531/18018 [37:28:00<55:11, 6.80s/step, loss=1.175, acc=0.720, acc_len=1.75, thru=21.0, I=23.1, T=38.9, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:58:14,040] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17530: forward=2335.2ms backward=913.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:14,230] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:58:21,859] loop.py:45 INFO TRAIN_STEP step=17532 loss=1.292418 Training: 97%|█████████▋| 17532/18018 [37:28:07<57:32, 7.10s/step, loss=1.292, acc=0.702, acc_len=1.69, thru=16.5, I=18.3, T=33.6, wait=4.0s, pool=61, epoch=2/2][2026-10-20 02:58:31,210] loop.py:45 INFO TRAIN_STEP step=17533 loss=1.201739 Training: 97%|█████████▋| 17533/18018 [37:28:17<1:02:52, 7.78s/step, loss=1.202, acc=0.720, acc_len=1.75, thru=12.5, I=13.2, T=34.0, wait=5.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:39,533] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:58:40,629] loop.py:45 INFO TRAIN_STEP step=17534 loss=1.176378 Training: 97%|█████████▋| 17534/18018 [37:28:26<1:06:43, 8.27s/step, loss=1.176, acc=0.723, acc_len=1.77, thru=16.2, I=15.6, T=33.0, wait=5.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:41,043] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:58:49,113] loop.py:380 INFO TIMING step=17535: step=3.495s data=1.443s compute=3.469s [fwd=2.507s bwd=0.947s opt=0.003s] dispatch=4.979s [2026-10-20 02:58:49,113] loop.py:45 INFO TRAIN_STEP step=17535 loss=1.342178 Training: 97%|█████████▋| 17535/18018 [37:28:35<1:07:06, 8.34s/step, loss=1.342, acc=0.691, acc_len=1.65, thru=11.7, I=13.8, T=36.6, wait=5.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:55,922] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:58:57,138] loop.py:45 INFO TRAIN_STEP step=17536 loss=1.276646 Training: 97%|█████████▋| 17536/18018 [37:28:43<1:06:12, 8.24s/step, loss=1.277, acc=0.696, acc_len=1.67, thru=17.6, I=18.4, T=26.3, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:58:57,133] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17535: forward=3438.0ms backward=1388.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:58:57,435] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 02:59:03,053] loop.py:45 INFO TRAIN_STEP step=17537 loss=1.256497 Training: 97%|█████████▋| 17537/18018 [37:28:49<1:00:28, 7.54s/step, loss=1.256, acc=0.706, acc_len=1.70, thru=17.3, I=19.8, T=37.3, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:03,007] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:03,508] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 02:59:10,077] loop.py:45 INFO TRAIN_STEP step=17538 loss=1.299439 Training: 97%|█████████▋| 17538/18018 [37:28:56<59:06, 7.39s/step, loss=1.299, acc=0.699, acc_len=1.68, thru=17.7, I=21.0, T=36.8, wait=3.5s, pool=49, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:16,156] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:59:17,262] loop.py:45 INFO TRAIN_STEP step=17539 loss=1.182599 Training: 97%|█████████▋| 17539/18018 [37:29:03<58:29, 7.33s/step, loss=1.183, acc=0.720, acc_len=1.75, thru=18.0, I=19.7, T=32.8, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:17,666] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:22,888] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:59:24,082] loop.py:380 INFO TIMING step=17540: step=3.403s data=1.376s compute=3.373s [fwd=2.436s bwd=0.923s opt=0.003s] dispatch=3.408s [2026-10-20 02:59:24,082] loop.py:45 INFO TRAIN_STEP step=17540 loss=1.236524 Training: 97%|█████████▋| 17540/18018 [37:29:10<57:09, 7.17s/step, loss=1.237, acc=0.709, acc_len=1.72, thru=18.1, I=22.9, T=37.6, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:24,401] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:29,053] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:59:30,124] loop.py:45 INFO TRAIN_STEP step=17541 loss=1.307998 Training: 97%|█████████▋| 17541/18018 [37:29:16<54:19, 6.83s/step, loss=1.308, acc=0.700, acc_len=1.68, thru=21.0, I=27.1, T=37.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 02:59:30,119] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17540: forward=2375.0ms backward=948.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:30,561] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:35,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:59:36,581] loop.py:45 INFO TRAIN_STEP step=17542 loss=1.099486 Training: 97%|█████████▋| 17542/18018 [37:29:22<53:19, 6.72s/step, loss=1.099, acc=0.737, acc_len=1.81, thru=19.9, I=25.4, T=34.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:36,661] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-20 02:59:45,183] loop.py:45 INFO TRAIN_STEP step=17543 loss=1.207430 Training: 97%|█████████▋| 17543/18018 [37:29:31<57:40, 7.29s/step, loss=1.207, acc=0.721, acc_len=1.75, thru=13.5, I=15.1, T=31.7, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:53,106] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 02:59:53,750] loop.py:45 INFO TRAIN_STEP step=17544 loss=1.177736 Training: 97%|█████████▋| 17544/18018 [37:29:39<1:00:35, 7.67s/step, loss=1.178, acc=0.718, acc_len=1.75, thru=16.2, I=16.3, T=33.8, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 02:59:54,112] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 02:59:59,841] loop.py:380 INFO TIMING step=17545: step=3.687s data=1.595s compute=3.655s [fwd=2.575s bwd=1.065s opt=0.003s] dispatch=2.395s [2026-10-20 02:59:59,841] loop.py:45 INFO TRAIN_STEP step=17545 loss=1.383824 Training: 97%|█████████▋| 17545/18018 [37:29:45<56:44, 7.20s/step, loss=1.384, acc=0.685, acc_len=1.63, thru=19.3, I=20.7, T=34.7, wait=2.4s, pool=61, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:04,475] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:00:06,254] loop.py:45 INFO TRAIN_STEP step=17546 loss=1.323456 Training: 97%|█████████▋| 17546/18018 [37:29:52<54:45, 6.96s/step, loss=1.323, acc=0.696, acc_len=1.66, thru=18.8, I=23.2, T=34.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:00:06,249] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17545: forward=2585.9ms backward=1021.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:06,485] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:10,882] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:00:12,009] loop.py:45 INFO TRAIN_STEP step=17547 loss=1.303067 Training: 97%|█████████▋| 17547/18018 [37:29:58<51:48, 6.60s/step, loss=1.303, acc=0.701, acc_len=1.68, thru=21.3, I=27.0, T=36.3, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:12,393] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:19,668] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:00:20,258] loop.py:45 INFO TRAIN_STEP step=17548 loss=1.336281 Training: 97%|█████████▋| 17548/18018 [37:30:06<55:34, 7.09s/step, loss=1.336, acc=0.691, acc_len=1.65, thru=14.9, I=19.0, T=34.1, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:20,672] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:27,920] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:00:28,050] loop.py:45 INFO TRAIN_STEP step=17549 loss=1.172018 Training: 97%|█████████▋| 17549/18018 [37:30:14<57:05, 7.30s/step, loss=1.172, acc=0.721, acc_len=1.75, thru=16.3, I=17.0, T=36.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:28,422] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:33,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:00:34,312] loop.py:380 INFO TIMING step=17550: step=3.524s data=1.521s compute=3.483s [fwd=2.499s bwd=0.969s opt=0.003s] dispatch=2.730s [2026-10-20 03:00:34,313] loop.py:45 INFO TRAIN_STEP step=17550 loss=1.254210 Training: 97%|█████████▋| 17550/18018 [37:30:20<54:31, 6.99s/step, loss=1.254, acc=0.710, acc_len=1.71, thru=20.0, I=23.0, T=36.3, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:34,642] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:40,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 03:00:41,445] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17550: forward=2529.0ms backward=1019.3ms [2026-10-20 03:00:41,450] loop.py:45 INFO TRAIN_STEP step=17551 loss=1.292864 Training: 97%|█████████▋| 17551/18018 [37:30:27<54:45, 7.04s/step, loss=1.293, acc=0.701, acc_len=1.69, thru=18.4, I=22.7, T=35.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:41,539] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-20 03:00:47,798] loop.py:45 INFO TRAIN_STEP step=17552 loss=1.359333 Training: 97%|█████████▋| 17552/18018 [37:30:33<53:02, 6.83s/step, loss=1.359, acc=0.688, acc_len=1.63, thru=19.2, I=21.9, T=39.3, wait=3.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:54,918] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:00:55,007] loop.py:45 INFO TRAIN_STEP step=17553 loss=1.296218 Training: 97%|█████████▋| 17553/18018 [37:30:41<53:48, 6.94s/step, loss=1.296, acc=0.698, acc_len=1.67, thru=18.6, I=18.4, T=35.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:00:55,421] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:01,517] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:01:02,454] loop.py:45 INFO TRAIN_STEP step=17554 loss=1.321788 Training: 97%|█████████▋| 17554/18018 [37:30:48<54:52, 7.09s/step, loss=1.322, acc=0.698, acc_len=1.67, thru=18.1, I=19.5, T=37.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:02,537] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-20 03:01:08,510] loop.py:380 INFO TIMING step=17555: step=3.406s data=1.374s compute=3.376s [fwd=2.434s bwd=0.927s opt=0.003s] dispatch=2.639s [2026-10-20 03:01:08,511] loop.py:45 INFO TRAIN_STEP step=17555 loss=1.251311 Training: 97%|█████████▋| 17555/18018 [37:30:54<52:20, 6.78s/step, loss=1.251, acc=0.707, acc_len=1.71, thru=19.7, I=22.7, T=37.6, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:15,956] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 03:01:16,882] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17555: forward=2714.6ms backward=1075.0ms [2026-10-20 03:01:16,887] loop.py:45 INFO TRAIN_STEP step=17556 loss=1.161630 Training: 97%|█████████▋| 17556/18018 [37:31:02<55:54, 7.26s/step, loss=1.162, acc=0.723, acc_len=1.76, thru=15.8, I=17.5, T=33.4, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:16,963] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:23,181] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:01:24,428] loop.py:45 INFO TRAIN_STEP step=17557 loss=1.150506 Training: 97%|█████████▋| 17557/18018 [37:31:10<56:25, 7.34s/step, loss=1.151, acc=0.730, acc_len=1.78, thru=17.3, I=21.3, T=34.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:24,688] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:28,723] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:01:30,456] loop.py:45 INFO TRAIN_STEP step=17558 loss=1.275133 Training: 97%|█████████▋| 17558/18018 [37:31:16<53:16, 6.95s/step, loss=1.275, acc=0.706, acc_len=1.71, thru=20.1, I=27.9, T=37.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:30,735] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:37,309] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:01:38,037] loop.py:45 INFO TRAIN_STEP step=17559 loss=1.215945 Training: 97%|█████████▋| 17559/18018 [37:31:24<54:36, 7.14s/step, loss=1.216, acc=0.713, acc_len=1.72, thru=15.3, I=20.9, T=35.4, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:38,316] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:01:45,011] loop.py:380 INFO TIMING step=17560: step=3.622s data=1.501s compute=3.593s [fwd=2.566s bwd=1.013s opt=0.003s] dispatch=3.342s [2026-10-20 03:01:45,011] loop.py:45 INFO TRAIN_STEP step=17560 loss=1.273891 Training: 97%|█████████▋| 17560/18018 [37:31:31<54:06, 7.09s/step, loss=1.274, acc=0.705, acc_len=1.69, thru=15.8, I=18.7, T=35.3, wait=3.3s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:51,962] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:01:52,815] loop.py:45 INFO TRAIN_STEP step=17561 loss=1.275577 Training: 97%|█████████▋| 17561/18018 [37:31:38<55:37, 7.30s/step, loss=1.276, acc=0.701, acc_len=1.68, thru=18.7, I=17.6, T=38.1, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:01:52,810] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17560: forward=2407.6ms backward=913.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:01:52,967] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:01:59,598] loop.py:45 INFO TRAIN_STEP step=17562 loss=1.302707 Training: 97%|█████████▋| 17562/18018 [37:31:45<54:19, 7.15s/step, loss=1.303, acc=0.701, acc_len=1.68, thru=17.1, I=22.0, T=36.5, wait=3.3s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:02:06,126] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:02:07,051] loop.py:45 INFO TRAIN_STEP step=17563 loss=1.291462 Training: 97%|█████████▋| 17563/18018 [37:31:53<54:53, 7.24s/step, loss=1.291, acc=0.702, acc_len=1.70, thru=17.2, I=18.7, T=35.4, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:02:07,131] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:02:12,408] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:02:12,576] loop.py:45 INFO TRAIN_STEP step=17564 loss=1.237250 Training: 97%|█████████▋| 17564/18018 [37:31:58<50:53, 6.72s/step, loss=1.237, acc=0.710, acc_len=1.71, thru=22.2, I=24.8, T=39.2, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:02:12,910] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:02:19,936] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:02:20,075] loop.py:380 INFO TIMING step=17565: step=3.715s data=1.482s compute=3.688s [fwd=2.604s bwd=1.070s opt=0.003s] dispatch=3.777s [2026-10-20 03:02:20,075] loop.py:45 INFO TRAIN_STEP step=17565 loss=1.177651 Training: 97%|█████████▋| 17565/18018 [37:32:06<52:31, 6.96s/step, loss=1.178, acc=0.723, acc_len=1.76, thru=19.3, I=18.3, T=34.5, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:02:20,437] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:02:27,487] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:02:28,804] loop.py:45 INFO TRAIN_STEP step=17566 loss=1.218829 Training: 97%|█████████▋| 17566/18018 [37:32:14<56:25, 7.49s/step, loss=1.219, acc=0.715, acc_len=1.73, thru=15.8, I=19.6, T=31.9, wait=4.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:02:28,799] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17565: forward=2818.0ms backward=1137.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:02:29,000] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=17/64 [2026-10-20 03:02:34,455] loop.py:45 INFO TRAIN_STEP step=17567 loss=1.173068 Training: 97%|█████████▋| 17567/18018 [37:32:20<52:08, 6.94s/step, loss=1.173, acc=0.726, acc_len=1.77, thru=19.1, I=26.5, T=35.5, wait=2.0s, pool=49, epoch=2/2][2026-10-20 03:02:42,725] loop.py:45 INFO TRAIN_STEP step=17568 loss=1.165321 Training: 98%|█████████▊| 17568/18018 [37:32:28<55:01, 7.34s/step, loss=1.165, acc=0.726, acc_len=1.77, thru=15.3, I=14.6, T=36.6, wait=4.8s, pool=57, epoch=2/2][2026-10-20 03:02:51,808] loop.py:45 INFO TRAIN_STEP step=17569 loss=1.174892 Training: 98%|█████████▊| 17569/18018 [37:32:37<58:49, 7.86s/step, loss=1.175, acc=0.722, acc_len=1.76, thru=14.1, I=13.9, T=35.0, wait=5.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:02:58,996] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:03:00,251] loop.py:380 INFO TIMING step=17570: step=3.692s data=1.490s compute=3.667s [fwd=2.672s bwd=0.981s opt=0.003s] dispatch=4.745s [2026-10-20 03:03:00,251] loop.py:45 INFO TRAIN_STEP step=17570 loss=1.176940 Training: 98%|█████████▊| 17570/18018 [37:32:46<59:59, 8.04s/step, loss=1.177, acc=0.721, acc_len=1.75, thru=17.1, I=17.6, T=34.7, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:00,504] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:05,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:03:06,554] loop.py:45 INFO TRAIN_STEP step=17571 loss=1.111447 Training: 98%|█████████▊| 17571/18018 [37:32:52<55:59, 7.52s/step, loss=1.111, acc=0.734, acc_len=1.80, thru=19.5, I=24.9, T=37.4, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:03:06,550] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17570: forward=2427.3ms backward=944.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:06,688] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-20 03:03:12,634] loop.py:45 INFO TRAIN_STEP step=17572 loss=1.154626 Training: 98%|█████████▊| 17572/18018 [37:32:58<52:40, 7.09s/step, loss=1.155, acc=0.727, acc_len=1.77, thru=20.2, I=22.4, T=36.2, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:19,215] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:03:19,287] loop.py:45 INFO TRAIN_STEP step=17573 loss=1.275567 Training: 98%|█████████▊| 17573/18018 [37:33:05<51:35, 6.96s/step, loss=1.276, acc=0.705, acc_len=1.69, thru=19.2, I=20.0, T=34.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:19,718] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:25,832] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:03:26,056] loop.py:45 INFO TRAIN_STEP step=17574 loss=1.248570 Training: 98%|█████████▊| 17574/18018 [37:33:12<51:03, 6.90s/step, loss=1.249, acc=0.709, acc_len=1.71, thru=18.6, I=19.5, T=36.4, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:26,334] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:03:34,401] loop.py:380 INFO TIMING step=17575: step=3.700s data=1.612s compute=3.668s [fwd=2.650s bwd=1.004s opt=0.003s] dispatch=4.635s [2026-10-20 03:03:34,401] loop.py:45 INFO TRAIN_STEP step=17575 loss=1.334899 Training: 98%|█████████▊| 17575/18018 [37:33:20<54:08, 7.33s/step, loss=1.335, acc=0.691, acc_len=1.65, thru=15.2, I=16.7, T=34.6, wait=4.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:41,559] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:03:41,591] loop.py:45 INFO TRAIN_STEP step=17576 loss=1.190589 Training: 98%|█████████▊| 17576/18018 [37:33:27<53:42, 7.29s/step, loss=1.191, acc=0.725, acc_len=1.77, thru=16.8, I=16.6, T=34.8, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:03:41,587] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17575: forward=2614.8ms backward=1025.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:42,060] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:03:49,133] loop.py:45 INFO TRAIN_STEP step=17577 loss=1.322029 Training: 98%|█████████▊| 17577/18018 [37:33:35<54:08, 7.37s/step, loss=1.322, acc=0.695, acc_len=1.67, thru=16.5, I=17.3, T=34.3, wait=3.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:55,548] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:03:56,190] loop.py:45 INFO TRAIN_STEP step=17578 loss=1.342640 Training: 98%|█████████▊| 17578/18018 [37:33:42<53:20, 7.27s/step, loss=1.343, acc=0.691, acc_len=1.65, thru=18.7, I=18.8, T=34.1, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:03:56,555] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:00,703] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:04:01,854] loop.py:45 INFO TRAIN_STEP step=17579 loss=1.263263 Training: 98%|█████████▊| 17579/18018 [37:33:47<49:40, 6.79s/step, loss=1.263, acc=0.708, acc_len=1.70, thru=20.9, I=25.8, T=37.4, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:02,213] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:04:08,458] loop.py:380 INFO TIMING step=17580: step=3.458s data=1.393s compute=3.423s [fwd=2.449s bwd=0.958s opt=0.003s] dispatch=3.138s [2026-10-20 03:04:08,458] loop.py:45 INFO TRAIN_STEP step=17580 loss=1.307601 Training: 98%|█████████▊| 17580/18018 [37:33:54<49:09, 6.73s/step, loss=1.308, acc=0.698, acc_len=1.67, thru=18.0, I=22.0, T=37.0, wait=3.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:19,230] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:04:19,487] loop.py:45 INFO TRAIN_STEP step=17581 loss=1.237246 Training: 98%|█████████▊| 17581/18018 [37:34:05<58:25, 8.02s/step, loss=1.237, acc=0.708, acc_len=1.71, thru=14.2, I=12.6, T=33.6, wait=7.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:04:19,482] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17580: forward=2704.1ms backward=1059.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:19,733] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:26,065] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:04:27,792] loop.py:45 INFO TRAIN_STEP step=17582 loss=1.286227 Training: 98%|█████████▊| 17582/18018 [37:34:13<58:55, 8.11s/step, loss=1.286, acc=0.702, acc_len=1.69, thru=15.7, I=20.0, T=28.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:28,082] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:32,980] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:04:34,241] loop.py:45 INFO TRAIN_STEP step=17583 loss=1.210220 Training: 98%|█████████▊| 17583/18018 [37:34:20<55:10, 7.61s/step, loss=1.210, acc=0.718, acc_len=1.74, thru=16.4, I=22.9, T=36.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:34,491] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:39,978] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:04:40,713] loop.py:45 INFO TRAIN_STEP step=17584 loss=1.256777 Training: 98%|█████████▊| 17584/18018 [37:34:26<52:34, 7.27s/step, loss=1.257, acc=0.706, acc_len=1.70, thru=19.6, I=23.6, T=38.4, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:40,985] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:47,520] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:04:48,402] loop.py:380 INFO TIMING step=17585: step=3.497s data=1.494s compute=3.453s [fwd=2.407s bwd=1.032s opt=0.003s] dispatch=4.183s [2026-10-20 03:04:48,402] loop.py:45 INFO TRAIN_STEP step=17585 loss=1.332822 Training: 98%|█████████▊| 17585/18018 [37:34:34<53:21, 7.39s/step, loss=1.333, acc=0.696, acc_len=1.66, thru=18.1, I=20.5, T=36.6, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:48,535] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:53,765] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:04:54,684] loop.py:45 INFO TRAIN_STEP step=17586 loss=1.271949 Training: 98%|█████████▊| 17586/18018 [37:34:40<50:50, 7.06s/step, loss=1.272, acc=0.704, acc_len=1.69, thru=19.9, I=23.8, T=35.7, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:04:54,679] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17585: forward=2495.4ms backward=1040.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:04:54,777] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-20 03:05:00,832] loop.py:45 INFO TRAIN_STEP step=17587 loss=1.383224 Training: 98%|█████████▊| 17587/18018 [37:34:46<48:44, 6.79s/step, loss=1.383, acc=0.687, acc_len=1.64, thru=18.8, I=20.4, T=38.5, wait=2.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:08,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:05:09,031] loop.py:45 INFO TRAIN_STEP step=17588 loss=1.328840 Training: 98%|█████████▊| 17588/18018 [37:34:55<51:40, 7.21s/step, loss=1.329, acc=0.693, acc_len=1.66, thru=17.4, I=17.5, T=35.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:09,372] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:05:14,969] loop.py:45 INFO TRAIN_STEP step=17589 loss=1.218814 Training: 98%|█████████▊| 17589/18018 [37:35:01<48:49, 6.83s/step, loss=1.219, acc=0.716, acc_len=1.73, thru=19.9, I=21.1, T=38.4, wait=2.6s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:23,664] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:05:23,782] loop.py:380 INFO TIMING step=17590: step=3.822s data=1.687s compute=3.782s [fwd=2.683s bwd=1.084s opt=0.003s] dispatch=4.980s [2026-10-20 03:05:23,782] loop.py:45 INFO TRAIN_STEP step=17590 loss=1.211100 Training: 98%|█████████▊| 17590/18018 [37:35:09<52:57, 7.42s/step, loss=1.211, acc=0.717, acc_len=1.74, thru=15.5, I=16.1, T=33.5, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:24,167] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:30,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:05:30,556] loop.py:45 INFO TRAIN_STEP step=17591 loss=1.290204 Training: 98%|█████████▊| 17591/18018 [37:35:16<51:26, 7.23s/step, loss=1.290, acc=0.701, acc_len=1.69, thru=18.5, I=18.8, T=37.3, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:05:30,551] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17590: forward=2434.0ms backward=958.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:30,793] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:05:37,088] loop.py:45 INFO TRAIN_STEP step=17592 loss=1.295287 Training: 98%|█████████▊| 17592/18018 [37:35:23<49:50, 7.02s/step, loss=1.295, acc=0.703, acc_len=1.69, thru=19.2, I=19.4, T=32.0, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:43,684] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:05:44,339] loop.py:45 INFO TRAIN_STEP step=17593 loss=1.275886 Training: 98%|█████████▊| 17593/18018 [37:35:30<50:12, 7.09s/step, loss=1.276, acc=0.706, acc_len=1.70, thru=18.6, I=17.8, T=35.6, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:44,692] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:52,085] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:05:52,978] loop.py:45 INFO TRAIN_STEP step=17594 loss=1.222064 Training: 98%|█████████▊| 17594/18018 [37:35:39<53:22, 7.55s/step, loss=1.222, acc=0.709, acc_len=1.72, thru=15.4, I=17.9, T=32.9, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:53,093] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:05:59,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:06:00,276] loop.py:380 INFO TIMING step=17595: step=3.957s data=1.694s compute=3.920s [fwd=2.742s bwd=1.163s opt=0.003s] dispatch=3.333s [2026-10-20 03:06:00,276] loop.py:45 INFO TRAIN_STEP step=17595 loss=1.304980 Training: 98%|█████████▊| 17595/18018 [37:35:46<52:42, 7.48s/step, loss=1.305, acc=0.695, acc_len=1.67, thru=17.1, I=18.2, T=32.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:00,391] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-20 03:06:06,653] loop.py:45 INFO TRAIN_STEP step=17596 loss=1.152873 Training: 98%|█████████▊| 17596/18018 [37:35:52<50:16, 7.15s/step, loss=1.153, acc=0.724, acc_len=1.76, thru=17.8, I=18.6, T=36.6, wait=2.9s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:06:06,648] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17595: forward=2444.6ms backward=992.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:12,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:06:13,290] loop.py:45 INFO TRAIN_STEP step=17597 loss=1.226085 Training: 98%|█████████▊| 17597/18018 [37:35:59<49:04, 6.99s/step, loss=1.226, acc=0.714, acc_len=1.73, thru=19.3, I=20.6, T=36.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:13,394] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:20,261] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:06:20,380] loop.py:45 INFO TRAIN_STEP step=17598 loss=1.312856 Training: 98%|█████████▊| 17598/18018 [37:36:06<49:09, 7.02s/step, loss=1.313, acc=0.699, acc_len=1.68, thru=17.8, I=20.2, T=33.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:20,764] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:27,001] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:06:28,825] loop.py:45 INFO TRAIN_STEP step=17599 loss=1.422685 Training: 98%|█████████▊| 17599/18018 [37:36:14<52:01, 7.45s/step, loss=1.423, acc=0.680, acc_len=1.60, thru=16.0, I=19.4, T=27.3, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:29,008] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:34,894] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:06:36,468] loop.py:380 INFO TIMING step=17600: step=4.465s data=2.369s compute=4.424s [fwd=3.386s bwd=1.024s opt=0.003s] dispatch=3.169s [2026-10-20 03:06:36,468] loop.py:45 INFO TRAIN_STEP step=17600 loss=1.258883 Training: 98%|█████████▊| 17600/18018 [37:36:22<52:18, 7.51s/step, loss=1.259, acc=0.706, acc_len=1.70, thru=14.4, I=21.0, T=28.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:36,904] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:43,544] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:06:45,129] loop.py:45 INFO TRAIN_STEP step=17601 loss=1.245795 Training: 98%|█████████▊| 17601/18018 [37:36:31<54:34, 7.85s/step, loss=1.246, acc=0.709, acc_len=1.71, thru=15.6, I=19.3, T=26.6, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:06:45,124] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17600: forward=3490.5ms backward=1281.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:45,555] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:51,227] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:06:51,687] loop.py:45 INFO TRAIN_STEP step=17602 loss=1.219340 Training: 98%|█████████▊| 17602/18018 [37:36:37<51:45, 7.47s/step, loss=1.219, acc=0.717, acc_len=1.74, thru=17.8, I=23.0, T=33.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:06:51,741] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-20 03:06:59,146] loop.py:45 INFO TRAIN_STEP step=17603 loss=1.264685 Training: 98%|█████████▊| 17603/18018 [37:36:45<51:37, 7.46s/step, loss=1.265, acc=0.706, acc_len=1.70, thru=15.7, I=16.9, T=33.6, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:04,895] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:07:06,252] loop.py:45 INFO TRAIN_STEP step=17604 loss=1.163340 Training: 98%|█████████▊| 17604/18018 [37:36:52<50:45, 7.36s/step, loss=1.163, acc=0.723, acc_len=1.77, thru=19.0, I=19.2, T=35.1, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:06,407] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:11,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:07:11,935] loop.py:380 INFO TIMING step=17605: step=3.324s data=1.328s compute=3.299s [fwd=2.357s bwd=0.928s opt=0.003s] dispatch=2.349s [2026-10-20 03:07:11,935] loop.py:45 INFO TRAIN_STEP step=17605 loss=1.130262 Training: 98%|█████████▊| 17605/18018 [37:36:58<47:11, 6.86s/step, loss=1.130, acc=0.728, acc_len=1.77, thru=20.8, I=23.3, T=38.5, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:12,351] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:17,634] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:07:18,571] loop.py:45 INFO TRAIN_STEP step=17606 loss=1.153944 Training: 98%|█████████▊| 17606/18018 [37:37:04<46:36, 6.79s/step, loss=1.154, acc=0.725, acc_len=1.76, thru=19.6, I=21.7, T=33.3, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:07:18,567] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17605: forward=2732.5ms backward=1070.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:18,642] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-20 03:07:25,949] loop.py:45 INFO TRAIN_STEP step=17607 loss=1.308826 Training: 98%|█████████▊| 17607/18018 [37:37:12<47:42, 6.96s/step, loss=1.309, acc=0.700, acc_len=1.68, thru=15.8, I=16.5, T=30.9, wait=3.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:32,067] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:07:32,927] loop.py:45 INFO TRAIN_STEP step=17608 loss=1.259912 Training: 98%|█████████▊| 17608/18018 [37:37:19<47:37, 6.97s/step, loss=1.260, acc=0.707, acc_len=1.70, thru=18.0, I=18.0, T=34.3, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:33,082] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:39,347] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:07:40,752] loop.py:45 INFO TRAIN_STEP step=17609 loss=1.275973 Training: 98%|█████████▊| 17609/18018 [37:37:26<49:15, 7.23s/step, loss=1.276, acc=0.704, acc_len=1.69, thru=16.5, I=20.6, T=35.4, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:40,869] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-20 03:07:47,063] loop.py:380 INFO TIMING step=17610: step=3.767s data=1.599s compute=3.737s [fwd=2.616s bwd=1.106s opt=0.003s] dispatch=2.532s [2026-10-20 03:07:47,063] loop.py:45 INFO TRAIN_STEP step=17610 loss=1.143102 Training: 98%|█████████▊| 17610/18018 [37:37:33<47:16, 6.95s/step, loss=1.143, acc=0.731, acc_len=1.79, thru=18.5, I=21.3, T=34.0, wait=2.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:55,474] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:07:57,135] loop.py:45 INFO TRAIN_STEP step=17611 loss=1.335158 Training: 98%|█████████▊| 17611/18018 [37:37:43<53:30, 7.89s/step, loss=1.335, acc=0.694, acc_len=1.66, thru=14.4, I=15.7, T=29.9, wait=5.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:07:57,130] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17610: forward=3015.4ms backward=1178.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:07:57,489] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:02,754] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:08:04,581] loop.py:45 INFO TRAIN_STEP step=17612 loss=1.240834 Training: 98%|█████████▊| 17612/18018 [37:37:50<52:28, 7.75s/step, loss=1.241, acc=0.710, acc_len=1.71, thru=16.2, I=25.0, T=28.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:04,769] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:10,466] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:08:11,063] loop.py:45 INFO TRAIN_STEP step=17613 loss=1.252940 Training: 98%|█████████▊| 17613/18018 [37:37:57<49:46, 7.37s/step, loss=1.253, acc=0.709, acc_len=1.71, thru=17.7, I=22.0, T=32.0, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:11,470] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:08:17,668] loop.py:45 INFO TRAIN_STEP step=17614 loss=1.250851 Training: 98%|█████████▊| 17614/18018 [37:38:03<48:05, 7.14s/step, loss=1.251, acc=0.710, acc_len=1.71, thru=17.4, I=20.5, T=35.4, wait=3.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:23,849] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:08:24,993] loop.py:380 INFO TIMING step=17615: step=3.934s data=1.692s compute=3.906s [fwd=2.812s bwd=1.079s opt=0.003s] dispatch=3.382s [2026-10-20 03:08:24,993] loop.py:45 INFO TRAIN_STEP step=17615 loss=1.370660 Training: 98%|█████████▊| 17615/18018 [37:38:11<48:20, 7.20s/step, loss=1.371, acc=0.689, acc_len=1.65, thru=17.6, I=20.4, T=32.5, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:25,357] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:30,688] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:08:31,388] loop.py:45 INFO TRAIN_STEP step=17616 loss=1.144449 Training: 98%|█████████▊| 17616/18018 [37:38:17<46:36, 6.96s/step, loss=1.144, acc=0.730, acc_len=1.78, thru=18.8, I=23.7, T=36.5, wait=2.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:08:31,384] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17615: forward=2478.0ms backward=989.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:31,697] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:08:37,702] loop.py:45 INFO TRAIN_STEP step=17617 loss=1.218344 Training: 98%|█████████▊| 17617/18018 [37:38:23<45:12, 6.76s/step, loss=1.218, acc=0.715, acc_len=1.73, thru=18.7, I=20.7, T=31.5, wait=2.2s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:43,051] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:08:44,317] loop.py:45 INFO TRAIN_STEP step=17618 loss=1.203477 Training: 98%|█████████▊| 17618/18018 [37:38:30<44:47, 6.72s/step, loss=1.203, acc=0.717, acc_len=1.74, thru=19.6, I=19.7, T=34.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:44,561] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:50,168] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:08:51,243] loop.py:45 INFO TRAIN_STEP step=17619 loss=1.316793 Training: 98%|█████████▊| 17619/18018 [37:38:37<45:05, 6.78s/step, loss=1.317, acc=0.698, acc_len=1.67, thru=18.2, I=22.7, T=32.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:51,678] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:57,133] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:08:57,560] loop.py:380 INFO TIMING step=17620: step=3.773s data=1.697s compute=3.744s [fwd=2.590s bwd=1.140s opt=0.003s] dispatch=2.532s [2026-10-20 03:08:57,560] loop.py:45 INFO TRAIN_STEP step=17620 loss=1.243858 Training: 98%|█████████▊| 17620/18018 [37:38:43<44:03, 6.64s/step, loss=1.244, acc=0.711, acc_len=1.72, thru=19.0, I=23.3, T=33.9, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:08:57,645] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:03,324] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:09:04,301] loop.py:45 INFO TRAIN_STEP step=17621 loss=1.268792 Training: 98%|█████████▊| 17621/18018 [37:38:50<44:08, 6.67s/step, loss=1.269, acc=0.708, acc_len=1.71, thru=18.9, I=21.4, T=33.9, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:09:04,295] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17620: forward=2642.0ms backward=1079.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:04,336] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-20 03:09:12,492] loop.py:45 INFO TRAIN_STEP step=17622 loss=1.251516 Training: 98%|█████████▊| 17622/18018 [37:38:58<47:02, 7.13s/step, loss=1.252, acc=0.707, acc_len=1.70, thru=13.0, I=14.7, T=33.8, wait=4.4s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:19,655] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:09:21,115] loop.py:45 INFO TRAIN_STEP step=17623 loss=1.209286 Training: 98%|█████████▊| 17623/18018 [37:39:07<49:52, 7.58s/step, loss=1.209, acc=0.720, acc_len=1.75, thru=16.8, I=17.6, T=31.7, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:21,164] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:27,586] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:09:28,554] loop.py:45 INFO TRAIN_STEP step=17624 loss=1.256572 Training: 98%|█████████▊| 17624/18018 [37:39:14<49:28, 7.54s/step, loss=1.257, acc=0.707, acc_len=1.71, thru=15.7, I=21.3, T=32.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:28,623] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:34,729] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:09:34,878] loop.py:380 INFO TIMING step=17625: step=3.582s data=1.538s compute=3.549s [fwd=2.551s bwd=0.983s opt=0.003s] dispatch=2.732s [2026-10-20 03:09:34,878] loop.py:45 INFO TRAIN_STEP step=17625 loss=1.274667 Training: 98%|█████████▊| 17625/18018 [37:39:20<46:58, 7.17s/step, loss=1.275, acc=0.705, acc_len=1.70, thru=18.9, I=20.7, T=35.7, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:35,231] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:40,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:09:41,742] loop.py:45 INFO TRAIN_STEP step=17626 loss=1.329538 Training: 98%|█████████▊| 17626/18018 [37:39:27<46:14, 7.08s/step, loss=1.330, acc=0.696, acc_len=1.67, thru=19.0, I=22.1, T=30.9, wait=2.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:09:41,737] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17625: forward=2796.2ms backward=1300.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:42,228] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:48,131] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:09:49,305] loop.py:45 INFO TRAIN_STEP step=17627 loss=1.210133 Training: 98%|█████████▊| 17627/18018 [37:39:35<47:04, 7.22s/step, loss=1.210, acc=0.716, acc_len=1.74, thru=16.3, I=22.6, T=33.7, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:09:49,638] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:09:55,879] loop.py:45 INFO TRAIN_STEP step=17628 loss=1.269907 Training: 98%|█████████▊| 17628/18018 [37:39:41<45:41, 7.03s/step, loss=1.270, acc=0.706, acc_len=1.70, thru=15.9, I=22.1, T=36.3, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:10:03,129] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:10:03,440] loop.py:45 INFO TRAIN_STEP step=17629 loss=1.262490 Training: 98%|█████████▊| 17629/18018 [37:39:49<46:36, 7.19s/step, loss=1.262, acc=0.705, acc_len=1.69, thru=17.8, I=18.4, T=36.0, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:10:03,633] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:10:07,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:10:09,267] loop.py:380 INFO TIMING step=17630: step=3.597s data=1.567s compute=3.566s [fwd=2.604s bwd=0.947s opt=0.003s] dispatch=2.219s [2026-10-20 03:10:09,267] loop.py:45 INFO TRAIN_STEP step=17630 loss=1.283581 Training: 98%|█████████▊| 17630/18018 [37:39:55<43:50, 6.78s/step, loss=1.284, acc=0.702, acc_len=1.69, thru=21.2, I=24.5, T=35.6, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:10:09,346] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 [2026-10-20 03:10:16,599] loop.py:45 INFO TRAIN_STEP step=17631 loss=1.304652 Training: 98%|█████████▊| 17631/18018 [37:40:02<44:48, 6.95s/step, loss=1.305, acc=0.696, acc_len=1.66, thru=13.8, I=22.0, T=37.7, wait=3.9s, pool=33, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:10:16,595] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17630: forward=2391.6ms backward=949.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:10:26,221] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:10:26,619] loop.py:45 INFO TRAIN_STEP step=17632 loss=1.238026 Training: 98%|█████████▊| 17632/18018 [37:40:12<50:37, 7.87s/step, loss=1.238, acc=0.710, acc_len=1.71, thru=16.2, I=13.8, T=35.6, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:10:26,729] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 [2026-10-20 03:10:33,295] loop.py:45 INFO TRAIN_STEP step=17633 loss=1.333269 Training: 98%|█████████▊| 17633/18018 [37:40:19<48:11, 7.51s/step, loss=1.333, acc=0.694, acc_len=1.66, thru=18.3, I=19.9, T=39.2, wait=3.4s, pool=57, epoch=2/2][2026-10-20 03:10:41,908] loop.py:45 INFO TRAIN_STEP step=17634 loss=1.342222 Training: 98%|█████████▊| 17634/18018 [37:40:27<50:11, 7.84s/step, loss=1.342, acc=0.691, acc_len=1.65, thru=15.2, I=15.6, T=34.1, wait=4.8s, pool=57, epoch=2/2][2026-10-20 03:10:50,936] loop.py:380 INFO TIMING step=17635: step=3.622s data=1.527s compute=3.570s [fwd=2.530s bwd=1.026s opt=0.003s] dispatch=5.396s [2026-10-20 03:10:50,936] loop.py:45 INFO TRAIN_STEP step=17635 loss=1.104040 Training: 98%|█████████▊| 17635/18018 [37:40:37<52:19, 8.20s/step, loss=1.104, acc=0.733, acc_len=1.79, thru=14.8, I=15.3, T=35.3, wait=5.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:10:57,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:10:58,179] loop.py:45 INFO TRAIN_STEP step=17636 loss=1.365385 Training: 98%|█████████▊| 17636/18018 [37:40:44<50:22, 7.91s/step, loss=1.365, acc=0.688, acc_len=1.64, thru=19.1, I=18.4, T=34.3, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:10:58,174] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17635: forward=2684.5ms backward=997.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:10:58,325] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:11:03,779] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:11:03,984] loop.py:45 INFO TRAIN_STEP step=17637 loss=1.226507 Training: 98%|█████████▊| 17637/18018 [37:40:50<46:13, 7.28s/step, loss=1.227, acc=0.713, acc_len=1.73, thru=20.9, I=22.3, T=37.5, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:11:04,280] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:11:11,131] loop.py:45 INFO TRAIN_STEP step=17638 loss=1.332749 Training: 98%|█████████▊| 17638/18018 [37:40:57<45:50, 7.24s/step, loss=1.333, acc=0.692, acc_len=1.65, thru=14.8, I=16.8, T=37.1, wait=3.7s, pool=29, epoch=2/2][2026-10-20 03:11:19,236] loop.py:45 INFO TRAIN_STEP step=17639 loss=1.282658 Training: 98%|█████████▊| 17639/18018 [37:41:05<47:22, 7.50s/step, loss=1.283, acc=0.701, acc_len=1.68, thru=16.8, I=15.1, T=35.7, wait=4.5s, pool=45, epoch=2/2][2026-10-20 03:11:27,457] loop.py:380 INFO TIMING step=17640: step=3.700s data=1.528s compute=3.661s [fwd=2.609s bwd=1.037s opt=0.003s] dispatch=4.511s [2026-10-20 03:11:27,458] loop.py:45 INFO TRAIN_STEP step=17640 loss=1.327583 Training: 98%|█████████▊| 17640/18018 [37:41:13<48:36, 7.72s/step, loss=1.328, acc=0.693, acc_len=1.66, thru=15.0, I=17.1, T=34.6, wait=4.5s, pool=41, epoch=2/2][2026-10-20 03:11:38,744] loop.py:45 INFO TRAIN_STEP step=17641 loss=1.293349 Training: 98%|█████████▊| 17641/18018 [37:41:24<55:12, 8.79s/step, loss=1.293, acc=0.698, acc_len=1.67, thru=14.2, I=12.3, T=33.1, wait=7.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:11:38,737] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 03:11:38,740] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17640: forward=2698.5ms backward=1109.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:11:39,239] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:11:44,189] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:11:45,521] loop.py:45 INFO TRAIN_STEP step=17642 loss=1.154504 Training: 98%|█████████▊| 17642/18018 [37:41:31<51:17, 8.18s/step, loss=1.155, acc=0.726, acc_len=1.77, thru=18.8, I=20.6, T=34.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:11:45,698] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:11:51,299] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:11:52,056] loop.py:45 INFO TRAIN_STEP step=17643 loss=1.233672 Training: 98%|█████████▊| 17643/18018 [37:41:38<48:03, 7.69s/step, loss=1.234, acc=0.713, acc_len=1.73, thru=19.1, I=22.7, T=37.1, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:11:52,306] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:11:58,801] loop.py:45 INFO TRAIN_STEP step=17644 loss=1.258128 Training: 98%|█████████▊| 17644/18018 [37:41:44<46:09, 7.41s/step, loss=1.258, acc=0.704, acc_len=1.70, thru=19.0, I=17.9, T=35.0, wait=3.1s, pool=57, epoch=2/2][2026-10-20 03:12:06,996] loop.py:380 INFO TIMING step=17645: step=3.776s data=1.586s compute=3.736s [fwd=2.640s bwd=1.081s opt=0.003s] dispatch=4.410s [2026-10-20 03:12:06,996] loop.py:45 INFO TRAIN_STEP step=17645 loss=1.185924 Training: 98%|█████████▊| 17645/18018 [37:41:53<47:30, 7.64s/step, loss=1.186, acc=0.723, acc_len=1.76, thru=15.3, I=15.0, T=33.9, wait=4.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:12:13,764] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:12:13,933] loop.py:45 INFO TRAIN_STEP step=17646 loss=1.205600 Training: 98%|█████████▊| 17646/18018 [37:42:00<46:04, 7.43s/step, loss=1.206, acc=0.716, acc_len=1.73, thru=18.5, I=18.3, T=35.1, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:12:13,928] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17645: forward=2623.1ms backward=969.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:12:14,265] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:12:20,298] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:12:20,578] loop.py:45 INFO TRAIN_STEP step=17647 loss=1.214004 Training: 98%|█████████▊| 17647/18018 [37:42:06<44:29, 7.20s/step, loss=1.214, acc=0.714, acc_len=1.73, thru=18.3, I=19.9, T=36.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:12:20,801] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:12:27,808] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:12:28,161] loop.py:45 INFO TRAIN_STEP step=17648 loss=1.305979 Training: 98%|█████████▊| 17648/18018 [37:42:14<45:05, 7.31s/step, loss=1.306, acc=0.699, acc_len=1.67, thru=17.3, I=18.5, T=35.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:12:28,310] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:12:33,907] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:12:35,084] loop.py:45 INFO TRAIN_STEP step=17649 loss=1.215309 Training: 98%|█████████▊| 17649/18018 [37:42:21<44:15, 7.20s/step, loss=1.215, acc=0.717, acc_len=1.74, thru=18.4, I=20.8, T=36.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:12:35,416] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:12:42,109] loop.py:380 INFO TIMING step=17650: step=3.446s data=1.458s compute=3.415s [fwd=2.448s bwd=0.952s opt=0.003s] dispatch=3.568s [2026-10-20 03:12:42,109] loop.py:45 INFO TRAIN_STEP step=17650 loss=1.162825 Training: 98%|█████████▊| 17650/18018 [37:42:28<43:49, 7.14s/step, loss=1.163, acc=0.724, acc_len=1.77, thru=15.6, I=17.8, T=37.1, wait=3.6s, pool=37, epoch=2/2][2026-10-20 03:12:51,483] loop.py:45 INFO TRAIN_STEP step=17651 loss=1.465619 Training: 98%|█████████▊| 17651/18018 [37:42:37<47:47, 7.81s/step, loss=1.466, acc=0.674, acc_len=1.59, thru=14.6, I=18.1, T=35.3, wait=5.7s, pool=37, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:12:51,479] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17650: forward=2581.6ms backward=998.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:12:58,730] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:13:00,501] loop.py:45 INFO TRAIN_STEP step=17652 loss=1.256251 Training: 98%|█████████▊| 17652/18018 [37:42:46<49:51, 8.17s/step, loss=1.256, acc=0.709, acc_len=1.72, thru=17.5, I=17.1, T=32.2, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:13:00,743] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:13:04,834] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:13:06,000] loop.py:45 INFO TRAIN_STEP step=17653 loss=1.350968 Training: 98%|█████████▊| 17653/18018 [37:42:52<44:50, 7.37s/step, loss=1.351, acc=0.691, acc_len=1.65, thru=20.9, I=29.2, T=38.7, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:13:06,344] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:13:11,684] loop.py:45 INFO TRAIN_STEP step=17654 loss=1.236153 Training: 98%|█████████▊| 17654/18018 [37:42:57<41:38, 6.87s/step, loss=1.236, acc=0.714, acc_len=1.73, thru=22.0, I=25.6, T=39.9, wait=2.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:13:19,220] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:13:19,408] loop.py:380 INFO TIMING step=17655: step=3.531s data=1.481s compute=3.506s [fwd=2.524s bwd=0.967s opt=0.003s] dispatch=4.184s [2026-10-20 03:13:19,408] loop.py:45 INFO TRAIN_STEP step=17655 loss=1.224133 Training: 98%|█████████▊| 17655/18018 [37:43:05<43:05, 7.12s/step, loss=1.224, acc=0.713, acc_len=1.72, thru=17.4, I=17.9, T=36.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:13:19,722] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:13:27,011] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:13:27,482] loop.py:45 INFO TRAIN_STEP step=17656 loss=1.346118 (TrainerActor pid=9310) [2026-10-20 03:13:27,477] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17655: forward=2529.5ms backward=1009.7ms Training: 98%|█████████▊| 17656/18018 [37:43:13<44:41, 7.41s/step, loss=1.346, acc=0.692, acc_len=1.65, thru=16.5, I=17.2, T=35.6, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:13:27,513] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 [2026-10-20 03:13:36,963] loop.py:45 INFO TRAIN_STEP step=17657 loss=1.301142 Training: 98%|█████████▊| 17657/18018 [37:43:23<48:19, 8.03s/step, loss=1.301, acc=0.697, acc_len=1.67, thru=12.9, I=12.9, T=32.9, wait=5.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:13:42,465] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:13:43,172] loop.py:45 INFO TRAIN_STEP step=17658 loss=1.208097 Training: 98%|█████████▊| 17658/18018 [37:43:29<44:54, 7.48s/step, loss=1.208, acc=0.717, acc_len=1.75, thru=19.8, I=17.5, T=35.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:13:43,470] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:13:49,123] loop.py:45 INFO TRAIN_STEP step=17659 loss=1.329031 Training: 98%|█████████▊| 17659/18018 [37:43:35<42:01, 7.02s/step, loss=1.329, acc=0.693, acc_len=1.65, thru=18.4, I=20.9, T=38.3, wait=2.6s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:00,515] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:14:00,847] loop.py:380 INFO TIMING step=17660: step=4.139s data=1.712s compute=4.106s [fwd=2.844s bwd=1.248s opt=0.003s] dispatch=7.576s [2026-10-20 03:14:00,848] loop.py:45 INFO TRAIN_STEP step=17660 loss=1.286701 Training: 98%|█████████▊| 17660/18018 [37:43:46<50:19, 8.43s/step, loss=1.287, acc=0.701, acc_len=1.68, thru=13.8, I=13.6, T=30.9, wait=7.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:01,016] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:06,169] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:14:07,356] loop.py:45 INFO TRAIN_STEP step=17661 loss=1.294943 Training: 98%|█████████▊| 17661/18018 [37:43:53<46:44, 7.86s/step, loss=1.295, acc=0.699, acc_len=1.68, thru=18.2, I=21.6, T=34.6, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:14:07,351] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17660: forward=2604.6ms backward=1023.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:07,676] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:13,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:14:14,130] loop.py:45 INFO TRAIN_STEP step=17662 loss=1.197696 Training: 98%|█████████▊| 17662/18018 [37:44:00<44:41, 7.53s/step, loss=1.198, acc=0.719, acc_len=1.74, thru=18.2, I=23.3, T=35.7, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:14,339] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:19,458] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:14:20,885] loop.py:45 INFO TRAIN_STEP step=17663 loss=1.163864 Training: 98%|█████████▊| 17663/18018 [37:44:06<43:11, 7.30s/step, loss=1.164, acc=0.722, acc_len=1.76, thru=18.6, I=24.2, T=35.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:20,990] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-20 03:14:26,716] loop.py:45 INFO TRAIN_STEP step=17664 loss=1.204904 Training: 98%|█████████▊| 17664/18018 [37:44:12<40:27, 6.86s/step, loss=1.205, acc=0.714, acc_len=1.72, thru=19.6, I=21.0, T=37.7, wait=2.4s, pool=49, epoch=2/2][2026-10-20 03:14:33,961] loop.py:380 INFO TIMING step=17665: step=3.442s data=1.374s compute=3.408s [fwd=2.442s bwd=0.952s opt=0.003s] dispatch=3.794s [2026-10-20 03:14:33,961] loop.py:45 INFO TRAIN_STEP step=17665 loss=1.229079 Training: 98%|█████████▊| 17665/18018 [37:44:20<41:01, 6.97s/step, loss=1.229, acc=0.711, acc_len=1.72, thru=18.5, I=18.4, T=37.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:33,877] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:34,379] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:39,536] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:14:40,661] loop.py:45 INFO TRAIN_STEP step=17666 loss=1.221453 Training: 98%|█████████▊| 17666/18018 [37:44:26<40:25, 6.89s/step, loss=1.221, acc=0.715, acc_len=1.73, thru=19.3, I=20.8, T=35.3, wait=3.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:14:40,657] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17665: forward=2550.5ms backward=1025.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:14:41,044] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:14:48,294] loop.py:45 INFO TRAIN_STEP step=17667 loss=1.193161 Training: 98%|█████████▊| 17667/18018 [37:44:34<41:37, 7.11s/step, loss=1.193, acc=0.718, acc_len=1.75, thru=15.6, I=18.3, T=36.1, wait=4.1s, pool=57, epoch=2/2][2026-10-20 03:14:55,633] loop.py:45 INFO TRAIN_STEP step=17668 loss=1.303488 Training: 98%|█████████▊| 17668/18018 [37:44:41<41:53, 7.18s/step, loss=1.303, acc=0.698, acc_len=1.67, thru=14.1, I=16.8, T=37.2, wait=3.9s, pool=25, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:06,803] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:15:07,694] loop.py:45 INFO TRAIN_STEP step=17669 loss=1.236048 Training: 98%|█████████▊| 17669/18018 [37:44:53<50:17, 8.65s/step, loss=1.236, acc=0.712, acc_len=1.72, thru=15.2, I=15.8, T=32.5, wait=8.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:07,810] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-20 03:15:15,863] loop.py:380 INFO TIMING step=17670: step=3.671s data=1.513s compute=3.640s [fwd=2.551s bwd=1.075s opt=0.003s] dispatch=4.488s [2026-10-20 03:15:15,863] loop.py:45 INFO TRAIN_STEP step=17670 loss=1.106829 Training: 98%|█████████▊| 17670/18018 [37:45:01<49:19, 8.50s/step, loss=1.107, acc=0.736, acc_len=1.80, thru=14.1, I=15.0, T=34.9, wait=4.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:24,450] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:15:26,078] loop.py:45 INFO TRAIN_STEP step=17671 loss=1.273820 Training: 98%|█████████▊| 17671/18018 [37:45:12<52:08, 9.02s/step, loss=1.274, acc=0.705, acc_len=1.69, thru=13.8, I=16.5, T=32.6, wait=6.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:15:26,074] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17670: forward=2865.4ms backward=1006.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:26,462] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 03:15:31,926] loop.py:45 INFO TRAIN_STEP step=17672 loss=1.191826 Training: 98%|█████████▊| 17672/18018 [37:45:18<46:30, 8.07s/step, loss=1.192, acc=0.717, acc_len=1.74, thru=17.5, I=24.1, T=37.1, wait=2.4s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:39,472] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:15:40,112] loop.py:45 INFO TRAIN_STEP step=17673 loss=1.218658 Training: 98%|█████████▊| 17673/18018 [37:45:26<46:35, 8.10s/step, loss=1.219, acc=0.713, acc_len=1.73, thru=17.7, I=17.8, T=36.3, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:40,477] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:46,310] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:15:46,406] loop.py:45 INFO TRAIN_STEP step=17674 loss=1.300269 Training: 98%|█████████▊| 17674/18018 [37:45:32<43:20, 7.56s/step, loss=1.300, acc=0.702, acc_len=1.68, thru=19.9, I=21.0, T=38.2, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:46,810] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:52,100] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:15:52,353] loop.py:380 INFO TIMING step=17675: step=3.272s data=1.316s compute=3.240s [fwd=2.340s bwd=0.886s opt=0.003s] dispatch=2.666s [2026-10-20 03:15:52,353] loop.py:45 INFO TRAIN_STEP step=17675 loss=1.241009 Training: 98%|█████████▊| 17675/18018 [37:45:38<40:27, 7.08s/step, loss=1.241, acc=0.713, acc_len=1.73, thru=21.6, I=22.5, T=39.1, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:52,603] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:56,835] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:15:57,936] loop.py:45 INFO TRAIN_STEP step=17676 loss=1.180103 Training: 98%|█████████▊| 17676/18018 [37:45:44<37:46, 6.63s/step, loss=1.180, acc=0.720, acc_len=1.75, thru=22.8, I=25.9, T=39.4, wait=2.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:15:57,931] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17675: forward=2307.9ms backward=898.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:15:58,344] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:16:04,378] loop.py:45 INFO TRAIN_STEP step=17677 loss=1.228788 Training: 98%|█████████▊| 17677/18018 [37:45:50<37:21, 6.57s/step, loss=1.229, acc=0.713, acc_len=1.72, thru=19.2, I=22.6, T=37.4, wait=3.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:16:10,906] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:16:11,192] loop.py:45 INFO TRAIN_STEP step=17678 loss=1.170295 Training: 98%|█████████▊| 17678/18018 [37:45:57<37:39, 6.64s/step, loss=1.170, acc=0.720, acc_len=1.75, thru=19.5, I=18.4, T=38.0, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:16:11,409] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:16:18,560] loop.py:45 INFO TRAIN_STEP step=17679 loss=1.373155 Training: 98%|█████████▊| 17679/18018 [37:46:04<38:46, 6.86s/step, loss=1.373, acc=0.690, acc_len=1.64, thru=16.9, I=18.7, T=36.7, wait=3.9s, pool=57, epoch=2/2][2026-10-20 03:16:27,722] loop.py:380 INFO TIMING step=17680: step=3.622s data=1.559s compute=3.577s [fwd=2.514s bwd=1.047s opt=0.003s] dispatch=5.530s [2026-10-20 03:16:27,722] loop.py:45 INFO TRAIN_STEP step=17680 loss=1.215534 Training: 98%|█████████▊| 17680/18018 [37:46:13<42:32, 7.55s/step, loss=1.216, acc=0.713, acc_len=1.73, thru=14.2, I=13.8, T=35.3, wait=5.5s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:16:34,089] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:16:35,454] loop.py:45 INFO TRAIN_STEP step=17681 loss=1.223958 Training: 98%|█████████▊| 17681/18018 [37:46:21<42:43, 7.61s/step, loss=1.224, acc=0.713, acc_len=1.73, thru=17.4, I=18.4, T=34.7, wait=4.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:16:35,449] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17680: forward=2628.4ms backward=1010.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:16:35,605] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:16:41,497] loop.py:45 INFO TRAIN_STEP step=17682 loss=1.188632 Training: 98%|█████████▊| 17682/18018 [37:46:27<39:58, 7.14s/step, loss=1.189, acc=0.718, acc_len=1.74, thru=18.9, I=21.8, T=37.4, wait=2.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:16:48,014] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:16:48,602] loop.py:45 INFO TRAIN_STEP step=17683 loss=1.180339 Training: 98%|█████████▊| 17683/18018 [37:46:34<39:47, 7.13s/step, loss=1.180, acc=0.724, acc_len=1.76, thru=19.5, I=18.7, T=36.1, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:16:49,019] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:16:54,695] loop.py:45 INFO TRAIN_STEP step=17684 loss=1.294037 Training: 98%|█████████▊| 17684/18018 [37:46:40<37:56, 6.82s/step, loss=1.294, acc=0.703, acc_len=1.69, thru=18.9, I=22.4, T=36.6, wait=2.6s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:17:05,824] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:17:06,706] loop.py:380 INFO TIMING step=17685: step=3.789s data=1.508s compute=3.755s [fwd=2.706s bwd=1.034s opt=0.003s] dispatch=8.213s [2026-10-20 03:17:06,707] loop.py:45 INFO TRAIN_STEP step=17685 loss=1.279741 Training: 98%|█████████▊| 17685/18018 [37:46:52<46:29, 8.38s/step, loss=1.280, acc=0.700, acc_len=1.68, thru=13.3, I=11.7, T=33.8, wait=8.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:17:06,836] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-20 03:17:13,723] loop.py:45 INFO TRAIN_STEP step=17686 loss=1.255153 Training: 98%|█████████▊| 17686/18018 [37:46:59<44:05, 7.97s/step, loss=1.255, acc=0.705, acc_len=1.70, thru=17.7, I=18.1, T=38.3, wait=3.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:17:13,719] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17685: forward=2374.3ms backward=922.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:17:26,766] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:17:28,291] loop.py:45 INFO TRAIN_STEP step=17687 loss=1.298749 Training: 98%|█████████▊| 17687/18018 [37:47:14<54:52, 9.95s/step, loss=1.299, acc=0.698, acc_len=1.67, thru=11.9, I=10.0, T=31.5, wait=10.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:17:28,778] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 03:17:35,089] loop.py:45 INFO TRAIN_STEP step=17688 loss=1.117263 Training: 98%|█████████▊| 17688/18018 [37:47:21<49:30, 9.00s/step, loss=1.117, acc=0.732, acc_len=1.80, thru=14.5, I=18.4, T=35.8, wait=3.2s, pool=41, epoch=2/2] [2026-10-20 03:17:45,092] loop.py:45 INFO TRAIN_STEP step=17689 loss=1.337197 Training: 98%|█████████▊| 17689/18018 [37:47:31<51:00, 9.30s/step, loss=1.337, acc=0.696, acc_len=1.66, thru=15.6, I=14.9, T=35.5, wait=6.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:17:45,080] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:17:45,582] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:17:50,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:17:51,776] loop.py:380 INFO TIMING step=17690: step=3.754s data=1.684s compute=3.732s [fwd=2.682s bwd=1.035s opt=0.003s] dispatch=2.922s [2026-10-20 03:17:51,777] loop.py:45 INFO TRAIN_STEP step=17690 loss=1.232297 Training: 98%|█████████▊| 17690/18018 [37:47:37<46:33, 8.52s/step, loss=1.232, acc=0.709, acc_len=1.72, thru=18.9, I=21.7, T=34.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:17:52,274] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 03:17:58,514] loop.py:45 INFO TRAIN_STEP step=17691 loss=1.247262 Training: 98%|█████████▊| 17691/18018 [37:47:44<43:30, 7.98s/step, loss=1.247, acc=0.710, acc_len=1.72, thru=17.0, I=21.9, T=37.8, wait=3.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:17:58,510] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17690: forward=2365.2ms backward=968.7ms [2026-10-20 03:18:05,568] loop.py:45 INFO TRAIN_STEP step=17692 loss=1.261829 Training: 98%|█████████▊| 17692/18018 [37:47:51<41:51, 7.70s/step, loss=1.262, acc=0.707, acc_len=1.70, thru=16.8, I=17.2, T=36.1, wait=3.5s, pool=49, epoch=2/2][2026-10-20 03:18:17,966] loop.py:45 INFO TRAIN_STEP step=17693 loss=1.306898 Training: 98%|█████████▊| 17693/18018 [37:48:04<49:21, 9.11s/step, loss=1.307, acc=0.701, acc_len=1.68, thru=15.8, I=9.9, T=34.2, wait=8.7s, pool=61, epoch=2/2] [2026-10-20 03:18:23,744] loop.py:45 INFO TRAIN_STEP step=17694 loss=1.306930 Training: 98%|█████████▊| 17694/18018 [37:48:09<43:48, 8.11s/step, loss=1.307, acc=0.702, acc_len=1.69, thru=19.6, I=21.8, T=36.4, wait=2.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:18:31,173] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:18:32,086] loop.py:380 INFO TIMING step=17695: step=3.729s data=1.546s compute=3.694s [fwd=2.640s bwd=1.039s opt=0.003s] dispatch=4.605s [2026-10-20 03:18:32,086] loop.py:45 INFO TRAIN_STEP step=17695 loss=1.313594 Training: 98%|█████████▊| 17695/18018 [37:48:18<44:02, 8.18s/step, loss=1.314, acc=0.701, acc_len=1.68, thru=15.0, I=16.2, T=34.3, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:18:32,184] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:18:38,538] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:18:39,516] loop.py:45 INFO TRAIN_STEP step=17696 loss=1.268062 Training: 98%|█████████▊| 17696/18018 [37:48:25<42:42, 7.96s/step, loss=1.268, acc=0.705, acc_len=1.70, thru=16.1, I=19.2, T=36.6, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:18:39,511] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17695: forward=2444.9ms backward=1001.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:18:39,551] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 [2026-10-20 03:18:45,444] loop.py:45 INFO TRAIN_STEP step=17697 loss=1.208617 Training: 98%|█████████▊| 17697/18018 [37:48:31<39:18, 7.35s/step, loss=1.209, acc=0.719, acc_len=1.75, thru=17.0, I=18.1, T=37.9, wait=2.5s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:18:52,969] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:18:54,148] loop.py:45 INFO TRAIN_STEP step=17698 loss=1.297072 Training: 98%|█████████▊| 17698/18018 [37:48:40<41:21, 7.75s/step, loss=1.297, acc=0.699, acc_len=1.67, thru=17.3, I=18.5, T=35.6, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:18:54,478] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:18:59,588] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:19:00,120] loop.py:45 INFO TRAIN_STEP step=17699 loss=1.258622 Training: 98%|█████████▊| 17699/18018 [37:48:46<38:22, 7.22s/step, loss=1.259, acc=0.708, acc_len=1.70, thru=21.3, I=25.1, T=37.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:19:00,594] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:19:06,930] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:19:07,423] loop.py:380 INFO TIMING step=17700: step=3.691s data=1.609s compute=3.661s [fwd=2.608s bwd=1.038s opt=0.003s] dispatch=3.603s [2026-10-20 03:19:07,423] loop.py:45 INFO TRAIN_STEP step=17700 loss=1.325489 Training: 98%|█████████▊| 17700/18018 [37:48:53<38:24, 7.25s/step, loss=1.325, acc=0.699, acc_len=1.67, thru=16.9, I=19.2, T=34.7, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:19:07,472] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:19:13,096] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:19:14,209] loop.py:45 INFO TRAIN_STEP step=17701 loss=1.203247 Training: 98%|█████████▊| 17701/18018 [37:49:00<37:32, 7.11s/step, loss=1.203, acc=0.720, acc_len=1.75, thru=18.6, I=21.2, T=36.7, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:19:14,205] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17700: forward=2462.2ms backward=973.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:19:14,606] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:19:20,692] loop.py:45 INFO TRAIN_STEP step=17702 loss=1.227733 Training: 98%|█████████▊| 17702/18018 [37:49:06<36:27, 6.92s/step, loss=1.228, acc=0.707, acc_len=1.70, thru=18.9, I=22.5, T=38.3, wait=3.1s, pool=61, epoch=2/2][2026-10-20 03:19:27,258] loop.py:45 INFO TRAIN_STEP step=17703 loss=1.248465 Training: 98%|█████████▊| 17703/18018 [37:49:13<35:46, 6.81s/step, loss=1.248, acc=0.713, acc_len=1.72, thru=16.5, I=19.5, T=34.7, wait=2.9s, pool=33, epoch=2/2][2026-10-20 03:19:36,217] loop.py:45 INFO TRAIN_STEP step=17704 loss=1.294804 Training: 98%|█████████▊| 17704/18018 [37:49:22<39:01, 7.46s/step, loss=1.295, acc=0.701, acc_len=1.69, thru=14.5, I=15.0, T=34.9, wait=5.3s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:19:44,652] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:19:45,474] loop.py:380 INFO TIMING step=17705: step=3.803s data=1.617s compute=3.775s [fwd=2.712s bwd=1.050s opt=0.003s] dispatch=5.446s [2026-10-20 03:19:45,474] loop.py:45 INFO TRAIN_STEP step=17705 loss=1.343189 Training: 98%|█████████▊| 17705/18018 [37:49:31<41:42, 8.00s/step, loss=1.343, acc=0.692, acc_len=1.65, thru=16.5, I=14.7, T=33.7, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:19:45,657] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:19:51,407] loop.py:45 INFO TRAIN_STEP step=17706 loss=1.146233 Training: 98%|█████████▊| 17706/18018 [37:49:37<38:21, 7.38s/step, loss=1.146, acc=0.726, acc_len=1.77, thru=17.6, I=19.0, T=37.2, wait=2.5s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:19:51,403] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17705: forward=2405.7ms backward=993.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:19:58,600] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:19:59,622] loop.py:45 INFO TRAIN_STEP step=17707 loss=1.187206 Training: 98%|█████████▊| 17707/18018 [37:49:45<39:32, 7.63s/step, loss=1.187, acc=0.721, acc_len=1.76, thru=17.5, I=18.5, T=33.9, wait=4.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:20:00,110] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:20:06,239] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:20:07,175] loop.py:45 INFO TRAIN_STEP step=17708 loss=1.159362 Training: 98%|█████████▊| 17708/18018 [37:49:53<39:17, 7.61s/step, loss=1.159, acc=0.726, acc_len=1.77, thru=16.1, I=20.9, T=37.7, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:20:07,247] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:20:15,508] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:20:16,264] loop.py:45 INFO TRAIN_STEP step=17709 loss=1.265505 Training: 98%|█████████▊| 17709/18018 [37:50:02<41:27, 8.05s/step, loss=1.266, acc=0.706, acc_len=1.70, thru=14.2, I=15.4, T=34.1, wait=5.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:20:16,515] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:20:24,132] loop.py:380 INFO TIMING step=17710: step=3.565s data=1.486s compute=3.537s [fwd=2.496s bwd=1.027s opt=0.003s] dispatch=4.294s [2026-10-20 03:20:24,133] loop.py:45 INFO TRAIN_STEP step=17710 loss=1.255994 Training: 98%|█████████▊| 17710/18018 [37:50:10<41:02, 8.00s/step, loss=1.256, acc=0.707, acc_len=1.71, thru=15.6, I=16.3, T=35.9, wait=4.3s, pool=49, epoch=2/2][2026-10-20 03:20:33,113] loop.py:45 INFO TRAIN_STEP step=17711 loss=1.264663 Training: 98%|█████████▊| 17711/18018 [37:50:19<42:25, 8.29s/step, loss=1.265, acc=0.709, acc_len=1.71, thru=15.4, I=15.0, T=29.6, wait=4.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:20:33,109] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17710: forward=3254.7ms backward=1023.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:20:39,659] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:20:40,793] loop.py:45 INFO TRAIN_STEP step=17712 loss=1.209188 Training: 98%|█████████▊| 17712/18018 [37:50:26<41:21, 8.11s/step, loss=1.209, acc=0.714, acc_len=1.73, thru=17.4, I=16.9, T=29.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:20:41,171] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:20:46,402] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:20:47,814] loop.py:45 INFO TRAIN_STEP step=17713 loss=1.247502 Training: 98%|█████████▊| 17713/18018 [37:50:33<39:33, 7.78s/step, loss=1.248, acc=0.709, acc_len=1.71, thru=17.0, I=21.3, T=30.2, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:20:47,929] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=25/64 [2026-10-20 03:20:53,966] loop.py:45 INFO TRAIN_STEP step=17714 loss=1.250262 Training: 98%|█████████▊| 17714/18018 [37:50:40<36:57, 7.29s/step, loss=1.250, acc=0.712, acc_len=1.72, thru=18.4, I=22.3, T=35.7, wait=2.6s, pool=61, epoch=2/2][2026-10-20 03:21:00,344] loop.py:380 INFO TIMING step=17715: step=3.504s data=1.433s compute=3.466s [fwd=2.483s bwd=0.969s opt=0.003s] dispatch=2.864s [2026-10-20 03:21:00,344] loop.py:45 INFO TRAIN_STEP step=17715 loss=1.330258 Training: 98%|█████████▊| 17715/18018 [37:50:46<35:26, 7.02s/step, loss=1.330, acc=0.694, acc_len=1.66, thru=17.6, I=19.2, T=36.5, wait=2.9s, pool=45, epoch=2/2][2026-10-20 03:21:07,670] loop.py:45 INFO TRAIN_STEP step=17716 loss=1.379113 Training: 98%|█████████▊| 17716/18018 [37:50:53<35:47, 7.11s/step, loss=1.379, acc=0.687, acc_len=1.64, thru=16.2, I=19.8, T=35.4, wait=3.7s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:21:07,665] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17715: forward=2485.4ms backward=1064.9ms [2026-10-20 03:21:16,989] loop.py:45 INFO TRAIN_STEP step=17717 loss=1.203134 Training: 98%|█████████▊| 17717/18018 [37:51:03<38:59, 7.77s/step, loss=1.203, acc=0.719, acc_len=1.75, thru=14.9, I=12.7, T=33.5, wait=5.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:21:25,646] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:21:26,169] loop.py:45 INFO TRAIN_STEP step=17718 loss=1.215495 Training: 98%|█████████▊| 17718/18018 [37:51:12<40:58, 8.20s/step, loss=1.215, acc=0.713, acc_len=1.73, thru=15.3, I=15.3, T=36.2, wait=5.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:21:26,652] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:21:33,565] loop.py:45 INFO TRAIN_STEP step=17719 loss=1.191491 Training: 98%|█████████▊| 17719/18018 [37:51:19<39:38, 7.96s/step, loss=1.191, acc=0.719, acc_len=1.75, thru=17.3, I=17.6, T=37.4, wait=4.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:21:41,604] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:21:41,721] loop.py:380 INFO TIMING step=17720: step=4.028s data=1.624s compute=3.999s [fwd=2.799s bwd=1.186s opt=0.003s] dispatch=4.118s [2026-10-20 03:21:41,721] loop.py:45 INFO TRAIN_STEP step=17720 loss=1.252436 Training: 98%|█████████▊| 17720/18018 [37:51:27<39:48, 8.02s/step, loss=1.252, acc=0.709, acc_len=1.71, thru=17.6, I=16.8, T=31.8, wait=4.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:21:42,106] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:21:49,368] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:21:50,033] loop.py:45 INFO TRAIN_STEP step=17721 loss=1.232122 Training: 98%|█████████▊| 17721/18018 [37:51:36<40:07, 8.11s/step, loss=1.232, acc=0.713, acc_len=1.72, thru=16.6, I=16.8, T=30.4, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:21:50,029] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17720: forward=2834.4ms backward=1306.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:21:50,373] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:21:55,287] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:21:56,109] loop.py:45 INFO TRAIN_STEP step=17722 loss=1.339993 Training: 98%|█████████▊| 17722/18018 [37:51:42<36:58, 7.50s/step, loss=1.340, acc=0.696, acc_len=1.67, thru=19.3, I=22.1, T=34.8, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:21:56,295] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:22:03,031] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:22:03,530] loop.py:45 INFO TRAIN_STEP step=17723 loss=1.227977 Training: 98%|█████████▊| 17723/18018 [37:51:49<36:44, 7.47s/step, loss=1.228, acc=0.713, acc_len=1.73, thru=15.6, I=18.6, T=35.2, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:22:04,038] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:22:09,293] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:22:09,751] loop.py:45 INFO TRAIN_STEP step=17724 loss=1.163254 Training: 98%|█████████▊| 17724/18018 [37:51:55<34:46, 7.10s/step, loss=1.163, acc=0.721, acc_len=1.76, thru=19.8, I=22.8, T=35.7, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:22:09,801] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-20 03:22:17,416] loop.py:380 INFO TIMING step=17725: step=3.451s data=1.407s compute=3.425s [fwd=2.482s bwd=0.928s opt=0.003s] dispatch=4.203s [2026-10-20 03:22:17,416] loop.py:45 INFO TRAIN_STEP step=17725 loss=1.268933 Training: 98%|█████████▊| 17725/18018 [37:52:03<35:29, 7.27s/step, loss=1.269, acc=0.707, acc_len=1.70, thru=14.6, I=15.5, T=37.1, wait=4.2s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:22:26,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:22:28,654] loop.py:45 INFO TRAIN_STEP step=17726 loss=1.243601 Training: 98%|█████████▊| 17726/18018 [37:52:14<41:09, 8.46s/step, loss=1.244, acc=0.713, acc_len=1.73, thru=14.9, I=15.3, T=32.0, wait=7.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:22:28,650] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17725: forward=2941.0ms backward=1013.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:22:28,750] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:22:34,985] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:22:35,189] loop.py:45 INFO TRAIN_STEP step=17727 loss=1.224941 Training: 98%|█████████▊| 17727/18018 [37:52:21<38:13, 7.88s/step, loss=1.225, acc=0.711, acc_len=1.72, thru=18.2, I=23.9, T=36.4, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:22:35,486] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:22:43,408] loop.py:45 INFO TRAIN_STEP step=17728 loss=1.279546 Training: 98%|█████████▊| 17728/18018 [37:52:29<38:35, 7.98s/step, loss=1.280, acc=0.706, acc_len=1.70, thru=16.1, I=16.2, T=33.9, wait=4.4s, pool=61, epoch=2/2][2026-10-20 03:22:49,505] loop.py:45 INFO TRAIN_STEP step=17729 loss=1.228826 Training: 98%|█████████▊| 17729/18018 [37:52:35<35:43, 7.42s/step, loss=1.229, acc=0.710, acc_len=1.72, thru=19.6, I=21.4, T=36.0, wait=2.5s, pool=61, epoch=2/2][2026-10-20 03:23:00,162] loop.py:380 INFO TIMING step=17730: step=4.462s data=2.394s compute=4.403s [fwd=3.138s bwd=1.251s opt=0.003s] dispatch=6.186s [2026-10-20 03:23:00,162] loop.py:45 INFO TRAIN_STEP step=17730 loss=1.353603 Training: 98%|█████████▊| 17730/18018 [37:52:46<40:15, 8.39s/step, loss=1.354, acc=0.693, acc_len=1.66, thru=11.7, I=12.1, T=28.7, wait=6.2s, pool=61, epoch=2/2][2026-10-20 03:23:08,139] loop.py:45 INFO TRAIN_STEP step=17731 loss=1.280658 Training: 98%|█████████▊| 17731/18018 [37:52:54<39:32, 8.27s/step, loss=1.281, acc=0.704, acc_len=1.69, thru=15.2, I=14.5, T=34.7, wait=4.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:23:08,135] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17730: forward=2593.0ms backward=1039.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:23:16,103] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:23:16,733] loop.py:45 INFO TRAIN_STEP step=17732 loss=1.231653 Training: 98%|█████████▊| 17732/18018 [37:53:02<39:52, 8.36s/step, loss=1.232, acc=0.712, acc_len=1.73, thru=16.5, I=17.1, T=35.7, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:23:17,111] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:23:22,750] loop.py:45 INFO TRAIN_STEP step=17733 loss=1.196357 Training: 98%|█████████▊| 17733/18018 [37:53:08<36:23, 7.66s/step, loss=1.196, acc=0.720, acc_len=1.75, thru=18.2, I=21.6, T=36.2, wait=2.5s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:23:32,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:23:32,406] loop.py:45 INFO TRAIN_STEP step=17734 loss=1.254039 Training: 98%|█████████▊| 17734/18018 [37:53:18<39:05, 8.26s/step, loss=1.254, acc=0.709, acc_len=1.71, thru=15.6, I=14.8, T=34.6, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:23:32,717] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:23:39,988] loop.py:380 INFO TIMING step=17735: step=3.947s data=1.768s compute=3.915s [fwd=2.791s bwd=1.109s opt=0.003s] dispatch=3.626s [2026-10-20 03:23:39,988] loop.py:45 INFO TRAIN_STEP step=17735 loss=1.101998 Training: 98%|█████████▊| 17735/18018 [37:53:26<37:59, 8.06s/step, loss=1.102, acc=0.732, acc_len=1.80, thru=16.9, I=17.5, T=32.4, wait=3.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:23:46,321] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:23:47,003] loop.py:45 INFO TRAIN_STEP step=17736 loss=1.245906 Training: 98%|█████████▊| 17736/18018 [37:53:33<36:23, 7.74s/step, loss=1.246, acc=0.712, acc_len=1.73, thru=17.8, I=18.4, T=34.0, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:23:46,998] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17735: forward=2566.0ms backward=1135.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:23:47,328] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:23:53,205] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:23:54,246] loop.py:45 INFO TRAIN_STEP step=17737 loss=1.344431 Training: 98%|█████████▊| 17737/18018 [37:53:40<35:33, 7.59s/step, loss=1.344, acc=0.691, acc_len=1.64, thru=18.5, I=20.1, T=37.3, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:23:54,713] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:24:00,041] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:24:01,351] loop.py:45 INFO TRAIN_STEP step=17738 loss=1.338894 Training: 98%|█████████▊| 17738/18018 [37:53:47<34:45, 7.45s/step, loss=1.339, acc=0.692, acc_len=1.65, thru=18.3, I=23.7, T=36.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:24:01,553] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:24:08,464] loop.py:45 INFO TRAIN_STEP step=17739 loss=1.167423 Training: 98%|█████████▊| 17739/18018 [37:53:54<34:09, 7.35s/step, loss=1.167, acc=0.728, acc_len=1.77, thru=15.2, I=18.7, T=38.5, wait=3.8s, pool=41, epoch=2/2][2026-10-20 03:24:18,564] loop.py:380 INFO TIMING step=17740: step=3.735s data=1.596s compute=3.704s [fwd=2.646s bwd=1.043s opt=0.003s] dispatch=6.357s [2026-10-20 03:24:18,565] loop.py:45 INFO TRAIN_STEP step=17740 loss=1.186149 Training: 98%|█████████▊| 17740/18018 [37:54:04<37:51, 8.17s/step, loss=1.186, acc=0.721, acc_len=1.75, thru=14.9, I=14.6, T=34.3, wait=6.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:24:25,611] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:24:27,037] loop.py:45 INFO TRAIN_STEP step=17741 loss=1.308293 Training: 98%|█████████▊| 17741/18018 [37:54:13<38:08, 8.26s/step, loss=1.308, acc=0.696, acc_len=1.67, thru=15.6, I=17.2, T=30.5, wait=4.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:24:27,032] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17740: forward=3089.8ms backward=1054.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:24:27,127] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:24:32,225] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:24:33,374] loop.py:45 INFO TRAIN_STEP step=17742 loss=1.321143 Training: 98%|█████████▊| 17742/18018 [37:54:19<35:21, 7.68s/step, loss=1.321, acc=0.693, acc_len=1.66, thru=19.0, I=21.7, T=37.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:24:33,732] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:24:39,284] loop.py:45 INFO TRAIN_STEP step=17743 loss=1.295638 Training: 98%|█████████▊| 17743/18018 [37:54:25<32:46, 7.15s/step, loss=1.296, acc=0.699, acc_len=1.68, thru=20.1, I=22.3, T=36.5, wait=2.4s, pool=57, epoch=2/2][2026-10-20 03:24:49,409] loop.py:45 INFO TRAIN_STEP step=17744 loss=1.272354 Training: 98%|█████████▊| 17744/18018 [37:54:35<36:43, 8.04s/step, loss=1.272, acc=0.705, acc_len=1.70, thru=12.3, I=12.3, T=31.3, wait=6.0s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:24:56,390] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:24:57,043] loop.py:380 INFO TIMING step=17745: step=3.651s data=1.493s compute=3.621s [fwd=2.632s bwd=0.974s opt=0.003s] dispatch=3.975s [2026-10-20 03:24:57,043] loop.py:45 INFO TRAIN_STEP step=17745 loss=1.221675 Training: 98%|█████████▊| 17745/18018 [37:54:43<36:02, 7.92s/step, loss=1.222, acc=0.709, acc_len=1.72, thru=17.6, I=17.6, T=35.1, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:24:57,395] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:25:04,130] loop.py:45 INFO TRAIN_STEP step=17746 loss=1.289187 Training: 98%|█████████▊| 17746/18018 [37:54:50<34:46, 7.67s/step, loss=1.289, acc=0.702, acc_len=1.69, thru=12.9, I=16.7, T=35.4, wait=3.5s, pool=25, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:25:04,126] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17745: forward=2573.2ms backward=1003.2ms (TrainerActor pid=1752, ip=10.0.3.62) [2026-10-20 03:25:04,192] eagle_store.py:517 WARNING batch_get_buffer missing keys (k567872-b4b852ea_hs, k567872-b4b852ea_ids, k567872-b4b852ea_lhs); sleeping 0.50s. [2026-10-20 03:25:14,727] loop.py:45 INFO TRAIN_STEP step=17747 loss=1.194864 Training: 98%|█████████▊| 17747/18018 [37:55:00<38:36, 8.55s/step, loss=1.195, acc=0.719, acc_len=1.75, thru=15.3, I=14.9, T=30.2, wait=6.4s, pool=61, epoch=2/2][2026-10-20 03:25:21,395] loop.py:45 INFO TRAIN_STEP step=17748 loss=1.166584 Training: 99%|█████████▊| 17748/18018 [37:55:07<35:55, 7.98s/step, loss=1.167, acc=0.720, acc_len=1.76, thru=17.2, I=16.7, T=35.6, wait=3.1s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:25:27,834] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-20 03:25:04,221] eagle_store.py:517 WARNING batch_get_buffer missing keys (k567873-9fac81b3_hs, k567873-9fac81b3_ids, k567873-9fac81b3_lhs); sleeping 0.50s. [repeated 2x across cluster] [2026-10-20 03:25:28,972] loop.py:45 INFO TRAIN_STEP step=17749 loss=1.313598 Training: 99%|█████████▊| 17749/18018 [37:55:15<35:15, 7.86s/step, loss=1.314, acc=0.699, acc_len=1.67, thru=18.3, I=19.8, T=33.9, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:25:29,343] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:25:35,290] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:25:36,006] loop.py:380 INFO TIMING step=17750: step=3.591s data=1.562s compute=3.563s [fwd=2.554s bwd=0.995s opt=0.003s] dispatch=3.431s [2026-10-20 03:25:36,006] loop.py:45 INFO TRAIN_STEP step=17750 loss=1.136226 Training: 99%|█████████▊| 17750/18018 [37:55:22<34:00, 7.61s/step, loss=1.136, acc=0.728, acc_len=1.79, thru=18.1, I=23.0, T=35.6, wait=3.4s, pool=65, epoch=2/2][2026-10-20 03:25:36,010] loop.py:419 INFO Saving checkpoint at step 17750... (TrainerActor pid=9310) [2026-10-20 03:25:39,406] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0017751 (TrainerActor pid=9310) [2026-10-20 03:25:39,676] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0017001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:25:40,332] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:25:40,834] inference_manager.py:408 INFO Pool capacity available after 5.5s, pool_size=1/64 [2026-10-20 03:25:48,222] loop.py:45 INFO TRAIN_STEP step=17751 loss=1.161832 Training: 99%|█████████▊| 17751/18018 [37:55:34<40:01, 8.99s/step, loss=1.162, acc=0.725, acc_len=1.77, thru=15.7, I=16.3, T=32.2, wait=3.7s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:25:48,217] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17750: forward=2770.0ms backward=1147.4ms (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 03:25:39,407] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0017751 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 03:25:40,559] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0017001 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:25:55,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:25:56,206] loop.py:45 INFO TRAIN_STEP step=17752 loss=1.298864 Training: 99%|█████████▊| 17752/18018 [37:55:42<38:31, 8.69s/step, loss=1.299, acc=0.698, acc_len=1.67, thru=16.4, I=18.2, T=34.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:25:56,589] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:26:02,128] loop.py:45 INFO TRAIN_STEP step=17753 loss=1.320016 Training: 99%|█████████▊| 17753/18018 [37:55:48<34:42, 7.86s/step, loss=1.320, acc=0.694, acc_len=1.66, thru=17.3, I=20.4, T=39.7, wait=2.7s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:26:08,896] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:26:09,958] loop.py:45 INFO TRAIN_STEP step=17754 loss=1.305981 Training: 99%|█████████▊| 17754/18018 [37:55:56<34:32, 7.85s/step, loss=1.306, acc=0.702, acc_len=1.69, thru=19.3, I=19.5, T=33.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:26:10,409] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:26:17,694] loop.py:380 INFO TIMING step=17755: step=3.746s data=1.599s compute=3.719s [fwd=2.685s bwd=1.019s opt=0.003s] dispatch=3.981s [2026-10-20 03:26:17,694] loop.py:45 INFO TRAIN_STEP step=17755 loss=1.186223 Training: 99%|█████████▊| 17755/18018 [37:56:03<34:15, 7.82s/step, loss=1.186, acc=0.721, acc_len=1.75, thru=14.4, I=17.5, T=34.2, wait=4.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:26:25,350] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 03:26:25,675] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17755: forward=2617.8ms backward=1074.6ms [2026-10-20 03:26:25,679] loop.py:45 INFO TRAIN_STEP step=17756 loss=1.301811 Training: 99%|█████████▊| 17756/18018 [37:56:11<34:21, 7.87s/step, loss=1.302, acc=0.699, acc_len=1.68, thru=15.7, I=16.6, T=34.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:26:25,852] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:26:33,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:26:34,227] loop.py:45 INFO TRAIN_STEP step=17757 loss=1.271520 Training: 99%|█████████▊| 17757/18018 [37:56:20<35:06, 8.07s/step, loss=1.272, acc=0.703, acc_len=1.69, thru=15.0, I=17.3, T=33.6, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:26:34,529] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:26:41,167] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:26:41,321] loop.py:45 INFO TRAIN_STEP step=17758 loss=1.160878 Training: 99%|█████████▊| 17758/18018 [37:56:27<33:42, 7.78s/step, loss=1.161, acc=0.724, acc_len=1.77, thru=16.6, I=17.3, T=35.9, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:26:41,669] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:26:51,308] loop.py:45 INFO TRAIN_STEP step=17759 loss=1.325169 Training: 99%|█████████▊| 17759/18018 [37:56:37<36:26, 8.44s/step, loss=1.325, acc=0.696, acc_len=1.67, thru=11.2, I=12.2, T=31.5, wait=5.9s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:26:59,869] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:27:01,575] loop.py:380 INFO TIMING step=17760: step=3.961s data=1.792s compute=3.914s [fwd=2.778s bwd=1.121s opt=0.003s] dispatch=6.298s [2026-10-20 03:27:01,576] loop.py:45 INFO TRAIN_STEP step=17760 loss=1.193236 Training: 99%|█████████▊| 17760/18018 [37:56:47<38:39, 8.99s/step, loss=1.193, acc=0.723, acc_len=1.76, thru=14.4, I=16.9, T=32.3, wait=6.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:27:01,883] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:27:07,013] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:27:08,214] loop.py:45 INFO TRAIN_STEP step=17761 loss=1.198777 Training: 99%|█████████▊| 17761/18018 [37:56:54<35:28, 8.28s/step, loss=1.199, acc=0.718, acc_len=1.74, thru=17.3, I=24.8, T=37.9, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:27:08,210] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17760: forward=2334.3ms backward=985.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:27:08,517] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:27:14,768] loop.py:45 INFO TRAIN_STEP step=17762 loss=1.188958 Training: 99%|█████████▊| 17762/18018 [37:57:00<33:07, 7.76s/step, loss=1.189, acc=0.719, acc_len=1.75, thru=18.9, I=21.1, T=39.4, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:27:20,072] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:27:21,249] loop.py:45 INFO TRAIN_STEP step=17763 loss=1.358529 Training: 99%|█████████▊| 17763/18018 [37:57:07<31:21, 7.38s/step, loss=1.359, acc=0.689, acc_len=1.64, thru=20.1, I=21.6, T=36.3, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:27:21,582] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:27:29,201] loop.py:45 INFO TRAIN_STEP step=17764 loss=1.230423 Training: 99%|█████████▊| 17764/18018 [37:57:15<31:57, 7.55s/step, loss=1.230, acc=0.715, acc_len=1.73, thru=12.8, I=16.0, T=31.5, wait=3.9s, pool=45, epoch=2/2][2026-10-20 03:27:37,286] loop.py:380 INFO TIMING step=17765: step=3.721s data=1.626s compute=3.690s [fwd=2.650s bwd=1.025s opt=0.003s] dispatch=4.357s [2026-10-20 03:27:37,286] loop.py:45 INFO TRAIN_STEP step=17765 loss=1.256913 Training: 99%|█████████▊| 17765/18018 [37:57:23<32:31, 7.71s/step, loss=1.257, acc=0.708, acc_len=1.70, thru=15.5, I=14.6, T=34.4, wait=4.4s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:27:44,879] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:27:45,989] loop.py:45 INFO TRAIN_STEP step=17766 loss=1.317574 Training: 99%|█████████▊| 17766/18018 [37:57:32<33:38, 8.01s/step, loss=1.318, acc=0.694, acc_len=1.66, thru=17.5, I=17.7, T=34.0, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:27:45,984] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17765: forward=2685.2ms backward=1023.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:27:46,391] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:27:53,721] loop.py:45 INFO TRAIN_STEP step=17767 loss=1.258688 Training: 99%|█████████▊| 17767/18018 [37:57:39<33:09, 7.93s/step, loss=1.259, acc=0.705, acc_len=1.70, thru=15.4, I=17.5, T=32.4, wait=3.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:28:00,760] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:28:02,087] loop.py:45 INFO TRAIN_STEP step=17768 loss=1.291277 Training: 99%|█████████▊| 17768/18018 [37:57:48<33:34, 8.06s/step, loss=1.291, acc=0.702, acc_len=1.68, thru=16.7, I=17.2, T=35.6, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:28:02,271] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:28:08,831] loop.py:45 INFO TRAIN_STEP step=17769 loss=1.182670 Training: 99%|█████████▊| 17769/18018 [37:57:54<31:48, 7.66s/step, loss=1.183, acc=0.720, acc_len=1.74, thru=14.8, I=18.3, T=36.0, wait=3.2s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:28:17,196] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:28:17,683] loop.py:380 INFO TIMING step=17770: step=3.895s data=1.640s compute=3.857s [fwd=2.786s bwd=1.055s opt=0.003s] dispatch=4.948s [2026-10-20 03:28:17,684] loop.py:45 INFO TRAIN_STEP step=17770 loss=1.229331 Training: 99%|█████████▊| 17770/18018 [37:58:03<33:09, 8.02s/step, loss=1.229, acc=0.720, acc_len=1.75, thru=17.6, I=17.3, T=32.9, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:28:17,719] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:28:24,585] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:28:25,587] loop.py:45 INFO TRAIN_STEP step=17771 loss=1.350837 Training: 99%|█████████▊| 17771/18018 [37:58:11<32:52, 7.99s/step, loss=1.351, acc=0.687, acc_len=1.63, thru=16.4, I=18.2, T=33.4, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:28:25,581] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17770: forward=2699.5ms backward=1084.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:28:26,094] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:28:32,307] loop.py:45 INFO TRAIN_STEP step=17772 loss=1.314054 Training: 99%|█████████▊| 17772/18018 [37:58:18<31:11, 7.61s/step, loss=1.314, acc=0.697, acc_len=1.67, thru=16.4, I=21.8, T=35.4, wait=3.1s, pool=53, epoch=2/2][2026-10-20 03:28:42,034] loop.py:45 INFO TRAIN_STEP step=17773 loss=1.245073 Training: 99%|█████████▊| 17773/18018 [37:58:28<33:39, 8.24s/step, loss=1.245, acc=0.710, acc_len=1.72, thru=13.3, I=12.2, T=32.4, wait=5.8s, pool=53, epoch=2/2][2026-10-20 03:28:53,341] loop.py:45 INFO TRAIN_STEP step=17774 loss=1.288926 Training: 99%|█████████▊| 17774/18018 [37:58:39<37:15, 9.16s/step, loss=1.289, acc=0.700, acc_len=1.68, thru=11.9, I=11.5, T=32.0, wait=7.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:28:59,847] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:29:01,019] loop.py:380 INFO TIMING step=17775: step=3.941s data=1.824s compute=3.876s [fwd=2.721s bwd=1.140s opt=0.003s] dispatch=3.727s [2026-10-20 03:29:01,019] loop.py:45 INFO TRAIN_STEP step=17775 loss=1.204774 Training: 99%|█████████▊| 17775/18018 [37:58:47<35:18, 8.72s/step, loss=1.205, acc=0.717, acc_len=1.73, thru=17.9, I=18.6, T=32.5, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:29:01,357] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:29:06,950] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:29:07,886] loop.py:45 INFO TRAIN_STEP step=17776 loss=1.316188 Training: 99%|█████████▊| 17776/18018 [37:58:53<32:54, 8.16s/step, loss=1.316, acc=0.695, acc_len=1.66, thru=17.9, I=23.0, T=36.3, wait=3.3s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:29:07,881] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17775: forward=2490.6ms backward=987.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:29:07,957] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:29:14,059] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:29:15,101] loop.py:45 INFO TRAIN_STEP step=17777 loss=1.244259 Training: 99%|█████████▊| 17777/18018 [37:59:01<31:38, 7.88s/step, loss=1.244, acc=0.706, acc_len=1.70, thru=17.3, I=20.8, T=33.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:29:15,571] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:29:22,179] loop.py:45 INFO TRAIN_STEP step=17778 loss=1.204329 Training: 99%|█████████▊| 17778/18018 [37:59:08<30:33, 7.64s/step, loss=1.204, acc=0.717, acc_len=1.73, thru=16.0, I=18.6, T=35.5, wait=3.5s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:29:26,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:29:28,426] loop.py:45 INFO TRAIN_STEP step=17779 loss=1.244512 Training: 99%|█████████▊| 17779/18018 [37:59:14<28:45, 7.22s/step, loss=1.245, acc=0.711, acc_len=1.72, thru=19.8, I=22.0, T=34.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:29:28,889] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:29:34,663] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:29:34,676] loop.py:380 INFO TIMING step=17780: step=3.290s data=1.303s compute=3.259s [fwd=2.340s bwd=0.904s opt=0.003s] dispatch=2.954s [2026-10-20 03:29:34,677] loop.py:45 INFO TRAIN_STEP step=17780 loss=1.233808 Training: 99%|█████████▊| 17780/18018 [37:59:20<27:29, 6.93s/step, loss=1.234, acc=0.709, acc_len=1.71, thru=19.6, I=24.7, T=38.9, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:29:35,164] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:29:43,279] loop.py:45 INFO TRAIN_STEP step=17781 loss=1.348417 Training: 99%|█████████▊| 17781/18018 [37:59:29<29:21, 7.43s/step, loss=1.348, acc=0.692, acc_len=1.65, thru=15.0, I=15.2, T=35.1, wait=4.9s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:29:43,275] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17780: forward=2573.7ms backward=1025.3ms [2026-10-20 03:29:51,210] loop.py:45 INFO TRAIN_STEP step=17782 loss=1.273341 Training: 99%|█████████▊| 17782/18018 [37:59:37<29:49, 7.58s/step, loss=1.273, acc=0.704, acc_len=1.70, thru=16.7, I=17.3, T=34.8, wait=4.2s, pool=53, epoch=2/2][2026-10-20 03:30:02,361] loop.py:45 INFO TRAIN_STEP step=17783 loss=1.273400 Training: 99%|█████████▊| 17783/18018 [37:59:48<33:53, 8.65s/step, loss=1.273, acc=0.703, acc_len=1.69, thru=12.7, I=10.5, T=33.7, wait=7.3s, pool=53, epoch=2/2][2026-10-20 03:30:10,323] loop.py:45 INFO TRAIN_STEP step=17784 loss=1.190205 Training: 99%|█████████▊| 17784/18018 [37:59:56<32:56, 8.45s/step, loss=1.190, acc=0.719, acc_len=1.75, thru=18.3, I=15.3, T=35.2, wait=4.3s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:30:15,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:30:16,943] loop.py:380 INFO TIMING step=17785: step=3.558s data=1.490s compute=3.529s [fwd=2.507s bwd=1.007s opt=0.003s] dispatch=3.052s [2026-10-20 03:30:16,943] loop.py:45 INFO TRAIN_STEP step=17785 loss=1.153435 Training: 99%|█████████▊| 17785/18018 [38:00:03<30:40, 7.90s/step, loss=1.153, acc=0.722, acc_len=1.76, thru=19.5, I=21.2, T=36.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:30:17,367] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:30:23,252] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:30:23,302] loop.py:45 INFO TRAIN_STEP step=17786 loss=1.291474 Training: 99%|█████████▊| 17786/18018 [38:00:09<28:45, 7.44s/step, loss=1.291, acc=0.700, acc_len=1.68, thru=19.7, I=22.9, T=36.4, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:30:23,297] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17785: forward=2517.9ms backward=956.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:30:23,753] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:30:30,317] loop.py:45 INFO TRAIN_STEP step=17787 loss=1.262538 Training: 99%|█████████▊| 17787/18018 [38:00:16<28:08, 7.31s/step, loss=1.263, acc=0.705, acc_len=1.70, thru=18.1, I=18.7, T=36.0, wait=3.5s, pool=57, epoch=2/2][2026-10-20 03:30:37,827] loop.py:45 INFO TRAIN_STEP step=17788 loss=1.260773 Training: 99%|█████████▊| 17788/18018 [38:00:23<28:15, 7.37s/step, loss=1.261, acc=0.708, acc_len=1.70, thru=17.4, I=17.1, T=36.2, wait=4.0s, pool=57, epoch=2/2][2026-10-20 03:30:45,340] loop.py:45 INFO TRAIN_STEP step=17789 loss=1.211004 Training: 99%|█████████▊| 17789/18018 [38:00:31<28:17, 7.41s/step, loss=1.211, acc=0.714, acc_len=1.73, thru=15.8, I=16.6, T=35.4, wait=3.9s, pool=37, epoch=2/2][2026-10-20 03:30:56,020] loop.py:380 INFO TIMING step=17790: step=3.883s data=1.631s compute=3.855s [fwd=2.750s bwd=1.090s opt=0.003s] dispatch=6.789s [2026-10-20 03:30:56,020] loop.py:45 INFO TRAIN_STEP step=17790 loss=1.193249 Training: 99%|█████████▊| 17790/18018 [38:00:42<31:53, 8.39s/step, loss=1.193, acc=0.718, acc_len=1.74, thru=16.4, I=14.1, T=33.0, wait=6.8s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:02,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:31:03,145] loop.py:45 INFO TRAIN_STEP step=17791 loss=1.277310 Training: 99%|█████████▊| 17791/18018 [38:00:49<30:18, 8.01s/step, loss=1.277, acc=0.708, acc_len=1.71, thru=16.1, I=17.2, T=36.0, wait=3.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:31:03,141] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17790: forward=2505.3ms backward=999.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:03,516] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:10,334] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:31:10,776] loop.py:45 INFO TRAIN_STEP step=17792 loss=1.081769 Training: 99%|█████████▊| 17792/18018 [38:00:56<29:45, 7.90s/step, loss=1.082, acc=0.740, acc_len=1.82, thru=16.5, I=18.6, T=35.6, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:10,842] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:16,476] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:31:17,177] loop.py:45 INFO TRAIN_STEP step=17793 loss=1.254146 Training: 99%|█████████▉| 17793/18018 [38:01:03<27:56, 7.45s/step, loss=1.254, acc=0.709, acc_len=1.71, thru=19.7, I=20.9, T=38.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:17,480] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:31:25,292] loop.py:45 INFO TRAIN_STEP step=17794 loss=1.212662 Training: 99%|█████████▉| 17794/18018 [38:01:11<28:33, 7.65s/step, loss=1.213, acc=0.710, acc_len=1.71, thru=14.1, I=17.1, T=36.4, wait=4.6s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:34,235] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:31:35,071] loop.py:380 INFO TIMING step=17795: step=3.729s data=1.590s compute=3.691s [fwd=2.685s bwd=0.991s opt=0.003s] dispatch=6.040s [2026-10-20 03:31:35,072] loop.py:45 INFO TRAIN_STEP step=17795 loss=1.323649 Training: 99%|█████████▉| 17795/18018 [38:01:21<30:48, 8.29s/step, loss=1.324, acc=0.696, acc_len=1.67, thru=15.4, I=14.5, T=34.3, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:35,241] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:42,317] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:31:43,198] loop.py:45 INFO TRAIN_STEP step=17796 loss=1.385356 Training: 99%|█████████▉| 17796/18018 [38:01:29<30:29, 8.24s/step, loss=1.385, acc=0.684, acc_len=1.62, thru=15.2, I=17.3, T=35.5, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:31:43,195] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17795: forward=2566.6ms backward=995.6ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:43,325] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:48,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:31:48,680] loop.py:45 INFO TRAIN_STEP step=17797 loss=1.246549 Training: 99%|█████████▉| 17797/18018 [38:01:34<27:18, 7.41s/step, loss=1.247, acc=0.705, acc_len=1.70, thru=21.0, I=21.4, T=36.9, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:31:48,911] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:31:57,554] loop.py:45 INFO TRAIN_STEP step=17798 loss=1.230296 Training: 99%|█████████▉| 17798/18018 [38:01:43<28:47, 7.85s/step, loss=1.230, acc=0.710, acc_len=1.71, thru=11.1, I=13.4, T=33.5, wait=5.0s, pool=33, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:08,620] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:32:09,025] loop.py:45 INFO TRAIN_STEP step=17799 loss=1.114038 Training: 99%|█████████▉| 17799/18018 [38:01:55<32:37, 8.94s/step, loss=1.114, acc=0.736, acc_len=1.81, thru=16.4, I=12.7, T=34.1, wait=7.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:09,123] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:15,406] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:32:15,803] loop.py:380 INFO TIMING step=17800: step=3.502s data=1.450s compute=3.473s [fwd=2.468s bwd=0.990s opt=0.003s] dispatch=3.267s [2026-10-20 03:32:15,803] loop.py:45 INFO TRAIN_STEP step=17800 loss=1.166860 Training: 99%|█████████▉| 17800/18018 [38:02:01<30:07, 8.29s/step, loss=1.167, acc=0.725, acc_len=1.77, thru=18.3, I=19.6, T=36.6, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:15,913] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:22,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:32:22,340] loop.py:45 INFO TRAIN_STEP step=17801 loss=1.227515 Training: 99%|█████████▉| 17801/18018 [38:02:08<28:04, 7.76s/step, loss=1.228, acc=0.713, acc_len=1.73, thru=19.8, I=21.6, T=34.7, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:32:22,336] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17800: forward=2649.6ms backward=985.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:22,716] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:29,944] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:32:31,006] loop.py:45 INFO TRAIN_STEP step=17802 loss=1.238861 Training: 99%|█████████▉| 17802/18018 [38:02:17<28:55, 8.03s/step, loss=1.239, acc=0.712, acc_len=1.72, thru=15.3, I=16.3, T=32.1, wait=4.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:31,453] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:37,546] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:32:38,221] loop.py:45 INFO TRAIN_STEP step=17803 loss=1.165699 Training: 99%|█████████▉| 17803/18018 [38:02:24<27:54, 7.79s/step, loss=1.166, acc=0.723, acc_len=1.76, thru=16.6, I=20.1, T=36.9, wait=3.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:32:38,551] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:32:46,694] loop.py:45 INFO TRAIN_STEP step=17804 loss=1.277883 Training: 99%|█████████▉| 17804/18018 [38:02:32<28:30, 7.99s/step, loss=1.278, acc=0.702, acc_len=1.69, thru=15.1, I=15.9, T=30.8, wait=4.3s, pool=61, epoch=2/2][2026-10-20 03:32:54,288] loop.py:380 INFO TIMING step=17805: step=3.513s data=1.462s compute=3.469s [fwd=2.468s bwd=0.987s opt=0.003s] dispatch=4.073s [2026-10-20 03:32:54,288] loop.py:45 INFO TRAIN_STEP step=17805 loss=1.243914 Training: 99%|█████████▉| 17805/18018 [38:02:40<27:57, 7.87s/step, loss=1.244, acc=0.710, acc_len=1.72, thru=15.2, I=16.3, T=36.4, wait=4.1s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:00,843] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:33:01,300] loop.py:45 INFO TRAIN_STEP step=17806 loss=1.206276 Training: 99%|█████████▉| 17806/18018 [38:02:47<26:54, 7.62s/step, loss=1.206, acc=0.717, acc_len=1.74, thru=18.3, I=19.4, T=36.6, wait=3.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:33:01,296] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17805: forward=2482.5ms backward=973.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:01,351] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:08,400] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:33:09,185] loop.py:45 INFO TRAIN_STEP step=17807 loss=1.120878 Training: 99%|█████████▉| 17807/18018 [38:02:55<27:03, 7.70s/step, loss=1.121, acc=0.733, acc_len=1.80, thru=17.2, I=19.4, T=36.0, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:09,405] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:14,859] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:33:15,493] loop.py:45 INFO TRAIN_STEP step=17808 loss=1.175113 Training: 99%|█████████▉| 17808/18018 [38:03:01<25:28, 7.28s/step, loss=1.175, acc=0.720, acc_len=1.75, thru=19.8, I=22.9, T=33.7, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:15,863] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:33:21,490] loop.py:45 INFO TRAIN_STEP step=17809 loss=1.229856 Training: 99%|█████████▉| 17809/18018 [38:03:07<24:01, 6.89s/step, loss=1.230, acc=0.711, acc_len=1.71, thru=19.8, I=22.8, T=37.6, wait=2.6s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:27,376] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:33:28,382] loop.py:380 INFO TIMING step=17810: step=3.350s data=1.360s compute=3.305s [fwd=2.351s bwd=0.941s opt=0.003s] dispatch=3.533s [2026-10-20 03:33:28,382] loop.py:45 INFO TRAIN_STEP step=17810 loss=1.212009 Training: 99%|█████████▉| 17810/18018 [38:03:14<23:54, 6.89s/step, loss=1.212, acc=0.716, acc_len=1.73, thru=18.6, I=20.5, T=38.2, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:28,889] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:33:34,393] loop.py:45 INFO TRAIN_STEP step=17811 loss=1.271664 Training: 99%|█████████▉| 17811/18018 [38:03:20<22:52, 6.63s/step, loss=1.272, acc=0.701, acc_len=1.69, thru=21.6, I=27.2, T=39.0, wait=2.7s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:33:34,388] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17810: forward=2351.2ms backward=890.6ms [2026-10-20 03:33:41,238] loop.py:45 INFO TRAIN_STEP step=17812 loss=1.137928 Training: 99%|█████████▉| 17812/18018 [38:03:27<22:58, 6.69s/step, loss=1.138, acc=0.727, acc_len=1.78, thru=17.9, I=19.0, T=37.1, wait=3.4s, pool=53, epoch=2/2][2026-10-20 03:33:48,205] loop.py:45 INFO TRAIN_STEP step=17813 loss=1.320186 Training: 99%|█████████▉| 17813/18018 [38:03:34<23:08, 6.78s/step, loss=1.320, acc=0.698, acc_len=1.67, thru=18.9, I=18.2, T=36.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:48,204] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:48,707] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:54,280] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:33:54,575] loop.py:45 INFO TRAIN_STEP step=17814 loss=1.195319 Training: 99%|█████████▉| 17814/18018 [38:03:40<22:37, 6.65s/step, loss=1.195, acc=0.717, acc_len=1.74, thru=19.7, I=20.2, T=38.1, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:33:54,782] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:34:00,553] loop.py:380 INFO TIMING step=17815: step=3.436s data=1.409s compute=3.402s [fwd=2.450s bwd=0.938s opt=0.003s] dispatch=2.532s [2026-10-20 03:34:00,553] loop.py:45 INFO TRAIN_STEP step=17815 loss=1.382013 Training: 99%|█████████▉| 17815/18018 [38:03:46<21:49, 6.45s/step, loss=1.382, acc=0.684, acc_len=1.63, thru=19.9, I=19.8, T=37.3, wait=2.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:07,155] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:34:07,865] loop.py:45 INFO TRAIN_STEP step=17816 loss=1.245235 Training: 99%|█████████▉| 17816/18018 [38:03:53<22:35, 6.71s/step, loss=1.245, acc=0.710, acc_len=1.72, thru=18.4, I=19.1, T=36.6, wait=3.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:34:07,861] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17815: forward=2520.3ms backward=932.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:08,161] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:34:16,330] loop.py:45 INFO TRAIN_STEP step=17817 loss=1.247586 Training: 99%|█████████▉| 17817/18018 [38:04:02<24:14, 7.24s/step, loss=1.248, acc=0.708, acc_len=1.71, thru=15.2, I=15.9, T=35.8, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:16,281] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:16,783] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:34:22,186] loop.py:45 INFO TRAIN_STEP step=17818 loss=1.301539 Training: 99%|█████████▉| 17818/18018 [38:04:08<22:44, 6.82s/step, loss=1.302, acc=0.699, acc_len=1.67, thru=19.8, I=19.8, T=38.4, wait=2.5s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:27,801] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:34:28,964] loop.py:45 INFO TRAIN_STEP step=17819 loss=1.234645 Training: 99%|█████████▉| 17819/18018 [38:04:15<22:34, 6.81s/step, loss=1.235, acc=0.711, acc_len=1.72, thru=19.5, I=21.6, T=36.7, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:29,308] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:34,402] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:34:35,373] loop.py:380 INFO TIMING step=17820: step=3.545s data=1.575s compute=3.514s [fwd=2.510s bwd=0.990s opt=0.003s] dispatch=2.855s [2026-10-20 03:34:35,373] loop.py:45 INFO TRAIN_STEP step=17820 loss=1.229492 Training: 99%|█████████▉| 17820/18018 [38:04:21<22:04, 6.69s/step, loss=1.229, acc=0.713, acc_len=1.72, thru=19.7, I=25.5, T=36.1, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:35,434] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=49/64 [2026-10-20 03:34:42,199] loop.py:45 INFO TRAIN_STEP step=17821 loss=1.173250 Training: 99%|█████████▉| 17821/18018 [38:04:28<22:05, 6.73s/step, loss=1.173, acc=0.723, acc_len=1.76, thru=18.2, I=19.9, T=37.5, wait=3.4s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:34:42,195] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17820: forward=2435.3ms backward=934.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:49,170] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:34:49,397] loop.py:45 INFO TRAIN_STEP step=17822 loss=1.225216 Training: 99%|█████████▉| 17822/18018 [38:04:35<22:26, 6.87s/step, loss=1.225, acc=0.710, acc_len=1.71, thru=18.1, I=18.3, T=34.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:49,674] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:57,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:34:57,486] loop.py:45 INFO TRAIN_STEP step=17823 loss=1.205856 Training: 99%|█████████▉| 17823/18018 [38:04:43<23:31, 7.24s/step, loss=1.206, acc=0.716, acc_len=1.73, thru=17.0, I=16.7, T=34.1, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:34:57,584] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:35:02,512] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:35:03,575] loop.py:45 INFO TRAIN_STEP step=17824 loss=1.208448 Training: 99%|█████████▉| 17824/18018 [38:04:49<22:16, 6.89s/step, loss=1.208, acc=0.714, acc_len=1.73, thru=19.8, I=21.5, T=35.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:35:04,019] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:35:10,566] loop.py:380 INFO TIMING step=17825: step=3.265s data=1.264s compute=3.234s [fwd=2.326s bwd=0.894s opt=0.003s] dispatch=3.717s [2026-10-20 03:35:10,566] loop.py:45 INFO TRAIN_STEP step=17825 loss=1.259035 Training: 99%|█████████▉| 17825/18018 [38:04:56<22:15, 6.92s/step, loss=1.259, acc=0.704, acc_len=1.69, thru=16.9, I=20.2, T=39.2, wait=3.7s, pool=57, epoch=2/2][2026-10-20 03:35:19,575] loop.py:45 INFO TRAIN_STEP step=17826 loss=1.112783 Training: 99%|█████████▉| 17826/18018 [38:05:05<24:09, 7.55s/step, loss=1.113, acc=0.732, acc_len=1.79, thru=14.6, I=16.0, T=35.9, wait=5.4s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:35:19,571] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17825: forward=2522.0ms backward=1003.9ms [2026-10-20 03:35:26,687] loop.py:45 INFO TRAIN_STEP step=17827 loss=1.265038 Training: 99%|█████████▉| 17827/18018 [38:05:12<23:36, 7.42s/step, loss=1.265, acc=0.703, acc_len=1.69, thru=17.2, I=18.1, T=37.0, wait=3.6s, pool=53, epoch=2/2][2026-10-20 03:35:36,853] loop.py:45 INFO TRAIN_STEP step=17828 loss=1.249843 Training: 99%|█████████▉| 17828/18018 [38:05:22<26:05, 8.24s/step, loss=1.250, acc=0.710, acc_len=1.71, thru=11.9, I=11.9, T=31.7, wait=6.1s, pool=45, epoch=2/2][2026-10-20 03:35:45,208] loop.py:45 INFO TRAIN_STEP step=17829 loss=1.270391 Training: 99%|█████████▉| 17829/18018 [38:05:31<26:04, 8.28s/step, loss=1.270, acc=0.707, acc_len=1.70, thru=13.6, I=13.5, T=36.5, wait=4.8s, pool=33, epoch=2/2][2026-10-20 03:35:53,890] loop.py:380 INFO TIMING step=17830: step=3.674s data=1.539s compute=3.644s [fwd=2.642s bwd=0.988s opt=0.003s] dispatch=5.000s [2026-10-20 03:35:53,890] loop.py:45 INFO TRAIN_STEP step=17830 loss=1.235745 Training: 99%|█████████▉| 17830/18018 [38:05:39<26:18, 8.40s/step, loss=1.236, acc=0.707, acc_len=1.71, thru=15.7, I=13.9, T=34.8, wait=5.0s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:00,222] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:36:00,689] loop.py:45 INFO TRAIN_STEP step=17831 loss=1.282412 Training: 99%|█████████▉| 17831/18018 [38:05:46<24:40, 7.92s/step, loss=1.282, acc=0.707, acc_len=1.70, thru=18.5, I=19.3, T=37.3, wait=3.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:36:00,684] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17830: forward=2381.9ms backward=993.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:00,744] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-20 03:36:05,872] loop.py:45 INFO TRAIN_STEP step=17832 loss=1.326388 Training: 99%|█████████▉| 17832/18018 [38:05:51<22:00, 7.10s/step, loss=1.326, acc=0.694, acc_len=1.66, thru=22.0, I=23.1, T=39.2, wait=1.9s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:13,050] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:36:13,400] loop.py:45 INFO TRAIN_STEP step=17833 loss=1.311354 Training: 99%|█████████▉| 17833/18018 [38:05:59<22:16, 7.23s/step, loss=1.311, acc=0.698, acc_len=1.67, thru=18.8, I=18.7, T=36.5, wait=4.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:13,551] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:20,112] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:36:21,021] loop.py:45 INFO TRAIN_STEP step=17834 loss=1.242812 Training: 99%|█████████▉| 17834/18018 [38:06:07<22:31, 7.35s/step, loss=1.243, acc=0.709, acc_len=1.71, thru=16.5, I=18.9, T=31.1, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:21,128] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=25/64 [2026-10-20 03:36:27,324] loop.py:380 INFO TIMING step=17835: step=3.616s data=1.598s compute=3.584s [fwd=2.576s bwd=0.992s opt=0.003s] dispatch=2.677s [2026-10-20 03:36:27,324] loop.py:45 INFO TRAIN_STEP step=17835 loss=1.330188 Training: 99%|█████████▉| 17835/18018 [38:06:13<21:27, 7.03s/step, loss=1.330, acc=0.694, acc_len=1.65, thru=18.7, I=21.5, T=35.4, wait=2.7s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:37,735] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:36:38,031] loop.py:45 INFO TRAIN_STEP step=17836 loss=1.314356 Training: 99%|█████████▉| 17836/18018 [38:06:24<24:40, 8.13s/step, loss=1.314, acc=0.698, acc_len=1.67, thru=13.9, I=12.9, T=31.1, wait=6.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:36:38,027] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17835: forward=2867.4ms backward=1196.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:38,236] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:42,702] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:36:44,139] loop.py:45 INFO TRAIN_STEP step=17837 loss=1.194679 Training: 99%|█████████▉| 17837/18018 [38:06:30<22:42, 7.53s/step, loss=1.195, acc=0.718, acc_len=1.75, thru=19.0, I=19.3, T=36.4, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:44,220] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-20 03:36:49,434] loop.py:45 INFO TRAIN_STEP step=17838 loss=1.176827 Training: 99%|█████████▉| 17838/18018 [38:06:35<20:34, 6.86s/step, loss=1.177, acc=0.722, acc_len=1.76, thru=22.0, I=26.6, T=39.5, wait=2.0s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:56,745] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:36:58,890] loop.py:45 INFO TRAIN_STEP step=17839 loss=1.370458 Training: 99%|█████████▉| 17839/18018 [38:06:44<22:46, 7.64s/step, loss=1.370, acc=0.688, acc_len=1.64, thru=14.9, I=18.8, T=31.5, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:36:59,258] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:37:04,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:37:04,692] loop.py:380 INFO TIMING step=17840: step=3.174s data=1.253s compute=3.145s [fwd=2.256s bwd=0.876s opt=0.003s] dispatch=2.622s [2026-10-20 03:37:04,693] loop.py:45 INFO TRAIN_STEP step=17840 loss=1.126405 Training: 99%|█████████▉| 17840/18018 [38:06:50<21:01, 7.09s/step, loss=1.126, acc=0.730, acc_len=1.78, thru=20.1, I=27.1, T=40.3, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:37:04,956] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:37:10,700] loop.py:45 INFO TRAIN_STEP step=17841 loss=1.341317 Training: 99%|█████████▉| 17841/18018 [38:06:56<19:57, 6.76s/step, loss=1.341, acc=0.692, acc_len=1.65, thru=18.5, I=22.0, T=36.0, wait=2.4s, pool=41, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:37:10,696] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17840: forward=2521.7ms backward=979.1ms [2026-10-20 03:37:17,918] loop.py:45 INFO TRAIN_STEP step=17842 loss=1.159893 Training: 99%|█████████▉| 17842/18018 [38:07:04<20:14, 6.90s/step, loss=1.160, acc=0.727, acc_len=1.78, thru=17.1, I=16.6, T=36.4, wait=3.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:37:25,842] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:37:26,973] loop.py:45 INFO TRAIN_STEP step=17843 loss=1.278924 Training: 99%|█████████▉| 17843/18018 [38:07:13<22:00, 7.55s/step, loss=1.279, acc=0.705, acc_len=1.69, thru=16.1, I=17.7, T=34.7, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:37:27,351] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:37:33,861] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:37:34,650] loop.py:45 INFO TRAIN_STEP step=17844 loss=1.356201 Training: 99%|█████████▉| 17844/18018 [38:07:20<21:59, 7.59s/step, loss=1.356, acc=0.689, acc_len=1.65, thru=16.0, I=20.8, T=37.2, wait=4.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:37:34,867] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:37:41,353] loop.py:380 INFO TIMING step=17845: step=3.292s data=1.347s compute=3.266s [fwd=2.357s bwd=0.895s opt=0.003s] dispatch=3.402s [2026-10-20 03:37:41,353] loop.py:45 INFO TRAIN_STEP step=17845 loss=1.242311 Training: 99%|█████████▉| 17845/18018 [38:07:27<21:06, 7.32s/step, loss=1.242, acc=0.710, acc_len=1.71, thru=19.4, I=21.2, T=38.9, wait=3.4s, pool=61, epoch=2/2][2026-10-20 03:37:49,533] loop.py:45 INFO TRAIN_STEP step=17846 loss=1.253840 Training: 99%|█████████▉| 17846/18018 [38:07:35<21:43, 7.58s/step, loss=1.254, acc=0.707, acc_len=1.70, thru=15.0, I=16.9, T=33.5, wait=4.3s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:37:49,529] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17845: forward=2665.0ms backward=1079.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:37:56,644] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:37:57,284] loop.py:45 INFO TRAIN_STEP step=17847 loss=1.162985 Training: 99%|█████████▉| 17847/18018 [38:07:43<21:44, 7.63s/step, loss=1.163, acc=0.728, acc_len=1.77, thru=17.7, I=18.1, T=30.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:37:57,648] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:02,566] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:38:03,699] loop.py:45 INFO TRAIN_STEP step=17848 loss=1.178628 Training: 99%|█████████▉| 17848/18018 [38:07:49<20:35, 7.27s/step, loss=1.179, acc=0.723, acc_len=1.76, thru=17.5, I=21.7, T=36.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:04,075] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:38:10,168] loop.py:45 INFO TRAIN_STEP step=17849 loss=1.229495 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:10,143] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 Training: 99%|█████████▉| 17849/18018 [38:07:56<19:47, 7.03s/step, loss=1.229, acc=0.708, acc_len=1.71, thru=19.6, I=23.0, T=35.8, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:10,645] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:38:17,108] loop.py:380 INFO TIMING step=17850: step=3.544s data=1.503s compute=3.514s [fwd=2.518s bwd=0.981s opt=0.003s] dispatch=3.389s [2026-10-20 03:38:17,108] loop.py:45 INFO TRAIN_STEP step=17850 loss=1.354490 Training: 99%|█████████▉| 17850/18018 [38:08:03<19:36, 7.00s/step, loss=1.354, acc=0.690, acc_len=1.64, thru=16.2, I=17.4, T=36.1, wait=3.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:23,603] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:38:23,847] loop.py:45 INFO TRAIN_STEP step=17851 loss=1.283975 Training: 99%|█████████▉| 17851/18018 [38:08:09<19:15, 6.92s/step, loss=1.284, acc=0.700, acc_len=1.67, thru=19.2, I=19.5, T=36.3, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:38:23,842] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17850: forward=2537.2ms backward=940.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:24,106] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:29,361] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:38:31,328] loop.py:45 INFO TRAIN_STEP step=17852 loss=1.259949 Training: 99%|█████████▉| 17852/18018 [38:08:17<19:36, 7.09s/step, loss=1.260, acc=0.704, acc_len=1.69, thru=17.7, I=20.9, T=30.9, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:31,384] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:37,551] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:38:39,095] loop.py:45 INFO TRAIN_STEP step=17853 loss=1.261714 Training: 99%|█████████▉| 17853/18018 [38:08:25<20:03, 7.29s/step, loss=1.262, acc=0.710, acc_len=1.71, thru=13.7, I=19.1, T=29.2, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:39,560] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 03:38:47,140] loop.py:45 INFO TRAIN_STEP step=17854 loss=1.191713 Training: 99%|█████████▉| 17854/18018 [38:08:33<20:33, 7.52s/step, loss=1.192, acc=0.718, acc_len=1.74, thru=11.8, I=16.5, T=35.5, wait=4.4s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:53,598] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:38:54,437] loop.py:380 INFO TIMING step=17855: step=3.706s data=1.537s compute=3.672s [fwd=2.668s bwd=0.990s opt=0.003s] dispatch=3.583s [2026-10-20 03:38:54,437] loop.py:45 INFO TRAIN_STEP step=17855 loss=1.199908 Training: 99%|█████████▉| 17855/18018 [38:08:40<20:14, 7.45s/step, loss=1.200, acc=0.718, acc_len=1.74, thru=18.8, I=19.4, T=34.5, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:38:54,603] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:02,690] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:39:02,899] loop.py:45 INFO TRAIN_STEP step=17856 loss=1.198948 Training: 99%|█████████▉| 17856/18018 [38:08:48<20:56, 7.76s/step, loss=1.199, acc=0.718, acc_len=1.74, thru=14.3, I=15.6, T=36.3, wait=4.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:39:02,894] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17855: forward=2522.2ms backward=958.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:03,192] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:39:09,150] loop.py:45 INFO TRAIN_STEP step=17857 loss=1.225748 Training: 99%|█████████▉| 17857/18018 [38:08:55<19:35, 7.30s/step, loss=1.226, acc=0.714, acc_len=1.73, thru=19.6, I=19.7, T=37.8, wait=2.9s, pool=61, epoch=2/2][2026-10-20 03:39:17,126] loop.py:45 INFO TRAIN_STEP step=17858 loss=1.189488 Training: 99%|█████████▉| 17858/18018 [38:09:03<20:00, 7.51s/step, loss=1.189, acc=0.713, acc_len=1.72, thru=15.8, I=15.8, T=37.2, wait=4.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:24,477] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:39:24,658] loop.py:45 INFO TRAIN_STEP step=17859 loss=1.270013 Training: 99%|█████████▉| 17859/18018 [38:09:10<19:54, 7.51s/step, loss=1.270, acc=0.705, acc_len=1.70, thru=18.9, I=17.0, T=35.0, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:24,980] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:39:31,407] loop.py:380 INFO TIMING step=17860: step=3.556s data=1.516s compute=3.518s [fwd=2.554s bwd=0.950s opt=0.003s] dispatch=3.185s [2026-10-20 03:39:31,407] loop.py:45 INFO TRAIN_STEP step=17860 loss=1.247631 Training: 99%|█████████▉| 17860/18018 [38:09:17<19:10, 7.28s/step, loss=1.248, acc=0.707, acc_len=1.70, thru=17.3, I=19.5, T=36.0, wait=3.2s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:40,788] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:39:41,069] loop.py:45 INFO TRAIN_STEP step=17861 loss=1.309374 Training: 99%|█████████▉| 17861/18018 [38:09:27<20:55, 8.00s/step, loss=1.309, acc=0.696, acc_len=1.67, thru=14.3, I=14.0, T=34.0, wait=5.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:39:41,065] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17860: forward=2593.9ms backward=1122.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:41,290] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:48,380] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:39:49,140] loop.py:45 INFO TRAIN_STEP step=17862 loss=1.183169 Training: 99%|█████████▉| 17862/18018 [38:09:35<20:51, 8.02s/step, loss=1.183, acc=0.718, acc_len=1.74, thru=15.7, I=17.2, T=33.7, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:49,384] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:54,315] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:39:55,116] loop.py:45 INFO TRAIN_STEP step=17863 loss=1.324141 Training: 99%|█████████▉| 17863/18018 [38:09:41<19:08, 7.41s/step, loss=1.324, acc=0.696, acc_len=1.67, thru=20.0, I=22.2, T=36.5, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:55,321] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:39:59,409] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:40:01,614] loop.py:45 INFO TRAIN_STEP step=17864 loss=1.264769 Training: 99%|█████████▉| 17864/18018 [38:09:47<18:18, 7.13s/step, loss=1.265, acc=0.705, acc_len=1.69, thru=19.4, I=26.6, T=28.9, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:01,927] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:07,675] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:40:08,133] loop.py:380 INFO TIMING step=17865: step=3.428s data=1.472s compute=3.382s [fwd=2.425s bwd=0.943s opt=0.003s] dispatch=3.080s [2026-10-20 03:40:08,133] loop.py:45 INFO TRAIN_STEP step=17865 loss=1.281258 Training: 99%|█████████▉| 17865/18018 [38:09:54<17:43, 6.95s/step, loss=1.281, acc=0.704, acc_len=1.70, thru=16.5, I=22.4, T=37.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:08,180] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=49/64 [2026-10-20 03:40:17,465] loop.py:45 INFO TRAIN_STEP step=17866 loss=1.388244 Training: 99%|█████████▉| 17866/18018 [38:10:03<19:25, 7.66s/step, loss=1.388, acc=0.686, acc_len=1.64, thru=13.5, I=13.0, T=31.3, wait=5.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:40:17,457] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17865: forward=2810.1ms backward=1200.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:17,471] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:17,973] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:24,154] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:40:24,496] loop.py:45 INFO TRAIN_STEP step=17867 loss=1.318303 Training: 99%|█████████▉| 17867/18018 [38:10:10<18:48, 7.47s/step, loss=1.318, acc=0.697, acc_len=1.68, thru=18.5, I=19.8, T=36.7, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:24,657] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:40:31,903] loop.py:45 INFO TRAIN_STEP step=17868 loss=1.142947 Training: 99%|█████████▉| 17868/18018 [38:10:17<18:38, 7.45s/step, loss=1.143, acc=0.728, acc_len=1.77, thru=15.7, I=15.5, T=34.1, wait=3.6s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:40,248] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:40:42,166] loop.py:45 INFO TRAIN_STEP step=17869 loss=1.397994 Training: 99%|█████████▉| 17869/18018 [38:10:28<20:36, 8.30s/step, loss=1.398, acc=0.682, acc_len=1.63, thru=13.6, I=16.9, T=29.3, wait=5.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:42,277] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=25/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:47,714] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:40:48,096] loop.py:380 INFO TIMING step=17870: step=3.606s data=1.543s compute=3.565s [fwd=2.499s bwd=1.052s opt=0.003s] dispatch=2.314s [2026-10-20 03:40:48,097] loop.py:45 INFO TRAIN_STEP step=17870 loss=1.213927 Training: 99%|█████████▉| 17870/18018 [38:10:34<18:42, 7.59s/step, loss=1.214, acc=0.717, acc_len=1.74, thru=19.1, I=23.1, T=35.5, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:48,221] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:55,428] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:40:56,143] loop.py:45 INFO TRAIN_STEP step=17871 loss=1.324853 Training: 99%|█████████▉| 17871/18018 [38:10:42<18:55, 7.73s/step, loss=1.325, acc=0.696, acc_len=1.66, thru=17.7, I=18.5, T=33.0, wait=4.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:40:56,138] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17870: forward=2751.1ms backward=1064.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:40:56,435] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:01,456] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:41:02,390] loop.py:45 INFO TRAIN_STEP step=17872 loss=1.283546 Training: 99%|█████████▉| 17872/18018 [38:10:48<17:43, 7.28s/step, loss=1.284, acc=0.704, acc_len=1.69, thru=18.8, I=22.8, T=36.2, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:02,462] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:07,446] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:41:07,987] loop.py:45 INFO TRAIN_STEP step=17873 loss=1.266227 Training: 99%|█████████▉| 17873/18018 [38:10:54<16:22, 6.78s/step, loss=1.266, acc=0.707, acc_len=1.70, thru=21.4, I=24.4, T=35.9, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:08,450] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:41:14,222] loop.py:45 INFO TRAIN_STEP step=17874 loss=1.316916 Training: 99%|█████████▉| 17874/18018 [38:11:00<15:52, 6.61s/step, loss=1.317, acc=0.693, acc_len=1.66, thru=19.6, I=21.4, T=37.6, wait=2.8s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:24,382] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:41:24,604] loop.py:380 INFO TIMING step=17875: step=3.544s data=1.470s compute=3.490s [fwd=2.444s bwd=1.031s opt=0.003s] dispatch=6.830s [2026-10-20 03:41:24,604] loop.py:45 INFO TRAIN_STEP step=17875 loss=1.379006 Training: 99%|█████████▉| 17875/18018 [38:11:10<18:27, 7.74s/step, loss=1.379, acc=0.690, acc_len=1.65, thru=13.1, I=13.0, T=36.1, wait=6.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:24,885] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:41:34,466] loop.py:45 INFO TRAIN_STEP step=17876 loss=1.350537 Training: 99%|█████████▉| 17876/18018 [38:11:20<19:49, 8.38s/step, loss=1.351, acc=0.691, acc_len=1.65, thru=12.9, I=13.9, T=35.3, wait=6.2s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:41:34,462] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17875: forward=2546.3ms backward=1040.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:43,039] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:41:43,458] loop.py:45 INFO TRAIN_STEP step=17877 loss=1.212221 Training: 99%|█████████▉| 17877/18018 [38:11:29<20:07, 8.56s/step, loss=1.212, acc=0.713, acc_len=1.73, thru=14.0, I=14.1, T=33.0, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:43,556] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:49,792] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:41:49,889] loop.py:45 INFO TRAIN_STEP step=17878 loss=1.297487 Training: 99%|█████████▉| 17878/18018 [38:11:35<18:29, 7.92s/step, loss=1.297, acc=0.702, acc_len=1.69, thru=19.6, I=19.6, T=37.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:41:50,293] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:41:58,274] loop.py:45 INFO TRAIN_STEP step=17879 loss=1.197297 Training: 99%|█████████▉| 17879/18018 [38:11:44<18:40, 8.06s/step, loss=1.197, acc=0.717, acc_len=1.74, thru=15.2, I=15.4, T=34.5, wait=4.7s, pool=49, epoch=2/2][2026-10-20 03:42:05,543] loop.py:380 INFO TIMING step=17880: step=3.431s data=1.439s compute=3.393s [fwd=2.443s bwd=0.936s opt=0.003s] dispatch=3.831s [2026-10-20 03:42:05,543] loop.py:45 INFO TRAIN_STEP step=17880 loss=1.243510 Training: 99%|█████████▉| 17880/18018 [38:11:51<17:59, 7.82s/step, loss=1.244, acc=0.709, acc_len=1.72, thru=17.9, I=16.7, T=37.3, wait=3.8s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:42:12,615] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:42:13,644] loop.py:45 INFO TRAIN_STEP step=17881 loss=1.223615 Training: 99%|█████████▉| 17881/18018 [38:11:59<18:03, 7.91s/step, loss=1.224, acc=0.717, acc_len=1.74, thru=17.8, I=19.3, T=35.4, wait=4.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:42:13,640] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17880: forward=2580.5ms backward=997.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:42:14,125] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:42:19,944] loop.py:45 INFO TRAIN_STEP step=17882 loss=1.229808 Training: 99%|█████████▉| 17882/18018 [38:12:06<16:49, 7.42s/step, loss=1.230, acc=0.709, acc_len=1.71, thru=19.3, I=24.8, T=37.3, wait=2.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:42:29,846] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:42:32,179] loop.py:45 INFO TRAIN_STEP step=17883 loss=1.331003 Training: 99%|█████████▉| 17883/18018 [38:12:18<19:57, 8.87s/step, loss=1.331, acc=0.694, acc_len=1.66, thru=13.3, I=12.3, T=32.1, wait=8.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:42:32,365] inference_manager.py:408 INFO Pool capacity available after 2.5s, pool_size=1/64 [2026-10-20 03:42:39,773] loop.py:45 INFO TRAIN_STEP step=17884 loss=1.202194 Training: 99%|█████████▉| 17884/18018 [38:12:25<18:57, 8.49s/step, loss=1.202, acc=0.720, acc_len=1.74, thru=9.7, I=16.7, T=36.8, wait=4.1s, pool=29, epoch=2/2] (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:42:47,078] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:42:48,255] loop.py:380 INFO TIMING step=17885: step=3.683s data=1.432s compute=3.657s [fwd=2.553s bwd=1.090s opt=0.003s] dispatch=4.786s [2026-10-20 03:42:48,256] loop.py:45 INFO TRAIN_STEP step=17885 loss=1.350449 Training: 99%|█████████▉| 17885/18018 [38:12:34<18:48, 8.48s/step, loss=1.350, acc=0.692, acc_len=1.65, thru=19.1, I=17.0, T=34.8, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:42:48,588] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:42:53,931] loop.py:45 INFO TRAIN_STEP step=17886 loss=1.246183 (TrainerActor pid=9310) [2026-10-20 03:42:53,926] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17885: forward=2302.1ms backward=884.4ms Training: 99%|█████████▉| 17886/18018 [38:12:40<16:48, 7.64s/step, loss=1.246, acc=0.707, acc_len=1.70, thru=20.5, I=25.5, T=39.6, wait=2.4s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:02,134] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:43:02,853] loop.py:45 INFO TRAIN_STEP step=17887 loss=1.285949 Training: 99%|█████████▉| 17887/18018 [38:12:48<17:31, 8.03s/step, loss=1.286, acc=0.702, acc_len=1.68, thru=15.0, I=15.6, T=33.9, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:03,138] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:08,329] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:43:08,704] loop.py:45 INFO TRAIN_STEP step=17888 loss=1.250185 Training: 99%|█████████▉| 17888/18018 [38:12:54<15:58, 7.37s/step, loss=1.250, acc=0.706, acc_len=1.70, thru=20.8, I=23.4, T=38.8, wait=2.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:08,843] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=9/64 [2026-10-20 03:43:17,875] loop.py:45 INFO TRAIN_STEP step=17889 loss=1.182454 Training: 99%|█████████▉| 17889/18018 [38:13:03<17:00, 7.91s/step, loss=1.182, acc=0.723, acc_len=1.76, thru=14.1, I=15.2, T=33.7, wait=5.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:17,802] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:18,304] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:24,153] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:43:25,190] loop.py:380 INFO TIMING step=17890: step=3.896s data=1.770s compute=3.849s [fwd=2.691s bwd=1.144s opt=0.003s] dispatch=3.413s [2026-10-20 03:43:25,191] loop.py:45 INFO TRAIN_STEP step=17890 loss=1.257340 Training: 99%|█████████▉| 17890/18018 [38:13:11<16:29, 7.73s/step, loss=1.257, acc=0.706, acc_len=1.70, thru=18.8, I=19.5, T=32.9, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:25,660] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:30,601] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:43:31,311] loop.py:45 INFO TRAIN_STEP step=17891 loss=1.305417 Training: 99%|█████████▉| 17891/18018 [38:13:17<15:20, 7.25s/step, loss=1.305, acc=0.698, acc_len=1.67, thru=19.1, I=25.0, T=35.6, wait=2.5s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:43:31,301] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17890: forward=2553.1ms backward=997.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:31,607] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:38,901] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:43:40,245] loop.py:45 INFO TRAIN_STEP step=17892 loss=1.202613 Training: 99%|█████████▉| 17892/18018 [38:13:26<16:17, 7.76s/step, loss=1.203, acc=0.718, acc_len=1.74, thru=13.7, I=16.6, T=34.0, wait=5.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:40,413] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:46,739] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:43:47,467] loop.py:45 INFO TRAIN_STEP step=17893 loss=1.311535 Training: 99%|█████████▉| 17893/18018 [38:13:33<15:49, 7.60s/step, loss=1.312, acc=0.695, acc_len=1.66, thru=16.9, I=21.7, T=35.8, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:43:47,743] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:43:53,867] loop.py:45 INFO TRAIN_STEP step=17894 loss=1.324217 Training: 99%|█████████▉| 17894/18018 [38:13:39<14:57, 7.24s/step, loss=1.324, acc=0.696, acc_len=1.66, thru=18.0, I=17.9, T=36.3, wait=2.9s, pool=49, epoch=2/2][2026-10-20 03:44:01,379] loop.py:380 INFO TIMING step=17895: step=3.485s data=1.358s compute=3.448s [fwd=2.451s bwd=0.982s opt=0.003s] dispatch=4.020s [2026-10-20 03:44:01,379] loop.py:45 INFO TRAIN_STEP step=17895 loss=1.322123 Training: 99%|█████████▉| 17895/18018 [38:13:47<15:00, 7.32s/step, loss=1.322, acc=0.697, acc_len=1.67, thru=15.7, I=15.3, T=36.7, wait=4.0s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:44:10,157] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:44:10,880] loop.py:45 INFO TRAIN_STEP step=17896 loss=1.412037 Training: 99%|█████████▉| 17896/18018 [38:13:56<16:12, 7.97s/step, loss=1.412, acc=0.682, acc_len=1.62, thru=17.1, I=16.7, T=33.7, wait=5.7s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:44:10,875] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17895: forward=2656.7ms backward=1093.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:44:11,162] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:44:17,028] loop.py:45 INFO TRAIN_STEP step=17897 loss=1.201065 Training: 99%|█████████▉| 17897/18018 [38:14:03<14:58, 7.43s/step, loss=1.201, acc=0.715, acc_len=1.73, thru=19.2, I=20.5, T=38.6, wait=2.8s, pool=61, epoch=2/2][2026-10-20 03:44:24,264] loop.py:45 INFO TRAIN_STEP step=17898 loss=1.199317 Training: 99%|█████████▉| 17898/18018 [38:14:10<14:44, 7.37s/step, loss=1.199, acc=0.719, acc_len=1.74, thru=16.4, I=17.3, T=34.8, wait=3.5s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:44:29,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:44:30,482] loop.py:45 INFO TRAIN_STEP step=17899 loss=1.239516 Training: 99%|█████████▉| 17899/18018 [38:14:16<13:55, 7.02s/step, loss=1.240, acc=0.713, acc_len=1.72, thru=19.7, I=20.1, T=36.5, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:44:30,630] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:44:36,264] loop.py:380 INFO TIMING step=17900: step=3.336s data=1.371s compute=3.305s [fwd=2.358s bwd=0.933s opt=0.003s] dispatch=2.436s [2026-10-20 03:44:36,264] loop.py:45 INFO TRAIN_STEP step=17900 loss=1.301044 Training: 99%|█████████▉| 17900/18018 [38:14:22<13:04, 6.65s/step, loss=1.301, acc=0.700, acc_len=1.68, thru=20.9, I=28.0, T=38.4, wait=2.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:44:41,799] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:44:42,750] loop.py:45 INFO TRAIN_STEP step=17901 loss=1.183609 Training: 99%|█████████▉| 17901/18018 [38:14:28<12:52, 6.60s/step, loss=1.184, acc=0.718, acc_len=1.74, thru=19.9, I=22.0, T=38.6, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:44:42,745] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17900: forward=2347.7ms backward=922.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:44:42,821] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=41/64 [2026-10-20 03:44:49,054] loop.py:45 INFO TRAIN_STEP step=17902 loss=1.353903 Training: 99%|█████████▉| 17902/18018 [38:14:35<12:35, 6.51s/step, loss=1.354, acc=0.688, acc_len=1.64, thru=17.5, I=23.7, T=37.8, wait=2.9s, pool=29, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:44:56,790] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:44:57,491] loop.py:45 INFO TRAIN_STEP step=17903 loss=1.203622 Training: 99%|█████████▉| 17903/18018 [38:14:43<13:35, 7.09s/step, loss=1.204, acc=0.717, acc_len=1.74, thru=17.8, I=17.4, T=37.0, wait=5.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:44:57,796] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:45:04,024] loop.py:45 INFO TRAIN_STEP step=17904 loss=1.243985 Training: 99%|█████████▉| 17904/18018 [38:14:50<13:09, 6.92s/step, loss=1.244, acc=0.707, acc_len=1.71, thru=18.4, I=19.8, T=37.1, wait=3.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:45:08,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:45:09,781] loop.py:380 INFO TIMING step=17905: step=3.437s data=1.381s compute=3.413s [fwd=2.414s bwd=0.986s opt=0.003s] dispatch=2.311s [2026-10-20 03:45:09,781] loop.py:45 INFO TRAIN_STEP step=17905 loss=1.175054 Training: 99%|█████████▉| 17905/18018 [38:14:55<12:22, 6.57s/step, loss=1.175, acc=0.722, acc_len=1.75, thru=21.5, I=23.2, T=37.2, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:45:10,103] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:45:15,534] loop.py:45 INFO TRAIN_STEP step=17906 loss=1.263576 Training: 99%|█████████▉| 17906/18018 [38:15:01<11:48, 6.33s/step, loss=1.264, acc=0.709, acc_len=1.71, thru=20.1, I=24.6, T=36.0, wait=2.2s, pool=45, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:45:15,529] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17905: forward=2558.6ms backward=954.9ms [2026-10-20 03:45:24,474] loop.py:45 INFO TRAIN_STEP step=17907 loss=1.182688 Training: 99%|█████████▉| 17907/18018 [38:15:10<13:09, 7.11s/step, loss=1.183, acc=0.719, acc_len=1.74, thru=14.8, I=13.5, T=35.0, wait=5.3s, pool=53, epoch=2/2][2026-10-20 03:45:32,099] loop.py:45 INFO TRAIN_STEP step=17908 loss=1.303604 Training: 99%|█████████▉| 17908/18018 [38:15:18<13:19, 7.26s/step, loss=1.304, acc=0.699, acc_len=1.68, thru=16.0, I=16.6, T=36.4, wait=4.1s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:45:41,308] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:45:41,891] loop.py:45 INFO TRAIN_STEP step=17909 loss=1.338848 Training: 99%|█████████▉| 17909/18018 [38:15:27<14:34, 8.02s/step, loss=1.339, acc=0.694, acc_len=1.66, thru=15.6, I=15.1, T=33.7, wait=6.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:45:42,313] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:45:49,126] loop.py:380 INFO TIMING step=17910: step=3.540s data=1.530s compute=3.512s [fwd=2.545s bwd=0.952s opt=0.003s] dispatch=3.686s [2026-10-20 03:45:49,127] loop.py:45 INFO TRAIN_STEP step=17910 loss=1.173930 Training: 99%|█████████▉| 17910/18018 [38:15:35<14:00, 7.79s/step, loss=1.174, acc=0.719, acc_len=1.74, thru=15.3, I=16.7, T=36.2, wait=3.7s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:45:57,183] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:45:57,230] loop.py:45 INFO TRAIN_STEP step=17911 loss=1.245272 Training: 99%|█████████▉| 17911/18018 [38:15:43<14:03, 7.88s/step, loss=1.245, acc=0.709, acc_len=1.71, thru=17.9, I=16.5, T=34.5, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:45:57,226] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17910: forward=2625.3ms backward=1042.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:45:57,687] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:46:02,815] loop.py:45 INFO TRAIN_STEP step=17912 loss=1.280923 Training: 99%|█████████▉| 17912/18018 [38:15:48<12:42, 7.19s/step, loss=1.281, acc=0.704, acc_len=1.69, thru=19.3, I=22.6, T=36.9, wait=2.1s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:11,960] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:46:13,750] loop.py:45 INFO TRAIN_STEP step=17913 loss=1.341131 Training: 99%|█████████▉| 17913/18018 [38:15:59<14:33, 8.32s/step, loss=1.341, acc=0.693, acc_len=1.65, thru=14.5, I=14.5, T=32.1, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:13,971] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:18,913] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:46:20,422] loop.py:45 INFO TRAIN_STEP step=17914 loss=1.300159 Training: 99%|█████████▉| 17914/18018 [38:16:06<13:33, 7.82s/step, loss=1.300, acc=0.697, acc_len=1.67, thru=17.3, I=26.7, T=36.8, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:20,925] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:26,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:46:26,260] loop.py:380 INFO TIMING step=17915: step=3.216s data=1.271s compute=3.186s [fwd=2.288s bwd=0.883s opt=0.003s] dispatch=2.613s [2026-10-20 03:46:26,260] loop.py:45 INFO TRAIN_STEP step=17915 loss=1.174081 Training: 99%|█████████▉| 17915/18018 [38:16:12<12:24, 7.23s/step, loss=1.174, acc=0.721, acc_len=1.75, thru=21.3, I=26.4, T=39.8, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:26,725] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:31,397] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:46:33,056] loop.py:45 INFO TRAIN_STEP step=17916 loss=1.199160 (TrainerActor pid=9310) [2026-10-20 03:46:33,052] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17915: forward=2788.6ms backward=1038.8ms Training: 99%|█████████▉| 17916/18018 [38:16:19<12:03, 7.10s/step, loss=1.199, acc=0.719, acc_len=1.74, thru=19.8, I=25.0, T=33.0, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:33,410] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 03:46:39,389] loop.py:45 INFO TRAIN_STEP step=17917 loss=1.243877 Training: 99%|█████████▉| 17917/18018 [38:16:25<11:33, 6.87s/step, loss=1.244, acc=0.709, acc_len=1.71, thru=15.7, I=22.4, T=35.7, wait=2.7s, pool=33, epoch=2/2][2026-10-20 03:46:53,198] loop.py:45 INFO TRAIN_STEP step=17918 loss=1.280883 Training: 99%|█████████▉| 17918/18018 [38:16:39<14:55, 8.95s/step, loss=1.281, acc=0.703, acc_len=1.69, thru=11.0, I=11.4, T=29.3, wait=9.4s, pool=53, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:57,656] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:46:59,121] loop.py:45 INFO TRAIN_STEP step=17919 loss=1.321723 Training: 99%|█████████▉| 17919/18018 [38:16:45<13:16, 8.04s/step, loss=1.322, acc=0.694, acc_len=1.66, thru=19.3, I=21.4, T=36.7, wait=2.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:46:59,177] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=49/64 [2026-10-20 03:47:05,717] loop.py:380 INFO TIMING step=17920: step=3.485s data=1.503s compute=3.454s [fwd=2.488s bwd=0.951s opt=0.003s] dispatch=3.101s [2026-10-20 03:47:05,718] loop.py:45 INFO TRAIN_STEP step=17920 loss=1.219382 Training: 99%|█████████▉| 17920/18018 [38:16:51<12:25, 7.61s/step, loss=1.219, acc=0.713, acc_len=1.72, thru=18.3, I=22.3, T=36.7, wait=3.1s, pool=57, epoch=2/2][2026-10-20 03:47:12,596] loop.py:45 INFO TRAIN_STEP step=17921 loss=1.256232 Training: 99%|█████████▉| 17921/18018 [38:16:58<11:56, 7.39s/step, loss=1.256, acc=0.706, acc_len=1.69, thru=17.7, I=17.2, T=35.9, wait=3.3s, pool=53, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:47:12,591] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17920: forward=2539.9ms backward=966.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:19,211] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:47:19,421] loop.py:45 INFO TRAIN_STEP step=17922 loss=1.216681 Training: 99%|█████████▉| 17922/18018 [38:17:05<11:33, 7.22s/step, loss=1.217, acc=0.717, acc_len=1.74, thru=20.2, I=19.5, T=34.3, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:19,712] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:26,758] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:47:27,886] loop.py:45 INFO TRAIN_STEP step=17923 loss=1.300564 Training: 99%|█████████▉| 17923/18018 [38:17:13<12:01, 7.59s/step, loss=1.301, acc=0.699, acc_len=1.68, thru=15.8, I=17.5, T=34.9, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:28,269] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:47:34,143] loop.py:45 INFO TRAIN_STEP step=17924 loss=1.280967 Training: 99%|█████████▉| 17924/18018 [38:17:20<11:16, 7.19s/step, loss=1.281, acc=0.705, acc_len=1.69, thru=19.7, I=21.1, T=38.8, wait=3.0s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:40,214] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:47:42,026] loop.py:380 INFO TIMING step=17925: step=3.948s data=1.790s compute=3.914s [fwd=2.776s bwd=1.123s opt=0.003s] dispatch=3.925s [2026-10-20 03:47:42,026] loop.py:45 INFO TRAIN_STEP step=17925 loss=1.267558 Training: 99%|█████████▉| 17925/18018 [38:17:28<11:28, 7.40s/step, loss=1.268, acc=0.705, acc_len=1.69, thru=18.5, I=20.9, T=32.4, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:42,228] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:47,178] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:47:48,221] loop.py:45 INFO TRAIN_STEP step=17926 loss=1.267116 Training: 99%|█████████▉| 17926/18018 [38:17:34<10:47, 7.04s/step, loss=1.267, acc=0.706, acc_len=1.70, thru=19.3, I=27.0, T=39.8, wait=3.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:47:48,216] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17925: forward=2291.7ms backward=881.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:48,688] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:52,182] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:47:53,597] loop.py:45 INFO TRAIN_STEP step=17927 loss=1.159353 Training: 99%|█████████▉| 17927/18018 [38:17:39<09:55, 6.54s/step, loss=1.159, acc=0.721, acc_len=1.76, thru=22.1, I=32.2, T=38.9, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:53,718] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=41/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:58,823] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:47:59,808] loop.py:45 INFO TRAIN_STEP step=17928 loss=1.185070 Training: 100%|█████████▉| 17928/18018 [38:17:45<09:39, 6.44s/step, loss=1.185, acc=0.720, acc_len=1.75, thru=21.7, I=27.9, T=37.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:47:59,841] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:48:05,687] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:48:05,732] loop.py:45 INFO TRAIN_STEP step=17929 loss=1.247495 Training: 100%|█████████▉| 17929/18018 [38:17:51<09:19, 6.29s/step, loss=1.247, acc=0.710, acc_len=1.71, thru=20.9, I=25.0, T=34.9, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:48:06,189] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:48:11,830] loop.py:380 INFO TIMING step=17930: step=3.522s data=1.475s compute=3.495s [fwd=2.485s bwd=0.996s opt=0.003s] dispatch=2.568s [2026-10-20 03:48:11,830] loop.py:45 INFO TRAIN_STEP step=17930 loss=1.318663 Training: 100%|█████████▉| 17930/18018 [38:17:57<09:08, 6.23s/step, loss=1.319, acc=0.692, acc_len=1.65, thru=18.7, I=22.0, T=36.3, wait=2.6s, pool=49, epoch=2/2][2026-10-20 03:48:20,447] loop.py:45 INFO TRAIN_STEP step=17931 loss=1.293648 Training: 100%|█████████▉| 17931/18018 [38:18:06<10:04, 6.95s/step, loss=1.294, acc=0.703, acc_len=1.69, thru=14.8, I=13.8, T=34.6, wait=4.9s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:48:20,444] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17930: forward=2500.1ms backward=1136.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:48:26,497] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:48:27,528] loop.py:45 INFO TRAIN_STEP step=17932 loss=1.204113 Training: 100%|█████████▉| 17932/18018 [38:18:13<10:00, 6.99s/step, loss=1.204, acc=0.717, acc_len=1.74, thru=19.7, I=18.6, T=33.5, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:48:28,006] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:48:34,596] loop.py:45 INFO TRAIN_STEP step=17933 loss=1.239208 Training: 100%|█████████▉| 17933/18018 [38:18:20<09:55, 7.01s/step, loss=1.239, acc=0.708, acc_len=1.71, thru=16.4, I=22.5, T=34.0, wait=3.3s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:48:42,399] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:48:43,424] loop.py:45 INFO TRAIN_STEP step=17934 loss=1.199930 Training: 100%|█████████▉| 17934/18018 [38:18:29<10:34, 7.56s/step, loss=1.200, acc=0.717, acc_len=1.74, thru=14.8, I=15.4, T=29.8, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:48:43,913] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:48:49,622] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:48:50,432] loop.py:380 INFO TIMING step=17935: step=4.260s data=2.085s compute=4.233s [fwd=3.100s bwd=1.118s opt=0.003s] dispatch=2.740s [2026-10-20 03:48:50,432] loop.py:45 INFO TRAIN_STEP step=17935 loss=1.261300 Training: 100%|█████████▉| 17935/18018 [38:18:36<10:13, 7.39s/step, loss=1.261, acc=0.708, acc_len=1.71, thru=16.8, I=21.5, T=30.0, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:48:50,633] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 [2026-10-20 03:48:59,638] loop.py:45 INFO TRAIN_STEP step=17936 loss=1.307936 Training: 100%|█████████▉| 17936/18018 [38:18:45<10:50, 7.94s/step, loss=1.308, acc=0.700, acc_len=1.67, thru=13.3, I=13.2, T=36.2, wait=5.7s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:48:59,634] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17935: forward=2520.0ms backward=973.7ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:49:06,026] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:49:07,560] loop.py:45 INFO TRAIN_STEP step=17937 loss=1.330725 Training: 100%|█████████▉| 17937/18018 [38:18:53<10:42, 7.93s/step, loss=1.331, acc=0.696, acc_len=1.67, thru=16.3, I=18.7, T=31.6, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:49:08,039] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 03:49:14,264] loop.py:45 INFO TRAIN_STEP step=17938 loss=1.304611 Training: 100%|█████████▉| 17938/18018 [38:19:00<10:05, 7.56s/step, loss=1.305, acc=0.704, acc_len=1.69, thru=17.7, I=19.5, T=36.1, wait=3.2s, pool=57, epoch=2/2][2026-10-20 03:49:20,955] loop.py:45 INFO TRAIN_STEP step=17939 loss=1.294705 Training: 100%|█████████▉| 17939/18018 [38:19:07<09:36, 7.30s/step, loss=1.295, acc=0.702, acc_len=1.68, thru=16.6, I=16.5, T=36.3, wait=3.2s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:49:27,816] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:49:28,447] loop.py:380 INFO TIMING step=17940: step=3.564s data=1.527s compute=3.537s [fwd=2.540s bwd=0.983s opt=0.003s] dispatch=3.917s [2026-10-20 03:49:28,447] loop.py:45 INFO TRAIN_STEP step=17940 loss=1.238185 Training: 100%|█████████▉| 17940/18018 [38:19:14<09:33, 7.36s/step, loss=1.238, acc=0.710, acc_len=1.71, thru=18.6, I=18.4, T=35.9, wait=3.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:49:28,820] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:49:33,564] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:49:34,997] loop.py:45 INFO TRAIN_STEP step=17941 loss=1.246109 Training: 100%|█████████▉| 17941/18018 [38:19:21<09:07, 7.12s/step, loss=1.246, acc=0.709, acc_len=1.71, thru=19.2, I=21.6, T=32.1, wait=2.6s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:49:34,993] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17940: forward=2945.2ms backward=995.0ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:49:35,084] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=33/64 [2026-10-20 03:49:42,109] loop.py:45 INFO TRAIN_STEP step=17942 loss=1.224650 Training: 100%|█████████▉| 17942/18018 [38:19:28<09:00, 7.11s/step, loss=1.225, acc=0.712, acc_len=1.72, thru=14.5, I=17.1, T=37.2, wait=3.7s, pool=49, epoch=2/2][2026-10-20 03:49:52,692] loop.py:45 INFO TRAIN_STEP step=17943 loss=1.370548 Training: 100%|█████████▉| 17943/18018 [38:19:38<10:11, 8.16s/step, loss=1.371, acc=0.686, acc_len=1.63, thru=12.6, I=11.5, T=34.0, wait=6.8s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:49:57,547] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:49:59,249] loop.py:45 INFO TRAIN_STEP step=17944 loss=1.373718 Training: 100%|█████████▉| 17944/18018 [38:19:45<09:28, 7.68s/step, loss=1.374, acc=0.686, acc_len=1.63, thru=19.2, I=22.4, T=37.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:49:59,560] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 03:50:05,453] loop.py:380 INFO TIMING step=17945: step=3.478s data=1.520s compute=3.445s [fwd=2.441s bwd=0.990s opt=0.003s] dispatch=2.717s [2026-10-20 03:50:05,453] loop.py:45 INFO TRAIN_STEP step=17945 loss=1.237308 Training: 100%|█████████▉| 17945/18018 [38:19:51<08:48, 7.23s/step, loss=1.237, acc=0.710, acc_len=1.72, thru=19.6, I=27.3, T=36.8, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:12,561] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:50:13,158] loop.py:45 INFO TRAIN_STEP step=17946 loss=1.296851 Training: 100%|█████████▉| 17946/18018 [38:19:59<08:51, 7.38s/step, loss=1.297, acc=0.701, acc_len=1.68, thru=17.1, I=17.7, T=35.2, wait=4.1s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:50:13,154] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17945: forward=2587.6ms backward=989.5ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:13,566] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:17,504] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:50:19,454] loop.py:45 INFO TRAIN_STEP step=17947 loss=1.303638 Training: 100%|█████████▉| 17947/18018 [38:20:05<08:20, 7.05s/step, loss=1.304, acc=0.698, acc_len=1.67, thru=19.6, I=24.4, T=31.7, wait=2.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:19,516] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=49/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:24,288] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:50:25,613] loop.py:45 INFO TRAIN_STEP step=17948 loss=1.183487 Training: 100%|█████████▉| 17948/18018 [38:20:11<07:54, 6.78s/step, loss=1.183, acc=0.719, acc_len=1.75, thru=17.9, I=25.8, T=36.2, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:25,796] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:32,667] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:50:34,314] loop.py:45 INFO TRAIN_STEP step=17949 loss=1.240136 Training: 100%|█████████▉| 17949/18018 [38:20:20<08:27, 7.36s/step, loss=1.240, acc=0.710, acc_len=1.71, thru=15.0, I=18.3, T=32.5, wait=4.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:34,679] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:40,523] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:50:41,120] loop.py:380 INFO TIMING step=17950: step=3.422s data=1.441s compute=3.392s [fwd=2.383s bwd=0.995s opt=0.003s] dispatch=3.375s [2026-10-20 03:50:41,120] loop.py:45 INFO TRAIN_STEP step=17950 loss=1.168657 Training: 100%|█████████▉| 17950/18018 [38:20:27<08:09, 7.19s/step, loss=1.169, acc=0.720, acc_len=1.75, thru=17.7, I=23.3, T=37.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:41,529] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:50:47,467] loop.py:45 INFO TRAIN_STEP step=17951 loss=1.323446 Training: 100%|█████████▉| 17951/18018 [38:20:33<07:44, 6.94s/step, loss=1.323, acc=0.696, acc_len=1.67, thru=19.8, I=22.1, T=36.7, wait=2.8s, pool=61, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:50:47,463] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17950: forward=2440.7ms backward=1009.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:53,194] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:50:54,714] loop.py:45 INFO TRAIN_STEP step=17952 loss=1.203696 Training: 100%|█████████▉| 17952/18018 [38:20:40<07:44, 7.03s/step, loss=1.204, acc=0.717, acc_len=1.74, thru=17.9, I=20.6, T=35.4, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:50:55,208] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:02,727] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:51:02,895] loop.py:45 INFO TRAIN_STEP step=17953 loss=1.179965 Training: 100%|█████████▉| 17953/18018 [38:20:48<07:59, 7.38s/step, loss=1.180, acc=0.719, acc_len=1.75, thru=13.8, I=17.2, T=36.2, wait=4.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:03,229] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:09,120] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:51:09,743] loop.py:45 INFO TRAIN_STEP step=17954 loss=1.295094 Training: 100%|█████████▉| 17954/18018 [38:20:55<07:41, 7.22s/step, loss=1.295, acc=0.702, acc_len=1.69, thru=19.6, I=20.3, T=38.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:10,126] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:51:17,952] loop.py:380 INFO TIMING step=17955: step=3.388s data=1.401s compute=3.357s [fwd=2.395s bwd=0.948s opt=0.003s] dispatch=4.812s [2026-10-20 03:51:17,952] loop.py:45 INFO TRAIN_STEP step=17955 loss=1.208712 Training: 100%|█████████▉| 17955/18018 [38:21:04<07:53, 7.51s/step, loss=1.209, acc=0.713, acc_len=1.73, thru=15.4, I=16.7, T=37.8, wait=4.8s, pool=53, epoch=2/2][2026-10-20 03:51:28,774] loop.py:45 INFO TRAIN_STEP step=17956 loss=1.163031 Training: 100%|█████████▉| 17956/18018 [38:21:14<08:47, 8.51s/step, loss=1.163, acc=0.730, acc_len=1.79, thru=13.2, I=13.1, T=32.5, wait=6.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:28,717] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 03:51:28,770] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17955: forward=2693.9ms backward=1181.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:29,220] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:34,613] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:51:35,142] loop.py:45 INFO TRAIN_STEP step=17957 loss=1.263259 Training: 100%|█████████▉| 17957/18018 [38:21:21<07:59, 7.87s/step, loss=1.263, acc=0.706, acc_len=1.70, thru=18.9, I=19.8, T=37.6, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:35,618] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:51:41,285] loop.py:45 INFO TRAIN_STEP step=17958 loss=1.239768 Training: 100%|█████████▉| 17958/18018 [38:21:27<07:20, 7.35s/step, loss=1.240, acc=0.710, acc_len=1.72, thru=20.1, I=21.8, T=39.0, wait=2.9s, pool=61, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:47,363] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:51:47,790] loop.py:45 INFO TRAIN_STEP step=17959 loss=1.275521 Training: 100%|█████████▉| 17959/18018 [38:21:33<06:58, 7.10s/step, loss=1.276, acc=0.704, acc_len=1.69, thru=20.2, I=20.8, T=39.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:51:47,870] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=33/64 [2026-10-20 03:51:53,932] loop.py:380 INFO TIMING step=17960: step=3.476s data=1.432s compute=3.439s [fwd=2.456s bwd=0.969s opt=0.003s] dispatch=2.658s [2026-10-20 03:51:53,932] loop.py:45 INFO TRAIN_STEP step=17960 loss=1.247309 Training: 100%|█████████▉| 17960/18018 [38:21:40<06:34, 6.81s/step, loss=1.247, acc=0.712, acc_len=1.72, thru=20.2, I=21.5, T=36.8, wait=2.7s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:00,271] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:52:01,819] loop.py:45 INFO TRAIN_STEP step=17961 loss=1.155974 Training: 100%|█████████▉| 17961/18018 [38:21:47<06:46, 7.13s/step, loss=1.156, acc=0.729, acc_len=1.78, thru=17.1, I=19.3, T=36.8, wait=4.4s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:52:01,814] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17960: forward=2459.5ms backward=967.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:02,281] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 03:52:09,514] loop.py:45 INFO TRAIN_STEP step=17962 loss=1.212264 Training: 100%|█████████▉| 17962/18018 [38:21:55<06:48, 7.30s/step, loss=1.212, acc=0.715, acc_len=1.73, thru=13.2, I=17.8, T=37.1, wait=4.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:18,017] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:52:19,549] loop.py:45 INFO TRAIN_STEP step=17963 loss=1.221373 Training: 100%|█████████▉| 17963/18018 [38:22:05<07:26, 8.12s/step, loss=1.221, acc=0.717, acc_len=1.74, thru=15.2, I=16.3, T=32.9, wait=6.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:20,032] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 [2026-10-20 03:52:26,059] loop.py:45 INFO TRAIN_STEP step=17964 loss=1.294953 Training: 100%|█████████▉| 17964/18018 [38:22:12<06:52, 7.64s/step, loss=1.295, acc=0.701, acc_len=1.68, thru=15.7, I=19.3, T=34.8, wait=2.8s, pool=41, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:32,405] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:52:33,331] loop.py:380 INFO TIMING step=17965: step=3.902s data=1.690s compute=3.868s [fwd=2.846s bwd=1.007s opt=0.003s] dispatch=3.361s [2026-10-20 03:52:33,331] loop.py:45 INFO TRAIN_STEP step=17965 loss=1.234518 Training: 100%|█████████▉| 17965/18018 [38:22:19<06:38, 7.53s/step, loss=1.235, acc=0.713, acc_len=1.72, thru=18.5, I=19.6, T=32.8, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:33,421] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=33/64 [2026-10-20 03:52:40,034] loop.py:45 INFO TRAIN_STEP step=17966 loss=1.137358 Training: 100%|█████████▉| 17966/18018 [38:22:26<06:18, 7.28s/step, loss=1.137, acc=0.727, acc_len=1.77, thru=16.6, I=17.9, T=37.5, wait=3.3s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:52:40,029] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17965: forward=2446.5ms backward=922.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:48,358] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:52:49,614] loop.py:45 INFO TRAIN_STEP step=17967 loss=1.382076 Training: 100%|█████████▉| 17967/18018 [38:22:35<06:46, 7.97s/step, loss=1.382, acc=0.682, acc_len=1.62, thru=15.5, I=16.2, T=28.8, wait=5.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:49,867] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:55,721] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:52:56,278] loop.py:45 INFO TRAIN_STEP step=17968 loss=1.119094 Training: 100%|█████████▉| 17968/18018 [38:22:42<06:18, 7.58s/step, loss=1.119, acc=0.731, acc_len=1.79, thru=18.4, I=20.6, T=36.0, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:52:56,726] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:53:02,631] loop.py:45 INFO TRAIN_STEP step=17969 loss=1.236327 Training: 100%|█████████▉| 17969/18018 [38:22:48<05:53, 7.21s/step, loss=1.236, acc=0.710, acc_len=1.71, thru=18.9, I=21.7, T=38.7, wait=3.0s, pool=53, epoch=2/2][2026-10-20 03:53:10,346] loop.py:380 INFO TIMING step=17970: step=3.627s data=1.535s compute=3.595s [fwd=2.565s bwd=1.016s opt=0.003s] dispatch=4.078s [2026-10-20 03:53:10,347] loop.py:45 INFO TRAIN_STEP step=17970 loss=1.304337 Training: 100%|█████████▉| 17970/18018 [38:22:56<05:53, 7.36s/step, loss=1.304, acc=0.696, acc_len=1.66, thru=16.1, I=16.3, T=35.3, wait=4.1s, pool=49, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:15,450] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:53:16,688] loop.py:45 INFO TRAIN_STEP step=17971 loss=1.215828 Training: 100%|█████████▉| 17971/18018 [38:23:02<05:31, 7.06s/step, loss=1.216, acc=0.717, acc_len=1.74, thru=20.3, I=21.8, T=36.7, wait=2.8s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:53:16,684] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17970: forward=2447.2ms backward=992.2ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:16,960] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:24,223] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:53:25,375] loop.py:45 INFO TRAIN_STEP step=17972 loss=1.234228 Training: 100%|█████████▉| 17972/18018 [38:23:11<05:47, 7.55s/step, loss=1.234, acc=0.711, acc_len=1.72, thru=13.7, I=17.8, T=33.6, wait=4.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:25,733] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:30,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:53:31,421] loop.py:45 INFO TRAIN_STEP step=17973 loss=1.229586 Training: 100%|█████████▉| 17973/18018 [38:23:17<05:19, 7.10s/step, loss=1.230, acc=0.712, acc_len=1.72, thru=19.8, I=22.8, T=38.9, wait=2.7s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:31,892] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:38,453] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:53:39,472] loop.py:45 INFO TRAIN_STEP step=17974 loss=1.334647 Training: 100%|█████████▉| 17974/18018 [38:23:25<05:24, 7.38s/step, loss=1.335, acc=0.696, acc_len=1.67, thru=15.8, I=19.0, T=34.6, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:39,961] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:44,231] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:53:45,806] loop.py:380 INFO TIMING step=17975: step=3.572s data=1.543s compute=3.540s [fwd=2.556s bwd=0.971s opt=0.003s] dispatch=2.753s [2026-10-20 03:53:45,806] loop.py:45 INFO TRAIN_STEP step=17975 loss=1.230248 Training: 100%|█████████▉| 17975/18018 [38:23:31<05:03, 7.07s/step, loss=1.230, acc=0.710, acc_len=1.72, thru=19.2, I=27.2, T=35.8, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:46,245] inference_manager.py:408 INFO Pool capacity available after 2.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:52,629] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:53:52,921] loop.py:45 INFO TRAIN_STEP step=17976 loss=1.290262 (TrainerActor pid=9310) [2026-10-20 03:53:52,917] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17975: forward=2507.6ms backward=1056.8ms Training: 100%|█████████▉| 17976/18018 [38:23:39<04:57, 7.08s/step, loss=1.290, acc=0.701, acc_len=1.68, thru=17.6, I=22.7, T=35.4, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:53,131] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:53:59,259] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:54:00,484] loop.py:45 INFO TRAIN_STEP step=17977 loss=1.341421 Training: 100%|█████████▉| 17977/18018 [38:23:46<04:56, 7.23s/step, loss=1.341, acc=0.690, acc_len=1.65, thru=17.5, I=19.4, T=30.4, wait=3.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:00,769] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:05,961] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:54:06,219] loop.py:45 INFO TRAIN_STEP step=17978 loss=1.317451 Training: 100%|█████████▉| 17978/18018 [38:23:52<04:31, 6.78s/step, loss=1.317, acc=0.698, acc_len=1.68, thru=19.6, I=22.6, T=36.3, wait=2.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:06,463] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:11,353] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:54:12,237] loop.py:45 INFO TRAIN_STEP step=17979 loss=1.242656 Training: 100%|█████████▉| 17979/18018 [38:23:58<04:15, 6.55s/step, loss=1.243, acc=0.707, acc_len=1.70, thru=21.0, I=22.9, T=38.0, wait=2.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:12,362] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:18,056] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:54:18,589] loop.py:380 INFO TIMING step=17980: step=3.269s data=1.277s compute=3.237s [fwd=2.341s bwd=0.881s opt=0.003s] dispatch=3.072s [2026-10-20 03:54:18,590] loop.py:45 INFO TRAIN_STEP step=17980 loss=1.148491 Training: 100%|█████████▉| 17980/18018 [38:24:04<04:06, 6.49s/step, loss=1.148, acc=0.728, acc_len=1.77, thru=20.2, I=23.3, T=39.2, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:19,064] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:24,420] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:54:25,605] loop.py:45 INFO TRAIN_STEP step=17981 loss=1.249132 Training: 100%|█████████▉| 17981/18018 [38:24:11<04:05, 6.65s/step, loss=1.249, acc=0.705, acc_len=1.69, thru=19.4, I=22.8, T=33.5, wait=3.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:54:25,601] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17980: forward=2693.7ms backward=1086.3ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:25,929] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:54:31,305] loop.py:45 INFO TRAIN_STEP step=17982 loss=1.219863 Training: 100%|█████████▉| 17982/18018 [38:24:17<03:49, 6.36s/step, loss=1.220, acc=0.712, acc_len=1.72, thru=19.3, I=22.1, T=37.2, wait=2.2s, pool=45, epoch=2/2][2026-10-20 03:54:38,972] loop.py:45 INFO TRAIN_STEP step=17983 loss=1.232616 Training: 100%|█████████▉| 17983/18018 [38:24:25<03:56, 6.75s/step, loss=1.233, acc=0.707, acc_len=1.70, thru=17.8, I=18.4, T=34.2, wait=3.9s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:45,251] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:54:46,248] loop.py:45 INFO TRAIN_STEP step=17984 loss=1.261937 Training: 100%|█████████▉| 17984/18018 [38:24:32<03:55, 6.91s/step, loss=1.262, acc=0.711, acc_len=1.72, thru=17.7, I=19.4, T=35.0, wait=3.6s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:46,281] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=57/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:50,762] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:54:51,632] loop.py:380 INFO TIMING step=17985: step=3.376s data=1.323s compute=3.350s [fwd=2.425s bwd=0.910s opt=0.003s] dispatch=1.997s [2026-10-20 03:54:51,632] loop.py:45 INFO TRAIN_STEP step=17985 loss=1.255516 Training: 100%|█████████▉| 17985/18018 [38:24:37<03:32, 6.45s/step, loss=1.256, acc=0.706, acc_len=1.70, thru=22.0, I=27.0, T=37.9, wait=2.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:51,771] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:58,044] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:54:58,932] loop.py:45 INFO TRAIN_STEP step=17986 loss=1.430176 Training: 100%|█████████▉| 17986/18018 [38:24:45<03:34, 6.71s/step, loss=1.430, acc=0.676, acc_len=1.60, thru=18.7, I=22.5, T=38.1, wait=3.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:54:58,927] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17985: forward=2396.4ms backward=909.8ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:54:59,064] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=9/64 [2026-10-20 03:55:04,617] loop.py:45 INFO TRAIN_STEP step=17987 loss=1.258578 Training: 100%|█████████▉| 17987/18018 [38:24:50<03:18, 6.40s/step, loss=1.259, acc=0.709, acc_len=1.71, thru=19.7, I=21.4, T=39.7, wait=2.5s, pool=53, epoch=2/2][2026-10-20 03:55:13,118] loop.py:45 INFO TRAIN_STEP step=17988 loss=1.383696 Training: 100%|█████████▉| 17988/18018 [38:24:59<03:30, 7.03s/step, loss=1.384, acc=0.686, acc_len=1.63, thru=16.1, I=15.3, T=36.4, wait=5.0s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:18,166] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:55:19,316] loop.py:45 INFO TRAIN_STEP step=17989 loss=1.202674 Training: 100%|█████████▉| 17989/18018 [38:25:05<03:16, 6.78s/step, loss=1.203, acc=0.720, acc_len=1.75, thru=20.8, I=22.7, T=38.5, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:19,672] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:26,091] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:55:27,268] loop.py:380 INFO TIMING step=17990: step=3.461s data=1.422s compute=3.428s [fwd=2.509s bwd=0.904s opt=0.003s] dispatch=4.481s [2026-10-20 03:55:27,269] loop.py:45 INFO TRAIN_STEP step=17990 loss=1.214268 Training: 100%|█████████▉| 17990/18018 [38:25:13<03:19, 7.13s/step, loss=1.214, acc=0.715, acc_len=1.73, thru=16.6, I=21.7, T=37.0, wait=4.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:27,598] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:31,884] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:55:32,738] loop.py:45 INFO TRAIN_STEP step=17991 loss=1.247565 Training: 100%|█████████▉| 17991/18018 [38:25:18<02:59, 6.63s/step, loss=1.248, acc=0.708, acc_len=1.70, thru=21.8, I=28.2, T=38.8, wait=2.2s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:55:32,733] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17990: forward=2382.5ms backward=875.1ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:32,897] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:55:40,537] loop.py:45 INFO TRAIN_STEP step=17992 loss=1.332972 Training: 100%|█████████▉| 17992/18018 [38:25:26<03:01, 6.99s/step, loss=1.333, acc=0.694, acc_len=1.67, thru=14.0, I=16.2, T=35.3, wait=4.2s, pool=45, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:46,596] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:55:47,817] loop.py:45 INFO TRAIN_STEP step=17993 loss=1.305709 Training: 100%|█████████▉| 17993/18018 [38:25:33<02:56, 7.07s/step, loss=1.306, acc=0.698, acc_len=1.67, thru=17.8, I=19.6, T=33.4, wait=3.4s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:48,106] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:52,722] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:55:53,328] loop.py:45 INFO TRAIN_STEP step=17994 loss=1.259251 Training: 100%|█████████▉| 17994/18018 [38:25:39<02:38, 6.60s/step, loss=1.259, acc=0.704, acc_len=1.69, thru=21.0, I=26.0, T=37.4, wait=2.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:53,728] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:55:59,255] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:55:59,891] loop.py:380 INFO TIMING step=17995: step=3.354s data=1.318s compute=3.319s [fwd=2.354s bwd=0.951s opt=0.003s] dispatch=3.199s [2026-10-20 03:55:59,891] loop.py:45 INFO TRAIN_STEP step=17995 loss=1.311693 Training: 100%|█████████▉| 17995/18018 [38:25:45<02:31, 6.59s/step, loss=1.312, acc=0.695, acc_len=1.66, thru=19.5, I=22.9, T=38.2, wait=3.2s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:00,261] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:56:08,539] loop.py:45 INFO TRAIN_STEP step=17996 loss=1.259297 Training: 100%|█████████▉| 17996/18018 [38:25:54<02:38, 7.21s/step, loss=1.259, acc=0.707, acc_len=1.70, thru=15.6, I=16.4, T=35.2, wait=5.0s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:56:08,535] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=17995: forward=2584.3ms backward=1003.4ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:08,521] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:09,022] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:56:15,298] loop.py:45 INFO TRAIN_STEP step=17997 loss=1.187522 Training: 100%|█████████▉| 17997/18018 [38:26:01<02:28, 7.07s/step, loss=1.188, acc=0.715, acc_len=1.73, thru=17.6, I=18.4, T=37.9, wait=3.4s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:21,336] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:56:22,596] loop.py:45 INFO TRAIN_STEP step=17998 loss=1.279693 Training: 100%|█████████▉| 17998/18018 [38:26:08<02:22, 7.14s/step, loss=1.280, acc=0.702, acc_len=1.69, thru=17.8, I=20.1, T=36.8, wait=3.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:22,846] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:56:29,597] loop.py:45 INFO TRAIN_STEP step=17999 loss=1.166209 Training: 100%|█████████▉| 17999/18018 [38:26:15<02:14, 7.10s/step, loss=1.166, acc=0.723, acc_len=1.76, thru=15.1, I=18.7, T=39.3, wait=3.7s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:36,272] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:56:36,809] loop.py:380 INFO TIMING step=18000: step=3.626s data=1.386s compute=3.600s [fwd=2.594s bwd=0.992s opt=0.003s] dispatch=3.576s [2026-10-20 03:56:36,809] loop.py:45 INFO TRAIN_STEP step=18000 loss=1.242472 Training: 100%|█████████▉| 18000/18018 [38:26:22<02:08, 7.13s/step, loss=1.242, acc=0.711, acc_len=1.72, thru=19.9, I=18.1, T=35.3, wait=3.6s, pool=65, epoch=2/2][2026-10-20 03:56:36,812] loop.py:419 INFO Saving checkpoint at step 18000... (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 03:56:40,197] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0018001 (TrainerActor pid=9310) [2026-10-20 03:56:40,540] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0017251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:41,316] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=33/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:41,316] inference_manager.py:408 INFO Pool capacity available after 5.0s, pool_size=33/64 [2026-10-20 03:56:47,106] loop.py:45 INFO TRAIN_STEP step=18001 loss=1.163974 Training: 100%|█████████▉| 18001/18018 [38:26:33<02:17, 8.08s/step, loss=1.164, acc=0.725, acc_len=1.76, thru=19.4, I=23.8, T=38.9, wait=2.6s, pool=49, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:56:47,101] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=18000: forward=2346.1ms backward=903.2ms (TrainerActor pid=9310) [2026-10-20 03:56:40,196] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0018001 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 03:56:41,217] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0017251 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:54,621] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:56:54,998] loop.py:45 INFO TRAIN_STEP step=18002 loss=1.270766 Training: 100%|█████████▉| 18002/18018 [38:26:41<02:08, 8.03s/step, loss=1.271, acc=0.704, acc_len=1.69, thru=17.0, I=17.4, T=36.2, wait=4.3s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:56:55,125] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:01,173] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:57:01,487] loop.py:45 INFO TRAIN_STEP step=18003 loss=1.154040 Training: 100%|█████████▉| 18003/18018 [38:26:47<01:53, 7.56s/step, loss=1.154, acc=0.725, acc_len=1.76, thru=20.1, I=19.4, T=36.3, wait=3.0s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:01,674] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:06,580] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:57:07,744] loop.py:45 INFO TRAIN_STEP step=18004 loss=1.227055 Training: 100%|█████████▉| 18004/18018 [38:26:53<01:40, 7.17s/step, loss=1.227, acc=0.711, acc_len=1.72, thru=19.9, I=23.2, T=37.5, wait=2.8s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:08,088] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 [2026-10-20 03:57:13,774] loop.py:380 INFO TIMING step=18005: step=3.417s data=1.418s compute=3.389s [fwd=2.432s bwd=0.942s opt=0.003s] dispatch=2.605s [2026-10-20 03:57:13,775] loop.py:45 INFO TRAIN_STEP step=18005 loss=1.121674 Training: 100%|█████████▉| 18005/18018 [38:26:59<01:28, 6.83s/step, loss=1.122, acc=0.730, acc_len=1.78, thru=18.3, I=20.6, T=37.5, wait=2.6s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:22,125] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:57:23,214] loop.py:45 INFO TRAIN_STEP step=18006 loss=1.302519 Training: 100%|█████████▉| 18006/18018 [38:27:09<01:31, 7.61s/step, loss=1.303, acc=0.698, acc_len=1.67, thru=16.9, I=16.8, T=36.5, wait=5.9s, pool=65, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:57:23,210] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=18005: forward=2514.4ms backward=955.9ms (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:23,637] inference_manager.py:408 INFO Pool capacity available after 1.5s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:29,025] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:57:29,420] loop.py:45 INFO TRAIN_STEP step=18007 loss=1.349970 Training: 100%|█████████▉| 18007/18018 [38:27:15<01:19, 7.19s/step, loss=1.350, acc=0.689, acc_len=1.64, thru=20.4, I=24.5, T=38.9, wait=2.9s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:29,540] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=17/64 [2026-10-20 03:57:36,070] loop.py:45 INFO TRAIN_STEP step=18008 loss=1.315907 Training: 100%|█████████▉| 18008/18018 [38:27:22<01:10, 7.03s/step, loss=1.316, acc=0.696, acc_len=1.66, thru=16.8, I=19.9, T=36.6, wait=3.1s, pool=37, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:42,583] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:57:43,331] loop.py:45 INFO TRAIN_STEP step=18009 loss=1.247862 Training: 100%|█████████▉| 18009/18018 [38:27:29<01:03, 7.10s/step, loss=1.248, acc=0.711, acc_len=1.71, thru=19.3, I=18.7, T=34.5, wait=3.5s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:43,589] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:49,716] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 [2026-10-20 03:57:50,080] loop.py:380 INFO TIMING step=18010: step=3.598s data=1.509s compute=3.560s [fwd=2.546s bwd=0.999s opt=0.003s] dispatch=3.142s [2026-10-20 03:57:50,080] loop.py:45 INFO TRAIN_STEP step=18010 loss=1.168660 Training: 100%|█████████▉| 18010/18018 [38:27:36<00:55, 6.99s/step, loss=1.169, acc=0.724, acc_len=1.77, thru=18.8, I=22.1, T=35.6, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:57:50,218] inference_manager.py:408 INFO Pool capacity available after 0.5s, pool_size=1/64 [2026-10-20 03:57:56,518] loop.py:45 INFO TRAIN_STEP step=18011 loss=1.259386 Training: 100%|█████████▉| 18011/18018 [38:27:42<00:47, 6.83s/step, loss=1.259, acc=0.708, acc_len=1.70, thru=18.5, I=20.6, T=35.2, wait=2.8s, pool=57, epoch=2/2](TrainerActor pid=9310) [2026-10-20 03:57:56,513] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=18010: forward=2587.9ms backward=1002.7ms [2026-10-20 03:58:03,130] loop.py:45 INFO TRAIN_STEP step=18012 loss=1.279141 Training: 100%|█████████▉| 18012/18018 [38:27:49<00:40, 6.76s/step, loss=1.279, acc=0.703, acc_len=1.68, thru=17.7, I=18.4, T=37.3, wait=3.2s, pool=49, epoch=2/2][2026-10-20 03:58:10,745] loop.py:45 INFO TRAIN_STEP step=18013 loss=1.290594 Training: 100%|█████████▉| 18013/18018 [38:27:56<00:35, 7.02s/step, loss=1.291, acc=0.705, acc_len=1.70, thru=16.7, I=18.6, T=35.1, wait=4.0s, pool=61, epoch=2/2][2026-10-20 03:58:17,751] loop.py:45 INFO TRAIN_STEP step=18014 loss=1.125798 Training: 100%|█████████▉| 18014/18018 [38:28:03<00:28, 7.01s/step, loss=1.126, acc=0.732, acc_len=1.80, thru=13.8, I=19.7, T=38.7, wait=3.7s, pool=25, epoch=2/2][2026-10-20 03:58:26,531] loop.py:380 INFO TIMING step=18015: step=3.627s data=1.478s compute=3.602s [fwd=2.608s bwd=0.980s opt=0.003s] dispatch=5.144s [2026-10-20 03:58:26,532] loop.py:45 INFO TRAIN_STEP step=18015 loss=1.275239 Training: 100%|█████████▉| 18015/18018 [38:28:12<00:22, 7.54s/step, loss=1.275, acc=0.705, acc_len=1.70, thru=16.4, I=14.8, T=35.3, wait=5.1s, pool=57, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:58:32,403] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=9310) [2026-10-20 03:58:33,201] dflash_trainer.py:668 INFO COMPUTE_BREAKDOWN step=18015: forward=2525.5ms backward=1031.4ms [2026-10-20 03:58:33,206] loop.py:45 INFO TRAIN_STEP step=18016 loss=1.277532 Training: 100%|█████████▉| 18016/18018 [38:28:19<00:14, 7.28s/step, loss=1.278, acc=0.705, acc_len=1.70, thru=18.7, I=20.2, T=35.5, wait=3.1s, pool=65, epoch=2/2](AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:58:33,412] inference_manager.py:408 INFO Pool capacity available after 1.0s, pool_size=1/64 [2026-10-20 03:58:40,760] loop.py:45 INFO TRAIN_STEP step=18017 loss=1.205326 Training: 100%|█████████▉| 18017/18018 [38:28:26<00:07, 7.36s/step, loss=1.205, acc=0.717, acc_len=1.74, thru=16.1, I=17.2, T=34.6, wait=3.9s, pool=49, epoch=2/2][2026-10-20 03:58:49,396] loop.py:45 INFO TRAIN_STEP step=18018 loss=1.236271 Training: 100%|██████████| 18018/18018 [38:28:35<00:00, 7.75s/step, loss=1.236, acc=0.707, acc_len=1.71, thru=16.3, I=15.1, T=37.2, wait=5.2s, pool=61, epoch=2/2][2026-10-20 03:58:49,399] loop.py:433 INFO Epoch 2 completed (9009 steps). Total steps: 18018/18018 [2026-10-20 03:58:49,399] loop.py:444 INFO Saving checkpoint at end of epoch 2 (step 18018)... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:58:49,593] inference_manager.py:387 WARNING Sample pool full, pausing generation: pool_size=65/64 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 03:58:53,441] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0018019 (TrainerActor pid=9310) [2026-10-20 03:58:53,660] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0017501 [2026-10-20 03:58:54,533] loop.py:463 INFO Max steps reached Training: 100%|██████████| 18018/18018 [38:28:40<00:00, 7.69s/step, loss=1.236, acc=0.707, acc_len=1.71, thru=16.3, I=15.1, T=37.2, wait=5.2s, pool=61, epoch=2/2] [2026-10-20 03:58:54,539] loop.py:500 INFO Training completed: 18018 steps in 138934.6s | avg inference=16.6 entries/s | avg training=16.6 entries/s (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:58:54,632] inference_manager.py:398 INFO Still waiting for pool capacity (waited 5.0s): pool_size=65/64 (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:58:54,632] inference_manager.py:406 INFO Pool wait aborted (shutdown requested) after 5.0s (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:58:54,641] inference_manager.py:326 INFO Drain: discarding 8 buffered prompts on shutdown [2026-10-20 03:58:55,297] loop.py:146 INFO Shutting down 4 inference engine(s)... (AsyncInferenceManager pid=3029, ip=10.0.3.60) [2026-10-20 03:58:55,296] inference_manager.py:219 INFO AsyncInferenceManager stopped (VllmEngine pid=1867, ip=10.0.3.60) [2026-10-20 03:59:01,053] vllm_engine.py:725 INFO VllmEngine rank 2: shutdown complete (TrainerActor pid=9310) [2026-10-20 03:58:53,440] checkpoint.py:541 INFO Saved checkpoint to /data/output/ckpts/checkpoints/iter_0018019 (TrainerActor pid=1751, ip=10.0.3.62) [2026-10-20 03:58:54,530] checkpoint.py:179 INFO Pruned 1 old checkpoint(s) on this node, keeping 3: iter_0017501 (TrainerActor pid=9310) [2026-10-20 03:59:02,085] data_fetcher.py:325 WARNING __iter__: Exception waiting for data: The actor died unexpectedly before finishing this task. (TrainerActor pid=9310) class_name: _QueueActor (TrainerActor pid=9310) actor_id: 6a55a92ecc5df2e57894ba6702000000 (TrainerActor pid=9310) pid: 8911 (TrainerActor pid=9310) namespace: c893011b-0e3f-44b4-9783-5621b93cd7b4 (TrainerActor pid=9310) ip: 10.0.3.61 (TrainerActor pid=9310) The actor is dead because its owner has died. Owner Id: b4d322228e05a3cd8f123754201b27c890d1b4d50932a68624f0a507 Owner Ip address: 10.0.3.61 Owner worker exit type: INTENDED_SYSTEM_EXIT Worker exit detail: Owner's worker process has crashed., timeout=None (VllmEngine pid=1809, ip=10.0.3.59) [2026-10-20 03:59:01,141] vllm_engine.py:725 INFO VllmEngine rank 1: shutdown complete [repeated 3x across cluster] (TrainerActor pid=1754, ip=10.0.3.62) [2026-10-20 03:59:02,085] data_fetcher.py:325 WARNING __iter__: Exception waiting for data: The actor died unexpectedly before finishing this task. [repeated 7x across cluster] (TrainerActor pid=1754, ip=10.0.3.62) class_name: _QueueActor [repeated 7x across cluster] (TrainerActor pid=1754, ip=10.0.3.62) actor_id: 2b602930b590f4dd5f47129902000000 [repeated 7x across cluster] (TrainerActor pid=1754, ip=10.0.3.62) pid: 1027 [repeated 7x across cluster] (TrainerActor pid=1754, ip=10.0.3.62) namespace: c893011b-0e3f-44b4-9783-5621b93cd7b4 [repeated 7x across cluster] (TrainerActor pid=1754, ip=10.0.3.62) ip: 10.0.3.59 [repeated 7x across cluster] (TrainerActor pid=1754, ip=10.0.3.62) The actor is dead because its owner has died. Owner Id: b4d322228e05a3cd8f123754201b27c890d1b4d50932a68624f0a507 Owner Ip address: 10.0.3.61 Owner worker exit type: INTENDED_SYSTEM_EXIT Worker exit detail: Owner's worker process has crashed., timeout=None [repeated 7x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:30:16 [jit_monitor.py:141] CuTeDSL JIT compilation during inference: FlashAttentionForwardSm100. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1751, ip=10.0.3.63) (EngineCore pid=1946) WARNING 10-18 13:30:04 [vllm.py:1388] Enforce eager set, disabling torch.compile and CUDAGraphs. This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none (VllmEngine pid=1751, ip=10.0.3.63) (EngineCore pid=1946) WARNING 10-18 13:30:04 [vllm.py:1423] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (VllmEngine pid=1751, ip=10.0.3.63) (EngineCore pid=1946) INFO 10-18 13:30:04 [kernel.py:310] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['vllm_c', 'native'], fused_add_rms_norm=['vllm_c', 'native']) (VllmEngine pid=1751, ip=10.0.3.63) (EngineCore pid=1946) INFO 10-18 13:30:04 [vllm.py:1602] Cudagraph is disabled under eager mode (VllmEngine pid=1751, ip=10.0.3.63) (EngineCore pid=1946) INFO 10-18 13:30:04 [compilation.py:329] Enabled custom fusions: norm_quant, act_quant, allreduce_rms (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:30:05 [base.py:235] Multi-modal warmup completed in 0.752s (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-18 13:30:06 [base.py:235] Readonly multi-modal warmup completed in 0.754s (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) WARNING 10-18 13:30:16 [jit_monitor.py:141] Triton kernel JIT compilation during inference: _convert_req_index_to_global_index_kernel. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) WARNING 10-18 13:30:16 [flashmla_sparse.py:850] Padding num_heads from 16 to 64 for FP8 sparse decode kernel (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) INFO 10-18 13:30:16 [gpu_model_runner.py:4736] extract_hidden_states prefill-only path: skipping logits, sampling, token proposal, and decode (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) WARNING 10-18 13:30:16 [jit_monitor.py:141] CuTeDSL JIT compilation during inference: FlashAttentionForwardSm100. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:30:25 [jit_monitor.py:141] Triton kernel JIT compilation during inference: _convert_req_index_to_global_index_kernel. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:30:25 [flashmla_sparse.py:850] Padding num_heads from 16 to 64 for FP8 sparse decode kernel (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-18 13:30:19 [gpu_model_runner.py:4736] extract_hidden_states prefill-only path: skipping logits, sampling, token proposal, and decode (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) WARNING 10-18 13:30:27 [jit_monitor.py:141] Triton kernel JIT compilation during inference: _convert_req_index_to_global_index_kernel. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) WARNING 10-18 13:30:27 [flashmla_sparse.py:850] Padding num_heads from 16 to 64 for FP8 sparse decode kernel (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) INFO 10-18 13:30:27 [gpu_model_runner.py:4736] extract_hidden_states prefill-only path: skipping logits, sampling, token proposal, and decode (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:30:27 [jit_monitor.py:141] CuTeDSL JIT compilation during inference: FlashAttentionForwardSm100. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:30:27 [jit_monitor.py:141] Triton kernel JIT compilation during inference: _convert_req_index_to_global_index_kernel. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:30:27 [flashmla_sparse.py:850] Padding num_heads from 16 to 64 for FP8 sparse decode kernel (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) INFO 10-18 13:30:27 [gpu_model_runner.py:4736] extract_hidden_states prefill-only path: skipping logits, sampling, token proposal, and decode (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) WARNING 10-18 13:31:19 [jit_monitor.py:141] CuTeDSL JIT compilation during inference: FlashAttentionForwardSm100. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1751, ip=10.0.3.63) (Worker_TP0 pid=1961) WARNING 10-18 13:31:22 [jit_monitor.py:141] Triton kernel JIT compilation during inference: BuildPrefillChunkMetadataKernel.kernel. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) WARNING 10-18 13:32:20 [jit_monitor.py:141] Triton kernel JIT compilation during inference: BuildPrefillChunkMetadataKernel.kernel. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1867, ip=10.0.3.60) (Worker_TP0 pid=2077) WARNING 10-18 13:32:46 [jit_monitor.py:141] Triton kernel JIT compilation during inference: BuildPrefillChunkMetadataKernel.kernel. This causes a latency spike; consider extending warmup to cover this shape/config. (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) WARNING 10-18 13:33:09 [jit_monitor.py:141] Triton kernel JIT compilation during inference: BuildPrefillChunkMetadataKernel.kernel. This causes a latency spike; consider extending warmup to cover this shape/config. (TrainerActor pid=9310) ET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 06/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 07/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 08/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 09/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 10/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 11/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 12/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 13/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 14/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 15/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 16/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 17/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 18/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 19/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 20/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 21/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 22/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 23/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 24/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 25/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 26/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 27/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 28/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 29/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 30/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 31/0 : 0[0] -> 4[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Connected NVLS tree (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO 4 coll channels, 4 collnet channels, 32 nvls channels, 4 p2p channels, 4 p2p channels per peer (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Symmetric memory is not supported. cuMemEnable 0, globalGinSupport 1, globalNicFused 0 cuMemGdrSupport 1 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO CC Off, workFifoBytes 1048576 (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO ncclCommInitRankConfig comm 0x4af1baa0 rank 0 nranks 8 cudaDev 0 nvmlDev 0 busId 806000 commId 0x5b058ae8201a6de4 - Init COMPLETE (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Init timings - ncclCommInitRankConfig: rank 0 nranks 8 total 20.89 (kernels 18.49, alloc 0.37, bootstrap 0.09, allgathers 0.10, topo 0.46, graphs 0.00, connections 1.37, rest 0.00) (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 00/1 : 7[3] -> 0[0] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 01/1 : 7[3] -> 0[0] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 02/1 : 7[3] -> 0[0] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 03/1 : 7[3] -> 0[0] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 00/1 : 6[2] -> 0[0] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 01/1 : 6[2] -> 0[0] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel (VllmEngine pid=1635, ip=10.0.3.58) INFO 10-20 03:58:55 [core_client.py:689] [shutdown] MPClient: start timeout=default (VllmEngine pid=1635, ip=10.0.3.58) INFO 10-20 03:58:55 [core_client.py:691] [shutdown] MPClient: stopping engine manager (VllmEngine pid=1635, ip=10.0.3.58) INFO 10-20 03:58:55 [utils.py:615] [shutdown] Process manager: send sigterm to process EngineCore (VllmEngine pid=1635, ip=10.0.3.58) (EngineCore pid=1830) INFO 10-20 03:58:55 [core.py:1344] [shutdown] EngineCore: trigger received signal=SIGTERM (VllmEngine pid=1635, ip=10.0.3.58) (EngineCore pid=1830) INFO 10-20 03:58:55 [core.py:1495] [shutdown] EngineCore: start mode=abort timeout=0s (VllmEngine pid=1635, ip=10.0.3.58) (EngineCore pid=1830) INFO 10-20 03:58:55 [core.py:1526] [shutdown] EngineCore: request processing complete; starting resource teardown (VllmEngine pid=1635, ip=10.0.3.58) (EngineCore pid=1830) INFO 10-20 03:58:55 [core.py:1357] [shutdown] EngineCore: exiting busy loop (VllmEngine pid=1635, ip=10.0.3.58) (Worker_TP0 pid=1845) INFO 10-20 03:58:55 [multiproc_executor.py:836] Parent process exited, terminating worker queues (VllmEngine pid=1635, ip=10.0.3.58) (EngineCore pid=1830) INFO 10-20 03:58:55 [multiproc_executor.py:472] [shutdown] Executor: waiting for worker exit count=4 (VllmEngine pid=1809, ip=10.0.3.59) CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 14/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 15/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 16/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 17/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 18/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 19/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 20/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 21/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 22/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 23/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 24/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 25/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 26/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 27/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 28/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 29/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 30/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Channel 31/0 : 2[2] -> 3[3] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO comm 0x38363320 rank 2 nRanks 4 cudaDev 2 busId 1806000 - Abort START (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:28 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:45 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:584 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:28 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:45 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:584 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2244 [2] NCCL INFO misc/socket.cc:659 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:28 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:45 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:584 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:28 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:45 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO misc/socket.cc:584 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2244 [2] NCCL INFO misc/socket.cc:659 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387839 [2] NCCL INFO comm 0x38363320 rank 2 nranks 4 cudaDev 2 busId 1806000 - Abort COMPLETE (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO comm 0x41554960 rank 2 nRanks 4 cudaDev 2 busId 1806000 - Abort START (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:28 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:45 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:584 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:28 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:45 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:584 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:28 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:45 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:584 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2361 [2] NCCL INFO misc/socket.cc:659 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:28 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:45 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO misc/socket.cc:584 -> 3 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:387843 [2] NCCL INFO comm 0x41554960 rank 2 nranks 4 cudaDev 2 busId 1806000 - Abort COMPLETE (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO comm 0x4e632390 rank 2 nranks 4 cudaDev 2 busId 1806000 - Destroy COMPLETE (VllmEngine pid=1867, ip=10.0.3.60) 2P/CUMEM (VllmEngine pid=1751, ip=10.0.3.63) EM (VllmEngine pid=1809, ip=10.0.3.59) 59:2019:2019 [0] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2489 [0] NCCL INFO [Proxy Service] Device 0 CPU core 29 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2492 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 32 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO threadThresholds 8/8/64 | 32/8/64 | 512 | 512 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO 32 coll channels, 32 collnet channels, 24 nvls channels, 32 p2p channels, 32 p2p channels per peer (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO CC Off, workFifoBytes 1048576 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO ncclCommInitRankConfig comm 0x5d36e700 rank 0 nranks 4 cudaDev 0 nvmlDev 0 busId 806000 commId 0xd159f757db5ec71a - Init COMPLETE (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO Init timings - ncclCommInitRankConfig: rank 0 nranks 4 total 3.68 (kernels 0.00, alloc 0.00, bootstrap 2.66, allgathers 0.00, topo 0.49, graphs 0.01, connections 0.52, rest 0.00) (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2019:2019 [0] NCCL INFO Symmetric VA size=280GB (VllmEngine pid=1809, ip=10.0.3.59) host-10-0 (VllmEngine pid=1751, ip=10.0.3.63) host-10-0 (VllmEngine pid=1635, ip=10.0.3.58) host-10-0 (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2021:2021 [2] NCCL INFO ENV/Plugin: Closing env plugin ncclEnvDefault (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3- (VllmEngine pid=1751, ip=10.0.3.63) CCL INFO Channel 13/0 : 3[3] -> 0[0] via P2P/CUMEM (VllmEngine pid=1809, ip=10.0.3.59) (EngineCore pid=2004) WARNING 10-20 03:59:00 [multiproc_executor.py:484] [shutdown] Executor: workers still running after grace period; sending SIGTERM count=3 (VllmEngine pid=1809, ip=10.0.3.59) WARNING 10-20 03:59:00 [utils.py:635] [shutdown] Process manager: force killing remaining processes count=1 (VllmEngine pid=1809, ip=10.0.3.59) WARNING 10-20 03:59:00 [utils.py:640] [shutdown] Process manager: force killing remaining process EngineCore pid 2004 (VllmEngine pid=1809, ip=10.0.3.59) INFO 10-20 03:59:00 [core_client.py:693] [shutdown] MPClient: engine manager stopped (VllmEngine pid=1809, ip=10.0.3.59) INFO 10-20 03:59:00 [core_client.py:694] [shutdown] MPClient: cleaning up background resources (VllmEngine pid=1809, ip=10.0.3.59) INFO 10-20 03:59:00 [core_client.py:696] [shutdown] MPClient: complete (VllmEngine pid=1809, ip=10.0.3.59) INFO 10-20 03:58:55 [core_client.py:689] [shutdown] MPClient: start timeout=default [repeated 3x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) INFO 10-20 03:58:55 [core_client.py:691] [shutdown] MPClient: stopping engine manager [repeated 3x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) INFO 10-20 03:58:55 [utils.py:615] [shutdown] Process manager: send sigterm to process EngineCore [repeated 3x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (EngineCore pid=2004) INFO 10-20 03:58:55 [core.py:1344] [shutdown] EngineCore: trigger received signal=SIGTERM [repeated 3x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (EngineCore pid=2004) INFO 10-20 03:58:55 [core.py:1495] [shutdown] EngineCore: start mode=abort timeout=0s [repeated 3x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (EngineCore pid=2004) INFO 10-20 03:58:55 [core.py:1526] [shutdown] EngineCore: request processing complete; starting resource teardown [repeated 3x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (EngineCore pid=2004) INFO 10-20 03:58:55 [core.py:1357] [shutdown] EngineCore: exiting busy loop [repeated 3x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (Worker_TP0 pid=2019) INFO 10-20 03:58:55 [multiproc_executor.py:836] Parent process exited, terminating worker queues [repeated 3x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) (EngineCore pid=2004) INFO 10-20 03:58:55 [multiproc_executor.py:472] [shutdown] Executor: waiting for worker exit count=4 [repeated 3x across cluster] (TrainerActor pid=9371) els, 32 nvls channels, 4 p2p channels, 4 p2p channels per peer (TrainerActor pid=9371) host-10-0-3-61:9371:9371 [2] NCCL INFO Symmetric memory is not supported. cuMemEnable 0, globalGinSupport 1, globalNicFused 0 cuMemGdrSupport 1 (TrainerActor pid=9370) mlDev 3 busId 1906000 commId 0x5b058ae8201a6de4 - Init COMPLETE (TrainerActor pid=1754, ip=10.0.3.62) er (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:1754 [2] NCCL INFO Channel 00/1 : 6[2] -> 0[0] [send] via NET/IB/0(4)/GDRDMA/Shared (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:1754 [2] NCCL INFO Channel 01/1 : 6[2] -> 0[0] [send] via NET/IB/0(4)/GDRDMA/Shared (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:1754 [2] NCCL INFO Channel 02/1 : 6[2] -> 0[0] [send] via NET/IB/0(4)/GDRDMA/Shared (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:1754 [2] NCCL INFO Channel 03/1 : 6[2] -> 0[0] [send] via NET/IB/0(4)/GDRDMA/Shared (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:1754 [2] NCCL INFO Channel 00/1 : 0[0] -> 6[2] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:1754 [2] NCCL INFO Channel 01/1 : 0[0] -> 6[2] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:1754 [2] NCCL INFO Channel 02/1 : 0[0] -> 6[2] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:1754 [2] NCCL INFO Channel 03/1 : 0[0] -> 6[2] [receive] via NET/IB/0/GDRDMA/Shared (TrainerActor pid=1754, ip=10.0.3.62) host-10-0-3-62:1754:2098 [2] NCCL INFO NCCL_IB_GID_INDEX set by environment to 3. (TrainerActor pid=1753, ip=10.0.3.62) ETE (TrainerActor pid=1752, ip=10.0.3.62) -1 nHeads 1 nvlsRanks 4 buffSize 1048576 nvlsPerRankSize 67108864 nvlsTotalSize 67108864 (TrainerActor pid=1752, ip=10.0.3.62) host-10-0-3-62:1752:1752 [1] NCCL INFO Connected NVLS tree (TrainerActor pid=1752, ip=10.0.3.62) host-10-0-3-62:1752:1752 [1] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so (TrainerActor pid=1751, ip=10.0.3.62) 09/0 : 4[0] -> 0[0] [send] via NET/IB/0/GDRDMA (TrainerActor pid=1751, ip=10.0.3.62) host-10-0-3-62:1751:2103 [0] NCCL INFO [Service thread] Connection closed by localRank 3 (TrainerActor pid=1751, ip=10.0.3.62) host-10-0-3-62:1751:2103 [0] NCCL INFO [Service thread] Connection closed by localRank 2 (TrainerActor pid=9310) 02/1 : 6[2] -> 0[0] [receive] via NET/IB/0/GDRDMA/Shared (VllmEngine pid=1635, ip=10.0.3.58) CUMEM (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 03/1 : 0[0] -> 3[3] via P2P/IPC [repeated 275x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO Connected all rings, use ring PXN 0 GDR 1 [repeated 11x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:387844 [1] NCCL INFO comm 0x23f6b330 rank 1 nRanks 4 cudaDev 1 busId 906000 - Abort START [repeated 18x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:387844 [1] NCCL INFO misc/socket.cc:584 -> 3 [repeated 182x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:387844 [1] NCCL INFO comm 0x23f6b330 rank 1 nranks 4 cudaDev 1 busId 906000 - Abort COMPLETE [repeated 14x across cluster] (VllmEngine pid=1809, ip=10.0.3.59) host-10-0-3-59:2020:2020 [1] NCCL INFO comm 0x30a5ccd0 rank 1 nranks 4 cudaDev 1 busId 906000 - Destroy COMPLETE [repeated 7x across cluster] (TrainerActor pid=9368) 3-61:9368:9368 [1] NCCL INFO NVLS comm 0x3ce20270 headRank -1 nHeads 1 nvlsRanks 4 buffSize 1048576 nvlsPerRankSize 67108864 nvlsTotalSize 67108864 (VllmEngine pid=1867, ip=10.0.3.60) 0:2077:2077 [0] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1 [repeated 3x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2523 [0] NCCL INFO [Proxy Service] Device 0 CPU core 2 [repeated 3x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2526 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 5 [repeated 3x across cluster] (TrainerActor pid=9368) host-10-0-3-61:9368:9368 [1] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512 [repeated 6x across cluster] (TrainerActor pid=9368) host-10-0-3-61:9368:9368 [1] NCCL INFO 4 coll channels, 4 collnet channels, 32 nvls channels, 4 p2p channels, 4 p2p channels per peer [repeated 6x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO Symmetric VA size=280GB [repeated 6x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2077:2077 [0] NCCL INFO CC Off, workFifoBytes 1048576 [repeated 3x across cluster] (TrainerActor pid=9368) host-10-0-3-61:9368:9368 [1] NCCL INFO ncclCommInitRankConfig comm 0x3ce20270 rank 1 nranks 8 cudaDev 1 nvmlDev 1 busId 906000 commId 0x5b058ae8201a6de4 - Init COMPLETE [repeated 8x across cluster] (TrainerActor pid=9368) host-10-0-3-61:9368:9368 [1] NCCL INFO Init timings - ncclCommInitRankConfig: rank 1 nranks 8 total 20.89 (kernels 18.88, alloc 0.07, bootstrap 0.00, allgathers 0.10, topo 0.46, graphs 0.00, connections 0.63, rest 0.74) [repeated 10x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) host-10-0-3-60:2079:2079 [2] NCCL INFO ENV/Plugin: Closing env plugin ncclEnvDefault [repeated 2x across cluster] (VllmEngine pid=1867, ip=10.0.3.60) (EngineCore pid=2062) WARNING 10-20 03:59:00 [multiproc_executor.py:484] [shutdown] Executor: workers still running after grace period; sending SIGTERM count=4 [repeated 2x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-20 03:59:00 [utils.py:635] [shutdown] Process manager: force killing remaining processes count=1 [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) WARNING 10-20 03:59:00 [utils.py:640] [shutdown] Process manager: force killing remaining process EngineCore pid 1946 [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-20 03:59:00 [core_client.py:693] [shutdown] MPClient: engine manager stopped [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-20 03:59:00 [core_client.py:694] [shutdown] MPClient: cleaning up background resources [repeated 3x across cluster] (VllmEngine pid=1751, ip=10.0.3.63) INFO 10-20 03:59:00 [core_client.py:696] [shutdown] MPClient: complete [repeated 3x across cluster] (TrainerActor pid=9368) host-10-0-3-61:9368:9368 [1] NCCL INFO Symmetric memory is not supported. cuMemEnable 0, globalGinSupport 1, globalNicFused 0 cuMemGdrSupport 1 [repeated 4x across cluster] (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 03/1 : 0[0] -> 7[3] [send] via NET/IB/0/GDRDMA/Shared [repeated 50x across cluster] (TrainerActor pid=9310) host-10-0-3-61:9310:9310 [0] NCCL INFO Channel 03/1 : 4[0] -> 0[0] [receive] via NET/IB/0/GDRDMA/Shared [repeated 21x across cluster] (TrainerActor pid=1753, ip=10.0.3.62) host-10-0-3-62:1753:2099 [3] NCCL INFO NCCL_IB_GID_INDEX set by environment to 3. (TrainerActor pid=9368) host-10-0-3-61:9368:9368 [1] NCCL INFO Connected NVLS tree [repeated 2x across cluster] (TrainerActor pid=9368) host-10-0-3-61:9368:9368 [1] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so [repeated 2x across cluster] (TrainerActor pid=9310) host-10-0-3-61:9310:9577 [0] NCCL INFO [Service thread] Connection closed by localRank 3 [repeated 2x across cluster] (TrainerActor pid=9368) host-10-0-3-61:9368:9368 [1] NCCL INFO Channel 03/1 : 1[1] -> 0[0] via P2P/IPC [repeated 4x across cluster]